1 //===- InstCombineSelect.cpp ----------------------------------------------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This file implements the visitSelect function.
12 //===----------------------------------------------------------------------===//
14 #include "InstCombineInternal.h"
15 #include "llvm/Analysis/ConstantFolding.h"
16 #include "llvm/Analysis/InstructionSimplify.h"
17 #include "llvm/Analysis/ValueTracking.h"
18 #include "llvm/IR/MDBuilder.h"
19 #include "llvm/IR/PatternMatch.h"
21 using namespace PatternMatch;
23 #define DEBUG_TYPE "instcombine"
25 static SelectPatternFlavor
26 getInverseMinMaxSelectPattern(SelectPatternFlavor SPF) {
29 llvm_unreachable("unhandled!");
42 static CmpInst::Predicate getCmpPredicateForMinMax(SelectPatternFlavor SPF,
46 llvm_unreachable("unhandled!");
49 return ICmpInst::ICMP_SLT;
51 return ICmpInst::ICMP_ULT;
53 return ICmpInst::ICMP_SGT;
55 return ICmpInst::ICMP_UGT;
57 return Ordered ? FCmpInst::FCMP_OLT : FCmpInst::FCMP_ULT;
59 return Ordered ? FCmpInst::FCMP_OGT : FCmpInst::FCMP_UGT;
63 static Value *generateMinMaxSelectPattern(InstCombiner::BuilderTy *Builder,
64 SelectPatternFlavor SPF, Value *A,
66 CmpInst::Predicate Pred = getCmpPredicateForMinMax(SPF);
67 assert(CmpInst::isIntPredicate(Pred));
68 return Builder->CreateSelect(Builder->CreateICmp(Pred, A, B), A, B);
71 /// We want to turn code that looks like this:
73 /// %D = select %cond, %C, %A
75 /// %C = select %cond, %B, 0
78 /// Assuming that the specified instruction is an operand to the select, return
79 /// a bitmask indicating which operands of this instruction are foldable if they
80 /// equal the other incoming value of the select.
82 static unsigned getSelectFoldableOperands(Instruction *I) {
83 switch (I->getOpcode()) {
84 case Instruction::Add:
85 case Instruction::Mul:
86 case Instruction::And:
88 case Instruction::Xor:
89 return 3; // Can fold through either operand.
90 case Instruction::Sub: // Can only fold on the amount subtracted.
91 case Instruction::Shl: // Can only fold on the shift amount.
92 case Instruction::LShr:
93 case Instruction::AShr:
96 return 0; // Cannot fold
100 /// For the same transformation as the previous function, return the identity
101 /// constant that goes into the select.
102 static Constant *getSelectFoldableConstant(Instruction *I) {
103 switch (I->getOpcode()) {
104 default: llvm_unreachable("This cannot happen!");
105 case Instruction::Add:
106 case Instruction::Sub:
107 case Instruction::Or:
108 case Instruction::Xor:
109 case Instruction::Shl:
110 case Instruction::LShr:
111 case Instruction::AShr:
112 return Constant::getNullValue(I->getType());
113 case Instruction::And:
114 return Constant::getAllOnesValue(I->getType());
115 case Instruction::Mul:
116 return ConstantInt::get(I->getType(), 1);
120 /// We have (select c, TI, FI), and we know that TI and FI have the same opcode.
121 Instruction *InstCombiner::foldSelectOpOp(SelectInst &SI, Instruction *TI,
123 // If this is a cast from the same type, merge.
124 if (TI->getNumOperands() == 1 && TI->isCast()) {
125 Type *FIOpndTy = FI->getOperand(0)->getType();
126 if (TI->getOperand(0)->getType() != FIOpndTy)
129 // The select condition may be a vector. We may only change the operand
130 // type if the vector width remains the same (and matches the condition).
131 Type *CondTy = SI.getCondition()->getType();
132 if (CondTy->isVectorTy()) {
133 if (!FIOpndTy->isVectorTy())
135 if (CondTy->getVectorNumElements() != FIOpndTy->getVectorNumElements())
138 // TODO: If the backend knew how to deal with casts better, we could
139 // remove this limitation. For now, there's too much potential to create
140 // worse codegen by promoting the select ahead of size-altering casts
143 // Note that ValueTracking's matchSelectPattern() looks through casts
144 // without checking 'hasOneUse' when it matches min/max patterns, so this
145 // transform may end up happening anyway.
146 if (TI->getOpcode() != Instruction::BitCast &&
147 (!TI->hasOneUse() || !FI->hasOneUse()))
150 } else if (!TI->hasOneUse() || !FI->hasOneUse()) {
151 // TODO: The one-use restrictions for a scalar select could be eased if
152 // the fold of a select in visitLoadInst() was enhanced to match a pattern
153 // that includes a cast.
157 // Fold this by inserting a select from the input values.
159 Builder->CreateSelect(SI.getCondition(), TI->getOperand(0),
160 FI->getOperand(0), SI.getName() + ".v", &SI);
161 return CastInst::Create(Instruction::CastOps(TI->getOpcode()), NewSI,
165 // Only handle binary operators with one-use here. As with the cast case
166 // above, it may be possible to relax the one-use constraint, but that needs
167 // be examined carefully since it may not reduce the total number of
169 BinaryOperator *BO = dyn_cast<BinaryOperator>(TI);
170 if (!BO || !TI->hasOneUse() || !FI->hasOneUse())
173 // Figure out if the operations have any operands in common.
174 Value *MatchOp, *OtherOpT, *OtherOpF;
176 if (TI->getOperand(0) == FI->getOperand(0)) {
177 MatchOp = TI->getOperand(0);
178 OtherOpT = TI->getOperand(1);
179 OtherOpF = FI->getOperand(1);
180 MatchIsOpZero = true;
181 } else if (TI->getOperand(1) == FI->getOperand(1)) {
182 MatchOp = TI->getOperand(1);
183 OtherOpT = TI->getOperand(0);
184 OtherOpF = FI->getOperand(0);
185 MatchIsOpZero = false;
186 } else if (!TI->isCommutative()) {
188 } else if (TI->getOperand(0) == FI->getOperand(1)) {
189 MatchOp = TI->getOperand(0);
190 OtherOpT = TI->getOperand(1);
191 OtherOpF = FI->getOperand(0);
192 MatchIsOpZero = true;
193 } else if (TI->getOperand(1) == FI->getOperand(0)) {
194 MatchOp = TI->getOperand(1);
195 OtherOpT = TI->getOperand(0);
196 OtherOpF = FI->getOperand(1);
197 MatchIsOpZero = true;
202 // If we reach here, they do have operations in common.
203 Value *NewSI = Builder->CreateSelect(SI.getCondition(), OtherOpT, OtherOpF,
204 SI.getName() + ".v", &SI);
205 Value *Op0 = MatchIsOpZero ? MatchOp : NewSI;
206 Value *Op1 = MatchIsOpZero ? NewSI : MatchOp;
207 return BinaryOperator::Create(BO->getOpcode(), Op0, Op1);
210 static bool isSelect01(Constant *C1, Constant *C2) {
211 ConstantInt *C1I = dyn_cast<ConstantInt>(C1);
214 ConstantInt *C2I = dyn_cast<ConstantInt>(C2);
217 if (!C1I->isZero() && !C2I->isZero()) // One side must be zero.
219 return C1I->isOne() || C1I->isAllOnesValue() ||
220 C2I->isOne() || C2I->isAllOnesValue();
223 /// Try to fold the select into one of the operands to allow further
225 Instruction *InstCombiner::foldSelectIntoOp(SelectInst &SI, Value *TrueVal,
227 // See the comment above GetSelectFoldableOperands for a description of the
228 // transformation we are doing here.
229 if (Instruction *TVI = dyn_cast<Instruction>(TrueVal)) {
230 if (TVI->hasOneUse() && TVI->getNumOperands() == 2 &&
231 !isa<Constant>(FalseVal)) {
232 if (unsigned SFO = getSelectFoldableOperands(TVI)) {
233 unsigned OpToFold = 0;
234 if ((SFO & 1) && FalseVal == TVI->getOperand(0)) {
236 } else if ((SFO & 2) && FalseVal == TVI->getOperand(1)) {
241 Constant *C = getSelectFoldableConstant(TVI);
242 Value *OOp = TVI->getOperand(2-OpToFold);
243 // Avoid creating select between 2 constants unless it's selecting
244 // between 0, 1 and -1.
245 if (!isa<Constant>(OOp) || isSelect01(C, cast<Constant>(OOp))) {
246 Value *NewSel = Builder->CreateSelect(SI.getCondition(), OOp, C);
247 NewSel->takeName(TVI);
248 BinaryOperator *TVI_BO = cast<BinaryOperator>(TVI);
249 BinaryOperator *BO = BinaryOperator::Create(TVI_BO->getOpcode(),
251 BO->copyIRFlags(TVI_BO);
259 if (Instruction *FVI = dyn_cast<Instruction>(FalseVal)) {
260 if (FVI->hasOneUse() && FVI->getNumOperands() == 2 &&
261 !isa<Constant>(TrueVal)) {
262 if (unsigned SFO = getSelectFoldableOperands(FVI)) {
263 unsigned OpToFold = 0;
264 if ((SFO & 1) && TrueVal == FVI->getOperand(0)) {
266 } else if ((SFO & 2) && TrueVal == FVI->getOperand(1)) {
271 Constant *C = getSelectFoldableConstant(FVI);
272 Value *OOp = FVI->getOperand(2-OpToFold);
273 // Avoid creating select between 2 constants unless it's selecting
274 // between 0, 1 and -1.
275 if (!isa<Constant>(OOp) || isSelect01(C, cast<Constant>(OOp))) {
276 Value *NewSel = Builder->CreateSelect(SI.getCondition(), C, OOp);
277 NewSel->takeName(FVI);
278 BinaryOperator *FVI_BO = cast<BinaryOperator>(FVI);
279 BinaryOperator *BO = BinaryOperator::Create(FVI_BO->getOpcode(),
281 BO->copyIRFlags(FVI_BO);
293 /// (select (icmp eq (and X, C1), 0), Y, (or Y, C2))
295 /// (or (shl (and X, C1), C3), y)
297 /// C1 and C2 are both powers of 2
299 /// C3 = Log(C2) - Log(C1)
301 /// This transform handles cases where:
302 /// 1. The icmp predicate is inverted
303 /// 2. The select operands are reversed
304 /// 3. The magnitude of C2 and C1 are flipped
305 static Value *foldSelectICmpAndOr(const SelectInst &SI, Value *TrueVal,
307 InstCombiner::BuilderTy *Builder) {
308 const ICmpInst *IC = dyn_cast<ICmpInst>(SI.getCondition());
309 if (!IC || !IC->isEquality() || !SI.getType()->isIntegerTy())
312 Value *CmpLHS = IC->getOperand(0);
313 Value *CmpRHS = IC->getOperand(1);
315 if (!match(CmpRHS, m_Zero()))
320 if (!match(CmpLHS, m_And(m_Value(X), m_Power2(C1))))
324 bool OrOnTrueVal = false;
325 bool OrOnFalseVal = match(FalseVal, m_Or(m_Specific(TrueVal), m_Power2(C2)));
327 OrOnTrueVal = match(TrueVal, m_Or(m_Specific(FalseVal), m_Power2(C2)));
329 if (!OrOnFalseVal && !OrOnTrueVal)
333 Value *Y = OrOnFalseVal ? TrueVal : FalseVal;
335 unsigned C1Log = C1->logBase2();
336 unsigned C2Log = C2->logBase2();
338 V = Builder->CreateZExtOrTrunc(V, Y->getType());
339 V = Builder->CreateShl(V, C2Log - C1Log);
340 } else if (C1Log > C2Log) {
341 V = Builder->CreateLShr(V, C1Log - C2Log);
342 V = Builder->CreateZExtOrTrunc(V, Y->getType());
344 V = Builder->CreateZExtOrTrunc(V, Y->getType());
346 ICmpInst::Predicate Pred = IC->getPredicate();
347 if ((Pred == ICmpInst::ICMP_NE && OrOnFalseVal) ||
348 (Pred == ICmpInst::ICMP_EQ && OrOnTrueVal))
349 V = Builder->CreateXor(V, *C2);
351 return Builder->CreateOr(V, Y);
354 /// Attempt to fold a cttz/ctlz followed by a icmp plus select into a single
355 /// call to cttz/ctlz with flag 'is_zero_undef' cleared.
357 /// For example, we can fold the following code sequence:
359 /// %0 = tail call i32 @llvm.cttz.i32(i32 %x, i1 true)
360 /// %1 = icmp ne i32 %x, 0
361 /// %2 = select i1 %1, i32 %0, i32 32
365 /// %0 = tail call i32 @llvm.cttz.i32(i32 %x, i1 false)
366 static Value *foldSelectCttzCtlz(ICmpInst *ICI, Value *TrueVal, Value *FalseVal,
367 InstCombiner::BuilderTy *Builder) {
368 ICmpInst::Predicate Pred = ICI->getPredicate();
369 Value *CmpLHS = ICI->getOperand(0);
370 Value *CmpRHS = ICI->getOperand(1);
372 // Check if the condition value compares a value for equality against zero.
373 if (!ICI->isEquality() || !match(CmpRHS, m_Zero()))
376 Value *Count = FalseVal;
377 Value *ValueOnZero = TrueVal;
378 if (Pred == ICmpInst::ICMP_NE)
379 std::swap(Count, ValueOnZero);
381 // Skip zero extend/truncate.
383 if (match(Count, m_ZExt(m_Value(V))) ||
384 match(Count, m_Trunc(m_Value(V))))
387 // Check if the value propagated on zero is a constant number equal to the
388 // sizeof in bits of 'Count'.
389 unsigned SizeOfInBits = Count->getType()->getScalarSizeInBits();
390 if (!match(ValueOnZero, m_SpecificInt(SizeOfInBits)))
393 // Check that 'Count' is a call to intrinsic cttz/ctlz. Also check that the
394 // input to the cttz/ctlz is used as LHS for the compare instruction.
395 if (match(Count, m_Intrinsic<Intrinsic::cttz>(m_Specific(CmpLHS))) ||
396 match(Count, m_Intrinsic<Intrinsic::ctlz>(m_Specific(CmpLHS)))) {
397 IntrinsicInst *II = cast<IntrinsicInst>(Count);
398 IRBuilder<> Builder(II);
399 // Explicitly clear the 'undef_on_zero' flag.
400 IntrinsicInst *NewI = cast<IntrinsicInst>(II->clone());
401 Type *Ty = NewI->getArgOperand(1)->getType();
402 NewI->setArgOperand(1, Constant::getNullValue(Ty));
403 Builder.Insert(NewI);
404 return Builder.CreateZExtOrTrunc(NewI, ValueOnZero->getType());
410 /// Return true if we find and adjust an icmp+select pattern where the compare
411 /// is with a constant that can be incremented or decremented to match the
412 /// minimum or maximum idiom.
413 static bool adjustMinMax(SelectInst &Sel, ICmpInst &Cmp) {
414 ICmpInst::Predicate Pred = Cmp.getPredicate();
415 Value *CmpLHS = Cmp.getOperand(0);
416 Value *CmpRHS = Cmp.getOperand(1);
417 Value *TrueVal = Sel.getTrueValue();
418 Value *FalseVal = Sel.getFalseValue();
420 // We may move or edit the compare, so make sure the select is the only user.
422 if (!Cmp.hasOneUse() || !match(CmpRHS, m_APInt(CmpC)))
425 // These transforms only work for selects of integers or vector selects of
427 Type *SelTy = Sel.getType();
428 auto *SelEltTy = dyn_cast<IntegerType>(SelTy->getScalarType());
429 if (!SelEltTy || SelTy->isVectorTy() != Cmp.getType()->isVectorTy())
432 Constant *AdjustedRHS;
433 if (Pred == ICmpInst::ICMP_UGT || Pred == ICmpInst::ICMP_SGT)
434 AdjustedRHS = ConstantInt::get(CmpRHS->getType(), *CmpC + 1);
435 else if (Pred == ICmpInst::ICMP_ULT || Pred == ICmpInst::ICMP_SLT)
436 AdjustedRHS = ConstantInt::get(CmpRHS->getType(), *CmpC - 1);
440 // X > C ? X : C+1 --> X < C+1 ? C+1 : X
441 // X < C ? X : C-1 --> X > C-1 ? C-1 : X
442 if ((CmpLHS == TrueVal && AdjustedRHS == FalseVal) ||
443 (CmpLHS == FalseVal && AdjustedRHS == TrueVal)) {
444 ; // Nothing to do here. Values match without any sign/zero extension.
446 // Types do not match. Instead of calculating this with mixed types, promote
447 // all to the larger type. This enables scalar evolution to analyze this
449 else if (CmpRHS->getType()->getScalarSizeInBits() < SelEltTy->getBitWidth()) {
450 Constant *SextRHS = ConstantExpr::getSExt(AdjustedRHS, SelTy);
452 // X = sext x; x >s c ? X : C+1 --> X = sext x; X <s C+1 ? C+1 : X
453 // X = sext x; x <s c ? X : C-1 --> X = sext x; X >s C-1 ? C-1 : X
454 // X = sext x; x >u c ? X : C+1 --> X = sext x; X <u C+1 ? C+1 : X
455 // X = sext x; x <u c ? X : C-1 --> X = sext x; X >u C-1 ? C-1 : X
456 if (match(TrueVal, m_SExt(m_Specific(CmpLHS))) && SextRHS == FalseVal) {
458 AdjustedRHS = SextRHS;
459 } else if (match(FalseVal, m_SExt(m_Specific(CmpLHS))) &&
460 SextRHS == TrueVal) {
462 AdjustedRHS = SextRHS;
463 } else if (Cmp.isUnsigned()) {
464 Constant *ZextRHS = ConstantExpr::getZExt(AdjustedRHS, SelTy);
465 // X = zext x; x >u c ? X : C+1 --> X = zext x; X <u C+1 ? C+1 : X
466 // X = zext x; x <u c ? X : C-1 --> X = zext x; X >u C-1 ? C-1 : X
467 // zext + signed compare cannot be changed:
468 // 0xff <s 0x00, but 0x00ff >s 0x0000
469 if (match(TrueVal, m_ZExt(m_Specific(CmpLHS))) && ZextRHS == FalseVal) {
471 AdjustedRHS = ZextRHS;
472 } else if (match(FalseVal, m_ZExt(m_Specific(CmpLHS))) &&
473 ZextRHS == TrueVal) {
475 AdjustedRHS = ZextRHS;
486 Pred = ICmpInst::getSwappedPredicate(Pred);
487 CmpRHS = AdjustedRHS;
488 std::swap(FalseVal, TrueVal);
489 Cmp.setPredicate(Pred);
490 Cmp.setOperand(0, CmpLHS);
491 Cmp.setOperand(1, CmpRHS);
492 Sel.setOperand(1, TrueVal);
493 Sel.setOperand(2, FalseVal);
494 Sel.swapProfMetadata();
496 // Move the compare instruction right before the select instruction. Otherwise
497 // the sext/zext value may be defined after the compare instruction uses it.
498 Cmp.moveBefore(&Sel);
503 /// If this is an integer min/max where the select's 'true' operand is a
504 /// constant, canonicalize that constant to the 'false' operand:
505 /// select (icmp Pred X, C), C, X --> select (icmp Pred' X, C), X, C
507 canonicalizeMinMaxWithConstant(SelectInst &Sel, ICmpInst &Cmp,
508 InstCombiner::BuilderTy &Builder) {
509 // TODO: We should also canonicalize min/max when the select has a different
510 // constant value than the cmp constant, but we need to fix the backend first.
511 if (!Cmp.hasOneUse() || !isa<Constant>(Cmp.getOperand(1)) ||
512 !isa<Constant>(Sel.getTrueValue()) ||
513 isa<Constant>(Sel.getFalseValue()) ||
514 Cmp.getOperand(1) != Sel.getTrueValue())
517 // Canonicalize the compare predicate based on whether we have min or max.
519 ICmpInst::Predicate NewPred;
520 SelectPatternResult SPR = matchSelectPattern(&Sel, LHS, RHS);
521 switch (SPR.Flavor) {
522 case SPF_SMIN: NewPred = ICmpInst::ICMP_SLT; break;
523 case SPF_UMIN: NewPred = ICmpInst::ICMP_ULT; break;
524 case SPF_SMAX: NewPred = ICmpInst::ICMP_SGT; break;
525 case SPF_UMAX: NewPred = ICmpInst::ICMP_UGT; break;
526 default: return nullptr;
529 // Canonicalize the constant to the right side.
530 if (isa<Constant>(LHS))
533 Value *NewCmp = Builder.CreateICmp(NewPred, LHS, RHS);
534 SelectInst *NewSel = SelectInst::Create(NewCmp, LHS, RHS, "", nullptr, &Sel);
536 // We swapped the select operands, so swap the metadata too.
537 NewSel->swapProfMetadata();
541 /// Visit a SelectInst that has an ICmpInst as its first operand.
542 Instruction *InstCombiner::foldSelectInstWithICmp(SelectInst &SI,
544 if (Instruction *NewSel = canonicalizeMinMaxWithConstant(SI, *ICI, *Builder))
547 bool Changed = adjustMinMax(SI, *ICI);
549 ICmpInst::Predicate Pred = ICI->getPredicate();
550 Value *CmpLHS = ICI->getOperand(0);
551 Value *CmpRHS = ICI->getOperand(1);
552 Value *TrueVal = SI.getTrueValue();
553 Value *FalseVal = SI.getFalseValue();
555 // Transform (X >s -1) ? C1 : C2 --> ((X >>s 31) & (C2 - C1)) + C1
556 // and (X <s 0) ? C2 : C1 --> ((X >>s 31) & (C2 - C1)) + C1
557 // FIXME: Type and constness constraints could be lifted, but we have to
558 // watch code size carefully. We should consider xor instead of
559 // sub/add when we decide to do that.
560 if (IntegerType *Ty = dyn_cast<IntegerType>(CmpLHS->getType())) {
561 if (TrueVal->getType() == Ty) {
562 if (ConstantInt *Cmp = dyn_cast<ConstantInt>(CmpRHS)) {
563 ConstantInt *C1 = nullptr, *C2 = nullptr;
564 if (Pred == ICmpInst::ICMP_SGT && Cmp->isAllOnesValue()) {
565 C1 = dyn_cast<ConstantInt>(TrueVal);
566 C2 = dyn_cast<ConstantInt>(FalseVal);
567 } else if (Pred == ICmpInst::ICMP_SLT && Cmp->isNullValue()) {
568 C1 = dyn_cast<ConstantInt>(FalseVal);
569 C2 = dyn_cast<ConstantInt>(TrueVal);
572 // This shift results in either -1 or 0.
573 Value *AShr = Builder->CreateAShr(CmpLHS, Ty->getBitWidth()-1);
575 // Check if we can express the operation with a single or.
576 if (C2->isAllOnesValue())
577 return replaceInstUsesWith(SI, Builder->CreateOr(AShr, C1));
579 Value *And = Builder->CreateAnd(AShr, C2->getValue()-C1->getValue());
580 return replaceInstUsesWith(SI, Builder->CreateAdd(And, C1));
586 // NOTE: if we wanted to, this is where to detect integer MIN/MAX
588 if (CmpRHS != CmpLHS && isa<Constant>(CmpRHS)) {
589 if (CmpLHS == TrueVal && Pred == ICmpInst::ICMP_EQ) {
590 // Transform (X == C) ? X : Y -> (X == C) ? C : Y
591 SI.setOperand(1, CmpRHS);
593 } else if (CmpLHS == FalseVal && Pred == ICmpInst::ICMP_NE) {
594 // Transform (X != C) ? Y : X -> (X != C) ? Y : C
595 SI.setOperand(2, CmpRHS);
600 // FIXME: This code is nearly duplicated in InstSimplify. Using/refactoring
601 // decomposeBitTestICmp() might help.
604 DL.getTypeSizeInBits(TrueVal->getType()->getScalarType());
605 APInt MinSignedValue = APInt::getSignBit(BitWidth);
609 bool IsBitTest = false;
610 if (ICmpInst::isEquality(Pred) &&
611 match(CmpLHS, m_And(m_Value(X), m_Power2(Y))) &&
612 match(CmpRHS, m_Zero())) {
614 TrueWhenUnset = Pred == ICmpInst::ICMP_EQ;
615 } else if (Pred == ICmpInst::ICMP_SLT && match(CmpRHS, m_Zero())) {
619 TrueWhenUnset = false;
620 } else if (Pred == ICmpInst::ICMP_SGT && match(CmpRHS, m_AllOnes())) {
624 TrueWhenUnset = true;
628 // (X & Y) == 0 ? X : X ^ Y --> X & ~Y
629 if (TrueWhenUnset && TrueVal == X &&
630 match(FalseVal, m_Xor(m_Specific(X), m_APInt(C))) && *Y == *C)
631 V = Builder->CreateAnd(X, ~(*Y));
632 // (X & Y) != 0 ? X ^ Y : X --> X & ~Y
633 else if (!TrueWhenUnset && FalseVal == X &&
634 match(TrueVal, m_Xor(m_Specific(X), m_APInt(C))) && *Y == *C)
635 V = Builder->CreateAnd(X, ~(*Y));
636 // (X & Y) == 0 ? X ^ Y : X --> X | Y
637 else if (TrueWhenUnset && FalseVal == X &&
638 match(TrueVal, m_Xor(m_Specific(X), m_APInt(C))) && *Y == *C)
639 V = Builder->CreateOr(X, *Y);
640 // (X & Y) != 0 ? X : X ^ Y --> X | Y
641 else if (!TrueWhenUnset && TrueVal == X &&
642 match(FalseVal, m_Xor(m_Specific(X), m_APInt(C))) && *Y == *C)
643 V = Builder->CreateOr(X, *Y);
646 return replaceInstUsesWith(SI, V);
650 if (Value *V = foldSelectICmpAndOr(SI, TrueVal, FalseVal, Builder))
651 return replaceInstUsesWith(SI, V);
653 if (Value *V = foldSelectCttzCtlz(ICI, TrueVal, FalseVal, Builder))
654 return replaceInstUsesWith(SI, V);
656 return Changed ? &SI : nullptr;
660 /// SI is a select whose condition is a PHI node (but the two may be in
661 /// different blocks). See if the true/false values (V) are live in all of the
662 /// predecessor blocks of the PHI. For example, cases like this can't be mapped:
664 /// X = phi [ C1, BB1], [C2, BB2]
666 /// Z = select X, Y, 0
668 /// because Y is not live in BB1/BB2.
670 static bool canSelectOperandBeMappingIntoPredBlock(const Value *V,
671 const SelectInst &SI) {
672 // If the value is a non-instruction value like a constant or argument, it
673 // can always be mapped.
674 const Instruction *I = dyn_cast<Instruction>(V);
677 // If V is a PHI node defined in the same block as the condition PHI, we can
678 // map the arguments.
679 const PHINode *CondPHI = cast<PHINode>(SI.getCondition());
681 if (const PHINode *VP = dyn_cast<PHINode>(I))
682 if (VP->getParent() == CondPHI->getParent())
685 // Otherwise, if the PHI and select are defined in the same block and if V is
686 // defined in a different block, then we can transform it.
687 if (SI.getParent() == CondPHI->getParent() &&
688 I->getParent() != CondPHI->getParent())
691 // Otherwise we have a 'hard' case and we can't tell without doing more
692 // detailed dominator based analysis, punt.
696 /// We have an SPF (e.g. a min or max) of an SPF of the form:
697 /// SPF2(SPF1(A, B), C)
698 Instruction *InstCombiner::foldSPFofSPF(Instruction *Inner,
699 SelectPatternFlavor SPF1,
702 SelectPatternFlavor SPF2, Value *C) {
703 if (Outer.getType() != Inner->getType())
706 if (C == A || C == B) {
707 // MAX(MAX(A, B), B) -> MAX(A, B)
708 // MIN(MIN(a, b), a) -> MIN(a, b)
710 return replaceInstUsesWith(Outer, Inner);
712 // MAX(MIN(a, b), a) -> a
713 // MIN(MAX(a, b), a) -> a
714 if ((SPF1 == SPF_SMIN && SPF2 == SPF_SMAX) ||
715 (SPF1 == SPF_SMAX && SPF2 == SPF_SMIN) ||
716 (SPF1 == SPF_UMIN && SPF2 == SPF_UMAX) ||
717 (SPF1 == SPF_UMAX && SPF2 == SPF_UMIN))
718 return replaceInstUsesWith(Outer, C);
722 const APInt *CB, *CC;
723 if (match(B, m_APInt(CB)) && match(C, m_APInt(CC))) {
724 // MIN(MIN(A, 23), 97) -> MIN(A, 23)
725 // MAX(MAX(A, 97), 23) -> MAX(A, 97)
726 if ((SPF1 == SPF_UMIN && CB->ule(*CC)) ||
727 (SPF1 == SPF_SMIN && CB->sle(*CC)) ||
728 (SPF1 == SPF_UMAX && CB->uge(*CC)) ||
729 (SPF1 == SPF_SMAX && CB->sge(*CC)))
730 return replaceInstUsesWith(Outer, Inner);
732 // MIN(MIN(A, 97), 23) -> MIN(A, 23)
733 // MAX(MAX(A, 23), 97) -> MAX(A, 97)
734 if ((SPF1 == SPF_UMIN && CB->ugt(*CC)) ||
735 (SPF1 == SPF_SMIN && CB->sgt(*CC)) ||
736 (SPF1 == SPF_UMAX && CB->ult(*CC)) ||
737 (SPF1 == SPF_SMAX && CB->slt(*CC))) {
738 Outer.replaceUsesOfWith(Inner, A);
744 // ABS(ABS(X)) -> ABS(X)
745 // NABS(NABS(X)) -> NABS(X)
746 if (SPF1 == SPF2 && (SPF1 == SPF_ABS || SPF1 == SPF_NABS)) {
747 return replaceInstUsesWith(Outer, Inner);
750 // ABS(NABS(X)) -> ABS(X)
751 // NABS(ABS(X)) -> NABS(X)
752 if ((SPF1 == SPF_ABS && SPF2 == SPF_NABS) ||
753 (SPF1 == SPF_NABS && SPF2 == SPF_ABS)) {
754 SelectInst *SI = cast<SelectInst>(Inner);
756 Builder->CreateSelect(SI->getCondition(), SI->getFalseValue(),
757 SI->getTrueValue(), SI->getName(), SI);
758 return replaceInstUsesWith(Outer, NewSI);
761 auto IsFreeOrProfitableToInvert =
762 [&](Value *V, Value *&NotV, bool &ElidesXor) {
763 if (match(V, m_Not(m_Value(NotV)))) {
764 // If V has at most 2 uses then we can get rid of the xor operation
766 ElidesXor |= !V->hasNUsesOrMore(3);
770 if (IsFreeToInvert(V, !V->hasNUsesOrMore(3))) {
778 Value *NotA, *NotB, *NotC;
779 bool ElidesXor = false;
781 // MIN(MIN(~A, ~B), ~C) == ~MAX(MAX(A, B), C)
782 // MIN(MAX(~A, ~B), ~C) == ~MAX(MIN(A, B), C)
783 // MAX(MIN(~A, ~B), ~C) == ~MIN(MAX(A, B), C)
784 // MAX(MAX(~A, ~B), ~C) == ~MIN(MIN(A, B), C)
786 // This transform is performance neutral if we can elide at least one xor from
787 // the set of three operands, since we'll be tacking on an xor at the very
789 if (IsFreeOrProfitableToInvert(A, NotA, ElidesXor) &&
790 IsFreeOrProfitableToInvert(B, NotB, ElidesXor) &&
791 IsFreeOrProfitableToInvert(C, NotC, ElidesXor) && ElidesXor) {
793 NotA = Builder->CreateNot(A);
795 NotB = Builder->CreateNot(B);
797 NotC = Builder->CreateNot(C);
799 Value *NewInner = generateMinMaxSelectPattern(
800 Builder, getInverseMinMaxSelectPattern(SPF1), NotA, NotB);
801 Value *NewOuter = Builder->CreateNot(generateMinMaxSelectPattern(
802 Builder, getInverseMinMaxSelectPattern(SPF2), NewInner, NotC));
803 return replaceInstUsesWith(Outer, NewOuter);
809 /// If one of the constants is zero (we know they can't both be) and we have an
810 /// icmp instruction with zero, and we have an 'and' with the non-constant value
811 /// and a power of two we can turn the select into a shift on the result of the
813 static Value *foldSelectICmpAnd(const SelectInst &SI, ConstantInt *TrueVal,
814 ConstantInt *FalseVal,
815 InstCombiner::BuilderTy *Builder) {
816 const ICmpInst *IC = dyn_cast<ICmpInst>(SI.getCondition());
817 if (!IC || !IC->isEquality() || !SI.getType()->isIntegerTy())
820 if (!match(IC->getOperand(1), m_Zero()))
824 Value *LHS = IC->getOperand(0);
825 if (!match(LHS, m_And(m_Value(), m_ConstantInt(AndRHS))))
828 // If both select arms are non-zero see if we have a select of the form
829 // 'x ? 2^n + C : C'. Then we can offset both arms by C, use the logic
830 // for 'x ? 2^n : 0' and fix the thing up at the end.
831 ConstantInt *Offset = nullptr;
832 if (!TrueVal->isZero() && !FalseVal->isZero()) {
833 if ((TrueVal->getValue() - FalseVal->getValue()).isPowerOf2())
835 else if ((FalseVal->getValue() - TrueVal->getValue()).isPowerOf2())
840 // Adjust TrueVal and FalseVal to the offset.
841 TrueVal = ConstantInt::get(Builder->getContext(),
842 TrueVal->getValue() - Offset->getValue());
843 FalseVal = ConstantInt::get(Builder->getContext(),
844 FalseVal->getValue() - Offset->getValue());
847 // Make sure the mask in the 'and' and one of the select arms is a power of 2.
848 if (!AndRHS->getValue().isPowerOf2() ||
849 (!TrueVal->getValue().isPowerOf2() &&
850 !FalseVal->getValue().isPowerOf2()))
853 // Determine which shift is needed to transform result of the 'and' into the
855 ConstantInt *ValC = !TrueVal->isZero() ? TrueVal : FalseVal;
856 unsigned ValZeros = ValC->getValue().logBase2();
857 unsigned AndZeros = AndRHS->getValue().logBase2();
859 // If types don't match we can still convert the select by introducing a zext
860 // or a trunc of the 'and'. The trunc case requires that all of the truncated
861 // bits are zero, we can figure that out by looking at the 'and' mask.
862 if (AndZeros >= ValC->getBitWidth())
865 Value *V = Builder->CreateZExtOrTrunc(LHS, SI.getType());
866 if (ValZeros > AndZeros)
867 V = Builder->CreateShl(V, ValZeros - AndZeros);
868 else if (ValZeros < AndZeros)
869 V = Builder->CreateLShr(V, AndZeros - ValZeros);
871 // Okay, now we know that everything is set up, we just don't know whether we
872 // have a icmp_ne or icmp_eq and whether the true or false val is the zero.
873 bool ShouldNotVal = !TrueVal->isZero();
874 ShouldNotVal ^= IC->getPredicate() == ICmpInst::ICMP_NE;
876 V = Builder->CreateXor(V, ValC);
878 // Apply an offset if needed.
880 V = Builder->CreateAdd(V, Offset);
884 /// Turn select C, (X + Y), (X - Y) --> (X + (select C, Y, (-Y))).
885 /// This is even legal for FP.
886 static Instruction *foldAddSubSelect(SelectInst &SI,
887 InstCombiner::BuilderTy &Builder) {
888 Value *CondVal = SI.getCondition();
889 Value *TrueVal = SI.getTrueValue();
890 Value *FalseVal = SI.getFalseValue();
891 auto *TI = dyn_cast<Instruction>(TrueVal);
892 auto *FI = dyn_cast<Instruction>(FalseVal);
893 if (!TI || !FI || !TI->hasOneUse() || !FI->hasOneUse())
896 Instruction *AddOp = nullptr, *SubOp = nullptr;
897 if ((TI->getOpcode() == Instruction::Sub &&
898 FI->getOpcode() == Instruction::Add) ||
899 (TI->getOpcode() == Instruction::FSub &&
900 FI->getOpcode() == Instruction::FAdd)) {
903 } else if ((FI->getOpcode() == Instruction::Sub &&
904 TI->getOpcode() == Instruction::Add) ||
905 (FI->getOpcode() == Instruction::FSub &&
906 TI->getOpcode() == Instruction::FAdd)) {
912 Value *OtherAddOp = nullptr;
913 if (SubOp->getOperand(0) == AddOp->getOperand(0)) {
914 OtherAddOp = AddOp->getOperand(1);
915 } else if (SubOp->getOperand(0) == AddOp->getOperand(1)) {
916 OtherAddOp = AddOp->getOperand(0);
920 // So at this point we know we have (Y -> OtherAddOp):
921 // select C, (add X, Y), (sub X, Z)
922 Value *NegVal; // Compute -Z
923 if (SI.getType()->isFPOrFPVectorTy()) {
924 NegVal = Builder.CreateFNeg(SubOp->getOperand(1));
925 if (Instruction *NegInst = dyn_cast<Instruction>(NegVal)) {
926 FastMathFlags Flags = AddOp->getFastMathFlags();
927 Flags &= SubOp->getFastMathFlags();
928 NegInst->setFastMathFlags(Flags);
931 NegVal = Builder.CreateNeg(SubOp->getOperand(1));
934 Value *NewTrueOp = OtherAddOp;
935 Value *NewFalseOp = NegVal;
937 std::swap(NewTrueOp, NewFalseOp);
938 Value *NewSel = Builder.CreateSelect(CondVal, NewTrueOp, NewFalseOp,
939 SI.getName() + ".p", &SI);
941 if (SI.getType()->isFPOrFPVectorTy()) {
943 BinaryOperator::CreateFAdd(SubOp->getOperand(0), NewSel);
945 FastMathFlags Flags = AddOp->getFastMathFlags();
946 Flags &= SubOp->getFastMathFlags();
947 RI->setFastMathFlags(Flags);
950 return BinaryOperator::CreateAdd(SubOp->getOperand(0), NewSel);
956 Instruction *InstCombiner::foldSelectExtConst(SelectInst &Sel) {
957 Instruction *ExtInst;
958 if (!match(Sel.getTrueValue(), m_Instruction(ExtInst)) &&
959 !match(Sel.getFalseValue(), m_Instruction(ExtInst)))
962 auto ExtOpcode = ExtInst->getOpcode();
963 if (ExtOpcode != Instruction::ZExt && ExtOpcode != Instruction::SExt)
966 // TODO: Handle larger types? That requires adjusting FoldOpIntoSelect too.
967 Value *X = ExtInst->getOperand(0);
968 Type *SmallType = X->getType();
969 if (!SmallType->getScalarType()->isIntegerTy(1))
973 if (!match(Sel.getTrueValue(), m_Constant(C)) &&
974 !match(Sel.getFalseValue(), m_Constant(C)))
977 // If the constant is the same after truncation to the smaller type and
978 // extension to the original type, we can narrow the select.
979 Value *Cond = Sel.getCondition();
980 Type *SelType = Sel.getType();
981 Constant *TruncC = ConstantExpr::getTrunc(C, SmallType);
982 Constant *ExtC = ConstantExpr::getCast(ExtOpcode, TruncC, SelType);
984 Value *TruncCVal = cast<Value>(TruncC);
985 if (ExtInst == Sel.getFalseValue())
986 std::swap(X, TruncCVal);
988 // select Cond, (ext X), C --> ext(select Cond, X, C')
989 // select Cond, C, (ext X) --> ext(select Cond, C', X)
990 Value *NewSel = Builder->CreateSelect(Cond, X, TruncCVal, "narrow", &Sel);
991 return CastInst::Create(Instruction::CastOps(ExtOpcode), NewSel, SelType);
994 // If one arm of the select is the extend of the condition, replace that arm
995 // with the extension of the appropriate known bool value.
997 if (ExtInst == Sel.getTrueValue()) {
998 // select X, (sext X), C --> select X, -1, C
999 // select X, (zext X), C --> select X, 1, C
1000 Constant *One = ConstantInt::getTrue(SmallType);
1001 Constant *AllOnesOrOne = ConstantExpr::getCast(ExtOpcode, One, SelType);
1002 return SelectInst::Create(Cond, AllOnesOrOne, C, "", nullptr, &Sel);
1004 // select X, C, (sext X) --> select X, C, 0
1005 // select X, C, (zext X) --> select X, C, 0
1006 Constant *Zero = ConstantInt::getNullValue(SelType);
1007 return SelectInst::Create(Cond, C, Zero, "", nullptr, &Sel);
1014 /// Try to transform a vector select with a constant condition vector into a
1015 /// shuffle for easier combining with other shuffles and insert/extract.
1016 static Instruction *canonicalizeSelectToShuffle(SelectInst &SI) {
1017 Value *CondVal = SI.getCondition();
1019 if (!CondVal->getType()->isVectorTy() || !match(CondVal, m_Constant(CondC)))
1022 unsigned NumElts = CondVal->getType()->getVectorNumElements();
1023 SmallVector<Constant *, 16> Mask;
1024 Mask.reserve(NumElts);
1025 Type *Int32Ty = Type::getInt32Ty(CondVal->getContext());
1026 for (unsigned i = 0; i != NumElts; ++i) {
1027 Constant *Elt = CondC->getAggregateElement(i);
1031 if (Elt->isOneValue()) {
1032 // If the select condition element is true, choose from the 1st vector.
1033 Mask.push_back(ConstantInt::get(Int32Ty, i));
1034 } else if (Elt->isNullValue()) {
1035 // If the select condition element is false, choose from the 2nd vector.
1036 Mask.push_back(ConstantInt::get(Int32Ty, i + NumElts));
1037 } else if (isa<UndefValue>(Elt)) {
1038 // If the select condition element is undef, the shuffle mask is undef.
1039 Mask.push_back(UndefValue::get(Int32Ty));
1041 // Bail out on a constant expression.
1046 return new ShuffleVectorInst(SI.getTrueValue(), SI.getFalseValue(),
1047 ConstantVector::get(Mask));
1050 /// Reuse bitcasted operands between a compare and select:
1051 /// select (cmp (bitcast C), (bitcast D)), (bitcast' C), (bitcast' D) -->
1052 /// bitcast (select (cmp (bitcast C), (bitcast D)), (bitcast C), (bitcast D))
1053 static Instruction *foldSelectCmpBitcasts(SelectInst &Sel,
1054 InstCombiner::BuilderTy &Builder) {
1055 Value *Cond = Sel.getCondition();
1056 Value *TVal = Sel.getTrueValue();
1057 Value *FVal = Sel.getFalseValue();
1059 CmpInst::Predicate Pred;
1061 if (!match(Cond, m_Cmp(Pred, m_Value(A), m_Value(B))))
1064 // The select condition is a compare instruction. If the select's true/false
1065 // values are already the same as the compare operands, there's nothing to do.
1066 if (TVal == A || TVal == B || FVal == A || FVal == B)
1070 if (!match(A, m_BitCast(m_Value(C))) || !match(B, m_BitCast(m_Value(D))))
1073 // select (cmp (bitcast C), (bitcast D)), (bitcast TSrc), (bitcast FSrc)
1075 if (!match(TVal, m_BitCast(m_Value(TSrc))) ||
1076 !match(FVal, m_BitCast(m_Value(FSrc))))
1079 // If the select true/false values are *different bitcasts* of the same source
1080 // operands, make the select operands the same as the compare operands and
1081 // cast the result. This is the canonical select form for min/max.
1083 if (TSrc == C && FSrc == D) {
1084 // select (cmp (bitcast C), (bitcast D)), (bitcast' C), (bitcast' D) -->
1085 // bitcast (select (cmp A, B), A, B)
1086 NewSel = Builder.CreateSelect(Cond, A, B, "", &Sel);
1087 } else if (TSrc == D && FSrc == C) {
1088 // select (cmp (bitcast C), (bitcast D)), (bitcast' D), (bitcast' C) -->
1089 // bitcast (select (cmp A, B), B, A)
1090 NewSel = Builder.CreateSelect(Cond, B, A, "", &Sel);
1094 return CastInst::CreateBitOrPointerCast(NewSel, Sel.getType());
1097 Instruction *InstCombiner::visitSelectInst(SelectInst &SI) {
1098 Value *CondVal = SI.getCondition();
1099 Value *TrueVal = SI.getTrueValue();
1100 Value *FalseVal = SI.getFalseValue();
1101 Type *SelType = SI.getType();
1104 SimplifySelectInst(CondVal, TrueVal, FalseVal, DL, &TLI, &DT, &AC))
1105 return replaceInstUsesWith(SI, V);
1107 if (Instruction *I = canonicalizeSelectToShuffle(SI))
1110 if (SelType->getScalarType()->isIntegerTy(1) &&
1111 TrueVal->getType() == CondVal->getType()) {
1112 if (match(TrueVal, m_One())) {
1113 // Change: A = select B, true, C --> A = or B, C
1114 return BinaryOperator::CreateOr(CondVal, FalseVal);
1116 if (match(TrueVal, m_Zero())) {
1117 // Change: A = select B, false, C --> A = and !B, C
1118 Value *NotCond = Builder->CreateNot(CondVal, "not." + CondVal->getName());
1119 return BinaryOperator::CreateAnd(NotCond, FalseVal);
1121 if (match(FalseVal, m_Zero())) {
1122 // Change: A = select B, C, false --> A = and B, C
1123 return BinaryOperator::CreateAnd(CondVal, TrueVal);
1125 if (match(FalseVal, m_One())) {
1126 // Change: A = select B, C, true --> A = or !B, C
1127 Value *NotCond = Builder->CreateNot(CondVal, "not." + CondVal->getName());
1128 return BinaryOperator::CreateOr(NotCond, TrueVal);
1131 // select a, a, b -> a | b
1132 // select a, b, a -> a & b
1133 if (CondVal == TrueVal)
1134 return BinaryOperator::CreateOr(CondVal, FalseVal);
1135 if (CondVal == FalseVal)
1136 return BinaryOperator::CreateAnd(CondVal, TrueVal);
1138 // select a, ~a, b -> (~a) & b
1139 // select a, b, ~a -> (~a) | b
1140 if (match(TrueVal, m_Not(m_Specific(CondVal))))
1141 return BinaryOperator::CreateAnd(TrueVal, FalseVal);
1142 if (match(FalseVal, m_Not(m_Specific(CondVal))))
1143 return BinaryOperator::CreateOr(TrueVal, FalseVal);
1146 // Selecting between two integer or vector splat integer constants?
1148 // Note that we don't handle a scalar select of vectors:
1149 // select i1 %c, <2 x i8> <1, 1>, <2 x i8> <0, 0>
1150 // because that may need 3 instructions to splat the condition value:
1151 // extend, insertelement, shufflevector.
1152 if (CondVal->getType()->isVectorTy() == SelType->isVectorTy()) {
1153 // select C, 1, 0 -> zext C to int
1154 if (match(TrueVal, m_One()) && match(FalseVal, m_Zero()))
1155 return new ZExtInst(CondVal, SelType);
1157 // select C, -1, 0 -> sext C to int
1158 if (match(TrueVal, m_AllOnes()) && match(FalseVal, m_Zero()))
1159 return new SExtInst(CondVal, SelType);
1161 // select C, 0, 1 -> zext !C to int
1162 if (match(TrueVal, m_Zero()) && match(FalseVal, m_One())) {
1163 Value *NotCond = Builder->CreateNot(CondVal, "not." + CondVal->getName());
1164 return new ZExtInst(NotCond, SelType);
1167 // select C, 0, -1 -> sext !C to int
1168 if (match(TrueVal, m_Zero()) && match(FalseVal, m_AllOnes())) {
1169 Value *NotCond = Builder->CreateNot(CondVal, "not." + CondVal->getName());
1170 return new SExtInst(NotCond, SelType);
1174 if (ConstantInt *TrueValC = dyn_cast<ConstantInt>(TrueVal))
1175 if (ConstantInt *FalseValC = dyn_cast<ConstantInt>(FalseVal))
1176 if (Value *V = foldSelectICmpAnd(SI, TrueValC, FalseValC, Builder))
1177 return replaceInstUsesWith(SI, V);
1179 // See if we are selecting two values based on a comparison of the two values.
1180 if (FCmpInst *FCI = dyn_cast<FCmpInst>(CondVal)) {
1181 if (FCI->getOperand(0) == TrueVal && FCI->getOperand(1) == FalseVal) {
1182 // Transform (X == Y) ? X : Y -> Y
1183 if (FCI->getPredicate() == FCmpInst::FCMP_OEQ) {
1184 // This is not safe in general for floating point:
1185 // consider X== -0, Y== +0.
1186 // It becomes safe if either operand is a nonzero constant.
1187 ConstantFP *CFPt, *CFPf;
1188 if (((CFPt = dyn_cast<ConstantFP>(TrueVal)) &&
1189 !CFPt->getValueAPF().isZero()) ||
1190 ((CFPf = dyn_cast<ConstantFP>(FalseVal)) &&
1191 !CFPf->getValueAPF().isZero()))
1192 return replaceInstUsesWith(SI, FalseVal);
1194 // Transform (X une Y) ? X : Y -> X
1195 if (FCI->getPredicate() == FCmpInst::FCMP_UNE) {
1196 // This is not safe in general for floating point:
1197 // consider X== -0, Y== +0.
1198 // It becomes safe if either operand is a nonzero constant.
1199 ConstantFP *CFPt, *CFPf;
1200 if (((CFPt = dyn_cast<ConstantFP>(TrueVal)) &&
1201 !CFPt->getValueAPF().isZero()) ||
1202 ((CFPf = dyn_cast<ConstantFP>(FalseVal)) &&
1203 !CFPf->getValueAPF().isZero()))
1204 return replaceInstUsesWith(SI, TrueVal);
1207 // Canonicalize to use ordered comparisons by swapping the select
1211 // (X ugt Y) ? X : Y -> (X ole Y) ? Y : X
1212 if (FCI->hasOneUse() && FCmpInst::isUnordered(FCI->getPredicate())) {
1213 FCmpInst::Predicate InvPred = FCI->getInversePredicate();
1214 IRBuilder<>::FastMathFlagGuard FMFG(*Builder);
1215 Builder->setFastMathFlags(FCI->getFastMathFlags());
1216 Value *NewCond = Builder->CreateFCmp(InvPred, TrueVal, FalseVal,
1217 FCI->getName() + ".inv");
1219 return SelectInst::Create(NewCond, FalseVal, TrueVal,
1220 SI.getName() + ".p");
1223 // NOTE: if we wanted to, this is where to detect MIN/MAX
1224 } else if (FCI->getOperand(0) == FalseVal && FCI->getOperand(1) == TrueVal){
1225 // Transform (X == Y) ? Y : X -> X
1226 if (FCI->getPredicate() == FCmpInst::FCMP_OEQ) {
1227 // This is not safe in general for floating point:
1228 // consider X== -0, Y== +0.
1229 // It becomes safe if either operand is a nonzero constant.
1230 ConstantFP *CFPt, *CFPf;
1231 if (((CFPt = dyn_cast<ConstantFP>(TrueVal)) &&
1232 !CFPt->getValueAPF().isZero()) ||
1233 ((CFPf = dyn_cast<ConstantFP>(FalseVal)) &&
1234 !CFPf->getValueAPF().isZero()))
1235 return replaceInstUsesWith(SI, FalseVal);
1237 // Transform (X une Y) ? Y : X -> Y
1238 if (FCI->getPredicate() == FCmpInst::FCMP_UNE) {
1239 // This is not safe in general for floating point:
1240 // consider X== -0, Y== +0.
1241 // It becomes safe if either operand is a nonzero constant.
1242 ConstantFP *CFPt, *CFPf;
1243 if (((CFPt = dyn_cast<ConstantFP>(TrueVal)) &&
1244 !CFPt->getValueAPF().isZero()) ||
1245 ((CFPf = dyn_cast<ConstantFP>(FalseVal)) &&
1246 !CFPf->getValueAPF().isZero()))
1247 return replaceInstUsesWith(SI, TrueVal);
1250 // Canonicalize to use ordered comparisons by swapping the select
1254 // (X ugt Y) ? X : Y -> (X ole Y) ? X : Y
1255 if (FCI->hasOneUse() && FCmpInst::isUnordered(FCI->getPredicate())) {
1256 FCmpInst::Predicate InvPred = FCI->getInversePredicate();
1257 IRBuilder<>::FastMathFlagGuard FMFG(*Builder);
1258 Builder->setFastMathFlags(FCI->getFastMathFlags());
1259 Value *NewCond = Builder->CreateFCmp(InvPred, FalseVal, TrueVal,
1260 FCI->getName() + ".inv");
1262 return SelectInst::Create(NewCond, FalseVal, TrueVal,
1263 SI.getName() + ".p");
1266 // NOTE: if we wanted to, this is where to detect MIN/MAX
1268 // NOTE: if we wanted to, this is where to detect ABS
1271 // See if we are selecting two values based on a comparison of the two values.
1272 if (ICmpInst *ICI = dyn_cast<ICmpInst>(CondVal))
1273 if (Instruction *Result = foldSelectInstWithICmp(SI, ICI))
1276 if (Instruction *Add = foldAddSubSelect(SI, *Builder))
1279 // Turn (select C, (op X, Y), (op X, Z)) -> (op X, (select C, Y, Z))
1280 auto *TI = dyn_cast<Instruction>(TrueVal);
1281 auto *FI = dyn_cast<Instruction>(FalseVal);
1282 if (TI && FI && TI->getOpcode() == FI->getOpcode())
1283 if (Instruction *IV = foldSelectOpOp(SI, TI, FI))
1286 if (Instruction *I = foldSelectExtConst(SI))
1289 // See if we can fold the select into one of our operands.
1290 if (SelType->isIntOrIntVectorTy() || SelType->isFPOrFPVectorTy()) {
1291 if (Instruction *FoldI = foldSelectIntoOp(SI, TrueVal, FalseVal))
1294 Value *LHS, *RHS, *LHS2, *RHS2;
1295 Instruction::CastOps CastOp;
1296 SelectPatternResult SPR = matchSelectPattern(&SI, LHS, RHS, &CastOp);
1297 auto SPF = SPR.Flavor;
1299 if (SelectPatternResult::isMinOrMax(SPF)) {
1300 // Canonicalize so that type casts are outside select patterns.
1301 if (LHS->getType()->getPrimitiveSizeInBits() !=
1302 SelType->getPrimitiveSizeInBits()) {
1303 CmpInst::Predicate Pred = getCmpPredicateForMinMax(SPF, SPR.Ordered);
1306 if (CmpInst::isIntPredicate(Pred)) {
1307 Cmp = Builder->CreateICmp(Pred, LHS, RHS);
1309 IRBuilder<>::FastMathFlagGuard FMFG(*Builder);
1310 auto FMF = cast<FPMathOperator>(SI.getCondition())->getFastMathFlags();
1311 Builder->setFastMathFlags(FMF);
1312 Cmp = Builder->CreateFCmp(Pred, LHS, RHS);
1315 Value *NewSI = Builder->CreateCast(
1316 CastOp, Builder->CreateSelect(Cmp, LHS, RHS, SI.getName(), &SI),
1318 return replaceInstUsesWith(SI, NewSI);
1323 // MAX(MAX(a, b), a) -> MAX(a, b)
1324 // MIN(MIN(a, b), a) -> MIN(a, b)
1325 // MAX(MIN(a, b), a) -> a
1326 // MIN(MAX(a, b), a) -> a
1327 // ABS(ABS(a)) -> ABS(a)
1328 // NABS(NABS(a)) -> NABS(a)
1329 if (SelectPatternFlavor SPF2 = matchSelectPattern(LHS, LHS2, RHS2).Flavor)
1330 if (Instruction *R = foldSPFofSPF(cast<Instruction>(LHS),SPF2,LHS2,RHS2,
1333 if (SelectPatternFlavor SPF2 = matchSelectPattern(RHS, LHS2, RHS2).Flavor)
1334 if (Instruction *R = foldSPFofSPF(cast<Instruction>(RHS),SPF2,LHS2,RHS2,
1339 // MAX(~a, ~b) -> ~MIN(a, b)
1340 if ((SPF == SPF_SMAX || SPF == SPF_UMAX) &&
1341 IsFreeToInvert(LHS, LHS->hasNUses(2)) &&
1342 IsFreeToInvert(RHS, RHS->hasNUses(2))) {
1343 // For this transform to be profitable, we need to eliminate at least two
1344 // 'not' instructions if we're going to add one 'not' instruction.
1346 (LHS->hasNUses(2) && match(LHS, m_Not(m_Value()))) +
1347 (RHS->hasNUses(2) && match(RHS, m_Not(m_Value()))) +
1348 (SI.hasOneUse() && match(*SI.user_begin(), m_Not(m_Value())));
1350 if (NumberOfNots >= 2) {
1351 Value *NewLHS = Builder->CreateNot(LHS);
1352 Value *NewRHS = Builder->CreateNot(RHS);
1353 Value *NewCmp = SPF == SPF_SMAX
1354 ? Builder->CreateICmpSLT(NewLHS, NewRHS)
1355 : Builder->CreateICmpULT(NewLHS, NewRHS);
1357 Builder->CreateNot(Builder->CreateSelect(NewCmp, NewLHS, NewRHS));
1358 return replaceInstUsesWith(SI, NewSI);
1363 // ABS(-X) -> ABS(X)
1366 // See if we can fold the select into a phi node if the condition is a select.
1367 if (isa<PHINode>(SI.getCondition()))
1368 // The true/false values have to be live in the PHI predecessor's blocks.
1369 if (canSelectOperandBeMappingIntoPredBlock(TrueVal, SI) &&
1370 canSelectOperandBeMappingIntoPredBlock(FalseVal, SI))
1371 if (Instruction *NV = FoldOpIntoPhi(SI))
1374 if (SelectInst *TrueSI = dyn_cast<SelectInst>(TrueVal)) {
1375 if (TrueSI->getCondition()->getType() == CondVal->getType()) {
1376 // select(C, select(C, a, b), c) -> select(C, a, c)
1377 if (TrueSI->getCondition() == CondVal) {
1378 if (SI.getTrueValue() == TrueSI->getTrueValue())
1380 SI.setOperand(1, TrueSI->getTrueValue());
1383 // select(C0, select(C1, a, b), b) -> select(C0&C1, a, b)
1384 // We choose this as normal form to enable folding on the And and shortening
1385 // paths for the values (this helps GetUnderlyingObjects() for example).
1386 if (TrueSI->getFalseValue() == FalseVal && TrueSI->hasOneUse()) {
1387 Value *And = Builder->CreateAnd(CondVal, TrueSI->getCondition());
1388 SI.setOperand(0, And);
1389 SI.setOperand(1, TrueSI->getTrueValue());
1394 if (SelectInst *FalseSI = dyn_cast<SelectInst>(FalseVal)) {
1395 if (FalseSI->getCondition()->getType() == CondVal->getType()) {
1396 // select(C, a, select(C, b, c)) -> select(C, a, c)
1397 if (FalseSI->getCondition() == CondVal) {
1398 if (SI.getFalseValue() == FalseSI->getFalseValue())
1400 SI.setOperand(2, FalseSI->getFalseValue());
1403 // select(C0, a, select(C1, a, b)) -> select(C0|C1, a, b)
1404 if (FalseSI->getTrueValue() == TrueVal && FalseSI->hasOneUse()) {
1405 Value *Or = Builder->CreateOr(CondVal, FalseSI->getCondition());
1406 SI.setOperand(0, Or);
1407 SI.setOperand(2, FalseSI->getFalseValue());
1413 if (BinaryOperator::isNot(CondVal)) {
1414 SI.setOperand(0, BinaryOperator::getNotArgument(CondVal));
1415 SI.setOperand(1, FalseVal);
1416 SI.setOperand(2, TrueVal);
1420 if (VectorType *VecTy = dyn_cast<VectorType>(SelType)) {
1421 unsigned VWidth = VecTy->getNumElements();
1422 APInt UndefElts(VWidth, 0);
1423 APInt AllOnesEltMask(APInt::getAllOnesValue(VWidth));
1424 if (Value *V = SimplifyDemandedVectorElts(&SI, AllOnesEltMask, UndefElts)) {
1426 return replaceInstUsesWith(SI, V);
1430 if (isa<ConstantAggregateZero>(CondVal)) {
1431 return replaceInstUsesWith(SI, FalseVal);
1435 // See if we can determine the result of this select based on a dominating
1437 BasicBlock *Parent = SI.getParent();
1438 if (BasicBlock *Dom = Parent->getSinglePredecessor()) {
1439 auto *PBI = dyn_cast_or_null<BranchInst>(Dom->getTerminator());
1440 if (PBI && PBI->isConditional() &&
1441 PBI->getSuccessor(0) != PBI->getSuccessor(1) &&
1442 (PBI->getSuccessor(0) == Parent || PBI->getSuccessor(1) == Parent)) {
1443 bool CondIsFalse = PBI->getSuccessor(1) == Parent;
1444 Optional<bool> Implication = isImpliedCondition(
1445 PBI->getCondition(), SI.getCondition(), DL, CondIsFalse);
1447 Value *V = *Implication ? TrueVal : FalseVal;
1448 return replaceInstUsesWith(SI, V);
1453 if (Instruction *BitCastSel = foldSelectCmpBitcasts(SI, *Builder))