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 "InstCombine.h"
15 #include "llvm/Analysis/ConstantFolding.h"
16 #include "llvm/Analysis/InstructionSimplify.h"
17 #include "llvm/IR/PatternMatch.h"
19 using namespace PatternMatch;
21 #define DEBUG_TYPE "instcombine"
23 /// MatchSelectPattern - Pattern match integer [SU]MIN, [SU]MAX, and ABS idioms,
24 /// returning the kind and providing the out parameter results if we
25 /// successfully match.
26 static SelectPatternFlavor
27 MatchSelectPattern(Value *V, Value *&LHS, Value *&RHS) {
28 SelectInst *SI = dyn_cast<SelectInst>(V);
29 if (!SI) return SPF_UNKNOWN;
31 ICmpInst *ICI = dyn_cast<ICmpInst>(SI->getCondition());
32 if (!ICI) return SPF_UNKNOWN;
34 ICmpInst::Predicate Pred = ICI->getPredicate();
35 Value *CmpLHS = ICI->getOperand(0);
36 Value *CmpRHS = ICI->getOperand(1);
37 Value *TrueVal = SI->getTrueValue();
38 Value *FalseVal = SI->getFalseValue();
43 // (icmp X, Y) ? X : Y
44 if (TrueVal == CmpLHS && FalseVal == CmpRHS) {
46 default: return SPF_UNKNOWN; // Equality.
47 case ICmpInst::ICMP_UGT:
48 case ICmpInst::ICMP_UGE: return SPF_UMAX;
49 case ICmpInst::ICMP_SGT:
50 case ICmpInst::ICMP_SGE: return SPF_SMAX;
51 case ICmpInst::ICMP_ULT:
52 case ICmpInst::ICMP_ULE: return SPF_UMIN;
53 case ICmpInst::ICMP_SLT:
54 case ICmpInst::ICMP_SLE: return SPF_SMIN;
58 // (icmp X, Y) ? Y : X
59 if (TrueVal == CmpRHS && FalseVal == CmpLHS) {
61 default: return SPF_UNKNOWN; // Equality.
62 case ICmpInst::ICMP_UGT:
63 case ICmpInst::ICMP_UGE: return SPF_UMIN;
64 case ICmpInst::ICMP_SGT:
65 case ICmpInst::ICMP_SGE: return SPF_SMIN;
66 case ICmpInst::ICMP_ULT:
67 case ICmpInst::ICMP_ULE: return SPF_UMAX;
68 case ICmpInst::ICMP_SLT:
69 case ICmpInst::ICMP_SLE: return SPF_SMAX;
73 if (ConstantInt *C1 = dyn_cast<ConstantInt>(CmpRHS)) {
74 if ((CmpLHS == TrueVal && match(FalseVal, m_Neg(m_Specific(CmpLHS)))) ||
75 (CmpLHS == FalseVal && match(TrueVal, m_Neg(m_Specific(CmpLHS))))) {
77 // ABS(X) ==> (X >s 0) ? X : -X and (X >s -1) ? X : -X
78 // NABS(X) ==> (X >s 0) ? -X : X and (X >s -1) ? -X : X
79 if (Pred == ICmpInst::ICMP_SGT && (C1->isZero() || C1->isMinusOne())) {
80 return (CmpLHS == TrueVal) ? SPF_ABS : SPF_NABS;
83 // ABS(X) ==> (X <s 0) ? -X : X and (X <s 1) ? -X : X
84 // NABS(X) ==> (X <s 0) ? X : -X and (X <s 1) ? X : -X
85 if (Pred == ICmpInst::ICMP_SLT && (C1->isZero() || C1->isOne())) {
86 return (CmpLHS == FalseVal) ? SPF_ABS : SPF_NABS;
91 // TODO: (X > 4) ? X : 5 --> (X >= 5) ? X : 5 --> MAX(X, 5)
97 /// GetSelectFoldableOperands - We want to turn code that looks like this:
99 /// %D = select %cond, %C, %A
101 /// %C = select %cond, %B, 0
104 /// Assuming that the specified instruction is an operand to the select, return
105 /// a bitmask indicating which operands of this instruction are foldable if they
106 /// equal the other incoming value of the select.
108 static unsigned GetSelectFoldableOperands(Instruction *I) {
109 switch (I->getOpcode()) {
110 case Instruction::Add:
111 case Instruction::Mul:
112 case Instruction::And:
113 case Instruction::Or:
114 case Instruction::Xor:
115 return 3; // Can fold through either operand.
116 case Instruction::Sub: // Can only fold on the amount subtracted.
117 case Instruction::Shl: // Can only fold on the shift amount.
118 case Instruction::LShr:
119 case Instruction::AShr:
122 return 0; // Cannot fold
126 /// GetSelectFoldableConstant - For the same transformation as the previous
127 /// function, return the identity constant that goes into the select.
128 static Constant *GetSelectFoldableConstant(Instruction *I) {
129 switch (I->getOpcode()) {
130 default: llvm_unreachable("This cannot happen!");
131 case Instruction::Add:
132 case Instruction::Sub:
133 case Instruction::Or:
134 case Instruction::Xor:
135 case Instruction::Shl:
136 case Instruction::LShr:
137 case Instruction::AShr:
138 return Constant::getNullValue(I->getType());
139 case Instruction::And:
140 return Constant::getAllOnesValue(I->getType());
141 case Instruction::Mul:
142 return ConstantInt::get(I->getType(), 1);
146 /// FoldSelectOpOp - Here we have (select c, TI, FI), and we know that TI and FI
147 /// have the same opcode and only one use each. Try to simplify this.
148 Instruction *InstCombiner::FoldSelectOpOp(SelectInst &SI, Instruction *TI,
150 if (TI->getNumOperands() == 1) {
151 // If this is a non-volatile load or a cast from the same type,
154 Type *FIOpndTy = FI->getOperand(0)->getType();
155 if (TI->getOperand(0)->getType() != FIOpndTy)
157 // The select condition may be a vector. We may only change the operand
158 // type if the vector width remains the same (and matches the condition).
159 Type *CondTy = SI.getCondition()->getType();
160 if (CondTy->isVectorTy() && (!FIOpndTy->isVectorTy() ||
161 CondTy->getVectorNumElements() != FIOpndTy->getVectorNumElements()))
164 return nullptr; // unknown unary op.
167 // Fold this by inserting a select from the input values.
168 Value *NewSI = Builder->CreateSelect(SI.getCondition(), TI->getOperand(0),
169 FI->getOperand(0), SI.getName()+".v");
170 return CastInst::Create(Instruction::CastOps(TI->getOpcode()), NewSI,
174 // Only handle binary operators here.
175 if (!isa<BinaryOperator>(TI))
178 // Figure out if the operations have any operands in common.
179 Value *MatchOp, *OtherOpT, *OtherOpF;
181 if (TI->getOperand(0) == FI->getOperand(0)) {
182 MatchOp = TI->getOperand(0);
183 OtherOpT = TI->getOperand(1);
184 OtherOpF = FI->getOperand(1);
185 MatchIsOpZero = true;
186 } else if (TI->getOperand(1) == FI->getOperand(1)) {
187 MatchOp = TI->getOperand(1);
188 OtherOpT = TI->getOperand(0);
189 OtherOpF = FI->getOperand(0);
190 MatchIsOpZero = false;
191 } else if (!TI->isCommutative()) {
193 } else if (TI->getOperand(0) == FI->getOperand(1)) {
194 MatchOp = TI->getOperand(0);
195 OtherOpT = TI->getOperand(1);
196 OtherOpF = FI->getOperand(0);
197 MatchIsOpZero = true;
198 } else if (TI->getOperand(1) == FI->getOperand(0)) {
199 MatchOp = TI->getOperand(1);
200 OtherOpT = TI->getOperand(0);
201 OtherOpF = FI->getOperand(1);
202 MatchIsOpZero = true;
207 // If we reach here, they do have operations in common.
208 Value *NewSI = Builder->CreateSelect(SI.getCondition(), OtherOpT,
209 OtherOpF, SI.getName()+".v");
211 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(TI)) {
213 return BinaryOperator::Create(BO->getOpcode(), MatchOp, NewSI);
215 return BinaryOperator::Create(BO->getOpcode(), NewSI, MatchOp);
217 llvm_unreachable("Shouldn't get here");
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 /// FoldSelectIntoOp - Try fold the select into one of the operands to
234 /// facilitate further optimization.
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 if (isa<PossiblyExactOperator>(BO))
262 BO->setIsExact(TVI_BO->isExact());
263 if (isa<OverflowingBinaryOperator>(BO)) {
264 BO->setHasNoUnsignedWrap(TVI_BO->hasNoUnsignedWrap());
265 BO->setHasNoSignedWrap(TVI_BO->hasNoSignedWrap());
274 if (Instruction *FVI = dyn_cast<Instruction>(FalseVal)) {
275 if (FVI->hasOneUse() && FVI->getNumOperands() == 2 &&
276 !isa<Constant>(TrueVal)) {
277 if (unsigned SFO = GetSelectFoldableOperands(FVI)) {
278 unsigned OpToFold = 0;
279 if ((SFO & 1) && TrueVal == FVI->getOperand(0)) {
281 } else if ((SFO & 2) && TrueVal == FVI->getOperand(1)) {
286 Constant *C = GetSelectFoldableConstant(FVI);
287 Value *OOp = FVI->getOperand(2-OpToFold);
288 // Avoid creating select between 2 constants unless it's selecting
289 // between 0, 1 and -1.
290 if (!isa<Constant>(OOp) || isSelect01(C, cast<Constant>(OOp))) {
291 Value *NewSel = Builder->CreateSelect(SI.getCondition(), C, OOp);
292 NewSel->takeName(FVI);
293 BinaryOperator *FVI_BO = cast<BinaryOperator>(FVI);
294 BinaryOperator *BO = BinaryOperator::Create(FVI_BO->getOpcode(),
296 if (isa<PossiblyExactOperator>(BO))
297 BO->setIsExact(FVI_BO->isExact());
298 if (isa<OverflowingBinaryOperator>(BO)) {
299 BO->setHasNoUnsignedWrap(FVI_BO->hasNoUnsignedWrap());
300 BO->setHasNoSignedWrap(FVI_BO->hasNoSignedWrap());
312 /// SimplifyWithOpReplaced - See if V simplifies when its operand Op is
313 /// replaced with RepOp.
314 static Value *SimplifyWithOpReplaced(Value *V, Value *Op, Value *RepOp,
315 const DataLayout *TD,
316 const TargetLibraryInfo *TLI,
318 AssumptionTracker *AT) {
319 // Trivial replacement.
323 Instruction *I = dyn_cast<Instruction>(V);
327 // If this is a binary operator, try to simplify it with the replaced op.
328 if (BinaryOperator *B = dyn_cast<BinaryOperator>(I)) {
329 if (B->getOperand(0) == Op)
330 return SimplifyBinOp(B->getOpcode(), RepOp, B->getOperand(1), TD, TLI);
331 if (B->getOperand(1) == Op)
332 return SimplifyBinOp(B->getOpcode(), B->getOperand(0), RepOp, TD, TLI);
335 // Same for CmpInsts.
336 if (CmpInst *C = dyn_cast<CmpInst>(I)) {
337 if (C->getOperand(0) == Op)
338 return SimplifyCmpInst(C->getPredicate(), RepOp, C->getOperand(1), TD,
340 if (C->getOperand(1) == Op)
341 return SimplifyCmpInst(C->getPredicate(), C->getOperand(0), RepOp, TD,
345 // TODO: We could hand off more cases to instsimplify here.
347 // If all operands are constant after substituting Op for RepOp then we can
348 // constant fold the instruction.
349 if (Constant *CRepOp = dyn_cast<Constant>(RepOp)) {
350 // Build a list of all constant operands.
351 SmallVector<Constant*, 8> ConstOps;
352 for (unsigned i = 0, e = I->getNumOperands(); i != e; ++i) {
353 if (I->getOperand(i) == Op)
354 ConstOps.push_back(CRepOp);
355 else if (Constant *COp = dyn_cast<Constant>(I->getOperand(i)))
356 ConstOps.push_back(COp);
361 // All operands were constants, fold it.
362 if (ConstOps.size() == I->getNumOperands()) {
363 if (CmpInst *C = dyn_cast<CmpInst>(I))
364 return ConstantFoldCompareInstOperands(C->getPredicate(), ConstOps[0],
365 ConstOps[1], TD, TLI);
367 if (LoadInst *LI = dyn_cast<LoadInst>(I))
368 if (!LI->isVolatile())
369 return ConstantFoldLoadFromConstPtr(ConstOps[0], TD);
371 return ConstantFoldInstOperands(I->getOpcode(), I->getType(),
379 /// foldSelectICmpAndOr - We want to turn:
380 /// (select (icmp eq (and X, C1), 0), Y, (or Y, C2))
382 /// (or (shl (and X, C1), C3), y)
384 /// C1 and C2 are both powers of 2
386 /// C3 = Log(C2) - Log(C1)
388 /// This transform handles cases where:
389 /// 1. The icmp predicate is inverted
390 /// 2. The select operands are reversed
391 /// 3. The magnitude of C2 and C1 are flipped
393 /// This also tries to turn
394 /// --- Single bit tests:
395 /// if ((x & C) == 0) x |= C to x |= C
396 /// if ((x & C) != 0) x ^= C to x &= ~C
397 /// if ((x & C) == 0) x ^= C to x |= C
398 /// if ((x & C) != 0) x &= ~C to x &= ~C
399 /// if ((x & C) == 0) x &= ~C to nothing
400 static Value *foldSelectICmpAndOr(SelectInst &SI, Value *TrueVal,
402 InstCombiner::BuilderTy *Builder) {
403 const ICmpInst *IC = dyn_cast<ICmpInst>(SI.getCondition());
404 if (!IC || !IC->isEquality() || !SI.getType()->isIntegerTy())
407 Value *CmpLHS = IC->getOperand(0);
408 Value *CmpRHS = IC->getOperand(1);
410 if (!match(CmpRHS, m_Zero()))
415 if (!match(CmpLHS, m_And(m_Value(X), m_Power2(C1))))
419 if (match(TrueVal, m_Specific(X))) {
420 // if ((X & C) != 0) X ^= C becomes X &= ~C
421 if (match(FalseVal, m_Xor(m_Specific(X), m_APInt(C2))) && C1 == C2)
422 return Builder->CreateAnd(X, ~(*C1));
423 // if ((X & C) != 0) X &= ~C becomes X &= ~C
424 if (match(FalseVal, m_And(m_Specific(X), m_APInt(C2))) && *C1 == ~(*C2))
426 } else if (match(FalseVal, m_Specific(X))) {
427 // if ((X & C) == 0) X ^= C becomes X |= C
428 if (match(TrueVal, m_Xor(m_Specific(X), m_APInt(C2))) && C1 == C2)
429 return Builder->CreateOr(X, *C1);
430 // if ((X & C) == 0) X &= ~C becomes nothing
431 if (match(TrueVal, m_And(m_Specific(X), m_APInt(C2))) && *C1 == ~(*C2))
433 // if ((X & C) == 0) X |= C becomes X |= C
434 if (match(TrueVal, m_Or(m_Specific(X), m_APInt(C2))) && C1 == C2)
438 bool OrOnTrueVal = false;
439 bool OrOnFalseVal = match(FalseVal, m_Or(m_Specific(TrueVal), m_Power2(C2)));
441 OrOnTrueVal = match(TrueVal, m_Or(m_Specific(FalseVal), m_Power2(C2)));
443 if (!OrOnFalseVal && !OrOnTrueVal)
447 Value *Y = OrOnFalseVal ? TrueVal : FalseVal;
449 unsigned C1Log = C1->logBase2();
450 unsigned C2Log = C2->logBase2();
452 V = Builder->CreateZExtOrTrunc(V, Y->getType());
453 V = Builder->CreateShl(V, C2Log - C1Log);
454 } else if (C1Log > C2Log) {
455 V = Builder->CreateLShr(V, C1Log - C2Log);
456 V = Builder->CreateZExtOrTrunc(V, Y->getType());
458 V = Builder->CreateZExtOrTrunc(V, Y->getType());
460 ICmpInst::Predicate Pred = IC->getPredicate();
461 if ((Pred == ICmpInst::ICMP_NE && OrOnFalseVal) ||
462 (Pred == ICmpInst::ICMP_EQ && OrOnTrueVal))
463 V = Builder->CreateXor(V, *C2);
465 return Builder->CreateOr(V, Y);
468 /// visitSelectInstWithICmp - Visit a SelectInst that has an
469 /// ICmpInst as its first operand.
471 Instruction *InstCombiner::visitSelectInstWithICmp(SelectInst &SI,
473 bool Changed = false;
474 ICmpInst::Predicate Pred = ICI->getPredicate();
475 Value *CmpLHS = ICI->getOperand(0);
476 Value *CmpRHS = ICI->getOperand(1);
477 Value *TrueVal = SI.getTrueValue();
478 Value *FalseVal = SI.getFalseValue();
480 // Check cases where the comparison is with a constant that
481 // can be adjusted to fit the min/max idiom. We may move or edit ICI
482 // here, so make sure the select is the only user.
483 if (ICI->hasOneUse())
484 if (ConstantInt *CI = dyn_cast<ConstantInt>(CmpRHS)) {
485 // X < MIN ? T : F --> F
486 if ((Pred == ICmpInst::ICMP_SLT || Pred == ICmpInst::ICMP_ULT)
487 && CI->isMinValue(Pred == ICmpInst::ICMP_SLT))
488 return ReplaceInstUsesWith(SI, FalseVal);
489 // X > MAX ? T : F --> F
490 else if ((Pred == ICmpInst::ICMP_SGT || Pred == ICmpInst::ICMP_UGT)
491 && CI->isMaxValue(Pred == ICmpInst::ICMP_SGT))
492 return ReplaceInstUsesWith(SI, FalseVal);
495 case ICmpInst::ICMP_ULT:
496 case ICmpInst::ICMP_SLT:
497 case ICmpInst::ICMP_UGT:
498 case ICmpInst::ICMP_SGT: {
499 // These transformations only work for selects over integers.
500 IntegerType *SelectTy = dyn_cast<IntegerType>(SI.getType());
504 Constant *AdjustedRHS;
505 if (Pred == ICmpInst::ICMP_UGT || Pred == ICmpInst::ICMP_SGT)
506 AdjustedRHS = ConstantInt::get(CI->getContext(), CI->getValue() + 1);
507 else // (Pred == ICmpInst::ICMP_ULT || Pred == ICmpInst::ICMP_SLT)
508 AdjustedRHS = ConstantInt::get(CI->getContext(), CI->getValue() - 1);
510 // X > C ? X : C+1 --> X < C+1 ? C+1 : X
511 // X < C ? X : C-1 --> X > C-1 ? C-1 : X
512 if ((CmpLHS == TrueVal && AdjustedRHS == FalseVal) ||
513 (CmpLHS == FalseVal && AdjustedRHS == TrueVal))
514 ; // Nothing to do here. Values match without any sign/zero extension.
516 // Types do not match. Instead of calculating this with mixed types
517 // promote all to the larger type. This enables scalar evolution to
518 // analyze this expression.
519 else if (CmpRHS->getType()->getScalarSizeInBits()
520 < SelectTy->getBitWidth()) {
521 Constant *sextRHS = ConstantExpr::getSExt(AdjustedRHS, SelectTy);
523 // X = sext x; x >s c ? X : C+1 --> X = sext x; X <s C+1 ? C+1 : X
524 // X = sext x; x <s c ? X : C-1 --> X = sext x; X >s C-1 ? C-1 : X
525 // X = sext x; x >u c ? X : C+1 --> X = sext x; X <u C+1 ? C+1 : X
526 // X = sext x; x <u c ? X : C-1 --> X = sext x; X >u C-1 ? C-1 : X
527 if (match(TrueVal, m_SExt(m_Specific(CmpLHS))) &&
528 sextRHS == FalseVal) {
530 AdjustedRHS = sextRHS;
531 } else if (match(FalseVal, m_SExt(m_Specific(CmpLHS))) &&
532 sextRHS == TrueVal) {
534 AdjustedRHS = sextRHS;
535 } else if (ICI->isUnsigned()) {
536 Constant *zextRHS = ConstantExpr::getZExt(AdjustedRHS, SelectTy);
537 // X = zext x; x >u c ? X : C+1 --> X = zext x; X <u C+1 ? C+1 : X
538 // X = zext x; x <u c ? X : C-1 --> X = zext x; X >u C-1 ? C-1 : X
539 // zext + signed compare cannot be changed:
540 // 0xff <s 0x00, but 0x00ff >s 0x0000
541 if (match(TrueVal, m_ZExt(m_Specific(CmpLHS))) &&
542 zextRHS == FalseVal) {
544 AdjustedRHS = zextRHS;
545 } else if (match(FalseVal, m_ZExt(m_Specific(CmpLHS))) &&
546 zextRHS == TrueVal) {
548 AdjustedRHS = zextRHS;
556 Pred = ICmpInst::getSwappedPredicate(Pred);
557 CmpRHS = AdjustedRHS;
558 std::swap(FalseVal, TrueVal);
559 ICI->setPredicate(Pred);
560 ICI->setOperand(0, CmpLHS);
561 ICI->setOperand(1, CmpRHS);
562 SI.setOperand(1, TrueVal);
563 SI.setOperand(2, FalseVal);
565 // Move ICI instruction right before the select instruction. Otherwise
566 // the sext/zext value may be defined after the ICI instruction uses it.
567 ICI->moveBefore(&SI);
575 // Transform (X >s -1) ? C1 : C2 --> ((X >>s 31) & (C2 - C1)) + C1
576 // and (X <s 0) ? C2 : C1 --> ((X >>s 31) & (C2 - C1)) + C1
577 // FIXME: Type and constness constraints could be lifted, but we have to
578 // watch code size carefully. We should consider xor instead of
579 // sub/add when we decide to do that.
580 if (IntegerType *Ty = dyn_cast<IntegerType>(CmpLHS->getType())) {
581 if (TrueVal->getType() == Ty) {
582 if (ConstantInt *Cmp = dyn_cast<ConstantInt>(CmpRHS)) {
583 ConstantInt *C1 = nullptr, *C2 = nullptr;
584 if (Pred == ICmpInst::ICMP_SGT && Cmp->isAllOnesValue()) {
585 C1 = dyn_cast<ConstantInt>(TrueVal);
586 C2 = dyn_cast<ConstantInt>(FalseVal);
587 } else if (Pred == ICmpInst::ICMP_SLT && Cmp->isNullValue()) {
588 C1 = dyn_cast<ConstantInt>(FalseVal);
589 C2 = dyn_cast<ConstantInt>(TrueVal);
592 // This shift results in either -1 or 0.
593 Value *AShr = Builder->CreateAShr(CmpLHS, Ty->getBitWidth()-1);
595 // Check if we can express the operation with a single or.
596 if (C2->isAllOnesValue())
597 return ReplaceInstUsesWith(SI, Builder->CreateOr(AShr, C1));
599 Value *And = Builder->CreateAnd(AShr, C2->getValue()-C1->getValue());
600 return ReplaceInstUsesWith(SI, Builder->CreateAdd(And, C1));
606 // If we have an equality comparison then we know the value in one of the
607 // arms of the select. See if substituting this value into the arm and
608 // simplifying the result yields the same value as the other arm.
609 if (Pred == ICmpInst::ICMP_EQ) {
610 if (SimplifyWithOpReplaced(FalseVal, CmpLHS, CmpRHS, DL, TLI,
611 DT, AT) == TrueVal ||
612 SimplifyWithOpReplaced(FalseVal, CmpRHS, CmpLHS, DL, TLI,
614 return ReplaceInstUsesWith(SI, FalseVal);
615 if (SimplifyWithOpReplaced(TrueVal, CmpLHS, CmpRHS, DL, TLI,
616 DT, AT) == FalseVal ||
617 SimplifyWithOpReplaced(TrueVal, CmpRHS, CmpLHS, DL, TLI,
619 return ReplaceInstUsesWith(SI, FalseVal);
620 } else if (Pred == ICmpInst::ICMP_NE) {
621 if (SimplifyWithOpReplaced(TrueVal, CmpLHS, CmpRHS, DL, TLI,
622 DT, AT) == FalseVal ||
623 SimplifyWithOpReplaced(TrueVal, CmpRHS, CmpLHS, DL, TLI,
625 return ReplaceInstUsesWith(SI, TrueVal);
626 if (SimplifyWithOpReplaced(FalseVal, CmpLHS, CmpRHS, DL, TLI,
627 DT, AT) == TrueVal ||
628 SimplifyWithOpReplaced(FalseVal, CmpRHS, CmpLHS, DL, TLI,
630 return ReplaceInstUsesWith(SI, TrueVal);
633 // NOTE: if we wanted to, this is where to detect integer MIN/MAX
635 if (CmpRHS != CmpLHS && isa<Constant>(CmpRHS)) {
636 if (CmpLHS == TrueVal && Pred == ICmpInst::ICMP_EQ) {
637 // Transform (X == C) ? X : Y -> (X == C) ? C : Y
638 SI.setOperand(1, CmpRHS);
640 } else if (CmpLHS == FalseVal && Pred == ICmpInst::ICMP_NE) {
641 // Transform (X != C) ? Y : X -> (X != C) ? Y : C
642 SI.setOperand(2, CmpRHS);
647 if (Value *V = foldSelectICmpAndOr(SI, TrueVal, FalseVal, Builder))
648 return ReplaceInstUsesWith(SI, V);
650 return Changed ? &SI : nullptr;
654 /// CanSelectOperandBeMappingIntoPredBlock - SI is a select whose condition is a
655 /// PHI node (but the two may be in different blocks). See if the true/false
656 /// values (V) are live in all of the predecessor blocks of the PHI. For
657 /// example, cases like this cannot be mapped:
659 /// X = phi [ C1, BB1], [C2, BB2]
661 /// Z = select X, Y, 0
663 /// because Y is not live in BB1/BB2.
665 static bool CanSelectOperandBeMappingIntoPredBlock(const Value *V,
666 const SelectInst &SI) {
667 // If the value is a non-instruction value like a constant or argument, it
668 // can always be mapped.
669 const Instruction *I = dyn_cast<Instruction>(V);
672 // If V is a PHI node defined in the same block as the condition PHI, we can
673 // map the arguments.
674 const PHINode *CondPHI = cast<PHINode>(SI.getCondition());
676 if (const PHINode *VP = dyn_cast<PHINode>(I))
677 if (VP->getParent() == CondPHI->getParent())
680 // Otherwise, if the PHI and select are defined in the same block and if V is
681 // defined in a different block, then we can transform it.
682 if (SI.getParent() == CondPHI->getParent() &&
683 I->getParent() != CondPHI->getParent())
686 // Otherwise we have a 'hard' case and we can't tell without doing more
687 // detailed dominator based analysis, punt.
691 /// FoldSPFofSPF - We have an SPF (e.g. a min or max) of an SPF of the form:
692 /// SPF2(SPF1(A, B), C)
693 Instruction *InstCombiner::FoldSPFofSPF(Instruction *Inner,
694 SelectPatternFlavor SPF1,
697 SelectPatternFlavor SPF2, Value *C) {
698 if (C == A || C == B) {
699 // MAX(MAX(A, B), B) -> MAX(A, B)
700 // MIN(MIN(a, b), a) -> MIN(a, b)
702 return ReplaceInstUsesWith(Outer, Inner);
704 // MAX(MIN(a, b), a) -> a
705 // MIN(MAX(a, b), a) -> a
706 if ((SPF1 == SPF_SMIN && SPF2 == SPF_SMAX) ||
707 (SPF1 == SPF_SMAX && SPF2 == SPF_SMIN) ||
708 (SPF1 == SPF_UMIN && SPF2 == SPF_UMAX) ||
709 (SPF1 == SPF_UMAX && SPF2 == SPF_UMIN))
710 return ReplaceInstUsesWith(Outer, C);
714 if (ConstantInt *CB = dyn_cast<ConstantInt>(B)) {
715 if (ConstantInt *CC = dyn_cast<ConstantInt>(C)) {
716 APInt ACB = CB->getValue();
717 APInt ACC = CC->getValue();
719 // MIN(MIN(A, 23), 97) -> MIN(A, 23)
720 // MAX(MAX(A, 97), 23) -> MAX(A, 97)
721 if ((SPF1 == SPF_UMIN && ACB.ule(ACC)) ||
722 (SPF1 == SPF_SMIN && ACB.sle(ACC)) ||
723 (SPF1 == SPF_UMAX && ACB.uge(ACC)) ||
724 (SPF1 == SPF_SMAX && ACB.sge(ACC)))
725 return ReplaceInstUsesWith(Outer, Inner);
727 // MIN(MIN(A, 97), 23) -> MIN(A, 23)
728 // MAX(MAX(A, 23), 97) -> MAX(A, 97)
729 if ((SPF1 == SPF_UMIN && ACB.ugt(ACC)) ||
730 (SPF1 == SPF_SMIN && ACB.sgt(ACC)) ||
731 (SPF1 == SPF_UMAX && ACB.ult(ACC)) ||
732 (SPF1 == SPF_SMAX && ACB.slt(ACC))) {
733 Outer.replaceUsesOfWith(Inner, A);
740 // ABS(ABS(X)) -> ABS(X)
741 // NABS(NABS(X)) -> NABS(X)
742 if (SPF1 == SPF2 && (SPF1 == SPF_ABS || SPF1 == SPF_NABS)) {
743 return ReplaceInstUsesWith(Outer, Inner);
746 // ABS(NABS(X)) -> ABS(X)
747 // NABS(ABS(X)) -> NABS(X)
748 if ((SPF1 == SPF_ABS && SPF2 == SPF_NABS) ||
749 (SPF1 == SPF_NABS && SPF2 == SPF_ABS)) {
750 SelectInst *SI = cast<SelectInst>(Inner);
751 Value *NewSI = Builder->CreateSelect(
752 SI->getCondition(), SI->getFalseValue(), SI->getTrueValue());
753 return ReplaceInstUsesWith(Outer, NewSI);
758 /// foldSelectICmpAnd - If one of the constants is zero (we know they can't
759 /// both be) and we have an icmp instruction with zero, and we have an 'and'
760 /// with the non-constant value and a power of two we can turn the select
761 /// into a shift on the result of the 'and'.
762 static Value *foldSelectICmpAnd(const SelectInst &SI, ConstantInt *TrueVal,
763 ConstantInt *FalseVal,
764 InstCombiner::BuilderTy *Builder) {
765 const ICmpInst *IC = dyn_cast<ICmpInst>(SI.getCondition());
766 if (!IC || !IC->isEquality() || !SI.getType()->isIntegerTy())
769 if (!match(IC->getOperand(1), m_Zero()))
773 Value *LHS = IC->getOperand(0);
774 if (!match(LHS, m_And(m_Value(), m_ConstantInt(AndRHS))))
777 // If both select arms are non-zero see if we have a select of the form
778 // 'x ? 2^n + C : C'. Then we can offset both arms by C, use the logic
779 // for 'x ? 2^n : 0' and fix the thing up at the end.
780 ConstantInt *Offset = nullptr;
781 if (!TrueVal->isZero() && !FalseVal->isZero()) {
782 if ((TrueVal->getValue() - FalseVal->getValue()).isPowerOf2())
784 else if ((FalseVal->getValue() - TrueVal->getValue()).isPowerOf2())
789 // Adjust TrueVal and FalseVal to the offset.
790 TrueVal = ConstantInt::get(Builder->getContext(),
791 TrueVal->getValue() - Offset->getValue());
792 FalseVal = ConstantInt::get(Builder->getContext(),
793 FalseVal->getValue() - Offset->getValue());
796 // Make sure the mask in the 'and' and one of the select arms is a power of 2.
797 if (!AndRHS->getValue().isPowerOf2() ||
798 (!TrueVal->getValue().isPowerOf2() &&
799 !FalseVal->getValue().isPowerOf2()))
802 // Determine which shift is needed to transform result of the 'and' into the
804 ConstantInt *ValC = !TrueVal->isZero() ? TrueVal : FalseVal;
805 unsigned ValZeros = ValC->getValue().logBase2();
806 unsigned AndZeros = AndRHS->getValue().logBase2();
808 // If types don't match we can still convert the select by introducing a zext
809 // or a trunc of the 'and'. The trunc case requires that all of the truncated
810 // bits are zero, we can figure that out by looking at the 'and' mask.
811 if (AndZeros >= ValC->getBitWidth())
814 Value *V = Builder->CreateZExtOrTrunc(LHS, SI.getType());
815 if (ValZeros > AndZeros)
816 V = Builder->CreateShl(V, ValZeros - AndZeros);
817 else if (ValZeros < AndZeros)
818 V = Builder->CreateLShr(V, AndZeros - ValZeros);
820 // Okay, now we know that everything is set up, we just don't know whether we
821 // have a icmp_ne or icmp_eq and whether the true or false val is the zero.
822 bool ShouldNotVal = !TrueVal->isZero();
823 ShouldNotVal ^= IC->getPredicate() == ICmpInst::ICMP_NE;
825 V = Builder->CreateXor(V, ValC);
827 // Apply an offset if needed.
829 V = Builder->CreateAdd(V, Offset);
833 Instruction *InstCombiner::visitSelectInst(SelectInst &SI) {
834 Value *CondVal = SI.getCondition();
835 Value *TrueVal = SI.getTrueValue();
836 Value *FalseVal = SI.getFalseValue();
838 if (Value *V = SimplifySelectInst(CondVal, TrueVal, FalseVal, DL, TLI,
840 return ReplaceInstUsesWith(SI, V);
842 if (SI.getType()->isIntegerTy(1)) {
843 if (ConstantInt *C = dyn_cast<ConstantInt>(TrueVal)) {
844 if (C->getZExtValue()) {
845 // Change: A = select B, true, C --> A = or B, C
846 return BinaryOperator::CreateOr(CondVal, FalseVal);
848 // Change: A = select B, false, C --> A = and !B, C
849 Value *NotCond = Builder->CreateNot(CondVal, "not."+CondVal->getName());
850 return BinaryOperator::CreateAnd(NotCond, FalseVal);
852 if (ConstantInt *C = dyn_cast<ConstantInt>(FalseVal)) {
853 if (C->getZExtValue() == false) {
854 // Change: A = select B, C, false --> A = and B, C
855 return BinaryOperator::CreateAnd(CondVal, TrueVal);
857 // Change: A = select B, C, true --> A = or !B, C
858 Value *NotCond = Builder->CreateNot(CondVal, "not."+CondVal->getName());
859 return BinaryOperator::CreateOr(NotCond, TrueVal);
862 // select a, b, a -> a&b
863 // select a, a, b -> a|b
864 if (CondVal == TrueVal)
865 return BinaryOperator::CreateOr(CondVal, FalseVal);
866 if (CondVal == FalseVal)
867 return BinaryOperator::CreateAnd(CondVal, TrueVal);
869 // select a, ~a, b -> (~a)&b
870 // select a, b, ~a -> (~a)|b
871 if (match(TrueVal, m_Not(m_Specific(CondVal))))
872 return BinaryOperator::CreateAnd(TrueVal, FalseVal);
873 if (match(FalseVal, m_Not(m_Specific(CondVal))))
874 return BinaryOperator::CreateOr(TrueVal, FalseVal);
877 // Selecting between two integer constants?
878 if (ConstantInt *TrueValC = dyn_cast<ConstantInt>(TrueVal))
879 if (ConstantInt *FalseValC = dyn_cast<ConstantInt>(FalseVal)) {
880 // select C, 1, 0 -> zext C to int
881 if (FalseValC->isZero() && TrueValC->getValue() == 1)
882 return new ZExtInst(CondVal, SI.getType());
884 // select C, -1, 0 -> sext C to int
885 if (FalseValC->isZero() && TrueValC->isAllOnesValue())
886 return new SExtInst(CondVal, SI.getType());
888 // select C, 0, 1 -> zext !C to int
889 if (TrueValC->isZero() && FalseValC->getValue() == 1) {
890 Value *NotCond = Builder->CreateNot(CondVal, "not."+CondVal->getName());
891 return new ZExtInst(NotCond, SI.getType());
894 // select C, 0, -1 -> sext !C to int
895 if (TrueValC->isZero() && FalseValC->isAllOnesValue()) {
896 Value *NotCond = Builder->CreateNot(CondVal, "not."+CondVal->getName());
897 return new SExtInst(NotCond, SI.getType());
900 if (Value *V = foldSelectICmpAnd(SI, TrueValC, FalseValC, Builder))
901 return ReplaceInstUsesWith(SI, V);
904 // See if we are selecting two values based on a comparison of the two values.
905 if (FCmpInst *FCI = dyn_cast<FCmpInst>(CondVal)) {
906 if (FCI->getOperand(0) == TrueVal && FCI->getOperand(1) == FalseVal) {
907 // Transform (X == Y) ? X : Y -> Y
908 if (FCI->getPredicate() == FCmpInst::FCMP_OEQ) {
909 // This is not safe in general for floating point:
910 // consider X== -0, Y== +0.
911 // It becomes safe if either operand is a nonzero constant.
912 ConstantFP *CFPt, *CFPf;
913 if (((CFPt = dyn_cast<ConstantFP>(TrueVal)) &&
914 !CFPt->getValueAPF().isZero()) ||
915 ((CFPf = dyn_cast<ConstantFP>(FalseVal)) &&
916 !CFPf->getValueAPF().isZero()))
917 return ReplaceInstUsesWith(SI, FalseVal);
919 // Transform (X une Y) ? X : Y -> X
920 if (FCI->getPredicate() == FCmpInst::FCMP_UNE) {
921 // This is not safe in general for floating point:
922 // consider X== -0, Y== +0.
923 // It becomes safe if either operand is a nonzero constant.
924 ConstantFP *CFPt, *CFPf;
925 if (((CFPt = dyn_cast<ConstantFP>(TrueVal)) &&
926 !CFPt->getValueAPF().isZero()) ||
927 ((CFPf = dyn_cast<ConstantFP>(FalseVal)) &&
928 !CFPf->getValueAPF().isZero()))
929 return ReplaceInstUsesWith(SI, TrueVal);
931 // NOTE: if we wanted to, this is where to detect MIN/MAX
933 } else if (FCI->getOperand(0) == FalseVal && FCI->getOperand(1) == TrueVal){
934 // Transform (X == Y) ? Y : X -> X
935 if (FCI->getPredicate() == FCmpInst::FCMP_OEQ) {
936 // This is not safe in general for floating point:
937 // consider X== -0, Y== +0.
938 // It becomes safe if either operand is a nonzero constant.
939 ConstantFP *CFPt, *CFPf;
940 if (((CFPt = dyn_cast<ConstantFP>(TrueVal)) &&
941 !CFPt->getValueAPF().isZero()) ||
942 ((CFPf = dyn_cast<ConstantFP>(FalseVal)) &&
943 !CFPf->getValueAPF().isZero()))
944 return ReplaceInstUsesWith(SI, FalseVal);
946 // Transform (X une Y) ? Y : X -> Y
947 if (FCI->getPredicate() == FCmpInst::FCMP_UNE) {
948 // This is not safe in general for floating point:
949 // consider X== -0, Y== +0.
950 // It becomes safe if either operand is a nonzero constant.
951 ConstantFP *CFPt, *CFPf;
952 if (((CFPt = dyn_cast<ConstantFP>(TrueVal)) &&
953 !CFPt->getValueAPF().isZero()) ||
954 ((CFPf = dyn_cast<ConstantFP>(FalseVal)) &&
955 !CFPf->getValueAPF().isZero()))
956 return ReplaceInstUsesWith(SI, TrueVal);
958 // NOTE: if we wanted to, this is where to detect MIN/MAX
960 // NOTE: if we wanted to, this is where to detect ABS
963 // See if we are selecting two values based on a comparison of the two values.
964 if (ICmpInst *ICI = dyn_cast<ICmpInst>(CondVal))
965 if (Instruction *Result = visitSelectInstWithICmp(SI, ICI))
968 if (Instruction *TI = dyn_cast<Instruction>(TrueVal))
969 if (Instruction *FI = dyn_cast<Instruction>(FalseVal))
970 if (TI->hasOneUse() && FI->hasOneUse()) {
971 Instruction *AddOp = nullptr, *SubOp = nullptr;
973 // Turn (select C, (op X, Y), (op X, Z)) -> (op X, (select C, Y, Z))
974 if (TI->getOpcode() == FI->getOpcode())
975 if (Instruction *IV = FoldSelectOpOp(SI, TI, FI))
978 // Turn select C, (X+Y), (X-Y) --> (X+(select C, Y, (-Y))). This is
979 // even legal for FP.
980 if ((TI->getOpcode() == Instruction::Sub &&
981 FI->getOpcode() == Instruction::Add) ||
982 (TI->getOpcode() == Instruction::FSub &&
983 FI->getOpcode() == Instruction::FAdd)) {
984 AddOp = FI; SubOp = TI;
985 } else if ((FI->getOpcode() == Instruction::Sub &&
986 TI->getOpcode() == Instruction::Add) ||
987 (FI->getOpcode() == Instruction::FSub &&
988 TI->getOpcode() == Instruction::FAdd)) {
989 AddOp = TI; SubOp = FI;
993 Value *OtherAddOp = nullptr;
994 if (SubOp->getOperand(0) == AddOp->getOperand(0)) {
995 OtherAddOp = AddOp->getOperand(1);
996 } else if (SubOp->getOperand(0) == AddOp->getOperand(1)) {
997 OtherAddOp = AddOp->getOperand(0);
1001 // So at this point we know we have (Y -> OtherAddOp):
1002 // select C, (add X, Y), (sub X, Z)
1003 Value *NegVal; // Compute -Z
1004 if (SI.getType()->isFPOrFPVectorTy()) {
1005 NegVal = Builder->CreateFNeg(SubOp->getOperand(1));
1006 if (Instruction *NegInst = dyn_cast<Instruction>(NegVal)) {
1007 FastMathFlags Flags = AddOp->getFastMathFlags();
1008 Flags &= SubOp->getFastMathFlags();
1009 NegInst->setFastMathFlags(Flags);
1012 NegVal = Builder->CreateNeg(SubOp->getOperand(1));
1015 Value *NewTrueOp = OtherAddOp;
1016 Value *NewFalseOp = NegVal;
1018 std::swap(NewTrueOp, NewFalseOp);
1020 Builder->CreateSelect(CondVal, NewTrueOp,
1021 NewFalseOp, SI.getName() + ".p");
1023 if (SI.getType()->isFPOrFPVectorTy()) {
1025 BinaryOperator::CreateFAdd(SubOp->getOperand(0), NewSel);
1027 FastMathFlags Flags = AddOp->getFastMathFlags();
1028 Flags &= SubOp->getFastMathFlags();
1029 RI->setFastMathFlags(Flags);
1032 return BinaryOperator::CreateAdd(SubOp->getOperand(0), NewSel);
1037 // See if we can fold the select into one of our operands.
1038 if (SI.getType()->isIntegerTy()) {
1039 if (Instruction *FoldI = FoldSelectIntoOp(SI, TrueVal, FalseVal))
1042 // MAX(MAX(a, b), a) -> MAX(a, b)
1043 // MIN(MIN(a, b), a) -> MIN(a, b)
1044 // MAX(MIN(a, b), a) -> a
1045 // MIN(MAX(a, b), a) -> a
1046 Value *LHS, *RHS, *LHS2, *RHS2;
1047 if (SelectPatternFlavor SPF = MatchSelectPattern(&SI, LHS, RHS)) {
1048 if (SelectPatternFlavor SPF2 = MatchSelectPattern(LHS, LHS2, RHS2))
1049 if (Instruction *R = FoldSPFofSPF(cast<Instruction>(LHS),SPF2,LHS2,RHS2,
1052 if (SelectPatternFlavor SPF2 = MatchSelectPattern(RHS, LHS2, RHS2))
1053 if (Instruction *R = FoldSPFofSPF(cast<Instruction>(RHS),SPF2,LHS2,RHS2,
1059 // ABS(-X) -> ABS(X)
1062 // See if we can fold the select into a phi node if the condition is a select.
1063 if (isa<PHINode>(SI.getCondition()))
1064 // The true/false values have to be live in the PHI predecessor's blocks.
1065 if (CanSelectOperandBeMappingIntoPredBlock(TrueVal, SI) &&
1066 CanSelectOperandBeMappingIntoPredBlock(FalseVal, SI))
1067 if (Instruction *NV = FoldOpIntoPhi(SI))
1070 if (SelectInst *TrueSI = dyn_cast<SelectInst>(TrueVal)) {
1071 if (TrueSI->getCondition() == CondVal) {
1072 if (SI.getTrueValue() == TrueSI->getTrueValue())
1074 SI.setOperand(1, TrueSI->getTrueValue());
1078 if (SelectInst *FalseSI = dyn_cast<SelectInst>(FalseVal)) {
1079 if (FalseSI->getCondition() == CondVal) {
1080 if (SI.getFalseValue() == FalseSI->getFalseValue())
1082 SI.setOperand(2, FalseSI->getFalseValue());
1087 if (BinaryOperator::isNot(CondVal)) {
1088 SI.setOperand(0, BinaryOperator::getNotArgument(CondVal));
1089 SI.setOperand(1, FalseVal);
1090 SI.setOperand(2, TrueVal);
1094 if (VectorType* VecTy = dyn_cast<VectorType>(SI.getType())) {
1095 unsigned VWidth = VecTy->getNumElements();
1096 APInt UndefElts(VWidth, 0);
1097 APInt AllOnesEltMask(APInt::getAllOnesValue(VWidth));
1098 if (Value *V = SimplifyDemandedVectorElts(&SI, AllOnesEltMask, UndefElts)) {
1100 return ReplaceInstUsesWith(SI, V);
1104 if (isa<ConstantAggregateZero>(CondVal)) {
1105 return ReplaceInstUsesWith(SI, FalseVal);