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/STLExtras.h" |
16 | #include "llvm/ADT/SmallVector.h" |
17 | #include "llvm/Analysis/AssumptionCache.h" |
18 | #include "llvm/Analysis/CmpInstAnalysis.h" |
19 | #include "llvm/Analysis/InstructionSimplify.h" |
20 | #include "llvm/Analysis/OverflowInstAnalysis.h" |
21 | #include "llvm/Analysis/ValueTracking.h" |
22 | #include "llvm/Analysis/VectorUtils.h" |
23 | #include "llvm/IR/BasicBlock.h" |
24 | #include "llvm/IR/Constant.h" |
25 | #include "llvm/IR/ConstantRange.h" |
26 | #include "llvm/IR/Constants.h" |
27 | #include "llvm/IR/DerivedTypes.h" |
28 | #include "llvm/IR/FMF.h" |
29 | #include "llvm/IR/IRBuilder.h" |
30 | #include "llvm/IR/InstrTypes.h" |
31 | #include "llvm/IR/Instruction.h" |
32 | #include "llvm/IR/Instructions.h" |
33 | #include "llvm/IR/IntrinsicInst.h" |
34 | #include "llvm/IR/Intrinsics.h" |
35 | #include "llvm/IR/Operator.h" |
36 | #include "llvm/IR/PatternMatch.h" |
37 | #include "llvm/IR/Type.h" |
38 | #include "llvm/IR/User.h" |
39 | #include "llvm/IR/Value.h" |
40 | #include "llvm/Support/Casting.h" |
41 | #include "llvm/Support/ErrorHandling.h" |
42 | #include "llvm/Support/KnownBits.h" |
43 | #include "llvm/Transforms/InstCombine/InstCombiner.h" |
44 | #include <cassert> |
45 | #include <utility> |
46 | |
47 | #define DEBUG_TYPE "instcombine" |
48 | #include "llvm/Transforms/Utils/InstructionWorklist.h" |
49 | |
50 | using namespace llvm; |
51 | using namespace PatternMatch; |
52 | |
53 | |
54 | /// Replace a select operand based on an equality comparison with the identity |
55 | /// constant of a binop. |
56 | static Instruction *foldSelectBinOpIdentity(SelectInst &Sel, |
57 | const TargetLibraryInfo &TLI, |
58 | InstCombinerImpl &IC) { |
59 | // The select condition must be an equality compare with a constant operand. |
60 | Value *X; |
61 | Constant *C; |
62 | CmpPredicate Pred; |
63 | if (!match(V: Sel.getCondition(), P: m_Cmp(Pred, L: m_Value(V&: X), R: m_Constant(C)))) |
64 | return nullptr; |
65 | |
66 | bool IsEq; |
67 | if (ICmpInst::isEquality(P: Pred)) |
68 | IsEq = Pred == ICmpInst::ICMP_EQ; |
69 | else if (Pred == FCmpInst::FCMP_OEQ) |
70 | IsEq = true; |
71 | else if (Pred == FCmpInst::FCMP_UNE) |
72 | IsEq = false; |
73 | else |
74 | return nullptr; |
75 | |
76 | // A select operand must be a binop. |
77 | BinaryOperator *BO; |
78 | if (!match(V: Sel.getOperand(i_nocapture: IsEq ? 1 : 2), P: m_BinOp(I&: BO))) |
79 | return nullptr; |
80 | |
81 | // The compare constant must be the identity constant for that binop. |
82 | // If this a floating-point compare with 0.0, any zero constant will do. |
83 | Type *Ty = BO->getType(); |
84 | Constant *IdC = ConstantExpr::getBinOpIdentity(Opcode: BO->getOpcode(), Ty, AllowRHSConstant: true); |
85 | if (IdC != C) { |
86 | if (!IdC || !CmpInst::isFPPredicate(P: Pred)) |
87 | return nullptr; |
88 | if (!match(V: IdC, P: m_AnyZeroFP()) || !match(V: C, P: m_AnyZeroFP())) |
89 | return nullptr; |
90 | } |
91 | |
92 | // Last, match the compare variable operand with a binop operand. |
93 | Value *Y; |
94 | if (!BO->isCommutative() && !match(V: BO, P: m_BinOp(L: m_Value(V&: Y), R: m_Specific(V: X)))) |
95 | return nullptr; |
96 | if (!match(V: BO, P: m_c_BinOp(L: m_Value(V&: Y), R: m_Specific(V: X)))) |
97 | return nullptr; |
98 | |
99 | // +0.0 compares equal to -0.0, and so it does not behave as required for this |
100 | // transform. Bail out if we can not exclude that possibility. |
101 | if (isa<FPMathOperator>(Val: BO)) |
102 | if (!BO->hasNoSignedZeros() && |
103 | !cannotBeNegativeZero(V: Y, |
104 | SQ: IC.getSimplifyQuery().getWithInstruction(I: &Sel))) |
105 | return nullptr; |
106 | |
107 | // BO = binop Y, X |
108 | // S = { select (cmp eq X, C), BO, ? } or { select (cmp ne X, C), ?, BO } |
109 | // => |
110 | // S = { select (cmp eq X, C), Y, ? } or { select (cmp ne X, C), ?, Y } |
111 | return IC.replaceOperand(I&: Sel, OpNum: IsEq ? 1 : 2, V: Y); |
112 | } |
113 | |
114 | /// This folds: |
115 | /// select (icmp eq (and X, C1)), TC, FC |
116 | /// iff C1 is a power 2 and the difference between TC and FC is a power-of-2. |
117 | /// To something like: |
118 | /// (shr (and (X, C1)), (log2(C1) - log2(TC-FC))) + FC |
119 | /// Or: |
120 | /// (shl (and (X, C1)), (log2(TC-FC) - log2(C1))) + FC |
121 | /// With some variations depending if FC is larger than TC, or the shift |
122 | /// isn't needed, or the bit widths don't match. |
123 | static Value *foldSelectICmpAnd(SelectInst &Sel, Value *CondVal, Value *TrueVal, |
124 | Value *FalseVal, Value *V, const APInt &AndMask, |
125 | bool CreateAnd, |
126 | InstCombiner::BuilderTy &Builder) { |
127 | const APInt *SelTC, *SelFC; |
128 | if (!match(V: TrueVal, P: m_APInt(Res&: SelTC)) || !match(V: FalseVal, P: m_APInt(Res&: SelFC))) |
129 | return nullptr; |
130 | |
131 | Type *SelType = Sel.getType(); |
132 | // In general, when both constants are non-zero, we would need an offset to |
133 | // replace the select. This would require more instructions than we started |
134 | // with. But there's one special-case that we handle here because it can |
135 | // simplify/reduce the instructions. |
136 | const APInt &TC = *SelTC; |
137 | const APInt &FC = *SelFC; |
138 | if (!TC.isZero() && !FC.isZero()) { |
139 | if (TC.getBitWidth() != AndMask.getBitWidth()) |
140 | return nullptr; |
141 | // If we have to create an 'and', then we must kill the cmp to not |
142 | // increase the instruction count. |
143 | if (CreateAnd && !CondVal->hasOneUse()) |
144 | return nullptr; |
145 | |
146 | // (V & AndMaskC) == 0 ? TC : FC --> TC | (V & AndMaskC) |
147 | // (V & AndMaskC) == 0 ? TC : FC --> TC ^ (V & AndMaskC) |
148 | // (V & AndMaskC) == 0 ? TC : FC --> TC + (V & AndMaskC) |
149 | // (V & AndMaskC) == 0 ? TC : FC --> TC - (V & AndMaskC) |
150 | Constant *TCC = ConstantInt::get(Ty: SelType, V: TC); |
151 | Constant *FCC = ConstantInt::get(Ty: SelType, V: FC); |
152 | Constant *MaskC = ConstantInt::get(Ty: SelType, V: AndMask); |
153 | for (auto Opc : {Instruction::Or, Instruction::Xor, Instruction::Add, |
154 | Instruction::Sub}) { |
155 | if (ConstantFoldBinaryOpOperands(Opcode: Opc, LHS: TCC, RHS: MaskC, DL: Sel.getDataLayout()) == |
156 | FCC) { |
157 | if (CreateAnd) |
158 | V = Builder.CreateAnd(LHS: V, RHS: MaskC); |
159 | return Builder.CreateBinOp(Opc, LHS: TCC, RHS: V); |
160 | } |
161 | } |
162 | |
163 | return nullptr; |
164 | } |
165 | |
166 | // Make sure one of the select arms is a power-of-2. |
167 | if (!TC.isPowerOf2() && !FC.isPowerOf2()) |
168 | return nullptr; |
169 | |
170 | // Determine which shift is needed to transform result of the 'and' into the |
171 | // desired result. |
172 | const APInt &ValC = !TC.isZero() ? TC : FC; |
173 | unsigned ValZeros = ValC.logBase2(); |
174 | unsigned AndZeros = AndMask.logBase2(); |
175 | bool ShouldNotVal = !TC.isZero(); |
176 | bool NeedShift = ValZeros != AndZeros; |
177 | bool NeedZExtTrunc = |
178 | SelType->getScalarSizeInBits() != V->getType()->getScalarSizeInBits(); |
179 | |
180 | // If we would need to create an 'and' + 'shift' + 'xor' + cast to replace |
181 | // a 'select' + 'icmp', then this transformation would result in more |
182 | // instructions and potentially interfere with other folding. |
183 | if (CreateAnd + ShouldNotVal + NeedShift + NeedZExtTrunc > |
184 | 1 + CondVal->hasOneUse()) |
185 | return nullptr; |
186 | |
187 | // Insert the 'and' instruction on the input to the truncate. |
188 | if (CreateAnd) |
189 | V = Builder.CreateAnd(LHS: V, RHS: ConstantInt::get(Ty: V->getType(), V: AndMask)); |
190 | |
191 | // If types don't match, we can still convert the select by introducing a zext |
192 | // or a trunc of the 'and'. |
193 | if (ValZeros > AndZeros) { |
194 | V = Builder.CreateZExtOrTrunc(V, DestTy: SelType); |
195 | V = Builder.CreateShl(LHS: V, RHS: ValZeros - AndZeros); |
196 | } else if (ValZeros < AndZeros) { |
197 | V = Builder.CreateLShr(LHS: V, RHS: AndZeros - ValZeros); |
198 | V = Builder.CreateZExtOrTrunc(V, DestTy: SelType); |
199 | } else { |
200 | V = Builder.CreateZExtOrTrunc(V, DestTy: SelType); |
201 | } |
202 | |
203 | // Okay, now we know that everything is set up, we just don't know whether we |
204 | // have a icmp_ne or icmp_eq and whether the true or false val is the zero. |
205 | if (ShouldNotVal) |
206 | V = Builder.CreateXor(LHS: V, RHS: ValC); |
207 | |
208 | return V; |
209 | } |
210 | |
211 | /// We want to turn code that looks like this: |
212 | /// %C = or %A, %B |
213 | /// %D = select %cond, %C, %A |
214 | /// into: |
215 | /// %C = select %cond, %B, 0 |
216 | /// %D = or %A, %C |
217 | /// |
218 | /// Assuming that the specified instruction is an operand to the select, return |
219 | /// a bitmask indicating which operands of this instruction are foldable if they |
220 | /// equal the other incoming value of the select. |
221 | static unsigned getSelectFoldableOperands(BinaryOperator *I) { |
222 | switch (I->getOpcode()) { |
223 | case Instruction::Add: |
224 | case Instruction::FAdd: |
225 | case Instruction::Mul: |
226 | case Instruction::FMul: |
227 | case Instruction::And: |
228 | case Instruction::Or: |
229 | case Instruction::Xor: |
230 | return 3; // Can fold through either operand. |
231 | case Instruction::Sub: // Can only fold on the amount subtracted. |
232 | case Instruction::FSub: |
233 | case Instruction::FDiv: // Can only fold on the divisor amount. |
234 | case Instruction::Shl: // Can only fold on the shift amount. |
235 | case Instruction::LShr: |
236 | case Instruction::AShr: |
237 | return 1; |
238 | default: |
239 | return 0; // Cannot fold |
240 | } |
241 | } |
242 | |
243 | /// We have (select c, TI, FI), and we know that TI and FI have the same opcode. |
244 | Instruction *InstCombinerImpl::foldSelectOpOp(SelectInst &SI, Instruction *TI, |
245 | Instruction *FI) { |
246 | // Don't break up min/max patterns. The hasOneUse checks below prevent that |
247 | // for most cases, but vector min/max with bitcasts can be transformed. If the |
248 | // one-use restrictions are eased for other patterns, we still don't want to |
249 | // obfuscate min/max. |
250 | if ((match(V: &SI, P: m_SMin(L: m_Value(), R: m_Value())) || |
251 | match(V: &SI, P: m_SMax(L: m_Value(), R: m_Value())) || |
252 | match(V: &SI, P: m_UMin(L: m_Value(), R: m_Value())) || |
253 | match(V: &SI, P: m_UMax(L: m_Value(), R: m_Value())))) |
254 | return nullptr; |
255 | |
256 | // If this is a cast from the same type, merge. |
257 | Value *Cond = SI.getCondition(); |
258 | Type *CondTy = Cond->getType(); |
259 | if (TI->getNumOperands() == 1 && TI->isCast()) { |
260 | Type *FIOpndTy = FI->getOperand(i: 0)->getType(); |
261 | if (TI->getOperand(i: 0)->getType() != FIOpndTy) |
262 | return nullptr; |
263 | |
264 | // The select condition may be a vector. We may only change the operand |
265 | // type if the vector width remains the same (and matches the condition). |
266 | if (auto *CondVTy = dyn_cast<VectorType>(Val: CondTy)) { |
267 | if (!FIOpndTy->isVectorTy() || |
268 | CondVTy->getElementCount() != |
269 | cast<VectorType>(Val: FIOpndTy)->getElementCount()) |
270 | return nullptr; |
271 | |
272 | // TODO: If the backend knew how to deal with casts better, we could |
273 | // remove this limitation. For now, there's too much potential to create |
274 | // worse codegen by promoting the select ahead of size-altering casts |
275 | // (PR28160). |
276 | // |
277 | // Note that ValueTracking's matchSelectPattern() looks through casts |
278 | // without checking 'hasOneUse' when it matches min/max patterns, so this |
279 | // transform may end up happening anyway. |
280 | if (TI->getOpcode() != Instruction::BitCast && |
281 | (!TI->hasOneUse() || !FI->hasOneUse())) |
282 | return nullptr; |
283 | } else if (!TI->hasOneUse() || !FI->hasOneUse()) { |
284 | // TODO: The one-use restrictions for a scalar select could be eased if |
285 | // the fold of a select in visitLoadInst() was enhanced to match a pattern |
286 | // that includes a cast. |
287 | return nullptr; |
288 | } |
289 | |
290 | // Fold this by inserting a select from the input values. |
291 | Value *NewSI = |
292 | Builder.CreateSelect(C: Cond, True: TI->getOperand(i: 0), False: FI->getOperand(i: 0), |
293 | Name: SI.getName() + ".v" , MDFrom: &SI); |
294 | return CastInst::Create(Instruction::CastOps(TI->getOpcode()), S: NewSI, |
295 | Ty: TI->getType()); |
296 | } |
297 | |
298 | Value *OtherOpT, *OtherOpF; |
299 | bool MatchIsOpZero; |
300 | auto getCommonOp = [&](Instruction *TI, Instruction *FI, bool Commute, |
301 | bool Swapped = false) -> Value * { |
302 | assert(!(Commute && Swapped) && |
303 | "Commute and Swapped can't set at the same time" ); |
304 | if (!Swapped) { |
305 | if (TI->getOperand(i: 0) == FI->getOperand(i: 0)) { |
306 | OtherOpT = TI->getOperand(i: 1); |
307 | OtherOpF = FI->getOperand(i: 1); |
308 | MatchIsOpZero = true; |
309 | return TI->getOperand(i: 0); |
310 | } else if (TI->getOperand(i: 1) == FI->getOperand(i: 1)) { |
311 | OtherOpT = TI->getOperand(i: 0); |
312 | OtherOpF = FI->getOperand(i: 0); |
313 | MatchIsOpZero = false; |
314 | return TI->getOperand(i: 1); |
315 | } |
316 | } |
317 | |
318 | if (!Commute && !Swapped) |
319 | return nullptr; |
320 | |
321 | // If we are allowing commute or swap of operands, then |
322 | // allow a cross-operand match. In that case, MatchIsOpZero |
323 | // means that TI's operand 0 (FI's operand 1) is the common op. |
324 | if (TI->getOperand(i: 0) == FI->getOperand(i: 1)) { |
325 | OtherOpT = TI->getOperand(i: 1); |
326 | OtherOpF = FI->getOperand(i: 0); |
327 | MatchIsOpZero = true; |
328 | return TI->getOperand(i: 0); |
329 | } else if (TI->getOperand(i: 1) == FI->getOperand(i: 0)) { |
330 | OtherOpT = TI->getOperand(i: 0); |
331 | OtherOpF = FI->getOperand(i: 1); |
332 | MatchIsOpZero = false; |
333 | return TI->getOperand(i: 1); |
334 | } |
335 | return nullptr; |
336 | }; |
337 | |
338 | if (TI->hasOneUse() || FI->hasOneUse()) { |
339 | // Cond ? -X : -Y --> -(Cond ? X : Y) |
340 | Value *X, *Y; |
341 | if (match(V: TI, P: m_FNeg(X: m_Value(V&: X))) && match(V: FI, P: m_FNeg(X: m_Value(V&: Y)))) { |
342 | // Intersect FMF from the fneg instructions and union those with the |
343 | // select. |
344 | FastMathFlags FMF = TI->getFastMathFlags(); |
345 | FMF &= FI->getFastMathFlags(); |
346 | FMF |= SI.getFastMathFlags(); |
347 | Value *NewSel = |
348 | Builder.CreateSelect(C: Cond, True: X, False: Y, Name: SI.getName() + ".v" , MDFrom: &SI); |
349 | if (auto *NewSelI = dyn_cast<Instruction>(Val: NewSel)) |
350 | NewSelI->setFastMathFlags(FMF); |
351 | Instruction *NewFNeg = UnaryOperator::CreateFNeg(V: NewSel); |
352 | NewFNeg->setFastMathFlags(FMF); |
353 | return NewFNeg; |
354 | } |
355 | |
356 | // Min/max intrinsic with a common operand can have the common operand |
357 | // pulled after the select. This is the same transform as below for binops, |
358 | // but specialized for intrinsic matching and without the restrictive uses |
359 | // clause. |
360 | auto *TII = dyn_cast<IntrinsicInst>(Val: TI); |
361 | auto *FII = dyn_cast<IntrinsicInst>(Val: FI); |
362 | if (TII && FII && TII->getIntrinsicID() == FII->getIntrinsicID()) { |
363 | if (match(V: TII, P: m_MaxOrMin(L: m_Value(), R: m_Value()))) { |
364 | if (Value *MatchOp = getCommonOp(TI, FI, true)) { |
365 | Value *NewSel = |
366 | Builder.CreateSelect(C: Cond, True: OtherOpT, False: OtherOpF, Name: "minmaxop" , MDFrom: &SI); |
367 | return CallInst::Create(Func: TII->getCalledFunction(), Args: {NewSel, MatchOp}); |
368 | } |
369 | } |
370 | |
371 | // select c, (ldexp v, e0), (ldexp v, e1) -> ldexp v, (select c, e0, e1) |
372 | // select c, (ldexp v0, e), (ldexp v1, e) -> ldexp (select c, v0, v1), e |
373 | // |
374 | // select c, (ldexp v0, e0), (ldexp v1, e1) -> |
375 | // ldexp (select c, v0, v1), (select c, e0, e1) |
376 | if (TII->getIntrinsicID() == Intrinsic::ldexp) { |
377 | Value *LdexpVal0 = TII->getArgOperand(i: 0); |
378 | Value *LdexpExp0 = TII->getArgOperand(i: 1); |
379 | Value *LdexpVal1 = FII->getArgOperand(i: 0); |
380 | Value *LdexpExp1 = FII->getArgOperand(i: 1); |
381 | if (LdexpExp0->getType() == LdexpExp1->getType()) { |
382 | FPMathOperator *SelectFPOp = cast<FPMathOperator>(Val: &SI); |
383 | FastMathFlags FMF = cast<FPMathOperator>(Val: TII)->getFastMathFlags(); |
384 | FMF &= cast<FPMathOperator>(Val: FII)->getFastMathFlags(); |
385 | FMF |= SelectFPOp->getFastMathFlags(); |
386 | |
387 | Value *SelectVal = Builder.CreateSelect(C: Cond, True: LdexpVal0, False: LdexpVal1); |
388 | Value *SelectExp = Builder.CreateSelect(C: Cond, True: LdexpExp0, False: LdexpExp1); |
389 | |
390 | CallInst *NewLdexp = Builder.CreateIntrinsic( |
391 | RetTy: TII->getType(), ID: Intrinsic::ldexp, Args: {SelectVal, SelectExp}); |
392 | NewLdexp->setFastMathFlags(FMF); |
393 | return replaceInstUsesWith(I&: SI, V: NewLdexp); |
394 | } |
395 | } |
396 | } |
397 | |
398 | auto CreateCmpSel = [&](std::optional<CmpPredicate> P, |
399 | bool Swapped) -> CmpInst * { |
400 | if (!P) |
401 | return nullptr; |
402 | auto *MatchOp = getCommonOp(TI, FI, ICmpInst::isEquality(P: *P), |
403 | ICmpInst::isRelational(P: *P) && Swapped); |
404 | if (!MatchOp) |
405 | return nullptr; |
406 | Value *NewSel = Builder.CreateSelect(C: Cond, True: OtherOpT, False: OtherOpF, |
407 | Name: SI.getName() + ".v" , MDFrom: &SI); |
408 | return new ICmpInst(MatchIsOpZero ? *P |
409 | : ICmpInst::getSwappedCmpPredicate(Pred: *P), |
410 | MatchOp, NewSel); |
411 | }; |
412 | |
413 | // icmp with a common operand also can have the common operand |
414 | // pulled after the select. |
415 | CmpPredicate TPred, FPred; |
416 | if (match(V: TI, P: m_ICmp(Pred&: TPred, L: m_Value(), R: m_Value())) && |
417 | match(V: FI, P: m_ICmp(Pred&: FPred, L: m_Value(), R: m_Value()))) { |
418 | if (auto *R = |
419 | CreateCmpSel(CmpPredicate::getMatching(A: TPred, B: FPred), false)) |
420 | return R; |
421 | if (auto *R = |
422 | CreateCmpSel(CmpPredicate::getMatching( |
423 | A: TPred, B: ICmpInst::getSwappedCmpPredicate(Pred: FPred)), |
424 | true)) |
425 | return R; |
426 | } |
427 | } |
428 | |
429 | // Only handle binary operators (including two-operand getelementptr) with |
430 | // one-use here. As with the cast case above, it may be possible to relax the |
431 | // one-use constraint, but that needs be examined carefully since it may not |
432 | // reduce the total number of instructions. |
433 | if (TI->getNumOperands() != 2 || FI->getNumOperands() != 2 || |
434 | !TI->isSameOperationAs(I: FI) || |
435 | (!isa<BinaryOperator>(Val: TI) && !isa<GetElementPtrInst>(Val: TI)) || |
436 | !TI->hasOneUse() || !FI->hasOneUse()) |
437 | return nullptr; |
438 | |
439 | // Figure out if the operations have any operands in common. |
440 | Value *MatchOp = getCommonOp(TI, FI, TI->isCommutative()); |
441 | if (!MatchOp) |
442 | return nullptr; |
443 | |
444 | // If the select condition is a vector, the operands of the original select's |
445 | // operands also must be vectors. This may not be the case for getelementptr |
446 | // for example. |
447 | if (CondTy->isVectorTy() && (!OtherOpT->getType()->isVectorTy() || |
448 | !OtherOpF->getType()->isVectorTy())) |
449 | return nullptr; |
450 | |
451 | // If we are sinking div/rem after a select, we may need to freeze the |
452 | // condition because div/rem may induce immediate UB with a poison operand. |
453 | // For example, the following transform is not safe if Cond can ever be poison |
454 | // because we can replace poison with zero and then we have div-by-zero that |
455 | // didn't exist in the original code: |
456 | // Cond ? x/y : x/z --> x / (Cond ? y : z) |
457 | auto *BO = dyn_cast<BinaryOperator>(Val: TI); |
458 | if (BO && BO->isIntDivRem() && !isGuaranteedNotToBePoison(V: Cond)) { |
459 | // A udiv/urem with a common divisor is safe because UB can only occur with |
460 | // div-by-zero, and that would be present in the original code. |
461 | if (BO->getOpcode() == Instruction::SDiv || |
462 | BO->getOpcode() == Instruction::SRem || MatchIsOpZero) |
463 | Cond = Builder.CreateFreeze(V: Cond); |
464 | } |
465 | |
466 | // If we reach here, they do have operations in common. |
467 | Value *NewSI = Builder.CreateSelect(C: Cond, True: OtherOpT, False: OtherOpF, |
468 | Name: SI.getName() + ".v" , MDFrom: &SI); |
469 | Value *Op0 = MatchIsOpZero ? MatchOp : NewSI; |
470 | Value *Op1 = MatchIsOpZero ? NewSI : MatchOp; |
471 | if (auto *BO = dyn_cast<BinaryOperator>(Val: TI)) { |
472 | BinaryOperator *NewBO = BinaryOperator::Create(Op: BO->getOpcode(), S1: Op0, S2: Op1); |
473 | NewBO->copyIRFlags(V: TI); |
474 | NewBO->andIRFlags(V: FI); |
475 | return NewBO; |
476 | } |
477 | if (auto *TGEP = dyn_cast<GetElementPtrInst>(Val: TI)) { |
478 | auto *FGEP = cast<GetElementPtrInst>(Val: FI); |
479 | Type *ElementType = TGEP->getSourceElementType(); |
480 | return GetElementPtrInst::Create( |
481 | PointeeType: ElementType, Ptr: Op0, IdxList: Op1, NW: TGEP->getNoWrapFlags() & FGEP->getNoWrapFlags()); |
482 | } |
483 | llvm_unreachable("Expected BinaryOperator or GEP" ); |
484 | return nullptr; |
485 | } |
486 | |
487 | static bool isSelect01(const APInt &C1I, const APInt &C2I) { |
488 | if (!C1I.isZero() && !C2I.isZero()) // One side must be zero. |
489 | return false; |
490 | return C1I.isOne() || C1I.isAllOnes() || C2I.isOne() || C2I.isAllOnes(); |
491 | } |
492 | |
493 | /// Try to fold the select into one of the operands to allow further |
494 | /// optimization. |
495 | Instruction *InstCombinerImpl::foldSelectIntoOp(SelectInst &SI, Value *TrueVal, |
496 | Value *FalseVal) { |
497 | // See the comment above getSelectFoldableOperands for a description of the |
498 | // transformation we are doing here. |
499 | auto TryFoldSelectIntoOp = [&](SelectInst &SI, Value *TrueVal, |
500 | Value *FalseVal, |
501 | bool Swapped) -> Instruction * { |
502 | auto *TVI = dyn_cast<BinaryOperator>(Val: TrueVal); |
503 | if (!TVI || !TVI->hasOneUse() || isa<Constant>(Val: FalseVal)) |
504 | return nullptr; |
505 | |
506 | unsigned SFO = getSelectFoldableOperands(I: TVI); |
507 | unsigned OpToFold = 0; |
508 | if ((SFO & 1) && FalseVal == TVI->getOperand(i_nocapture: 0)) |
509 | OpToFold = 1; |
510 | else if ((SFO & 2) && FalseVal == TVI->getOperand(i_nocapture: 1)) |
511 | OpToFold = 2; |
512 | |
513 | if (!OpToFold) |
514 | return nullptr; |
515 | |
516 | FastMathFlags FMF; |
517 | if (isa<FPMathOperator>(Val: &SI)) |
518 | FMF = SI.getFastMathFlags(); |
519 | Constant *C = ConstantExpr::getBinOpIdentity( |
520 | Opcode: TVI->getOpcode(), Ty: TVI->getType(), AllowRHSConstant: true, NSZ: FMF.noSignedZeros()); |
521 | Value *OOp = TVI->getOperand(i_nocapture: 2 - OpToFold); |
522 | // Avoid creating select between 2 constants unless it's selecting |
523 | // between 0, 1 and -1. |
524 | const APInt *OOpC; |
525 | bool OOpIsAPInt = match(V: OOp, P: m_APInt(Res&: OOpC)); |
526 | if (isa<Constant>(Val: OOp) && |
527 | (!OOpIsAPInt || !isSelect01(C1I: C->getUniqueInteger(), C2I: *OOpC))) |
528 | return nullptr; |
529 | |
530 | // If the false value is a NaN then we have that the floating point math |
531 | // operation in the transformed code may not preserve the exact NaN |
532 | // bit-pattern -- e.g. `fadd sNaN, 0.0 -> qNaN`. |
533 | // This makes the transformation incorrect since the original program would |
534 | // have preserved the exact NaN bit-pattern. |
535 | // Avoid the folding if the false value might be a NaN. |
536 | if (isa<FPMathOperator>(Val: &SI) && |
537 | !computeKnownFPClass(Val: FalseVal, FMF, Interested: fcNan, CtxI: &SI).isKnownNeverNaN()) |
538 | return nullptr; |
539 | |
540 | Value *NewSel = Builder.CreateSelect(C: SI.getCondition(), True: Swapped ? C : OOp, |
541 | False: Swapped ? OOp : C, Name: "" , MDFrom: &SI); |
542 | if (isa<FPMathOperator>(Val: &SI)) |
543 | cast<Instruction>(Val: NewSel)->setFastMathFlags(FMF); |
544 | NewSel->takeName(V: TVI); |
545 | BinaryOperator *BO = |
546 | BinaryOperator::Create(Op: TVI->getOpcode(), S1: FalseVal, S2: NewSel); |
547 | BO->copyIRFlags(V: TVI); |
548 | if (isa<FPMathOperator>(Val: &SI)) { |
549 | // Merge poison generating flags from the select. |
550 | BO->setHasNoNaNs(BO->hasNoNaNs() && FMF.noNaNs()); |
551 | BO->setHasNoInfs(BO->hasNoInfs() && FMF.noInfs()); |
552 | // Merge no-signed-zeros flag from the select. |
553 | // Otherwise we may produce zeros with different sign. |
554 | BO->setHasNoSignedZeros(BO->hasNoSignedZeros() && FMF.noSignedZeros()); |
555 | } |
556 | return BO; |
557 | }; |
558 | |
559 | if (Instruction *R = TryFoldSelectIntoOp(SI, TrueVal, FalseVal, false)) |
560 | return R; |
561 | |
562 | if (Instruction *R = TryFoldSelectIntoOp(SI, FalseVal, TrueVal, true)) |
563 | return R; |
564 | |
565 | return nullptr; |
566 | } |
567 | |
568 | /// Try to fold a select to a min/max intrinsic. Many cases are already handled |
569 | /// by matchDecomposedSelectPattern but here we handle the cases where more |
570 | /// extensive modification of the IR is required. |
571 | static Value *foldSelectICmpMinMax(const ICmpInst *Cmp, Value *TVal, |
572 | Value *FVal, |
573 | InstCombiner::BuilderTy &Builder, |
574 | const SimplifyQuery &SQ) { |
575 | const Value *CmpLHS = Cmp->getOperand(i_nocapture: 0); |
576 | const Value *CmpRHS = Cmp->getOperand(i_nocapture: 1); |
577 | ICmpInst::Predicate Pred = Cmp->getPredicate(); |
578 | |
579 | // (X > Y) ? X : (Y - 1) ==> MIN(X, Y - 1) |
580 | // (X < Y) ? X : (Y + 1) ==> MAX(X, Y + 1) |
581 | // This transformation is valid when overflow corresponding to the sign of |
582 | // the comparison is poison and we must drop the non-matching overflow flag. |
583 | if (CmpRHS == TVal) { |
584 | std::swap(a&: CmpLHS, b&: CmpRHS); |
585 | Pred = CmpInst::getSwappedPredicate(pred: Pred); |
586 | } |
587 | |
588 | // TODO: consider handling 'or disjoint' as well, though these would need to |
589 | // be converted to 'add' instructions. |
590 | if (!(CmpLHS == TVal && isa<Instruction>(Val: FVal))) |
591 | return nullptr; |
592 | |
593 | if (Pred == CmpInst::ICMP_SGT && |
594 | match(V: FVal, P: m_NSWAdd(L: m_Specific(V: CmpRHS), R: m_One()))) { |
595 | cast<Instruction>(Val: FVal)->setHasNoUnsignedWrap(false); |
596 | return Builder.CreateBinaryIntrinsic(ID: Intrinsic::smax, LHS: TVal, RHS: FVal); |
597 | } |
598 | |
599 | if (Pred == CmpInst::ICMP_SLT && |
600 | match(V: FVal, P: m_NSWAdd(L: m_Specific(V: CmpRHS), R: m_AllOnes()))) { |
601 | cast<Instruction>(Val: FVal)->setHasNoUnsignedWrap(false); |
602 | return Builder.CreateBinaryIntrinsic(ID: Intrinsic::smin, LHS: TVal, RHS: FVal); |
603 | } |
604 | |
605 | if (Pred == CmpInst::ICMP_UGT && |
606 | match(V: FVal, P: m_NUWAdd(L: m_Specific(V: CmpRHS), R: m_One()))) { |
607 | cast<Instruction>(Val: FVal)->setHasNoSignedWrap(false); |
608 | return Builder.CreateBinaryIntrinsic(ID: Intrinsic::umax, LHS: TVal, RHS: FVal); |
609 | } |
610 | |
611 | // Note: We must use isKnownNonZero here because "sub nuw %x, 1" will be |
612 | // canonicalized to "add %x, -1" discarding the nuw flag. |
613 | if (Pred == CmpInst::ICMP_ULT && |
614 | match(V: FVal, P: m_Add(L: m_Specific(V: CmpRHS), R: m_AllOnes())) && |
615 | isKnownNonZero(V: CmpRHS, Q: SQ)) { |
616 | cast<Instruction>(Val: FVal)->setHasNoSignedWrap(false); |
617 | cast<Instruction>(Val: FVal)->setHasNoUnsignedWrap(false); |
618 | return Builder.CreateBinaryIntrinsic(ID: Intrinsic::umin, LHS: TVal, RHS: FVal); |
619 | } |
620 | |
621 | return nullptr; |
622 | } |
623 | |
624 | /// We want to turn: |
625 | /// (select (icmp eq (and X, Y), 0), (and (lshr X, Z), 1), 1) |
626 | /// into: |
627 | /// zext (icmp ne i32 (and X, (or Y, (shl 1, Z))), 0) |
628 | /// Note: |
629 | /// Z may be 0 if lshr is missing. |
630 | /// Worst-case scenario is that we will replace 5 instructions with 5 different |
631 | /// instructions, but we got rid of select. |
632 | static Instruction *foldSelectICmpAndAnd(Type *SelType, const ICmpInst *Cmp, |
633 | Value *TVal, Value *FVal, |
634 | InstCombiner::BuilderTy &Builder) { |
635 | if (!(Cmp->hasOneUse() && Cmp->getOperand(i_nocapture: 0)->hasOneUse() && |
636 | Cmp->getPredicate() == ICmpInst::ICMP_EQ && |
637 | match(V: Cmp->getOperand(i_nocapture: 1), P: m_Zero()) && match(V: FVal, P: m_One()))) |
638 | return nullptr; |
639 | |
640 | // The TrueVal has general form of: and %B, 1 |
641 | Value *B; |
642 | if (!match(V: TVal, P: m_OneUse(SubPattern: m_And(L: m_Value(V&: B), R: m_One())))) |
643 | return nullptr; |
644 | |
645 | // Where %B may be optionally shifted: lshr %X, %Z. |
646 | Value *X, *Z; |
647 | const bool HasShift = match(V: B, P: m_OneUse(SubPattern: m_LShr(L: m_Value(V&: X), R: m_Value(V&: Z)))); |
648 | |
649 | // The shift must be valid. |
650 | // TODO: This restricts the fold to constant shift amounts. Is there a way to |
651 | // handle variable shifts safely? PR47012 |
652 | if (HasShift && |
653 | !match(V: Z, P: m_SpecificInt_ICMP(Predicate: CmpInst::ICMP_ULT, |
654 | Threshold: APInt(SelType->getScalarSizeInBits(), |
655 | SelType->getScalarSizeInBits())))) |
656 | return nullptr; |
657 | |
658 | if (!HasShift) |
659 | X = B; |
660 | |
661 | Value *Y; |
662 | if (!match(V: Cmp->getOperand(i_nocapture: 0), P: m_c_And(L: m_Specific(V: X), R: m_Value(V&: Y)))) |
663 | return nullptr; |
664 | |
665 | // ((X & Y) == 0) ? ((X >> Z) & 1) : 1 --> (X & (Y | (1 << Z))) != 0 |
666 | // ((X & Y) == 0) ? (X & 1) : 1 --> (X & (Y | 1)) != 0 |
667 | Constant *One = ConstantInt::get(Ty: SelType, V: 1); |
668 | Value *MaskB = HasShift ? Builder.CreateShl(LHS: One, RHS: Z) : One; |
669 | Value *FullMask = Builder.CreateOr(LHS: Y, RHS: MaskB); |
670 | Value *MaskedX = Builder.CreateAnd(LHS: X, RHS: FullMask); |
671 | Value *ICmpNeZero = Builder.CreateIsNotNull(Arg: MaskedX); |
672 | return new ZExtInst(ICmpNeZero, SelType); |
673 | } |
674 | |
675 | /// We want to turn: |
676 | /// (select (icmp eq (and X, C1), 0), 0, (shl [nsw/nuw] X, C2)); |
677 | /// iff C1 is a mask and the number of its leading zeros is equal to C2 |
678 | /// into: |
679 | /// shl X, C2 |
680 | static Value *foldSelectICmpAndZeroShl(const ICmpInst *Cmp, Value *TVal, |
681 | Value *FVal, |
682 | InstCombiner::BuilderTy &Builder) { |
683 | CmpPredicate Pred; |
684 | Value *AndVal; |
685 | if (!match(V: Cmp, P: m_ICmp(Pred, L: m_Value(V&: AndVal), R: m_Zero()))) |
686 | return nullptr; |
687 | |
688 | if (Pred == ICmpInst::ICMP_NE) { |
689 | Pred = ICmpInst::ICMP_EQ; |
690 | std::swap(a&: TVal, b&: FVal); |
691 | } |
692 | |
693 | Value *X; |
694 | const APInt *C2, *C1; |
695 | if (Pred != ICmpInst::ICMP_EQ || |
696 | !match(V: AndVal, P: m_And(L: m_Value(V&: X), R: m_APInt(Res&: C1))) || |
697 | !match(V: TVal, P: m_Zero()) || !match(V: FVal, P: m_Shl(L: m_Specific(V: X), R: m_APInt(Res&: C2)))) |
698 | return nullptr; |
699 | |
700 | if (!C1->isMask() || |
701 | C1->countLeadingZeros() != static_cast<unsigned>(C2->getZExtValue())) |
702 | return nullptr; |
703 | |
704 | auto *FI = dyn_cast<Instruction>(Val: FVal); |
705 | if (!FI) |
706 | return nullptr; |
707 | |
708 | FI->setHasNoSignedWrap(false); |
709 | FI->setHasNoUnsignedWrap(false); |
710 | return FVal; |
711 | } |
712 | |
713 | /// We want to turn: |
714 | /// (select (icmp sgt x, C), lshr (X, Y), ashr (X, Y)); iff C s>= -1 |
715 | /// (select (icmp slt x, C), ashr (X, Y), lshr (X, Y)); iff C s>= 0 |
716 | /// into: |
717 | /// ashr (X, Y) |
718 | static Value *foldSelectICmpLshrAshr(const ICmpInst *IC, Value *TrueVal, |
719 | Value *FalseVal, |
720 | InstCombiner::BuilderTy &Builder) { |
721 | ICmpInst::Predicate Pred = IC->getPredicate(); |
722 | Value *CmpLHS = IC->getOperand(i_nocapture: 0); |
723 | Value *CmpRHS = IC->getOperand(i_nocapture: 1); |
724 | if (!CmpRHS->getType()->isIntOrIntVectorTy()) |
725 | return nullptr; |
726 | |
727 | Value *X, *Y; |
728 | unsigned Bitwidth = CmpRHS->getType()->getScalarSizeInBits(); |
729 | if ((Pred != ICmpInst::ICMP_SGT || |
730 | !match(V: CmpRHS, P: m_SpecificInt_ICMP(Predicate: ICmpInst::ICMP_SGE, |
731 | Threshold: APInt::getAllOnes(numBits: Bitwidth)))) && |
732 | (Pred != ICmpInst::ICMP_SLT || |
733 | !match(V: CmpRHS, P: m_SpecificInt_ICMP(Predicate: ICmpInst::ICMP_SGE, |
734 | Threshold: APInt::getZero(numBits: Bitwidth))))) |
735 | return nullptr; |
736 | |
737 | // Canonicalize so that ashr is in FalseVal. |
738 | if (Pred == ICmpInst::ICMP_SLT) |
739 | std::swap(a&: TrueVal, b&: FalseVal); |
740 | |
741 | if (match(V: TrueVal, P: m_LShr(L: m_Value(V&: X), R: m_Value(V&: Y))) && |
742 | match(V: FalseVal, P: m_AShr(L: m_Specific(V: X), R: m_Specific(V: Y))) && |
743 | match(V: CmpLHS, P: m_Specific(V: X))) { |
744 | const auto *Ashr = cast<Instruction>(Val: FalseVal); |
745 | // if lshr is not exact and ashr is, this new ashr must not be exact. |
746 | bool IsExact = Ashr->isExact() && cast<Instruction>(Val: TrueVal)->isExact(); |
747 | return Builder.CreateAShr(LHS: X, RHS: Y, Name: IC->getName(), isExact: IsExact); |
748 | } |
749 | |
750 | return nullptr; |
751 | } |
752 | |
753 | /// We want to turn: |
754 | /// (select (icmp eq (and X, C1), 0), Y, (BinOp Y, C2)) |
755 | /// into: |
756 | /// IF C2 u>= C1 |
757 | /// (BinOp Y, (shl (and X, C1), C3)) |
758 | /// ELSE |
759 | /// (BinOp Y, (lshr (and X, C1), C3)) |
760 | /// iff: |
761 | /// 0 on the RHS is the identity value (i.e add, xor, shl, etc...) |
762 | /// C1 and C2 are both powers of 2 |
763 | /// where: |
764 | /// IF C2 u>= C1 |
765 | /// C3 = Log(C2) - Log(C1) |
766 | /// ELSE |
767 | /// C3 = Log(C1) - Log(C2) |
768 | /// |
769 | /// This transform handles cases where: |
770 | /// 1. The icmp predicate is inverted |
771 | /// 2. The select operands are reversed |
772 | /// 3. The magnitude of C2 and C1 are flipped |
773 | static Value *foldSelectICmpAndBinOp(Value *CondVal, Value *TrueVal, |
774 | Value *FalseVal, Value *V, |
775 | const APInt &AndMask, bool CreateAnd, |
776 | InstCombiner::BuilderTy &Builder) { |
777 | // Only handle integer compares. |
778 | if (!TrueVal->getType()->isIntOrIntVectorTy()) |
779 | return nullptr; |
780 | |
781 | unsigned C1Log = AndMask.logBase2(); |
782 | Value *Y; |
783 | BinaryOperator *BinOp; |
784 | const APInt *C2; |
785 | bool NeedXor; |
786 | if (match(V: FalseVal, P: m_BinOp(L: m_Specific(V: TrueVal), R: m_Power2(V&: C2)))) { |
787 | Y = TrueVal; |
788 | BinOp = cast<BinaryOperator>(Val: FalseVal); |
789 | NeedXor = false; |
790 | } else if (match(V: TrueVal, P: m_BinOp(L: m_Specific(V: FalseVal), R: m_Power2(V&: C2)))) { |
791 | Y = FalseVal; |
792 | BinOp = cast<BinaryOperator>(Val: TrueVal); |
793 | NeedXor = true; |
794 | } else { |
795 | return nullptr; |
796 | } |
797 | |
798 | // Check that 0 on RHS is identity value for this binop. |
799 | auto *IdentityC = |
800 | ConstantExpr::getBinOpIdentity(Opcode: BinOp->getOpcode(), Ty: BinOp->getType(), |
801 | /*AllowRHSConstant*/ true); |
802 | if (IdentityC == nullptr || !IdentityC->isNullValue()) |
803 | return nullptr; |
804 | |
805 | unsigned C2Log = C2->logBase2(); |
806 | |
807 | bool NeedShift = C1Log != C2Log; |
808 | bool NeedZExtTrunc = Y->getType()->getScalarSizeInBits() != |
809 | V->getType()->getScalarSizeInBits(); |
810 | |
811 | // Make sure we don't create more instructions than we save. |
812 | if ((NeedShift + NeedXor + NeedZExtTrunc + CreateAnd) > |
813 | (CondVal->hasOneUse() + BinOp->hasOneUse())) |
814 | return nullptr; |
815 | |
816 | if (CreateAnd) { |
817 | // Insert the AND instruction on the input to the truncate. |
818 | V = Builder.CreateAnd(LHS: V, RHS: ConstantInt::get(Ty: V->getType(), V: AndMask)); |
819 | } |
820 | |
821 | if (C2Log > C1Log) { |
822 | V = Builder.CreateZExtOrTrunc(V, DestTy: Y->getType()); |
823 | V = Builder.CreateShl(LHS: V, RHS: C2Log - C1Log); |
824 | } else if (C1Log > C2Log) { |
825 | V = Builder.CreateLShr(LHS: V, RHS: C1Log - C2Log); |
826 | V = Builder.CreateZExtOrTrunc(V, DestTy: Y->getType()); |
827 | } else |
828 | V = Builder.CreateZExtOrTrunc(V, DestTy: Y->getType()); |
829 | |
830 | if (NeedXor) |
831 | V = Builder.CreateXor(LHS: V, RHS: *C2); |
832 | |
833 | auto *Res = Builder.CreateBinOp(Opc: BinOp->getOpcode(), LHS: Y, RHS: V); |
834 | if (auto *BO = dyn_cast<BinaryOperator>(Val: Res)) |
835 | BO->copyIRFlags(V: BinOp); |
836 | return Res; |
837 | } |
838 | |
839 | /// Canonicalize a set or clear of a masked set of constant bits to |
840 | /// select-of-constants form. |
841 | static Instruction *foldSetClearBits(SelectInst &Sel, |
842 | InstCombiner::BuilderTy &Builder) { |
843 | Value *Cond = Sel.getCondition(); |
844 | Value *T = Sel.getTrueValue(); |
845 | Value *F = Sel.getFalseValue(); |
846 | Type *Ty = Sel.getType(); |
847 | Value *X; |
848 | const APInt *NotC, *C; |
849 | |
850 | // Cond ? (X & ~C) : (X | C) --> (X & ~C) | (Cond ? 0 : C) |
851 | if (match(V: T, P: m_And(L: m_Value(V&: X), R: m_APInt(Res&: NotC))) && |
852 | match(V: F, P: m_OneUse(SubPattern: m_Or(L: m_Specific(V: X), R: m_APInt(Res&: C)))) && *NotC == ~(*C)) { |
853 | Constant *Zero = ConstantInt::getNullValue(Ty); |
854 | Constant *OrC = ConstantInt::get(Ty, V: *C); |
855 | Value *NewSel = Builder.CreateSelect(C: Cond, True: Zero, False: OrC, Name: "masksel" , MDFrom: &Sel); |
856 | return BinaryOperator::CreateOr(V1: T, V2: NewSel); |
857 | } |
858 | |
859 | // Cond ? (X | C) : (X & ~C) --> (X & ~C) | (Cond ? C : 0) |
860 | if (match(V: F, P: m_And(L: m_Value(V&: X), R: m_APInt(Res&: NotC))) && |
861 | match(V: T, P: m_OneUse(SubPattern: m_Or(L: m_Specific(V: X), R: m_APInt(Res&: C)))) && *NotC == ~(*C)) { |
862 | Constant *Zero = ConstantInt::getNullValue(Ty); |
863 | Constant *OrC = ConstantInt::get(Ty, V: *C); |
864 | Value *NewSel = Builder.CreateSelect(C: Cond, True: OrC, False: Zero, Name: "masksel" , MDFrom: &Sel); |
865 | return BinaryOperator::CreateOr(V1: F, V2: NewSel); |
866 | } |
867 | |
868 | return nullptr; |
869 | } |
870 | |
871 | // select (x == 0), 0, x * y --> freeze(y) * x |
872 | // select (y == 0), 0, x * y --> freeze(x) * y |
873 | // select (x == 0), undef, x * y --> freeze(y) * x |
874 | // select (x == undef), 0, x * y --> freeze(y) * x |
875 | // Usage of mul instead of 0 will make the result more poisonous, |
876 | // so the operand that was not checked in the condition should be frozen. |
877 | // The latter folding is applied only when a constant compared with x is |
878 | // is a vector consisting of 0 and undefs. If a constant compared with x |
879 | // is a scalar undefined value or undefined vector then an expression |
880 | // should be already folded into a constant. |
881 | static Instruction *foldSelectZeroOrMul(SelectInst &SI, InstCombinerImpl &IC) { |
882 | auto *CondVal = SI.getCondition(); |
883 | auto *TrueVal = SI.getTrueValue(); |
884 | auto *FalseVal = SI.getFalseValue(); |
885 | Value *X, *Y; |
886 | CmpPredicate Predicate; |
887 | |
888 | // Assuming that constant compared with zero is not undef (but it may be |
889 | // a vector with some undef elements). Otherwise (when a constant is undef) |
890 | // the select expression should be already simplified. |
891 | if (!match(V: CondVal, P: m_ICmp(Pred&: Predicate, L: m_Value(V&: X), R: m_Zero())) || |
892 | !ICmpInst::isEquality(P: Predicate)) |
893 | return nullptr; |
894 | |
895 | if (Predicate == ICmpInst::ICMP_NE) |
896 | std::swap(a&: TrueVal, b&: FalseVal); |
897 | |
898 | // Check that TrueVal is a constant instead of matching it with m_Zero() |
899 | // to handle the case when it is a scalar undef value or a vector containing |
900 | // non-zero elements that are masked by undef elements in the compare |
901 | // constant. |
902 | auto *TrueValC = dyn_cast<Constant>(Val: TrueVal); |
903 | if (TrueValC == nullptr || |
904 | !match(V: FalseVal, P: m_c_Mul(L: m_Specific(V: X), R: m_Value(V&: Y))) || |
905 | !isa<Instruction>(Val: FalseVal)) |
906 | return nullptr; |
907 | |
908 | auto *ZeroC = cast<Constant>(Val: cast<Instruction>(Val: CondVal)->getOperand(i: 1)); |
909 | auto *MergedC = Constant::mergeUndefsWith(C: TrueValC, Other: ZeroC); |
910 | // If X is compared with 0 then TrueVal could be either zero or undef. |
911 | // m_Zero match vectors containing some undef elements, but for scalars |
912 | // m_Undef should be used explicitly. |
913 | if (!match(V: MergedC, P: m_Zero()) && !match(V: MergedC, P: m_Undef())) |
914 | return nullptr; |
915 | |
916 | auto *FalseValI = cast<Instruction>(Val: FalseVal); |
917 | auto *FrY = IC.InsertNewInstBefore(New: new FreezeInst(Y, Y->getName() + ".fr" ), |
918 | Old: FalseValI->getIterator()); |
919 | IC.replaceOperand(I&: *FalseValI, OpNum: FalseValI->getOperand(i: 0) == Y ? 0 : 1, V: FrY); |
920 | return IC.replaceInstUsesWith(I&: SI, V: FalseValI); |
921 | } |
922 | |
923 | /// Transform patterns such as (a > b) ? a - b : 0 into usub.sat(a, b). |
924 | /// There are 8 commuted/swapped variants of this pattern. |
925 | static Value *canonicalizeSaturatedSubtract(const ICmpInst *ICI, |
926 | const Value *TrueVal, |
927 | const Value *FalseVal, |
928 | InstCombiner::BuilderTy &Builder) { |
929 | ICmpInst::Predicate Pred = ICI->getPredicate(); |
930 | Value *A = ICI->getOperand(i_nocapture: 0); |
931 | Value *B = ICI->getOperand(i_nocapture: 1); |
932 | |
933 | // (b > a) ? 0 : a - b -> (b <= a) ? a - b : 0 |
934 | // (a == 0) ? 0 : a - 1 -> (a != 0) ? a - 1 : 0 |
935 | if (match(V: TrueVal, P: m_Zero())) { |
936 | Pred = ICmpInst::getInversePredicate(pred: Pred); |
937 | std::swap(a&: TrueVal, b&: FalseVal); |
938 | } |
939 | |
940 | if (!match(V: FalseVal, P: m_Zero())) |
941 | return nullptr; |
942 | |
943 | // ugt 0 is canonicalized to ne 0 and requires special handling |
944 | // (a != 0) ? a + -1 : 0 -> usub.sat(a, 1) |
945 | if (Pred == ICmpInst::ICMP_NE) { |
946 | if (match(V: B, P: m_Zero()) && match(V: TrueVal, P: m_Add(L: m_Specific(V: A), R: m_AllOnes()))) |
947 | return Builder.CreateBinaryIntrinsic(ID: Intrinsic::usub_sat, LHS: A, |
948 | RHS: ConstantInt::get(Ty: A->getType(), V: 1)); |
949 | return nullptr; |
950 | } |
951 | |
952 | if (!ICmpInst::isUnsigned(predicate: Pred)) |
953 | return nullptr; |
954 | |
955 | if (Pred == ICmpInst::ICMP_ULE || Pred == ICmpInst::ICMP_ULT) { |
956 | // (b < a) ? a - b : 0 -> (a > b) ? a - b : 0 |
957 | std::swap(a&: A, b&: B); |
958 | Pred = ICmpInst::getSwappedPredicate(pred: Pred); |
959 | } |
960 | |
961 | assert((Pred == ICmpInst::ICMP_UGE || Pred == ICmpInst::ICMP_UGT) && |
962 | "Unexpected isUnsigned predicate!" ); |
963 | |
964 | // Ensure the sub is of the form: |
965 | // (a > b) ? a - b : 0 -> usub.sat(a, b) |
966 | // (a > b) ? b - a : 0 -> -usub.sat(a, b) |
967 | // Checking for both a-b and a+(-b) as a constant. |
968 | bool IsNegative = false; |
969 | const APInt *C; |
970 | if (match(V: TrueVal, P: m_Sub(L: m_Specific(V: B), R: m_Specific(V: A))) || |
971 | (match(V: A, P: m_APInt(Res&: C)) && |
972 | match(V: TrueVal, P: m_Add(L: m_Specific(V: B), R: m_SpecificInt(V: -*C))))) |
973 | IsNegative = true; |
974 | else if (!match(V: TrueVal, P: m_Sub(L: m_Specific(V: A), R: m_Specific(V: B))) && |
975 | !(match(V: B, P: m_APInt(Res&: C)) && |
976 | match(V: TrueVal, P: m_Add(L: m_Specific(V: A), R: m_SpecificInt(V: -*C))))) |
977 | return nullptr; |
978 | |
979 | // If we are adding a negate and the sub and icmp are used anywhere else, we |
980 | // would end up with more instructions. |
981 | if (IsNegative && !TrueVal->hasOneUse() && !ICI->hasOneUse()) |
982 | return nullptr; |
983 | |
984 | // (a > b) ? a - b : 0 -> usub.sat(a, b) |
985 | // (a > b) ? b - a : 0 -> -usub.sat(a, b) |
986 | Value *Result = Builder.CreateBinaryIntrinsic(ID: Intrinsic::usub_sat, LHS: A, RHS: B); |
987 | if (IsNegative) |
988 | Result = Builder.CreateNeg(V: Result); |
989 | return Result; |
990 | } |
991 | |
992 | static Value *canonicalizeSaturatedAdd(ICmpInst *Cmp, Value *TVal, Value *FVal, |
993 | InstCombiner::BuilderTy &Builder) { |
994 | if (!Cmp->hasOneUse()) |
995 | return nullptr; |
996 | |
997 | // Match unsigned saturated add with constant. |
998 | Value *Cmp0 = Cmp->getOperand(i_nocapture: 0); |
999 | Value *Cmp1 = Cmp->getOperand(i_nocapture: 1); |
1000 | ICmpInst::Predicate Pred = Cmp->getPredicate(); |
1001 | Value *X; |
1002 | const APInt *C; |
1003 | |
1004 | // Match unsigned saturated add of 2 variables with an unnecessary 'not'. |
1005 | // There are 8 commuted variants. |
1006 | // Canonicalize -1 (saturated result) to true value of the select. |
1007 | if (match(V: FVal, P: m_AllOnes())) { |
1008 | std::swap(a&: TVal, b&: FVal); |
1009 | Pred = CmpInst::getInversePredicate(pred: Pred); |
1010 | } |
1011 | if (!match(V: TVal, P: m_AllOnes())) |
1012 | return nullptr; |
1013 | |
1014 | // uge -1 is canonicalized to eq -1 and requires special handling |
1015 | // (a == -1) ? -1 : a + 1 -> uadd.sat(a, 1) |
1016 | if (Pred == ICmpInst::ICMP_EQ) { |
1017 | if (match(V: FVal, P: m_Add(L: m_Specific(V: Cmp0), R: m_One())) && |
1018 | match(V: Cmp1, P: m_AllOnes())) { |
1019 | return Builder.CreateBinaryIntrinsic( |
1020 | ID: Intrinsic::uadd_sat, LHS: Cmp0, RHS: ConstantInt::get(Ty: Cmp0->getType(), V: 1)); |
1021 | } |
1022 | return nullptr; |
1023 | } |
1024 | |
1025 | if ((Pred == ICmpInst::ICMP_UGE || Pred == ICmpInst::ICMP_UGT) && |
1026 | match(V: FVal, P: m_Add(L: m_Specific(V: Cmp0), R: m_APIntAllowPoison(Res&: C))) && |
1027 | match(V: Cmp1, P: m_SpecificIntAllowPoison(V: ~*C))) { |
1028 | // (X u> ~C) ? -1 : (X + C) --> uadd.sat(X, C) |
1029 | // (X u>= ~C)? -1 : (X + C) --> uadd.sat(X, C) |
1030 | return Builder.CreateBinaryIntrinsic(ID: Intrinsic::uadd_sat, LHS: Cmp0, |
1031 | RHS: ConstantInt::get(Ty: Cmp0->getType(), V: *C)); |
1032 | } |
1033 | |
1034 | // Negative one does not work here because X u> -1 ? -1, X + -1 is not a |
1035 | // saturated add. |
1036 | if (Pred == ICmpInst::ICMP_UGT && |
1037 | match(V: FVal, P: m_Add(L: m_Specific(V: Cmp0), R: m_APIntAllowPoison(Res&: C))) && |
1038 | match(V: Cmp1, P: m_SpecificIntAllowPoison(V: ~*C - 1)) && !C->isAllOnes()) { |
1039 | // (X u> ~C - 1) ? -1 : (X + C) --> uadd.sat(X, C) |
1040 | return Builder.CreateBinaryIntrinsic(ID: Intrinsic::uadd_sat, LHS: Cmp0, |
1041 | RHS: ConstantInt::get(Ty: Cmp0->getType(), V: *C)); |
1042 | } |
1043 | |
1044 | // Zero does not work here because X u>= 0 ? -1 : X -> is always -1, which is |
1045 | // not a saturated add. |
1046 | if (Pred == ICmpInst::ICMP_UGE && |
1047 | match(V: FVal, P: m_Add(L: m_Specific(V: Cmp0), R: m_APIntAllowPoison(Res&: C))) && |
1048 | match(V: Cmp1, P: m_SpecificIntAllowPoison(V: -*C)) && !C->isZero()) { |
1049 | // (X u >= -C) ? -1 : (X + C) --> uadd.sat(X, C) |
1050 | return Builder.CreateBinaryIntrinsic(ID: Intrinsic::uadd_sat, LHS: Cmp0, |
1051 | RHS: ConstantInt::get(Ty: Cmp0->getType(), V: *C)); |
1052 | } |
1053 | |
1054 | // Canonicalize predicate to less-than or less-or-equal-than. |
1055 | if (Pred == ICmpInst::ICMP_UGT || Pred == ICmpInst::ICMP_UGE) { |
1056 | std::swap(a&: Cmp0, b&: Cmp1); |
1057 | Pred = CmpInst::getSwappedPredicate(pred: Pred); |
1058 | } |
1059 | if (Pred != ICmpInst::ICMP_ULT && Pred != ICmpInst::ICMP_ULE) |
1060 | return nullptr; |
1061 | |
1062 | // Match unsigned saturated add of 2 variables with an unnecessary 'not'. |
1063 | // Strictness of the comparison is irrelevant. |
1064 | Value *Y; |
1065 | if (match(V: Cmp0, P: m_Not(V: m_Value(V&: X))) && |
1066 | match(V: FVal, P: m_c_Add(L: m_Specific(V: X), R: m_Value(V&: Y))) && Y == Cmp1) { |
1067 | // (~X u< Y) ? -1 : (X + Y) --> uadd.sat(X, Y) |
1068 | // (~X u< Y) ? -1 : (Y + X) --> uadd.sat(X, Y) |
1069 | return Builder.CreateBinaryIntrinsic(ID: Intrinsic::uadd_sat, LHS: X, RHS: Y); |
1070 | } |
1071 | // The 'not' op may be included in the sum but not the compare. |
1072 | // Strictness of the comparison is irrelevant. |
1073 | X = Cmp0; |
1074 | Y = Cmp1; |
1075 | if (match(V: FVal, P: m_c_Add(L: m_NotForbidPoison(V: m_Specific(V: X)), R: m_Specific(V: Y)))) { |
1076 | // (X u< Y) ? -1 : (~X + Y) --> uadd.sat(~X, Y) |
1077 | // (X u< Y) ? -1 : (Y + ~X) --> uadd.sat(Y, ~X) |
1078 | BinaryOperator *BO = cast<BinaryOperator>(Val: FVal); |
1079 | return Builder.CreateBinaryIntrinsic( |
1080 | ID: Intrinsic::uadd_sat, LHS: BO->getOperand(i_nocapture: 0), RHS: BO->getOperand(i_nocapture: 1)); |
1081 | } |
1082 | // The overflow may be detected via the add wrapping round. |
1083 | // This is only valid for strict comparison! |
1084 | if (Pred == ICmpInst::ICMP_ULT && |
1085 | match(V: Cmp0, P: m_c_Add(L: m_Specific(V: Cmp1), R: m_Value(V&: Y))) && |
1086 | match(V: FVal, P: m_c_Add(L: m_Specific(V: Cmp1), R: m_Specific(V: Y)))) { |
1087 | // ((X + Y) u< X) ? -1 : (X + Y) --> uadd.sat(X, Y) |
1088 | // ((X + Y) u< Y) ? -1 : (X + Y) --> uadd.sat(X, Y) |
1089 | return Builder.CreateBinaryIntrinsic(ID: Intrinsic::uadd_sat, LHS: Cmp1, RHS: Y); |
1090 | } |
1091 | |
1092 | return nullptr; |
1093 | } |
1094 | |
1095 | /// Try to match patterns with select and subtract as absolute difference. |
1096 | static Value *foldAbsDiff(ICmpInst *Cmp, Value *TVal, Value *FVal, |
1097 | InstCombiner::BuilderTy &Builder) { |
1098 | auto *TI = dyn_cast<Instruction>(Val: TVal); |
1099 | auto *FI = dyn_cast<Instruction>(Val: FVal); |
1100 | if (!TI || !FI) |
1101 | return nullptr; |
1102 | |
1103 | // Normalize predicate to gt/lt rather than ge/le. |
1104 | ICmpInst::Predicate Pred = Cmp->getStrictPredicate(); |
1105 | Value *A = Cmp->getOperand(i_nocapture: 0); |
1106 | Value *B = Cmp->getOperand(i_nocapture: 1); |
1107 | |
1108 | // Normalize "A - B" as the true value of the select. |
1109 | if (match(V: FI, P: m_Sub(L: m_Specific(V: A), R: m_Specific(V: B)))) { |
1110 | std::swap(a&: FI, b&: TI); |
1111 | Pred = ICmpInst::getSwappedPredicate(pred: Pred); |
1112 | } |
1113 | |
1114 | // With any pair of no-wrap subtracts: |
1115 | // (A > B) ? (A - B) : (B - A) --> abs(A - B) |
1116 | if (Pred == CmpInst::ICMP_SGT && |
1117 | match(V: TI, P: m_Sub(L: m_Specific(V: A), R: m_Specific(V: B))) && |
1118 | match(V: FI, P: m_Sub(L: m_Specific(V: B), R: m_Specific(V: A))) && |
1119 | (TI->hasNoSignedWrap() || TI->hasNoUnsignedWrap()) && |
1120 | (FI->hasNoSignedWrap() || FI->hasNoUnsignedWrap())) { |
1121 | // The remaining subtract is not "nuw" any more. |
1122 | // If there's one use of the subtract (no other use than the use we are |
1123 | // about to replace), then we know that the sub is "nsw" in this context |
1124 | // even if it was only "nuw" before. If there's another use, then we can't |
1125 | // add "nsw" to the existing instruction because it may not be safe in the |
1126 | // other user's context. |
1127 | TI->setHasNoUnsignedWrap(false); |
1128 | if (!TI->hasNoSignedWrap()) |
1129 | TI->setHasNoSignedWrap(TI->hasOneUse()); |
1130 | return Builder.CreateBinaryIntrinsic(ID: Intrinsic::abs, LHS: TI, RHS: Builder.getTrue()); |
1131 | } |
1132 | |
1133 | return nullptr; |
1134 | } |
1135 | |
1136 | /// Fold the following code sequence: |
1137 | /// \code |
1138 | /// int a = ctlz(x & -x); |
1139 | // x ? 31 - a : a; |
1140 | // // or |
1141 | // x ? 31 - a : 32; |
1142 | /// \code |
1143 | /// |
1144 | /// into: |
1145 | /// cttz(x) |
1146 | static Instruction *foldSelectCtlzToCttz(ICmpInst *ICI, Value *TrueVal, |
1147 | Value *FalseVal, |
1148 | InstCombiner::BuilderTy &Builder) { |
1149 | unsigned BitWidth = TrueVal->getType()->getScalarSizeInBits(); |
1150 | if (!ICI->isEquality() || !match(V: ICI->getOperand(i_nocapture: 1), P: m_Zero())) |
1151 | return nullptr; |
1152 | |
1153 | if (ICI->getPredicate() == ICmpInst::ICMP_NE) |
1154 | std::swap(a&: TrueVal, b&: FalseVal); |
1155 | |
1156 | Value *Ctlz; |
1157 | if (!match(V: FalseVal, |
1158 | P: m_Xor(L: m_Value(V&: Ctlz), R: m_SpecificInt(V: BitWidth - 1)))) |
1159 | return nullptr; |
1160 | |
1161 | if (!match(V: Ctlz, P: m_Intrinsic<Intrinsic::ctlz>())) |
1162 | return nullptr; |
1163 | |
1164 | if (TrueVal != Ctlz && !match(V: TrueVal, P: m_SpecificInt(V: BitWidth))) |
1165 | return nullptr; |
1166 | |
1167 | Value *X = ICI->getOperand(i_nocapture: 0); |
1168 | auto *II = cast<IntrinsicInst>(Val: Ctlz); |
1169 | if (!match(V: II->getOperand(i_nocapture: 0), P: m_c_And(L: m_Specific(V: X), R: m_Neg(V: m_Specific(V: X))))) |
1170 | return nullptr; |
1171 | |
1172 | Function *F = Intrinsic::getOrInsertDeclaration( |
1173 | M: II->getModule(), id: Intrinsic::cttz, Tys: II->getType()); |
1174 | return CallInst::Create(Func: F, Args: {X, II->getArgOperand(i: 1)}); |
1175 | } |
1176 | |
1177 | /// Attempt to fold a cttz/ctlz followed by a icmp plus select into a single |
1178 | /// call to cttz/ctlz with flag 'is_zero_poison' cleared. |
1179 | /// |
1180 | /// For example, we can fold the following code sequence: |
1181 | /// \code |
1182 | /// %0 = tail call i32 @llvm.cttz.i32(i32 %x, i1 true) |
1183 | /// %1 = icmp ne i32 %x, 0 |
1184 | /// %2 = select i1 %1, i32 %0, i32 32 |
1185 | /// \code |
1186 | /// |
1187 | /// into: |
1188 | /// %0 = tail call i32 @llvm.cttz.i32(i32 %x, i1 false) |
1189 | static Value *foldSelectCttzCtlz(ICmpInst *ICI, Value *TrueVal, Value *FalseVal, |
1190 | InstCombinerImpl &IC) { |
1191 | ICmpInst::Predicate Pred = ICI->getPredicate(); |
1192 | Value *CmpLHS = ICI->getOperand(i_nocapture: 0); |
1193 | Value *CmpRHS = ICI->getOperand(i_nocapture: 1); |
1194 | |
1195 | // Check if the select condition compares a value for equality. |
1196 | if (!ICI->isEquality()) |
1197 | return nullptr; |
1198 | |
1199 | Value *SelectArg = FalseVal; |
1200 | Value *ValueOnZero = TrueVal; |
1201 | if (Pred == ICmpInst::ICMP_NE) |
1202 | std::swap(a&: SelectArg, b&: ValueOnZero); |
1203 | |
1204 | // Skip zero extend/truncate. |
1205 | Value *Count = nullptr; |
1206 | if (!match(V: SelectArg, P: m_ZExt(Op: m_Value(V&: Count))) && |
1207 | !match(V: SelectArg, P: m_Trunc(Op: m_Value(V&: Count)))) |
1208 | Count = SelectArg; |
1209 | |
1210 | // Check that 'Count' is a call to intrinsic cttz/ctlz. Also check that the |
1211 | // input to the cttz/ctlz is used as LHS for the compare instruction. |
1212 | Value *X; |
1213 | if (!match(V: Count, P: m_Intrinsic<Intrinsic::cttz>(Op0: m_Value(V&: X))) && |
1214 | !match(V: Count, P: m_Intrinsic<Intrinsic::ctlz>(Op0: m_Value(V&: X)))) |
1215 | return nullptr; |
1216 | |
1217 | // (X == 0) ? BitWidth : ctz(X) |
1218 | // (X == -1) ? BitWidth : ctz(~X) |
1219 | // (X == Y) ? BitWidth : ctz(X ^ Y) |
1220 | if ((X != CmpLHS || !match(V: CmpRHS, P: m_Zero())) && |
1221 | (!match(V: X, P: m_Not(V: m_Specific(V: CmpLHS))) || !match(V: CmpRHS, P: m_AllOnes())) && |
1222 | !match(V: X, P: m_c_Xor(L: m_Specific(V: CmpLHS), R: m_Specific(V: CmpRHS)))) |
1223 | return nullptr; |
1224 | |
1225 | IntrinsicInst *II = cast<IntrinsicInst>(Val: Count); |
1226 | |
1227 | // Check if the value propagated on zero is a constant number equal to the |
1228 | // sizeof in bits of 'Count'. |
1229 | unsigned SizeOfInBits = Count->getType()->getScalarSizeInBits(); |
1230 | if (match(V: ValueOnZero, P: m_SpecificInt(V: SizeOfInBits))) { |
1231 | // Explicitly clear the 'is_zero_poison' flag. It's always valid to go from |
1232 | // true to false on this flag, so we can replace it for all users. |
1233 | II->setArgOperand(i: 1, v: ConstantInt::getFalse(Context&: II->getContext())); |
1234 | // A range annotation on the intrinsic may no longer be valid. |
1235 | II->dropPoisonGeneratingAnnotations(); |
1236 | IC.addToWorklist(I: II); |
1237 | return SelectArg; |
1238 | } |
1239 | |
1240 | // The ValueOnZero is not the bitwidth. But if the cttz/ctlz (and optional |
1241 | // zext/trunc) have one use (ending at the select), the cttz/ctlz result will |
1242 | // not be used if the input is zero. Relax to 'zero is poison' for that case. |
1243 | if (II->hasOneUse() && SelectArg->hasOneUse() && |
1244 | !match(V: II->getArgOperand(i: 1), P: m_One())) { |
1245 | II->setArgOperand(i: 1, v: ConstantInt::getTrue(Context&: II->getContext())); |
1246 | // noundef attribute on the intrinsic may no longer be valid. |
1247 | II->dropUBImplyingAttrsAndMetadata(); |
1248 | IC.addToWorklist(I: II); |
1249 | } |
1250 | |
1251 | return nullptr; |
1252 | } |
1253 | |
1254 | static Value *canonicalizeSPF(ICmpInst &Cmp, Value *TrueVal, Value *FalseVal, |
1255 | InstCombinerImpl &IC) { |
1256 | Value *LHS, *RHS; |
1257 | // TODO: What to do with pointer min/max patterns? |
1258 | if (!TrueVal->getType()->isIntOrIntVectorTy()) |
1259 | return nullptr; |
1260 | |
1261 | SelectPatternFlavor SPF = |
1262 | matchDecomposedSelectPattern(CmpI: &Cmp, TrueVal, FalseVal, LHS, RHS).Flavor; |
1263 | if (SPF == SelectPatternFlavor::SPF_ABS || |
1264 | SPF == SelectPatternFlavor::SPF_NABS) { |
1265 | if (!Cmp.hasOneUse() && !RHS->hasOneUse()) |
1266 | return nullptr; // TODO: Relax this restriction. |
1267 | |
1268 | // Note that NSW flag can only be propagated for normal, non-negated abs! |
1269 | bool IntMinIsPoison = SPF == SelectPatternFlavor::SPF_ABS && |
1270 | match(V: RHS, P: m_NSWNeg(V: m_Specific(V: LHS))); |
1271 | Constant *IntMinIsPoisonC = |
1272 | ConstantInt::get(Ty: Type::getInt1Ty(C&: Cmp.getContext()), V: IntMinIsPoison); |
1273 | Value *Abs = |
1274 | IC.Builder.CreateBinaryIntrinsic(ID: Intrinsic::abs, LHS, RHS: IntMinIsPoisonC); |
1275 | |
1276 | if (SPF == SelectPatternFlavor::SPF_NABS) |
1277 | return IC.Builder.CreateNeg(V: Abs); // Always without NSW flag! |
1278 | return Abs; |
1279 | } |
1280 | |
1281 | if (SelectPatternResult::isMinOrMax(SPF)) { |
1282 | Intrinsic::ID IntrinsicID = getMinMaxIntrinsic(SPF); |
1283 | return IC.Builder.CreateBinaryIntrinsic(ID: IntrinsicID, LHS, RHS); |
1284 | } |
1285 | |
1286 | return nullptr; |
1287 | } |
1288 | |
1289 | bool InstCombinerImpl::replaceInInstruction(Value *V, Value *Old, Value *New, |
1290 | unsigned Depth) { |
1291 | // Conservatively limit replacement to two instructions upwards. |
1292 | if (Depth == 2) |
1293 | return false; |
1294 | |
1295 | assert(!isa<Constant>(Old) && "Only replace non-constant values" ); |
1296 | |
1297 | auto *I = dyn_cast<Instruction>(Val: V); |
1298 | if (!I || !I->hasOneUse() || |
1299 | !isSafeToSpeculativelyExecuteWithVariableReplaced(I)) |
1300 | return false; |
1301 | |
1302 | // Forbid potentially lane-crossing instructions. |
1303 | if (Old->getType()->isVectorTy() && !isNotCrossLaneOperation(I)) |
1304 | return false; |
1305 | |
1306 | bool Changed = false; |
1307 | for (Use &U : I->operands()) { |
1308 | if (U == Old) { |
1309 | replaceUse(U, NewValue: New); |
1310 | Worklist.add(I); |
1311 | Changed = true; |
1312 | } else { |
1313 | Changed |= replaceInInstruction(V: U, Old, New, Depth: Depth + 1); |
1314 | } |
1315 | } |
1316 | return Changed; |
1317 | } |
1318 | |
1319 | /// If we have a select with an equality comparison, then we know the value in |
1320 | /// one of the arms of the select. See if substituting this value into an arm |
1321 | /// and simplifying the result yields the same value as the other arm. |
1322 | /// |
1323 | /// To make this transform safe, we must drop poison-generating flags |
1324 | /// (nsw, etc) if we simplified to a binop because the select may be guarding |
1325 | /// that poison from propagating. If the existing binop already had no |
1326 | /// poison-generating flags, then this transform can be done by instsimplify. |
1327 | /// |
1328 | /// Consider: |
1329 | /// %cmp = icmp eq i32 %x, 2147483647 |
1330 | /// %add = add nsw i32 %x, 1 |
1331 | /// %sel = select i1 %cmp, i32 -2147483648, i32 %add |
1332 | /// |
1333 | /// We can't replace %sel with %add unless we strip away the flags. |
1334 | /// TODO: Wrapping flags could be preserved in some cases with better analysis. |
1335 | Instruction *InstCombinerImpl::foldSelectValueEquivalence(SelectInst &Sel, |
1336 | CmpInst &Cmp) { |
1337 | // Canonicalize the pattern to an equivalence on the predicate by swapping the |
1338 | // select operands. |
1339 | Value *TrueVal = Sel.getTrueValue(), *FalseVal = Sel.getFalseValue(); |
1340 | bool Swapped = false; |
1341 | if (Cmp.isEquivalence(/*Invert=*/true)) { |
1342 | std::swap(a&: TrueVal, b&: FalseVal); |
1343 | Swapped = true; |
1344 | } else if (!Cmp.isEquivalence()) { |
1345 | return nullptr; |
1346 | } |
1347 | |
1348 | Value *CmpLHS = Cmp.getOperand(i_nocapture: 0), *CmpRHS = Cmp.getOperand(i_nocapture: 1); |
1349 | auto ReplaceOldOpWithNewOp = [&](Value *OldOp, |
1350 | Value *NewOp) -> Instruction * { |
1351 | // In X == Y ? f(X) : Z, try to evaluate f(Y) and replace the operand. |
1352 | // Take care to avoid replacing X == Y ? X : Z with X == Y ? Y : Z, as that |
1353 | // would lead to an infinite replacement cycle. |
1354 | // If we will be able to evaluate f(Y) to a constant, we can allow undef, |
1355 | // otherwise Y cannot be undef as we might pick different values for undef |
1356 | // in the cmp and in f(Y). |
1357 | if (TrueVal == OldOp && (isa<Constant>(Val: OldOp) || !isa<Constant>(Val: NewOp))) |
1358 | return nullptr; |
1359 | |
1360 | if (Value *V = simplifyWithOpReplaced(V: TrueVal, Op: OldOp, RepOp: NewOp, Q: SQ, |
1361 | /* AllowRefinement=*/true)) { |
1362 | // Need some guarantees about the new simplified op to ensure we don't inf |
1363 | // loop. |
1364 | // If we simplify to a constant, replace if we aren't creating new undef. |
1365 | if (match(V, P: m_ImmConstant()) && |
1366 | isGuaranteedNotToBeUndef(V, AC: SQ.AC, CtxI: &Sel, DT: &DT)) |
1367 | return replaceOperand(I&: Sel, OpNum: Swapped ? 2 : 1, V); |
1368 | |
1369 | // If NewOp is a constant and OldOp is not replace iff NewOp doesn't |
1370 | // contain and undef elements. |
1371 | // Make sure that V is always simpler than TrueVal, otherwise we might |
1372 | // end up in an infinite loop. |
1373 | if (match(V: NewOp, P: m_ImmConstant()) || |
1374 | (isa<Instruction>(Val: TrueVal) && |
1375 | is_contained(Range: cast<Instruction>(Val: TrueVal)->operands(), Element: V))) { |
1376 | if (isGuaranteedNotToBeUndef(V: NewOp, AC: SQ.AC, CtxI: &Sel, DT: &DT)) |
1377 | return replaceOperand(I&: Sel, OpNum: Swapped ? 2 : 1, V); |
1378 | return nullptr; |
1379 | } |
1380 | } |
1381 | |
1382 | // Even if TrueVal does not simplify, we can directly replace a use of |
1383 | // CmpLHS with CmpRHS, as long as the instruction is not used anywhere |
1384 | // else and is safe to speculatively execute (we may end up executing it |
1385 | // with different operands, which should not cause side-effects or trigger |
1386 | // undefined behavior). Only do this if CmpRHS is a constant, as |
1387 | // profitability is not clear for other cases. |
1388 | if (OldOp == CmpLHS && match(V: NewOp, P: m_ImmConstant()) && |
1389 | !match(V: OldOp, P: m_Constant()) && |
1390 | isGuaranteedNotToBeUndef(V: NewOp, AC: SQ.AC, CtxI: &Sel, DT: &DT)) |
1391 | if (replaceInInstruction(V: TrueVal, Old: OldOp, New: NewOp)) |
1392 | return &Sel; |
1393 | return nullptr; |
1394 | }; |
1395 | |
1396 | if (Instruction *R = ReplaceOldOpWithNewOp(CmpLHS, CmpRHS)) |
1397 | return R; |
1398 | if (Instruction *R = ReplaceOldOpWithNewOp(CmpRHS, CmpLHS)) |
1399 | return R; |
1400 | |
1401 | auto *FalseInst = dyn_cast<Instruction>(Val: FalseVal); |
1402 | if (!FalseInst) |
1403 | return nullptr; |
1404 | |
1405 | // InstSimplify already performed this fold if it was possible subject to |
1406 | // current poison-generating flags. Check whether dropping poison-generating |
1407 | // flags enables the transform. |
1408 | |
1409 | // Try each equivalence substitution possibility. |
1410 | // We have an 'EQ' comparison, so the select's false value will propagate. |
1411 | // Example: |
1412 | // (X == 42) ? 43 : (X + 1) --> (X == 42) ? (X + 1) : (X + 1) --> X + 1 |
1413 | SmallVector<Instruction *> DropFlags; |
1414 | if (simplifyWithOpReplaced(V: FalseVal, Op: CmpLHS, RepOp: CmpRHS, Q: SQ, |
1415 | /* AllowRefinement */ false, |
1416 | DropFlags: &DropFlags) == TrueVal || |
1417 | simplifyWithOpReplaced(V: FalseVal, Op: CmpRHS, RepOp: CmpLHS, Q: SQ, |
1418 | /* AllowRefinement */ false, |
1419 | DropFlags: &DropFlags) == TrueVal) { |
1420 | for (Instruction *I : DropFlags) { |
1421 | I->dropPoisonGeneratingAnnotations(); |
1422 | Worklist.add(I); |
1423 | } |
1424 | |
1425 | return replaceInstUsesWith(I&: Sel, V: FalseVal); |
1426 | } |
1427 | |
1428 | return nullptr; |
1429 | } |
1430 | |
1431 | /// Fold the following code sequence: |
1432 | /// \code |
1433 | /// %XeqZ = icmp eq i64 %X, %Z |
1434 | /// %YeqZ = icmp eq i64 %Y, %Z |
1435 | /// %XeqY = icmp eq i64 %X, %Y |
1436 | /// %not.YeqZ = xor i1 %YeqZ, true |
1437 | /// %and = select i1 %not.YeqZ, i1 %XeqY, i1 false |
1438 | /// %equal = select i1 %XeqZ, i1 %YeqZ, i1 %and |
1439 | /// \code |
1440 | /// |
1441 | /// into: |
1442 | /// %equal = icmp eq i64 %X, %Y |
1443 | Instruction *InstCombinerImpl::foldSelectEqualityTest(SelectInst &Sel) { |
1444 | Value *X, *Y, *Z; |
1445 | Value *XeqY, *XeqZ = Sel.getCondition(), *YeqZ = Sel.getTrueValue(); |
1446 | |
1447 | if (!match(V: XeqZ, P: m_SpecificICmp(MatchPred: ICmpInst::ICMP_EQ, L: m_Value(V&: X), R: m_Value(V&: Z)))) |
1448 | return nullptr; |
1449 | |
1450 | if (!match(V: YeqZ, |
1451 | P: m_c_SpecificICmp(MatchPred: ICmpInst::ICMP_EQ, L: m_Value(V&: Y), R: m_Specific(V: Z)))) |
1452 | std::swap(a&: X, b&: Z); |
1453 | |
1454 | if (!match(V: YeqZ, |
1455 | P: m_c_SpecificICmp(MatchPred: ICmpInst::ICMP_EQ, L: m_Value(V&: Y), R: m_Specific(V: Z)))) |
1456 | return nullptr; |
1457 | |
1458 | if (!match(V: Sel.getFalseValue(), |
1459 | P: m_c_LogicalAnd(L: m_Not(V: m_Specific(V: YeqZ)), R: m_Value(V&: XeqY)))) |
1460 | return nullptr; |
1461 | |
1462 | if (!match(V: XeqY, |
1463 | P: m_c_SpecificICmp(MatchPred: ICmpInst::ICMP_EQ, L: m_Specific(V: X), R: m_Specific(V: Y)))) |
1464 | return nullptr; |
1465 | |
1466 | cast<ICmpInst>(Val: XeqY)->setSameSign(false); |
1467 | return replaceInstUsesWith(I&: Sel, V: XeqY); |
1468 | } |
1469 | |
1470 | // See if this is a pattern like: |
1471 | // %old_cmp1 = icmp slt i32 %x, C2 |
1472 | // %old_replacement = select i1 %old_cmp1, i32 %target_low, i32 %target_high |
1473 | // %old_x_offseted = add i32 %x, C1 |
1474 | // %old_cmp0 = icmp ult i32 %old_x_offseted, C0 |
1475 | // %r = select i1 %old_cmp0, i32 %x, i32 %old_replacement |
1476 | // This can be rewritten as more canonical pattern: |
1477 | // %new_cmp1 = icmp slt i32 %x, -C1 |
1478 | // %new_cmp2 = icmp sge i32 %x, C0-C1 |
1479 | // %new_clamped_low = select i1 %new_cmp1, i32 %target_low, i32 %x |
1480 | // %r = select i1 %new_cmp2, i32 %target_high, i32 %new_clamped_low |
1481 | // Iff -C1 s<= C2 s<= C0-C1 |
1482 | // Also ULT predicate can also be UGT iff C0 != -1 (+invert result) |
1483 | // SLT predicate can also be SGT iff C2 != INT_MAX (+invert res.) |
1484 | static Value *canonicalizeClampLike(SelectInst &Sel0, ICmpInst &Cmp0, |
1485 | InstCombiner::BuilderTy &Builder, |
1486 | InstCombiner &IC) { |
1487 | Value *X = Sel0.getTrueValue(); |
1488 | Value *Sel1 = Sel0.getFalseValue(); |
1489 | |
1490 | // First match the condition of the outermost select. |
1491 | // Said condition must be one-use. |
1492 | if (!Cmp0.hasOneUse()) |
1493 | return nullptr; |
1494 | ICmpInst::Predicate Pred0 = Cmp0.getPredicate(); |
1495 | Value *Cmp00 = Cmp0.getOperand(i_nocapture: 0); |
1496 | Constant *C0; |
1497 | if (!match(V: Cmp0.getOperand(i_nocapture: 1), |
1498 | P: m_CombineAnd(L: m_AnyIntegralConstant(), R: m_Constant(C&: C0)))) |
1499 | return nullptr; |
1500 | |
1501 | if (!isa<SelectInst>(Val: Sel1)) { |
1502 | Pred0 = ICmpInst::getInversePredicate(pred: Pred0); |
1503 | std::swap(a&: X, b&: Sel1); |
1504 | } |
1505 | |
1506 | // Canonicalize Cmp0 into ult or uge. |
1507 | // FIXME: we shouldn't care about lanes that are 'undef' in the end? |
1508 | switch (Pred0) { |
1509 | case ICmpInst::Predicate::ICMP_ULT: |
1510 | case ICmpInst::Predicate::ICMP_UGE: |
1511 | // Although icmp ult %x, 0 is an unusual thing to try and should generally |
1512 | // have been simplified, it does not verify with undef inputs so ensure we |
1513 | // are not in a strange state. |
1514 | if (!match(V: C0, P: m_SpecificInt_ICMP( |
1515 | Predicate: ICmpInst::Predicate::ICMP_NE, |
1516 | Threshold: APInt::getZero(numBits: C0->getType()->getScalarSizeInBits())))) |
1517 | return nullptr; |
1518 | break; // Great! |
1519 | case ICmpInst::Predicate::ICMP_ULE: |
1520 | case ICmpInst::Predicate::ICMP_UGT: |
1521 | // We want to canonicalize it to 'ult' or 'uge', so we'll need to increment |
1522 | // C0, which again means it must not have any all-ones elements. |
1523 | if (!match(V: C0, |
1524 | P: m_SpecificInt_ICMP( |
1525 | Predicate: ICmpInst::Predicate::ICMP_NE, |
1526 | Threshold: APInt::getAllOnes(numBits: C0->getType()->getScalarSizeInBits())))) |
1527 | return nullptr; // Can't do, have all-ones element[s]. |
1528 | Pred0 = ICmpInst::getFlippedStrictnessPredicate(pred: Pred0); |
1529 | C0 = InstCombiner::AddOne(C: C0); |
1530 | break; |
1531 | default: |
1532 | return nullptr; // Unknown predicate. |
1533 | } |
1534 | |
1535 | // Now that we've canonicalized the ICmp, we know the X we expect; |
1536 | // the select in other hand should be one-use. |
1537 | if (!Sel1->hasOneUse()) |
1538 | return nullptr; |
1539 | |
1540 | // If the types do not match, look through any truncs to the underlying |
1541 | // instruction. |
1542 | if (Cmp00->getType() != X->getType() && X->hasOneUse()) |
1543 | match(V: X, P: m_TruncOrSelf(Op: m_Value(V&: X))); |
1544 | |
1545 | // We now can finish matching the condition of the outermost select: |
1546 | // it should either be the X itself, or an addition of some constant to X. |
1547 | Constant *C1; |
1548 | if (Cmp00 == X) |
1549 | C1 = ConstantInt::getNullValue(Ty: X->getType()); |
1550 | else if (!match(V: Cmp00, |
1551 | P: m_Add(L: m_Specific(V: X), |
1552 | R: m_CombineAnd(L: m_AnyIntegralConstant(), R: m_Constant(C&: C1))))) |
1553 | return nullptr; |
1554 | |
1555 | Value *Cmp1; |
1556 | CmpPredicate Pred1; |
1557 | Constant *C2; |
1558 | Value *ReplacementLow, *ReplacementHigh; |
1559 | if (!match(V: Sel1, P: m_Select(C: m_Value(V&: Cmp1), L: m_Value(V&: ReplacementLow), |
1560 | R: m_Value(V&: ReplacementHigh))) || |
1561 | !match(V: Cmp1, |
1562 | P: m_ICmp(Pred&: Pred1, L: m_Specific(V: X), |
1563 | R: m_CombineAnd(L: m_AnyIntegralConstant(), R: m_Constant(C&: C2))))) |
1564 | return nullptr; |
1565 | |
1566 | if (!Cmp1->hasOneUse() && (Cmp00 == X || !Cmp00->hasOneUse())) |
1567 | return nullptr; // Not enough one-use instructions for the fold. |
1568 | // FIXME: this restriction could be relaxed if Cmp1 can be reused as one of |
1569 | // two comparisons we'll need to build. |
1570 | |
1571 | // Canonicalize Cmp1 into the form we expect. |
1572 | // FIXME: we shouldn't care about lanes that are 'undef' in the end? |
1573 | switch (Pred1) { |
1574 | case ICmpInst::Predicate::ICMP_SLT: |
1575 | break; |
1576 | case ICmpInst::Predicate::ICMP_SLE: |
1577 | // We'd have to increment C2 by one, and for that it must not have signed |
1578 | // max element, but then it would have been canonicalized to 'slt' before |
1579 | // we get here. So we can't do anything useful with 'sle'. |
1580 | return nullptr; |
1581 | case ICmpInst::Predicate::ICMP_SGT: |
1582 | // We want to canonicalize it to 'slt', so we'll need to increment C2, |
1583 | // which again means it must not have any signed max elements. |
1584 | if (!match(V: C2, |
1585 | P: m_SpecificInt_ICMP(Predicate: ICmpInst::Predicate::ICMP_NE, |
1586 | Threshold: APInt::getSignedMaxValue( |
1587 | numBits: C2->getType()->getScalarSizeInBits())))) |
1588 | return nullptr; // Can't do, have signed max element[s]. |
1589 | C2 = InstCombiner::AddOne(C: C2); |
1590 | [[fallthrough]]; |
1591 | case ICmpInst::Predicate::ICMP_SGE: |
1592 | // Also non-canonical, but here we don't need to change C2, |
1593 | // so we don't have any restrictions on C2, so we can just handle it. |
1594 | Pred1 = ICmpInst::Predicate::ICMP_SLT; |
1595 | std::swap(a&: ReplacementLow, b&: ReplacementHigh); |
1596 | break; |
1597 | default: |
1598 | return nullptr; // Unknown predicate. |
1599 | } |
1600 | assert(Pred1 == ICmpInst::Predicate::ICMP_SLT && |
1601 | "Unexpected predicate type." ); |
1602 | |
1603 | // The thresholds of this clamp-like pattern. |
1604 | auto *ThresholdLowIncl = ConstantExpr::getNeg(C: C1); |
1605 | auto *ThresholdHighExcl = ConstantExpr::getSub(C1: C0, C2: C1); |
1606 | |
1607 | assert((Pred0 == ICmpInst::Predicate::ICMP_ULT || |
1608 | Pred0 == ICmpInst::Predicate::ICMP_UGE) && |
1609 | "Unexpected predicate type." ); |
1610 | if (Pred0 == ICmpInst::Predicate::ICMP_UGE) |
1611 | std::swap(a&: ThresholdLowIncl, b&: ThresholdHighExcl); |
1612 | |
1613 | // The fold has a precondition 1: C2 s>= ThresholdLow |
1614 | auto *Precond1 = ConstantFoldCompareInstOperands( |
1615 | Predicate: ICmpInst::Predicate::ICMP_SGE, LHS: C2, RHS: ThresholdLowIncl, DL: IC.getDataLayout()); |
1616 | if (!Precond1 || !match(V: Precond1, P: m_One())) |
1617 | return nullptr; |
1618 | // The fold has a precondition 2: C2 s<= ThresholdHigh |
1619 | auto *Precond2 = ConstantFoldCompareInstOperands( |
1620 | Predicate: ICmpInst::Predicate::ICMP_SLE, LHS: C2, RHS: ThresholdHighExcl, DL: IC.getDataLayout()); |
1621 | if (!Precond2 || !match(V: Precond2, P: m_One())) |
1622 | return nullptr; |
1623 | |
1624 | // If we are matching from a truncated input, we need to sext the |
1625 | // ReplacementLow and ReplacementHigh values. Only do the transform if they |
1626 | // are free to extend due to being constants. |
1627 | if (X->getType() != Sel0.getType()) { |
1628 | Constant *LowC, *HighC; |
1629 | if (!match(V: ReplacementLow, P: m_ImmConstant(C&: LowC)) || |
1630 | !match(V: ReplacementHigh, P: m_ImmConstant(C&: HighC))) |
1631 | return nullptr; |
1632 | const DataLayout &DL = Sel0.getDataLayout(); |
1633 | ReplacementLow = |
1634 | ConstantFoldCastOperand(Opcode: Instruction::SExt, C: LowC, DestTy: X->getType(), DL); |
1635 | ReplacementHigh = |
1636 | ConstantFoldCastOperand(Opcode: Instruction::SExt, C: HighC, DestTy: X->getType(), DL); |
1637 | assert(ReplacementLow && ReplacementHigh && |
1638 | "Constant folding of ImmConstant cannot fail" ); |
1639 | } |
1640 | |
1641 | // All good, finally emit the new pattern. |
1642 | Value *ShouldReplaceLow = Builder.CreateICmpSLT(LHS: X, RHS: ThresholdLowIncl); |
1643 | Value *ShouldReplaceHigh = Builder.CreateICmpSGE(LHS: X, RHS: ThresholdHighExcl); |
1644 | Value *MaybeReplacedLow = |
1645 | Builder.CreateSelect(C: ShouldReplaceLow, True: ReplacementLow, False: X); |
1646 | |
1647 | // Create the final select. If we looked through a truncate above, we will |
1648 | // need to retruncate the result. |
1649 | Value *MaybeReplacedHigh = Builder.CreateSelect( |
1650 | C: ShouldReplaceHigh, True: ReplacementHigh, False: MaybeReplacedLow); |
1651 | return Builder.CreateTrunc(V: MaybeReplacedHigh, DestTy: Sel0.getType()); |
1652 | } |
1653 | |
1654 | // If we have |
1655 | // %cmp = icmp [canonical predicate] i32 %x, C0 |
1656 | // %r = select i1 %cmp, i32 %y, i32 C1 |
1657 | // Where C0 != C1 and %x may be different from %y, see if the constant that we |
1658 | // will have if we flip the strictness of the predicate (i.e. without changing |
1659 | // the result) is identical to the C1 in select. If it matches we can change |
1660 | // original comparison to one with swapped predicate, reuse the constant, |
1661 | // and swap the hands of select. |
1662 | static Instruction * |
1663 | tryToReuseConstantFromSelectInComparison(SelectInst &Sel, ICmpInst &Cmp, |
1664 | InstCombinerImpl &IC) { |
1665 | CmpPredicate Pred; |
1666 | Value *X; |
1667 | Constant *C0; |
1668 | if (!match(V: &Cmp, P: m_OneUse(SubPattern: m_ICmp( |
1669 | Pred, L: m_Value(V&: X), |
1670 | R: m_CombineAnd(L: m_AnyIntegralConstant(), R: m_Constant(C&: C0)))))) |
1671 | return nullptr; |
1672 | |
1673 | // If comparison predicate is non-relational, we won't be able to do anything. |
1674 | if (ICmpInst::isEquality(P: Pred)) |
1675 | return nullptr; |
1676 | |
1677 | // If comparison predicate is non-canonical, then we certainly won't be able |
1678 | // to make it canonical; canonicalizeCmpWithConstant() already tried. |
1679 | if (!InstCombiner::isCanonicalPredicate(Pred)) |
1680 | return nullptr; |
1681 | |
1682 | // If the [input] type of comparison and select type are different, lets abort |
1683 | // for now. We could try to compare constants with trunc/[zs]ext though. |
1684 | if (C0->getType() != Sel.getType()) |
1685 | return nullptr; |
1686 | |
1687 | // ULT with 'add' of a constant is canonical. See foldICmpAddConstant(). |
1688 | // FIXME: Are there more magic icmp predicate+constant pairs we must avoid? |
1689 | // Or should we just abandon this transform entirely? |
1690 | if (Pred == CmpInst::ICMP_ULT && match(V: X, P: m_Add(L: m_Value(), R: m_Constant()))) |
1691 | return nullptr; |
1692 | |
1693 | |
1694 | Value *SelVal0, *SelVal1; // We do not care which one is from where. |
1695 | match(V: &Sel, P: m_Select(C: m_Value(), L: m_Value(V&: SelVal0), R: m_Value(V&: SelVal1))); |
1696 | // At least one of these values we are selecting between must be a constant |
1697 | // else we'll never succeed. |
1698 | if (!match(V: SelVal0, P: m_AnyIntegralConstant()) && |
1699 | !match(V: SelVal1, P: m_AnyIntegralConstant())) |
1700 | return nullptr; |
1701 | |
1702 | // Does this constant C match any of the `select` values? |
1703 | auto MatchesSelectValue = [SelVal0, SelVal1](Constant *C) { |
1704 | return C->isElementWiseEqual(Y: SelVal0) || C->isElementWiseEqual(Y: SelVal1); |
1705 | }; |
1706 | |
1707 | // If C0 *already* matches true/false value of select, we are done. |
1708 | if (MatchesSelectValue(C0)) |
1709 | return nullptr; |
1710 | |
1711 | // Check the constant we'd have with flipped-strictness predicate. |
1712 | auto FlippedStrictness = getFlippedStrictnessPredicateAndConstant(Pred, C: C0); |
1713 | if (!FlippedStrictness) |
1714 | return nullptr; |
1715 | |
1716 | // If said constant doesn't match either, then there is no hope, |
1717 | if (!MatchesSelectValue(FlippedStrictness->second)) |
1718 | return nullptr; |
1719 | |
1720 | // It matched! Lets insert the new comparison just before select. |
1721 | InstCombiner::BuilderTy::InsertPointGuard Guard(IC.Builder); |
1722 | IC.Builder.SetInsertPoint(&Sel); |
1723 | |
1724 | Pred = ICmpInst::getSwappedPredicate(pred: Pred); // Yes, swapped. |
1725 | Value *NewCmp = IC.Builder.CreateICmp(P: Pred, LHS: X, RHS: FlippedStrictness->second, |
1726 | Name: Cmp.getName() + ".inv" ); |
1727 | IC.replaceOperand(I&: Sel, OpNum: 0, V: NewCmp); |
1728 | Sel.swapValues(); |
1729 | Sel.swapProfMetadata(); |
1730 | |
1731 | return &Sel; |
1732 | } |
1733 | |
1734 | static Instruction *foldSelectZeroOrOnes(ICmpInst *Cmp, Value *TVal, |
1735 | Value *FVal, |
1736 | InstCombiner::BuilderTy &Builder) { |
1737 | if (!Cmp->hasOneUse()) |
1738 | return nullptr; |
1739 | |
1740 | const APInt *CmpC; |
1741 | if (!match(V: Cmp->getOperand(i_nocapture: 1), P: m_APIntAllowPoison(Res&: CmpC))) |
1742 | return nullptr; |
1743 | |
1744 | // (X u< 2) ? -X : -1 --> sext (X != 0) |
1745 | Value *X = Cmp->getOperand(i_nocapture: 0); |
1746 | if (Cmp->getPredicate() == ICmpInst::ICMP_ULT && *CmpC == 2 && |
1747 | match(V: TVal, P: m_Neg(V: m_Specific(V: X))) && match(V: FVal, P: m_AllOnes())) |
1748 | return new SExtInst(Builder.CreateIsNotNull(Arg: X), TVal->getType()); |
1749 | |
1750 | // (X u> 1) ? -1 : -X --> sext (X != 0) |
1751 | if (Cmp->getPredicate() == ICmpInst::ICMP_UGT && *CmpC == 1 && |
1752 | match(V: FVal, P: m_Neg(V: m_Specific(V: X))) && match(V: TVal, P: m_AllOnes())) |
1753 | return new SExtInst(Builder.CreateIsNotNull(Arg: X), TVal->getType()); |
1754 | |
1755 | return nullptr; |
1756 | } |
1757 | |
1758 | static Value *foldSelectInstWithICmpConst(SelectInst &SI, ICmpInst *ICI, |
1759 | InstCombiner::BuilderTy &Builder) { |
1760 | const APInt *CmpC; |
1761 | Value *V; |
1762 | CmpPredicate Pred; |
1763 | if (!match(V: ICI, P: m_ICmp(Pred, L: m_Value(V), R: m_APInt(Res&: CmpC)))) |
1764 | return nullptr; |
1765 | |
1766 | // Match clamp away from min/max value as a max/min operation. |
1767 | Value *TVal = SI.getTrueValue(); |
1768 | Value *FVal = SI.getFalseValue(); |
1769 | if (Pred == ICmpInst::ICMP_EQ && V == FVal) { |
1770 | // (V == UMIN) ? UMIN+1 : V --> umax(V, UMIN+1) |
1771 | if (CmpC->isMinValue() && match(V: TVal, P: m_SpecificInt(V: *CmpC + 1))) |
1772 | return Builder.CreateBinaryIntrinsic(ID: Intrinsic::umax, LHS: V, RHS: TVal); |
1773 | // (V == UMAX) ? UMAX-1 : V --> umin(V, UMAX-1) |
1774 | if (CmpC->isMaxValue() && match(V: TVal, P: m_SpecificInt(V: *CmpC - 1))) |
1775 | return Builder.CreateBinaryIntrinsic(ID: Intrinsic::umin, LHS: V, RHS: TVal); |
1776 | // (V == SMIN) ? SMIN+1 : V --> smax(V, SMIN+1) |
1777 | if (CmpC->isMinSignedValue() && match(V: TVal, P: m_SpecificInt(V: *CmpC + 1))) |
1778 | return Builder.CreateBinaryIntrinsic(ID: Intrinsic::smax, LHS: V, RHS: TVal); |
1779 | // (V == SMAX) ? SMAX-1 : V --> smin(V, SMAX-1) |
1780 | if (CmpC->isMaxSignedValue() && match(V: TVal, P: m_SpecificInt(V: *CmpC - 1))) |
1781 | return Builder.CreateBinaryIntrinsic(ID: Intrinsic::smin, LHS: V, RHS: TVal); |
1782 | } |
1783 | |
1784 | // Fold icmp(X) ? f(X) : C to f(X) when f(X) is guaranteed to be equal to C |
1785 | // for all X in the exact range of the inverse predicate. |
1786 | Instruction *Op; |
1787 | const APInt *C; |
1788 | CmpInst::Predicate CPred; |
1789 | if (match(V: &SI, P: m_Select(C: m_Specific(V: ICI), L: m_APInt(Res&: C), R: m_Instruction(I&: Op)))) |
1790 | CPred = ICI->getPredicate(); |
1791 | else if (match(V: &SI, P: m_Select(C: m_Specific(V: ICI), L: m_Instruction(I&: Op), R: m_APInt(Res&: C)))) |
1792 | CPred = ICI->getInversePredicate(); |
1793 | else |
1794 | return nullptr; |
1795 | |
1796 | ConstantRange InvDomCR = ConstantRange::makeExactICmpRegion(Pred: CPred, Other: *CmpC); |
1797 | const APInt *OpC; |
1798 | if (match(V: Op, P: m_BinOp(L: m_Specific(V), R: m_APInt(Res&: OpC)))) { |
1799 | ConstantRange R = InvDomCR.binaryOp( |
1800 | BinOp: static_cast<Instruction::BinaryOps>(Op->getOpcode()), Other: *OpC); |
1801 | if (R == *C) { |
1802 | Op->dropPoisonGeneratingFlags(); |
1803 | return Op; |
1804 | } |
1805 | } |
1806 | if (auto *MMI = dyn_cast<MinMaxIntrinsic>(Val: Op); |
1807 | MMI && MMI->getLHS() == V && match(V: MMI->getRHS(), P: m_APInt(Res&: OpC))) { |
1808 | ConstantRange R = ConstantRange::intrinsic(IntrinsicID: MMI->getIntrinsicID(), |
1809 | Ops: {InvDomCR, ConstantRange(*OpC)}); |
1810 | if (R == *C) { |
1811 | MMI->dropPoisonGeneratingAnnotations(); |
1812 | return MMI; |
1813 | } |
1814 | } |
1815 | |
1816 | return nullptr; |
1817 | } |
1818 | |
1819 | /// `A == MIN_INT ? B != MIN_INT : A < B` --> `A < B` |
1820 | /// `A == MAX_INT ? B != MAX_INT : A > B` --> `A > B` |
1821 | static Instruction *foldSelectWithExtremeEqCond(Value *CmpLHS, Value *CmpRHS, |
1822 | Value *TrueVal, |
1823 | Value *FalseVal) { |
1824 | Type *Ty = CmpLHS->getType(); |
1825 | |
1826 | if (Ty->isPtrOrPtrVectorTy()) |
1827 | return nullptr; |
1828 | |
1829 | CmpPredicate Pred; |
1830 | Value *B; |
1831 | |
1832 | if (!match(V: FalseVal, P: m_c_ICmp(Pred, L: m_Specific(V: CmpLHS), R: m_Value(V&: B)))) |
1833 | return nullptr; |
1834 | |
1835 | Value *TValRHS; |
1836 | if (!match(V: TrueVal, P: m_SpecificICmp(MatchPred: ICmpInst::ICMP_NE, L: m_Specific(V: B), |
1837 | R: m_Value(V&: TValRHS)))) |
1838 | return nullptr; |
1839 | |
1840 | APInt C; |
1841 | unsigned BitWidth = Ty->getScalarSizeInBits(); |
1842 | |
1843 | if (ICmpInst::isLT(P: Pred)) { |
1844 | C = CmpInst::isSigned(predicate: Pred) ? APInt::getSignedMinValue(numBits: BitWidth) |
1845 | : APInt::getMinValue(numBits: BitWidth); |
1846 | } else if (ICmpInst::isGT(P: Pred)) { |
1847 | C = CmpInst::isSigned(predicate: Pred) ? APInt::getSignedMaxValue(numBits: BitWidth) |
1848 | : APInt::getMaxValue(numBits: BitWidth); |
1849 | } else { |
1850 | return nullptr; |
1851 | } |
1852 | |
1853 | if (!match(V: CmpRHS, P: m_SpecificInt(V: C)) || !match(V: TValRHS, P: m_SpecificInt(V: C))) |
1854 | return nullptr; |
1855 | |
1856 | return new ICmpInst(Pred, CmpLHS, B); |
1857 | } |
1858 | |
1859 | static Instruction *foldSelectICmpEq(SelectInst &SI, ICmpInst *ICI, |
1860 | InstCombinerImpl &IC) { |
1861 | ICmpInst::Predicate Pred = ICI->getPredicate(); |
1862 | if (!ICmpInst::isEquality(P: Pred)) |
1863 | return nullptr; |
1864 | |
1865 | Value *TrueVal = SI.getTrueValue(); |
1866 | Value *FalseVal = SI.getFalseValue(); |
1867 | Value *CmpLHS = ICI->getOperand(i_nocapture: 0); |
1868 | Value *CmpRHS = ICI->getOperand(i_nocapture: 1); |
1869 | |
1870 | if (Pred == ICmpInst::ICMP_NE) |
1871 | std::swap(a&: TrueVal, b&: FalseVal); |
1872 | |
1873 | if (Instruction *Res = |
1874 | foldSelectWithExtremeEqCond(CmpLHS, CmpRHS, TrueVal, FalseVal)) |
1875 | return Res; |
1876 | |
1877 | return nullptr; |
1878 | } |
1879 | |
1880 | /// Fold `X Pred C1 ? X BOp C2 : C1 BOp C2` to `min/max(X, C1) BOp C2`. |
1881 | /// This allows for better canonicalization. |
1882 | Value *InstCombinerImpl::foldSelectWithConstOpToBinOp(ICmpInst *Cmp, |
1883 | Value *TrueVal, |
1884 | Value *FalseVal) { |
1885 | Constant *C1, *C2, *C3; |
1886 | Value *X; |
1887 | CmpPredicate Predicate; |
1888 | |
1889 | if (!match(V: Cmp, P: m_ICmp(Pred&: Predicate, L: m_Value(V&: X), R: m_Constant(C&: C1)))) |
1890 | return nullptr; |
1891 | |
1892 | if (!ICmpInst::isRelational(P: Predicate)) |
1893 | return nullptr; |
1894 | |
1895 | if (match(V: TrueVal, P: m_Constant())) { |
1896 | std::swap(a&: FalseVal, b&: TrueVal); |
1897 | Predicate = ICmpInst::getInversePredicate(pred: Predicate); |
1898 | } |
1899 | |
1900 | if (!match(V: FalseVal, P: m_Constant(C&: C3)) || !TrueVal->hasOneUse()) |
1901 | return nullptr; |
1902 | |
1903 | bool IsIntrinsic; |
1904 | unsigned Opcode; |
1905 | if (BinaryOperator *BOp = dyn_cast<BinaryOperator>(Val: TrueVal)) { |
1906 | Opcode = BOp->getOpcode(); |
1907 | IsIntrinsic = false; |
1908 | |
1909 | // This fold causes some regressions and is primarily intended for |
1910 | // add and sub. So we early exit for div and rem to minimize the |
1911 | // regressions. |
1912 | if (Instruction::isIntDivRem(Opcode)) |
1913 | return nullptr; |
1914 | |
1915 | if (!match(V: BOp, P: m_BinOp(L: m_Specific(V: X), R: m_Constant(C&: C2)))) |
1916 | return nullptr; |
1917 | |
1918 | } else if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(Val: TrueVal)) { |
1919 | if (!match(V: II, P: m_MaxOrMin(L: m_Specific(V: X), R: m_Constant(C&: C2)))) |
1920 | return nullptr; |
1921 | Opcode = II->getIntrinsicID(); |
1922 | IsIntrinsic = true; |
1923 | } else { |
1924 | return nullptr; |
1925 | } |
1926 | |
1927 | Value *RHS; |
1928 | SelectPatternFlavor SPF; |
1929 | const DataLayout &DL = Cmp->getDataLayout(); |
1930 | auto Flipped = getFlippedStrictnessPredicateAndConstant(Pred: Predicate, C: C1); |
1931 | |
1932 | auto FoldBinaryOpOrIntrinsic = [&](Constant *LHS, Constant *RHS) { |
1933 | return IsIntrinsic ? ConstantFoldBinaryIntrinsic(ID: Opcode, LHS, RHS, |
1934 | Ty: LHS->getType(), FMFSource: nullptr) |
1935 | : ConstantFoldBinaryOpOperands(Opcode, LHS, RHS, DL); |
1936 | }; |
1937 | |
1938 | if (C3 == FoldBinaryOpOrIntrinsic(C1, C2)) { |
1939 | SPF = getSelectPattern(Pred: Predicate).Flavor; |
1940 | RHS = C1; |
1941 | } else if (Flipped && C3 == FoldBinaryOpOrIntrinsic(Flipped->second, C2)) { |
1942 | SPF = getSelectPattern(Pred: Flipped->first).Flavor; |
1943 | RHS = Flipped->second; |
1944 | } else { |
1945 | return nullptr; |
1946 | } |
1947 | |
1948 | Intrinsic::ID MinMaxID = getMinMaxIntrinsic(SPF); |
1949 | Value *MinMax = Builder.CreateBinaryIntrinsic(ID: MinMaxID, LHS: X, RHS); |
1950 | if (IsIntrinsic) |
1951 | return Builder.CreateBinaryIntrinsic(ID: Opcode, LHS: MinMax, RHS: C2); |
1952 | |
1953 | const auto BinOpc = Instruction::BinaryOps(Opcode); |
1954 | Value *BinOp = Builder.CreateBinOp(Opc: BinOpc, LHS: MinMax, RHS: C2); |
1955 | |
1956 | // If we can attach no-wrap flags to the new instruction, do so if the |
1957 | // old instruction had them and C1 BinOp C2 does not overflow. |
1958 | if (Instruction *BinOpInst = dyn_cast<Instruction>(Val: BinOp)) { |
1959 | if (BinOpc == Instruction::Add || BinOpc == Instruction::Sub || |
1960 | BinOpc == Instruction::Mul) { |
1961 | Instruction *OldBinOp = cast<BinaryOperator>(Val: TrueVal); |
1962 | if (OldBinOp->hasNoSignedWrap() && |
1963 | willNotOverflow(Opcode: BinOpc, LHS: RHS, RHS: C2, CxtI: *BinOpInst, /*IsSigned=*/true)) |
1964 | BinOpInst->setHasNoSignedWrap(); |
1965 | if (OldBinOp->hasNoUnsignedWrap() && |
1966 | willNotOverflow(Opcode: BinOpc, LHS: RHS, RHS: C2, CxtI: *BinOpInst, /*IsSigned=*/false)) |
1967 | BinOpInst->setHasNoUnsignedWrap(); |
1968 | } |
1969 | } |
1970 | return BinOp; |
1971 | } |
1972 | |
1973 | /// Visit a SelectInst that has an ICmpInst as its first operand. |
1974 | Instruction *InstCombinerImpl::foldSelectInstWithICmp(SelectInst &SI, |
1975 | ICmpInst *ICI) { |
1976 | if (Value *V = |
1977 | canonicalizeSPF(Cmp&: *ICI, TrueVal: SI.getTrueValue(), FalseVal: SI.getFalseValue(), IC&: *this)) |
1978 | return replaceInstUsesWith(I&: SI, V); |
1979 | |
1980 | if (Value *V = foldSelectInstWithICmpConst(SI, ICI, Builder)) |
1981 | return replaceInstUsesWith(I&: SI, V); |
1982 | |
1983 | if (Value *V = canonicalizeClampLike(Sel0&: SI, Cmp0&: *ICI, Builder, IC&: *this)) |
1984 | return replaceInstUsesWith(I&: SI, V); |
1985 | |
1986 | if (Instruction *NewSel = |
1987 | tryToReuseConstantFromSelectInComparison(Sel&: SI, Cmp&: *ICI, IC&: *this)) |
1988 | return NewSel; |
1989 | |
1990 | // NOTE: if we wanted to, this is where to detect integer MIN/MAX |
1991 | bool Changed = false; |
1992 | Value *TrueVal = SI.getTrueValue(); |
1993 | Value *FalseVal = SI.getFalseValue(); |
1994 | ICmpInst::Predicate Pred = ICI->getPredicate(); |
1995 | Value *CmpLHS = ICI->getOperand(i_nocapture: 0); |
1996 | Value *CmpRHS = ICI->getOperand(i_nocapture: 1); |
1997 | |
1998 | if (Instruction *NewSel = foldSelectICmpEq(SI, ICI, IC&: *this)) |
1999 | return NewSel; |
2000 | |
2001 | // Canonicalize a signbit condition to use zero constant by swapping: |
2002 | // (CmpLHS > -1) ? TV : FV --> (CmpLHS < 0) ? FV : TV |
2003 | // To avoid conflicts (infinite loops) with other canonicalizations, this is |
2004 | // not applied with any constant select arm. |
2005 | if (Pred == ICmpInst::ICMP_SGT && match(V: CmpRHS, P: m_AllOnes()) && |
2006 | !match(V: TrueVal, P: m_Constant()) && !match(V: FalseVal, P: m_Constant()) && |
2007 | ICI->hasOneUse()) { |
2008 | InstCombiner::BuilderTy::InsertPointGuard Guard(Builder); |
2009 | Builder.SetInsertPoint(&SI); |
2010 | Value *IsNeg = Builder.CreateIsNeg(Arg: CmpLHS, Name: ICI->getName()); |
2011 | replaceOperand(I&: SI, OpNum: 0, V: IsNeg); |
2012 | SI.swapValues(); |
2013 | SI.swapProfMetadata(); |
2014 | return &SI; |
2015 | } |
2016 | |
2017 | if (Value *V = foldSelectICmpMinMax(Cmp: ICI, TVal: TrueVal, FVal: FalseVal, Builder, SQ)) |
2018 | return replaceInstUsesWith(I&: SI, V); |
2019 | |
2020 | if (Instruction *V = |
2021 | foldSelectICmpAndAnd(SelType: SI.getType(), Cmp: ICI, TVal: TrueVal, FVal: FalseVal, Builder)) |
2022 | return V; |
2023 | |
2024 | if (Value *V = foldSelectICmpAndZeroShl(Cmp: ICI, TVal: TrueVal, FVal: FalseVal, Builder)) |
2025 | return replaceInstUsesWith(I&: SI, V); |
2026 | |
2027 | if (Instruction *V = foldSelectCtlzToCttz(ICI, TrueVal, FalseVal, Builder)) |
2028 | return V; |
2029 | |
2030 | if (Instruction *V = foldSelectZeroOrOnes(Cmp: ICI, TVal: TrueVal, FVal: FalseVal, Builder)) |
2031 | return V; |
2032 | |
2033 | if (Value *V = foldSelectICmpLshrAshr(IC: ICI, TrueVal, FalseVal, Builder)) |
2034 | return replaceInstUsesWith(I&: SI, V); |
2035 | |
2036 | if (Value *V = foldSelectCttzCtlz(ICI, TrueVal, FalseVal, IC&: *this)) |
2037 | return replaceInstUsesWith(I&: SI, V); |
2038 | |
2039 | if (Value *V = canonicalizeSaturatedSubtract(ICI, TrueVal, FalseVal, Builder)) |
2040 | return replaceInstUsesWith(I&: SI, V); |
2041 | |
2042 | if (Value *V = canonicalizeSaturatedAdd(Cmp: ICI, TVal: TrueVal, FVal: FalseVal, Builder)) |
2043 | return replaceInstUsesWith(I&: SI, V); |
2044 | |
2045 | if (Value *V = foldAbsDiff(Cmp: ICI, TVal: TrueVal, FVal: FalseVal, Builder)) |
2046 | return replaceInstUsesWith(I&: SI, V); |
2047 | |
2048 | if (Value *V = foldSelectWithConstOpToBinOp(Cmp: ICI, TrueVal, FalseVal)) |
2049 | return replaceInstUsesWith(I&: SI, V); |
2050 | |
2051 | return Changed ? &SI : nullptr; |
2052 | } |
2053 | |
2054 | /// We have an SPF (e.g. a min or max) of an SPF of the form: |
2055 | /// SPF2(SPF1(A, B), C) |
2056 | Instruction *InstCombinerImpl::foldSPFofSPF(Instruction *Inner, |
2057 | SelectPatternFlavor SPF1, Value *A, |
2058 | Value *B, Instruction &Outer, |
2059 | SelectPatternFlavor SPF2, |
2060 | Value *C) { |
2061 | if (Outer.getType() != Inner->getType()) |
2062 | return nullptr; |
2063 | |
2064 | if (C == A || C == B) { |
2065 | // MAX(MAX(A, B), B) -> MAX(A, B) |
2066 | // MIN(MIN(a, b), a) -> MIN(a, b) |
2067 | // TODO: This could be done in instsimplify. |
2068 | if (SPF1 == SPF2 && SelectPatternResult::isMinOrMax(SPF: SPF1)) |
2069 | return replaceInstUsesWith(I&: Outer, V: Inner); |
2070 | } |
2071 | |
2072 | return nullptr; |
2073 | } |
2074 | |
2075 | /// Turn select C, (X + Y), (X - Y) --> (X + (select C, Y, (-Y))). |
2076 | /// This is even legal for FP. |
2077 | static Instruction *foldAddSubSelect(SelectInst &SI, |
2078 | InstCombiner::BuilderTy &Builder) { |
2079 | Value *CondVal = SI.getCondition(); |
2080 | Value *TrueVal = SI.getTrueValue(); |
2081 | Value *FalseVal = SI.getFalseValue(); |
2082 | auto *TI = dyn_cast<Instruction>(Val: TrueVal); |
2083 | auto *FI = dyn_cast<Instruction>(Val: FalseVal); |
2084 | if (!TI || !FI || !TI->hasOneUse() || !FI->hasOneUse()) |
2085 | return nullptr; |
2086 | |
2087 | Instruction *AddOp = nullptr, *SubOp = nullptr; |
2088 | if ((TI->getOpcode() == Instruction::Sub && |
2089 | FI->getOpcode() == Instruction::Add) || |
2090 | (TI->getOpcode() == Instruction::FSub && |
2091 | FI->getOpcode() == Instruction::FAdd)) { |
2092 | AddOp = FI; |
2093 | SubOp = TI; |
2094 | } else if ((FI->getOpcode() == Instruction::Sub && |
2095 | TI->getOpcode() == Instruction::Add) || |
2096 | (FI->getOpcode() == Instruction::FSub && |
2097 | TI->getOpcode() == Instruction::FAdd)) { |
2098 | AddOp = TI; |
2099 | SubOp = FI; |
2100 | } |
2101 | |
2102 | if (AddOp) { |
2103 | Value *OtherAddOp = nullptr; |
2104 | if (SubOp->getOperand(i: 0) == AddOp->getOperand(i: 0)) { |
2105 | OtherAddOp = AddOp->getOperand(i: 1); |
2106 | } else if (SubOp->getOperand(i: 0) == AddOp->getOperand(i: 1)) { |
2107 | OtherAddOp = AddOp->getOperand(i: 0); |
2108 | } |
2109 | |
2110 | if (OtherAddOp) { |
2111 | // So at this point we know we have (Y -> OtherAddOp): |
2112 | // select C, (add X, Y), (sub X, Z) |
2113 | Value *NegVal; // Compute -Z |
2114 | if (SI.getType()->isFPOrFPVectorTy()) { |
2115 | NegVal = Builder.CreateFNeg(V: SubOp->getOperand(i: 1)); |
2116 | if (Instruction *NegInst = dyn_cast<Instruction>(Val: NegVal)) { |
2117 | FastMathFlags Flags = AddOp->getFastMathFlags(); |
2118 | Flags &= SubOp->getFastMathFlags(); |
2119 | NegInst->setFastMathFlags(Flags); |
2120 | } |
2121 | } else { |
2122 | NegVal = Builder.CreateNeg(V: SubOp->getOperand(i: 1)); |
2123 | } |
2124 | |
2125 | Value *NewTrueOp = OtherAddOp; |
2126 | Value *NewFalseOp = NegVal; |
2127 | if (AddOp != TI) |
2128 | std::swap(a&: NewTrueOp, b&: NewFalseOp); |
2129 | Value *NewSel = Builder.CreateSelect(C: CondVal, True: NewTrueOp, False: NewFalseOp, |
2130 | Name: SI.getName() + ".p" , MDFrom: &SI); |
2131 | |
2132 | if (SI.getType()->isFPOrFPVectorTy()) { |
2133 | Instruction *RI = |
2134 | BinaryOperator::CreateFAdd(V1: SubOp->getOperand(i: 0), V2: NewSel); |
2135 | |
2136 | FastMathFlags Flags = AddOp->getFastMathFlags(); |
2137 | Flags &= SubOp->getFastMathFlags(); |
2138 | RI->setFastMathFlags(Flags); |
2139 | return RI; |
2140 | } else |
2141 | return BinaryOperator::CreateAdd(V1: SubOp->getOperand(i: 0), V2: NewSel); |
2142 | } |
2143 | } |
2144 | return nullptr; |
2145 | } |
2146 | |
2147 | /// Turn X + Y overflows ? -1 : X + Y -> uadd_sat X, Y |
2148 | /// And X - Y overflows ? 0 : X - Y -> usub_sat X, Y |
2149 | /// Along with a number of patterns similar to: |
2150 | /// X + Y overflows ? (X < 0 ? INTMIN : INTMAX) : X + Y --> sadd_sat X, Y |
2151 | /// X - Y overflows ? (X > 0 ? INTMAX : INTMIN) : X - Y --> ssub_sat X, Y |
2152 | static Instruction * |
2153 | foldOverflowingAddSubSelect(SelectInst &SI, InstCombiner::BuilderTy &Builder) { |
2154 | Value *CondVal = SI.getCondition(); |
2155 | Value *TrueVal = SI.getTrueValue(); |
2156 | Value *FalseVal = SI.getFalseValue(); |
2157 | |
2158 | WithOverflowInst *II; |
2159 | if (!match(V: CondVal, P: m_ExtractValue<1>(V: m_WithOverflowInst(I&: II))) || |
2160 | !match(V: FalseVal, P: m_ExtractValue<0>(V: m_Specific(V: II)))) |
2161 | return nullptr; |
2162 | |
2163 | Value *X = II->getLHS(); |
2164 | Value *Y = II->getRHS(); |
2165 | |
2166 | auto IsSignedSaturateLimit = [&](Value *Limit, bool IsAdd) { |
2167 | Type *Ty = Limit->getType(); |
2168 | |
2169 | CmpPredicate Pred; |
2170 | Value *TrueVal, *FalseVal, *Op; |
2171 | const APInt *C; |
2172 | if (!match(V: Limit, P: m_Select(C: m_ICmp(Pred, L: m_Value(V&: Op), R: m_APInt(Res&: C)), |
2173 | L: m_Value(V&: TrueVal), R: m_Value(V&: FalseVal)))) |
2174 | return false; |
2175 | |
2176 | auto IsZeroOrOne = [](const APInt &C) { return C.isZero() || C.isOne(); }; |
2177 | auto IsMinMax = [&](Value *Min, Value *Max) { |
2178 | APInt MinVal = APInt::getSignedMinValue(numBits: Ty->getScalarSizeInBits()); |
2179 | APInt MaxVal = APInt::getSignedMaxValue(numBits: Ty->getScalarSizeInBits()); |
2180 | return match(V: Min, P: m_SpecificInt(V: MinVal)) && |
2181 | match(V: Max, P: m_SpecificInt(V: MaxVal)); |
2182 | }; |
2183 | |
2184 | if (Op != X && Op != Y) |
2185 | return false; |
2186 | |
2187 | if (IsAdd) { |
2188 | // X + Y overflows ? (X <s 0 ? INTMIN : INTMAX) : X + Y --> sadd_sat X, Y |
2189 | // X + Y overflows ? (X <s 1 ? INTMIN : INTMAX) : X + Y --> sadd_sat X, Y |
2190 | // X + Y overflows ? (Y <s 0 ? INTMIN : INTMAX) : X + Y --> sadd_sat X, Y |
2191 | // X + Y overflows ? (Y <s 1 ? INTMIN : INTMAX) : X + Y --> sadd_sat X, Y |
2192 | if (Pred == ICmpInst::ICMP_SLT && IsZeroOrOne(*C) && |
2193 | IsMinMax(TrueVal, FalseVal)) |
2194 | return true; |
2195 | // X + Y overflows ? (X >s 0 ? INTMAX : INTMIN) : X + Y --> sadd_sat X, Y |
2196 | // X + Y overflows ? (X >s -1 ? INTMAX : INTMIN) : X + Y --> sadd_sat X, Y |
2197 | // X + Y overflows ? (Y >s 0 ? INTMAX : INTMIN) : X + Y --> sadd_sat X, Y |
2198 | // X + Y overflows ? (Y >s -1 ? INTMAX : INTMIN) : X + Y --> sadd_sat X, Y |
2199 | if (Pred == ICmpInst::ICMP_SGT && IsZeroOrOne(*C + 1) && |
2200 | IsMinMax(FalseVal, TrueVal)) |
2201 | return true; |
2202 | } else { |
2203 | // X - Y overflows ? (X <s 0 ? INTMIN : INTMAX) : X - Y --> ssub_sat X, Y |
2204 | // X - Y overflows ? (X <s -1 ? INTMIN : INTMAX) : X - Y --> ssub_sat X, Y |
2205 | if (Op == X && Pred == ICmpInst::ICMP_SLT && IsZeroOrOne(*C + 1) && |
2206 | IsMinMax(TrueVal, FalseVal)) |
2207 | return true; |
2208 | // X - Y overflows ? (X >s -1 ? INTMAX : INTMIN) : X - Y --> ssub_sat X, Y |
2209 | // X - Y overflows ? (X >s -2 ? INTMAX : INTMIN) : X - Y --> ssub_sat X, Y |
2210 | if (Op == X && Pred == ICmpInst::ICMP_SGT && IsZeroOrOne(*C + 2) && |
2211 | IsMinMax(FalseVal, TrueVal)) |
2212 | return true; |
2213 | // X - Y overflows ? (Y <s 0 ? INTMAX : INTMIN) : X - Y --> ssub_sat X, Y |
2214 | // X - Y overflows ? (Y <s 1 ? INTMAX : INTMIN) : X - Y --> ssub_sat X, Y |
2215 | if (Op == Y && Pred == ICmpInst::ICMP_SLT && IsZeroOrOne(*C) && |
2216 | IsMinMax(FalseVal, TrueVal)) |
2217 | return true; |
2218 | // X - Y overflows ? (Y >s 0 ? INTMIN : INTMAX) : X - Y --> ssub_sat X, Y |
2219 | // X - Y overflows ? (Y >s -1 ? INTMIN : INTMAX) : X - Y --> ssub_sat X, Y |
2220 | if (Op == Y && Pred == ICmpInst::ICMP_SGT && IsZeroOrOne(*C + 1) && |
2221 | IsMinMax(TrueVal, FalseVal)) |
2222 | return true; |
2223 | } |
2224 | |
2225 | return false; |
2226 | }; |
2227 | |
2228 | Intrinsic::ID NewIntrinsicID; |
2229 | if (II->getIntrinsicID() == Intrinsic::uadd_with_overflow && |
2230 | match(V: TrueVal, P: m_AllOnes())) |
2231 | // X + Y overflows ? -1 : X + Y -> uadd_sat X, Y |
2232 | NewIntrinsicID = Intrinsic::uadd_sat; |
2233 | else if (II->getIntrinsicID() == Intrinsic::usub_with_overflow && |
2234 | match(V: TrueVal, P: m_Zero())) |
2235 | // X - Y overflows ? 0 : X - Y -> usub_sat X, Y |
2236 | NewIntrinsicID = Intrinsic::usub_sat; |
2237 | else if (II->getIntrinsicID() == Intrinsic::sadd_with_overflow && |
2238 | IsSignedSaturateLimit(TrueVal, /*IsAdd=*/true)) |
2239 | // X + Y overflows ? (X <s 0 ? INTMIN : INTMAX) : X + Y --> sadd_sat X, Y |
2240 | // X + Y overflows ? (X <s 1 ? INTMIN : INTMAX) : X + Y --> sadd_sat X, Y |
2241 | // X + Y overflows ? (X >s 0 ? INTMAX : INTMIN) : X + Y --> sadd_sat X, Y |
2242 | // X + Y overflows ? (X >s -1 ? INTMAX : INTMIN) : X + Y --> sadd_sat X, Y |
2243 | // X + Y overflows ? (Y <s 0 ? INTMIN : INTMAX) : X + Y --> sadd_sat X, Y |
2244 | // X + Y overflows ? (Y <s 1 ? INTMIN : INTMAX) : X + Y --> sadd_sat X, Y |
2245 | // X + Y overflows ? (Y >s 0 ? INTMAX : INTMIN) : X + Y --> sadd_sat X, Y |
2246 | // X + Y overflows ? (Y >s -1 ? INTMAX : INTMIN) : X + Y --> sadd_sat X, Y |
2247 | NewIntrinsicID = Intrinsic::sadd_sat; |
2248 | else if (II->getIntrinsicID() == Intrinsic::ssub_with_overflow && |
2249 | IsSignedSaturateLimit(TrueVal, /*IsAdd=*/false)) |
2250 | // X - Y overflows ? (X <s 0 ? INTMIN : INTMAX) : X - Y --> ssub_sat X, Y |
2251 | // X - Y overflows ? (X <s -1 ? INTMIN : INTMAX) : X - Y --> ssub_sat X, Y |
2252 | // X - Y overflows ? (X >s -1 ? INTMAX : INTMIN) : X - Y --> ssub_sat X, Y |
2253 | // X - Y overflows ? (X >s -2 ? INTMAX : INTMIN) : X - Y --> ssub_sat X, Y |
2254 | // X - Y overflows ? (Y <s 0 ? INTMAX : INTMIN) : X - Y --> ssub_sat X, Y |
2255 | // X - Y overflows ? (Y <s 1 ? INTMAX : INTMIN) : X - Y --> ssub_sat X, Y |
2256 | // X - Y overflows ? (Y >s 0 ? INTMIN : INTMAX) : X - Y --> ssub_sat X, Y |
2257 | // X - Y overflows ? (Y >s -1 ? INTMIN : INTMAX) : X - Y --> ssub_sat X, Y |
2258 | NewIntrinsicID = Intrinsic::ssub_sat; |
2259 | else |
2260 | return nullptr; |
2261 | |
2262 | Function *F = Intrinsic::getOrInsertDeclaration(M: SI.getModule(), |
2263 | id: NewIntrinsicID, Tys: SI.getType()); |
2264 | return CallInst::Create(Func: F, Args: {X, Y}); |
2265 | } |
2266 | |
2267 | Instruction *InstCombinerImpl::foldSelectExtConst(SelectInst &Sel) { |
2268 | Constant *C; |
2269 | if (!match(V: Sel.getTrueValue(), P: m_Constant(C)) && |
2270 | !match(V: Sel.getFalseValue(), P: m_Constant(C))) |
2271 | return nullptr; |
2272 | |
2273 | Instruction *ExtInst; |
2274 | if (!match(V: Sel.getTrueValue(), P: m_Instruction(I&: ExtInst)) && |
2275 | !match(V: Sel.getFalseValue(), P: m_Instruction(I&: ExtInst))) |
2276 | return nullptr; |
2277 | |
2278 | auto ExtOpcode = ExtInst->getOpcode(); |
2279 | if (ExtOpcode != Instruction::ZExt && ExtOpcode != Instruction::SExt) |
2280 | return nullptr; |
2281 | |
2282 | // If we are extending from a boolean type or if we can create a select that |
2283 | // has the same size operands as its condition, try to narrow the select. |
2284 | Value *X = ExtInst->getOperand(i: 0); |
2285 | Type *SmallType = X->getType(); |
2286 | Value *Cond = Sel.getCondition(); |
2287 | auto *Cmp = dyn_cast<CmpInst>(Val: Cond); |
2288 | if (!SmallType->isIntOrIntVectorTy(BitWidth: 1) && |
2289 | (!Cmp || Cmp->getOperand(i_nocapture: 0)->getType() != SmallType)) |
2290 | return nullptr; |
2291 | |
2292 | // If the constant is the same after truncation to the smaller type and |
2293 | // extension to the original type, we can narrow the select. |
2294 | Type *SelType = Sel.getType(); |
2295 | Constant *TruncC = getLosslessTrunc(C, TruncTy: SmallType, ExtOp: ExtOpcode); |
2296 | if (TruncC && ExtInst->hasOneUse()) { |
2297 | Value *TruncCVal = cast<Value>(Val: TruncC); |
2298 | if (ExtInst == Sel.getFalseValue()) |
2299 | std::swap(a&: X, b&: TruncCVal); |
2300 | |
2301 | // select Cond, (ext X), C --> ext(select Cond, X, C') |
2302 | // select Cond, C, (ext X) --> ext(select Cond, C', X) |
2303 | Value *NewSel = Builder.CreateSelect(C: Cond, True: X, False: TruncCVal, Name: "narrow" , MDFrom: &Sel); |
2304 | return CastInst::Create(Instruction::CastOps(ExtOpcode), S: NewSel, Ty: SelType); |
2305 | } |
2306 | |
2307 | return nullptr; |
2308 | } |
2309 | |
2310 | /// Try to transform a vector select with a constant condition vector into a |
2311 | /// shuffle for easier combining with other shuffles and insert/extract. |
2312 | static Instruction *canonicalizeSelectToShuffle(SelectInst &SI) { |
2313 | Value *CondVal = SI.getCondition(); |
2314 | Constant *CondC; |
2315 | auto *CondValTy = dyn_cast<FixedVectorType>(Val: CondVal->getType()); |
2316 | if (!CondValTy || !match(V: CondVal, P: m_Constant(C&: CondC))) |
2317 | return nullptr; |
2318 | |
2319 | unsigned NumElts = CondValTy->getNumElements(); |
2320 | SmallVector<int, 16> Mask; |
2321 | Mask.reserve(N: NumElts); |
2322 | for (unsigned i = 0; i != NumElts; ++i) { |
2323 | Constant *Elt = CondC->getAggregateElement(Elt: i); |
2324 | if (!Elt) |
2325 | return nullptr; |
2326 | |
2327 | if (Elt->isOneValue()) { |
2328 | // If the select condition element is true, choose from the 1st vector. |
2329 | Mask.push_back(Elt: i); |
2330 | } else if (Elt->isNullValue()) { |
2331 | // If the select condition element is false, choose from the 2nd vector. |
2332 | Mask.push_back(Elt: i + NumElts); |
2333 | } else if (isa<UndefValue>(Val: Elt)) { |
2334 | // Undef in a select condition (choose one of the operands) does not mean |
2335 | // the same thing as undef in a shuffle mask (any value is acceptable), so |
2336 | // give up. |
2337 | return nullptr; |
2338 | } else { |
2339 | // Bail out on a constant expression. |
2340 | return nullptr; |
2341 | } |
2342 | } |
2343 | |
2344 | return new ShuffleVectorInst(SI.getTrueValue(), SI.getFalseValue(), Mask); |
2345 | } |
2346 | |
2347 | /// If we have a select of vectors with a scalar condition, try to convert that |
2348 | /// to a vector select by splatting the condition. A splat may get folded with |
2349 | /// other operations in IR and having all operands of a select be vector types |
2350 | /// is likely better for vector codegen. |
2351 | static Instruction *canonicalizeScalarSelectOfVecs(SelectInst &Sel, |
2352 | InstCombinerImpl &IC) { |
2353 | auto *Ty = dyn_cast<VectorType>(Val: Sel.getType()); |
2354 | if (!Ty) |
2355 | return nullptr; |
2356 | |
2357 | // We can replace a single-use extract with constant index. |
2358 | Value *Cond = Sel.getCondition(); |
2359 | if (!match(V: Cond, P: m_OneUse(SubPattern: m_ExtractElt(Val: m_Value(), Idx: m_ConstantInt())))) |
2360 | return nullptr; |
2361 | |
2362 | // select (extelt V, Index), T, F --> select (splat V, Index), T, F |
2363 | // Splatting the extracted condition reduces code (we could directly create a |
2364 | // splat shuffle of the source vector to eliminate the intermediate step). |
2365 | return IC.replaceOperand( |
2366 | I&: Sel, OpNum: 0, V: IC.Builder.CreateVectorSplat(EC: Ty->getElementCount(), V: Cond)); |
2367 | } |
2368 | |
2369 | /// Reuse bitcasted operands between a compare and select: |
2370 | /// select (cmp (bitcast C), (bitcast D)), (bitcast' C), (bitcast' D) --> |
2371 | /// bitcast (select (cmp (bitcast C), (bitcast D)), (bitcast C), (bitcast D)) |
2372 | static Instruction *foldSelectCmpBitcasts(SelectInst &Sel, |
2373 | InstCombiner::BuilderTy &Builder) { |
2374 | Value *Cond = Sel.getCondition(); |
2375 | Value *TVal = Sel.getTrueValue(); |
2376 | Value *FVal = Sel.getFalseValue(); |
2377 | |
2378 | CmpPredicate Pred; |
2379 | Value *A, *B; |
2380 | if (!match(V: Cond, P: m_Cmp(Pred, L: m_Value(V&: A), R: m_Value(V&: B)))) |
2381 | return nullptr; |
2382 | |
2383 | // The select condition is a compare instruction. If the select's true/false |
2384 | // values are already the same as the compare operands, there's nothing to do. |
2385 | if (TVal == A || TVal == B || FVal == A || FVal == B) |
2386 | return nullptr; |
2387 | |
2388 | Value *C, *D; |
2389 | if (!match(V: A, P: m_BitCast(Op: m_Value(V&: C))) || !match(V: B, P: m_BitCast(Op: m_Value(V&: D)))) |
2390 | return nullptr; |
2391 | |
2392 | // select (cmp (bitcast C), (bitcast D)), (bitcast TSrc), (bitcast FSrc) |
2393 | Value *TSrc, *FSrc; |
2394 | if (!match(V: TVal, P: m_BitCast(Op: m_Value(V&: TSrc))) || |
2395 | !match(V: FVal, P: m_BitCast(Op: m_Value(V&: FSrc)))) |
2396 | return nullptr; |
2397 | |
2398 | // If the select true/false values are *different bitcasts* of the same source |
2399 | // operands, make the select operands the same as the compare operands and |
2400 | // cast the result. This is the canonical select form for min/max. |
2401 | Value *NewSel; |
2402 | if (TSrc == C && FSrc == D) { |
2403 | // select (cmp (bitcast C), (bitcast D)), (bitcast' C), (bitcast' D) --> |
2404 | // bitcast (select (cmp A, B), A, B) |
2405 | NewSel = Builder.CreateSelect(C: Cond, True: A, False: B, Name: "" , MDFrom: &Sel); |
2406 | } else if (TSrc == D && FSrc == C) { |
2407 | // select (cmp (bitcast C), (bitcast D)), (bitcast' D), (bitcast' C) --> |
2408 | // bitcast (select (cmp A, B), B, A) |
2409 | NewSel = Builder.CreateSelect(C: Cond, True: B, False: A, Name: "" , MDFrom: &Sel); |
2410 | } else { |
2411 | return nullptr; |
2412 | } |
2413 | return new BitCastInst(NewSel, Sel.getType()); |
2414 | } |
2415 | |
2416 | /// Try to eliminate select instructions that test the returned flag of cmpxchg |
2417 | /// instructions. |
2418 | /// |
2419 | /// If a select instruction tests the returned flag of a cmpxchg instruction and |
2420 | /// selects between the returned value of the cmpxchg instruction its compare |
2421 | /// operand, the result of the select will always be equal to its false value. |
2422 | /// For example: |
2423 | /// |
2424 | /// %cmpxchg = cmpxchg ptr %ptr, i64 %compare, i64 %new_value seq_cst seq_cst |
2425 | /// %val = extractvalue { i64, i1 } %cmpxchg, 0 |
2426 | /// %success = extractvalue { i64, i1 } %cmpxchg, 1 |
2427 | /// %sel = select i1 %success, i64 %compare, i64 %val |
2428 | /// ret i64 %sel |
2429 | /// |
2430 | /// The returned value of the cmpxchg instruction (%val) is the original value |
2431 | /// located at %ptr prior to any update. If the cmpxchg operation succeeds, %val |
2432 | /// must have been equal to %compare. Thus, the result of the select is always |
2433 | /// equal to %val, and the code can be simplified to: |
2434 | /// |
2435 | /// %cmpxchg = cmpxchg ptr %ptr, i64 %compare, i64 %new_value seq_cst seq_cst |
2436 | /// %val = extractvalue { i64, i1 } %cmpxchg, 0 |
2437 | /// ret i64 %val |
2438 | /// |
2439 | static Value *foldSelectCmpXchg(SelectInst &SI) { |
2440 | // A helper that determines if V is an extractvalue instruction whose |
2441 | // aggregate operand is a cmpxchg instruction and whose single index is equal |
2442 | // to I. If such conditions are true, the helper returns the cmpxchg |
2443 | // instruction; otherwise, a nullptr is returned. |
2444 | auto = [](Value *V, unsigned I) -> AtomicCmpXchgInst * { |
2445 | auto * = dyn_cast<ExtractValueInst>(Val: V); |
2446 | if (!Extract) |
2447 | return nullptr; |
2448 | if (Extract->getIndices()[0] != I) |
2449 | return nullptr; |
2450 | return dyn_cast<AtomicCmpXchgInst>(Val: Extract->getAggregateOperand()); |
2451 | }; |
2452 | |
2453 | // If the select has a single user, and this user is a select instruction that |
2454 | // we can simplify, skip the cmpxchg simplification for now. |
2455 | if (SI.hasOneUse()) |
2456 | if (auto *Select = dyn_cast<SelectInst>(Val: SI.user_back())) |
2457 | if (Select->getCondition() == SI.getCondition()) |
2458 | if (Select->getFalseValue() == SI.getTrueValue() || |
2459 | Select->getTrueValue() == SI.getFalseValue()) |
2460 | return nullptr; |
2461 | |
2462 | // Ensure the select condition is the returned flag of a cmpxchg instruction. |
2463 | auto *CmpXchg = isExtractFromCmpXchg(SI.getCondition(), 1); |
2464 | if (!CmpXchg) |
2465 | return nullptr; |
2466 | |
2467 | // Check the true value case: The true value of the select is the returned |
2468 | // value of the same cmpxchg used by the condition, and the false value is the |
2469 | // cmpxchg instruction's compare operand. |
2470 | if (auto *X = isExtractFromCmpXchg(SI.getTrueValue(), 0)) |
2471 | if (X == CmpXchg && X->getCompareOperand() == SI.getFalseValue()) |
2472 | return SI.getFalseValue(); |
2473 | |
2474 | // Check the false value case: The false value of the select is the returned |
2475 | // value of the same cmpxchg used by the condition, and the true value is the |
2476 | // cmpxchg instruction's compare operand. |
2477 | if (auto *X = isExtractFromCmpXchg(SI.getFalseValue(), 0)) |
2478 | if (X == CmpXchg && X->getCompareOperand() == SI.getTrueValue()) |
2479 | return SI.getFalseValue(); |
2480 | |
2481 | return nullptr; |
2482 | } |
2483 | |
2484 | /// Try to reduce a funnel/rotate pattern that includes a compare and select |
2485 | /// into a funnel shift intrinsic. Example: |
2486 | /// rotl32(a, b) --> (b == 0 ? a : ((a >> (32 - b)) | (a << b))) |
2487 | /// --> call llvm.fshl.i32(a, a, b) |
2488 | /// fshl32(a, b, c) --> (c == 0 ? a : ((b >> (32 - c)) | (a << c))) |
2489 | /// --> call llvm.fshl.i32(a, b, c) |
2490 | /// fshr32(a, b, c) --> (c == 0 ? b : ((a >> (32 - c)) | (b << c))) |
2491 | /// --> call llvm.fshr.i32(a, b, c) |
2492 | static Instruction *foldSelectFunnelShift(SelectInst &Sel, |
2493 | InstCombiner::BuilderTy &Builder) { |
2494 | // This must be a power-of-2 type for a bitmasking transform to be valid. |
2495 | unsigned Width = Sel.getType()->getScalarSizeInBits(); |
2496 | if (!isPowerOf2_32(Value: Width)) |
2497 | return nullptr; |
2498 | |
2499 | BinaryOperator *Or0, *Or1; |
2500 | if (!match(V: Sel.getFalseValue(), P: m_OneUse(SubPattern: m_Or(L: m_BinOp(I&: Or0), R: m_BinOp(I&: Or1))))) |
2501 | return nullptr; |
2502 | |
2503 | Value *SV0, *SV1, *SA0, *SA1; |
2504 | if (!match(V: Or0, P: m_OneUse(SubPattern: m_LogicalShift(L: m_Value(V&: SV0), |
2505 | R: m_ZExtOrSelf(Op: m_Value(V&: SA0))))) || |
2506 | !match(V: Or1, P: m_OneUse(SubPattern: m_LogicalShift(L: m_Value(V&: SV1), |
2507 | R: m_ZExtOrSelf(Op: m_Value(V&: SA1))))) || |
2508 | Or0->getOpcode() == Or1->getOpcode()) |
2509 | return nullptr; |
2510 | |
2511 | // Canonicalize to or(shl(SV0, SA0), lshr(SV1, SA1)). |
2512 | if (Or0->getOpcode() == BinaryOperator::LShr) { |
2513 | std::swap(a&: Or0, b&: Or1); |
2514 | std::swap(a&: SV0, b&: SV1); |
2515 | std::swap(a&: SA0, b&: SA1); |
2516 | } |
2517 | assert(Or0->getOpcode() == BinaryOperator::Shl && |
2518 | Or1->getOpcode() == BinaryOperator::LShr && |
2519 | "Illegal or(shift,shift) pair" ); |
2520 | |
2521 | // Check the shift amounts to see if they are an opposite pair. |
2522 | Value *ShAmt; |
2523 | if (match(V: SA1, P: m_OneUse(SubPattern: m_Sub(L: m_SpecificInt(V: Width), R: m_Specific(V: SA0))))) |
2524 | ShAmt = SA0; |
2525 | else if (match(V: SA0, P: m_OneUse(SubPattern: m_Sub(L: m_SpecificInt(V: Width), R: m_Specific(V: SA1))))) |
2526 | ShAmt = SA1; |
2527 | else |
2528 | return nullptr; |
2529 | |
2530 | // We should now have this pattern: |
2531 | // select ?, TVal, (or (shl SV0, SA0), (lshr SV1, SA1)) |
2532 | // The false value of the select must be a funnel-shift of the true value: |
2533 | // IsFShl -> TVal must be SV0 else TVal must be SV1. |
2534 | bool IsFshl = (ShAmt == SA0); |
2535 | Value *TVal = Sel.getTrueValue(); |
2536 | if ((IsFshl && TVal != SV0) || (!IsFshl && TVal != SV1)) |
2537 | return nullptr; |
2538 | |
2539 | // Finally, see if the select is filtering out a shift-by-zero. |
2540 | Value *Cond = Sel.getCondition(); |
2541 | if (!match(V: Cond, P: m_OneUse(SubPattern: m_SpecificICmp(MatchPred: ICmpInst::ICMP_EQ, L: m_Specific(V: ShAmt), |
2542 | R: m_ZeroInt())))) |
2543 | return nullptr; |
2544 | |
2545 | // If this is not a rotate then the select was blocking poison from the |
2546 | // 'shift-by-zero' non-TVal, but a funnel shift won't - so freeze it. |
2547 | if (SV0 != SV1) { |
2548 | if (IsFshl && !llvm::isGuaranteedNotToBePoison(V: SV1)) |
2549 | SV1 = Builder.CreateFreeze(V: SV1); |
2550 | else if (!IsFshl && !llvm::isGuaranteedNotToBePoison(V: SV0)) |
2551 | SV0 = Builder.CreateFreeze(V: SV0); |
2552 | } |
2553 | |
2554 | // This is a funnel/rotate that avoids shift-by-bitwidth UB in a suboptimal way. |
2555 | // Convert to funnel shift intrinsic. |
2556 | Intrinsic::ID IID = IsFshl ? Intrinsic::fshl : Intrinsic::fshr; |
2557 | Function *F = |
2558 | Intrinsic::getOrInsertDeclaration(M: Sel.getModule(), id: IID, Tys: Sel.getType()); |
2559 | ShAmt = Builder.CreateZExt(V: ShAmt, DestTy: Sel.getType()); |
2560 | return CallInst::Create(Func: F, Args: { SV0, SV1, ShAmt }); |
2561 | } |
2562 | |
2563 | static Instruction *foldSelectToCopysign(SelectInst &Sel, |
2564 | InstCombiner::BuilderTy &Builder) { |
2565 | Value *Cond = Sel.getCondition(); |
2566 | Value *TVal = Sel.getTrueValue(); |
2567 | Value *FVal = Sel.getFalseValue(); |
2568 | Type *SelType = Sel.getType(); |
2569 | |
2570 | // Match select ?, TC, FC where the constants are equal but negated. |
2571 | // TODO: Generalize to handle a negated variable operand? |
2572 | const APFloat *TC, *FC; |
2573 | if (!match(V: TVal, P: m_APFloatAllowPoison(Res&: TC)) || |
2574 | !match(V: FVal, P: m_APFloatAllowPoison(Res&: FC)) || |
2575 | !abs(X: *TC).bitwiseIsEqual(RHS: abs(X: *FC))) |
2576 | return nullptr; |
2577 | |
2578 | assert(TC != FC && "Expected equal select arms to simplify" ); |
2579 | |
2580 | Value *X; |
2581 | const APInt *C; |
2582 | bool IsTrueIfSignSet; |
2583 | CmpPredicate Pred; |
2584 | if (!match(V: Cond, P: m_OneUse(SubPattern: m_ICmp(Pred, L: m_ElementWiseBitCast(Op: m_Value(V&: X)), |
2585 | R: m_APInt(Res&: C)))) || |
2586 | !isSignBitCheck(Pred, RHS: *C, TrueIfSigned&: IsTrueIfSignSet) || X->getType() != SelType) |
2587 | return nullptr; |
2588 | |
2589 | // If needed, negate the value that will be the sign argument of the copysign: |
2590 | // (bitcast X) < 0 ? -TC : TC --> copysign(TC, X) |
2591 | // (bitcast X) < 0 ? TC : -TC --> copysign(TC, -X) |
2592 | // (bitcast X) >= 0 ? -TC : TC --> copysign(TC, -X) |
2593 | // (bitcast X) >= 0 ? TC : -TC --> copysign(TC, X) |
2594 | // Note: FMF from the select can not be propagated to the new instructions. |
2595 | if (IsTrueIfSignSet ^ TC->isNegative()) |
2596 | X = Builder.CreateFNeg(V: X); |
2597 | |
2598 | // Canonicalize the magnitude argument as the positive constant since we do |
2599 | // not care about its sign. |
2600 | Value *MagArg = ConstantFP::get(Ty: SelType, V: abs(X: *TC)); |
2601 | Function *F = Intrinsic::getOrInsertDeclaration( |
2602 | M: Sel.getModule(), id: Intrinsic::copysign, Tys: Sel.getType()); |
2603 | return CallInst::Create(Func: F, Args: { MagArg, X }); |
2604 | } |
2605 | |
2606 | Instruction *InstCombinerImpl::foldVectorSelect(SelectInst &Sel) { |
2607 | if (!isa<VectorType>(Val: Sel.getType())) |
2608 | return nullptr; |
2609 | |
2610 | Value *Cond = Sel.getCondition(); |
2611 | Value *TVal = Sel.getTrueValue(); |
2612 | Value *FVal = Sel.getFalseValue(); |
2613 | Value *C, *X, *Y; |
2614 | |
2615 | if (match(V: Cond, P: m_VecReverse(Op0: m_Value(V&: C)))) { |
2616 | auto createSelReverse = [&](Value *C, Value *X, Value *Y) { |
2617 | Value *V = Builder.CreateSelect(C, True: X, False: Y, Name: Sel.getName(), MDFrom: &Sel); |
2618 | if (auto *I = dyn_cast<Instruction>(Val: V)) |
2619 | I->copyIRFlags(V: &Sel); |
2620 | Module *M = Sel.getModule(); |
2621 | Function *F = Intrinsic::getOrInsertDeclaration( |
2622 | M, id: Intrinsic::vector_reverse, Tys: V->getType()); |
2623 | return CallInst::Create(Func: F, Args: V); |
2624 | }; |
2625 | |
2626 | if (match(V: TVal, P: m_VecReverse(Op0: m_Value(V&: X)))) { |
2627 | // select rev(C), rev(X), rev(Y) --> rev(select C, X, Y) |
2628 | if (match(V: FVal, P: m_VecReverse(Op0: m_Value(V&: Y))) && |
2629 | (Cond->hasOneUse() || TVal->hasOneUse() || FVal->hasOneUse())) |
2630 | return createSelReverse(C, X, Y); |
2631 | |
2632 | // select rev(C), rev(X), FValSplat --> rev(select C, X, FValSplat) |
2633 | if ((Cond->hasOneUse() || TVal->hasOneUse()) && isSplatValue(V: FVal)) |
2634 | return createSelReverse(C, X, FVal); |
2635 | } |
2636 | // select rev(C), TValSplat, rev(Y) --> rev(select C, TValSplat, Y) |
2637 | else if (isSplatValue(V: TVal) && match(V: FVal, P: m_VecReverse(Op0: m_Value(V&: Y))) && |
2638 | (Cond->hasOneUse() || FVal->hasOneUse())) |
2639 | return createSelReverse(C, TVal, Y); |
2640 | } |
2641 | |
2642 | auto *VecTy = dyn_cast<FixedVectorType>(Val: Sel.getType()); |
2643 | if (!VecTy) |
2644 | return nullptr; |
2645 | |
2646 | unsigned NumElts = VecTy->getNumElements(); |
2647 | APInt PoisonElts(NumElts, 0); |
2648 | APInt AllOnesEltMask(APInt::getAllOnes(numBits: NumElts)); |
2649 | if (Value *V = SimplifyDemandedVectorElts(V: &Sel, DemandedElts: AllOnesEltMask, PoisonElts)) { |
2650 | if (V != &Sel) |
2651 | return replaceInstUsesWith(I&: Sel, V); |
2652 | return &Sel; |
2653 | } |
2654 | |
2655 | // A select of a "select shuffle" with a common operand can be rearranged |
2656 | // to select followed by "select shuffle". Because of poison, this only works |
2657 | // in the case of a shuffle with no undefined mask elements. |
2658 | ArrayRef<int> Mask; |
2659 | if (match(V: TVal, P: m_OneUse(SubPattern: m_Shuffle(v1: m_Value(V&: X), v2: m_Value(V&: Y), mask: m_Mask(Mask)))) && |
2660 | !is_contained(Range&: Mask, Element: PoisonMaskElem) && |
2661 | cast<ShuffleVectorInst>(Val: TVal)->isSelect()) { |
2662 | if (X == FVal) { |
2663 | // select Cond, (shuf_sel X, Y), X --> shuf_sel X, (select Cond, Y, X) |
2664 | Value *NewSel = Builder.CreateSelect(C: Cond, True: Y, False: X, Name: "sel" , MDFrom: &Sel); |
2665 | return new ShuffleVectorInst(X, NewSel, Mask); |
2666 | } |
2667 | if (Y == FVal) { |
2668 | // select Cond, (shuf_sel X, Y), Y --> shuf_sel (select Cond, X, Y), Y |
2669 | Value *NewSel = Builder.CreateSelect(C: Cond, True: X, False: Y, Name: "sel" , MDFrom: &Sel); |
2670 | return new ShuffleVectorInst(NewSel, Y, Mask); |
2671 | } |
2672 | } |
2673 | if (match(V: FVal, P: m_OneUse(SubPattern: m_Shuffle(v1: m_Value(V&: X), v2: m_Value(V&: Y), mask: m_Mask(Mask)))) && |
2674 | !is_contained(Range&: Mask, Element: PoisonMaskElem) && |
2675 | cast<ShuffleVectorInst>(Val: FVal)->isSelect()) { |
2676 | if (X == TVal) { |
2677 | // select Cond, X, (shuf_sel X, Y) --> shuf_sel X, (select Cond, X, Y) |
2678 | Value *NewSel = Builder.CreateSelect(C: Cond, True: X, False: Y, Name: "sel" , MDFrom: &Sel); |
2679 | return new ShuffleVectorInst(X, NewSel, Mask); |
2680 | } |
2681 | if (Y == TVal) { |
2682 | // select Cond, Y, (shuf_sel X, Y) --> shuf_sel (select Cond, Y, X), Y |
2683 | Value *NewSel = Builder.CreateSelect(C: Cond, True: Y, False: X, Name: "sel" , MDFrom: &Sel); |
2684 | return new ShuffleVectorInst(NewSel, Y, Mask); |
2685 | } |
2686 | } |
2687 | |
2688 | return nullptr; |
2689 | } |
2690 | |
2691 | static Instruction *foldSelectToPhiImpl(SelectInst &Sel, BasicBlock *BB, |
2692 | const DominatorTree &DT, |
2693 | InstCombiner::BuilderTy &Builder) { |
2694 | // Find the block's immediate dominator that ends with a conditional branch |
2695 | // that matches select's condition (maybe inverted). |
2696 | auto *IDomNode = DT[BB]->getIDom(); |
2697 | if (!IDomNode) |
2698 | return nullptr; |
2699 | BasicBlock *IDom = IDomNode->getBlock(); |
2700 | |
2701 | Value *Cond = Sel.getCondition(); |
2702 | Value *IfTrue, *IfFalse; |
2703 | BasicBlock *TrueSucc, *FalseSucc; |
2704 | if (match(V: IDom->getTerminator(), |
2705 | P: m_Br(C: m_Specific(V: Cond), T: m_BasicBlock(V&: TrueSucc), |
2706 | F: m_BasicBlock(V&: FalseSucc)))) { |
2707 | IfTrue = Sel.getTrueValue(); |
2708 | IfFalse = Sel.getFalseValue(); |
2709 | } else if (match(V: IDom->getTerminator(), |
2710 | P: m_Br(C: m_Not(V: m_Specific(V: Cond)), T: m_BasicBlock(V&: TrueSucc), |
2711 | F: m_BasicBlock(V&: FalseSucc)))) { |
2712 | IfTrue = Sel.getFalseValue(); |
2713 | IfFalse = Sel.getTrueValue(); |
2714 | } else |
2715 | return nullptr; |
2716 | |
2717 | // Make sure the branches are actually different. |
2718 | if (TrueSucc == FalseSucc) |
2719 | return nullptr; |
2720 | |
2721 | // We want to replace select %cond, %a, %b with a phi that takes value %a |
2722 | // for all incoming edges that are dominated by condition `%cond == true`, |
2723 | // and value %b for edges dominated by condition `%cond == false`. If %a |
2724 | // or %b are also phis from the same basic block, we can go further and take |
2725 | // their incoming values from the corresponding blocks. |
2726 | BasicBlockEdge TrueEdge(IDom, TrueSucc); |
2727 | BasicBlockEdge FalseEdge(IDom, FalseSucc); |
2728 | DenseMap<BasicBlock *, Value *> Inputs; |
2729 | for (auto *Pred : predecessors(BB)) { |
2730 | // Check implication. |
2731 | BasicBlockEdge Incoming(Pred, BB); |
2732 | if (DT.dominates(BBE1: TrueEdge, BBE2: Incoming)) |
2733 | Inputs[Pred] = IfTrue->DoPHITranslation(CurBB: BB, PredBB: Pred); |
2734 | else if (DT.dominates(BBE1: FalseEdge, BBE2: Incoming)) |
2735 | Inputs[Pred] = IfFalse->DoPHITranslation(CurBB: BB, PredBB: Pred); |
2736 | else |
2737 | return nullptr; |
2738 | // Check availability. |
2739 | if (auto *Insn = dyn_cast<Instruction>(Val: Inputs[Pred])) |
2740 | if (!DT.dominates(Def: Insn, User: Pred->getTerminator())) |
2741 | return nullptr; |
2742 | } |
2743 | |
2744 | Builder.SetInsertPoint(TheBB: BB, IP: BB->begin()); |
2745 | auto *PN = Builder.CreatePHI(Ty: Sel.getType(), NumReservedValues: Inputs.size()); |
2746 | for (auto *Pred : predecessors(BB)) |
2747 | PN->addIncoming(V: Inputs[Pred], BB: Pred); |
2748 | PN->takeName(V: &Sel); |
2749 | return PN; |
2750 | } |
2751 | |
2752 | static Instruction *foldSelectToPhi(SelectInst &Sel, const DominatorTree &DT, |
2753 | InstCombiner::BuilderTy &Builder) { |
2754 | // Try to replace this select with Phi in one of these blocks. |
2755 | SmallSetVector<BasicBlock *, 4> CandidateBlocks; |
2756 | CandidateBlocks.insert(X: Sel.getParent()); |
2757 | for (Value *V : Sel.operands()) |
2758 | if (auto *I = dyn_cast<Instruction>(Val: V)) |
2759 | CandidateBlocks.insert(X: I->getParent()); |
2760 | |
2761 | for (BasicBlock *BB : CandidateBlocks) |
2762 | if (auto *PN = foldSelectToPhiImpl(Sel, BB, DT, Builder)) |
2763 | return PN; |
2764 | return nullptr; |
2765 | } |
2766 | |
2767 | /// Tries to reduce a pattern that arises when calculating the remainder of the |
2768 | /// Euclidean division. When the divisor is a power of two and is guaranteed not |
2769 | /// to be negative, a signed remainder can be folded with a bitwise and. |
2770 | /// |
2771 | /// (x % n) < 0 ? (x % n) + n : (x % n) |
2772 | /// -> x & (n - 1) |
2773 | static Instruction *foldSelectWithSRem(SelectInst &SI, InstCombinerImpl &IC, |
2774 | IRBuilderBase &Builder) { |
2775 | Value *CondVal = SI.getCondition(); |
2776 | Value *TrueVal = SI.getTrueValue(); |
2777 | Value *FalseVal = SI.getFalseValue(); |
2778 | |
2779 | CmpPredicate Pred; |
2780 | Value *Op, *RemRes, *Remainder; |
2781 | const APInt *C; |
2782 | bool TrueIfSigned = false; |
2783 | |
2784 | if (!(match(V: CondVal, P: m_ICmp(Pred, L: m_Value(V&: RemRes), R: m_APInt(Res&: C))) && |
2785 | isSignBitCheck(Pred, RHS: *C, TrueIfSigned))) |
2786 | return nullptr; |
2787 | |
2788 | // If the sign bit is not set, we have a SGE/SGT comparison, and the operands |
2789 | // of the select are inverted. |
2790 | if (!TrueIfSigned) |
2791 | std::swap(a&: TrueVal, b&: FalseVal); |
2792 | |
2793 | auto FoldToBitwiseAnd = [&](Value *Remainder) -> Instruction * { |
2794 | Value *Add = Builder.CreateAdd( |
2795 | LHS: Remainder, RHS: Constant::getAllOnesValue(Ty: RemRes->getType())); |
2796 | return BinaryOperator::CreateAnd(V1: Op, V2: Add); |
2797 | }; |
2798 | |
2799 | // Match the general case: |
2800 | // %rem = srem i32 %x, %n |
2801 | // %cnd = icmp slt i32 %rem, 0 |
2802 | // %add = add i32 %rem, %n |
2803 | // %sel = select i1 %cnd, i32 %add, i32 %rem |
2804 | if (match(V: TrueVal, P: m_c_Add(L: m_Specific(V: RemRes), R: m_Value(V&: Remainder))) && |
2805 | match(V: RemRes, P: m_SRem(L: m_Value(V&: Op), R: m_Specific(V: Remainder))) && |
2806 | IC.isKnownToBeAPowerOfTwo(V: Remainder, /*OrZero=*/true) && |
2807 | FalseVal == RemRes) |
2808 | return FoldToBitwiseAnd(Remainder); |
2809 | |
2810 | // Match the case where the one arm has been replaced by constant 1: |
2811 | // %rem = srem i32 %n, 2 |
2812 | // %cnd = icmp slt i32 %rem, 0 |
2813 | // %sel = select i1 %cnd, i32 1, i32 %rem |
2814 | if (match(V: TrueVal, P: m_One()) && |
2815 | match(V: RemRes, P: m_SRem(L: m_Value(V&: Op), R: m_SpecificInt(V: 2))) && |
2816 | FalseVal == RemRes) |
2817 | return FoldToBitwiseAnd(ConstantInt::get(Ty: RemRes->getType(), V: 2)); |
2818 | |
2819 | return nullptr; |
2820 | } |
2821 | |
2822 | static Value *foldSelectWithFrozenICmp(SelectInst &Sel, InstCombiner::BuilderTy &Builder) { |
2823 | FreezeInst *FI = dyn_cast<FreezeInst>(Val: Sel.getCondition()); |
2824 | if (!FI) |
2825 | return nullptr; |
2826 | |
2827 | Value *Cond = FI->getOperand(i_nocapture: 0); |
2828 | Value *TrueVal = Sel.getTrueValue(), *FalseVal = Sel.getFalseValue(); |
2829 | |
2830 | // select (freeze(x == y)), x, y --> y |
2831 | // select (freeze(x != y)), x, y --> x |
2832 | // The freeze should be only used by this select. Otherwise, remaining uses of |
2833 | // the freeze can observe a contradictory value. |
2834 | // c = freeze(x == y) ; Let's assume that y = poison & x = 42; c is 0 or 1 |
2835 | // a = select c, x, y ; |
2836 | // f(a, c) ; f(poison, 1) cannot happen, but if a is folded |
2837 | // ; to y, this can happen. |
2838 | CmpPredicate Pred; |
2839 | if (FI->hasOneUse() && |
2840 | match(V: Cond, P: m_c_ICmp(Pred, L: m_Specific(V: TrueVal), R: m_Specific(V: FalseVal))) && |
2841 | (Pred == ICmpInst::ICMP_EQ || Pred == ICmpInst::ICMP_NE)) { |
2842 | return Pred == ICmpInst::ICMP_EQ ? FalseVal : TrueVal; |
2843 | } |
2844 | |
2845 | return nullptr; |
2846 | } |
2847 | |
2848 | /// Given that \p CondVal is known to be \p CondIsTrue, try to simplify \p SI. |
2849 | static Value *simplifyNestedSelectsUsingImpliedCond(SelectInst &SI, |
2850 | Value *CondVal, |
2851 | bool CondIsTrue, |
2852 | const DataLayout &DL) { |
2853 | Value *InnerCondVal = SI.getCondition(); |
2854 | Value *InnerTrueVal = SI.getTrueValue(); |
2855 | Value *InnerFalseVal = SI.getFalseValue(); |
2856 | assert(CondVal->getType() == InnerCondVal->getType() && |
2857 | "The type of inner condition must match with the outer." ); |
2858 | if (auto Implied = isImpliedCondition(LHS: CondVal, RHS: InnerCondVal, DL, LHSIsTrue: CondIsTrue)) |
2859 | return *Implied ? InnerTrueVal : InnerFalseVal; |
2860 | return nullptr; |
2861 | } |
2862 | |
2863 | Instruction *InstCombinerImpl::foldAndOrOfSelectUsingImpliedCond(Value *Op, |
2864 | SelectInst &SI, |
2865 | bool IsAnd) { |
2866 | assert(Op->getType()->isIntOrIntVectorTy(1) && |
2867 | "Op must be either i1 or vector of i1." ); |
2868 | if (SI.getCondition()->getType() != Op->getType()) |
2869 | return nullptr; |
2870 | if (Value *V = simplifyNestedSelectsUsingImpliedCond(SI, CondVal: Op, CondIsTrue: IsAnd, DL)) |
2871 | return SelectInst::Create(C: Op, |
2872 | S1: IsAnd ? V : ConstantInt::getTrue(Ty: Op->getType()), |
2873 | S2: IsAnd ? ConstantInt::getFalse(Ty: Op->getType()) : V); |
2874 | return nullptr; |
2875 | } |
2876 | |
2877 | // Canonicalize select with fcmp to fabs(). -0.0 makes this tricky. We need |
2878 | // fast-math-flags (nsz) or fsub with +0.0 (not fneg) for this to work. |
2879 | static Instruction *foldSelectWithFCmpToFabs(SelectInst &SI, |
2880 | InstCombinerImpl &IC) { |
2881 | Value *CondVal = SI.getCondition(); |
2882 | |
2883 | bool ChangedFMF = false; |
2884 | for (bool Swap : {false, true}) { |
2885 | Value *TrueVal = SI.getTrueValue(); |
2886 | Value *X = SI.getFalseValue(); |
2887 | CmpPredicate Pred; |
2888 | |
2889 | if (Swap) |
2890 | std::swap(a&: TrueVal, b&: X); |
2891 | |
2892 | if (!match(V: CondVal, P: m_FCmp(Pred, L: m_Specific(V: X), R: m_AnyZeroFP()))) |
2893 | continue; |
2894 | |
2895 | // fold (X <= +/-0.0) ? (0.0 - X) : X to fabs(X), when 'Swap' is false |
2896 | // fold (X > +/-0.0) ? X : (0.0 - X) to fabs(X), when 'Swap' is true |
2897 | // Note: We require "nnan" for this fold because fcmp ignores the signbit |
2898 | // of NAN, but IEEE-754 specifies the signbit of NAN values with |
2899 | // fneg/fabs operations. |
2900 | if (match(V: TrueVal, P: m_FSub(L: m_PosZeroFP(), R: m_Specific(V: X))) && |
2901 | (cast<FPMathOperator>(Val: CondVal)->hasNoNaNs() || SI.hasNoNaNs() || |
2902 | (SI.hasOneUse() && canIgnoreSignBitOfNaN(U: *SI.use_begin())) || |
2903 | isKnownNeverNaN(V: X, SQ: IC.getSimplifyQuery().getWithInstruction( |
2904 | I: cast<Instruction>(Val: CondVal))))) { |
2905 | if (!Swap && (Pred == FCmpInst::FCMP_OLE || Pred == FCmpInst::FCMP_ULE)) { |
2906 | Value *Fabs = IC.Builder.CreateUnaryIntrinsic(ID: Intrinsic::fabs, V: X, FMFSource: &SI); |
2907 | return IC.replaceInstUsesWith(I&: SI, V: Fabs); |
2908 | } |
2909 | if (Swap && (Pred == FCmpInst::FCMP_OGT || Pred == FCmpInst::FCMP_UGT)) { |
2910 | Value *Fabs = IC.Builder.CreateUnaryIntrinsic(ID: Intrinsic::fabs, V: X, FMFSource: &SI); |
2911 | return IC.replaceInstUsesWith(I&: SI, V: Fabs); |
2912 | } |
2913 | } |
2914 | |
2915 | if (!match(V: TrueVal, P: m_FNeg(X: m_Specific(V: X)))) |
2916 | return nullptr; |
2917 | |
2918 | // Forward-propagate nnan and ninf from the fcmp to the select. |
2919 | // If all inputs are not those values, then the select is not either. |
2920 | // Note: nsz is defined differently, so it may not be correct to propagate. |
2921 | FastMathFlags FMF = cast<FPMathOperator>(Val: CondVal)->getFastMathFlags(); |
2922 | if (FMF.noNaNs() && !SI.hasNoNaNs()) { |
2923 | SI.setHasNoNaNs(true); |
2924 | ChangedFMF = true; |
2925 | } |
2926 | if (FMF.noInfs() && !SI.hasNoInfs()) { |
2927 | SI.setHasNoInfs(true); |
2928 | ChangedFMF = true; |
2929 | } |
2930 | // Forward-propagate nnan from the fneg to the select. |
2931 | // The nnan flag can be propagated iff fneg is selected when X is NaN. |
2932 | if (!SI.hasNoNaNs() && cast<FPMathOperator>(Val: TrueVal)->hasNoNaNs() && |
2933 | (Swap ? FCmpInst::isOrdered(predicate: Pred) : FCmpInst::isUnordered(predicate: Pred))) { |
2934 | SI.setHasNoNaNs(true); |
2935 | ChangedFMF = true; |
2936 | } |
2937 | |
2938 | // With nsz, when 'Swap' is false: |
2939 | // fold (X < +/-0.0) ? -X : X or (X <= +/-0.0) ? -X : X to fabs(X) |
2940 | // fold (X > +/-0.0) ? -X : X or (X >= +/-0.0) ? -X : X to -fabs(x) |
2941 | // when 'Swap' is true: |
2942 | // fold (X > +/-0.0) ? X : -X or (X >= +/-0.0) ? X : -X to fabs(X) |
2943 | // fold (X < +/-0.0) ? X : -X or (X <= +/-0.0) ? X : -X to -fabs(X) |
2944 | // |
2945 | // Note: We require "nnan" for this fold because fcmp ignores the signbit |
2946 | // of NAN, but IEEE-754 specifies the signbit of NAN values with |
2947 | // fneg/fabs operations. |
2948 | if (!SI.hasNoSignedZeros() && |
2949 | (!SI.hasOneUse() || !canIgnoreSignBitOfZero(U: *SI.use_begin()))) |
2950 | return nullptr; |
2951 | if (!SI.hasNoNaNs() && |
2952 | (!SI.hasOneUse() || !canIgnoreSignBitOfNaN(U: *SI.use_begin()))) |
2953 | return nullptr; |
2954 | |
2955 | if (Swap) |
2956 | Pred = FCmpInst::getSwappedPredicate(pred: Pred); |
2957 | |
2958 | bool IsLTOrLE = Pred == FCmpInst::FCMP_OLT || Pred == FCmpInst::FCMP_OLE || |
2959 | Pred == FCmpInst::FCMP_ULT || Pred == FCmpInst::FCMP_ULE; |
2960 | bool IsGTOrGE = Pred == FCmpInst::FCMP_OGT || Pred == FCmpInst::FCMP_OGE || |
2961 | Pred == FCmpInst::FCMP_UGT || Pred == FCmpInst::FCMP_UGE; |
2962 | |
2963 | if (IsLTOrLE) { |
2964 | Value *Fabs = IC.Builder.CreateUnaryIntrinsic(ID: Intrinsic::fabs, V: X, FMFSource: &SI); |
2965 | return IC.replaceInstUsesWith(I&: SI, V: Fabs); |
2966 | } |
2967 | if (IsGTOrGE) { |
2968 | Value *Fabs = IC.Builder.CreateUnaryIntrinsic(ID: Intrinsic::fabs, V: X, FMFSource: &SI); |
2969 | Instruction *NewFNeg = UnaryOperator::CreateFNeg(V: Fabs); |
2970 | NewFNeg->setFastMathFlags(SI.getFastMathFlags()); |
2971 | return NewFNeg; |
2972 | } |
2973 | } |
2974 | |
2975 | // Match select with (icmp slt (bitcast X to int), 0) |
2976 | // or (icmp sgt (bitcast X to int), -1) |
2977 | |
2978 | for (bool Swap : {false, true}) { |
2979 | Value *TrueVal = SI.getTrueValue(); |
2980 | Value *X = SI.getFalseValue(); |
2981 | |
2982 | if (Swap) |
2983 | std::swap(a&: TrueVal, b&: X); |
2984 | |
2985 | CmpPredicate Pred; |
2986 | const APInt *C; |
2987 | bool TrueIfSigned; |
2988 | if (!match(V: CondVal, |
2989 | P: m_ICmp(Pred, L: m_ElementWiseBitCast(Op: m_Specific(V: X)), R: m_APInt(Res&: C))) || |
2990 | !isSignBitCheck(Pred, RHS: *C, TrueIfSigned)) |
2991 | continue; |
2992 | if (!match(V: TrueVal, P: m_FNeg(X: m_Specific(V: X)))) |
2993 | return nullptr; |
2994 | if (Swap == TrueIfSigned && !CondVal->hasOneUse() && !TrueVal->hasOneUse()) |
2995 | return nullptr; |
2996 | |
2997 | // Fold (IsNeg ? -X : X) or (!IsNeg ? X : -X) to fabs(X) |
2998 | // Fold (IsNeg ? X : -X) or (!IsNeg ? -X : X) to -fabs(X) |
2999 | Value *Fabs = IC.Builder.CreateUnaryIntrinsic(ID: Intrinsic::fabs, V: X, FMFSource: &SI); |
3000 | if (Swap != TrueIfSigned) |
3001 | return IC.replaceInstUsesWith(I&: SI, V: Fabs); |
3002 | return UnaryOperator::CreateFNegFMF(Op: Fabs, FMFSource: &SI); |
3003 | } |
3004 | |
3005 | return ChangedFMF ? &SI : nullptr; |
3006 | } |
3007 | |
3008 | // Match the following IR pattern: |
3009 | // %x.lowbits = and i8 %x, %lowbitmask |
3010 | // %x.lowbits.are.zero = icmp eq i8 %x.lowbits, 0 |
3011 | // %x.biased = add i8 %x, %bias |
3012 | // %x.biased.highbits = and i8 %x.biased, %highbitmask |
3013 | // %x.roundedup = select i1 %x.lowbits.are.zero, i8 %x, i8 %x.biased.highbits |
3014 | // Define: |
3015 | // %alignment = add i8 %lowbitmask, 1 |
3016 | // Iff 1. an %alignment is a power-of-two (aka, %lowbitmask is a low bit mask) |
3017 | // and 2. %bias is equal to either %lowbitmask or %alignment, |
3018 | // and 3. %highbitmask is equal to ~%lowbitmask (aka, to -%alignment) |
3019 | // then this pattern can be transformed into: |
3020 | // %x.offset = add i8 %x, %lowbitmask |
3021 | // %x.roundedup = and i8 %x.offset, %highbitmask |
3022 | static Value * |
3023 | foldRoundUpIntegerWithPow2Alignment(SelectInst &SI, |
3024 | InstCombiner::BuilderTy &Builder) { |
3025 | Value *Cond = SI.getCondition(); |
3026 | Value *X = SI.getTrueValue(); |
3027 | Value *XBiasedHighBits = SI.getFalseValue(); |
3028 | |
3029 | CmpPredicate Pred; |
3030 | Value *XLowBits; |
3031 | if (!match(V: Cond, P: m_ICmp(Pred, L: m_Value(V&: XLowBits), R: m_ZeroInt())) || |
3032 | !ICmpInst::isEquality(P: Pred)) |
3033 | return nullptr; |
3034 | |
3035 | if (Pred == ICmpInst::Predicate::ICMP_NE) |
3036 | std::swap(a&: X, b&: XBiasedHighBits); |
3037 | |
3038 | // FIXME: we could support non non-splats here. |
3039 | |
3040 | const APInt *LowBitMaskCst; |
3041 | if (!match(V: XLowBits, P: m_And(L: m_Specific(V: X), R: m_APIntAllowPoison(Res&: LowBitMaskCst)))) |
3042 | return nullptr; |
3043 | |
3044 | // Match even if the AND and ADD are swapped. |
3045 | const APInt *BiasCst, *HighBitMaskCst; |
3046 | if (!match(V: XBiasedHighBits, |
3047 | P: m_And(L: m_Add(L: m_Specific(V: X), R: m_APIntAllowPoison(Res&: BiasCst)), |
3048 | R: m_APIntAllowPoison(Res&: HighBitMaskCst))) && |
3049 | !match(V: XBiasedHighBits, |
3050 | P: m_Add(L: m_And(L: m_Specific(V: X), R: m_APIntAllowPoison(Res&: HighBitMaskCst)), |
3051 | R: m_APIntAllowPoison(Res&: BiasCst)))) |
3052 | return nullptr; |
3053 | |
3054 | if (!LowBitMaskCst->isMask()) |
3055 | return nullptr; |
3056 | |
3057 | APInt InvertedLowBitMaskCst = ~*LowBitMaskCst; |
3058 | if (InvertedLowBitMaskCst != *HighBitMaskCst) |
3059 | return nullptr; |
3060 | |
3061 | APInt AlignmentCst = *LowBitMaskCst + 1; |
3062 | |
3063 | if (*BiasCst != AlignmentCst && *BiasCst != *LowBitMaskCst) |
3064 | return nullptr; |
3065 | |
3066 | if (!XBiasedHighBits->hasOneUse()) { |
3067 | // We can't directly return XBiasedHighBits if it is more poisonous. |
3068 | if (*BiasCst == *LowBitMaskCst && impliesPoison(ValAssumedPoison: XBiasedHighBits, V: X)) |
3069 | return XBiasedHighBits; |
3070 | return nullptr; |
3071 | } |
3072 | |
3073 | // FIXME: could we preserve undef's here? |
3074 | Type *Ty = X->getType(); |
3075 | Value *XOffset = Builder.CreateAdd(LHS: X, RHS: ConstantInt::get(Ty, V: *LowBitMaskCst), |
3076 | Name: X->getName() + ".biased" ); |
3077 | Value *R = Builder.CreateAnd(LHS: XOffset, RHS: ConstantInt::get(Ty, V: *HighBitMaskCst)); |
3078 | R->takeName(V: &SI); |
3079 | return R; |
3080 | } |
3081 | |
3082 | namespace { |
3083 | struct DecomposedSelect { |
3084 | Value *Cond = nullptr; |
3085 | Value *TrueVal = nullptr; |
3086 | Value *FalseVal = nullptr; |
3087 | }; |
3088 | } // namespace |
3089 | |
3090 | /// Folds patterns like: |
3091 | /// select c2 (select c1 a b) (select c1 b a) |
3092 | /// into: |
3093 | /// select (xor c1 c2) b a |
3094 | static Instruction * |
3095 | foldSelectOfSymmetricSelect(SelectInst &OuterSelVal, |
3096 | InstCombiner::BuilderTy &Builder) { |
3097 | |
3098 | Value *OuterCond, *InnerCond, *InnerTrueVal, *InnerFalseVal; |
3099 | if (!match( |
3100 | V: &OuterSelVal, |
3101 | P: m_Select(C: m_Value(V&: OuterCond), |
3102 | L: m_OneUse(SubPattern: m_Select(C: m_Value(V&: InnerCond), L: m_Value(V&: InnerTrueVal), |
3103 | R: m_Value(V&: InnerFalseVal))), |
3104 | R: m_OneUse(SubPattern: m_Select(C: m_Deferred(V: InnerCond), |
3105 | L: m_Deferred(V: InnerFalseVal), |
3106 | R: m_Deferred(V: InnerTrueVal)))))) |
3107 | return nullptr; |
3108 | |
3109 | if (OuterCond->getType() != InnerCond->getType()) |
3110 | return nullptr; |
3111 | |
3112 | Value *Xor = Builder.CreateXor(LHS: InnerCond, RHS: OuterCond); |
3113 | return SelectInst::Create(C: Xor, S1: InnerFalseVal, S2: InnerTrueVal); |
3114 | } |
3115 | |
3116 | /// Look for patterns like |
3117 | /// %outer.cond = select i1 %inner.cond, i1 %alt.cond, i1 false |
3118 | /// %inner.sel = select i1 %inner.cond, i8 %inner.sel.t, i8 %inner.sel.f |
3119 | /// %outer.sel = select i1 %outer.cond, i8 %outer.sel.t, i8 %inner.sel |
3120 | /// and rewrite it as |
3121 | /// %inner.sel = select i1 %cond.alternative, i8 %sel.outer.t, i8 %sel.inner.t |
3122 | /// %sel.outer = select i1 %cond.inner, i8 %inner.sel, i8 %sel.inner.f |
3123 | static Instruction *foldNestedSelects(SelectInst &OuterSelVal, |
3124 | InstCombiner::BuilderTy &Builder) { |
3125 | // We must start with a `select`. |
3126 | DecomposedSelect OuterSel; |
3127 | match(V: &OuterSelVal, |
3128 | P: m_Select(C: m_Value(V&: OuterSel.Cond), L: m_Value(V&: OuterSel.TrueVal), |
3129 | R: m_Value(V&: OuterSel.FalseVal))); |
3130 | |
3131 | // Canonicalize inversion of the outermost `select`'s condition. |
3132 | if (match(V: OuterSel.Cond, P: m_Not(V: m_Value(V&: OuterSel.Cond)))) |
3133 | std::swap(a&: OuterSel.TrueVal, b&: OuterSel.FalseVal); |
3134 | |
3135 | // The condition of the outermost select must be an `and`/`or`. |
3136 | if (!match(V: OuterSel.Cond, P: m_c_LogicalOp(L: m_Value(), R: m_Value()))) |
3137 | return nullptr; |
3138 | |
3139 | // Depending on the logical op, inner select might be in different hand. |
3140 | bool IsAndVariant = match(V: OuterSel.Cond, P: m_LogicalAnd()); |
3141 | Value *InnerSelVal = IsAndVariant ? OuterSel.FalseVal : OuterSel.TrueVal; |
3142 | |
3143 | // Profitability check - avoid increasing instruction count. |
3144 | if (none_of(Range: ArrayRef<Value *>({OuterSelVal.getCondition(), InnerSelVal}), |
3145 | P: [](Value *V) { return V->hasOneUse(); })) |
3146 | return nullptr; |
3147 | |
3148 | // The appropriate hand of the outermost `select` must be a select itself. |
3149 | DecomposedSelect InnerSel; |
3150 | if (!match(V: InnerSelVal, |
3151 | P: m_Select(C: m_Value(V&: InnerSel.Cond), L: m_Value(V&: InnerSel.TrueVal), |
3152 | R: m_Value(V&: InnerSel.FalseVal)))) |
3153 | return nullptr; |
3154 | |
3155 | // Canonicalize inversion of the innermost `select`'s condition. |
3156 | if (match(V: InnerSel.Cond, P: m_Not(V: m_Value(V&: InnerSel.Cond)))) |
3157 | std::swap(a&: InnerSel.TrueVal, b&: InnerSel.FalseVal); |
3158 | |
3159 | Value *AltCond = nullptr; |
3160 | auto matchOuterCond = [OuterSel, IsAndVariant, &AltCond](auto m_InnerCond) { |
3161 | // An unsimplified select condition can match both LogicalAnd and LogicalOr |
3162 | // (select true, true, false). Since below we assume that LogicalAnd implies |
3163 | // InnerSel match the FVal and vice versa for LogicalOr, we can't match the |
3164 | // alternative pattern here. |
3165 | return IsAndVariant ? match(OuterSel.Cond, |
3166 | m_c_LogicalAnd(m_InnerCond, m_Value(V&: AltCond))) |
3167 | : match(OuterSel.Cond, |
3168 | m_c_LogicalOr(m_InnerCond, m_Value(V&: AltCond))); |
3169 | }; |
3170 | |
3171 | // Finally, match the condition that was driving the outermost `select`, |
3172 | // it should be a logical operation between the condition that was driving |
3173 | // the innermost `select` (after accounting for the possible inversions |
3174 | // of the condition), and some other condition. |
3175 | if (matchOuterCond(m_Specific(V: InnerSel.Cond))) { |
3176 | // Done! |
3177 | } else if (Value * NotInnerCond; matchOuterCond(m_CombineAnd( |
3178 | L: m_Not(V: m_Specific(V: InnerSel.Cond)), R: m_Value(V&: NotInnerCond)))) { |
3179 | // Done! |
3180 | std::swap(a&: InnerSel.TrueVal, b&: InnerSel.FalseVal); |
3181 | InnerSel.Cond = NotInnerCond; |
3182 | } else // Not the pattern we were looking for. |
3183 | return nullptr; |
3184 | |
3185 | Value *SelInner = Builder.CreateSelect( |
3186 | C: AltCond, True: IsAndVariant ? OuterSel.TrueVal : InnerSel.FalseVal, |
3187 | False: IsAndVariant ? InnerSel.TrueVal : OuterSel.FalseVal); |
3188 | SelInner->takeName(V: InnerSelVal); |
3189 | return SelectInst::Create(C: InnerSel.Cond, |
3190 | S1: IsAndVariant ? SelInner : InnerSel.TrueVal, |
3191 | S2: !IsAndVariant ? SelInner : InnerSel.FalseVal); |
3192 | } |
3193 | |
3194 | /// Return true if V is poison or \p Expected given that ValAssumedPoison is |
3195 | /// already poison. For example, if ValAssumedPoison is `icmp samesign X, 10` |
3196 | /// and V is `icmp ne X, 5`, impliesPoisonOrCond returns true. |
3197 | static bool impliesPoisonOrCond(const Value *ValAssumedPoison, const Value *V, |
3198 | bool Expected) { |
3199 | if (impliesPoison(ValAssumedPoison, V)) |
3200 | return true; |
3201 | |
3202 | // Handle the case that ValAssumedPoison is `icmp samesign pred X, C1` and V |
3203 | // is `icmp pred X, C2`, where C1 is well-defined. |
3204 | if (auto *ICmp = dyn_cast<ICmpInst>(Val: ValAssumedPoison)) { |
3205 | Value *LHS = ICmp->getOperand(i_nocapture: 0); |
3206 | const APInt *RHSC1; |
3207 | const APInt *RHSC2; |
3208 | CmpPredicate Pred; |
3209 | if (ICmp->hasSameSign() && |
3210 | match(V: ICmp->getOperand(i_nocapture: 1), P: m_APIntForbidPoison(Res&: RHSC1)) && |
3211 | match(V, P: m_ICmp(Pred, L: m_Specific(V: LHS), R: m_APIntAllowPoison(Res&: RHSC2)))) { |
3212 | unsigned BitWidth = RHSC1->getBitWidth(); |
3213 | ConstantRange CRX = |
3214 | RHSC1->isNonNegative() |
3215 | ? ConstantRange(APInt::getSignedMinValue(numBits: BitWidth), |
3216 | APInt::getZero(numBits: BitWidth)) |
3217 | : ConstantRange(APInt::getZero(numBits: BitWidth), |
3218 | APInt::getSignedMinValue(numBits: BitWidth)); |
3219 | return CRX.icmp(Pred: Expected ? Pred : ICmpInst::getInverseCmpPredicate(Pred), |
3220 | Other: *RHSC2); |
3221 | } |
3222 | } |
3223 | |
3224 | return false; |
3225 | } |
3226 | |
3227 | Instruction *InstCombinerImpl::foldSelectOfBools(SelectInst &SI) { |
3228 | Value *CondVal = SI.getCondition(); |
3229 | Value *TrueVal = SI.getTrueValue(); |
3230 | Value *FalseVal = SI.getFalseValue(); |
3231 | Type *SelType = SI.getType(); |
3232 | |
3233 | // Avoid potential infinite loops by checking for non-constant condition. |
3234 | // TODO: Can we assert instead by improving canonicalizeSelectToShuffle()? |
3235 | // Scalar select must have simplified? |
3236 | if (!SelType->isIntOrIntVectorTy(BitWidth: 1) || isa<Constant>(Val: CondVal) || |
3237 | TrueVal->getType() != CondVal->getType()) |
3238 | return nullptr; |
3239 | |
3240 | auto *One = ConstantInt::getTrue(Ty: SelType); |
3241 | auto *Zero = ConstantInt::getFalse(Ty: SelType); |
3242 | Value *A, *B, *C, *D; |
3243 | |
3244 | // Folding select to and/or i1 isn't poison safe in general. impliesPoison |
3245 | // checks whether folding it does not convert a well-defined value into |
3246 | // poison. |
3247 | if (match(V: TrueVal, P: m_One())) { |
3248 | if (impliesPoisonOrCond(ValAssumedPoison: FalseVal, V: CondVal, /*Expected=*/false)) { |
3249 | // Change: A = select B, true, C --> A = or B, C |
3250 | return BinaryOperator::CreateOr(V1: CondVal, V2: FalseVal); |
3251 | } |
3252 | |
3253 | if (match(V: CondVal, P: m_OneUse(SubPattern: m_Select(C: m_Value(V&: A), L: m_One(), R: m_Value(V&: B)))) && |
3254 | impliesPoisonOrCond(ValAssumedPoison: FalseVal, V: B, /*Expected=*/false)) { |
3255 | // (A || B) || C --> A || (B | C) |
3256 | return replaceInstUsesWith( |
3257 | I&: SI, V: Builder.CreateLogicalOr(Cond1: A, Cond2: Builder.CreateOr(LHS: B, RHS: FalseVal))); |
3258 | } |
3259 | |
3260 | // (A && B) || (C && B) --> (A || C) && B |
3261 | if (match(V: CondVal, P: m_LogicalAnd(L: m_Value(V&: A), R: m_Value(V&: B))) && |
3262 | match(V: FalseVal, P: m_LogicalAnd(L: m_Value(V&: C), R: m_Value(V&: D))) && |
3263 | (CondVal->hasOneUse() || FalseVal->hasOneUse())) { |
3264 | bool CondLogicAnd = isa<SelectInst>(Val: CondVal); |
3265 | bool FalseLogicAnd = isa<SelectInst>(Val: FalseVal); |
3266 | auto AndFactorization = [&](Value *Common, Value *InnerCond, |
3267 | Value *InnerVal, |
3268 | bool SelFirst = false) -> Instruction * { |
3269 | Value *InnerSel = Builder.CreateSelect(C: InnerCond, True: One, False: InnerVal); |
3270 | if (SelFirst) |
3271 | std::swap(a&: Common, b&: InnerSel); |
3272 | if (FalseLogicAnd || (CondLogicAnd && Common == A)) |
3273 | return SelectInst::Create(C: Common, S1: InnerSel, S2: Zero); |
3274 | else |
3275 | return BinaryOperator::CreateAnd(V1: Common, V2: InnerSel); |
3276 | }; |
3277 | |
3278 | if (A == C) |
3279 | return AndFactorization(A, B, D); |
3280 | if (A == D) |
3281 | return AndFactorization(A, B, C); |
3282 | if (B == C) |
3283 | return AndFactorization(B, A, D); |
3284 | if (B == D) |
3285 | return AndFactorization(B, A, C, CondLogicAnd && FalseLogicAnd); |
3286 | } |
3287 | } |
3288 | |
3289 | if (match(V: FalseVal, P: m_Zero())) { |
3290 | if (impliesPoisonOrCond(ValAssumedPoison: TrueVal, V: CondVal, /*Expected=*/true)) { |
3291 | // Change: A = select B, C, false --> A = and B, C |
3292 | return BinaryOperator::CreateAnd(V1: CondVal, V2: TrueVal); |
3293 | } |
3294 | |
3295 | if (match(V: CondVal, P: m_OneUse(SubPattern: m_Select(C: m_Value(V&: A), L: m_Value(V&: B), R: m_Zero()))) && |
3296 | impliesPoisonOrCond(ValAssumedPoison: TrueVal, V: B, /*Expected=*/true)) { |
3297 | // (A && B) && C --> A && (B & C) |
3298 | return replaceInstUsesWith( |
3299 | I&: SI, V: Builder.CreateLogicalAnd(Cond1: A, Cond2: Builder.CreateAnd(LHS: B, RHS: TrueVal))); |
3300 | } |
3301 | |
3302 | // (A || B) && (C || B) --> (A && C) || B |
3303 | if (match(V: CondVal, P: m_LogicalOr(L: m_Value(V&: A), R: m_Value(V&: B))) && |
3304 | match(V: TrueVal, P: m_LogicalOr(L: m_Value(V&: C), R: m_Value(V&: D))) && |
3305 | (CondVal->hasOneUse() || TrueVal->hasOneUse())) { |
3306 | bool CondLogicOr = isa<SelectInst>(Val: CondVal); |
3307 | bool TrueLogicOr = isa<SelectInst>(Val: TrueVal); |
3308 | auto OrFactorization = [&](Value *Common, Value *InnerCond, |
3309 | Value *InnerVal, |
3310 | bool SelFirst = false) -> Instruction * { |
3311 | Value *InnerSel = Builder.CreateSelect(C: InnerCond, True: InnerVal, False: Zero); |
3312 | if (SelFirst) |
3313 | std::swap(a&: Common, b&: InnerSel); |
3314 | if (TrueLogicOr || (CondLogicOr && Common == A)) |
3315 | return SelectInst::Create(C: Common, S1: One, S2: InnerSel); |
3316 | else |
3317 | return BinaryOperator::CreateOr(V1: Common, V2: InnerSel); |
3318 | }; |
3319 | |
3320 | if (A == C) |
3321 | return OrFactorization(A, B, D); |
3322 | if (A == D) |
3323 | return OrFactorization(A, B, C); |
3324 | if (B == C) |
3325 | return OrFactorization(B, A, D); |
3326 | if (B == D) |
3327 | return OrFactorization(B, A, C, CondLogicOr && TrueLogicOr); |
3328 | } |
3329 | } |
3330 | |
3331 | // We match the "full" 0 or 1 constant here to avoid a potential infinite |
3332 | // loop with vectors that may have undefined/poison elements. |
3333 | // select a, false, b -> select !a, b, false |
3334 | if (match(V: TrueVal, P: m_Specific(V: Zero))) { |
3335 | Value *NotCond = Builder.CreateNot(V: CondVal, Name: "not." + CondVal->getName()); |
3336 | return SelectInst::Create(C: NotCond, S1: FalseVal, S2: Zero); |
3337 | } |
3338 | // select a, b, true -> select !a, true, b |
3339 | if (match(V: FalseVal, P: m_Specific(V: One))) { |
3340 | Value *NotCond = Builder.CreateNot(V: CondVal, Name: "not." + CondVal->getName()); |
3341 | return SelectInst::Create(C: NotCond, S1: One, S2: TrueVal); |
3342 | } |
3343 | |
3344 | // DeMorgan in select form: !a && !b --> !(a || b) |
3345 | // select !a, !b, false --> not (select a, true, b) |
3346 | if (match(V: &SI, P: m_LogicalAnd(L: m_Not(V: m_Value(V&: A)), R: m_Not(V: m_Value(V&: B)))) && |
3347 | (CondVal->hasOneUse() || TrueVal->hasOneUse()) && |
3348 | !match(V: A, P: m_ConstantExpr()) && !match(V: B, P: m_ConstantExpr())) |
3349 | return BinaryOperator::CreateNot(Op: Builder.CreateSelect(C: A, True: One, False: B)); |
3350 | |
3351 | // DeMorgan in select form: !a || !b --> !(a && b) |
3352 | // select !a, true, !b --> not (select a, b, false) |
3353 | if (match(V: &SI, P: m_LogicalOr(L: m_Not(V: m_Value(V&: A)), R: m_Not(V: m_Value(V&: B)))) && |
3354 | (CondVal->hasOneUse() || FalseVal->hasOneUse()) && |
3355 | !match(V: A, P: m_ConstantExpr()) && !match(V: B, P: m_ConstantExpr())) |
3356 | return BinaryOperator::CreateNot(Op: Builder.CreateSelect(C: A, True: B, False: Zero)); |
3357 | |
3358 | // select (select a, true, b), true, b -> select a, true, b |
3359 | if (match(V: CondVal, P: m_Select(C: m_Value(V&: A), L: m_One(), R: m_Value(V&: B))) && |
3360 | match(V: TrueVal, P: m_One()) && match(V: FalseVal, P: m_Specific(V: B))) |
3361 | return replaceOperand(I&: SI, OpNum: 0, V: A); |
3362 | // select (select a, b, false), b, false -> select a, b, false |
3363 | if (match(V: CondVal, P: m_Select(C: m_Value(V&: A), L: m_Value(V&: B), R: m_Zero())) && |
3364 | match(V: TrueVal, P: m_Specific(V: B)) && match(V: FalseVal, P: m_Zero())) |
3365 | return replaceOperand(I&: SI, OpNum: 0, V: A); |
3366 | |
3367 | // ~(A & B) & (A | B) --> A ^ B |
3368 | if (match(V: &SI, P: m_c_LogicalAnd(L: m_Not(V: m_LogicalAnd(L: m_Value(V&: A), R: m_Value(V&: B))), |
3369 | R: m_c_LogicalOr(L: m_Deferred(V: A), R: m_Deferred(V: B))))) |
3370 | return BinaryOperator::CreateXor(V1: A, V2: B); |
3371 | |
3372 | // select (~a | c), a, b -> select a, (select c, true, b), false |
3373 | if (match(V: CondVal, |
3374 | P: m_OneUse(SubPattern: m_c_Or(L: m_Not(V: m_Specific(V: TrueVal)), R: m_Value(V&: C))))) { |
3375 | Value *OrV = Builder.CreateSelect(C, True: One, False: FalseVal); |
3376 | return SelectInst::Create(C: TrueVal, S1: OrV, S2: Zero); |
3377 | } |
3378 | // select (c & b), a, b -> select b, (select ~c, true, a), false |
3379 | if (match(V: CondVal, P: m_OneUse(SubPattern: m_c_And(L: m_Value(V&: C), R: m_Specific(V: FalseVal))))) { |
3380 | if (Value *NotC = getFreelyInverted(V: C, WillInvertAllUses: C->hasOneUse(), Builder: &Builder)) { |
3381 | Value *OrV = Builder.CreateSelect(C: NotC, True: One, False: TrueVal); |
3382 | return SelectInst::Create(C: FalseVal, S1: OrV, S2: Zero); |
3383 | } |
3384 | } |
3385 | // select (a | c), a, b -> select a, true, (select ~c, b, false) |
3386 | if (match(V: CondVal, P: m_OneUse(SubPattern: m_c_Or(L: m_Specific(V: TrueVal), R: m_Value(V&: C))))) { |
3387 | if (Value *NotC = getFreelyInverted(V: C, WillInvertAllUses: C->hasOneUse(), Builder: &Builder)) { |
3388 | Value *AndV = Builder.CreateSelect(C: NotC, True: FalseVal, False: Zero); |
3389 | return SelectInst::Create(C: TrueVal, S1: One, S2: AndV); |
3390 | } |
3391 | } |
3392 | // select (c & ~b), a, b -> select b, true, (select c, a, false) |
3393 | if (match(V: CondVal, |
3394 | P: m_OneUse(SubPattern: m_c_And(L: m_Value(V&: C), R: m_Not(V: m_Specific(V: FalseVal)))))) { |
3395 | Value *AndV = Builder.CreateSelect(C, True: TrueVal, False: Zero); |
3396 | return SelectInst::Create(C: FalseVal, S1: One, S2: AndV); |
3397 | } |
3398 | |
3399 | if (match(V: FalseVal, P: m_Zero()) || match(V: TrueVal, P: m_One())) { |
3400 | Use *Y = nullptr; |
3401 | bool IsAnd = match(V: FalseVal, P: m_Zero()) ? true : false; |
3402 | Value *Op1 = IsAnd ? TrueVal : FalseVal; |
3403 | if (isCheckForZeroAndMulWithOverflow(Op0: CondVal, Op1, IsAnd, Y)) { |
3404 | auto *FI = new FreezeInst(*Y, (*Y)->getName() + ".fr" ); |
3405 | InsertNewInstBefore(New: FI, Old: cast<Instruction>(Val: Y->getUser())->getIterator()); |
3406 | replaceUse(U&: *Y, NewValue: FI); |
3407 | return replaceInstUsesWith(I&: SI, V: Op1); |
3408 | } |
3409 | |
3410 | if (auto *V = foldBooleanAndOr(LHS: CondVal, RHS: Op1, I&: SI, IsAnd, |
3411 | /*IsLogical=*/true)) |
3412 | return replaceInstUsesWith(I&: SI, V); |
3413 | } |
3414 | |
3415 | // select (a || b), c, false -> select a, c, false |
3416 | // select c, (a || b), false -> select c, a, false |
3417 | // if c implies that b is false. |
3418 | if (match(V: CondVal, P: m_LogicalOr(L: m_Value(V&: A), R: m_Value(V&: B))) && |
3419 | match(V: FalseVal, P: m_Zero())) { |
3420 | std::optional<bool> Res = isImpliedCondition(LHS: TrueVal, RHS: B, DL); |
3421 | if (Res && *Res == false) |
3422 | return replaceOperand(I&: SI, OpNum: 0, V: A); |
3423 | } |
3424 | if (match(V: TrueVal, P: m_LogicalOr(L: m_Value(V&: A), R: m_Value(V&: B))) && |
3425 | match(V: FalseVal, P: m_Zero())) { |
3426 | std::optional<bool> Res = isImpliedCondition(LHS: CondVal, RHS: B, DL); |
3427 | if (Res && *Res == false) |
3428 | return replaceOperand(I&: SI, OpNum: 1, V: A); |
3429 | } |
3430 | // select c, true, (a && b) -> select c, true, a |
3431 | // select (a && b), true, c -> select a, true, c |
3432 | // if c = false implies that b = true |
3433 | if (match(V: TrueVal, P: m_One()) && |
3434 | match(V: FalseVal, P: m_LogicalAnd(L: m_Value(V&: A), R: m_Value(V&: B)))) { |
3435 | std::optional<bool> Res = isImpliedCondition(LHS: CondVal, RHS: B, DL, LHSIsTrue: false); |
3436 | if (Res && *Res == true) |
3437 | return replaceOperand(I&: SI, OpNum: 2, V: A); |
3438 | } |
3439 | if (match(V: CondVal, P: m_LogicalAnd(L: m_Value(V&: A), R: m_Value(V&: B))) && |
3440 | match(V: TrueVal, P: m_One())) { |
3441 | std::optional<bool> Res = isImpliedCondition(LHS: FalseVal, RHS: B, DL, LHSIsTrue: false); |
3442 | if (Res && *Res == true) |
3443 | return replaceOperand(I&: SI, OpNum: 0, V: A); |
3444 | } |
3445 | |
3446 | if (match(V: TrueVal, P: m_One())) { |
3447 | Value *C; |
3448 | |
3449 | // (C && A) || (!C && B) --> sel C, A, B |
3450 | // (A && C) || (!C && B) --> sel C, A, B |
3451 | // (C && A) || (B && !C) --> sel C, A, B |
3452 | // (A && C) || (B && !C) --> sel C, A, B (may require freeze) |
3453 | if (match(V: FalseVal, P: m_c_LogicalAnd(L: m_Not(V: m_Value(V&: C)), R: m_Value(V&: B))) && |
3454 | match(V: CondVal, P: m_c_LogicalAnd(L: m_Specific(V: C), R: m_Value(V&: A)))) { |
3455 | auto *SelCond = dyn_cast<SelectInst>(Val: CondVal); |
3456 | auto *SelFVal = dyn_cast<SelectInst>(Val: FalseVal); |
3457 | bool MayNeedFreeze = SelCond && SelFVal && |
3458 | match(V: SelFVal->getTrueValue(), |
3459 | P: m_Not(V: m_Specific(V: SelCond->getTrueValue()))); |
3460 | if (MayNeedFreeze) |
3461 | C = Builder.CreateFreeze(V: C); |
3462 | return SelectInst::Create(C, S1: A, S2: B); |
3463 | } |
3464 | |
3465 | // (!C && A) || (C && B) --> sel C, B, A |
3466 | // (A && !C) || (C && B) --> sel C, B, A |
3467 | // (!C && A) || (B && C) --> sel C, B, A |
3468 | // (A && !C) || (B && C) --> sel C, B, A (may require freeze) |
3469 | if (match(V: CondVal, P: m_c_LogicalAnd(L: m_Not(V: m_Value(V&: C)), R: m_Value(V&: A))) && |
3470 | match(V: FalseVal, P: m_c_LogicalAnd(L: m_Specific(V: C), R: m_Value(V&: B)))) { |
3471 | auto *SelCond = dyn_cast<SelectInst>(Val: CondVal); |
3472 | auto *SelFVal = dyn_cast<SelectInst>(Val: FalseVal); |
3473 | bool MayNeedFreeze = SelCond && SelFVal && |
3474 | match(V: SelCond->getTrueValue(), |
3475 | P: m_Not(V: m_Specific(V: SelFVal->getTrueValue()))); |
3476 | if (MayNeedFreeze) |
3477 | C = Builder.CreateFreeze(V: C); |
3478 | return SelectInst::Create(C, S1: B, S2: A); |
3479 | } |
3480 | } |
3481 | |
3482 | return nullptr; |
3483 | } |
3484 | |
3485 | // Return true if we can safely remove the select instruction for std::bit_ceil |
3486 | // pattern. |
3487 | static bool isSafeToRemoveBitCeilSelect(ICmpInst::Predicate Pred, Value *Cond0, |
3488 | const APInt *Cond1, Value *CtlzOp, |
3489 | unsigned BitWidth, |
3490 | bool &ShouldDropNoWrap) { |
3491 | // The challenge in recognizing std::bit_ceil(X) is that the operand is used |
3492 | // for the CTLZ proper and select condition, each possibly with some |
3493 | // operation like add and sub. |
3494 | // |
3495 | // Our aim is to make sure that -ctlz & (BitWidth - 1) == 0 even when the |
3496 | // select instruction would select 1, which allows us to get rid of the select |
3497 | // instruction. |
3498 | // |
3499 | // To see if we can do so, we do some symbolic execution with ConstantRange. |
3500 | // Specifically, we compute the range of values that Cond0 could take when |
3501 | // Cond == false. Then we successively transform the range until we obtain |
3502 | // the range of values that CtlzOp could take. |
3503 | // |
3504 | // Conceptually, we follow the def-use chain backward from Cond0 while |
3505 | // transforming the range for Cond0 until we meet the common ancestor of Cond0 |
3506 | // and CtlzOp. Then we follow the def-use chain forward until we obtain the |
3507 | // range for CtlzOp. That said, we only follow at most one ancestor from |
3508 | // Cond0. Likewise, we only follow at most one ancestor from CtrlOp. |
3509 | |
3510 | ConstantRange CR = ConstantRange::makeExactICmpRegion( |
3511 | Pred: CmpInst::getInversePredicate(pred: Pred), Other: *Cond1); |
3512 | |
3513 | ShouldDropNoWrap = false; |
3514 | |
3515 | // Match the operation that's used to compute CtlzOp from CommonAncestor. If |
3516 | // CtlzOp == CommonAncestor, return true as no operation is needed. If a |
3517 | // match is found, execute the operation on CR, update CR, and return true. |
3518 | // Otherwise, return false. |
3519 | auto MatchForward = [&](Value *CommonAncestor) { |
3520 | const APInt *C = nullptr; |
3521 | if (CtlzOp == CommonAncestor) |
3522 | return true; |
3523 | if (match(V: CtlzOp, P: m_Add(L: m_Specific(V: CommonAncestor), R: m_APInt(Res&: C)))) { |
3524 | ShouldDropNoWrap = true; |
3525 | CR = CR.add(Other: *C); |
3526 | return true; |
3527 | } |
3528 | if (match(V: CtlzOp, P: m_Sub(L: m_APInt(Res&: C), R: m_Specific(V: CommonAncestor)))) { |
3529 | ShouldDropNoWrap = true; |
3530 | CR = ConstantRange(*C).sub(Other: CR); |
3531 | return true; |
3532 | } |
3533 | if (match(V: CtlzOp, P: m_Not(V: m_Specific(V: CommonAncestor)))) { |
3534 | CR = CR.binaryNot(); |
3535 | return true; |
3536 | } |
3537 | return false; |
3538 | }; |
3539 | |
3540 | const APInt *C = nullptr; |
3541 | Value *CommonAncestor; |
3542 | if (MatchForward(Cond0)) { |
3543 | // Cond0 is either CtlzOp or CtlzOp's parent. CR has been updated. |
3544 | } else if (match(V: Cond0, P: m_Add(L: m_Value(V&: CommonAncestor), R: m_APInt(Res&: C)))) { |
3545 | CR = CR.sub(Other: *C); |
3546 | if (!MatchForward(CommonAncestor)) |
3547 | return false; |
3548 | // Cond0's parent is either CtlzOp or CtlzOp's parent. CR has been updated. |
3549 | } else { |
3550 | return false; |
3551 | } |
3552 | |
3553 | // Return true if all the values in the range are either 0 or negative (if |
3554 | // treated as signed). We do so by evaluating: |
3555 | // |
3556 | // CR - 1 u>= (1 << BitWidth) - 1. |
3557 | APInt IntMax = APInt::getSignMask(BitWidth) - 1; |
3558 | CR = CR.sub(Other: APInt(BitWidth, 1)); |
3559 | return CR.icmp(Pred: ICmpInst::ICMP_UGE, Other: IntMax); |
3560 | } |
3561 | |
3562 | // Transform the std::bit_ceil(X) pattern like: |
3563 | // |
3564 | // %dec = add i32 %x, -1 |
3565 | // %ctlz = tail call i32 @llvm.ctlz.i32(i32 %dec, i1 false) |
3566 | // %sub = sub i32 32, %ctlz |
3567 | // %shl = shl i32 1, %sub |
3568 | // %ugt = icmp ugt i32 %x, 1 |
3569 | // %sel = select i1 %ugt, i32 %shl, i32 1 |
3570 | // |
3571 | // into: |
3572 | // |
3573 | // %dec = add i32 %x, -1 |
3574 | // %ctlz = tail call i32 @llvm.ctlz.i32(i32 %dec, i1 false) |
3575 | // %neg = sub i32 0, %ctlz |
3576 | // %masked = and i32 %ctlz, 31 |
3577 | // %shl = shl i32 1, %sub |
3578 | // |
3579 | // Note that the select is optimized away while the shift count is masked with |
3580 | // 31. We handle some variations of the input operand like std::bit_ceil(X + |
3581 | // 1). |
3582 | static Instruction *foldBitCeil(SelectInst &SI, IRBuilderBase &Builder, |
3583 | InstCombinerImpl &IC) { |
3584 | Type *SelType = SI.getType(); |
3585 | unsigned BitWidth = SelType->getScalarSizeInBits(); |
3586 | |
3587 | Value *FalseVal = SI.getFalseValue(); |
3588 | Value *TrueVal = SI.getTrueValue(); |
3589 | CmpPredicate Pred; |
3590 | const APInt *Cond1; |
3591 | Value *Cond0, *Ctlz, *CtlzOp; |
3592 | if (!match(V: SI.getCondition(), P: m_ICmp(Pred, L: m_Value(V&: Cond0), R: m_APInt(Res&: Cond1)))) |
3593 | return nullptr; |
3594 | |
3595 | if (match(V: TrueVal, P: m_One())) { |
3596 | std::swap(a&: FalseVal, b&: TrueVal); |
3597 | Pred = CmpInst::getInversePredicate(pred: Pred); |
3598 | } |
3599 | |
3600 | bool ShouldDropNoWrap; |
3601 | |
3602 | if (!match(V: FalseVal, P: m_One()) || |
3603 | !match(V: TrueVal, |
3604 | P: m_OneUse(SubPattern: m_Shl(L: m_One(), R: m_OneUse(SubPattern: m_Sub(L: m_SpecificInt(V: BitWidth), |
3605 | R: m_Value(V&: Ctlz)))))) || |
3606 | !match(V: Ctlz, P: m_Intrinsic<Intrinsic::ctlz>(Op0: m_Value(V&: CtlzOp), Op1: m_Value())) || |
3607 | !isSafeToRemoveBitCeilSelect(Pred, Cond0, Cond1, CtlzOp, BitWidth, |
3608 | ShouldDropNoWrap)) |
3609 | return nullptr; |
3610 | |
3611 | if (ShouldDropNoWrap) { |
3612 | cast<Instruction>(Val: CtlzOp)->setHasNoUnsignedWrap(false); |
3613 | cast<Instruction>(Val: CtlzOp)->setHasNoSignedWrap(false); |
3614 | } |
3615 | |
3616 | // Build 1 << (-CTLZ & (BitWidth-1)). The negation likely corresponds to a |
3617 | // single hardware instruction as opposed to BitWidth - CTLZ, where BitWidth |
3618 | // is an integer constant. Masking with BitWidth-1 comes free on some |
3619 | // hardware as part of the shift instruction. |
3620 | |
3621 | // Drop range attributes and re-infer them in the next iteration. |
3622 | cast<Instruction>(Val: Ctlz)->dropPoisonGeneratingAnnotations(); |
3623 | // Set is_zero_poison to false and re-infer them in the next iteration. |
3624 | cast<Instruction>(Val: Ctlz)->setOperand(i: 1, Val: Builder.getFalse()); |
3625 | IC.addToWorklist(I: cast<Instruction>(Val: Ctlz)); |
3626 | Value *Neg = Builder.CreateNeg(V: Ctlz); |
3627 | Value *Masked = |
3628 | Builder.CreateAnd(LHS: Neg, RHS: ConstantInt::get(Ty: SelType, V: BitWidth - 1)); |
3629 | return BinaryOperator::Create(Op: Instruction::Shl, S1: ConstantInt::get(Ty: SelType, V: 1), |
3630 | S2: Masked); |
3631 | } |
3632 | |
3633 | // This function tries to fold the following operations: |
3634 | // (x < y) ? -1 : zext(x != y) |
3635 | // (x < y) ? -1 : zext(x > y) |
3636 | // (x > y) ? 1 : sext(x != y) |
3637 | // (x > y) ? 1 : sext(x < y) |
3638 | // Into ucmp/scmp(x, y), where signedness is determined by the signedness |
3639 | // of the comparison in the original sequence. |
3640 | Instruction *InstCombinerImpl::foldSelectToCmp(SelectInst &SI) { |
3641 | Value *TV = SI.getTrueValue(); |
3642 | Value *FV = SI.getFalseValue(); |
3643 | |
3644 | CmpPredicate Pred; |
3645 | Value *LHS, *RHS; |
3646 | if (!match(V: SI.getCondition(), P: m_ICmp(Pred, L: m_Value(V&: LHS), R: m_Value(V&: RHS)))) |
3647 | return nullptr; |
3648 | |
3649 | if (!LHS->getType()->isIntOrIntVectorTy()) |
3650 | return nullptr; |
3651 | |
3652 | // If there is no -1, 0 or 1 at TV, then invert the select statement and try |
3653 | // to canonicalize to one of the forms above |
3654 | if (!isa<Constant>(Val: TV)) { |
3655 | if (!isa<Constant>(Val: FV)) |
3656 | return nullptr; |
3657 | Pred = ICmpInst::getInverseCmpPredicate(Pred); |
3658 | std::swap(a&: TV, b&: FV); |
3659 | } |
3660 | |
3661 | if (ICmpInst::isNonStrictPredicate(predicate: Pred)) { |
3662 | if (Constant *C = dyn_cast<Constant>(Val: RHS)) { |
3663 | auto FlippedPredAndConst = |
3664 | getFlippedStrictnessPredicateAndConstant(Pred, C); |
3665 | if (!FlippedPredAndConst) |
3666 | return nullptr; |
3667 | Pred = FlippedPredAndConst->first; |
3668 | RHS = FlippedPredAndConst->second; |
3669 | } else { |
3670 | return nullptr; |
3671 | } |
3672 | } |
3673 | |
3674 | // Try to swap operands and the predicate. We need to be careful when doing |
3675 | // so because two of the patterns have opposite predicates, so use the |
3676 | // constant inside select to determine if swapping operands would be |
3677 | // beneficial to us. |
3678 | if ((ICmpInst::isGT(P: Pred) && match(V: TV, P: m_AllOnes())) || |
3679 | (ICmpInst::isLT(P: Pred) && match(V: TV, P: m_One()))) { |
3680 | Pred = ICmpInst::getSwappedPredicate(pred: Pred); |
3681 | std::swap(a&: LHS, b&: RHS); |
3682 | } |
3683 | bool IsSigned = ICmpInst::isSigned(predicate: Pred); |
3684 | |
3685 | bool Replace = false; |
3686 | CmpPredicate ExtendedCmpPredicate; |
3687 | // (x < y) ? -1 : zext(x != y) |
3688 | // (x < y) ? -1 : zext(x > y) |
3689 | if (ICmpInst::isLT(P: Pred) && match(V: TV, P: m_AllOnes()) && |
3690 | match(V: FV, P: m_ZExt(Op: m_c_ICmp(Pred&: ExtendedCmpPredicate, L: m_Specific(V: LHS), |
3691 | R: m_Specific(V: RHS)))) && |
3692 | (ExtendedCmpPredicate == ICmpInst::ICMP_NE || |
3693 | ICmpInst::getSwappedPredicate(pred: ExtendedCmpPredicate) == Pred)) |
3694 | Replace = true; |
3695 | |
3696 | // (x > y) ? 1 : sext(x != y) |
3697 | // (x > y) ? 1 : sext(x < y) |
3698 | if (ICmpInst::isGT(P: Pred) && match(V: TV, P: m_One()) && |
3699 | match(V: FV, P: m_SExt(Op: m_c_ICmp(Pred&: ExtendedCmpPredicate, L: m_Specific(V: LHS), |
3700 | R: m_Specific(V: RHS)))) && |
3701 | (ExtendedCmpPredicate == ICmpInst::ICMP_NE || |
3702 | ICmpInst::getSwappedPredicate(pred: ExtendedCmpPredicate) == Pred)) |
3703 | Replace = true; |
3704 | |
3705 | // (x == y) ? 0 : (x > y ? 1 : -1) |
3706 | CmpPredicate FalseBranchSelectPredicate; |
3707 | const APInt *InnerTV, *InnerFV; |
3708 | if (Pred == ICmpInst::ICMP_EQ && match(V: TV, P: m_Zero()) && |
3709 | match(V: FV, P: m_Select(C: m_c_ICmp(Pred&: FalseBranchSelectPredicate, L: m_Specific(V: LHS), |
3710 | R: m_Specific(V: RHS)), |
3711 | L: m_APInt(Res&: InnerTV), R: m_APInt(Res&: InnerFV)))) { |
3712 | if (!ICmpInst::isGT(P: FalseBranchSelectPredicate)) { |
3713 | FalseBranchSelectPredicate = |
3714 | ICmpInst::getSwappedPredicate(pred: FalseBranchSelectPredicate); |
3715 | std::swap(a&: LHS, b&: RHS); |
3716 | } |
3717 | |
3718 | if (!InnerTV->isOne()) { |
3719 | std::swap(a&: InnerTV, b&: InnerFV); |
3720 | std::swap(a&: LHS, b&: RHS); |
3721 | } |
3722 | |
3723 | if (ICmpInst::isGT(P: FalseBranchSelectPredicate) && InnerTV->isOne() && |
3724 | InnerFV->isAllOnes()) { |
3725 | IsSigned = ICmpInst::isSigned(predicate: FalseBranchSelectPredicate); |
3726 | Replace = true; |
3727 | } |
3728 | } |
3729 | |
3730 | Intrinsic::ID IID = IsSigned ? Intrinsic::scmp : Intrinsic::ucmp; |
3731 | if (Replace) |
3732 | return replaceInstUsesWith( |
3733 | I&: SI, V: Builder.CreateIntrinsic(RetTy: SI.getType(), ID: IID, Args: {LHS, RHS})); |
3734 | return nullptr; |
3735 | } |
3736 | |
3737 | bool InstCombinerImpl::fmulByZeroIsZero(Value *MulVal, FastMathFlags FMF, |
3738 | const Instruction *CtxI) const { |
3739 | KnownFPClass Known = computeKnownFPClass(Val: MulVal, FMF, Interested: fcNegative, CtxI); |
3740 | |
3741 | return Known.isKnownNeverNaN() && Known.isKnownNeverInfinity() && |
3742 | (FMF.noSignedZeros() || Known.signBitIsZeroOrNaN()); |
3743 | } |
3744 | |
3745 | static bool matchFMulByZeroIfResultEqZero(InstCombinerImpl &IC, Value *Cmp0, |
3746 | Value *Cmp1, Value *TrueVal, |
3747 | Value *FalseVal, Instruction &CtxI, |
3748 | bool SelectIsNSZ) { |
3749 | Value *MulRHS; |
3750 | if (match(V: Cmp1, P: m_PosZeroFP()) && |
3751 | match(V: TrueVal, P: m_c_FMul(L: m_Specific(V: Cmp0), R: m_Value(V&: MulRHS)))) { |
3752 | FastMathFlags FMF = cast<FPMathOperator>(Val: TrueVal)->getFastMathFlags(); |
3753 | // nsz must be on the select, it must be ignored on the multiply. We |
3754 | // need nnan and ninf on the multiply for the other value. |
3755 | FMF.setNoSignedZeros(SelectIsNSZ); |
3756 | return IC.fmulByZeroIsZero(MulVal: MulRHS, FMF, CtxI: &CtxI); |
3757 | } |
3758 | |
3759 | return false; |
3760 | } |
3761 | |
3762 | /// Check whether the KnownBits of a select arm may be affected by the |
3763 | /// select condition. |
3764 | static bool hasAffectedValue(Value *V, SmallPtrSetImpl<Value *> &Affected, |
3765 | unsigned Depth) { |
3766 | if (Depth == MaxAnalysisRecursionDepth) |
3767 | return false; |
3768 | |
3769 | // Ignore the case where the select arm itself is affected. These cases |
3770 | // are handled more efficiently by other optimizations. |
3771 | if (Depth != 0 && Affected.contains(Ptr: V)) |
3772 | return true; |
3773 | |
3774 | if (auto *I = dyn_cast<Instruction>(Val: V)) { |
3775 | if (isa<PHINode>(Val: I)) { |
3776 | if (Depth == MaxAnalysisRecursionDepth - 1) |
3777 | return false; |
3778 | Depth = MaxAnalysisRecursionDepth - 2; |
3779 | } |
3780 | return any_of(Range: I->operands(), P: [&](Value *Op) { |
3781 | return Op->getType()->isIntOrIntVectorTy() && |
3782 | hasAffectedValue(V: Op, Affected, Depth: Depth + 1); |
3783 | }); |
3784 | } |
3785 | |
3786 | return false; |
3787 | } |
3788 | |
3789 | // This transformation enables the possibility of transforming fcmp + sel into |
3790 | // a fmaxnum/fminnum intrinsic. |
3791 | static Value *foldSelectIntoAddConstant(SelectInst &SI, |
3792 | InstCombiner::BuilderTy &Builder) { |
3793 | // Do this transformation only when select instruction gives NaN and NSZ |
3794 | // guarantee. |
3795 | auto *SIFOp = dyn_cast<FPMathOperator>(Val: &SI); |
3796 | if (!SIFOp || !SIFOp->hasNoSignedZeros() || !SIFOp->hasNoNaNs()) |
3797 | return nullptr; |
3798 | |
3799 | auto TryFoldIntoAddConstant = |
3800 | [&Builder, &SI](CmpInst::Predicate Pred, Value *X, Value *Z, |
3801 | Instruction *FAdd, Constant *C, bool Swapped) -> Value * { |
3802 | // Only these relational predicates can be transformed into maxnum/minnum |
3803 | // intrinsic. |
3804 | if (!CmpInst::isRelational(P: Pred) || !match(V: Z, P: m_AnyZeroFP())) |
3805 | return nullptr; |
3806 | |
3807 | if (!match(V: FAdd, P: m_FAdd(L: m_Specific(V: X), R: m_Specific(V: C)))) |
3808 | return nullptr; |
3809 | |
3810 | Value *NewSelect = Builder.CreateSelect(C: SI.getCondition(), True: Swapped ? Z : X, |
3811 | False: Swapped ? X : Z, Name: "" , MDFrom: &SI); |
3812 | NewSelect->takeName(V: &SI); |
3813 | |
3814 | Value *NewFAdd = Builder.CreateFAdd(L: NewSelect, R: C); |
3815 | NewFAdd->takeName(V: FAdd); |
3816 | |
3817 | // Propagate FastMath flags |
3818 | FastMathFlags SelectFMF = SI.getFastMathFlags(); |
3819 | FastMathFlags FAddFMF = FAdd->getFastMathFlags(); |
3820 | FastMathFlags NewFMF = FastMathFlags::intersectRewrite(LHS: SelectFMF, RHS: FAddFMF) | |
3821 | FastMathFlags::unionValue(LHS: SelectFMF, RHS: FAddFMF); |
3822 | cast<Instruction>(Val: NewFAdd)->setFastMathFlags(NewFMF); |
3823 | cast<Instruction>(Val: NewSelect)->setFastMathFlags(NewFMF); |
3824 | |
3825 | return NewFAdd; |
3826 | }; |
3827 | |
3828 | // select((fcmp Pred, X, 0), (fadd X, C), C) |
3829 | // => fadd((select (fcmp Pred, X, 0), X, 0), C) |
3830 | // |
3831 | // Pred := OGT, OGE, OLT, OLE, UGT, UGE, ULT, and ULE |
3832 | Instruction *FAdd; |
3833 | Constant *C; |
3834 | Value *X, *Z; |
3835 | CmpPredicate Pred; |
3836 | |
3837 | // Note: OneUse check for `Cmp` is necessary because it makes sure that other |
3838 | // InstCombine folds don't undo this transformation and cause an infinite |
3839 | // loop. Furthermore, it could also increase the operation count. |
3840 | if (match(V: &SI, P: m_Select(C: m_OneUse(SubPattern: m_FCmp(Pred, L: m_Value(V&: X), R: m_Value(V&: Z))), |
3841 | L: m_OneUse(SubPattern: m_Instruction(I&: FAdd)), R: m_Constant(C)))) |
3842 | return TryFoldIntoAddConstant(Pred, X, Z, FAdd, C, /*Swapped=*/false); |
3843 | |
3844 | if (match(V: &SI, P: m_Select(C: m_OneUse(SubPattern: m_FCmp(Pred, L: m_Value(V&: X), R: m_Value(V&: Z))), |
3845 | L: m_Constant(C), R: m_OneUse(SubPattern: m_Instruction(I&: FAdd))))) |
3846 | return TryFoldIntoAddConstant(Pred, X, Z, FAdd, C, /*Swapped=*/true); |
3847 | |
3848 | return nullptr; |
3849 | } |
3850 | |
3851 | static Value *foldSelectBitTest(SelectInst &Sel, Value *CondVal, Value *TrueVal, |
3852 | Value *FalseVal, |
3853 | InstCombiner::BuilderTy &Builder, |
3854 | const SimplifyQuery &SQ) { |
3855 | // If this is a vector select, we need a vector compare. |
3856 | Type *SelType = Sel.getType(); |
3857 | if (SelType->isVectorTy() != CondVal->getType()->isVectorTy()) |
3858 | return nullptr; |
3859 | |
3860 | Value *V; |
3861 | APInt AndMask; |
3862 | bool CreateAnd = false; |
3863 | CmpPredicate Pred; |
3864 | Value *CmpLHS, *CmpRHS; |
3865 | |
3866 | if (match(V: CondVal, P: m_ICmp(Pred, L: m_Value(V&: CmpLHS), R: m_Value(V&: CmpRHS)))) { |
3867 | if (ICmpInst::isEquality(P: Pred)) { |
3868 | if (!match(V: CmpRHS, P: m_Zero())) |
3869 | return nullptr; |
3870 | |
3871 | V = CmpLHS; |
3872 | const APInt *AndRHS; |
3873 | if (!match(V: CmpLHS, P: m_And(L: m_Value(), R: m_Power2(V&: AndRHS)))) |
3874 | return nullptr; |
3875 | |
3876 | AndMask = *AndRHS; |
3877 | } else if (auto Res = decomposeBitTestICmp(LHS: CmpLHS, RHS: CmpRHS, Pred)) { |
3878 | assert(ICmpInst::isEquality(Res->Pred) && "Not equality test?" ); |
3879 | AndMask = Res->Mask; |
3880 | V = Res->X; |
3881 | KnownBits Known = computeKnownBits(V, Q: SQ.getWithInstruction(I: &Sel)); |
3882 | AndMask &= Known.getMaxValue(); |
3883 | if (!AndMask.isPowerOf2()) |
3884 | return nullptr; |
3885 | |
3886 | Pred = Res->Pred; |
3887 | CreateAnd = true; |
3888 | } else { |
3889 | return nullptr; |
3890 | } |
3891 | } else if (auto *Trunc = dyn_cast<TruncInst>(Val: CondVal)) { |
3892 | V = Trunc->getOperand(i_nocapture: 0); |
3893 | AndMask = APInt(V->getType()->getScalarSizeInBits(), 1); |
3894 | Pred = ICmpInst::ICMP_NE; |
3895 | CreateAnd = !Trunc->hasNoUnsignedWrap(); |
3896 | } else { |
3897 | return nullptr; |
3898 | } |
3899 | |
3900 | if (Pred == ICmpInst::ICMP_NE) |
3901 | std::swap(a&: TrueVal, b&: FalseVal); |
3902 | |
3903 | if (Value *X = foldSelectICmpAnd(Sel, CondVal, TrueVal, FalseVal, V, AndMask, |
3904 | CreateAnd, Builder)) |
3905 | return X; |
3906 | |
3907 | if (Value *X = foldSelectICmpAndBinOp(CondVal, TrueVal, FalseVal, V, AndMask, |
3908 | CreateAnd, Builder)) |
3909 | return X; |
3910 | |
3911 | return nullptr; |
3912 | } |
3913 | |
3914 | Instruction *InstCombinerImpl::visitSelectInst(SelectInst &SI) { |
3915 | Value *CondVal = SI.getCondition(); |
3916 | Value *TrueVal = SI.getTrueValue(); |
3917 | Value *FalseVal = SI.getFalseValue(); |
3918 | Type *SelType = SI.getType(); |
3919 | |
3920 | if (Value *V = simplifySelectInst(Cond: CondVal, TrueVal, FalseVal, |
3921 | Q: SQ.getWithInstruction(I: &SI))) |
3922 | return replaceInstUsesWith(I&: SI, V); |
3923 | |
3924 | if (Instruction *I = canonicalizeSelectToShuffle(SI)) |
3925 | return I; |
3926 | |
3927 | if (Instruction *I = canonicalizeScalarSelectOfVecs(Sel&: SI, IC&: *this)) |
3928 | return I; |
3929 | |
3930 | // If the type of select is not an integer type or if the condition and |
3931 | // the selection type are not both scalar nor both vector types, there is no |
3932 | // point in attempting to match these patterns. |
3933 | Type *CondType = CondVal->getType(); |
3934 | if (!isa<Constant>(Val: CondVal) && SelType->isIntOrIntVectorTy() && |
3935 | CondType->isVectorTy() == SelType->isVectorTy()) { |
3936 | if (Value *S = simplifyWithOpReplaced(V: TrueVal, Op: CondVal, |
3937 | RepOp: ConstantInt::getTrue(Ty: CondType), Q: SQ, |
3938 | /* AllowRefinement */ true)) |
3939 | return replaceOperand(I&: SI, OpNum: 1, V: S); |
3940 | |
3941 | if (Value *S = simplifyWithOpReplaced(V: FalseVal, Op: CondVal, |
3942 | RepOp: ConstantInt::getFalse(Ty: CondType), Q: SQ, |
3943 | /* AllowRefinement */ true)) |
3944 | return replaceOperand(I&: SI, OpNum: 2, V: S); |
3945 | |
3946 | if (replaceInInstruction(V: TrueVal, Old: CondVal, |
3947 | New: ConstantInt::getTrue(Ty: CondType)) || |
3948 | replaceInInstruction(V: FalseVal, Old: CondVal, |
3949 | New: ConstantInt::getFalse(Ty: CondType))) |
3950 | return &SI; |
3951 | } |
3952 | |
3953 | if (Instruction *R = foldSelectOfBools(SI)) |
3954 | return R; |
3955 | |
3956 | // Selecting between two integer or vector splat integer constants? |
3957 | // |
3958 | // Note that we don't handle a scalar select of vectors: |
3959 | // select i1 %c, <2 x i8> <1, 1>, <2 x i8> <0, 0> |
3960 | // because that may need 3 instructions to splat the condition value: |
3961 | // extend, insertelement, shufflevector. |
3962 | // |
3963 | // Do not handle i1 TrueVal and FalseVal otherwise would result in |
3964 | // zext/sext i1 to i1. |
3965 | if (SelType->isIntOrIntVectorTy() && !SelType->isIntOrIntVectorTy(BitWidth: 1) && |
3966 | CondVal->getType()->isVectorTy() == SelType->isVectorTy()) { |
3967 | // select C, 1, 0 -> zext C to int |
3968 | if (match(V: TrueVal, P: m_One()) && match(V: FalseVal, P: m_Zero())) |
3969 | return new ZExtInst(CondVal, SelType); |
3970 | |
3971 | // select C, -1, 0 -> sext C to int |
3972 | if (match(V: TrueVal, P: m_AllOnes()) && match(V: FalseVal, P: m_Zero())) |
3973 | return new SExtInst(CondVal, SelType); |
3974 | |
3975 | // select C, 0, 1 -> zext !C to int |
3976 | if (match(V: TrueVal, P: m_Zero()) && match(V: FalseVal, P: m_One())) { |
3977 | Value *NotCond = Builder.CreateNot(V: CondVal, Name: "not." + CondVal->getName()); |
3978 | return new ZExtInst(NotCond, SelType); |
3979 | } |
3980 | |
3981 | // select C, 0, -1 -> sext !C to int |
3982 | if (match(V: TrueVal, P: m_Zero()) && match(V: FalseVal, P: m_AllOnes())) { |
3983 | Value *NotCond = Builder.CreateNot(V: CondVal, Name: "not." + CondVal->getName()); |
3984 | return new SExtInst(NotCond, SelType); |
3985 | } |
3986 | } |
3987 | |
3988 | auto *SIFPOp = dyn_cast<FPMathOperator>(Val: &SI); |
3989 | |
3990 | if (auto *FCmp = dyn_cast<FCmpInst>(Val: CondVal)) { |
3991 | FCmpInst::Predicate Pred = FCmp->getPredicate(); |
3992 | Value *Cmp0 = FCmp->getOperand(i_nocapture: 0), *Cmp1 = FCmp->getOperand(i_nocapture: 1); |
3993 | // Are we selecting a value based on a comparison of the two values? |
3994 | if ((Cmp0 == TrueVal && Cmp1 == FalseVal) || |
3995 | (Cmp0 == FalseVal && Cmp1 == TrueVal)) { |
3996 | // Canonicalize to use ordered comparisons by swapping the select |
3997 | // operands. |
3998 | // |
3999 | // e.g. |
4000 | // (X ugt Y) ? X : Y -> (X ole Y) ? Y : X |
4001 | if (FCmp->hasOneUse() && FCmpInst::isUnordered(predicate: Pred)) { |
4002 | FCmpInst::Predicate InvPred = FCmp->getInversePredicate(); |
4003 | Value *NewCond = Builder.CreateFCmpFMF(P: InvPred, LHS: Cmp0, RHS: Cmp1, FMFSource: FCmp, |
4004 | Name: FCmp->getName() + ".inv" ); |
4005 | // Propagate ninf/nnan from fcmp to select. |
4006 | FastMathFlags FMF = SI.getFastMathFlags(); |
4007 | if (FCmp->hasNoNaNs()) |
4008 | FMF.setNoNaNs(true); |
4009 | if (FCmp->hasNoInfs()) |
4010 | FMF.setNoInfs(true); |
4011 | Value *NewSel = |
4012 | Builder.CreateSelectFMF(C: NewCond, True: FalseVal, False: TrueVal, FMFSource: FMF); |
4013 | return replaceInstUsesWith(I&: SI, V: NewSel); |
4014 | } |
4015 | } |
4016 | |
4017 | if (SIFPOp) { |
4018 | // Fold out scale-if-equals-zero pattern. |
4019 | // |
4020 | // This pattern appears in code with denormal range checks after it's |
4021 | // assumed denormals are treated as zero. This drops a canonicalization. |
4022 | |
4023 | // TODO: Could relax the signed zero logic. We just need to know the sign |
4024 | // of the result matches (fmul x, y has the same sign as x). |
4025 | // |
4026 | // TODO: Handle always-canonicalizing variant that selects some value or 1 |
4027 | // scaling factor in the fmul visitor. |
4028 | |
4029 | // TODO: Handle ldexp too |
4030 | |
4031 | Value *MatchCmp0 = nullptr; |
4032 | Value *MatchCmp1 = nullptr; |
4033 | |
4034 | // (select (fcmp [ou]eq x, 0.0), (fmul x, K), x => x |
4035 | // (select (fcmp [ou]ne x, 0.0), x, (fmul x, K) => x |
4036 | if (Pred == CmpInst::FCMP_OEQ || Pred == CmpInst::FCMP_UEQ) { |
4037 | MatchCmp0 = FalseVal; |
4038 | MatchCmp1 = TrueVal; |
4039 | } else if (Pred == CmpInst::FCMP_ONE || Pred == CmpInst::FCMP_UNE) { |
4040 | MatchCmp0 = TrueVal; |
4041 | MatchCmp1 = FalseVal; |
4042 | } |
4043 | |
4044 | if (Cmp0 == MatchCmp0 && |
4045 | matchFMulByZeroIfResultEqZero(IC&: *this, Cmp0, Cmp1, TrueVal: MatchCmp1, FalseVal: MatchCmp0, |
4046 | CtxI&: SI, SelectIsNSZ: SIFPOp->hasNoSignedZeros())) |
4047 | return replaceInstUsesWith(I&: SI, V: Cmp0); |
4048 | } |
4049 | } |
4050 | |
4051 | if (SIFPOp) { |
4052 | // TODO: Try to forward-propagate FMF from select arms to the select. |
4053 | |
4054 | auto *FCmp = dyn_cast<FCmpInst>(Val: CondVal); |
4055 | |
4056 | // Canonicalize select of FP values where NaN and -0.0 are not valid as |
4057 | // minnum/maxnum intrinsics. |
4058 | if (SIFPOp->hasNoNaNs() && |
4059 | (SIFPOp->hasNoSignedZeros() || |
4060 | (SIFPOp->hasOneUse() && |
4061 | canIgnoreSignBitOfZero(U: *SIFPOp->use_begin())))) { |
4062 | Value *X, *Y; |
4063 | if (match(V: &SI, P: m_OrdOrUnordFMax(L: m_Value(V&: X), R: m_Value(V&: Y)))) { |
4064 | Value *BinIntr = |
4065 | Builder.CreateBinaryIntrinsic(ID: Intrinsic::maxnum, LHS: X, RHS: Y, FMFSource: &SI); |
4066 | if (auto *BinIntrInst = dyn_cast<Instruction>(Val: BinIntr)) { |
4067 | BinIntrInst->setHasNoNaNs(FCmp->hasNoNaNs()); |
4068 | BinIntrInst->setHasNoInfs(FCmp->hasNoInfs()); |
4069 | } |
4070 | return replaceInstUsesWith(I&: SI, V: BinIntr); |
4071 | } |
4072 | |
4073 | if (match(V: &SI, P: m_OrdOrUnordFMin(L: m_Value(V&: X), R: m_Value(V&: Y)))) { |
4074 | Value *BinIntr = |
4075 | Builder.CreateBinaryIntrinsic(ID: Intrinsic::minnum, LHS: X, RHS: Y, FMFSource: &SI); |
4076 | if (auto *BinIntrInst = dyn_cast<Instruction>(Val: BinIntr)) { |
4077 | BinIntrInst->setHasNoNaNs(FCmp->hasNoNaNs()); |
4078 | BinIntrInst->setHasNoInfs(FCmp->hasNoInfs()); |
4079 | } |
4080 | return replaceInstUsesWith(I&: SI, V: BinIntr); |
4081 | } |
4082 | } |
4083 | } |
4084 | |
4085 | // Fold selecting to fabs. |
4086 | if (Instruction *Fabs = foldSelectWithFCmpToFabs(SI, IC&: *this)) |
4087 | return Fabs; |
4088 | |
4089 | // See if we are selecting two values based on a comparison of the two values. |
4090 | if (CmpInst *CI = dyn_cast<CmpInst>(Val: CondVal)) |
4091 | if (Instruction *NewSel = foldSelectValueEquivalence(Sel&: SI, Cmp&: *CI)) |
4092 | return NewSel; |
4093 | |
4094 | if (ICmpInst *ICI = dyn_cast<ICmpInst>(Val: CondVal)) |
4095 | if (Instruction *Result = foldSelectInstWithICmp(SI, ICI)) |
4096 | return Result; |
4097 | |
4098 | if (Value *V = foldSelectBitTest(Sel&: SI, CondVal, TrueVal, FalseVal, Builder, SQ)) |
4099 | return replaceInstUsesWith(I&: SI, V); |
4100 | |
4101 | if (Instruction *Add = foldAddSubSelect(SI, Builder)) |
4102 | return Add; |
4103 | if (Instruction *Add = foldOverflowingAddSubSelect(SI, Builder)) |
4104 | return Add; |
4105 | if (Instruction *Or = foldSetClearBits(Sel&: SI, Builder)) |
4106 | return Or; |
4107 | if (Instruction *Mul = foldSelectZeroOrMul(SI, IC&: *this)) |
4108 | return Mul; |
4109 | |
4110 | // Turn (select C, (op X, Y), (op X, Z)) -> (op X, (select C, Y, Z)) |
4111 | auto *TI = dyn_cast<Instruction>(Val: TrueVal); |
4112 | auto *FI = dyn_cast<Instruction>(Val: FalseVal); |
4113 | if (TI && FI && TI->getOpcode() == FI->getOpcode()) |
4114 | if (Instruction *IV = foldSelectOpOp(SI, TI, FI)) |
4115 | return IV; |
4116 | |
4117 | if (Instruction *I = foldSelectExtConst(Sel&: SI)) |
4118 | return I; |
4119 | |
4120 | if (Instruction *I = foldSelectWithSRem(SI, IC&: *this, Builder)) |
4121 | return I; |
4122 | |
4123 | // Fold (select C, (gep Ptr, Idx), Ptr) -> (gep Ptr, (select C, Idx, 0)) |
4124 | // Fold (select C, Ptr, (gep Ptr, Idx)) -> (gep Ptr, (select C, 0, Idx)) |
4125 | auto SelectGepWithBase = [&](GetElementPtrInst *Gep, Value *Base, |
4126 | bool Swap) -> GetElementPtrInst * { |
4127 | Value *Ptr = Gep->getPointerOperand(); |
4128 | if (Gep->getNumOperands() != 2 || Gep->getPointerOperand() != Base || |
4129 | !Gep->hasOneUse()) |
4130 | return nullptr; |
4131 | Value *Idx = Gep->getOperand(i_nocapture: 1); |
4132 | if (isa<VectorType>(Val: CondVal->getType()) && !isa<VectorType>(Val: Idx->getType())) |
4133 | return nullptr; |
4134 | Type *ElementType = Gep->getSourceElementType(); |
4135 | Value *NewT = Idx; |
4136 | Value *NewF = Constant::getNullValue(Ty: Idx->getType()); |
4137 | if (Swap) |
4138 | std::swap(a&: NewT, b&: NewF); |
4139 | Value *NewSI = |
4140 | Builder.CreateSelect(C: CondVal, True: NewT, False: NewF, Name: SI.getName() + ".idx" , MDFrom: &SI); |
4141 | return GetElementPtrInst::Create(PointeeType: ElementType, Ptr, IdxList: NewSI, |
4142 | NW: Gep->getNoWrapFlags()); |
4143 | }; |
4144 | if (auto *TrueGep = dyn_cast<GetElementPtrInst>(Val: TrueVal)) |
4145 | if (auto *NewGep = SelectGepWithBase(TrueGep, FalseVal, false)) |
4146 | return NewGep; |
4147 | if (auto *FalseGep = dyn_cast<GetElementPtrInst>(Val: FalseVal)) |
4148 | if (auto *NewGep = SelectGepWithBase(FalseGep, TrueVal, true)) |
4149 | return NewGep; |
4150 | |
4151 | // See if we can fold the select into one of our operands. |
4152 | if (SelType->isIntOrIntVectorTy() || SelType->isFPOrFPVectorTy()) { |
4153 | if (Instruction *FoldI = foldSelectIntoOp(SI, TrueVal, FalseVal)) |
4154 | return FoldI; |
4155 | |
4156 | Value *LHS, *RHS; |
4157 | Instruction::CastOps CastOp; |
4158 | SelectPatternResult SPR = matchSelectPattern(V: &SI, LHS, RHS, CastOp: &CastOp); |
4159 | auto SPF = SPR.Flavor; |
4160 | if (SPF) { |
4161 | Value *LHS2, *RHS2; |
4162 | if (SelectPatternFlavor SPF2 = matchSelectPattern(V: LHS, LHS&: LHS2, RHS&: RHS2).Flavor) |
4163 | if (Instruction *R = foldSPFofSPF(Inner: cast<Instruction>(Val: LHS), SPF1: SPF2, A: LHS2, |
4164 | B: RHS2, Outer&: SI, SPF2: SPF, C: RHS)) |
4165 | return R; |
4166 | if (SelectPatternFlavor SPF2 = matchSelectPattern(V: RHS, LHS&: LHS2, RHS&: RHS2).Flavor) |
4167 | if (Instruction *R = foldSPFofSPF(Inner: cast<Instruction>(Val: RHS), SPF1: SPF2, A: LHS2, |
4168 | B: RHS2, Outer&: SI, SPF2: SPF, C: LHS)) |
4169 | return R; |
4170 | } |
4171 | |
4172 | if (SelectPatternResult::isMinOrMax(SPF)) { |
4173 | // Canonicalize so that |
4174 | // - type casts are outside select patterns. |
4175 | // - float clamp is transformed to min/max pattern |
4176 | |
4177 | bool IsCastNeeded = LHS->getType() != SelType; |
4178 | Value *CmpLHS = cast<CmpInst>(Val: CondVal)->getOperand(i_nocapture: 0); |
4179 | Value *CmpRHS = cast<CmpInst>(Val: CondVal)->getOperand(i_nocapture: 1); |
4180 | if (IsCastNeeded || |
4181 | (LHS->getType()->isFPOrFPVectorTy() && |
4182 | ((CmpLHS != LHS && CmpLHS != RHS) || |
4183 | (CmpRHS != LHS && CmpRHS != RHS)))) { |
4184 | CmpInst::Predicate MinMaxPred = getMinMaxPred(SPF, Ordered: SPR.Ordered); |
4185 | |
4186 | Value *Cmp; |
4187 | if (CmpInst::isIntPredicate(P: MinMaxPred)) |
4188 | Cmp = Builder.CreateICmp(P: MinMaxPred, LHS, RHS); |
4189 | else |
4190 | Cmp = Builder.CreateFCmpFMF(P: MinMaxPred, LHS, RHS, |
4191 | FMFSource: cast<Instruction>(Val: SI.getCondition())); |
4192 | |
4193 | Value *NewSI = Builder.CreateSelect(C: Cmp, True: LHS, False: RHS, Name: SI.getName(), MDFrom: &SI); |
4194 | if (!IsCastNeeded) |
4195 | return replaceInstUsesWith(I&: SI, V: NewSI); |
4196 | |
4197 | Value *NewCast = Builder.CreateCast(Op: CastOp, V: NewSI, DestTy: SelType); |
4198 | return replaceInstUsesWith(I&: SI, V: NewCast); |
4199 | } |
4200 | } |
4201 | } |
4202 | |
4203 | // See if we can fold the select into a phi node if the condition is a select. |
4204 | if (auto *PN = dyn_cast<PHINode>(Val: SI.getCondition())) |
4205 | if (Instruction *NV = foldOpIntoPhi(I&: SI, PN)) |
4206 | return NV; |
4207 | |
4208 | if (SelectInst *TrueSI = dyn_cast<SelectInst>(Val: TrueVal)) { |
4209 | if (TrueSI->getCondition()->getType() == CondVal->getType()) { |
4210 | // Fold nested selects if the inner condition can be implied by the outer |
4211 | // condition. |
4212 | if (Value *V = simplifyNestedSelectsUsingImpliedCond( |
4213 | SI&: *TrueSI, CondVal, /*CondIsTrue=*/true, DL)) |
4214 | return replaceOperand(I&: SI, OpNum: 1, V); |
4215 | |
4216 | // select(C0, select(C1, a, b), b) -> select(C0&C1, a, b) |
4217 | // We choose this as normal form to enable folding on the And and |
4218 | // shortening paths for the values (this helps getUnderlyingObjects() for |
4219 | // example). |
4220 | if (TrueSI->getFalseValue() == FalseVal && TrueSI->hasOneUse()) { |
4221 | Value *And = Builder.CreateLogicalAnd(Cond1: CondVal, Cond2: TrueSI->getCondition()); |
4222 | replaceOperand(I&: SI, OpNum: 0, V: And); |
4223 | replaceOperand(I&: SI, OpNum: 1, V: TrueSI->getTrueValue()); |
4224 | return &SI; |
4225 | } |
4226 | } |
4227 | } |
4228 | if (SelectInst *FalseSI = dyn_cast<SelectInst>(Val: FalseVal)) { |
4229 | if (FalseSI->getCondition()->getType() == CondVal->getType()) { |
4230 | // Fold nested selects if the inner condition can be implied by the outer |
4231 | // condition. |
4232 | if (Value *V = simplifyNestedSelectsUsingImpliedCond( |
4233 | SI&: *FalseSI, CondVal, /*CondIsTrue=*/false, DL)) |
4234 | return replaceOperand(I&: SI, OpNum: 2, V); |
4235 | |
4236 | // select(C0, a, select(C1, a, b)) -> select(C0|C1, a, b) |
4237 | if (FalseSI->getTrueValue() == TrueVal && FalseSI->hasOneUse()) { |
4238 | Value *Or = Builder.CreateLogicalOr(Cond1: CondVal, Cond2: FalseSI->getCondition()); |
4239 | replaceOperand(I&: SI, OpNum: 0, V: Or); |
4240 | replaceOperand(I&: SI, OpNum: 2, V: FalseSI->getFalseValue()); |
4241 | return &SI; |
4242 | } |
4243 | } |
4244 | } |
4245 | |
4246 | // Try to simplify a binop sandwiched between 2 selects with the same |
4247 | // condition. This is not valid for div/rem because the select might be |
4248 | // preventing a division-by-zero. |
4249 | // TODO: A div/rem restriction is conservative; use something like |
4250 | // isSafeToSpeculativelyExecute(). |
4251 | // select(C, binop(select(C, X, Y), W), Z) -> select(C, binop(X, W), Z) |
4252 | BinaryOperator *TrueBO; |
4253 | if (match(V: TrueVal, P: m_OneUse(SubPattern: m_BinOp(I&: TrueBO))) && !TrueBO->isIntDivRem()) { |
4254 | if (auto *TrueBOSI = dyn_cast<SelectInst>(Val: TrueBO->getOperand(i_nocapture: 0))) { |
4255 | if (TrueBOSI->getCondition() == CondVal) { |
4256 | replaceOperand(I&: *TrueBO, OpNum: 0, V: TrueBOSI->getTrueValue()); |
4257 | Worklist.push(I: TrueBO); |
4258 | return &SI; |
4259 | } |
4260 | } |
4261 | if (auto *TrueBOSI = dyn_cast<SelectInst>(Val: TrueBO->getOperand(i_nocapture: 1))) { |
4262 | if (TrueBOSI->getCondition() == CondVal) { |
4263 | replaceOperand(I&: *TrueBO, OpNum: 1, V: TrueBOSI->getTrueValue()); |
4264 | Worklist.push(I: TrueBO); |
4265 | return &SI; |
4266 | } |
4267 | } |
4268 | } |
4269 | |
4270 | // select(C, Z, binop(select(C, X, Y), W)) -> select(C, Z, binop(Y, W)) |
4271 | BinaryOperator *FalseBO; |
4272 | if (match(V: FalseVal, P: m_OneUse(SubPattern: m_BinOp(I&: FalseBO))) && !FalseBO->isIntDivRem()) { |
4273 | if (auto *FalseBOSI = dyn_cast<SelectInst>(Val: FalseBO->getOperand(i_nocapture: 0))) { |
4274 | if (FalseBOSI->getCondition() == CondVal) { |
4275 | replaceOperand(I&: *FalseBO, OpNum: 0, V: FalseBOSI->getFalseValue()); |
4276 | Worklist.push(I: FalseBO); |
4277 | return &SI; |
4278 | } |
4279 | } |
4280 | if (auto *FalseBOSI = dyn_cast<SelectInst>(Val: FalseBO->getOperand(i_nocapture: 1))) { |
4281 | if (FalseBOSI->getCondition() == CondVal) { |
4282 | replaceOperand(I&: *FalseBO, OpNum: 1, V: FalseBOSI->getFalseValue()); |
4283 | Worklist.push(I: FalseBO); |
4284 | return &SI; |
4285 | } |
4286 | } |
4287 | } |
4288 | |
4289 | Value *NotCond; |
4290 | if (match(V: CondVal, P: m_Not(V: m_Value(V&: NotCond))) && |
4291 | !InstCombiner::shouldAvoidAbsorbingNotIntoSelect(SI)) { |
4292 | replaceOperand(I&: SI, OpNum: 0, V: NotCond); |
4293 | SI.swapValues(); |
4294 | SI.swapProfMetadata(); |
4295 | return &SI; |
4296 | } |
4297 | |
4298 | if (Instruction *I = foldVectorSelect(Sel&: SI)) |
4299 | return I; |
4300 | |
4301 | // If we can compute the condition, there's no need for a select. |
4302 | // Like the above fold, we are attempting to reduce compile-time cost by |
4303 | // putting this fold here with limitations rather than in InstSimplify. |
4304 | // The motivation for this call into value tracking is to take advantage of |
4305 | // the assumption cache, so make sure that is populated. |
4306 | if (!CondVal->getType()->isVectorTy() && !AC.assumptions().empty()) { |
4307 | KnownBits Known(1); |
4308 | computeKnownBits(V: CondVal, Known, CxtI: &SI); |
4309 | if (Known.One.isOne()) |
4310 | return replaceInstUsesWith(I&: SI, V: TrueVal); |
4311 | if (Known.Zero.isOne()) |
4312 | return replaceInstUsesWith(I&: SI, V: FalseVal); |
4313 | } |
4314 | |
4315 | if (Instruction *BitCastSel = foldSelectCmpBitcasts(Sel&: SI, Builder)) |
4316 | return BitCastSel; |
4317 | |
4318 | // Simplify selects that test the returned flag of cmpxchg instructions. |
4319 | if (Value *V = foldSelectCmpXchg(SI)) |
4320 | return replaceInstUsesWith(I&: SI, V); |
4321 | |
4322 | if (Instruction *Select = foldSelectBinOpIdentity(Sel&: SI, TLI, IC&: *this)) |
4323 | return Select; |
4324 | |
4325 | if (Instruction *Funnel = foldSelectFunnelShift(Sel&: SI, Builder)) |
4326 | return Funnel; |
4327 | |
4328 | if (Instruction *Copysign = foldSelectToCopysign(Sel&: SI, Builder)) |
4329 | return Copysign; |
4330 | |
4331 | if (Instruction *PN = foldSelectToPhi(Sel&: SI, DT, Builder)) |
4332 | return replaceInstUsesWith(I&: SI, V: PN); |
4333 | |
4334 | if (Value *Fr = foldSelectWithFrozenICmp(Sel&: SI, Builder)) |
4335 | return replaceInstUsesWith(I&: SI, V: Fr); |
4336 | |
4337 | if (Value *V = foldRoundUpIntegerWithPow2Alignment(SI, Builder)) |
4338 | return replaceInstUsesWith(I&: SI, V); |
4339 | |
4340 | if (Value *V = foldSelectIntoAddConstant(SI, Builder)) |
4341 | return replaceInstUsesWith(I&: SI, V); |
4342 | |
4343 | // select(mask, mload(,,mask,0), 0) -> mload(,,mask,0) |
4344 | // Load inst is intentionally not checked for hasOneUse() |
4345 | if (match(V: FalseVal, P: m_Zero()) && |
4346 | (match(V: TrueVal, P: m_MaskedLoad(Op0: m_Value(), Op1: m_Value(), Op2: m_Specific(V: CondVal), |
4347 | Op3: m_CombineOr(L: m_Undef(), R: m_Zero()))) || |
4348 | match(V: TrueVal, P: m_MaskedGather(Op0: m_Value(), Op1: m_Value(), Op2: m_Specific(V: CondVal), |
4349 | Op3: m_CombineOr(L: m_Undef(), R: m_Zero()))))) { |
4350 | auto *MaskedInst = cast<IntrinsicInst>(Val: TrueVal); |
4351 | if (isa<UndefValue>(Val: MaskedInst->getArgOperand(i: 3))) |
4352 | MaskedInst->setArgOperand(i: 3, v: FalseVal /* Zero */); |
4353 | return replaceInstUsesWith(I&: SI, V: MaskedInst); |
4354 | } |
4355 | |
4356 | Value *Mask; |
4357 | if (match(V: TrueVal, P: m_Zero()) && |
4358 | (match(V: FalseVal, P: m_MaskedLoad(Op0: m_Value(), Op1: m_Value(), Op2: m_Value(V&: Mask), |
4359 | Op3: m_CombineOr(L: m_Undef(), R: m_Zero()))) || |
4360 | match(V: FalseVal, P: m_MaskedGather(Op0: m_Value(), Op1: m_Value(), Op2: m_Value(V&: Mask), |
4361 | Op3: m_CombineOr(L: m_Undef(), R: m_Zero())))) && |
4362 | (CondVal->getType() == Mask->getType())) { |
4363 | // We can remove the select by ensuring the load zeros all lanes the |
4364 | // select would have. We determine this by proving there is no overlap |
4365 | // between the load and select masks. |
4366 | // (i.e (load_mask & select_mask) == 0 == no overlap) |
4367 | bool CanMergeSelectIntoLoad = false; |
4368 | if (Value *V = simplifyAndInst(LHS: CondVal, RHS: Mask, Q: SQ.getWithInstruction(I: &SI))) |
4369 | CanMergeSelectIntoLoad = match(V, P: m_Zero()); |
4370 | |
4371 | if (CanMergeSelectIntoLoad) { |
4372 | auto *MaskedInst = cast<IntrinsicInst>(Val: FalseVal); |
4373 | if (isa<UndefValue>(Val: MaskedInst->getArgOperand(i: 3))) |
4374 | MaskedInst->setArgOperand(i: 3, v: TrueVal /* Zero */); |
4375 | return replaceInstUsesWith(I&: SI, V: MaskedInst); |
4376 | } |
4377 | } |
4378 | |
4379 | if (Instruction *I = foldSelectOfSymmetricSelect(OuterSelVal&: SI, Builder)) |
4380 | return I; |
4381 | |
4382 | if (Instruction *I = foldNestedSelects(OuterSelVal&: SI, Builder)) |
4383 | return I; |
4384 | |
4385 | // Match logical variants of the pattern, |
4386 | // and transform them iff that gets rid of inversions. |
4387 | // (~x) | y --> ~(x & (~y)) |
4388 | // (~x) & y --> ~(x | (~y)) |
4389 | if (sinkNotIntoOtherHandOfLogicalOp(I&: SI)) |
4390 | return &SI; |
4391 | |
4392 | if (Instruction *I = foldBitCeil(SI, Builder, IC&: *this)) |
4393 | return I; |
4394 | |
4395 | if (Instruction *I = foldSelectToCmp(SI)) |
4396 | return I; |
4397 | |
4398 | if (Instruction *I = foldSelectEqualityTest(Sel&: SI)) |
4399 | return I; |
4400 | |
4401 | // Fold: |
4402 | // (select A && B, T, F) -> (select A, (select B, T, F), F) |
4403 | // (select A || B, T, F) -> (select A, T, (select B, T, F)) |
4404 | // if (select B, T, F) is foldable. |
4405 | // TODO: preserve FMF flags |
4406 | auto FoldSelectWithAndOrCond = [&](bool IsAnd, Value *A, |
4407 | Value *B) -> Instruction * { |
4408 | if (Value *V = simplifySelectInst(Cond: B, TrueVal, FalseVal, |
4409 | Q: SQ.getWithInstruction(I: &SI))) |
4410 | return SelectInst::Create(C: A, S1: IsAnd ? V : TrueVal, S2: IsAnd ? FalseVal : V); |
4411 | |
4412 | // Is (select B, T, F) a SPF? |
4413 | if (CondVal->hasOneUse() && SelType->isIntOrIntVectorTy()) { |
4414 | if (ICmpInst *Cmp = dyn_cast<ICmpInst>(Val: B)) |
4415 | if (Value *V = canonicalizeSPF(Cmp&: *Cmp, TrueVal, FalseVal, IC&: *this)) |
4416 | return SelectInst::Create(C: A, S1: IsAnd ? V : TrueVal, |
4417 | S2: IsAnd ? FalseVal : V); |
4418 | } |
4419 | |
4420 | return nullptr; |
4421 | }; |
4422 | |
4423 | Value *LHS, *RHS; |
4424 | if (match(V: CondVal, P: m_And(L: m_Value(V&: LHS), R: m_Value(V&: RHS)))) { |
4425 | if (Instruction *I = FoldSelectWithAndOrCond(/*IsAnd*/ true, LHS, RHS)) |
4426 | return I; |
4427 | if (Instruction *I = FoldSelectWithAndOrCond(/*IsAnd*/ true, RHS, LHS)) |
4428 | return I; |
4429 | } else if (match(V: CondVal, P: m_Or(L: m_Value(V&: LHS), R: m_Value(V&: RHS)))) { |
4430 | if (Instruction *I = FoldSelectWithAndOrCond(/*IsAnd*/ false, LHS, RHS)) |
4431 | return I; |
4432 | if (Instruction *I = FoldSelectWithAndOrCond(/*IsAnd*/ false, RHS, LHS)) |
4433 | return I; |
4434 | } else { |
4435 | // We cannot swap the operands of logical and/or. |
4436 | // TODO: Can we swap the operands by inserting a freeze? |
4437 | if (match(V: CondVal, P: m_LogicalAnd(L: m_Value(V&: LHS), R: m_Value(V&: RHS)))) { |
4438 | if (Instruction *I = FoldSelectWithAndOrCond(/*IsAnd*/ true, LHS, RHS)) |
4439 | return I; |
4440 | } else if (match(V: CondVal, P: m_LogicalOr(L: m_Value(V&: LHS), R: m_Value(V&: RHS)))) { |
4441 | if (Instruction *I = FoldSelectWithAndOrCond(/*IsAnd*/ false, LHS, RHS)) |
4442 | return I; |
4443 | } |
4444 | } |
4445 | |
4446 | // select Cond, !X, X -> xor Cond, X |
4447 | if (CondVal->getType() == SI.getType() && isKnownInversion(X: FalseVal, Y: TrueVal)) |
4448 | return BinaryOperator::CreateXor(V1: CondVal, V2: FalseVal); |
4449 | |
4450 | // For vectors, this transform is only safe if the simplification does not |
4451 | // look through any lane-crossing operations. For now, limit to scalars only. |
4452 | if (SelType->isIntegerTy() && |
4453 | (!isa<Constant>(Val: TrueVal) || !isa<Constant>(Val: FalseVal))) { |
4454 | // Try to simplify select arms based on KnownBits implied by the condition. |
4455 | CondContext CC(CondVal); |
4456 | findValuesAffectedByCondition(Cond: CondVal, /*IsAssume=*/false, InsertAffected: [&](Value *V) { |
4457 | CC.AffectedValues.insert(Ptr: V); |
4458 | }); |
4459 | SimplifyQuery Q = SQ.getWithInstruction(I: &SI).getWithCondContext(CC); |
4460 | if (!CC.AffectedValues.empty()) { |
4461 | if (!isa<Constant>(Val: TrueVal) && |
4462 | hasAffectedValue(V: TrueVal, Affected&: CC.AffectedValues, /*Depth=*/0)) { |
4463 | KnownBits Known = llvm::computeKnownBits(V: TrueVal, Q); |
4464 | if (Known.isConstant()) |
4465 | return replaceOperand(I&: SI, OpNum: 1, |
4466 | V: ConstantInt::get(Ty: SelType, V: Known.getConstant())); |
4467 | } |
4468 | |
4469 | CC.Invert = true; |
4470 | if (!isa<Constant>(Val: FalseVal) && |
4471 | hasAffectedValue(V: FalseVal, Affected&: CC.AffectedValues, /*Depth=*/0)) { |
4472 | KnownBits Known = llvm::computeKnownBits(V: FalseVal, Q); |
4473 | if (Known.isConstant()) |
4474 | return replaceOperand(I&: SI, OpNum: 2, |
4475 | V: ConstantInt::get(Ty: SelType, V: Known.getConstant())); |
4476 | } |
4477 | } |
4478 | } |
4479 | |
4480 | // select (trunc nuw X to i1), X, Y --> select (trunc nuw X to i1), 1, Y |
4481 | // select (trunc nuw X to i1), Y, X --> select (trunc nuw X to i1), Y, 0 |
4482 | // select (trunc nsw X to i1), X, Y --> select (trunc nsw X to i1), -1, Y |
4483 | // select (trunc nsw X to i1), Y, X --> select (trunc nsw X to i1), Y, 0 |
4484 | Value *Trunc; |
4485 | if (match(V: CondVal, P: m_NUWTrunc(Op: m_Value(V&: Trunc)))) { |
4486 | if (TrueVal == Trunc) |
4487 | return replaceOperand(I&: SI, OpNum: 1, V: ConstantInt::get(Ty: TrueVal->getType(), V: 1)); |
4488 | if (FalseVal == Trunc) |
4489 | return replaceOperand(I&: SI, OpNum: 2, V: ConstantInt::get(Ty: FalseVal->getType(), V: 0)); |
4490 | } |
4491 | if (match(V: CondVal, P: m_NSWTrunc(Op: m_Value(V&: Trunc)))) { |
4492 | if (TrueVal == Trunc) |
4493 | return replaceOperand(I&: SI, OpNum: 1, |
4494 | V: Constant::getAllOnesValue(Ty: TrueVal->getType())); |
4495 | if (FalseVal == Trunc) |
4496 | return replaceOperand(I&: SI, OpNum: 2, V: ConstantInt::get(Ty: FalseVal->getType(), V: 0)); |
4497 | } |
4498 | |
4499 | return nullptr; |
4500 | } |
4501 | |