1 //===- InstCombineSelect.cpp ----------------------------------------------===//
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
7 //===----------------------------------------------------------------------===//
9 // This file implements the visitSelect function.
11 //===----------------------------------------------------------------------===//
13 #include "InstCombineInternal.h"
14 #include "llvm/ADT/APInt.h"
15 #include "llvm/ADT/Optional.h"
16 #include "llvm/ADT/STLExtras.h"
17 #include "llvm/ADT/SmallVector.h"
18 #include "llvm/Analysis/AssumptionCache.h"
19 #include "llvm/Analysis/CmpInstAnalysis.h"
20 #include "llvm/Analysis/InstructionSimplify.h"
21 #include "llvm/Analysis/OverflowInstAnalysis.h"
22 #include "llvm/Analysis/ValueTracking.h"
23 #include "llvm/IR/BasicBlock.h"
24 #include "llvm/IR/Constant.h"
25 #include "llvm/IR/ConstantRange.h"
26 #include "llvm/IR/Constants.h"
27 #include "llvm/IR/DerivedTypes.h"
28 #include "llvm/IR/IRBuilder.h"
29 #include "llvm/IR/InstrTypes.h"
30 #include "llvm/IR/Instruction.h"
31 #include "llvm/IR/Instructions.h"
32 #include "llvm/IR/IntrinsicInst.h"
33 #include "llvm/IR/Intrinsics.h"
34 #include "llvm/IR/Operator.h"
35 #include "llvm/IR/PatternMatch.h"
36 #include "llvm/IR/Type.h"
37 #include "llvm/IR/User.h"
38 #include "llvm/IR/Value.h"
39 #include "llvm/Support/Casting.h"
40 #include "llvm/Support/ErrorHandling.h"
41 #include "llvm/Support/KnownBits.h"
42 #include "llvm/Transforms/InstCombine/InstCombiner.h"
46 #define DEBUG_TYPE "instcombine"
47 #include "llvm/Transforms/Utils/InstructionWorklist.h"
50 using namespace PatternMatch;
53 /// Replace a select operand based on an equality comparison with the identity
54 /// constant of a binop.
55 static Instruction *foldSelectBinOpIdentity(SelectInst &Sel,
56 const TargetLibraryInfo &TLI,
57 InstCombinerImpl &IC) {
58 // The select condition must be an equality compare with a constant operand.
61 CmpInst::Predicate Pred;
62 if (!match(Sel.getCondition(), m_Cmp(Pred, m_Value(X), m_Constant(C))))
66 if (ICmpInst::isEquality(Pred))
67 IsEq = Pred == ICmpInst::ICMP_EQ;
68 else if (Pred == FCmpInst::FCMP_OEQ)
70 else if (Pred == FCmpInst::FCMP_UNE)
75 // A select operand must be a binop.
77 if (!match(Sel.getOperand(IsEq ? 1 : 2), m_BinOp(BO)))
80 // The compare constant must be the identity constant for that binop.
81 // If this a floating-point compare with 0.0, any zero constant will do.
82 Type *Ty = BO->getType();
83 Constant *IdC = ConstantExpr::getBinOpIdentity(BO->getOpcode(), Ty, true);
85 if (!IdC || !CmpInst::isFPPredicate(Pred))
87 if (!match(IdC, m_AnyZeroFP()) || !match(C, m_AnyZeroFP()))
91 // Last, match the compare variable operand with a binop operand.
93 if (!BO->isCommutative() && !match(BO, m_BinOp(m_Value(Y), m_Specific(X))))
95 if (!match(BO, m_c_BinOp(m_Value(Y), m_Specific(X))))
98 // +0.0 compares equal to -0.0, and so it does not behave as required for this
99 // transform. Bail out if we can not exclude that possibility.
100 if (isa<FPMathOperator>(BO))
101 if (!BO->hasNoSignedZeros() && !CannotBeNegativeZero(Y, &TLI))
105 // S = { select (cmp eq X, C), BO, ? } or { select (cmp ne X, C), ?, BO }
107 // S = { select (cmp eq X, C), Y, ? } or { select (cmp ne X, C), ?, Y }
108 return IC.replaceOperand(Sel, IsEq ? 1 : 2, Y);
112 /// select (icmp eq (and X, C1)), TC, FC
113 /// iff C1 is a power 2 and the difference between TC and FC is a power-of-2.
114 /// To something like:
115 /// (shr (and (X, C1)), (log2(C1) - log2(TC-FC))) + FC
117 /// (shl (and (X, C1)), (log2(TC-FC) - log2(C1))) + FC
118 /// With some variations depending if FC is larger than TC, or the shift
119 /// isn't needed, or the bit widths don't match.
120 static Value *foldSelectICmpAnd(SelectInst &Sel, ICmpInst *Cmp,
121 InstCombiner::BuilderTy &Builder) {
122 const APInt *SelTC, *SelFC;
123 if (!match(Sel.getTrueValue(), m_APInt(SelTC)) ||
124 !match(Sel.getFalseValue(), m_APInt(SelFC)))
127 // If this is a vector select, we need a vector compare.
128 Type *SelType = Sel.getType();
129 if (SelType->isVectorTy() != Cmp->getType()->isVectorTy())
134 bool CreateAnd = false;
135 ICmpInst::Predicate Pred = Cmp->getPredicate();
136 if (ICmpInst::isEquality(Pred)) {
137 if (!match(Cmp->getOperand(1), m_Zero()))
140 V = Cmp->getOperand(0);
142 if (!match(V, m_And(m_Value(), m_Power2(AndRHS))))
146 } else if (decomposeBitTestICmp(Cmp->getOperand(0), Cmp->getOperand(1),
148 assert(ICmpInst::isEquality(Pred) && "Not equality test?");
149 if (!AndMask.isPowerOf2())
157 // In general, when both constants are non-zero, we would need an offset to
158 // replace the select. This would require more instructions than we started
159 // with. But there's one special-case that we handle here because it can
160 // simplify/reduce the instructions.
163 if (!TC.isZero() && !FC.isZero()) {
164 // If the select constants differ by exactly one bit and that's the same
165 // bit that is masked and checked by the select condition, the select can
166 // be replaced by bitwise logic to set/clear one bit of the constant result.
167 if (TC.getBitWidth() != AndMask.getBitWidth() || (TC ^ FC) != AndMask)
170 // If we have to create an 'and', then we must kill the cmp to not
171 // increase the instruction count.
172 if (!Cmp->hasOneUse())
174 V = Builder.CreateAnd(V, ConstantInt::get(SelType, AndMask));
176 bool ExtraBitInTC = TC.ugt(FC);
177 if (Pred == ICmpInst::ICMP_EQ) {
178 // If the masked bit in V is clear, clear or set the bit in the result:
179 // (V & AndMaskC) == 0 ? TC : FC --> (V & AndMaskC) ^ TC
180 // (V & AndMaskC) == 0 ? TC : FC --> (V & AndMaskC) | TC
181 Constant *C = ConstantInt::get(SelType, TC);
182 return ExtraBitInTC ? Builder.CreateXor(V, C) : Builder.CreateOr(V, C);
184 if (Pred == ICmpInst::ICMP_NE) {
185 // If the masked bit in V is set, set or clear the bit in the result:
186 // (V & AndMaskC) != 0 ? TC : FC --> (V & AndMaskC) | FC
187 // (V & AndMaskC) != 0 ? TC : FC --> (V & AndMaskC) ^ FC
188 Constant *C = ConstantInt::get(SelType, FC);
189 return ExtraBitInTC ? Builder.CreateOr(V, C) : Builder.CreateXor(V, C);
191 llvm_unreachable("Only expecting equality predicates");
194 // Make sure one of the select arms is a power-of-2.
195 if (!TC.isPowerOf2() && !FC.isPowerOf2())
198 // Determine which shift is needed to transform result of the 'and' into the
200 const APInt &ValC = !TC.isZero() ? TC : FC;
201 unsigned ValZeros = ValC.logBase2();
202 unsigned AndZeros = AndMask.logBase2();
204 // Insert the 'and' instruction on the input to the truncate.
206 V = Builder.CreateAnd(V, ConstantInt::get(V->getType(), AndMask));
208 // If types don't match, we can still convert the select by introducing a zext
209 // or a trunc of the 'and'.
210 if (ValZeros > AndZeros) {
211 V = Builder.CreateZExtOrTrunc(V, SelType);
212 V = Builder.CreateShl(V, ValZeros - AndZeros);
213 } else if (ValZeros < AndZeros) {
214 V = Builder.CreateLShr(V, AndZeros - ValZeros);
215 V = Builder.CreateZExtOrTrunc(V, SelType);
217 V = Builder.CreateZExtOrTrunc(V, SelType);
220 // Okay, now we know that everything is set up, we just don't know whether we
221 // have a icmp_ne or icmp_eq and whether the true or false val is the zero.
222 bool ShouldNotVal = !TC.isZero();
223 ShouldNotVal ^= Pred == ICmpInst::ICMP_NE;
225 V = Builder.CreateXor(V, ValC);
230 /// We want to turn code that looks like this:
232 /// %D = select %cond, %C, %A
234 /// %C = select %cond, %B, 0
237 /// Assuming that the specified instruction is an operand to the select, return
238 /// a bitmask indicating which operands of this instruction are foldable if they
239 /// equal the other incoming value of the select.
240 static unsigned getSelectFoldableOperands(BinaryOperator *I) {
241 switch (I->getOpcode()) {
242 case Instruction::Add:
243 case Instruction::FAdd:
244 case Instruction::Mul:
245 case Instruction::FMul:
246 case Instruction::And:
247 case Instruction::Or:
248 case Instruction::Xor:
249 return 3; // Can fold through either operand.
250 case Instruction::Sub: // Can only fold on the amount subtracted.
251 case Instruction::FSub:
252 case Instruction::FDiv: // Can only fold on the divisor amount.
253 case Instruction::Shl: // Can only fold on the shift amount.
254 case Instruction::LShr:
255 case Instruction::AShr:
258 return 0; // Cannot fold
262 /// We have (select c, TI, FI), and we know that TI and FI have the same opcode.
263 Instruction *InstCombinerImpl::foldSelectOpOp(SelectInst &SI, Instruction *TI,
265 // Don't break up min/max patterns. The hasOneUse checks below prevent that
266 // for most cases, but vector min/max with bitcasts can be transformed. If the
267 // one-use restrictions are eased for other patterns, we still don't want to
268 // obfuscate min/max.
269 if ((match(&SI, m_SMin(m_Value(), m_Value())) ||
270 match(&SI, m_SMax(m_Value(), m_Value())) ||
271 match(&SI, m_UMin(m_Value(), m_Value())) ||
272 match(&SI, m_UMax(m_Value(), m_Value()))))
275 // If this is a cast from the same type, merge.
276 Value *Cond = SI.getCondition();
277 Type *CondTy = Cond->getType();
278 if (TI->getNumOperands() == 1 && TI->isCast()) {
279 Type *FIOpndTy = FI->getOperand(0)->getType();
280 if (TI->getOperand(0)->getType() != FIOpndTy)
283 // The select condition may be a vector. We may only change the operand
284 // type if the vector width remains the same (and matches the condition).
285 if (auto *CondVTy = dyn_cast<VectorType>(CondTy)) {
286 if (!FIOpndTy->isVectorTy() ||
287 CondVTy->getElementCount() !=
288 cast<VectorType>(FIOpndTy)->getElementCount())
291 // TODO: If the backend knew how to deal with casts better, we could
292 // remove this limitation. For now, there's too much potential to create
293 // worse codegen by promoting the select ahead of size-altering casts
296 // Note that ValueTracking's matchSelectPattern() looks through casts
297 // without checking 'hasOneUse' when it matches min/max patterns, so this
298 // transform may end up happening anyway.
299 if (TI->getOpcode() != Instruction::BitCast &&
300 (!TI->hasOneUse() || !FI->hasOneUse()))
302 } else if (!TI->hasOneUse() || !FI->hasOneUse()) {
303 // TODO: The one-use restrictions for a scalar select could be eased if
304 // the fold of a select in visitLoadInst() was enhanced to match a pattern
305 // that includes a cast.
309 // Fold this by inserting a select from the input values.
311 Builder.CreateSelect(Cond, TI->getOperand(0), FI->getOperand(0),
312 SI.getName() + ".v", &SI);
313 return CastInst::Create(Instruction::CastOps(TI->getOpcode()), NewSI,
317 // Cond ? -X : -Y --> -(Cond ? X : Y)
319 if (match(TI, m_FNeg(m_Value(X))) && match(FI, m_FNeg(m_Value(Y))) &&
320 (TI->hasOneUse() || FI->hasOneUse())) {
321 // Intersect FMF from the fneg instructions and union those with the select.
322 FastMathFlags FMF = TI->getFastMathFlags();
323 FMF &= FI->getFastMathFlags();
324 FMF |= SI.getFastMathFlags();
325 Value *NewSel = Builder.CreateSelect(Cond, X, Y, SI.getName() + ".v", &SI);
326 if (auto *NewSelI = dyn_cast<Instruction>(NewSel))
327 NewSelI->setFastMathFlags(FMF);
328 Instruction *NewFNeg = UnaryOperator::CreateFNeg(NewSel);
329 NewFNeg->setFastMathFlags(FMF);
333 // Min/max intrinsic with a common operand can have the common operand pulled
334 // after the select. This is the same transform as below for binops, but
335 // specialized for intrinsic matching and without the restrictive uses clause.
336 auto *TII = dyn_cast<IntrinsicInst>(TI);
337 auto *FII = dyn_cast<IntrinsicInst>(FI);
338 if (TII && FII && TII->getIntrinsicID() == FII->getIntrinsicID() &&
339 (TII->hasOneUse() || FII->hasOneUse())) {
340 Value *T0, *T1, *F0, *F1;
341 if (match(TII, m_MaxOrMin(m_Value(T0), m_Value(T1))) &&
342 match(FII, m_MaxOrMin(m_Value(F0), m_Value(F1)))) {
344 Value *NewSel = Builder.CreateSelect(Cond, T1, F1, "minmaxop", &SI);
345 return CallInst::Create(TII->getCalledFunction(), {NewSel, T0});
348 Value *NewSel = Builder.CreateSelect(Cond, T1, F0, "minmaxop", &SI);
349 return CallInst::Create(TII->getCalledFunction(), {NewSel, T0});
352 Value *NewSel = Builder.CreateSelect(Cond, T0, F1, "minmaxop", &SI);
353 return CallInst::Create(TII->getCalledFunction(), {NewSel, T1});
356 Value *NewSel = Builder.CreateSelect(Cond, T0, F0, "minmaxop", &SI);
357 return CallInst::Create(TII->getCalledFunction(), {NewSel, T1});
362 // Only handle binary operators (including two-operand getelementptr) with
363 // one-use here. As with the cast case above, it may be possible to relax the
364 // one-use constraint, but that needs be examined carefully since it may not
365 // reduce the total number of instructions.
366 if (TI->getNumOperands() != 2 || FI->getNumOperands() != 2 ||
367 !TI->isSameOperationAs(FI) ||
368 (!isa<BinaryOperator>(TI) && !isa<GetElementPtrInst>(TI)) ||
369 !TI->hasOneUse() || !FI->hasOneUse())
372 // Figure out if the operations have any operands in common.
373 Value *MatchOp, *OtherOpT, *OtherOpF;
375 if (TI->getOperand(0) == FI->getOperand(0)) {
376 MatchOp = TI->getOperand(0);
377 OtherOpT = TI->getOperand(1);
378 OtherOpF = FI->getOperand(1);
379 MatchIsOpZero = true;
380 } else if (TI->getOperand(1) == FI->getOperand(1)) {
381 MatchOp = TI->getOperand(1);
382 OtherOpT = TI->getOperand(0);
383 OtherOpF = FI->getOperand(0);
384 MatchIsOpZero = false;
385 } else if (!TI->isCommutative()) {
387 } else if (TI->getOperand(0) == FI->getOperand(1)) {
388 MatchOp = TI->getOperand(0);
389 OtherOpT = TI->getOperand(1);
390 OtherOpF = FI->getOperand(0);
391 MatchIsOpZero = true;
392 } else if (TI->getOperand(1) == FI->getOperand(0)) {
393 MatchOp = TI->getOperand(1);
394 OtherOpT = TI->getOperand(0);
395 OtherOpF = FI->getOperand(1);
396 MatchIsOpZero = true;
401 // If the select condition is a vector, the operands of the original select's
402 // operands also must be vectors. This may not be the case for getelementptr
404 if (CondTy->isVectorTy() && (!OtherOpT->getType()->isVectorTy() ||
405 !OtherOpF->getType()->isVectorTy()))
408 // If we reach here, they do have operations in common.
409 Value *NewSI = Builder.CreateSelect(Cond, OtherOpT, OtherOpF,
410 SI.getName() + ".v", &SI);
411 Value *Op0 = MatchIsOpZero ? MatchOp : NewSI;
412 Value *Op1 = MatchIsOpZero ? NewSI : MatchOp;
413 if (auto *BO = dyn_cast<BinaryOperator>(TI)) {
414 BinaryOperator *NewBO = BinaryOperator::Create(BO->getOpcode(), Op0, Op1);
415 NewBO->copyIRFlags(TI);
416 NewBO->andIRFlags(FI);
419 if (auto *TGEP = dyn_cast<GetElementPtrInst>(TI)) {
420 auto *FGEP = cast<GetElementPtrInst>(FI);
421 Type *ElementType = TGEP->getResultElementType();
422 return TGEP->isInBounds() && FGEP->isInBounds()
423 ? GetElementPtrInst::CreateInBounds(ElementType, Op0, {Op1})
424 : GetElementPtrInst::Create(ElementType, Op0, {Op1});
426 llvm_unreachable("Expected BinaryOperator or GEP");
430 static bool isSelect01(const APInt &C1I, const APInt &C2I) {
431 if (!C1I.isZero() && !C2I.isZero()) // One side must be zero.
433 return C1I.isOne() || C1I.isAllOnes() || C2I.isOne() || C2I.isAllOnes();
436 /// Try to fold the select into one of the operands to allow further
438 Instruction *InstCombinerImpl::foldSelectIntoOp(SelectInst &SI, Value *TrueVal,
440 // See the comment above GetSelectFoldableOperands for a description of the
441 // transformation we are doing here.
442 auto TryFoldSelectIntoOp = [&](SelectInst &SI, Value *TrueVal,
444 bool Swapped) -> Instruction * {
445 if (auto *TVI = dyn_cast<BinaryOperator>(TrueVal)) {
446 if (TVI->hasOneUse() && !isa<Constant>(FalseVal)) {
447 if (unsigned SFO = getSelectFoldableOperands(TVI)) {
448 unsigned OpToFold = 0;
449 if ((SFO & 1) && FalseVal == TVI->getOperand(0))
451 else if ((SFO & 2) && FalseVal == TVI->getOperand(1))
456 // TODO: We probably ought to revisit cases where the select and FP
457 // instructions have different flags and add tests to ensure the
458 // behaviour is correct.
459 if (isa<FPMathOperator>(&SI))
460 FMF = SI.getFastMathFlags();
461 Constant *C = ConstantExpr::getBinOpIdentity(
462 TVI->getOpcode(), TVI->getType(), true, FMF.noSignedZeros());
463 Value *OOp = TVI->getOperand(2 - OpToFold);
464 // Avoid creating select between 2 constants unless it's selecting
465 // between 0, 1 and -1.
467 bool OOpIsAPInt = match(OOp, m_APInt(OOpC));
468 if (!isa<Constant>(OOp) ||
469 (OOpIsAPInt && isSelect01(C->getUniqueInteger(), *OOpC))) {
470 Value *NewSel = Builder.CreateSelect(
471 SI.getCondition(), Swapped ? C : OOp, Swapped ? OOp : C);
472 if (isa<FPMathOperator>(&SI))
473 cast<Instruction>(NewSel)->setFastMathFlags(FMF);
474 NewSel->takeName(TVI);
476 BinaryOperator::Create(TVI->getOpcode(), FalseVal, NewSel);
477 BO->copyIRFlags(TVI);
487 if (Instruction *R = TryFoldSelectIntoOp(SI, TrueVal, FalseVal, false))
490 if (Instruction *R = TryFoldSelectIntoOp(SI, FalseVal, TrueVal, true))
497 /// (select (icmp eq (and X, Y), 0), (and (lshr X, Z), 1), 1)
499 /// zext (icmp ne i32 (and X, (or Y, (shl 1, Z))), 0)
501 /// Z may be 0 if lshr is missing.
502 /// Worst-case scenario is that we will replace 5 instructions with 5 different
503 /// instructions, but we got rid of select.
504 static Instruction *foldSelectICmpAndAnd(Type *SelType, const ICmpInst *Cmp,
505 Value *TVal, Value *FVal,
506 InstCombiner::BuilderTy &Builder) {
507 if (!(Cmp->hasOneUse() && Cmp->getOperand(0)->hasOneUse() &&
508 Cmp->getPredicate() == ICmpInst::ICMP_EQ &&
509 match(Cmp->getOperand(1), m_Zero()) && match(FVal, m_One())))
512 // The TrueVal has general form of: and %B, 1
514 if (!match(TVal, m_OneUse(m_And(m_Value(B), m_One()))))
517 // Where %B may be optionally shifted: lshr %X, %Z.
519 const bool HasShift = match(B, m_OneUse(m_LShr(m_Value(X), m_Value(Z))));
521 // The shift must be valid.
522 // TODO: This restricts the fold to constant shift amounts. Is there a way to
523 // handle variable shifts safely? PR47012
525 !match(Z, m_SpecificInt_ICMP(CmpInst::ICMP_ULT,
526 APInt(SelType->getScalarSizeInBits(),
527 SelType->getScalarSizeInBits()))))
534 if (!match(Cmp->getOperand(0), m_c_And(m_Specific(X), m_Value(Y))))
537 // ((X & Y) == 0) ? ((X >> Z) & 1) : 1 --> (X & (Y | (1 << Z))) != 0
538 // ((X & Y) == 0) ? (X & 1) : 1 --> (X & (Y | 1)) != 0
539 Constant *One = ConstantInt::get(SelType, 1);
540 Value *MaskB = HasShift ? Builder.CreateShl(One, Z) : One;
541 Value *FullMask = Builder.CreateOr(Y, MaskB);
542 Value *MaskedX = Builder.CreateAnd(X, FullMask);
543 Value *ICmpNeZero = Builder.CreateIsNotNull(MaskedX);
544 return new ZExtInst(ICmpNeZero, SelType);
548 /// (select (icmp sgt x, C), lshr (X, Y), ashr (X, Y)); iff C s>= -1
549 /// (select (icmp slt x, C), ashr (X, Y), lshr (X, Y)); iff C s>= 0
552 static Value *foldSelectICmpLshrAshr(const ICmpInst *IC, Value *TrueVal,
554 InstCombiner::BuilderTy &Builder) {
555 ICmpInst::Predicate Pred = IC->getPredicate();
556 Value *CmpLHS = IC->getOperand(0);
557 Value *CmpRHS = IC->getOperand(1);
558 if (!CmpRHS->getType()->isIntOrIntVectorTy())
562 unsigned Bitwidth = CmpRHS->getType()->getScalarSizeInBits();
563 if ((Pred != ICmpInst::ICMP_SGT ||
565 m_SpecificInt_ICMP(ICmpInst::ICMP_SGE, APInt(Bitwidth, -1)))) &&
566 (Pred != ICmpInst::ICMP_SLT ||
568 m_SpecificInt_ICMP(ICmpInst::ICMP_SGE, APInt(Bitwidth, 0)))))
571 // Canonicalize so that ashr is in FalseVal.
572 if (Pred == ICmpInst::ICMP_SLT)
573 std::swap(TrueVal, FalseVal);
575 if (match(TrueVal, m_LShr(m_Value(X), m_Value(Y))) &&
576 match(FalseVal, m_AShr(m_Specific(X), m_Specific(Y))) &&
577 match(CmpLHS, m_Specific(X))) {
578 const auto *Ashr = cast<Instruction>(FalseVal);
579 // if lshr is not exact and ashr is, this new ashr must not be exact.
580 bool IsExact = Ashr->isExact() && cast<Instruction>(TrueVal)->isExact();
581 return Builder.CreateAShr(X, Y, IC->getName(), IsExact);
588 /// (select (icmp eq (and X, C1), 0), Y, (or Y, C2))
590 /// (or (shl (and X, C1), C3), Y)
592 /// C1 and C2 are both powers of 2
594 /// C3 = Log(C2) - Log(C1)
596 /// This transform handles cases where:
597 /// 1. The icmp predicate is inverted
598 /// 2. The select operands are reversed
599 /// 3. The magnitude of C2 and C1 are flipped
600 static Value *foldSelectICmpAndOr(const ICmpInst *IC, Value *TrueVal,
602 InstCombiner::BuilderTy &Builder) {
603 // Only handle integer compares. Also, if this is a vector select, we need a
605 if (!TrueVal->getType()->isIntOrIntVectorTy() ||
606 TrueVal->getType()->isVectorTy() != IC->getType()->isVectorTy())
609 Value *CmpLHS = IC->getOperand(0);
610 Value *CmpRHS = IC->getOperand(1);
615 bool NeedAnd = false;
616 if (IC->isEquality()) {
617 if (!match(CmpRHS, m_Zero()))
621 if (!match(CmpLHS, m_And(m_Value(), m_Power2(C1))))
625 C1Log = C1->logBase2();
626 IsEqualZero = IC->getPredicate() == ICmpInst::ICMP_EQ;
627 } else if (IC->getPredicate() == ICmpInst::ICMP_SLT ||
628 IC->getPredicate() == ICmpInst::ICMP_SGT) {
629 // We also need to recognize (icmp slt (trunc (X)), 0) and
630 // (icmp sgt (trunc (X)), -1).
631 IsEqualZero = IC->getPredicate() == ICmpInst::ICMP_SGT;
632 if ((IsEqualZero && !match(CmpRHS, m_AllOnes())) ||
633 (!IsEqualZero && !match(CmpRHS, m_Zero())))
636 if (!match(CmpLHS, m_OneUse(m_Trunc(m_Value(V)))))
639 C1Log = CmpLHS->getType()->getScalarSizeInBits() - 1;
646 bool OrOnTrueVal = false;
647 bool OrOnFalseVal = match(FalseVal, m_Or(m_Specific(TrueVal), m_Power2(C2)));
649 OrOnTrueVal = match(TrueVal, m_Or(m_Specific(FalseVal), m_Power2(C2)));
651 if (!OrOnFalseVal && !OrOnTrueVal)
654 Value *Y = OrOnFalseVal ? TrueVal : FalseVal;
656 unsigned C2Log = C2->logBase2();
658 bool NeedXor = (!IsEqualZero && OrOnFalseVal) || (IsEqualZero && OrOnTrueVal);
659 bool NeedShift = C1Log != C2Log;
660 bool NeedZExtTrunc = Y->getType()->getScalarSizeInBits() !=
661 V->getType()->getScalarSizeInBits();
663 // Make sure we don't create more instructions than we save.
664 Value *Or = OrOnFalseVal ? FalseVal : TrueVal;
665 if ((NeedShift + NeedXor + NeedZExtTrunc) >
666 (IC->hasOneUse() + Or->hasOneUse()))
670 // Insert the AND instruction on the input to the truncate.
671 APInt C1 = APInt::getOneBitSet(V->getType()->getScalarSizeInBits(), C1Log);
672 V = Builder.CreateAnd(V, ConstantInt::get(V->getType(), C1));
676 V = Builder.CreateZExtOrTrunc(V, Y->getType());
677 V = Builder.CreateShl(V, C2Log - C1Log);
678 } else if (C1Log > C2Log) {
679 V = Builder.CreateLShr(V, C1Log - C2Log);
680 V = Builder.CreateZExtOrTrunc(V, Y->getType());
682 V = Builder.CreateZExtOrTrunc(V, Y->getType());
685 V = Builder.CreateXor(V, *C2);
687 return Builder.CreateOr(V, Y);
690 /// Canonicalize a set or clear of a masked set of constant bits to
691 /// select-of-constants form.
692 static Instruction *foldSetClearBits(SelectInst &Sel,
693 InstCombiner::BuilderTy &Builder) {
694 Value *Cond = Sel.getCondition();
695 Value *T = Sel.getTrueValue();
696 Value *F = Sel.getFalseValue();
697 Type *Ty = Sel.getType();
699 const APInt *NotC, *C;
701 // Cond ? (X & ~C) : (X | C) --> (X & ~C) | (Cond ? 0 : C)
702 if (match(T, m_And(m_Value(X), m_APInt(NotC))) &&
703 match(F, m_OneUse(m_Or(m_Specific(X), m_APInt(C)))) && *NotC == ~(*C)) {
704 Constant *Zero = ConstantInt::getNullValue(Ty);
705 Constant *OrC = ConstantInt::get(Ty, *C);
706 Value *NewSel = Builder.CreateSelect(Cond, Zero, OrC, "masksel", &Sel);
707 return BinaryOperator::CreateOr(T, NewSel);
710 // Cond ? (X | C) : (X & ~C) --> (X & ~C) | (Cond ? C : 0)
711 if (match(F, m_And(m_Value(X), m_APInt(NotC))) &&
712 match(T, m_OneUse(m_Or(m_Specific(X), m_APInt(C)))) && *NotC == ~(*C)) {
713 Constant *Zero = ConstantInt::getNullValue(Ty);
714 Constant *OrC = ConstantInt::get(Ty, *C);
715 Value *NewSel = Builder.CreateSelect(Cond, OrC, Zero, "masksel", &Sel);
716 return BinaryOperator::CreateOr(F, NewSel);
722 // select (x == 0), 0, x * y --> freeze(y) * x
723 // select (y == 0), 0, x * y --> freeze(x) * y
724 // select (x == 0), undef, x * y --> freeze(y) * x
725 // select (x == undef), 0, x * y --> freeze(y) * x
726 // Usage of mul instead of 0 will make the result more poisonous,
727 // so the operand that was not checked in the condition should be frozen.
728 // The latter folding is applied only when a constant compared with x is
729 // is a vector consisting of 0 and undefs. If a constant compared with x
730 // is a scalar undefined value or undefined vector then an expression
731 // should be already folded into a constant.
732 static Instruction *foldSelectZeroOrMul(SelectInst &SI, InstCombinerImpl &IC) {
733 auto *CondVal = SI.getCondition();
734 auto *TrueVal = SI.getTrueValue();
735 auto *FalseVal = SI.getFalseValue();
737 ICmpInst::Predicate Predicate;
739 // Assuming that constant compared with zero is not undef (but it may be
740 // a vector with some undef elements). Otherwise (when a constant is undef)
741 // the select expression should be already simplified.
742 if (!match(CondVal, m_ICmp(Predicate, m_Value(X), m_Zero())) ||
743 !ICmpInst::isEquality(Predicate))
746 if (Predicate == ICmpInst::ICMP_NE)
747 std::swap(TrueVal, FalseVal);
749 // Check that TrueVal is a constant instead of matching it with m_Zero()
750 // to handle the case when it is a scalar undef value or a vector containing
751 // non-zero elements that are masked by undef elements in the compare
753 auto *TrueValC = dyn_cast<Constant>(TrueVal);
754 if (TrueValC == nullptr ||
755 !match(FalseVal, m_c_Mul(m_Specific(X), m_Value(Y))) ||
756 !isa<Instruction>(FalseVal))
759 auto *ZeroC = cast<Constant>(cast<Instruction>(CondVal)->getOperand(1));
760 auto *MergedC = Constant::mergeUndefsWith(TrueValC, ZeroC);
761 // If X is compared with 0 then TrueVal could be either zero or undef.
762 // m_Zero match vectors containing some undef elements, but for scalars
763 // m_Undef should be used explicitly.
764 if (!match(MergedC, m_Zero()) && !match(MergedC, m_Undef()))
767 auto *FalseValI = cast<Instruction>(FalseVal);
768 auto *FrY = IC.InsertNewInstBefore(new FreezeInst(Y, Y->getName() + ".fr"),
770 IC.replaceOperand(*FalseValI, FalseValI->getOperand(0) == Y ? 0 : 1, FrY);
771 return IC.replaceInstUsesWith(SI, FalseValI);
774 /// Transform patterns such as (a > b) ? a - b : 0 into usub.sat(a, b).
775 /// There are 8 commuted/swapped variants of this pattern.
776 /// TODO: Also support a - UMIN(a,b) patterns.
777 static Value *canonicalizeSaturatedSubtract(const ICmpInst *ICI,
778 const Value *TrueVal,
779 const Value *FalseVal,
780 InstCombiner::BuilderTy &Builder) {
781 ICmpInst::Predicate Pred = ICI->getPredicate();
782 if (!ICmpInst::isUnsigned(Pred))
785 // (b > a) ? 0 : a - b -> (b <= a) ? a - b : 0
786 if (match(TrueVal, m_Zero())) {
787 Pred = ICmpInst::getInversePredicate(Pred);
788 std::swap(TrueVal, FalseVal);
790 if (!match(FalseVal, m_Zero()))
793 Value *A = ICI->getOperand(0);
794 Value *B = ICI->getOperand(1);
795 if (Pred == ICmpInst::ICMP_ULE || Pred == ICmpInst::ICMP_ULT) {
796 // (b < a) ? a - b : 0 -> (a > b) ? a - b : 0
798 Pred = ICmpInst::getSwappedPredicate(Pred);
801 assert((Pred == ICmpInst::ICMP_UGE || Pred == ICmpInst::ICMP_UGT) &&
802 "Unexpected isUnsigned predicate!");
804 // Ensure the sub is of the form:
805 // (a > b) ? a - b : 0 -> usub.sat(a, b)
806 // (a > b) ? b - a : 0 -> -usub.sat(a, b)
807 // Checking for both a-b and a+(-b) as a constant.
808 bool IsNegative = false;
810 if (match(TrueVal, m_Sub(m_Specific(B), m_Specific(A))) ||
811 (match(A, m_APInt(C)) &&
812 match(TrueVal, m_Add(m_Specific(B), m_SpecificInt(-*C)))))
814 else if (!match(TrueVal, m_Sub(m_Specific(A), m_Specific(B))) &&
815 !(match(B, m_APInt(C)) &&
816 match(TrueVal, m_Add(m_Specific(A), m_SpecificInt(-*C)))))
819 // If we are adding a negate and the sub and icmp are used anywhere else, we
820 // would end up with more instructions.
821 if (IsNegative && !TrueVal->hasOneUse() && !ICI->hasOneUse())
824 // (a > b) ? a - b : 0 -> usub.sat(a, b)
825 // (a > b) ? b - a : 0 -> -usub.sat(a, b)
826 Value *Result = Builder.CreateBinaryIntrinsic(Intrinsic::usub_sat, A, B);
828 Result = Builder.CreateNeg(Result);
832 static Value *canonicalizeSaturatedAdd(ICmpInst *Cmp, Value *TVal, Value *FVal,
833 InstCombiner::BuilderTy &Builder) {
834 if (!Cmp->hasOneUse())
837 // Match unsigned saturated add with constant.
838 Value *Cmp0 = Cmp->getOperand(0);
839 Value *Cmp1 = Cmp->getOperand(1);
840 ICmpInst::Predicate Pred = Cmp->getPredicate();
842 const APInt *C, *CmpC;
843 if (Pred == ICmpInst::ICMP_ULT &&
844 match(TVal, m_Add(m_Value(X), m_APInt(C))) && X == Cmp0 &&
845 match(FVal, m_AllOnes()) && match(Cmp1, m_APInt(CmpC)) && *CmpC == ~*C) {
846 // (X u< ~C) ? (X + C) : -1 --> uadd.sat(X, C)
847 return Builder.CreateBinaryIntrinsic(
848 Intrinsic::uadd_sat, X, ConstantInt::get(X->getType(), *C));
851 // Match unsigned saturated add of 2 variables with an unnecessary 'not'.
852 // There are 8 commuted variants.
853 // Canonicalize -1 (saturated result) to true value of the select.
854 if (match(FVal, m_AllOnes())) {
855 std::swap(TVal, FVal);
856 Pred = CmpInst::getInversePredicate(Pred);
858 if (!match(TVal, m_AllOnes()))
861 // Canonicalize predicate to less-than or less-or-equal-than.
862 if (Pred == ICmpInst::ICMP_UGT || Pred == ICmpInst::ICMP_UGE) {
863 std::swap(Cmp0, Cmp1);
864 Pred = CmpInst::getSwappedPredicate(Pred);
866 if (Pred != ICmpInst::ICMP_ULT && Pred != ICmpInst::ICMP_ULE)
869 // Match unsigned saturated add of 2 variables with an unnecessary 'not'.
870 // Strictness of the comparison is irrelevant.
872 if (match(Cmp0, m_Not(m_Value(X))) &&
873 match(FVal, m_c_Add(m_Specific(X), m_Value(Y))) && Y == Cmp1) {
874 // (~X u< Y) ? -1 : (X + Y) --> uadd.sat(X, Y)
875 // (~X u< Y) ? -1 : (Y + X) --> uadd.sat(X, Y)
876 return Builder.CreateBinaryIntrinsic(Intrinsic::uadd_sat, X, Y);
878 // The 'not' op may be included in the sum but not the compare.
879 // Strictness of the comparison is irrelevant.
882 if (match(FVal, m_c_Add(m_Not(m_Specific(X)), m_Specific(Y)))) {
883 // (X u< Y) ? -1 : (~X + Y) --> uadd.sat(~X, Y)
884 // (X u< Y) ? -1 : (Y + ~X) --> uadd.sat(Y, ~X)
885 BinaryOperator *BO = cast<BinaryOperator>(FVal);
886 return Builder.CreateBinaryIntrinsic(
887 Intrinsic::uadd_sat, BO->getOperand(0), BO->getOperand(1));
889 // The overflow may be detected via the add wrapping round.
890 // This is only valid for strict comparison!
891 if (Pred == ICmpInst::ICMP_ULT &&
892 match(Cmp0, m_c_Add(m_Specific(Cmp1), m_Value(Y))) &&
893 match(FVal, m_c_Add(m_Specific(Cmp1), m_Specific(Y)))) {
894 // ((X + Y) u< X) ? -1 : (X + Y) --> uadd.sat(X, Y)
895 // ((X + Y) u< Y) ? -1 : (X + Y) --> uadd.sat(X, Y)
896 return Builder.CreateBinaryIntrinsic(Intrinsic::uadd_sat, Cmp1, Y);
902 /// Fold the following code sequence:
904 /// int a = ctlz(x & -x);
910 static Instruction *foldSelectCtlzToCttz(ICmpInst *ICI, Value *TrueVal,
912 InstCombiner::BuilderTy &Builder) {
913 unsigned BitWidth = TrueVal->getType()->getScalarSizeInBits();
914 if (!ICI->isEquality() || !match(ICI->getOperand(1), m_Zero()))
917 if (ICI->getPredicate() == ICmpInst::ICMP_NE)
918 std::swap(TrueVal, FalseVal);
921 m_Xor(m_Deferred(TrueVal), m_SpecificInt(BitWidth - 1))))
924 if (!match(TrueVal, m_Intrinsic<Intrinsic::ctlz>()))
927 Value *X = ICI->getOperand(0);
928 auto *II = cast<IntrinsicInst>(TrueVal);
929 if (!match(II->getOperand(0), m_c_And(m_Specific(X), m_Neg(m_Specific(X)))))
932 Function *F = Intrinsic::getDeclaration(II->getModule(), Intrinsic::cttz,
934 return CallInst::Create(F, {X, II->getArgOperand(1)});
937 /// Attempt to fold a cttz/ctlz followed by a icmp plus select into a single
938 /// call to cttz/ctlz with flag 'is_zero_poison' cleared.
940 /// For example, we can fold the following code sequence:
942 /// %0 = tail call i32 @llvm.cttz.i32(i32 %x, i1 true)
943 /// %1 = icmp ne i32 %x, 0
944 /// %2 = select i1 %1, i32 %0, i32 32
948 /// %0 = tail call i32 @llvm.cttz.i32(i32 %x, i1 false)
949 static Value *foldSelectCttzCtlz(ICmpInst *ICI, Value *TrueVal, Value *FalseVal,
950 InstCombiner::BuilderTy &Builder) {
951 ICmpInst::Predicate Pred = ICI->getPredicate();
952 Value *CmpLHS = ICI->getOperand(0);
953 Value *CmpRHS = ICI->getOperand(1);
955 // Check if the condition value compares a value for equality against zero.
956 if (!ICI->isEquality() || !match(CmpRHS, m_Zero()))
959 Value *SelectArg = FalseVal;
960 Value *ValueOnZero = TrueVal;
961 if (Pred == ICmpInst::ICMP_NE)
962 std::swap(SelectArg, ValueOnZero);
964 // Skip zero extend/truncate.
965 Value *Count = nullptr;
966 if (!match(SelectArg, m_ZExt(m_Value(Count))) &&
967 !match(SelectArg, m_Trunc(m_Value(Count))))
970 // Check that 'Count' is a call to intrinsic cttz/ctlz. Also check that the
971 // input to the cttz/ctlz is used as LHS for the compare instruction.
972 if (!match(Count, m_Intrinsic<Intrinsic::cttz>(m_Specific(CmpLHS))) &&
973 !match(Count, m_Intrinsic<Intrinsic::ctlz>(m_Specific(CmpLHS))))
976 IntrinsicInst *II = cast<IntrinsicInst>(Count);
978 // Check if the value propagated on zero is a constant number equal to the
979 // sizeof in bits of 'Count'.
980 unsigned SizeOfInBits = Count->getType()->getScalarSizeInBits();
981 if (match(ValueOnZero, m_SpecificInt(SizeOfInBits))) {
982 // Explicitly clear the 'is_zero_poison' flag. It's always valid to go from
983 // true to false on this flag, so we can replace it for all users.
984 II->setArgOperand(1, ConstantInt::getFalse(II->getContext()));
988 // The ValueOnZero is not the bitwidth. But if the cttz/ctlz (and optional
989 // zext/trunc) have one use (ending at the select), the cttz/ctlz result will
990 // not be used if the input is zero. Relax to 'zero is poison' for that case.
991 if (II->hasOneUse() && SelectArg->hasOneUse() &&
992 !match(II->getArgOperand(1), m_One()))
993 II->setArgOperand(1, ConstantInt::getTrue(II->getContext()));
998 /// Return true if we find and adjust an icmp+select pattern where the compare
999 /// is with a constant that can be incremented or decremented to match the
1000 /// minimum or maximum idiom.
1001 static bool adjustMinMax(SelectInst &Sel, ICmpInst &Cmp) {
1002 ICmpInst::Predicate Pred = Cmp.getPredicate();
1003 Value *CmpLHS = Cmp.getOperand(0);
1004 Value *CmpRHS = Cmp.getOperand(1);
1005 Value *TrueVal = Sel.getTrueValue();
1006 Value *FalseVal = Sel.getFalseValue();
1008 // We may move or edit the compare, so make sure the select is the only user.
1010 if (!Cmp.hasOneUse() || !match(CmpRHS, m_APInt(CmpC)))
1013 // These transforms only work for selects of integers or vector selects of
1015 Type *SelTy = Sel.getType();
1016 auto *SelEltTy = dyn_cast<IntegerType>(SelTy->getScalarType());
1017 if (!SelEltTy || SelTy->isVectorTy() != Cmp.getType()->isVectorTy())
1020 Constant *AdjustedRHS;
1021 if (Pred == ICmpInst::ICMP_UGT || Pred == ICmpInst::ICMP_SGT)
1022 AdjustedRHS = ConstantInt::get(CmpRHS->getType(), *CmpC + 1);
1023 else if (Pred == ICmpInst::ICMP_ULT || Pred == ICmpInst::ICMP_SLT)
1024 AdjustedRHS = ConstantInt::get(CmpRHS->getType(), *CmpC - 1);
1028 // X > C ? X : C+1 --> X < C+1 ? C+1 : X
1029 // X < C ? X : C-1 --> X > C-1 ? C-1 : X
1030 if ((CmpLHS == TrueVal && AdjustedRHS == FalseVal) ||
1031 (CmpLHS == FalseVal && AdjustedRHS == TrueVal)) {
1032 ; // Nothing to do here. Values match without any sign/zero extension.
1034 // Types do not match. Instead of calculating this with mixed types, promote
1035 // all to the larger type. This enables scalar evolution to analyze this
1037 else if (CmpRHS->getType()->getScalarSizeInBits() < SelEltTy->getBitWidth()) {
1038 Constant *SextRHS = ConstantExpr::getSExt(AdjustedRHS, SelTy);
1040 // X = sext x; x >s c ? X : C+1 --> X = sext x; X <s C+1 ? C+1 : X
1041 // X = sext x; x <s c ? X : C-1 --> X = sext x; X >s C-1 ? C-1 : X
1042 // X = sext x; x >u c ? X : C+1 --> X = sext x; X <u C+1 ? C+1 : X
1043 // X = sext x; x <u c ? X : C-1 --> X = sext x; X >u C-1 ? C-1 : X
1044 if (match(TrueVal, m_SExt(m_Specific(CmpLHS))) && SextRHS == FalseVal) {
1046 AdjustedRHS = SextRHS;
1047 } else if (match(FalseVal, m_SExt(m_Specific(CmpLHS))) &&
1048 SextRHS == TrueVal) {
1050 AdjustedRHS = SextRHS;
1051 } else if (Cmp.isUnsigned()) {
1052 Constant *ZextRHS = ConstantExpr::getZExt(AdjustedRHS, SelTy);
1053 // X = zext x; x >u c ? X : C+1 --> X = zext x; X <u C+1 ? C+1 : X
1054 // X = zext x; x <u c ? X : C-1 --> X = zext x; X >u C-1 ? C-1 : X
1055 // zext + signed compare cannot be changed:
1056 // 0xff <s 0x00, but 0x00ff >s 0x0000
1057 if (match(TrueVal, m_ZExt(m_Specific(CmpLHS))) && ZextRHS == FalseVal) {
1059 AdjustedRHS = ZextRHS;
1060 } else if (match(FalseVal, m_ZExt(m_Specific(CmpLHS))) &&
1061 ZextRHS == TrueVal) {
1063 AdjustedRHS = ZextRHS;
1074 Pred = ICmpInst::getSwappedPredicate(Pred);
1075 CmpRHS = AdjustedRHS;
1076 std::swap(FalseVal, TrueVal);
1077 Cmp.setPredicate(Pred);
1078 Cmp.setOperand(0, CmpLHS);
1079 Cmp.setOperand(1, CmpRHS);
1080 Sel.setOperand(1, TrueVal);
1081 Sel.setOperand(2, FalseVal);
1082 Sel.swapProfMetadata();
1084 // Move the compare instruction right before the select instruction. Otherwise
1085 // the sext/zext value may be defined after the compare instruction uses it.
1086 Cmp.moveBefore(&Sel);
1091 static Instruction *canonicalizeSPF(SelectInst &Sel, ICmpInst &Cmp,
1092 InstCombinerImpl &IC) {
1094 // TODO: What to do with pointer min/max patterns?
1095 if (!Sel.getType()->isIntOrIntVectorTy())
1098 SelectPatternFlavor SPF = matchSelectPattern(&Sel, LHS, RHS).Flavor;
1099 if (SPF == SelectPatternFlavor::SPF_ABS ||
1100 SPF == SelectPatternFlavor::SPF_NABS) {
1101 if (!Cmp.hasOneUse() && !RHS->hasOneUse())
1102 return nullptr; // TODO: Relax this restriction.
1104 // Note that NSW flag can only be propagated for normal, non-negated abs!
1105 bool IntMinIsPoison = SPF == SelectPatternFlavor::SPF_ABS &&
1106 match(RHS, m_NSWNeg(m_Specific(LHS)));
1107 Constant *IntMinIsPoisonC =
1108 ConstantInt::get(Type::getInt1Ty(Sel.getContext()), IntMinIsPoison);
1110 IC.Builder.CreateBinaryIntrinsic(Intrinsic::abs, LHS, IntMinIsPoisonC);
1112 if (SPF == SelectPatternFlavor::SPF_NABS)
1113 return BinaryOperator::CreateNeg(Abs); // Always without NSW flag!
1114 return IC.replaceInstUsesWith(Sel, Abs);
1117 if (SelectPatternResult::isMinOrMax(SPF)) {
1118 Intrinsic::ID IntrinsicID;
1120 case SelectPatternFlavor::SPF_UMIN:
1121 IntrinsicID = Intrinsic::umin;
1123 case SelectPatternFlavor::SPF_UMAX:
1124 IntrinsicID = Intrinsic::umax;
1126 case SelectPatternFlavor::SPF_SMIN:
1127 IntrinsicID = Intrinsic::smin;
1129 case SelectPatternFlavor::SPF_SMAX:
1130 IntrinsicID = Intrinsic::smax;
1133 llvm_unreachable("Unexpected SPF");
1135 return IC.replaceInstUsesWith(
1136 Sel, IC.Builder.CreateBinaryIntrinsic(IntrinsicID, LHS, RHS));
1142 /// If we have a select with an equality comparison, then we know the value in
1143 /// one of the arms of the select. See if substituting this value into an arm
1144 /// and simplifying the result yields the same value as the other arm.
1146 /// To make this transform safe, we must drop poison-generating flags
1147 /// (nsw, etc) if we simplified to a binop because the select may be guarding
1148 /// that poison from propagating. If the existing binop already had no
1149 /// poison-generating flags, then this transform can be done by instsimplify.
1152 /// %cmp = icmp eq i32 %x, 2147483647
1153 /// %add = add nsw i32 %x, 1
1154 /// %sel = select i1 %cmp, i32 -2147483648, i32 %add
1156 /// We can't replace %sel with %add unless we strip away the flags.
1157 /// TODO: Wrapping flags could be preserved in some cases with better analysis.
1158 Instruction *InstCombinerImpl::foldSelectValueEquivalence(SelectInst &Sel,
1160 // Value equivalence substitution requires an all-or-nothing replacement.
1161 // It does not make sense for a vector compare where each lane is chosen
1163 if (!Cmp.isEquality() || Cmp.getType()->isVectorTy())
1166 // Canonicalize the pattern to ICMP_EQ by swapping the select operands.
1167 Value *TrueVal = Sel.getTrueValue(), *FalseVal = Sel.getFalseValue();
1168 bool Swapped = false;
1169 if (Cmp.getPredicate() == ICmpInst::ICMP_NE) {
1170 std::swap(TrueVal, FalseVal);
1174 // In X == Y ? f(X) : Z, try to evaluate f(Y) and replace the operand.
1175 // Make sure Y cannot be undef though, as we might pick different values for
1176 // undef in the icmp and in f(Y). Additionally, take care to avoid replacing
1177 // X == Y ? X : Z with X == Y ? Y : Z, as that would lead to an infinite
1178 // replacement cycle.
1179 Value *CmpLHS = Cmp.getOperand(0), *CmpRHS = Cmp.getOperand(1);
1180 if (TrueVal != CmpLHS &&
1181 isGuaranteedNotToBeUndefOrPoison(CmpRHS, SQ.AC, &Sel, &DT)) {
1182 if (Value *V = simplifyWithOpReplaced(TrueVal, CmpLHS, CmpRHS, SQ,
1183 /* AllowRefinement */ true))
1184 return replaceOperand(Sel, Swapped ? 2 : 1, V);
1186 // Even if TrueVal does not simplify, we can directly replace a use of
1187 // CmpLHS with CmpRHS, as long as the instruction is not used anywhere
1188 // else and is safe to speculatively execute (we may end up executing it
1189 // with different operands, which should not cause side-effects or trigger
1190 // undefined behavior). Only do this if CmpRHS is a constant, as
1191 // profitability is not clear for other cases.
1192 // FIXME: The replacement could be performed recursively.
1193 if (match(CmpRHS, m_ImmConstant()) && !match(CmpLHS, m_ImmConstant()))
1194 if (auto *I = dyn_cast<Instruction>(TrueVal))
1195 if (I->hasOneUse() && isSafeToSpeculativelyExecute(I))
1196 for (Use &U : I->operands())
1198 replaceUse(U, CmpRHS);
1202 if (TrueVal != CmpRHS &&
1203 isGuaranteedNotToBeUndefOrPoison(CmpLHS, SQ.AC, &Sel, &DT))
1204 if (Value *V = simplifyWithOpReplaced(TrueVal, CmpRHS, CmpLHS, SQ,
1205 /* AllowRefinement */ true))
1206 return replaceOperand(Sel, Swapped ? 2 : 1, V);
1208 auto *FalseInst = dyn_cast<Instruction>(FalseVal);
1212 // InstSimplify already performed this fold if it was possible subject to
1213 // current poison-generating flags. Try the transform again with
1214 // poison-generating flags temporarily dropped.
1215 bool WasNUW = false, WasNSW = false, WasExact = false, WasInBounds = false;
1216 if (auto *OBO = dyn_cast<OverflowingBinaryOperator>(FalseVal)) {
1217 WasNUW = OBO->hasNoUnsignedWrap();
1218 WasNSW = OBO->hasNoSignedWrap();
1219 FalseInst->setHasNoUnsignedWrap(false);
1220 FalseInst->setHasNoSignedWrap(false);
1222 if (auto *PEO = dyn_cast<PossiblyExactOperator>(FalseVal)) {
1223 WasExact = PEO->isExact();
1224 FalseInst->setIsExact(false);
1226 if (auto *GEP = dyn_cast<GetElementPtrInst>(FalseVal)) {
1227 WasInBounds = GEP->isInBounds();
1228 GEP->setIsInBounds(false);
1231 // Try each equivalence substitution possibility.
1232 // We have an 'EQ' comparison, so the select's false value will propagate.
1234 // (X == 42) ? 43 : (X + 1) --> (X == 42) ? (X + 1) : (X + 1) --> X + 1
1235 if (simplifyWithOpReplaced(FalseVal, CmpLHS, CmpRHS, SQ,
1236 /* AllowRefinement */ false) == TrueVal ||
1237 simplifyWithOpReplaced(FalseVal, CmpRHS, CmpLHS, SQ,
1238 /* AllowRefinement */ false) == TrueVal) {
1239 return replaceInstUsesWith(Sel, FalseVal);
1242 // Restore poison-generating flags if the transform did not apply.
1244 FalseInst->setHasNoUnsignedWrap();
1246 FalseInst->setHasNoSignedWrap();
1248 FalseInst->setIsExact();
1250 cast<GetElementPtrInst>(FalseInst)->setIsInBounds();
1255 // See if this is a pattern like:
1256 // %old_cmp1 = icmp slt i32 %x, C2
1257 // %old_replacement = select i1 %old_cmp1, i32 %target_low, i32 %target_high
1258 // %old_x_offseted = add i32 %x, C1
1259 // %old_cmp0 = icmp ult i32 %old_x_offseted, C0
1260 // %r = select i1 %old_cmp0, i32 %x, i32 %old_replacement
1261 // This can be rewritten as more canonical pattern:
1262 // %new_cmp1 = icmp slt i32 %x, -C1
1263 // %new_cmp2 = icmp sge i32 %x, C0-C1
1264 // %new_clamped_low = select i1 %new_cmp1, i32 %target_low, i32 %x
1265 // %r = select i1 %new_cmp2, i32 %target_high, i32 %new_clamped_low
1266 // Iff -C1 s<= C2 s<= C0-C1
1267 // Also ULT predicate can also be UGT iff C0 != -1 (+invert result)
1268 // SLT predicate can also be SGT iff C2 != INT_MAX (+invert res.)
1269 static Value *canonicalizeClampLike(SelectInst &Sel0, ICmpInst &Cmp0,
1270 InstCombiner::BuilderTy &Builder) {
1271 Value *X = Sel0.getTrueValue();
1272 Value *Sel1 = Sel0.getFalseValue();
1274 // First match the condition of the outermost select.
1275 // Said condition must be one-use.
1276 if (!Cmp0.hasOneUse())
1278 ICmpInst::Predicate Pred0 = Cmp0.getPredicate();
1279 Value *Cmp00 = Cmp0.getOperand(0);
1281 if (!match(Cmp0.getOperand(1),
1282 m_CombineAnd(m_AnyIntegralConstant(), m_Constant(C0))))
1285 if (!isa<SelectInst>(Sel1)) {
1286 Pred0 = ICmpInst::getInversePredicate(Pred0);
1290 // Canonicalize Cmp0 into ult or uge.
1291 // FIXME: we shouldn't care about lanes that are 'undef' in the end?
1293 case ICmpInst::Predicate::ICMP_ULT:
1294 case ICmpInst::Predicate::ICMP_UGE:
1295 // Although icmp ult %x, 0 is an unusual thing to try and should generally
1296 // have been simplified, it does not verify with undef inputs so ensure we
1297 // are not in a strange state.
1298 if (!match(C0, m_SpecificInt_ICMP(
1299 ICmpInst::Predicate::ICMP_NE,
1300 APInt::getZero(C0->getType()->getScalarSizeInBits()))))
1303 case ICmpInst::Predicate::ICMP_ULE:
1304 case ICmpInst::Predicate::ICMP_UGT:
1305 // We want to canonicalize it to 'ult' or 'uge', so we'll need to increment
1306 // C0, which again means it must not have any all-ones elements.
1309 ICmpInst::Predicate::ICMP_NE,
1310 APInt::getAllOnes(C0->getType()->getScalarSizeInBits()))))
1311 return nullptr; // Can't do, have all-ones element[s].
1312 Pred0 = ICmpInst::getFlippedStrictnessPredicate(Pred0);
1313 C0 = InstCombiner::AddOne(C0);
1316 return nullptr; // Unknown predicate.
1319 // Now that we've canonicalized the ICmp, we know the X we expect;
1320 // the select in other hand should be one-use.
1321 if (!Sel1->hasOneUse())
1324 // If the types do not match, look through any truncs to the underlying
1326 if (Cmp00->getType() != X->getType() && X->hasOneUse())
1327 match(X, m_TruncOrSelf(m_Value(X)));
1329 // We now can finish matching the condition of the outermost select:
1330 // it should either be the X itself, or an addition of some constant to X.
1333 C1 = ConstantInt::getNullValue(X->getType());
1334 else if (!match(Cmp00,
1335 m_Add(m_Specific(X),
1336 m_CombineAnd(m_AnyIntegralConstant(), m_Constant(C1)))))
1340 ICmpInst::Predicate Pred1;
1342 Value *ReplacementLow, *ReplacementHigh;
1343 if (!match(Sel1, m_Select(m_Value(Cmp1), m_Value(ReplacementLow),
1344 m_Value(ReplacementHigh))) ||
1346 m_ICmp(Pred1, m_Specific(X),
1347 m_CombineAnd(m_AnyIntegralConstant(), m_Constant(C2)))))
1350 if (!Cmp1->hasOneUse() && (Cmp00 == X || !Cmp00->hasOneUse()))
1351 return nullptr; // Not enough one-use instructions for the fold.
1352 // FIXME: this restriction could be relaxed if Cmp1 can be reused as one of
1353 // two comparisons we'll need to build.
1355 // Canonicalize Cmp1 into the form we expect.
1356 // FIXME: we shouldn't care about lanes that are 'undef' in the end?
1358 case ICmpInst::Predicate::ICMP_SLT:
1360 case ICmpInst::Predicate::ICMP_SLE:
1361 // We'd have to increment C2 by one, and for that it must not have signed
1362 // max element, but then it would have been canonicalized to 'slt' before
1363 // we get here. So we can't do anything useful with 'sle'.
1365 case ICmpInst::Predicate::ICMP_SGT:
1366 // We want to canonicalize it to 'slt', so we'll need to increment C2,
1367 // which again means it must not have any signed max elements.
1369 m_SpecificInt_ICMP(ICmpInst::Predicate::ICMP_NE,
1370 APInt::getSignedMaxValue(
1371 C2->getType()->getScalarSizeInBits()))))
1372 return nullptr; // Can't do, have signed max element[s].
1373 C2 = InstCombiner::AddOne(C2);
1375 case ICmpInst::Predicate::ICMP_SGE:
1376 // Also non-canonical, but here we don't need to change C2,
1377 // so we don't have any restrictions on C2, so we can just handle it.
1378 Pred1 = ICmpInst::Predicate::ICMP_SLT;
1379 std::swap(ReplacementLow, ReplacementHigh);
1382 return nullptr; // Unknown predicate.
1384 assert(Pred1 == ICmpInst::Predicate::ICMP_SLT &&
1385 "Unexpected predicate type.");
1387 // The thresholds of this clamp-like pattern.
1388 auto *ThresholdLowIncl = ConstantExpr::getNeg(C1);
1389 auto *ThresholdHighExcl = ConstantExpr::getSub(C0, C1);
1391 assert((Pred0 == ICmpInst::Predicate::ICMP_ULT ||
1392 Pred0 == ICmpInst::Predicate::ICMP_UGE) &&
1393 "Unexpected predicate type.");
1394 if (Pred0 == ICmpInst::Predicate::ICMP_UGE)
1395 std::swap(ThresholdLowIncl, ThresholdHighExcl);
1397 // The fold has a precondition 1: C2 s>= ThresholdLow
1398 auto *Precond1 = ConstantExpr::getICmp(ICmpInst::Predicate::ICMP_SGE, C2,
1400 if (!match(Precond1, m_One()))
1402 // The fold has a precondition 2: C2 s<= ThresholdHigh
1403 auto *Precond2 = ConstantExpr::getICmp(ICmpInst::Predicate::ICMP_SLE, C2,
1405 if (!match(Precond2, m_One()))
1408 // If we are matching from a truncated input, we need to sext the
1409 // ReplacementLow and ReplacementHigh values. Only do the transform if they
1410 // are free to extend due to being constants.
1411 if (X->getType() != Sel0.getType()) {
1412 Constant *LowC, *HighC;
1413 if (!match(ReplacementLow, m_ImmConstant(LowC)) ||
1414 !match(ReplacementHigh, m_ImmConstant(HighC)))
1416 ReplacementLow = ConstantExpr::getSExt(LowC, X->getType());
1417 ReplacementHigh = ConstantExpr::getSExt(HighC, X->getType());
1420 // All good, finally emit the new pattern.
1421 Value *ShouldReplaceLow = Builder.CreateICmpSLT(X, ThresholdLowIncl);
1422 Value *ShouldReplaceHigh = Builder.CreateICmpSGE(X, ThresholdHighExcl);
1423 Value *MaybeReplacedLow =
1424 Builder.CreateSelect(ShouldReplaceLow, ReplacementLow, X);
1426 // Create the final select. If we looked through a truncate above, we will
1427 // need to retruncate the result.
1428 Value *MaybeReplacedHigh = Builder.CreateSelect(
1429 ShouldReplaceHigh, ReplacementHigh, MaybeReplacedLow);
1430 return Builder.CreateTrunc(MaybeReplacedHigh, Sel0.getType());
1434 // %cmp = icmp [canonical predicate] i32 %x, C0
1435 // %r = select i1 %cmp, i32 %y, i32 C1
1436 // Where C0 != C1 and %x may be different from %y, see if the constant that we
1437 // will have if we flip the strictness of the predicate (i.e. without changing
1438 // the result) is identical to the C1 in select. If it matches we can change
1439 // original comparison to one with swapped predicate, reuse the constant,
1440 // and swap the hands of select.
1441 static Instruction *
1442 tryToReuseConstantFromSelectInComparison(SelectInst &Sel, ICmpInst &Cmp,
1443 InstCombinerImpl &IC) {
1444 ICmpInst::Predicate Pred;
1447 if (!match(&Cmp, m_OneUse(m_ICmp(
1449 m_CombineAnd(m_AnyIntegralConstant(), m_Constant(C0))))))
1452 // If comparison predicate is non-relational, we won't be able to do anything.
1453 if (ICmpInst::isEquality(Pred))
1456 // If comparison predicate is non-canonical, then we certainly won't be able
1457 // to make it canonical; canonicalizeCmpWithConstant() already tried.
1458 if (!InstCombiner::isCanonicalPredicate(Pred))
1461 // If the [input] type of comparison and select type are different, lets abort
1462 // for now. We could try to compare constants with trunc/[zs]ext though.
1463 if (C0->getType() != Sel.getType())
1466 // ULT with 'add' of a constant is canonical. See foldICmpAddConstant().
1467 // FIXME: Are there more magic icmp predicate+constant pairs we must avoid?
1468 // Or should we just abandon this transform entirely?
1469 if (Pred == CmpInst::ICMP_ULT && match(X, m_Add(m_Value(), m_Constant())))
1473 Value *SelVal0, *SelVal1; // We do not care which one is from where.
1474 match(&Sel, m_Select(m_Value(), m_Value(SelVal0), m_Value(SelVal1)));
1475 // At least one of these values we are selecting between must be a constant
1476 // else we'll never succeed.
1477 if (!match(SelVal0, m_AnyIntegralConstant()) &&
1478 !match(SelVal1, m_AnyIntegralConstant()))
1481 // Does this constant C match any of the `select` values?
1482 auto MatchesSelectValue = [SelVal0, SelVal1](Constant *C) {
1483 return C->isElementWiseEqual(SelVal0) || C->isElementWiseEqual(SelVal1);
1486 // If C0 *already* matches true/false value of select, we are done.
1487 if (MatchesSelectValue(C0))
1490 // Check the constant we'd have with flipped-strictness predicate.
1491 auto FlippedStrictness =
1492 InstCombiner::getFlippedStrictnessPredicateAndConstant(Pred, C0);
1493 if (!FlippedStrictness)
1496 // If said constant doesn't match either, then there is no hope,
1497 if (!MatchesSelectValue(FlippedStrictness->second))
1500 // It matched! Lets insert the new comparison just before select.
1501 InstCombiner::BuilderTy::InsertPointGuard Guard(IC.Builder);
1502 IC.Builder.SetInsertPoint(&Sel);
1504 Pred = ICmpInst::getSwappedPredicate(Pred); // Yes, swapped.
1505 Value *NewCmp = IC.Builder.CreateICmp(Pred, X, FlippedStrictness->second,
1506 Cmp.getName() + ".inv");
1507 IC.replaceOperand(Sel, 0, NewCmp);
1509 Sel.swapProfMetadata();
1514 static Instruction *foldSelectZeroOrOnes(ICmpInst *Cmp, Value *TVal,
1516 InstCombiner::BuilderTy &Builder) {
1517 if (!Cmp->hasOneUse())
1521 if (!match(Cmp->getOperand(1), m_APIntAllowUndef(CmpC)))
1524 // (X u< 2) ? -X : -1 --> sext (X != 0)
1525 Value *X = Cmp->getOperand(0);
1526 if (Cmp->getPredicate() == ICmpInst::ICMP_ULT && *CmpC == 2 &&
1527 match(TVal, m_Neg(m_Specific(X))) && match(FVal, m_AllOnes()))
1528 return new SExtInst(Builder.CreateIsNotNull(X), TVal->getType());
1530 // (X u> 1) ? -1 : -X --> sext (X != 0)
1531 if (Cmp->getPredicate() == ICmpInst::ICMP_UGT && *CmpC == 1 &&
1532 match(FVal, m_Neg(m_Specific(X))) && match(TVal, m_AllOnes()))
1533 return new SExtInst(Builder.CreateIsNotNull(X), TVal->getType());
1538 static Value *foldSelectInstWithICmpConst(SelectInst &SI, ICmpInst *ICI) {
1541 CmpInst::Predicate Pred;
1542 if (!match(ICI, m_ICmp(Pred, m_Value(V), m_APInt(CmpC))))
1547 CmpInst::Predicate CPred;
1548 if (match(&SI, m_Select(m_Specific(ICI), m_APInt(C), m_BinOp(BO))))
1549 CPred = ICI->getPredicate();
1550 else if (match(&SI, m_Select(m_Specific(ICI), m_BinOp(BO), m_APInt(C))))
1551 CPred = ICI->getInversePredicate();
1555 const APInt *BinOpC;
1556 if (!match(BO, m_BinOp(m_Specific(V), m_APInt(BinOpC))))
1559 ConstantRange R = ConstantRange::makeExactICmpRegion(CPred, *CmpC)
1560 .binaryOp(BO->getOpcode(), *BinOpC);
1562 BO->dropPoisonGeneratingFlags();
1568 /// Visit a SelectInst that has an ICmpInst as its first operand.
1569 Instruction *InstCombinerImpl::foldSelectInstWithICmp(SelectInst &SI,
1571 if (Instruction *NewSel = foldSelectValueEquivalence(SI, *ICI))
1574 if (Instruction *NewSPF = canonicalizeSPF(SI, *ICI, *this))
1577 if (Value *V = foldSelectInstWithICmpConst(SI, ICI))
1578 return replaceInstUsesWith(SI, V);
1580 if (Value *V = canonicalizeClampLike(SI, *ICI, Builder))
1581 return replaceInstUsesWith(SI, V);
1583 if (Instruction *NewSel =
1584 tryToReuseConstantFromSelectInComparison(SI, *ICI, *this))
1587 bool Changed = adjustMinMax(SI, *ICI);
1589 if (Value *V = foldSelectICmpAnd(SI, ICI, Builder))
1590 return replaceInstUsesWith(SI, V);
1592 // NOTE: if we wanted to, this is where to detect integer MIN/MAX
1593 Value *TrueVal = SI.getTrueValue();
1594 Value *FalseVal = SI.getFalseValue();
1595 ICmpInst::Predicate Pred = ICI->getPredicate();
1596 Value *CmpLHS = ICI->getOperand(0);
1597 Value *CmpRHS = ICI->getOperand(1);
1598 if (CmpRHS != CmpLHS && isa<Constant>(CmpRHS)) {
1599 if (CmpLHS == TrueVal && Pred == ICmpInst::ICMP_EQ) {
1600 // Transform (X == C) ? X : Y -> (X == C) ? C : Y
1601 SI.setOperand(1, CmpRHS);
1603 } else if (CmpLHS == FalseVal && Pred == ICmpInst::ICMP_NE) {
1604 // Transform (X != C) ? Y : X -> (X != C) ? Y : C
1605 SI.setOperand(2, CmpRHS);
1610 // Canonicalize a signbit condition to use zero constant by swapping:
1611 // (CmpLHS > -1) ? TV : FV --> (CmpLHS < 0) ? FV : TV
1612 // To avoid conflicts (infinite loops) with other canonicalizations, this is
1613 // not applied with any constant select arm.
1614 if (Pred == ICmpInst::ICMP_SGT && match(CmpRHS, m_AllOnes()) &&
1615 !match(TrueVal, m_Constant()) && !match(FalseVal, m_Constant()) &&
1617 InstCombiner::BuilderTy::InsertPointGuard Guard(Builder);
1618 Builder.SetInsertPoint(&SI);
1619 Value *IsNeg = Builder.CreateIsNeg(CmpLHS, ICI->getName());
1620 replaceOperand(SI, 0, IsNeg);
1622 SI.swapProfMetadata();
1626 // FIXME: This code is nearly duplicated in InstSimplify. Using/refactoring
1627 // decomposeBitTestICmp() might help.
1630 DL.getTypeSizeInBits(TrueVal->getType()->getScalarType());
1631 APInt MinSignedValue = APInt::getSignedMinValue(BitWidth);
1635 bool IsBitTest = false;
1636 if (ICmpInst::isEquality(Pred) &&
1637 match(CmpLHS, m_And(m_Value(X), m_Power2(Y))) &&
1638 match(CmpRHS, m_Zero())) {
1640 TrueWhenUnset = Pred == ICmpInst::ICMP_EQ;
1641 } else if (Pred == ICmpInst::ICMP_SLT && match(CmpRHS, m_Zero())) {
1643 Y = &MinSignedValue;
1645 TrueWhenUnset = false;
1646 } else if (Pred == ICmpInst::ICMP_SGT && match(CmpRHS, m_AllOnes())) {
1648 Y = &MinSignedValue;
1650 TrueWhenUnset = true;
1654 // (X & Y) == 0 ? X : X ^ Y --> X & ~Y
1655 if (TrueWhenUnset && TrueVal == X &&
1656 match(FalseVal, m_Xor(m_Specific(X), m_APInt(C))) && *Y == *C)
1657 V = Builder.CreateAnd(X, ~(*Y));
1658 // (X & Y) != 0 ? X ^ Y : X --> X & ~Y
1659 else if (!TrueWhenUnset && FalseVal == X &&
1660 match(TrueVal, m_Xor(m_Specific(X), m_APInt(C))) && *Y == *C)
1661 V = Builder.CreateAnd(X, ~(*Y));
1662 // (X & Y) == 0 ? X ^ Y : X --> X | Y
1663 else if (TrueWhenUnset && FalseVal == X &&
1664 match(TrueVal, m_Xor(m_Specific(X), m_APInt(C))) && *Y == *C)
1665 V = Builder.CreateOr(X, *Y);
1666 // (X & Y) != 0 ? X : X ^ Y --> X | Y
1667 else if (!TrueWhenUnset && TrueVal == X &&
1668 match(FalseVal, m_Xor(m_Specific(X), m_APInt(C))) && *Y == *C)
1669 V = Builder.CreateOr(X, *Y);
1672 return replaceInstUsesWith(SI, V);
1676 if (Instruction *V =
1677 foldSelectICmpAndAnd(SI.getType(), ICI, TrueVal, FalseVal, Builder))
1680 if (Instruction *V = foldSelectCtlzToCttz(ICI, TrueVal, FalseVal, Builder))
1683 if (Instruction *V = foldSelectZeroOrOnes(ICI, TrueVal, FalseVal, Builder))
1686 if (Value *V = foldSelectICmpAndOr(ICI, TrueVal, FalseVal, Builder))
1687 return replaceInstUsesWith(SI, V);
1689 if (Value *V = foldSelectICmpLshrAshr(ICI, TrueVal, FalseVal, Builder))
1690 return replaceInstUsesWith(SI, V);
1692 if (Value *V = foldSelectCttzCtlz(ICI, TrueVal, FalseVal, Builder))
1693 return replaceInstUsesWith(SI, V);
1695 if (Value *V = canonicalizeSaturatedSubtract(ICI, TrueVal, FalseVal, Builder))
1696 return replaceInstUsesWith(SI, V);
1698 if (Value *V = canonicalizeSaturatedAdd(ICI, TrueVal, FalseVal, Builder))
1699 return replaceInstUsesWith(SI, V);
1701 return Changed ? &SI : nullptr;
1704 /// SI is a select whose condition is a PHI node (but the two may be in
1705 /// different blocks). See if the true/false values (V) are live in all of the
1706 /// predecessor blocks of the PHI. For example, cases like this can't be mapped:
1708 /// X = phi [ C1, BB1], [C2, BB2]
1710 /// Z = select X, Y, 0
1712 /// because Y is not live in BB1/BB2.
1713 static bool canSelectOperandBeMappingIntoPredBlock(const Value *V,
1714 const SelectInst &SI) {
1715 // If the value is a non-instruction value like a constant or argument, it
1716 // can always be mapped.
1717 const Instruction *I = dyn_cast<Instruction>(V);
1718 if (!I) return true;
1720 // If V is a PHI node defined in the same block as the condition PHI, we can
1721 // map the arguments.
1722 const PHINode *CondPHI = cast<PHINode>(SI.getCondition());
1724 if (const PHINode *VP = dyn_cast<PHINode>(I))
1725 if (VP->getParent() == CondPHI->getParent())
1728 // Otherwise, if the PHI and select are defined in the same block and if V is
1729 // defined in a different block, then we can transform it.
1730 if (SI.getParent() == CondPHI->getParent() &&
1731 I->getParent() != CondPHI->getParent())
1734 // Otherwise we have a 'hard' case and we can't tell without doing more
1735 // detailed dominator based analysis, punt.
1739 /// We have an SPF (e.g. a min or max) of an SPF of the form:
1740 /// SPF2(SPF1(A, B), C)
1741 Instruction *InstCombinerImpl::foldSPFofSPF(Instruction *Inner,
1742 SelectPatternFlavor SPF1, Value *A,
1743 Value *B, Instruction &Outer,
1744 SelectPatternFlavor SPF2,
1746 if (Outer.getType() != Inner->getType())
1749 if (C == A || C == B) {
1750 // MAX(MAX(A, B), B) -> MAX(A, B)
1751 // MIN(MIN(a, b), a) -> MIN(a, b)
1752 // TODO: This could be done in instsimplify.
1753 if (SPF1 == SPF2 && SelectPatternResult::isMinOrMax(SPF1))
1754 return replaceInstUsesWith(Outer, Inner);
1760 /// Turn select C, (X + Y), (X - Y) --> (X + (select C, Y, (-Y))).
1761 /// This is even legal for FP.
1762 static Instruction *foldAddSubSelect(SelectInst &SI,
1763 InstCombiner::BuilderTy &Builder) {
1764 Value *CondVal = SI.getCondition();
1765 Value *TrueVal = SI.getTrueValue();
1766 Value *FalseVal = SI.getFalseValue();
1767 auto *TI = dyn_cast<Instruction>(TrueVal);
1768 auto *FI = dyn_cast<Instruction>(FalseVal);
1769 if (!TI || !FI || !TI->hasOneUse() || !FI->hasOneUse())
1772 Instruction *AddOp = nullptr, *SubOp = nullptr;
1773 if ((TI->getOpcode() == Instruction::Sub &&
1774 FI->getOpcode() == Instruction::Add) ||
1775 (TI->getOpcode() == Instruction::FSub &&
1776 FI->getOpcode() == Instruction::FAdd)) {
1779 } else if ((FI->getOpcode() == Instruction::Sub &&
1780 TI->getOpcode() == Instruction::Add) ||
1781 (FI->getOpcode() == Instruction::FSub &&
1782 TI->getOpcode() == Instruction::FAdd)) {
1788 Value *OtherAddOp = nullptr;
1789 if (SubOp->getOperand(0) == AddOp->getOperand(0)) {
1790 OtherAddOp = AddOp->getOperand(1);
1791 } else if (SubOp->getOperand(0) == AddOp->getOperand(1)) {
1792 OtherAddOp = AddOp->getOperand(0);
1796 // So at this point we know we have (Y -> OtherAddOp):
1797 // select C, (add X, Y), (sub X, Z)
1798 Value *NegVal; // Compute -Z
1799 if (SI.getType()->isFPOrFPVectorTy()) {
1800 NegVal = Builder.CreateFNeg(SubOp->getOperand(1));
1801 if (Instruction *NegInst = dyn_cast<Instruction>(NegVal)) {
1802 FastMathFlags Flags = AddOp->getFastMathFlags();
1803 Flags &= SubOp->getFastMathFlags();
1804 NegInst->setFastMathFlags(Flags);
1807 NegVal = Builder.CreateNeg(SubOp->getOperand(1));
1810 Value *NewTrueOp = OtherAddOp;
1811 Value *NewFalseOp = NegVal;
1813 std::swap(NewTrueOp, NewFalseOp);
1814 Value *NewSel = Builder.CreateSelect(CondVal, NewTrueOp, NewFalseOp,
1815 SI.getName() + ".p", &SI);
1817 if (SI.getType()->isFPOrFPVectorTy()) {
1819 BinaryOperator::CreateFAdd(SubOp->getOperand(0), NewSel);
1821 FastMathFlags Flags = AddOp->getFastMathFlags();
1822 Flags &= SubOp->getFastMathFlags();
1823 RI->setFastMathFlags(Flags);
1826 return BinaryOperator::CreateAdd(SubOp->getOperand(0), NewSel);
1832 /// Turn X + Y overflows ? -1 : X + Y -> uadd_sat X, Y
1833 /// And X - Y overflows ? 0 : X - Y -> usub_sat X, Y
1834 /// Along with a number of patterns similar to:
1835 /// X + Y overflows ? (X < 0 ? INTMIN : INTMAX) : X + Y --> sadd_sat X, Y
1836 /// X - Y overflows ? (X > 0 ? INTMAX : INTMIN) : X - Y --> ssub_sat X, Y
1837 static Instruction *
1838 foldOverflowingAddSubSelect(SelectInst &SI, InstCombiner::BuilderTy &Builder) {
1839 Value *CondVal = SI.getCondition();
1840 Value *TrueVal = SI.getTrueValue();
1841 Value *FalseVal = SI.getFalseValue();
1843 WithOverflowInst *II;
1844 if (!match(CondVal, m_ExtractValue<1>(m_WithOverflowInst(II))) ||
1845 !match(FalseVal, m_ExtractValue<0>(m_Specific(II))))
1848 Value *X = II->getLHS();
1849 Value *Y = II->getRHS();
1851 auto IsSignedSaturateLimit = [&](Value *Limit, bool IsAdd) {
1852 Type *Ty = Limit->getType();
1854 ICmpInst::Predicate Pred;
1855 Value *TrueVal, *FalseVal, *Op;
1857 if (!match(Limit, m_Select(m_ICmp(Pred, m_Value(Op), m_APInt(C)),
1858 m_Value(TrueVal), m_Value(FalseVal))))
1861 auto IsZeroOrOne = [](const APInt &C) { return C.isZero() || C.isOne(); };
1862 auto IsMinMax = [&](Value *Min, Value *Max) {
1863 APInt MinVal = APInt::getSignedMinValue(Ty->getScalarSizeInBits());
1864 APInt MaxVal = APInt::getSignedMaxValue(Ty->getScalarSizeInBits());
1865 return match(Min, m_SpecificInt(MinVal)) &&
1866 match(Max, m_SpecificInt(MaxVal));
1869 if (Op != X && Op != Y)
1873 // X + Y overflows ? (X <s 0 ? INTMIN : INTMAX) : X + Y --> sadd_sat X, Y
1874 // X + Y overflows ? (X <s 1 ? INTMIN : INTMAX) : X + Y --> sadd_sat X, Y
1875 // X + Y overflows ? (Y <s 0 ? INTMIN : INTMAX) : X + Y --> sadd_sat X, Y
1876 // X + Y overflows ? (Y <s 1 ? INTMIN : INTMAX) : X + Y --> sadd_sat X, Y
1877 if (Pred == ICmpInst::ICMP_SLT && IsZeroOrOne(*C) &&
1878 IsMinMax(TrueVal, FalseVal))
1880 // X + Y overflows ? (X >s 0 ? INTMAX : INTMIN) : X + Y --> sadd_sat X, Y
1881 // X + Y overflows ? (X >s -1 ? INTMAX : INTMIN) : X + Y --> sadd_sat X, Y
1882 // X + Y overflows ? (Y >s 0 ? INTMAX : INTMIN) : X + Y --> sadd_sat X, Y
1883 // X + Y overflows ? (Y >s -1 ? INTMAX : INTMIN) : X + Y --> sadd_sat X, Y
1884 if (Pred == ICmpInst::ICMP_SGT && IsZeroOrOne(*C + 1) &&
1885 IsMinMax(FalseVal, TrueVal))
1888 // X - Y overflows ? (X <s 0 ? INTMIN : INTMAX) : X - Y --> ssub_sat X, Y
1889 // X - Y overflows ? (X <s -1 ? INTMIN : INTMAX) : X - Y --> ssub_sat X, Y
1890 if (Op == X && Pred == ICmpInst::ICMP_SLT && IsZeroOrOne(*C + 1) &&
1891 IsMinMax(TrueVal, FalseVal))
1893 // X - Y overflows ? (X >s -1 ? INTMAX : INTMIN) : X - Y --> ssub_sat X, Y
1894 // X - Y overflows ? (X >s -2 ? INTMAX : INTMIN) : X - Y --> ssub_sat X, Y
1895 if (Op == X && Pred == ICmpInst::ICMP_SGT && IsZeroOrOne(*C + 2) &&
1896 IsMinMax(FalseVal, TrueVal))
1898 // X - Y overflows ? (Y <s 0 ? INTMAX : INTMIN) : X - Y --> ssub_sat X, Y
1899 // X - Y overflows ? (Y <s 1 ? INTMAX : INTMIN) : X - Y --> ssub_sat X, Y
1900 if (Op == Y && Pred == ICmpInst::ICMP_SLT && IsZeroOrOne(*C) &&
1901 IsMinMax(FalseVal, TrueVal))
1903 // X - Y overflows ? (Y >s 0 ? INTMIN : INTMAX) : X - Y --> ssub_sat X, Y
1904 // X - Y overflows ? (Y >s -1 ? INTMIN : INTMAX) : X - Y --> ssub_sat X, Y
1905 if (Op == Y && Pred == ICmpInst::ICMP_SGT && IsZeroOrOne(*C + 1) &&
1906 IsMinMax(TrueVal, FalseVal))
1913 Intrinsic::ID NewIntrinsicID;
1914 if (II->getIntrinsicID() == Intrinsic::uadd_with_overflow &&
1915 match(TrueVal, m_AllOnes()))
1916 // X + Y overflows ? -1 : X + Y -> uadd_sat X, Y
1917 NewIntrinsicID = Intrinsic::uadd_sat;
1918 else if (II->getIntrinsicID() == Intrinsic::usub_with_overflow &&
1919 match(TrueVal, m_Zero()))
1920 // X - Y overflows ? 0 : X - Y -> usub_sat X, Y
1921 NewIntrinsicID = Intrinsic::usub_sat;
1922 else if (II->getIntrinsicID() == Intrinsic::sadd_with_overflow &&
1923 IsSignedSaturateLimit(TrueVal, /*IsAdd=*/true))
1924 // X + Y overflows ? (X <s 0 ? INTMIN : INTMAX) : X + Y --> sadd_sat X, Y
1925 // X + Y overflows ? (X <s 1 ? INTMIN : INTMAX) : X + Y --> sadd_sat X, Y
1926 // X + Y overflows ? (X >s 0 ? INTMAX : INTMIN) : X + Y --> sadd_sat X, Y
1927 // X + Y overflows ? (X >s -1 ? INTMAX : INTMIN) : X + Y --> sadd_sat X, Y
1928 // X + Y overflows ? (Y <s 0 ? INTMIN : INTMAX) : X + Y --> sadd_sat X, Y
1929 // X + Y overflows ? (Y <s 1 ? INTMIN : INTMAX) : X + Y --> sadd_sat X, Y
1930 // X + Y overflows ? (Y >s 0 ? INTMAX : INTMIN) : X + Y --> sadd_sat X, Y
1931 // X + Y overflows ? (Y >s -1 ? INTMAX : INTMIN) : X + Y --> sadd_sat X, Y
1932 NewIntrinsicID = Intrinsic::sadd_sat;
1933 else if (II->getIntrinsicID() == Intrinsic::ssub_with_overflow &&
1934 IsSignedSaturateLimit(TrueVal, /*IsAdd=*/false))
1935 // X - Y overflows ? (X <s 0 ? INTMIN : INTMAX) : X - Y --> ssub_sat X, Y
1936 // X - Y overflows ? (X <s -1 ? INTMIN : INTMAX) : X - Y --> ssub_sat X, Y
1937 // X - Y overflows ? (X >s -1 ? INTMAX : INTMIN) : X - Y --> ssub_sat X, Y
1938 // X - Y overflows ? (X >s -2 ? INTMAX : INTMIN) : X - Y --> ssub_sat X, Y
1939 // X - Y overflows ? (Y <s 0 ? INTMAX : INTMIN) : X - Y --> ssub_sat X, Y
1940 // X - Y overflows ? (Y <s 1 ? INTMAX : INTMIN) : X - Y --> ssub_sat X, Y
1941 // X - Y overflows ? (Y >s 0 ? INTMIN : INTMAX) : X - Y --> ssub_sat X, Y
1942 // X - Y overflows ? (Y >s -1 ? INTMIN : INTMAX) : X - Y --> ssub_sat X, Y
1943 NewIntrinsicID = Intrinsic::ssub_sat;
1948 Intrinsic::getDeclaration(SI.getModule(), NewIntrinsicID, SI.getType());
1949 return CallInst::Create(F, {X, Y});
1952 Instruction *InstCombinerImpl::foldSelectExtConst(SelectInst &Sel) {
1954 if (!match(Sel.getTrueValue(), m_Constant(C)) &&
1955 !match(Sel.getFalseValue(), m_Constant(C)))
1958 Instruction *ExtInst;
1959 if (!match(Sel.getTrueValue(), m_Instruction(ExtInst)) &&
1960 !match(Sel.getFalseValue(), m_Instruction(ExtInst)))
1963 auto ExtOpcode = ExtInst->getOpcode();
1964 if (ExtOpcode != Instruction::ZExt && ExtOpcode != Instruction::SExt)
1967 // If we are extending from a boolean type or if we can create a select that
1968 // has the same size operands as its condition, try to narrow the select.
1969 Value *X = ExtInst->getOperand(0);
1970 Type *SmallType = X->getType();
1971 Value *Cond = Sel.getCondition();
1972 auto *Cmp = dyn_cast<CmpInst>(Cond);
1973 if (!SmallType->isIntOrIntVectorTy(1) &&
1974 (!Cmp || Cmp->getOperand(0)->getType() != SmallType))
1977 // If the constant is the same after truncation to the smaller type and
1978 // extension to the original type, we can narrow the select.
1979 Type *SelType = Sel.getType();
1980 Constant *TruncC = ConstantExpr::getTrunc(C, SmallType);
1981 Constant *ExtC = ConstantExpr::getCast(ExtOpcode, TruncC, SelType);
1982 if (ExtC == C && ExtInst->hasOneUse()) {
1983 Value *TruncCVal = cast<Value>(TruncC);
1984 if (ExtInst == Sel.getFalseValue())
1985 std::swap(X, TruncCVal);
1987 // select Cond, (ext X), C --> ext(select Cond, X, C')
1988 // select Cond, C, (ext X) --> ext(select Cond, C', X)
1989 Value *NewSel = Builder.CreateSelect(Cond, X, TruncCVal, "narrow", &Sel);
1990 return CastInst::Create(Instruction::CastOps(ExtOpcode), NewSel, SelType);
1993 // If one arm of the select is the extend of the condition, replace that arm
1994 // with the extension of the appropriate known bool value.
1996 if (ExtInst == Sel.getTrueValue()) {
1997 // select X, (sext X), C --> select X, -1, C
1998 // select X, (zext X), C --> select X, 1, C
1999 Constant *One = ConstantInt::getTrue(SmallType);
2000 Constant *AllOnesOrOne = ConstantExpr::getCast(ExtOpcode, One, SelType);
2001 return SelectInst::Create(Cond, AllOnesOrOne, C, "", nullptr, &Sel);
2003 // select X, C, (sext X) --> select X, C, 0
2004 // select X, C, (zext X) --> select X, C, 0
2005 Constant *Zero = ConstantInt::getNullValue(SelType);
2006 return SelectInst::Create(Cond, C, Zero, "", nullptr, &Sel);
2013 /// Try to transform a vector select with a constant condition vector into a
2014 /// shuffle for easier combining with other shuffles and insert/extract.
2015 static Instruction *canonicalizeSelectToShuffle(SelectInst &SI) {
2016 Value *CondVal = SI.getCondition();
2018 auto *CondValTy = dyn_cast<FixedVectorType>(CondVal->getType());
2019 if (!CondValTy || !match(CondVal, m_Constant(CondC)))
2022 unsigned NumElts = CondValTy->getNumElements();
2023 SmallVector<int, 16> Mask;
2024 Mask.reserve(NumElts);
2025 for (unsigned i = 0; i != NumElts; ++i) {
2026 Constant *Elt = CondC->getAggregateElement(i);
2030 if (Elt->isOneValue()) {
2031 // If the select condition element is true, choose from the 1st vector.
2033 } else if (Elt->isNullValue()) {
2034 // If the select condition element is false, choose from the 2nd vector.
2035 Mask.push_back(i + NumElts);
2036 } else if (isa<UndefValue>(Elt)) {
2037 // Undef in a select condition (choose one of the operands) does not mean
2038 // the same thing as undef in a shuffle mask (any value is acceptable), so
2042 // Bail out on a constant expression.
2047 return new ShuffleVectorInst(SI.getTrueValue(), SI.getFalseValue(), Mask);
2050 /// If we have a select of vectors with a scalar condition, try to convert that
2051 /// to a vector select by splatting the condition. A splat may get folded with
2052 /// other operations in IR and having all operands of a select be vector types
2053 /// is likely better for vector codegen.
2054 static Instruction *canonicalizeScalarSelectOfVecs(SelectInst &Sel,
2055 InstCombinerImpl &IC) {
2056 auto *Ty = dyn_cast<VectorType>(Sel.getType());
2060 // We can replace a single-use extract with constant index.
2061 Value *Cond = Sel.getCondition();
2062 if (!match(Cond, m_OneUse(m_ExtractElt(m_Value(), m_ConstantInt()))))
2065 // select (extelt V, Index), T, F --> select (splat V, Index), T, F
2066 // Splatting the extracted condition reduces code (we could directly create a
2067 // splat shuffle of the source vector to eliminate the intermediate step).
2068 return IC.replaceOperand(
2069 Sel, 0, IC.Builder.CreateVectorSplat(Ty->getElementCount(), Cond));
2072 /// Reuse bitcasted operands between a compare and select:
2073 /// select (cmp (bitcast C), (bitcast D)), (bitcast' C), (bitcast' D) -->
2074 /// bitcast (select (cmp (bitcast C), (bitcast D)), (bitcast C), (bitcast D))
2075 static Instruction *foldSelectCmpBitcasts(SelectInst &Sel,
2076 InstCombiner::BuilderTy &Builder) {
2077 Value *Cond = Sel.getCondition();
2078 Value *TVal = Sel.getTrueValue();
2079 Value *FVal = Sel.getFalseValue();
2081 CmpInst::Predicate Pred;
2083 if (!match(Cond, m_Cmp(Pred, m_Value(A), m_Value(B))))
2086 // The select condition is a compare instruction. If the select's true/false
2087 // values are already the same as the compare operands, there's nothing to do.
2088 if (TVal == A || TVal == B || FVal == A || FVal == B)
2092 if (!match(A, m_BitCast(m_Value(C))) || !match(B, m_BitCast(m_Value(D))))
2095 // select (cmp (bitcast C), (bitcast D)), (bitcast TSrc), (bitcast FSrc)
2097 if (!match(TVal, m_BitCast(m_Value(TSrc))) ||
2098 !match(FVal, m_BitCast(m_Value(FSrc))))
2101 // If the select true/false values are *different bitcasts* of the same source
2102 // operands, make the select operands the same as the compare operands and
2103 // cast the result. This is the canonical select form for min/max.
2105 if (TSrc == C && FSrc == D) {
2106 // select (cmp (bitcast C), (bitcast D)), (bitcast' C), (bitcast' D) -->
2107 // bitcast (select (cmp A, B), A, B)
2108 NewSel = Builder.CreateSelect(Cond, A, B, "", &Sel);
2109 } else if (TSrc == D && FSrc == C) {
2110 // select (cmp (bitcast C), (bitcast D)), (bitcast' D), (bitcast' C) -->
2111 // bitcast (select (cmp A, B), B, A)
2112 NewSel = Builder.CreateSelect(Cond, B, A, "", &Sel);
2116 return CastInst::CreateBitOrPointerCast(NewSel, Sel.getType());
2119 /// Try to eliminate select instructions that test the returned flag of cmpxchg
2122 /// If a select instruction tests the returned flag of a cmpxchg instruction and
2123 /// selects between the returned value of the cmpxchg instruction its compare
2124 /// operand, the result of the select will always be equal to its false value.
2127 /// %0 = cmpxchg i64* %ptr, i64 %compare, i64 %new_value seq_cst seq_cst
2128 /// %1 = extractvalue { i64, i1 } %0, 1
2129 /// %2 = extractvalue { i64, i1 } %0, 0
2130 /// %3 = select i1 %1, i64 %compare, i64 %2
2133 /// The returned value of the cmpxchg instruction (%2) is the original value
2134 /// located at %ptr prior to any update. If the cmpxchg operation succeeds, %2
2135 /// must have been equal to %compare. Thus, the result of the select is always
2136 /// equal to %2, and the code can be simplified to:
2138 /// %0 = cmpxchg i64* %ptr, i64 %compare, i64 %new_value seq_cst seq_cst
2139 /// %1 = extractvalue { i64, i1 } %0, 0
2142 static Value *foldSelectCmpXchg(SelectInst &SI) {
2143 // A helper that determines if V is an extractvalue instruction whose
2144 // aggregate operand is a cmpxchg instruction and whose single index is equal
2145 // to I. If such conditions are true, the helper returns the cmpxchg
2146 // instruction; otherwise, a nullptr is returned.
2147 auto isExtractFromCmpXchg = [](Value *V, unsigned I) -> AtomicCmpXchgInst * {
2148 auto *Extract = dyn_cast<ExtractValueInst>(V);
2151 if (Extract->getIndices()[0] != I)
2153 return dyn_cast<AtomicCmpXchgInst>(Extract->getAggregateOperand());
2156 // If the select has a single user, and this user is a select instruction that
2157 // we can simplify, skip the cmpxchg simplification for now.
2159 if (auto *Select = dyn_cast<SelectInst>(SI.user_back()))
2160 if (Select->getCondition() == SI.getCondition())
2161 if (Select->getFalseValue() == SI.getTrueValue() ||
2162 Select->getTrueValue() == SI.getFalseValue())
2165 // Ensure the select condition is the returned flag of a cmpxchg instruction.
2166 auto *CmpXchg = isExtractFromCmpXchg(SI.getCondition(), 1);
2170 // Check the true value case: The true value of the select is the returned
2171 // value of the same cmpxchg used by the condition, and the false value is the
2172 // cmpxchg instruction's compare operand.
2173 if (auto *X = isExtractFromCmpXchg(SI.getTrueValue(), 0))
2174 if (X == CmpXchg && X->getCompareOperand() == SI.getFalseValue())
2175 return SI.getFalseValue();
2177 // Check the false value case: The false value of the select is the returned
2178 // value of the same cmpxchg used by the condition, and the true value is the
2179 // cmpxchg instruction's compare operand.
2180 if (auto *X = isExtractFromCmpXchg(SI.getFalseValue(), 0))
2181 if (X == CmpXchg && X->getCompareOperand() == SI.getTrueValue())
2182 return SI.getFalseValue();
2187 /// Try to reduce a funnel/rotate pattern that includes a compare and select
2188 /// into a funnel shift intrinsic. Example:
2189 /// rotl32(a, b) --> (b == 0 ? a : ((a >> (32 - b)) | (a << b)))
2190 /// --> call llvm.fshl.i32(a, a, b)
2191 /// fshl32(a, b, c) --> (c == 0 ? a : ((b >> (32 - c)) | (a << c)))
2192 /// --> call llvm.fshl.i32(a, b, c)
2193 /// fshr32(a, b, c) --> (c == 0 ? b : ((a >> (32 - c)) | (b << c)))
2194 /// --> call llvm.fshr.i32(a, b, c)
2195 static Instruction *foldSelectFunnelShift(SelectInst &Sel,
2196 InstCombiner::BuilderTy &Builder) {
2197 // This must be a power-of-2 type for a bitmasking transform to be valid.
2198 unsigned Width = Sel.getType()->getScalarSizeInBits();
2199 if (!isPowerOf2_32(Width))
2202 BinaryOperator *Or0, *Or1;
2203 if (!match(Sel.getFalseValue(), m_OneUse(m_Or(m_BinOp(Or0), m_BinOp(Or1)))))
2206 Value *SV0, *SV1, *SA0, *SA1;
2207 if (!match(Or0, m_OneUse(m_LogicalShift(m_Value(SV0),
2208 m_ZExtOrSelf(m_Value(SA0))))) ||
2209 !match(Or1, m_OneUse(m_LogicalShift(m_Value(SV1),
2210 m_ZExtOrSelf(m_Value(SA1))))) ||
2211 Or0->getOpcode() == Or1->getOpcode())
2214 // Canonicalize to or(shl(SV0, SA0), lshr(SV1, SA1)).
2215 if (Or0->getOpcode() == BinaryOperator::LShr) {
2216 std::swap(Or0, Or1);
2217 std::swap(SV0, SV1);
2218 std::swap(SA0, SA1);
2220 assert(Or0->getOpcode() == BinaryOperator::Shl &&
2221 Or1->getOpcode() == BinaryOperator::LShr &&
2222 "Illegal or(shift,shift) pair");
2224 // Check the shift amounts to see if they are an opposite pair.
2226 if (match(SA1, m_OneUse(m_Sub(m_SpecificInt(Width), m_Specific(SA0)))))
2228 else if (match(SA0, m_OneUse(m_Sub(m_SpecificInt(Width), m_Specific(SA1)))))
2233 // We should now have this pattern:
2234 // select ?, TVal, (or (shl SV0, SA0), (lshr SV1, SA1))
2235 // The false value of the select must be a funnel-shift of the true value:
2236 // IsFShl -> TVal must be SV0 else TVal must be SV1.
2237 bool IsFshl = (ShAmt == SA0);
2238 Value *TVal = Sel.getTrueValue();
2239 if ((IsFshl && TVal != SV0) || (!IsFshl && TVal != SV1))
2242 // Finally, see if the select is filtering out a shift-by-zero.
2243 Value *Cond = Sel.getCondition();
2244 ICmpInst::Predicate Pred;
2245 if (!match(Cond, m_OneUse(m_ICmp(Pred, m_Specific(ShAmt), m_ZeroInt()))) ||
2246 Pred != ICmpInst::ICMP_EQ)
2249 // If this is not a rotate then the select was blocking poison from the
2250 // 'shift-by-zero' non-TVal, but a funnel shift won't - so freeze it.
2252 if (IsFshl && !llvm::isGuaranteedNotToBePoison(SV1))
2253 SV1 = Builder.CreateFreeze(SV1);
2254 else if (!IsFshl && !llvm::isGuaranteedNotToBePoison(SV0))
2255 SV0 = Builder.CreateFreeze(SV0);
2258 // This is a funnel/rotate that avoids shift-by-bitwidth UB in a suboptimal way.
2259 // Convert to funnel shift intrinsic.
2260 Intrinsic::ID IID = IsFshl ? Intrinsic::fshl : Intrinsic::fshr;
2261 Function *F = Intrinsic::getDeclaration(Sel.getModule(), IID, Sel.getType());
2262 ShAmt = Builder.CreateZExt(ShAmt, Sel.getType());
2263 return CallInst::Create(F, { SV0, SV1, ShAmt });
2266 static Instruction *foldSelectToCopysign(SelectInst &Sel,
2267 InstCombiner::BuilderTy &Builder) {
2268 Value *Cond = Sel.getCondition();
2269 Value *TVal = Sel.getTrueValue();
2270 Value *FVal = Sel.getFalseValue();
2271 Type *SelType = Sel.getType();
2273 // Match select ?, TC, FC where the constants are equal but negated.
2274 // TODO: Generalize to handle a negated variable operand?
2275 const APFloat *TC, *FC;
2276 if (!match(TVal, m_APFloatAllowUndef(TC)) ||
2277 !match(FVal, m_APFloatAllowUndef(FC)) ||
2278 !abs(*TC).bitwiseIsEqual(abs(*FC)))
2281 assert(TC != FC && "Expected equal select arms to simplify");
2285 bool IsTrueIfSignSet;
2286 ICmpInst::Predicate Pred;
2287 if (!match(Cond, m_OneUse(m_ICmp(Pred, m_BitCast(m_Value(X)), m_APInt(C)))) ||
2288 !InstCombiner::isSignBitCheck(Pred, *C, IsTrueIfSignSet) ||
2289 X->getType() != SelType)
2292 // If needed, negate the value that will be the sign argument of the copysign:
2293 // (bitcast X) < 0 ? -TC : TC --> copysign(TC, X)
2294 // (bitcast X) < 0 ? TC : -TC --> copysign(TC, -X)
2295 // (bitcast X) >= 0 ? -TC : TC --> copysign(TC, -X)
2296 // (bitcast X) >= 0 ? TC : -TC --> copysign(TC, X)
2297 // Note: FMF from the select can not be propagated to the new instructions.
2298 if (IsTrueIfSignSet ^ TC->isNegative())
2299 X = Builder.CreateFNeg(X);
2301 // Canonicalize the magnitude argument as the positive constant since we do
2302 // not care about its sign.
2303 Value *MagArg = ConstantFP::get(SelType, abs(*TC));
2304 Function *F = Intrinsic::getDeclaration(Sel.getModule(), Intrinsic::copysign,
2306 return CallInst::Create(F, { MagArg, X });
2309 Instruction *InstCombinerImpl::foldVectorSelect(SelectInst &Sel) {
2310 auto *VecTy = dyn_cast<FixedVectorType>(Sel.getType());
2314 unsigned NumElts = VecTy->getNumElements();
2315 APInt UndefElts(NumElts, 0);
2316 APInt AllOnesEltMask(APInt::getAllOnes(NumElts));
2317 if (Value *V = SimplifyDemandedVectorElts(&Sel, AllOnesEltMask, UndefElts)) {
2319 return replaceInstUsesWith(Sel, V);
2323 // A select of a "select shuffle" with a common operand can be rearranged
2324 // to select followed by "select shuffle". Because of poison, this only works
2325 // in the case of a shuffle with no undefined mask elements.
2326 Value *Cond = Sel.getCondition();
2327 Value *TVal = Sel.getTrueValue();
2328 Value *FVal = Sel.getFalseValue();
2331 if (match(TVal, m_OneUse(m_Shuffle(m_Value(X), m_Value(Y), m_Mask(Mask)))) &&
2332 !is_contained(Mask, UndefMaskElem) &&
2333 cast<ShuffleVectorInst>(TVal)->isSelect()) {
2335 // select Cond, (shuf_sel X, Y), X --> shuf_sel X, (select Cond, Y, X)
2336 Value *NewSel = Builder.CreateSelect(Cond, Y, X, "sel", &Sel);
2337 return new ShuffleVectorInst(X, NewSel, Mask);
2340 // select Cond, (shuf_sel X, Y), Y --> shuf_sel (select Cond, X, Y), Y
2341 Value *NewSel = Builder.CreateSelect(Cond, X, Y, "sel", &Sel);
2342 return new ShuffleVectorInst(NewSel, Y, Mask);
2345 if (match(FVal, m_OneUse(m_Shuffle(m_Value(X), m_Value(Y), m_Mask(Mask)))) &&
2346 !is_contained(Mask, UndefMaskElem) &&
2347 cast<ShuffleVectorInst>(FVal)->isSelect()) {
2349 // select Cond, X, (shuf_sel X, Y) --> shuf_sel X, (select Cond, X, Y)
2350 Value *NewSel = Builder.CreateSelect(Cond, X, Y, "sel", &Sel);
2351 return new ShuffleVectorInst(X, NewSel, Mask);
2354 // select Cond, Y, (shuf_sel X, Y) --> shuf_sel (select Cond, Y, X), Y
2355 Value *NewSel = Builder.CreateSelect(Cond, Y, X, "sel", &Sel);
2356 return new ShuffleVectorInst(NewSel, Y, Mask);
2363 static Instruction *foldSelectToPhiImpl(SelectInst &Sel, BasicBlock *BB,
2364 const DominatorTree &DT,
2365 InstCombiner::BuilderTy &Builder) {
2366 // Find the block's immediate dominator that ends with a conditional branch
2367 // that matches select's condition (maybe inverted).
2368 auto *IDomNode = DT[BB]->getIDom();
2371 BasicBlock *IDom = IDomNode->getBlock();
2373 Value *Cond = Sel.getCondition();
2374 Value *IfTrue, *IfFalse;
2375 BasicBlock *TrueSucc, *FalseSucc;
2376 if (match(IDom->getTerminator(),
2377 m_Br(m_Specific(Cond), m_BasicBlock(TrueSucc),
2378 m_BasicBlock(FalseSucc)))) {
2379 IfTrue = Sel.getTrueValue();
2380 IfFalse = Sel.getFalseValue();
2381 } else if (match(IDom->getTerminator(),
2382 m_Br(m_Not(m_Specific(Cond)), m_BasicBlock(TrueSucc),
2383 m_BasicBlock(FalseSucc)))) {
2384 IfTrue = Sel.getFalseValue();
2385 IfFalse = Sel.getTrueValue();
2389 // Make sure the branches are actually different.
2390 if (TrueSucc == FalseSucc)
2393 // We want to replace select %cond, %a, %b with a phi that takes value %a
2394 // for all incoming edges that are dominated by condition `%cond == true`,
2395 // and value %b for edges dominated by condition `%cond == false`. If %a
2396 // or %b are also phis from the same basic block, we can go further and take
2397 // their incoming values from the corresponding blocks.
2398 BasicBlockEdge TrueEdge(IDom, TrueSucc);
2399 BasicBlockEdge FalseEdge(IDom, FalseSucc);
2400 DenseMap<BasicBlock *, Value *> Inputs;
2401 for (auto *Pred : predecessors(BB)) {
2402 // Check implication.
2403 BasicBlockEdge Incoming(Pred, BB);
2404 if (DT.dominates(TrueEdge, Incoming))
2405 Inputs[Pred] = IfTrue->DoPHITranslation(BB, Pred);
2406 else if (DT.dominates(FalseEdge, Incoming))
2407 Inputs[Pred] = IfFalse->DoPHITranslation(BB, Pred);
2410 // Check availability.
2411 if (auto *Insn = dyn_cast<Instruction>(Inputs[Pred]))
2412 if (!DT.dominates(Insn, Pred->getTerminator()))
2416 Builder.SetInsertPoint(&*BB->begin());
2417 auto *PN = Builder.CreatePHI(Sel.getType(), Inputs.size());
2418 for (auto *Pred : predecessors(BB))
2419 PN->addIncoming(Inputs[Pred], Pred);
2424 static Instruction *foldSelectToPhi(SelectInst &Sel, const DominatorTree &DT,
2425 InstCombiner::BuilderTy &Builder) {
2426 // Try to replace this select with Phi in one of these blocks.
2427 SmallSetVector<BasicBlock *, 4> CandidateBlocks;
2428 CandidateBlocks.insert(Sel.getParent());
2429 for (Value *V : Sel.operands())
2430 if (auto *I = dyn_cast<Instruction>(V))
2431 CandidateBlocks.insert(I->getParent());
2433 for (BasicBlock *BB : CandidateBlocks)
2434 if (auto *PN = foldSelectToPhiImpl(Sel, BB, DT, Builder))
2439 static Value *foldSelectWithFrozenICmp(SelectInst &Sel, InstCombiner::BuilderTy &Builder) {
2440 FreezeInst *FI = dyn_cast<FreezeInst>(Sel.getCondition());
2444 Value *Cond = FI->getOperand(0);
2445 Value *TrueVal = Sel.getTrueValue(), *FalseVal = Sel.getFalseValue();
2447 // select (freeze(x == y)), x, y --> y
2448 // select (freeze(x != y)), x, y --> x
2449 // The freeze should be only used by this select. Otherwise, remaining uses of
2450 // the freeze can observe a contradictory value.
2451 // c = freeze(x == y) ; Let's assume that y = poison & x = 42; c is 0 or 1
2452 // a = select c, x, y ;
2453 // f(a, c) ; f(poison, 1) cannot happen, but if a is folded
2454 // ; to y, this can happen.
2455 CmpInst::Predicate Pred;
2456 if (FI->hasOneUse() &&
2457 match(Cond, m_c_ICmp(Pred, m_Specific(TrueVal), m_Specific(FalseVal))) &&
2458 (Pred == ICmpInst::ICMP_EQ || Pred == ICmpInst::ICMP_NE)) {
2459 return Pred == ICmpInst::ICMP_EQ ? FalseVal : TrueVal;
2465 Instruction *InstCombinerImpl::foldAndOrOfSelectUsingImpliedCond(Value *Op,
2468 Value *CondVal = SI.getCondition();
2469 Value *A = SI.getTrueValue();
2470 Value *B = SI.getFalseValue();
2472 assert(Op->getType()->isIntOrIntVectorTy(1) &&
2473 "Op must be either i1 or vector of i1.");
2475 Optional<bool> Res = isImpliedCondition(Op, CondVal, DL, IsAnd);
2479 Value *Zero = Constant::getNullValue(A->getType());
2480 Value *One = Constant::getAllOnesValue(A->getType());
2484 // select op, (select cond, A, B), false => select op, A, false
2485 // and op, (select cond, A, B) => select op, A, false
2486 // if op = true implies condval = true.
2487 return SelectInst::Create(Op, A, Zero);
2489 // select op, true, (select cond, A, B) => select op, true, A
2490 // or op, (select cond, A, B) => select op, true, A
2491 // if op = false implies condval = true.
2492 return SelectInst::Create(Op, One, A);
2495 // select op, (select cond, A, B), false => select op, B, false
2496 // and op, (select cond, A, B) => select op, B, false
2497 // if op = true implies condval = false.
2498 return SelectInst::Create(Op, B, Zero);
2500 // select op, true, (select cond, A, B) => select op, true, B
2501 // or op, (select cond, A, B) => select op, true, B
2502 // if op = false implies condval = false.
2503 return SelectInst::Create(Op, One, B);
2507 // Canonicalize select with fcmp to fabs(). -0.0 makes this tricky. We need
2508 // fast-math-flags (nsz) or fsub with +0.0 (not fneg) for this to work.
2509 static Instruction *foldSelectWithFCmpToFabs(SelectInst &SI,
2510 InstCombinerImpl &IC) {
2511 Value *CondVal = SI.getCondition();
2513 for (bool Swap : {false, true}) {
2514 Value *TrueVal = SI.getTrueValue();
2515 Value *X = SI.getFalseValue();
2516 CmpInst::Predicate Pred;
2519 std::swap(TrueVal, X);
2521 if (!match(CondVal, m_FCmp(Pred, m_Specific(X), m_AnyZeroFP())))
2524 // fold (X <= +/-0.0) ? (0.0 - X) : X to fabs(X), when 'Swap' is false
2525 // fold (X > +/-0.0) ? X : (0.0 - X) to fabs(X), when 'Swap' is true
2526 if (match(TrueVal, m_FSub(m_PosZeroFP(), m_Specific(X)))) {
2527 if (!Swap && (Pred == FCmpInst::FCMP_OLE || Pred == FCmpInst::FCMP_ULE)) {
2528 Value *Fabs = IC.Builder.CreateUnaryIntrinsic(Intrinsic::fabs, X, &SI);
2529 return IC.replaceInstUsesWith(SI, Fabs);
2531 if (Swap && (Pred == FCmpInst::FCMP_OGT || Pred == FCmpInst::FCMP_UGT)) {
2532 Value *Fabs = IC.Builder.CreateUnaryIntrinsic(Intrinsic::fabs, X, &SI);
2533 return IC.replaceInstUsesWith(SI, Fabs);
2537 // With nsz, when 'Swap' is false:
2538 // fold (X < +/-0.0) ? -X : X or (X <= +/-0.0) ? -X : X to fabs(X)
2539 // fold (X > +/-0.0) ? -X : X or (X >= +/-0.0) ? -X : X to -fabs(x)
2540 // when 'Swap' is true:
2541 // fold (X > +/-0.0) ? X : -X or (X >= +/-0.0) ? X : -X to fabs(X)
2542 // fold (X < +/-0.0) ? X : -X or (X <= +/-0.0) ? X : -X to -fabs(X)
2543 if (!match(TrueVal, m_FNeg(m_Specific(X))) || !SI.hasNoSignedZeros())
2547 Pred = FCmpInst::getSwappedPredicate(Pred);
2549 bool IsLTOrLE = Pred == FCmpInst::FCMP_OLT || Pred == FCmpInst::FCMP_OLE ||
2550 Pred == FCmpInst::FCMP_ULT || Pred == FCmpInst::FCMP_ULE;
2551 bool IsGTOrGE = Pred == FCmpInst::FCMP_OGT || Pred == FCmpInst::FCMP_OGE ||
2552 Pred == FCmpInst::FCMP_UGT || Pred == FCmpInst::FCMP_UGE;
2555 Value *Fabs = IC.Builder.CreateUnaryIntrinsic(Intrinsic::fabs, X, &SI);
2556 return IC.replaceInstUsesWith(SI, Fabs);
2559 Value *Fabs = IC.Builder.CreateUnaryIntrinsic(Intrinsic::fabs, X, &SI);
2560 Instruction *NewFNeg = UnaryOperator::CreateFNeg(Fabs);
2561 NewFNeg->setFastMathFlags(SI.getFastMathFlags());
2569 // Match the following IR pattern:
2570 // %x.lowbits = and i8 %x, %lowbitmask
2571 // %x.lowbits.are.zero = icmp eq i8 %x.lowbits, 0
2572 // %x.biased = add i8 %x, %bias
2573 // %x.biased.highbits = and i8 %x.biased, %highbitmask
2574 // %x.roundedup = select i1 %x.lowbits.are.zero, i8 %x, i8 %x.biased.highbits
2576 // %alignment = add i8 %lowbitmask, 1
2577 // Iff 1. an %alignment is a power-of-two (aka, %lowbitmask is a low bit mask)
2578 // and 2. %bias is equal to either %lowbitmask or %alignment,
2579 // and 3. %highbitmask is equal to ~%lowbitmask (aka, to -%alignment)
2580 // then this pattern can be transformed into:
2581 // %x.offset = add i8 %x, %lowbitmask
2582 // %x.roundedup = and i8 %x.offset, %highbitmask
2584 foldRoundUpIntegerWithPow2Alignment(SelectInst &SI,
2585 InstCombiner::BuilderTy &Builder) {
2586 Value *Cond = SI.getCondition();
2587 Value *X = SI.getTrueValue();
2588 Value *XBiasedHighBits = SI.getFalseValue();
2590 ICmpInst::Predicate Pred;
2592 if (!match(Cond, m_ICmp(Pred, m_Value(XLowBits), m_ZeroInt())) ||
2593 !ICmpInst::isEquality(Pred))
2596 if (Pred == ICmpInst::Predicate::ICMP_NE)
2597 std::swap(X, XBiasedHighBits);
2599 // FIXME: we could support non non-splats here.
2601 const APInt *LowBitMaskCst;
2602 if (!match(XLowBits, m_And(m_Specific(X), m_APIntAllowUndef(LowBitMaskCst))))
2605 const APInt *BiasCst, *HighBitMaskCst;
2606 if (!match(XBiasedHighBits,
2607 m_And(m_Add(m_Specific(X), m_APIntAllowUndef(BiasCst)),
2608 m_APIntAllowUndef(HighBitMaskCst))))
2611 if (!LowBitMaskCst->isMask())
2614 APInt InvertedLowBitMaskCst = ~*LowBitMaskCst;
2615 if (InvertedLowBitMaskCst != *HighBitMaskCst)
2618 APInt AlignmentCst = *LowBitMaskCst + 1;
2620 if (*BiasCst != AlignmentCst && *BiasCst != *LowBitMaskCst)
2623 if (!XBiasedHighBits->hasOneUse()) {
2624 if (*BiasCst == *LowBitMaskCst)
2625 return XBiasedHighBits;
2629 // FIXME: could we preserve undef's here?
2630 Type *Ty = X->getType();
2631 Value *XOffset = Builder.CreateAdd(X, ConstantInt::get(Ty, *LowBitMaskCst),
2632 X->getName() + ".biased");
2633 Value *R = Builder.CreateAnd(XOffset, ConstantInt::get(Ty, *HighBitMaskCst));
2638 Instruction *InstCombinerImpl::visitSelectInst(SelectInst &SI) {
2639 Value *CondVal = SI.getCondition();
2640 Value *TrueVal = SI.getTrueValue();
2641 Value *FalseVal = SI.getFalseValue();
2642 Type *SelType = SI.getType();
2644 if (Value *V = simplifySelectInst(CondVal, TrueVal, FalseVal,
2645 SQ.getWithInstruction(&SI)))
2646 return replaceInstUsesWith(SI, V);
2648 if (Instruction *I = canonicalizeSelectToShuffle(SI))
2651 if (Instruction *I = canonicalizeScalarSelectOfVecs(SI, *this))
2654 // Avoid potential infinite loops by checking for non-constant condition.
2655 // TODO: Can we assert instead by improving canonicalizeSelectToShuffle()?
2656 // Scalar select must have simplified?
2657 if (SelType->isIntOrIntVectorTy(1) && !isa<Constant>(CondVal) &&
2658 TrueVal->getType() == CondVal->getType()) {
2659 // Folding select to and/or i1 isn't poison safe in general. impliesPoison
2660 // checks whether folding it does not convert a well-defined value into
2662 if (match(TrueVal, m_One())) {
2663 if (impliesPoison(FalseVal, CondVal)) {
2664 // Change: A = select B, true, C --> A = or B, C
2665 return BinaryOperator::CreateOr(CondVal, FalseVal);
2668 if (auto *LHS = dyn_cast<FCmpInst>(CondVal))
2669 if (auto *RHS = dyn_cast<FCmpInst>(FalseVal))
2670 if (Value *V = foldLogicOfFCmps(LHS, RHS, /*IsAnd*/ false,
2671 /*IsSelectLogical*/ true))
2672 return replaceInstUsesWith(SI, V);
2674 if (match(FalseVal, m_Zero())) {
2675 if (impliesPoison(TrueVal, CondVal)) {
2676 // Change: A = select B, C, false --> A = and B, C
2677 return BinaryOperator::CreateAnd(CondVal, TrueVal);
2680 if (auto *LHS = dyn_cast<FCmpInst>(CondVal))
2681 if (auto *RHS = dyn_cast<FCmpInst>(TrueVal))
2682 if (Value *V = foldLogicOfFCmps(LHS, RHS, /*IsAnd*/ true,
2683 /*IsSelectLogical*/ true))
2684 return replaceInstUsesWith(SI, V);
2687 auto *One = ConstantInt::getTrue(SelType);
2688 auto *Zero = ConstantInt::getFalse(SelType);
2690 // We match the "full" 0 or 1 constant here to avoid a potential infinite
2691 // loop with vectors that may have undefined/poison elements.
2692 // select a, false, b -> select !a, b, false
2693 if (match(TrueVal, m_Specific(Zero))) {
2694 Value *NotCond = Builder.CreateNot(CondVal, "not." + CondVal->getName());
2695 return SelectInst::Create(NotCond, FalseVal, Zero);
2697 // select a, b, true -> select !a, true, b
2698 if (match(FalseVal, m_Specific(One))) {
2699 Value *NotCond = Builder.CreateNot(CondVal, "not." + CondVal->getName());
2700 return SelectInst::Create(NotCond, One, TrueVal);
2703 // select a, a, b -> select a, true, b
2704 if (CondVal == TrueVal)
2705 return replaceOperand(SI, 1, One);
2706 // select a, b, a -> select a, b, false
2707 if (CondVal == FalseVal)
2708 return replaceOperand(SI, 2, Zero);
2710 // select a, !a, b -> select !a, b, false
2711 if (match(TrueVal, m_Not(m_Specific(CondVal))))
2712 return SelectInst::Create(TrueVal, FalseVal, Zero);
2713 // select a, b, !a -> select !a, true, b
2714 if (match(FalseVal, m_Not(m_Specific(CondVal))))
2715 return SelectInst::Create(FalseVal, One, TrueVal);
2719 // DeMorgan in select form: !a && !b --> !(a || b)
2720 // select !a, !b, false --> not (select a, true, b)
2721 if (match(&SI, m_LogicalAnd(m_Not(m_Value(A)), m_Not(m_Value(B)))) &&
2722 (CondVal->hasOneUse() || TrueVal->hasOneUse()) &&
2723 !match(A, m_ConstantExpr()) && !match(B, m_ConstantExpr()))
2724 return BinaryOperator::CreateNot(Builder.CreateSelect(A, One, B));
2726 // DeMorgan in select form: !a || !b --> !(a && b)
2727 // select !a, true, !b --> not (select a, b, false)
2728 if (match(&SI, m_LogicalOr(m_Not(m_Value(A)), m_Not(m_Value(B)))) &&
2729 (CondVal->hasOneUse() || FalseVal->hasOneUse()) &&
2730 !match(A, m_ConstantExpr()) && !match(B, m_ConstantExpr()))
2731 return BinaryOperator::CreateNot(Builder.CreateSelect(A, B, Zero));
2733 // select (select a, true, b), true, b -> select a, true, b
2734 if (match(CondVal, m_Select(m_Value(A), m_One(), m_Value(B))) &&
2735 match(TrueVal, m_One()) && match(FalseVal, m_Specific(B)))
2736 return replaceOperand(SI, 0, A);
2737 // select (select a, b, false), b, false -> select a, b, false
2738 if (match(CondVal, m_Select(m_Value(A), m_Value(B), m_Zero())) &&
2739 match(TrueVal, m_Specific(B)) && match(FalseVal, m_Zero()))
2740 return replaceOperand(SI, 0, A);
2743 // select (~a | c), a, b -> and a, (or c, freeze(b))
2744 if (match(CondVal, m_c_Or(m_Not(m_Specific(TrueVal)), m_Value(C))) &&
2745 CondVal->hasOneUse()) {
2746 FalseVal = Builder.CreateFreeze(FalseVal);
2747 return BinaryOperator::CreateAnd(TrueVal, Builder.CreateOr(C, FalseVal));
2749 // select (~c & b), a, b -> and b, (or freeze(a), c)
2750 if (match(CondVal, m_c_And(m_Not(m_Value(C)), m_Specific(FalseVal))) &&
2751 CondVal->hasOneUse()) {
2752 TrueVal = Builder.CreateFreeze(TrueVal);
2753 return BinaryOperator::CreateAnd(FalseVal, Builder.CreateOr(C, TrueVal));
2756 if (!SelType->isVectorTy()) {
2757 if (Value *S = simplifyWithOpReplaced(TrueVal, CondVal, One, SQ,
2758 /* AllowRefinement */ true))
2759 return replaceOperand(SI, 1, S);
2760 if (Value *S = simplifyWithOpReplaced(FalseVal, CondVal, Zero, SQ,
2761 /* AllowRefinement */ true))
2762 return replaceOperand(SI, 2, S);
2765 if (match(FalseVal, m_Zero()) || match(TrueVal, m_One())) {
2767 bool IsAnd = match(FalseVal, m_Zero()) ? true : false;
2768 Value *Op1 = IsAnd ? TrueVal : FalseVal;
2769 if (isCheckForZeroAndMulWithOverflow(CondVal, Op1, IsAnd, Y)) {
2770 auto *FI = new FreezeInst(*Y, (*Y)->getName() + ".fr");
2771 InsertNewInstBefore(FI, *cast<Instruction>(Y->getUser()));
2773 return replaceInstUsesWith(SI, Op1);
2776 if (auto *Op1SI = dyn_cast<SelectInst>(Op1))
2777 if (auto *I = foldAndOrOfSelectUsingImpliedCond(CondVal, *Op1SI,
2781 if (auto *ICmp0 = dyn_cast<ICmpInst>(CondVal))
2782 if (auto *ICmp1 = dyn_cast<ICmpInst>(Op1))
2783 if (auto *V = foldAndOrOfICmps(ICmp0, ICmp1, SI, IsAnd,
2784 /* IsLogical */ true))
2785 return replaceInstUsesWith(SI, V);
2788 // select (select a, true, b), c, false -> select a, c, false
2789 // select c, (select a, true, b), false -> select c, a, false
2790 // if c implies that b is false.
2791 if (match(CondVal, m_Select(m_Value(A), m_One(), m_Value(B))) &&
2792 match(FalseVal, m_Zero())) {
2793 Optional<bool> Res = isImpliedCondition(TrueVal, B, DL);
2794 if (Res && *Res == false)
2795 return replaceOperand(SI, 0, A);
2797 if (match(TrueVal, m_Select(m_Value(A), m_One(), m_Value(B))) &&
2798 match(FalseVal, m_Zero())) {
2799 Optional<bool> Res = isImpliedCondition(CondVal, B, DL);
2800 if (Res && *Res == false)
2801 return replaceOperand(SI, 1, A);
2803 // select c, true, (select a, b, false) -> select c, true, a
2804 // select (select a, b, false), true, c -> select a, true, c
2805 // if c = false implies that b = true
2806 if (match(TrueVal, m_One()) &&
2807 match(FalseVal, m_Select(m_Value(A), m_Value(B), m_Zero()))) {
2808 Optional<bool> Res = isImpliedCondition(CondVal, B, DL, false);
2809 if (Res && *Res == true)
2810 return replaceOperand(SI, 2, A);
2812 if (match(CondVal, m_Select(m_Value(A), m_Value(B), m_Zero())) &&
2813 match(TrueVal, m_One())) {
2814 Optional<bool> Res = isImpliedCondition(FalseVal, B, DL, false);
2815 if (Res && *Res == true)
2816 return replaceOperand(SI, 0, A);
2819 // sel (sel c, a, false), true, (sel !c, b, false) -> sel c, a, b
2820 // sel (sel !c, a, false), true, (sel c, b, false) -> sel c, b, a
2822 if (match(CondVal, m_Select(m_Value(C1), m_Value(A), m_Zero())) &&
2823 match(TrueVal, m_One()) &&
2824 match(FalseVal, m_Select(m_Value(C2), m_Value(B), m_Zero()))) {
2825 if (match(C2, m_Not(m_Specific(C1)))) // first case
2826 return SelectInst::Create(C1, A, B);
2827 else if (match(C1, m_Not(m_Specific(C2)))) // second case
2828 return SelectInst::Create(C2, B, A);
2832 // Selecting between two integer or vector splat integer constants?
2834 // Note that we don't handle a scalar select of vectors:
2835 // select i1 %c, <2 x i8> <1, 1>, <2 x i8> <0, 0>
2836 // because that may need 3 instructions to splat the condition value:
2837 // extend, insertelement, shufflevector.
2839 // Do not handle i1 TrueVal and FalseVal otherwise would result in
2840 // zext/sext i1 to i1.
2841 if (SelType->isIntOrIntVectorTy() && !SelType->isIntOrIntVectorTy(1) &&
2842 CondVal->getType()->isVectorTy() == SelType->isVectorTy()) {
2843 // select C, 1, 0 -> zext C to int
2844 if (match(TrueVal, m_One()) && match(FalseVal, m_Zero()))
2845 return new ZExtInst(CondVal, SelType);
2847 // select C, -1, 0 -> sext C to int
2848 if (match(TrueVal, m_AllOnes()) && match(FalseVal, m_Zero()))
2849 return new SExtInst(CondVal, SelType);
2851 // select C, 0, 1 -> zext !C to int
2852 if (match(TrueVal, m_Zero()) && match(FalseVal, m_One())) {
2853 Value *NotCond = Builder.CreateNot(CondVal, "not." + CondVal->getName());
2854 return new ZExtInst(NotCond, SelType);
2857 // select C, 0, -1 -> sext !C to int
2858 if (match(TrueVal, m_Zero()) && match(FalseVal, m_AllOnes())) {
2859 Value *NotCond = Builder.CreateNot(CondVal, "not." + CondVal->getName());
2860 return new SExtInst(NotCond, SelType);
2864 if (auto *FCmp = dyn_cast<FCmpInst>(CondVal)) {
2865 Value *Cmp0 = FCmp->getOperand(0), *Cmp1 = FCmp->getOperand(1);
2866 // Are we selecting a value based on a comparison of the two values?
2867 if ((Cmp0 == TrueVal && Cmp1 == FalseVal) ||
2868 (Cmp0 == FalseVal && Cmp1 == TrueVal)) {
2869 // Canonicalize to use ordered comparisons by swapping the select
2873 // (X ugt Y) ? X : Y -> (X ole Y) ? Y : X
2874 if (FCmp->hasOneUse() && FCmpInst::isUnordered(FCmp->getPredicate())) {
2875 FCmpInst::Predicate InvPred = FCmp->getInversePredicate();
2876 IRBuilder<>::FastMathFlagGuard FMFG(Builder);
2877 // FIXME: The FMF should propagate from the select, not the fcmp.
2878 Builder.setFastMathFlags(FCmp->getFastMathFlags());
2879 Value *NewCond = Builder.CreateFCmp(InvPred, Cmp0, Cmp1,
2880 FCmp->getName() + ".inv");
2881 Value *NewSel = Builder.CreateSelect(NewCond, FalseVal, TrueVal);
2882 return replaceInstUsesWith(SI, NewSel);
2885 // NOTE: if we wanted to, this is where to detect MIN/MAX
2889 // Fold selecting to fabs.
2890 if (Instruction *Fabs = foldSelectWithFCmpToFabs(SI, *this))
2893 // See if we are selecting two values based on a comparison of the two values.
2894 if (ICmpInst *ICI = dyn_cast<ICmpInst>(CondVal))
2895 if (Instruction *Result = foldSelectInstWithICmp(SI, ICI))
2898 if (Instruction *Add = foldAddSubSelect(SI, Builder))
2900 if (Instruction *Add = foldOverflowingAddSubSelect(SI, Builder))
2902 if (Instruction *Or = foldSetClearBits(SI, Builder))
2904 if (Instruction *Mul = foldSelectZeroOrMul(SI, *this))
2907 // Turn (select C, (op X, Y), (op X, Z)) -> (op X, (select C, Y, Z))
2908 auto *TI = dyn_cast<Instruction>(TrueVal);
2909 auto *FI = dyn_cast<Instruction>(FalseVal);
2910 if (TI && FI && TI->getOpcode() == FI->getOpcode())
2911 if (Instruction *IV = foldSelectOpOp(SI, TI, FI))
2914 if (Instruction *I = foldSelectExtConst(SI))
2917 // Fold (select C, (gep Ptr, Idx), Ptr) -> (gep Ptr, (select C, Idx, 0))
2918 // Fold (select C, Ptr, (gep Ptr, Idx)) -> (gep Ptr, (select C, 0, Idx))
2919 auto SelectGepWithBase = [&](GetElementPtrInst *Gep, Value *Base,
2920 bool Swap) -> GetElementPtrInst * {
2921 Value *Ptr = Gep->getPointerOperand();
2922 if (Gep->getNumOperands() != 2 || Gep->getPointerOperand() != Base ||
2925 Value *Idx = Gep->getOperand(1);
2926 if (isa<VectorType>(CondVal->getType()) && !isa<VectorType>(Idx->getType()))
2928 Type *ElementType = Gep->getResultElementType();
2930 Value *NewF = Constant::getNullValue(Idx->getType());
2932 std::swap(NewT, NewF);
2934 Builder.CreateSelect(CondVal, NewT, NewF, SI.getName() + ".idx", &SI);
2935 return GetElementPtrInst::Create(ElementType, Ptr, {NewSI});
2937 if (auto *TrueGep = dyn_cast<GetElementPtrInst>(TrueVal))
2938 if (auto *NewGep = SelectGepWithBase(TrueGep, FalseVal, false))
2940 if (auto *FalseGep = dyn_cast<GetElementPtrInst>(FalseVal))
2941 if (auto *NewGep = SelectGepWithBase(FalseGep, TrueVal, true))
2944 // See if we can fold the select into one of our operands.
2945 if (SelType->isIntOrIntVectorTy() || SelType->isFPOrFPVectorTy()) {
2946 if (Instruction *FoldI = foldSelectIntoOp(SI, TrueVal, FalseVal))
2950 Instruction::CastOps CastOp;
2951 SelectPatternResult SPR = matchSelectPattern(&SI, LHS, RHS, &CastOp);
2952 auto SPF = SPR.Flavor;
2955 if (SelectPatternFlavor SPF2 = matchSelectPattern(LHS, LHS2, RHS2).Flavor)
2956 if (Instruction *R = foldSPFofSPF(cast<Instruction>(LHS), SPF2, LHS2,
2957 RHS2, SI, SPF, RHS))
2959 if (SelectPatternFlavor SPF2 = matchSelectPattern(RHS, LHS2, RHS2).Flavor)
2960 if (Instruction *R = foldSPFofSPF(cast<Instruction>(RHS), SPF2, LHS2,
2961 RHS2, SI, SPF, LHS))
2965 if (SelectPatternResult::isMinOrMax(SPF)) {
2966 // Canonicalize so that
2967 // - type casts are outside select patterns.
2968 // - float clamp is transformed to min/max pattern
2970 bool IsCastNeeded = LHS->getType() != SelType;
2971 Value *CmpLHS = cast<CmpInst>(CondVal)->getOperand(0);
2972 Value *CmpRHS = cast<CmpInst>(CondVal)->getOperand(1);
2974 (LHS->getType()->isFPOrFPVectorTy() &&
2975 ((CmpLHS != LHS && CmpLHS != RHS) ||
2976 (CmpRHS != LHS && CmpRHS != RHS)))) {
2977 CmpInst::Predicate MinMaxPred = getMinMaxPred(SPF, SPR.Ordered);
2980 if (CmpInst::isIntPredicate(MinMaxPred)) {
2981 Cmp = Builder.CreateICmp(MinMaxPred, LHS, RHS);
2983 IRBuilder<>::FastMathFlagGuard FMFG(Builder);
2985 cast<FPMathOperator>(SI.getCondition())->getFastMathFlags();
2986 Builder.setFastMathFlags(FMF);
2987 Cmp = Builder.CreateFCmp(MinMaxPred, LHS, RHS);
2990 Value *NewSI = Builder.CreateSelect(Cmp, LHS, RHS, SI.getName(), &SI);
2992 return replaceInstUsesWith(SI, NewSI);
2994 Value *NewCast = Builder.CreateCast(CastOp, NewSI, SelType);
2995 return replaceInstUsesWith(SI, NewCast);
3000 // Canonicalize select of FP values where NaN and -0.0 are not valid as
3001 // minnum/maxnum intrinsics.
3002 if (isa<FPMathOperator>(SI) && SI.hasNoNaNs() && SI.hasNoSignedZeros()) {
3004 if (match(&SI, m_OrdFMax(m_Value(X), m_Value(Y))))
3005 return replaceInstUsesWith(
3006 SI, Builder.CreateBinaryIntrinsic(Intrinsic::maxnum, X, Y, &SI));
3008 if (match(&SI, m_OrdFMin(m_Value(X), m_Value(Y))))
3009 return replaceInstUsesWith(
3010 SI, Builder.CreateBinaryIntrinsic(Intrinsic::minnum, X, Y, &SI));
3013 // See if we can fold the select into a phi node if the condition is a select.
3014 if (auto *PN = dyn_cast<PHINode>(SI.getCondition()))
3015 // The true/false values have to be live in the PHI predecessor's blocks.
3016 if (canSelectOperandBeMappingIntoPredBlock(TrueVal, SI) &&
3017 canSelectOperandBeMappingIntoPredBlock(FalseVal, SI))
3018 if (Instruction *NV = foldOpIntoPhi(SI, PN))
3021 if (SelectInst *TrueSI = dyn_cast<SelectInst>(TrueVal)) {
3022 if (TrueSI->getCondition()->getType() == CondVal->getType()) {
3023 // select(C, select(C, a, b), c) -> select(C, a, c)
3024 if (TrueSI->getCondition() == CondVal) {
3025 if (SI.getTrueValue() == TrueSI->getTrueValue())
3027 return replaceOperand(SI, 1, TrueSI->getTrueValue());
3029 // select(C0, select(C1, a, b), b) -> select(C0&C1, a, b)
3030 // We choose this as normal form to enable folding on the And and
3031 // shortening paths for the values (this helps getUnderlyingObjects() for
3033 if (TrueSI->getFalseValue() == FalseVal && TrueSI->hasOneUse()) {
3034 Value *And = Builder.CreateLogicalAnd(CondVal, TrueSI->getCondition());
3035 replaceOperand(SI, 0, And);
3036 replaceOperand(SI, 1, TrueSI->getTrueValue());
3041 if (SelectInst *FalseSI = dyn_cast<SelectInst>(FalseVal)) {
3042 if (FalseSI->getCondition()->getType() == CondVal->getType()) {
3043 // select(C, a, select(C, b, c)) -> select(C, a, c)
3044 if (FalseSI->getCondition() == CondVal) {
3045 if (SI.getFalseValue() == FalseSI->getFalseValue())
3047 return replaceOperand(SI, 2, FalseSI->getFalseValue());
3049 // select(C0, a, select(C1, a, b)) -> select(C0|C1, a, b)
3050 if (FalseSI->getTrueValue() == TrueVal && FalseSI->hasOneUse()) {
3051 Value *Or = Builder.CreateLogicalOr(CondVal, FalseSI->getCondition());
3052 replaceOperand(SI, 0, Or);
3053 replaceOperand(SI, 2, FalseSI->getFalseValue());
3059 auto canMergeSelectThroughBinop = [](BinaryOperator *BO) {
3060 // The select might be preventing a division by 0.
3061 switch (BO->getOpcode()) {
3064 case Instruction::SRem:
3065 case Instruction::URem:
3066 case Instruction::SDiv:
3067 case Instruction::UDiv:
3072 // Try to simplify a binop sandwiched between 2 selects with the same
3074 // select(C, binop(select(C, X, Y), W), Z) -> select(C, binop(X, W), Z)
3075 BinaryOperator *TrueBO;
3076 if (match(TrueVal, m_OneUse(m_BinOp(TrueBO))) &&
3077 canMergeSelectThroughBinop(TrueBO)) {
3078 if (auto *TrueBOSI = dyn_cast<SelectInst>(TrueBO->getOperand(0))) {
3079 if (TrueBOSI->getCondition() == CondVal) {
3080 replaceOperand(*TrueBO, 0, TrueBOSI->getTrueValue());
3081 Worklist.push(TrueBO);
3085 if (auto *TrueBOSI = dyn_cast<SelectInst>(TrueBO->getOperand(1))) {
3086 if (TrueBOSI->getCondition() == CondVal) {
3087 replaceOperand(*TrueBO, 1, TrueBOSI->getTrueValue());
3088 Worklist.push(TrueBO);
3094 // select(C, Z, binop(select(C, X, Y), W)) -> select(C, Z, binop(Y, W))
3095 BinaryOperator *FalseBO;
3096 if (match(FalseVal, m_OneUse(m_BinOp(FalseBO))) &&
3097 canMergeSelectThroughBinop(FalseBO)) {
3098 if (auto *FalseBOSI = dyn_cast<SelectInst>(FalseBO->getOperand(0))) {
3099 if (FalseBOSI->getCondition() == CondVal) {
3100 replaceOperand(*FalseBO, 0, FalseBOSI->getFalseValue());
3101 Worklist.push(FalseBO);
3105 if (auto *FalseBOSI = dyn_cast<SelectInst>(FalseBO->getOperand(1))) {
3106 if (FalseBOSI->getCondition() == CondVal) {
3107 replaceOperand(*FalseBO, 1, FalseBOSI->getFalseValue());
3108 Worklist.push(FalseBO);
3115 if (match(CondVal, m_Not(m_Value(NotCond))) &&
3116 !InstCombiner::shouldAvoidAbsorbingNotIntoSelect(SI)) {
3117 replaceOperand(SI, 0, NotCond);
3119 SI.swapProfMetadata();
3123 if (Instruction *I = foldVectorSelect(SI))
3126 // If we can compute the condition, there's no need for a select.
3127 // Like the above fold, we are attempting to reduce compile-time cost by
3128 // putting this fold here with limitations rather than in InstSimplify.
3129 // The motivation for this call into value tracking is to take advantage of
3130 // the assumption cache, so make sure that is populated.
3131 if (!CondVal->getType()->isVectorTy() && !AC.assumptions().empty()) {
3133 computeKnownBits(CondVal, Known, 0, &SI);
3134 if (Known.One.isOne())
3135 return replaceInstUsesWith(SI, TrueVal);
3136 if (Known.Zero.isOne())
3137 return replaceInstUsesWith(SI, FalseVal);
3140 if (Instruction *BitCastSel = foldSelectCmpBitcasts(SI, Builder))
3143 // Simplify selects that test the returned flag of cmpxchg instructions.
3144 if (Value *V = foldSelectCmpXchg(SI))
3145 return replaceInstUsesWith(SI, V);
3147 if (Instruction *Select = foldSelectBinOpIdentity(SI, TLI, *this))
3150 if (Instruction *Funnel = foldSelectFunnelShift(SI, Builder))
3153 if (Instruction *Copysign = foldSelectToCopysign(SI, Builder))
3156 if (Instruction *PN = foldSelectToPhi(SI, DT, Builder))
3157 return replaceInstUsesWith(SI, PN);
3159 if (Value *Fr = foldSelectWithFrozenICmp(SI, Builder))
3160 return replaceInstUsesWith(SI, Fr);
3162 if (Value *V = foldRoundUpIntegerWithPow2Alignment(SI, Builder))
3163 return replaceInstUsesWith(SI, V);
3165 // select(mask, mload(,,mask,0), 0) -> mload(,,mask,0)
3166 // Load inst is intentionally not checked for hasOneUse()
3167 if (match(FalseVal, m_Zero()) &&
3168 (match(TrueVal, m_MaskedLoad(m_Value(), m_Value(), m_Specific(CondVal),
3169 m_CombineOr(m_Undef(), m_Zero()))) ||
3170 match(TrueVal, m_MaskedGather(m_Value(), m_Value(), m_Specific(CondVal),
3171 m_CombineOr(m_Undef(), m_Zero()))))) {
3172 auto *MaskedInst = cast<IntrinsicInst>(TrueVal);
3173 if (isa<UndefValue>(MaskedInst->getArgOperand(3)))
3174 MaskedInst->setArgOperand(3, FalseVal /* Zero */);
3175 return replaceInstUsesWith(SI, MaskedInst);
3179 if (match(TrueVal, m_Zero()) &&
3180 (match(FalseVal, m_MaskedLoad(m_Value(), m_Value(), m_Value(Mask),
3181 m_CombineOr(m_Undef(), m_Zero()))) ||
3182 match(FalseVal, m_MaskedGather(m_Value(), m_Value(), m_Value(Mask),
3183 m_CombineOr(m_Undef(), m_Zero())))) &&
3184 (CondVal->getType() == Mask->getType())) {
3185 // We can remove the select by ensuring the load zeros all lanes the
3186 // select would have. We determine this by proving there is no overlap
3187 // between the load and select masks.
3188 // (i.e (load_mask & select_mask) == 0 == no overlap)
3189 bool CanMergeSelectIntoLoad = false;
3190 if (Value *V = simplifyAndInst(CondVal, Mask, SQ.getWithInstruction(&SI)))
3191 CanMergeSelectIntoLoad = match(V, m_Zero());
3193 if (CanMergeSelectIntoLoad) {
3194 auto *MaskedInst = cast<IntrinsicInst>(FalseVal);
3195 if (isa<UndefValue>(MaskedInst->getArgOperand(3)))
3196 MaskedInst->setArgOperand(3, TrueVal /* Zero */);
3197 return replaceInstUsesWith(SI, MaskedInst);