[InstCombine] Signed saturation patterns
[llvm-core.git] / lib / Transforms / InstCombine / InstCombineSelect.cpp
blob9fc871e49b303076dc584cca9ca0b6a01156b2b5
1 //===- InstCombineSelect.cpp ----------------------------------------------===//
2 //
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
6 //
7 //===----------------------------------------------------------------------===//
8 //
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"
41 #include <cassert>
42 #include <utility>
44 using namespace llvm;
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.
61 Value *X;
62 Constant *C;
63 CmpInst::Predicate Pred;
64 if (!match(Sel.getCondition(), m_Cmp(Pred, m_Value(X), m_Constant(C))))
65 return nullptr;
67 bool IsEq;
68 if (ICmpInst::isEquality(Pred))
69 IsEq = Pred == ICmpInst::ICMP_EQ;
70 else if (Pred == FCmpInst::FCMP_OEQ)
71 IsEq = true;
72 else if (Pred == FCmpInst::FCMP_UNE)
73 IsEq = false;
74 else
75 return nullptr;
77 // A select operand must be a binop.
78 BinaryOperator *BO;
79 if (!match(Sel.getOperand(IsEq ? 1 : 2), m_BinOp(BO)))
80 return nullptr;
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);
86 if (IdC != C) {
87 if (!IdC || !CmpInst::isFPPredicate(Pred))
88 return nullptr;
89 if (!match(IdC, m_AnyZeroFP()) || !match(C, m_AnyZeroFP()))
90 return nullptr;
93 // Last, match the compare variable operand with a binop operand.
94 Value *Y;
95 if (!BO->isCommutative() && !match(BO, m_BinOp(m_Value(Y), m_Specific(X))))
96 return nullptr;
97 if (!match(BO, m_c_BinOp(m_Value(Y), m_Specific(X))))
98 return nullptr;
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))
104 return nullptr;
106 // BO = binop Y, X
107 // S = { select (cmp eq X, C), BO, ? } or { select (cmp ne X, C), ?, BO }
108 // =>
109 // S = { select (cmp eq X, C), Y, ? } or { select (cmp ne X, C), ?, Y }
110 Sel.setOperand(IsEq ? 1 : 2, Y);
111 return &Sel;
114 /// This folds:
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
119 /// Or:
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)))
128 return nullptr;
130 // If this is a vector select, we need a vector compare.
131 Type *SelType = Sel.getType();
132 if (SelType->isVectorTy() != Cmp->getType()->isVectorTy())
133 return nullptr;
135 Value *V;
136 APInt AndMask;
137 bool CreateAnd = false;
138 ICmpInst::Predicate Pred = Cmp->getPredicate();
139 if (ICmpInst::isEquality(Pred)) {
140 if (!match(Cmp->getOperand(1), m_Zero()))
141 return nullptr;
143 V = Cmp->getOperand(0);
144 const APInt *AndRHS;
145 if (!match(V, m_And(m_Value(), m_Power2(AndRHS))))
146 return nullptr;
148 AndMask = *AndRHS;
149 } else if (decomposeBitTestICmp(Cmp->getOperand(0), Cmp->getOperand(1),
150 Pred, V, AndMask)) {
151 assert(ICmpInst::isEquality(Pred) && "Not equality test?");
152 if (!AndMask.isPowerOf2())
153 return nullptr;
155 CreateAnd = true;
156 } else {
157 return nullptr;
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.
164 APInt TC = *SelTC;
165 APInt FC = *SelFC;
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)
171 return nullptr;
172 if (CreateAnd) {
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())
176 return nullptr;
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())
199 return nullptr;
201 // Determine which shift is needed to transform result of the 'and' into the
202 // desired result.
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.
208 if (CreateAnd)
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);
219 } else {
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;
227 if (ShouldNotVal)
228 V = Builder.CreateXor(V, ValC);
230 return V;
233 /// We want to turn code that looks like this:
234 /// %C = or %A, %B
235 /// %D = select %cond, %C, %A
236 /// into:
237 /// %C = select %cond, %B, 0
238 /// %D = or %A, %C
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:
255 return 1;
256 default:
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,
283 Instruction *FI) {
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()))))
292 return nullptr;
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)
300 return nullptr;
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())
306 return nullptr;
307 if (CondTy->getVectorNumElements() != FIOpndTy->getVectorNumElements())
308 return nullptr;
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
313 // (PR28160).
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()))
320 return nullptr;
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.
325 return nullptr;
328 // Fold this by inserting a select from the input values.
329 Value *NewSI =
330 Builder.CreateSelect(Cond, TI->getOperand(0), FI->getOperand(0),
331 SI.getName() + ".v", &SI);
332 return CastInst::Create(Instruction::CastOps(TI->getOpcode()), NewSI,
333 TI->getType());
336 // Cond ? -X : -Y --> -(Cond ? X : Y)
337 Value *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())
355 return nullptr;
357 // Figure out if the operations have any operands in common.
358 Value *MatchOp, *OtherOpT, *OtherOpF;
359 bool MatchIsOpZero;
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()) {
371 return nullptr;
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;
382 } else {
383 return nullptr;
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
388 // for example.
389 if (CondTy->isVectorTy() && (!OtherOpT->getType()->isVectorTy() ||
390 !OtherOpF->getType()->isVectorTy()))
391 return nullptr;
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);
402 return NewBO;
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");
412 return nullptr;
415 static bool isSelect01(const APInt &C1I, const APInt &C2I) {
416 if (!C1I.isNullValue() && !C2I.isNullValue()) // One side must be zero.
417 return false;
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
423 /// optimization.
424 Instruction *InstCombiner::foldSelectIntoOp(SelectInst &SI, Value *TrueVal,
425 Value *FalseVal) {
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)) {
433 OpToFold = 1;
434 } else if ((SFO & 2) && FalseVal == TVI->getOperand(1)) {
435 OpToFold = 2;
438 if (OpToFold) {
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.
443 const APInt *OOpC;
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(),
450 FalseVal, NewSel);
451 BO->copyIRFlags(TVI);
452 return BO;
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)) {
464 OpToFold = 1;
465 } else if ((SFO & 2) && TrueVal == FVI->getOperand(1)) {
466 OpToFold = 2;
469 if (OpToFold) {
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.
474 const APInt *OOpC;
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(),
481 TrueVal, NewSel);
482 BO->copyIRFlags(FVI);
483 return BO;
490 return nullptr;
493 /// We want to turn:
494 /// (select (icmp eq (and X, Y), 0), (and (lshr X, Z), 1), 1)
495 /// into:
496 /// zext (icmp ne i32 (and X, (or Y, (shl 1, Z))), 0)
497 /// Note:
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())))
507 return nullptr;
509 // The TrueVal has general form of: and %B, 1
510 Value *B;
511 if (!match(TVal, m_OneUse(m_And(m_Value(B), m_One()))))
512 return nullptr;
514 // Where %B may be optionally shifted: lshr %X, %Z.
515 Value *X, *Z;
516 const bool HasShift = match(B, m_OneUse(m_LShr(m_Value(X), m_Value(Z))));
517 if (!HasShift)
518 X = B;
520 Value *Y;
521 if (!match(Cmp->getOperand(0), m_c_And(m_Specific(X), m_Value(Y))))
522 return nullptr;
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);
534 /// We want to turn:
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
537 /// into:
538 /// ashr (X, Y)
539 static Value *foldSelectICmpLshrAshr(const ICmpInst *IC, Value *TrueVal,
540 Value *FalseVal,
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())
546 return nullptr;
548 Value *X, *Y;
549 unsigned Bitwidth = CmpRHS->getType()->getScalarSizeInBits();
550 if ((Pred != ICmpInst::ICMP_SGT ||
551 !match(CmpRHS,
552 m_SpecificInt_ICMP(ICmpInst::ICMP_SGE, APInt(Bitwidth, -1)))) &&
553 (Pred != ICmpInst::ICMP_SLT ||
554 !match(CmpRHS,
555 m_SpecificInt_ICMP(ICmpInst::ICMP_SGE, APInt(Bitwidth, 0)))))
556 return nullptr;
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);
571 return nullptr;
574 /// We want to turn:
575 /// (select (icmp eq (and X, C1), 0), Y, (or Y, C2))
576 /// into:
577 /// (or (shl (and X, C1), C3), Y)
578 /// iff:
579 /// C1 and C2 are both powers of 2
580 /// where:
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,
588 Value *FalseVal,
589 InstCombiner::BuilderTy &Builder) {
590 // Only handle integer compares. Also, if this is a vector select, we need a
591 // vector compare.
592 if (!TrueVal->getType()->isIntOrIntVectorTy() ||
593 TrueVal->getType()->isVectorTy() != IC->getType()->isVectorTy())
594 return nullptr;
596 Value *CmpLHS = IC->getOperand(0);
597 Value *CmpRHS = IC->getOperand(1);
599 Value *V;
600 unsigned C1Log;
601 bool IsEqualZero;
602 bool NeedAnd = false;
603 if (IC->isEquality()) {
604 if (!match(CmpRHS, m_Zero()))
605 return nullptr;
607 const APInt *C1;
608 if (!match(CmpLHS, m_And(m_Value(), m_Power2(C1))))
609 return nullptr;
611 V = CmpLHS;
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())))
621 return nullptr;
623 if (!match(CmpLHS, m_OneUse(m_Trunc(m_Value(V)))))
624 return nullptr;
626 C1Log = CmpLHS->getType()->getScalarSizeInBits() - 1;
627 NeedAnd = true;
628 } else {
629 return nullptr;
632 const APInt *C2;
633 bool OrOnTrueVal = false;
634 bool OrOnFalseVal = match(FalseVal, m_Or(m_Specific(TrueVal), m_Power2(C2)));
635 if (!OrOnFalseVal)
636 OrOnTrueVal = match(TrueVal, m_Or(m_Specific(FalseVal), m_Power2(C2)));
638 if (!OrOnFalseVal && !OrOnTrueVal)
639 return nullptr;
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()))
654 return nullptr;
656 if (NeedAnd) {
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));
662 if (C2Log > C1Log) {
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());
668 } else
669 V = Builder.CreateZExtOrTrunc(V, Y->getType());
671 if (NeedXor)
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))
686 return nullptr;
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()))
694 return nullptr;
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
700 std::swap(A, B);
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))))
710 IsNegative = true;
711 else if (!match(TrueVal, m_Sub(m_Specific(A), m_Specific(B))))
712 return nullptr;
714 // If sub is used anywhere else, we wouldn't be able to eliminate it
715 // afterwards.
716 if (!TrueVal->hasOneUse())
717 return nullptr;
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);
722 if (IsNegative)
723 Result = Builder.CreateNeg(Result);
724 return Result;
727 static Value *canonicalizeSaturatedAdd(ICmpInst *Cmp, Value *TVal, Value *FVal,
728 InstCombiner::BuilderTy &Builder) {
729 if (!Cmp->hasOneUse())
730 return nullptr;
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();
736 Value *X;
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()))
756 return nullptr;
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)
764 return nullptr;
766 // Match unsigned saturated add of 2 variables with an unnecessary 'not'.
767 Value *Y;
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.
775 X = Cmp0;
776 Y = Cmp1;
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));
785 return nullptr;
788 /// Fold the following code sequence:
789 /// \code
790 /// int a = ctlz(x & -x);
791 // x ? 31 - a : a;
792 /// \code
794 /// into:
795 /// cttz(x)
796 static Instruction *foldSelectCtlzToCttz(ICmpInst *ICI, Value *TrueVal,
797 Value *FalseVal,
798 InstCombiner::BuilderTy &Builder) {
799 unsigned BitWidth = TrueVal->getType()->getScalarSizeInBits();
800 if (!ICI->isEquality() || !match(ICI->getOperand(1), m_Zero()))
801 return nullptr;
803 if (ICI->getPredicate() == ICmpInst::ICMP_NE)
804 std::swap(TrueVal, FalseVal);
806 if (!match(FalseVal,
807 m_Xor(m_Deferred(TrueVal), m_SpecificInt(BitWidth - 1))))
808 return nullptr;
810 if (!match(TrueVal, m_Intrinsic<Intrinsic::ctlz>()))
811 return nullptr;
813 Value *X = ICI->getOperand(0);
814 auto *II = cast<IntrinsicInst>(TrueVal);
815 if (!match(II->getOperand(0), m_c_And(m_Specific(X), m_Neg(m_Specific(X)))))
816 return nullptr;
818 Function *F = Intrinsic::getDeclaration(II->getModule(), Intrinsic::cttz,
819 II->getType());
820 return CallInst::Create(F, {X, II->getArgOperand(1)});
823 /// Attempt to fold a cttz/ctlz followed by a icmp plus select into a single
824 /// call to cttz/ctlz with flag 'is_zero_undef' cleared.
826 /// For example, we can fold the following code sequence:
827 /// \code
828 /// %0 = tail call i32 @llvm.cttz.i32(i32 %x, i1 true)
829 /// %1 = icmp ne i32 %x, 0
830 /// %2 = select i1 %1, i32 %0, i32 32
831 /// \code
833 /// into:
834 /// %0 = tail call i32 @llvm.cttz.i32(i32 %x, i1 false)
835 static Value *foldSelectCttzCtlz(ICmpInst *ICI, Value *TrueVal, Value *FalseVal,
836 InstCombiner::BuilderTy &Builder) {
837 ICmpInst::Predicate Pred = ICI->getPredicate();
838 Value *CmpLHS = ICI->getOperand(0);
839 Value *CmpRHS = ICI->getOperand(1);
841 // Check if the condition value compares a value for equality against zero.
842 if (!ICI->isEquality() || !match(CmpRHS, m_Zero()))
843 return nullptr;
845 Value *Count = FalseVal;
846 Value *ValueOnZero = TrueVal;
847 if (Pred == ICmpInst::ICMP_NE)
848 std::swap(Count, ValueOnZero);
850 // Skip zero extend/truncate.
851 Value *V = nullptr;
852 if (match(Count, m_ZExt(m_Value(V))) ||
853 match(Count, m_Trunc(m_Value(V))))
854 Count = V;
856 // Check that 'Count' is a call to intrinsic cttz/ctlz. Also check that the
857 // input to the cttz/ctlz is used as LHS for the compare instruction.
858 if (!match(Count, m_Intrinsic<Intrinsic::cttz>(m_Specific(CmpLHS))) &&
859 !match(Count, m_Intrinsic<Intrinsic::ctlz>(m_Specific(CmpLHS))))
860 return nullptr;
862 IntrinsicInst *II = cast<IntrinsicInst>(Count);
864 // Check if the value propagated on zero is a constant number equal to the
865 // sizeof in bits of 'Count'.
866 unsigned SizeOfInBits = Count->getType()->getScalarSizeInBits();
867 if (match(ValueOnZero, m_SpecificInt(SizeOfInBits))) {
868 // Explicitly clear the 'undef_on_zero' flag.
869 IntrinsicInst *NewI = cast<IntrinsicInst>(II->clone());
870 NewI->setArgOperand(1, ConstantInt::getFalse(NewI->getContext()));
871 Builder.Insert(NewI);
872 return Builder.CreateZExtOrTrunc(NewI, ValueOnZero->getType());
875 // If the ValueOnZero is not the bitwidth, we can at least make use of the
876 // fact that the cttz/ctlz result will not be used if the input is zero, so
877 // it's okay to relax it to undef for that case.
878 if (II->hasOneUse() && !match(II->getArgOperand(1), m_One()))
879 II->setArgOperand(1, ConstantInt::getTrue(II->getContext()));
881 return nullptr;
884 /// Return true if we find and adjust an icmp+select pattern where the compare
885 /// is with a constant that can be incremented or decremented to match the
886 /// minimum or maximum idiom.
887 static bool adjustMinMax(SelectInst &Sel, ICmpInst &Cmp) {
888 ICmpInst::Predicate Pred = Cmp.getPredicate();
889 Value *CmpLHS = Cmp.getOperand(0);
890 Value *CmpRHS = Cmp.getOperand(1);
891 Value *TrueVal = Sel.getTrueValue();
892 Value *FalseVal = Sel.getFalseValue();
894 // We may move or edit the compare, so make sure the select is the only user.
895 const APInt *CmpC;
896 if (!Cmp.hasOneUse() || !match(CmpRHS, m_APInt(CmpC)))
897 return false;
899 // These transforms only work for selects of integers or vector selects of
900 // integer vectors.
901 Type *SelTy = Sel.getType();
902 auto *SelEltTy = dyn_cast<IntegerType>(SelTy->getScalarType());
903 if (!SelEltTy || SelTy->isVectorTy() != Cmp.getType()->isVectorTy())
904 return false;
906 Constant *AdjustedRHS;
907 if (Pred == ICmpInst::ICMP_UGT || Pred == ICmpInst::ICMP_SGT)
908 AdjustedRHS = ConstantInt::get(CmpRHS->getType(), *CmpC + 1);
909 else if (Pred == ICmpInst::ICMP_ULT || Pred == ICmpInst::ICMP_SLT)
910 AdjustedRHS = ConstantInt::get(CmpRHS->getType(), *CmpC - 1);
911 else
912 return false;
914 // X > C ? X : C+1 --> X < C+1 ? C+1 : X
915 // X < C ? X : C-1 --> X > C-1 ? C-1 : X
916 if ((CmpLHS == TrueVal && AdjustedRHS == FalseVal) ||
917 (CmpLHS == FalseVal && AdjustedRHS == TrueVal)) {
918 ; // Nothing to do here. Values match without any sign/zero extension.
920 // Types do not match. Instead of calculating this with mixed types, promote
921 // all to the larger type. This enables scalar evolution to analyze this
922 // expression.
923 else if (CmpRHS->getType()->getScalarSizeInBits() < SelEltTy->getBitWidth()) {
924 Constant *SextRHS = ConstantExpr::getSExt(AdjustedRHS, SelTy);
926 // X = sext x; x >s c ? X : C+1 --> X = sext x; X <s C+1 ? C+1 : X
927 // X = sext x; x <s c ? X : C-1 --> X = sext x; X >s C-1 ? C-1 : X
928 // X = sext x; x >u c ? X : C+1 --> X = sext x; X <u C+1 ? C+1 : X
929 // X = sext x; x <u c ? X : C-1 --> X = sext x; X >u C-1 ? C-1 : X
930 if (match(TrueVal, m_SExt(m_Specific(CmpLHS))) && SextRHS == FalseVal) {
931 CmpLHS = TrueVal;
932 AdjustedRHS = SextRHS;
933 } else if (match(FalseVal, m_SExt(m_Specific(CmpLHS))) &&
934 SextRHS == TrueVal) {
935 CmpLHS = FalseVal;
936 AdjustedRHS = SextRHS;
937 } else if (Cmp.isUnsigned()) {
938 Constant *ZextRHS = ConstantExpr::getZExt(AdjustedRHS, SelTy);
939 // X = zext x; x >u c ? X : C+1 --> X = zext x; X <u C+1 ? C+1 : X
940 // X = zext x; x <u c ? X : C-1 --> X = zext x; X >u C-1 ? C-1 : X
941 // zext + signed compare cannot be changed:
942 // 0xff <s 0x00, but 0x00ff >s 0x0000
943 if (match(TrueVal, m_ZExt(m_Specific(CmpLHS))) && ZextRHS == FalseVal) {
944 CmpLHS = TrueVal;
945 AdjustedRHS = ZextRHS;
946 } else if (match(FalseVal, m_ZExt(m_Specific(CmpLHS))) &&
947 ZextRHS == TrueVal) {
948 CmpLHS = FalseVal;
949 AdjustedRHS = ZextRHS;
950 } else {
951 return false;
953 } else {
954 return false;
956 } else {
957 return false;
960 Pred = ICmpInst::getSwappedPredicate(Pred);
961 CmpRHS = AdjustedRHS;
962 std::swap(FalseVal, TrueVal);
963 Cmp.setPredicate(Pred);
964 Cmp.setOperand(0, CmpLHS);
965 Cmp.setOperand(1, CmpRHS);
966 Sel.setOperand(1, TrueVal);
967 Sel.setOperand(2, FalseVal);
968 Sel.swapProfMetadata();
970 // Move the compare instruction right before the select instruction. Otherwise
971 // the sext/zext value may be defined after the compare instruction uses it.
972 Cmp.moveBefore(&Sel);
974 return true;
977 /// If this is an integer min/max (icmp + select) with a constant operand,
978 /// create the canonical icmp for the min/max operation and canonicalize the
979 /// constant to the 'false' operand of the select:
980 /// select (icmp Pred X, C1), C2, X --> select (icmp Pred' X, C2), X, C2
981 /// Note: if C1 != C2, this will change the icmp constant to the existing
982 /// constant operand of the select.
983 static Instruction *
984 canonicalizeMinMaxWithConstant(SelectInst &Sel, ICmpInst &Cmp,
985 InstCombiner::BuilderTy &Builder) {
986 if (!Cmp.hasOneUse() || !isa<Constant>(Cmp.getOperand(1)))
987 return nullptr;
989 // Canonicalize the compare predicate based on whether we have min or max.
990 Value *LHS, *RHS;
991 SelectPatternResult SPR = matchSelectPattern(&Sel, LHS, RHS);
992 if (!SelectPatternResult::isMinOrMax(SPR.Flavor))
993 return nullptr;
995 // Is this already canonical?
996 ICmpInst::Predicate CanonicalPred = getMinMaxPred(SPR.Flavor);
997 if (Cmp.getOperand(0) == LHS && Cmp.getOperand(1) == RHS &&
998 Cmp.getPredicate() == CanonicalPred)
999 return nullptr;
1001 // Create the canonical compare and plug it into the select.
1002 Sel.setCondition(Builder.CreateICmp(CanonicalPred, LHS, RHS));
1004 // If the select operands did not change, we're done.
1005 if (Sel.getTrueValue() == LHS && Sel.getFalseValue() == RHS)
1006 return &Sel;
1008 // If we are swapping the select operands, swap the metadata too.
1009 assert(Sel.getTrueValue() == RHS && Sel.getFalseValue() == LHS &&
1010 "Unexpected results from matchSelectPattern");
1011 Sel.swapValues();
1012 Sel.swapProfMetadata();
1013 return &Sel;
1016 /// There are many select variants for each of ABS/NABS.
1017 /// In matchSelectPattern(), there are different compare constants, compare
1018 /// predicates/operands and select operands.
1019 /// In isKnownNegation(), there are different formats of negated operands.
1020 /// Canonicalize all these variants to 1 pattern.
1021 /// This makes CSE more likely.
1022 static Instruction *canonicalizeAbsNabs(SelectInst &Sel, ICmpInst &Cmp,
1023 InstCombiner::BuilderTy &Builder) {
1024 if (!Cmp.hasOneUse() || !isa<Constant>(Cmp.getOperand(1)))
1025 return nullptr;
1027 // Choose a sign-bit check for the compare (likely simpler for codegen).
1028 // ABS: (X <s 0) ? -X : X
1029 // NABS: (X <s 0) ? X : -X
1030 Value *LHS, *RHS;
1031 SelectPatternFlavor SPF = matchSelectPattern(&Sel, LHS, RHS).Flavor;
1032 if (SPF != SelectPatternFlavor::SPF_ABS &&
1033 SPF != SelectPatternFlavor::SPF_NABS)
1034 return nullptr;
1036 Value *TVal = Sel.getTrueValue();
1037 Value *FVal = Sel.getFalseValue();
1038 assert(isKnownNegation(TVal, FVal) &&
1039 "Unexpected result from matchSelectPattern");
1041 // The compare may use the negated abs()/nabs() operand, or it may use
1042 // negation in non-canonical form such as: sub A, B.
1043 bool CmpUsesNegatedOp = match(Cmp.getOperand(0), m_Neg(m_Specific(TVal))) ||
1044 match(Cmp.getOperand(0), m_Neg(m_Specific(FVal)));
1046 bool CmpCanonicalized = !CmpUsesNegatedOp &&
1047 match(Cmp.getOperand(1), m_ZeroInt()) &&
1048 Cmp.getPredicate() == ICmpInst::ICMP_SLT;
1049 bool RHSCanonicalized = match(RHS, m_Neg(m_Specific(LHS)));
1051 // Is this already canonical?
1052 if (CmpCanonicalized && RHSCanonicalized)
1053 return nullptr;
1055 // If RHS is used by other instructions except compare and select, don't
1056 // canonicalize it to not increase the instruction count.
1057 if (!(RHS->hasOneUse() || (RHS->hasNUses(2) && CmpUsesNegatedOp)))
1058 return nullptr;
1060 // Create the canonical compare: icmp slt LHS 0.
1061 if (!CmpCanonicalized) {
1062 Cmp.setPredicate(ICmpInst::ICMP_SLT);
1063 Cmp.setOperand(1, ConstantInt::getNullValue(Cmp.getOperand(0)->getType()));
1064 if (CmpUsesNegatedOp)
1065 Cmp.setOperand(0, LHS);
1068 // Create the canonical RHS: RHS = sub (0, LHS).
1069 if (!RHSCanonicalized) {
1070 assert(RHS->hasOneUse() && "RHS use number is not right");
1071 RHS = Builder.CreateNeg(LHS);
1072 if (TVal == LHS) {
1073 Sel.setFalseValue(RHS);
1074 FVal = RHS;
1075 } else {
1076 Sel.setTrueValue(RHS);
1077 TVal = RHS;
1081 // If the select operands do not change, we're done.
1082 if (SPF == SelectPatternFlavor::SPF_NABS) {
1083 if (TVal == LHS)
1084 return &Sel;
1085 assert(FVal == LHS && "Unexpected results from matchSelectPattern");
1086 } else {
1087 if (FVal == LHS)
1088 return &Sel;
1089 assert(TVal == LHS && "Unexpected results from matchSelectPattern");
1092 // We are swapping the select operands, so swap the metadata too.
1093 Sel.swapValues();
1094 Sel.swapProfMetadata();
1095 return &Sel;
1098 static Value *simplifyWithOpReplaced(Value *V, Value *Op, Value *ReplaceOp,
1099 const SimplifyQuery &Q) {
1100 // If this is a binary operator, try to simplify it with the replaced op
1101 // because we know Op and ReplaceOp are equivalant.
1102 // For example: V = X + 1, Op = X, ReplaceOp = 42
1103 // Simplifies as: add(42, 1) --> 43
1104 if (auto *BO = dyn_cast<BinaryOperator>(V)) {
1105 if (BO->getOperand(0) == Op)
1106 return SimplifyBinOp(BO->getOpcode(), ReplaceOp, BO->getOperand(1), Q);
1107 if (BO->getOperand(1) == Op)
1108 return SimplifyBinOp(BO->getOpcode(), BO->getOperand(0), ReplaceOp, Q);
1111 return nullptr;
1114 /// If we have a select with an equality comparison, then we know the value in
1115 /// one of the arms of the select. See if substituting this value into an arm
1116 /// and simplifying the result yields the same value as the other arm.
1118 /// To make this transform safe, we must drop poison-generating flags
1119 /// (nsw, etc) if we simplified to a binop because the select may be guarding
1120 /// that poison from propagating. If the existing binop already had no
1121 /// poison-generating flags, then this transform can be done by instsimplify.
1123 /// Consider:
1124 /// %cmp = icmp eq i32 %x, 2147483647
1125 /// %add = add nsw i32 %x, 1
1126 /// %sel = select i1 %cmp, i32 -2147483648, i32 %add
1128 /// We can't replace %sel with %add unless we strip away the flags.
1129 /// TODO: Wrapping flags could be preserved in some cases with better analysis.
1130 static Value *foldSelectValueEquivalence(SelectInst &Sel, ICmpInst &Cmp,
1131 const SimplifyQuery &Q) {
1132 if (!Cmp.isEquality())
1133 return nullptr;
1135 // Canonicalize the pattern to ICMP_EQ by swapping the select operands.
1136 Value *TrueVal = Sel.getTrueValue(), *FalseVal = Sel.getFalseValue();
1137 if (Cmp.getPredicate() == ICmpInst::ICMP_NE)
1138 std::swap(TrueVal, FalseVal);
1140 // Try each equivalence substitution possibility.
1141 // We have an 'EQ' comparison, so the select's false value will propagate.
1142 // Example:
1143 // (X == 42) ? 43 : (X + 1) --> (X == 42) ? (X + 1) : (X + 1) --> X + 1
1144 // (X == 42) ? (X + 1) : 43 --> (X == 42) ? (42 + 1) : 43 --> 43
1145 Value *CmpLHS = Cmp.getOperand(0), *CmpRHS = Cmp.getOperand(1);
1146 if (simplifyWithOpReplaced(FalseVal, CmpLHS, CmpRHS, Q) == TrueVal ||
1147 simplifyWithOpReplaced(FalseVal, CmpRHS, CmpLHS, Q) == TrueVal ||
1148 simplifyWithOpReplaced(TrueVal, CmpLHS, CmpRHS, Q) == FalseVal ||
1149 simplifyWithOpReplaced(TrueVal, CmpRHS, CmpLHS, Q) == FalseVal) {
1150 if (auto *FalseInst = dyn_cast<Instruction>(FalseVal))
1151 FalseInst->dropPoisonGeneratingFlags();
1152 return FalseVal;
1154 return nullptr;
1157 // See if this is a pattern like:
1158 // %old_cmp1 = icmp slt i32 %x, C2
1159 // %old_replacement = select i1 %old_cmp1, i32 %target_low, i32 %target_high
1160 // %old_x_offseted = add i32 %x, C1
1161 // %old_cmp0 = icmp ult i32 %old_x_offseted, C0
1162 // %r = select i1 %old_cmp0, i32 %x, i32 %old_replacement
1163 // This can be rewritten as more canonical pattern:
1164 // %new_cmp1 = icmp slt i32 %x, -C1
1165 // %new_cmp2 = icmp sge i32 %x, C0-C1
1166 // %new_clamped_low = select i1 %new_cmp1, i32 %target_low, i32 %x
1167 // %r = select i1 %new_cmp2, i32 %target_high, i32 %new_clamped_low
1168 // Iff -C1 s<= C2 s<= C0-C1
1169 // Also ULT predicate can also be UGT iff C0 != -1 (+invert result)
1170 // SLT predicate can also be SGT iff C2 != INT_MAX (+invert res.)
1171 static Instruction *canonicalizeClampLike(SelectInst &Sel0, ICmpInst &Cmp0,
1172 InstCombiner::BuilderTy &Builder) {
1173 Value *X = Sel0.getTrueValue();
1174 Value *Sel1 = Sel0.getFalseValue();
1176 // First match the condition of the outermost select.
1177 // Said condition must be one-use.
1178 if (!Cmp0.hasOneUse())
1179 return nullptr;
1180 Value *Cmp00 = Cmp0.getOperand(0);
1181 Constant *C0;
1182 if (!match(Cmp0.getOperand(1),
1183 m_CombineAnd(m_AnyIntegralConstant(), m_Constant(C0))))
1184 return nullptr;
1185 // Canonicalize Cmp0 into the form we expect.
1186 // FIXME: we shouldn't care about lanes that are 'undef' in the end?
1187 switch (Cmp0.getPredicate()) {
1188 case ICmpInst::Predicate::ICMP_ULT:
1189 break; // Great!
1190 case ICmpInst::Predicate::ICMP_ULE:
1191 // We'd have to increment C0 by one, and for that it must not have all-ones
1192 // element, but then it would have been canonicalized to 'ult' before
1193 // we get here. So we can't do anything useful with 'ule'.
1194 return nullptr;
1195 case ICmpInst::Predicate::ICMP_UGT:
1196 // We want to canonicalize it to 'ult', so we'll need to increment C0,
1197 // which again means it must not have any all-ones elements.
1198 if (!match(C0,
1199 m_SpecificInt_ICMP(ICmpInst::Predicate::ICMP_NE,
1200 APInt::getAllOnesValue(
1201 C0->getType()->getScalarSizeInBits()))))
1202 return nullptr; // Can't do, have all-ones element[s].
1203 C0 = AddOne(C0);
1204 std::swap(X, Sel1);
1205 break;
1206 case ICmpInst::Predicate::ICMP_UGE:
1207 // The only way we'd get this predicate if this `icmp` has extra uses,
1208 // but then we won't be able to do this fold.
1209 return nullptr;
1210 default:
1211 return nullptr; // Unknown predicate.
1214 // Now that we've canonicalized the ICmp, we know the X we expect;
1215 // the select in other hand should be one-use.
1216 if (!Sel1->hasOneUse())
1217 return nullptr;
1219 // We now can finish matching the condition of the outermost select:
1220 // it should either be the X itself, or an addition of some constant to X.
1221 Constant *C1;
1222 if (Cmp00 == X)
1223 C1 = ConstantInt::getNullValue(Sel0.getType());
1224 else if (!match(Cmp00,
1225 m_Add(m_Specific(X),
1226 m_CombineAnd(m_AnyIntegralConstant(), m_Constant(C1)))))
1227 return nullptr;
1229 Value *Cmp1;
1230 ICmpInst::Predicate Pred1;
1231 Constant *C2;
1232 Value *ReplacementLow, *ReplacementHigh;
1233 if (!match(Sel1, m_Select(m_Value(Cmp1), m_Value(ReplacementLow),
1234 m_Value(ReplacementHigh))) ||
1235 !match(Cmp1,
1236 m_ICmp(Pred1, m_Specific(X),
1237 m_CombineAnd(m_AnyIntegralConstant(), m_Constant(C2)))))
1238 return nullptr;
1240 if (!Cmp1->hasOneUse() && (Cmp00 == X || !Cmp00->hasOneUse()))
1241 return nullptr; // Not enough one-use instructions for the fold.
1242 // FIXME: this restriction could be relaxed if Cmp1 can be reused as one of
1243 // two comparisons we'll need to build.
1245 // Canonicalize Cmp1 into the form we expect.
1246 // FIXME: we shouldn't care about lanes that are 'undef' in the end?
1247 switch (Pred1) {
1248 case ICmpInst::Predicate::ICMP_SLT:
1249 break;
1250 case ICmpInst::Predicate::ICMP_SLE:
1251 // We'd have to increment C2 by one, and for that it must not have signed
1252 // max element, but then it would have been canonicalized to 'slt' before
1253 // we get here. So we can't do anything useful with 'sle'.
1254 return nullptr;
1255 case ICmpInst::Predicate::ICMP_SGT:
1256 // We want to canonicalize it to 'slt', so we'll need to increment C2,
1257 // which again means it must not have any signed max elements.
1258 if (!match(C2,
1259 m_SpecificInt_ICMP(ICmpInst::Predicate::ICMP_NE,
1260 APInt::getSignedMaxValue(
1261 C2->getType()->getScalarSizeInBits()))))
1262 return nullptr; // Can't do, have signed max element[s].
1263 C2 = AddOne(C2);
1264 LLVM_FALLTHROUGH;
1265 case ICmpInst::Predicate::ICMP_SGE:
1266 // Also non-canonical, but here we don't need to change C2,
1267 // so we don't have any restrictions on C2, so we can just handle it.
1268 std::swap(ReplacementLow, ReplacementHigh);
1269 break;
1270 default:
1271 return nullptr; // Unknown predicate.
1274 // The thresholds of this clamp-like pattern.
1275 auto *ThresholdLowIncl = ConstantExpr::getNeg(C1);
1276 auto *ThresholdHighExcl = ConstantExpr::getSub(C0, C1);
1278 // The fold has a precondition 1: C2 s>= ThresholdLow
1279 auto *Precond1 = ConstantExpr::getICmp(ICmpInst::Predicate::ICMP_SGE, C2,
1280 ThresholdLowIncl);
1281 if (!match(Precond1, m_One()))
1282 return nullptr;
1283 // The fold has a precondition 2: C2 s<= ThresholdHigh
1284 auto *Precond2 = ConstantExpr::getICmp(ICmpInst::Predicate::ICMP_SLE, C2,
1285 ThresholdHighExcl);
1286 if (!match(Precond2, m_One()))
1287 return nullptr;
1289 // All good, finally emit the new pattern.
1290 Value *ShouldReplaceLow = Builder.CreateICmpSLT(X, ThresholdLowIncl);
1291 Value *ShouldReplaceHigh = Builder.CreateICmpSGE(X, ThresholdHighExcl);
1292 Value *MaybeReplacedLow =
1293 Builder.CreateSelect(ShouldReplaceLow, ReplacementLow, X);
1294 Instruction *MaybeReplacedHigh =
1295 SelectInst::Create(ShouldReplaceHigh, ReplacementHigh, MaybeReplacedLow);
1297 return MaybeReplacedHigh;
1300 // If we have
1301 // %cmp = icmp [canonical predicate] i32 %x, C0
1302 // %r = select i1 %cmp, i32 %y, i32 C1
1303 // Where C0 != C1 and %x may be different from %y, see if the constant that we
1304 // will have if we flip the strictness of the predicate (i.e. without changing
1305 // the result) is identical to the C1 in select. If it matches we can change
1306 // original comparison to one with swapped predicate, reuse the constant,
1307 // and swap the hands of select.
1308 static Instruction *
1309 tryToReuseConstantFromSelectInComparison(SelectInst &Sel, ICmpInst &Cmp,
1310 InstCombiner::BuilderTy &Builder) {
1311 ICmpInst::Predicate Pred;
1312 Value *X;
1313 Constant *C0;
1314 if (!match(&Cmp, m_OneUse(m_ICmp(
1315 Pred, m_Value(X),
1316 m_CombineAnd(m_AnyIntegralConstant(), m_Constant(C0))))))
1317 return nullptr;
1319 // If comparison predicate is non-relational, we won't be able to do anything.
1320 if (ICmpInst::isEquality(Pred))
1321 return nullptr;
1323 // If comparison predicate is non-canonical, then we certainly won't be able
1324 // to make it canonical; canonicalizeCmpWithConstant() already tried.
1325 if (!isCanonicalPredicate(Pred))
1326 return nullptr;
1328 // If the [input] type of comparison and select type are different, lets abort
1329 // for now. We could try to compare constants with trunc/[zs]ext though.
1330 if (C0->getType() != Sel.getType())
1331 return nullptr;
1333 // FIXME: are there any magic icmp predicate+constant pairs we must not touch?
1335 Value *SelVal0, *SelVal1; // We do not care which one is from where.
1336 match(&Sel, m_Select(m_Value(), m_Value(SelVal0), m_Value(SelVal1)));
1337 // At least one of these values we are selecting between must be a constant
1338 // else we'll never succeed.
1339 if (!match(SelVal0, m_AnyIntegralConstant()) &&
1340 !match(SelVal1, m_AnyIntegralConstant()))
1341 return nullptr;
1343 // Does this constant C match any of the `select` values?
1344 auto MatchesSelectValue = [SelVal0, SelVal1](Constant *C) {
1345 return C->isElementWiseEqual(SelVal0) || C->isElementWiseEqual(SelVal1);
1348 // If C0 *already* matches true/false value of select, we are done.
1349 if (MatchesSelectValue(C0))
1350 return nullptr;
1352 // Check the constant we'd have with flipped-strictness predicate.
1353 auto FlippedStrictness = getFlippedStrictnessPredicateAndConstant(Pred, C0);
1354 if (!FlippedStrictness)
1355 return nullptr;
1357 // If said constant doesn't match either, then there is no hope,
1358 if (!MatchesSelectValue(FlippedStrictness->second))
1359 return nullptr;
1361 // It matched! Lets insert the new comparison just before select.
1362 InstCombiner::BuilderTy::InsertPointGuard Guard(Builder);
1363 Builder.SetInsertPoint(&Sel);
1365 Pred = ICmpInst::getSwappedPredicate(Pred); // Yes, swapped.
1366 Value *NewCmp = Builder.CreateICmp(Pred, X, FlippedStrictness->second,
1367 Cmp.getName() + ".inv");
1368 Sel.setCondition(NewCmp);
1369 Sel.swapValues();
1370 Sel.swapProfMetadata();
1372 return &Sel;
1375 /// Visit a SelectInst that has an ICmpInst as its first operand.
1376 Instruction *InstCombiner::foldSelectInstWithICmp(SelectInst &SI,
1377 ICmpInst *ICI) {
1378 if (Value *V = foldSelectValueEquivalence(SI, *ICI, SQ))
1379 return replaceInstUsesWith(SI, V);
1381 if (Instruction *NewSel = canonicalizeMinMaxWithConstant(SI, *ICI, Builder))
1382 return NewSel;
1384 if (Instruction *NewAbs = canonicalizeAbsNabs(SI, *ICI, Builder))
1385 return NewAbs;
1387 if (Instruction *NewAbs = canonicalizeClampLike(SI, *ICI, Builder))
1388 return NewAbs;
1390 if (Instruction *NewSel =
1391 tryToReuseConstantFromSelectInComparison(SI, *ICI, Builder))
1392 return NewSel;
1394 bool Changed = adjustMinMax(SI, *ICI);
1396 if (Value *V = foldSelectICmpAnd(SI, ICI, Builder))
1397 return replaceInstUsesWith(SI, V);
1399 // NOTE: if we wanted to, this is where to detect integer MIN/MAX
1400 Value *TrueVal = SI.getTrueValue();
1401 Value *FalseVal = SI.getFalseValue();
1402 ICmpInst::Predicate Pred = ICI->getPredicate();
1403 Value *CmpLHS = ICI->getOperand(0);
1404 Value *CmpRHS = ICI->getOperand(1);
1405 if (CmpRHS != CmpLHS && isa<Constant>(CmpRHS)) {
1406 if (CmpLHS == TrueVal && Pred == ICmpInst::ICMP_EQ) {
1407 // Transform (X == C) ? X : Y -> (X == C) ? C : Y
1408 SI.setOperand(1, CmpRHS);
1409 Changed = true;
1410 } else if (CmpLHS == FalseVal && Pred == ICmpInst::ICMP_NE) {
1411 // Transform (X != C) ? Y : X -> (X != C) ? Y : C
1412 SI.setOperand(2, CmpRHS);
1413 Changed = true;
1417 // FIXME: This code is nearly duplicated in InstSimplify. Using/refactoring
1418 // decomposeBitTestICmp() might help.
1420 unsigned BitWidth =
1421 DL.getTypeSizeInBits(TrueVal->getType()->getScalarType());
1422 APInt MinSignedValue = APInt::getSignedMinValue(BitWidth);
1423 Value *X;
1424 const APInt *Y, *C;
1425 bool TrueWhenUnset;
1426 bool IsBitTest = false;
1427 if (ICmpInst::isEquality(Pred) &&
1428 match(CmpLHS, m_And(m_Value(X), m_Power2(Y))) &&
1429 match(CmpRHS, m_Zero())) {
1430 IsBitTest = true;
1431 TrueWhenUnset = Pred == ICmpInst::ICMP_EQ;
1432 } else if (Pred == ICmpInst::ICMP_SLT && match(CmpRHS, m_Zero())) {
1433 X = CmpLHS;
1434 Y = &MinSignedValue;
1435 IsBitTest = true;
1436 TrueWhenUnset = false;
1437 } else if (Pred == ICmpInst::ICMP_SGT && match(CmpRHS, m_AllOnes())) {
1438 X = CmpLHS;
1439 Y = &MinSignedValue;
1440 IsBitTest = true;
1441 TrueWhenUnset = true;
1443 if (IsBitTest) {
1444 Value *V = nullptr;
1445 // (X & Y) == 0 ? X : X ^ Y --> X & ~Y
1446 if (TrueWhenUnset && TrueVal == X &&
1447 match(FalseVal, m_Xor(m_Specific(X), m_APInt(C))) && *Y == *C)
1448 V = Builder.CreateAnd(X, ~(*Y));
1449 // (X & Y) != 0 ? X ^ Y : X --> X & ~Y
1450 else if (!TrueWhenUnset && FalseVal == X &&
1451 match(TrueVal, m_Xor(m_Specific(X), m_APInt(C))) && *Y == *C)
1452 V = Builder.CreateAnd(X, ~(*Y));
1453 // (X & Y) == 0 ? X ^ Y : X --> X | Y
1454 else if (TrueWhenUnset && FalseVal == X &&
1455 match(TrueVal, m_Xor(m_Specific(X), m_APInt(C))) && *Y == *C)
1456 V = Builder.CreateOr(X, *Y);
1457 // (X & Y) != 0 ? X : X ^ Y --> X | Y
1458 else if (!TrueWhenUnset && TrueVal == X &&
1459 match(FalseVal, m_Xor(m_Specific(X), m_APInt(C))) && *Y == *C)
1460 V = Builder.CreateOr(X, *Y);
1462 if (V)
1463 return replaceInstUsesWith(SI, V);
1467 if (Instruction *V =
1468 foldSelectICmpAndAnd(SI.getType(), ICI, TrueVal, FalseVal, Builder))
1469 return V;
1471 if (Instruction *V = foldSelectCtlzToCttz(ICI, TrueVal, FalseVal, Builder))
1472 return V;
1474 if (Value *V = foldSelectICmpAndOr(ICI, TrueVal, FalseVal, Builder))
1475 return replaceInstUsesWith(SI, V);
1477 if (Value *V = foldSelectICmpLshrAshr(ICI, TrueVal, FalseVal, Builder))
1478 return replaceInstUsesWith(SI, V);
1480 if (Value *V = foldSelectCttzCtlz(ICI, TrueVal, FalseVal, Builder))
1481 return replaceInstUsesWith(SI, V);
1483 if (Value *V = canonicalizeSaturatedSubtract(ICI, TrueVal, FalseVal, Builder))
1484 return replaceInstUsesWith(SI, V);
1486 if (Value *V = canonicalizeSaturatedAdd(ICI, TrueVal, FalseVal, Builder))
1487 return replaceInstUsesWith(SI, V);
1489 return Changed ? &SI : nullptr;
1492 /// SI is a select whose condition is a PHI node (but the two may be in
1493 /// different blocks). See if the true/false values (V) are live in all of the
1494 /// predecessor blocks of the PHI. For example, cases like this can't be mapped:
1496 /// X = phi [ C1, BB1], [C2, BB2]
1497 /// Y = add
1498 /// Z = select X, Y, 0
1500 /// because Y is not live in BB1/BB2.
1501 static bool canSelectOperandBeMappingIntoPredBlock(const Value *V,
1502 const SelectInst &SI) {
1503 // If the value is a non-instruction value like a constant or argument, it
1504 // can always be mapped.
1505 const Instruction *I = dyn_cast<Instruction>(V);
1506 if (!I) return true;
1508 // If V is a PHI node defined in the same block as the condition PHI, we can
1509 // map the arguments.
1510 const PHINode *CondPHI = cast<PHINode>(SI.getCondition());
1512 if (const PHINode *VP = dyn_cast<PHINode>(I))
1513 if (VP->getParent() == CondPHI->getParent())
1514 return true;
1516 // Otherwise, if the PHI and select are defined in the same block and if V is
1517 // defined in a different block, then we can transform it.
1518 if (SI.getParent() == CondPHI->getParent() &&
1519 I->getParent() != CondPHI->getParent())
1520 return true;
1522 // Otherwise we have a 'hard' case and we can't tell without doing more
1523 // detailed dominator based analysis, punt.
1524 return false;
1527 /// We have an SPF (e.g. a min or max) of an SPF of the form:
1528 /// SPF2(SPF1(A, B), C)
1529 Instruction *InstCombiner::foldSPFofSPF(Instruction *Inner,
1530 SelectPatternFlavor SPF1,
1531 Value *A, Value *B,
1532 Instruction &Outer,
1533 SelectPatternFlavor SPF2, Value *C) {
1534 if (Outer.getType() != Inner->getType())
1535 return nullptr;
1537 if (C == A || C == B) {
1538 // MAX(MAX(A, B), B) -> MAX(A, B)
1539 // MIN(MIN(a, b), a) -> MIN(a, b)
1540 // TODO: This could be done in instsimplify.
1541 if (SPF1 == SPF2 && SelectPatternResult::isMinOrMax(SPF1))
1542 return replaceInstUsesWith(Outer, Inner);
1544 // MAX(MIN(a, b), a) -> a
1545 // MIN(MAX(a, b), a) -> a
1546 // TODO: This could be done in instsimplify.
1547 if ((SPF1 == SPF_SMIN && SPF2 == SPF_SMAX) ||
1548 (SPF1 == SPF_SMAX && SPF2 == SPF_SMIN) ||
1549 (SPF1 == SPF_UMIN && SPF2 == SPF_UMAX) ||
1550 (SPF1 == SPF_UMAX && SPF2 == SPF_UMIN))
1551 return replaceInstUsesWith(Outer, C);
1554 if (SPF1 == SPF2) {
1555 const APInt *CB, *CC;
1556 if (match(B, m_APInt(CB)) && match(C, m_APInt(CC))) {
1557 // MIN(MIN(A, 23), 97) -> MIN(A, 23)
1558 // MAX(MAX(A, 97), 23) -> MAX(A, 97)
1559 // TODO: This could be done in instsimplify.
1560 if ((SPF1 == SPF_UMIN && CB->ule(*CC)) ||
1561 (SPF1 == SPF_SMIN && CB->sle(*CC)) ||
1562 (SPF1 == SPF_UMAX && CB->uge(*CC)) ||
1563 (SPF1 == SPF_SMAX && CB->sge(*CC)))
1564 return replaceInstUsesWith(Outer, Inner);
1566 // MIN(MIN(A, 97), 23) -> MIN(A, 23)
1567 // MAX(MAX(A, 23), 97) -> MAX(A, 97)
1568 if ((SPF1 == SPF_UMIN && CB->ugt(*CC)) ||
1569 (SPF1 == SPF_SMIN && CB->sgt(*CC)) ||
1570 (SPF1 == SPF_UMAX && CB->ult(*CC)) ||
1571 (SPF1 == SPF_SMAX && CB->slt(*CC))) {
1572 Outer.replaceUsesOfWith(Inner, A);
1573 return &Outer;
1578 // max(max(A, B), min(A, B)) --> max(A, B)
1579 // min(min(A, B), max(A, B)) --> min(A, B)
1580 // TODO: This could be done in instsimplify.
1581 if (SPF1 == SPF2 &&
1582 ((SPF1 == SPF_UMIN && match(C, m_c_UMax(m_Specific(A), m_Specific(B)))) ||
1583 (SPF1 == SPF_SMIN && match(C, m_c_SMax(m_Specific(A), m_Specific(B)))) ||
1584 (SPF1 == SPF_UMAX && match(C, m_c_UMin(m_Specific(A), m_Specific(B)))) ||
1585 (SPF1 == SPF_SMAX && match(C, m_c_SMin(m_Specific(A), m_Specific(B))))))
1586 return replaceInstUsesWith(Outer, Inner);
1588 // ABS(ABS(X)) -> ABS(X)
1589 // NABS(NABS(X)) -> NABS(X)
1590 // TODO: This could be done in instsimplify.
1591 if (SPF1 == SPF2 && (SPF1 == SPF_ABS || SPF1 == SPF_NABS)) {
1592 return replaceInstUsesWith(Outer, Inner);
1595 // ABS(NABS(X)) -> ABS(X)
1596 // NABS(ABS(X)) -> NABS(X)
1597 if ((SPF1 == SPF_ABS && SPF2 == SPF_NABS) ||
1598 (SPF1 == SPF_NABS && SPF2 == SPF_ABS)) {
1599 SelectInst *SI = cast<SelectInst>(Inner);
1600 Value *NewSI =
1601 Builder.CreateSelect(SI->getCondition(), SI->getFalseValue(),
1602 SI->getTrueValue(), SI->getName(), SI);
1603 return replaceInstUsesWith(Outer, NewSI);
1606 auto IsFreeOrProfitableToInvert =
1607 [&](Value *V, Value *&NotV, bool &ElidesXor) {
1608 if (match(V, m_Not(m_Value(NotV)))) {
1609 // If V has at most 2 uses then we can get rid of the xor operation
1610 // entirely.
1611 ElidesXor |= !V->hasNUsesOrMore(3);
1612 return true;
1615 if (isFreeToInvert(V, !V->hasNUsesOrMore(3))) {
1616 NotV = nullptr;
1617 return true;
1620 return false;
1623 Value *NotA, *NotB, *NotC;
1624 bool ElidesXor = false;
1626 // MIN(MIN(~A, ~B), ~C) == ~MAX(MAX(A, B), C)
1627 // MIN(MAX(~A, ~B), ~C) == ~MAX(MIN(A, B), C)
1628 // MAX(MIN(~A, ~B), ~C) == ~MIN(MAX(A, B), C)
1629 // MAX(MAX(~A, ~B), ~C) == ~MIN(MIN(A, B), C)
1631 // This transform is performance neutral if we can elide at least one xor from
1632 // the set of three operands, since we'll be tacking on an xor at the very
1633 // end.
1634 if (SelectPatternResult::isMinOrMax(SPF1) &&
1635 SelectPatternResult::isMinOrMax(SPF2) &&
1636 IsFreeOrProfitableToInvert(A, NotA, ElidesXor) &&
1637 IsFreeOrProfitableToInvert(B, NotB, ElidesXor) &&
1638 IsFreeOrProfitableToInvert(C, NotC, ElidesXor) && ElidesXor) {
1639 if (!NotA)
1640 NotA = Builder.CreateNot(A);
1641 if (!NotB)
1642 NotB = Builder.CreateNot(B);
1643 if (!NotC)
1644 NotC = Builder.CreateNot(C);
1646 Value *NewInner = createMinMax(Builder, getInverseMinMaxFlavor(SPF1), NotA,
1647 NotB);
1648 Value *NewOuter = Builder.CreateNot(
1649 createMinMax(Builder, getInverseMinMaxFlavor(SPF2), NewInner, NotC));
1650 return replaceInstUsesWith(Outer, NewOuter);
1653 return nullptr;
1656 /// Turn select C, (X + Y), (X - Y) --> (X + (select C, Y, (-Y))).
1657 /// This is even legal for FP.
1658 static Instruction *foldAddSubSelect(SelectInst &SI,
1659 InstCombiner::BuilderTy &Builder) {
1660 Value *CondVal = SI.getCondition();
1661 Value *TrueVal = SI.getTrueValue();
1662 Value *FalseVal = SI.getFalseValue();
1663 auto *TI = dyn_cast<Instruction>(TrueVal);
1664 auto *FI = dyn_cast<Instruction>(FalseVal);
1665 if (!TI || !FI || !TI->hasOneUse() || !FI->hasOneUse())
1666 return nullptr;
1668 Instruction *AddOp = nullptr, *SubOp = nullptr;
1669 if ((TI->getOpcode() == Instruction::Sub &&
1670 FI->getOpcode() == Instruction::Add) ||
1671 (TI->getOpcode() == Instruction::FSub &&
1672 FI->getOpcode() == Instruction::FAdd)) {
1673 AddOp = FI;
1674 SubOp = TI;
1675 } else if ((FI->getOpcode() == Instruction::Sub &&
1676 TI->getOpcode() == Instruction::Add) ||
1677 (FI->getOpcode() == Instruction::FSub &&
1678 TI->getOpcode() == Instruction::FAdd)) {
1679 AddOp = TI;
1680 SubOp = FI;
1683 if (AddOp) {
1684 Value *OtherAddOp = nullptr;
1685 if (SubOp->getOperand(0) == AddOp->getOperand(0)) {
1686 OtherAddOp = AddOp->getOperand(1);
1687 } else if (SubOp->getOperand(0) == AddOp->getOperand(1)) {
1688 OtherAddOp = AddOp->getOperand(0);
1691 if (OtherAddOp) {
1692 // So at this point we know we have (Y -> OtherAddOp):
1693 // select C, (add X, Y), (sub X, Z)
1694 Value *NegVal; // Compute -Z
1695 if (SI.getType()->isFPOrFPVectorTy()) {
1696 NegVal = Builder.CreateFNeg(SubOp->getOperand(1));
1697 if (Instruction *NegInst = dyn_cast<Instruction>(NegVal)) {
1698 FastMathFlags Flags = AddOp->getFastMathFlags();
1699 Flags &= SubOp->getFastMathFlags();
1700 NegInst->setFastMathFlags(Flags);
1702 } else {
1703 NegVal = Builder.CreateNeg(SubOp->getOperand(1));
1706 Value *NewTrueOp = OtherAddOp;
1707 Value *NewFalseOp = NegVal;
1708 if (AddOp != TI)
1709 std::swap(NewTrueOp, NewFalseOp);
1710 Value *NewSel = Builder.CreateSelect(CondVal, NewTrueOp, NewFalseOp,
1711 SI.getName() + ".p", &SI);
1713 if (SI.getType()->isFPOrFPVectorTy()) {
1714 Instruction *RI =
1715 BinaryOperator::CreateFAdd(SubOp->getOperand(0), NewSel);
1717 FastMathFlags Flags = AddOp->getFastMathFlags();
1718 Flags &= SubOp->getFastMathFlags();
1719 RI->setFastMathFlags(Flags);
1720 return RI;
1721 } else
1722 return BinaryOperator::CreateAdd(SubOp->getOperand(0), NewSel);
1725 return nullptr;
1728 Instruction *InstCombiner::foldSelectExtConst(SelectInst &Sel) {
1729 Constant *C;
1730 if (!match(Sel.getTrueValue(), m_Constant(C)) &&
1731 !match(Sel.getFalseValue(), m_Constant(C)))
1732 return nullptr;
1734 Instruction *ExtInst;
1735 if (!match(Sel.getTrueValue(), m_Instruction(ExtInst)) &&
1736 !match(Sel.getFalseValue(), m_Instruction(ExtInst)))
1737 return nullptr;
1739 auto ExtOpcode = ExtInst->getOpcode();
1740 if (ExtOpcode != Instruction::ZExt && ExtOpcode != Instruction::SExt)
1741 return nullptr;
1743 // If we are extending from a boolean type or if we can create a select that
1744 // has the same size operands as its condition, try to narrow the select.
1745 Value *X = ExtInst->getOperand(0);
1746 Type *SmallType = X->getType();
1747 Value *Cond = Sel.getCondition();
1748 auto *Cmp = dyn_cast<CmpInst>(Cond);
1749 if (!SmallType->isIntOrIntVectorTy(1) &&
1750 (!Cmp || Cmp->getOperand(0)->getType() != SmallType))
1751 return nullptr;
1753 // If the constant is the same after truncation to the smaller type and
1754 // extension to the original type, we can narrow the select.
1755 Type *SelType = Sel.getType();
1756 Constant *TruncC = ConstantExpr::getTrunc(C, SmallType);
1757 Constant *ExtC = ConstantExpr::getCast(ExtOpcode, TruncC, SelType);
1758 if (ExtC == C) {
1759 Value *TruncCVal = cast<Value>(TruncC);
1760 if (ExtInst == Sel.getFalseValue())
1761 std::swap(X, TruncCVal);
1763 // select Cond, (ext X), C --> ext(select Cond, X, C')
1764 // select Cond, C, (ext X) --> ext(select Cond, C', X)
1765 Value *NewSel = Builder.CreateSelect(Cond, X, TruncCVal, "narrow", &Sel);
1766 return CastInst::Create(Instruction::CastOps(ExtOpcode), NewSel, SelType);
1769 // If one arm of the select is the extend of the condition, replace that arm
1770 // with the extension of the appropriate known bool value.
1771 if (Cond == X) {
1772 if (ExtInst == Sel.getTrueValue()) {
1773 // select X, (sext X), C --> select X, -1, C
1774 // select X, (zext X), C --> select X, 1, C
1775 Constant *One = ConstantInt::getTrue(SmallType);
1776 Constant *AllOnesOrOne = ConstantExpr::getCast(ExtOpcode, One, SelType);
1777 return SelectInst::Create(Cond, AllOnesOrOne, C, "", nullptr, &Sel);
1778 } else {
1779 // select X, C, (sext X) --> select X, C, 0
1780 // select X, C, (zext X) --> select X, C, 0
1781 Constant *Zero = ConstantInt::getNullValue(SelType);
1782 return SelectInst::Create(Cond, C, Zero, "", nullptr, &Sel);
1786 return nullptr;
1789 /// Try to transform a vector select with a constant condition vector into a
1790 /// shuffle for easier combining with other shuffles and insert/extract.
1791 static Instruction *canonicalizeSelectToShuffle(SelectInst &SI) {
1792 Value *CondVal = SI.getCondition();
1793 Constant *CondC;
1794 if (!CondVal->getType()->isVectorTy() || !match(CondVal, m_Constant(CondC)))
1795 return nullptr;
1797 unsigned NumElts = CondVal->getType()->getVectorNumElements();
1798 SmallVector<Constant *, 16> Mask;
1799 Mask.reserve(NumElts);
1800 Type *Int32Ty = Type::getInt32Ty(CondVal->getContext());
1801 for (unsigned i = 0; i != NumElts; ++i) {
1802 Constant *Elt = CondC->getAggregateElement(i);
1803 if (!Elt)
1804 return nullptr;
1806 if (Elt->isOneValue()) {
1807 // If the select condition element is true, choose from the 1st vector.
1808 Mask.push_back(ConstantInt::get(Int32Ty, i));
1809 } else if (Elt->isNullValue()) {
1810 // If the select condition element is false, choose from the 2nd vector.
1811 Mask.push_back(ConstantInt::get(Int32Ty, i + NumElts));
1812 } else if (isa<UndefValue>(Elt)) {
1813 // Undef in a select condition (choose one of the operands) does not mean
1814 // the same thing as undef in a shuffle mask (any value is acceptable), so
1815 // give up.
1816 return nullptr;
1817 } else {
1818 // Bail out on a constant expression.
1819 return nullptr;
1823 return new ShuffleVectorInst(SI.getTrueValue(), SI.getFalseValue(),
1824 ConstantVector::get(Mask));
1827 /// If we have a select of vectors with a scalar condition, try to convert that
1828 /// to a vector select by splatting the condition. A splat may get folded with
1829 /// other operations in IR and having all operands of a select be vector types
1830 /// is likely better for vector codegen.
1831 static Instruction *canonicalizeScalarSelectOfVecs(
1832 SelectInst &Sel, InstCombiner::BuilderTy &Builder) {
1833 Type *Ty = Sel.getType();
1834 if (!Ty->isVectorTy())
1835 return nullptr;
1837 // We can replace a single-use extract with constant index.
1838 Value *Cond = Sel.getCondition();
1839 if (!match(Cond, m_OneUse(m_ExtractElement(m_Value(), m_ConstantInt()))))
1840 return nullptr;
1842 // select (extelt V, Index), T, F --> select (splat V, Index), T, F
1843 // Splatting the extracted condition reduces code (we could directly create a
1844 // splat shuffle of the source vector to eliminate the intermediate step).
1845 unsigned NumElts = Ty->getVectorNumElements();
1846 Value *SplatCond = Builder.CreateVectorSplat(NumElts, Cond);
1847 Sel.setCondition(SplatCond);
1848 return &Sel;
1851 /// Reuse bitcasted operands between a compare and select:
1852 /// select (cmp (bitcast C), (bitcast D)), (bitcast' C), (bitcast' D) -->
1853 /// bitcast (select (cmp (bitcast C), (bitcast D)), (bitcast C), (bitcast D))
1854 static Instruction *foldSelectCmpBitcasts(SelectInst &Sel,
1855 InstCombiner::BuilderTy &Builder) {
1856 Value *Cond = Sel.getCondition();
1857 Value *TVal = Sel.getTrueValue();
1858 Value *FVal = Sel.getFalseValue();
1860 CmpInst::Predicate Pred;
1861 Value *A, *B;
1862 if (!match(Cond, m_Cmp(Pred, m_Value(A), m_Value(B))))
1863 return nullptr;
1865 // The select condition is a compare instruction. If the select's true/false
1866 // values are already the same as the compare operands, there's nothing to do.
1867 if (TVal == A || TVal == B || FVal == A || FVal == B)
1868 return nullptr;
1870 Value *C, *D;
1871 if (!match(A, m_BitCast(m_Value(C))) || !match(B, m_BitCast(m_Value(D))))
1872 return nullptr;
1874 // select (cmp (bitcast C), (bitcast D)), (bitcast TSrc), (bitcast FSrc)
1875 Value *TSrc, *FSrc;
1876 if (!match(TVal, m_BitCast(m_Value(TSrc))) ||
1877 !match(FVal, m_BitCast(m_Value(FSrc))))
1878 return nullptr;
1880 // If the select true/false values are *different bitcasts* of the same source
1881 // operands, make the select operands the same as the compare operands and
1882 // cast the result. This is the canonical select form for min/max.
1883 Value *NewSel;
1884 if (TSrc == C && FSrc == D) {
1885 // select (cmp (bitcast C), (bitcast D)), (bitcast' C), (bitcast' D) -->
1886 // bitcast (select (cmp A, B), A, B)
1887 NewSel = Builder.CreateSelect(Cond, A, B, "", &Sel);
1888 } else if (TSrc == D && FSrc == C) {
1889 // select (cmp (bitcast C), (bitcast D)), (bitcast' D), (bitcast' C) -->
1890 // bitcast (select (cmp A, B), B, A)
1891 NewSel = Builder.CreateSelect(Cond, B, A, "", &Sel);
1892 } else {
1893 return nullptr;
1895 return CastInst::CreateBitOrPointerCast(NewSel, Sel.getType());
1898 /// Try to eliminate select instructions that test the returned flag of cmpxchg
1899 /// instructions.
1901 /// If a select instruction tests the returned flag of a cmpxchg instruction and
1902 /// selects between the returned value of the cmpxchg instruction its compare
1903 /// operand, the result of the select will always be equal to its false value.
1904 /// For example:
1906 /// %0 = cmpxchg i64* %ptr, i64 %compare, i64 %new_value seq_cst seq_cst
1907 /// %1 = extractvalue { i64, i1 } %0, 1
1908 /// %2 = extractvalue { i64, i1 } %0, 0
1909 /// %3 = select i1 %1, i64 %compare, i64 %2
1910 /// ret i64 %3
1912 /// The returned value of the cmpxchg instruction (%2) is the original value
1913 /// located at %ptr prior to any update. If the cmpxchg operation succeeds, %2
1914 /// must have been equal to %compare. Thus, the result of the select is always
1915 /// equal to %2, and the code can be simplified to:
1917 /// %0 = cmpxchg i64* %ptr, i64 %compare, i64 %new_value seq_cst seq_cst
1918 /// %1 = extractvalue { i64, i1 } %0, 0
1919 /// ret i64 %1
1921 static Instruction *foldSelectCmpXchg(SelectInst &SI) {
1922 // A helper that determines if V is an extractvalue instruction whose
1923 // aggregate operand is a cmpxchg instruction and whose single index is equal
1924 // to I. If such conditions are true, the helper returns the cmpxchg
1925 // instruction; otherwise, a nullptr is returned.
1926 auto isExtractFromCmpXchg = [](Value *V, unsigned I) -> AtomicCmpXchgInst * {
1927 auto *Extract = dyn_cast<ExtractValueInst>(V);
1928 if (!Extract)
1929 return nullptr;
1930 if (Extract->getIndices()[0] != I)
1931 return nullptr;
1932 return dyn_cast<AtomicCmpXchgInst>(Extract->getAggregateOperand());
1935 // If the select has a single user, and this user is a select instruction that
1936 // we can simplify, skip the cmpxchg simplification for now.
1937 if (SI.hasOneUse())
1938 if (auto *Select = dyn_cast<SelectInst>(SI.user_back()))
1939 if (Select->getCondition() == SI.getCondition())
1940 if (Select->getFalseValue() == SI.getTrueValue() ||
1941 Select->getTrueValue() == SI.getFalseValue())
1942 return nullptr;
1944 // Ensure the select condition is the returned flag of a cmpxchg instruction.
1945 auto *CmpXchg = isExtractFromCmpXchg(SI.getCondition(), 1);
1946 if (!CmpXchg)
1947 return nullptr;
1949 // Check the true value case: The true value of the select is the returned
1950 // value of the same cmpxchg used by the condition, and the false value is the
1951 // cmpxchg instruction's compare operand.
1952 if (auto *X = isExtractFromCmpXchg(SI.getTrueValue(), 0))
1953 if (X == CmpXchg && X->getCompareOperand() == SI.getFalseValue()) {
1954 SI.setTrueValue(SI.getFalseValue());
1955 return &SI;
1958 // Check the false value case: The false value of the select is the returned
1959 // value of the same cmpxchg used by the condition, and the true value is the
1960 // cmpxchg instruction's compare operand.
1961 if (auto *X = isExtractFromCmpXchg(SI.getFalseValue(), 0))
1962 if (X == CmpXchg && X->getCompareOperand() == SI.getTrueValue()) {
1963 SI.setTrueValue(SI.getFalseValue());
1964 return &SI;
1967 return nullptr;
1970 static Instruction *moveAddAfterMinMax(SelectPatternFlavor SPF, Value *X,
1971 Value *Y,
1972 InstCombiner::BuilderTy &Builder) {
1973 assert(SelectPatternResult::isMinOrMax(SPF) && "Expected min/max pattern");
1974 bool IsUnsigned = SPF == SelectPatternFlavor::SPF_UMIN ||
1975 SPF == SelectPatternFlavor::SPF_UMAX;
1976 // TODO: If InstSimplify could fold all cases where C2 <= C1, we could change
1977 // the constant value check to an assert.
1978 Value *A;
1979 const APInt *C1, *C2;
1980 if (IsUnsigned && match(X, m_NUWAdd(m_Value(A), m_APInt(C1))) &&
1981 match(Y, m_APInt(C2)) && C2->uge(*C1) && X->hasNUses(2)) {
1982 // umin (add nuw A, C1), C2 --> add nuw (umin A, C2 - C1), C1
1983 // umax (add nuw A, C1), C2 --> add nuw (umax A, C2 - C1), C1
1984 Value *NewMinMax = createMinMax(Builder, SPF, A,
1985 ConstantInt::get(X->getType(), *C2 - *C1));
1986 return BinaryOperator::CreateNUW(BinaryOperator::Add, NewMinMax,
1987 ConstantInt::get(X->getType(), *C1));
1990 if (!IsUnsigned && match(X, m_NSWAdd(m_Value(A), m_APInt(C1))) &&
1991 match(Y, m_APInt(C2)) && X->hasNUses(2)) {
1992 bool Overflow;
1993 APInt Diff = C2->ssub_ov(*C1, Overflow);
1994 if (!Overflow) {
1995 // smin (add nsw A, C1), C2 --> add nsw (smin A, C2 - C1), C1
1996 // smax (add nsw A, C1), C2 --> add nsw (smax A, C2 - C1), C1
1997 Value *NewMinMax = createMinMax(Builder, SPF, A,
1998 ConstantInt::get(X->getType(), Diff));
1999 return BinaryOperator::CreateNSW(BinaryOperator::Add, NewMinMax,
2000 ConstantInt::get(X->getType(), *C1));
2004 return nullptr;
2007 /// Match a sadd_sat or ssub_sat which is using min/max to clamp the value.
2008 Instruction *InstCombiner::matchSAddSubSat(SelectInst &MinMax1) {
2009 Type *Ty = MinMax1.getType();
2011 // We are looking for a tree of:
2012 // max(INT_MIN, min(INT_MAX, add(sext(A), sext(B))))
2013 // Where the min and max could be reversed
2014 Instruction *MinMax2;
2015 BinaryOperator *AddSub;
2016 const APInt *MinValue, *MaxValue;
2017 if (match(&MinMax1, m_SMin(m_Instruction(MinMax2), m_APInt(MaxValue)))) {
2018 if (!match(MinMax2, m_SMax(m_BinOp(AddSub), m_APInt(MinValue))))
2019 return nullptr;
2020 } else if (match(&MinMax1,
2021 m_SMax(m_Instruction(MinMax2), m_APInt(MinValue)))) {
2022 if (!match(MinMax2, m_SMin(m_BinOp(AddSub), m_APInt(MaxValue))))
2023 return nullptr;
2024 } else
2025 return nullptr;
2027 // Check that the constants clamp a saturate, and that the new type would be
2028 // sensible to convert to.
2029 if (!(*MaxValue + 1).isPowerOf2() || -*MinValue != *MaxValue + 1)
2030 return nullptr;
2031 // In what bitwidth can this be treated as saturating arithmetics?
2032 unsigned NewBitWidth = (*MaxValue + 1).logBase2() + 1;
2033 // FIXME: This isn't quite right for vectors, but using the scalar type is a
2034 // good first approximation for what should be done there.
2035 if (!shouldChangeType(Ty->getScalarType()->getIntegerBitWidth(), NewBitWidth))
2036 return nullptr;
2038 // Also make sure that the number of uses is as expected. The "3"s are for the
2039 // the two items of min/max (the compare and the select).
2040 if (MinMax2->hasNUsesOrMore(3) || AddSub->hasNUsesOrMore(3))
2041 return nullptr;
2043 // Create the new type (which can be a vector type)
2044 Type *NewTy = Ty->getWithNewBitWidth(NewBitWidth);
2045 // Match the two extends from the add/sub
2046 Value *A, *B;
2047 if(!match(AddSub, m_BinOp(m_SExt(m_Value(A)), m_SExt(m_Value(B)))))
2048 return nullptr;
2049 // And check the incoming values are of a type smaller than or equal to the
2050 // size of the saturation. Otherwise the higher bits can cause different
2051 // results.
2052 if (A->getType()->getScalarSizeInBits() > NewBitWidth ||
2053 B->getType()->getScalarSizeInBits() > NewBitWidth)
2054 return nullptr;
2056 Intrinsic::ID IntrinsicID;
2057 if (AddSub->getOpcode() == Instruction::Add)
2058 IntrinsicID = Intrinsic::sadd_sat;
2059 else if (AddSub->getOpcode() == Instruction::Sub)
2060 IntrinsicID = Intrinsic::ssub_sat;
2061 else
2062 return nullptr;
2064 // Finally create and return the sat intrinsic, truncated to the new type
2065 Function *F = Intrinsic::getDeclaration(MinMax1.getModule(), IntrinsicID, NewTy);
2066 Value *AT = Builder.CreateSExt(A, NewTy);
2067 Value *BT = Builder.CreateSExt(B, NewTy);
2068 Value *Sat = Builder.CreateCall(F, {AT, BT});
2069 return CastInst::Create(Instruction::SExt, Sat, Ty);
2072 /// Reduce a sequence of min/max with a common operand.
2073 static Instruction *factorizeMinMaxTree(SelectPatternFlavor SPF, Value *LHS,
2074 Value *RHS,
2075 InstCombiner::BuilderTy &Builder) {
2076 assert(SelectPatternResult::isMinOrMax(SPF) && "Expected a min/max");
2077 // TODO: Allow FP min/max with nnan/nsz.
2078 if (!LHS->getType()->isIntOrIntVectorTy())
2079 return nullptr;
2081 // Match 3 of the same min/max ops. Example: umin(umin(), umin()).
2082 Value *A, *B, *C, *D;
2083 SelectPatternResult L = matchSelectPattern(LHS, A, B);
2084 SelectPatternResult R = matchSelectPattern(RHS, C, D);
2085 if (SPF != L.Flavor || L.Flavor != R.Flavor)
2086 return nullptr;
2088 // Look for a common operand. The use checks are different than usual because
2089 // a min/max pattern typically has 2 uses of each op: 1 by the cmp and 1 by
2090 // the select.
2091 Value *MinMaxOp = nullptr;
2092 Value *ThirdOp = nullptr;
2093 if (!LHS->hasNUsesOrMore(3) && RHS->hasNUsesOrMore(3)) {
2094 // If the LHS is only used in this chain and the RHS is used outside of it,
2095 // reuse the RHS min/max because that will eliminate the LHS.
2096 if (D == A || C == A) {
2097 // min(min(a, b), min(c, a)) --> min(min(c, a), b)
2098 // min(min(a, b), min(a, d)) --> min(min(a, d), b)
2099 MinMaxOp = RHS;
2100 ThirdOp = B;
2101 } else if (D == B || C == B) {
2102 // min(min(a, b), min(c, b)) --> min(min(c, b), a)
2103 // min(min(a, b), min(b, d)) --> min(min(b, d), a)
2104 MinMaxOp = RHS;
2105 ThirdOp = A;
2107 } else if (!RHS->hasNUsesOrMore(3)) {
2108 // Reuse the LHS. This will eliminate the RHS.
2109 if (D == A || D == B) {
2110 // min(min(a, b), min(c, a)) --> min(min(a, b), c)
2111 // min(min(a, b), min(c, b)) --> min(min(a, b), c)
2112 MinMaxOp = LHS;
2113 ThirdOp = C;
2114 } else if (C == A || C == B) {
2115 // min(min(a, b), min(b, d)) --> min(min(a, b), d)
2116 // min(min(a, b), min(c, b)) --> min(min(a, b), d)
2117 MinMaxOp = LHS;
2118 ThirdOp = D;
2121 if (!MinMaxOp || !ThirdOp)
2122 return nullptr;
2124 CmpInst::Predicate P = getMinMaxPred(SPF);
2125 Value *CmpABC = Builder.CreateICmp(P, MinMaxOp, ThirdOp);
2126 return SelectInst::Create(CmpABC, MinMaxOp, ThirdOp);
2129 /// Try to reduce a rotate pattern that includes a compare and select into a
2130 /// funnel shift intrinsic. Example:
2131 /// rotl32(a, b) --> (b == 0 ? a : ((a >> (32 - b)) | (a << b)))
2132 /// --> call llvm.fshl.i32(a, a, b)
2133 static Instruction *foldSelectRotate(SelectInst &Sel) {
2134 // The false value of the select must be a rotate of the true value.
2135 Value *Or0, *Or1;
2136 if (!match(Sel.getFalseValue(), m_OneUse(m_Or(m_Value(Or0), m_Value(Or1)))))
2137 return nullptr;
2139 Value *TVal = Sel.getTrueValue();
2140 Value *SA0, *SA1;
2141 if (!match(Or0, m_OneUse(m_LogicalShift(m_Specific(TVal), m_Value(SA0)))) ||
2142 !match(Or1, m_OneUse(m_LogicalShift(m_Specific(TVal), m_Value(SA1)))))
2143 return nullptr;
2145 auto ShiftOpcode0 = cast<BinaryOperator>(Or0)->getOpcode();
2146 auto ShiftOpcode1 = cast<BinaryOperator>(Or1)->getOpcode();
2147 if (ShiftOpcode0 == ShiftOpcode1)
2148 return nullptr;
2150 // We have one of these patterns so far:
2151 // select ?, TVal, (or (lshr TVal, SA0), (shl TVal, SA1))
2152 // select ?, TVal, (or (shl TVal, SA0), (lshr TVal, SA1))
2153 // This must be a power-of-2 rotate for a bitmasking transform to be valid.
2154 unsigned Width = Sel.getType()->getScalarSizeInBits();
2155 if (!isPowerOf2_32(Width))
2156 return nullptr;
2158 // Check the shift amounts to see if they are an opposite pair.
2159 Value *ShAmt;
2160 if (match(SA1, m_OneUse(m_Sub(m_SpecificInt(Width), m_Specific(SA0)))))
2161 ShAmt = SA0;
2162 else if (match(SA0, m_OneUse(m_Sub(m_SpecificInt(Width), m_Specific(SA1)))))
2163 ShAmt = SA1;
2164 else
2165 return nullptr;
2167 // Finally, see if the select is filtering out a shift-by-zero.
2168 Value *Cond = Sel.getCondition();
2169 ICmpInst::Predicate Pred;
2170 if (!match(Cond, m_OneUse(m_ICmp(Pred, m_Specific(ShAmt), m_ZeroInt()))) ||
2171 Pred != ICmpInst::ICMP_EQ)
2172 return nullptr;
2174 // This is a rotate that avoids shift-by-bitwidth UB in a suboptimal way.
2175 // Convert to funnel shift intrinsic.
2176 bool IsFshl = (ShAmt == SA0 && ShiftOpcode0 == BinaryOperator::Shl) ||
2177 (ShAmt == SA1 && ShiftOpcode1 == BinaryOperator::Shl);
2178 Intrinsic::ID IID = IsFshl ? Intrinsic::fshl : Intrinsic::fshr;
2179 Function *F = Intrinsic::getDeclaration(Sel.getModule(), IID, Sel.getType());
2180 return IntrinsicInst::Create(F, { TVal, TVal, ShAmt });
2183 Instruction *InstCombiner::visitSelectInst(SelectInst &SI) {
2184 Value *CondVal = SI.getCondition();
2185 Value *TrueVal = SI.getTrueValue();
2186 Value *FalseVal = SI.getFalseValue();
2187 Type *SelType = SI.getType();
2189 // FIXME: Remove this workaround when freeze related patches are done.
2190 // For select with undef operand which feeds into an equality comparison,
2191 // don't simplify it so loop unswitch can know the equality comparison
2192 // may have an undef operand. This is a workaround for PR31652 caused by
2193 // descrepancy about branch on undef between LoopUnswitch and GVN.
2194 if (isa<UndefValue>(TrueVal) || isa<UndefValue>(FalseVal)) {
2195 if (llvm::any_of(SI.users(), [&](User *U) {
2196 ICmpInst *CI = dyn_cast<ICmpInst>(U);
2197 if (CI && CI->isEquality())
2198 return true;
2199 return false;
2200 })) {
2201 return nullptr;
2205 if (Value *V = SimplifySelectInst(CondVal, TrueVal, FalseVal,
2206 SQ.getWithInstruction(&SI)))
2207 return replaceInstUsesWith(SI, V);
2209 if (Instruction *I = canonicalizeSelectToShuffle(SI))
2210 return I;
2212 if (Instruction *I = canonicalizeScalarSelectOfVecs(SI, Builder))
2213 return I;
2215 // Canonicalize a one-use integer compare with a non-canonical predicate by
2216 // inverting the predicate and swapping the select operands. This matches a
2217 // compare canonicalization for conditional branches.
2218 // TODO: Should we do the same for FP compares?
2219 CmpInst::Predicate Pred;
2220 if (match(CondVal, m_OneUse(m_ICmp(Pred, m_Value(), m_Value()))) &&
2221 !isCanonicalPredicate(Pred)) {
2222 // Swap true/false values and condition.
2223 CmpInst *Cond = cast<CmpInst>(CondVal);
2224 Cond->setPredicate(CmpInst::getInversePredicate(Pred));
2225 SI.setOperand(1, FalseVal);
2226 SI.setOperand(2, TrueVal);
2227 SI.swapProfMetadata();
2228 Worklist.Add(Cond);
2229 return &SI;
2232 if (SelType->isIntOrIntVectorTy(1) &&
2233 TrueVal->getType() == CondVal->getType()) {
2234 if (match(TrueVal, m_One())) {
2235 // Change: A = select B, true, C --> A = or B, C
2236 return BinaryOperator::CreateOr(CondVal, FalseVal);
2238 if (match(TrueVal, m_Zero())) {
2239 // Change: A = select B, false, C --> A = and !B, C
2240 Value *NotCond = Builder.CreateNot(CondVal, "not." + CondVal->getName());
2241 return BinaryOperator::CreateAnd(NotCond, FalseVal);
2243 if (match(FalseVal, m_Zero())) {
2244 // Change: A = select B, C, false --> A = and B, C
2245 return BinaryOperator::CreateAnd(CondVal, TrueVal);
2247 if (match(FalseVal, m_One())) {
2248 // Change: A = select B, C, true --> A = or !B, C
2249 Value *NotCond = Builder.CreateNot(CondVal, "not." + CondVal->getName());
2250 return BinaryOperator::CreateOr(NotCond, TrueVal);
2253 // select a, a, b -> a | b
2254 // select a, b, a -> a & b
2255 if (CondVal == TrueVal)
2256 return BinaryOperator::CreateOr(CondVal, FalseVal);
2257 if (CondVal == FalseVal)
2258 return BinaryOperator::CreateAnd(CondVal, TrueVal);
2260 // select a, ~a, b -> (~a) & b
2261 // select a, b, ~a -> (~a) | b
2262 if (match(TrueVal, m_Not(m_Specific(CondVal))))
2263 return BinaryOperator::CreateAnd(TrueVal, FalseVal);
2264 if (match(FalseVal, m_Not(m_Specific(CondVal))))
2265 return BinaryOperator::CreateOr(TrueVal, FalseVal);
2268 // Selecting between two integer or vector splat integer constants?
2270 // Note that we don't handle a scalar select of vectors:
2271 // select i1 %c, <2 x i8> <1, 1>, <2 x i8> <0, 0>
2272 // because that may need 3 instructions to splat the condition value:
2273 // extend, insertelement, shufflevector.
2274 if (SelType->isIntOrIntVectorTy() &&
2275 CondVal->getType()->isVectorTy() == SelType->isVectorTy()) {
2276 // select C, 1, 0 -> zext C to int
2277 if (match(TrueVal, m_One()) && match(FalseVal, m_Zero()))
2278 return new ZExtInst(CondVal, SelType);
2280 // select C, -1, 0 -> sext C to int
2281 if (match(TrueVal, m_AllOnes()) && match(FalseVal, m_Zero()))
2282 return new SExtInst(CondVal, SelType);
2284 // select C, 0, 1 -> zext !C to int
2285 if (match(TrueVal, m_Zero()) && match(FalseVal, m_One())) {
2286 Value *NotCond = Builder.CreateNot(CondVal, "not." + CondVal->getName());
2287 return new ZExtInst(NotCond, SelType);
2290 // select C, 0, -1 -> sext !C to int
2291 if (match(TrueVal, m_Zero()) && match(FalseVal, m_AllOnes())) {
2292 Value *NotCond = Builder.CreateNot(CondVal, "not." + CondVal->getName());
2293 return new SExtInst(NotCond, SelType);
2297 // See if we are selecting two values based on a comparison of the two values.
2298 if (FCmpInst *FCI = dyn_cast<FCmpInst>(CondVal)) {
2299 if (FCI->getOperand(0) == TrueVal && FCI->getOperand(1) == FalseVal) {
2300 // Canonicalize to use ordered comparisons by swapping the select
2301 // operands.
2303 // e.g.
2304 // (X ugt Y) ? X : Y -> (X ole Y) ? Y : X
2305 if (FCI->hasOneUse() && FCmpInst::isUnordered(FCI->getPredicate())) {
2306 FCmpInst::Predicate InvPred = FCI->getInversePredicate();
2307 IRBuilder<>::FastMathFlagGuard FMFG(Builder);
2308 Builder.setFastMathFlags(FCI->getFastMathFlags());
2309 Value *NewCond = Builder.CreateFCmp(InvPred, TrueVal, FalseVal,
2310 FCI->getName() + ".inv");
2312 return SelectInst::Create(NewCond, FalseVal, TrueVal,
2313 SI.getName() + ".p");
2316 // NOTE: if we wanted to, this is where to detect MIN/MAX
2317 } else if (FCI->getOperand(0) == FalseVal && FCI->getOperand(1) == TrueVal){
2318 // Canonicalize to use ordered comparisons by swapping the select
2319 // operands.
2321 // e.g.
2322 // (X ugt Y) ? X : Y -> (X ole Y) ? X : Y
2323 if (FCI->hasOneUse() && FCmpInst::isUnordered(FCI->getPredicate())) {
2324 FCmpInst::Predicate InvPred = FCI->getInversePredicate();
2325 IRBuilder<>::FastMathFlagGuard FMFG(Builder);
2326 Builder.setFastMathFlags(FCI->getFastMathFlags());
2327 Value *NewCond = Builder.CreateFCmp(InvPred, FalseVal, TrueVal,
2328 FCI->getName() + ".inv");
2330 return SelectInst::Create(NewCond, FalseVal, TrueVal,
2331 SI.getName() + ".p");
2334 // NOTE: if we wanted to, this is where to detect MIN/MAX
2338 // Canonicalize select with fcmp to fabs(). -0.0 makes this tricky. We need
2339 // fast-math-flags (nsz) or fsub with +0.0 (not fneg) for this to work. We
2340 // also require nnan because we do not want to unintentionally change the
2341 // sign of a NaN value.
2342 // FIXME: These folds should test/propagate FMF from the select, not the
2343 // fsub or fneg.
2344 // (X <= +/-0.0) ? (0.0 - X) : X --> fabs(X)
2345 Instruction *FSub;
2346 if (match(CondVal, m_FCmp(Pred, m_Specific(FalseVal), m_AnyZeroFP())) &&
2347 match(TrueVal, m_FSub(m_PosZeroFP(), m_Specific(FalseVal))) &&
2348 match(TrueVal, m_Instruction(FSub)) && FSub->hasNoNaNs() &&
2349 (Pred == FCmpInst::FCMP_OLE || Pred == FCmpInst::FCMP_ULE)) {
2350 Value *Fabs = Builder.CreateUnaryIntrinsic(Intrinsic::fabs, FalseVal, FSub);
2351 return replaceInstUsesWith(SI, Fabs);
2353 // (X > +/-0.0) ? X : (0.0 - X) --> fabs(X)
2354 if (match(CondVal, m_FCmp(Pred, m_Specific(TrueVal), m_AnyZeroFP())) &&
2355 match(FalseVal, m_FSub(m_PosZeroFP(), m_Specific(TrueVal))) &&
2356 match(FalseVal, m_Instruction(FSub)) && FSub->hasNoNaNs() &&
2357 (Pred == FCmpInst::FCMP_OGT || Pred == FCmpInst::FCMP_UGT)) {
2358 Value *Fabs = Builder.CreateUnaryIntrinsic(Intrinsic::fabs, TrueVal, FSub);
2359 return replaceInstUsesWith(SI, Fabs);
2361 // With nnan and nsz:
2362 // (X < +/-0.0) ? -X : X --> fabs(X)
2363 // (X <= +/-0.0) ? -X : X --> fabs(X)
2364 Instruction *FNeg;
2365 if (match(CondVal, m_FCmp(Pred, m_Specific(FalseVal), m_AnyZeroFP())) &&
2366 match(TrueVal, m_FNeg(m_Specific(FalseVal))) &&
2367 match(TrueVal, m_Instruction(FNeg)) &&
2368 FNeg->hasNoNaNs() && FNeg->hasNoSignedZeros() &&
2369 (Pred == FCmpInst::FCMP_OLT || Pred == FCmpInst::FCMP_OLE ||
2370 Pred == FCmpInst::FCMP_ULT || Pred == FCmpInst::FCMP_ULE)) {
2371 Value *Fabs = Builder.CreateUnaryIntrinsic(Intrinsic::fabs, FalseVal, FNeg);
2372 return replaceInstUsesWith(SI, Fabs);
2374 // With nnan and nsz:
2375 // (X > +/-0.0) ? X : -X --> fabs(X)
2376 // (X >= +/-0.0) ? X : -X --> fabs(X)
2377 if (match(CondVal, m_FCmp(Pred, m_Specific(TrueVal), m_AnyZeroFP())) &&
2378 match(FalseVal, m_FNeg(m_Specific(TrueVal))) &&
2379 match(FalseVal, m_Instruction(FNeg)) &&
2380 FNeg->hasNoNaNs() && FNeg->hasNoSignedZeros() &&
2381 (Pred == FCmpInst::FCMP_OGT || Pred == FCmpInst::FCMP_OGE ||
2382 Pred == FCmpInst::FCMP_UGT || Pred == FCmpInst::FCMP_UGE)) {
2383 Value *Fabs = Builder.CreateUnaryIntrinsic(Intrinsic::fabs, TrueVal, FNeg);
2384 return replaceInstUsesWith(SI, Fabs);
2387 // See if we are selecting two values based on a comparison of the two values.
2388 if (ICmpInst *ICI = dyn_cast<ICmpInst>(CondVal))
2389 if (Instruction *Result = foldSelectInstWithICmp(SI, ICI))
2390 return Result;
2392 if (Instruction *Add = foldAddSubSelect(SI, Builder))
2393 return Add;
2395 // Turn (select C, (op X, Y), (op X, Z)) -> (op X, (select C, Y, Z))
2396 auto *TI = dyn_cast<Instruction>(TrueVal);
2397 auto *FI = dyn_cast<Instruction>(FalseVal);
2398 if (TI && FI && TI->getOpcode() == FI->getOpcode())
2399 if (Instruction *IV = foldSelectOpOp(SI, TI, FI))
2400 return IV;
2402 if (Instruction *I = foldSelectExtConst(SI))
2403 return I;
2405 // See if we can fold the select into one of our operands.
2406 if (SelType->isIntOrIntVectorTy() || SelType->isFPOrFPVectorTy()) {
2407 if (Instruction *FoldI = foldSelectIntoOp(SI, TrueVal, FalseVal))
2408 return FoldI;
2410 Value *LHS, *RHS;
2411 Instruction::CastOps CastOp;
2412 SelectPatternResult SPR = matchSelectPattern(&SI, LHS, RHS, &CastOp);
2413 auto SPF = SPR.Flavor;
2414 if (SPF) {
2415 Value *LHS2, *RHS2;
2416 if (SelectPatternFlavor SPF2 = matchSelectPattern(LHS, LHS2, RHS2).Flavor)
2417 if (Instruction *R = foldSPFofSPF(cast<Instruction>(LHS), SPF2, LHS2,
2418 RHS2, SI, SPF, RHS))
2419 return R;
2420 if (SelectPatternFlavor SPF2 = matchSelectPattern(RHS, LHS2, RHS2).Flavor)
2421 if (Instruction *R = foldSPFofSPF(cast<Instruction>(RHS), SPF2, LHS2,
2422 RHS2, SI, SPF, LHS))
2423 return R;
2424 // TODO.
2425 // ABS(-X) -> ABS(X)
2428 if (SelectPatternResult::isMinOrMax(SPF)) {
2429 // Canonicalize so that
2430 // - type casts are outside select patterns.
2431 // - float clamp is transformed to min/max pattern
2433 bool IsCastNeeded = LHS->getType() != SelType;
2434 Value *CmpLHS = cast<CmpInst>(CondVal)->getOperand(0);
2435 Value *CmpRHS = cast<CmpInst>(CondVal)->getOperand(1);
2436 if (IsCastNeeded ||
2437 (LHS->getType()->isFPOrFPVectorTy() &&
2438 ((CmpLHS != LHS && CmpLHS != RHS) ||
2439 (CmpRHS != LHS && CmpRHS != RHS)))) {
2440 CmpInst::Predicate MinMaxPred = getMinMaxPred(SPF, SPR.Ordered);
2442 Value *Cmp;
2443 if (CmpInst::isIntPredicate(MinMaxPred)) {
2444 Cmp = Builder.CreateICmp(MinMaxPred, LHS, RHS);
2445 } else {
2446 IRBuilder<>::FastMathFlagGuard FMFG(Builder);
2447 auto FMF =
2448 cast<FPMathOperator>(SI.getCondition())->getFastMathFlags();
2449 Builder.setFastMathFlags(FMF);
2450 Cmp = Builder.CreateFCmp(MinMaxPred, LHS, RHS);
2453 Value *NewSI = Builder.CreateSelect(Cmp, LHS, RHS, SI.getName(), &SI);
2454 if (!IsCastNeeded)
2455 return replaceInstUsesWith(SI, NewSI);
2457 Value *NewCast = Builder.CreateCast(CastOp, NewSI, SelType);
2458 return replaceInstUsesWith(SI, NewCast);
2461 // MAX(~a, ~b) -> ~MIN(a, b)
2462 // MAX(~a, C) -> ~MIN(a, ~C)
2463 // MIN(~a, ~b) -> ~MAX(a, b)
2464 // MIN(~a, C) -> ~MAX(a, ~C)
2465 auto moveNotAfterMinMax = [&](Value *X, Value *Y) -> Instruction * {
2466 Value *A;
2467 if (match(X, m_Not(m_Value(A))) && !X->hasNUsesOrMore(3) &&
2468 !isFreeToInvert(A, A->hasOneUse()) &&
2469 // Passing false to only consider m_Not and constants.
2470 isFreeToInvert(Y, false)) {
2471 Value *B = Builder.CreateNot(Y);
2472 Value *NewMinMax = createMinMax(Builder, getInverseMinMaxFlavor(SPF),
2473 A, B);
2474 // Copy the profile metadata.
2475 if (MDNode *MD = SI.getMetadata(LLVMContext::MD_prof)) {
2476 cast<SelectInst>(NewMinMax)->setMetadata(LLVMContext::MD_prof, MD);
2477 // Swap the metadata if the operands are swapped.
2478 if (X == SI.getFalseValue() && Y == SI.getTrueValue())
2479 cast<SelectInst>(NewMinMax)->swapProfMetadata();
2482 return BinaryOperator::CreateNot(NewMinMax);
2485 return nullptr;
2488 if (Instruction *I = moveNotAfterMinMax(LHS, RHS))
2489 return I;
2490 if (Instruction *I = moveNotAfterMinMax(RHS, LHS))
2491 return I;
2493 if (Instruction *I = moveAddAfterMinMax(SPF, LHS, RHS, Builder))
2494 return I;
2496 if (Instruction *I = factorizeMinMaxTree(SPF, LHS, RHS, Builder))
2497 return I;
2498 if (Instruction *I = matchSAddSubSat(SI))
2499 return I;
2503 // Canonicalize select of FP values where NaN and -0.0 are not valid as
2504 // minnum/maxnum intrinsics.
2505 if (isa<FPMathOperator>(SI) && SI.hasNoNaNs() && SI.hasNoSignedZeros()) {
2506 Value *X, *Y;
2507 if (match(&SI, m_OrdFMax(m_Value(X), m_Value(Y))))
2508 return replaceInstUsesWith(
2509 SI, Builder.CreateBinaryIntrinsic(Intrinsic::maxnum, X, Y, &SI));
2511 if (match(&SI, m_OrdFMin(m_Value(X), m_Value(Y))))
2512 return replaceInstUsesWith(
2513 SI, Builder.CreateBinaryIntrinsic(Intrinsic::minnum, X, Y, &SI));
2516 // See if we can fold the select into a phi node if the condition is a select.
2517 if (auto *PN = dyn_cast<PHINode>(SI.getCondition()))
2518 // The true/false values have to be live in the PHI predecessor's blocks.
2519 if (canSelectOperandBeMappingIntoPredBlock(TrueVal, SI) &&
2520 canSelectOperandBeMappingIntoPredBlock(FalseVal, SI))
2521 if (Instruction *NV = foldOpIntoPhi(SI, PN))
2522 return NV;
2524 if (SelectInst *TrueSI = dyn_cast<SelectInst>(TrueVal)) {
2525 if (TrueSI->getCondition()->getType() == CondVal->getType()) {
2526 // select(C, select(C, a, b), c) -> select(C, a, c)
2527 if (TrueSI->getCondition() == CondVal) {
2528 if (SI.getTrueValue() == TrueSI->getTrueValue())
2529 return nullptr;
2530 SI.setOperand(1, TrueSI->getTrueValue());
2531 return &SI;
2533 // select(C0, select(C1, a, b), b) -> select(C0&C1, a, b)
2534 // We choose this as normal form to enable folding on the And and shortening
2535 // paths for the values (this helps GetUnderlyingObjects() for example).
2536 if (TrueSI->getFalseValue() == FalseVal && TrueSI->hasOneUse()) {
2537 Value *And = Builder.CreateAnd(CondVal, TrueSI->getCondition());
2538 SI.setOperand(0, And);
2539 SI.setOperand(1, TrueSI->getTrueValue());
2540 return &SI;
2544 if (SelectInst *FalseSI = dyn_cast<SelectInst>(FalseVal)) {
2545 if (FalseSI->getCondition()->getType() == CondVal->getType()) {
2546 // select(C, a, select(C, b, c)) -> select(C, a, c)
2547 if (FalseSI->getCondition() == CondVal) {
2548 if (SI.getFalseValue() == FalseSI->getFalseValue())
2549 return nullptr;
2550 SI.setOperand(2, FalseSI->getFalseValue());
2551 return &SI;
2553 // select(C0, a, select(C1, a, b)) -> select(C0|C1, a, b)
2554 if (FalseSI->getTrueValue() == TrueVal && FalseSI->hasOneUse()) {
2555 Value *Or = Builder.CreateOr(CondVal, FalseSI->getCondition());
2556 SI.setOperand(0, Or);
2557 SI.setOperand(2, FalseSI->getFalseValue());
2558 return &SI;
2563 auto canMergeSelectThroughBinop = [](BinaryOperator *BO) {
2564 // The select might be preventing a division by 0.
2565 switch (BO->getOpcode()) {
2566 default:
2567 return true;
2568 case Instruction::SRem:
2569 case Instruction::URem:
2570 case Instruction::SDiv:
2571 case Instruction::UDiv:
2572 return false;
2576 // Try to simplify a binop sandwiched between 2 selects with the same
2577 // condition.
2578 // select(C, binop(select(C, X, Y), W), Z) -> select(C, binop(X, W), Z)
2579 BinaryOperator *TrueBO;
2580 if (match(TrueVal, m_OneUse(m_BinOp(TrueBO))) &&
2581 canMergeSelectThroughBinop(TrueBO)) {
2582 if (auto *TrueBOSI = dyn_cast<SelectInst>(TrueBO->getOperand(0))) {
2583 if (TrueBOSI->getCondition() == CondVal) {
2584 TrueBO->setOperand(0, TrueBOSI->getTrueValue());
2585 Worklist.Add(TrueBO);
2586 return &SI;
2589 if (auto *TrueBOSI = dyn_cast<SelectInst>(TrueBO->getOperand(1))) {
2590 if (TrueBOSI->getCondition() == CondVal) {
2591 TrueBO->setOperand(1, TrueBOSI->getTrueValue());
2592 Worklist.Add(TrueBO);
2593 return &SI;
2598 // select(C, Z, binop(select(C, X, Y), W)) -> select(C, Z, binop(Y, W))
2599 BinaryOperator *FalseBO;
2600 if (match(FalseVal, m_OneUse(m_BinOp(FalseBO))) &&
2601 canMergeSelectThroughBinop(FalseBO)) {
2602 if (auto *FalseBOSI = dyn_cast<SelectInst>(FalseBO->getOperand(0))) {
2603 if (FalseBOSI->getCondition() == CondVal) {
2604 FalseBO->setOperand(0, FalseBOSI->getFalseValue());
2605 Worklist.Add(FalseBO);
2606 return &SI;
2609 if (auto *FalseBOSI = dyn_cast<SelectInst>(FalseBO->getOperand(1))) {
2610 if (FalseBOSI->getCondition() == CondVal) {
2611 FalseBO->setOperand(1, FalseBOSI->getFalseValue());
2612 Worklist.Add(FalseBO);
2613 return &SI;
2618 Value *NotCond;
2619 if (match(CondVal, m_Not(m_Value(NotCond)))) {
2620 SI.setOperand(0, NotCond);
2621 SI.setOperand(1, FalseVal);
2622 SI.setOperand(2, TrueVal);
2623 SI.swapProfMetadata();
2624 return &SI;
2627 if (VectorType *VecTy = dyn_cast<VectorType>(SelType)) {
2628 unsigned VWidth = VecTy->getNumElements();
2629 APInt UndefElts(VWidth, 0);
2630 APInt AllOnesEltMask(APInt::getAllOnesValue(VWidth));
2631 if (Value *V = SimplifyDemandedVectorElts(&SI, AllOnesEltMask, UndefElts)) {
2632 if (V != &SI)
2633 return replaceInstUsesWith(SI, V);
2634 return &SI;
2638 // If we can compute the condition, there's no need for a select.
2639 // Like the above fold, we are attempting to reduce compile-time cost by
2640 // putting this fold here with limitations rather than in InstSimplify.
2641 // The motivation for this call into value tracking is to take advantage of
2642 // the assumption cache, so make sure that is populated.
2643 if (!CondVal->getType()->isVectorTy() && !AC.assumptions().empty()) {
2644 KnownBits Known(1);
2645 computeKnownBits(CondVal, Known, 0, &SI);
2646 if (Known.One.isOneValue())
2647 return replaceInstUsesWith(SI, TrueVal);
2648 if (Known.Zero.isOneValue())
2649 return replaceInstUsesWith(SI, FalseVal);
2652 if (Instruction *BitCastSel = foldSelectCmpBitcasts(SI, Builder))
2653 return BitCastSel;
2655 // Simplify selects that test the returned flag of cmpxchg instructions.
2656 if (Instruction *Select = foldSelectCmpXchg(SI))
2657 return Select;
2659 if (Instruction *Select = foldSelectBinOpIdentity(SI, TLI))
2660 return Select;
2662 if (Instruction *Rot = foldSelectRotate(SI))
2663 return Rot;
2665 return nullptr;