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" |
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(numBits: ShAmt0->getType()->getScalarSizeInBits()); |
44 | return MaximalRepresentableShiftAmount.uge(RHS: 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 ) { |
61 | // Look for a shift of some instruction, ignore zext of shift amount if any. |
62 | Instruction *Sh0Op0; |
63 | Value *ShAmt0; |
64 | if (!match(V: Sh0, |
65 | P: m_Shift(L: m_Instruction(I&: Sh0Op0), R: m_ZExtOrSelf(Op: m_Value(V&: 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(V: Sh0Op0, |
74 | P: m_CombineOr(L: m_CombineAnd(L: m_Trunc(Op: m_Instruction(I&: Sh1)), R: m_Value(V&: Trunc)), |
75 | R: m_Instruction(I&: 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(V: Sh1, P: m_Shift(L: m_Value(V&: X), R: m_ZExtOrSelf(Op: m_Value(V&: 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(V: Sh0, P: m_Shr(L: m_Value(), R: m_Value())) && |
89 | match(V: Sh1, P: m_Shr(L: m_Value(), R: 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(V: Sh0, P: m_c_BinOp(L: m_OneUse(SubPattern: m_Value()), R: m_Value()))) |
106 | return nullptr; |
107 | |
108 | // Can we fold (ShAmt0+ShAmt1) ? |
109 | auto *NewShAmt = dyn_cast_or_null<Constant>( |
110 | Val: simplifyAddInst(LHS: ShAmt0, RHS: ShAmt1, /*isNSW=*/IsNSW: false, /*isNUW=*/IsNUW: false, |
111 | Q: SQ.getWithInstruction(I: 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(V: NewShAmt, P: m_SpecificInt_ICMP(Predicate: ICmpInst::Predicate::ICMP_ULT, |
118 | Threshold: 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(V: NewShAmt, |
129 | P: m_SpecificInt_ICMP(Predicate: ICmpInst::Predicate::ICMP_EQ, |
130 | Threshold: 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." ); |
138 | |
139 | if (NewShAmt->getType() != X->getType()) { |
140 | NewShAmt = ConstantFoldCastOperand(Opcode: Instruction::ZExt, C: NewShAmt, |
141 | DestTy: X->getType(), DL: SQ.DL); |
142 | if (!NewShAmt) |
143 | return nullptr; |
144 | } |
145 | |
146 | // All good, we can do this fold. |
147 | BinaryOperator *NewShift = BinaryOperator::Create(Op: ShiftOpcode, S1: X, S2: NewShAmt); |
148 | |
149 | // The flags can only be propagated if there wasn't a trunc. |
150 | if (!Trunc) { |
151 | // If the pattern did not involve trunc, and both of the original shifts |
152 | // had the same flag set, preserve the flag. |
153 | if (ShiftOpcode == Instruction::BinaryOps::Shl) { |
154 | NewShift->setHasNoUnsignedWrap(Sh0->hasNoUnsignedWrap() && |
155 | Sh1->hasNoUnsignedWrap()); |
156 | NewShift->setHasNoSignedWrap(Sh0->hasNoSignedWrap() && |
157 | Sh1->hasNoSignedWrap()); |
158 | } else { |
159 | NewShift->setIsExact(Sh0->isExact() && Sh1->isExact()); |
160 | } |
161 | } |
162 | |
163 | Instruction *Ret = NewShift; |
164 | if (Trunc) { |
165 | Builder.Insert(I: NewShift); |
166 | Ret = CastInst::Create(Instruction::Trunc, S: NewShift, Ty: Sh0->getType()); |
167 | } |
168 | |
169 | return Ret; |
170 | } |
171 | |
172 | // If we have some pattern that leaves only some low bits set, and then performs |
173 | // left-shift of those bits, if none of the bits that are left after the final |
174 | // shift are modified by the mask, we can omit the mask. |
175 | // |
176 | // There are many variants to this pattern: |
177 | // a) (x & ((1 << MaskShAmt) - 1)) << ShiftShAmt |
178 | // b) (x & (~(-1 << MaskShAmt))) << ShiftShAmt |
179 | // c) (x & (-1 l>> MaskShAmt)) << ShiftShAmt |
180 | // d) (x & ((-1 << MaskShAmt) l>> MaskShAmt)) << ShiftShAmt |
181 | // e) ((x << MaskShAmt) l>> MaskShAmt) << ShiftShAmt |
182 | // f) ((x << MaskShAmt) a>> MaskShAmt) << ShiftShAmt |
183 | // All these patterns can be simplified to just: |
184 | // x << ShiftShAmt |
185 | // iff: |
186 | // a,b) (MaskShAmt+ShiftShAmt) u>= bitwidth(x) |
187 | // c,d,e,f) (ShiftShAmt-MaskShAmt) s>= 0 (i.e. ShiftShAmt u>= MaskShAmt) |
188 | static Instruction * |
189 | dropRedundantMaskingOfLeftShiftInput(BinaryOperator *OuterShift, |
190 | const SimplifyQuery &Q, |
191 | InstCombiner::BuilderTy &Builder) { |
192 | assert(OuterShift->getOpcode() == Instruction::BinaryOps::Shl && |
193 | "The input must be 'shl'!" ); |
194 | |
195 | Value *Masked, *ShiftShAmt; |
196 | match(V: OuterShift, |
197 | P: m_Shift(L: m_Value(V&: Masked), R: m_ZExtOrSelf(Op: m_Value(V&: ShiftShAmt)))); |
198 | |
199 | // *If* there is a truncation between an outer shift and a possibly-mask, |
200 | // then said truncation *must* be one-use, else we can't perform the fold. |
201 | Value *Trunc; |
202 | if (match(V: Masked, P: m_CombineAnd(L: m_Trunc(Op: m_Value(V&: Masked)), R: m_Value(V&: Trunc))) && |
203 | !Trunc->hasOneUse()) |
204 | return nullptr; |
205 | |
206 | Type *NarrowestTy = OuterShift->getType(); |
207 | Type *WidestTy = Masked->getType(); |
208 | bool HadTrunc = WidestTy != NarrowestTy; |
209 | |
210 | // The mask must be computed in a type twice as wide to ensure |
211 | // that no bits are lost if the sum-of-shifts is wider than the base type. |
212 | Type *ExtendedTy = WidestTy->getExtendedType(); |
213 | |
214 | Value *MaskShAmt; |
215 | |
216 | // ((1 << MaskShAmt) - 1) |
217 | auto MaskA = m_Add(L: m_Shl(L: m_One(), R: m_Value(V&: MaskShAmt)), R: m_AllOnes()); |
218 | // (~(-1 << maskNbits)) |
219 | auto MaskB = m_Not(V: m_Shl(L: m_AllOnes(), R: m_Value(V&: MaskShAmt))); |
220 | // (-1 l>> MaskShAmt) |
221 | auto MaskC = m_LShr(L: m_AllOnes(), R: m_Value(V&: MaskShAmt)); |
222 | // ((-1 << MaskShAmt) l>> MaskShAmt) |
223 | auto MaskD = |
224 | m_LShr(L: m_Shl(L: m_AllOnes(), R: m_Value(V&: MaskShAmt)), R: m_Deferred(V: MaskShAmt)); |
225 | |
226 | Value *X; |
227 | Constant *NewMask; |
228 | |
229 | if (match(V: Masked, P: m_c_And(L: m_CombineOr(L: MaskA, R: MaskB), R: m_Value(V&: X)))) { |
230 | // Peek through an optional zext of the shift amount. |
231 | match(V: MaskShAmt, P: m_ZExtOrSelf(Op: m_Value(V&: MaskShAmt))); |
232 | |
233 | // Verify that it would be safe to try to add those two shift amounts. |
234 | if (!canTryToConstantAddTwoShiftAmounts(Sh0: OuterShift, ShAmt0: ShiftShAmt, Sh1: Masked, |
235 | ShAmt1: MaskShAmt)) |
236 | return nullptr; |
237 | |
238 | // Can we simplify (MaskShAmt+ShiftShAmt) ? |
239 | auto *SumOfShAmts = dyn_cast_or_null<Constant>(Val: simplifyAddInst( |
240 | LHS: MaskShAmt, RHS: ShiftShAmt, /*IsNSW=*/false, /*IsNUW=*/false, Q)); |
241 | if (!SumOfShAmts) |
242 | return nullptr; // Did not simplify. |
243 | // In this pattern SumOfShAmts correlates with the number of low bits |
244 | // that shall remain in the root value (OuterShift). |
245 | |
246 | // An extend of an undef value becomes zero because the high bits are never |
247 | // completely unknown. Replace the `undef` shift amounts with final |
248 | // shift bitwidth to ensure that the value remains undef when creating the |
249 | // subsequent shift op. |
250 | SumOfShAmts = Constant::replaceUndefsWith( |
251 | C: SumOfShAmts, Replacement: ConstantInt::get(Ty: SumOfShAmts->getType()->getScalarType(), |
252 | V: ExtendedTy->getScalarSizeInBits())); |
253 | auto *ExtendedSumOfShAmts = ConstantFoldCastOperand( |
254 | Opcode: Instruction::ZExt, C: SumOfShAmts, DestTy: ExtendedTy, DL: Q.DL); |
255 | if (!ExtendedSumOfShAmts) |
256 | return nullptr; |
257 | |
258 | // And compute the mask as usual: ~(-1 << (SumOfShAmts)) |
259 | auto *ExtendedAllOnes = ConstantExpr::getAllOnesValue(Ty: ExtendedTy); |
260 | Constant *ExtendedInvertedMask = ConstantFoldBinaryOpOperands( |
261 | Opcode: Instruction::Shl, LHS: ExtendedAllOnes, RHS: ExtendedSumOfShAmts, DL: Q.DL); |
262 | if (!ExtendedInvertedMask) |
263 | return nullptr; |
264 | |
265 | NewMask = ConstantExpr::getNot(C: ExtendedInvertedMask); |
266 | } else if (match(V: Masked, P: m_c_And(L: m_CombineOr(L: MaskC, R: MaskD), R: m_Value(V&: X))) || |
267 | match(V: Masked, P: m_Shr(L: m_Shl(L: m_Value(V&: X), R: m_Value(V&: MaskShAmt)), |
268 | R: m_Deferred(V: MaskShAmt)))) { |
269 | // Peek through an optional zext of the shift amount. |
270 | match(V: MaskShAmt, P: m_ZExtOrSelf(Op: m_Value(V&: MaskShAmt))); |
271 | |
272 | // Verify that it would be safe to try to add those two shift amounts. |
273 | if (!canTryToConstantAddTwoShiftAmounts(Sh0: OuterShift, ShAmt0: ShiftShAmt, Sh1: Masked, |
274 | ShAmt1: MaskShAmt)) |
275 | return nullptr; |
276 | |
277 | // Can we simplify (ShiftShAmt-MaskShAmt) ? |
278 | auto *ShAmtsDiff = dyn_cast_or_null<Constant>(Val: simplifySubInst( |
279 | LHS: ShiftShAmt, RHS: MaskShAmt, /*IsNSW=*/false, /*IsNUW=*/false, Q)); |
280 | if (!ShAmtsDiff) |
281 | return nullptr; // Did not simplify. |
282 | // In this pattern ShAmtsDiff correlates with the number of high bits that |
283 | // shall be unset in the root value (OuterShift). |
284 | |
285 | // An extend of an undef value becomes zero because the high bits are never |
286 | // completely unknown. Replace the `undef` shift amounts with negated |
287 | // bitwidth of innermost shift to ensure that the value remains undef when |
288 | // creating the subsequent shift op. |
289 | unsigned WidestTyBitWidth = WidestTy->getScalarSizeInBits(); |
290 | ShAmtsDiff = Constant::replaceUndefsWith( |
291 | C: ShAmtsDiff, Replacement: ConstantInt::get(Ty: ShAmtsDiff->getType()->getScalarType(), |
292 | V: -WidestTyBitWidth)); |
293 | auto *ExtendedNumHighBitsToClear = ConstantFoldCastOperand( |
294 | Opcode: Instruction::ZExt, |
295 | C: ConstantExpr::getSub(C1: ConstantInt::get(Ty: ShAmtsDiff->getType(), |
296 | V: WidestTyBitWidth, |
297 | /*isSigned=*/IsSigned: false), |
298 | C2: ShAmtsDiff), |
299 | DestTy: ExtendedTy, DL: Q.DL); |
300 | if (!ExtendedNumHighBitsToClear) |
301 | return nullptr; |
302 | |
303 | // And compute the mask as usual: (-1 l>> (NumHighBitsToClear)) |
304 | auto *ExtendedAllOnes = ConstantExpr::getAllOnesValue(Ty: ExtendedTy); |
305 | NewMask = ConstantFoldBinaryOpOperands(Opcode: Instruction::LShr, LHS: ExtendedAllOnes, |
306 | RHS: ExtendedNumHighBitsToClear, DL: Q.DL); |
307 | if (!NewMask) |
308 | return nullptr; |
309 | } else |
310 | return nullptr; // Don't know anything about this pattern. |
311 | |
312 | NewMask = ConstantExpr::getTrunc(C: NewMask, Ty: NarrowestTy); |
313 | |
314 | // Does this mask has any unset bits? If not then we can just not apply it. |
315 | bool NeedMask = !match(V: NewMask, P: m_AllOnes()); |
316 | |
317 | // If we need to apply a mask, there are several more restrictions we have. |
318 | if (NeedMask) { |
319 | // The old masking instruction must go away. |
320 | if (!Masked->hasOneUse()) |
321 | return nullptr; |
322 | // The original "masking" instruction must not have been`ashr`. |
323 | if (match(V: Masked, P: m_AShr(L: m_Value(), R: m_Value()))) |
324 | return nullptr; |
325 | } |
326 | |
327 | // If we need to apply truncation, let's do it first, since we can. |
328 | // We have already ensured that the old truncation will go away. |
329 | if (HadTrunc) |
330 | X = Builder.CreateTrunc(V: X, DestTy: NarrowestTy); |
331 | |
332 | // No 'NUW'/'NSW'! We no longer know that we won't shift-out non-0 bits. |
333 | // We didn't change the Type of this outermost shift, so we can just do it. |
334 | auto *NewShift = BinaryOperator::Create(Op: OuterShift->getOpcode(), S1: X, |
335 | S2: OuterShift->getOperand(i_nocapture: 1)); |
336 | if (!NeedMask) |
337 | return NewShift; |
338 | |
339 | Builder.Insert(I: NewShift); |
340 | return BinaryOperator::Create(Op: Instruction::And, S1: NewShift, S2: NewMask); |
341 | } |
342 | |
343 | /// If we have a shift-by-constant of a bin op (bitwise logic op or add/sub w/ |
344 | /// shl) that itself has a shift-by-constant operand with identical opcode, we |
345 | /// may be able to convert that into 2 independent shifts followed by the logic |
346 | /// op. This eliminates a use of an intermediate value (reduces dependency |
347 | /// chain). |
348 | static Instruction *foldShiftOfShiftedBinOp(BinaryOperator &I, |
349 | InstCombiner::BuilderTy &Builder) { |
350 | assert(I.isShift() && "Expected a shift as input" ); |
351 | auto *BinInst = dyn_cast<BinaryOperator>(Val: I.getOperand(i_nocapture: 0)); |
352 | if (!BinInst || |
353 | (!BinInst->isBitwiseLogicOp() && |
354 | BinInst->getOpcode() != Instruction::Add && |
355 | BinInst->getOpcode() != Instruction::Sub) || |
356 | !BinInst->hasOneUse()) |
357 | return nullptr; |
358 | |
359 | Constant *C0, *C1; |
360 | if (!match(V: I.getOperand(i_nocapture: 1), P: m_Constant(C&: C1))) |
361 | return nullptr; |
362 | |
363 | Instruction::BinaryOps ShiftOpcode = I.getOpcode(); |
364 | // Transform for add/sub only works with shl. |
365 | if ((BinInst->getOpcode() == Instruction::Add || |
366 | BinInst->getOpcode() == Instruction::Sub) && |
367 | ShiftOpcode != Instruction::Shl) |
368 | return nullptr; |
369 | |
370 | Type *Ty = I.getType(); |
371 | |
372 | // Find a matching shift by constant. The fold is not valid if the sum |
373 | // of the shift values equals or exceeds bitwidth. |
374 | Value *X, *Y; |
375 | auto matchFirstShift = [&](Value *V, Value *W) { |
376 | unsigned Size = Ty->getScalarSizeInBits(); |
377 | APInt Threshold(Size, Size); |
378 | return match(V, P: m_BinOp(Opcode: ShiftOpcode, L: m_Value(V&: X), R: m_Constant(C&: C0))) && |
379 | (V->hasOneUse() || match(V: W, P: m_ImmConstant())) && |
380 | match(V: ConstantExpr::getAdd(C1: C0, C2: C1), |
381 | P: m_SpecificInt_ICMP(Predicate: ICmpInst::ICMP_ULT, Threshold)); |
382 | }; |
383 | |
384 | // Logic ops and Add are commutative, so check each operand for a match. Sub |
385 | // is not so we cannot reoder if we match operand(1) and need to keep the |
386 | // operands in their original positions. |
387 | bool FirstShiftIsOp1 = false; |
388 | if (matchFirstShift(BinInst->getOperand(i_nocapture: 0), BinInst->getOperand(i_nocapture: 1))) |
389 | Y = BinInst->getOperand(i_nocapture: 1); |
390 | else if (matchFirstShift(BinInst->getOperand(i_nocapture: 1), BinInst->getOperand(i_nocapture: 0))) { |
391 | Y = BinInst->getOperand(i_nocapture: 0); |
392 | FirstShiftIsOp1 = BinInst->getOpcode() == Instruction::Sub; |
393 | } else |
394 | return nullptr; |
395 | |
396 | // shift (binop (shift X, C0), Y), C1 -> binop (shift X, C0+C1), (shift Y, C1) |
397 | Constant *ShiftSumC = ConstantExpr::getAdd(C1: C0, C2: C1); |
398 | Value *NewShift1 = Builder.CreateBinOp(Opc: ShiftOpcode, LHS: X, RHS: ShiftSumC); |
399 | Value *NewShift2 = Builder.CreateBinOp(Opc: ShiftOpcode, LHS: Y, RHS: C1); |
400 | Value *Op1 = FirstShiftIsOp1 ? NewShift2 : NewShift1; |
401 | Value *Op2 = FirstShiftIsOp1 ? NewShift1 : NewShift2; |
402 | return BinaryOperator::Create(Op: BinInst->getOpcode(), S1: Op1, S2: Op2); |
403 | } |
404 | |
405 | Instruction *InstCombinerImpl::commonShiftTransforms(BinaryOperator &I) { |
406 | if (Instruction *Phi = foldBinopWithPhiOperands(BO&: I)) |
407 | return Phi; |
408 | |
409 | Value *Op0 = I.getOperand(i_nocapture: 0), *Op1 = I.getOperand(i_nocapture: 1); |
410 | assert(Op0->getType() == Op1->getType()); |
411 | Type *Ty = I.getType(); |
412 | |
413 | // If the shift amount is a one-use `sext`, we can demote it to `zext`. |
414 | Value *Y; |
415 | if (match(V: Op1, P: m_OneUse(SubPattern: m_SExt(Op: m_Value(V&: Y))))) { |
416 | Value *NewExt = Builder.CreateZExt(V: Y, DestTy: Ty, Name: Op1->getName()); |
417 | return BinaryOperator::Create(Op: I.getOpcode(), S1: Op0, S2: NewExt); |
418 | } |
419 | |
420 | // See if we can fold away this shift. |
421 | if (SimplifyDemandedInstructionBits(Inst&: I)) |
422 | return &I; |
423 | |
424 | // Try to fold constant and into select arguments. |
425 | if (isa<Constant>(Val: Op0)) |
426 | if (SelectInst *SI = dyn_cast<SelectInst>(Val: Op1)) |
427 | if (Instruction *R = FoldOpIntoSelect(Op&: I, SI)) |
428 | return R; |
429 | |
430 | Constant *CUI; |
431 | if (match(V: Op1, P: m_ImmConstant(C&: CUI))) |
432 | if (Instruction *Res = FoldShiftByConstant(Op0, Op1: CUI, I)) |
433 | return Res; |
434 | |
435 | if (auto *NewShift = cast_or_null<Instruction>( |
436 | Val: reassociateShiftAmtsOfTwoSameDirectionShifts(Sh0: &I, SQ))) |
437 | return NewShift; |
438 | |
439 | // Pre-shift a constant shifted by a variable amount with constant offset: |
440 | // C shift (A add nuw C1) --> (C shift C1) shift A |
441 | Value *A; |
442 | Constant *C, *C1; |
443 | if (match(V: Op0, P: m_Constant(C)) && |
444 | match(V: Op1, P: m_NUWAddLike(L: m_Value(V&: A), R: m_Constant(C&: C1)))) { |
445 | Value *NewC = Builder.CreateBinOp(Opc: I.getOpcode(), LHS: C, RHS: C1); |
446 | BinaryOperator *NewShiftOp = BinaryOperator::Create(Op: I.getOpcode(), S1: NewC, S2: A); |
447 | if (I.getOpcode() == Instruction::Shl) { |
448 | NewShiftOp->setHasNoSignedWrap(I.hasNoSignedWrap()); |
449 | NewShiftOp->setHasNoUnsignedWrap(I.hasNoUnsignedWrap()); |
450 | } else { |
451 | NewShiftOp->setIsExact(I.isExact()); |
452 | } |
453 | return NewShiftOp; |
454 | } |
455 | |
456 | unsigned BitWidth = Ty->getScalarSizeInBits(); |
457 | |
458 | const APInt *AC, *AddC; |
459 | // Try to pre-shift a constant shifted by a variable amount added with a |
460 | // negative number: |
461 | // C << (X - AddC) --> (C >> AddC) << X |
462 | // and |
463 | // C >> (X - AddC) --> (C << AddC) >> X |
464 | if (match(V: Op0, P: m_APInt(Res&: AC)) && match(V: Op1, P: m_Add(L: m_Value(V&: A), R: m_APInt(Res&: AddC))) && |
465 | AddC->isNegative() && (-*AddC).ult(RHS: BitWidth)) { |
466 | assert(!AC->isZero() && "Expected simplify of shifted zero" ); |
467 | unsigned PosOffset = (-*AddC).getZExtValue(); |
468 | |
469 | auto isSuitableForPreShift = [PosOffset, &I, AC]() { |
470 | switch (I.getOpcode()) { |
471 | default: |
472 | return false; |
473 | case Instruction::Shl: |
474 | return (I.hasNoSignedWrap() || I.hasNoUnsignedWrap()) && |
475 | AC->eq(RHS: AC->lshr(shiftAmt: PosOffset).shl(shiftAmt: PosOffset)); |
476 | case Instruction::LShr: |
477 | return I.isExact() && AC->eq(RHS: AC->shl(shiftAmt: PosOffset).lshr(shiftAmt: PosOffset)); |
478 | case Instruction::AShr: |
479 | return I.isExact() && AC->eq(RHS: AC->shl(shiftAmt: PosOffset).ashr(ShiftAmt: PosOffset)); |
480 | } |
481 | }; |
482 | if (isSuitableForPreShift()) { |
483 | Constant *NewC = ConstantInt::get(Ty, V: I.getOpcode() == Instruction::Shl |
484 | ? AC->lshr(shiftAmt: PosOffset) |
485 | : AC->shl(shiftAmt: PosOffset)); |
486 | BinaryOperator *NewShiftOp = |
487 | BinaryOperator::Create(Op: I.getOpcode(), S1: NewC, S2: A); |
488 | if (I.getOpcode() == Instruction::Shl) { |
489 | NewShiftOp->setHasNoUnsignedWrap(I.hasNoUnsignedWrap()); |
490 | } else { |
491 | NewShiftOp->setIsExact(); |
492 | } |
493 | return NewShiftOp; |
494 | } |
495 | } |
496 | |
497 | // X shift (A srem C) -> X shift (A and (C - 1)) iff C is a power of 2. |
498 | // Because shifts by negative values (which could occur if A were negative) |
499 | // are undefined. |
500 | if (Op1->hasOneUse() && match(V: Op1, P: m_SRem(L: m_Value(V&: A), R: m_Constant(C))) && |
501 | match(V: C, P: m_Power2())) { |
502 | // FIXME: Should this get moved into SimplifyDemandedBits by saying we don't |
503 | // demand the sign bit (and many others) here?? |
504 | Constant *Mask = ConstantExpr::getSub(C1: C, C2: ConstantInt::get(Ty, V: 1)); |
505 | Value *Rem = Builder.CreateAnd(LHS: A, RHS: Mask, Name: Op1->getName()); |
506 | return replaceOperand(I, OpNum: 1, V: Rem); |
507 | } |
508 | |
509 | if (Instruction *Logic = foldShiftOfShiftedBinOp(I, Builder)) |
510 | return Logic; |
511 | |
512 | if (match(V: Op1, P: m_Or(L: m_Value(), R: m_SpecificInt(V: BitWidth - 1)))) |
513 | return replaceOperand(I, OpNum: 1, V: ConstantInt::get(Ty, V: BitWidth - 1)); |
514 | |
515 | Instruction *CmpIntr; |
516 | if ((I.getOpcode() == Instruction::LShr || |
517 | I.getOpcode() == Instruction::AShr) && |
518 | match(V: Op0, P: m_OneUse(SubPattern: m_Instruction(I&: CmpIntr))) && |
519 | isa<CmpIntrinsic>(Val: CmpIntr) && |
520 | match(V: Op1, P: m_SpecificInt(V: Ty->getScalarSizeInBits() - 1))) { |
521 | Value *Cmp = |
522 | Builder.CreateICmp(P: cast<CmpIntrinsic>(Val: CmpIntr)->getLTPredicate(), |
523 | LHS: CmpIntr->getOperand(i: 0), RHS: CmpIntr->getOperand(i: 1)); |
524 | return CastInst::Create(I.getOpcode() == Instruction::LShr |
525 | ? Instruction::ZExt |
526 | : Instruction::SExt, |
527 | S: Cmp, Ty); |
528 | } |
529 | |
530 | return nullptr; |
531 | } |
532 | |
533 | /// Return true if we can simplify two logical (either left or right) shifts |
534 | /// that have constant shift amounts: OuterShift (InnerShift X, C1), C2. |
535 | static bool canEvaluateShiftedShift(unsigned OuterShAmt, bool IsOuterShl, |
536 | Instruction *InnerShift, |
537 | InstCombinerImpl &IC, Instruction *CxtI) { |
538 | assert(InnerShift->isLogicalShift() && "Unexpected instruction type" ); |
539 | |
540 | // We need constant scalar or constant splat shifts. |
541 | const APInt *InnerShiftConst; |
542 | if (!match(V: InnerShift->getOperand(i: 1), P: m_APInt(Res&: InnerShiftConst))) |
543 | return false; |
544 | |
545 | // Two logical shifts in the same direction: |
546 | // shl (shl X, C1), C2 --> shl X, C1 + C2 |
547 | // lshr (lshr X, C1), C2 --> lshr X, C1 + C2 |
548 | bool IsInnerShl = InnerShift->getOpcode() == Instruction::Shl; |
549 | if (IsInnerShl == IsOuterShl) |
550 | return true; |
551 | |
552 | // Equal shift amounts in opposite directions become bitwise 'and': |
553 | // lshr (shl X, C), C --> and X, C' |
554 | // shl (lshr X, C), C --> and X, C' |
555 | if (*InnerShiftConst == OuterShAmt) |
556 | return true; |
557 | |
558 | // If the 2nd shift is bigger than the 1st, we can fold: |
559 | // lshr (shl X, C1), C2 --> and (shl X, C1 - C2), C3 |
560 | // shl (lshr X, C1), C2 --> and (lshr X, C1 - C2), C3 |
561 | // but it isn't profitable unless we know the and'd out bits are already zero. |
562 | // Also, check that the inner shift is valid (less than the type width) or |
563 | // we'll crash trying to produce the bit mask for the 'and'. |
564 | unsigned TypeWidth = InnerShift->getType()->getScalarSizeInBits(); |
565 | if (InnerShiftConst->ugt(RHS: OuterShAmt) && InnerShiftConst->ult(RHS: TypeWidth)) { |
566 | unsigned InnerShAmt = InnerShiftConst->getZExtValue(); |
567 | unsigned MaskShift = |
568 | IsInnerShl ? TypeWidth - InnerShAmt : InnerShAmt - OuterShAmt; |
569 | APInt Mask = APInt::getLowBitsSet(numBits: TypeWidth, loBitsSet: OuterShAmt) << MaskShift; |
570 | if (IC.MaskedValueIsZero(V: InnerShift->getOperand(i: 0), Mask, CxtI)) |
571 | return true; |
572 | } |
573 | |
574 | return false; |
575 | } |
576 | |
577 | /// See if we can compute the specified value, but shifted logically to the left |
578 | /// or right by some number of bits. This should return true if the expression |
579 | /// can be computed for the same cost as the current expression tree. This is |
580 | /// used to eliminate extraneous shifting from things like: |
581 | /// %C = shl i128 %A, 64 |
582 | /// %D = shl i128 %B, 96 |
583 | /// %E = or i128 %C, %D |
584 | /// %F = lshr i128 %E, 64 |
585 | /// where the client will ask if E can be computed shifted right by 64-bits. If |
586 | /// this succeeds, getShiftedValue() will be called to produce the value. |
587 | static bool canEvaluateShifted(Value *V, unsigned NumBits, bool IsLeftShift, |
588 | InstCombinerImpl &IC, Instruction *CxtI) { |
589 | // We can always evaluate immediate constants. |
590 | if (match(V, P: m_ImmConstant())) |
591 | return true; |
592 | |
593 | Instruction *I = dyn_cast<Instruction>(Val: V); |
594 | if (!I) return false; |
595 | |
596 | // We can't mutate something that has multiple uses: doing so would |
597 | // require duplicating the instruction in general, which isn't profitable. |
598 | if (!I->hasOneUse()) return false; |
599 | |
600 | switch (I->getOpcode()) { |
601 | default: return false; |
602 | case Instruction::And: |
603 | case Instruction::Or: |
604 | case Instruction::Xor: |
605 | // Bitwise operators can all arbitrarily be arbitrarily evaluated shifted. |
606 | return canEvaluateShifted(V: I->getOperand(i: 0), NumBits, IsLeftShift, IC, CxtI: I) && |
607 | canEvaluateShifted(V: I->getOperand(i: 1), NumBits, IsLeftShift, IC, CxtI: I); |
608 | |
609 | case Instruction::Shl: |
610 | case Instruction::LShr: |
611 | return canEvaluateShiftedShift(OuterShAmt: NumBits, IsOuterShl: IsLeftShift, InnerShift: I, IC, CxtI); |
612 | |
613 | case Instruction::Select: { |
614 | SelectInst *SI = cast<SelectInst>(Val: I); |
615 | Value *TrueVal = SI->getTrueValue(); |
616 | Value *FalseVal = SI->getFalseValue(); |
617 | return canEvaluateShifted(V: TrueVal, NumBits, IsLeftShift, IC, CxtI: SI) && |
618 | canEvaluateShifted(V: FalseVal, NumBits, IsLeftShift, IC, CxtI: SI); |
619 | } |
620 | case Instruction::PHI: { |
621 | // We can change a phi if we can change all operands. Note that we never |
622 | // get into trouble with cyclic PHIs here because we only consider |
623 | // instructions with a single use. |
624 | PHINode *PN = cast<PHINode>(Val: I); |
625 | for (Value *IncValue : PN->incoming_values()) |
626 | if (!canEvaluateShifted(V: IncValue, NumBits, IsLeftShift, IC, CxtI: PN)) |
627 | return false; |
628 | return true; |
629 | } |
630 | case Instruction::Mul: { |
631 | const APInt *MulConst; |
632 | // We can fold (shr (mul X, -(1 << C)), C) -> (and (neg X), C`) |
633 | return !IsLeftShift && match(V: I->getOperand(i: 1), P: m_APInt(Res&: MulConst)) && |
634 | MulConst->isNegatedPowerOf2() && MulConst->countr_zero() == NumBits; |
635 | } |
636 | } |
637 | } |
638 | |
639 | /// Fold OuterShift (InnerShift X, C1), C2. |
640 | /// See canEvaluateShiftedShift() for the constraints on these instructions. |
641 | static Value *foldShiftedShift(BinaryOperator *InnerShift, unsigned OuterShAmt, |
642 | bool IsOuterShl, |
643 | InstCombiner::BuilderTy &Builder) { |
644 | bool IsInnerShl = InnerShift->getOpcode() == Instruction::Shl; |
645 | Type *ShType = InnerShift->getType(); |
646 | unsigned TypeWidth = ShType->getScalarSizeInBits(); |
647 | |
648 | // We only accept shifts-by-a-constant in canEvaluateShifted(). |
649 | const APInt *C1; |
650 | match(V: InnerShift->getOperand(i_nocapture: 1), P: m_APInt(Res&: C1)); |
651 | unsigned InnerShAmt = C1->getZExtValue(); |
652 | |
653 | // Change the shift amount and clear the appropriate IR flags. |
654 | auto NewInnerShift = [&](unsigned ShAmt) { |
655 | InnerShift->setOperand(i_nocapture: 1, Val_nocapture: ConstantInt::get(Ty: ShType, V: ShAmt)); |
656 | if (IsInnerShl) { |
657 | InnerShift->setHasNoUnsignedWrap(false); |
658 | InnerShift->setHasNoSignedWrap(false); |
659 | } else { |
660 | InnerShift->setIsExact(false); |
661 | } |
662 | return InnerShift; |
663 | }; |
664 | |
665 | // Two logical shifts in the same direction: |
666 | // shl (shl X, C1), C2 --> shl X, C1 + C2 |
667 | // lshr (lshr X, C1), C2 --> lshr X, C1 + C2 |
668 | if (IsInnerShl == IsOuterShl) { |
669 | // If this is an oversized composite shift, then unsigned shifts get 0. |
670 | if (InnerShAmt + OuterShAmt >= TypeWidth) |
671 | return Constant::getNullValue(Ty: ShType); |
672 | |
673 | return NewInnerShift(InnerShAmt + OuterShAmt); |
674 | } |
675 | |
676 | // Equal shift amounts in opposite directions become bitwise 'and': |
677 | // lshr (shl X, C), C --> and X, C' |
678 | // shl (lshr X, C), C --> and X, C' |
679 | if (InnerShAmt == OuterShAmt) { |
680 | APInt Mask = IsInnerShl |
681 | ? APInt::getLowBitsSet(numBits: TypeWidth, loBitsSet: TypeWidth - OuterShAmt) |
682 | : APInt::getHighBitsSet(numBits: TypeWidth, hiBitsSet: TypeWidth - OuterShAmt); |
683 | Value *And = Builder.CreateAnd(LHS: InnerShift->getOperand(i_nocapture: 0), |
684 | RHS: ConstantInt::get(Ty: ShType, V: Mask)); |
685 | if (auto *AndI = dyn_cast<Instruction>(Val: And)) { |
686 | AndI->moveBefore(InsertPos: InnerShift->getIterator()); |
687 | AndI->takeName(V: InnerShift); |
688 | } |
689 | return And; |
690 | } |
691 | |
692 | assert(InnerShAmt > OuterShAmt && |
693 | "Unexpected opposite direction logical shift pair" ); |
694 | |
695 | // In general, we would need an 'and' for this transform, but |
696 | // canEvaluateShiftedShift() guarantees that the masked-off bits are not used. |
697 | // lshr (shl X, C1), C2 --> shl X, C1 - C2 |
698 | // shl (lshr X, C1), C2 --> lshr X, C1 - C2 |
699 | return NewInnerShift(InnerShAmt - OuterShAmt); |
700 | } |
701 | |
702 | /// When canEvaluateShifted() returns true for an expression, this function |
703 | /// inserts the new computation that produces the shifted value. |
704 | static Value *getShiftedValue(Value *V, unsigned NumBits, bool isLeftShift, |
705 | InstCombinerImpl &IC, const DataLayout &DL) { |
706 | // We can always evaluate constants shifted. |
707 | if (Constant *C = dyn_cast<Constant>(Val: V)) { |
708 | if (isLeftShift) |
709 | return IC.Builder.CreateShl(LHS: C, RHS: NumBits); |
710 | else |
711 | return IC.Builder.CreateLShr(LHS: C, RHS: NumBits); |
712 | } |
713 | |
714 | Instruction *I = cast<Instruction>(Val: V); |
715 | IC.addToWorklist(I); |
716 | |
717 | switch (I->getOpcode()) { |
718 | default: llvm_unreachable("Inconsistency with CanEvaluateShifted" ); |
719 | case Instruction::And: |
720 | case Instruction::Or: |
721 | case Instruction::Xor: |
722 | // Bitwise operators can all arbitrarily be arbitrarily evaluated shifted. |
723 | I->setOperand( |
724 | i: 0, Val: getShiftedValue(V: I->getOperand(i: 0), NumBits, isLeftShift, IC, DL)); |
725 | I->setOperand( |
726 | i: 1, Val: getShiftedValue(V: I->getOperand(i: 1), NumBits, isLeftShift, IC, DL)); |
727 | return I; |
728 | |
729 | case Instruction::Shl: |
730 | case Instruction::LShr: |
731 | return foldShiftedShift(InnerShift: cast<BinaryOperator>(Val: I), OuterShAmt: NumBits, IsOuterShl: isLeftShift, |
732 | Builder&: IC.Builder); |
733 | |
734 | case Instruction::Select: |
735 | I->setOperand( |
736 | i: 1, Val: getShiftedValue(V: I->getOperand(i: 1), NumBits, isLeftShift, IC, DL)); |
737 | I->setOperand( |
738 | i: 2, Val: getShiftedValue(V: I->getOperand(i: 2), NumBits, isLeftShift, IC, DL)); |
739 | return I; |
740 | case Instruction::PHI: { |
741 | // We can change a phi if we can change all operands. Note that we never |
742 | // get into trouble with cyclic PHIs here because we only consider |
743 | // instructions with a single use. |
744 | PHINode *PN = cast<PHINode>(Val: I); |
745 | for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) |
746 | PN->setIncomingValue(i, V: getShiftedValue(V: PN->getIncomingValue(i), NumBits, |
747 | isLeftShift, IC, DL)); |
748 | return PN; |
749 | } |
750 | case Instruction::Mul: { |
751 | assert(!isLeftShift && "Unexpected shift direction!" ); |
752 | auto *Neg = BinaryOperator::CreateNeg(Op: I->getOperand(i: 0)); |
753 | IC.InsertNewInstWith(New: Neg, Old: I->getIterator()); |
754 | unsigned TypeWidth = I->getType()->getScalarSizeInBits(); |
755 | APInt Mask = APInt::getLowBitsSet(numBits: TypeWidth, loBitsSet: TypeWidth - NumBits); |
756 | auto *And = BinaryOperator::CreateAnd(V1: Neg, |
757 | V2: ConstantInt::get(Ty: I->getType(), V: Mask)); |
758 | And->takeName(V: I); |
759 | return IC.InsertNewInstWith(New: And, Old: I->getIterator()); |
760 | } |
761 | } |
762 | } |
763 | |
764 | // If this is a bitwise operator or add with a constant RHS we might be able |
765 | // to pull it through a shift. |
766 | static bool canShiftBinOpWithConstantRHS(BinaryOperator &Shift, |
767 | BinaryOperator *BO) { |
768 | switch (BO->getOpcode()) { |
769 | default: |
770 | return false; // Do not perform transform! |
771 | case Instruction::Add: |
772 | return Shift.getOpcode() == Instruction::Shl; |
773 | case Instruction::Or: |
774 | case Instruction::And: |
775 | return true; |
776 | case Instruction::Xor: |
777 | // Do not change a 'not' of logical shift because that would create a normal |
778 | // 'xor'. The 'not' is likely better for analysis, SCEV, and codegen. |
779 | return !(Shift.isLogicalShift() && match(V: BO, P: m_Not(V: m_Value()))); |
780 | } |
781 | } |
782 | |
783 | Instruction *InstCombinerImpl::FoldShiftByConstant(Value *Op0, Constant *C1, |
784 | BinaryOperator &I) { |
785 | // (C2 << X) << C1 --> (C2 << C1) << X |
786 | // (C2 >> X) >> C1 --> (C2 >> C1) >> X |
787 | Constant *C2; |
788 | Value *X; |
789 | bool IsLeftShift = I.getOpcode() == Instruction::Shl; |
790 | if (match(V: Op0, P: m_BinOp(Opcode: I.getOpcode(), L: m_ImmConstant(C&: C2), R: m_Value(V&: X)))) { |
791 | Instruction *R = BinaryOperator::Create( |
792 | Op: I.getOpcode(), S1: Builder.CreateBinOp(Opc: I.getOpcode(), LHS: C2, RHS: C1), S2: X); |
793 | BinaryOperator *BO0 = cast<BinaryOperator>(Val: Op0); |
794 | if (IsLeftShift) { |
795 | R->setHasNoUnsignedWrap(I.hasNoUnsignedWrap() && |
796 | BO0->hasNoUnsignedWrap()); |
797 | R->setHasNoSignedWrap(I.hasNoSignedWrap() && BO0->hasNoSignedWrap()); |
798 | } else |
799 | R->setIsExact(I.isExact() && BO0->isExact()); |
800 | return R; |
801 | } |
802 | |
803 | Type *Ty = I.getType(); |
804 | unsigned TypeBits = Ty->getScalarSizeInBits(); |
805 | |
806 | // (X / +DivC) >> (Width - 1) --> ext (X <= -DivC) |
807 | // (X / -DivC) >> (Width - 1) --> ext (X >= +DivC) |
808 | const APInt *DivC; |
809 | if (!IsLeftShift && match(V: C1, P: m_SpecificIntAllowPoison(V: TypeBits - 1)) && |
810 | match(V: Op0, P: m_SDiv(L: m_Value(V&: X), R: m_APInt(Res&: DivC))) && !DivC->isZero() && |
811 | !DivC->isMinSignedValue()) { |
812 | Constant *NegDivC = ConstantInt::get(Ty, V: -(*DivC)); |
813 | ICmpInst::Predicate Pred = |
814 | DivC->isNegative() ? ICmpInst::ICMP_SGE : ICmpInst::ICMP_SLE; |
815 | Value *Cmp = Builder.CreateICmp(P: Pred, LHS: X, RHS: NegDivC); |
816 | auto ExtOpcode = (I.getOpcode() == Instruction::AShr) ? Instruction::SExt |
817 | : Instruction::ZExt; |
818 | return CastInst::Create(ExtOpcode, S: Cmp, Ty); |
819 | } |
820 | |
821 | const APInt *Op1C; |
822 | if (!match(V: C1, P: m_APInt(Res&: Op1C))) |
823 | return nullptr; |
824 | |
825 | assert(!Op1C->uge(TypeBits) && |
826 | "Shift over the type width should have been removed already" ); |
827 | |
828 | // See if we can propagate this shift into the input, this covers the trivial |
829 | // cast of lshr(shl(x,c1),c2) as well as other more complex cases. |
830 | if (I.getOpcode() != Instruction::AShr && |
831 | canEvaluateShifted(V: Op0, NumBits: Op1C->getZExtValue(), IsLeftShift, IC&: *this, CxtI: &I)) { |
832 | LLVM_DEBUG( |
833 | dbgs() << "ICE: GetShiftedValue propagating shift through expression" |
834 | " to eliminate shift:\n IN: " |
835 | << *Op0 << "\n SH: " << I << "\n" ); |
836 | |
837 | return replaceInstUsesWith( |
838 | I, V: getShiftedValue(V: Op0, NumBits: Op1C->getZExtValue(), isLeftShift: IsLeftShift, IC&: *this, DL)); |
839 | } |
840 | |
841 | if (Instruction *FoldedShift = foldBinOpIntoSelectOrPhi(I)) |
842 | return FoldedShift; |
843 | |
844 | if (!Op0->hasOneUse()) |
845 | return nullptr; |
846 | |
847 | if (auto *Op0BO = dyn_cast<BinaryOperator>(Val: Op0)) { |
848 | // If the operand is a bitwise operator with a constant RHS, and the |
849 | // shift is the only use, we can pull it out of the shift. |
850 | const APInt *Op0C; |
851 | if (match(V: Op0BO->getOperand(i_nocapture: 1), P: m_APInt(Res&: Op0C))) { |
852 | if (canShiftBinOpWithConstantRHS(Shift&: I, BO: Op0BO)) { |
853 | Value *NewRHS = |
854 | Builder.CreateBinOp(Opc: I.getOpcode(), LHS: Op0BO->getOperand(i_nocapture: 1), RHS: C1); |
855 | |
856 | Value *NewShift = |
857 | Builder.CreateBinOp(Opc: I.getOpcode(), LHS: Op0BO->getOperand(i_nocapture: 0), RHS: C1); |
858 | NewShift->takeName(V: Op0BO); |
859 | |
860 | return BinaryOperator::Create(Op: Op0BO->getOpcode(), S1: NewShift, S2: NewRHS); |
861 | } |
862 | } |
863 | } |
864 | |
865 | // If we have a select that conditionally executes some binary operator, |
866 | // see if we can pull it the select and operator through the shift. |
867 | // |
868 | // For example, turning: |
869 | // shl (select C, (add X, C1), X), C2 |
870 | // Into: |
871 | // Y = shl X, C2 |
872 | // select C, (add Y, C1 << C2), Y |
873 | Value *Cond; |
874 | BinaryOperator *TBO; |
875 | Value *FalseVal; |
876 | if (match(V: Op0, P: m_Select(C: m_Value(V&: Cond), L: m_OneUse(SubPattern: m_BinOp(I&: TBO)), |
877 | R: m_Value(V&: FalseVal)))) { |
878 | const APInt *C; |
879 | if (!isa<Constant>(Val: FalseVal) && TBO->getOperand(i_nocapture: 0) == FalseVal && |
880 | match(V: TBO->getOperand(i_nocapture: 1), P: m_APInt(Res&: C)) && |
881 | canShiftBinOpWithConstantRHS(Shift&: I, BO: TBO)) { |
882 | Value *NewRHS = |
883 | Builder.CreateBinOp(Opc: I.getOpcode(), LHS: TBO->getOperand(i_nocapture: 1), RHS: C1); |
884 | |
885 | Value *NewShift = Builder.CreateBinOp(Opc: I.getOpcode(), LHS: FalseVal, RHS: C1); |
886 | Value *NewOp = Builder.CreateBinOp(Opc: TBO->getOpcode(), LHS: NewShift, RHS: NewRHS); |
887 | return SelectInst::Create(C: Cond, S1: NewOp, S2: NewShift); |
888 | } |
889 | } |
890 | |
891 | BinaryOperator *FBO; |
892 | Value *TrueVal; |
893 | if (match(V: Op0, P: m_Select(C: m_Value(V&: Cond), L: m_Value(V&: TrueVal), |
894 | R: m_OneUse(SubPattern: m_BinOp(I&: FBO))))) { |
895 | const APInt *C; |
896 | if (!isa<Constant>(Val: TrueVal) && FBO->getOperand(i_nocapture: 0) == TrueVal && |
897 | match(V: FBO->getOperand(i_nocapture: 1), P: m_APInt(Res&: C)) && |
898 | canShiftBinOpWithConstantRHS(Shift&: I, BO: FBO)) { |
899 | Value *NewRHS = |
900 | Builder.CreateBinOp(Opc: I.getOpcode(), LHS: FBO->getOperand(i_nocapture: 1), RHS: C1); |
901 | |
902 | Value *NewShift = Builder.CreateBinOp(Opc: I.getOpcode(), LHS: TrueVal, RHS: C1); |
903 | Value *NewOp = Builder.CreateBinOp(Opc: FBO->getOpcode(), LHS: NewShift, RHS: NewRHS); |
904 | return SelectInst::Create(C: Cond, S1: NewShift, S2: NewOp); |
905 | } |
906 | } |
907 | |
908 | return nullptr; |
909 | } |
910 | |
911 | // Tries to perform |
912 | // (lshr (add (zext X), (zext Y)), K) |
913 | // -> (icmp ult (add X, Y), X) |
914 | // where |
915 | // - The add's operands are zexts from a K-bits integer to a bigger type. |
916 | // - The add is only used by the shr, or by iK (or narrower) truncates. |
917 | // - The lshr type has more than 2 bits (other types are boolean math). |
918 | // - K > 1 |
919 | // note that |
920 | // - The resulting add cannot have nuw/nsw, else on overflow we get a |
921 | // poison value and the transform isn't legal anymore. |
922 | Instruction *InstCombinerImpl::foldLShrOverflowBit(BinaryOperator &I) { |
923 | assert(I.getOpcode() == Instruction::LShr); |
924 | |
925 | Value *Add = I.getOperand(i_nocapture: 0); |
926 | Value *ShiftAmt = I.getOperand(i_nocapture: 1); |
927 | Type *Ty = I.getType(); |
928 | |
929 | if (Ty->getScalarSizeInBits() < 3) |
930 | return nullptr; |
931 | |
932 | const APInt *ShAmtAPInt = nullptr; |
933 | Value *X = nullptr, *Y = nullptr; |
934 | if (!match(V: ShiftAmt, P: m_APInt(Res&: ShAmtAPInt)) || |
935 | !match(V: Add, |
936 | P: m_Add(L: m_OneUse(SubPattern: m_ZExt(Op: m_Value(V&: X))), R: m_OneUse(SubPattern: m_ZExt(Op: m_Value(V&: Y)))))) |
937 | return nullptr; |
938 | |
939 | const unsigned ShAmt = ShAmtAPInt->getZExtValue(); |
940 | if (ShAmt == 1) |
941 | return nullptr; |
942 | |
943 | // X/Y are zexts from `ShAmt`-sized ints. |
944 | if (X->getType()->getScalarSizeInBits() != ShAmt || |
945 | Y->getType()->getScalarSizeInBits() != ShAmt) |
946 | return nullptr; |
947 | |
948 | // Make sure that `Add` is only used by `I` and `ShAmt`-truncates. |
949 | if (!Add->hasOneUse()) { |
950 | for (User *U : Add->users()) { |
951 | if (U == &I) |
952 | continue; |
953 | |
954 | TruncInst *Trunc = dyn_cast<TruncInst>(Val: U); |
955 | if (!Trunc || Trunc->getType()->getScalarSizeInBits() > ShAmt) |
956 | return nullptr; |
957 | } |
958 | } |
959 | |
960 | // Insert at Add so that the newly created `NarrowAdd` will dominate it's |
961 | // users (i.e. `Add`'s users). |
962 | Instruction *AddInst = cast<Instruction>(Val: Add); |
963 | Builder.SetInsertPoint(AddInst); |
964 | |
965 | Value *NarrowAdd = Builder.CreateAdd(LHS: X, RHS: Y, Name: "add.narrowed" ); |
966 | Value *Overflow = |
967 | Builder.CreateICmpULT(LHS: NarrowAdd, RHS: X, Name: "add.narrowed.overflow" ); |
968 | |
969 | // Replace the uses of the original add with a zext of the |
970 | // NarrowAdd's result. Note that all users at this stage are known to |
971 | // be ShAmt-sized truncs, or the lshr itself. |
972 | if (!Add->hasOneUse()) { |
973 | replaceInstUsesWith(I&: *AddInst, V: Builder.CreateZExt(V: NarrowAdd, DestTy: Ty)); |
974 | eraseInstFromFunction(I&: *AddInst); |
975 | } |
976 | |
977 | // Replace the LShr with a zext of the overflow check. |
978 | return new ZExtInst(Overflow, Ty); |
979 | } |
980 | |
981 | // Try to set nuw/nsw flags on shl or exact flag on lshr/ashr using knownbits. |
982 | static bool setShiftFlags(BinaryOperator &I, const SimplifyQuery &Q) { |
983 | assert(I.isShift() && "Expected a shift as input" ); |
984 | // We already have all the flags. |
985 | if (I.getOpcode() == Instruction::Shl) { |
986 | if (I.hasNoUnsignedWrap() && I.hasNoSignedWrap()) |
987 | return false; |
988 | } else { |
989 | if (I.isExact()) |
990 | return false; |
991 | |
992 | // shr (shl X, Y), Y |
993 | if (match(V: I.getOperand(i_nocapture: 0), P: m_Shl(L: m_Value(), R: m_Specific(V: I.getOperand(i_nocapture: 1))))) { |
994 | I.setIsExact(); |
995 | return true; |
996 | } |
997 | // Infer 'exact' flag if shift amount is cttz(x) on the same operand. |
998 | if (match(V: I.getOperand(i_nocapture: 1), P: m_Intrinsic<Intrinsic::cttz>( |
999 | Op0: m_Specific(V: I.getOperand(i_nocapture: 0)), Op1: m_Value()))) { |
1000 | I.setIsExact(); |
1001 | return true; |
1002 | } |
1003 | } |
1004 | |
1005 | // Compute what we know about shift count. |
1006 | KnownBits KnownCnt = computeKnownBits(V: I.getOperand(i_nocapture: 1), Q); |
1007 | unsigned BitWidth = KnownCnt.getBitWidth(); |
1008 | // Since shift produces a poison value if RHS is equal to or larger than the |
1009 | // bit width, we can safely assume that RHS is less than the bit width. |
1010 | uint64_t MaxCnt = KnownCnt.getMaxValue().getLimitedValue(Limit: BitWidth - 1); |
1011 | |
1012 | KnownBits KnownAmt = computeKnownBits(V: I.getOperand(i_nocapture: 0), Q); |
1013 | bool Changed = false; |
1014 | |
1015 | if (I.getOpcode() == Instruction::Shl) { |
1016 | // If we have as many leading zeros than maximum shift cnt we have nuw. |
1017 | if (!I.hasNoUnsignedWrap() && MaxCnt <= KnownAmt.countMinLeadingZeros()) { |
1018 | I.setHasNoUnsignedWrap(); |
1019 | Changed = true; |
1020 | } |
1021 | // If we have more sign bits than maximum shift cnt we have nsw. |
1022 | if (!I.hasNoSignedWrap()) { |
1023 | if (MaxCnt < KnownAmt.countMinSignBits() || |
1024 | MaxCnt < |
1025 | ComputeNumSignBits(Op: I.getOperand(i_nocapture: 0), DL: Q.DL, AC: Q.AC, CxtI: Q.CxtI, DT: Q.DT)) { |
1026 | I.setHasNoSignedWrap(); |
1027 | Changed = true; |
1028 | } |
1029 | } |
1030 | return Changed; |
1031 | } |
1032 | |
1033 | // If we have at least as many trailing zeros as maximum count then we have |
1034 | // exact. |
1035 | Changed = MaxCnt <= KnownAmt.countMinTrailingZeros(); |
1036 | I.setIsExact(Changed); |
1037 | |
1038 | return Changed; |
1039 | } |
1040 | |
1041 | Instruction *InstCombinerImpl::visitShl(BinaryOperator &I) { |
1042 | const SimplifyQuery Q = SQ.getWithInstruction(I: &I); |
1043 | |
1044 | if (Value *V = simplifyShlInst(Op0: I.getOperand(i_nocapture: 0), Op1: I.getOperand(i_nocapture: 1), |
1045 | IsNSW: I.hasNoSignedWrap(), IsNUW: I.hasNoUnsignedWrap(), Q)) |
1046 | return replaceInstUsesWith(I, V); |
1047 | |
1048 | if (Instruction *X = foldVectorBinop(Inst&: I)) |
1049 | return X; |
1050 | |
1051 | if (Instruction *V = commonShiftTransforms(I)) |
1052 | return V; |
1053 | |
1054 | if (Instruction *V = dropRedundantMaskingOfLeftShiftInput(OuterShift: &I, Q, Builder)) |
1055 | return V; |
1056 | |
1057 | Value *Op0 = I.getOperand(i_nocapture: 0), *Op1 = I.getOperand(i_nocapture: 1); |
1058 | Type *Ty = I.getType(); |
1059 | unsigned BitWidth = Ty->getScalarSizeInBits(); |
1060 | |
1061 | const APInt *C; |
1062 | if (match(V: Op1, P: m_APInt(Res&: C))) { |
1063 | unsigned ShAmtC = C->getZExtValue(); |
1064 | |
1065 | // shl (zext X), C --> zext (shl X, C) |
1066 | // This is only valid if X would have zeros shifted out. |
1067 | Value *X; |
1068 | if (match(V: Op0, P: m_OneUse(SubPattern: m_ZExt(Op: m_Value(V&: X))))) { |
1069 | unsigned SrcWidth = X->getType()->getScalarSizeInBits(); |
1070 | if (ShAmtC < SrcWidth && |
1071 | MaskedValueIsZero(V: X, Mask: APInt::getHighBitsSet(numBits: SrcWidth, hiBitsSet: ShAmtC), CxtI: &I)) |
1072 | return new ZExtInst(Builder.CreateShl(LHS: X, RHS: ShAmtC), Ty); |
1073 | } |
1074 | |
1075 | // (X >> C) << C --> X & (-1 << C) |
1076 | if (match(V: Op0, P: m_Shr(L: m_Value(V&: X), R: m_Specific(V: Op1)))) { |
1077 | APInt Mask(APInt::getHighBitsSet(numBits: BitWidth, hiBitsSet: BitWidth - ShAmtC)); |
1078 | return BinaryOperator::CreateAnd(V1: X, V2: ConstantInt::get(Ty, V: Mask)); |
1079 | } |
1080 | |
1081 | const APInt *C1; |
1082 | if (match(V: Op0, P: m_Exact(SubPattern: m_Shr(L: m_Value(V&: X), R: m_APInt(Res&: C1)))) && |
1083 | C1->ult(RHS: BitWidth)) { |
1084 | unsigned ShrAmt = C1->getZExtValue(); |
1085 | if (ShrAmt < ShAmtC) { |
1086 | // If C1 < C: (X >>?,exact C1) << C --> X << (C - C1) |
1087 | Constant *ShiftDiff = ConstantInt::get(Ty, V: ShAmtC - ShrAmt); |
1088 | auto *NewShl = BinaryOperator::CreateShl(V1: X, V2: ShiftDiff); |
1089 | NewShl->setHasNoUnsignedWrap( |
1090 | I.hasNoUnsignedWrap() || |
1091 | (ShrAmt && |
1092 | cast<Instruction>(Val: Op0)->getOpcode() == Instruction::LShr && |
1093 | I.hasNoSignedWrap())); |
1094 | NewShl->setHasNoSignedWrap(I.hasNoSignedWrap()); |
1095 | return NewShl; |
1096 | } |
1097 | if (ShrAmt > ShAmtC) { |
1098 | // If C1 > C: (X >>?exact C1) << C --> X >>?exact (C1 - C) |
1099 | Constant *ShiftDiff = ConstantInt::get(Ty, V: ShrAmt - ShAmtC); |
1100 | auto *NewShr = BinaryOperator::Create( |
1101 | Op: cast<BinaryOperator>(Val: Op0)->getOpcode(), S1: X, S2: ShiftDiff); |
1102 | NewShr->setIsExact(true); |
1103 | return NewShr; |
1104 | } |
1105 | } |
1106 | |
1107 | if (match(V: Op0, P: m_OneUse(SubPattern: m_Shr(L: m_Value(V&: X), R: m_APInt(Res&: C1)))) && |
1108 | C1->ult(RHS: BitWidth)) { |
1109 | unsigned ShrAmt = C1->getZExtValue(); |
1110 | if (ShrAmt < ShAmtC) { |
1111 | // If C1 < C: (X >>? C1) << C --> (X << (C - C1)) & (-1 << C) |
1112 | Constant *ShiftDiff = ConstantInt::get(Ty, V: ShAmtC - ShrAmt); |
1113 | auto *NewShl = BinaryOperator::CreateShl(V1: X, V2: ShiftDiff); |
1114 | NewShl->setHasNoUnsignedWrap( |
1115 | I.hasNoUnsignedWrap() || |
1116 | (ShrAmt && |
1117 | cast<Instruction>(Val: Op0)->getOpcode() == Instruction::LShr && |
1118 | I.hasNoSignedWrap())); |
1119 | NewShl->setHasNoSignedWrap(I.hasNoSignedWrap()); |
1120 | Builder.Insert(I: NewShl); |
1121 | APInt Mask(APInt::getHighBitsSet(numBits: BitWidth, hiBitsSet: BitWidth - ShAmtC)); |
1122 | return BinaryOperator::CreateAnd(V1: NewShl, V2: ConstantInt::get(Ty, V: Mask)); |
1123 | } |
1124 | if (ShrAmt > ShAmtC) { |
1125 | // If C1 > C: (X >>? C1) << C --> (X >>? (C1 - C)) & (-1 << C) |
1126 | Constant *ShiftDiff = ConstantInt::get(Ty, V: ShrAmt - ShAmtC); |
1127 | auto *OldShr = cast<BinaryOperator>(Val: Op0); |
1128 | auto *NewShr = |
1129 | BinaryOperator::Create(Op: OldShr->getOpcode(), S1: X, S2: ShiftDiff); |
1130 | NewShr->setIsExact(OldShr->isExact()); |
1131 | Builder.Insert(I: NewShr); |
1132 | APInt Mask(APInt::getHighBitsSet(numBits: BitWidth, hiBitsSet: BitWidth - ShAmtC)); |
1133 | return BinaryOperator::CreateAnd(V1: NewShr, V2: ConstantInt::get(Ty, V: Mask)); |
1134 | } |
1135 | } |
1136 | |
1137 | // Similar to above, but look through an intermediate trunc instruction. |
1138 | BinaryOperator *Shr; |
1139 | if (match(V: Op0, P: m_OneUse(SubPattern: m_Trunc(Op: m_OneUse(SubPattern: m_BinOp(I&: Shr))))) && |
1140 | match(V: Shr, P: m_Shr(L: m_Value(V&: X), R: m_APInt(Res&: C1)))) { |
1141 | // The larger shift direction survives through the transform. |
1142 | unsigned ShrAmtC = C1->getZExtValue(); |
1143 | unsigned ShDiff = ShrAmtC > ShAmtC ? ShrAmtC - ShAmtC : ShAmtC - ShrAmtC; |
1144 | Constant *ShiftDiffC = ConstantInt::get(Ty: X->getType(), V: ShDiff); |
1145 | auto ShiftOpc = ShrAmtC > ShAmtC ? Shr->getOpcode() : Instruction::Shl; |
1146 | |
1147 | // If C1 > C: |
1148 | // (trunc (X >> C1)) << C --> (trunc (X >> (C1 - C))) && (-1 << C) |
1149 | // If C > C1: |
1150 | // (trunc (X >> C1)) << C --> (trunc (X << (C - C1))) && (-1 << C) |
1151 | Value *NewShift = Builder.CreateBinOp(Opc: ShiftOpc, LHS: X, RHS: ShiftDiffC, Name: "sh.diff" ); |
1152 | Value *Trunc = Builder.CreateTrunc(V: NewShift, DestTy: Ty, Name: "tr.sh.diff" ); |
1153 | APInt Mask(APInt::getHighBitsSet(numBits: BitWidth, hiBitsSet: BitWidth - ShAmtC)); |
1154 | return BinaryOperator::CreateAnd(V1: Trunc, V2: ConstantInt::get(Ty, V: Mask)); |
1155 | } |
1156 | |
1157 | // If we have an opposite shift by the same amount, we may be able to |
1158 | // reorder binops and shifts to eliminate math/logic. |
1159 | auto isSuitableBinOpcode = [](Instruction::BinaryOps BinOpcode) { |
1160 | switch (BinOpcode) { |
1161 | default: |
1162 | return false; |
1163 | case Instruction::Add: |
1164 | case Instruction::And: |
1165 | case Instruction::Or: |
1166 | case Instruction::Xor: |
1167 | case Instruction::Sub: |
1168 | // NOTE: Sub is not commutable and the tranforms below may not be valid |
1169 | // when the shift-right is operand 1 (RHS) of the sub. |
1170 | return true; |
1171 | } |
1172 | }; |
1173 | BinaryOperator *Op0BO; |
1174 | if (match(V: Op0, P: m_OneUse(SubPattern: m_BinOp(I&: Op0BO))) && |
1175 | isSuitableBinOpcode(Op0BO->getOpcode())) { |
1176 | // Commute so shift-right is on LHS of the binop. |
1177 | // (Y bop (X >> C)) << C -> ((X >> C) bop Y) << C |
1178 | // (Y bop ((X >> C) & CC)) << C -> (((X >> C) & CC) bop Y) << C |
1179 | Value *Shr = Op0BO->getOperand(i_nocapture: 0); |
1180 | Value *Y = Op0BO->getOperand(i_nocapture: 1); |
1181 | Value *X; |
1182 | const APInt *CC; |
1183 | if (Op0BO->isCommutative() && Y->hasOneUse() && |
1184 | (match(V: Y, P: m_Shr(L: m_Value(), R: m_Specific(V: Op1))) || |
1185 | match(V: Y, P: m_And(L: m_OneUse(SubPattern: m_Shr(L: m_Value(), R: m_Specific(V: Op1))), |
1186 | R: m_APInt(Res&: CC))))) |
1187 | std::swap(a&: Shr, b&: Y); |
1188 | |
1189 | // ((X >> C) bop Y) << C -> (X bop (Y << C)) & (~0 << C) |
1190 | if (match(V: Shr, P: m_OneUse(SubPattern: m_Shr(L: m_Value(V&: X), R: m_Specific(V: Op1))))) { |
1191 | // Y << C |
1192 | Value *YS = Builder.CreateShl(LHS: Y, RHS: Op1, Name: Op0BO->getName()); |
1193 | // (X bop (Y << C)) |
1194 | Value *B = |
1195 | Builder.CreateBinOp(Opc: Op0BO->getOpcode(), LHS: X, RHS: YS, Name: Shr->getName()); |
1196 | unsigned Op1Val = C->getLimitedValue(Limit: BitWidth); |
1197 | APInt Bits = APInt::getHighBitsSet(numBits: BitWidth, hiBitsSet: BitWidth - Op1Val); |
1198 | Constant *Mask = ConstantInt::get(Ty, V: Bits); |
1199 | return BinaryOperator::CreateAnd(V1: B, V2: Mask); |
1200 | } |
1201 | |
1202 | // (((X >> C) & CC) bop Y) << C -> (X & (CC << C)) bop (Y << C) |
1203 | if (match(V: Shr, |
1204 | P: m_OneUse(SubPattern: m_And(L: m_OneUse(SubPattern: m_Shr(L: m_Value(V&: X), R: m_Specific(V: Op1))), |
1205 | R: m_APInt(Res&: CC))))) { |
1206 | // Y << C |
1207 | Value *YS = Builder.CreateShl(LHS: Y, RHS: Op1, Name: Op0BO->getName()); |
1208 | // X & (CC << C) |
1209 | Value *M = Builder.CreateAnd(LHS: X, RHS: ConstantInt::get(Ty, V: CC->shl(ShiftAmt: *C)), |
1210 | Name: X->getName() + ".mask" ); |
1211 | auto *NewOp = BinaryOperator::Create(Op: Op0BO->getOpcode(), S1: M, S2: YS); |
1212 | if (auto *Disjoint = dyn_cast<PossiblyDisjointInst>(Val: Op0BO); |
1213 | Disjoint && Disjoint->isDisjoint()) |
1214 | cast<PossiblyDisjointInst>(Val: NewOp)->setIsDisjoint(true); |
1215 | return NewOp; |
1216 | } |
1217 | } |
1218 | |
1219 | // (C1 - X) << C --> (C1 << C) - (X << C) |
1220 | if (match(V: Op0, P: m_OneUse(SubPattern: m_Sub(L: m_APInt(Res&: C1), R: m_Value(V&: X))))) { |
1221 | Constant *NewLHS = ConstantInt::get(Ty, V: C1->shl(ShiftAmt: *C)); |
1222 | Value *NewShift = Builder.CreateShl(LHS: X, RHS: Op1); |
1223 | return BinaryOperator::CreateSub(V1: NewLHS, V2: NewShift); |
1224 | } |
1225 | } |
1226 | |
1227 | if (setShiftFlags(I, Q)) |
1228 | return &I; |
1229 | |
1230 | // Transform (x >> y) << y to x & (-1 << y) |
1231 | // Valid for any type of right-shift. |
1232 | Value *X; |
1233 | if (match(V: Op0, P: m_OneUse(SubPattern: m_Shr(L: m_Value(V&: X), R: m_Specific(V: Op1))))) { |
1234 | Constant *AllOnes = ConstantInt::getAllOnesValue(Ty); |
1235 | Value *Mask = Builder.CreateShl(LHS: AllOnes, RHS: Op1); |
1236 | return BinaryOperator::CreateAnd(V1: Mask, V2: X); |
1237 | } |
1238 | |
1239 | // Transform (-1 >> y) << y to -1 << y |
1240 | if (match(V: Op0, P: m_LShr(L: m_AllOnes(), R: m_Specific(V: Op1)))) { |
1241 | Constant *AllOnes = ConstantInt::getAllOnesValue(Ty); |
1242 | return BinaryOperator::CreateShl(V1: AllOnes, V2: Op1); |
1243 | } |
1244 | |
1245 | Constant *C1; |
1246 | if (match(V: Op1, P: m_ImmConstant(C&: C1))) { |
1247 | Constant *C2; |
1248 | Value *X; |
1249 | // (X * C2) << C1 --> X * (C2 << C1) |
1250 | if (match(V: Op0, P: m_Mul(L: m_Value(V&: X), R: m_ImmConstant(C&: C2)))) |
1251 | return BinaryOperator::CreateMul(V1: X, V2: Builder.CreateShl(LHS: C2, RHS: C1)); |
1252 | |
1253 | // shl (zext i1 X), C1 --> select (X, 1 << C1, 0) |
1254 | if (match(V: Op0, P: m_ZExt(Op: m_Value(V&: X))) && X->getType()->isIntOrIntVectorTy(BitWidth: 1)) { |
1255 | auto *NewC = Builder.CreateShl(LHS: ConstantInt::get(Ty, V: 1), RHS: C1); |
1256 | return SelectInst::Create(C: X, S1: NewC, S2: ConstantInt::getNullValue(Ty)); |
1257 | } |
1258 | } |
1259 | |
1260 | if (match(V: Op0, P: m_One())) { |
1261 | // (1 << (C - x)) -> ((1 << C) >> x) if C is bitwidth - 1 |
1262 | if (match(V: Op1, P: m_Sub(L: m_SpecificInt(V: BitWidth - 1), R: m_Value(V&: X)))) |
1263 | return BinaryOperator::CreateLShr( |
1264 | V1: ConstantInt::get(Ty, V: APInt::getSignMask(BitWidth)), V2: X); |
1265 | |
1266 | // Canonicalize "extract lowest set bit" using cttz to and-with-negate: |
1267 | // 1 << (cttz X) --> -X & X |
1268 | if (match(V: Op1, |
1269 | P: m_OneUse(SubPattern: m_Intrinsic<Intrinsic::cttz>(Op0: m_Value(V&: X), Op1: m_Value())))) { |
1270 | Value *NegX = Builder.CreateNeg(V: X, Name: "neg" ); |
1271 | return BinaryOperator::CreateAnd(V1: NegX, V2: X); |
1272 | } |
1273 | } |
1274 | |
1275 | return nullptr; |
1276 | } |
1277 | |
1278 | Instruction *InstCombinerImpl::visitLShr(BinaryOperator &I) { |
1279 | if (Value *V = simplifyLShrInst(Op0: I.getOperand(i_nocapture: 0), Op1: I.getOperand(i_nocapture: 1), IsExact: I.isExact(), |
1280 | Q: SQ.getWithInstruction(I: &I))) |
1281 | return replaceInstUsesWith(I, V); |
1282 | |
1283 | if (Instruction *X = foldVectorBinop(Inst&: I)) |
1284 | return X; |
1285 | |
1286 | if (Instruction *R = commonShiftTransforms(I)) |
1287 | return R; |
1288 | |
1289 | Value *Op0 = I.getOperand(i_nocapture: 0), *Op1 = I.getOperand(i_nocapture: 1); |
1290 | Type *Ty = I.getType(); |
1291 | Value *X; |
1292 | const APInt *C; |
1293 | unsigned BitWidth = Ty->getScalarSizeInBits(); |
1294 | |
1295 | // (iN (~X) u>> (N - 1)) --> zext (X > -1) |
1296 | if (match(V: Op0, P: m_OneUse(SubPattern: m_Not(V: m_Value(V&: X)))) && |
1297 | match(V: Op1, P: m_SpecificIntAllowPoison(V: BitWidth - 1))) |
1298 | return new ZExtInst(Builder.CreateIsNotNeg(Arg: X, Name: "isnotneg" ), Ty); |
1299 | |
1300 | // ((X << nuw Z) sub nuw Y) >>u exact Z --> X sub nuw (Y >>u exact Z) |
1301 | Value *Y; |
1302 | if (I.isExact() && |
1303 | match(V: Op0, P: m_OneUse(SubPattern: m_NUWSub(L: m_NUWShl(L: m_Value(V&: X), R: m_Specific(V: Op1)), |
1304 | R: m_Value(V&: Y))))) { |
1305 | Value *NewLshr = Builder.CreateLShr(LHS: Y, RHS: Op1, Name: "" , /*isExact=*/true); |
1306 | auto *NewSub = BinaryOperator::CreateNUWSub(V1: X, V2: NewLshr); |
1307 | NewSub->setHasNoSignedWrap( |
1308 | cast<OverflowingBinaryOperator>(Val: Op0)->hasNoSignedWrap()); |
1309 | return NewSub; |
1310 | } |
1311 | |
1312 | // Fold (X + Y) / 2 --> (X & Y) iff (X u<= 1) && (Y u<= 1) |
1313 | if (match(V: Op0, P: m_Add(L: m_Value(V&: X), R: m_Value(V&: Y))) && match(V: Op1, P: m_One()) && |
1314 | computeKnownBits(V: X, CxtI: &I).countMaxActiveBits() <= 1 && |
1315 | computeKnownBits(V: Y, CxtI: &I).countMaxActiveBits() <= 1) |
1316 | return BinaryOperator::CreateAnd(V1: X, V2: Y); |
1317 | |
1318 | // (sub nuw X, (Y << nuw Z)) >>u exact Z --> (X >>u exact Z) sub nuw Y |
1319 | if (I.isExact() && |
1320 | match(V: Op0, P: m_OneUse(SubPattern: m_NUWSub(L: m_Value(V&: X), |
1321 | R: m_NUWShl(L: m_Value(V&: Y), R: m_Specific(V: Op1)))))) { |
1322 | Value *NewLshr = Builder.CreateLShr(LHS: X, RHS: Op1, Name: "" , /*isExact=*/true); |
1323 | auto *NewSub = BinaryOperator::CreateNUWSub(V1: NewLshr, V2: Y); |
1324 | NewSub->setHasNoSignedWrap( |
1325 | cast<OverflowingBinaryOperator>(Val: Op0)->hasNoSignedWrap()); |
1326 | return NewSub; |
1327 | } |
1328 | |
1329 | auto isSuitableBinOpcode = [](Instruction::BinaryOps BinOpcode) { |
1330 | switch (BinOpcode) { |
1331 | default: |
1332 | return false; |
1333 | case Instruction::Add: |
1334 | case Instruction::And: |
1335 | case Instruction::Or: |
1336 | case Instruction::Xor: |
1337 | // Sub is handled separately. |
1338 | return true; |
1339 | } |
1340 | }; |
1341 | |
1342 | // If both the binop and the shift are nuw, then: |
1343 | // ((X << nuw Z) binop nuw Y) >>u Z --> X binop nuw (Y >>u Z) |
1344 | if (match(V: Op0, P: m_OneUse(SubPattern: m_c_BinOp(L: m_NUWShl(L: m_Value(V&: X), R: m_Specific(V: Op1)), |
1345 | R: m_Value(V&: Y))))) { |
1346 | BinaryOperator *Op0OB = cast<BinaryOperator>(Val: Op0); |
1347 | if (isSuitableBinOpcode(Op0OB->getOpcode())) { |
1348 | if (auto *OBO = dyn_cast<OverflowingBinaryOperator>(Val: Op0); |
1349 | !OBO || OBO->hasNoUnsignedWrap()) { |
1350 | Value *NewLshr = Builder.CreateLShr( |
1351 | LHS: Y, RHS: Op1, Name: "" , isExact: I.isExact() && Op0OB->getOpcode() != Instruction::And); |
1352 | auto *NewBinOp = BinaryOperator::Create(Op: Op0OB->getOpcode(), S1: NewLshr, S2: X); |
1353 | if (OBO) { |
1354 | NewBinOp->setHasNoUnsignedWrap(true); |
1355 | NewBinOp->setHasNoSignedWrap(OBO->hasNoSignedWrap()); |
1356 | } else if (auto *Disjoint = dyn_cast<PossiblyDisjointInst>(Val: Op0)) { |
1357 | cast<PossiblyDisjointInst>(Val: NewBinOp)->setIsDisjoint( |
1358 | Disjoint->isDisjoint()); |
1359 | } |
1360 | return NewBinOp; |
1361 | } |
1362 | } |
1363 | } |
1364 | |
1365 | if (match(V: Op1, P: m_APInt(Res&: C))) { |
1366 | unsigned ShAmtC = C->getZExtValue(); |
1367 | auto *II = dyn_cast<IntrinsicInst>(Val: Op0); |
1368 | if (II && isPowerOf2_32(Value: BitWidth) && Log2_32(Value: BitWidth) == ShAmtC && |
1369 | (II->getIntrinsicID() == Intrinsic::ctlz || |
1370 | II->getIntrinsicID() == Intrinsic::cttz || |
1371 | II->getIntrinsicID() == Intrinsic::ctpop)) { |
1372 | // ctlz.i32(x)>>5 --> zext(x == 0) |
1373 | // cttz.i32(x)>>5 --> zext(x == 0) |
1374 | // ctpop.i32(x)>>5 --> zext(x == -1) |
1375 | bool IsPop = II->getIntrinsicID() == Intrinsic::ctpop; |
1376 | Constant *RHS = ConstantInt::getSigned(Ty, V: IsPop ? -1 : 0); |
1377 | Value *Cmp = Builder.CreateICmpEQ(LHS: II->getArgOperand(i: 0), RHS); |
1378 | return new ZExtInst(Cmp, Ty); |
1379 | } |
1380 | |
1381 | const APInt *C1; |
1382 | if (match(V: Op0, P: m_Shl(L: m_Value(V&: X), R: m_APInt(Res&: C1))) && C1->ult(RHS: BitWidth)) { |
1383 | if (C1->ult(RHS: ShAmtC)) { |
1384 | unsigned ShlAmtC = C1->getZExtValue(); |
1385 | Constant *ShiftDiff = ConstantInt::get(Ty, V: ShAmtC - ShlAmtC); |
1386 | if (cast<BinaryOperator>(Val: Op0)->hasNoUnsignedWrap()) { |
1387 | // (X <<nuw C1) >>u C --> X >>u (C - C1) |
1388 | auto *NewLShr = BinaryOperator::CreateLShr(V1: X, V2: ShiftDiff); |
1389 | NewLShr->setIsExact(I.isExact()); |
1390 | return NewLShr; |
1391 | } |
1392 | if (Op0->hasOneUse()) { |
1393 | // (X << C1) >>u C --> (X >>u (C - C1)) & (-1 >> C) |
1394 | Value *NewLShr = Builder.CreateLShr(LHS: X, RHS: ShiftDiff, Name: "" , isExact: I.isExact()); |
1395 | APInt Mask(APInt::getLowBitsSet(numBits: BitWidth, loBitsSet: BitWidth - ShAmtC)); |
1396 | return BinaryOperator::CreateAnd(V1: NewLShr, V2: ConstantInt::get(Ty, V: Mask)); |
1397 | } |
1398 | } else if (C1->ugt(RHS: ShAmtC)) { |
1399 | unsigned ShlAmtC = C1->getZExtValue(); |
1400 | Constant *ShiftDiff = ConstantInt::get(Ty, V: ShlAmtC - ShAmtC); |
1401 | if (cast<BinaryOperator>(Val: Op0)->hasNoUnsignedWrap()) { |
1402 | // (X <<nuw C1) >>u C --> X <<nuw/nsw (C1 - C) |
1403 | auto *NewShl = BinaryOperator::CreateShl(V1: X, V2: ShiftDiff); |
1404 | NewShl->setHasNoUnsignedWrap(true); |
1405 | NewShl->setHasNoSignedWrap(ShAmtC > 0); |
1406 | return NewShl; |
1407 | } |
1408 | if (Op0->hasOneUse()) { |
1409 | // (X << C1) >>u C --> X << (C1 - C) & (-1 >> C) |
1410 | Value *NewShl = Builder.CreateShl(LHS: X, RHS: ShiftDiff); |
1411 | APInt Mask(APInt::getLowBitsSet(numBits: BitWidth, loBitsSet: BitWidth - ShAmtC)); |
1412 | return BinaryOperator::CreateAnd(V1: NewShl, V2: ConstantInt::get(Ty, V: Mask)); |
1413 | } |
1414 | } else { |
1415 | assert(*C1 == ShAmtC); |
1416 | // (X << C) >>u C --> X & (-1 >>u C) |
1417 | APInt Mask(APInt::getLowBitsSet(numBits: BitWidth, loBitsSet: BitWidth - ShAmtC)); |
1418 | return BinaryOperator::CreateAnd(V1: X, V2: ConstantInt::get(Ty, V: Mask)); |
1419 | } |
1420 | } |
1421 | |
1422 | // ((X << C) + Y) >>u C --> (X + (Y >>u C)) & (-1 >>u C) |
1423 | // TODO: Consolidate with the more general transform that starts from shl |
1424 | // (the shifts are in the opposite order). |
1425 | if (match(V: Op0, |
1426 | P: m_OneUse(SubPattern: m_c_Add(L: m_OneUse(SubPattern: m_Shl(L: m_Value(V&: X), R: m_Specific(V: Op1))), |
1427 | R: m_Value(V&: Y))))) { |
1428 | Value *NewLshr = Builder.CreateLShr(LHS: Y, RHS: Op1); |
1429 | Value *NewAdd = Builder.CreateAdd(LHS: NewLshr, RHS: X); |
1430 | unsigned Op1Val = C->getLimitedValue(Limit: BitWidth); |
1431 | APInt Bits = APInt::getLowBitsSet(numBits: BitWidth, loBitsSet: BitWidth - Op1Val); |
1432 | Constant *Mask = ConstantInt::get(Ty, V: Bits); |
1433 | return BinaryOperator::CreateAnd(V1: NewAdd, V2: Mask); |
1434 | } |
1435 | |
1436 | if (match(V: Op0, P: m_OneUse(SubPattern: m_ZExt(Op: m_Value(V&: X)))) && |
1437 | (!Ty->isIntegerTy() || shouldChangeType(From: Ty, To: X->getType()))) { |
1438 | assert(ShAmtC < X->getType()->getScalarSizeInBits() && |
1439 | "Big shift not simplified to zero?" ); |
1440 | // lshr (zext iM X to iN), C --> zext (lshr X, C) to iN |
1441 | Value *NewLShr = Builder.CreateLShr(LHS: X, RHS: ShAmtC); |
1442 | return new ZExtInst(NewLShr, Ty); |
1443 | } |
1444 | |
1445 | if (match(V: Op0, P: m_SExt(Op: m_Value(V&: X)))) { |
1446 | unsigned SrcTyBitWidth = X->getType()->getScalarSizeInBits(); |
1447 | // lshr (sext i1 X to iN), C --> select (X, -1 >> C, 0) |
1448 | if (SrcTyBitWidth == 1) { |
1449 | auto *NewC = ConstantInt::get( |
1450 | Ty, V: APInt::getLowBitsSet(numBits: BitWidth, loBitsSet: BitWidth - ShAmtC)); |
1451 | return SelectInst::Create(C: X, S1: NewC, S2: ConstantInt::getNullValue(Ty)); |
1452 | } |
1453 | |
1454 | if ((!Ty->isIntegerTy() || shouldChangeType(From: Ty, To: X->getType())) && |
1455 | Op0->hasOneUse()) { |
1456 | // Are we moving the sign bit to the low bit and widening with high |
1457 | // zeros? lshr (sext iM X to iN), N-1 --> zext (lshr X, M-1) to iN |
1458 | if (ShAmtC == BitWidth - 1) { |
1459 | Value *NewLShr = Builder.CreateLShr(LHS: X, RHS: SrcTyBitWidth - 1); |
1460 | return new ZExtInst(NewLShr, Ty); |
1461 | } |
1462 | |
1463 | // lshr (sext iM X to iN), N-M --> zext (ashr X, min(N-M, M-1)) to iN |
1464 | if (ShAmtC == BitWidth - SrcTyBitWidth) { |
1465 | // The new shift amount can't be more than the narrow source type. |
1466 | unsigned NewShAmt = std::min(a: ShAmtC, b: SrcTyBitWidth - 1); |
1467 | Value *AShr = Builder.CreateAShr(LHS: X, RHS: NewShAmt); |
1468 | return new ZExtInst(AShr, Ty); |
1469 | } |
1470 | } |
1471 | } |
1472 | |
1473 | if (ShAmtC == BitWidth - 1) { |
1474 | // lshr i32 or(X,-X), 31 --> zext (X != 0) |
1475 | if (match(V: Op0, P: m_OneUse(SubPattern: m_c_Or(L: m_Neg(V: m_Value(V&: X)), R: m_Deferred(V: X))))) |
1476 | return new ZExtInst(Builder.CreateIsNotNull(Arg: X), Ty); |
1477 | |
1478 | // lshr i32 (X -nsw Y), 31 --> zext (X < Y) |
1479 | if (match(V: Op0, P: m_OneUse(SubPattern: m_NSWSub(L: m_Value(V&: X), R: m_Value(V&: Y))))) |
1480 | return new ZExtInst(Builder.CreateICmpSLT(LHS: X, RHS: Y), Ty); |
1481 | |
1482 | // Check if a number is negative and odd: |
1483 | // lshr i32 (srem X, 2), 31 --> and (X >> 31), X |
1484 | if (match(V: Op0, P: m_OneUse(SubPattern: m_SRem(L: m_Value(V&: X), R: m_SpecificInt(V: 2))))) { |
1485 | Value *Signbit = Builder.CreateLShr(LHS: X, RHS: ShAmtC); |
1486 | return BinaryOperator::CreateAnd(V1: Signbit, V2: X); |
1487 | } |
1488 | |
1489 | // lshr iN (X - 1) & ~X, N-1 --> zext (X == 0) |
1490 | if (match(V: Op0, P: m_OneUse(SubPattern: m_c_And(L: m_Add(L: m_Value(V&: X), R: m_AllOnes()), |
1491 | R: m_Not(V: m_Deferred(V: X)))))) |
1492 | return new ZExtInst(Builder.CreateIsNull(Arg: X), Ty); |
1493 | } |
1494 | |
1495 | Instruction *TruncSrc; |
1496 | if (match(V: Op0, P: m_OneUse(SubPattern: m_Trunc(Op: m_Instruction(I&: TruncSrc)))) && |
1497 | match(V: TruncSrc, P: m_LShr(L: m_Value(V&: X), R: m_APInt(Res&: C1)))) { |
1498 | unsigned SrcWidth = X->getType()->getScalarSizeInBits(); |
1499 | unsigned AmtSum = ShAmtC + C1->getZExtValue(); |
1500 | |
1501 | // If the combined shift fits in the source width: |
1502 | // (trunc (X >>u C1)) >>u C --> and (trunc (X >>u (C1 + C)), MaskC |
1503 | // |
1504 | // If the first shift covers the number of bits truncated, then the |
1505 | // mask instruction is eliminated (and so the use check is relaxed). |
1506 | if (AmtSum < SrcWidth && |
1507 | (TruncSrc->hasOneUse() || C1->uge(RHS: SrcWidth - BitWidth))) { |
1508 | Value *SumShift = Builder.CreateLShr(LHS: X, RHS: AmtSum, Name: "sum.shift" ); |
1509 | Value *Trunc = Builder.CreateTrunc(V: SumShift, DestTy: Ty, Name: I.getName()); |
1510 | |
1511 | // If the first shift does not cover the number of bits truncated, then |
1512 | // we require a mask to get rid of high bits in the result. |
1513 | APInt MaskC = APInt::getAllOnes(numBits: BitWidth).lshr(shiftAmt: ShAmtC); |
1514 | return BinaryOperator::CreateAnd(V1: Trunc, V2: ConstantInt::get(Ty, V: MaskC)); |
1515 | } |
1516 | } |
1517 | |
1518 | const APInt *MulC; |
1519 | if (match(V: Op0, P: m_NUWMul(L: m_Value(V&: X), R: m_APInt(Res&: MulC)))) { |
1520 | if (BitWidth > 2 && (*MulC - 1).isPowerOf2() && |
1521 | MulC->logBase2() == ShAmtC) { |
1522 | // Look for a "splat" mul pattern - it replicates bits across each half |
1523 | // of a value, so a right shift simplifies back to just X: |
1524 | // lshr i[2N] (mul nuw X, (2^N)+1), N --> X |
1525 | if (ShAmtC * 2 == BitWidth) |
1526 | return replaceInstUsesWith(I, V: X); |
1527 | |
1528 | // lshr (mul nuw (X, 2^N + 1)), N -> add nuw (X, lshr(X, N)) |
1529 | if (Op0->hasOneUse()) { |
1530 | auto *NewAdd = BinaryOperator::CreateNUWAdd( |
1531 | V1: X, V2: Builder.CreateLShr(LHS: X, RHS: ConstantInt::get(Ty, V: ShAmtC), Name: "" , |
1532 | isExact: I.isExact())); |
1533 | NewAdd->setHasNoSignedWrap( |
1534 | cast<OverflowingBinaryOperator>(Val: Op0)->hasNoSignedWrap()); |
1535 | return NewAdd; |
1536 | } |
1537 | } |
1538 | |
1539 | // The one-use check is not strictly necessary, but codegen may not be |
1540 | // able to invert the transform and perf may suffer with an extra mul |
1541 | // instruction. |
1542 | if (Op0->hasOneUse()) { |
1543 | APInt NewMulC = MulC->lshr(shiftAmt: ShAmtC); |
1544 | // if c is divisible by (1 << ShAmtC): |
1545 | // lshr (mul nuw x, MulC), ShAmtC -> mul nuw nsw x, (MulC >> ShAmtC) |
1546 | if (MulC->eq(RHS: NewMulC.shl(shiftAmt: ShAmtC))) { |
1547 | auto *NewMul = |
1548 | BinaryOperator::CreateNUWMul(V1: X, V2: ConstantInt::get(Ty, V: NewMulC)); |
1549 | assert(ShAmtC != 0 && |
1550 | "lshr X, 0 should be handled by simplifyLShrInst." ); |
1551 | NewMul->setHasNoSignedWrap(true); |
1552 | return NewMul; |
1553 | } |
1554 | } |
1555 | } |
1556 | |
1557 | // lshr (mul nsw (X, 2^N + 1)), N -> add nsw (X, lshr(X, N)) |
1558 | if (match(V: Op0, P: m_OneUse(SubPattern: m_NSWMul(L: m_Value(V&: X), R: m_APInt(Res&: MulC))))) { |
1559 | if (BitWidth > 2 && (*MulC - 1).isPowerOf2() && |
1560 | MulC->logBase2() == ShAmtC) { |
1561 | return BinaryOperator::CreateNSWAdd( |
1562 | V1: X, V2: Builder.CreateLShr(LHS: X, RHS: ConstantInt::get(Ty, V: ShAmtC), Name: "" , |
1563 | isExact: I.isExact())); |
1564 | } |
1565 | } |
1566 | |
1567 | // Try to narrow bswap. |
1568 | // In the case where the shift amount equals the bitwidth difference, the |
1569 | // shift is eliminated. |
1570 | if (match(V: Op0, P: m_OneUse(SubPattern: m_Intrinsic<Intrinsic::bswap>( |
1571 | Op0: m_OneUse(SubPattern: m_ZExt(Op: m_Value(V&: X))))))) { |
1572 | unsigned SrcWidth = X->getType()->getScalarSizeInBits(); |
1573 | unsigned WidthDiff = BitWidth - SrcWidth; |
1574 | if (SrcWidth % 16 == 0) { |
1575 | Value *NarrowSwap = Builder.CreateUnaryIntrinsic(ID: Intrinsic::bswap, V: X); |
1576 | if (ShAmtC >= WidthDiff) { |
1577 | // (bswap (zext X)) >> C --> zext (bswap X >> C') |
1578 | Value *NewShift = Builder.CreateLShr(LHS: NarrowSwap, RHS: ShAmtC - WidthDiff); |
1579 | return new ZExtInst(NewShift, Ty); |
1580 | } else { |
1581 | // (bswap (zext X)) >> C --> (zext (bswap X)) << C' |
1582 | Value *NewZExt = Builder.CreateZExt(V: NarrowSwap, DestTy: Ty); |
1583 | Constant *ShiftDiff = ConstantInt::get(Ty, V: WidthDiff - ShAmtC); |
1584 | return BinaryOperator::CreateShl(V1: NewZExt, V2: ShiftDiff); |
1585 | } |
1586 | } |
1587 | } |
1588 | |
1589 | // Reduce add-carry of bools to logic: |
1590 | // ((zext BoolX) + (zext BoolY)) >> 1 --> zext (BoolX && BoolY) |
1591 | Value *BoolX, *BoolY; |
1592 | if (ShAmtC == 1 && match(V: Op0, P: m_Add(L: m_Value(V&: X), R: m_Value(V&: Y))) && |
1593 | match(V: X, P: m_ZExt(Op: m_Value(V&: BoolX))) && match(V: Y, P: m_ZExt(Op: m_Value(V&: BoolY))) && |
1594 | BoolX->getType()->isIntOrIntVectorTy(BitWidth: 1) && |
1595 | BoolY->getType()->isIntOrIntVectorTy(BitWidth: 1) && |
1596 | (X->hasOneUse() || Y->hasOneUse() || Op0->hasOneUse())) { |
1597 | Value *And = Builder.CreateAnd(LHS: BoolX, RHS: BoolY); |
1598 | return new ZExtInst(And, Ty); |
1599 | } |
1600 | } |
1601 | |
1602 | const SimplifyQuery Q = SQ.getWithInstruction(I: &I); |
1603 | if (setShiftFlags(I, Q)) |
1604 | return &I; |
1605 | |
1606 | // Transform (x << y) >> y to x & (-1 >> y) |
1607 | if (match(V: Op0, P: m_OneUse(SubPattern: m_Shl(L: m_Value(V&: X), R: m_Specific(V: Op1))))) { |
1608 | Constant *AllOnes = ConstantInt::getAllOnesValue(Ty); |
1609 | Value *Mask = Builder.CreateLShr(LHS: AllOnes, RHS: Op1); |
1610 | return BinaryOperator::CreateAnd(V1: Mask, V2: X); |
1611 | } |
1612 | |
1613 | // Transform (-1 << y) >> y to -1 >> y |
1614 | if (match(V: Op0, P: m_Shl(L: m_AllOnes(), R: m_Specific(V: Op1)))) { |
1615 | Constant *AllOnes = ConstantInt::getAllOnesValue(Ty); |
1616 | return BinaryOperator::CreateLShr(V1: AllOnes, V2: Op1); |
1617 | } |
1618 | |
1619 | if (Instruction *Overflow = foldLShrOverflowBit(I)) |
1620 | return Overflow; |
1621 | |
1622 | // Transform ((pow2 << x) >> cttz(pow2 << y)) -> ((1 << x) >> y) |
1623 | Value *Shl0_Op0, *Shl0_Op1, *Shl1_Op1; |
1624 | BinaryOperator *Shl1; |
1625 | if (match(V: Op0, P: m_Shl(L: m_Value(V&: Shl0_Op0), R: m_Value(V&: Shl0_Op1))) && |
1626 | match(V: Op1, P: m_Intrinsic<Intrinsic::cttz>(Op0: m_BinOp(I&: Shl1))) && |
1627 | match(V: Shl1, P: m_Shl(L: m_Specific(V: Shl0_Op0), R: m_Value(V&: Shl1_Op1))) && |
1628 | isKnownToBeAPowerOfTwo(V: Shl0_Op0, /*OrZero=*/true, CxtI: &I)) { |
1629 | auto *Shl0 = cast<BinaryOperator>(Val: Op0); |
1630 | bool HasNUW = Shl0->hasNoUnsignedWrap() && Shl1->hasNoUnsignedWrap(); |
1631 | bool HasNSW = Shl0->hasNoSignedWrap() && Shl1->hasNoSignedWrap(); |
1632 | if (HasNUW || HasNSW) { |
1633 | Value *NewShl = Builder.CreateShl(LHS: ConstantInt::get(Ty: Shl1->getType(), V: 1), |
1634 | RHS: Shl0_Op1, Name: "" , HasNUW, HasNSW); |
1635 | return BinaryOperator::CreateLShr(V1: NewShl, V2: Shl1_Op1); |
1636 | } |
1637 | } |
1638 | return nullptr; |
1639 | } |
1640 | |
1641 | Instruction * |
1642 | InstCombinerImpl::( |
1643 | BinaryOperator &OldAShr) { |
1644 | assert(OldAShr.getOpcode() == Instruction::AShr && |
1645 | "Must be called with arithmetic right-shift instruction only." ); |
1646 | |
1647 | // Check that constant C is a splat of the element-wise bitwidth of V. |
1648 | auto BitWidthSplat = [](Constant *C, Value *V) { |
1649 | return match( |
1650 | V: C, P: m_SpecificInt_ICMP(Predicate: ICmpInst::Predicate::ICMP_EQ, |
1651 | Threshold: APInt(C->getType()->getScalarSizeInBits(), |
1652 | V->getType()->getScalarSizeInBits()))); |
1653 | }; |
1654 | |
1655 | // It should look like variable-length sign-extension on the outside: |
1656 | // (Val << (bitwidth(Val)-Nbits)) a>> (bitwidth(Val)-Nbits) |
1657 | Value *NBits; |
1658 | Instruction *MaybeTrunc; |
1659 | Constant *C1, *C2; |
1660 | if (!match(V: &OldAShr, |
1661 | P: m_AShr(L: m_Shl(L: m_Instruction(I&: MaybeTrunc), |
1662 | R: m_ZExtOrSelf(Op: m_Sub(L: m_Constant(C&: C1), |
1663 | R: m_ZExtOrSelf(Op: m_Value(V&: NBits))))), |
1664 | R: m_ZExtOrSelf(Op: m_Sub(L: m_Constant(C&: C2), |
1665 | R: m_ZExtOrSelf(Op: m_Deferred(V: NBits)))))) || |
1666 | !BitWidthSplat(C1, &OldAShr) || !BitWidthSplat(C2, &OldAShr)) |
1667 | return nullptr; |
1668 | |
1669 | // There may or may not be a truncation after outer two shifts. |
1670 | Instruction *; |
1671 | match(V: MaybeTrunc, P: m_TruncOrSelf(Op: m_Instruction(I&: HighBitExtract))); |
1672 | bool HadTrunc = MaybeTrunc != HighBitExtract; |
1673 | |
1674 | // And finally, the innermost part of the pattern must be a right-shift. |
1675 | Value *X, *NumLowBitsToSkip; |
1676 | if (!match(V: HighBitExtract, P: m_Shr(L: m_Value(V&: X), R: m_Value(V&: NumLowBitsToSkip)))) |
1677 | return nullptr; |
1678 | |
1679 | // Said right-shift must extract high NBits bits - C0 must be it's bitwidth. |
1680 | Constant *C0; |
1681 | if (!match(V: NumLowBitsToSkip, |
1682 | P: m_ZExtOrSelf( |
1683 | Op: m_Sub(L: m_Constant(C&: C0), R: m_ZExtOrSelf(Op: m_Specific(V: NBits))))) || |
1684 | !BitWidthSplat(C0, HighBitExtract)) |
1685 | return nullptr; |
1686 | |
1687 | // Since the NBits is identical for all shifts, if the outermost and |
1688 | // innermost shifts are identical, then outermost shifts are redundant. |
1689 | // If we had truncation, do keep it though. |
1690 | if (HighBitExtract->getOpcode() == OldAShr.getOpcode()) |
1691 | return replaceInstUsesWith(I&: OldAShr, V: MaybeTrunc); |
1692 | |
1693 | // Else, if there was a truncation, then we need to ensure that one |
1694 | // instruction will go away. |
1695 | if (HadTrunc && !match(V: &OldAShr, P: m_c_BinOp(L: m_OneUse(SubPattern: m_Value()), R: m_Value()))) |
1696 | return nullptr; |
1697 | |
1698 | // Finally, bypass two innermost shifts, and perform the outermost shift on |
1699 | // the operands of the innermost shift. |
1700 | Instruction *NewAShr = |
1701 | BinaryOperator::Create(Op: OldAShr.getOpcode(), S1: X, S2: NumLowBitsToSkip); |
1702 | NewAShr->copyIRFlags(V: HighBitExtract); // We can preserve 'exact'-ness. |
1703 | if (!HadTrunc) |
1704 | return NewAShr; |
1705 | |
1706 | Builder.Insert(I: NewAShr); |
1707 | return TruncInst::CreateTruncOrBitCast(S: NewAShr, Ty: OldAShr.getType()); |
1708 | } |
1709 | |
1710 | Instruction *InstCombinerImpl::visitAShr(BinaryOperator &I) { |
1711 | if (Value *V = simplifyAShrInst(Op0: I.getOperand(i_nocapture: 0), Op1: I.getOperand(i_nocapture: 1), IsExact: I.isExact(), |
1712 | Q: SQ.getWithInstruction(I: &I))) |
1713 | return replaceInstUsesWith(I, V); |
1714 | |
1715 | if (Instruction *X = foldVectorBinop(Inst&: I)) |
1716 | return X; |
1717 | |
1718 | if (Instruction *R = commonShiftTransforms(I)) |
1719 | return R; |
1720 | |
1721 | Value *Op0 = I.getOperand(i_nocapture: 0), *Op1 = I.getOperand(i_nocapture: 1); |
1722 | Type *Ty = I.getType(); |
1723 | unsigned BitWidth = Ty->getScalarSizeInBits(); |
1724 | const APInt *ShAmtAPInt; |
1725 | if (match(V: Op1, P: m_APInt(Res&: ShAmtAPInt)) && ShAmtAPInt->ult(RHS: BitWidth)) { |
1726 | unsigned ShAmt = ShAmtAPInt->getZExtValue(); |
1727 | |
1728 | // If the shift amount equals the difference in width of the destination |
1729 | // and source scalar types: |
1730 | // ashr (shl (zext X), C), C --> sext X |
1731 | Value *X; |
1732 | if (match(V: Op0, P: m_Shl(L: m_ZExt(Op: m_Value(V&: X)), R: m_Specific(V: Op1))) && |
1733 | ShAmt == BitWidth - X->getType()->getScalarSizeInBits()) |
1734 | return new SExtInst(X, Ty); |
1735 | |
1736 | // We can't handle (X << C1) >>s C2. It shifts arbitrary bits in. However, |
1737 | // we can handle (X <<nsw C1) >>s C2 since it only shifts in sign bits. |
1738 | const APInt *ShOp1; |
1739 | if (match(V: Op0, P: m_NSWShl(L: m_Value(V&: X), R: m_APInt(Res&: ShOp1))) && |
1740 | ShOp1->ult(RHS: BitWidth)) { |
1741 | unsigned ShlAmt = ShOp1->getZExtValue(); |
1742 | if (ShlAmt < ShAmt) { |
1743 | // (X <<nsw C1) >>s C2 --> X >>s (C2 - C1) |
1744 | Constant *ShiftDiff = ConstantInt::get(Ty, V: ShAmt - ShlAmt); |
1745 | auto *NewAShr = BinaryOperator::CreateAShr(V1: X, V2: ShiftDiff); |
1746 | NewAShr->setIsExact(I.isExact()); |
1747 | return NewAShr; |
1748 | } |
1749 | if (ShlAmt > ShAmt) { |
1750 | // (X <<nsw C1) >>s C2 --> X <<nsw (C1 - C2) |
1751 | Constant *ShiftDiff = ConstantInt::get(Ty, V: ShlAmt - ShAmt); |
1752 | auto *NewShl = BinaryOperator::Create(Op: Instruction::Shl, S1: X, S2: ShiftDiff); |
1753 | NewShl->setHasNoSignedWrap(true); |
1754 | return NewShl; |
1755 | } |
1756 | } |
1757 | |
1758 | if (match(V: Op0, P: m_AShr(L: m_Value(V&: X), R: m_APInt(Res&: ShOp1))) && |
1759 | ShOp1->ult(RHS: BitWidth)) { |
1760 | unsigned AmtSum = ShAmt + ShOp1->getZExtValue(); |
1761 | // Oversized arithmetic shifts replicate the sign bit. |
1762 | AmtSum = std::min(a: AmtSum, b: BitWidth - 1); |
1763 | // (X >>s C1) >>s C2 --> X >>s (C1 + C2) |
1764 | return BinaryOperator::CreateAShr(V1: X, V2: ConstantInt::get(Ty, V: AmtSum)); |
1765 | } |
1766 | |
1767 | if (match(V: Op0, P: m_OneUse(SubPattern: m_SExt(Op: m_Value(V&: X)))) && |
1768 | (Ty->isVectorTy() || shouldChangeType(From: Ty, To: X->getType()))) { |
1769 | // ashr (sext X), C --> sext (ashr X, C') |
1770 | Type *SrcTy = X->getType(); |
1771 | ShAmt = std::min(a: ShAmt, b: SrcTy->getScalarSizeInBits() - 1); |
1772 | Value *NewSh = Builder.CreateAShr(LHS: X, RHS: ConstantInt::get(Ty: SrcTy, V: ShAmt)); |
1773 | return new SExtInst(NewSh, Ty); |
1774 | } |
1775 | |
1776 | if (ShAmt == BitWidth - 1) { |
1777 | // ashr i32 or(X,-X), 31 --> sext (X != 0) |
1778 | if (match(V: Op0, P: m_OneUse(SubPattern: m_c_Or(L: m_Neg(V: m_Value(V&: X)), R: m_Deferred(V: X))))) |
1779 | return new SExtInst(Builder.CreateIsNotNull(Arg: X), Ty); |
1780 | |
1781 | // ashr i32 (X -nsw Y), 31 --> sext (X < Y) |
1782 | Value *Y; |
1783 | if (match(V: Op0, P: m_OneUse(SubPattern: m_NSWSub(L: m_Value(V&: X), R: m_Value(V&: Y))))) |
1784 | return new SExtInst(Builder.CreateICmpSLT(LHS: X, RHS: Y), Ty); |
1785 | |
1786 | // ashr iN (X - 1) & ~X, N-1 --> sext (X == 0) |
1787 | if (match(V: Op0, P: m_OneUse(SubPattern: m_c_And(L: m_Add(L: m_Value(V&: X), R: m_AllOnes()), |
1788 | R: m_Not(V: m_Deferred(V: X)))))) |
1789 | return new SExtInst(Builder.CreateIsNull(Arg: X), Ty); |
1790 | } |
1791 | |
1792 | const APInt *MulC; |
1793 | if (match(V: Op0, P: m_OneUse(SubPattern: m_NSWMul(L: m_Value(V&: X), R: m_APInt(Res&: MulC)))) && |
1794 | (BitWidth > 2 && (*MulC - 1).isPowerOf2() && |
1795 | MulC->logBase2() == ShAmt && |
1796 | (ShAmt < BitWidth - 1))) /* Minus 1 for the sign bit */ { |
1797 | |
1798 | // ashr (mul nsw (X, 2^N + 1)), N -> add nsw (X, ashr(X, N)) |
1799 | auto *NewAdd = BinaryOperator::CreateNSWAdd( |
1800 | V1: X, |
1801 | V2: Builder.CreateAShr(LHS: X, RHS: ConstantInt::get(Ty, V: ShAmt), Name: "" , isExact: I.isExact())); |
1802 | NewAdd->setHasNoUnsignedWrap( |
1803 | cast<OverflowingBinaryOperator>(Val: Op0)->hasNoUnsignedWrap()); |
1804 | return NewAdd; |
1805 | } |
1806 | } |
1807 | |
1808 | const SimplifyQuery Q = SQ.getWithInstruction(I: &I); |
1809 | if (setShiftFlags(I, Q)) |
1810 | return &I; |
1811 | |
1812 | // Prefer `-(x & 1)` over `(x << (bitwidth(x)-1)) a>> (bitwidth(x)-1)` |
1813 | // as the pattern to splat the lowest bit. |
1814 | // FIXME: iff X is already masked, we don't need the one-use check. |
1815 | Value *X; |
1816 | if (match(V: Op1, P: m_SpecificIntAllowPoison(V: BitWidth - 1)) && |
1817 | match(V: Op0, P: m_OneUse(SubPattern: m_Shl(L: m_Value(V&: X), |
1818 | R: m_SpecificIntAllowPoison(V: BitWidth - 1))))) { |
1819 | Constant *Mask = ConstantInt::get(Ty, V: 1); |
1820 | // Retain the knowledge about the ignored lanes. |
1821 | Mask = Constant::mergeUndefsWith( |
1822 | C: Constant::mergeUndefsWith(C: Mask, Other: cast<Constant>(Val: Op1)), |
1823 | Other: cast<Constant>(Val: cast<Instruction>(Val: Op0)->getOperand(i: 1))); |
1824 | X = Builder.CreateAnd(LHS: X, RHS: Mask); |
1825 | return BinaryOperator::CreateNeg(Op: X); |
1826 | } |
1827 | |
1828 | if (Instruction *R = foldVariableSignZeroExtensionOfVariableHighBitExtract(OldAShr&: I)) |
1829 | return R; |
1830 | |
1831 | // See if we can turn a signed shr into an unsigned shr. |
1832 | if (MaskedValueIsZero(V: Op0, Mask: APInt::getSignMask(BitWidth), CxtI: &I)) { |
1833 | Instruction *Lshr = BinaryOperator::CreateLShr(V1: Op0, V2: Op1); |
1834 | Lshr->setIsExact(I.isExact()); |
1835 | return Lshr; |
1836 | } |
1837 | |
1838 | // ashr (xor %x, -1), %y --> xor (ashr %x, %y), -1 |
1839 | if (match(V: Op0, P: m_OneUse(SubPattern: m_Not(V: m_Value(V&: X))))) { |
1840 | // Note that we must drop 'exact'-ness of the shift! |
1841 | // Note that we can't keep undef's in -1 vector constant! |
1842 | auto *NewAShr = Builder.CreateAShr(LHS: X, RHS: Op1, Name: Op0->getName() + ".not" ); |
1843 | return BinaryOperator::CreateNot(Op: NewAShr); |
1844 | } |
1845 | |
1846 | return nullptr; |
1847 | } |
1848 | |