File: | llvm/lib/Transforms/InstCombine/InstCombineSelect.cpp |
Warning: | line 750, column 5 Value stored to 'Pred' is never read |
Press '?' to see keyboard shortcuts
Keyboard shortcuts:
1 | //===- InstCombineSelect.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 visitSelect function. |
10 | // |
11 | //===----------------------------------------------------------------------===// |
12 | |
13 | #include "InstCombineInternal.h" |
14 | #include "llvm/ADT/APInt.h" |
15 | #include "llvm/ADT/Optional.h" |
16 | #include "llvm/ADT/STLExtras.h" |
17 | #include "llvm/ADT/SmallVector.h" |
18 | #include "llvm/Analysis/AssumptionCache.h" |
19 | #include "llvm/Analysis/CmpInstAnalysis.h" |
20 | #include "llvm/Analysis/InstructionSimplify.h" |
21 | #include "llvm/Analysis/OverflowInstAnalysis.h" |
22 | #include "llvm/Analysis/ValueTracking.h" |
23 | #include "llvm/IR/BasicBlock.h" |
24 | #include "llvm/IR/Constant.h" |
25 | #include "llvm/IR/Constants.h" |
26 | #include "llvm/IR/DerivedTypes.h" |
27 | #include "llvm/IR/IRBuilder.h" |
28 | #include "llvm/IR/InstrTypes.h" |
29 | #include "llvm/IR/Instruction.h" |
30 | #include "llvm/IR/Instructions.h" |
31 | #include "llvm/IR/IntrinsicInst.h" |
32 | #include "llvm/IR/Intrinsics.h" |
33 | #include "llvm/IR/Operator.h" |
34 | #include "llvm/IR/PatternMatch.h" |
35 | #include "llvm/IR/Type.h" |
36 | #include "llvm/IR/User.h" |
37 | #include "llvm/IR/Value.h" |
38 | #include "llvm/Support/Casting.h" |
39 | #include "llvm/Support/ErrorHandling.h" |
40 | #include "llvm/Support/KnownBits.h" |
41 | #include "llvm/Transforms/InstCombine/InstCombineWorklist.h" |
42 | #include "llvm/Transforms/InstCombine/InstCombiner.h" |
43 | #include <cassert> |
44 | #include <utility> |
45 | |
46 | using namespace llvm; |
47 | using namespace PatternMatch; |
48 | |
49 | #define DEBUG_TYPE"instcombine" "instcombine" |
50 | |
51 | static Value *createMinMax(InstCombiner::BuilderTy &Builder, |
52 | SelectPatternFlavor SPF, Value *A, Value *B) { |
53 | CmpInst::Predicate Pred = getMinMaxPred(SPF); |
54 | assert(CmpInst::isIntPredicate(Pred) && "Expected integer predicate")(static_cast<void> (0)); |
55 | return Builder.CreateSelect(Builder.CreateICmp(Pred, A, B), A, B); |
56 | } |
57 | |
58 | /// Replace a select operand based on an equality comparison with the identity |
59 | /// constant of a binop. |
60 | static Instruction *foldSelectBinOpIdentity(SelectInst &Sel, |
61 | const TargetLibraryInfo &TLI, |
62 | InstCombinerImpl &IC) { |
63 | // The select condition must be an equality compare with a constant operand. |
64 | Value *X; |
65 | Constant *C; |
66 | CmpInst::Predicate Pred; |
67 | if (!match(Sel.getCondition(), m_Cmp(Pred, m_Value(X), m_Constant(C)))) |
68 | return nullptr; |
69 | |
70 | bool IsEq; |
71 | if (ICmpInst::isEquality(Pred)) |
72 | IsEq = Pred == ICmpInst::ICMP_EQ; |
73 | else if (Pred == FCmpInst::FCMP_OEQ) |
74 | IsEq = true; |
75 | else if (Pred == FCmpInst::FCMP_UNE) |
76 | IsEq = false; |
77 | else |
78 | return nullptr; |
79 | |
80 | // A select operand must be a binop. |
81 | BinaryOperator *BO; |
82 | if (!match(Sel.getOperand(IsEq ? 1 : 2), m_BinOp(BO))) |
83 | return nullptr; |
84 | |
85 | // The compare constant must be the identity constant for that binop. |
86 | // If this a floating-point compare with 0.0, any zero constant will do. |
87 | Type *Ty = BO->getType(); |
88 | Constant *IdC = ConstantExpr::getBinOpIdentity(BO->getOpcode(), Ty, true); |
89 | if (IdC != C) { |
90 | if (!IdC || !CmpInst::isFPPredicate(Pred)) |
91 | return nullptr; |
92 | if (!match(IdC, m_AnyZeroFP()) || !match(C, m_AnyZeroFP())) |
93 | return nullptr; |
94 | } |
95 | |
96 | // Last, match the compare variable operand with a binop operand. |
97 | Value *Y; |
98 | if (!BO->isCommutative() && !match(BO, m_BinOp(m_Value(Y), m_Specific(X)))) |
99 | return nullptr; |
100 | if (!match(BO, m_c_BinOp(m_Value(Y), m_Specific(X)))) |
101 | return nullptr; |
102 | |
103 | // +0.0 compares equal to -0.0, and so it does not behave as required for this |
104 | // transform. Bail out if we can not exclude that possibility. |
105 | if (isa<FPMathOperator>(BO)) |
106 | if (!BO->hasNoSignedZeros() && !CannotBeNegativeZero(Y, &TLI)) |
107 | return nullptr; |
108 | |
109 | // BO = binop Y, X |
110 | // S = { select (cmp eq X, C), BO, ? } or { select (cmp ne X, C), ?, BO } |
111 | // => |
112 | // S = { select (cmp eq X, C), Y, ? } or { select (cmp ne X, C), ?, Y } |
113 | return IC.replaceOperand(Sel, IsEq ? 1 : 2, Y); |
114 | } |
115 | |
116 | /// This folds: |
117 | /// select (icmp eq (and X, C1)), TC, FC |
118 | /// iff C1 is a power 2 and the difference between TC and FC is a power-of-2. |
119 | /// To something like: |
120 | /// (shr (and (X, C1)), (log2(C1) - log2(TC-FC))) + FC |
121 | /// Or: |
122 | /// (shl (and (X, C1)), (log2(TC-FC) - log2(C1))) + FC |
123 | /// With some variations depending if FC is larger than TC, or the shift |
124 | /// isn't needed, or the bit widths don't match. |
125 | static Value *foldSelectICmpAnd(SelectInst &Sel, ICmpInst *Cmp, |
126 | InstCombiner::BuilderTy &Builder) { |
127 | const APInt *SelTC, *SelFC; |
128 | if (!match(Sel.getTrueValue(), m_APInt(SelTC)) || |
129 | !match(Sel.getFalseValue(), m_APInt(SelFC))) |
130 | return nullptr; |
131 | |
132 | // If this is a vector select, we need a vector compare. |
133 | Type *SelType = Sel.getType(); |
134 | if (SelType->isVectorTy() != Cmp->getType()->isVectorTy()) |
135 | return nullptr; |
136 | |
137 | Value *V; |
138 | APInt AndMask; |
139 | bool CreateAnd = false; |
140 | ICmpInst::Predicate Pred = Cmp->getPredicate(); |
141 | if (ICmpInst::isEquality(Pred)) { |
142 | if (!match(Cmp->getOperand(1), m_Zero())) |
143 | return nullptr; |
144 | |
145 | V = Cmp->getOperand(0); |
146 | const APInt *AndRHS; |
147 | if (!match(V, m_And(m_Value(), m_Power2(AndRHS)))) |
148 | return nullptr; |
149 | |
150 | AndMask = *AndRHS; |
151 | } else if (decomposeBitTestICmp(Cmp->getOperand(0), Cmp->getOperand(1), |
152 | Pred, V, AndMask)) { |
153 | assert(ICmpInst::isEquality(Pred) && "Not equality test?")(static_cast<void> (0)); |
154 | if (!AndMask.isPowerOf2()) |
155 | return nullptr; |
156 | |
157 | CreateAnd = true; |
158 | } else { |
159 | return nullptr; |
160 | } |
161 | |
162 | // In general, when both constants are non-zero, we would need an offset to |
163 | // replace the select. This would require more instructions than we started |
164 | // with. But there's one special-case that we handle here because it can |
165 | // simplify/reduce the instructions. |
166 | APInt TC = *SelTC; |
167 | APInt FC = *SelFC; |
168 | if (!TC.isNullValue() && !FC.isNullValue()) { |
169 | // If the select constants differ by exactly one bit and that's the same |
170 | // bit that is masked and checked by the select condition, the select can |
171 | // be replaced by bitwise logic to set/clear one bit of the constant result. |
172 | if (TC.getBitWidth() != AndMask.getBitWidth() || (TC ^ FC) != AndMask) |
173 | return nullptr; |
174 | if (CreateAnd) { |
175 | // If we have to create an 'and', then we must kill the cmp to not |
176 | // increase the instruction count. |
177 | if (!Cmp->hasOneUse()) |
178 | return nullptr; |
179 | V = Builder.CreateAnd(V, ConstantInt::get(SelType, AndMask)); |
180 | } |
181 | bool ExtraBitInTC = TC.ugt(FC); |
182 | if (Pred == ICmpInst::ICMP_EQ) { |
183 | // If the masked bit in V is clear, clear or set the bit in the result: |
184 | // (V & AndMaskC) == 0 ? TC : FC --> (V & AndMaskC) ^ TC |
185 | // (V & AndMaskC) == 0 ? TC : FC --> (V & AndMaskC) | TC |
186 | Constant *C = ConstantInt::get(SelType, TC); |
187 | return ExtraBitInTC ? Builder.CreateXor(V, C) : Builder.CreateOr(V, C); |
188 | } |
189 | if (Pred == ICmpInst::ICMP_NE) { |
190 | // If the masked bit in V is set, set or clear the bit in the result: |
191 | // (V & AndMaskC) != 0 ? TC : FC --> (V & AndMaskC) | FC |
192 | // (V & AndMaskC) != 0 ? TC : FC --> (V & AndMaskC) ^ FC |
193 | Constant *C = ConstantInt::get(SelType, FC); |
194 | return ExtraBitInTC ? Builder.CreateOr(V, C) : Builder.CreateXor(V, C); |
195 | } |
196 | llvm_unreachable("Only expecting equality predicates")__builtin_unreachable(); |
197 | } |
198 | |
199 | // Make sure one of the select arms is a power-of-2. |
200 | if (!TC.isPowerOf2() && !FC.isPowerOf2()) |
201 | return nullptr; |
202 | |
203 | // Determine which shift is needed to transform result of the 'and' into the |
204 | // desired result. |
205 | const APInt &ValC = !TC.isNullValue() ? TC : FC; |
206 | unsigned ValZeros = ValC.logBase2(); |
207 | unsigned AndZeros = AndMask.logBase2(); |
208 | |
209 | // Insert the 'and' instruction on the input to the truncate. |
210 | if (CreateAnd) |
211 | V = Builder.CreateAnd(V, ConstantInt::get(V->getType(), AndMask)); |
212 | |
213 | // If types don't match, we can still convert the select by introducing a zext |
214 | // or a trunc of the 'and'. |
215 | if (ValZeros > AndZeros) { |
216 | V = Builder.CreateZExtOrTrunc(V, SelType); |
217 | V = Builder.CreateShl(V, ValZeros - AndZeros); |
218 | } else if (ValZeros < AndZeros) { |
219 | V = Builder.CreateLShr(V, AndZeros - ValZeros); |
220 | V = Builder.CreateZExtOrTrunc(V, SelType); |
221 | } else { |
222 | V = Builder.CreateZExtOrTrunc(V, SelType); |
223 | } |
224 | |
225 | // Okay, now we know that everything is set up, we just don't know whether we |
226 | // have a icmp_ne or icmp_eq and whether the true or false val is the zero. |
227 | bool ShouldNotVal = !TC.isNullValue(); |
228 | ShouldNotVal ^= Pred == ICmpInst::ICMP_NE; |
229 | if (ShouldNotVal) |
230 | V = Builder.CreateXor(V, ValC); |
231 | |
232 | return V; |
233 | } |
234 | |
235 | /// We want to turn code that looks like this: |
236 | /// %C = or %A, %B |
237 | /// %D = select %cond, %C, %A |
238 | /// into: |
239 | /// %C = select %cond, %B, 0 |
240 | /// %D = or %A, %C |
241 | /// |
242 | /// Assuming that the specified instruction is an operand to the select, return |
243 | /// a bitmask indicating which operands of this instruction are foldable if they |
244 | /// equal the other incoming value of the select. |
245 | static unsigned getSelectFoldableOperands(BinaryOperator *I) { |
246 | switch (I->getOpcode()) { |
247 | case Instruction::Add: |
248 | case Instruction::Mul: |
249 | case Instruction::And: |
250 | case Instruction::Or: |
251 | case Instruction::Xor: |
252 | return 3; // Can fold through either operand. |
253 | case Instruction::Sub: // Can only fold on the amount subtracted. |
254 | case Instruction::Shl: // Can only fold on the shift amount. |
255 | case Instruction::LShr: |
256 | case Instruction::AShr: |
257 | return 1; |
258 | default: |
259 | return 0; // Cannot fold |
260 | } |
261 | } |
262 | |
263 | /// We have (select c, TI, FI), and we know that TI and FI have the same opcode. |
264 | Instruction *InstCombinerImpl::foldSelectOpOp(SelectInst &SI, Instruction *TI, |
265 | Instruction *FI) { |
266 | // Don't break up min/max patterns. The hasOneUse checks below prevent that |
267 | // for most cases, but vector min/max with bitcasts can be transformed. If the |
268 | // one-use restrictions are eased for other patterns, we still don't want to |
269 | // obfuscate min/max. |
270 | if ((match(&SI, m_SMin(m_Value(), m_Value())) || |
271 | match(&SI, m_SMax(m_Value(), m_Value())) || |
272 | match(&SI, m_UMin(m_Value(), m_Value())) || |
273 | match(&SI, m_UMax(m_Value(), m_Value())))) |
274 | return nullptr; |
275 | |
276 | // If this is a cast from the same type, merge. |
277 | Value *Cond = SI.getCondition(); |
278 | Type *CondTy = Cond->getType(); |
279 | if (TI->getNumOperands() == 1 && TI->isCast()) { |
280 | Type *FIOpndTy = FI->getOperand(0)->getType(); |
281 | if (TI->getOperand(0)->getType() != FIOpndTy) |
282 | return nullptr; |
283 | |
284 | // The select condition may be a vector. We may only change the operand |
285 | // type if the vector width remains the same (and matches the condition). |
286 | if (auto *CondVTy = dyn_cast<VectorType>(CondTy)) { |
287 | if (!FIOpndTy->isVectorTy() || |
288 | CondVTy->getElementCount() != |
289 | cast<VectorType>(FIOpndTy)->getElementCount()) |
290 | return nullptr; |
291 | |
292 | // TODO: If the backend knew how to deal with casts better, we could |
293 | // remove this limitation. For now, there's too much potential to create |
294 | // worse codegen by promoting the select ahead of size-altering casts |
295 | // (PR28160). |
296 | // |
297 | // Note that ValueTracking's matchSelectPattern() looks through casts |
298 | // without checking 'hasOneUse' when it matches min/max patterns, so this |
299 | // transform may end up happening anyway. |
300 | if (TI->getOpcode() != Instruction::BitCast && |
301 | (!TI->hasOneUse() || !FI->hasOneUse())) |
302 | return nullptr; |
303 | } else if (!TI->hasOneUse() || !FI->hasOneUse()) { |
304 | // TODO: The one-use restrictions for a scalar select could be eased if |
305 | // the fold of a select in visitLoadInst() was enhanced to match a pattern |
306 | // that includes a cast. |
307 | return nullptr; |
308 | } |
309 | |
310 | // Fold this by inserting a select from the input values. |
311 | Value *NewSI = |
312 | Builder.CreateSelect(Cond, TI->getOperand(0), FI->getOperand(0), |
313 | SI.getName() + ".v", &SI); |
314 | return CastInst::Create(Instruction::CastOps(TI->getOpcode()), NewSI, |
315 | TI->getType()); |
316 | } |
317 | |
318 | // Cond ? -X : -Y --> -(Cond ? X : Y) |
319 | Value *X, *Y; |
320 | if (match(TI, m_FNeg(m_Value(X))) && match(FI, m_FNeg(m_Value(Y))) && |
321 | (TI->hasOneUse() || FI->hasOneUse())) { |
322 | // Intersect FMF from the fneg instructions and union those with the select. |
323 | FastMathFlags FMF = TI->getFastMathFlags(); |
324 | FMF &= FI->getFastMathFlags(); |
325 | FMF |= SI.getFastMathFlags(); |
326 | Value *NewSel = Builder.CreateSelect(Cond, X, Y, SI.getName() + ".v", &SI); |
327 | if (auto *NewSelI = dyn_cast<Instruction>(NewSel)) |
328 | NewSelI->setFastMathFlags(FMF); |
329 | Instruction *NewFNeg = UnaryOperator::CreateFNeg(NewSel); |
330 | NewFNeg->setFastMathFlags(FMF); |
331 | return NewFNeg; |
332 | } |
333 | |
334 | // Min/max intrinsic with a common operand can have the common operand pulled |
335 | // after the select. This is the same transform as below for binops, but |
336 | // specialized for intrinsic matching and without the restrictive uses clause. |
337 | auto *TII = dyn_cast<IntrinsicInst>(TI); |
338 | auto *FII = dyn_cast<IntrinsicInst>(FI); |
339 | if (TII && FII && TII->getIntrinsicID() == FII->getIntrinsicID() && |
340 | (TII->hasOneUse() || FII->hasOneUse())) { |
341 | Value *T0, *T1, *F0, *F1; |
342 | if (match(TII, m_MaxOrMin(m_Value(T0), m_Value(T1))) && |
343 | match(FII, m_MaxOrMin(m_Value(F0), m_Value(F1)))) { |
344 | if (T0 == F0) { |
345 | Value *NewSel = Builder.CreateSelect(Cond, T1, F1, "minmaxop", &SI); |
346 | return CallInst::Create(TII->getCalledFunction(), {NewSel, T0}); |
347 | } |
348 | if (T0 == F1) { |
349 | Value *NewSel = Builder.CreateSelect(Cond, T1, F0, "minmaxop", &SI); |
350 | return CallInst::Create(TII->getCalledFunction(), {NewSel, T0}); |
351 | } |
352 | if (T1 == F0) { |
353 | Value *NewSel = Builder.CreateSelect(Cond, T0, F1, "minmaxop", &SI); |
354 | return CallInst::Create(TII->getCalledFunction(), {NewSel, T1}); |
355 | } |
356 | if (T1 == F1) { |
357 | Value *NewSel = Builder.CreateSelect(Cond, T0, F0, "minmaxop", &SI); |
358 | return CallInst::Create(TII->getCalledFunction(), {NewSel, T1}); |
359 | } |
360 | } |
361 | } |
362 | |
363 | // Only handle binary operators (including two-operand getelementptr) with |
364 | // one-use here. As with the cast case above, it may be possible to relax the |
365 | // one-use constraint, but that needs be examined carefully since it may not |
366 | // reduce the total number of instructions. |
367 | if (TI->getNumOperands() != 2 || FI->getNumOperands() != 2 || |
368 | (!isa<BinaryOperator>(TI) && !isa<GetElementPtrInst>(TI)) || |
369 | !TI->hasOneUse() || !FI->hasOneUse()) |
370 | return nullptr; |
371 | |
372 | // Figure out if the operations have any operands in common. |
373 | Value *MatchOp, *OtherOpT, *OtherOpF; |
374 | bool MatchIsOpZero; |
375 | if (TI->getOperand(0) == FI->getOperand(0)) { |
376 | MatchOp = TI->getOperand(0); |
377 | OtherOpT = TI->getOperand(1); |
378 | OtherOpF = FI->getOperand(1); |
379 | MatchIsOpZero = true; |
380 | } else if (TI->getOperand(1) == FI->getOperand(1)) { |
381 | MatchOp = TI->getOperand(1); |
382 | OtherOpT = TI->getOperand(0); |
383 | OtherOpF = FI->getOperand(0); |
384 | MatchIsOpZero = false; |
385 | } else if (!TI->isCommutative()) { |
386 | return nullptr; |
387 | } else if (TI->getOperand(0) == FI->getOperand(1)) { |
388 | MatchOp = TI->getOperand(0); |
389 | OtherOpT = TI->getOperand(1); |
390 | OtherOpF = FI->getOperand(0); |
391 | MatchIsOpZero = true; |
392 | } else if (TI->getOperand(1) == FI->getOperand(0)) { |
393 | MatchOp = TI->getOperand(1); |
394 | OtherOpT = TI->getOperand(0); |
395 | OtherOpF = FI->getOperand(1); |
396 | MatchIsOpZero = true; |
397 | } else { |
398 | return nullptr; |
399 | } |
400 | |
401 | // If the select condition is a vector, the operands of the original select's |
402 | // operands also must be vectors. This may not be the case for getelementptr |
403 | // for example. |
404 | if (CondTy->isVectorTy() && (!OtherOpT->getType()->isVectorTy() || |
405 | !OtherOpF->getType()->isVectorTy())) |
406 | return nullptr; |
407 | |
408 | // If we reach here, they do have operations in common. |
409 | Value *NewSI = Builder.CreateSelect(Cond, OtherOpT, OtherOpF, |
410 | SI.getName() + ".v", &SI); |
411 | Value *Op0 = MatchIsOpZero ? MatchOp : NewSI; |
412 | Value *Op1 = MatchIsOpZero ? NewSI : MatchOp; |
413 | if (auto *BO = dyn_cast<BinaryOperator>(TI)) { |
414 | BinaryOperator *NewBO = BinaryOperator::Create(BO->getOpcode(), Op0, Op1); |
415 | NewBO->copyIRFlags(TI); |
416 | NewBO->andIRFlags(FI); |
417 | return NewBO; |
418 | } |
419 | if (auto *TGEP = dyn_cast<GetElementPtrInst>(TI)) { |
420 | auto *FGEP = cast<GetElementPtrInst>(FI); |
421 | Type *ElementType = TGEP->getResultElementType(); |
422 | return TGEP->isInBounds() && FGEP->isInBounds() |
423 | ? GetElementPtrInst::CreateInBounds(ElementType, Op0, {Op1}) |
424 | : GetElementPtrInst::Create(ElementType, Op0, {Op1}); |
425 | } |
426 | llvm_unreachable("Expected BinaryOperator or GEP")__builtin_unreachable(); |
427 | return nullptr; |
428 | } |
429 | |
430 | static bool isSelect01(const APInt &C1I, const APInt &C2I) { |
431 | if (!C1I.isNullValue() && !C2I.isNullValue()) // One side must be zero. |
432 | return false; |
433 | return C1I.isOneValue() || C1I.isAllOnesValue() || |
434 | C2I.isOneValue() || C2I.isAllOnesValue(); |
435 | } |
436 | |
437 | /// Try to fold the select into one of the operands to allow further |
438 | /// optimization. |
439 | Instruction *InstCombinerImpl::foldSelectIntoOp(SelectInst &SI, Value *TrueVal, |
440 | Value *FalseVal) { |
441 | // See the comment above GetSelectFoldableOperands for a description of the |
442 | // transformation we are doing here. |
443 | if (auto *TVI = dyn_cast<BinaryOperator>(TrueVal)) { |
444 | if (TVI->hasOneUse() && !isa<Constant>(FalseVal)) { |
445 | if (unsigned SFO = getSelectFoldableOperands(TVI)) { |
446 | unsigned OpToFold = 0; |
447 | if ((SFO & 1) && FalseVal == TVI->getOperand(0)) { |
448 | OpToFold = 1; |
449 | } else if ((SFO & 2) && FalseVal == TVI->getOperand(1)) { |
450 | OpToFold = 2; |
451 | } |
452 | |
453 | if (OpToFold) { |
454 | Constant *C = ConstantExpr::getBinOpIdentity(TVI->getOpcode(), |
455 | TVI->getType(), true); |
456 | Value *OOp = TVI->getOperand(2-OpToFold); |
457 | // Avoid creating select between 2 constants unless it's selecting |
458 | // between 0, 1 and -1. |
459 | const APInt *OOpC; |
460 | bool OOpIsAPInt = match(OOp, m_APInt(OOpC)); |
461 | if (!isa<Constant>(OOp) || |
462 | (OOpIsAPInt && isSelect01(C->getUniqueInteger(), *OOpC))) { |
463 | Value *NewSel = Builder.CreateSelect(SI.getCondition(), OOp, C); |
464 | NewSel->takeName(TVI); |
465 | BinaryOperator *BO = BinaryOperator::Create(TVI->getOpcode(), |
466 | FalseVal, NewSel); |
467 | BO->copyIRFlags(TVI); |
468 | return BO; |
469 | } |
470 | } |
471 | } |
472 | } |
473 | } |
474 | |
475 | if (auto *FVI = dyn_cast<BinaryOperator>(FalseVal)) { |
476 | if (FVI->hasOneUse() && !isa<Constant>(TrueVal)) { |
477 | if (unsigned SFO = getSelectFoldableOperands(FVI)) { |
478 | unsigned OpToFold = 0; |
479 | if ((SFO & 1) && TrueVal == FVI->getOperand(0)) { |
480 | OpToFold = 1; |
481 | } else if ((SFO & 2) && TrueVal == FVI->getOperand(1)) { |
482 | OpToFold = 2; |
483 | } |
484 | |
485 | if (OpToFold) { |
486 | Constant *C = ConstantExpr::getBinOpIdentity(FVI->getOpcode(), |
487 | FVI->getType(), true); |
488 | Value *OOp = FVI->getOperand(2-OpToFold); |
489 | // Avoid creating select between 2 constants unless it's selecting |
490 | // between 0, 1 and -1. |
491 | const APInt *OOpC; |
492 | bool OOpIsAPInt = match(OOp, m_APInt(OOpC)); |
493 | if (!isa<Constant>(OOp) || |
494 | (OOpIsAPInt && isSelect01(C->getUniqueInteger(), *OOpC))) { |
495 | Value *NewSel = Builder.CreateSelect(SI.getCondition(), C, OOp); |
496 | NewSel->takeName(FVI); |
497 | BinaryOperator *BO = BinaryOperator::Create(FVI->getOpcode(), |
498 | TrueVal, NewSel); |
499 | BO->copyIRFlags(FVI); |
500 | return BO; |
501 | } |
502 | } |
503 | } |
504 | } |
505 | } |
506 | |
507 | return nullptr; |
508 | } |
509 | |
510 | /// We want to turn: |
511 | /// (select (icmp eq (and X, Y), 0), (and (lshr X, Z), 1), 1) |
512 | /// into: |
513 | /// zext (icmp ne i32 (and X, (or Y, (shl 1, Z))), 0) |
514 | /// Note: |
515 | /// Z may be 0 if lshr is missing. |
516 | /// Worst-case scenario is that we will replace 5 instructions with 5 different |
517 | /// instructions, but we got rid of select. |
518 | static Instruction *foldSelectICmpAndAnd(Type *SelType, const ICmpInst *Cmp, |
519 | Value *TVal, Value *FVal, |
520 | InstCombiner::BuilderTy &Builder) { |
521 | if (!(Cmp->hasOneUse() && Cmp->getOperand(0)->hasOneUse() && |
522 | Cmp->getPredicate() == ICmpInst::ICMP_EQ && |
523 | match(Cmp->getOperand(1), m_Zero()) && match(FVal, m_One()))) |
524 | return nullptr; |
525 | |
526 | // The TrueVal has general form of: and %B, 1 |
527 | Value *B; |
528 | if (!match(TVal, m_OneUse(m_And(m_Value(B), m_One())))) |
529 | return nullptr; |
530 | |
531 | // Where %B may be optionally shifted: lshr %X, %Z. |
532 | Value *X, *Z; |
533 | const bool HasShift = match(B, m_OneUse(m_LShr(m_Value(X), m_Value(Z)))); |
534 | if (!HasShift) |
535 | X = B; |
536 | |
537 | Value *Y; |
538 | if (!match(Cmp->getOperand(0), m_c_And(m_Specific(X), m_Value(Y)))) |
539 | return nullptr; |
540 | |
541 | // ((X & Y) == 0) ? ((X >> Z) & 1) : 1 --> (X & (Y | (1 << Z))) != 0 |
542 | // ((X & Y) == 0) ? (X & 1) : 1 --> (X & (Y | 1)) != 0 |
543 | Constant *One = ConstantInt::get(SelType, 1); |
544 | Value *MaskB = HasShift ? Builder.CreateShl(One, Z) : One; |
545 | Value *FullMask = Builder.CreateOr(Y, MaskB); |
546 | Value *MaskedX = Builder.CreateAnd(X, FullMask); |
547 | Value *ICmpNeZero = Builder.CreateIsNotNull(MaskedX); |
548 | return new ZExtInst(ICmpNeZero, SelType); |
549 | } |
550 | |
551 | /// We want to turn: |
552 | /// (select (icmp sgt x, C), lshr (X, Y), ashr (X, Y)); iff C s>= -1 |
553 | /// (select (icmp slt x, C), ashr (X, Y), lshr (X, Y)); iff C s>= 0 |
554 | /// into: |
555 | /// ashr (X, Y) |
556 | static Value *foldSelectICmpLshrAshr(const ICmpInst *IC, Value *TrueVal, |
557 | Value *FalseVal, |
558 | InstCombiner::BuilderTy &Builder) { |
559 | ICmpInst::Predicate Pred = IC->getPredicate(); |
560 | Value *CmpLHS = IC->getOperand(0); |
561 | Value *CmpRHS = IC->getOperand(1); |
562 | if (!CmpRHS->getType()->isIntOrIntVectorTy()) |
563 | return nullptr; |
564 | |
565 | Value *X, *Y; |
566 | unsigned Bitwidth = CmpRHS->getType()->getScalarSizeInBits(); |
567 | if ((Pred != ICmpInst::ICMP_SGT || |
568 | !match(CmpRHS, |
569 | m_SpecificInt_ICMP(ICmpInst::ICMP_SGE, APInt(Bitwidth, -1)))) && |
570 | (Pred != ICmpInst::ICMP_SLT || |
571 | !match(CmpRHS, |
572 | m_SpecificInt_ICMP(ICmpInst::ICMP_SGE, APInt(Bitwidth, 0))))) |
573 | return nullptr; |
574 | |
575 | // Canonicalize so that ashr is in FalseVal. |
576 | if (Pred == ICmpInst::ICMP_SLT) |
577 | std::swap(TrueVal, FalseVal); |
578 | |
579 | if (match(TrueVal, m_LShr(m_Value(X), m_Value(Y))) && |
580 | match(FalseVal, m_AShr(m_Specific(X), m_Specific(Y))) && |
581 | match(CmpLHS, m_Specific(X))) { |
582 | const auto *Ashr = cast<Instruction>(FalseVal); |
583 | // if lshr is not exact and ashr is, this new ashr must not be exact. |
584 | bool IsExact = Ashr->isExact() && cast<Instruction>(TrueVal)->isExact(); |
585 | return Builder.CreateAShr(X, Y, IC->getName(), IsExact); |
586 | } |
587 | |
588 | return nullptr; |
589 | } |
590 | |
591 | /// We want to turn: |
592 | /// (select (icmp eq (and X, C1), 0), Y, (or Y, C2)) |
593 | /// into: |
594 | /// (or (shl (and X, C1), C3), Y) |
595 | /// iff: |
596 | /// C1 and C2 are both powers of 2 |
597 | /// where: |
598 | /// C3 = Log(C2) - Log(C1) |
599 | /// |
600 | /// This transform handles cases where: |
601 | /// 1. The icmp predicate is inverted |
602 | /// 2. The select operands are reversed |
603 | /// 3. The magnitude of C2 and C1 are flipped |
604 | static Value *foldSelectICmpAndOr(const ICmpInst *IC, Value *TrueVal, |
605 | Value *FalseVal, |
606 | InstCombiner::BuilderTy &Builder) { |
607 | // Only handle integer compares. Also, if this is a vector select, we need a |
608 | // vector compare. |
609 | if (!TrueVal->getType()->isIntOrIntVectorTy() || |
610 | TrueVal->getType()->isVectorTy() != IC->getType()->isVectorTy()) |
611 | return nullptr; |
612 | |
613 | Value *CmpLHS = IC->getOperand(0); |
614 | Value *CmpRHS = IC->getOperand(1); |
615 | |
616 | Value *V; |
617 | unsigned C1Log; |
618 | bool IsEqualZero; |
619 | bool NeedAnd = false; |
620 | if (IC->isEquality()) { |
621 | if (!match(CmpRHS, m_Zero())) |
622 | return nullptr; |
623 | |
624 | const APInt *C1; |
625 | if (!match(CmpLHS, m_And(m_Value(), m_Power2(C1)))) |
626 | return nullptr; |
627 | |
628 | V = CmpLHS; |
629 | C1Log = C1->logBase2(); |
630 | IsEqualZero = IC->getPredicate() == ICmpInst::ICMP_EQ; |
631 | } else if (IC->getPredicate() == ICmpInst::ICMP_SLT || |
632 | IC->getPredicate() == ICmpInst::ICMP_SGT) { |
633 | // We also need to recognize (icmp slt (trunc (X)), 0) and |
634 | // (icmp sgt (trunc (X)), -1). |
635 | IsEqualZero = IC->getPredicate() == ICmpInst::ICMP_SGT; |
636 | if ((IsEqualZero && !match(CmpRHS, m_AllOnes())) || |
637 | (!IsEqualZero && !match(CmpRHS, m_Zero()))) |
638 | return nullptr; |
639 | |
640 | if (!match(CmpLHS, m_OneUse(m_Trunc(m_Value(V))))) |
641 | return nullptr; |
642 | |
643 | C1Log = CmpLHS->getType()->getScalarSizeInBits() - 1; |
644 | NeedAnd = true; |
645 | } else { |
646 | return nullptr; |
647 | } |
648 | |
649 | const APInt *C2; |
650 | bool OrOnTrueVal = false; |
651 | bool OrOnFalseVal = match(FalseVal, m_Or(m_Specific(TrueVal), m_Power2(C2))); |
652 | if (!OrOnFalseVal) |
653 | OrOnTrueVal = match(TrueVal, m_Or(m_Specific(FalseVal), m_Power2(C2))); |
654 | |
655 | if (!OrOnFalseVal && !OrOnTrueVal) |
656 | return nullptr; |
657 | |
658 | Value *Y = OrOnFalseVal ? TrueVal : FalseVal; |
659 | |
660 | unsigned C2Log = C2->logBase2(); |
661 | |
662 | bool NeedXor = (!IsEqualZero && OrOnFalseVal) || (IsEqualZero && OrOnTrueVal); |
663 | bool NeedShift = C1Log != C2Log; |
664 | bool NeedZExtTrunc = Y->getType()->getScalarSizeInBits() != |
665 | V->getType()->getScalarSizeInBits(); |
666 | |
667 | // Make sure we don't create more instructions than we save. |
668 | Value *Or = OrOnFalseVal ? FalseVal : TrueVal; |
669 | if ((NeedShift + NeedXor + NeedZExtTrunc) > |
670 | (IC->hasOneUse() + Or->hasOneUse())) |
671 | return nullptr; |
672 | |
673 | if (NeedAnd) { |
674 | // Insert the AND instruction on the input to the truncate. |
675 | APInt C1 = APInt::getOneBitSet(V->getType()->getScalarSizeInBits(), C1Log); |
676 | V = Builder.CreateAnd(V, ConstantInt::get(V->getType(), C1)); |
677 | } |
678 | |
679 | if (C2Log > C1Log) { |
680 | V = Builder.CreateZExtOrTrunc(V, Y->getType()); |
681 | V = Builder.CreateShl(V, C2Log - C1Log); |
682 | } else if (C1Log > C2Log) { |
683 | V = Builder.CreateLShr(V, C1Log - C2Log); |
684 | V = Builder.CreateZExtOrTrunc(V, Y->getType()); |
685 | } else |
686 | V = Builder.CreateZExtOrTrunc(V, Y->getType()); |
687 | |
688 | if (NeedXor) |
689 | V = Builder.CreateXor(V, *C2); |
690 | |
691 | return Builder.CreateOr(V, Y); |
692 | } |
693 | |
694 | /// Canonicalize a set or clear of a masked set of constant bits to |
695 | /// select-of-constants form. |
696 | static Instruction *foldSetClearBits(SelectInst &Sel, |
697 | InstCombiner::BuilderTy &Builder) { |
698 | Value *Cond = Sel.getCondition(); |
699 | Value *T = Sel.getTrueValue(); |
700 | Value *F = Sel.getFalseValue(); |
701 | Type *Ty = Sel.getType(); |
702 | Value *X; |
703 | const APInt *NotC, *C; |
704 | |
705 | // Cond ? (X & ~C) : (X | C) --> (X & ~C) | (Cond ? 0 : C) |
706 | if (match(T, m_And(m_Value(X), m_APInt(NotC))) && |
707 | match(F, m_OneUse(m_Or(m_Specific(X), m_APInt(C)))) && *NotC == ~(*C)) { |
708 | Constant *Zero = ConstantInt::getNullValue(Ty); |
709 | Constant *OrC = ConstantInt::get(Ty, *C); |
710 | Value *NewSel = Builder.CreateSelect(Cond, Zero, OrC, "masksel", &Sel); |
711 | return BinaryOperator::CreateOr(T, NewSel); |
712 | } |
713 | |
714 | // Cond ? (X | C) : (X & ~C) --> (X & ~C) | (Cond ? C : 0) |
715 | if (match(F, m_And(m_Value(X), m_APInt(NotC))) && |
716 | match(T, m_OneUse(m_Or(m_Specific(X), m_APInt(C)))) && *NotC == ~(*C)) { |
717 | Constant *Zero = ConstantInt::getNullValue(Ty); |
718 | Constant *OrC = ConstantInt::get(Ty, *C); |
719 | Value *NewSel = Builder.CreateSelect(Cond, OrC, Zero, "masksel", &Sel); |
720 | return BinaryOperator::CreateOr(F, NewSel); |
721 | } |
722 | |
723 | return nullptr; |
724 | } |
725 | |
726 | /// Transform patterns such as (a > b) ? a - b : 0 into usub.sat(a, b). |
727 | /// There are 8 commuted/swapped variants of this pattern. |
728 | /// TODO: Also support a - UMIN(a,b) patterns. |
729 | static Value *canonicalizeSaturatedSubtract(const ICmpInst *ICI, |
730 | const Value *TrueVal, |
731 | const Value *FalseVal, |
732 | InstCombiner::BuilderTy &Builder) { |
733 | ICmpInst::Predicate Pred = ICI->getPredicate(); |
734 | if (!ICmpInst::isUnsigned(Pred)) |
735 | return nullptr; |
736 | |
737 | // (b > a) ? 0 : a - b -> (b <= a) ? a - b : 0 |
738 | if (match(TrueVal, m_Zero())) { |
739 | Pred = ICmpInst::getInversePredicate(Pred); |
740 | std::swap(TrueVal, FalseVal); |
741 | } |
742 | if (!match(FalseVal, m_Zero())) |
743 | return nullptr; |
744 | |
745 | Value *A = ICI->getOperand(0); |
746 | Value *B = ICI->getOperand(1); |
747 | if (Pred == ICmpInst::ICMP_ULE || Pred == ICmpInst::ICMP_ULT) { |
748 | // (b < a) ? a - b : 0 -> (a > b) ? a - b : 0 |
749 | std::swap(A, B); |
750 | Pred = ICmpInst::getSwappedPredicate(Pred); |
Value stored to 'Pred' is never read | |
751 | } |
752 | |
753 | assert((Pred == ICmpInst::ICMP_UGE || Pred == ICmpInst::ICMP_UGT) &&(static_cast<void> (0)) |
754 | "Unexpected isUnsigned predicate!")(static_cast<void> (0)); |
755 | |
756 | // Ensure the sub is of the form: |
757 | // (a > b) ? a - b : 0 -> usub.sat(a, b) |
758 | // (a > b) ? b - a : 0 -> -usub.sat(a, b) |
759 | // Checking for both a-b and a+(-b) as a constant. |
760 | bool IsNegative = false; |
761 | const APInt *C; |
762 | if (match(TrueVal, m_Sub(m_Specific(B), m_Specific(A))) || |
763 | (match(A, m_APInt(C)) && |
764 | match(TrueVal, m_Add(m_Specific(B), m_SpecificInt(-*C))))) |
765 | IsNegative = true; |
766 | else if (!match(TrueVal, m_Sub(m_Specific(A), m_Specific(B))) && |
767 | !(match(B, m_APInt(C)) && |
768 | match(TrueVal, m_Add(m_Specific(A), m_SpecificInt(-*C))))) |
769 | return nullptr; |
770 | |
771 | // If we are adding a negate and the sub and icmp are used anywhere else, we |
772 | // would end up with more instructions. |
773 | if (IsNegative && !TrueVal->hasOneUse() && !ICI->hasOneUse()) |
774 | return nullptr; |
775 | |
776 | // (a > b) ? a - b : 0 -> usub.sat(a, b) |
777 | // (a > b) ? b - a : 0 -> -usub.sat(a, b) |
778 | Value *Result = Builder.CreateBinaryIntrinsic(Intrinsic::usub_sat, A, B); |
779 | if (IsNegative) |
780 | Result = Builder.CreateNeg(Result); |
781 | return Result; |
782 | } |
783 | |
784 | static Value *canonicalizeSaturatedAdd(ICmpInst *Cmp, Value *TVal, Value *FVal, |
785 | InstCombiner::BuilderTy &Builder) { |
786 | if (!Cmp->hasOneUse()) |
787 | return nullptr; |
788 | |
789 | // Match unsigned saturated add with constant. |
790 | Value *Cmp0 = Cmp->getOperand(0); |
791 | Value *Cmp1 = Cmp->getOperand(1); |
792 | ICmpInst::Predicate Pred = Cmp->getPredicate(); |
793 | Value *X; |
794 | const APInt *C, *CmpC; |
795 | if (Pred == ICmpInst::ICMP_ULT && |
796 | match(TVal, m_Add(m_Value(X), m_APInt(C))) && X == Cmp0 && |
797 | match(FVal, m_AllOnes()) && match(Cmp1, m_APInt(CmpC)) && *CmpC == ~*C) { |
798 | // (X u< ~C) ? (X + C) : -1 --> uadd.sat(X, C) |
799 | return Builder.CreateBinaryIntrinsic( |
800 | Intrinsic::uadd_sat, X, ConstantInt::get(X->getType(), *C)); |
801 | } |
802 | |
803 | // Match unsigned saturated add of 2 variables with an unnecessary 'not'. |
804 | // There are 8 commuted variants. |
805 | // Canonicalize -1 (saturated result) to true value of the select. |
806 | if (match(FVal, m_AllOnes())) { |
807 | std::swap(TVal, FVal); |
808 | Pred = CmpInst::getInversePredicate(Pred); |
809 | } |
810 | if (!match(TVal, m_AllOnes())) |
811 | return nullptr; |
812 | |
813 | // Canonicalize predicate to less-than or less-or-equal-than. |
814 | if (Pred == ICmpInst::ICMP_UGT || Pred == ICmpInst::ICMP_UGE) { |
815 | std::swap(Cmp0, Cmp1); |
816 | Pred = CmpInst::getSwappedPredicate(Pred); |
817 | } |
818 | if (Pred != ICmpInst::ICMP_ULT && Pred != ICmpInst::ICMP_ULE) |
819 | return nullptr; |
820 | |
821 | // Match unsigned saturated add of 2 variables with an unnecessary 'not'. |
822 | // Strictness of the comparison is irrelevant. |
823 | Value *Y; |
824 | if (match(Cmp0, m_Not(m_Value(X))) && |
825 | match(FVal, m_c_Add(m_Specific(X), m_Value(Y))) && Y == Cmp1) { |
826 | // (~X u< Y) ? -1 : (X + Y) --> uadd.sat(X, Y) |
827 | // (~X u< Y) ? -1 : (Y + X) --> uadd.sat(X, Y) |
828 | return Builder.CreateBinaryIntrinsic(Intrinsic::uadd_sat, X, Y); |
829 | } |
830 | // The 'not' op may be included in the sum but not the compare. |
831 | // Strictness of the comparison is irrelevant. |
832 | X = Cmp0; |
833 | Y = Cmp1; |
834 | if (match(FVal, m_c_Add(m_Not(m_Specific(X)), m_Specific(Y)))) { |
835 | // (X u< Y) ? -1 : (~X + Y) --> uadd.sat(~X, Y) |
836 | // (X u< Y) ? -1 : (Y + ~X) --> uadd.sat(Y, ~X) |
837 | BinaryOperator *BO = cast<BinaryOperator>(FVal); |
838 | return Builder.CreateBinaryIntrinsic( |
839 | Intrinsic::uadd_sat, BO->getOperand(0), BO->getOperand(1)); |
840 | } |
841 | // The overflow may be detected via the add wrapping round. |
842 | // This is only valid for strict comparison! |
843 | if (Pred == ICmpInst::ICMP_ULT && |
844 | match(Cmp0, m_c_Add(m_Specific(Cmp1), m_Value(Y))) && |
845 | match(FVal, m_c_Add(m_Specific(Cmp1), m_Specific(Y)))) { |
846 | // ((X + Y) u< X) ? -1 : (X + Y) --> uadd.sat(X, Y) |
847 | // ((X + Y) u< Y) ? -1 : (X + Y) --> uadd.sat(X, Y) |
848 | return Builder.CreateBinaryIntrinsic(Intrinsic::uadd_sat, Cmp1, Y); |
849 | } |
850 | |
851 | return nullptr; |
852 | } |
853 | |
854 | /// Fold the following code sequence: |
855 | /// \code |
856 | /// int a = ctlz(x & -x); |
857 | // x ? 31 - a : a; |
858 | /// \code |
859 | /// |
860 | /// into: |
861 | /// cttz(x) |
862 | static Instruction *foldSelectCtlzToCttz(ICmpInst *ICI, Value *TrueVal, |
863 | Value *FalseVal, |
864 | InstCombiner::BuilderTy &Builder) { |
865 | unsigned BitWidth = TrueVal->getType()->getScalarSizeInBits(); |
866 | if (!ICI->isEquality() || !match(ICI->getOperand(1), m_Zero())) |
867 | return nullptr; |
868 | |
869 | if (ICI->getPredicate() == ICmpInst::ICMP_NE) |
870 | std::swap(TrueVal, FalseVal); |
871 | |
872 | if (!match(FalseVal, |
873 | m_Xor(m_Deferred(TrueVal), m_SpecificInt(BitWidth - 1)))) |
874 | return nullptr; |
875 | |
876 | if (!match(TrueVal, m_Intrinsic<Intrinsic::ctlz>())) |
877 | return nullptr; |
878 | |
879 | Value *X = ICI->getOperand(0); |
880 | auto *II = cast<IntrinsicInst>(TrueVal); |
881 | if (!match(II->getOperand(0), m_c_And(m_Specific(X), m_Neg(m_Specific(X))))) |
882 | return nullptr; |
883 | |
884 | Function *F = Intrinsic::getDeclaration(II->getModule(), Intrinsic::cttz, |
885 | II->getType()); |
886 | return CallInst::Create(F, {X, II->getArgOperand(1)}); |
887 | } |
888 | |
889 | /// Attempt to fold a cttz/ctlz followed by a icmp plus select into a single |
890 | /// call to cttz/ctlz with flag 'is_zero_undef' cleared. |
891 | /// |
892 | /// For example, we can fold the following code sequence: |
893 | /// \code |
894 | /// %0 = tail call i32 @llvm.cttz.i32(i32 %x, i1 true) |
895 | /// %1 = icmp ne i32 %x, 0 |
896 | /// %2 = select i1 %1, i32 %0, i32 32 |
897 | /// \code |
898 | /// |
899 | /// into: |
900 | /// %0 = tail call i32 @llvm.cttz.i32(i32 %x, i1 false) |
901 | static Value *foldSelectCttzCtlz(ICmpInst *ICI, Value *TrueVal, Value *FalseVal, |
902 | InstCombiner::BuilderTy &Builder) { |
903 | ICmpInst::Predicate Pred = ICI->getPredicate(); |
904 | Value *CmpLHS = ICI->getOperand(0); |
905 | Value *CmpRHS = ICI->getOperand(1); |
906 | |
907 | // Check if the condition value compares a value for equality against zero. |
908 | if (!ICI->isEquality() || !match(CmpRHS, m_Zero())) |
909 | return nullptr; |
910 | |
911 | Value *SelectArg = FalseVal; |
912 | Value *ValueOnZero = TrueVal; |
913 | if (Pred == ICmpInst::ICMP_NE) |
914 | std::swap(SelectArg, ValueOnZero); |
915 | |
916 | // Skip zero extend/truncate. |
917 | Value *Count = nullptr; |
918 | if (!match(SelectArg, m_ZExt(m_Value(Count))) && |
919 | !match(SelectArg, m_Trunc(m_Value(Count)))) |
920 | Count = SelectArg; |
921 | |
922 | // Check that 'Count' is a call to intrinsic cttz/ctlz. Also check that the |
923 | // input to the cttz/ctlz is used as LHS for the compare instruction. |
924 | if (!match(Count, m_Intrinsic<Intrinsic::cttz>(m_Specific(CmpLHS))) && |
925 | !match(Count, m_Intrinsic<Intrinsic::ctlz>(m_Specific(CmpLHS)))) |
926 | return nullptr; |
927 | |
928 | IntrinsicInst *II = cast<IntrinsicInst>(Count); |
929 | |
930 | // Check if the value propagated on zero is a constant number equal to the |
931 | // sizeof in bits of 'Count'. |
932 | unsigned SizeOfInBits = Count->getType()->getScalarSizeInBits(); |
933 | if (match(ValueOnZero, m_SpecificInt(SizeOfInBits))) { |
934 | // Explicitly clear the 'undef_on_zero' flag. It's always valid to go from |
935 | // true to false on this flag, so we can replace it for all users. |
936 | II->setArgOperand(1, ConstantInt::getFalse(II->getContext())); |
937 | return SelectArg; |
938 | } |
939 | |
940 | // The ValueOnZero is not the bitwidth. But if the cttz/ctlz (and optional |
941 | // zext/trunc) have one use (ending at the select), the cttz/ctlz result will |
942 | // not be used if the input is zero. Relax to 'undef_on_zero' for that case. |
943 | if (II->hasOneUse() && SelectArg->hasOneUse() && |
944 | !match(II->getArgOperand(1), m_One())) |
945 | II->setArgOperand(1, ConstantInt::getTrue(II->getContext())); |
946 | |
947 | return nullptr; |
948 | } |
949 | |
950 | /// Return true if we find and adjust an icmp+select pattern where the compare |
951 | /// is with a constant that can be incremented or decremented to match the |
952 | /// minimum or maximum idiom. |
953 | static bool adjustMinMax(SelectInst &Sel, ICmpInst &Cmp) { |
954 | ICmpInst::Predicate Pred = Cmp.getPredicate(); |
955 | Value *CmpLHS = Cmp.getOperand(0); |
956 | Value *CmpRHS = Cmp.getOperand(1); |
957 | Value *TrueVal = Sel.getTrueValue(); |
958 | Value *FalseVal = Sel.getFalseValue(); |
959 | |
960 | // We may move or edit the compare, so make sure the select is the only user. |
961 | const APInt *CmpC; |
962 | if (!Cmp.hasOneUse() || !match(CmpRHS, m_APInt(CmpC))) |
963 | return false; |
964 | |
965 | // These transforms only work for selects of integers or vector selects of |
966 | // integer vectors. |
967 | Type *SelTy = Sel.getType(); |
968 | auto *SelEltTy = dyn_cast<IntegerType>(SelTy->getScalarType()); |
969 | if (!SelEltTy || SelTy->isVectorTy() != Cmp.getType()->isVectorTy()) |
970 | return false; |
971 | |
972 | Constant *AdjustedRHS; |
973 | if (Pred == ICmpInst::ICMP_UGT || Pred == ICmpInst::ICMP_SGT) |
974 | AdjustedRHS = ConstantInt::get(CmpRHS->getType(), *CmpC + 1); |
975 | else if (Pred == ICmpInst::ICMP_ULT || Pred == ICmpInst::ICMP_SLT) |
976 | AdjustedRHS = ConstantInt::get(CmpRHS->getType(), *CmpC - 1); |
977 | else |
978 | return false; |
979 | |
980 | // X > C ? X : C+1 --> X < C+1 ? C+1 : X |
981 | // X < C ? X : C-1 --> X > C-1 ? C-1 : X |
982 | if ((CmpLHS == TrueVal && AdjustedRHS == FalseVal) || |
983 | (CmpLHS == FalseVal && AdjustedRHS == TrueVal)) { |
984 | ; // Nothing to do here. Values match without any sign/zero extension. |
985 | } |
986 | // Types do not match. Instead of calculating this with mixed types, promote |
987 | // all to the larger type. This enables scalar evolution to analyze this |
988 | // expression. |
989 | else if (CmpRHS->getType()->getScalarSizeInBits() < SelEltTy->getBitWidth()) { |
990 | Constant *SextRHS = ConstantExpr::getSExt(AdjustedRHS, SelTy); |
991 | |
992 | // X = sext x; x >s c ? X : C+1 --> X = sext x; X <s C+1 ? C+1 : X |
993 | // X = sext x; x <s c ? X : C-1 --> X = sext x; X >s C-1 ? C-1 : X |
994 | // X = sext x; x >u c ? X : C+1 --> X = sext x; X <u C+1 ? C+1 : X |
995 | // X = sext x; x <u c ? X : C-1 --> X = sext x; X >u C-1 ? C-1 : X |
996 | if (match(TrueVal, m_SExt(m_Specific(CmpLHS))) && SextRHS == FalseVal) { |
997 | CmpLHS = TrueVal; |
998 | AdjustedRHS = SextRHS; |
999 | } else if (match(FalseVal, m_SExt(m_Specific(CmpLHS))) && |
1000 | SextRHS == TrueVal) { |
1001 | CmpLHS = FalseVal; |
1002 | AdjustedRHS = SextRHS; |
1003 | } else if (Cmp.isUnsigned()) { |
1004 | Constant *ZextRHS = ConstantExpr::getZExt(AdjustedRHS, SelTy); |
1005 | // X = zext x; x >u c ? X : C+1 --> X = zext x; X <u C+1 ? C+1 : X |
1006 | // X = zext x; x <u c ? X : C-1 --> X = zext x; X >u C-1 ? C-1 : X |
1007 | // zext + signed compare cannot be changed: |
1008 | // 0xff <s 0x00, but 0x00ff >s 0x0000 |
1009 | if (match(TrueVal, m_ZExt(m_Specific(CmpLHS))) && ZextRHS == FalseVal) { |
1010 | CmpLHS = TrueVal; |
1011 | AdjustedRHS = ZextRHS; |
1012 | } else if (match(FalseVal, m_ZExt(m_Specific(CmpLHS))) && |
1013 | ZextRHS == TrueVal) { |
1014 | CmpLHS = FalseVal; |
1015 | AdjustedRHS = ZextRHS; |
1016 | } else { |
1017 | return false; |
1018 | } |
1019 | } else { |
1020 | return false; |
1021 | } |
1022 | } else { |
1023 | return false; |
1024 | } |
1025 | |
1026 | Pred = ICmpInst::getSwappedPredicate(Pred); |
1027 | CmpRHS = AdjustedRHS; |
1028 | std::swap(FalseVal, TrueVal); |
1029 | Cmp.setPredicate(Pred); |
1030 | Cmp.setOperand(0, CmpLHS); |
1031 | Cmp.setOperand(1, CmpRHS); |
1032 | Sel.setOperand(1, TrueVal); |
1033 | Sel.setOperand(2, FalseVal); |
1034 | Sel.swapProfMetadata(); |
1035 | |
1036 | // Move the compare instruction right before the select instruction. Otherwise |
1037 | // the sext/zext value may be defined after the compare instruction uses it. |
1038 | Cmp.moveBefore(&Sel); |
1039 | |
1040 | return true; |
1041 | } |
1042 | |
1043 | /// If this is an integer min/max (icmp + select) with a constant operand, |
1044 | /// create the canonical icmp for the min/max operation and canonicalize the |
1045 | /// constant to the 'false' operand of the select: |
1046 | /// select (icmp Pred X, C1), C2, X --> select (icmp Pred' X, C2), X, C2 |
1047 | /// Note: if C1 != C2, this will change the icmp constant to the existing |
1048 | /// constant operand of the select. |
1049 | static Instruction *canonicalizeMinMaxWithConstant(SelectInst &Sel, |
1050 | ICmpInst &Cmp, |
1051 | InstCombinerImpl &IC) { |
1052 | if (!Cmp.hasOneUse() || !isa<Constant>(Cmp.getOperand(1))) |
1053 | return nullptr; |
1054 | |
1055 | // Canonicalize the compare predicate based on whether we have min or max. |
1056 | Value *LHS, *RHS; |
1057 | SelectPatternResult SPR = matchSelectPattern(&Sel, LHS, RHS); |
1058 | if (!SelectPatternResult::isMinOrMax(SPR.Flavor)) |
1059 | return nullptr; |
1060 | |
1061 | // Is this already canonical? |
1062 | ICmpInst::Predicate CanonicalPred = getMinMaxPred(SPR.Flavor); |
1063 | if (Cmp.getOperand(0) == LHS && Cmp.getOperand(1) == RHS && |
1064 | Cmp.getPredicate() == CanonicalPred) |
1065 | return nullptr; |
1066 | |
1067 | // Bail out on unsimplified X-0 operand (due to some worklist management bug), |
1068 | // as this may cause an infinite combine loop. Let the sub be folded first. |
1069 | if (match(LHS, m_Sub(m_Value(), m_Zero())) || |
1070 | match(RHS, m_Sub(m_Value(), m_Zero()))) |
1071 | return nullptr; |
1072 | |
1073 | // Create the canonical compare and plug it into the select. |
1074 | IC.replaceOperand(Sel, 0, IC.Builder.CreateICmp(CanonicalPred, LHS, RHS)); |
1075 | |
1076 | // If the select operands did not change, we're done. |
1077 | if (Sel.getTrueValue() == LHS && Sel.getFalseValue() == RHS) |
1078 | return &Sel; |
1079 | |
1080 | // If we are swapping the select operands, swap the metadata too. |
1081 | assert(Sel.getTrueValue() == RHS && Sel.getFalseValue() == LHS &&(static_cast<void> (0)) |
1082 | "Unexpected results from matchSelectPattern")(static_cast<void> (0)); |
1083 | Sel.swapValues(); |
1084 | Sel.swapProfMetadata(); |
1085 | return &Sel; |
1086 | } |
1087 | |
1088 | static Instruction *canonicalizeAbsNabs(SelectInst &Sel, ICmpInst &Cmp, |
1089 | InstCombinerImpl &IC) { |
1090 | if (!Cmp.hasOneUse() || !isa<Constant>(Cmp.getOperand(1))) |
1091 | return nullptr; |
1092 | |
1093 | Value *LHS, *RHS; |
1094 | SelectPatternFlavor SPF = matchSelectPattern(&Sel, LHS, RHS).Flavor; |
1095 | if (SPF != SelectPatternFlavor::SPF_ABS && |
1096 | SPF != SelectPatternFlavor::SPF_NABS) |
1097 | return nullptr; |
1098 | |
1099 | // Note that NSW flag can only be propagated for normal, non-negated abs! |
1100 | bool IntMinIsPoison = SPF == SelectPatternFlavor::SPF_ABS && |
1101 | match(RHS, m_NSWNeg(m_Specific(LHS))); |
1102 | Constant *IntMinIsPoisonC = |
1103 | ConstantInt::get(Type::getInt1Ty(Sel.getContext()), IntMinIsPoison); |
1104 | Instruction *Abs = |
1105 | IC.Builder.CreateBinaryIntrinsic(Intrinsic::abs, LHS, IntMinIsPoisonC); |
1106 | |
1107 | if (SPF == SelectPatternFlavor::SPF_NABS) |
1108 | return BinaryOperator::CreateNeg(Abs); // Always without NSW flag! |
1109 | |
1110 | return IC.replaceInstUsesWith(Sel, Abs); |
1111 | } |
1112 | |
1113 | /// If we have a select with an equality comparison, then we know the value in |
1114 | /// one of the arms of the select. See if substituting this value into an arm |
1115 | /// and simplifying the result yields the same value as the other arm. |
1116 | /// |
1117 | /// To make this transform safe, we must drop poison-generating flags |
1118 | /// (nsw, etc) if we simplified to a binop because the select may be guarding |
1119 | /// that poison from propagating. If the existing binop already had no |
1120 | /// poison-generating flags, then this transform can be done by instsimplify. |
1121 | /// |
1122 | /// Consider: |
1123 | /// %cmp = icmp eq i32 %x, 2147483647 |
1124 | /// %add = add nsw i32 %x, 1 |
1125 | /// %sel = select i1 %cmp, i32 -2147483648, i32 %add |
1126 | /// |
1127 | /// We can't replace %sel with %add unless we strip away the flags. |
1128 | /// TODO: Wrapping flags could be preserved in some cases with better analysis. |
1129 | Instruction *InstCombinerImpl::foldSelectValueEquivalence(SelectInst &Sel, |
1130 | ICmpInst &Cmp) { |
1131 | // Value equivalence substitution requires an all-or-nothing replacement. |
1132 | // It does not make sense for a vector compare where each lane is chosen |
1133 | // independently. |
1134 | if (!Cmp.isEquality() || Cmp.getType()->isVectorTy()) |
1135 | return nullptr; |
1136 | |
1137 | // Canonicalize the pattern to ICMP_EQ by swapping the select operands. |
1138 | Value *TrueVal = Sel.getTrueValue(), *FalseVal = Sel.getFalseValue(); |
1139 | bool Swapped = false; |
1140 | if (Cmp.getPredicate() == ICmpInst::ICMP_NE) { |
1141 | std::swap(TrueVal, FalseVal); |
1142 | Swapped = true; |
1143 | } |
1144 | |
1145 | // In X == Y ? f(X) : Z, try to evaluate f(Y) and replace the operand. |
1146 | // Make sure Y cannot be undef though, as we might pick different values for |
1147 | // undef in the icmp and in f(Y). Additionally, take care to avoid replacing |
1148 | // X == Y ? X : Z with X == Y ? Y : Z, as that would lead to an infinite |
1149 | // replacement cycle. |
1150 | Value *CmpLHS = Cmp.getOperand(0), *CmpRHS = Cmp.getOperand(1); |
1151 | if (TrueVal != CmpLHS && |
1152 | isGuaranteedNotToBeUndefOrPoison(CmpRHS, SQ.AC, &Sel, &DT)) { |
1153 | if (Value *V = simplifyWithOpReplaced(TrueVal, CmpLHS, CmpRHS, SQ, |
1154 | /* AllowRefinement */ true)) |
1155 | return replaceOperand(Sel, Swapped ? 2 : 1, V); |
1156 | |
1157 | // Even if TrueVal does not simplify, we can directly replace a use of |
1158 | // CmpLHS with CmpRHS, as long as the instruction is not used anywhere |
1159 | // else and is safe to speculatively execute (we may end up executing it |
1160 | // with different operands, which should not cause side-effects or trigger |
1161 | // undefined behavior). Only do this if CmpRHS is a constant, as |
1162 | // profitability is not clear for other cases. |
1163 | // FIXME: The replacement could be performed recursively. |
1164 | if (match(CmpRHS, m_ImmConstant()) && !match(CmpLHS, m_ImmConstant())) |
1165 | if (auto *I = dyn_cast<Instruction>(TrueVal)) |
1166 | if (I->hasOneUse() && isSafeToSpeculativelyExecute(I)) |
1167 | for (Use &U : I->operands()) |
1168 | if (U == CmpLHS) { |
1169 | replaceUse(U, CmpRHS); |
1170 | return &Sel; |
1171 | } |
1172 | } |
1173 | if (TrueVal != CmpRHS && |
1174 | isGuaranteedNotToBeUndefOrPoison(CmpLHS, SQ.AC, &Sel, &DT)) |
1175 | if (Value *V = simplifyWithOpReplaced(TrueVal, CmpRHS, CmpLHS, SQ, |
1176 | /* AllowRefinement */ true)) |
1177 | return replaceOperand(Sel, Swapped ? 2 : 1, V); |
1178 | |
1179 | auto *FalseInst = dyn_cast<Instruction>(FalseVal); |
1180 | if (!FalseInst) |
1181 | return nullptr; |
1182 | |
1183 | // InstSimplify already performed this fold if it was possible subject to |
1184 | // current poison-generating flags. Try the transform again with |
1185 | // poison-generating flags temporarily dropped. |
1186 | bool WasNUW = false, WasNSW = false, WasExact = false, WasInBounds = false; |
1187 | if (auto *OBO = dyn_cast<OverflowingBinaryOperator>(FalseVal)) { |
1188 | WasNUW = OBO->hasNoUnsignedWrap(); |
1189 | WasNSW = OBO->hasNoSignedWrap(); |
1190 | FalseInst->setHasNoUnsignedWrap(false); |
1191 | FalseInst->setHasNoSignedWrap(false); |
1192 | } |
1193 | if (auto *PEO = dyn_cast<PossiblyExactOperator>(FalseVal)) { |
1194 | WasExact = PEO->isExact(); |
1195 | FalseInst->setIsExact(false); |
1196 | } |
1197 | if (auto *GEP = dyn_cast<GetElementPtrInst>(FalseVal)) { |
1198 | WasInBounds = GEP->isInBounds(); |
1199 | GEP->setIsInBounds(false); |
1200 | } |
1201 | |
1202 | // Try each equivalence substitution possibility. |
1203 | // We have an 'EQ' comparison, so the select's false value will propagate. |
1204 | // Example: |
1205 | // (X == 42) ? 43 : (X + 1) --> (X == 42) ? (X + 1) : (X + 1) --> X + 1 |
1206 | if (simplifyWithOpReplaced(FalseVal, CmpLHS, CmpRHS, SQ, |
1207 | /* AllowRefinement */ false) == TrueVal || |
1208 | simplifyWithOpReplaced(FalseVal, CmpRHS, CmpLHS, SQ, |
1209 | /* AllowRefinement */ false) == TrueVal) { |
1210 | return replaceInstUsesWith(Sel, FalseVal); |
1211 | } |
1212 | |
1213 | // Restore poison-generating flags if the transform did not apply. |
1214 | if (WasNUW) |
1215 | FalseInst->setHasNoUnsignedWrap(); |
1216 | if (WasNSW) |
1217 | FalseInst->setHasNoSignedWrap(); |
1218 | if (WasExact) |
1219 | FalseInst->setIsExact(); |
1220 | if (WasInBounds) |
1221 | cast<GetElementPtrInst>(FalseInst)->setIsInBounds(); |
1222 | |
1223 | return nullptr; |
1224 | } |
1225 | |
1226 | // See if this is a pattern like: |
1227 | // %old_cmp1 = icmp slt i32 %x, C2 |
1228 | // %old_replacement = select i1 %old_cmp1, i32 %target_low, i32 %target_high |
1229 | // %old_x_offseted = add i32 %x, C1 |
1230 | // %old_cmp0 = icmp ult i32 %old_x_offseted, C0 |
1231 | // %r = select i1 %old_cmp0, i32 %x, i32 %old_replacement |
1232 | // This can be rewritten as more canonical pattern: |
1233 | // %new_cmp1 = icmp slt i32 %x, -C1 |
1234 | // %new_cmp2 = icmp sge i32 %x, C0-C1 |
1235 | // %new_clamped_low = select i1 %new_cmp1, i32 %target_low, i32 %x |
1236 | // %r = select i1 %new_cmp2, i32 %target_high, i32 %new_clamped_low |
1237 | // Iff -C1 s<= C2 s<= C0-C1 |
1238 | // Also ULT predicate can also be UGT iff C0 != -1 (+invert result) |
1239 | // SLT predicate can also be SGT iff C2 != INT_MAX (+invert res.) |
1240 | static Instruction *canonicalizeClampLike(SelectInst &Sel0, ICmpInst &Cmp0, |
1241 | InstCombiner::BuilderTy &Builder) { |
1242 | Value *X = Sel0.getTrueValue(); |
1243 | Value *Sel1 = Sel0.getFalseValue(); |
1244 | |
1245 | // First match the condition of the outermost select. |
1246 | // Said condition must be one-use. |
1247 | if (!Cmp0.hasOneUse()) |
1248 | return nullptr; |
1249 | Value *Cmp00 = Cmp0.getOperand(0); |
1250 | Constant *C0; |
1251 | if (!match(Cmp0.getOperand(1), |
1252 | m_CombineAnd(m_AnyIntegralConstant(), m_Constant(C0)))) |
1253 | return nullptr; |
1254 | // Canonicalize Cmp0 into the form we expect. |
1255 | // FIXME: we shouldn't care about lanes that are 'undef' in the end? |
1256 | switch (Cmp0.getPredicate()) { |
1257 | case ICmpInst::Predicate::ICMP_ULT: |
1258 | break; // Great! |
1259 | case ICmpInst::Predicate::ICMP_ULE: |
1260 | // We'd have to increment C0 by one, and for that it must not have all-ones |
1261 | // element, but then it would have been canonicalized to 'ult' before |
1262 | // we get here. So we can't do anything useful with 'ule'. |
1263 | return nullptr; |
1264 | case ICmpInst::Predicate::ICMP_UGT: |
1265 | // We want to canonicalize it to 'ult', so we'll need to increment C0, |
1266 | // which again means it must not have any all-ones elements. |
1267 | if (!match(C0, |
1268 | m_SpecificInt_ICMP(ICmpInst::Predicate::ICMP_NE, |
1269 | APInt::getAllOnesValue( |
1270 | C0->getType()->getScalarSizeInBits())))) |
1271 | return nullptr; // Can't do, have all-ones element[s]. |
1272 | C0 = InstCombiner::AddOne(C0); |
1273 | std::swap(X, Sel1); |
1274 | break; |
1275 | case ICmpInst::Predicate::ICMP_UGE: |
1276 | // The only way we'd get this predicate if this `icmp` has extra uses, |
1277 | // but then we won't be able to do this fold. |
1278 | return nullptr; |
1279 | default: |
1280 | return nullptr; // Unknown predicate. |
1281 | } |
1282 | |
1283 | // Now that we've canonicalized the ICmp, we know the X we expect; |
1284 | // the select in other hand should be one-use. |
1285 | if (!Sel1->hasOneUse()) |
1286 | return nullptr; |
1287 | |
1288 | // We now can finish matching the condition of the outermost select: |
1289 | // it should either be the X itself, or an addition of some constant to X. |
1290 | Constant *C1; |
1291 | if (Cmp00 == X) |
1292 | C1 = ConstantInt::getNullValue(Sel0.getType()); |
1293 | else if (!match(Cmp00, |
1294 | m_Add(m_Specific(X), |
1295 | m_CombineAnd(m_AnyIntegralConstant(), m_Constant(C1))))) |
1296 | return nullptr; |
1297 | |
1298 | Value *Cmp1; |
1299 | ICmpInst::Predicate Pred1; |
1300 | Constant *C2; |
1301 | Value *ReplacementLow, *ReplacementHigh; |
1302 | if (!match(Sel1, m_Select(m_Value(Cmp1), m_Value(ReplacementLow), |
1303 | m_Value(ReplacementHigh))) || |
1304 | !match(Cmp1, |
1305 | m_ICmp(Pred1, m_Specific(X), |
1306 | m_CombineAnd(m_AnyIntegralConstant(), m_Constant(C2))))) |
1307 | return nullptr; |
1308 | |
1309 | if (!Cmp1->hasOneUse() && (Cmp00 == X || !Cmp00->hasOneUse())) |
1310 | return nullptr; // Not enough one-use instructions for the fold. |
1311 | // FIXME: this restriction could be relaxed if Cmp1 can be reused as one of |
1312 | // two comparisons we'll need to build. |
1313 | |
1314 | // Canonicalize Cmp1 into the form we expect. |
1315 | // FIXME: we shouldn't care about lanes that are 'undef' in the end? |
1316 | switch (Pred1) { |
1317 | case ICmpInst::Predicate::ICMP_SLT: |
1318 | break; |
1319 | case ICmpInst::Predicate::ICMP_SLE: |
1320 | // We'd have to increment C2 by one, and for that it must not have signed |
1321 | // max element, but then it would have been canonicalized to 'slt' before |
1322 | // we get here. So we can't do anything useful with 'sle'. |
1323 | return nullptr; |
1324 | case ICmpInst::Predicate::ICMP_SGT: |
1325 | // We want to canonicalize it to 'slt', so we'll need to increment C2, |
1326 | // which again means it must not have any signed max elements. |
1327 | if (!match(C2, |
1328 | m_SpecificInt_ICMP(ICmpInst::Predicate::ICMP_NE, |
1329 | APInt::getSignedMaxValue( |
1330 | C2->getType()->getScalarSizeInBits())))) |
1331 | return nullptr; // Can't do, have signed max element[s]. |
1332 | C2 = InstCombiner::AddOne(C2); |
1333 | LLVM_FALLTHROUGH[[gnu::fallthrough]]; |
1334 | case ICmpInst::Predicate::ICMP_SGE: |
1335 | // Also non-canonical, but here we don't need to change C2, |
1336 | // so we don't have any restrictions on C2, so we can just handle it. |
1337 | std::swap(ReplacementLow, ReplacementHigh); |
1338 | break; |
1339 | default: |
1340 | return nullptr; // Unknown predicate. |
1341 | } |
1342 | |
1343 | // The thresholds of this clamp-like pattern. |
1344 | auto *ThresholdLowIncl = ConstantExpr::getNeg(C1); |
1345 | auto *ThresholdHighExcl = ConstantExpr::getSub(C0, C1); |
1346 | |
1347 | // The fold has a precondition 1: C2 s>= ThresholdLow |
1348 | auto *Precond1 = ConstantExpr::getICmp(ICmpInst::Predicate::ICMP_SGE, C2, |
1349 | ThresholdLowIncl); |
1350 | if (!match(Precond1, m_One())) |
1351 | return nullptr; |
1352 | // The fold has a precondition 2: C2 s<= ThresholdHigh |
1353 | auto *Precond2 = ConstantExpr::getICmp(ICmpInst::Predicate::ICMP_SLE, C2, |
1354 | ThresholdHighExcl); |
1355 | if (!match(Precond2, m_One())) |
1356 | return nullptr; |
1357 | |
1358 | // All good, finally emit the new pattern. |
1359 | Value *ShouldReplaceLow = Builder.CreateICmpSLT(X, ThresholdLowIncl); |
1360 | Value *ShouldReplaceHigh = Builder.CreateICmpSGE(X, ThresholdHighExcl); |
1361 | Value *MaybeReplacedLow = |
1362 | Builder.CreateSelect(ShouldReplaceLow, ReplacementLow, X); |
1363 | Instruction *MaybeReplacedHigh = |
1364 | SelectInst::Create(ShouldReplaceHigh, ReplacementHigh, MaybeReplacedLow); |
1365 | |
1366 | return MaybeReplacedHigh; |
1367 | } |
1368 | |
1369 | // If we have |
1370 | // %cmp = icmp [canonical predicate] i32 %x, C0 |
1371 | // %r = select i1 %cmp, i32 %y, i32 C1 |
1372 | // Where C0 != C1 and %x may be different from %y, see if the constant that we |
1373 | // will have if we flip the strictness of the predicate (i.e. without changing |
1374 | // the result) is identical to the C1 in select. If it matches we can change |
1375 | // original comparison to one with swapped predicate, reuse the constant, |
1376 | // and swap the hands of select. |
1377 | static Instruction * |
1378 | tryToReuseConstantFromSelectInComparison(SelectInst &Sel, ICmpInst &Cmp, |
1379 | InstCombinerImpl &IC) { |
1380 | ICmpInst::Predicate Pred; |
1381 | Value *X; |
1382 | Constant *C0; |
1383 | if (!match(&Cmp, m_OneUse(m_ICmp( |
1384 | Pred, m_Value(X), |
1385 | m_CombineAnd(m_AnyIntegralConstant(), m_Constant(C0)))))) |
1386 | return nullptr; |
1387 | |
1388 | // If comparison predicate is non-relational, we won't be able to do anything. |
1389 | if (ICmpInst::isEquality(Pred)) |
1390 | return nullptr; |
1391 | |
1392 | // If comparison predicate is non-canonical, then we certainly won't be able |
1393 | // to make it canonical; canonicalizeCmpWithConstant() already tried. |
1394 | if (!InstCombiner::isCanonicalPredicate(Pred)) |
1395 | return nullptr; |
1396 | |
1397 | // If the [input] type of comparison and select type are different, lets abort |
1398 | // for now. We could try to compare constants with trunc/[zs]ext though. |
1399 | if (C0->getType() != Sel.getType()) |
1400 | return nullptr; |
1401 | |
1402 | // FIXME: are there any magic icmp predicate+constant pairs we must not touch? |
1403 | |
1404 | Value *SelVal0, *SelVal1; // We do not care which one is from where. |
1405 | match(&Sel, m_Select(m_Value(), m_Value(SelVal0), m_Value(SelVal1))); |
1406 | // At least one of these values we are selecting between must be a constant |
1407 | // else we'll never succeed. |
1408 | if (!match(SelVal0, m_AnyIntegralConstant()) && |
1409 | !match(SelVal1, m_AnyIntegralConstant())) |
1410 | return nullptr; |
1411 | |
1412 | // Does this constant C match any of the `select` values? |
1413 | auto MatchesSelectValue = [SelVal0, SelVal1](Constant *C) { |
1414 | return C->isElementWiseEqual(SelVal0) || C->isElementWiseEqual(SelVal1); |
1415 | }; |
1416 | |
1417 | // If C0 *already* matches true/false value of select, we are done. |
1418 | if (MatchesSelectValue(C0)) |
1419 | return nullptr; |
1420 | |
1421 | // Check the constant we'd have with flipped-strictness predicate. |
1422 | auto FlippedStrictness = |
1423 | InstCombiner::getFlippedStrictnessPredicateAndConstant(Pred, C0); |
1424 | if (!FlippedStrictness) |
1425 | return nullptr; |
1426 | |
1427 | // If said constant doesn't match either, then there is no hope, |
1428 | if (!MatchesSelectValue(FlippedStrictness->second)) |
1429 | return nullptr; |
1430 | |
1431 | // It matched! Lets insert the new comparison just before select. |
1432 | InstCombiner::BuilderTy::InsertPointGuard Guard(IC.Builder); |
1433 | IC.Builder.SetInsertPoint(&Sel); |
1434 | |
1435 | Pred = ICmpInst::getSwappedPredicate(Pred); // Yes, swapped. |
1436 | Value *NewCmp = IC.Builder.CreateICmp(Pred, X, FlippedStrictness->second, |
1437 | Cmp.getName() + ".inv"); |
1438 | IC.replaceOperand(Sel, 0, NewCmp); |
1439 | Sel.swapValues(); |
1440 | Sel.swapProfMetadata(); |
1441 | |
1442 | return &Sel; |
1443 | } |
1444 | |
1445 | /// Visit a SelectInst that has an ICmpInst as its first operand. |
1446 | Instruction *InstCombinerImpl::foldSelectInstWithICmp(SelectInst &SI, |
1447 | ICmpInst *ICI) { |
1448 | if (Instruction *NewSel = foldSelectValueEquivalence(SI, *ICI)) |
1449 | return NewSel; |
1450 | |
1451 | if (Instruction *NewSel = canonicalizeMinMaxWithConstant(SI, *ICI, *this)) |
1452 | return NewSel; |
1453 | |
1454 | if (Instruction *NewAbs = canonicalizeAbsNabs(SI, *ICI, *this)) |
1455 | return NewAbs; |
1456 | |
1457 | if (Instruction *NewAbs = canonicalizeClampLike(SI, *ICI, Builder)) |
1458 | return NewAbs; |
1459 | |
1460 | if (Instruction *NewSel = |
1461 | tryToReuseConstantFromSelectInComparison(SI, *ICI, *this)) |
1462 | return NewSel; |
1463 | |
1464 | bool Changed = adjustMinMax(SI, *ICI); |
1465 | |
1466 | if (Value *V = foldSelectICmpAnd(SI, ICI, Builder)) |
1467 | return replaceInstUsesWith(SI, V); |
1468 | |
1469 | // NOTE: if we wanted to, this is where to detect integer MIN/MAX |
1470 | Value *TrueVal = SI.getTrueValue(); |
1471 | Value *FalseVal = SI.getFalseValue(); |
1472 | ICmpInst::Predicate Pred = ICI->getPredicate(); |
1473 | Value *CmpLHS = ICI->getOperand(0); |
1474 | Value *CmpRHS = ICI->getOperand(1); |
1475 | if (CmpRHS != CmpLHS && isa<Constant>(CmpRHS)) { |
1476 | if (CmpLHS == TrueVal && Pred == ICmpInst::ICMP_EQ) { |
1477 | // Transform (X == C) ? X : Y -> (X == C) ? C : Y |
1478 | SI.setOperand(1, CmpRHS); |
1479 | Changed = true; |
1480 | } else if (CmpLHS == FalseVal && Pred == ICmpInst::ICMP_NE) { |
1481 | // Transform (X != C) ? Y : X -> (X != C) ? Y : C |
1482 | SI.setOperand(2, CmpRHS); |
1483 | Changed = true; |
1484 | } |
1485 | } |
1486 | |
1487 | // FIXME: This code is nearly duplicated in InstSimplify. Using/refactoring |
1488 | // decomposeBitTestICmp() might help. |
1489 | { |
1490 | unsigned BitWidth = |
1491 | DL.getTypeSizeInBits(TrueVal->getType()->getScalarType()); |
1492 | APInt MinSignedValue = APInt::getSignedMinValue(BitWidth); |
1493 | Value *X; |
1494 | const APInt *Y, *C; |
1495 | bool TrueWhenUnset; |
1496 | bool IsBitTest = false; |
1497 | if (ICmpInst::isEquality(Pred) && |
1498 | match(CmpLHS, m_And(m_Value(X), m_Power2(Y))) && |
1499 | match(CmpRHS, m_Zero())) { |
1500 | IsBitTest = true; |
1501 | TrueWhenUnset = Pred == ICmpInst::ICMP_EQ; |
1502 | } else if (Pred == ICmpInst::ICMP_SLT && match(CmpRHS, m_Zero())) { |
1503 | X = CmpLHS; |
1504 | Y = &MinSignedValue; |
1505 | IsBitTest = true; |
1506 | TrueWhenUnset = false; |
1507 | } else if (Pred == ICmpInst::ICMP_SGT && match(CmpRHS, m_AllOnes())) { |
1508 | X = CmpLHS; |
1509 | Y = &MinSignedValue; |
1510 | IsBitTest = true; |
1511 | TrueWhenUnset = true; |
1512 | } |
1513 | if (IsBitTest) { |
1514 | Value *V = nullptr; |
1515 | // (X & Y) == 0 ? X : X ^ Y --> X & ~Y |
1516 | if (TrueWhenUnset && TrueVal == X && |
1517 | match(FalseVal, m_Xor(m_Specific(X), m_APInt(C))) && *Y == *C) |
1518 | V = Builder.CreateAnd(X, ~(*Y)); |
1519 | // (X & Y) != 0 ? X ^ Y : X --> X & ~Y |
1520 | else if (!TrueWhenUnset && FalseVal == X && |
1521 | match(TrueVal, m_Xor(m_Specific(X), m_APInt(C))) && *Y == *C) |
1522 | V = Builder.CreateAnd(X, ~(*Y)); |
1523 | // (X & Y) == 0 ? X ^ Y : X --> X | Y |
1524 | else if (TrueWhenUnset && FalseVal == X && |
1525 | match(TrueVal, m_Xor(m_Specific(X), m_APInt(C))) && *Y == *C) |
1526 | V = Builder.CreateOr(X, *Y); |
1527 | // (X & Y) != 0 ? X : X ^ Y --> X | Y |
1528 | else if (!TrueWhenUnset && TrueVal == X && |
1529 | match(FalseVal, m_Xor(m_Specific(X), m_APInt(C))) && *Y == *C) |
1530 | V = Builder.CreateOr(X, *Y); |
1531 | |
1532 | if (V) |
1533 | return replaceInstUsesWith(SI, V); |
1534 | } |
1535 | } |
1536 | |
1537 | if (Instruction *V = |
1538 | foldSelectICmpAndAnd(SI.getType(), ICI, TrueVal, FalseVal, Builder)) |
1539 | return V; |
1540 | |
1541 | if (Instruction *V = foldSelectCtlzToCttz(ICI, TrueVal, FalseVal, Builder)) |
1542 | return V; |
1543 | |
1544 | if (Value *V = foldSelectICmpAndOr(ICI, TrueVal, FalseVal, Builder)) |
1545 | return replaceInstUsesWith(SI, V); |
1546 | |
1547 | if (Value *V = foldSelectICmpLshrAshr(ICI, TrueVal, FalseVal, Builder)) |
1548 | return replaceInstUsesWith(SI, V); |
1549 | |
1550 | if (Value *V = foldSelectCttzCtlz(ICI, TrueVal, FalseVal, Builder)) |
1551 | return replaceInstUsesWith(SI, V); |
1552 | |
1553 | if (Value *V = canonicalizeSaturatedSubtract(ICI, TrueVal, FalseVal, Builder)) |
1554 | return replaceInstUsesWith(SI, V); |
1555 | |
1556 | if (Value *V = canonicalizeSaturatedAdd(ICI, TrueVal, FalseVal, Builder)) |
1557 | return replaceInstUsesWith(SI, V); |
1558 | |
1559 | return Changed ? &SI : nullptr; |
1560 | } |
1561 | |
1562 | /// SI is a select whose condition is a PHI node (but the two may be in |
1563 | /// different blocks). See if the true/false values (V) are live in all of the |
1564 | /// predecessor blocks of the PHI. For example, cases like this can't be mapped: |
1565 | /// |
1566 | /// X = phi [ C1, BB1], [C2, BB2] |
1567 | /// Y = add |
1568 | /// Z = select X, Y, 0 |
1569 | /// |
1570 | /// because Y is not live in BB1/BB2. |
1571 | static bool canSelectOperandBeMappingIntoPredBlock(const Value *V, |
1572 | const SelectInst &SI) { |
1573 | // If the value is a non-instruction value like a constant or argument, it |
1574 | // can always be mapped. |
1575 | const Instruction *I = dyn_cast<Instruction>(V); |
1576 | if (!I) return true; |
1577 | |
1578 | // If V is a PHI node defined in the same block as the condition PHI, we can |
1579 | // map the arguments. |
1580 | const PHINode *CondPHI = cast<PHINode>(SI.getCondition()); |
1581 | |
1582 | if (const PHINode *VP = dyn_cast<PHINode>(I)) |
1583 | if (VP->getParent() == CondPHI->getParent()) |
1584 | return true; |
1585 | |
1586 | // Otherwise, if the PHI and select are defined in the same block and if V is |
1587 | // defined in a different block, then we can transform it. |
1588 | if (SI.getParent() == CondPHI->getParent() && |
1589 | I->getParent() != CondPHI->getParent()) |
1590 | return true; |
1591 | |
1592 | // Otherwise we have a 'hard' case and we can't tell without doing more |
1593 | // detailed dominator based analysis, punt. |
1594 | return false; |
1595 | } |
1596 | |
1597 | /// We have an SPF (e.g. a min or max) of an SPF of the form: |
1598 | /// SPF2(SPF1(A, B), C) |
1599 | Instruction *InstCombinerImpl::foldSPFofSPF(Instruction *Inner, |
1600 | SelectPatternFlavor SPF1, Value *A, |
1601 | Value *B, Instruction &Outer, |
1602 | SelectPatternFlavor SPF2, |
1603 | Value *C) { |
1604 | if (Outer.getType() != Inner->getType()) |
1605 | return nullptr; |
1606 | |
1607 | if (C == A || C == B) { |
1608 | // MAX(MAX(A, B), B) -> MAX(A, B) |
1609 | // MIN(MIN(a, b), a) -> MIN(a, b) |
1610 | // TODO: This could be done in instsimplify. |
1611 | if (SPF1 == SPF2 && SelectPatternResult::isMinOrMax(SPF1)) |
1612 | return replaceInstUsesWith(Outer, Inner); |
1613 | |
1614 | // MAX(MIN(a, b), a) -> a |
1615 | // MIN(MAX(a, b), a) -> a |
1616 | // TODO: This could be done in instsimplify. |
1617 | if ((SPF1 == SPF_SMIN && SPF2 == SPF_SMAX) || |
1618 | (SPF1 == SPF_SMAX && SPF2 == SPF_SMIN) || |
1619 | (SPF1 == SPF_UMIN && SPF2 == SPF_UMAX) || |
1620 | (SPF1 == SPF_UMAX && SPF2 == SPF_UMIN)) |
1621 | return replaceInstUsesWith(Outer, C); |
1622 | } |
1623 | |
1624 | if (SPF1 == SPF2) { |
1625 | const APInt *CB, *CC; |
1626 | if (match(B, m_APInt(CB)) && match(C, m_APInt(CC))) { |
1627 | // MIN(MIN(A, 23), 97) -> MIN(A, 23) |
1628 | // MAX(MAX(A, 97), 23) -> MAX(A, 97) |
1629 | // TODO: This could be done in instsimplify. |
1630 | if ((SPF1 == SPF_UMIN && CB->ule(*CC)) || |
1631 | (SPF1 == SPF_SMIN && CB->sle(*CC)) || |
1632 | (SPF1 == SPF_UMAX && CB->uge(*CC)) || |
1633 | (SPF1 == SPF_SMAX && CB->sge(*CC))) |
1634 | return replaceInstUsesWith(Outer, Inner); |
1635 | |
1636 | // MIN(MIN(A, 97), 23) -> MIN(A, 23) |
1637 | // MAX(MAX(A, 23), 97) -> MAX(A, 97) |
1638 | if ((SPF1 == SPF_UMIN && CB->ugt(*CC)) || |
1639 | (SPF1 == SPF_SMIN && CB->sgt(*CC)) || |
1640 | (SPF1 == SPF_UMAX && CB->ult(*CC)) || |
1641 | (SPF1 == SPF_SMAX && CB->slt(*CC))) { |
1642 | Outer.replaceUsesOfWith(Inner, A); |
1643 | return &Outer; |
1644 | } |
1645 | } |
1646 | } |
1647 | |
1648 | // max(max(A, B), min(A, B)) --> max(A, B) |
1649 | // min(min(A, B), max(A, B)) --> min(A, B) |
1650 | // TODO: This could be done in instsimplify. |
1651 | if (SPF1 == SPF2 && |
1652 | ((SPF1 == SPF_UMIN && match(C, m_c_UMax(m_Specific(A), m_Specific(B)))) || |
1653 | (SPF1 == SPF_SMIN && match(C, m_c_SMax(m_Specific(A), m_Specific(B)))) || |
1654 | (SPF1 == SPF_UMAX && match(C, m_c_UMin(m_Specific(A), m_Specific(B)))) || |
1655 | (SPF1 == SPF_SMAX && match(C, m_c_SMin(m_Specific(A), m_Specific(B)))))) |
1656 | return replaceInstUsesWith(Outer, Inner); |
1657 | |
1658 | // ABS(ABS(X)) -> ABS(X) |
1659 | // NABS(NABS(X)) -> NABS(X) |
1660 | // TODO: This could be done in instsimplify. |
1661 | if (SPF1 == SPF2 && (SPF1 == SPF_ABS || SPF1 == SPF_NABS)) { |
1662 | return replaceInstUsesWith(Outer, Inner); |
1663 | } |
1664 | |
1665 | // ABS(NABS(X)) -> ABS(X) |
1666 | // NABS(ABS(X)) -> NABS(X) |
1667 | if ((SPF1 == SPF_ABS && SPF2 == SPF_NABS) || |
1668 | (SPF1 == SPF_NABS && SPF2 == SPF_ABS)) { |
1669 | SelectInst *SI = cast<SelectInst>(Inner); |
1670 | Value *NewSI = |
1671 | Builder.CreateSelect(SI->getCondition(), SI->getFalseValue(), |
1672 | SI->getTrueValue(), SI->getName(), SI); |
1673 | return replaceInstUsesWith(Outer, NewSI); |
1674 | } |
1675 | |
1676 | auto IsFreeOrProfitableToInvert = |
1677 | [&](Value *V, Value *&NotV, bool &ElidesXor) { |
1678 | if (match(V, m_Not(m_Value(NotV)))) { |
1679 | // If V has at most 2 uses then we can get rid of the xor operation |
1680 | // entirely. |
1681 | ElidesXor |= !V->hasNUsesOrMore(3); |
1682 | return true; |
1683 | } |
1684 | |
1685 | if (isFreeToInvert(V, !V->hasNUsesOrMore(3))) { |
1686 | NotV = nullptr; |
1687 | return true; |
1688 | } |
1689 | |
1690 | return false; |
1691 | }; |
1692 | |
1693 | Value *NotA, *NotB, *NotC; |
1694 | bool ElidesXor = false; |
1695 | |
1696 | // MIN(MIN(~A, ~B), ~C) == ~MAX(MAX(A, B), C) |
1697 | // MIN(MAX(~A, ~B), ~C) == ~MAX(MIN(A, B), C) |
1698 | // MAX(MIN(~A, ~B), ~C) == ~MIN(MAX(A, B), C) |
1699 | // MAX(MAX(~A, ~B), ~C) == ~MIN(MIN(A, B), C) |
1700 | // |
1701 | // This transform is performance neutral if we can elide at least one xor from |
1702 | // the set of three operands, since we'll be tacking on an xor at the very |
1703 | // end. |
1704 | if (SelectPatternResult::isMinOrMax(SPF1) && |
1705 | SelectPatternResult::isMinOrMax(SPF2) && |
1706 | IsFreeOrProfitableToInvert(A, NotA, ElidesXor) && |
1707 | IsFreeOrProfitableToInvert(B, NotB, ElidesXor) && |
1708 | IsFreeOrProfitableToInvert(C, NotC, ElidesXor) && ElidesXor) { |
1709 | if (!NotA) |
1710 | NotA = Builder.CreateNot(A); |
1711 | if (!NotB) |
1712 | NotB = Builder.CreateNot(B); |
1713 | if (!NotC) |
1714 | NotC = Builder.CreateNot(C); |
1715 | |
1716 | Value *NewInner = createMinMax(Builder, getInverseMinMaxFlavor(SPF1), NotA, |
1717 | NotB); |
1718 | Value *NewOuter = Builder.CreateNot( |
1719 | createMinMax(Builder, getInverseMinMaxFlavor(SPF2), NewInner, NotC)); |
1720 | return replaceInstUsesWith(Outer, NewOuter); |
1721 | } |
1722 | |
1723 | return nullptr; |
1724 | } |
1725 | |
1726 | /// Turn select C, (X + Y), (X - Y) --> (X + (select C, Y, (-Y))). |
1727 | /// This is even legal for FP. |
1728 | static Instruction *foldAddSubSelect(SelectInst &SI, |
1729 | InstCombiner::BuilderTy &Builder) { |
1730 | Value *CondVal = SI.getCondition(); |
1731 | Value *TrueVal = SI.getTrueValue(); |
1732 | Value *FalseVal = SI.getFalseValue(); |
1733 | auto *TI = dyn_cast<Instruction>(TrueVal); |
1734 | auto *FI = dyn_cast<Instruction>(FalseVal); |
1735 | if (!TI || !FI || !TI->hasOneUse() || !FI->hasOneUse()) |
1736 | return nullptr; |
1737 | |
1738 | Instruction *AddOp = nullptr, *SubOp = nullptr; |
1739 | if ((TI->getOpcode() == Instruction::Sub && |
1740 | FI->getOpcode() == Instruction::Add) || |
1741 | (TI->getOpcode() == Instruction::FSub && |
1742 | FI->getOpcode() == Instruction::FAdd)) { |
1743 | AddOp = FI; |
1744 | SubOp = TI; |
1745 | } else if ((FI->getOpcode() == Instruction::Sub && |
1746 | TI->getOpcode() == Instruction::Add) || |
1747 | (FI->getOpcode() == Instruction::FSub && |
1748 | TI->getOpcode() == Instruction::FAdd)) { |
1749 | AddOp = TI; |
1750 | SubOp = FI; |
1751 | } |
1752 | |
1753 | if (AddOp) { |
1754 | Value *OtherAddOp = nullptr; |
1755 | if (SubOp->getOperand(0) == AddOp->getOperand(0)) { |
1756 | OtherAddOp = AddOp->getOperand(1); |
1757 | } else if (SubOp->getOperand(0) == AddOp->getOperand(1)) { |
1758 | OtherAddOp = AddOp->getOperand(0); |
1759 | } |
1760 | |
1761 | if (OtherAddOp) { |
1762 | // So at this point we know we have (Y -> OtherAddOp): |
1763 | // select C, (add X, Y), (sub X, Z) |
1764 | Value *NegVal; // Compute -Z |
1765 | if (SI.getType()->isFPOrFPVectorTy()) { |
1766 | NegVal = Builder.CreateFNeg(SubOp->getOperand(1)); |
1767 | if (Instruction *NegInst = dyn_cast<Instruction>(NegVal)) { |
1768 | FastMathFlags Flags = AddOp->getFastMathFlags(); |
1769 | Flags &= SubOp->getFastMathFlags(); |
1770 | NegInst->setFastMathFlags(Flags); |
1771 | } |
1772 | } else { |
1773 | NegVal = Builder.CreateNeg(SubOp->getOperand(1)); |
1774 | } |
1775 | |
1776 | Value *NewTrueOp = OtherAddOp; |
1777 | Value *NewFalseOp = NegVal; |
1778 | if (AddOp != TI) |
1779 | std::swap(NewTrueOp, NewFalseOp); |
1780 | Value *NewSel = Builder.CreateSelect(CondVal, NewTrueOp, NewFalseOp, |
1781 | SI.getName() + ".p", &SI); |
1782 | |
1783 | if (SI.getType()->isFPOrFPVectorTy()) { |
1784 | Instruction *RI = |
1785 | BinaryOperator::CreateFAdd(SubOp->getOperand(0), NewSel); |
1786 | |
1787 | FastMathFlags Flags = AddOp->getFastMathFlags(); |
1788 | Flags &= SubOp->getFastMathFlags(); |
1789 | RI->setFastMathFlags(Flags); |
1790 | return RI; |
1791 | } else |
1792 | return BinaryOperator::CreateAdd(SubOp->getOperand(0), NewSel); |
1793 | } |
1794 | } |
1795 | return nullptr; |
1796 | } |
1797 | |
1798 | /// Turn X + Y overflows ? -1 : X + Y -> uadd_sat X, Y |
1799 | /// And X - Y overflows ? 0 : X - Y -> usub_sat X, Y |
1800 | /// Along with a number of patterns similar to: |
1801 | /// X + Y overflows ? (X < 0 ? INTMIN : INTMAX) : X + Y --> sadd_sat X, Y |
1802 | /// X - Y overflows ? (X > 0 ? INTMAX : INTMIN) : X - Y --> ssub_sat X, Y |
1803 | static Instruction * |
1804 | foldOverflowingAddSubSelect(SelectInst &SI, InstCombiner::BuilderTy &Builder) { |
1805 | Value *CondVal = SI.getCondition(); |
1806 | Value *TrueVal = SI.getTrueValue(); |
1807 | Value *FalseVal = SI.getFalseValue(); |
1808 | |
1809 | WithOverflowInst *II; |
1810 | if (!match(CondVal, m_ExtractValue<1>(m_WithOverflowInst(II))) || |
1811 | !match(FalseVal, m_ExtractValue<0>(m_Specific(II)))) |
1812 | return nullptr; |
1813 | |
1814 | Value *X = II->getLHS(); |
1815 | Value *Y = II->getRHS(); |
1816 | |
1817 | auto IsSignedSaturateLimit = [&](Value *Limit, bool IsAdd) { |
1818 | Type *Ty = Limit->getType(); |
1819 | |
1820 | ICmpInst::Predicate Pred; |
1821 | Value *TrueVal, *FalseVal, *Op; |
1822 | const APInt *C; |
1823 | if (!match(Limit, m_Select(m_ICmp(Pred, m_Value(Op), m_APInt(C)), |
1824 | m_Value(TrueVal), m_Value(FalseVal)))) |
1825 | return false; |
1826 | |
1827 | auto IsZeroOrOne = [](const APInt &C) { |
1828 | return C.isNullValue() || C.isOneValue(); |
1829 | }; |
1830 | auto IsMinMax = [&](Value *Min, Value *Max) { |
1831 | APInt MinVal = APInt::getSignedMinValue(Ty->getScalarSizeInBits()); |
1832 | APInt MaxVal = APInt::getSignedMaxValue(Ty->getScalarSizeInBits()); |
1833 | return match(Min, m_SpecificInt(MinVal)) && |
1834 | match(Max, m_SpecificInt(MaxVal)); |
1835 | }; |
1836 | |
1837 | if (Op != X && Op != Y) |
1838 | return false; |
1839 | |
1840 | if (IsAdd) { |
1841 | // X + Y overflows ? (X <s 0 ? INTMIN : INTMAX) : X + Y --> sadd_sat X, Y |
1842 | // X + Y overflows ? (X <s 1 ? INTMIN : INTMAX) : X + Y --> sadd_sat X, Y |
1843 | // X + Y overflows ? (Y <s 0 ? INTMIN : INTMAX) : X + Y --> sadd_sat X, Y |
1844 | // X + Y overflows ? (Y <s 1 ? INTMIN : INTMAX) : X + Y --> sadd_sat X, Y |
1845 | if (Pred == ICmpInst::ICMP_SLT && IsZeroOrOne(*C) && |
1846 | IsMinMax(TrueVal, FalseVal)) |
1847 | return true; |
1848 | // X + Y overflows ? (X >s 0 ? INTMAX : INTMIN) : X + Y --> sadd_sat X, Y |
1849 | // X + Y overflows ? (X >s -1 ? INTMAX : INTMIN) : X + Y --> sadd_sat X, Y |
1850 | // X + Y overflows ? (Y >s 0 ? INTMAX : INTMIN) : X + Y --> sadd_sat X, Y |
1851 | // X + Y overflows ? (Y >s -1 ? INTMAX : INTMIN) : X + Y --> sadd_sat X, Y |
1852 | if (Pred == ICmpInst::ICMP_SGT && IsZeroOrOne(*C + 1) && |
1853 | IsMinMax(FalseVal, TrueVal)) |
1854 | return true; |
1855 | } else { |
1856 | // X - Y overflows ? (X <s 0 ? INTMIN : INTMAX) : X - Y --> ssub_sat X, Y |
1857 | // X - Y overflows ? (X <s -1 ? INTMIN : INTMAX) : X - Y --> ssub_sat X, Y |
1858 | if (Op == X && Pred == ICmpInst::ICMP_SLT && IsZeroOrOne(*C + 1) && |
1859 | IsMinMax(TrueVal, FalseVal)) |
1860 | return true; |
1861 | // X - Y overflows ? (X >s -1 ? INTMAX : INTMIN) : X - Y --> ssub_sat X, Y |
1862 | // X - Y overflows ? (X >s -2 ? INTMAX : INTMIN) : X - Y --> ssub_sat X, Y |
1863 | if (Op == X && Pred == ICmpInst::ICMP_SGT && IsZeroOrOne(*C + 2) && |
1864 | IsMinMax(FalseVal, TrueVal)) |
1865 | return true; |
1866 | // X - Y overflows ? (Y <s 0 ? INTMAX : INTMIN) : X - Y --> ssub_sat X, Y |
1867 | // X - Y overflows ? (Y <s 1 ? INTMAX : INTMIN) : X - Y --> ssub_sat X, Y |
1868 | if (Op == Y && Pred == ICmpInst::ICMP_SLT && IsZeroOrOne(*C) && |
1869 | IsMinMax(FalseVal, TrueVal)) |
1870 | return true; |
1871 | // X - Y overflows ? (Y >s 0 ? INTMIN : INTMAX) : X - Y --> ssub_sat X, Y |
1872 | // X - Y overflows ? (Y >s -1 ? INTMIN : INTMAX) : X - Y --> ssub_sat X, Y |
1873 | if (Op == Y && Pred == ICmpInst::ICMP_SGT && IsZeroOrOne(*C + 1) && |
1874 | IsMinMax(TrueVal, FalseVal)) |
1875 | return true; |
1876 | } |
1877 | |
1878 | return false; |
1879 | }; |
1880 | |
1881 | Intrinsic::ID NewIntrinsicID; |
1882 | if (II->getIntrinsicID() == Intrinsic::uadd_with_overflow && |
1883 | match(TrueVal, m_AllOnes())) |
1884 | // X + Y overflows ? -1 : X + Y -> uadd_sat X, Y |
1885 | NewIntrinsicID = Intrinsic::uadd_sat; |
1886 | else if (II->getIntrinsicID() == Intrinsic::usub_with_overflow && |
1887 | match(TrueVal, m_Zero())) |
1888 | // X - Y overflows ? 0 : X - Y -> usub_sat X, Y |
1889 | NewIntrinsicID = Intrinsic::usub_sat; |
1890 | else if (II->getIntrinsicID() == Intrinsic::sadd_with_overflow && |
1891 | IsSignedSaturateLimit(TrueVal, /*IsAdd=*/true)) |
1892 | // X + Y overflows ? (X <s 0 ? INTMIN : INTMAX) : X + Y --> sadd_sat X, Y |
1893 | // X + Y overflows ? (X <s 1 ? INTMIN : INTMAX) : X + Y --> sadd_sat X, Y |
1894 | // X + Y overflows ? (X >s 0 ? INTMAX : INTMIN) : X + Y --> sadd_sat X, Y |
1895 | // X + Y overflows ? (X >s -1 ? INTMAX : INTMIN) : X + Y --> sadd_sat X, Y |
1896 | // X + Y overflows ? (Y <s 0 ? INTMIN : INTMAX) : X + Y --> sadd_sat X, Y |
1897 | // X + Y overflows ? (Y <s 1 ? INTMIN : INTMAX) : X + Y --> sadd_sat X, Y |
1898 | // X + Y overflows ? (Y >s 0 ? INTMAX : INTMIN) : X + Y --> sadd_sat X, Y |
1899 | // X + Y overflows ? (Y >s -1 ? INTMAX : INTMIN) : X + Y --> sadd_sat X, Y |
1900 | NewIntrinsicID = Intrinsic::sadd_sat; |
1901 | else if (II->getIntrinsicID() == Intrinsic::ssub_with_overflow && |
1902 | IsSignedSaturateLimit(TrueVal, /*IsAdd=*/false)) |
1903 | // X - Y overflows ? (X <s 0 ? INTMIN : INTMAX) : X - Y --> ssub_sat X, Y |
1904 | // X - Y overflows ? (X <s -1 ? INTMIN : INTMAX) : X - Y --> ssub_sat X, Y |
1905 | // X - Y overflows ? (X >s -1 ? INTMAX : INTMIN) : X - Y --> ssub_sat X, Y |
1906 | // X - Y overflows ? (X >s -2 ? INTMAX : INTMIN) : X - Y --> ssub_sat X, Y |
1907 | // X - Y overflows ? (Y <s 0 ? INTMAX : INTMIN) : X - Y --> ssub_sat X, Y |
1908 | // X - Y overflows ? (Y <s 1 ? INTMAX : INTMIN) : X - Y --> ssub_sat X, Y |
1909 | // X - Y overflows ? (Y >s 0 ? INTMIN : INTMAX) : X - Y --> ssub_sat X, Y |
1910 | // X - Y overflows ? (Y >s -1 ? INTMIN : INTMAX) : X - Y --> ssub_sat X, Y |
1911 | NewIntrinsicID = Intrinsic::ssub_sat; |
1912 | else |
1913 | return nullptr; |
1914 | |
1915 | Function *F = |
1916 | Intrinsic::getDeclaration(SI.getModule(), NewIntrinsicID, SI.getType()); |
1917 | return CallInst::Create(F, {X, Y}); |
1918 | } |
1919 | |
1920 | Instruction *InstCombinerImpl::foldSelectExtConst(SelectInst &Sel) { |
1921 | Constant *C; |
1922 | if (!match(Sel.getTrueValue(), m_Constant(C)) && |
1923 | !match(Sel.getFalseValue(), m_Constant(C))) |
1924 | return nullptr; |
1925 | |
1926 | Instruction *ExtInst; |
1927 | if (!match(Sel.getTrueValue(), m_Instruction(ExtInst)) && |
1928 | !match(Sel.getFalseValue(), m_Instruction(ExtInst))) |
1929 | return nullptr; |
1930 | |
1931 | auto ExtOpcode = ExtInst->getOpcode(); |
1932 | if (ExtOpcode != Instruction::ZExt && ExtOpcode != Instruction::SExt) |
1933 | return nullptr; |
1934 | |
1935 | // If we are extending from a boolean type or if we can create a select that |
1936 | // has the same size operands as its condition, try to narrow the select. |
1937 | Value *X = ExtInst->getOperand(0); |
1938 | Type *SmallType = X->getType(); |
1939 | Value *Cond = Sel.getCondition(); |
1940 | auto *Cmp = dyn_cast<CmpInst>(Cond); |
1941 | if (!SmallType->isIntOrIntVectorTy(1) && |
1942 | (!Cmp || Cmp->getOperand(0)->getType() != SmallType)) |
1943 | return nullptr; |
1944 | |
1945 | // If the constant is the same after truncation to the smaller type and |
1946 | // extension to the original type, we can narrow the select. |
1947 | Type *SelType = Sel.getType(); |
1948 | Constant *TruncC = ConstantExpr::getTrunc(C, SmallType); |
1949 | Constant *ExtC = ConstantExpr::getCast(ExtOpcode, TruncC, SelType); |
1950 | if (ExtC == C && ExtInst->hasOneUse()) { |
1951 | Value *TruncCVal = cast<Value>(TruncC); |
1952 | if (ExtInst == Sel.getFalseValue()) |
1953 | std::swap(X, TruncCVal); |
1954 | |
1955 | // select Cond, (ext X), C --> ext(select Cond, X, C') |
1956 | // select Cond, C, (ext X) --> ext(select Cond, C', X) |
1957 | Value *NewSel = Builder.CreateSelect(Cond, X, TruncCVal, "narrow", &Sel); |
1958 | return CastInst::Create(Instruction::CastOps(ExtOpcode), NewSel, SelType); |
1959 | } |
1960 | |
1961 | // If one arm of the select is the extend of the condition, replace that arm |
1962 | // with the extension of the appropriate known bool value. |
1963 | if (Cond == X) { |
1964 | if (ExtInst == Sel.getTrueValue()) { |
1965 | // select X, (sext X), C --> select X, -1, C |
1966 | // select X, (zext X), C --> select X, 1, C |
1967 | Constant *One = ConstantInt::getTrue(SmallType); |
1968 | Constant *AllOnesOrOne = ConstantExpr::getCast(ExtOpcode, One, SelType); |
1969 | return SelectInst::Create(Cond, AllOnesOrOne, C, "", nullptr, &Sel); |
1970 | } else { |
1971 | // select X, C, (sext X) --> select X, C, 0 |
1972 | // select X, C, (zext X) --> select X, C, 0 |
1973 | Constant *Zero = ConstantInt::getNullValue(SelType); |
1974 | return SelectInst::Create(Cond, C, Zero, "", nullptr, &Sel); |
1975 | } |
1976 | } |
1977 | |
1978 | return nullptr; |
1979 | } |
1980 | |
1981 | /// Try to transform a vector select with a constant condition vector into a |
1982 | /// shuffle for easier combining with other shuffles and insert/extract. |
1983 | static Instruction *canonicalizeSelectToShuffle(SelectInst &SI) { |
1984 | Value *CondVal = SI.getCondition(); |
1985 | Constant *CondC; |
1986 | auto *CondValTy = dyn_cast<FixedVectorType>(CondVal->getType()); |
1987 | if (!CondValTy || !match(CondVal, m_Constant(CondC))) |
1988 | return nullptr; |
1989 | |
1990 | unsigned NumElts = CondValTy->getNumElements(); |
1991 | SmallVector<int, 16> Mask; |
1992 | Mask.reserve(NumElts); |
1993 | for (unsigned i = 0; i != NumElts; ++i) { |
1994 | Constant *Elt = CondC->getAggregateElement(i); |
1995 | if (!Elt) |
1996 | return nullptr; |
1997 | |
1998 | if (Elt->isOneValue()) { |
1999 | // If the select condition element is true, choose from the 1st vector. |
2000 | Mask.push_back(i); |
2001 | } else if (Elt->isNullValue()) { |
2002 | // If the select condition element is false, choose from the 2nd vector. |
2003 | Mask.push_back(i + NumElts); |
2004 | } else if (isa<UndefValue>(Elt)) { |
2005 | // Undef in a select condition (choose one of the operands) does not mean |
2006 | // the same thing as undef in a shuffle mask (any value is acceptable), so |
2007 | // give up. |
2008 | return nullptr; |
2009 | } else { |
2010 | // Bail out on a constant expression. |
2011 | return nullptr; |
2012 | } |
2013 | } |
2014 | |
2015 | return new ShuffleVectorInst(SI.getTrueValue(), SI.getFalseValue(), Mask); |
2016 | } |
2017 | |
2018 | /// If we have a select of vectors with a scalar condition, try to convert that |
2019 | /// to a vector select by splatting the condition. A splat may get folded with |
2020 | /// other operations in IR and having all operands of a select be vector types |
2021 | /// is likely better for vector codegen. |
2022 | static Instruction *canonicalizeScalarSelectOfVecs(SelectInst &Sel, |
2023 | InstCombinerImpl &IC) { |
2024 | auto *Ty = dyn_cast<VectorType>(Sel.getType()); |
2025 | if (!Ty) |
2026 | return nullptr; |
2027 | |
2028 | // We can replace a single-use extract with constant index. |
2029 | Value *Cond = Sel.getCondition(); |
2030 | if (!match(Cond, m_OneUse(m_ExtractElt(m_Value(), m_ConstantInt())))) |
2031 | return nullptr; |
2032 | |
2033 | // select (extelt V, Index), T, F --> select (splat V, Index), T, F |
2034 | // Splatting the extracted condition reduces code (we could directly create a |
2035 | // splat shuffle of the source vector to eliminate the intermediate step). |
2036 | return IC.replaceOperand( |
2037 | Sel, 0, IC.Builder.CreateVectorSplat(Ty->getElementCount(), Cond)); |
2038 | } |
2039 | |
2040 | /// Reuse bitcasted operands between a compare and select: |
2041 | /// select (cmp (bitcast C), (bitcast D)), (bitcast' C), (bitcast' D) --> |
2042 | /// bitcast (select (cmp (bitcast C), (bitcast D)), (bitcast C), (bitcast D)) |
2043 | static Instruction *foldSelectCmpBitcasts(SelectInst &Sel, |
2044 | InstCombiner::BuilderTy &Builder) { |
2045 | Value *Cond = Sel.getCondition(); |
2046 | Value *TVal = Sel.getTrueValue(); |
2047 | Value *FVal = Sel.getFalseValue(); |
2048 | |
2049 | CmpInst::Predicate Pred; |
2050 | Value *A, *B; |
2051 | if (!match(Cond, m_Cmp(Pred, m_Value(A), m_Value(B)))) |
2052 | return nullptr; |
2053 | |
2054 | // The select condition is a compare instruction. If the select's true/false |
2055 | // values are already the same as the compare operands, there's nothing to do. |
2056 | if (TVal == A || TVal == B || FVal == A || FVal == B) |
2057 | return nullptr; |
2058 | |
2059 | Value *C, *D; |
2060 | if (!match(A, m_BitCast(m_Value(C))) || !match(B, m_BitCast(m_Value(D)))) |
2061 | return nullptr; |
2062 | |
2063 | // select (cmp (bitcast C), (bitcast D)), (bitcast TSrc), (bitcast FSrc) |
2064 | Value *TSrc, *FSrc; |
2065 | if (!match(TVal, m_BitCast(m_Value(TSrc))) || |
2066 | !match(FVal, m_BitCast(m_Value(FSrc)))) |
2067 | return nullptr; |
2068 | |
2069 | // If the select true/false values are *different bitcasts* of the same source |
2070 | // operands, make the select operands the same as the compare operands and |
2071 | // cast the result. This is the canonical select form for min/max. |
2072 | Value *NewSel; |
2073 | if (TSrc == C && FSrc == D) { |
2074 | // select (cmp (bitcast C), (bitcast D)), (bitcast' C), (bitcast' D) --> |
2075 | // bitcast (select (cmp A, B), A, B) |
2076 | NewSel = Builder.CreateSelect(Cond, A, B, "", &Sel); |
2077 | } else if (TSrc == D && FSrc == C) { |
2078 | // select (cmp (bitcast C), (bitcast D)), (bitcast' D), (bitcast' C) --> |
2079 | // bitcast (select (cmp A, B), B, A) |
2080 | NewSel = Builder.CreateSelect(Cond, B, A, "", &Sel); |
2081 | } else { |
2082 | return nullptr; |
2083 | } |
2084 | return CastInst::CreateBitOrPointerCast(NewSel, Sel.getType()); |
2085 | } |
2086 | |
2087 | /// Try to eliminate select instructions that test the returned flag of cmpxchg |
2088 | /// instructions. |
2089 | /// |
2090 | /// If a select instruction tests the returned flag of a cmpxchg instruction and |
2091 | /// selects between the returned value of the cmpxchg instruction its compare |
2092 | /// operand, the result of the select will always be equal to its false value. |
2093 | /// For example: |
2094 | /// |
2095 | /// %0 = cmpxchg i64* %ptr, i64 %compare, i64 %new_value seq_cst seq_cst |
2096 | /// %1 = extractvalue { i64, i1 } %0, 1 |
2097 | /// %2 = extractvalue { i64, i1 } %0, 0 |
2098 | /// %3 = select i1 %1, i64 %compare, i64 %2 |
2099 | /// ret i64 %3 |
2100 | /// |
2101 | /// The returned value of the cmpxchg instruction (%2) is the original value |
2102 | /// located at %ptr prior to any update. If the cmpxchg operation succeeds, %2 |
2103 | /// must have been equal to %compare. Thus, the result of the select is always |
2104 | /// equal to %2, and the code can be simplified to: |
2105 | /// |
2106 | /// %0 = cmpxchg i64* %ptr, i64 %compare, i64 %new_value seq_cst seq_cst |
2107 | /// %1 = extractvalue { i64, i1 } %0, 0 |
2108 | /// ret i64 %1 |
2109 | /// |
2110 | static Value *foldSelectCmpXchg(SelectInst &SI) { |
2111 | // A helper that determines if V is an extractvalue instruction whose |
2112 | // aggregate operand is a cmpxchg instruction and whose single index is equal |
2113 | // to I. If such conditions are true, the helper returns the cmpxchg |
2114 | // instruction; otherwise, a nullptr is returned. |
2115 | auto isExtractFromCmpXchg = [](Value *V, unsigned I) -> AtomicCmpXchgInst * { |
2116 | auto *Extract = dyn_cast<ExtractValueInst>(V); |
2117 | if (!Extract) |
2118 | return nullptr; |
2119 | if (Extract->getIndices()[0] != I) |
2120 | return nullptr; |
2121 | return dyn_cast<AtomicCmpXchgInst>(Extract->getAggregateOperand()); |
2122 | }; |
2123 | |
2124 | // If the select has a single user, and this user is a select instruction that |
2125 | // we can simplify, skip the cmpxchg simplification for now. |
2126 | if (SI.hasOneUse()) |
2127 | if (auto *Select = dyn_cast<SelectInst>(SI.user_back())) |
2128 | if (Select->getCondition() == SI.getCondition()) |
2129 | if (Select->getFalseValue() == SI.getTrueValue() || |
2130 | Select->getTrueValue() == SI.getFalseValue()) |
2131 | return nullptr; |
2132 | |
2133 | // Ensure the select condition is the returned flag of a cmpxchg instruction. |
2134 | auto *CmpXchg = isExtractFromCmpXchg(SI.getCondition(), 1); |
2135 | if (!CmpXchg) |
2136 | return nullptr; |
2137 | |
2138 | // Check the true value case: The true value of the select is the returned |
2139 | // value of the same cmpxchg used by the condition, and the false value is the |
2140 | // cmpxchg instruction's compare operand. |
2141 | if (auto *X = isExtractFromCmpXchg(SI.getTrueValue(), 0)) |
2142 | if (X == CmpXchg && X->getCompareOperand() == SI.getFalseValue()) |
2143 | return SI.getFalseValue(); |
2144 | |
2145 | // Check the false value case: The false value of the select is the returned |
2146 | // value of the same cmpxchg used by the condition, and the true value is the |
2147 | // cmpxchg instruction's compare operand. |
2148 | if (auto *X = isExtractFromCmpXchg(SI.getFalseValue(), 0)) |
2149 | if (X == CmpXchg && X->getCompareOperand() == SI.getTrueValue()) |
2150 | return SI.getFalseValue(); |
2151 | |
2152 | return nullptr; |
2153 | } |
2154 | |
2155 | static Instruction *moveAddAfterMinMax(SelectPatternFlavor SPF, Value *X, |
2156 | Value *Y, |
2157 | InstCombiner::BuilderTy &Builder) { |
2158 | assert(SelectPatternResult::isMinOrMax(SPF) && "Expected min/max pattern")(static_cast<void> (0)); |
2159 | bool IsUnsigned = SPF == SelectPatternFlavor::SPF_UMIN || |
2160 | SPF == SelectPatternFlavor::SPF_UMAX; |
2161 | // TODO: If InstSimplify could fold all cases where C2 <= C1, we could change |
2162 | // the constant value check to an assert. |
2163 | Value *A; |
2164 | const APInt *C1, *C2; |
2165 | if (IsUnsigned && match(X, m_NUWAdd(m_Value(A), m_APInt(C1))) && |
2166 | match(Y, m_APInt(C2)) && C2->uge(*C1) && X->hasNUses(2)) { |
2167 | // umin (add nuw A, C1), C2 --> add nuw (umin A, C2 - C1), C1 |
2168 | // umax (add nuw A, C1), C2 --> add nuw (umax A, C2 - C1), C1 |
2169 | Value *NewMinMax = createMinMax(Builder, SPF, A, |
2170 | ConstantInt::get(X->getType(), *C2 - *C1)); |
2171 | return BinaryOperator::CreateNUW(BinaryOperator::Add, NewMinMax, |
2172 | ConstantInt::get(X->getType(), *C1)); |
2173 | } |
2174 | |
2175 | if (!IsUnsigned && match(X, m_NSWAdd(m_Value(A), m_APInt(C1))) && |
2176 | match(Y, m_APInt(C2)) && X->hasNUses(2)) { |
2177 | bool Overflow; |
2178 | APInt Diff = C2->ssub_ov(*C1, Overflow); |
2179 | if (!Overflow) { |
2180 | // smin (add nsw A, C1), C2 --> add nsw (smin A, C2 - C1), C1 |
2181 | // smax (add nsw A, C1), C2 --> add nsw (smax A, C2 - C1), C1 |
2182 | Value *NewMinMax = createMinMax(Builder, SPF, A, |
2183 | ConstantInt::get(X->getType(), Diff)); |
2184 | return BinaryOperator::CreateNSW(BinaryOperator::Add, NewMinMax, |
2185 | ConstantInt::get(X->getType(), *C1)); |
2186 | } |
2187 | } |
2188 | |
2189 | return nullptr; |
2190 | } |
2191 | |
2192 | /// Match a sadd_sat or ssub_sat which is using min/max to clamp the value. |
2193 | Instruction *InstCombinerImpl::matchSAddSubSat(Instruction &MinMax1) { |
2194 | Type *Ty = MinMax1.getType(); |
2195 | |
2196 | // We are looking for a tree of: |
2197 | // max(INT_MIN, min(INT_MAX, add(sext(A), sext(B)))) |
2198 | // Where the min and max could be reversed |
2199 | Instruction *MinMax2; |
2200 | BinaryOperator *AddSub; |
2201 | const APInt *MinValue, *MaxValue; |
2202 | if (match(&MinMax1, m_SMin(m_Instruction(MinMax2), m_APInt(MaxValue)))) { |
2203 | if (!match(MinMax2, m_SMax(m_BinOp(AddSub), m_APInt(MinValue)))) |
2204 | return nullptr; |
2205 | } else if (match(&MinMax1, |
2206 | m_SMax(m_Instruction(MinMax2), m_APInt(MinValue)))) { |
2207 | if (!match(MinMax2, m_SMin(m_BinOp(AddSub), m_APInt(MaxValue)))) |
2208 | return nullptr; |
2209 | } else |
2210 | return nullptr; |
2211 | |
2212 | // Check that the constants clamp a saturate, and that the new type would be |
2213 | // sensible to convert to. |
2214 | if (!(*MaxValue + 1).isPowerOf2() || -*MinValue != *MaxValue + 1) |
2215 | return nullptr; |
2216 | // In what bitwidth can this be treated as saturating arithmetics? |
2217 | unsigned NewBitWidth = (*MaxValue + 1).logBase2() + 1; |
2218 | // FIXME: This isn't quite right for vectors, but using the scalar type is a |
2219 | // good first approximation for what should be done there. |
2220 | if (!shouldChangeType(Ty->getScalarType()->getIntegerBitWidth(), NewBitWidth)) |
2221 | return nullptr; |
2222 | |
2223 | // Also make sure that the number of uses is as expected. The 3 is for the |
2224 | // the two items of the compare and the select, or 2 from a min/max. |
2225 | unsigned ExpUses = isa<IntrinsicInst>(MinMax1) ? 2 : 3; |
2226 | if (MinMax2->hasNUsesOrMore(ExpUses) || AddSub->hasNUsesOrMore(ExpUses)) |
2227 | return nullptr; |
2228 | |
2229 | // Create the new type (which can be a vector type) |
2230 | Type *NewTy = Ty->getWithNewBitWidth(NewBitWidth); |
2231 | // Match the two extends from the add/sub |
2232 | Value *A, *B; |
2233 | if(!match(AddSub, m_BinOp(m_SExt(m_Value(A)), m_SExt(m_Value(B))))) |
2234 | return nullptr; |
2235 | // And check the incoming values are of a type smaller than or equal to the |
2236 | // size of the saturation. Otherwise the higher bits can cause different |
2237 | // results. |
2238 | if (A->getType()->getScalarSizeInBits() > NewBitWidth || |
2239 | B->getType()->getScalarSizeInBits() > NewBitWidth) |
2240 | return nullptr; |
2241 | |
2242 | Intrinsic::ID IntrinsicID; |
2243 | if (AddSub->getOpcode() == Instruction::Add) |
2244 | IntrinsicID = Intrinsic::sadd_sat; |
2245 | else if (AddSub->getOpcode() == Instruction::Sub) |
2246 | IntrinsicID = Intrinsic::ssub_sat; |
2247 | else |
2248 | return nullptr; |
2249 | |
2250 | // Finally create and return the sat intrinsic, truncated to the new type |
2251 | Function *F = Intrinsic::getDeclaration(MinMax1.getModule(), IntrinsicID, NewTy); |
2252 | Value *AT = Builder.CreateSExt(A, NewTy); |
2253 | Value *BT = Builder.CreateSExt(B, NewTy); |
2254 | Value *Sat = Builder.CreateCall(F, {AT, BT}); |
2255 | return CastInst::Create(Instruction::SExt, Sat, Ty); |
2256 | } |
2257 | |
2258 | /// Reduce a sequence of min/max with a common operand. |
2259 | static Instruction *factorizeMinMaxTree(SelectPatternFlavor SPF, Value *LHS, |
2260 | Value *RHS, |
2261 | InstCombiner::BuilderTy &Builder) { |
2262 | assert(SelectPatternResult::isMinOrMax(SPF) && "Expected a min/max")(static_cast<void> (0)); |
2263 | // TODO: Allow FP min/max with nnan/nsz. |
2264 | if (!LHS->getType()->isIntOrIntVectorTy()) |
2265 | return nullptr; |
2266 | |
2267 | // Match 3 of the same min/max ops. Example: umin(umin(), umin()). |
2268 | Value *A, *B, *C, *D; |
2269 | SelectPatternResult L = matchSelectPattern(LHS, A, B); |
2270 | SelectPatternResult R = matchSelectPattern(RHS, C, D); |
2271 | if (SPF != L.Flavor || L.Flavor != R.Flavor) |
2272 | return nullptr; |
2273 | |
2274 | // Look for a common operand. The use checks are different than usual because |
2275 | // a min/max pattern typically has 2 uses of each op: 1 by the cmp and 1 by |
2276 | // the select. |
2277 | Value *MinMaxOp = nullptr; |
2278 | Value *ThirdOp = nullptr; |
2279 | if (!LHS->hasNUsesOrMore(3) && RHS->hasNUsesOrMore(3)) { |
2280 | // If the LHS is only used in this chain and the RHS is used outside of it, |
2281 | // reuse the RHS min/max because that will eliminate the LHS. |
2282 | if (D == A || C == A) { |
2283 | // min(min(a, b), min(c, a)) --> min(min(c, a), b) |
2284 | // min(min(a, b), min(a, d)) --> min(min(a, d), b) |
2285 | MinMaxOp = RHS; |
2286 | ThirdOp = B; |
2287 | } else if (D == B || C == B) { |
2288 | // min(min(a, b), min(c, b)) --> min(min(c, b), a) |
2289 | // min(min(a, b), min(b, d)) --> min(min(b, d), a) |
2290 | MinMaxOp = RHS; |
2291 | ThirdOp = A; |
2292 | } |
2293 | } else if (!RHS->hasNUsesOrMore(3)) { |
2294 | // Reuse the LHS. This will eliminate the RHS. |
2295 | if (D == A || D == B) { |
2296 | // min(min(a, b), min(c, a)) --> min(min(a, b), c) |
2297 | // min(min(a, b), min(c, b)) --> min(min(a, b), c) |
2298 | MinMaxOp = LHS; |
2299 | ThirdOp = C; |
2300 | } else if (C == A || C == B) { |
2301 | // min(min(a, b), min(b, d)) --> min(min(a, b), d) |
2302 | // min(min(a, b), min(c, b)) --> min(min(a, b), d) |
2303 | MinMaxOp = LHS; |
2304 | ThirdOp = D; |
2305 | } |
2306 | } |
2307 | if (!MinMaxOp || !ThirdOp) |
2308 | return nullptr; |
2309 | |
2310 | CmpInst::Predicate P = getMinMaxPred(SPF); |
2311 | Value *CmpABC = Builder.CreateICmp(P, MinMaxOp, ThirdOp); |
2312 | return SelectInst::Create(CmpABC, MinMaxOp, ThirdOp); |
2313 | } |
2314 | |
2315 | /// Try to reduce a funnel/rotate pattern that includes a compare and select |
2316 | /// into a funnel shift intrinsic. Example: |
2317 | /// rotl32(a, b) --> (b == 0 ? a : ((a >> (32 - b)) | (a << b))) |
2318 | /// --> call llvm.fshl.i32(a, a, b) |
2319 | /// fshl32(a, b, c) --> (c == 0 ? a : ((b >> (32 - c)) | (a << c))) |
2320 | /// --> call llvm.fshl.i32(a, b, c) |
2321 | /// fshr32(a, b, c) --> (c == 0 ? b : ((a >> (32 - c)) | (b << c))) |
2322 | /// --> call llvm.fshr.i32(a, b, c) |
2323 | static Instruction *foldSelectFunnelShift(SelectInst &Sel, |
2324 | InstCombiner::BuilderTy &Builder) { |
2325 | // This must be a power-of-2 type for a bitmasking transform to be valid. |
2326 | unsigned Width = Sel.getType()->getScalarSizeInBits(); |
2327 | if (!isPowerOf2_32(Width)) |
2328 | return nullptr; |
2329 | |
2330 | BinaryOperator *Or0, *Or1; |
2331 | if (!match(Sel.getFalseValue(), m_OneUse(m_Or(m_BinOp(Or0), m_BinOp(Or1))))) |
2332 | return nullptr; |
2333 | |
2334 | Value *SV0, *SV1, *SA0, *SA1; |
2335 | if (!match(Or0, m_OneUse(m_LogicalShift(m_Value(SV0), |
2336 | m_ZExtOrSelf(m_Value(SA0))))) || |
2337 | !match(Or1, m_OneUse(m_LogicalShift(m_Value(SV1), |
2338 | m_ZExtOrSelf(m_Value(SA1))))) || |
2339 | Or0->getOpcode() == Or1->getOpcode()) |
2340 | return nullptr; |
2341 | |
2342 | // Canonicalize to or(shl(SV0, SA0), lshr(SV1, SA1)). |
2343 | if (Or0->getOpcode() == BinaryOperator::LShr) { |
2344 | std::swap(Or0, Or1); |
2345 | std::swap(SV0, SV1); |
2346 | std::swap(SA0, SA1); |
2347 | } |
2348 | assert(Or0->getOpcode() == BinaryOperator::Shl &&(static_cast<void> (0)) |
2349 | Or1->getOpcode() == BinaryOperator::LShr &&(static_cast<void> (0)) |
2350 | "Illegal or(shift,shift) pair")(static_cast<void> (0)); |
2351 | |
2352 | // Check the shift amounts to see if they are an opposite pair. |
2353 | Value *ShAmt; |
2354 | if (match(SA1, m_OneUse(m_Sub(m_SpecificInt(Width), m_Specific(SA0))))) |
2355 | ShAmt = SA0; |
2356 | else if (match(SA0, m_OneUse(m_Sub(m_SpecificInt(Width), m_Specific(SA1))))) |
2357 | ShAmt = SA1; |
2358 | else |
2359 | return nullptr; |
2360 | |
2361 | // We should now have this pattern: |
2362 | // select ?, TVal, (or (shl SV0, SA0), (lshr SV1, SA1)) |
2363 | // The false value of the select must be a funnel-shift of the true value: |
2364 | // IsFShl -> TVal must be SV0 else TVal must be SV1. |
2365 | bool IsFshl = (ShAmt == SA0); |
2366 | Value *TVal = Sel.getTrueValue(); |
2367 | if ((IsFshl && TVal != SV0) || (!IsFshl && TVal != SV1)) |
2368 | return nullptr; |
2369 | |
2370 | // Finally, see if the select is filtering out a shift-by-zero. |
2371 | Value *Cond = Sel.getCondition(); |
2372 | ICmpInst::Predicate Pred; |
2373 | if (!match(Cond, m_OneUse(m_ICmp(Pred, m_Specific(ShAmt), m_ZeroInt()))) || |
2374 | Pred != ICmpInst::ICMP_EQ) |
2375 | return nullptr; |
2376 | |
2377 | // If this is not a rotate then the select was blocking poison from the |
2378 | // 'shift-by-zero' non-TVal, but a funnel shift won't - so freeze it. |
2379 | if (SV0 != SV1) { |
2380 | if (IsFshl && !llvm::isGuaranteedNotToBePoison(SV1)) |
2381 | SV1 = Builder.CreateFreeze(SV1); |
2382 | else if (!IsFshl && !llvm::isGuaranteedNotToBePoison(SV0)) |
2383 | SV0 = Builder.CreateFreeze(SV0); |
2384 | } |
2385 | |
2386 | // This is a funnel/rotate that avoids shift-by-bitwidth UB in a suboptimal way. |
2387 | // Convert to funnel shift intrinsic. |
2388 | Intrinsic::ID IID = IsFshl ? Intrinsic::fshl : Intrinsic::fshr; |
2389 | Function *F = Intrinsic::getDeclaration(Sel.getModule(), IID, Sel.getType()); |
2390 | ShAmt = Builder.CreateZExt(ShAmt, Sel.getType()); |
2391 | return CallInst::Create(F, { SV0, SV1, ShAmt }); |
2392 | } |
2393 | |
2394 | static Instruction *foldSelectToCopysign(SelectInst &Sel, |
2395 | InstCombiner::BuilderTy &Builder) { |
2396 | Value *Cond = Sel.getCondition(); |
2397 | Value *TVal = Sel.getTrueValue(); |
2398 | Value *FVal = Sel.getFalseValue(); |
2399 | Type *SelType = Sel.getType(); |
2400 | |
2401 | // Match select ?, TC, FC where the constants are equal but negated. |
2402 | // TODO: Generalize to handle a negated variable operand? |
2403 | const APFloat *TC, *FC; |
2404 | if (!match(TVal, m_APFloat(TC)) || !match(FVal, m_APFloat(FC)) || |
2405 | !abs(*TC).bitwiseIsEqual(abs(*FC))) |
2406 | return nullptr; |
2407 | |
2408 | assert(TC != FC && "Expected equal select arms to simplify")(static_cast<void> (0)); |
2409 | |
2410 | Value *X; |
2411 | const APInt *C; |
2412 | bool IsTrueIfSignSet; |
2413 | ICmpInst::Predicate Pred; |
2414 | if (!match(Cond, m_OneUse(m_ICmp(Pred, m_BitCast(m_Value(X)), m_APInt(C)))) || |
2415 | !InstCombiner::isSignBitCheck(Pred, *C, IsTrueIfSignSet) || |
2416 | X->getType() != SelType) |
2417 | return nullptr; |
2418 | |
2419 | // If needed, negate the value that will be the sign argument of the copysign: |
2420 | // (bitcast X) < 0 ? -TC : TC --> copysign(TC, X) |
2421 | // (bitcast X) < 0 ? TC : -TC --> copysign(TC, -X) |
2422 | // (bitcast X) >= 0 ? -TC : TC --> copysign(TC, -X) |
2423 | // (bitcast X) >= 0 ? TC : -TC --> copysign(TC, X) |
2424 | if (IsTrueIfSignSet ^ TC->isNegative()) |
2425 | X = Builder.CreateFNegFMF(X, &Sel); |
2426 | |
2427 | // Canonicalize the magnitude argument as the positive constant since we do |
2428 | // not care about its sign. |
2429 | Value *MagArg = TC->isNegative() ? FVal : TVal; |
2430 | Function *F = Intrinsic::getDeclaration(Sel.getModule(), Intrinsic::copysign, |
2431 | Sel.getType()); |
2432 | Instruction *CopySign = CallInst::Create(F, { MagArg, X }); |
2433 | CopySign->setFastMathFlags(Sel.getFastMathFlags()); |
2434 | return CopySign; |
2435 | } |
2436 | |
2437 | Instruction *InstCombinerImpl::foldVectorSelect(SelectInst &Sel) { |
2438 | auto *VecTy = dyn_cast<FixedVectorType>(Sel.getType()); |
2439 | if (!VecTy) |
2440 | return nullptr; |
2441 | |
2442 | unsigned NumElts = VecTy->getNumElements(); |
2443 | APInt UndefElts(NumElts, 0); |
2444 | APInt AllOnesEltMask(APInt::getAllOnesValue(NumElts)); |
2445 | if (Value *V = SimplifyDemandedVectorElts(&Sel, AllOnesEltMask, UndefElts)) { |
2446 | if (V != &Sel) |
2447 | return replaceInstUsesWith(Sel, V); |
2448 | return &Sel; |
2449 | } |
2450 | |
2451 | // A select of a "select shuffle" with a common operand can be rearranged |
2452 | // to select followed by "select shuffle". Because of poison, this only works |
2453 | // in the case of a shuffle with no undefined mask elements. |
2454 | Value *Cond = Sel.getCondition(); |
2455 | Value *TVal = Sel.getTrueValue(); |
2456 | Value *FVal = Sel.getFalseValue(); |
2457 | Value *X, *Y; |
2458 | ArrayRef<int> Mask; |
2459 | if (match(TVal, m_OneUse(m_Shuffle(m_Value(X), m_Value(Y), m_Mask(Mask)))) && |
2460 | !is_contained(Mask, UndefMaskElem) && |
2461 | cast<ShuffleVectorInst>(TVal)->isSelect()) { |
2462 | if (X == FVal) { |
2463 | // select Cond, (shuf_sel X, Y), X --> shuf_sel X, (select Cond, Y, X) |
2464 | Value *NewSel = Builder.CreateSelect(Cond, Y, X, "sel", &Sel); |
2465 | return new ShuffleVectorInst(X, NewSel, Mask); |
2466 | } |
2467 | if (Y == FVal) { |
2468 | // select Cond, (shuf_sel X, Y), Y --> shuf_sel (select Cond, X, Y), Y |
2469 | Value *NewSel = Builder.CreateSelect(Cond, X, Y, "sel", &Sel); |
2470 | return new ShuffleVectorInst(NewSel, Y, Mask); |
2471 | } |
2472 | } |
2473 | if (match(FVal, m_OneUse(m_Shuffle(m_Value(X), m_Value(Y), m_Mask(Mask)))) && |
2474 | !is_contained(Mask, UndefMaskElem) && |
2475 | cast<ShuffleVectorInst>(FVal)->isSelect()) { |
2476 | if (X == TVal) { |
2477 | // select Cond, X, (shuf_sel X, Y) --> shuf_sel X, (select Cond, X, Y) |
2478 | Value *NewSel = Builder.CreateSelect(Cond, X, Y, "sel", &Sel); |
2479 | return new ShuffleVectorInst(X, NewSel, Mask); |
2480 | } |
2481 | if (Y == TVal) { |
2482 | // select Cond, Y, (shuf_sel X, Y) --> shuf_sel (select Cond, Y, X), Y |
2483 | Value *NewSel = Builder.CreateSelect(Cond, Y, X, "sel", &Sel); |
2484 | return new ShuffleVectorInst(NewSel, Y, Mask); |
2485 | } |
2486 | } |
2487 | |
2488 | return nullptr; |
2489 | } |
2490 | |
2491 | static Instruction *foldSelectToPhiImpl(SelectInst &Sel, BasicBlock *BB, |
2492 | const DominatorTree &DT, |
2493 | InstCombiner::BuilderTy &Builder) { |
2494 | // Find the block's immediate dominator that ends with a conditional branch |
2495 | // that matches select's condition (maybe inverted). |
2496 | auto *IDomNode = DT[BB]->getIDom(); |
2497 | if (!IDomNode) |
2498 | return nullptr; |
2499 | BasicBlock *IDom = IDomNode->getBlock(); |
2500 | |
2501 | Value *Cond = Sel.getCondition(); |
2502 | Value *IfTrue, *IfFalse; |
2503 | BasicBlock *TrueSucc, *FalseSucc; |
2504 | if (match(IDom->getTerminator(), |
2505 | m_Br(m_Specific(Cond), m_BasicBlock(TrueSucc), |
2506 | m_BasicBlock(FalseSucc)))) { |
2507 | IfTrue = Sel.getTrueValue(); |
2508 | IfFalse = Sel.getFalseValue(); |
2509 | } else if (match(IDom->getTerminator(), |
2510 | m_Br(m_Not(m_Specific(Cond)), m_BasicBlock(TrueSucc), |
2511 | m_BasicBlock(FalseSucc)))) { |
2512 | IfTrue = Sel.getFalseValue(); |
2513 | IfFalse = Sel.getTrueValue(); |
2514 | } else |
2515 | return nullptr; |
2516 | |
2517 | // Make sure the branches are actually different. |
2518 | if (TrueSucc == FalseSucc) |
2519 | return nullptr; |
2520 | |
2521 | // We want to replace select %cond, %a, %b with a phi that takes value %a |
2522 | // for all incoming edges that are dominated by condition `%cond == true`, |
2523 | // and value %b for edges dominated by condition `%cond == false`. If %a |
2524 | // or %b are also phis from the same basic block, we can go further and take |
2525 | // their incoming values from the corresponding blocks. |
2526 | BasicBlockEdge TrueEdge(IDom, TrueSucc); |
2527 | BasicBlockEdge FalseEdge(IDom, FalseSucc); |
2528 | DenseMap<BasicBlock *, Value *> Inputs; |
2529 | for (auto *Pred : predecessors(BB)) { |
2530 | // Check implication. |
2531 | BasicBlockEdge Incoming(Pred, BB); |
2532 | if (DT.dominates(TrueEdge, Incoming)) |
2533 | Inputs[Pred] = IfTrue->DoPHITranslation(BB, Pred); |
2534 | else if (DT.dominates(FalseEdge, Incoming)) |
2535 | Inputs[Pred] = IfFalse->DoPHITranslation(BB, Pred); |
2536 | else |
2537 | return nullptr; |
2538 | // Check availability. |
2539 | if (auto *Insn = dyn_cast<Instruction>(Inputs[Pred])) |
2540 | if (!DT.dominates(Insn, Pred->getTerminator())) |
2541 | return nullptr; |
2542 | } |
2543 | |
2544 | Builder.SetInsertPoint(&*BB->begin()); |
2545 | auto *PN = Builder.CreatePHI(Sel.getType(), Inputs.size()); |
2546 | for (auto *Pred : predecessors(BB)) |
2547 | PN->addIncoming(Inputs[Pred], Pred); |
2548 | PN->takeName(&Sel); |
2549 | return PN; |
2550 | } |
2551 | |
2552 | static Instruction *foldSelectToPhi(SelectInst &Sel, const DominatorTree &DT, |
2553 | InstCombiner::BuilderTy &Builder) { |
2554 | // Try to replace this select with Phi in one of these blocks. |
2555 | SmallSetVector<BasicBlock *, 4> CandidateBlocks; |
2556 | CandidateBlocks.insert(Sel.getParent()); |
2557 | for (Value *V : Sel.operands()) |
2558 | if (auto *I = dyn_cast<Instruction>(V)) |
2559 | CandidateBlocks.insert(I->getParent()); |
2560 | |
2561 | for (BasicBlock *BB : CandidateBlocks) |
2562 | if (auto *PN = foldSelectToPhiImpl(Sel, BB, DT, Builder)) |
2563 | return PN; |
2564 | return nullptr; |
2565 | } |
2566 | |
2567 | static Value *foldSelectWithFrozenICmp(SelectInst &Sel, InstCombiner::BuilderTy &Builder) { |
2568 | FreezeInst *FI = dyn_cast<FreezeInst>(Sel.getCondition()); |
2569 | if (!FI) |
2570 | return nullptr; |
2571 | |
2572 | Value *Cond = FI->getOperand(0); |
2573 | Value *TrueVal = Sel.getTrueValue(), *FalseVal = Sel.getFalseValue(); |
2574 | |
2575 | // select (freeze(x == y)), x, y --> y |
2576 | // select (freeze(x != y)), x, y --> x |
2577 | // The freeze should be only used by this select. Otherwise, remaining uses of |
2578 | // the freeze can observe a contradictory value. |
2579 | // c = freeze(x == y) ; Let's assume that y = poison & x = 42; c is 0 or 1 |
2580 | // a = select c, x, y ; |
2581 | // f(a, c) ; f(poison, 1) cannot happen, but if a is folded |
2582 | // ; to y, this can happen. |
2583 | CmpInst::Predicate Pred; |
2584 | if (FI->hasOneUse() && |
2585 | match(Cond, m_c_ICmp(Pred, m_Specific(TrueVal), m_Specific(FalseVal))) && |
2586 | (Pred == ICmpInst::ICMP_EQ || Pred == ICmpInst::ICMP_NE)) { |
2587 | return Pred == ICmpInst::ICMP_EQ ? FalseVal : TrueVal; |
2588 | } |
2589 | |
2590 | return nullptr; |
2591 | } |
2592 | |
2593 | Instruction *InstCombinerImpl::foldAndOrOfSelectUsingImpliedCond(Value *Op, |
2594 | SelectInst &SI, |
2595 | bool IsAnd) { |
2596 | Value *CondVal = SI.getCondition(); |
2597 | Value *A = SI.getTrueValue(); |
2598 | Value *B = SI.getFalseValue(); |
2599 | |
2600 | assert(Op->getType()->isIntOrIntVectorTy(1) &&(static_cast<void> (0)) |
2601 | "Op must be either i1 or vector of i1.")(static_cast<void> (0)); |
2602 | |
2603 | Optional<bool> Res = isImpliedCondition(Op, CondVal, DL, IsAnd); |
2604 | if (!Res) |
2605 | return nullptr; |
2606 | |
2607 | Value *Zero = Constant::getNullValue(A->getType()); |
2608 | Value *One = Constant::getAllOnesValue(A->getType()); |
2609 | |
2610 | if (*Res == true) { |
2611 | if (IsAnd) |
2612 | // select op, (select cond, A, B), false => select op, A, false |
2613 | // and op, (select cond, A, B) => select op, A, false |
2614 | // if op = true implies condval = true. |
2615 | return SelectInst::Create(Op, A, Zero); |
2616 | else |
2617 | // select op, true, (select cond, A, B) => select op, true, A |
2618 | // or op, (select cond, A, B) => select op, true, A |
2619 | // if op = false implies condval = true. |
2620 | return SelectInst::Create(Op, One, A); |
2621 | } else { |
2622 | if (IsAnd) |
2623 | // select op, (select cond, A, B), false => select op, B, false |
2624 | // and op, (select cond, A, B) => select op, B, false |
2625 | // if op = true implies condval = false. |
2626 | return SelectInst::Create(Op, B, Zero); |
2627 | else |
2628 | // select op, true, (select cond, A, B) => select op, true, B |
2629 | // or op, (select cond, A, B) => select op, true, B |
2630 | // if op = false implies condval = false. |
2631 | return SelectInst::Create(Op, One, B); |
2632 | } |
2633 | } |
2634 | |
2635 | Instruction *InstCombinerImpl::visitSelectInst(SelectInst &SI) { |
2636 | Value *CondVal = SI.getCondition(); |
2637 | Value *TrueVal = SI.getTrueValue(); |
2638 | Value *FalseVal = SI.getFalseValue(); |
2639 | Type *SelType = SI.getType(); |
2640 | |
2641 | // FIXME: Remove this workaround when freeze related patches are done. |
2642 | // For select with undef operand which feeds into an equality comparison, |
2643 | // don't simplify it so loop unswitch can know the equality comparison |
2644 | // may have an undef operand. This is a workaround for PR31652 caused by |
2645 | // descrepancy about branch on undef between LoopUnswitch and GVN. |
2646 | if (match(TrueVal, m_Undef()) || match(FalseVal, m_Undef())) { |
2647 | if (llvm::any_of(SI.users(), [&](User *U) { |
2648 | ICmpInst *CI = dyn_cast<ICmpInst>(U); |
2649 | if (CI && CI->isEquality()) |
2650 | return true; |
2651 | return false; |
2652 | })) { |
2653 | return nullptr; |
2654 | } |
2655 | } |
2656 | |
2657 | if (Value *V = SimplifySelectInst(CondVal, TrueVal, FalseVal, |
2658 | SQ.getWithInstruction(&SI))) |
2659 | return replaceInstUsesWith(SI, V); |
2660 | |
2661 | if (Instruction *I = canonicalizeSelectToShuffle(SI)) |
2662 | return I; |
2663 | |
2664 | if (Instruction *I = canonicalizeScalarSelectOfVecs(SI, *this)) |
2665 | return I; |
2666 | |
2667 | CmpInst::Predicate Pred; |
2668 | |
2669 | // Avoid potential infinite loops by checking for non-constant condition. |
2670 | // TODO: Can we assert instead by improving canonicalizeSelectToShuffle()? |
2671 | // Scalar select must have simplified? |
2672 | if (SelType->isIntOrIntVectorTy(1) && !isa<Constant>(CondVal) && |
2673 | TrueVal->getType() == CondVal->getType()) { |
2674 | // Folding select to and/or i1 isn't poison safe in general. impliesPoison |
2675 | // checks whether folding it does not convert a well-defined value into |
2676 | // poison. |
2677 | if (match(TrueVal, m_One()) && impliesPoison(FalseVal, CondVal)) { |
2678 | // Change: A = select B, true, C --> A = or B, C |
2679 | return BinaryOperator::CreateOr(CondVal, FalseVal); |
2680 | } |
2681 | if (match(FalseVal, m_Zero()) && impliesPoison(TrueVal, CondVal)) { |
2682 | // Change: A = select B, C, false --> A = and B, C |
2683 | return BinaryOperator::CreateAnd(CondVal, TrueVal); |
2684 | } |
2685 | |
2686 | auto *One = ConstantInt::getTrue(SelType); |
2687 | auto *Zero = ConstantInt::getFalse(SelType); |
2688 | |
2689 | // We match the "full" 0 or 1 constant here to avoid a potential infinite |
2690 | // loop with vectors that may have undefined/poison elements. |
2691 | // select a, false, b -> select !a, b, false |
2692 | if (match(TrueVal, m_Specific(Zero))) { |
2693 | Value *NotCond = Builder.CreateNot(CondVal, "not." + CondVal->getName()); |
2694 | return SelectInst::Create(NotCond, FalseVal, Zero); |
2695 | } |
2696 | // select a, b, true -> select !a, true, b |
2697 | if (match(FalseVal, m_Specific(One))) { |
2698 | Value *NotCond = Builder.CreateNot(CondVal, "not." + CondVal->getName()); |
2699 | return SelectInst::Create(NotCond, One, TrueVal); |
2700 | } |
2701 | |
2702 | // select a, a, b -> select a, true, b |
2703 | if (CondVal == TrueVal) |
2704 | return replaceOperand(SI, 1, One); |
2705 | // select a, b, a -> select a, b, false |
2706 | if (CondVal == FalseVal) |
2707 | return replaceOperand(SI, 2, Zero); |
2708 | |
2709 | // select a, !a, b -> select !a, b, false |
2710 | if (match(TrueVal, m_Not(m_Specific(CondVal)))) |
2711 | return SelectInst::Create(TrueVal, FalseVal, Zero); |
2712 | // select a, b, !a -> select !a, true, b |
2713 | if (match(FalseVal, m_Not(m_Specific(CondVal)))) |
2714 | return SelectInst::Create(FalseVal, One, TrueVal); |
2715 | |
2716 | Value *A, *B; |
2717 | |
2718 | // DeMorgan in select form: !a && !b --> !(a || b) |
2719 | // select !a, !b, false --> not (select a, true, b) |
2720 | if (match(&SI, m_LogicalAnd(m_Not(m_Value(A)), m_Not(m_Value(B)))) && |
2721 | (CondVal->hasOneUse() || TrueVal->hasOneUse()) && |
2722 | !match(A, m_ConstantExpr()) && !match(B, m_ConstantExpr())) |
2723 | return BinaryOperator::CreateNot(Builder.CreateSelect(A, One, B)); |
2724 | |
2725 | // DeMorgan in select form: !a || !b --> !(a && b) |
2726 | // select !a, true, !b --> not (select a, b, false) |
2727 | if (match(&SI, m_LogicalOr(m_Not(m_Value(A)), m_Not(m_Value(B)))) && |
2728 | (CondVal->hasOneUse() || FalseVal->hasOneUse()) && |
2729 | !match(A, m_ConstantExpr()) && !match(B, m_ConstantExpr())) |
2730 | return BinaryOperator::CreateNot(Builder.CreateSelect(A, B, Zero)); |
2731 | |
2732 | // select (select a, true, b), true, b -> select a, true, b |
2733 | if (match(CondVal, m_Select(m_Value(A), m_One(), m_Value(B))) && |
2734 | match(TrueVal, m_One()) && match(FalseVal, m_Specific(B))) |
2735 | return replaceOperand(SI, 0, A); |
2736 | // select (select a, b, false), b, false -> select a, b, false |
2737 | if (match(CondVal, m_Select(m_Value(A), m_Value(B), m_Zero())) && |
2738 | match(TrueVal, m_Specific(B)) && match(FalseVal, m_Zero())) |
2739 | return replaceOperand(SI, 0, A); |
2740 | |
2741 | if (!SelType->isVectorTy()) { |
2742 | if (Value *S = simplifyWithOpReplaced(TrueVal, CondVal, One, SQ, |
2743 | /* AllowRefinement */ true)) |
2744 | return replaceOperand(SI, 1, S); |
2745 | if (Value *S = simplifyWithOpReplaced(FalseVal, CondVal, Zero, SQ, |
2746 | /* AllowRefinement */ true)) |
2747 | return replaceOperand(SI, 2, S); |
2748 | } |
2749 | |
2750 | if (match(FalseVal, m_Zero()) || match(TrueVal, m_One())) { |
2751 | Use *Y = nullptr; |
2752 | bool IsAnd = match(FalseVal, m_Zero()) ? true : false; |
2753 | Value *Op1 = IsAnd ? TrueVal : FalseVal; |
2754 | if (isCheckForZeroAndMulWithOverflow(CondVal, Op1, IsAnd, Y)) { |
2755 | auto *FI = new FreezeInst(*Y, (*Y)->getName() + ".fr"); |
2756 | InsertNewInstBefore(FI, *cast<Instruction>(Y->getUser())); |
2757 | replaceUse(*Y, FI); |
2758 | return replaceInstUsesWith(SI, Op1); |
2759 | } |
2760 | |
2761 | if (auto *Op1SI = dyn_cast<SelectInst>(Op1)) |
2762 | if (auto *I = foldAndOrOfSelectUsingImpliedCond(CondVal, *Op1SI, |
2763 | /* IsAnd */ IsAnd)) |
2764 | return I; |
2765 | |
2766 | if (auto *ICmp0 = dyn_cast<ICmpInst>(CondVal)) { |
2767 | if (auto *ICmp1 = dyn_cast<ICmpInst>(Op1)) { |
2768 | if (auto *V = foldAndOrOfICmpsOfAndWithPow2(ICmp0, ICmp1, &SI, IsAnd, |
2769 | /* IsLogical */ true)) |
2770 | return replaceInstUsesWith(SI, V); |
2771 | |
2772 | if (auto *V = foldEqOfParts(ICmp0, ICmp1, IsAnd)) |
2773 | return replaceInstUsesWith(SI, V); |
2774 | } |
2775 | } |
2776 | } |
2777 | |
2778 | // select (select a, true, b), c, false -> select a, c, false |
2779 | // select c, (select a, true, b), false -> select c, a, false |
2780 | // if c implies that b is false. |
2781 | if (match(CondVal, m_Select(m_Value(A), m_One(), m_Value(B))) && |
2782 | match(FalseVal, m_Zero())) { |
2783 | Optional<bool> Res = isImpliedCondition(TrueVal, B, DL); |
2784 | if (Res && *Res == false) |
2785 | return replaceOperand(SI, 0, A); |
2786 | } |
2787 | if (match(TrueVal, m_Select(m_Value(A), m_One(), m_Value(B))) && |
2788 | match(FalseVal, m_Zero())) { |
2789 | Optional<bool> Res = isImpliedCondition(CondVal, B, DL); |
2790 | if (Res && *Res == false) |
2791 | return replaceOperand(SI, 1, A); |
2792 | } |
2793 | // select c, true, (select a, b, false) -> select c, true, a |
2794 | // select (select a, b, false), true, c -> select a, true, c |
2795 | // if c = false implies that b = true |
2796 | if (match(TrueVal, m_One()) && |
2797 | match(FalseVal, m_Select(m_Value(A), m_Value(B), m_Zero()))) { |
2798 | Optional<bool> Res = isImpliedCondition(CondVal, B, DL, false); |
2799 | if (Res && *Res == true) |
2800 | return replaceOperand(SI, 2, A); |
2801 | } |
2802 | if (match(CondVal, m_Select(m_Value(A), m_Value(B), m_Zero())) && |
2803 | match(TrueVal, m_One())) { |
2804 | Optional<bool> Res = isImpliedCondition(FalseVal, B, DL, false); |
2805 | if (Res && *Res == true) |
2806 | return replaceOperand(SI, 0, A); |
2807 | } |
2808 | |
2809 | // sel (sel c, a, false), true, (sel !c, b, false) -> sel c, a, b |
2810 | // sel (sel !c, a, false), true, (sel c, b, false) -> sel c, b, a |
2811 | Value *C1, *C2; |
2812 | if (match(CondVal, m_Select(m_Value(C1), m_Value(A), m_Zero())) && |
2813 | match(TrueVal, m_One()) && |
2814 | match(FalseVal, m_Select(m_Value(C2), m_Value(B), m_Zero()))) { |
2815 | if (match(C2, m_Not(m_Specific(C1)))) // first case |
2816 | return SelectInst::Create(C1, A, B); |
2817 | else if (match(C1, m_Not(m_Specific(C2)))) // second case |
2818 | return SelectInst::Create(C2, B, A); |
2819 | } |
2820 | } |
2821 | |
2822 | // Selecting between two integer or vector splat integer constants? |
2823 | // |
2824 | // Note that we don't handle a scalar select of vectors: |
2825 | // select i1 %c, <2 x i8> <1, 1>, <2 x i8> <0, 0> |
2826 | // because that may need 3 instructions to splat the condition value: |
2827 | // extend, insertelement, shufflevector. |
2828 | // |
2829 | // Do not handle i1 TrueVal and FalseVal otherwise would result in |
2830 | // zext/sext i1 to i1. |
2831 | if (SelType->isIntOrIntVectorTy() && !SelType->isIntOrIntVectorTy(1) && |
2832 | CondVal->getType()->isVectorTy() == SelType->isVectorTy()) { |
2833 | // select C, 1, 0 -> zext C to int |
2834 | if (match(TrueVal, m_One()) && match(FalseVal, m_Zero())) |
2835 | return new ZExtInst(CondVal, SelType); |
2836 | |
2837 | // select C, -1, 0 -> sext C to int |
2838 | if (match(TrueVal, m_AllOnes()) && match(FalseVal, m_Zero())) |
2839 | return new SExtInst(CondVal, SelType); |
2840 | |
2841 | // select C, 0, 1 -> zext !C to int |
2842 | if (match(TrueVal, m_Zero()) && match(FalseVal, m_One())) { |
2843 | Value *NotCond = Builder.CreateNot(CondVal, "not." + CondVal->getName()); |
2844 | return new ZExtInst(NotCond, SelType); |
2845 | } |
2846 | |
2847 | // select C, 0, -1 -> sext !C to int |
2848 | if (match(TrueVal, m_Zero()) && match(FalseVal, m_AllOnes())) { |
2849 | Value *NotCond = Builder.CreateNot(CondVal, "not." + CondVal->getName()); |
2850 | return new SExtInst(NotCond, SelType); |
2851 | } |
2852 | } |
2853 | |
2854 | if (auto *FCmp = dyn_cast<FCmpInst>(CondVal)) { |
2855 | Value *Cmp0 = FCmp->getOperand(0), *Cmp1 = FCmp->getOperand(1); |
2856 | // Are we selecting a value based on a comparison of the two values? |
2857 | if ((Cmp0 == TrueVal && Cmp1 == FalseVal) || |
2858 | (Cmp0 == FalseVal && Cmp1 == TrueVal)) { |
2859 | // Canonicalize to use ordered comparisons by swapping the select |
2860 | // operands. |
2861 | // |
2862 | // e.g. |
2863 | // (X ugt Y) ? X : Y -> (X ole Y) ? Y : X |
2864 | if (FCmp->hasOneUse() && FCmpInst::isUnordered(FCmp->getPredicate())) { |
2865 | FCmpInst::Predicate InvPred = FCmp->getInversePredicate(); |
2866 | IRBuilder<>::FastMathFlagGuard FMFG(Builder); |
2867 | // FIXME: The FMF should propagate from the select, not the fcmp. |
2868 | Builder.setFastMathFlags(FCmp->getFastMathFlags()); |
2869 | Value *NewCond = Builder.CreateFCmp(InvPred, Cmp0, Cmp1, |
2870 | FCmp->getName() + ".inv"); |
2871 | Value *NewSel = Builder.CreateSelect(NewCond, FalseVal, TrueVal); |
2872 | return replaceInstUsesWith(SI, NewSel); |
2873 | } |
2874 | |
2875 | // NOTE: if we wanted to, this is where to detect MIN/MAX |
2876 | } |
2877 | } |
2878 | |
2879 | // Canonicalize select with fcmp to fabs(). -0.0 makes this tricky. We need |
2880 | // fast-math-flags (nsz) or fsub with +0.0 (not fneg) for this to work. |
2881 | // (X <= +/-0.0) ? (0.0 - X) : X --> fabs(X) |
2882 | Instruction *FSub; |
2883 | if (match(CondVal, m_FCmp(Pred, m_Specific(FalseVal), m_AnyZeroFP())) && |
2884 | match(TrueVal, m_FSub(m_PosZeroFP(), m_Specific(FalseVal))) && |
2885 | match(TrueVal, m_Instruction(FSub)) && |
2886 | (Pred == FCmpInst::FCMP_OLE || Pred == FCmpInst::FCMP_ULE)) { |
2887 | Value *Fabs = Builder.CreateUnaryIntrinsic(Intrinsic::fabs, FalseVal, &SI); |
2888 | return replaceInstUsesWith(SI, Fabs); |
2889 | } |
2890 | // (X > +/-0.0) ? X : (0.0 - X) --> fabs(X) |
2891 | if (match(CondVal, m_FCmp(Pred, m_Specific(TrueVal), m_AnyZeroFP())) && |
2892 | match(FalseVal, m_FSub(m_PosZeroFP(), m_Specific(TrueVal))) && |
2893 | match(FalseVal, m_Instruction(FSub)) && |
2894 | (Pred == FCmpInst::FCMP_OGT || Pred == FCmpInst::FCMP_UGT)) { |
2895 | Value *Fabs = Builder.CreateUnaryIntrinsic(Intrinsic::fabs, TrueVal, &SI); |
2896 | return replaceInstUsesWith(SI, Fabs); |
2897 | } |
2898 | // With nnan and nsz: |
2899 | // (X < +/-0.0) ? -X : X --> fabs(X) |
2900 | // (X <= +/-0.0) ? -X : X --> fabs(X) |
2901 | Instruction *FNeg; |
2902 | if (match(CondVal, m_FCmp(Pred, m_Specific(FalseVal), m_AnyZeroFP())) && |
2903 | match(TrueVal, m_FNeg(m_Specific(FalseVal))) && |
2904 | match(TrueVal, m_Instruction(FNeg)) && SI.hasNoSignedZeros() && |
2905 | (Pred == FCmpInst::FCMP_OLT || Pred == FCmpInst::FCMP_OLE || |
2906 | Pred == FCmpInst::FCMP_ULT || Pred == FCmpInst::FCMP_ULE)) { |
2907 | Value *Fabs = Builder.CreateUnaryIntrinsic(Intrinsic::fabs, FalseVal, &SI); |
2908 | return replaceInstUsesWith(SI, Fabs); |
2909 | } |
2910 | // With nnan and nsz: |
2911 | // (X > +/-0.0) ? X : -X --> fabs(X) |
2912 | // (X >= +/-0.0) ? X : -X --> fabs(X) |
2913 | if (match(CondVal, m_FCmp(Pred, m_Specific(TrueVal), m_AnyZeroFP())) && |
2914 | match(FalseVal, m_FNeg(m_Specific(TrueVal))) && |
2915 | match(FalseVal, m_Instruction(FNeg)) && SI.hasNoSignedZeros() && |
2916 | (Pred == FCmpInst::FCMP_OGT || Pred == FCmpInst::FCMP_OGE || |
2917 | Pred == FCmpInst::FCMP_UGT || Pred == FCmpInst::FCMP_UGE)) { |
2918 | Value *Fabs = Builder.CreateUnaryIntrinsic(Intrinsic::fabs, TrueVal, &SI); |
2919 | return replaceInstUsesWith(SI, Fabs); |
2920 | } |
2921 | |
2922 | // See if we are selecting two values based on a comparison of the two values. |
2923 | if (ICmpInst *ICI = dyn_cast<ICmpInst>(CondVal)) |
2924 | if (Instruction *Result = foldSelectInstWithICmp(SI, ICI)) |
2925 | return Result; |
2926 | |
2927 | if (Instruction *Add = foldAddSubSelect(SI, Builder)) |
2928 | return Add; |
2929 | if (Instruction *Add = foldOverflowingAddSubSelect(SI, Builder)) |
2930 | return Add; |
2931 | if (Instruction *Or = foldSetClearBits(SI, Builder)) |
2932 | return Or; |
2933 | |
2934 | // Turn (select C, (op X, Y), (op X, Z)) -> (op X, (select C, Y, Z)) |
2935 | auto *TI = dyn_cast<Instruction>(TrueVal); |
2936 | auto *FI = dyn_cast<Instruction>(FalseVal); |
2937 | if (TI && FI && TI->getOpcode() == FI->getOpcode()) |
2938 | if (Instruction *IV = foldSelectOpOp(SI, TI, FI)) |
2939 | return IV; |
2940 | |
2941 | if (Instruction *I = foldSelectExtConst(SI)) |
2942 | return I; |
2943 | |
2944 | // Fold (select C, (gep Ptr, Idx), Ptr) -> (gep Ptr, (select C, Idx, 0)) |
2945 | // Fold (select C, Ptr, (gep Ptr, Idx)) -> (gep Ptr, (select C, 0, Idx)) |
2946 | auto SelectGepWithBase = [&](GetElementPtrInst *Gep, Value *Base, |
2947 | bool Swap) -> GetElementPtrInst * { |
2948 | Value *Ptr = Gep->getPointerOperand(); |
2949 | if (Gep->getNumOperands() != 2 || Gep->getPointerOperand() != Base || |
2950 | !Gep->hasOneUse()) |
2951 | return nullptr; |
2952 | Type *ElementType = Gep->getResultElementType(); |
2953 | Value *Idx = Gep->getOperand(1); |
2954 | Value *NewT = Idx; |
2955 | Value *NewF = Constant::getNullValue(Idx->getType()); |
2956 | if (Swap) |
2957 | std::swap(NewT, NewF); |
2958 | Value *NewSI = |
2959 | Builder.CreateSelect(CondVal, NewT, NewF, SI.getName() + ".idx", &SI); |
2960 | return GetElementPtrInst::Create(ElementType, Ptr, {NewSI}); |
2961 | }; |
2962 | if (auto *TrueGep = dyn_cast<GetElementPtrInst>(TrueVal)) |
2963 | if (auto *NewGep = SelectGepWithBase(TrueGep, FalseVal, false)) |
2964 | return NewGep; |
2965 | if (auto *FalseGep = dyn_cast<GetElementPtrInst>(FalseVal)) |
2966 | if (auto *NewGep = SelectGepWithBase(FalseGep, TrueVal, true)) |
2967 | return NewGep; |
2968 | |
2969 | // See if we can fold the select into one of our operands. |
2970 | if (SelType->isIntOrIntVectorTy() || SelType->isFPOrFPVectorTy()) { |
2971 | if (Instruction *FoldI = foldSelectIntoOp(SI, TrueVal, FalseVal)) |
2972 | return FoldI; |
2973 | |
2974 | Value *LHS, *RHS; |
2975 | Instruction::CastOps CastOp; |
2976 | SelectPatternResult SPR = matchSelectPattern(&SI, LHS, RHS, &CastOp); |
2977 | auto SPF = SPR.Flavor; |
2978 | if (SPF) { |
2979 | Value *LHS2, *RHS2; |
2980 | if (SelectPatternFlavor SPF2 = matchSelectPattern(LHS, LHS2, RHS2).Flavor) |
2981 | if (Instruction *R = foldSPFofSPF(cast<Instruction>(LHS), SPF2, LHS2, |
2982 | RHS2, SI, SPF, RHS)) |
2983 | return R; |
2984 | if (SelectPatternFlavor SPF2 = matchSelectPattern(RHS, LHS2, RHS2).Flavor) |
2985 | if (Instruction *R = foldSPFofSPF(cast<Instruction>(RHS), SPF2, LHS2, |
2986 | RHS2, SI, SPF, LHS)) |
2987 | return R; |
2988 | // TODO. |
2989 | // ABS(-X) -> ABS(X) |
2990 | } |
2991 | |
2992 | if (SelectPatternResult::isMinOrMax(SPF)) { |
2993 | // Canonicalize so that |
2994 | // - type casts are outside select patterns. |
2995 | // - float clamp is transformed to min/max pattern |
2996 | |
2997 | bool IsCastNeeded = LHS->getType() != SelType; |
2998 | Value *CmpLHS = cast<CmpInst>(CondVal)->getOperand(0); |
2999 | Value *CmpRHS = cast<CmpInst>(CondVal)->getOperand(1); |
3000 | if (IsCastNeeded || |
3001 | (LHS->getType()->isFPOrFPVectorTy() && |
3002 | ((CmpLHS != LHS && CmpLHS != RHS) || |
3003 | (CmpRHS != LHS && CmpRHS != RHS)))) { |
3004 | CmpInst::Predicate MinMaxPred = getMinMaxPred(SPF, SPR.Ordered); |
3005 | |
3006 | Value *Cmp; |
3007 | if (CmpInst::isIntPredicate(MinMaxPred)) { |
3008 | Cmp = Builder.CreateICmp(MinMaxPred, LHS, RHS); |
3009 | } else { |
3010 | IRBuilder<>::FastMathFlagGuard FMFG(Builder); |
3011 | auto FMF = |
3012 | cast<FPMathOperator>(SI.getCondition())->getFastMathFlags(); |
3013 | Builder.setFastMathFlags(FMF); |
3014 | Cmp = Builder.CreateFCmp(MinMaxPred, LHS, RHS); |
3015 | } |
3016 | |
3017 | Value *NewSI = Builder.CreateSelect(Cmp, LHS, RHS, SI.getName(), &SI); |
3018 | if (!IsCastNeeded) |
3019 | return replaceInstUsesWith(SI, NewSI); |
3020 | |
3021 | Value *NewCast = Builder.CreateCast(CastOp, NewSI, SelType); |
3022 | return replaceInstUsesWith(SI, NewCast); |
3023 | } |
3024 | |
3025 | // MAX(~a, ~b) -> ~MIN(a, b) |
3026 | // MAX(~a, C) -> ~MIN(a, ~C) |
3027 | // MIN(~a, ~b) -> ~MAX(a, b) |
3028 | // MIN(~a, C) -> ~MAX(a, ~C) |
3029 | auto moveNotAfterMinMax = [&](Value *X, Value *Y) -> Instruction * { |
3030 | Value *A; |
3031 | if (match(X, m_Not(m_Value(A))) && !X->hasNUsesOrMore(3) && |
3032 | !isFreeToInvert(A, A->hasOneUse()) && |
3033 | // Passing false to only consider m_Not and constants. |
3034 | isFreeToInvert(Y, false)) { |
3035 | Value *B = Builder.CreateNot(Y); |
3036 | Value *NewMinMax = createMinMax(Builder, getInverseMinMaxFlavor(SPF), |
3037 | A, B); |
3038 | // Copy the profile metadata. |
3039 | if (MDNode *MD = SI.getMetadata(LLVMContext::MD_prof)) { |
3040 | cast<SelectInst>(NewMinMax)->setMetadata(LLVMContext::MD_prof, MD); |
3041 | // Swap the metadata if the operands are swapped. |
3042 | if (X == SI.getFalseValue() && Y == SI.getTrueValue()) |
3043 | cast<SelectInst>(NewMinMax)->swapProfMetadata(); |
3044 | } |
3045 | |
3046 | return BinaryOperator::CreateNot(NewMinMax); |
3047 | } |
3048 | |
3049 | return nullptr; |
3050 | }; |
3051 | |
3052 | if (Instruction *I = moveNotAfterMinMax(LHS, RHS)) |
3053 | return I; |
3054 | if (Instruction *I = moveNotAfterMinMax(RHS, LHS)) |
3055 | return I; |
3056 | |
3057 | if (Instruction *I = moveAddAfterMinMax(SPF, LHS, RHS, Builder)) |
3058 | return I; |
3059 | |
3060 | if (Instruction *I = factorizeMinMaxTree(SPF, LHS, RHS, Builder)) |
3061 | return I; |
3062 | if (Instruction *I = matchSAddSubSat(SI)) |
3063 | return I; |
3064 | } |
3065 | } |
3066 | |
3067 | // Canonicalize select of FP values where NaN and -0.0 are not valid as |
3068 | // minnum/maxnum intrinsics. |
3069 | if (isa<FPMathOperator>(SI) && SI.hasNoNaNs() && SI.hasNoSignedZeros()) { |
3070 | Value *X, *Y; |
3071 | if (match(&SI, m_OrdFMax(m_Value(X), m_Value(Y)))) |
3072 | return replaceInstUsesWith( |
3073 | SI, Builder.CreateBinaryIntrinsic(Intrinsic::maxnum, X, Y, &SI)); |
3074 | |
3075 | if (match(&SI, m_OrdFMin(m_Value(X), m_Value(Y)))) |
3076 | return replaceInstUsesWith( |
3077 | SI, Builder.CreateBinaryIntrinsic(Intrinsic::minnum, X, Y, &SI)); |
3078 | } |
3079 | |
3080 | // See if we can fold the select into a phi node if the condition is a select. |
3081 | if (auto *PN = dyn_cast<PHINode>(SI.getCondition())) |
3082 | // The true/false values have to be live in the PHI predecessor's blocks. |
3083 | if (canSelectOperandBeMappingIntoPredBlock(TrueVal, SI) && |
3084 | canSelectOperandBeMappingIntoPredBlock(FalseVal, SI)) |
3085 | if (Instruction *NV = foldOpIntoPhi(SI, PN)) |
3086 | return NV; |
3087 | |
3088 | if (SelectInst *TrueSI = dyn_cast<SelectInst>(TrueVal)) { |
3089 | if (TrueSI->getCondition()->getType() == CondVal->getType()) { |
3090 | // select(C, select(C, a, b), c) -> select(C, a, c) |
3091 | if (TrueSI->getCondition() == CondVal) { |
3092 | if (SI.getTrueValue() == TrueSI->getTrueValue()) |
3093 | return nullptr; |
3094 | return replaceOperand(SI, 1, TrueSI->getTrueValue()); |
3095 | } |
3096 | // select(C0, select(C1, a, b), b) -> select(C0&C1, a, b) |
3097 | // We choose this as normal form to enable folding on the And and |
3098 | // shortening paths for the values (this helps getUnderlyingObjects() for |
3099 | // example). |
3100 | if (TrueSI->getFalseValue() == FalseVal && TrueSI->hasOneUse()) { |
3101 | Value *And = Builder.CreateLogicalAnd(CondVal, TrueSI->getCondition()); |
3102 | replaceOperand(SI, 0, And); |
3103 | replaceOperand(SI, 1, TrueSI->getTrueValue()); |
3104 | return &SI; |
3105 | } |
3106 | } |
3107 | } |
3108 | if (SelectInst *FalseSI = dyn_cast<SelectInst>(FalseVal)) { |
3109 | if (FalseSI->getCondition()->getType() == CondVal->getType()) { |
3110 | // select(C, a, select(C, b, c)) -> select(C, a, c) |
3111 | if (FalseSI->getCondition() == CondVal) { |
3112 | if (SI.getFalseValue() == FalseSI->getFalseValue()) |
3113 | return nullptr; |
3114 | return replaceOperand(SI, 2, FalseSI->getFalseValue()); |
3115 | } |
3116 | // select(C0, a, select(C1, a, b)) -> select(C0|C1, a, b) |
3117 | if (FalseSI->getTrueValue() == TrueVal && FalseSI->hasOneUse()) { |
3118 | Value *Or = Builder.CreateLogicalOr(CondVal, FalseSI->getCondition()); |
3119 | replaceOperand(SI, 0, Or); |
3120 | replaceOperand(SI, 2, FalseSI->getFalseValue()); |
3121 | return &SI; |
3122 | } |
3123 | } |
3124 | } |
3125 | |
3126 | auto canMergeSelectThroughBinop = [](BinaryOperator *BO) { |
3127 | // The select might be preventing a division by 0. |
3128 | switch (BO->getOpcode()) { |
3129 | default: |
3130 | return true; |
3131 | case Instruction::SRem: |
3132 | case Instruction::URem: |
3133 | case Instruction::SDiv: |
3134 | case Instruction::UDiv: |
3135 | return false; |
3136 | } |
3137 | }; |
3138 | |
3139 | // Try to simplify a binop sandwiched between 2 selects with the same |
3140 | // condition. |
3141 | // select(C, binop(select(C, X, Y), W), Z) -> select(C, binop(X, W), Z) |
3142 | BinaryOperator *TrueBO; |
3143 | if (match(TrueVal, m_OneUse(m_BinOp(TrueBO))) && |
3144 | canMergeSelectThroughBinop(TrueBO)) { |
3145 | if (auto *TrueBOSI = dyn_cast<SelectInst>(TrueBO->getOperand(0))) { |
3146 | if (TrueBOSI->getCondition() == CondVal) { |
3147 | replaceOperand(*TrueBO, 0, TrueBOSI->getTrueValue()); |
3148 | Worklist.push(TrueBO); |
3149 | return &SI; |
3150 | } |
3151 | } |
3152 | if (auto *TrueBOSI = dyn_cast<SelectInst>(TrueBO->getOperand(1))) { |
3153 | if (TrueBOSI->getCondition() == CondVal) { |
3154 | replaceOperand(*TrueBO, 1, TrueBOSI->getTrueValue()); |
3155 | Worklist.push(TrueBO); |
3156 | return &SI; |
3157 | } |
3158 | } |
3159 | } |
3160 | |
3161 | // select(C, Z, binop(select(C, X, Y), W)) -> select(C, Z, binop(Y, W)) |
3162 | BinaryOperator *FalseBO; |
3163 | if (match(FalseVal, m_OneUse(m_BinOp(FalseBO))) && |
3164 | canMergeSelectThroughBinop(FalseBO)) { |
3165 | if (auto *FalseBOSI = dyn_cast<SelectInst>(FalseBO->getOperand(0))) { |
3166 | if (FalseBOSI->getCondition() == CondVal) { |
3167 | replaceOperand(*FalseBO, 0, FalseBOSI->getFalseValue()); |
3168 | Worklist.push(FalseBO); |
3169 | return &SI; |
3170 | } |
3171 | } |
3172 | if (auto *FalseBOSI = dyn_cast<SelectInst>(FalseBO->getOperand(1))) { |
3173 | if (FalseBOSI->getCondition() == CondVal) { |
3174 | replaceOperand(*FalseBO, 1, FalseBOSI->getFalseValue()); |
3175 | Worklist.push(FalseBO); |
3176 | return &SI; |
3177 | } |
3178 | } |
3179 | } |
3180 | |
3181 | Value *NotCond; |
3182 | if (match(CondVal, m_Not(m_Value(NotCond))) && |
3183 | !InstCombiner::shouldAvoidAbsorbingNotIntoSelect(SI)) { |
3184 | replaceOperand(SI, 0, NotCond); |
3185 | SI.swapValues(); |
3186 | SI.swapProfMetadata(); |
3187 | return &SI; |
3188 | } |
3189 | |
3190 | if (Instruction *I = foldVectorSelect(SI)) |
3191 | return I; |
3192 | |
3193 | // If we can compute the condition, there's no need for a select. |
3194 | // Like the above fold, we are attempting to reduce compile-time cost by |
3195 | // putting this fold here with limitations rather than in InstSimplify. |
3196 | // The motivation for this call into value tracking is to take advantage of |
3197 | // the assumption cache, so make sure that is populated. |
3198 | if (!CondVal->getType()->isVectorTy() && !AC.assumptions().empty()) { |
3199 | KnownBits Known(1); |
3200 | computeKnownBits(CondVal, Known, 0, &SI); |
3201 | if (Known.One.isOneValue()) |
3202 | return replaceInstUsesWith(SI, TrueVal); |
3203 | if (Known.Zero.isOneValue()) |
3204 | return replaceInstUsesWith(SI, FalseVal); |
3205 | } |
3206 | |
3207 | if (Instruction *BitCastSel = foldSelectCmpBitcasts(SI, Builder)) |
3208 | return BitCastSel; |
3209 | |
3210 | // Simplify selects that test the returned flag of cmpxchg instructions. |
3211 | if (Value *V = foldSelectCmpXchg(SI)) |
3212 | return replaceInstUsesWith(SI, V); |
3213 | |
3214 | if (Instruction *Select = foldSelectBinOpIdentity(SI, TLI, *this)) |
3215 | return Select; |
3216 | |
3217 | if (Instruction *Funnel = foldSelectFunnelShift(SI, Builder)) |
3218 | return Funnel; |
3219 | |
3220 | if (Instruction *Copysign = foldSelectToCopysign(SI, Builder)) |
3221 | return Copysign; |
3222 | |
3223 | if (Instruction *PN = foldSelectToPhi(SI, DT, Builder)) |
3224 | return replaceInstUsesWith(SI, PN); |
3225 | |
3226 | if (Value *Fr = foldSelectWithFrozenICmp(SI, Builder)) |
3227 | return replaceInstUsesWith(SI, Fr); |
3228 | |
3229 | // select(mask, mload(,,mask,0), 0) -> mload(,,mask,0) |
3230 | // Load inst is intentionally not checked for hasOneUse() |
3231 | if (match(FalseVal, m_Zero()) && |
3232 | match(TrueVal, m_MaskedLoad(m_Value(), m_Value(), m_Specific(CondVal), |
3233 | m_CombineOr(m_Undef(), m_Zero())))) { |
3234 | auto *MaskedLoad = cast<IntrinsicInst>(TrueVal); |
3235 | if (isa<UndefValue>(MaskedLoad->getArgOperand(3))) |
3236 | MaskedLoad->setArgOperand(3, FalseVal /* Zero */); |
3237 | return replaceInstUsesWith(SI, MaskedLoad); |
3238 | } |
3239 | |
3240 | Value *Mask; |
3241 | if (match(TrueVal, m_Zero()) && |
3242 | match(FalseVal, m_MaskedLoad(m_Value(), m_Value(), m_Value(Mask), |
3243 | m_CombineOr(m_Undef(), m_Zero()))) && |
3244 | (CondVal->getType() == Mask->getType())) { |
3245 | // We can remove the select by ensuring the load zeros all lanes the |
3246 | // select would have. We determine this by proving there is no overlap |
3247 | // between the load and select masks. |
3248 | // (i.e (load_mask & select_mask) == 0 == no overlap) |
3249 | bool CanMergeSelectIntoLoad = false; |
3250 | if (Value *V = SimplifyAndInst(CondVal, Mask, SQ.getWithInstruction(&SI))) |
3251 | CanMergeSelectIntoLoad = match(V, m_Zero()); |
3252 | |
3253 | if (CanMergeSelectIntoLoad) { |
3254 | auto *MaskedLoad = cast<IntrinsicInst>(FalseVal); |
3255 | if (isa<UndefValue>(MaskedLoad->getArgOperand(3))) |
3256 | MaskedLoad->setArgOperand(3, TrueVal /* Zero */); |
3257 | return replaceInstUsesWith(SI, MaskedLoad); |
3258 | } |
3259 | } |
3260 | |
3261 | return nullptr; |
3262 | } |