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 // Don't break up min/max patterns. The hasOneUse checks below prevent that
124 // for most cases, but vector min/max with bitcasts can be transformed. If the
125 // one-use restrictions are eased for other patterns, we still don't want to
126 // obfuscate min/max.
127 if ((match(&SI, m_SMin(m_Value(), m_Value())) ||
128 match(&SI, m_SMax(m_Value(), m_Value())) ||
129 match(&SI, m_UMin(m_Value(), m_Value())) ||
130 match(&SI, m_UMax(m_Value(), m_Value()))))
133 // If this is a cast from the same type, merge.
134 if (TI->getNumOperands() == 1 && TI->isCast()) {
135 Type *FIOpndTy = FI->getOperand(0)->getType();
136 if (TI->getOperand(0)->getType() != FIOpndTy)
139 // The select condition may be a vector. We may only change the operand
140 // type if the vector width remains the same (and matches the condition).
141 Type *CondTy = SI.getCondition()->getType();
142 if (CondTy->isVectorTy()) {
143 if (!FIOpndTy->isVectorTy())
145 if (CondTy->getVectorNumElements() != FIOpndTy->getVectorNumElements())
148 // TODO: If the backend knew how to deal with casts better, we could
149 // remove this limitation. For now, there's too much potential to create
150 // worse codegen by promoting the select ahead of size-altering casts
153 // Note that ValueTracking's matchSelectPattern() looks through casts
154 // without checking 'hasOneUse' when it matches min/max patterns, so this
155 // transform may end up happening anyway.
156 if (TI->getOpcode() != Instruction::BitCast &&
157 (!TI->hasOneUse() || !FI->hasOneUse()))
160 } else if (!TI->hasOneUse() || !FI->hasOneUse()) {
161 // TODO: The one-use restrictions for a scalar select could be eased if
162 // the fold of a select in visitLoadInst() was enhanced to match a pattern
163 // that includes a cast.
167 // Fold this by inserting a select from the input values.
169 Builder->CreateSelect(SI.getCondition(), TI->getOperand(0),
170 FI->getOperand(0), SI.getName() + ".v", &SI);
171 return CastInst::Create(Instruction::CastOps(TI->getOpcode()), NewSI,
175 // Only handle binary operators with one-use here. As with the cast case
176 // above, it may be possible to relax the one-use constraint, but that needs
177 // be examined carefully since it may not reduce the total number of
179 BinaryOperator *BO = dyn_cast<BinaryOperator>(TI);
180 if (!BO || !TI->hasOneUse() || !FI->hasOneUse())
183 // Figure out if the operations have any operands in common.
184 Value *MatchOp, *OtherOpT, *OtherOpF;
186 if (TI->getOperand(0) == FI->getOperand(0)) {
187 MatchOp = TI->getOperand(0);
188 OtherOpT = TI->getOperand(1);
189 OtherOpF = FI->getOperand(1);
190 MatchIsOpZero = true;
191 } else if (TI->getOperand(1) == FI->getOperand(1)) {
192 MatchOp = TI->getOperand(1);
193 OtherOpT = TI->getOperand(0);
194 OtherOpF = FI->getOperand(0);
195 MatchIsOpZero = false;
196 } else if (!TI->isCommutative()) {
198 } else if (TI->getOperand(0) == FI->getOperand(1)) {
199 MatchOp = TI->getOperand(0);
200 OtherOpT = TI->getOperand(1);
201 OtherOpF = FI->getOperand(0);
202 MatchIsOpZero = true;
203 } else if (TI->getOperand(1) == FI->getOperand(0)) {
204 MatchOp = TI->getOperand(1);
205 OtherOpT = TI->getOperand(0);
206 OtherOpF = FI->getOperand(1);
207 MatchIsOpZero = true;
212 // If we reach here, they do have operations in common.
213 Value *NewSI = Builder->CreateSelect(SI.getCondition(), OtherOpT, OtherOpF,
214 SI.getName() + ".v", &SI);
215 Value *Op0 = MatchIsOpZero ? MatchOp : NewSI;
216 Value *Op1 = MatchIsOpZero ? NewSI : MatchOp;
217 return BinaryOperator::Create(BO->getOpcode(), Op0, Op1);
220 static bool isSelect01(Constant *C1, Constant *C2) {
221 ConstantInt *C1I = dyn_cast<ConstantInt>(C1);
224 ConstantInt *C2I = dyn_cast<ConstantInt>(C2);
227 if (!C1I->isZero() && !C2I->isZero()) // One side must be zero.
229 return C1I->isOne() || C1I->isAllOnesValue() ||
230 C2I->isOne() || C2I->isAllOnesValue();
233 /// Try to fold the select into one of the operands to allow further
235 Instruction *InstCombiner::foldSelectIntoOp(SelectInst &SI, Value *TrueVal,
237 // See the comment above GetSelectFoldableOperands for a description of the
238 // transformation we are doing here.
239 if (Instruction *TVI = dyn_cast<Instruction>(TrueVal)) {
240 if (TVI->hasOneUse() && TVI->getNumOperands() == 2 &&
241 !isa<Constant>(FalseVal)) {
242 if (unsigned SFO = getSelectFoldableOperands(TVI)) {
243 unsigned OpToFold = 0;
244 if ((SFO & 1) && FalseVal == TVI->getOperand(0)) {
246 } else if ((SFO & 2) && FalseVal == TVI->getOperand(1)) {
251 Constant *C = getSelectFoldableConstant(TVI);
252 Value *OOp = TVI->getOperand(2-OpToFold);
253 // Avoid creating select between 2 constants unless it's selecting
254 // between 0, 1 and -1.
255 if (!isa<Constant>(OOp) || isSelect01(C, cast<Constant>(OOp))) {
256 Value *NewSel = Builder->CreateSelect(SI.getCondition(), OOp, C);
257 NewSel->takeName(TVI);
258 BinaryOperator *TVI_BO = cast<BinaryOperator>(TVI);
259 BinaryOperator *BO = BinaryOperator::Create(TVI_BO->getOpcode(),
261 BO->copyIRFlags(TVI_BO);
269 if (Instruction *FVI = dyn_cast<Instruction>(FalseVal)) {
270 if (FVI->hasOneUse() && FVI->getNumOperands() == 2 &&
271 !isa<Constant>(TrueVal)) {
272 if (unsigned SFO = getSelectFoldableOperands(FVI)) {
273 unsigned OpToFold = 0;
274 if ((SFO & 1) && TrueVal == FVI->getOperand(0)) {
276 } else if ((SFO & 2) && TrueVal == FVI->getOperand(1)) {
281 Constant *C = getSelectFoldableConstant(FVI);
282 Value *OOp = FVI->getOperand(2-OpToFold);
283 // Avoid creating select between 2 constants unless it's selecting
284 // between 0, 1 and -1.
285 if (!isa<Constant>(OOp) || isSelect01(C, cast<Constant>(OOp))) {
286 Value *NewSel = Builder->CreateSelect(SI.getCondition(), C, OOp);
287 NewSel->takeName(FVI);
288 BinaryOperator *FVI_BO = cast<BinaryOperator>(FVI);
289 BinaryOperator *BO = BinaryOperator::Create(FVI_BO->getOpcode(),
291 BO->copyIRFlags(FVI_BO);
303 /// (select (icmp eq (and X, C1), 0), Y, (or Y, C2))
305 /// (or (shl (and X, C1), C3), y)
307 /// C1 and C2 are both powers of 2
309 /// C3 = Log(C2) - Log(C1)
311 /// This transform handles cases where:
312 /// 1. The icmp predicate is inverted
313 /// 2. The select operands are reversed
314 /// 3. The magnitude of C2 and C1 are flipped
315 static Value *foldSelectICmpAndOr(const SelectInst &SI, Value *TrueVal,
317 InstCombiner::BuilderTy *Builder) {
318 const ICmpInst *IC = dyn_cast<ICmpInst>(SI.getCondition());
319 if (!IC || !IC->isEquality() || !SI.getType()->isIntegerTy())
322 Value *CmpLHS = IC->getOperand(0);
323 Value *CmpRHS = IC->getOperand(1);
325 if (!match(CmpRHS, m_Zero()))
330 if (!match(CmpLHS, m_And(m_Value(X), m_Power2(C1))))
334 bool OrOnTrueVal = false;
335 bool OrOnFalseVal = match(FalseVal, m_Or(m_Specific(TrueVal), m_Power2(C2)));
337 OrOnTrueVal = match(TrueVal, m_Or(m_Specific(FalseVal), m_Power2(C2)));
339 if (!OrOnFalseVal && !OrOnTrueVal)
343 Value *Y = OrOnFalseVal ? TrueVal : FalseVal;
345 unsigned C1Log = C1->logBase2();
346 unsigned C2Log = C2->logBase2();
348 V = Builder->CreateZExtOrTrunc(V, Y->getType());
349 V = Builder->CreateShl(V, C2Log - C1Log);
350 } else if (C1Log > C2Log) {
351 V = Builder->CreateLShr(V, C1Log - C2Log);
352 V = Builder->CreateZExtOrTrunc(V, Y->getType());
354 V = Builder->CreateZExtOrTrunc(V, Y->getType());
356 ICmpInst::Predicate Pred = IC->getPredicate();
357 if ((Pred == ICmpInst::ICMP_NE && OrOnFalseVal) ||
358 (Pred == ICmpInst::ICMP_EQ && OrOnTrueVal))
359 V = Builder->CreateXor(V, *C2);
361 return Builder->CreateOr(V, Y);
364 /// Attempt to fold a cttz/ctlz followed by a icmp plus select into a single
365 /// call to cttz/ctlz with flag 'is_zero_undef' cleared.
367 /// For example, we can fold the following code sequence:
369 /// %0 = tail call i32 @llvm.cttz.i32(i32 %x, i1 true)
370 /// %1 = icmp ne i32 %x, 0
371 /// %2 = select i1 %1, i32 %0, i32 32
375 /// %0 = tail call i32 @llvm.cttz.i32(i32 %x, i1 false)
376 static Value *foldSelectCttzCtlz(ICmpInst *ICI, Value *TrueVal, Value *FalseVal,
377 InstCombiner::BuilderTy *Builder) {
378 ICmpInst::Predicate Pred = ICI->getPredicate();
379 Value *CmpLHS = ICI->getOperand(0);
380 Value *CmpRHS = ICI->getOperand(1);
382 // Check if the condition value compares a value for equality against zero.
383 if (!ICI->isEquality() || !match(CmpRHS, m_Zero()))
386 Value *Count = FalseVal;
387 Value *ValueOnZero = TrueVal;
388 if (Pred == ICmpInst::ICMP_NE)
389 std::swap(Count, ValueOnZero);
391 // Skip zero extend/truncate.
393 if (match(Count, m_ZExt(m_Value(V))) ||
394 match(Count, m_Trunc(m_Value(V))))
397 // Check if the value propagated on zero is a constant number equal to the
398 // sizeof in bits of 'Count'.
399 unsigned SizeOfInBits = Count->getType()->getScalarSizeInBits();
400 if (!match(ValueOnZero, m_SpecificInt(SizeOfInBits)))
403 // Check that 'Count' is a call to intrinsic cttz/ctlz. Also check that the
404 // input to the cttz/ctlz is used as LHS for the compare instruction.
405 if (match(Count, m_Intrinsic<Intrinsic::cttz>(m_Specific(CmpLHS))) ||
406 match(Count, m_Intrinsic<Intrinsic::ctlz>(m_Specific(CmpLHS)))) {
407 IntrinsicInst *II = cast<IntrinsicInst>(Count);
408 // Explicitly clear the 'undef_on_zero' flag.
409 IntrinsicInst *NewI = cast<IntrinsicInst>(II->clone());
410 Type *Ty = NewI->getArgOperand(1)->getType();
411 NewI->setArgOperand(1, Constant::getNullValue(Ty));
412 Builder->Insert(NewI);
413 return Builder->CreateZExtOrTrunc(NewI, ValueOnZero->getType());
419 /// Return true if we find and adjust an icmp+select pattern where the compare
420 /// is with a constant that can be incremented or decremented to match the
421 /// minimum or maximum idiom.
422 static bool adjustMinMax(SelectInst &Sel, ICmpInst &Cmp) {
423 ICmpInst::Predicate Pred = Cmp.getPredicate();
424 Value *CmpLHS = Cmp.getOperand(0);
425 Value *CmpRHS = Cmp.getOperand(1);
426 Value *TrueVal = Sel.getTrueValue();
427 Value *FalseVal = Sel.getFalseValue();
429 // We may move or edit the compare, so make sure the select is the only user.
431 if (!Cmp.hasOneUse() || !match(CmpRHS, m_APInt(CmpC)))
434 // These transforms only work for selects of integers or vector selects of
436 Type *SelTy = Sel.getType();
437 auto *SelEltTy = dyn_cast<IntegerType>(SelTy->getScalarType());
438 if (!SelEltTy || SelTy->isVectorTy() != Cmp.getType()->isVectorTy())
441 Constant *AdjustedRHS;
442 if (Pred == ICmpInst::ICMP_UGT || Pred == ICmpInst::ICMP_SGT)
443 AdjustedRHS = ConstantInt::get(CmpRHS->getType(), *CmpC + 1);
444 else if (Pred == ICmpInst::ICMP_ULT || Pred == ICmpInst::ICMP_SLT)
445 AdjustedRHS = ConstantInt::get(CmpRHS->getType(), *CmpC - 1);
449 // X > C ? X : C+1 --> X < C+1 ? C+1 : X
450 // X < C ? X : C-1 --> X > C-1 ? C-1 : X
451 if ((CmpLHS == TrueVal && AdjustedRHS == FalseVal) ||
452 (CmpLHS == FalseVal && AdjustedRHS == TrueVal)) {
453 ; // Nothing to do here. Values match without any sign/zero extension.
455 // Types do not match. Instead of calculating this with mixed types, promote
456 // all to the larger type. This enables scalar evolution to analyze this
458 else if (CmpRHS->getType()->getScalarSizeInBits() < SelEltTy->getBitWidth()) {
459 Constant *SextRHS = ConstantExpr::getSExt(AdjustedRHS, SelTy);
461 // X = sext x; x >s c ? X : C+1 --> X = sext x; X <s C+1 ? C+1 : X
462 // X = sext x; x <s c ? X : C-1 --> X = sext x; X >s C-1 ? C-1 : X
463 // X = sext x; x >u c ? X : C+1 --> X = sext x; X <u C+1 ? C+1 : X
464 // X = sext x; x <u c ? X : C-1 --> X = sext x; X >u C-1 ? C-1 : X
465 if (match(TrueVal, m_SExt(m_Specific(CmpLHS))) && SextRHS == FalseVal) {
467 AdjustedRHS = SextRHS;
468 } else if (match(FalseVal, m_SExt(m_Specific(CmpLHS))) &&
469 SextRHS == TrueVal) {
471 AdjustedRHS = SextRHS;
472 } else if (Cmp.isUnsigned()) {
473 Constant *ZextRHS = ConstantExpr::getZExt(AdjustedRHS, SelTy);
474 // X = zext x; x >u c ? X : C+1 --> X = zext x; X <u C+1 ? C+1 : X
475 // X = zext x; x <u c ? X : C-1 --> X = zext x; X >u C-1 ? C-1 : X
476 // zext + signed compare cannot be changed:
477 // 0xff <s 0x00, but 0x00ff >s 0x0000
478 if (match(TrueVal, m_ZExt(m_Specific(CmpLHS))) && ZextRHS == FalseVal) {
480 AdjustedRHS = ZextRHS;
481 } else if (match(FalseVal, m_ZExt(m_Specific(CmpLHS))) &&
482 ZextRHS == TrueVal) {
484 AdjustedRHS = ZextRHS;
495 Pred = ICmpInst::getSwappedPredicate(Pred);
496 CmpRHS = AdjustedRHS;
497 std::swap(FalseVal, TrueVal);
498 Cmp.setPredicate(Pred);
499 Cmp.setOperand(0, CmpLHS);
500 Cmp.setOperand(1, CmpRHS);
501 Sel.setOperand(1, TrueVal);
502 Sel.setOperand(2, FalseVal);
503 Sel.swapProfMetadata();
505 // Move the compare instruction right before the select instruction. Otherwise
506 // the sext/zext value may be defined after the compare instruction uses it.
507 Cmp.moveBefore(&Sel);
512 /// If this is an integer min/max (icmp + select) with a constant operand,
513 /// create the canonical icmp for the min/max operation and canonicalize the
514 /// constant to the 'false' operand of the select:
515 /// select (icmp Pred X, C1), C2, X --> select (icmp Pred' X, C2), X, C2
516 /// Note: if C1 != C2, this will change the icmp constant to the existing
517 /// constant operand of the select.
519 canonicalizeMinMaxWithConstant(SelectInst &Sel, ICmpInst &Cmp,
520 InstCombiner::BuilderTy &Builder) {
521 if (!Cmp.hasOneUse() || !isa<Constant>(Cmp.getOperand(1)))
524 // Canonicalize the compare predicate based on whether we have min or max.
526 ICmpInst::Predicate NewPred;
527 SelectPatternResult SPR = matchSelectPattern(&Sel, LHS, RHS);
528 switch (SPR.Flavor) {
529 case SPF_SMIN: NewPred = ICmpInst::ICMP_SLT; break;
530 case SPF_UMIN: NewPred = ICmpInst::ICMP_ULT; break;
531 case SPF_SMAX: NewPred = ICmpInst::ICMP_SGT; break;
532 case SPF_UMAX: NewPred = ICmpInst::ICMP_UGT; break;
533 default: return nullptr;
536 // Is this already canonical?
537 if (Cmp.getOperand(0) == LHS && Cmp.getOperand(1) == RHS &&
538 Cmp.getPredicate() == NewPred)
541 // Create the canonical compare and plug it into the select.
542 Sel.setCondition(Builder.CreateICmp(NewPred, LHS, RHS));
544 // If the select operands did not change, we're done.
545 if (Sel.getTrueValue() == LHS && Sel.getFalseValue() == RHS)
548 // If we are swapping the select operands, swap the metadata too.
549 assert(Sel.getTrueValue() == RHS && Sel.getFalseValue() == LHS &&
550 "Unexpected results from matchSelectPattern");
551 Sel.setTrueValue(LHS);
552 Sel.setFalseValue(RHS);
553 Sel.swapProfMetadata();
557 /// Visit a SelectInst that has an ICmpInst as its first operand.
558 Instruction *InstCombiner::foldSelectInstWithICmp(SelectInst &SI,
560 if (Instruction *NewSel = canonicalizeMinMaxWithConstant(SI, *ICI, *Builder))
563 bool Changed = adjustMinMax(SI, *ICI);
565 ICmpInst::Predicate Pred = ICI->getPredicate();
566 Value *CmpLHS = ICI->getOperand(0);
567 Value *CmpRHS = ICI->getOperand(1);
568 Value *TrueVal = SI.getTrueValue();
569 Value *FalseVal = SI.getFalseValue();
571 // Transform (X >s -1) ? C1 : C2 --> ((X >>s 31) & (C2 - C1)) + C1
572 // and (X <s 0) ? C2 : C1 --> ((X >>s 31) & (C2 - C1)) + C1
573 // FIXME: Type and constness constraints could be lifted, but we have to
574 // watch code size carefully. We should consider xor instead of
575 // sub/add when we decide to do that.
576 if (IntegerType *Ty = dyn_cast<IntegerType>(CmpLHS->getType())) {
577 if (TrueVal->getType() == Ty) {
578 if (ConstantInt *Cmp = dyn_cast<ConstantInt>(CmpRHS)) {
579 ConstantInt *C1 = nullptr, *C2 = nullptr;
580 if (Pred == ICmpInst::ICMP_SGT && Cmp->isAllOnesValue()) {
581 C1 = dyn_cast<ConstantInt>(TrueVal);
582 C2 = dyn_cast<ConstantInt>(FalseVal);
583 } else if (Pred == ICmpInst::ICMP_SLT && Cmp->isNullValue()) {
584 C1 = dyn_cast<ConstantInt>(FalseVal);
585 C2 = dyn_cast<ConstantInt>(TrueVal);
588 // This shift results in either -1 or 0.
589 Value *AShr = Builder->CreateAShr(CmpLHS, Ty->getBitWidth()-1);
591 // Check if we can express the operation with a single or.
592 if (C2->isAllOnesValue())
593 return replaceInstUsesWith(SI, Builder->CreateOr(AShr, C1));
595 Value *And = Builder->CreateAnd(AShr, C2->getValue()-C1->getValue());
596 return replaceInstUsesWith(SI, Builder->CreateAdd(And, C1));
602 // NOTE: if we wanted to, this is where to detect integer MIN/MAX
604 if (CmpRHS != CmpLHS && isa<Constant>(CmpRHS)) {
605 if (CmpLHS == TrueVal && Pred == ICmpInst::ICMP_EQ) {
606 // Transform (X == C) ? X : Y -> (X == C) ? C : Y
607 SI.setOperand(1, CmpRHS);
609 } else if (CmpLHS == FalseVal && Pred == ICmpInst::ICMP_NE) {
610 // Transform (X != C) ? Y : X -> (X != C) ? Y : C
611 SI.setOperand(2, CmpRHS);
616 // FIXME: This code is nearly duplicated in InstSimplify. Using/refactoring
617 // decomposeBitTestICmp() might help.
620 DL.getTypeSizeInBits(TrueVal->getType()->getScalarType());
621 APInt MinSignedValue = APInt::getSignBit(BitWidth);
625 bool IsBitTest = false;
626 if (ICmpInst::isEquality(Pred) &&
627 match(CmpLHS, m_And(m_Value(X), m_Power2(Y))) &&
628 match(CmpRHS, m_Zero())) {
630 TrueWhenUnset = Pred == ICmpInst::ICMP_EQ;
631 } else if (Pred == ICmpInst::ICMP_SLT && match(CmpRHS, m_Zero())) {
635 TrueWhenUnset = false;
636 } else if (Pred == ICmpInst::ICMP_SGT && match(CmpRHS, m_AllOnes())) {
640 TrueWhenUnset = true;
644 // (X & Y) == 0 ? X : X ^ Y --> X & ~Y
645 if (TrueWhenUnset && TrueVal == X &&
646 match(FalseVal, m_Xor(m_Specific(X), m_APInt(C))) && *Y == *C)
647 V = Builder->CreateAnd(X, ~(*Y));
648 // (X & Y) != 0 ? X ^ Y : X --> X & ~Y
649 else if (!TrueWhenUnset && FalseVal == X &&
650 match(TrueVal, m_Xor(m_Specific(X), m_APInt(C))) && *Y == *C)
651 V = Builder->CreateAnd(X, ~(*Y));
652 // (X & Y) == 0 ? X ^ Y : X --> X | Y
653 else if (TrueWhenUnset && FalseVal == X &&
654 match(TrueVal, m_Xor(m_Specific(X), m_APInt(C))) && *Y == *C)
655 V = Builder->CreateOr(X, *Y);
656 // (X & Y) != 0 ? X : X ^ Y --> X | Y
657 else if (!TrueWhenUnset && TrueVal == X &&
658 match(FalseVal, m_Xor(m_Specific(X), m_APInt(C))) && *Y == *C)
659 V = Builder->CreateOr(X, *Y);
662 return replaceInstUsesWith(SI, V);
666 if (Value *V = foldSelectICmpAndOr(SI, TrueVal, FalseVal, Builder))
667 return replaceInstUsesWith(SI, V);
669 if (Value *V = foldSelectCttzCtlz(ICI, TrueVal, FalseVal, Builder))
670 return replaceInstUsesWith(SI, V);
672 return Changed ? &SI : nullptr;
676 /// SI is a select whose condition is a PHI node (but the two may be in
677 /// different blocks). See if the true/false values (V) are live in all of the
678 /// predecessor blocks of the PHI. For example, cases like this can't be mapped:
680 /// X = phi [ C1, BB1], [C2, BB2]
682 /// Z = select X, Y, 0
684 /// because Y is not live in BB1/BB2.
686 static bool canSelectOperandBeMappingIntoPredBlock(const Value *V,
687 const SelectInst &SI) {
688 // If the value is a non-instruction value like a constant or argument, it
689 // can always be mapped.
690 const Instruction *I = dyn_cast<Instruction>(V);
693 // If V is a PHI node defined in the same block as the condition PHI, we can
694 // map the arguments.
695 const PHINode *CondPHI = cast<PHINode>(SI.getCondition());
697 if (const PHINode *VP = dyn_cast<PHINode>(I))
698 if (VP->getParent() == CondPHI->getParent())
701 // Otherwise, if the PHI and select are defined in the same block and if V is
702 // defined in a different block, then we can transform it.
703 if (SI.getParent() == CondPHI->getParent() &&
704 I->getParent() != CondPHI->getParent())
707 // Otherwise we have a 'hard' case and we can't tell without doing more
708 // detailed dominator based analysis, punt.
712 /// We have an SPF (e.g. a min or max) of an SPF of the form:
713 /// SPF2(SPF1(A, B), C)
714 Instruction *InstCombiner::foldSPFofSPF(Instruction *Inner,
715 SelectPatternFlavor SPF1,
718 SelectPatternFlavor SPF2, Value *C) {
719 if (Outer.getType() != Inner->getType())
722 if (C == A || C == B) {
723 // MAX(MAX(A, B), B) -> MAX(A, B)
724 // MIN(MIN(a, b), a) -> MIN(a, b)
726 return replaceInstUsesWith(Outer, Inner);
728 // MAX(MIN(a, b), a) -> a
729 // MIN(MAX(a, b), a) -> a
730 if ((SPF1 == SPF_SMIN && SPF2 == SPF_SMAX) ||
731 (SPF1 == SPF_SMAX && SPF2 == SPF_SMIN) ||
732 (SPF1 == SPF_UMIN && SPF2 == SPF_UMAX) ||
733 (SPF1 == SPF_UMAX && SPF2 == SPF_UMIN))
734 return replaceInstUsesWith(Outer, C);
738 const APInt *CB, *CC;
739 if (match(B, m_APInt(CB)) && match(C, m_APInt(CC))) {
740 // MIN(MIN(A, 23), 97) -> MIN(A, 23)
741 // MAX(MAX(A, 97), 23) -> MAX(A, 97)
742 if ((SPF1 == SPF_UMIN && CB->ule(*CC)) ||
743 (SPF1 == SPF_SMIN && CB->sle(*CC)) ||
744 (SPF1 == SPF_UMAX && CB->uge(*CC)) ||
745 (SPF1 == SPF_SMAX && CB->sge(*CC)))
746 return replaceInstUsesWith(Outer, Inner);
748 // MIN(MIN(A, 97), 23) -> MIN(A, 23)
749 // MAX(MAX(A, 23), 97) -> MAX(A, 97)
750 if ((SPF1 == SPF_UMIN && CB->ugt(*CC)) ||
751 (SPF1 == SPF_SMIN && CB->sgt(*CC)) ||
752 (SPF1 == SPF_UMAX && CB->ult(*CC)) ||
753 (SPF1 == SPF_SMAX && CB->slt(*CC))) {
754 Outer.replaceUsesOfWith(Inner, A);
760 // ABS(ABS(X)) -> ABS(X)
761 // NABS(NABS(X)) -> NABS(X)
762 if (SPF1 == SPF2 && (SPF1 == SPF_ABS || SPF1 == SPF_NABS)) {
763 return replaceInstUsesWith(Outer, Inner);
766 // ABS(NABS(X)) -> ABS(X)
767 // NABS(ABS(X)) -> NABS(X)
768 if ((SPF1 == SPF_ABS && SPF2 == SPF_NABS) ||
769 (SPF1 == SPF_NABS && SPF2 == SPF_ABS)) {
770 SelectInst *SI = cast<SelectInst>(Inner);
772 Builder->CreateSelect(SI->getCondition(), SI->getFalseValue(),
773 SI->getTrueValue(), SI->getName(), SI);
774 return replaceInstUsesWith(Outer, NewSI);
777 auto IsFreeOrProfitableToInvert =
778 [&](Value *V, Value *&NotV, bool &ElidesXor) {
779 if (match(V, m_Not(m_Value(NotV)))) {
780 // If V has at most 2 uses then we can get rid of the xor operation
782 ElidesXor |= !V->hasNUsesOrMore(3);
786 if (IsFreeToInvert(V, !V->hasNUsesOrMore(3))) {
794 Value *NotA, *NotB, *NotC;
795 bool ElidesXor = false;
797 // MIN(MIN(~A, ~B), ~C) == ~MAX(MAX(A, B), C)
798 // MIN(MAX(~A, ~B), ~C) == ~MAX(MIN(A, B), C)
799 // MAX(MIN(~A, ~B), ~C) == ~MIN(MAX(A, B), C)
800 // MAX(MAX(~A, ~B), ~C) == ~MIN(MIN(A, B), C)
802 // This transform is performance neutral if we can elide at least one xor from
803 // the set of three operands, since we'll be tacking on an xor at the very
805 if (SelectPatternResult::isMinOrMax(SPF1) &&
806 SelectPatternResult::isMinOrMax(SPF2) &&
807 IsFreeOrProfitableToInvert(A, NotA, ElidesXor) &&
808 IsFreeOrProfitableToInvert(B, NotB, ElidesXor) &&
809 IsFreeOrProfitableToInvert(C, NotC, ElidesXor) && ElidesXor) {
811 NotA = Builder->CreateNot(A);
813 NotB = Builder->CreateNot(B);
815 NotC = Builder->CreateNot(C);
817 Value *NewInner = generateMinMaxSelectPattern(
818 Builder, getInverseMinMaxSelectPattern(SPF1), NotA, NotB);
819 Value *NewOuter = Builder->CreateNot(generateMinMaxSelectPattern(
820 Builder, getInverseMinMaxSelectPattern(SPF2), NewInner, NotC));
821 return replaceInstUsesWith(Outer, NewOuter);
827 /// If one of the constants is zero (we know they can't both be) and we have an
828 /// icmp instruction with zero, and we have an 'and' with the non-constant value
829 /// and a power of two we can turn the select into a shift on the result of the
831 static Value *foldSelectICmpAnd(const SelectInst &SI, ConstantInt *TrueVal,
832 ConstantInt *FalseVal,
833 InstCombiner::BuilderTy *Builder) {
834 const ICmpInst *IC = dyn_cast<ICmpInst>(SI.getCondition());
835 if (!IC || !IC->isEquality() || !SI.getType()->isIntegerTy())
838 if (!match(IC->getOperand(1), m_Zero()))
842 Value *LHS = IC->getOperand(0);
843 if (!match(LHS, m_And(m_Value(), m_ConstantInt(AndRHS))))
846 // If both select arms are non-zero see if we have a select of the form
847 // 'x ? 2^n + C : C'. Then we can offset both arms by C, use the logic
848 // for 'x ? 2^n : 0' and fix the thing up at the end.
849 ConstantInt *Offset = nullptr;
850 if (!TrueVal->isZero() && !FalseVal->isZero()) {
851 if ((TrueVal->getValue() - FalseVal->getValue()).isPowerOf2())
853 else if ((FalseVal->getValue() - TrueVal->getValue()).isPowerOf2())
858 // Adjust TrueVal and FalseVal to the offset.
859 TrueVal = ConstantInt::get(Builder->getContext(),
860 TrueVal->getValue() - Offset->getValue());
861 FalseVal = ConstantInt::get(Builder->getContext(),
862 FalseVal->getValue() - Offset->getValue());
865 // Make sure the mask in the 'and' and one of the select arms is a power of 2.
866 if (!AndRHS->getValue().isPowerOf2() ||
867 (!TrueVal->getValue().isPowerOf2() &&
868 !FalseVal->getValue().isPowerOf2()))
871 // Determine which shift is needed to transform result of the 'and' into the
873 ConstantInt *ValC = !TrueVal->isZero() ? TrueVal : FalseVal;
874 unsigned ValZeros = ValC->getValue().logBase2();
875 unsigned AndZeros = AndRHS->getValue().logBase2();
877 // If types don't match we can still convert the select by introducing a zext
878 // or a trunc of the 'and'. The trunc case requires that all of the truncated
879 // bits are zero, we can figure that out by looking at the 'and' mask.
880 if (AndZeros >= ValC->getBitWidth())
883 Value *V = Builder->CreateZExtOrTrunc(LHS, SI.getType());
884 if (ValZeros > AndZeros)
885 V = Builder->CreateShl(V, ValZeros - AndZeros);
886 else if (ValZeros < AndZeros)
887 V = Builder->CreateLShr(V, AndZeros - ValZeros);
889 // Okay, now we know that everything is set up, we just don't know whether we
890 // have a icmp_ne or icmp_eq and whether the true or false val is the zero.
891 bool ShouldNotVal = !TrueVal->isZero();
892 ShouldNotVal ^= IC->getPredicate() == ICmpInst::ICMP_NE;
894 V = Builder->CreateXor(V, ValC);
896 // Apply an offset if needed.
898 V = Builder->CreateAdd(V, Offset);
902 /// Turn select C, (X + Y), (X - Y) --> (X + (select C, Y, (-Y))).
903 /// This is even legal for FP.
904 static Instruction *foldAddSubSelect(SelectInst &SI,
905 InstCombiner::BuilderTy &Builder) {
906 Value *CondVal = SI.getCondition();
907 Value *TrueVal = SI.getTrueValue();
908 Value *FalseVal = SI.getFalseValue();
909 auto *TI = dyn_cast<Instruction>(TrueVal);
910 auto *FI = dyn_cast<Instruction>(FalseVal);
911 if (!TI || !FI || !TI->hasOneUse() || !FI->hasOneUse())
914 Instruction *AddOp = nullptr, *SubOp = nullptr;
915 if ((TI->getOpcode() == Instruction::Sub &&
916 FI->getOpcode() == Instruction::Add) ||
917 (TI->getOpcode() == Instruction::FSub &&
918 FI->getOpcode() == Instruction::FAdd)) {
921 } else if ((FI->getOpcode() == Instruction::Sub &&
922 TI->getOpcode() == Instruction::Add) ||
923 (FI->getOpcode() == Instruction::FSub &&
924 TI->getOpcode() == Instruction::FAdd)) {
930 Value *OtherAddOp = nullptr;
931 if (SubOp->getOperand(0) == AddOp->getOperand(0)) {
932 OtherAddOp = AddOp->getOperand(1);
933 } else if (SubOp->getOperand(0) == AddOp->getOperand(1)) {
934 OtherAddOp = AddOp->getOperand(0);
938 // So at this point we know we have (Y -> OtherAddOp):
939 // select C, (add X, Y), (sub X, Z)
940 Value *NegVal; // Compute -Z
941 if (SI.getType()->isFPOrFPVectorTy()) {
942 NegVal = Builder.CreateFNeg(SubOp->getOperand(1));
943 if (Instruction *NegInst = dyn_cast<Instruction>(NegVal)) {
944 FastMathFlags Flags = AddOp->getFastMathFlags();
945 Flags &= SubOp->getFastMathFlags();
946 NegInst->setFastMathFlags(Flags);
949 NegVal = Builder.CreateNeg(SubOp->getOperand(1));
952 Value *NewTrueOp = OtherAddOp;
953 Value *NewFalseOp = NegVal;
955 std::swap(NewTrueOp, NewFalseOp);
956 Value *NewSel = Builder.CreateSelect(CondVal, NewTrueOp, NewFalseOp,
957 SI.getName() + ".p", &SI);
959 if (SI.getType()->isFPOrFPVectorTy()) {
961 BinaryOperator::CreateFAdd(SubOp->getOperand(0), NewSel);
963 FastMathFlags Flags = AddOp->getFastMathFlags();
964 Flags &= SubOp->getFastMathFlags();
965 RI->setFastMathFlags(Flags);
968 return BinaryOperator::CreateAdd(SubOp->getOperand(0), NewSel);
974 Instruction *InstCombiner::foldSelectExtConst(SelectInst &Sel) {
975 Instruction *ExtInst;
976 if (!match(Sel.getTrueValue(), m_Instruction(ExtInst)) &&
977 !match(Sel.getFalseValue(), m_Instruction(ExtInst)))
980 auto ExtOpcode = ExtInst->getOpcode();
981 if (ExtOpcode != Instruction::ZExt && ExtOpcode != Instruction::SExt)
984 // TODO: Handle larger types? That requires adjusting FoldOpIntoSelect too.
985 Value *X = ExtInst->getOperand(0);
986 Type *SmallType = X->getType();
987 if (!SmallType->getScalarType()->isIntegerTy(1))
991 if (!match(Sel.getTrueValue(), m_Constant(C)) &&
992 !match(Sel.getFalseValue(), m_Constant(C)))
995 // If the constant is the same after truncation to the smaller type and
996 // extension to the original type, we can narrow the select.
997 Value *Cond = Sel.getCondition();
998 Type *SelType = Sel.getType();
999 Constant *TruncC = ConstantExpr::getTrunc(C, SmallType);
1000 Constant *ExtC = ConstantExpr::getCast(ExtOpcode, TruncC, SelType);
1002 Value *TruncCVal = cast<Value>(TruncC);
1003 if (ExtInst == Sel.getFalseValue())
1004 std::swap(X, TruncCVal);
1006 // select Cond, (ext X), C --> ext(select Cond, X, C')
1007 // select Cond, C, (ext X) --> ext(select Cond, C', X)
1008 Value *NewSel = Builder->CreateSelect(Cond, X, TruncCVal, "narrow", &Sel);
1009 return CastInst::Create(Instruction::CastOps(ExtOpcode), NewSel, SelType);
1012 // If one arm of the select is the extend of the condition, replace that arm
1013 // with the extension of the appropriate known bool value.
1015 if (ExtInst == Sel.getTrueValue()) {
1016 // select X, (sext X), C --> select X, -1, C
1017 // select X, (zext X), C --> select X, 1, C
1018 Constant *One = ConstantInt::getTrue(SmallType);
1019 Constant *AllOnesOrOne = ConstantExpr::getCast(ExtOpcode, One, SelType);
1020 return SelectInst::Create(Cond, AllOnesOrOne, C, "", nullptr, &Sel);
1022 // select X, C, (sext X) --> select X, C, 0
1023 // select X, C, (zext X) --> select X, C, 0
1024 Constant *Zero = ConstantInt::getNullValue(SelType);
1025 return SelectInst::Create(Cond, C, Zero, "", nullptr, &Sel);
1032 /// Try to transform a vector select with a constant condition vector into a
1033 /// shuffle for easier combining with other shuffles and insert/extract.
1034 static Instruction *canonicalizeSelectToShuffle(SelectInst &SI) {
1035 Value *CondVal = SI.getCondition();
1037 if (!CondVal->getType()->isVectorTy() || !match(CondVal, m_Constant(CondC)))
1040 unsigned NumElts = CondVal->getType()->getVectorNumElements();
1041 SmallVector<Constant *, 16> Mask;
1042 Mask.reserve(NumElts);
1043 Type *Int32Ty = Type::getInt32Ty(CondVal->getContext());
1044 for (unsigned i = 0; i != NumElts; ++i) {
1045 Constant *Elt = CondC->getAggregateElement(i);
1049 if (Elt->isOneValue()) {
1050 // If the select condition element is true, choose from the 1st vector.
1051 Mask.push_back(ConstantInt::get(Int32Ty, i));
1052 } else if (Elt->isNullValue()) {
1053 // If the select condition element is false, choose from the 2nd vector.
1054 Mask.push_back(ConstantInt::get(Int32Ty, i + NumElts));
1055 } else if (isa<UndefValue>(Elt)) {
1056 // Undef in a select condition (choose one of the operands) does not mean
1057 // the same thing as undef in a shuffle mask (any value is acceptable), so
1061 // Bail out on a constant expression.
1066 return new ShuffleVectorInst(SI.getTrueValue(), SI.getFalseValue(),
1067 ConstantVector::get(Mask));
1070 /// Reuse bitcasted operands between a compare and select:
1071 /// select (cmp (bitcast C), (bitcast D)), (bitcast' C), (bitcast' D) -->
1072 /// bitcast (select (cmp (bitcast C), (bitcast D)), (bitcast C), (bitcast D))
1073 static Instruction *foldSelectCmpBitcasts(SelectInst &Sel,
1074 InstCombiner::BuilderTy &Builder) {
1075 Value *Cond = Sel.getCondition();
1076 Value *TVal = Sel.getTrueValue();
1077 Value *FVal = Sel.getFalseValue();
1079 CmpInst::Predicate Pred;
1081 if (!match(Cond, m_Cmp(Pred, m_Value(A), m_Value(B))))
1084 // The select condition is a compare instruction. If the select's true/false
1085 // values are already the same as the compare operands, there's nothing to do.
1086 if (TVal == A || TVal == B || FVal == A || FVal == B)
1090 if (!match(A, m_BitCast(m_Value(C))) || !match(B, m_BitCast(m_Value(D))))
1093 // select (cmp (bitcast C), (bitcast D)), (bitcast TSrc), (bitcast FSrc)
1095 if (!match(TVal, m_BitCast(m_Value(TSrc))) ||
1096 !match(FVal, m_BitCast(m_Value(FSrc))))
1099 // If the select true/false values are *different bitcasts* of the same source
1100 // operands, make the select operands the same as the compare operands and
1101 // cast the result. This is the canonical select form for min/max.
1103 if (TSrc == C && FSrc == D) {
1104 // select (cmp (bitcast C), (bitcast D)), (bitcast' C), (bitcast' D) -->
1105 // bitcast (select (cmp A, B), A, B)
1106 NewSel = Builder.CreateSelect(Cond, A, B, "", &Sel);
1107 } else if (TSrc == D && FSrc == C) {
1108 // select (cmp (bitcast C), (bitcast D)), (bitcast' D), (bitcast' C) -->
1109 // bitcast (select (cmp A, B), B, A)
1110 NewSel = Builder.CreateSelect(Cond, B, A, "", &Sel);
1114 return CastInst::CreateBitOrPointerCast(NewSel, Sel.getType());
1117 Instruction *InstCombiner::visitSelectInst(SelectInst &SI) {
1118 Value *CondVal = SI.getCondition();
1119 Value *TrueVal = SI.getTrueValue();
1120 Value *FalseVal = SI.getFalseValue();
1121 Type *SelType = SI.getType();
1124 SimplifySelectInst(CondVal, TrueVal, FalseVal, DL, &TLI, &DT, &AC))
1125 return replaceInstUsesWith(SI, V);
1127 if (Instruction *I = canonicalizeSelectToShuffle(SI))
1130 if (SelType->getScalarType()->isIntegerTy(1) &&
1131 TrueVal->getType() == CondVal->getType()) {
1132 if (match(TrueVal, m_One())) {
1133 // Change: A = select B, true, C --> A = or B, C
1134 return BinaryOperator::CreateOr(CondVal, FalseVal);
1136 if (match(TrueVal, m_Zero())) {
1137 // Change: A = select B, false, C --> A = and !B, C
1138 Value *NotCond = Builder->CreateNot(CondVal, "not." + CondVal->getName());
1139 return BinaryOperator::CreateAnd(NotCond, FalseVal);
1141 if (match(FalseVal, m_Zero())) {
1142 // Change: A = select B, C, false --> A = and B, C
1143 return BinaryOperator::CreateAnd(CondVal, TrueVal);
1145 if (match(FalseVal, m_One())) {
1146 // Change: A = select B, C, true --> A = or !B, C
1147 Value *NotCond = Builder->CreateNot(CondVal, "not." + CondVal->getName());
1148 return BinaryOperator::CreateOr(NotCond, TrueVal);
1151 // select a, a, b -> a | b
1152 // select a, b, a -> a & b
1153 if (CondVal == TrueVal)
1154 return BinaryOperator::CreateOr(CondVal, FalseVal);
1155 if (CondVal == FalseVal)
1156 return BinaryOperator::CreateAnd(CondVal, TrueVal);
1158 // select a, ~a, b -> (~a) & b
1159 // select a, b, ~a -> (~a) | b
1160 if (match(TrueVal, m_Not(m_Specific(CondVal))))
1161 return BinaryOperator::CreateAnd(TrueVal, FalseVal);
1162 if (match(FalseVal, m_Not(m_Specific(CondVal))))
1163 return BinaryOperator::CreateOr(TrueVal, FalseVal);
1166 // Selecting between two integer or vector splat integer constants?
1168 // Note that we don't handle a scalar select of vectors:
1169 // select i1 %c, <2 x i8> <1, 1>, <2 x i8> <0, 0>
1170 // because that may need 3 instructions to splat the condition value:
1171 // extend, insertelement, shufflevector.
1172 if (CondVal->getType()->isVectorTy() == SelType->isVectorTy()) {
1173 // select C, 1, 0 -> zext C to int
1174 if (match(TrueVal, m_One()) && match(FalseVal, m_Zero()))
1175 return new ZExtInst(CondVal, SelType);
1177 // select C, -1, 0 -> sext C to int
1178 if (match(TrueVal, m_AllOnes()) && match(FalseVal, m_Zero()))
1179 return new SExtInst(CondVal, SelType);
1181 // select C, 0, 1 -> zext !C to int
1182 if (match(TrueVal, m_Zero()) && match(FalseVal, m_One())) {
1183 Value *NotCond = Builder->CreateNot(CondVal, "not." + CondVal->getName());
1184 return new ZExtInst(NotCond, SelType);
1187 // select C, 0, -1 -> sext !C to int
1188 if (match(TrueVal, m_Zero()) && match(FalseVal, m_AllOnes())) {
1189 Value *NotCond = Builder->CreateNot(CondVal, "not." + CondVal->getName());
1190 return new SExtInst(NotCond, SelType);
1194 if (ConstantInt *TrueValC = dyn_cast<ConstantInt>(TrueVal))
1195 if (ConstantInt *FalseValC = dyn_cast<ConstantInt>(FalseVal))
1196 if (Value *V = foldSelectICmpAnd(SI, TrueValC, FalseValC, Builder))
1197 return replaceInstUsesWith(SI, V);
1199 // See if we are selecting two values based on a comparison of the two values.
1200 if (FCmpInst *FCI = dyn_cast<FCmpInst>(CondVal)) {
1201 if (FCI->getOperand(0) == TrueVal && FCI->getOperand(1) == FalseVal) {
1202 // Transform (X == Y) ? X : Y -> Y
1203 if (FCI->getPredicate() == FCmpInst::FCMP_OEQ) {
1204 // This is not safe in general for floating point:
1205 // consider X== -0, Y== +0.
1206 // It becomes safe if either operand is a nonzero constant.
1207 ConstantFP *CFPt, *CFPf;
1208 if (((CFPt = dyn_cast<ConstantFP>(TrueVal)) &&
1209 !CFPt->getValueAPF().isZero()) ||
1210 ((CFPf = dyn_cast<ConstantFP>(FalseVal)) &&
1211 !CFPf->getValueAPF().isZero()))
1212 return replaceInstUsesWith(SI, FalseVal);
1214 // Transform (X une Y) ? X : Y -> X
1215 if (FCI->getPredicate() == FCmpInst::FCMP_UNE) {
1216 // This is not safe in general for floating point:
1217 // consider X== -0, Y== +0.
1218 // It becomes safe if either operand is a nonzero constant.
1219 ConstantFP *CFPt, *CFPf;
1220 if (((CFPt = dyn_cast<ConstantFP>(TrueVal)) &&
1221 !CFPt->getValueAPF().isZero()) ||
1222 ((CFPf = dyn_cast<ConstantFP>(FalseVal)) &&
1223 !CFPf->getValueAPF().isZero()))
1224 return replaceInstUsesWith(SI, TrueVal);
1227 // Canonicalize to use ordered comparisons by swapping the select
1231 // (X ugt Y) ? X : Y -> (X ole Y) ? Y : X
1232 if (FCI->hasOneUse() && FCmpInst::isUnordered(FCI->getPredicate())) {
1233 FCmpInst::Predicate InvPred = FCI->getInversePredicate();
1234 IRBuilder<>::FastMathFlagGuard FMFG(*Builder);
1235 Builder->setFastMathFlags(FCI->getFastMathFlags());
1236 Value *NewCond = Builder->CreateFCmp(InvPred, TrueVal, FalseVal,
1237 FCI->getName() + ".inv");
1239 return SelectInst::Create(NewCond, FalseVal, TrueVal,
1240 SI.getName() + ".p");
1243 // NOTE: if we wanted to, this is where to detect MIN/MAX
1244 } else if (FCI->getOperand(0) == FalseVal && FCI->getOperand(1) == TrueVal){
1245 // Transform (X == Y) ? Y : X -> X
1246 if (FCI->getPredicate() == FCmpInst::FCMP_OEQ) {
1247 // This is not safe in general for floating point:
1248 // consider X== -0, Y== +0.
1249 // It becomes safe if either operand is a nonzero constant.
1250 ConstantFP *CFPt, *CFPf;
1251 if (((CFPt = dyn_cast<ConstantFP>(TrueVal)) &&
1252 !CFPt->getValueAPF().isZero()) ||
1253 ((CFPf = dyn_cast<ConstantFP>(FalseVal)) &&
1254 !CFPf->getValueAPF().isZero()))
1255 return replaceInstUsesWith(SI, FalseVal);
1257 // Transform (X une Y) ? Y : X -> Y
1258 if (FCI->getPredicate() == FCmpInst::FCMP_UNE) {
1259 // This is not safe in general for floating point:
1260 // consider X== -0, Y== +0.
1261 // It becomes safe if either operand is a nonzero constant.
1262 ConstantFP *CFPt, *CFPf;
1263 if (((CFPt = dyn_cast<ConstantFP>(TrueVal)) &&
1264 !CFPt->getValueAPF().isZero()) ||
1265 ((CFPf = dyn_cast<ConstantFP>(FalseVal)) &&
1266 !CFPf->getValueAPF().isZero()))
1267 return replaceInstUsesWith(SI, TrueVal);
1270 // Canonicalize to use ordered comparisons by swapping the select
1274 // (X ugt Y) ? X : Y -> (X ole Y) ? X : Y
1275 if (FCI->hasOneUse() && FCmpInst::isUnordered(FCI->getPredicate())) {
1276 FCmpInst::Predicate InvPred = FCI->getInversePredicate();
1277 IRBuilder<>::FastMathFlagGuard FMFG(*Builder);
1278 Builder->setFastMathFlags(FCI->getFastMathFlags());
1279 Value *NewCond = Builder->CreateFCmp(InvPred, FalseVal, TrueVal,
1280 FCI->getName() + ".inv");
1282 return SelectInst::Create(NewCond, FalseVal, TrueVal,
1283 SI.getName() + ".p");
1286 // NOTE: if we wanted to, this is where to detect MIN/MAX
1288 // NOTE: if we wanted to, this is where to detect ABS
1291 // See if we are selecting two values based on a comparison of the two values.
1292 if (ICmpInst *ICI = dyn_cast<ICmpInst>(CondVal))
1293 if (Instruction *Result = foldSelectInstWithICmp(SI, ICI))
1296 if (Instruction *Add = foldAddSubSelect(SI, *Builder))
1299 // Turn (select C, (op X, Y), (op X, Z)) -> (op X, (select C, Y, Z))
1300 auto *TI = dyn_cast<Instruction>(TrueVal);
1301 auto *FI = dyn_cast<Instruction>(FalseVal);
1302 if (TI && FI && TI->getOpcode() == FI->getOpcode())
1303 if (Instruction *IV = foldSelectOpOp(SI, TI, FI))
1306 if (Instruction *I = foldSelectExtConst(SI))
1309 // See if we can fold the select into one of our operands.
1310 if (SelType->isIntOrIntVectorTy() || SelType->isFPOrFPVectorTy()) {
1311 if (Instruction *FoldI = foldSelectIntoOp(SI, TrueVal, FalseVal))
1314 Value *LHS, *RHS, *LHS2, *RHS2;
1315 Instruction::CastOps CastOp;
1316 SelectPatternResult SPR = matchSelectPattern(&SI, LHS, RHS, &CastOp);
1317 auto SPF = SPR.Flavor;
1319 if (SelectPatternResult::isMinOrMax(SPF)) {
1320 // Canonicalize so that type casts are outside select patterns.
1321 if (LHS->getType()->getPrimitiveSizeInBits() !=
1322 SelType->getPrimitiveSizeInBits()) {
1323 CmpInst::Predicate Pred = getCmpPredicateForMinMax(SPF, SPR.Ordered);
1326 if (CmpInst::isIntPredicate(Pred)) {
1327 Cmp = Builder->CreateICmp(Pred, LHS, RHS);
1329 IRBuilder<>::FastMathFlagGuard FMFG(*Builder);
1330 auto FMF = cast<FPMathOperator>(SI.getCondition())->getFastMathFlags();
1331 Builder->setFastMathFlags(FMF);
1332 Cmp = Builder->CreateFCmp(Pred, LHS, RHS);
1335 Value *NewSI = Builder->CreateCast(
1336 CastOp, Builder->CreateSelect(Cmp, LHS, RHS, SI.getName(), &SI),
1338 return replaceInstUsesWith(SI, NewSI);
1343 // MAX(MAX(a, b), a) -> MAX(a, b)
1344 // MIN(MIN(a, b), a) -> MIN(a, b)
1345 // MAX(MIN(a, b), a) -> a
1346 // MIN(MAX(a, b), a) -> a
1347 // ABS(ABS(a)) -> ABS(a)
1348 // NABS(NABS(a)) -> NABS(a)
1349 if (SelectPatternFlavor SPF2 = matchSelectPattern(LHS, LHS2, RHS2).Flavor)
1350 if (Instruction *R = foldSPFofSPF(cast<Instruction>(LHS),SPF2,LHS2,RHS2,
1353 if (SelectPatternFlavor SPF2 = matchSelectPattern(RHS, LHS2, RHS2).Flavor)
1354 if (Instruction *R = foldSPFofSPF(cast<Instruction>(RHS),SPF2,LHS2,RHS2,
1359 // MAX(~a, ~b) -> ~MIN(a, b)
1360 if ((SPF == SPF_SMAX || SPF == SPF_UMAX) &&
1361 IsFreeToInvert(LHS, LHS->hasNUses(2)) &&
1362 IsFreeToInvert(RHS, RHS->hasNUses(2))) {
1363 // For this transform to be profitable, we need to eliminate at least two
1364 // 'not' instructions if we're going to add one 'not' instruction.
1366 (LHS->hasNUses(2) && match(LHS, m_Not(m_Value()))) +
1367 (RHS->hasNUses(2) && match(RHS, m_Not(m_Value()))) +
1368 (SI.hasOneUse() && match(*SI.user_begin(), m_Not(m_Value())));
1370 if (NumberOfNots >= 2) {
1371 Value *NewLHS = Builder->CreateNot(LHS);
1372 Value *NewRHS = Builder->CreateNot(RHS);
1373 Value *NewCmp = SPF == SPF_SMAX
1374 ? Builder->CreateICmpSLT(NewLHS, NewRHS)
1375 : Builder->CreateICmpULT(NewLHS, NewRHS);
1377 Builder->CreateNot(Builder->CreateSelect(NewCmp, NewLHS, NewRHS));
1378 return replaceInstUsesWith(SI, NewSI);
1383 // ABS(-X) -> ABS(X)
1386 // See if we can fold the select into a phi node if the condition is a select.
1387 if (auto *PN = dyn_cast<PHINode>(SI.getCondition()))
1388 // The true/false values have to be live in the PHI predecessor's blocks.
1389 if (canSelectOperandBeMappingIntoPredBlock(TrueVal, SI) &&
1390 canSelectOperandBeMappingIntoPredBlock(FalseVal, SI))
1391 if (Instruction *NV = foldOpIntoPhi(SI, PN))
1394 if (SelectInst *TrueSI = dyn_cast<SelectInst>(TrueVal)) {
1395 if (TrueSI->getCondition()->getType() == CondVal->getType()) {
1396 // select(C, select(C, a, b), c) -> select(C, a, c)
1397 if (TrueSI->getCondition() == CondVal) {
1398 if (SI.getTrueValue() == TrueSI->getTrueValue())
1400 SI.setOperand(1, TrueSI->getTrueValue());
1403 // select(C0, select(C1, a, b), b) -> select(C0&C1, a, b)
1404 // We choose this as normal form to enable folding on the And and shortening
1405 // paths for the values (this helps GetUnderlyingObjects() for example).
1406 if (TrueSI->getFalseValue() == FalseVal && TrueSI->hasOneUse()) {
1407 Value *And = Builder->CreateAnd(CondVal, TrueSI->getCondition());
1408 SI.setOperand(0, And);
1409 SI.setOperand(1, TrueSI->getTrueValue());
1414 if (SelectInst *FalseSI = dyn_cast<SelectInst>(FalseVal)) {
1415 if (FalseSI->getCondition()->getType() == CondVal->getType()) {
1416 // select(C, a, select(C, b, c)) -> select(C, a, c)
1417 if (FalseSI->getCondition() == CondVal) {
1418 if (SI.getFalseValue() == FalseSI->getFalseValue())
1420 SI.setOperand(2, FalseSI->getFalseValue());
1423 // select(C0, a, select(C1, a, b)) -> select(C0|C1, a, b)
1424 if (FalseSI->getTrueValue() == TrueVal && FalseSI->hasOneUse()) {
1425 Value *Or = Builder->CreateOr(CondVal, FalseSI->getCondition());
1426 SI.setOperand(0, Or);
1427 SI.setOperand(2, FalseSI->getFalseValue());
1433 if (BinaryOperator::isNot(CondVal)) {
1434 SI.setOperand(0, BinaryOperator::getNotArgument(CondVal));
1435 SI.setOperand(1, FalseVal);
1436 SI.setOperand(2, TrueVal);
1440 if (VectorType *VecTy = dyn_cast<VectorType>(SelType)) {
1441 unsigned VWidth = VecTy->getNumElements();
1442 APInt UndefElts(VWidth, 0);
1443 APInt AllOnesEltMask(APInt::getAllOnesValue(VWidth));
1444 if (Value *V = SimplifyDemandedVectorElts(&SI, AllOnesEltMask, UndefElts)) {
1446 return replaceInstUsesWith(SI, V);
1450 if (isa<ConstantAggregateZero>(CondVal)) {
1451 return replaceInstUsesWith(SI, FalseVal);
1455 // See if we can determine the result of this select based on a dominating
1457 BasicBlock *Parent = SI.getParent();
1458 if (BasicBlock *Dom = Parent->getSinglePredecessor()) {
1459 auto *PBI = dyn_cast_or_null<BranchInst>(Dom->getTerminator());
1460 if (PBI && PBI->isConditional() &&
1461 PBI->getSuccessor(0) != PBI->getSuccessor(1) &&
1462 (PBI->getSuccessor(0) == Parent || PBI->getSuccessor(1) == Parent)) {
1463 bool CondIsFalse = PBI->getSuccessor(1) == Parent;
1464 Optional<bool> Implication = isImpliedCondition(
1465 PBI->getCondition(), SI.getCondition(), DL, CondIsFalse);
1467 Value *V = *Implication ? TrueVal : FalseVal;
1468 return replaceInstUsesWith(SI, V);
1473 // If we can compute the condition, there's no need for a select.
1474 // Like the above fold, we are attempting to reduce compile-time cost by
1475 // putting this fold here with limitations rather than in InstSimplify.
1476 // The motivation for this call into value tracking is to take advantage of
1477 // the assumption cache, so make sure that is populated.
1478 if (!CondVal->getType()->isVectorTy() && !AC.assumptions().empty()) {
1479 APInt KnownOne(1, 0), KnownZero(1, 0);
1480 computeKnownBits(CondVal, KnownZero, KnownOne, 0, &SI);
1482 return replaceInstUsesWith(SI, TrueVal);
1484 return replaceInstUsesWith(SI, FalseVal);
1487 if (Instruction *BitCastSel = foldSelectCmpBitcasts(SI, *Builder))