1 //===- InstCombineSelect.cpp ----------------------------------------------===//
3 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
4 // See https://llvm.org/LICENSE.txt for license information.
5 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
7 //===----------------------------------------------------------------------===//
9 // This file implements the visitSelect function.
11 //===----------------------------------------------------------------------===//
13 #include "InstCombineInternal.h"
14 #include "llvm/ADT/APInt.h"
15 #include "llvm/ADT/Optional.h"
16 #include "llvm/ADT/STLExtras.h"
17 #include "llvm/ADT/SmallVector.h"
18 #include "llvm/Analysis/AssumptionCache.h"
19 #include "llvm/Analysis/CmpInstAnalysis.h"
20 #include "llvm/Analysis/InstructionSimplify.h"
21 #include "llvm/Analysis/ValueTracking.h"
22 #include "llvm/IR/BasicBlock.h"
23 #include "llvm/IR/Constant.h"
24 #include "llvm/IR/Constants.h"
25 #include "llvm/IR/DerivedTypes.h"
26 #include "llvm/IR/IRBuilder.h"
27 #include "llvm/IR/InstrTypes.h"
28 #include "llvm/IR/Instruction.h"
29 #include "llvm/IR/Instructions.h"
30 #include "llvm/IR/IntrinsicInst.h"
31 #include "llvm/IR/Intrinsics.h"
32 #include "llvm/IR/Operator.h"
33 #include "llvm/IR/PatternMatch.h"
34 #include "llvm/IR/Type.h"
35 #include "llvm/IR/User.h"
36 #include "llvm/IR/Value.h"
37 #include "llvm/Support/Casting.h"
38 #include "llvm/Support/ErrorHandling.h"
39 #include "llvm/Support/KnownBits.h"
40 #include "llvm/Transforms/InstCombine/InstCombineWorklist.h"
45 using namespace PatternMatch;
47 #define DEBUG_TYPE "instcombine"
49 static Value *createMinMax(InstCombiner::BuilderTy &Builder,
50 SelectPatternFlavor SPF, Value *A, Value *B) {
51 CmpInst::Predicate Pred = getMinMaxPred(SPF);
52 assert(CmpInst::isIntPredicate(Pred) && "Expected integer predicate");
53 return Builder.CreateSelect(Builder.CreateICmp(Pred, A, B), A, B);
56 /// Replace a select operand based on an equality comparison with the identity
57 /// constant of a binop.
58 static Instruction *foldSelectBinOpIdentity(SelectInst &Sel,
59 const TargetLibraryInfo &TLI) {
60 // The select condition must be an equality compare with a constant operand.
63 CmpInst::Predicate Pred;
64 if (!match(Sel.getCondition(), m_Cmp(Pred, m_Value(X), m_Constant(C))))
68 if (ICmpInst::isEquality(Pred))
69 IsEq = Pred == ICmpInst::ICMP_EQ;
70 else if (Pred == FCmpInst::FCMP_OEQ)
72 else if (Pred == FCmpInst::FCMP_UNE)
77 // A select operand must be a binop.
79 if (!match(Sel.getOperand(IsEq ? 1 : 2), m_BinOp(BO)))
82 // The compare constant must be the identity constant for that binop.
83 // If this a floating-point compare with 0.0, any zero constant will do.
84 Type *Ty = BO->getType();
85 Constant *IdC = ConstantExpr::getBinOpIdentity(BO->getOpcode(), Ty, true);
87 if (!IdC || !CmpInst::isFPPredicate(Pred))
89 if (!match(IdC, m_AnyZeroFP()) || !match(C, m_AnyZeroFP()))
93 // Last, match the compare variable operand with a binop operand.
95 if (!BO->isCommutative() && !match(BO, m_BinOp(m_Value(Y), m_Specific(X))))
97 if (!match(BO, m_c_BinOp(m_Value(Y), m_Specific(X))))
100 // +0.0 compares equal to -0.0, and so it does not behave as required for this
101 // transform. Bail out if we can not exclude that possibility.
102 if (isa<FPMathOperator>(BO))
103 if (!BO->hasNoSignedZeros() && !CannotBeNegativeZero(Y, &TLI))
107 // S = { select (cmp eq X, C), BO, ? } or { select (cmp ne X, C), ?, BO }
109 // S = { select (cmp eq X, C), Y, ? } or { select (cmp ne X, C), ?, Y }
110 Sel.setOperand(IsEq ? 1 : 2, Y);
115 /// select (icmp eq (and X, C1)), TC, FC
116 /// iff C1 is a power 2 and the difference between TC and FC is a power-of-2.
117 /// To something like:
118 /// (shr (and (X, C1)), (log2(C1) - log2(TC-FC))) + FC
120 /// (shl (and (X, C1)), (log2(TC-FC) - log2(C1))) + FC
121 /// With some variations depending if FC is larger than TC, or the shift
122 /// isn't needed, or the bit widths don't match.
123 static Value *foldSelectICmpAnd(SelectInst &Sel, ICmpInst *Cmp,
124 InstCombiner::BuilderTy &Builder) {
125 const APInt *SelTC, *SelFC;
126 if (!match(Sel.getTrueValue(), m_APInt(SelTC)) ||
127 !match(Sel.getFalseValue(), m_APInt(SelFC)))
130 // If this is a vector select, we need a vector compare.
131 Type *SelType = Sel.getType();
132 if (SelType->isVectorTy() != Cmp->getType()->isVectorTy())
137 bool CreateAnd = false;
138 ICmpInst::Predicate Pred = Cmp->getPredicate();
139 if (ICmpInst::isEquality(Pred)) {
140 if (!match(Cmp->getOperand(1), m_Zero()))
143 V = Cmp->getOperand(0);
145 if (!match(V, m_And(m_Value(), m_Power2(AndRHS))))
149 } else if (decomposeBitTestICmp(Cmp->getOperand(0), Cmp->getOperand(1),
151 assert(ICmpInst::isEquality(Pred) && "Not equality test?");
152 if (!AndMask.isPowerOf2())
160 // In general, when both constants are non-zero, we would need an offset to
161 // replace the select. This would require more instructions than we started
162 // with. But there's one special-case that we handle here because it can
163 // simplify/reduce the instructions.
166 if (!TC.isNullValue() && !FC.isNullValue()) {
167 // If the select constants differ by exactly one bit and that's the same
168 // bit that is masked and checked by the select condition, the select can
169 // be replaced by bitwise logic to set/clear one bit of the constant result.
170 if (TC.getBitWidth() != AndMask.getBitWidth() || (TC ^ FC) != AndMask)
173 // If we have to create an 'and', then we must kill the cmp to not
174 // increase the instruction count.
175 if (!Cmp->hasOneUse())
177 V = Builder.CreateAnd(V, ConstantInt::get(SelType, AndMask));
179 bool ExtraBitInTC = TC.ugt(FC);
180 if (Pred == ICmpInst::ICMP_EQ) {
181 // If the masked bit in V is clear, clear or set the bit in the result:
182 // (V & AndMaskC) == 0 ? TC : FC --> (V & AndMaskC) ^ TC
183 // (V & AndMaskC) == 0 ? TC : FC --> (V & AndMaskC) | TC
184 Constant *C = ConstantInt::get(SelType, TC);
185 return ExtraBitInTC ? Builder.CreateXor(V, C) : Builder.CreateOr(V, C);
187 if (Pred == ICmpInst::ICMP_NE) {
188 // If the masked bit in V is set, set or clear the bit in the result:
189 // (V & AndMaskC) != 0 ? TC : FC --> (V & AndMaskC) | FC
190 // (V & AndMaskC) != 0 ? TC : FC --> (V & AndMaskC) ^ FC
191 Constant *C = ConstantInt::get(SelType, FC);
192 return ExtraBitInTC ? Builder.CreateOr(V, C) : Builder.CreateXor(V, C);
194 llvm_unreachable("Only expecting equality predicates");
197 // Make sure one of the select arms is a power-of-2.
198 if (!TC.isPowerOf2() && !FC.isPowerOf2())
201 // Determine which shift is needed to transform result of the 'and' into the
203 const APInt &ValC = !TC.isNullValue() ? TC : FC;
204 unsigned ValZeros = ValC.logBase2();
205 unsigned AndZeros = AndMask.logBase2();
207 // Insert the 'and' instruction on the input to the truncate.
209 V = Builder.CreateAnd(V, ConstantInt::get(V->getType(), AndMask));
211 // If types don't match, we can still convert the select by introducing a zext
212 // or a trunc of the 'and'.
213 if (ValZeros > AndZeros) {
214 V = Builder.CreateZExtOrTrunc(V, SelType);
215 V = Builder.CreateShl(V, ValZeros - AndZeros);
216 } else if (ValZeros < AndZeros) {
217 V = Builder.CreateLShr(V, AndZeros - ValZeros);
218 V = Builder.CreateZExtOrTrunc(V, SelType);
220 V = Builder.CreateZExtOrTrunc(V, SelType);
223 // Okay, now we know that everything is set up, we just don't know whether we
224 // have a icmp_ne or icmp_eq and whether the true or false val is the zero.
225 bool ShouldNotVal = !TC.isNullValue();
226 ShouldNotVal ^= Pred == ICmpInst::ICMP_NE;
228 V = Builder.CreateXor(V, ValC);
233 /// We want to turn code that looks like this:
235 /// %D = select %cond, %C, %A
237 /// %C = select %cond, %B, 0
240 /// Assuming that the specified instruction is an operand to the select, return
241 /// a bitmask indicating which operands of this instruction are foldable if they
242 /// equal the other incoming value of the select.
243 static unsigned getSelectFoldableOperands(BinaryOperator *I) {
244 switch (I->getOpcode()) {
245 case Instruction::Add:
246 case Instruction::Mul:
247 case Instruction::And:
248 case Instruction::Or:
249 case Instruction::Xor:
250 return 3; // Can fold through either operand.
251 case Instruction::Sub: // Can only fold on the amount subtracted.
252 case Instruction::Shl: // Can only fold on the shift amount.
253 case Instruction::LShr:
254 case Instruction::AShr:
257 return 0; // Cannot fold
261 /// For the same transformation as the previous function, return the identity
262 /// constant that goes into the select.
263 static APInt getSelectFoldableConstant(BinaryOperator *I) {
264 switch (I->getOpcode()) {
265 default: llvm_unreachable("This cannot happen!");
266 case Instruction::Add:
267 case Instruction::Sub:
268 case Instruction::Or:
269 case Instruction::Xor:
270 case Instruction::Shl:
271 case Instruction::LShr:
272 case Instruction::AShr:
273 return APInt::getNullValue(I->getType()->getScalarSizeInBits());
274 case Instruction::And:
275 return APInt::getAllOnesValue(I->getType()->getScalarSizeInBits());
276 case Instruction::Mul:
277 return APInt(I->getType()->getScalarSizeInBits(), 1);
281 /// We have (select c, TI, FI), and we know that TI and FI have the same opcode.
282 Instruction *InstCombiner::foldSelectOpOp(SelectInst &SI, Instruction *TI,
284 // Don't break up min/max patterns. The hasOneUse checks below prevent that
285 // for most cases, but vector min/max with bitcasts can be transformed. If the
286 // one-use restrictions are eased for other patterns, we still don't want to
287 // obfuscate min/max.
288 if ((match(&SI, m_SMin(m_Value(), m_Value())) ||
289 match(&SI, m_SMax(m_Value(), m_Value())) ||
290 match(&SI, m_UMin(m_Value(), m_Value())) ||
291 match(&SI, m_UMax(m_Value(), m_Value()))))
294 // If this is a cast from the same type, merge.
295 Value *Cond = SI.getCondition();
296 Type *CondTy = Cond->getType();
297 if (TI->getNumOperands() == 1 && TI->isCast()) {
298 Type *FIOpndTy = FI->getOperand(0)->getType();
299 if (TI->getOperand(0)->getType() != FIOpndTy)
302 // The select condition may be a vector. We may only change the operand
303 // type if the vector width remains the same (and matches the condition).
304 if (CondTy->isVectorTy()) {
305 if (!FIOpndTy->isVectorTy())
307 if (CondTy->getVectorNumElements() != FIOpndTy->getVectorNumElements())
310 // TODO: If the backend knew how to deal with casts better, we could
311 // remove this limitation. For now, there's too much potential to create
312 // worse codegen by promoting the select ahead of size-altering casts
315 // Note that ValueTracking's matchSelectPattern() looks through casts
316 // without checking 'hasOneUse' when it matches min/max patterns, so this
317 // transform may end up happening anyway.
318 if (TI->getOpcode() != Instruction::BitCast &&
319 (!TI->hasOneUse() || !FI->hasOneUse()))
321 } else if (!TI->hasOneUse() || !FI->hasOneUse()) {
322 // TODO: The one-use restrictions for a scalar select could be eased if
323 // the fold of a select in visitLoadInst() was enhanced to match a pattern
324 // that includes a cast.
328 // Fold this by inserting a select from the input values.
330 Builder.CreateSelect(Cond, TI->getOperand(0), FI->getOperand(0),
331 SI.getName() + ".v", &SI);
332 return CastInst::Create(Instruction::CastOps(TI->getOpcode()), NewSI,
336 // Cond ? -X : -Y --> -(Cond ? X : Y)
338 if (match(TI, m_FNeg(m_Value(X))) && match(FI, m_FNeg(m_Value(Y))) &&
339 (TI->hasOneUse() || FI->hasOneUse())) {
340 Value *NewSel = Builder.CreateSelect(Cond, X, Y, SI.getName() + ".v", &SI);
341 // TODO: Remove the hack for the binop form when the unary op is optimized
342 // properly with all IR passes.
343 if (TI->getOpcode() != Instruction::FNeg)
344 return BinaryOperator::CreateFNegFMF(NewSel, cast<BinaryOperator>(TI));
345 return UnaryOperator::CreateFNeg(NewSel);
348 // Only handle binary operators (including two-operand getelementptr) with
349 // one-use here. As with the cast case above, it may be possible to relax the
350 // one-use constraint, but that needs be examined carefully since it may not
351 // reduce the total number of instructions.
352 if (TI->getNumOperands() != 2 || FI->getNumOperands() != 2 ||
353 (!isa<BinaryOperator>(TI) && !isa<GetElementPtrInst>(TI)) ||
354 !TI->hasOneUse() || !FI->hasOneUse())
357 // Figure out if the operations have any operands in common.
358 Value *MatchOp, *OtherOpT, *OtherOpF;
360 if (TI->getOperand(0) == FI->getOperand(0)) {
361 MatchOp = TI->getOperand(0);
362 OtherOpT = TI->getOperand(1);
363 OtherOpF = FI->getOperand(1);
364 MatchIsOpZero = true;
365 } else if (TI->getOperand(1) == FI->getOperand(1)) {
366 MatchOp = TI->getOperand(1);
367 OtherOpT = TI->getOperand(0);
368 OtherOpF = FI->getOperand(0);
369 MatchIsOpZero = false;
370 } else if (!TI->isCommutative()) {
372 } else if (TI->getOperand(0) == FI->getOperand(1)) {
373 MatchOp = TI->getOperand(0);
374 OtherOpT = TI->getOperand(1);
375 OtherOpF = FI->getOperand(0);
376 MatchIsOpZero = true;
377 } else if (TI->getOperand(1) == FI->getOperand(0)) {
378 MatchOp = TI->getOperand(1);
379 OtherOpT = TI->getOperand(0);
380 OtherOpF = FI->getOperand(1);
381 MatchIsOpZero = true;
386 // If the select condition is a vector, the operands of the original select's
387 // operands also must be vectors. This may not be the case for getelementptr
389 if (CondTy->isVectorTy() && (!OtherOpT->getType()->isVectorTy() ||
390 !OtherOpF->getType()->isVectorTy()))
393 // If we reach here, they do have operations in common.
394 Value *NewSI = Builder.CreateSelect(Cond, OtherOpT, OtherOpF,
395 SI.getName() + ".v", &SI);
396 Value *Op0 = MatchIsOpZero ? MatchOp : NewSI;
397 Value *Op1 = MatchIsOpZero ? NewSI : MatchOp;
398 if (auto *BO = dyn_cast<BinaryOperator>(TI)) {
399 BinaryOperator *NewBO = BinaryOperator::Create(BO->getOpcode(), Op0, Op1);
400 NewBO->copyIRFlags(TI);
401 NewBO->andIRFlags(FI);
404 if (auto *TGEP = dyn_cast<GetElementPtrInst>(TI)) {
405 auto *FGEP = cast<GetElementPtrInst>(FI);
406 Type *ElementType = TGEP->getResultElementType();
407 return TGEP->isInBounds() && FGEP->isInBounds()
408 ? GetElementPtrInst::CreateInBounds(ElementType, Op0, {Op1})
409 : GetElementPtrInst::Create(ElementType, Op0, {Op1});
411 llvm_unreachable("Expected BinaryOperator or GEP");
415 static bool isSelect01(const APInt &C1I, const APInt &C2I) {
416 if (!C1I.isNullValue() && !C2I.isNullValue()) // One side must be zero.
418 return C1I.isOneValue() || C1I.isAllOnesValue() ||
419 C2I.isOneValue() || C2I.isAllOnesValue();
422 /// Try to fold the select into one of the operands to allow further
424 Instruction *InstCombiner::foldSelectIntoOp(SelectInst &SI, Value *TrueVal,
426 // See the comment above GetSelectFoldableOperands for a description of the
427 // transformation we are doing here.
428 if (auto *TVI = dyn_cast<BinaryOperator>(TrueVal)) {
429 if (TVI->hasOneUse() && !isa<Constant>(FalseVal)) {
430 if (unsigned SFO = getSelectFoldableOperands(TVI)) {
431 unsigned OpToFold = 0;
432 if ((SFO & 1) && FalseVal == TVI->getOperand(0)) {
434 } else if ((SFO & 2) && FalseVal == TVI->getOperand(1)) {
439 APInt CI = getSelectFoldableConstant(TVI);
440 Value *OOp = TVI->getOperand(2-OpToFold);
441 // Avoid creating select between 2 constants unless it's selecting
442 // between 0, 1 and -1.
444 bool OOpIsAPInt = match(OOp, m_APInt(OOpC));
445 if (!isa<Constant>(OOp) || (OOpIsAPInt && isSelect01(CI, *OOpC))) {
446 Value *C = ConstantInt::get(OOp->getType(), CI);
447 Value *NewSel = Builder.CreateSelect(SI.getCondition(), OOp, C);
448 NewSel->takeName(TVI);
449 BinaryOperator *BO = BinaryOperator::Create(TVI->getOpcode(),
451 BO->copyIRFlags(TVI);
459 if (auto *FVI = dyn_cast<BinaryOperator>(FalseVal)) {
460 if (FVI->hasOneUse() && !isa<Constant>(TrueVal)) {
461 if (unsigned SFO = getSelectFoldableOperands(FVI)) {
462 unsigned OpToFold = 0;
463 if ((SFO & 1) && TrueVal == FVI->getOperand(0)) {
465 } else if ((SFO & 2) && TrueVal == FVI->getOperand(1)) {
470 APInt CI = getSelectFoldableConstant(FVI);
471 Value *OOp = FVI->getOperand(2-OpToFold);
472 // Avoid creating select between 2 constants unless it's selecting
473 // between 0, 1 and -1.
475 bool OOpIsAPInt = match(OOp, m_APInt(OOpC));
476 if (!isa<Constant>(OOp) || (OOpIsAPInt && isSelect01(CI, *OOpC))) {
477 Value *C = ConstantInt::get(OOp->getType(), CI);
478 Value *NewSel = Builder.CreateSelect(SI.getCondition(), C, OOp);
479 NewSel->takeName(FVI);
480 BinaryOperator *BO = BinaryOperator::Create(FVI->getOpcode(),
482 BO->copyIRFlags(FVI);
494 /// (select (icmp eq (and X, Y), 0), (and (lshr X, Z), 1), 1)
496 /// zext (icmp ne i32 (and X, (or Y, (shl 1, Z))), 0)
498 /// Z may be 0 if lshr is missing.
499 /// Worst-case scenario is that we will replace 5 instructions with 5 different
500 /// instructions, but we got rid of select.
501 static Instruction *foldSelectICmpAndAnd(Type *SelType, const ICmpInst *Cmp,
502 Value *TVal, Value *FVal,
503 InstCombiner::BuilderTy &Builder) {
504 if (!(Cmp->hasOneUse() && Cmp->getOperand(0)->hasOneUse() &&
505 Cmp->getPredicate() == ICmpInst::ICMP_EQ &&
506 match(Cmp->getOperand(1), m_Zero()) && match(FVal, m_One())))
509 // The TrueVal has general form of: and %B, 1
511 if (!match(TVal, m_OneUse(m_And(m_Value(B), m_One()))))
514 // Where %B may be optionally shifted: lshr %X, %Z.
516 const bool HasShift = match(B, m_OneUse(m_LShr(m_Value(X), m_Value(Z))));
521 if (!match(Cmp->getOperand(0), m_c_And(m_Specific(X), m_Value(Y))))
524 // ((X & Y) == 0) ? ((X >> Z) & 1) : 1 --> (X & (Y | (1 << Z))) != 0
525 // ((X & Y) == 0) ? (X & 1) : 1 --> (X & (Y | 1)) != 0
526 Constant *One = ConstantInt::get(SelType, 1);
527 Value *MaskB = HasShift ? Builder.CreateShl(One, Z) : One;
528 Value *FullMask = Builder.CreateOr(Y, MaskB);
529 Value *MaskedX = Builder.CreateAnd(X, FullMask);
530 Value *ICmpNeZero = Builder.CreateIsNotNull(MaskedX);
531 return new ZExtInst(ICmpNeZero, SelType);
535 /// (select (icmp sgt x, C), lshr (X, Y), ashr (X, Y)); iff C s>= -1
536 /// (select (icmp slt x, C), ashr (X, Y), lshr (X, Y)); iff C s>= 0
539 static Value *foldSelectICmpLshrAshr(const ICmpInst *IC, Value *TrueVal,
541 InstCombiner::BuilderTy &Builder) {
542 ICmpInst::Predicate Pred = IC->getPredicate();
543 Value *CmpLHS = IC->getOperand(0);
544 Value *CmpRHS = IC->getOperand(1);
545 if (!CmpRHS->getType()->isIntOrIntVectorTy())
549 unsigned Bitwidth = CmpRHS->getType()->getScalarSizeInBits();
550 if ((Pred != ICmpInst::ICMP_SGT ||
552 m_SpecificInt_ICMP(ICmpInst::ICMP_SGE, APInt(Bitwidth, -1)))) &&
553 (Pred != ICmpInst::ICMP_SLT ||
555 m_SpecificInt_ICMP(ICmpInst::ICMP_SGE, APInt(Bitwidth, 0)))))
558 // Canonicalize so that ashr is in FalseVal.
559 if (Pred == ICmpInst::ICMP_SLT)
560 std::swap(TrueVal, FalseVal);
562 if (match(TrueVal, m_LShr(m_Value(X), m_Value(Y))) &&
563 match(FalseVal, m_AShr(m_Specific(X), m_Specific(Y))) &&
564 match(CmpLHS, m_Specific(X))) {
565 const auto *Ashr = cast<Instruction>(FalseVal);
566 // if lshr is not exact and ashr is, this new ashr must not be exact.
567 bool IsExact = Ashr->isExact() && cast<Instruction>(TrueVal)->isExact();
568 return Builder.CreateAShr(X, Y, IC->getName(), IsExact);
575 /// (select (icmp eq (and X, C1), 0), Y, (or Y, C2))
577 /// (or (shl (and X, C1), C3), Y)
579 /// C1 and C2 are both powers of 2
581 /// C3 = Log(C2) - Log(C1)
583 /// This transform handles cases where:
584 /// 1. The icmp predicate is inverted
585 /// 2. The select operands are reversed
586 /// 3. The magnitude of C2 and C1 are flipped
587 static Value *foldSelectICmpAndOr(const ICmpInst *IC, Value *TrueVal,
589 InstCombiner::BuilderTy &Builder) {
590 // Only handle integer compares. Also, if this is a vector select, we need a
592 if (!TrueVal->getType()->isIntOrIntVectorTy() ||
593 TrueVal->getType()->isVectorTy() != IC->getType()->isVectorTy())
596 Value *CmpLHS = IC->getOperand(0);
597 Value *CmpRHS = IC->getOperand(1);
602 bool NeedAnd = false;
603 if (IC->isEquality()) {
604 if (!match(CmpRHS, m_Zero()))
608 if (!match(CmpLHS, m_And(m_Value(), m_Power2(C1))))
612 C1Log = C1->logBase2();
613 IsEqualZero = IC->getPredicate() == ICmpInst::ICMP_EQ;
614 } else if (IC->getPredicate() == ICmpInst::ICMP_SLT ||
615 IC->getPredicate() == ICmpInst::ICMP_SGT) {
616 // We also need to recognize (icmp slt (trunc (X)), 0) and
617 // (icmp sgt (trunc (X)), -1).
618 IsEqualZero = IC->getPredicate() == ICmpInst::ICMP_SGT;
619 if ((IsEqualZero && !match(CmpRHS, m_AllOnes())) ||
620 (!IsEqualZero && !match(CmpRHS, m_Zero())))
623 if (!match(CmpLHS, m_OneUse(m_Trunc(m_Value(V)))))
626 C1Log = CmpLHS->getType()->getScalarSizeInBits() - 1;
633 bool OrOnTrueVal = false;
634 bool OrOnFalseVal = match(FalseVal, m_Or(m_Specific(TrueVal), m_Power2(C2)));
636 OrOnTrueVal = match(TrueVal, m_Or(m_Specific(FalseVal), m_Power2(C2)));
638 if (!OrOnFalseVal && !OrOnTrueVal)
641 Value *Y = OrOnFalseVal ? TrueVal : FalseVal;
643 unsigned C2Log = C2->logBase2();
645 bool NeedXor = (!IsEqualZero && OrOnFalseVal) || (IsEqualZero && OrOnTrueVal);
646 bool NeedShift = C1Log != C2Log;
647 bool NeedZExtTrunc = Y->getType()->getScalarSizeInBits() !=
648 V->getType()->getScalarSizeInBits();
650 // Make sure we don't create more instructions than we save.
651 Value *Or = OrOnFalseVal ? FalseVal : TrueVal;
652 if ((NeedShift + NeedXor + NeedZExtTrunc) >
653 (IC->hasOneUse() + Or->hasOneUse()))
657 // Insert the AND instruction on the input to the truncate.
658 APInt C1 = APInt::getOneBitSet(V->getType()->getScalarSizeInBits(), C1Log);
659 V = Builder.CreateAnd(V, ConstantInt::get(V->getType(), C1));
663 V = Builder.CreateZExtOrTrunc(V, Y->getType());
664 V = Builder.CreateShl(V, C2Log - C1Log);
665 } else if (C1Log > C2Log) {
666 V = Builder.CreateLShr(V, C1Log - C2Log);
667 V = Builder.CreateZExtOrTrunc(V, Y->getType());
669 V = Builder.CreateZExtOrTrunc(V, Y->getType());
672 V = Builder.CreateXor(V, *C2);
674 return Builder.CreateOr(V, Y);
677 /// Transform patterns such as (a > b) ? a - b : 0 into usub.sat(a, b).
678 /// There are 8 commuted/swapped variants of this pattern.
679 /// TODO: Also support a - UMIN(a,b) patterns.
680 static Value *canonicalizeSaturatedSubtract(const ICmpInst *ICI,
681 const Value *TrueVal,
682 const Value *FalseVal,
683 InstCombiner::BuilderTy &Builder) {
684 ICmpInst::Predicate Pred = ICI->getPredicate();
685 if (!ICmpInst::isUnsigned(Pred))
688 // (b > a) ? 0 : a - b -> (b <= a) ? a - b : 0
689 if (match(TrueVal, m_Zero())) {
690 Pred = ICmpInst::getInversePredicate(Pred);
691 std::swap(TrueVal, FalseVal);
693 if (!match(FalseVal, m_Zero()))
696 Value *A = ICI->getOperand(0);
697 Value *B = ICI->getOperand(1);
698 if (Pred == ICmpInst::ICMP_ULE || Pred == ICmpInst::ICMP_ULT) {
699 // (b < a) ? a - b : 0 -> (a > b) ? a - b : 0
701 Pred = ICmpInst::getSwappedPredicate(Pred);
704 assert((Pred == ICmpInst::ICMP_UGE || Pred == ICmpInst::ICMP_UGT) &&
705 "Unexpected isUnsigned predicate!");
707 // Account for swapped form of subtraction: ((a > b) ? b - a : 0).
708 bool IsNegative = false;
709 if (match(TrueVal, m_Sub(m_Specific(B), m_Specific(A))))
711 else if (!match(TrueVal, m_Sub(m_Specific(A), m_Specific(B))))
714 // If sub is used anywhere else, we wouldn't be able to eliminate it
716 if (!TrueVal->hasOneUse())
719 // (a > b) ? a - b : 0 -> usub.sat(a, b)
720 // (a > b) ? b - a : 0 -> -usub.sat(a, b)
721 Value *Result = Builder.CreateBinaryIntrinsic(Intrinsic::usub_sat, A, B);
723 Result = Builder.CreateNeg(Result);
727 static Value *canonicalizeSaturatedAdd(ICmpInst *Cmp, Value *TVal, Value *FVal,
728 InstCombiner::BuilderTy &Builder) {
729 if (!Cmp->hasOneUse())
732 // Match unsigned saturated add with constant.
733 Value *Cmp0 = Cmp->getOperand(0);
734 Value *Cmp1 = Cmp->getOperand(1);
735 ICmpInst::Predicate Pred = Cmp->getPredicate();
737 const APInt *C, *CmpC;
738 if (Pred == ICmpInst::ICMP_ULT &&
739 match(TVal, m_Add(m_Value(X), m_APInt(C))) && X == Cmp0 &&
740 match(FVal, m_AllOnes()) && match(Cmp1, m_APInt(CmpC)) && *CmpC == ~*C) {
741 // (X u< ~C) ? (X + C) : -1 --> uadd.sat(X, C)
742 return Builder.CreateBinaryIntrinsic(
743 Intrinsic::uadd_sat, X, ConstantInt::get(X->getType(), *C));
746 // Match unsigned saturated add of 2 variables with an unnecessary 'not'.
747 // There are 8 commuted variants.
748 // Canonicalize -1 (saturated result) to true value of the select. Just
749 // swapping the compare operands is legal, because the selected value is the
750 // same in case of equality, so we can interchange u< and u<=.
751 if (match(FVal, m_AllOnes())) {
752 std::swap(TVal, FVal);
753 std::swap(Cmp0, Cmp1);
755 if (!match(TVal, m_AllOnes()))
758 // Canonicalize predicate to 'ULT'.
759 if (Pred == ICmpInst::ICMP_UGT) {
760 Pred = ICmpInst::ICMP_ULT;
761 std::swap(Cmp0, Cmp1);
763 if (Pred != ICmpInst::ICMP_ULT)
766 // Match unsigned saturated add of 2 variables with an unnecessary 'not'.
768 if (match(Cmp0, m_Not(m_Value(X))) &&
769 match(FVal, m_c_Add(m_Specific(X), m_Value(Y))) && Y == Cmp1) {
770 // (~X u< Y) ? -1 : (X + Y) --> uadd.sat(X, Y)
771 // (~X u< Y) ? -1 : (Y + X) --> uadd.sat(X, Y)
772 return Builder.CreateBinaryIntrinsic(Intrinsic::uadd_sat, X, Y);
774 // The 'not' op may be included in the sum but not the compare.
777 if (match(FVal, m_c_Add(m_Not(m_Specific(X)), m_Specific(Y)))) {
778 // (X u< Y) ? -1 : (~X + Y) --> uadd.sat(~X, Y)
779 // (X u< Y) ? -1 : (Y + ~X) --> uadd.sat(Y, ~X)
780 BinaryOperator *BO = cast<BinaryOperator>(FVal);
781 return Builder.CreateBinaryIntrinsic(
782 Intrinsic::uadd_sat, BO->getOperand(0), BO->getOperand(1));
788 /// Attempt to fold a cttz/ctlz followed by a icmp plus select into a single
789 /// call to cttz/ctlz with flag 'is_zero_undef' cleared.
791 /// For example, we can fold the following code sequence:
793 /// %0 = tail call i32 @llvm.cttz.i32(i32 %x, i1 true)
794 /// %1 = icmp ne i32 %x, 0
795 /// %2 = select i1 %1, i32 %0, i32 32
799 /// %0 = tail call i32 @llvm.cttz.i32(i32 %x, i1 false)
800 static Value *foldSelectCttzCtlz(ICmpInst *ICI, Value *TrueVal, Value *FalseVal,
801 InstCombiner::BuilderTy &Builder) {
802 ICmpInst::Predicate Pred = ICI->getPredicate();
803 Value *CmpLHS = ICI->getOperand(0);
804 Value *CmpRHS = ICI->getOperand(1);
806 // Check if the condition value compares a value for equality against zero.
807 if (!ICI->isEquality() || !match(CmpRHS, m_Zero()))
810 Value *Count = FalseVal;
811 Value *ValueOnZero = TrueVal;
812 if (Pred == ICmpInst::ICMP_NE)
813 std::swap(Count, ValueOnZero);
815 // Skip zero extend/truncate.
817 if (match(Count, m_ZExt(m_Value(V))) ||
818 match(Count, m_Trunc(m_Value(V))))
821 // Check that 'Count' is a call to intrinsic cttz/ctlz. Also check that the
822 // input to the cttz/ctlz is used as LHS for the compare instruction.
823 if (!match(Count, m_Intrinsic<Intrinsic::cttz>(m_Specific(CmpLHS))) &&
824 !match(Count, m_Intrinsic<Intrinsic::ctlz>(m_Specific(CmpLHS))))
827 IntrinsicInst *II = cast<IntrinsicInst>(Count);
829 // Check if the value propagated on zero is a constant number equal to the
830 // sizeof in bits of 'Count'.
831 unsigned SizeOfInBits = Count->getType()->getScalarSizeInBits();
832 if (match(ValueOnZero, m_SpecificInt(SizeOfInBits))) {
833 // Explicitly clear the 'undef_on_zero' flag.
834 IntrinsicInst *NewI = cast<IntrinsicInst>(II->clone());
835 NewI->setArgOperand(1, ConstantInt::getFalse(NewI->getContext()));
836 Builder.Insert(NewI);
837 return Builder.CreateZExtOrTrunc(NewI, ValueOnZero->getType());
840 // If the ValueOnZero is not the bitwidth, we can at least make use of the
841 // fact that the cttz/ctlz result will not be used if the input is zero, so
842 // it's okay to relax it to undef for that case.
843 if (II->hasOneUse() && !match(II->getArgOperand(1), m_One()))
844 II->setArgOperand(1, ConstantInt::getTrue(II->getContext()));
849 /// Return true if we find and adjust an icmp+select pattern where the compare
850 /// is with a constant that can be incremented or decremented to match the
851 /// minimum or maximum idiom.
852 static bool adjustMinMax(SelectInst &Sel, ICmpInst &Cmp) {
853 ICmpInst::Predicate Pred = Cmp.getPredicate();
854 Value *CmpLHS = Cmp.getOperand(0);
855 Value *CmpRHS = Cmp.getOperand(1);
856 Value *TrueVal = Sel.getTrueValue();
857 Value *FalseVal = Sel.getFalseValue();
859 // We may move or edit the compare, so make sure the select is the only user.
861 if (!Cmp.hasOneUse() || !match(CmpRHS, m_APInt(CmpC)))
864 // These transforms only work for selects of integers or vector selects of
866 Type *SelTy = Sel.getType();
867 auto *SelEltTy = dyn_cast<IntegerType>(SelTy->getScalarType());
868 if (!SelEltTy || SelTy->isVectorTy() != Cmp.getType()->isVectorTy())
871 Constant *AdjustedRHS;
872 if (Pred == ICmpInst::ICMP_UGT || Pred == ICmpInst::ICMP_SGT)
873 AdjustedRHS = ConstantInt::get(CmpRHS->getType(), *CmpC + 1);
874 else if (Pred == ICmpInst::ICMP_ULT || Pred == ICmpInst::ICMP_SLT)
875 AdjustedRHS = ConstantInt::get(CmpRHS->getType(), *CmpC - 1);
879 // X > C ? X : C+1 --> X < C+1 ? C+1 : X
880 // X < C ? X : C-1 --> X > C-1 ? C-1 : X
881 if ((CmpLHS == TrueVal && AdjustedRHS == FalseVal) ||
882 (CmpLHS == FalseVal && AdjustedRHS == TrueVal)) {
883 ; // Nothing to do here. Values match without any sign/zero extension.
885 // Types do not match. Instead of calculating this with mixed types, promote
886 // all to the larger type. This enables scalar evolution to analyze this
888 else if (CmpRHS->getType()->getScalarSizeInBits() < SelEltTy->getBitWidth()) {
889 Constant *SextRHS = ConstantExpr::getSExt(AdjustedRHS, SelTy);
891 // X = sext x; x >s c ? X : C+1 --> X = sext x; X <s C+1 ? C+1 : X
892 // X = sext x; x <s c ? X : C-1 --> X = sext x; X >s C-1 ? C-1 : X
893 // X = sext x; x >u c ? X : C+1 --> X = sext x; X <u C+1 ? C+1 : X
894 // X = sext x; x <u c ? X : C-1 --> X = sext x; X >u C-1 ? C-1 : X
895 if (match(TrueVal, m_SExt(m_Specific(CmpLHS))) && SextRHS == FalseVal) {
897 AdjustedRHS = SextRHS;
898 } else if (match(FalseVal, m_SExt(m_Specific(CmpLHS))) &&
899 SextRHS == TrueVal) {
901 AdjustedRHS = SextRHS;
902 } else if (Cmp.isUnsigned()) {
903 Constant *ZextRHS = ConstantExpr::getZExt(AdjustedRHS, SelTy);
904 // X = zext x; x >u c ? X : C+1 --> X = zext x; X <u C+1 ? C+1 : X
905 // X = zext x; x <u c ? X : C-1 --> X = zext x; X >u C-1 ? C-1 : X
906 // zext + signed compare cannot be changed:
907 // 0xff <s 0x00, but 0x00ff >s 0x0000
908 if (match(TrueVal, m_ZExt(m_Specific(CmpLHS))) && ZextRHS == FalseVal) {
910 AdjustedRHS = ZextRHS;
911 } else if (match(FalseVal, m_ZExt(m_Specific(CmpLHS))) &&
912 ZextRHS == TrueVal) {
914 AdjustedRHS = ZextRHS;
925 Pred = ICmpInst::getSwappedPredicate(Pred);
926 CmpRHS = AdjustedRHS;
927 std::swap(FalseVal, TrueVal);
928 Cmp.setPredicate(Pred);
929 Cmp.setOperand(0, CmpLHS);
930 Cmp.setOperand(1, CmpRHS);
931 Sel.setOperand(1, TrueVal);
932 Sel.setOperand(2, FalseVal);
933 Sel.swapProfMetadata();
935 // Move the compare instruction right before the select instruction. Otherwise
936 // the sext/zext value may be defined after the compare instruction uses it.
937 Cmp.moveBefore(&Sel);
942 /// If this is an integer min/max (icmp + select) with a constant operand,
943 /// create the canonical icmp for the min/max operation and canonicalize the
944 /// constant to the 'false' operand of the select:
945 /// select (icmp Pred X, C1), C2, X --> select (icmp Pred' X, C2), X, C2
946 /// Note: if C1 != C2, this will change the icmp constant to the existing
947 /// constant operand of the select.
949 canonicalizeMinMaxWithConstant(SelectInst &Sel, ICmpInst &Cmp,
950 InstCombiner::BuilderTy &Builder) {
951 if (!Cmp.hasOneUse() || !isa<Constant>(Cmp.getOperand(1)))
954 // Canonicalize the compare predicate based on whether we have min or max.
956 SelectPatternResult SPR = matchSelectPattern(&Sel, LHS, RHS);
957 if (!SelectPatternResult::isMinOrMax(SPR.Flavor))
960 // Is this already canonical?
961 ICmpInst::Predicate CanonicalPred = getMinMaxPred(SPR.Flavor);
962 if (Cmp.getOperand(0) == LHS && Cmp.getOperand(1) == RHS &&
963 Cmp.getPredicate() == CanonicalPred)
966 // Create the canonical compare and plug it into the select.
967 Sel.setCondition(Builder.CreateICmp(CanonicalPred, LHS, RHS));
969 // If the select operands did not change, we're done.
970 if (Sel.getTrueValue() == LHS && Sel.getFalseValue() == RHS)
973 // If we are swapping the select operands, swap the metadata too.
974 assert(Sel.getTrueValue() == RHS && Sel.getFalseValue() == LHS &&
975 "Unexpected results from matchSelectPattern");
976 Sel.setTrueValue(LHS);
977 Sel.setFalseValue(RHS);
978 Sel.swapProfMetadata();
982 /// There are many select variants for each of ABS/NABS.
983 /// In matchSelectPattern(), there are different compare constants, compare
984 /// predicates/operands and select operands.
985 /// In isKnownNegation(), there are different formats of negated operands.
986 /// Canonicalize all these variants to 1 pattern.
987 /// This makes CSE more likely.
988 static Instruction *canonicalizeAbsNabs(SelectInst &Sel, ICmpInst &Cmp,
989 InstCombiner::BuilderTy &Builder) {
990 if (!Cmp.hasOneUse() || !isa<Constant>(Cmp.getOperand(1)))
993 // Choose a sign-bit check for the compare (likely simpler for codegen).
994 // ABS: (X <s 0) ? -X : X
995 // NABS: (X <s 0) ? X : -X
997 SelectPatternFlavor SPF = matchSelectPattern(&Sel, LHS, RHS).Flavor;
998 if (SPF != SelectPatternFlavor::SPF_ABS &&
999 SPF != SelectPatternFlavor::SPF_NABS)
1002 Value *TVal = Sel.getTrueValue();
1003 Value *FVal = Sel.getFalseValue();
1004 assert(isKnownNegation(TVal, FVal) &&
1005 "Unexpected result from matchSelectPattern");
1007 // The compare may use the negated abs()/nabs() operand, or it may use
1008 // negation in non-canonical form such as: sub A, B.
1009 bool CmpUsesNegatedOp = match(Cmp.getOperand(0), m_Neg(m_Specific(TVal))) ||
1010 match(Cmp.getOperand(0), m_Neg(m_Specific(FVal)));
1012 bool CmpCanonicalized = !CmpUsesNegatedOp &&
1013 match(Cmp.getOperand(1), m_ZeroInt()) &&
1014 Cmp.getPredicate() == ICmpInst::ICMP_SLT;
1015 bool RHSCanonicalized = match(RHS, m_Neg(m_Specific(LHS)));
1017 // Is this already canonical?
1018 if (CmpCanonicalized && RHSCanonicalized)
1021 // If RHS is used by other instructions except compare and select, don't
1022 // canonicalize it to not increase the instruction count.
1023 if (!(RHS->hasOneUse() || (RHS->hasNUses(2) && CmpUsesNegatedOp)))
1026 // Create the canonical compare: icmp slt LHS 0.
1027 if (!CmpCanonicalized) {
1028 Cmp.setPredicate(ICmpInst::ICMP_SLT);
1029 Cmp.setOperand(1, ConstantInt::getNullValue(Cmp.getOperand(0)->getType()));
1030 if (CmpUsesNegatedOp)
1031 Cmp.setOperand(0, LHS);
1034 // Create the canonical RHS: RHS = sub (0, LHS).
1035 if (!RHSCanonicalized) {
1036 assert(RHS->hasOneUse() && "RHS use number is not right");
1037 RHS = Builder.CreateNeg(LHS);
1039 Sel.setFalseValue(RHS);
1042 Sel.setTrueValue(RHS);
1047 // If the select operands do not change, we're done.
1048 if (SPF == SelectPatternFlavor::SPF_NABS) {
1051 assert(FVal == LHS && "Unexpected results from matchSelectPattern");
1055 assert(TVal == LHS && "Unexpected results from matchSelectPattern");
1058 // We are swapping the select operands, so swap the metadata too.
1059 Sel.setTrueValue(FVal);
1060 Sel.setFalseValue(TVal);
1061 Sel.swapProfMetadata();
1065 /// Visit a SelectInst that has an ICmpInst as its first operand.
1066 Instruction *InstCombiner::foldSelectInstWithICmp(SelectInst &SI,
1068 Value *TrueVal = SI.getTrueValue();
1069 Value *FalseVal = SI.getFalseValue();
1071 if (Instruction *NewSel = canonicalizeMinMaxWithConstant(SI, *ICI, Builder))
1074 if (Instruction *NewAbs = canonicalizeAbsNabs(SI, *ICI, Builder))
1077 bool Changed = adjustMinMax(SI, *ICI);
1079 if (Value *V = foldSelectICmpAnd(SI, ICI, Builder))
1080 return replaceInstUsesWith(SI, V);
1082 // NOTE: if we wanted to, this is where to detect integer MIN/MAX
1083 ICmpInst::Predicate Pred = ICI->getPredicate();
1084 Value *CmpLHS = ICI->getOperand(0);
1085 Value *CmpRHS = ICI->getOperand(1);
1086 if (CmpRHS != CmpLHS && isa<Constant>(CmpRHS)) {
1087 if (CmpLHS == TrueVal && Pred == ICmpInst::ICMP_EQ) {
1088 // Transform (X == C) ? X : Y -> (X == C) ? C : Y
1089 SI.setOperand(1, CmpRHS);
1091 } else if (CmpLHS == FalseVal && Pred == ICmpInst::ICMP_NE) {
1092 // Transform (X != C) ? Y : X -> (X != C) ? Y : C
1093 SI.setOperand(2, CmpRHS);
1098 // FIXME: This code is nearly duplicated in InstSimplify. Using/refactoring
1099 // decomposeBitTestICmp() might help.
1102 DL.getTypeSizeInBits(TrueVal->getType()->getScalarType());
1103 APInt MinSignedValue = APInt::getSignedMinValue(BitWidth);
1107 bool IsBitTest = false;
1108 if (ICmpInst::isEquality(Pred) &&
1109 match(CmpLHS, m_And(m_Value(X), m_Power2(Y))) &&
1110 match(CmpRHS, m_Zero())) {
1112 TrueWhenUnset = Pred == ICmpInst::ICMP_EQ;
1113 } else if (Pred == ICmpInst::ICMP_SLT && match(CmpRHS, m_Zero())) {
1115 Y = &MinSignedValue;
1117 TrueWhenUnset = false;
1118 } else if (Pred == ICmpInst::ICMP_SGT && match(CmpRHS, m_AllOnes())) {
1120 Y = &MinSignedValue;
1122 TrueWhenUnset = true;
1126 // (X & Y) == 0 ? X : X ^ Y --> X & ~Y
1127 if (TrueWhenUnset && TrueVal == X &&
1128 match(FalseVal, m_Xor(m_Specific(X), m_APInt(C))) && *Y == *C)
1129 V = Builder.CreateAnd(X, ~(*Y));
1130 // (X & Y) != 0 ? X ^ Y : X --> X & ~Y
1131 else if (!TrueWhenUnset && FalseVal == X &&
1132 match(TrueVal, m_Xor(m_Specific(X), m_APInt(C))) && *Y == *C)
1133 V = Builder.CreateAnd(X, ~(*Y));
1134 // (X & Y) == 0 ? X ^ Y : X --> X | Y
1135 else if (TrueWhenUnset && FalseVal == X &&
1136 match(TrueVal, m_Xor(m_Specific(X), m_APInt(C))) && *Y == *C)
1137 V = Builder.CreateOr(X, *Y);
1138 // (X & Y) != 0 ? X : X ^ Y --> X | Y
1139 else if (!TrueWhenUnset && TrueVal == X &&
1140 match(FalseVal, m_Xor(m_Specific(X), m_APInt(C))) && *Y == *C)
1141 V = Builder.CreateOr(X, *Y);
1144 return replaceInstUsesWith(SI, V);
1148 if (Instruction *V =
1149 foldSelectICmpAndAnd(SI.getType(), ICI, TrueVal, FalseVal, Builder))
1152 if (Value *V = foldSelectICmpAndOr(ICI, TrueVal, FalseVal, Builder))
1153 return replaceInstUsesWith(SI, V);
1155 if (Value *V = foldSelectICmpLshrAshr(ICI, TrueVal, FalseVal, Builder))
1156 return replaceInstUsesWith(SI, V);
1158 if (Value *V = foldSelectCttzCtlz(ICI, TrueVal, FalseVal, Builder))
1159 return replaceInstUsesWith(SI, V);
1161 if (Value *V = canonicalizeSaturatedSubtract(ICI, TrueVal, FalseVal, Builder))
1162 return replaceInstUsesWith(SI, V);
1164 if (Value *V = canonicalizeSaturatedAdd(ICI, TrueVal, FalseVal, Builder))
1165 return replaceInstUsesWith(SI, V);
1167 return Changed ? &SI : nullptr;
1170 /// SI is a select whose condition is a PHI node (but the two may be in
1171 /// different blocks). See if the true/false values (V) are live in all of the
1172 /// predecessor blocks of the PHI. For example, cases like this can't be mapped:
1174 /// X = phi [ C1, BB1], [C2, BB2]
1176 /// Z = select X, Y, 0
1178 /// because Y is not live in BB1/BB2.
1179 static bool canSelectOperandBeMappingIntoPredBlock(const Value *V,
1180 const SelectInst &SI) {
1181 // If the value is a non-instruction value like a constant or argument, it
1182 // can always be mapped.
1183 const Instruction *I = dyn_cast<Instruction>(V);
1184 if (!I) return true;
1186 // If V is a PHI node defined in the same block as the condition PHI, we can
1187 // map the arguments.
1188 const PHINode *CondPHI = cast<PHINode>(SI.getCondition());
1190 if (const PHINode *VP = dyn_cast<PHINode>(I))
1191 if (VP->getParent() == CondPHI->getParent())
1194 // Otherwise, if the PHI and select are defined in the same block and if V is
1195 // defined in a different block, then we can transform it.
1196 if (SI.getParent() == CondPHI->getParent() &&
1197 I->getParent() != CondPHI->getParent())
1200 // Otherwise we have a 'hard' case and we can't tell without doing more
1201 // detailed dominator based analysis, punt.
1205 /// We have an SPF (e.g. a min or max) of an SPF of the form:
1206 /// SPF2(SPF1(A, B), C)
1207 Instruction *InstCombiner::foldSPFofSPF(Instruction *Inner,
1208 SelectPatternFlavor SPF1,
1211 SelectPatternFlavor SPF2, Value *C) {
1212 if (Outer.getType() != Inner->getType())
1215 if (C == A || C == B) {
1216 // MAX(MAX(A, B), B) -> MAX(A, B)
1217 // MIN(MIN(a, b), a) -> MIN(a, b)
1218 // TODO: This could be done in instsimplify.
1219 if (SPF1 == SPF2 && SelectPatternResult::isMinOrMax(SPF1))
1220 return replaceInstUsesWith(Outer, Inner);
1222 // MAX(MIN(a, b), a) -> a
1223 // MIN(MAX(a, b), a) -> a
1224 // TODO: This could be done in instsimplify.
1225 if ((SPF1 == SPF_SMIN && SPF2 == SPF_SMAX) ||
1226 (SPF1 == SPF_SMAX && SPF2 == SPF_SMIN) ||
1227 (SPF1 == SPF_UMIN && SPF2 == SPF_UMAX) ||
1228 (SPF1 == SPF_UMAX && SPF2 == SPF_UMIN))
1229 return replaceInstUsesWith(Outer, C);
1233 const APInt *CB, *CC;
1234 if (match(B, m_APInt(CB)) && match(C, m_APInt(CC))) {
1235 // MIN(MIN(A, 23), 97) -> MIN(A, 23)
1236 // MAX(MAX(A, 97), 23) -> MAX(A, 97)
1237 // TODO: This could be done in instsimplify.
1238 if ((SPF1 == SPF_UMIN && CB->ule(*CC)) ||
1239 (SPF1 == SPF_SMIN && CB->sle(*CC)) ||
1240 (SPF1 == SPF_UMAX && CB->uge(*CC)) ||
1241 (SPF1 == SPF_SMAX && CB->sge(*CC)))
1242 return replaceInstUsesWith(Outer, Inner);
1244 // MIN(MIN(A, 97), 23) -> MIN(A, 23)
1245 // MAX(MAX(A, 23), 97) -> MAX(A, 97)
1246 if ((SPF1 == SPF_UMIN && CB->ugt(*CC)) ||
1247 (SPF1 == SPF_SMIN && CB->sgt(*CC)) ||
1248 (SPF1 == SPF_UMAX && CB->ult(*CC)) ||
1249 (SPF1 == SPF_SMAX && CB->slt(*CC))) {
1250 Outer.replaceUsesOfWith(Inner, A);
1256 // ABS(ABS(X)) -> ABS(X)
1257 // NABS(NABS(X)) -> NABS(X)
1258 // TODO: This could be done in instsimplify.
1259 if (SPF1 == SPF2 && (SPF1 == SPF_ABS || SPF1 == SPF_NABS)) {
1260 return replaceInstUsesWith(Outer, Inner);
1263 // ABS(NABS(X)) -> ABS(X)
1264 // NABS(ABS(X)) -> NABS(X)
1265 if ((SPF1 == SPF_ABS && SPF2 == SPF_NABS) ||
1266 (SPF1 == SPF_NABS && SPF2 == SPF_ABS)) {
1267 SelectInst *SI = cast<SelectInst>(Inner);
1269 Builder.CreateSelect(SI->getCondition(), SI->getFalseValue(),
1270 SI->getTrueValue(), SI->getName(), SI);
1271 return replaceInstUsesWith(Outer, NewSI);
1274 auto IsFreeOrProfitableToInvert =
1275 [&](Value *V, Value *&NotV, bool &ElidesXor) {
1276 if (match(V, m_Not(m_Value(NotV)))) {
1277 // If V has at most 2 uses then we can get rid of the xor operation
1279 ElidesXor |= !V->hasNUsesOrMore(3);
1283 if (IsFreeToInvert(V, !V->hasNUsesOrMore(3))) {
1291 Value *NotA, *NotB, *NotC;
1292 bool ElidesXor = false;
1294 // MIN(MIN(~A, ~B), ~C) == ~MAX(MAX(A, B), C)
1295 // MIN(MAX(~A, ~B), ~C) == ~MAX(MIN(A, B), C)
1296 // MAX(MIN(~A, ~B), ~C) == ~MIN(MAX(A, B), C)
1297 // MAX(MAX(~A, ~B), ~C) == ~MIN(MIN(A, B), C)
1299 // This transform is performance neutral if we can elide at least one xor from
1300 // the set of three operands, since we'll be tacking on an xor at the very
1302 if (SelectPatternResult::isMinOrMax(SPF1) &&
1303 SelectPatternResult::isMinOrMax(SPF2) &&
1304 IsFreeOrProfitableToInvert(A, NotA, ElidesXor) &&
1305 IsFreeOrProfitableToInvert(B, NotB, ElidesXor) &&
1306 IsFreeOrProfitableToInvert(C, NotC, ElidesXor) && ElidesXor) {
1308 NotA = Builder.CreateNot(A);
1310 NotB = Builder.CreateNot(B);
1312 NotC = Builder.CreateNot(C);
1314 Value *NewInner = createMinMax(Builder, getInverseMinMaxFlavor(SPF1), NotA,
1316 Value *NewOuter = Builder.CreateNot(
1317 createMinMax(Builder, getInverseMinMaxFlavor(SPF2), NewInner, NotC));
1318 return replaceInstUsesWith(Outer, NewOuter);
1324 /// Turn select C, (X + Y), (X - Y) --> (X + (select C, Y, (-Y))).
1325 /// This is even legal for FP.
1326 static Instruction *foldAddSubSelect(SelectInst &SI,
1327 InstCombiner::BuilderTy &Builder) {
1328 Value *CondVal = SI.getCondition();
1329 Value *TrueVal = SI.getTrueValue();
1330 Value *FalseVal = SI.getFalseValue();
1331 auto *TI = dyn_cast<Instruction>(TrueVal);
1332 auto *FI = dyn_cast<Instruction>(FalseVal);
1333 if (!TI || !FI || !TI->hasOneUse() || !FI->hasOneUse())
1336 Instruction *AddOp = nullptr, *SubOp = nullptr;
1337 if ((TI->getOpcode() == Instruction::Sub &&
1338 FI->getOpcode() == Instruction::Add) ||
1339 (TI->getOpcode() == Instruction::FSub &&
1340 FI->getOpcode() == Instruction::FAdd)) {
1343 } else if ((FI->getOpcode() == Instruction::Sub &&
1344 TI->getOpcode() == Instruction::Add) ||
1345 (FI->getOpcode() == Instruction::FSub &&
1346 TI->getOpcode() == Instruction::FAdd)) {
1352 Value *OtherAddOp = nullptr;
1353 if (SubOp->getOperand(0) == AddOp->getOperand(0)) {
1354 OtherAddOp = AddOp->getOperand(1);
1355 } else if (SubOp->getOperand(0) == AddOp->getOperand(1)) {
1356 OtherAddOp = AddOp->getOperand(0);
1360 // So at this point we know we have (Y -> OtherAddOp):
1361 // select C, (add X, Y), (sub X, Z)
1362 Value *NegVal; // Compute -Z
1363 if (SI.getType()->isFPOrFPVectorTy()) {
1364 NegVal = Builder.CreateFNeg(SubOp->getOperand(1));
1365 if (Instruction *NegInst = dyn_cast<Instruction>(NegVal)) {
1366 FastMathFlags Flags = AddOp->getFastMathFlags();
1367 Flags &= SubOp->getFastMathFlags();
1368 NegInst->setFastMathFlags(Flags);
1371 NegVal = Builder.CreateNeg(SubOp->getOperand(1));
1374 Value *NewTrueOp = OtherAddOp;
1375 Value *NewFalseOp = NegVal;
1377 std::swap(NewTrueOp, NewFalseOp);
1378 Value *NewSel = Builder.CreateSelect(CondVal, NewTrueOp, NewFalseOp,
1379 SI.getName() + ".p", &SI);
1381 if (SI.getType()->isFPOrFPVectorTy()) {
1383 BinaryOperator::CreateFAdd(SubOp->getOperand(0), NewSel);
1385 FastMathFlags Flags = AddOp->getFastMathFlags();
1386 Flags &= SubOp->getFastMathFlags();
1387 RI->setFastMathFlags(Flags);
1390 return BinaryOperator::CreateAdd(SubOp->getOperand(0), NewSel);
1396 Instruction *InstCombiner::foldSelectExtConst(SelectInst &Sel) {
1398 if (!match(Sel.getTrueValue(), m_Constant(C)) &&
1399 !match(Sel.getFalseValue(), m_Constant(C)))
1402 Instruction *ExtInst;
1403 if (!match(Sel.getTrueValue(), m_Instruction(ExtInst)) &&
1404 !match(Sel.getFalseValue(), m_Instruction(ExtInst)))
1407 auto ExtOpcode = ExtInst->getOpcode();
1408 if (ExtOpcode != Instruction::ZExt && ExtOpcode != Instruction::SExt)
1411 // If we are extending from a boolean type or if we can create a select that
1412 // has the same size operands as its condition, try to narrow the select.
1413 Value *X = ExtInst->getOperand(0);
1414 Type *SmallType = X->getType();
1415 Value *Cond = Sel.getCondition();
1416 auto *Cmp = dyn_cast<CmpInst>(Cond);
1417 if (!SmallType->isIntOrIntVectorTy(1) &&
1418 (!Cmp || Cmp->getOperand(0)->getType() != SmallType))
1421 // If the constant is the same after truncation to the smaller type and
1422 // extension to the original type, we can narrow the select.
1423 Type *SelType = Sel.getType();
1424 Constant *TruncC = ConstantExpr::getTrunc(C, SmallType);
1425 Constant *ExtC = ConstantExpr::getCast(ExtOpcode, TruncC, SelType);
1427 Value *TruncCVal = cast<Value>(TruncC);
1428 if (ExtInst == Sel.getFalseValue())
1429 std::swap(X, TruncCVal);
1431 // select Cond, (ext X), C --> ext(select Cond, X, C')
1432 // select Cond, C, (ext X) --> ext(select Cond, C', X)
1433 Value *NewSel = Builder.CreateSelect(Cond, X, TruncCVal, "narrow", &Sel);
1434 return CastInst::Create(Instruction::CastOps(ExtOpcode), NewSel, SelType);
1437 // If one arm of the select is the extend of the condition, replace that arm
1438 // with the extension of the appropriate known bool value.
1440 if (ExtInst == Sel.getTrueValue()) {
1441 // select X, (sext X), C --> select X, -1, C
1442 // select X, (zext X), C --> select X, 1, C
1443 Constant *One = ConstantInt::getTrue(SmallType);
1444 Constant *AllOnesOrOne = ConstantExpr::getCast(ExtOpcode, One, SelType);
1445 return SelectInst::Create(Cond, AllOnesOrOne, C, "", nullptr, &Sel);
1447 // select X, C, (sext X) --> select X, C, 0
1448 // select X, C, (zext X) --> select X, C, 0
1449 Constant *Zero = ConstantInt::getNullValue(SelType);
1450 return SelectInst::Create(Cond, C, Zero, "", nullptr, &Sel);
1457 /// Try to transform a vector select with a constant condition vector into a
1458 /// shuffle for easier combining with other shuffles and insert/extract.
1459 static Instruction *canonicalizeSelectToShuffle(SelectInst &SI) {
1460 Value *CondVal = SI.getCondition();
1462 if (!CondVal->getType()->isVectorTy() || !match(CondVal, m_Constant(CondC)))
1465 unsigned NumElts = CondVal->getType()->getVectorNumElements();
1466 SmallVector<Constant *, 16> Mask;
1467 Mask.reserve(NumElts);
1468 Type *Int32Ty = Type::getInt32Ty(CondVal->getContext());
1469 for (unsigned i = 0; i != NumElts; ++i) {
1470 Constant *Elt = CondC->getAggregateElement(i);
1474 if (Elt->isOneValue()) {
1475 // If the select condition element is true, choose from the 1st vector.
1476 Mask.push_back(ConstantInt::get(Int32Ty, i));
1477 } else if (Elt->isNullValue()) {
1478 // If the select condition element is false, choose from the 2nd vector.
1479 Mask.push_back(ConstantInt::get(Int32Ty, i + NumElts));
1480 } else if (isa<UndefValue>(Elt)) {
1481 // Undef in a select condition (choose one of the operands) does not mean
1482 // the same thing as undef in a shuffle mask (any value is acceptable), so
1486 // Bail out on a constant expression.
1491 return new ShuffleVectorInst(SI.getTrueValue(), SI.getFalseValue(),
1492 ConstantVector::get(Mask));
1495 /// Reuse bitcasted operands between a compare and select:
1496 /// select (cmp (bitcast C), (bitcast D)), (bitcast' C), (bitcast' D) -->
1497 /// bitcast (select (cmp (bitcast C), (bitcast D)), (bitcast C), (bitcast D))
1498 static Instruction *foldSelectCmpBitcasts(SelectInst &Sel,
1499 InstCombiner::BuilderTy &Builder) {
1500 Value *Cond = Sel.getCondition();
1501 Value *TVal = Sel.getTrueValue();
1502 Value *FVal = Sel.getFalseValue();
1504 CmpInst::Predicate Pred;
1506 if (!match(Cond, m_Cmp(Pred, m_Value(A), m_Value(B))))
1509 // The select condition is a compare instruction. If the select's true/false
1510 // values are already the same as the compare operands, there's nothing to do.
1511 if (TVal == A || TVal == B || FVal == A || FVal == B)
1515 if (!match(A, m_BitCast(m_Value(C))) || !match(B, m_BitCast(m_Value(D))))
1518 // select (cmp (bitcast C), (bitcast D)), (bitcast TSrc), (bitcast FSrc)
1520 if (!match(TVal, m_BitCast(m_Value(TSrc))) ||
1521 !match(FVal, m_BitCast(m_Value(FSrc))))
1524 // If the select true/false values are *different bitcasts* of the same source
1525 // operands, make the select operands the same as the compare operands and
1526 // cast the result. This is the canonical select form for min/max.
1528 if (TSrc == C && FSrc == D) {
1529 // select (cmp (bitcast C), (bitcast D)), (bitcast' C), (bitcast' D) -->
1530 // bitcast (select (cmp A, B), A, B)
1531 NewSel = Builder.CreateSelect(Cond, A, B, "", &Sel);
1532 } else if (TSrc == D && FSrc == C) {
1533 // select (cmp (bitcast C), (bitcast D)), (bitcast' D), (bitcast' C) -->
1534 // bitcast (select (cmp A, B), B, A)
1535 NewSel = Builder.CreateSelect(Cond, B, A, "", &Sel);
1539 return CastInst::CreateBitOrPointerCast(NewSel, Sel.getType());
1542 /// Try to eliminate select instructions that test the returned flag of cmpxchg
1545 /// If a select instruction tests the returned flag of a cmpxchg instruction and
1546 /// selects between the returned value of the cmpxchg instruction its compare
1547 /// operand, the result of the select will always be equal to its false value.
1550 /// %0 = cmpxchg i64* %ptr, i64 %compare, i64 %new_value seq_cst seq_cst
1551 /// %1 = extractvalue { i64, i1 } %0, 1
1552 /// %2 = extractvalue { i64, i1 } %0, 0
1553 /// %3 = select i1 %1, i64 %compare, i64 %2
1556 /// The returned value of the cmpxchg instruction (%2) is the original value
1557 /// located at %ptr prior to any update. If the cmpxchg operation succeeds, %2
1558 /// must have been equal to %compare. Thus, the result of the select is always
1559 /// equal to %2, and the code can be simplified to:
1561 /// %0 = cmpxchg i64* %ptr, i64 %compare, i64 %new_value seq_cst seq_cst
1562 /// %1 = extractvalue { i64, i1 } %0, 0
1565 static Instruction *foldSelectCmpXchg(SelectInst &SI) {
1566 // A helper that determines if V is an extractvalue instruction whose
1567 // aggregate operand is a cmpxchg instruction and whose single index is equal
1568 // to I. If such conditions are true, the helper returns the cmpxchg
1569 // instruction; otherwise, a nullptr is returned.
1570 auto isExtractFromCmpXchg = [](Value *V, unsigned I) -> AtomicCmpXchgInst * {
1571 auto *Extract = dyn_cast<ExtractValueInst>(V);
1574 if (Extract->getIndices()[0] != I)
1576 return dyn_cast<AtomicCmpXchgInst>(Extract->getAggregateOperand());
1579 // If the select has a single user, and this user is a select instruction that
1580 // we can simplify, skip the cmpxchg simplification for now.
1582 if (auto *Select = dyn_cast<SelectInst>(SI.user_back()))
1583 if (Select->getCondition() == SI.getCondition())
1584 if (Select->getFalseValue() == SI.getTrueValue() ||
1585 Select->getTrueValue() == SI.getFalseValue())
1588 // Ensure the select condition is the returned flag of a cmpxchg instruction.
1589 auto *CmpXchg = isExtractFromCmpXchg(SI.getCondition(), 1);
1593 // Check the true value case: The true value of the select is the returned
1594 // value of the same cmpxchg used by the condition, and the false value is the
1595 // cmpxchg instruction's compare operand.
1596 if (auto *X = isExtractFromCmpXchg(SI.getTrueValue(), 0))
1597 if (X == CmpXchg && X->getCompareOperand() == SI.getFalseValue()) {
1598 SI.setTrueValue(SI.getFalseValue());
1602 // Check the false value case: The false value of the select is the returned
1603 // value of the same cmpxchg used by the condition, and the true value is the
1604 // cmpxchg instruction's compare operand.
1605 if (auto *X = isExtractFromCmpXchg(SI.getFalseValue(), 0))
1606 if (X == CmpXchg && X->getCompareOperand() == SI.getTrueValue()) {
1607 SI.setTrueValue(SI.getFalseValue());
1614 static Instruction *moveAddAfterMinMax(SelectPatternFlavor SPF, Value *X,
1616 InstCombiner::BuilderTy &Builder) {
1617 assert(SelectPatternResult::isMinOrMax(SPF) && "Expected min/max pattern");
1618 bool IsUnsigned = SPF == SelectPatternFlavor::SPF_UMIN ||
1619 SPF == SelectPatternFlavor::SPF_UMAX;
1620 // TODO: If InstSimplify could fold all cases where C2 <= C1, we could change
1621 // the constant value check to an assert.
1623 const APInt *C1, *C2;
1624 if (IsUnsigned && match(X, m_NUWAdd(m_Value(A), m_APInt(C1))) &&
1625 match(Y, m_APInt(C2)) && C2->uge(*C1) && X->hasNUses(2)) {
1626 // umin (add nuw A, C1), C2 --> add nuw (umin A, C2 - C1), C1
1627 // umax (add nuw A, C1), C2 --> add nuw (umax A, C2 - C1), C1
1628 Value *NewMinMax = createMinMax(Builder, SPF, A,
1629 ConstantInt::get(X->getType(), *C2 - *C1));
1630 return BinaryOperator::CreateNUW(BinaryOperator::Add, NewMinMax,
1631 ConstantInt::get(X->getType(), *C1));
1634 if (!IsUnsigned && match(X, m_NSWAdd(m_Value(A), m_APInt(C1))) &&
1635 match(Y, m_APInt(C2)) && X->hasNUses(2)) {
1637 APInt Diff = C2->ssub_ov(*C1, Overflow);
1639 // smin (add nsw A, C1), C2 --> add nsw (smin A, C2 - C1), C1
1640 // smax (add nsw A, C1), C2 --> add nsw (smax A, C2 - C1), C1
1641 Value *NewMinMax = createMinMax(Builder, SPF, A,
1642 ConstantInt::get(X->getType(), Diff));
1643 return BinaryOperator::CreateNSW(BinaryOperator::Add, NewMinMax,
1644 ConstantInt::get(X->getType(), *C1));
1651 /// Reduce a sequence of min/max with a common operand.
1652 static Instruction *factorizeMinMaxTree(SelectPatternFlavor SPF, Value *LHS,
1654 InstCombiner::BuilderTy &Builder) {
1655 assert(SelectPatternResult::isMinOrMax(SPF) && "Expected a min/max");
1656 // TODO: Allow FP min/max with nnan/nsz.
1657 if (!LHS->getType()->isIntOrIntVectorTy())
1660 // Match 3 of the same min/max ops. Example: umin(umin(), umin()).
1661 Value *A, *B, *C, *D;
1662 SelectPatternResult L = matchSelectPattern(LHS, A, B);
1663 SelectPatternResult R = matchSelectPattern(RHS, C, D);
1664 if (SPF != L.Flavor || L.Flavor != R.Flavor)
1667 // Look for a common operand. The use checks are different than usual because
1668 // a min/max pattern typically has 2 uses of each op: 1 by the cmp and 1 by
1670 Value *MinMaxOp = nullptr;
1671 Value *ThirdOp = nullptr;
1672 if (!LHS->hasNUsesOrMore(3) && RHS->hasNUsesOrMore(3)) {
1673 // If the LHS is only used in this chain and the RHS is used outside of it,
1674 // reuse the RHS min/max because that will eliminate the LHS.
1675 if (D == A || C == A) {
1676 // min(min(a, b), min(c, a)) --> min(min(c, a), b)
1677 // min(min(a, b), min(a, d)) --> min(min(a, d), b)
1680 } else if (D == B || C == B) {
1681 // min(min(a, b), min(c, b)) --> min(min(c, b), a)
1682 // min(min(a, b), min(b, d)) --> min(min(b, d), a)
1686 } else if (!RHS->hasNUsesOrMore(3)) {
1687 // Reuse the LHS. This will eliminate the RHS.
1688 if (D == A || D == B) {
1689 // min(min(a, b), min(c, a)) --> min(min(a, b), c)
1690 // min(min(a, b), min(c, b)) --> min(min(a, b), c)
1693 } else if (C == A || C == B) {
1694 // min(min(a, b), min(b, d)) --> min(min(a, b), d)
1695 // min(min(a, b), min(c, b)) --> min(min(a, b), d)
1700 if (!MinMaxOp || !ThirdOp)
1703 CmpInst::Predicate P = getMinMaxPred(SPF);
1704 Value *CmpABC = Builder.CreateICmp(P, MinMaxOp, ThirdOp);
1705 return SelectInst::Create(CmpABC, MinMaxOp, ThirdOp);
1708 /// Try to reduce a rotate pattern that includes a compare and select into a
1709 /// funnel shift intrinsic. Example:
1710 /// rotl32(a, b) --> (b == 0 ? a : ((a >> (32 - b)) | (a << b)))
1711 /// --> call llvm.fshl.i32(a, a, b)
1712 static Instruction *foldSelectRotate(SelectInst &Sel) {
1713 // The false value of the select must be a rotate of the true value.
1715 if (!match(Sel.getFalseValue(), m_OneUse(m_Or(m_Value(Or0), m_Value(Or1)))))
1718 Value *TVal = Sel.getTrueValue();
1720 if (!match(Or0, m_OneUse(m_LogicalShift(m_Specific(TVal), m_Value(SA0)))) ||
1721 !match(Or1, m_OneUse(m_LogicalShift(m_Specific(TVal), m_Value(SA1)))))
1724 auto ShiftOpcode0 = cast<BinaryOperator>(Or0)->getOpcode();
1725 auto ShiftOpcode1 = cast<BinaryOperator>(Or1)->getOpcode();
1726 if (ShiftOpcode0 == ShiftOpcode1)
1729 // We have one of these patterns so far:
1730 // select ?, TVal, (or (lshr TVal, SA0), (shl TVal, SA1))
1731 // select ?, TVal, (or (shl TVal, SA0), (lshr TVal, SA1))
1732 // This must be a power-of-2 rotate for a bitmasking transform to be valid.
1733 unsigned Width = Sel.getType()->getScalarSizeInBits();
1734 if (!isPowerOf2_32(Width))
1737 // Check the shift amounts to see if they are an opposite pair.
1739 if (match(SA1, m_OneUse(m_Sub(m_SpecificInt(Width), m_Specific(SA0)))))
1741 else if (match(SA0, m_OneUse(m_Sub(m_SpecificInt(Width), m_Specific(SA1)))))
1746 // Finally, see if the select is filtering out a shift-by-zero.
1747 Value *Cond = Sel.getCondition();
1748 ICmpInst::Predicate Pred;
1749 if (!match(Cond, m_OneUse(m_ICmp(Pred, m_Specific(ShAmt), m_ZeroInt()))) ||
1750 Pred != ICmpInst::ICMP_EQ)
1753 // This is a rotate that avoids shift-by-bitwidth UB in a suboptimal way.
1754 // Convert to funnel shift intrinsic.
1755 bool IsFshl = (ShAmt == SA0 && ShiftOpcode0 == BinaryOperator::Shl) ||
1756 (ShAmt == SA1 && ShiftOpcode1 == BinaryOperator::Shl);
1757 Intrinsic::ID IID = IsFshl ? Intrinsic::fshl : Intrinsic::fshr;
1758 Function *F = Intrinsic::getDeclaration(Sel.getModule(), IID, Sel.getType());
1759 return IntrinsicInst::Create(F, { TVal, TVal, ShAmt });
1762 Instruction *InstCombiner::visitSelectInst(SelectInst &SI) {
1763 Value *CondVal = SI.getCondition();
1764 Value *TrueVal = SI.getTrueValue();
1765 Value *FalseVal = SI.getFalseValue();
1766 Type *SelType = SI.getType();
1768 // FIXME: Remove this workaround when freeze related patches are done.
1769 // For select with undef operand which feeds into an equality comparison,
1770 // don't simplify it so loop unswitch can know the equality comparison
1771 // may have an undef operand. This is a workaround for PR31652 caused by
1772 // descrepancy about branch on undef between LoopUnswitch and GVN.
1773 if (isa<UndefValue>(TrueVal) || isa<UndefValue>(FalseVal)) {
1774 if (llvm::any_of(SI.users(), [&](User *U) {
1775 ICmpInst *CI = dyn_cast<ICmpInst>(U);
1776 if (CI && CI->isEquality())
1784 if (Value *V = SimplifySelectInst(CondVal, TrueVal, FalseVal,
1785 SQ.getWithInstruction(&SI)))
1786 return replaceInstUsesWith(SI, V);
1788 if (Instruction *I = canonicalizeSelectToShuffle(SI))
1791 // Canonicalize a one-use integer compare with a non-canonical predicate by
1792 // inverting the predicate and swapping the select operands. This matches a
1793 // compare canonicalization for conditional branches.
1794 // TODO: Should we do the same for FP compares?
1795 CmpInst::Predicate Pred;
1796 if (match(CondVal, m_OneUse(m_ICmp(Pred, m_Value(), m_Value()))) &&
1797 !isCanonicalPredicate(Pred)) {
1798 // Swap true/false values and condition.
1799 CmpInst *Cond = cast<CmpInst>(CondVal);
1800 Cond->setPredicate(CmpInst::getInversePredicate(Pred));
1801 SI.setOperand(1, FalseVal);
1802 SI.setOperand(2, TrueVal);
1803 SI.swapProfMetadata();
1808 if (SelType->isIntOrIntVectorTy(1) &&
1809 TrueVal->getType() == CondVal->getType()) {
1810 if (match(TrueVal, m_One())) {
1811 // Change: A = select B, true, C --> A = or B, C
1812 return BinaryOperator::CreateOr(CondVal, FalseVal);
1814 if (match(TrueVal, m_Zero())) {
1815 // Change: A = select B, false, C --> A = and !B, C
1816 Value *NotCond = Builder.CreateNot(CondVal, "not." + CondVal->getName());
1817 return BinaryOperator::CreateAnd(NotCond, FalseVal);
1819 if (match(FalseVal, m_Zero())) {
1820 // Change: A = select B, C, false --> A = and B, C
1821 return BinaryOperator::CreateAnd(CondVal, TrueVal);
1823 if (match(FalseVal, m_One())) {
1824 // Change: A = select B, C, true --> A = or !B, C
1825 Value *NotCond = Builder.CreateNot(CondVal, "not." + CondVal->getName());
1826 return BinaryOperator::CreateOr(NotCond, TrueVal);
1829 // select a, a, b -> a | b
1830 // select a, b, a -> a & b
1831 if (CondVal == TrueVal)
1832 return BinaryOperator::CreateOr(CondVal, FalseVal);
1833 if (CondVal == FalseVal)
1834 return BinaryOperator::CreateAnd(CondVal, TrueVal);
1836 // select a, ~a, b -> (~a) & b
1837 // select a, b, ~a -> (~a) | b
1838 if (match(TrueVal, m_Not(m_Specific(CondVal))))
1839 return BinaryOperator::CreateAnd(TrueVal, FalseVal);
1840 if (match(FalseVal, m_Not(m_Specific(CondVal))))
1841 return BinaryOperator::CreateOr(TrueVal, FalseVal);
1844 // Selecting between two integer or vector splat integer constants?
1846 // Note that we don't handle a scalar select of vectors:
1847 // select i1 %c, <2 x i8> <1, 1>, <2 x i8> <0, 0>
1848 // because that may need 3 instructions to splat the condition value:
1849 // extend, insertelement, shufflevector.
1850 if (SelType->isIntOrIntVectorTy() &&
1851 CondVal->getType()->isVectorTy() == SelType->isVectorTy()) {
1852 // select C, 1, 0 -> zext C to int
1853 if (match(TrueVal, m_One()) && match(FalseVal, m_Zero()))
1854 return new ZExtInst(CondVal, SelType);
1856 // select C, -1, 0 -> sext C to int
1857 if (match(TrueVal, m_AllOnes()) && match(FalseVal, m_Zero()))
1858 return new SExtInst(CondVal, SelType);
1860 // select C, 0, 1 -> zext !C to int
1861 if (match(TrueVal, m_Zero()) && match(FalseVal, m_One())) {
1862 Value *NotCond = Builder.CreateNot(CondVal, "not." + CondVal->getName());
1863 return new ZExtInst(NotCond, SelType);
1866 // select C, 0, -1 -> sext !C to int
1867 if (match(TrueVal, m_Zero()) && match(FalseVal, m_AllOnes())) {
1868 Value *NotCond = Builder.CreateNot(CondVal, "not." + CondVal->getName());
1869 return new SExtInst(NotCond, SelType);
1873 // See if we are selecting two values based on a comparison of the two values.
1874 if (FCmpInst *FCI = dyn_cast<FCmpInst>(CondVal)) {
1875 if (FCI->getOperand(0) == TrueVal && FCI->getOperand(1) == FalseVal) {
1876 // Canonicalize to use ordered comparisons by swapping the select
1880 // (X ugt Y) ? X : Y -> (X ole Y) ? Y : X
1881 if (FCI->hasOneUse() && FCmpInst::isUnordered(FCI->getPredicate())) {
1882 FCmpInst::Predicate InvPred = FCI->getInversePredicate();
1883 IRBuilder<>::FastMathFlagGuard FMFG(Builder);
1884 Builder.setFastMathFlags(FCI->getFastMathFlags());
1885 Value *NewCond = Builder.CreateFCmp(InvPred, TrueVal, FalseVal,
1886 FCI->getName() + ".inv");
1888 return SelectInst::Create(NewCond, FalseVal, TrueVal,
1889 SI.getName() + ".p");
1892 // NOTE: if we wanted to, this is where to detect MIN/MAX
1893 } else if (FCI->getOperand(0) == FalseVal && FCI->getOperand(1) == TrueVal){
1894 // Canonicalize to use ordered comparisons by swapping the select
1898 // (X ugt Y) ? X : Y -> (X ole Y) ? X : Y
1899 if (FCI->hasOneUse() && FCmpInst::isUnordered(FCI->getPredicate())) {
1900 FCmpInst::Predicate InvPred = FCI->getInversePredicate();
1901 IRBuilder<>::FastMathFlagGuard FMFG(Builder);
1902 Builder.setFastMathFlags(FCI->getFastMathFlags());
1903 Value *NewCond = Builder.CreateFCmp(InvPred, FalseVal, TrueVal,
1904 FCI->getName() + ".inv");
1906 return SelectInst::Create(NewCond, FalseVal, TrueVal,
1907 SI.getName() + ".p");
1910 // NOTE: if we wanted to, this is where to detect MIN/MAX
1914 // Canonicalize select with fcmp to fabs(). -0.0 makes this tricky. We need
1915 // fast-math-flags (nsz) or fsub with +0.0 (not fneg) for this to work. We
1916 // also require nnan because we do not want to unintentionally change the
1917 // sign of a NaN value.
1918 // FIXME: These folds should test/propagate FMF from the select, not the
1920 // (X <= +/-0.0) ? (0.0 - X) : X --> fabs(X)
1922 if (match(CondVal, m_FCmp(Pred, m_Specific(FalseVal), m_AnyZeroFP())) &&
1923 match(TrueVal, m_FSub(m_PosZeroFP(), m_Specific(FalseVal))) &&
1924 match(TrueVal, m_Instruction(FSub)) && FSub->hasNoNaNs() &&
1925 (Pred == FCmpInst::FCMP_OLE || Pred == FCmpInst::FCMP_ULE)) {
1926 Value *Fabs = Builder.CreateUnaryIntrinsic(Intrinsic::fabs, FalseVal, FSub);
1927 return replaceInstUsesWith(SI, Fabs);
1929 // (X > +/-0.0) ? X : (0.0 - X) --> fabs(X)
1930 if (match(CondVal, m_FCmp(Pred, m_Specific(TrueVal), m_AnyZeroFP())) &&
1931 match(FalseVal, m_FSub(m_PosZeroFP(), m_Specific(TrueVal))) &&
1932 match(FalseVal, m_Instruction(FSub)) && FSub->hasNoNaNs() &&
1933 (Pred == FCmpInst::FCMP_OGT || Pred == FCmpInst::FCMP_UGT)) {
1934 Value *Fabs = Builder.CreateUnaryIntrinsic(Intrinsic::fabs, TrueVal, FSub);
1935 return replaceInstUsesWith(SI, Fabs);
1937 // With nnan and nsz:
1938 // (X < +/-0.0) ? -X : X --> fabs(X)
1939 // (X <= +/-0.0) ? -X : X --> fabs(X)
1941 if (match(CondVal, m_FCmp(Pred, m_Specific(FalseVal), m_AnyZeroFP())) &&
1942 match(TrueVal, m_FNeg(m_Specific(FalseVal))) &&
1943 match(TrueVal, m_Instruction(FNeg)) &&
1944 FNeg->hasNoNaNs() && FNeg->hasNoSignedZeros() &&
1945 (Pred == FCmpInst::FCMP_OLT || Pred == FCmpInst::FCMP_OLE ||
1946 Pred == FCmpInst::FCMP_ULT || Pred == FCmpInst::FCMP_ULE)) {
1947 Value *Fabs = Builder.CreateUnaryIntrinsic(Intrinsic::fabs, FalseVal, FNeg);
1948 return replaceInstUsesWith(SI, Fabs);
1950 // With nnan and nsz:
1951 // (X > +/-0.0) ? X : -X --> fabs(X)
1952 // (X >= +/-0.0) ? X : -X --> fabs(X)
1953 if (match(CondVal, m_FCmp(Pred, m_Specific(TrueVal), m_AnyZeroFP())) &&
1954 match(FalseVal, m_FNeg(m_Specific(TrueVal))) &&
1955 match(FalseVal, m_Instruction(FNeg)) &&
1956 FNeg->hasNoNaNs() && FNeg->hasNoSignedZeros() &&
1957 (Pred == FCmpInst::FCMP_OGT || Pred == FCmpInst::FCMP_OGE ||
1958 Pred == FCmpInst::FCMP_UGT || Pred == FCmpInst::FCMP_UGE)) {
1959 Value *Fabs = Builder.CreateUnaryIntrinsic(Intrinsic::fabs, TrueVal, FNeg);
1960 return replaceInstUsesWith(SI, Fabs);
1963 // See if we are selecting two values based on a comparison of the two values.
1964 if (ICmpInst *ICI = dyn_cast<ICmpInst>(CondVal))
1965 if (Instruction *Result = foldSelectInstWithICmp(SI, ICI))
1968 if (Instruction *Add = foldAddSubSelect(SI, Builder))
1971 // Turn (select C, (op X, Y), (op X, Z)) -> (op X, (select C, Y, Z))
1972 auto *TI = dyn_cast<Instruction>(TrueVal);
1973 auto *FI = dyn_cast<Instruction>(FalseVal);
1974 if (TI && FI && TI->getOpcode() == FI->getOpcode())
1975 if (Instruction *IV = foldSelectOpOp(SI, TI, FI))
1978 if (Instruction *I = foldSelectExtConst(SI))
1981 // See if we can fold the select into one of our operands.
1982 if (SelType->isIntOrIntVectorTy() || SelType->isFPOrFPVectorTy()) {
1983 if (Instruction *FoldI = foldSelectIntoOp(SI, TrueVal, FalseVal))
1987 Instruction::CastOps CastOp;
1988 SelectPatternResult SPR = matchSelectPattern(&SI, LHS, RHS, &CastOp);
1989 auto SPF = SPR.Flavor;
1992 if (SelectPatternFlavor SPF2 = matchSelectPattern(LHS, LHS2, RHS2).Flavor)
1993 if (Instruction *R = foldSPFofSPF(cast<Instruction>(LHS), SPF2, LHS2,
1994 RHS2, SI, SPF, RHS))
1996 if (SelectPatternFlavor SPF2 = matchSelectPattern(RHS, LHS2, RHS2).Flavor)
1997 if (Instruction *R = foldSPFofSPF(cast<Instruction>(RHS), SPF2, LHS2,
1998 RHS2, SI, SPF, LHS))
2001 // ABS(-X) -> ABS(X)
2004 if (SelectPatternResult::isMinOrMax(SPF)) {
2005 // Canonicalize so that
2006 // - type casts are outside select patterns.
2007 // - float clamp is transformed to min/max pattern
2009 bool IsCastNeeded = LHS->getType() != SelType;
2010 Value *CmpLHS = cast<CmpInst>(CondVal)->getOperand(0);
2011 Value *CmpRHS = cast<CmpInst>(CondVal)->getOperand(1);
2013 (LHS->getType()->isFPOrFPVectorTy() &&
2014 ((CmpLHS != LHS && CmpLHS != RHS) ||
2015 (CmpRHS != LHS && CmpRHS != RHS)))) {
2016 CmpInst::Predicate Pred = getMinMaxPred(SPF, SPR.Ordered);
2019 if (CmpInst::isIntPredicate(Pred)) {
2020 Cmp = Builder.CreateICmp(Pred, LHS, RHS);
2022 IRBuilder<>::FastMathFlagGuard FMFG(Builder);
2023 auto FMF = cast<FPMathOperator>(SI.getCondition())->getFastMathFlags();
2024 Builder.setFastMathFlags(FMF);
2025 Cmp = Builder.CreateFCmp(Pred, LHS, RHS);
2028 Value *NewSI = Builder.CreateSelect(Cmp, LHS, RHS, SI.getName(), &SI);
2030 return replaceInstUsesWith(SI, NewSI);
2032 Value *NewCast = Builder.CreateCast(CastOp, NewSI, SelType);
2033 return replaceInstUsesWith(SI, NewCast);
2036 // MAX(~a, ~b) -> ~MIN(a, b)
2037 // MAX(~a, C) -> ~MIN(a, ~C)
2038 // MIN(~a, ~b) -> ~MAX(a, b)
2039 // MIN(~a, C) -> ~MAX(a, ~C)
2040 auto moveNotAfterMinMax = [&](Value *X, Value *Y) -> Instruction * {
2042 if (match(X, m_Not(m_Value(A))) && !X->hasNUsesOrMore(3) &&
2043 !IsFreeToInvert(A, A->hasOneUse()) &&
2044 // Passing false to only consider m_Not and constants.
2045 IsFreeToInvert(Y, false)) {
2046 Value *B = Builder.CreateNot(Y);
2047 Value *NewMinMax = createMinMax(Builder, getInverseMinMaxFlavor(SPF),
2049 // Copy the profile metadata.
2050 if (MDNode *MD = SI.getMetadata(LLVMContext::MD_prof)) {
2051 cast<SelectInst>(NewMinMax)->setMetadata(LLVMContext::MD_prof, MD);
2052 // Swap the metadata if the operands are swapped.
2053 if (X == SI.getFalseValue() && Y == SI.getTrueValue())
2054 cast<SelectInst>(NewMinMax)->swapProfMetadata();
2057 return BinaryOperator::CreateNot(NewMinMax);
2063 if (Instruction *I = moveNotAfterMinMax(LHS, RHS))
2065 if (Instruction *I = moveNotAfterMinMax(RHS, LHS))
2068 if (Instruction *I = moveAddAfterMinMax(SPF, LHS, RHS, Builder))
2071 if (Instruction *I = factorizeMinMaxTree(SPF, LHS, RHS, Builder))
2076 // Canonicalize select of FP values where NaN and -0.0 are not valid as
2077 // minnum/maxnum intrinsics.
2078 if (isa<FPMathOperator>(SI) && SI.hasNoNaNs() && SI.hasNoSignedZeros()) {
2080 if (match(&SI, m_OrdFMax(m_Value(X), m_Value(Y))))
2081 return replaceInstUsesWith(
2082 SI, Builder.CreateBinaryIntrinsic(Intrinsic::maxnum, X, Y, &SI));
2084 if (match(&SI, m_OrdFMin(m_Value(X), m_Value(Y))))
2085 return replaceInstUsesWith(
2086 SI, Builder.CreateBinaryIntrinsic(Intrinsic::minnum, X, Y, &SI));
2089 // See if we can fold the select into a phi node if the condition is a select.
2090 if (auto *PN = dyn_cast<PHINode>(SI.getCondition()))
2091 // The true/false values have to be live in the PHI predecessor's blocks.
2092 if (canSelectOperandBeMappingIntoPredBlock(TrueVal, SI) &&
2093 canSelectOperandBeMappingIntoPredBlock(FalseVal, SI))
2094 if (Instruction *NV = foldOpIntoPhi(SI, PN))
2097 if (SelectInst *TrueSI = dyn_cast<SelectInst>(TrueVal)) {
2098 if (TrueSI->getCondition()->getType() == CondVal->getType()) {
2099 // select(C, select(C, a, b), c) -> select(C, a, c)
2100 if (TrueSI->getCondition() == CondVal) {
2101 if (SI.getTrueValue() == TrueSI->getTrueValue())
2103 SI.setOperand(1, TrueSI->getTrueValue());
2106 // select(C0, select(C1, a, b), b) -> select(C0&C1, a, b)
2107 // We choose this as normal form to enable folding on the And and shortening
2108 // paths for the values (this helps GetUnderlyingObjects() for example).
2109 if (TrueSI->getFalseValue() == FalseVal && TrueSI->hasOneUse()) {
2110 Value *And = Builder.CreateAnd(CondVal, TrueSI->getCondition());
2111 SI.setOperand(0, And);
2112 SI.setOperand(1, TrueSI->getTrueValue());
2117 if (SelectInst *FalseSI = dyn_cast<SelectInst>(FalseVal)) {
2118 if (FalseSI->getCondition()->getType() == CondVal->getType()) {
2119 // select(C, a, select(C, b, c)) -> select(C, a, c)
2120 if (FalseSI->getCondition() == CondVal) {
2121 if (SI.getFalseValue() == FalseSI->getFalseValue())
2123 SI.setOperand(2, FalseSI->getFalseValue());
2126 // select(C0, a, select(C1, a, b)) -> select(C0|C1, a, b)
2127 if (FalseSI->getTrueValue() == TrueVal && FalseSI->hasOneUse()) {
2128 Value *Or = Builder.CreateOr(CondVal, FalseSI->getCondition());
2129 SI.setOperand(0, Or);
2130 SI.setOperand(2, FalseSI->getFalseValue());
2136 auto canMergeSelectThroughBinop = [](BinaryOperator *BO) {
2137 // The select might be preventing a division by 0.
2138 switch (BO->getOpcode()) {
2141 case Instruction::SRem:
2142 case Instruction::URem:
2143 case Instruction::SDiv:
2144 case Instruction::UDiv:
2149 // Try to simplify a binop sandwiched between 2 selects with the same
2151 // select(C, binop(select(C, X, Y), W), Z) -> select(C, binop(X, W), Z)
2152 BinaryOperator *TrueBO;
2153 if (match(TrueVal, m_OneUse(m_BinOp(TrueBO))) &&
2154 canMergeSelectThroughBinop(TrueBO)) {
2155 if (auto *TrueBOSI = dyn_cast<SelectInst>(TrueBO->getOperand(0))) {
2156 if (TrueBOSI->getCondition() == CondVal) {
2157 TrueBO->setOperand(0, TrueBOSI->getTrueValue());
2158 Worklist.Add(TrueBO);
2162 if (auto *TrueBOSI = dyn_cast<SelectInst>(TrueBO->getOperand(1))) {
2163 if (TrueBOSI->getCondition() == CondVal) {
2164 TrueBO->setOperand(1, TrueBOSI->getTrueValue());
2165 Worklist.Add(TrueBO);
2171 // select(C, Z, binop(select(C, X, Y), W)) -> select(C, Z, binop(Y, W))
2172 BinaryOperator *FalseBO;
2173 if (match(FalseVal, m_OneUse(m_BinOp(FalseBO))) &&
2174 canMergeSelectThroughBinop(FalseBO)) {
2175 if (auto *FalseBOSI = dyn_cast<SelectInst>(FalseBO->getOperand(0))) {
2176 if (FalseBOSI->getCondition() == CondVal) {
2177 FalseBO->setOperand(0, FalseBOSI->getFalseValue());
2178 Worklist.Add(FalseBO);
2182 if (auto *FalseBOSI = dyn_cast<SelectInst>(FalseBO->getOperand(1))) {
2183 if (FalseBOSI->getCondition() == CondVal) {
2184 FalseBO->setOperand(1, FalseBOSI->getFalseValue());
2185 Worklist.Add(FalseBO);
2192 if (match(CondVal, m_Not(m_Value(NotCond)))) {
2193 SI.setOperand(0, NotCond);
2194 SI.setOperand(1, FalseVal);
2195 SI.setOperand(2, TrueVal);
2196 SI.swapProfMetadata();
2200 if (VectorType *VecTy = dyn_cast<VectorType>(SelType)) {
2201 unsigned VWidth = VecTy->getNumElements();
2202 APInt UndefElts(VWidth, 0);
2203 APInt AllOnesEltMask(APInt::getAllOnesValue(VWidth));
2204 if (Value *V = SimplifyDemandedVectorElts(&SI, AllOnesEltMask, UndefElts)) {
2206 return replaceInstUsesWith(SI, V);
2211 // If we can compute the condition, there's no need for a select.
2212 // Like the above fold, we are attempting to reduce compile-time cost by
2213 // putting this fold here with limitations rather than in InstSimplify.
2214 // The motivation for this call into value tracking is to take advantage of
2215 // the assumption cache, so make sure that is populated.
2216 if (!CondVal->getType()->isVectorTy() && !AC.assumptions().empty()) {
2218 computeKnownBits(CondVal, Known, 0, &SI);
2219 if (Known.One.isOneValue())
2220 return replaceInstUsesWith(SI, TrueVal);
2221 if (Known.Zero.isOneValue())
2222 return replaceInstUsesWith(SI, FalseVal);
2225 if (Instruction *BitCastSel = foldSelectCmpBitcasts(SI, Builder))
2228 // Simplify selects that test the returned flag of cmpxchg instructions.
2229 if (Instruction *Select = foldSelectCmpXchg(SI))
2232 if (Instruction *Select = foldSelectBinOpIdentity(SI, TLI))
2235 if (Instruction *Rot = foldSelectRotate(SI))