xref: /freebsd/contrib/llvm-project/llvm/lib/Transforms/InstCombine/InstCombineSelect.cpp (revision 1323ec571215a77ddd21294f0871979d5ad6b992)
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.
10 //
11 //===----------------------------------------------------------------------===//
12 
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/OverflowInstAnalysis.h"
22 #include "llvm/Analysis/ValueTracking.h"
23 #include "llvm/IR/BasicBlock.h"
24 #include "llvm/IR/Constant.h"
25 #include "llvm/IR/Constants.h"
26 #include "llvm/IR/DerivedTypes.h"
27 #include "llvm/IR/IRBuilder.h"
28 #include "llvm/IR/InstrTypes.h"
29 #include "llvm/IR/Instruction.h"
30 #include "llvm/IR/Instructions.h"
31 #include "llvm/IR/IntrinsicInst.h"
32 #include "llvm/IR/Intrinsics.h"
33 #include "llvm/IR/Operator.h"
34 #include "llvm/IR/PatternMatch.h"
35 #include "llvm/IR/Type.h"
36 #include "llvm/IR/User.h"
37 #include "llvm/IR/Value.h"
38 #include "llvm/Support/Casting.h"
39 #include "llvm/Support/ErrorHandling.h"
40 #include "llvm/Support/KnownBits.h"
41 #include "llvm/Transforms/InstCombine/InstCombineWorklist.h"
42 #include "llvm/Transforms/InstCombine/InstCombiner.h"
43 #include <cassert>
44 #include <utility>
45 
46 using namespace llvm;
47 using namespace PatternMatch;
48 
49 #define DEBUG_TYPE "instcombine"
50 
51 static Value *createMinMax(InstCombiner::BuilderTy &Builder,
52                            SelectPatternFlavor SPF, Value *A, Value *B) {
53   CmpInst::Predicate Pred = getMinMaxPred(SPF);
54   assert(CmpInst::isIntPredicate(Pred) && "Expected integer predicate");
55   return Builder.CreateSelect(Builder.CreateICmp(Pred, A, B), A, B);
56 }
57 
58 /// Replace a select operand based on an equality comparison with the identity
59 /// constant of a binop.
60 static Instruction *foldSelectBinOpIdentity(SelectInst &Sel,
61                                             const TargetLibraryInfo &TLI,
62                                             InstCombinerImpl &IC) {
63   // The select condition must be an equality compare with a constant operand.
64   Value *X;
65   Constant *C;
66   CmpInst::Predicate Pred;
67   if (!match(Sel.getCondition(), m_Cmp(Pred, m_Value(X), m_Constant(C))))
68     return nullptr;
69 
70   bool IsEq;
71   if (ICmpInst::isEquality(Pred))
72     IsEq = Pred == ICmpInst::ICMP_EQ;
73   else if (Pred == FCmpInst::FCMP_OEQ)
74     IsEq = true;
75   else if (Pred == FCmpInst::FCMP_UNE)
76     IsEq = false;
77   else
78     return nullptr;
79 
80   // A select operand must be a binop.
81   BinaryOperator *BO;
82   if (!match(Sel.getOperand(IsEq ? 1 : 2), m_BinOp(BO)))
83     return nullptr;
84 
85   // The compare constant must be the identity constant for that binop.
86   // If this a floating-point compare with 0.0, any zero constant will do.
87   Type *Ty = BO->getType();
88   Constant *IdC = ConstantExpr::getBinOpIdentity(BO->getOpcode(), Ty, true);
89   if (IdC != C) {
90     if (!IdC || !CmpInst::isFPPredicate(Pred))
91       return nullptr;
92     if (!match(IdC, m_AnyZeroFP()) || !match(C, m_AnyZeroFP()))
93       return nullptr;
94   }
95 
96   // Last, match the compare variable operand with a binop operand.
97   Value *Y;
98   if (!BO->isCommutative() && !match(BO, m_BinOp(m_Value(Y), m_Specific(X))))
99     return nullptr;
100   if (!match(BO, m_c_BinOp(m_Value(Y), m_Specific(X))))
101     return nullptr;
102 
103   // +0.0 compares equal to -0.0, and so it does not behave as required for this
104   // transform. Bail out if we can not exclude that possibility.
105   if (isa<FPMathOperator>(BO))
106     if (!BO->hasNoSignedZeros() && !CannotBeNegativeZero(Y, &TLI))
107       return nullptr;
108 
109   // BO = binop Y, X
110   // S = { select (cmp eq X, C), BO, ? } or { select (cmp ne X, C), ?, BO }
111   // =>
112   // S = { select (cmp eq X, C),  Y, ? } or { select (cmp ne X, C), ?,  Y }
113   return IC.replaceOperand(Sel, IsEq ? 1 : 2, Y);
114 }
115 
116 /// This folds:
117 ///  select (icmp eq (and X, C1)), TC, FC
118 ///    iff C1 is a power 2 and the difference between TC and FC is a power-of-2.
119 /// To something like:
120 ///  (shr (and (X, C1)), (log2(C1) - log2(TC-FC))) + FC
121 /// Or:
122 ///  (shl (and (X, C1)), (log2(TC-FC) - log2(C1))) + FC
123 /// With some variations depending if FC is larger than TC, or the shift
124 /// isn't needed, or the bit widths don't match.
125 static Value *foldSelectICmpAnd(SelectInst &Sel, ICmpInst *Cmp,
126                                 InstCombiner::BuilderTy &Builder) {
127   const APInt *SelTC, *SelFC;
128   if (!match(Sel.getTrueValue(), m_APInt(SelTC)) ||
129       !match(Sel.getFalseValue(), m_APInt(SelFC)))
130     return nullptr;
131 
132   // If this is a vector select, we need a vector compare.
133   Type *SelType = Sel.getType();
134   if (SelType->isVectorTy() != Cmp->getType()->isVectorTy())
135     return nullptr;
136 
137   Value *V;
138   APInt AndMask;
139   bool CreateAnd = false;
140   ICmpInst::Predicate Pred = Cmp->getPredicate();
141   if (ICmpInst::isEquality(Pred)) {
142     if (!match(Cmp->getOperand(1), m_Zero()))
143       return nullptr;
144 
145     V = Cmp->getOperand(0);
146     const APInt *AndRHS;
147     if (!match(V, m_And(m_Value(), m_Power2(AndRHS))))
148       return nullptr;
149 
150     AndMask = *AndRHS;
151   } else if (decomposeBitTestICmp(Cmp->getOperand(0), Cmp->getOperand(1),
152                                   Pred, V, AndMask)) {
153     assert(ICmpInst::isEquality(Pred) && "Not equality test?");
154     if (!AndMask.isPowerOf2())
155       return nullptr;
156 
157     CreateAnd = true;
158   } else {
159     return nullptr;
160   }
161 
162   // In general, when both constants are non-zero, we would need an offset to
163   // replace the select. This would require more instructions than we started
164   // with. But there's one special-case that we handle here because it can
165   // simplify/reduce the instructions.
166   APInt TC = *SelTC;
167   APInt FC = *SelFC;
168   if (!TC.isNullValue() && !FC.isNullValue()) {
169     // If the select constants differ by exactly one bit and that's the same
170     // bit that is masked and checked by the select condition, the select can
171     // be replaced by bitwise logic to set/clear one bit of the constant result.
172     if (TC.getBitWidth() != AndMask.getBitWidth() || (TC ^ FC) != AndMask)
173       return nullptr;
174     if (CreateAnd) {
175       // If we have to create an 'and', then we must kill the cmp to not
176       // increase the instruction count.
177       if (!Cmp->hasOneUse())
178         return nullptr;
179       V = Builder.CreateAnd(V, ConstantInt::get(SelType, AndMask));
180     }
181     bool ExtraBitInTC = TC.ugt(FC);
182     if (Pred == ICmpInst::ICMP_EQ) {
183       // If the masked bit in V is clear, clear or set the bit in the result:
184       // (V & AndMaskC) == 0 ? TC : FC --> (V & AndMaskC) ^ TC
185       // (V & AndMaskC) == 0 ? TC : FC --> (V & AndMaskC) | TC
186       Constant *C = ConstantInt::get(SelType, TC);
187       return ExtraBitInTC ? Builder.CreateXor(V, C) : Builder.CreateOr(V, C);
188     }
189     if (Pred == ICmpInst::ICMP_NE) {
190       // If the masked bit in V is set, set or clear the bit in the result:
191       // (V & AndMaskC) != 0 ? TC : FC --> (V & AndMaskC) | FC
192       // (V & AndMaskC) != 0 ? TC : FC --> (V & AndMaskC) ^ FC
193       Constant *C = ConstantInt::get(SelType, FC);
194       return ExtraBitInTC ? Builder.CreateOr(V, C) : Builder.CreateXor(V, C);
195     }
196     llvm_unreachable("Only expecting equality predicates");
197   }
198 
199   // Make sure one of the select arms is a power-of-2.
200   if (!TC.isPowerOf2() && !FC.isPowerOf2())
201     return nullptr;
202 
203   // Determine which shift is needed to transform result of the 'and' into the
204   // desired result.
205   const APInt &ValC = !TC.isNullValue() ? TC : FC;
206   unsigned ValZeros = ValC.logBase2();
207   unsigned AndZeros = AndMask.logBase2();
208 
209   // Insert the 'and' instruction on the input to the truncate.
210   if (CreateAnd)
211     V = Builder.CreateAnd(V, ConstantInt::get(V->getType(), AndMask));
212 
213   // If types don't match, we can still convert the select by introducing a zext
214   // or a trunc of the 'and'.
215   if (ValZeros > AndZeros) {
216     V = Builder.CreateZExtOrTrunc(V, SelType);
217     V = Builder.CreateShl(V, ValZeros - AndZeros);
218   } else if (ValZeros < AndZeros) {
219     V = Builder.CreateLShr(V, AndZeros - ValZeros);
220     V = Builder.CreateZExtOrTrunc(V, SelType);
221   } else {
222     V = Builder.CreateZExtOrTrunc(V, SelType);
223   }
224 
225   // Okay, now we know that everything is set up, we just don't know whether we
226   // have a icmp_ne or icmp_eq and whether the true or false val is the zero.
227   bool ShouldNotVal = !TC.isNullValue();
228   ShouldNotVal ^= Pred == ICmpInst::ICMP_NE;
229   if (ShouldNotVal)
230     V = Builder.CreateXor(V, ValC);
231 
232   return V;
233 }
234 
235 /// We want to turn code that looks like this:
236 ///   %C = or %A, %B
237 ///   %D = select %cond, %C, %A
238 /// into:
239 ///   %C = select %cond, %B, 0
240 ///   %D = or %A, %C
241 ///
242 /// Assuming that the specified instruction is an operand to the select, return
243 /// a bitmask indicating which operands of this instruction are foldable if they
244 /// equal the other incoming value of the select.
245 static unsigned getSelectFoldableOperands(BinaryOperator *I) {
246   switch (I->getOpcode()) {
247   case Instruction::Add:
248   case Instruction::Mul:
249   case Instruction::And:
250   case Instruction::Or:
251   case Instruction::Xor:
252     return 3;              // Can fold through either operand.
253   case Instruction::Sub:   // Can only fold on the amount subtracted.
254   case Instruction::Shl:   // Can only fold on the shift amount.
255   case Instruction::LShr:
256   case Instruction::AShr:
257     return 1;
258   default:
259     return 0;              // Cannot fold
260   }
261 }
262 
263 /// We have (select c, TI, FI), and we know that TI and FI have the same opcode.
264 Instruction *InstCombinerImpl::foldSelectOpOp(SelectInst &SI, Instruction *TI,
265                                               Instruction *FI) {
266   // Don't break up min/max patterns. The hasOneUse checks below prevent that
267   // for most cases, but vector min/max with bitcasts can be transformed. If the
268   // one-use restrictions are eased for other patterns, we still don't want to
269   // obfuscate min/max.
270   if ((match(&SI, m_SMin(m_Value(), m_Value())) ||
271        match(&SI, m_SMax(m_Value(), m_Value())) ||
272        match(&SI, m_UMin(m_Value(), m_Value())) ||
273        match(&SI, m_UMax(m_Value(), m_Value()))))
274     return nullptr;
275 
276   // If this is a cast from the same type, merge.
277   Value *Cond = SI.getCondition();
278   Type *CondTy = Cond->getType();
279   if (TI->getNumOperands() == 1 && TI->isCast()) {
280     Type *FIOpndTy = FI->getOperand(0)->getType();
281     if (TI->getOperand(0)->getType() != FIOpndTy)
282       return nullptr;
283 
284     // The select condition may be a vector. We may only change the operand
285     // type if the vector width remains the same (and matches the condition).
286     if (auto *CondVTy = dyn_cast<VectorType>(CondTy)) {
287       if (!FIOpndTy->isVectorTy() ||
288           CondVTy->getElementCount() !=
289               cast<VectorType>(FIOpndTy)->getElementCount())
290         return nullptr;
291 
292       // TODO: If the backend knew how to deal with casts better, we could
293       // remove this limitation. For now, there's too much potential to create
294       // worse codegen by promoting the select ahead of size-altering casts
295       // (PR28160).
296       //
297       // Note that ValueTracking's matchSelectPattern() looks through casts
298       // without checking 'hasOneUse' when it matches min/max patterns, so this
299       // transform may end up happening anyway.
300       if (TI->getOpcode() != Instruction::BitCast &&
301           (!TI->hasOneUse() || !FI->hasOneUse()))
302         return nullptr;
303     } else if (!TI->hasOneUse() || !FI->hasOneUse()) {
304       // TODO: The one-use restrictions for a scalar select could be eased if
305       // the fold of a select in visitLoadInst() was enhanced to match a pattern
306       // that includes a cast.
307       return nullptr;
308     }
309 
310     // Fold this by inserting a select from the input values.
311     Value *NewSI =
312         Builder.CreateSelect(Cond, TI->getOperand(0), FI->getOperand(0),
313                              SI.getName() + ".v", &SI);
314     return CastInst::Create(Instruction::CastOps(TI->getOpcode()), NewSI,
315                             TI->getType());
316   }
317 
318   // Cond ? -X : -Y --> -(Cond ? X : Y)
319   Value *X, *Y;
320   if (match(TI, m_FNeg(m_Value(X))) && match(FI, m_FNeg(m_Value(Y))) &&
321       (TI->hasOneUse() || FI->hasOneUse())) {
322     Value *NewSel = Builder.CreateSelect(Cond, X, Y, SI.getName() + ".v", &SI);
323     return UnaryOperator::CreateFNegFMF(NewSel, TI);
324   }
325 
326   // Min/max intrinsic with a common operand can have the common operand pulled
327   // after the select. This is the same transform as below for binops, but
328   // specialized for intrinsic matching and without the restrictive uses clause.
329   auto *TII = dyn_cast<IntrinsicInst>(TI);
330   auto *FII = dyn_cast<IntrinsicInst>(FI);
331   if (TII && FII && TII->getIntrinsicID() == FII->getIntrinsicID() &&
332       (TII->hasOneUse() || FII->hasOneUse())) {
333     Value *T0, *T1, *F0, *F1;
334     if (match(TII, m_MaxOrMin(m_Value(T0), m_Value(T1))) &&
335         match(FII, m_MaxOrMin(m_Value(F0), m_Value(F1)))) {
336       if (T0 == F0) {
337         Value *NewSel = Builder.CreateSelect(Cond, T1, F1, "minmaxop", &SI);
338         return CallInst::Create(TII->getCalledFunction(), {NewSel, T0});
339       }
340       if (T0 == F1) {
341         Value *NewSel = Builder.CreateSelect(Cond, T1, F0, "minmaxop", &SI);
342         return CallInst::Create(TII->getCalledFunction(), {NewSel, T0});
343       }
344       if (T1 == F0) {
345         Value *NewSel = Builder.CreateSelect(Cond, T0, F1, "minmaxop", &SI);
346         return CallInst::Create(TII->getCalledFunction(), {NewSel, T1});
347       }
348       if (T1 == F1) {
349         Value *NewSel = Builder.CreateSelect(Cond, T0, F0, "minmaxop", &SI);
350         return CallInst::Create(TII->getCalledFunction(), {NewSel, T1});
351       }
352     }
353   }
354 
355   // Only handle binary operators (including two-operand getelementptr) with
356   // one-use here. As with the cast case above, it may be possible to relax the
357   // one-use constraint, but that needs be examined carefully since it may not
358   // reduce the total number of instructions.
359   if (TI->getNumOperands() != 2 || FI->getNumOperands() != 2 ||
360       (!isa<BinaryOperator>(TI) && !isa<GetElementPtrInst>(TI)) ||
361       !TI->hasOneUse() || !FI->hasOneUse())
362     return nullptr;
363 
364   // Figure out if the operations have any operands in common.
365   Value *MatchOp, *OtherOpT, *OtherOpF;
366   bool MatchIsOpZero;
367   if (TI->getOperand(0) == FI->getOperand(0)) {
368     MatchOp  = TI->getOperand(0);
369     OtherOpT = TI->getOperand(1);
370     OtherOpF = FI->getOperand(1);
371     MatchIsOpZero = true;
372   } else if (TI->getOperand(1) == FI->getOperand(1)) {
373     MatchOp  = TI->getOperand(1);
374     OtherOpT = TI->getOperand(0);
375     OtherOpF = FI->getOperand(0);
376     MatchIsOpZero = false;
377   } else if (!TI->isCommutative()) {
378     return nullptr;
379   } else if (TI->getOperand(0) == FI->getOperand(1)) {
380     MatchOp  = TI->getOperand(0);
381     OtherOpT = TI->getOperand(1);
382     OtherOpF = FI->getOperand(0);
383     MatchIsOpZero = true;
384   } else if (TI->getOperand(1) == FI->getOperand(0)) {
385     MatchOp  = TI->getOperand(1);
386     OtherOpT = TI->getOperand(0);
387     OtherOpF = FI->getOperand(1);
388     MatchIsOpZero = true;
389   } else {
390     return nullptr;
391   }
392 
393   // If the select condition is a vector, the operands of the original select's
394   // operands also must be vectors. This may not be the case for getelementptr
395   // for example.
396   if (CondTy->isVectorTy() && (!OtherOpT->getType()->isVectorTy() ||
397                                !OtherOpF->getType()->isVectorTy()))
398     return nullptr;
399 
400   // If we reach here, they do have operations in common.
401   Value *NewSI = Builder.CreateSelect(Cond, OtherOpT, OtherOpF,
402                                       SI.getName() + ".v", &SI);
403   Value *Op0 = MatchIsOpZero ? MatchOp : NewSI;
404   Value *Op1 = MatchIsOpZero ? NewSI : MatchOp;
405   if (auto *BO = dyn_cast<BinaryOperator>(TI)) {
406     BinaryOperator *NewBO = BinaryOperator::Create(BO->getOpcode(), Op0, Op1);
407     NewBO->copyIRFlags(TI);
408     NewBO->andIRFlags(FI);
409     return NewBO;
410   }
411   if (auto *TGEP = dyn_cast<GetElementPtrInst>(TI)) {
412     auto *FGEP = cast<GetElementPtrInst>(FI);
413     Type *ElementType = TGEP->getResultElementType();
414     return TGEP->isInBounds() && FGEP->isInBounds()
415                ? GetElementPtrInst::CreateInBounds(ElementType, Op0, {Op1})
416                : GetElementPtrInst::Create(ElementType, Op0, {Op1});
417   }
418   llvm_unreachable("Expected BinaryOperator or GEP");
419   return nullptr;
420 }
421 
422 static bool isSelect01(const APInt &C1I, const APInt &C2I) {
423   if (!C1I.isNullValue() && !C2I.isNullValue()) // One side must be zero.
424     return false;
425   return C1I.isOneValue() || C1I.isAllOnesValue() ||
426          C2I.isOneValue() || C2I.isAllOnesValue();
427 }
428 
429 /// Try to fold the select into one of the operands to allow further
430 /// optimization.
431 Instruction *InstCombinerImpl::foldSelectIntoOp(SelectInst &SI, Value *TrueVal,
432                                                 Value *FalseVal) {
433   // See the comment above GetSelectFoldableOperands for a description of the
434   // transformation we are doing here.
435   if (auto *TVI = dyn_cast<BinaryOperator>(TrueVal)) {
436     if (TVI->hasOneUse() && !isa<Constant>(FalseVal)) {
437       if (unsigned SFO = getSelectFoldableOperands(TVI)) {
438         unsigned OpToFold = 0;
439         if ((SFO & 1) && FalseVal == TVI->getOperand(0)) {
440           OpToFold = 1;
441         } else if ((SFO & 2) && FalseVal == TVI->getOperand(1)) {
442           OpToFold = 2;
443         }
444 
445         if (OpToFold) {
446           Constant *C = ConstantExpr::getBinOpIdentity(TVI->getOpcode(),
447                                                        TVI->getType(), true);
448           Value *OOp = TVI->getOperand(2-OpToFold);
449           // Avoid creating select between 2 constants unless it's selecting
450           // between 0, 1 and -1.
451           const APInt *OOpC;
452           bool OOpIsAPInt = match(OOp, m_APInt(OOpC));
453           if (!isa<Constant>(OOp) ||
454               (OOpIsAPInt && isSelect01(C->getUniqueInteger(), *OOpC))) {
455             Value *NewSel = Builder.CreateSelect(SI.getCondition(), OOp, C);
456             NewSel->takeName(TVI);
457             BinaryOperator *BO = BinaryOperator::Create(TVI->getOpcode(),
458                                                         FalseVal, NewSel);
459             BO->copyIRFlags(TVI);
460             return BO;
461           }
462         }
463       }
464     }
465   }
466 
467   if (auto *FVI = dyn_cast<BinaryOperator>(FalseVal)) {
468     if (FVI->hasOneUse() && !isa<Constant>(TrueVal)) {
469       if (unsigned SFO = getSelectFoldableOperands(FVI)) {
470         unsigned OpToFold = 0;
471         if ((SFO & 1) && TrueVal == FVI->getOperand(0)) {
472           OpToFold = 1;
473         } else if ((SFO & 2) && TrueVal == FVI->getOperand(1)) {
474           OpToFold = 2;
475         }
476 
477         if (OpToFold) {
478           Constant *C = ConstantExpr::getBinOpIdentity(FVI->getOpcode(),
479                                                        FVI->getType(), true);
480           Value *OOp = FVI->getOperand(2-OpToFold);
481           // Avoid creating select between 2 constants unless it's selecting
482           // between 0, 1 and -1.
483           const APInt *OOpC;
484           bool OOpIsAPInt = match(OOp, m_APInt(OOpC));
485           if (!isa<Constant>(OOp) ||
486               (OOpIsAPInt && isSelect01(C->getUniqueInteger(), *OOpC))) {
487             Value *NewSel = Builder.CreateSelect(SI.getCondition(), C, OOp);
488             NewSel->takeName(FVI);
489             BinaryOperator *BO = BinaryOperator::Create(FVI->getOpcode(),
490                                                         TrueVal, NewSel);
491             BO->copyIRFlags(FVI);
492             return BO;
493           }
494         }
495       }
496     }
497   }
498 
499   return nullptr;
500 }
501 
502 /// We want to turn:
503 ///   (select (icmp eq (and X, Y), 0), (and (lshr X, Z), 1), 1)
504 /// into:
505 ///   zext (icmp ne i32 (and X, (or Y, (shl 1, Z))), 0)
506 /// Note:
507 ///   Z may be 0 if lshr is missing.
508 /// Worst-case scenario is that we will replace 5 instructions with 5 different
509 /// instructions, but we got rid of select.
510 static Instruction *foldSelectICmpAndAnd(Type *SelType, const ICmpInst *Cmp,
511                                          Value *TVal, Value *FVal,
512                                          InstCombiner::BuilderTy &Builder) {
513   if (!(Cmp->hasOneUse() && Cmp->getOperand(0)->hasOneUse() &&
514         Cmp->getPredicate() == ICmpInst::ICMP_EQ &&
515         match(Cmp->getOperand(1), m_Zero()) && match(FVal, m_One())))
516     return nullptr;
517 
518   // The TrueVal has general form of:  and %B, 1
519   Value *B;
520   if (!match(TVal, m_OneUse(m_And(m_Value(B), m_One()))))
521     return nullptr;
522 
523   // Where %B may be optionally shifted:  lshr %X, %Z.
524   Value *X, *Z;
525   const bool HasShift = match(B, m_OneUse(m_LShr(m_Value(X), m_Value(Z))));
526   if (!HasShift)
527     X = B;
528 
529   Value *Y;
530   if (!match(Cmp->getOperand(0), m_c_And(m_Specific(X), m_Value(Y))))
531     return nullptr;
532 
533   // ((X & Y) == 0) ? ((X >> Z) & 1) : 1 --> (X & (Y | (1 << Z))) != 0
534   // ((X & Y) == 0) ? (X & 1) : 1 --> (X & (Y | 1)) != 0
535   Constant *One = ConstantInt::get(SelType, 1);
536   Value *MaskB = HasShift ? Builder.CreateShl(One, Z) : One;
537   Value *FullMask = Builder.CreateOr(Y, MaskB);
538   Value *MaskedX = Builder.CreateAnd(X, FullMask);
539   Value *ICmpNeZero = Builder.CreateIsNotNull(MaskedX);
540   return new ZExtInst(ICmpNeZero, SelType);
541 }
542 
543 /// We want to turn:
544 ///   (select (icmp sgt x, C), lshr (X, Y), ashr (X, Y)); iff C s>= -1
545 ///   (select (icmp slt x, C), ashr (X, Y), lshr (X, Y)); iff C s>= 0
546 /// into:
547 ///   ashr (X, Y)
548 static Value *foldSelectICmpLshrAshr(const ICmpInst *IC, Value *TrueVal,
549                                      Value *FalseVal,
550                                      InstCombiner::BuilderTy &Builder) {
551   ICmpInst::Predicate Pred = IC->getPredicate();
552   Value *CmpLHS = IC->getOperand(0);
553   Value *CmpRHS = IC->getOperand(1);
554   if (!CmpRHS->getType()->isIntOrIntVectorTy())
555     return nullptr;
556 
557   Value *X, *Y;
558   unsigned Bitwidth = CmpRHS->getType()->getScalarSizeInBits();
559   if ((Pred != ICmpInst::ICMP_SGT ||
560        !match(CmpRHS,
561               m_SpecificInt_ICMP(ICmpInst::ICMP_SGE, APInt(Bitwidth, -1)))) &&
562       (Pred != ICmpInst::ICMP_SLT ||
563        !match(CmpRHS,
564               m_SpecificInt_ICMP(ICmpInst::ICMP_SGE, APInt(Bitwidth, 0)))))
565     return nullptr;
566 
567   // Canonicalize so that ashr is in FalseVal.
568   if (Pred == ICmpInst::ICMP_SLT)
569     std::swap(TrueVal, FalseVal);
570 
571   if (match(TrueVal, m_LShr(m_Value(X), m_Value(Y))) &&
572       match(FalseVal, m_AShr(m_Specific(X), m_Specific(Y))) &&
573       match(CmpLHS, m_Specific(X))) {
574     const auto *Ashr = cast<Instruction>(FalseVal);
575     // if lshr is not exact and ashr is, this new ashr must not be exact.
576     bool IsExact = Ashr->isExact() && cast<Instruction>(TrueVal)->isExact();
577     return Builder.CreateAShr(X, Y, IC->getName(), IsExact);
578   }
579 
580   return nullptr;
581 }
582 
583 /// We want to turn:
584 ///   (select (icmp eq (and X, C1), 0), Y, (or Y, C2))
585 /// into:
586 ///   (or (shl (and X, C1), C3), Y)
587 /// iff:
588 ///   C1 and C2 are both powers of 2
589 /// where:
590 ///   C3 = Log(C2) - Log(C1)
591 ///
592 /// This transform handles cases where:
593 /// 1. The icmp predicate is inverted
594 /// 2. The select operands are reversed
595 /// 3. The magnitude of C2 and C1 are flipped
596 static Value *foldSelectICmpAndOr(const ICmpInst *IC, Value *TrueVal,
597                                   Value *FalseVal,
598                                   InstCombiner::BuilderTy &Builder) {
599   // Only handle integer compares. Also, if this is a vector select, we need a
600   // vector compare.
601   if (!TrueVal->getType()->isIntOrIntVectorTy() ||
602       TrueVal->getType()->isVectorTy() != IC->getType()->isVectorTy())
603     return nullptr;
604 
605   Value *CmpLHS = IC->getOperand(0);
606   Value *CmpRHS = IC->getOperand(1);
607 
608   Value *V;
609   unsigned C1Log;
610   bool IsEqualZero;
611   bool NeedAnd = false;
612   if (IC->isEquality()) {
613     if (!match(CmpRHS, m_Zero()))
614       return nullptr;
615 
616     const APInt *C1;
617     if (!match(CmpLHS, m_And(m_Value(), m_Power2(C1))))
618       return nullptr;
619 
620     V = CmpLHS;
621     C1Log = C1->logBase2();
622     IsEqualZero = IC->getPredicate() == ICmpInst::ICMP_EQ;
623   } else if (IC->getPredicate() == ICmpInst::ICMP_SLT ||
624              IC->getPredicate() == ICmpInst::ICMP_SGT) {
625     // We also need to recognize (icmp slt (trunc (X)), 0) and
626     // (icmp sgt (trunc (X)), -1).
627     IsEqualZero = IC->getPredicate() == ICmpInst::ICMP_SGT;
628     if ((IsEqualZero && !match(CmpRHS, m_AllOnes())) ||
629         (!IsEqualZero && !match(CmpRHS, m_Zero())))
630       return nullptr;
631 
632     if (!match(CmpLHS, m_OneUse(m_Trunc(m_Value(V)))))
633       return nullptr;
634 
635     C1Log = CmpLHS->getType()->getScalarSizeInBits() - 1;
636     NeedAnd = true;
637   } else {
638     return nullptr;
639   }
640 
641   const APInt *C2;
642   bool OrOnTrueVal = false;
643   bool OrOnFalseVal = match(FalseVal, m_Or(m_Specific(TrueVal), m_Power2(C2)));
644   if (!OrOnFalseVal)
645     OrOnTrueVal = match(TrueVal, m_Or(m_Specific(FalseVal), m_Power2(C2)));
646 
647   if (!OrOnFalseVal && !OrOnTrueVal)
648     return nullptr;
649 
650   Value *Y = OrOnFalseVal ? TrueVal : FalseVal;
651 
652   unsigned C2Log = C2->logBase2();
653 
654   bool NeedXor = (!IsEqualZero && OrOnFalseVal) || (IsEqualZero && OrOnTrueVal);
655   bool NeedShift = C1Log != C2Log;
656   bool NeedZExtTrunc = Y->getType()->getScalarSizeInBits() !=
657                        V->getType()->getScalarSizeInBits();
658 
659   // Make sure we don't create more instructions than we save.
660   Value *Or = OrOnFalseVal ? FalseVal : TrueVal;
661   if ((NeedShift + NeedXor + NeedZExtTrunc) >
662       (IC->hasOneUse() + Or->hasOneUse()))
663     return nullptr;
664 
665   if (NeedAnd) {
666     // Insert the AND instruction on the input to the truncate.
667     APInt C1 = APInt::getOneBitSet(V->getType()->getScalarSizeInBits(), C1Log);
668     V = Builder.CreateAnd(V, ConstantInt::get(V->getType(), C1));
669   }
670 
671   if (C2Log > C1Log) {
672     V = Builder.CreateZExtOrTrunc(V, Y->getType());
673     V = Builder.CreateShl(V, C2Log - C1Log);
674   } else if (C1Log > C2Log) {
675     V = Builder.CreateLShr(V, C1Log - C2Log);
676     V = Builder.CreateZExtOrTrunc(V, Y->getType());
677   } else
678     V = Builder.CreateZExtOrTrunc(V, Y->getType());
679 
680   if (NeedXor)
681     V = Builder.CreateXor(V, *C2);
682 
683   return Builder.CreateOr(V, Y);
684 }
685 
686 /// Canonicalize a set or clear of a masked set of constant bits to
687 /// select-of-constants form.
688 static Instruction *foldSetClearBits(SelectInst &Sel,
689                                      InstCombiner::BuilderTy &Builder) {
690   Value *Cond = Sel.getCondition();
691   Value *T = Sel.getTrueValue();
692   Value *F = Sel.getFalseValue();
693   Type *Ty = Sel.getType();
694   Value *X;
695   const APInt *NotC, *C;
696 
697   // Cond ? (X & ~C) : (X | C) --> (X & ~C) | (Cond ? 0 : C)
698   if (match(T, m_And(m_Value(X), m_APInt(NotC))) &&
699       match(F, m_OneUse(m_Or(m_Specific(X), m_APInt(C)))) && *NotC == ~(*C)) {
700     Constant *Zero = ConstantInt::getNullValue(Ty);
701     Constant *OrC = ConstantInt::get(Ty, *C);
702     Value *NewSel = Builder.CreateSelect(Cond, Zero, OrC, "masksel", &Sel);
703     return BinaryOperator::CreateOr(T, NewSel);
704   }
705 
706   // Cond ? (X | C) : (X & ~C) --> (X & ~C) | (Cond ? C : 0)
707   if (match(F, m_And(m_Value(X), m_APInt(NotC))) &&
708       match(T, m_OneUse(m_Or(m_Specific(X), m_APInt(C)))) && *NotC == ~(*C)) {
709     Constant *Zero = ConstantInt::getNullValue(Ty);
710     Constant *OrC = ConstantInt::get(Ty, *C);
711     Value *NewSel = Builder.CreateSelect(Cond, OrC, Zero, "masksel", &Sel);
712     return BinaryOperator::CreateOr(F, NewSel);
713   }
714 
715   return nullptr;
716 }
717 
718 /// Transform patterns such as (a > b) ? a - b : 0 into usub.sat(a, b).
719 /// There are 8 commuted/swapped variants of this pattern.
720 /// TODO: Also support a - UMIN(a,b) patterns.
721 static Value *canonicalizeSaturatedSubtract(const ICmpInst *ICI,
722                                             const Value *TrueVal,
723                                             const Value *FalseVal,
724                                             InstCombiner::BuilderTy &Builder) {
725   ICmpInst::Predicate Pred = ICI->getPredicate();
726   if (!ICmpInst::isUnsigned(Pred))
727     return nullptr;
728 
729   // (b > a) ? 0 : a - b -> (b <= a) ? a - b : 0
730   if (match(TrueVal, m_Zero())) {
731     Pred = ICmpInst::getInversePredicate(Pred);
732     std::swap(TrueVal, FalseVal);
733   }
734   if (!match(FalseVal, m_Zero()))
735     return nullptr;
736 
737   Value *A = ICI->getOperand(0);
738   Value *B = ICI->getOperand(1);
739   if (Pred == ICmpInst::ICMP_ULE || Pred == ICmpInst::ICMP_ULT) {
740     // (b < a) ? a - b : 0 -> (a > b) ? a - b : 0
741     std::swap(A, B);
742     Pred = ICmpInst::getSwappedPredicate(Pred);
743   }
744 
745   assert((Pred == ICmpInst::ICMP_UGE || Pred == ICmpInst::ICMP_UGT) &&
746          "Unexpected isUnsigned predicate!");
747 
748   // Ensure the sub is of the form:
749   //  (a > b) ? a - b : 0 -> usub.sat(a, b)
750   //  (a > b) ? b - a : 0 -> -usub.sat(a, b)
751   // Checking for both a-b and a+(-b) as a constant.
752   bool IsNegative = false;
753   const APInt *C;
754   if (match(TrueVal, m_Sub(m_Specific(B), m_Specific(A))) ||
755       (match(A, m_APInt(C)) &&
756        match(TrueVal, m_Add(m_Specific(B), m_SpecificInt(-*C)))))
757     IsNegative = true;
758   else if (!match(TrueVal, m_Sub(m_Specific(A), m_Specific(B))) &&
759            !(match(B, m_APInt(C)) &&
760              match(TrueVal, m_Add(m_Specific(A), m_SpecificInt(-*C)))))
761     return nullptr;
762 
763   // If we are adding a negate and the sub and icmp are used anywhere else, we
764   // would end up with more instructions.
765   if (IsNegative && !TrueVal->hasOneUse() && !ICI->hasOneUse())
766     return nullptr;
767 
768   // (a > b) ? a - b : 0 -> usub.sat(a, b)
769   // (a > b) ? b - a : 0 -> -usub.sat(a, b)
770   Value *Result = Builder.CreateBinaryIntrinsic(Intrinsic::usub_sat, A, B);
771   if (IsNegative)
772     Result = Builder.CreateNeg(Result);
773   return Result;
774 }
775 
776 static Value *canonicalizeSaturatedAdd(ICmpInst *Cmp, Value *TVal, Value *FVal,
777                                        InstCombiner::BuilderTy &Builder) {
778   if (!Cmp->hasOneUse())
779     return nullptr;
780 
781   // Match unsigned saturated add with constant.
782   Value *Cmp0 = Cmp->getOperand(0);
783   Value *Cmp1 = Cmp->getOperand(1);
784   ICmpInst::Predicate Pred = Cmp->getPredicate();
785   Value *X;
786   const APInt *C, *CmpC;
787   if (Pred == ICmpInst::ICMP_ULT &&
788       match(TVal, m_Add(m_Value(X), m_APInt(C))) && X == Cmp0 &&
789       match(FVal, m_AllOnes()) && match(Cmp1, m_APInt(CmpC)) && *CmpC == ~*C) {
790     // (X u< ~C) ? (X + C) : -1 --> uadd.sat(X, C)
791     return Builder.CreateBinaryIntrinsic(
792         Intrinsic::uadd_sat, X, ConstantInt::get(X->getType(), *C));
793   }
794 
795   // Match unsigned saturated add of 2 variables with an unnecessary 'not'.
796   // There are 8 commuted variants.
797   // Canonicalize -1 (saturated result) to true value of the select.
798   if (match(FVal, m_AllOnes())) {
799     std::swap(TVal, FVal);
800     Pred = CmpInst::getInversePredicate(Pred);
801   }
802   if (!match(TVal, m_AllOnes()))
803     return nullptr;
804 
805   // Canonicalize predicate to less-than or less-or-equal-than.
806   if (Pred == ICmpInst::ICMP_UGT || Pred == ICmpInst::ICMP_UGE) {
807     std::swap(Cmp0, Cmp1);
808     Pred = CmpInst::getSwappedPredicate(Pred);
809   }
810   if (Pred != ICmpInst::ICMP_ULT && Pred != ICmpInst::ICMP_ULE)
811     return nullptr;
812 
813   // Match unsigned saturated add of 2 variables with an unnecessary 'not'.
814   // Strictness of the comparison is irrelevant.
815   Value *Y;
816   if (match(Cmp0, m_Not(m_Value(X))) &&
817       match(FVal, m_c_Add(m_Specific(X), m_Value(Y))) && Y == Cmp1) {
818     // (~X u< Y) ? -1 : (X + Y) --> uadd.sat(X, Y)
819     // (~X u< Y) ? -1 : (Y + X) --> uadd.sat(X, Y)
820     return Builder.CreateBinaryIntrinsic(Intrinsic::uadd_sat, X, Y);
821   }
822   // The 'not' op may be included in the sum but not the compare.
823   // Strictness of the comparison is irrelevant.
824   X = Cmp0;
825   Y = Cmp1;
826   if (match(FVal, m_c_Add(m_Not(m_Specific(X)), m_Specific(Y)))) {
827     // (X u< Y) ? -1 : (~X + Y) --> uadd.sat(~X, Y)
828     // (X u< Y) ? -1 : (Y + ~X) --> uadd.sat(Y, ~X)
829     BinaryOperator *BO = cast<BinaryOperator>(FVal);
830     return Builder.CreateBinaryIntrinsic(
831         Intrinsic::uadd_sat, BO->getOperand(0), BO->getOperand(1));
832   }
833   // The overflow may be detected via the add wrapping round.
834   // This is only valid for strict comparison!
835   if (Pred == ICmpInst::ICMP_ULT &&
836       match(Cmp0, m_c_Add(m_Specific(Cmp1), m_Value(Y))) &&
837       match(FVal, m_c_Add(m_Specific(Cmp1), m_Specific(Y)))) {
838     // ((X + Y) u< X) ? -1 : (X + Y) --> uadd.sat(X, Y)
839     // ((X + Y) u< Y) ? -1 : (X + Y) --> uadd.sat(X, Y)
840     return Builder.CreateBinaryIntrinsic(Intrinsic::uadd_sat, Cmp1, Y);
841   }
842 
843   return nullptr;
844 }
845 
846 /// Fold the following code sequence:
847 /// \code
848 ///   int a = ctlz(x & -x);
849 //    x ? 31 - a : a;
850 /// \code
851 ///
852 /// into:
853 ///   cttz(x)
854 static Instruction *foldSelectCtlzToCttz(ICmpInst *ICI, Value *TrueVal,
855                                          Value *FalseVal,
856                                          InstCombiner::BuilderTy &Builder) {
857   unsigned BitWidth = TrueVal->getType()->getScalarSizeInBits();
858   if (!ICI->isEquality() || !match(ICI->getOperand(1), m_Zero()))
859     return nullptr;
860 
861   if (ICI->getPredicate() == ICmpInst::ICMP_NE)
862     std::swap(TrueVal, FalseVal);
863 
864   if (!match(FalseVal,
865              m_Xor(m_Deferred(TrueVal), m_SpecificInt(BitWidth - 1))))
866     return nullptr;
867 
868   if (!match(TrueVal, m_Intrinsic<Intrinsic::ctlz>()))
869     return nullptr;
870 
871   Value *X = ICI->getOperand(0);
872   auto *II = cast<IntrinsicInst>(TrueVal);
873   if (!match(II->getOperand(0), m_c_And(m_Specific(X), m_Neg(m_Specific(X)))))
874     return nullptr;
875 
876   Function *F = Intrinsic::getDeclaration(II->getModule(), Intrinsic::cttz,
877                                           II->getType());
878   return CallInst::Create(F, {X, II->getArgOperand(1)});
879 }
880 
881 /// Attempt to fold a cttz/ctlz followed by a icmp plus select into a single
882 /// call to cttz/ctlz with flag 'is_zero_undef' cleared.
883 ///
884 /// For example, we can fold the following code sequence:
885 /// \code
886 ///   %0 = tail call i32 @llvm.cttz.i32(i32 %x, i1 true)
887 ///   %1 = icmp ne i32 %x, 0
888 ///   %2 = select i1 %1, i32 %0, i32 32
889 /// \code
890 ///
891 /// into:
892 ///   %0 = tail call i32 @llvm.cttz.i32(i32 %x, i1 false)
893 static Value *foldSelectCttzCtlz(ICmpInst *ICI, Value *TrueVal, Value *FalseVal,
894                                  InstCombiner::BuilderTy &Builder) {
895   ICmpInst::Predicate Pred = ICI->getPredicate();
896   Value *CmpLHS = ICI->getOperand(0);
897   Value *CmpRHS = ICI->getOperand(1);
898 
899   // Check if the condition value compares a value for equality against zero.
900   if (!ICI->isEquality() || !match(CmpRHS, m_Zero()))
901     return nullptr;
902 
903   Value *SelectArg = FalseVal;
904   Value *ValueOnZero = TrueVal;
905   if (Pred == ICmpInst::ICMP_NE)
906     std::swap(SelectArg, ValueOnZero);
907 
908   // Skip zero extend/truncate.
909   Value *Count = nullptr;
910   if (!match(SelectArg, m_ZExt(m_Value(Count))) &&
911       !match(SelectArg, m_Trunc(m_Value(Count))))
912     Count = SelectArg;
913 
914   // Check that 'Count' is a call to intrinsic cttz/ctlz. Also check that the
915   // input to the cttz/ctlz is used as LHS for the compare instruction.
916   if (!match(Count, m_Intrinsic<Intrinsic::cttz>(m_Specific(CmpLHS))) &&
917       !match(Count, m_Intrinsic<Intrinsic::ctlz>(m_Specific(CmpLHS))))
918     return nullptr;
919 
920   IntrinsicInst *II = cast<IntrinsicInst>(Count);
921 
922   // Check if the value propagated on zero is a constant number equal to the
923   // sizeof in bits of 'Count'.
924   unsigned SizeOfInBits = Count->getType()->getScalarSizeInBits();
925   if (match(ValueOnZero, m_SpecificInt(SizeOfInBits))) {
926     // Explicitly clear the 'undef_on_zero' flag. It's always valid to go from
927     // true to false on this flag, so we can replace it for all users.
928     II->setArgOperand(1, ConstantInt::getFalse(II->getContext()));
929     return SelectArg;
930   }
931 
932   // The ValueOnZero is not the bitwidth. But if the cttz/ctlz (and optional
933   // zext/trunc) have one use (ending at the select), the cttz/ctlz result will
934   // not be used if the input is zero. Relax to 'undef_on_zero' for that case.
935   if (II->hasOneUse() && SelectArg->hasOneUse() &&
936       !match(II->getArgOperand(1), m_One()))
937     II->setArgOperand(1, ConstantInt::getTrue(II->getContext()));
938 
939   return nullptr;
940 }
941 
942 /// Return true if we find and adjust an icmp+select pattern where the compare
943 /// is with a constant that can be incremented or decremented to match the
944 /// minimum or maximum idiom.
945 static bool adjustMinMax(SelectInst &Sel, ICmpInst &Cmp) {
946   ICmpInst::Predicate Pred = Cmp.getPredicate();
947   Value *CmpLHS = Cmp.getOperand(0);
948   Value *CmpRHS = Cmp.getOperand(1);
949   Value *TrueVal = Sel.getTrueValue();
950   Value *FalseVal = Sel.getFalseValue();
951 
952   // We may move or edit the compare, so make sure the select is the only user.
953   const APInt *CmpC;
954   if (!Cmp.hasOneUse() || !match(CmpRHS, m_APInt(CmpC)))
955     return false;
956 
957   // These transforms only work for selects of integers or vector selects of
958   // integer vectors.
959   Type *SelTy = Sel.getType();
960   auto *SelEltTy = dyn_cast<IntegerType>(SelTy->getScalarType());
961   if (!SelEltTy || SelTy->isVectorTy() != Cmp.getType()->isVectorTy())
962     return false;
963 
964   Constant *AdjustedRHS;
965   if (Pred == ICmpInst::ICMP_UGT || Pred == ICmpInst::ICMP_SGT)
966     AdjustedRHS = ConstantInt::get(CmpRHS->getType(), *CmpC + 1);
967   else if (Pred == ICmpInst::ICMP_ULT || Pred == ICmpInst::ICMP_SLT)
968     AdjustedRHS = ConstantInt::get(CmpRHS->getType(), *CmpC - 1);
969   else
970     return false;
971 
972   // X > C ? X : C+1  -->  X < C+1 ? C+1 : X
973   // X < C ? X : C-1  -->  X > C-1 ? C-1 : X
974   if ((CmpLHS == TrueVal && AdjustedRHS == FalseVal) ||
975       (CmpLHS == FalseVal && AdjustedRHS == TrueVal)) {
976     ; // Nothing to do here. Values match without any sign/zero extension.
977   }
978   // Types do not match. Instead of calculating this with mixed types, promote
979   // all to the larger type. This enables scalar evolution to analyze this
980   // expression.
981   else if (CmpRHS->getType()->getScalarSizeInBits() < SelEltTy->getBitWidth()) {
982     Constant *SextRHS = ConstantExpr::getSExt(AdjustedRHS, SelTy);
983 
984     // X = sext x; x >s c ? X : C+1 --> X = sext x; X <s C+1 ? C+1 : X
985     // X = sext x; x <s c ? X : C-1 --> X = sext x; X >s C-1 ? C-1 : X
986     // X = sext x; x >u c ? X : C+1 --> X = sext x; X <u C+1 ? C+1 : X
987     // X = sext x; x <u c ? X : C-1 --> X = sext x; X >u C-1 ? C-1 : X
988     if (match(TrueVal, m_SExt(m_Specific(CmpLHS))) && SextRHS == FalseVal) {
989       CmpLHS = TrueVal;
990       AdjustedRHS = SextRHS;
991     } else if (match(FalseVal, m_SExt(m_Specific(CmpLHS))) &&
992                SextRHS == TrueVal) {
993       CmpLHS = FalseVal;
994       AdjustedRHS = SextRHS;
995     } else if (Cmp.isUnsigned()) {
996       Constant *ZextRHS = ConstantExpr::getZExt(AdjustedRHS, SelTy);
997       // X = zext x; x >u c ? X : C+1 --> X = zext x; X <u C+1 ? C+1 : X
998       // X = zext x; x <u c ? X : C-1 --> X = zext x; X >u C-1 ? C-1 : X
999       // zext + signed compare cannot be changed:
1000       //    0xff <s 0x00, but 0x00ff >s 0x0000
1001       if (match(TrueVal, m_ZExt(m_Specific(CmpLHS))) && ZextRHS == FalseVal) {
1002         CmpLHS = TrueVal;
1003         AdjustedRHS = ZextRHS;
1004       } else if (match(FalseVal, m_ZExt(m_Specific(CmpLHS))) &&
1005                  ZextRHS == TrueVal) {
1006         CmpLHS = FalseVal;
1007         AdjustedRHS = ZextRHS;
1008       } else {
1009         return false;
1010       }
1011     } else {
1012       return false;
1013     }
1014   } else {
1015     return false;
1016   }
1017 
1018   Pred = ICmpInst::getSwappedPredicate(Pred);
1019   CmpRHS = AdjustedRHS;
1020   std::swap(FalseVal, TrueVal);
1021   Cmp.setPredicate(Pred);
1022   Cmp.setOperand(0, CmpLHS);
1023   Cmp.setOperand(1, CmpRHS);
1024   Sel.setOperand(1, TrueVal);
1025   Sel.setOperand(2, FalseVal);
1026   Sel.swapProfMetadata();
1027 
1028   // Move the compare instruction right before the select instruction. Otherwise
1029   // the sext/zext value may be defined after the compare instruction uses it.
1030   Cmp.moveBefore(&Sel);
1031 
1032   return true;
1033 }
1034 
1035 /// If this is an integer min/max (icmp + select) with a constant operand,
1036 /// create the canonical icmp for the min/max operation and canonicalize the
1037 /// constant to the 'false' operand of the select:
1038 /// select (icmp Pred X, C1), C2, X --> select (icmp Pred' X, C2), X, C2
1039 /// Note: if C1 != C2, this will change the icmp constant to the existing
1040 /// constant operand of the select.
1041 static Instruction *canonicalizeMinMaxWithConstant(SelectInst &Sel,
1042                                                    ICmpInst &Cmp,
1043                                                    InstCombinerImpl &IC) {
1044   if (!Cmp.hasOneUse() || !isa<Constant>(Cmp.getOperand(1)))
1045     return nullptr;
1046 
1047   // Canonicalize the compare predicate based on whether we have min or max.
1048   Value *LHS, *RHS;
1049   SelectPatternResult SPR = matchSelectPattern(&Sel, LHS, RHS);
1050   if (!SelectPatternResult::isMinOrMax(SPR.Flavor))
1051     return nullptr;
1052 
1053   // Is this already canonical?
1054   ICmpInst::Predicate CanonicalPred = getMinMaxPred(SPR.Flavor);
1055   if (Cmp.getOperand(0) == LHS && Cmp.getOperand(1) == RHS &&
1056       Cmp.getPredicate() == CanonicalPred)
1057     return nullptr;
1058 
1059   // Bail out on unsimplified X-0 operand (due to some worklist management bug),
1060   // as this may cause an infinite combine loop. Let the sub be folded first.
1061   if (match(LHS, m_Sub(m_Value(), m_Zero())) ||
1062       match(RHS, m_Sub(m_Value(), m_Zero())))
1063     return nullptr;
1064 
1065   // Create the canonical compare and plug it into the select.
1066   IC.replaceOperand(Sel, 0, IC.Builder.CreateICmp(CanonicalPred, LHS, RHS));
1067 
1068   // If the select operands did not change, we're done.
1069   if (Sel.getTrueValue() == LHS && Sel.getFalseValue() == RHS)
1070     return &Sel;
1071 
1072   // If we are swapping the select operands, swap the metadata too.
1073   assert(Sel.getTrueValue() == RHS && Sel.getFalseValue() == LHS &&
1074          "Unexpected results from matchSelectPattern");
1075   Sel.swapValues();
1076   Sel.swapProfMetadata();
1077   return &Sel;
1078 }
1079 
1080 static Instruction *canonicalizeAbsNabs(SelectInst &Sel, ICmpInst &Cmp,
1081                                         InstCombinerImpl &IC) {
1082   if (!Cmp.hasOneUse() || !isa<Constant>(Cmp.getOperand(1)))
1083     return nullptr;
1084 
1085   Value *LHS, *RHS;
1086   SelectPatternFlavor SPF = matchSelectPattern(&Sel, LHS, RHS).Flavor;
1087   if (SPF != SelectPatternFlavor::SPF_ABS &&
1088       SPF != SelectPatternFlavor::SPF_NABS)
1089     return nullptr;
1090 
1091   // Note that NSW flag can only be propagated for normal, non-negated abs!
1092   bool IntMinIsPoison = SPF == SelectPatternFlavor::SPF_ABS &&
1093                         match(RHS, m_NSWNeg(m_Specific(LHS)));
1094   Constant *IntMinIsPoisonC =
1095       ConstantInt::get(Type::getInt1Ty(Sel.getContext()), IntMinIsPoison);
1096   Instruction *Abs =
1097       IC.Builder.CreateBinaryIntrinsic(Intrinsic::abs, LHS, IntMinIsPoisonC);
1098 
1099   if (SPF == SelectPatternFlavor::SPF_NABS)
1100     return BinaryOperator::CreateNeg(Abs); // Always without NSW flag!
1101 
1102   return IC.replaceInstUsesWith(Sel, Abs);
1103 }
1104 
1105 /// If we have a select with an equality comparison, then we know the value in
1106 /// one of the arms of the select. See if substituting this value into an arm
1107 /// and simplifying the result yields the same value as the other arm.
1108 ///
1109 /// To make this transform safe, we must drop poison-generating flags
1110 /// (nsw, etc) if we simplified to a binop because the select may be guarding
1111 /// that poison from propagating. If the existing binop already had no
1112 /// poison-generating flags, then this transform can be done by instsimplify.
1113 ///
1114 /// Consider:
1115 ///   %cmp = icmp eq i32 %x, 2147483647
1116 ///   %add = add nsw i32 %x, 1
1117 ///   %sel = select i1 %cmp, i32 -2147483648, i32 %add
1118 ///
1119 /// We can't replace %sel with %add unless we strip away the flags.
1120 /// TODO: Wrapping flags could be preserved in some cases with better analysis.
1121 Instruction *InstCombinerImpl::foldSelectValueEquivalence(SelectInst &Sel,
1122                                                           ICmpInst &Cmp) {
1123   // Value equivalence substitution requires an all-or-nothing replacement.
1124   // It does not make sense for a vector compare where each lane is chosen
1125   // independently.
1126   if (!Cmp.isEquality() || Cmp.getType()->isVectorTy())
1127     return nullptr;
1128 
1129   // Canonicalize the pattern to ICMP_EQ by swapping the select operands.
1130   Value *TrueVal = Sel.getTrueValue(), *FalseVal = Sel.getFalseValue();
1131   bool Swapped = false;
1132   if (Cmp.getPredicate() == ICmpInst::ICMP_NE) {
1133     std::swap(TrueVal, FalseVal);
1134     Swapped = true;
1135   }
1136 
1137   // In X == Y ? f(X) : Z, try to evaluate f(Y) and replace the operand.
1138   // Make sure Y cannot be undef though, as we might pick different values for
1139   // undef in the icmp and in f(Y). Additionally, take care to avoid replacing
1140   // X == Y ? X : Z with X == Y ? Y : Z, as that would lead to an infinite
1141   // replacement cycle.
1142   Value *CmpLHS = Cmp.getOperand(0), *CmpRHS = Cmp.getOperand(1);
1143   if (TrueVal != CmpLHS &&
1144       isGuaranteedNotToBeUndefOrPoison(CmpRHS, SQ.AC, &Sel, &DT)) {
1145     if (Value *V = simplifyWithOpReplaced(TrueVal, CmpLHS, CmpRHS, SQ,
1146                                           /* AllowRefinement */ true))
1147       return replaceOperand(Sel, Swapped ? 2 : 1, V);
1148 
1149     // Even if TrueVal does not simplify, we can directly replace a use of
1150     // CmpLHS with CmpRHS, as long as the instruction is not used anywhere
1151     // else and is safe to speculatively execute (we may end up executing it
1152     // with different operands, which should not cause side-effects or trigger
1153     // undefined behavior). Only do this if CmpRHS is a constant, as
1154     // profitability is not clear for other cases.
1155     // FIXME: The replacement could be performed recursively.
1156     if (match(CmpRHS, m_ImmConstant()) && !match(CmpLHS, m_ImmConstant()))
1157       if (auto *I = dyn_cast<Instruction>(TrueVal))
1158         if (I->hasOneUse() && isSafeToSpeculativelyExecute(I))
1159           for (Use &U : I->operands())
1160             if (U == CmpLHS) {
1161               replaceUse(U, CmpRHS);
1162               return &Sel;
1163             }
1164   }
1165   if (TrueVal != CmpRHS &&
1166       isGuaranteedNotToBeUndefOrPoison(CmpLHS, SQ.AC, &Sel, &DT))
1167     if (Value *V = simplifyWithOpReplaced(TrueVal, CmpRHS, CmpLHS, SQ,
1168                                           /* AllowRefinement */ true))
1169       return replaceOperand(Sel, Swapped ? 2 : 1, V);
1170 
1171   auto *FalseInst = dyn_cast<Instruction>(FalseVal);
1172   if (!FalseInst)
1173     return nullptr;
1174 
1175   // InstSimplify already performed this fold if it was possible subject to
1176   // current poison-generating flags. Try the transform again with
1177   // poison-generating flags temporarily dropped.
1178   bool WasNUW = false, WasNSW = false, WasExact = false, WasInBounds = false;
1179   if (auto *OBO = dyn_cast<OverflowingBinaryOperator>(FalseVal)) {
1180     WasNUW = OBO->hasNoUnsignedWrap();
1181     WasNSW = OBO->hasNoSignedWrap();
1182     FalseInst->setHasNoUnsignedWrap(false);
1183     FalseInst->setHasNoSignedWrap(false);
1184   }
1185   if (auto *PEO = dyn_cast<PossiblyExactOperator>(FalseVal)) {
1186     WasExact = PEO->isExact();
1187     FalseInst->setIsExact(false);
1188   }
1189   if (auto *GEP = dyn_cast<GetElementPtrInst>(FalseVal)) {
1190     WasInBounds = GEP->isInBounds();
1191     GEP->setIsInBounds(false);
1192   }
1193 
1194   // Try each equivalence substitution possibility.
1195   // We have an 'EQ' comparison, so the select's false value will propagate.
1196   // Example:
1197   // (X == 42) ? 43 : (X + 1) --> (X == 42) ? (X + 1) : (X + 1) --> X + 1
1198   if (simplifyWithOpReplaced(FalseVal, CmpLHS, CmpRHS, SQ,
1199                              /* AllowRefinement */ false) == TrueVal ||
1200       simplifyWithOpReplaced(FalseVal, CmpRHS, CmpLHS, SQ,
1201                              /* AllowRefinement */ false) == TrueVal) {
1202     return replaceInstUsesWith(Sel, FalseVal);
1203   }
1204 
1205   // Restore poison-generating flags if the transform did not apply.
1206   if (WasNUW)
1207     FalseInst->setHasNoUnsignedWrap();
1208   if (WasNSW)
1209     FalseInst->setHasNoSignedWrap();
1210   if (WasExact)
1211     FalseInst->setIsExact();
1212   if (WasInBounds)
1213     cast<GetElementPtrInst>(FalseInst)->setIsInBounds();
1214 
1215   return nullptr;
1216 }
1217 
1218 // See if this is a pattern like:
1219 //   %old_cmp1 = icmp slt i32 %x, C2
1220 //   %old_replacement = select i1 %old_cmp1, i32 %target_low, i32 %target_high
1221 //   %old_x_offseted = add i32 %x, C1
1222 //   %old_cmp0 = icmp ult i32 %old_x_offseted, C0
1223 //   %r = select i1 %old_cmp0, i32 %x, i32 %old_replacement
1224 // This can be rewritten as more canonical pattern:
1225 //   %new_cmp1 = icmp slt i32 %x, -C1
1226 //   %new_cmp2 = icmp sge i32 %x, C0-C1
1227 //   %new_clamped_low = select i1 %new_cmp1, i32 %target_low, i32 %x
1228 //   %r = select i1 %new_cmp2, i32 %target_high, i32 %new_clamped_low
1229 // Iff -C1 s<= C2 s<= C0-C1
1230 // Also ULT predicate can also be UGT iff C0 != -1 (+invert result)
1231 //      SLT predicate can also be SGT iff C2 != INT_MAX (+invert res.)
1232 static Instruction *canonicalizeClampLike(SelectInst &Sel0, ICmpInst &Cmp0,
1233                                           InstCombiner::BuilderTy &Builder) {
1234   Value *X = Sel0.getTrueValue();
1235   Value *Sel1 = Sel0.getFalseValue();
1236 
1237   // First match the condition of the outermost select.
1238   // Said condition must be one-use.
1239   if (!Cmp0.hasOneUse())
1240     return nullptr;
1241   Value *Cmp00 = Cmp0.getOperand(0);
1242   Constant *C0;
1243   if (!match(Cmp0.getOperand(1),
1244              m_CombineAnd(m_AnyIntegralConstant(), m_Constant(C0))))
1245     return nullptr;
1246   // Canonicalize Cmp0 into the form we expect.
1247   // FIXME: we shouldn't care about lanes that are 'undef' in the end?
1248   switch (Cmp0.getPredicate()) {
1249   case ICmpInst::Predicate::ICMP_ULT:
1250     break; // Great!
1251   case ICmpInst::Predicate::ICMP_ULE:
1252     // We'd have to increment C0 by one, and for that it must not have all-ones
1253     // element, but then it would have been canonicalized to 'ult' before
1254     // we get here. So we can't do anything useful with 'ule'.
1255     return nullptr;
1256   case ICmpInst::Predicate::ICMP_UGT:
1257     // We want to canonicalize it to 'ult', so we'll need to increment C0,
1258     // which again means it must not have any all-ones elements.
1259     if (!match(C0,
1260                m_SpecificInt_ICMP(ICmpInst::Predicate::ICMP_NE,
1261                                   APInt::getAllOnesValue(
1262                                       C0->getType()->getScalarSizeInBits()))))
1263       return nullptr; // Can't do, have all-ones element[s].
1264     C0 = InstCombiner::AddOne(C0);
1265     std::swap(X, Sel1);
1266     break;
1267   case ICmpInst::Predicate::ICMP_UGE:
1268     // The only way we'd get this predicate if this `icmp` has extra uses,
1269     // but then we won't be able to do this fold.
1270     return nullptr;
1271   default:
1272     return nullptr; // Unknown predicate.
1273   }
1274 
1275   // Now that we've canonicalized the ICmp, we know the X we expect;
1276   // the select in other hand should be one-use.
1277   if (!Sel1->hasOneUse())
1278     return nullptr;
1279 
1280   // We now can finish matching the condition of the outermost select:
1281   // it should either be the X itself, or an addition of some constant to X.
1282   Constant *C1;
1283   if (Cmp00 == X)
1284     C1 = ConstantInt::getNullValue(Sel0.getType());
1285   else if (!match(Cmp00,
1286                   m_Add(m_Specific(X),
1287                         m_CombineAnd(m_AnyIntegralConstant(), m_Constant(C1)))))
1288     return nullptr;
1289 
1290   Value *Cmp1;
1291   ICmpInst::Predicate Pred1;
1292   Constant *C2;
1293   Value *ReplacementLow, *ReplacementHigh;
1294   if (!match(Sel1, m_Select(m_Value(Cmp1), m_Value(ReplacementLow),
1295                             m_Value(ReplacementHigh))) ||
1296       !match(Cmp1,
1297              m_ICmp(Pred1, m_Specific(X),
1298                     m_CombineAnd(m_AnyIntegralConstant(), m_Constant(C2)))))
1299     return nullptr;
1300 
1301   if (!Cmp1->hasOneUse() && (Cmp00 == X || !Cmp00->hasOneUse()))
1302     return nullptr; // Not enough one-use instructions for the fold.
1303   // FIXME: this restriction could be relaxed if Cmp1 can be reused as one of
1304   //        two comparisons we'll need to build.
1305 
1306   // Canonicalize Cmp1 into the form we expect.
1307   // FIXME: we shouldn't care about lanes that are 'undef' in the end?
1308   switch (Pred1) {
1309   case ICmpInst::Predicate::ICMP_SLT:
1310     break;
1311   case ICmpInst::Predicate::ICMP_SLE:
1312     // We'd have to increment C2 by one, and for that it must not have signed
1313     // max element, but then it would have been canonicalized to 'slt' before
1314     // we get here. So we can't do anything useful with 'sle'.
1315     return nullptr;
1316   case ICmpInst::Predicate::ICMP_SGT:
1317     // We want to canonicalize it to 'slt', so we'll need to increment C2,
1318     // which again means it must not have any signed max elements.
1319     if (!match(C2,
1320                m_SpecificInt_ICMP(ICmpInst::Predicate::ICMP_NE,
1321                                   APInt::getSignedMaxValue(
1322                                       C2->getType()->getScalarSizeInBits()))))
1323       return nullptr; // Can't do, have signed max element[s].
1324     C2 = InstCombiner::AddOne(C2);
1325     LLVM_FALLTHROUGH;
1326   case ICmpInst::Predicate::ICMP_SGE:
1327     // Also non-canonical, but here we don't need to change C2,
1328     // so we don't have any restrictions on C2, so we can just handle it.
1329     std::swap(ReplacementLow, ReplacementHigh);
1330     break;
1331   default:
1332     return nullptr; // Unknown predicate.
1333   }
1334 
1335   // The thresholds of this clamp-like pattern.
1336   auto *ThresholdLowIncl = ConstantExpr::getNeg(C1);
1337   auto *ThresholdHighExcl = ConstantExpr::getSub(C0, C1);
1338 
1339   // The fold has a precondition 1: C2 s>= ThresholdLow
1340   auto *Precond1 = ConstantExpr::getICmp(ICmpInst::Predicate::ICMP_SGE, C2,
1341                                          ThresholdLowIncl);
1342   if (!match(Precond1, m_One()))
1343     return nullptr;
1344   // The fold has a precondition 2: C2 s<= ThresholdHigh
1345   auto *Precond2 = ConstantExpr::getICmp(ICmpInst::Predicate::ICMP_SLE, C2,
1346                                          ThresholdHighExcl);
1347   if (!match(Precond2, m_One()))
1348     return nullptr;
1349 
1350   // All good, finally emit the new pattern.
1351   Value *ShouldReplaceLow = Builder.CreateICmpSLT(X, ThresholdLowIncl);
1352   Value *ShouldReplaceHigh = Builder.CreateICmpSGE(X, ThresholdHighExcl);
1353   Value *MaybeReplacedLow =
1354       Builder.CreateSelect(ShouldReplaceLow, ReplacementLow, X);
1355   Instruction *MaybeReplacedHigh =
1356       SelectInst::Create(ShouldReplaceHigh, ReplacementHigh, MaybeReplacedLow);
1357 
1358   return MaybeReplacedHigh;
1359 }
1360 
1361 // If we have
1362 //  %cmp = icmp [canonical predicate] i32 %x, C0
1363 //  %r = select i1 %cmp, i32 %y, i32 C1
1364 // Where C0 != C1 and %x may be different from %y, see if the constant that we
1365 // will have if we flip the strictness of the predicate (i.e. without changing
1366 // the result) is identical to the C1 in select. If it matches we can change
1367 // original comparison to one with swapped predicate, reuse the constant,
1368 // and swap the hands of select.
1369 static Instruction *
1370 tryToReuseConstantFromSelectInComparison(SelectInst &Sel, ICmpInst &Cmp,
1371                                          InstCombinerImpl &IC) {
1372   ICmpInst::Predicate Pred;
1373   Value *X;
1374   Constant *C0;
1375   if (!match(&Cmp, m_OneUse(m_ICmp(
1376                        Pred, m_Value(X),
1377                        m_CombineAnd(m_AnyIntegralConstant(), m_Constant(C0))))))
1378     return nullptr;
1379 
1380   // If comparison predicate is non-relational, we won't be able to do anything.
1381   if (ICmpInst::isEquality(Pred))
1382     return nullptr;
1383 
1384   // If comparison predicate is non-canonical, then we certainly won't be able
1385   // to make it canonical; canonicalizeCmpWithConstant() already tried.
1386   if (!InstCombiner::isCanonicalPredicate(Pred))
1387     return nullptr;
1388 
1389   // If the [input] type of comparison and select type are different, lets abort
1390   // for now. We could try to compare constants with trunc/[zs]ext though.
1391   if (C0->getType() != Sel.getType())
1392     return nullptr;
1393 
1394   // FIXME: are there any magic icmp predicate+constant pairs we must not touch?
1395 
1396   Value *SelVal0, *SelVal1; // We do not care which one is from where.
1397   match(&Sel, m_Select(m_Value(), m_Value(SelVal0), m_Value(SelVal1)));
1398   // At least one of these values we are selecting between must be a constant
1399   // else we'll never succeed.
1400   if (!match(SelVal0, m_AnyIntegralConstant()) &&
1401       !match(SelVal1, m_AnyIntegralConstant()))
1402     return nullptr;
1403 
1404   // Does this constant C match any of the `select` values?
1405   auto MatchesSelectValue = [SelVal0, SelVal1](Constant *C) {
1406     return C->isElementWiseEqual(SelVal0) || C->isElementWiseEqual(SelVal1);
1407   };
1408 
1409   // If C0 *already* matches true/false value of select, we are done.
1410   if (MatchesSelectValue(C0))
1411     return nullptr;
1412 
1413   // Check the constant we'd have with flipped-strictness predicate.
1414   auto FlippedStrictness =
1415       InstCombiner::getFlippedStrictnessPredicateAndConstant(Pred, C0);
1416   if (!FlippedStrictness)
1417     return nullptr;
1418 
1419   // If said constant doesn't match either, then there is no hope,
1420   if (!MatchesSelectValue(FlippedStrictness->second))
1421     return nullptr;
1422 
1423   // It matched! Lets insert the new comparison just before select.
1424   InstCombiner::BuilderTy::InsertPointGuard Guard(IC.Builder);
1425   IC.Builder.SetInsertPoint(&Sel);
1426 
1427   Pred = ICmpInst::getSwappedPredicate(Pred); // Yes, swapped.
1428   Value *NewCmp = IC.Builder.CreateICmp(Pred, X, FlippedStrictness->second,
1429                                         Cmp.getName() + ".inv");
1430   IC.replaceOperand(Sel, 0, NewCmp);
1431   Sel.swapValues();
1432   Sel.swapProfMetadata();
1433 
1434   return &Sel;
1435 }
1436 
1437 /// Visit a SelectInst that has an ICmpInst as its first operand.
1438 Instruction *InstCombinerImpl::foldSelectInstWithICmp(SelectInst &SI,
1439                                                       ICmpInst *ICI) {
1440   if (Instruction *NewSel = foldSelectValueEquivalence(SI, *ICI))
1441     return NewSel;
1442 
1443   if (Instruction *NewSel = canonicalizeMinMaxWithConstant(SI, *ICI, *this))
1444     return NewSel;
1445 
1446   if (Instruction *NewAbs = canonicalizeAbsNabs(SI, *ICI, *this))
1447     return NewAbs;
1448 
1449   if (Instruction *NewAbs = canonicalizeClampLike(SI, *ICI, Builder))
1450     return NewAbs;
1451 
1452   if (Instruction *NewSel =
1453           tryToReuseConstantFromSelectInComparison(SI, *ICI, *this))
1454     return NewSel;
1455 
1456   bool Changed = adjustMinMax(SI, *ICI);
1457 
1458   if (Value *V = foldSelectICmpAnd(SI, ICI, Builder))
1459     return replaceInstUsesWith(SI, V);
1460 
1461   // NOTE: if we wanted to, this is where to detect integer MIN/MAX
1462   Value *TrueVal = SI.getTrueValue();
1463   Value *FalseVal = SI.getFalseValue();
1464   ICmpInst::Predicate Pred = ICI->getPredicate();
1465   Value *CmpLHS = ICI->getOperand(0);
1466   Value *CmpRHS = ICI->getOperand(1);
1467   if (CmpRHS != CmpLHS && isa<Constant>(CmpRHS)) {
1468     if (CmpLHS == TrueVal && Pred == ICmpInst::ICMP_EQ) {
1469       // Transform (X == C) ? X : Y -> (X == C) ? C : Y
1470       SI.setOperand(1, CmpRHS);
1471       Changed = true;
1472     } else if (CmpLHS == FalseVal && Pred == ICmpInst::ICMP_NE) {
1473       // Transform (X != C) ? Y : X -> (X != C) ? Y : C
1474       SI.setOperand(2, CmpRHS);
1475       Changed = true;
1476     }
1477   }
1478 
1479   // FIXME: This code is nearly duplicated in InstSimplify. Using/refactoring
1480   // decomposeBitTestICmp() might help.
1481   {
1482     unsigned BitWidth =
1483         DL.getTypeSizeInBits(TrueVal->getType()->getScalarType());
1484     APInt MinSignedValue = APInt::getSignedMinValue(BitWidth);
1485     Value *X;
1486     const APInt *Y, *C;
1487     bool TrueWhenUnset;
1488     bool IsBitTest = false;
1489     if (ICmpInst::isEquality(Pred) &&
1490         match(CmpLHS, m_And(m_Value(X), m_Power2(Y))) &&
1491         match(CmpRHS, m_Zero())) {
1492       IsBitTest = true;
1493       TrueWhenUnset = Pred == ICmpInst::ICMP_EQ;
1494     } else if (Pred == ICmpInst::ICMP_SLT && match(CmpRHS, m_Zero())) {
1495       X = CmpLHS;
1496       Y = &MinSignedValue;
1497       IsBitTest = true;
1498       TrueWhenUnset = false;
1499     } else if (Pred == ICmpInst::ICMP_SGT && match(CmpRHS, m_AllOnes())) {
1500       X = CmpLHS;
1501       Y = &MinSignedValue;
1502       IsBitTest = true;
1503       TrueWhenUnset = true;
1504     }
1505     if (IsBitTest) {
1506       Value *V = nullptr;
1507       // (X & Y) == 0 ? X : X ^ Y  --> X & ~Y
1508       if (TrueWhenUnset && TrueVal == X &&
1509           match(FalseVal, m_Xor(m_Specific(X), m_APInt(C))) && *Y == *C)
1510         V = Builder.CreateAnd(X, ~(*Y));
1511       // (X & Y) != 0 ? X ^ Y : X  --> X & ~Y
1512       else if (!TrueWhenUnset && FalseVal == X &&
1513                match(TrueVal, m_Xor(m_Specific(X), m_APInt(C))) && *Y == *C)
1514         V = Builder.CreateAnd(X, ~(*Y));
1515       // (X & Y) == 0 ? X ^ Y : X  --> X | Y
1516       else if (TrueWhenUnset && FalseVal == X &&
1517                match(TrueVal, m_Xor(m_Specific(X), m_APInt(C))) && *Y == *C)
1518         V = Builder.CreateOr(X, *Y);
1519       // (X & Y) != 0 ? X : X ^ Y  --> X | Y
1520       else if (!TrueWhenUnset && TrueVal == X &&
1521                match(FalseVal, m_Xor(m_Specific(X), m_APInt(C))) && *Y == *C)
1522         V = Builder.CreateOr(X, *Y);
1523 
1524       if (V)
1525         return replaceInstUsesWith(SI, V);
1526     }
1527   }
1528 
1529   if (Instruction *V =
1530           foldSelectICmpAndAnd(SI.getType(), ICI, TrueVal, FalseVal, Builder))
1531     return V;
1532 
1533   if (Instruction *V = foldSelectCtlzToCttz(ICI, TrueVal, FalseVal, Builder))
1534     return V;
1535 
1536   if (Value *V = foldSelectICmpAndOr(ICI, TrueVal, FalseVal, Builder))
1537     return replaceInstUsesWith(SI, V);
1538 
1539   if (Value *V = foldSelectICmpLshrAshr(ICI, TrueVal, FalseVal, Builder))
1540     return replaceInstUsesWith(SI, V);
1541 
1542   if (Value *V = foldSelectCttzCtlz(ICI, TrueVal, FalseVal, Builder))
1543     return replaceInstUsesWith(SI, V);
1544 
1545   if (Value *V = canonicalizeSaturatedSubtract(ICI, TrueVal, FalseVal, Builder))
1546     return replaceInstUsesWith(SI, V);
1547 
1548   if (Value *V = canonicalizeSaturatedAdd(ICI, TrueVal, FalseVal, Builder))
1549     return replaceInstUsesWith(SI, V);
1550 
1551   return Changed ? &SI : nullptr;
1552 }
1553 
1554 /// SI is a select whose condition is a PHI node (but the two may be in
1555 /// different blocks). See if the true/false values (V) are live in all of the
1556 /// predecessor blocks of the PHI. For example, cases like this can't be mapped:
1557 ///
1558 ///   X = phi [ C1, BB1], [C2, BB2]
1559 ///   Y = add
1560 ///   Z = select X, Y, 0
1561 ///
1562 /// because Y is not live in BB1/BB2.
1563 static bool canSelectOperandBeMappingIntoPredBlock(const Value *V,
1564                                                    const SelectInst &SI) {
1565   // If the value is a non-instruction value like a constant or argument, it
1566   // can always be mapped.
1567   const Instruction *I = dyn_cast<Instruction>(V);
1568   if (!I) return true;
1569 
1570   // If V is a PHI node defined in the same block as the condition PHI, we can
1571   // map the arguments.
1572   const PHINode *CondPHI = cast<PHINode>(SI.getCondition());
1573 
1574   if (const PHINode *VP = dyn_cast<PHINode>(I))
1575     if (VP->getParent() == CondPHI->getParent())
1576       return true;
1577 
1578   // Otherwise, if the PHI and select are defined in the same block and if V is
1579   // defined in a different block, then we can transform it.
1580   if (SI.getParent() == CondPHI->getParent() &&
1581       I->getParent() != CondPHI->getParent())
1582     return true;
1583 
1584   // Otherwise we have a 'hard' case and we can't tell without doing more
1585   // detailed dominator based analysis, punt.
1586   return false;
1587 }
1588 
1589 /// We have an SPF (e.g. a min or max) of an SPF of the form:
1590 ///   SPF2(SPF1(A, B), C)
1591 Instruction *InstCombinerImpl::foldSPFofSPF(Instruction *Inner,
1592                                             SelectPatternFlavor SPF1, Value *A,
1593                                             Value *B, Instruction &Outer,
1594                                             SelectPatternFlavor SPF2,
1595                                             Value *C) {
1596   if (Outer.getType() != Inner->getType())
1597     return nullptr;
1598 
1599   if (C == A || C == B) {
1600     // MAX(MAX(A, B), B) -> MAX(A, B)
1601     // MIN(MIN(a, b), a) -> MIN(a, b)
1602     // TODO: This could be done in instsimplify.
1603     if (SPF1 == SPF2 && SelectPatternResult::isMinOrMax(SPF1))
1604       return replaceInstUsesWith(Outer, Inner);
1605 
1606     // MAX(MIN(a, b), a) -> a
1607     // MIN(MAX(a, b), a) -> a
1608     // TODO: This could be done in instsimplify.
1609     if ((SPF1 == SPF_SMIN && SPF2 == SPF_SMAX) ||
1610         (SPF1 == SPF_SMAX && SPF2 == SPF_SMIN) ||
1611         (SPF1 == SPF_UMIN && SPF2 == SPF_UMAX) ||
1612         (SPF1 == SPF_UMAX && SPF2 == SPF_UMIN))
1613       return replaceInstUsesWith(Outer, C);
1614   }
1615 
1616   if (SPF1 == SPF2) {
1617     const APInt *CB, *CC;
1618     if (match(B, m_APInt(CB)) && match(C, m_APInt(CC))) {
1619       // MIN(MIN(A, 23), 97) -> MIN(A, 23)
1620       // MAX(MAX(A, 97), 23) -> MAX(A, 97)
1621       // TODO: This could be done in instsimplify.
1622       if ((SPF1 == SPF_UMIN && CB->ule(*CC)) ||
1623           (SPF1 == SPF_SMIN && CB->sle(*CC)) ||
1624           (SPF1 == SPF_UMAX && CB->uge(*CC)) ||
1625           (SPF1 == SPF_SMAX && CB->sge(*CC)))
1626         return replaceInstUsesWith(Outer, Inner);
1627 
1628       // MIN(MIN(A, 97), 23) -> MIN(A, 23)
1629       // MAX(MAX(A, 23), 97) -> MAX(A, 97)
1630       if ((SPF1 == SPF_UMIN && CB->ugt(*CC)) ||
1631           (SPF1 == SPF_SMIN && CB->sgt(*CC)) ||
1632           (SPF1 == SPF_UMAX && CB->ult(*CC)) ||
1633           (SPF1 == SPF_SMAX && CB->slt(*CC))) {
1634         Outer.replaceUsesOfWith(Inner, A);
1635         return &Outer;
1636       }
1637     }
1638   }
1639 
1640   // max(max(A, B), min(A, B)) --> max(A, B)
1641   // min(min(A, B), max(A, B)) --> min(A, B)
1642   // TODO: This could be done in instsimplify.
1643   if (SPF1 == SPF2 &&
1644       ((SPF1 == SPF_UMIN && match(C, m_c_UMax(m_Specific(A), m_Specific(B)))) ||
1645        (SPF1 == SPF_SMIN && match(C, m_c_SMax(m_Specific(A), m_Specific(B)))) ||
1646        (SPF1 == SPF_UMAX && match(C, m_c_UMin(m_Specific(A), m_Specific(B)))) ||
1647        (SPF1 == SPF_SMAX && match(C, m_c_SMin(m_Specific(A), m_Specific(B))))))
1648     return replaceInstUsesWith(Outer, Inner);
1649 
1650   // ABS(ABS(X)) -> ABS(X)
1651   // NABS(NABS(X)) -> NABS(X)
1652   // TODO: This could be done in instsimplify.
1653   if (SPF1 == SPF2 && (SPF1 == SPF_ABS || SPF1 == SPF_NABS)) {
1654     return replaceInstUsesWith(Outer, Inner);
1655   }
1656 
1657   // ABS(NABS(X)) -> ABS(X)
1658   // NABS(ABS(X)) -> NABS(X)
1659   if ((SPF1 == SPF_ABS && SPF2 == SPF_NABS) ||
1660       (SPF1 == SPF_NABS && SPF2 == SPF_ABS)) {
1661     SelectInst *SI = cast<SelectInst>(Inner);
1662     Value *NewSI =
1663         Builder.CreateSelect(SI->getCondition(), SI->getFalseValue(),
1664                              SI->getTrueValue(), SI->getName(), SI);
1665     return replaceInstUsesWith(Outer, NewSI);
1666   }
1667 
1668   auto IsFreeOrProfitableToInvert =
1669       [&](Value *V, Value *&NotV, bool &ElidesXor) {
1670     if (match(V, m_Not(m_Value(NotV)))) {
1671       // If V has at most 2 uses then we can get rid of the xor operation
1672       // entirely.
1673       ElidesXor |= !V->hasNUsesOrMore(3);
1674       return true;
1675     }
1676 
1677     if (isFreeToInvert(V, !V->hasNUsesOrMore(3))) {
1678       NotV = nullptr;
1679       return true;
1680     }
1681 
1682     return false;
1683   };
1684 
1685   Value *NotA, *NotB, *NotC;
1686   bool ElidesXor = false;
1687 
1688   // MIN(MIN(~A, ~B), ~C) == ~MAX(MAX(A, B), C)
1689   // MIN(MAX(~A, ~B), ~C) == ~MAX(MIN(A, B), C)
1690   // MAX(MIN(~A, ~B), ~C) == ~MIN(MAX(A, B), C)
1691   // MAX(MAX(~A, ~B), ~C) == ~MIN(MIN(A, B), C)
1692   //
1693   // This transform is performance neutral if we can elide at least one xor from
1694   // the set of three operands, since we'll be tacking on an xor at the very
1695   // end.
1696   if (SelectPatternResult::isMinOrMax(SPF1) &&
1697       SelectPatternResult::isMinOrMax(SPF2) &&
1698       IsFreeOrProfitableToInvert(A, NotA, ElidesXor) &&
1699       IsFreeOrProfitableToInvert(B, NotB, ElidesXor) &&
1700       IsFreeOrProfitableToInvert(C, NotC, ElidesXor) && ElidesXor) {
1701     if (!NotA)
1702       NotA = Builder.CreateNot(A);
1703     if (!NotB)
1704       NotB = Builder.CreateNot(B);
1705     if (!NotC)
1706       NotC = Builder.CreateNot(C);
1707 
1708     Value *NewInner = createMinMax(Builder, getInverseMinMaxFlavor(SPF1), NotA,
1709                                    NotB);
1710     Value *NewOuter = Builder.CreateNot(
1711         createMinMax(Builder, getInverseMinMaxFlavor(SPF2), NewInner, NotC));
1712     return replaceInstUsesWith(Outer, NewOuter);
1713   }
1714 
1715   return nullptr;
1716 }
1717 
1718 /// Turn select C, (X + Y), (X - Y) --> (X + (select C, Y, (-Y))).
1719 /// This is even legal for FP.
1720 static Instruction *foldAddSubSelect(SelectInst &SI,
1721                                      InstCombiner::BuilderTy &Builder) {
1722   Value *CondVal = SI.getCondition();
1723   Value *TrueVal = SI.getTrueValue();
1724   Value *FalseVal = SI.getFalseValue();
1725   auto *TI = dyn_cast<Instruction>(TrueVal);
1726   auto *FI = dyn_cast<Instruction>(FalseVal);
1727   if (!TI || !FI || !TI->hasOneUse() || !FI->hasOneUse())
1728     return nullptr;
1729 
1730   Instruction *AddOp = nullptr, *SubOp = nullptr;
1731   if ((TI->getOpcode() == Instruction::Sub &&
1732        FI->getOpcode() == Instruction::Add) ||
1733       (TI->getOpcode() == Instruction::FSub &&
1734        FI->getOpcode() == Instruction::FAdd)) {
1735     AddOp = FI;
1736     SubOp = TI;
1737   } else if ((FI->getOpcode() == Instruction::Sub &&
1738               TI->getOpcode() == Instruction::Add) ||
1739              (FI->getOpcode() == Instruction::FSub &&
1740               TI->getOpcode() == Instruction::FAdd)) {
1741     AddOp = TI;
1742     SubOp = FI;
1743   }
1744 
1745   if (AddOp) {
1746     Value *OtherAddOp = nullptr;
1747     if (SubOp->getOperand(0) == AddOp->getOperand(0)) {
1748       OtherAddOp = AddOp->getOperand(1);
1749     } else if (SubOp->getOperand(0) == AddOp->getOperand(1)) {
1750       OtherAddOp = AddOp->getOperand(0);
1751     }
1752 
1753     if (OtherAddOp) {
1754       // So at this point we know we have (Y -> OtherAddOp):
1755       //        select C, (add X, Y), (sub X, Z)
1756       Value *NegVal; // Compute -Z
1757       if (SI.getType()->isFPOrFPVectorTy()) {
1758         NegVal = Builder.CreateFNeg(SubOp->getOperand(1));
1759         if (Instruction *NegInst = dyn_cast<Instruction>(NegVal)) {
1760           FastMathFlags Flags = AddOp->getFastMathFlags();
1761           Flags &= SubOp->getFastMathFlags();
1762           NegInst->setFastMathFlags(Flags);
1763         }
1764       } else {
1765         NegVal = Builder.CreateNeg(SubOp->getOperand(1));
1766       }
1767 
1768       Value *NewTrueOp = OtherAddOp;
1769       Value *NewFalseOp = NegVal;
1770       if (AddOp != TI)
1771         std::swap(NewTrueOp, NewFalseOp);
1772       Value *NewSel = Builder.CreateSelect(CondVal, NewTrueOp, NewFalseOp,
1773                                            SI.getName() + ".p", &SI);
1774 
1775       if (SI.getType()->isFPOrFPVectorTy()) {
1776         Instruction *RI =
1777             BinaryOperator::CreateFAdd(SubOp->getOperand(0), NewSel);
1778 
1779         FastMathFlags Flags = AddOp->getFastMathFlags();
1780         Flags &= SubOp->getFastMathFlags();
1781         RI->setFastMathFlags(Flags);
1782         return RI;
1783       } else
1784         return BinaryOperator::CreateAdd(SubOp->getOperand(0), NewSel);
1785     }
1786   }
1787   return nullptr;
1788 }
1789 
1790 /// Turn X + Y overflows ? -1 : X + Y -> uadd_sat X, Y
1791 /// And X - Y overflows ? 0 : X - Y -> usub_sat X, Y
1792 /// Along with a number of patterns similar to:
1793 /// X + Y overflows ? (X < 0 ? INTMIN : INTMAX) : X + Y --> sadd_sat X, Y
1794 /// X - Y overflows ? (X > 0 ? INTMAX : INTMIN) : X - Y --> ssub_sat X, Y
1795 static Instruction *
1796 foldOverflowingAddSubSelect(SelectInst &SI, InstCombiner::BuilderTy &Builder) {
1797   Value *CondVal = SI.getCondition();
1798   Value *TrueVal = SI.getTrueValue();
1799   Value *FalseVal = SI.getFalseValue();
1800 
1801   WithOverflowInst *II;
1802   if (!match(CondVal, m_ExtractValue<1>(m_WithOverflowInst(II))) ||
1803       !match(FalseVal, m_ExtractValue<0>(m_Specific(II))))
1804     return nullptr;
1805 
1806   Value *X = II->getLHS();
1807   Value *Y = II->getRHS();
1808 
1809   auto IsSignedSaturateLimit = [&](Value *Limit, bool IsAdd) {
1810     Type *Ty = Limit->getType();
1811 
1812     ICmpInst::Predicate Pred;
1813     Value *TrueVal, *FalseVal, *Op;
1814     const APInt *C;
1815     if (!match(Limit, m_Select(m_ICmp(Pred, m_Value(Op), m_APInt(C)),
1816                                m_Value(TrueVal), m_Value(FalseVal))))
1817       return false;
1818 
1819     auto IsZeroOrOne = [](const APInt &C) {
1820       return C.isNullValue() || C.isOneValue();
1821     };
1822     auto IsMinMax = [&](Value *Min, Value *Max) {
1823       APInt MinVal = APInt::getSignedMinValue(Ty->getScalarSizeInBits());
1824       APInt MaxVal = APInt::getSignedMaxValue(Ty->getScalarSizeInBits());
1825       return match(Min, m_SpecificInt(MinVal)) &&
1826              match(Max, m_SpecificInt(MaxVal));
1827     };
1828 
1829     if (Op != X && Op != Y)
1830       return false;
1831 
1832     if (IsAdd) {
1833       // X + Y overflows ? (X <s 0 ? INTMIN : INTMAX) : X + Y --> sadd_sat X, Y
1834       // X + Y overflows ? (X <s 1 ? INTMIN : INTMAX) : X + Y --> sadd_sat X, Y
1835       // X + Y overflows ? (Y <s 0 ? INTMIN : INTMAX) : X + Y --> sadd_sat X, Y
1836       // X + Y overflows ? (Y <s 1 ? INTMIN : INTMAX) : X + Y --> sadd_sat X, Y
1837       if (Pred == ICmpInst::ICMP_SLT && IsZeroOrOne(*C) &&
1838           IsMinMax(TrueVal, FalseVal))
1839         return true;
1840       // X + Y overflows ? (X >s 0 ? INTMAX : INTMIN) : X + Y --> sadd_sat X, Y
1841       // X + Y overflows ? (X >s -1 ? INTMAX : INTMIN) : X + Y --> sadd_sat X, Y
1842       // X + Y overflows ? (Y >s 0 ? INTMAX : INTMIN) : X + Y --> sadd_sat X, Y
1843       // X + Y overflows ? (Y >s -1 ? INTMAX : INTMIN) : X + Y --> sadd_sat X, Y
1844       if (Pred == ICmpInst::ICMP_SGT && IsZeroOrOne(*C + 1) &&
1845           IsMinMax(FalseVal, TrueVal))
1846         return true;
1847     } else {
1848       // X - Y overflows ? (X <s 0 ? INTMIN : INTMAX) : X - Y --> ssub_sat X, Y
1849       // X - Y overflows ? (X <s -1 ? INTMIN : INTMAX) : X - Y --> ssub_sat X, Y
1850       if (Op == X && Pred == ICmpInst::ICMP_SLT && IsZeroOrOne(*C + 1) &&
1851           IsMinMax(TrueVal, FalseVal))
1852         return true;
1853       // X - Y overflows ? (X >s -1 ? INTMAX : INTMIN) : X - Y --> ssub_sat X, Y
1854       // X - Y overflows ? (X >s -2 ? INTMAX : INTMIN) : X - Y --> ssub_sat X, Y
1855       if (Op == X && Pred == ICmpInst::ICMP_SGT && IsZeroOrOne(*C + 2) &&
1856           IsMinMax(FalseVal, TrueVal))
1857         return true;
1858       // X - Y overflows ? (Y <s 0 ? INTMAX : INTMIN) : X - Y --> ssub_sat X, Y
1859       // X - Y overflows ? (Y <s 1 ? INTMAX : INTMIN) : X - Y --> ssub_sat X, Y
1860       if (Op == Y && Pred == ICmpInst::ICMP_SLT && IsZeroOrOne(*C) &&
1861           IsMinMax(FalseVal, TrueVal))
1862         return true;
1863       // X - Y overflows ? (Y >s 0 ? INTMIN : INTMAX) : X - Y --> ssub_sat X, Y
1864       // X - Y overflows ? (Y >s -1 ? INTMIN : INTMAX) : X - Y --> ssub_sat X, Y
1865       if (Op == Y && Pred == ICmpInst::ICMP_SGT && IsZeroOrOne(*C + 1) &&
1866           IsMinMax(TrueVal, FalseVal))
1867         return true;
1868     }
1869 
1870     return false;
1871   };
1872 
1873   Intrinsic::ID NewIntrinsicID;
1874   if (II->getIntrinsicID() == Intrinsic::uadd_with_overflow &&
1875       match(TrueVal, m_AllOnes()))
1876     // X + Y overflows ? -1 : X + Y -> uadd_sat X, Y
1877     NewIntrinsicID = Intrinsic::uadd_sat;
1878   else if (II->getIntrinsicID() == Intrinsic::usub_with_overflow &&
1879            match(TrueVal, m_Zero()))
1880     // X - Y overflows ? 0 : X - Y -> usub_sat X, Y
1881     NewIntrinsicID = Intrinsic::usub_sat;
1882   else if (II->getIntrinsicID() == Intrinsic::sadd_with_overflow &&
1883            IsSignedSaturateLimit(TrueVal, /*IsAdd=*/true))
1884     // X + Y overflows ? (X <s 0 ? INTMIN : INTMAX) : X + Y --> sadd_sat X, Y
1885     // X + Y overflows ? (X <s 1 ? INTMIN : INTMAX) : X + Y --> sadd_sat X, Y
1886     // X + Y overflows ? (X >s 0 ? INTMAX : INTMIN) : X + Y --> sadd_sat X, Y
1887     // X + Y overflows ? (X >s -1 ? INTMAX : INTMIN) : X + Y --> sadd_sat X, Y
1888     // X + Y overflows ? (Y <s 0 ? INTMIN : INTMAX) : X + Y --> sadd_sat X, Y
1889     // X + Y overflows ? (Y <s 1 ? INTMIN : INTMAX) : X + Y --> sadd_sat X, Y
1890     // X + Y overflows ? (Y >s 0 ? INTMAX : INTMIN) : X + Y --> sadd_sat X, Y
1891     // X + Y overflows ? (Y >s -1 ? INTMAX : INTMIN) : X + Y --> sadd_sat X, Y
1892     NewIntrinsicID = Intrinsic::sadd_sat;
1893   else if (II->getIntrinsicID() == Intrinsic::ssub_with_overflow &&
1894            IsSignedSaturateLimit(TrueVal, /*IsAdd=*/false))
1895     // X - Y overflows ? (X <s 0 ? INTMIN : INTMAX) : X - Y --> ssub_sat X, Y
1896     // X - Y overflows ? (X <s -1 ? INTMIN : INTMAX) : X - Y --> ssub_sat X, Y
1897     // X - Y overflows ? (X >s -1 ? INTMAX : INTMIN) : X - Y --> ssub_sat X, Y
1898     // X - Y overflows ? (X >s -2 ? INTMAX : INTMIN) : X - Y --> ssub_sat X, Y
1899     // X - Y overflows ? (Y <s 0 ? INTMAX : INTMIN) : X - Y --> ssub_sat X, Y
1900     // X - Y overflows ? (Y <s 1 ? INTMAX : INTMIN) : X - Y --> ssub_sat X, Y
1901     // X - Y overflows ? (Y >s 0 ? INTMIN : INTMAX) : X - Y --> ssub_sat X, Y
1902     // X - Y overflows ? (Y >s -1 ? INTMIN : INTMAX) : X - Y --> ssub_sat X, Y
1903     NewIntrinsicID = Intrinsic::ssub_sat;
1904   else
1905     return nullptr;
1906 
1907   Function *F =
1908       Intrinsic::getDeclaration(SI.getModule(), NewIntrinsicID, SI.getType());
1909   return CallInst::Create(F, {X, Y});
1910 }
1911 
1912 Instruction *InstCombinerImpl::foldSelectExtConst(SelectInst &Sel) {
1913   Constant *C;
1914   if (!match(Sel.getTrueValue(), m_Constant(C)) &&
1915       !match(Sel.getFalseValue(), m_Constant(C)))
1916     return nullptr;
1917 
1918   Instruction *ExtInst;
1919   if (!match(Sel.getTrueValue(), m_Instruction(ExtInst)) &&
1920       !match(Sel.getFalseValue(), m_Instruction(ExtInst)))
1921     return nullptr;
1922 
1923   auto ExtOpcode = ExtInst->getOpcode();
1924   if (ExtOpcode != Instruction::ZExt && ExtOpcode != Instruction::SExt)
1925     return nullptr;
1926 
1927   // If we are extending from a boolean type or if we can create a select that
1928   // has the same size operands as its condition, try to narrow the select.
1929   Value *X = ExtInst->getOperand(0);
1930   Type *SmallType = X->getType();
1931   Value *Cond = Sel.getCondition();
1932   auto *Cmp = dyn_cast<CmpInst>(Cond);
1933   if (!SmallType->isIntOrIntVectorTy(1) &&
1934       (!Cmp || Cmp->getOperand(0)->getType() != SmallType))
1935     return nullptr;
1936 
1937   // If the constant is the same after truncation to the smaller type and
1938   // extension to the original type, we can narrow the select.
1939   Type *SelType = Sel.getType();
1940   Constant *TruncC = ConstantExpr::getTrunc(C, SmallType);
1941   Constant *ExtC = ConstantExpr::getCast(ExtOpcode, TruncC, SelType);
1942   if (ExtC == C && ExtInst->hasOneUse()) {
1943     Value *TruncCVal = cast<Value>(TruncC);
1944     if (ExtInst == Sel.getFalseValue())
1945       std::swap(X, TruncCVal);
1946 
1947     // select Cond, (ext X), C --> ext(select Cond, X, C')
1948     // select Cond, C, (ext X) --> ext(select Cond, C', X)
1949     Value *NewSel = Builder.CreateSelect(Cond, X, TruncCVal, "narrow", &Sel);
1950     return CastInst::Create(Instruction::CastOps(ExtOpcode), NewSel, SelType);
1951   }
1952 
1953   // If one arm of the select is the extend of the condition, replace that arm
1954   // with the extension of the appropriate known bool value.
1955   if (Cond == X) {
1956     if (ExtInst == Sel.getTrueValue()) {
1957       // select X, (sext X), C --> select X, -1, C
1958       // select X, (zext X), C --> select X,  1, C
1959       Constant *One = ConstantInt::getTrue(SmallType);
1960       Constant *AllOnesOrOne = ConstantExpr::getCast(ExtOpcode, One, SelType);
1961       return SelectInst::Create(Cond, AllOnesOrOne, C, "", nullptr, &Sel);
1962     } else {
1963       // select X, C, (sext X) --> select X, C, 0
1964       // select X, C, (zext X) --> select X, C, 0
1965       Constant *Zero = ConstantInt::getNullValue(SelType);
1966       return SelectInst::Create(Cond, C, Zero, "", nullptr, &Sel);
1967     }
1968   }
1969 
1970   return nullptr;
1971 }
1972 
1973 /// Try to transform a vector select with a constant condition vector into a
1974 /// shuffle for easier combining with other shuffles and insert/extract.
1975 static Instruction *canonicalizeSelectToShuffle(SelectInst &SI) {
1976   Value *CondVal = SI.getCondition();
1977   Constant *CondC;
1978   auto *CondValTy = dyn_cast<FixedVectorType>(CondVal->getType());
1979   if (!CondValTy || !match(CondVal, m_Constant(CondC)))
1980     return nullptr;
1981 
1982   unsigned NumElts = CondValTy->getNumElements();
1983   SmallVector<int, 16> Mask;
1984   Mask.reserve(NumElts);
1985   for (unsigned i = 0; i != NumElts; ++i) {
1986     Constant *Elt = CondC->getAggregateElement(i);
1987     if (!Elt)
1988       return nullptr;
1989 
1990     if (Elt->isOneValue()) {
1991       // If the select condition element is true, choose from the 1st vector.
1992       Mask.push_back(i);
1993     } else if (Elt->isNullValue()) {
1994       // If the select condition element is false, choose from the 2nd vector.
1995       Mask.push_back(i + NumElts);
1996     } else if (isa<UndefValue>(Elt)) {
1997       // Undef in a select condition (choose one of the operands) does not mean
1998       // the same thing as undef in a shuffle mask (any value is acceptable), so
1999       // give up.
2000       return nullptr;
2001     } else {
2002       // Bail out on a constant expression.
2003       return nullptr;
2004     }
2005   }
2006 
2007   return new ShuffleVectorInst(SI.getTrueValue(), SI.getFalseValue(), Mask);
2008 }
2009 
2010 /// If we have a select of vectors with a scalar condition, try to convert that
2011 /// to a vector select by splatting the condition. A splat may get folded with
2012 /// other operations in IR and having all operands of a select be vector types
2013 /// is likely better for vector codegen.
2014 static Instruction *canonicalizeScalarSelectOfVecs(SelectInst &Sel,
2015                                                    InstCombinerImpl &IC) {
2016   auto *Ty = dyn_cast<VectorType>(Sel.getType());
2017   if (!Ty)
2018     return nullptr;
2019 
2020   // We can replace a single-use extract with constant index.
2021   Value *Cond = Sel.getCondition();
2022   if (!match(Cond, m_OneUse(m_ExtractElt(m_Value(), m_ConstantInt()))))
2023     return nullptr;
2024 
2025   // select (extelt V, Index), T, F --> select (splat V, Index), T, F
2026   // Splatting the extracted condition reduces code (we could directly create a
2027   // splat shuffle of the source vector to eliminate the intermediate step).
2028   return IC.replaceOperand(
2029       Sel, 0, IC.Builder.CreateVectorSplat(Ty->getElementCount(), Cond));
2030 }
2031 
2032 /// Reuse bitcasted operands between a compare and select:
2033 /// select (cmp (bitcast C), (bitcast D)), (bitcast' C), (bitcast' D) -->
2034 /// bitcast (select (cmp (bitcast C), (bitcast D)), (bitcast C), (bitcast D))
2035 static Instruction *foldSelectCmpBitcasts(SelectInst &Sel,
2036                                           InstCombiner::BuilderTy &Builder) {
2037   Value *Cond = Sel.getCondition();
2038   Value *TVal = Sel.getTrueValue();
2039   Value *FVal = Sel.getFalseValue();
2040 
2041   CmpInst::Predicate Pred;
2042   Value *A, *B;
2043   if (!match(Cond, m_Cmp(Pred, m_Value(A), m_Value(B))))
2044     return nullptr;
2045 
2046   // The select condition is a compare instruction. If the select's true/false
2047   // values are already the same as the compare operands, there's nothing to do.
2048   if (TVal == A || TVal == B || FVal == A || FVal == B)
2049     return nullptr;
2050 
2051   Value *C, *D;
2052   if (!match(A, m_BitCast(m_Value(C))) || !match(B, m_BitCast(m_Value(D))))
2053     return nullptr;
2054 
2055   // select (cmp (bitcast C), (bitcast D)), (bitcast TSrc), (bitcast FSrc)
2056   Value *TSrc, *FSrc;
2057   if (!match(TVal, m_BitCast(m_Value(TSrc))) ||
2058       !match(FVal, m_BitCast(m_Value(FSrc))))
2059     return nullptr;
2060 
2061   // If the select true/false values are *different bitcasts* of the same source
2062   // operands, make the select operands the same as the compare operands and
2063   // cast the result. This is the canonical select form for min/max.
2064   Value *NewSel;
2065   if (TSrc == C && FSrc == D) {
2066     // select (cmp (bitcast C), (bitcast D)), (bitcast' C), (bitcast' D) -->
2067     // bitcast (select (cmp A, B), A, B)
2068     NewSel = Builder.CreateSelect(Cond, A, B, "", &Sel);
2069   } else if (TSrc == D && FSrc == C) {
2070     // select (cmp (bitcast C), (bitcast D)), (bitcast' D), (bitcast' C) -->
2071     // bitcast (select (cmp A, B), B, A)
2072     NewSel = Builder.CreateSelect(Cond, B, A, "", &Sel);
2073   } else {
2074     return nullptr;
2075   }
2076   return CastInst::CreateBitOrPointerCast(NewSel, Sel.getType());
2077 }
2078 
2079 /// Try to eliminate select instructions that test the returned flag of cmpxchg
2080 /// instructions.
2081 ///
2082 /// If a select instruction tests the returned flag of a cmpxchg instruction and
2083 /// selects between the returned value of the cmpxchg instruction its compare
2084 /// operand, the result of the select will always be equal to its false value.
2085 /// For example:
2086 ///
2087 ///   %0 = cmpxchg i64* %ptr, i64 %compare, i64 %new_value seq_cst seq_cst
2088 ///   %1 = extractvalue { i64, i1 } %0, 1
2089 ///   %2 = extractvalue { i64, i1 } %0, 0
2090 ///   %3 = select i1 %1, i64 %compare, i64 %2
2091 ///   ret i64 %3
2092 ///
2093 /// The returned value of the cmpxchg instruction (%2) is the original value
2094 /// located at %ptr prior to any update. If the cmpxchg operation succeeds, %2
2095 /// must have been equal to %compare. Thus, the result of the select is always
2096 /// equal to %2, and the code can be simplified to:
2097 ///
2098 ///   %0 = cmpxchg i64* %ptr, i64 %compare, i64 %new_value seq_cst seq_cst
2099 ///   %1 = extractvalue { i64, i1 } %0, 0
2100 ///   ret i64 %1
2101 ///
2102 static Value *foldSelectCmpXchg(SelectInst &SI) {
2103   // A helper that determines if V is an extractvalue instruction whose
2104   // aggregate operand is a cmpxchg instruction and whose single index is equal
2105   // to I. If such conditions are true, the helper returns the cmpxchg
2106   // instruction; otherwise, a nullptr is returned.
2107   auto isExtractFromCmpXchg = [](Value *V, unsigned I) -> AtomicCmpXchgInst * {
2108     auto *Extract = dyn_cast<ExtractValueInst>(V);
2109     if (!Extract)
2110       return nullptr;
2111     if (Extract->getIndices()[0] != I)
2112       return nullptr;
2113     return dyn_cast<AtomicCmpXchgInst>(Extract->getAggregateOperand());
2114   };
2115 
2116   // If the select has a single user, and this user is a select instruction that
2117   // we can simplify, skip the cmpxchg simplification for now.
2118   if (SI.hasOneUse())
2119     if (auto *Select = dyn_cast<SelectInst>(SI.user_back()))
2120       if (Select->getCondition() == SI.getCondition())
2121         if (Select->getFalseValue() == SI.getTrueValue() ||
2122             Select->getTrueValue() == SI.getFalseValue())
2123           return nullptr;
2124 
2125   // Ensure the select condition is the returned flag of a cmpxchg instruction.
2126   auto *CmpXchg = isExtractFromCmpXchg(SI.getCondition(), 1);
2127   if (!CmpXchg)
2128     return nullptr;
2129 
2130   // Check the true value case: The true value of the select is the returned
2131   // value of the same cmpxchg used by the condition, and the false value is the
2132   // cmpxchg instruction's compare operand.
2133   if (auto *X = isExtractFromCmpXchg(SI.getTrueValue(), 0))
2134     if (X == CmpXchg && X->getCompareOperand() == SI.getFalseValue())
2135       return SI.getFalseValue();
2136 
2137   // Check the false value case: The false value of the select is the returned
2138   // value of the same cmpxchg used by the condition, and the true value is the
2139   // cmpxchg instruction's compare operand.
2140   if (auto *X = isExtractFromCmpXchg(SI.getFalseValue(), 0))
2141     if (X == CmpXchg && X->getCompareOperand() == SI.getTrueValue())
2142       return SI.getFalseValue();
2143 
2144   return nullptr;
2145 }
2146 
2147 static Instruction *moveAddAfterMinMax(SelectPatternFlavor SPF, Value *X,
2148                                        Value *Y,
2149                                        InstCombiner::BuilderTy &Builder) {
2150   assert(SelectPatternResult::isMinOrMax(SPF) && "Expected min/max pattern");
2151   bool IsUnsigned = SPF == SelectPatternFlavor::SPF_UMIN ||
2152                     SPF == SelectPatternFlavor::SPF_UMAX;
2153   // TODO: If InstSimplify could fold all cases where C2 <= C1, we could change
2154   // the constant value check to an assert.
2155   Value *A;
2156   const APInt *C1, *C2;
2157   if (IsUnsigned && match(X, m_NUWAdd(m_Value(A), m_APInt(C1))) &&
2158       match(Y, m_APInt(C2)) && C2->uge(*C1) && X->hasNUses(2)) {
2159     // umin (add nuw A, C1), C2 --> add nuw (umin A, C2 - C1), C1
2160     // umax (add nuw A, C1), C2 --> add nuw (umax A, C2 - C1), C1
2161     Value *NewMinMax = createMinMax(Builder, SPF, A,
2162                                     ConstantInt::get(X->getType(), *C2 - *C1));
2163     return BinaryOperator::CreateNUW(BinaryOperator::Add, NewMinMax,
2164                                      ConstantInt::get(X->getType(), *C1));
2165   }
2166 
2167   if (!IsUnsigned && match(X, m_NSWAdd(m_Value(A), m_APInt(C1))) &&
2168       match(Y, m_APInt(C2)) && X->hasNUses(2)) {
2169     bool Overflow;
2170     APInt Diff = C2->ssub_ov(*C1, Overflow);
2171     if (!Overflow) {
2172       // smin (add nsw A, C1), C2 --> add nsw (smin A, C2 - C1), C1
2173       // smax (add nsw A, C1), C2 --> add nsw (smax A, C2 - C1), C1
2174       Value *NewMinMax = createMinMax(Builder, SPF, A,
2175                                       ConstantInt::get(X->getType(), Diff));
2176       return BinaryOperator::CreateNSW(BinaryOperator::Add, NewMinMax,
2177                                        ConstantInt::get(X->getType(), *C1));
2178     }
2179   }
2180 
2181   return nullptr;
2182 }
2183 
2184 /// Match a sadd_sat or ssub_sat which is using min/max to clamp the value.
2185 Instruction *InstCombinerImpl::matchSAddSubSat(SelectInst &MinMax1) {
2186   Type *Ty = MinMax1.getType();
2187 
2188   // We are looking for a tree of:
2189   // max(INT_MIN, min(INT_MAX, add(sext(A), sext(B))))
2190   // Where the min and max could be reversed
2191   Instruction *MinMax2;
2192   BinaryOperator *AddSub;
2193   const APInt *MinValue, *MaxValue;
2194   if (match(&MinMax1, m_SMin(m_Instruction(MinMax2), m_APInt(MaxValue)))) {
2195     if (!match(MinMax2, m_SMax(m_BinOp(AddSub), m_APInt(MinValue))))
2196       return nullptr;
2197   } else if (match(&MinMax1,
2198                    m_SMax(m_Instruction(MinMax2), m_APInt(MinValue)))) {
2199     if (!match(MinMax2, m_SMin(m_BinOp(AddSub), m_APInt(MaxValue))))
2200       return nullptr;
2201   } else
2202     return nullptr;
2203 
2204   // Check that the constants clamp a saturate, and that the new type would be
2205   // sensible to convert to.
2206   if (!(*MaxValue + 1).isPowerOf2() || -*MinValue != *MaxValue + 1)
2207     return nullptr;
2208   // In what bitwidth can this be treated as saturating arithmetics?
2209   unsigned NewBitWidth = (*MaxValue + 1).logBase2() + 1;
2210   // FIXME: This isn't quite right for vectors, but using the scalar type is a
2211   // good first approximation for what should be done there.
2212   if (!shouldChangeType(Ty->getScalarType()->getIntegerBitWidth(), NewBitWidth))
2213     return nullptr;
2214 
2215   // Also make sure that the number of uses is as expected. The "3"s are for the
2216   // the two items of min/max (the compare and the select).
2217   if (MinMax2->hasNUsesOrMore(3) || AddSub->hasNUsesOrMore(3))
2218     return nullptr;
2219 
2220   // Create the new type (which can be a vector type)
2221   Type *NewTy = Ty->getWithNewBitWidth(NewBitWidth);
2222   // Match the two extends from the add/sub
2223   Value *A, *B;
2224   if(!match(AddSub, m_BinOp(m_SExt(m_Value(A)), m_SExt(m_Value(B)))))
2225     return nullptr;
2226   // And check the incoming values are of a type smaller than or equal to the
2227   // size of the saturation. Otherwise the higher bits can cause different
2228   // results.
2229   if (A->getType()->getScalarSizeInBits() > NewBitWidth ||
2230       B->getType()->getScalarSizeInBits() > NewBitWidth)
2231     return nullptr;
2232 
2233   Intrinsic::ID IntrinsicID;
2234   if (AddSub->getOpcode() == Instruction::Add)
2235     IntrinsicID = Intrinsic::sadd_sat;
2236   else if (AddSub->getOpcode() == Instruction::Sub)
2237     IntrinsicID = Intrinsic::ssub_sat;
2238   else
2239     return nullptr;
2240 
2241   // Finally create and return the sat intrinsic, truncated to the new type
2242   Function *F = Intrinsic::getDeclaration(MinMax1.getModule(), IntrinsicID, NewTy);
2243   Value *AT = Builder.CreateSExt(A, NewTy);
2244   Value *BT = Builder.CreateSExt(B, NewTy);
2245   Value *Sat = Builder.CreateCall(F, {AT, BT});
2246   return CastInst::Create(Instruction::SExt, Sat, Ty);
2247 }
2248 
2249 /// Reduce a sequence of min/max with a common operand.
2250 static Instruction *factorizeMinMaxTree(SelectPatternFlavor SPF, Value *LHS,
2251                                         Value *RHS,
2252                                         InstCombiner::BuilderTy &Builder) {
2253   assert(SelectPatternResult::isMinOrMax(SPF) && "Expected a min/max");
2254   // TODO: Allow FP min/max with nnan/nsz.
2255   if (!LHS->getType()->isIntOrIntVectorTy())
2256     return nullptr;
2257 
2258   // Match 3 of the same min/max ops. Example: umin(umin(), umin()).
2259   Value *A, *B, *C, *D;
2260   SelectPatternResult L = matchSelectPattern(LHS, A, B);
2261   SelectPatternResult R = matchSelectPattern(RHS, C, D);
2262   if (SPF != L.Flavor || L.Flavor != R.Flavor)
2263     return nullptr;
2264 
2265   // Look for a common operand. The use checks are different than usual because
2266   // a min/max pattern typically has 2 uses of each op: 1 by the cmp and 1 by
2267   // the select.
2268   Value *MinMaxOp = nullptr;
2269   Value *ThirdOp = nullptr;
2270   if (!LHS->hasNUsesOrMore(3) && RHS->hasNUsesOrMore(3)) {
2271     // If the LHS is only used in this chain and the RHS is used outside of it,
2272     // reuse the RHS min/max because that will eliminate the LHS.
2273     if (D == A || C == A) {
2274       // min(min(a, b), min(c, a)) --> min(min(c, a), b)
2275       // min(min(a, b), min(a, d)) --> min(min(a, d), b)
2276       MinMaxOp = RHS;
2277       ThirdOp = B;
2278     } else if (D == B || C == B) {
2279       // min(min(a, b), min(c, b)) --> min(min(c, b), a)
2280       // min(min(a, b), min(b, d)) --> min(min(b, d), a)
2281       MinMaxOp = RHS;
2282       ThirdOp = A;
2283     }
2284   } else if (!RHS->hasNUsesOrMore(3)) {
2285     // Reuse the LHS. This will eliminate the RHS.
2286     if (D == A || D == B) {
2287       // min(min(a, b), min(c, a)) --> min(min(a, b), c)
2288       // min(min(a, b), min(c, b)) --> min(min(a, b), c)
2289       MinMaxOp = LHS;
2290       ThirdOp = C;
2291     } else if (C == A || C == B) {
2292       // min(min(a, b), min(b, d)) --> min(min(a, b), d)
2293       // min(min(a, b), min(c, b)) --> min(min(a, b), d)
2294       MinMaxOp = LHS;
2295       ThirdOp = D;
2296     }
2297   }
2298   if (!MinMaxOp || !ThirdOp)
2299     return nullptr;
2300 
2301   CmpInst::Predicate P = getMinMaxPred(SPF);
2302   Value *CmpABC = Builder.CreateICmp(P, MinMaxOp, ThirdOp);
2303   return SelectInst::Create(CmpABC, MinMaxOp, ThirdOp);
2304 }
2305 
2306 /// Try to reduce a funnel/rotate pattern that includes a compare and select
2307 /// into a funnel shift intrinsic. Example:
2308 /// rotl32(a, b) --> (b == 0 ? a : ((a >> (32 - b)) | (a << b)))
2309 ///              --> call llvm.fshl.i32(a, a, b)
2310 /// fshl32(a, b, c) --> (c == 0 ? a : ((b >> (32 - c)) | (a << c)))
2311 ///                 --> call llvm.fshl.i32(a, b, c)
2312 /// fshr32(a, b, c) --> (c == 0 ? b : ((a >> (32 - c)) | (b << c)))
2313 ///                 --> call llvm.fshr.i32(a, b, c)
2314 static Instruction *foldSelectFunnelShift(SelectInst &Sel,
2315                                           InstCombiner::BuilderTy &Builder) {
2316   // This must be a power-of-2 type for a bitmasking transform to be valid.
2317   unsigned Width = Sel.getType()->getScalarSizeInBits();
2318   if (!isPowerOf2_32(Width))
2319     return nullptr;
2320 
2321   BinaryOperator *Or0, *Or1;
2322   if (!match(Sel.getFalseValue(), m_OneUse(m_Or(m_BinOp(Or0), m_BinOp(Or1)))))
2323     return nullptr;
2324 
2325   Value *SV0, *SV1, *SA0, *SA1;
2326   if (!match(Or0, m_OneUse(m_LogicalShift(m_Value(SV0),
2327                                           m_ZExtOrSelf(m_Value(SA0))))) ||
2328       !match(Or1, m_OneUse(m_LogicalShift(m_Value(SV1),
2329                                           m_ZExtOrSelf(m_Value(SA1))))) ||
2330       Or0->getOpcode() == Or1->getOpcode())
2331     return nullptr;
2332 
2333   // Canonicalize to or(shl(SV0, SA0), lshr(SV1, SA1)).
2334   if (Or0->getOpcode() == BinaryOperator::LShr) {
2335     std::swap(Or0, Or1);
2336     std::swap(SV0, SV1);
2337     std::swap(SA0, SA1);
2338   }
2339   assert(Or0->getOpcode() == BinaryOperator::Shl &&
2340          Or1->getOpcode() == BinaryOperator::LShr &&
2341          "Illegal or(shift,shift) pair");
2342 
2343   // Check the shift amounts to see if they are an opposite pair.
2344   Value *ShAmt;
2345   if (match(SA1, m_OneUse(m_Sub(m_SpecificInt(Width), m_Specific(SA0)))))
2346     ShAmt = SA0;
2347   else if (match(SA0, m_OneUse(m_Sub(m_SpecificInt(Width), m_Specific(SA1)))))
2348     ShAmt = SA1;
2349   else
2350     return nullptr;
2351 
2352   // We should now have this pattern:
2353   // select ?, TVal, (or (shl SV0, SA0), (lshr SV1, SA1))
2354   // The false value of the select must be a funnel-shift of the true value:
2355   // IsFShl -> TVal must be SV0 else TVal must be SV1.
2356   bool IsFshl = (ShAmt == SA0);
2357   Value *TVal = Sel.getTrueValue();
2358   if ((IsFshl && TVal != SV0) || (!IsFshl && TVal != SV1))
2359     return nullptr;
2360 
2361   // Finally, see if the select is filtering out a shift-by-zero.
2362   Value *Cond = Sel.getCondition();
2363   ICmpInst::Predicate Pred;
2364   if (!match(Cond, m_OneUse(m_ICmp(Pred, m_Specific(ShAmt), m_ZeroInt()))) ||
2365       Pred != ICmpInst::ICMP_EQ)
2366     return nullptr;
2367 
2368   // If this is not a rotate then the select was blocking poison from the
2369   // 'shift-by-zero' non-TVal, but a funnel shift won't - so freeze it.
2370   if (SV0 != SV1) {
2371     if (IsFshl && !llvm::isGuaranteedNotToBePoison(SV1))
2372       SV1 = Builder.CreateFreeze(SV1);
2373     else if (!IsFshl && !llvm::isGuaranteedNotToBePoison(SV0))
2374       SV0 = Builder.CreateFreeze(SV0);
2375   }
2376 
2377   // This is a funnel/rotate that avoids shift-by-bitwidth UB in a suboptimal way.
2378   // Convert to funnel shift intrinsic.
2379   Intrinsic::ID IID = IsFshl ? Intrinsic::fshl : Intrinsic::fshr;
2380   Function *F = Intrinsic::getDeclaration(Sel.getModule(), IID, Sel.getType());
2381   ShAmt = Builder.CreateZExt(ShAmt, Sel.getType());
2382   return CallInst::Create(F, { SV0, SV1, ShAmt });
2383 }
2384 
2385 static Instruction *foldSelectToCopysign(SelectInst &Sel,
2386                                          InstCombiner::BuilderTy &Builder) {
2387   Value *Cond = Sel.getCondition();
2388   Value *TVal = Sel.getTrueValue();
2389   Value *FVal = Sel.getFalseValue();
2390   Type *SelType = Sel.getType();
2391 
2392   // Match select ?, TC, FC where the constants are equal but negated.
2393   // TODO: Generalize to handle a negated variable operand?
2394   const APFloat *TC, *FC;
2395   if (!match(TVal, m_APFloat(TC)) || !match(FVal, m_APFloat(FC)) ||
2396       !abs(*TC).bitwiseIsEqual(abs(*FC)))
2397     return nullptr;
2398 
2399   assert(TC != FC && "Expected equal select arms to simplify");
2400 
2401   Value *X;
2402   const APInt *C;
2403   bool IsTrueIfSignSet;
2404   ICmpInst::Predicate Pred;
2405   if (!match(Cond, m_OneUse(m_ICmp(Pred, m_BitCast(m_Value(X)), m_APInt(C)))) ||
2406       !InstCombiner::isSignBitCheck(Pred, *C, IsTrueIfSignSet) ||
2407       X->getType() != SelType)
2408     return nullptr;
2409 
2410   // If needed, negate the value that will be the sign argument of the copysign:
2411   // (bitcast X) <  0 ? -TC :  TC --> copysign(TC,  X)
2412   // (bitcast X) <  0 ?  TC : -TC --> copysign(TC, -X)
2413   // (bitcast X) >= 0 ? -TC :  TC --> copysign(TC, -X)
2414   // (bitcast X) >= 0 ?  TC : -TC --> copysign(TC,  X)
2415   if (IsTrueIfSignSet ^ TC->isNegative())
2416     X = Builder.CreateFNegFMF(X, &Sel);
2417 
2418   // Canonicalize the magnitude argument as the positive constant since we do
2419   // not care about its sign.
2420   Value *MagArg = TC->isNegative() ? FVal : TVal;
2421   Function *F = Intrinsic::getDeclaration(Sel.getModule(), Intrinsic::copysign,
2422                                           Sel.getType());
2423   Instruction *CopySign = CallInst::Create(F, { MagArg, X });
2424   CopySign->setFastMathFlags(Sel.getFastMathFlags());
2425   return CopySign;
2426 }
2427 
2428 Instruction *InstCombinerImpl::foldVectorSelect(SelectInst &Sel) {
2429   auto *VecTy = dyn_cast<FixedVectorType>(Sel.getType());
2430   if (!VecTy)
2431     return nullptr;
2432 
2433   unsigned NumElts = VecTy->getNumElements();
2434   APInt UndefElts(NumElts, 0);
2435   APInt AllOnesEltMask(APInt::getAllOnesValue(NumElts));
2436   if (Value *V = SimplifyDemandedVectorElts(&Sel, AllOnesEltMask, UndefElts)) {
2437     if (V != &Sel)
2438       return replaceInstUsesWith(Sel, V);
2439     return &Sel;
2440   }
2441 
2442   // A select of a "select shuffle" with a common operand can be rearranged
2443   // to select followed by "select shuffle". Because of poison, this only works
2444   // in the case of a shuffle with no undefined mask elements.
2445   Value *Cond = Sel.getCondition();
2446   Value *TVal = Sel.getTrueValue();
2447   Value *FVal = Sel.getFalseValue();
2448   Value *X, *Y;
2449   ArrayRef<int> Mask;
2450   if (match(TVal, m_OneUse(m_Shuffle(m_Value(X), m_Value(Y), m_Mask(Mask)))) &&
2451       !is_contained(Mask, UndefMaskElem) &&
2452       cast<ShuffleVectorInst>(TVal)->isSelect()) {
2453     if (X == FVal) {
2454       // select Cond, (shuf_sel X, Y), X --> shuf_sel X, (select Cond, Y, X)
2455       Value *NewSel = Builder.CreateSelect(Cond, Y, X, "sel", &Sel);
2456       return new ShuffleVectorInst(X, NewSel, Mask);
2457     }
2458     if (Y == FVal) {
2459       // select Cond, (shuf_sel X, Y), Y --> shuf_sel (select Cond, X, Y), Y
2460       Value *NewSel = Builder.CreateSelect(Cond, X, Y, "sel", &Sel);
2461       return new ShuffleVectorInst(NewSel, Y, Mask);
2462     }
2463   }
2464   if (match(FVal, m_OneUse(m_Shuffle(m_Value(X), m_Value(Y), m_Mask(Mask)))) &&
2465       !is_contained(Mask, UndefMaskElem) &&
2466       cast<ShuffleVectorInst>(FVal)->isSelect()) {
2467     if (X == TVal) {
2468       // select Cond, X, (shuf_sel X, Y) --> shuf_sel X, (select Cond, X, Y)
2469       Value *NewSel = Builder.CreateSelect(Cond, X, Y, "sel", &Sel);
2470       return new ShuffleVectorInst(X, NewSel, Mask);
2471     }
2472     if (Y == TVal) {
2473       // select Cond, Y, (shuf_sel X, Y) --> shuf_sel (select Cond, Y, X), Y
2474       Value *NewSel = Builder.CreateSelect(Cond, Y, X, "sel", &Sel);
2475       return new ShuffleVectorInst(NewSel, Y, Mask);
2476     }
2477   }
2478 
2479   return nullptr;
2480 }
2481 
2482 static Instruction *foldSelectToPhiImpl(SelectInst &Sel, BasicBlock *BB,
2483                                         const DominatorTree &DT,
2484                                         InstCombiner::BuilderTy &Builder) {
2485   // Find the block's immediate dominator that ends with a conditional branch
2486   // that matches select's condition (maybe inverted).
2487   auto *IDomNode = DT[BB]->getIDom();
2488   if (!IDomNode)
2489     return nullptr;
2490   BasicBlock *IDom = IDomNode->getBlock();
2491 
2492   Value *Cond = Sel.getCondition();
2493   Value *IfTrue, *IfFalse;
2494   BasicBlock *TrueSucc, *FalseSucc;
2495   if (match(IDom->getTerminator(),
2496             m_Br(m_Specific(Cond), m_BasicBlock(TrueSucc),
2497                  m_BasicBlock(FalseSucc)))) {
2498     IfTrue = Sel.getTrueValue();
2499     IfFalse = Sel.getFalseValue();
2500   } else if (match(IDom->getTerminator(),
2501                    m_Br(m_Not(m_Specific(Cond)), m_BasicBlock(TrueSucc),
2502                         m_BasicBlock(FalseSucc)))) {
2503     IfTrue = Sel.getFalseValue();
2504     IfFalse = Sel.getTrueValue();
2505   } else
2506     return nullptr;
2507 
2508   // Make sure the branches are actually different.
2509   if (TrueSucc == FalseSucc)
2510     return nullptr;
2511 
2512   // We want to replace select %cond, %a, %b with a phi that takes value %a
2513   // for all incoming edges that are dominated by condition `%cond == true`,
2514   // and value %b for edges dominated by condition `%cond == false`. If %a
2515   // or %b are also phis from the same basic block, we can go further and take
2516   // their incoming values from the corresponding blocks.
2517   BasicBlockEdge TrueEdge(IDom, TrueSucc);
2518   BasicBlockEdge FalseEdge(IDom, FalseSucc);
2519   DenseMap<BasicBlock *, Value *> Inputs;
2520   for (auto *Pred : predecessors(BB)) {
2521     // Check implication.
2522     BasicBlockEdge Incoming(Pred, BB);
2523     if (DT.dominates(TrueEdge, Incoming))
2524       Inputs[Pred] = IfTrue->DoPHITranslation(BB, Pred);
2525     else if (DT.dominates(FalseEdge, Incoming))
2526       Inputs[Pred] = IfFalse->DoPHITranslation(BB, Pred);
2527     else
2528       return nullptr;
2529     // Check availability.
2530     if (auto *Insn = dyn_cast<Instruction>(Inputs[Pred]))
2531       if (!DT.dominates(Insn, Pred->getTerminator()))
2532         return nullptr;
2533   }
2534 
2535   Builder.SetInsertPoint(&*BB->begin());
2536   auto *PN = Builder.CreatePHI(Sel.getType(), Inputs.size());
2537   for (auto *Pred : predecessors(BB))
2538     PN->addIncoming(Inputs[Pred], Pred);
2539   PN->takeName(&Sel);
2540   return PN;
2541 }
2542 
2543 static Instruction *foldSelectToPhi(SelectInst &Sel, const DominatorTree &DT,
2544                                     InstCombiner::BuilderTy &Builder) {
2545   // Try to replace this select with Phi in one of these blocks.
2546   SmallSetVector<BasicBlock *, 4> CandidateBlocks;
2547   CandidateBlocks.insert(Sel.getParent());
2548   for (Value *V : Sel.operands())
2549     if (auto *I = dyn_cast<Instruction>(V))
2550       CandidateBlocks.insert(I->getParent());
2551 
2552   for (BasicBlock *BB : CandidateBlocks)
2553     if (auto *PN = foldSelectToPhiImpl(Sel, BB, DT, Builder))
2554       return PN;
2555   return nullptr;
2556 }
2557 
2558 static Value *foldSelectWithFrozenICmp(SelectInst &Sel, InstCombiner::BuilderTy &Builder) {
2559   FreezeInst *FI = dyn_cast<FreezeInst>(Sel.getCondition());
2560   if (!FI)
2561     return nullptr;
2562 
2563   Value *Cond = FI->getOperand(0);
2564   Value *TrueVal = Sel.getTrueValue(), *FalseVal = Sel.getFalseValue();
2565 
2566   //   select (freeze(x == y)), x, y --> y
2567   //   select (freeze(x != y)), x, y --> x
2568   // The freeze should be only used by this select. Otherwise, remaining uses of
2569   // the freeze can observe a contradictory value.
2570   //   c = freeze(x == y)   ; Let's assume that y = poison & x = 42; c is 0 or 1
2571   //   a = select c, x, y   ;
2572   //   f(a, c)              ; f(poison, 1) cannot happen, but if a is folded
2573   //                        ; to y, this can happen.
2574   CmpInst::Predicate Pred;
2575   if (FI->hasOneUse() &&
2576       match(Cond, m_c_ICmp(Pred, m_Specific(TrueVal), m_Specific(FalseVal))) &&
2577       (Pred == ICmpInst::ICMP_EQ || Pred == ICmpInst::ICMP_NE)) {
2578     return Pred == ICmpInst::ICMP_EQ ? FalseVal : TrueVal;
2579   }
2580 
2581   return nullptr;
2582 }
2583 
2584 Instruction *InstCombinerImpl::foldAndOrOfSelectUsingImpliedCond(Value *Op,
2585                                                                  SelectInst &SI,
2586                                                                  bool IsAnd) {
2587   Value *CondVal = SI.getCondition();
2588   Value *A = SI.getTrueValue();
2589   Value *B = SI.getFalseValue();
2590 
2591   assert(Op->getType()->isIntOrIntVectorTy(1) &&
2592          "Op must be either i1 or vector of i1.");
2593 
2594   Optional<bool> Res = isImpliedCondition(Op, CondVal, DL, IsAnd);
2595   if (!Res)
2596     return nullptr;
2597 
2598   Value *Zero = Constant::getNullValue(A->getType());
2599   Value *One = Constant::getAllOnesValue(A->getType());
2600 
2601   if (*Res == true) {
2602     if (IsAnd)
2603       // select op, (select cond, A, B), false => select op, A, false
2604       // and    op, (select cond, A, B)        => select op, A, false
2605       //   if op = true implies condval = true.
2606       return SelectInst::Create(Op, A, Zero);
2607     else
2608       // select op, true, (select cond, A, B) => select op, true, A
2609       // or     op, (select cond, A, B)       => select op, true, A
2610       //   if op = false implies condval = true.
2611       return SelectInst::Create(Op, One, A);
2612   } else {
2613     if (IsAnd)
2614       // select op, (select cond, A, B), false => select op, B, false
2615       // and    op, (select cond, A, B)        => select op, B, false
2616       //   if op = true implies condval = false.
2617       return SelectInst::Create(Op, B, Zero);
2618     else
2619       // select op, true, (select cond, A, B) => select op, true, B
2620       // or     op, (select cond, A, B)       => select op, true, B
2621       //   if op = false implies condval = false.
2622       return SelectInst::Create(Op, One, B);
2623   }
2624 }
2625 
2626 Instruction *InstCombinerImpl::visitSelectInst(SelectInst &SI) {
2627   Value *CondVal = SI.getCondition();
2628   Value *TrueVal = SI.getTrueValue();
2629   Value *FalseVal = SI.getFalseValue();
2630   Type *SelType = SI.getType();
2631 
2632   // FIXME: Remove this workaround when freeze related patches are done.
2633   // For select with undef operand which feeds into an equality comparison,
2634   // don't simplify it so loop unswitch can know the equality comparison
2635   // may have an undef operand. This is a workaround for PR31652 caused by
2636   // descrepancy about branch on undef between LoopUnswitch and GVN.
2637   if (match(TrueVal, m_Undef()) || match(FalseVal, m_Undef())) {
2638     if (llvm::any_of(SI.users(), [&](User *U) {
2639           ICmpInst *CI = dyn_cast<ICmpInst>(U);
2640           if (CI && CI->isEquality())
2641             return true;
2642           return false;
2643         })) {
2644       return nullptr;
2645     }
2646   }
2647 
2648   if (Value *V = SimplifySelectInst(CondVal, TrueVal, FalseVal,
2649                                     SQ.getWithInstruction(&SI)))
2650     return replaceInstUsesWith(SI, V);
2651 
2652   if (Instruction *I = canonicalizeSelectToShuffle(SI))
2653     return I;
2654 
2655   if (Instruction *I = canonicalizeScalarSelectOfVecs(SI, *this))
2656     return I;
2657 
2658   CmpInst::Predicate Pred;
2659 
2660   // Avoid potential infinite loops by checking for non-constant condition.
2661   // TODO: Can we assert instead by improving canonicalizeSelectToShuffle()?
2662   //       Scalar select must have simplified?
2663   if (SelType->isIntOrIntVectorTy(1) && !isa<Constant>(CondVal) &&
2664       TrueVal->getType() == CondVal->getType()) {
2665     // Folding select to and/or i1 isn't poison safe in general. impliesPoison
2666     // checks whether folding it does not convert a well-defined value into
2667     // poison.
2668     if (match(TrueVal, m_One()) && impliesPoison(FalseVal, CondVal)) {
2669       // Change: A = select B, true, C --> A = or B, C
2670       return BinaryOperator::CreateOr(CondVal, FalseVal);
2671     }
2672     if (match(FalseVal, m_Zero()) && impliesPoison(TrueVal, CondVal)) {
2673       // Change: A = select B, C, false --> A = and B, C
2674       return BinaryOperator::CreateAnd(CondVal, TrueVal);
2675     }
2676 
2677     auto *One = ConstantInt::getTrue(SelType);
2678     auto *Zero = ConstantInt::getFalse(SelType);
2679 
2680     // We match the "full" 0 or 1 constant here to avoid a potential infinite
2681     // loop with vectors that may have undefined/poison elements.
2682     // select a, false, b -> select !a, b, false
2683     if (match(TrueVal, m_Specific(Zero))) {
2684       Value *NotCond = Builder.CreateNot(CondVal, "not." + CondVal->getName());
2685       return SelectInst::Create(NotCond, FalseVal, Zero);
2686     }
2687     // select a, b, true -> select !a, true, b
2688     if (match(FalseVal, m_Specific(One))) {
2689       Value *NotCond = Builder.CreateNot(CondVal, "not." + CondVal->getName());
2690       return SelectInst::Create(NotCond, One, TrueVal);
2691     }
2692 
2693     // select a, a, b -> select a, true, b
2694     if (CondVal == TrueVal)
2695       return replaceOperand(SI, 1, One);
2696     // select a, b, a -> select a, b, false
2697     if (CondVal == FalseVal)
2698       return replaceOperand(SI, 2, Zero);
2699 
2700     // select a, !a, b -> select !a, b, false
2701     if (match(TrueVal, m_Not(m_Specific(CondVal))))
2702       return SelectInst::Create(TrueVal, FalseVal, Zero);
2703     // select a, b, !a -> select !a, true, b
2704     if (match(FalseVal, m_Not(m_Specific(CondVal))))
2705       return SelectInst::Create(FalseVal, One, TrueVal);
2706 
2707     Value *A, *B;
2708 
2709     // DeMorgan in select form: !a && !b --> !(a || b)
2710     // select !a, !b, false --> not (select a, true, b)
2711     if (match(&SI, m_LogicalAnd(m_Not(m_Value(A)), m_Not(m_Value(B)))) &&
2712         (CondVal->hasOneUse() || TrueVal->hasOneUse()) &&
2713         !match(A, m_ConstantExpr()) && !match(B, m_ConstantExpr()))
2714       return BinaryOperator::CreateNot(Builder.CreateSelect(A, One, B));
2715 
2716     // DeMorgan in select form: !a || !b --> !(a && b)
2717     // select !a, true, !b --> not (select a, b, false)
2718     if (match(&SI, m_LogicalOr(m_Not(m_Value(A)), m_Not(m_Value(B)))) &&
2719         (CondVal->hasOneUse() || FalseVal->hasOneUse()) &&
2720         !match(A, m_ConstantExpr()) && !match(B, m_ConstantExpr()))
2721       return BinaryOperator::CreateNot(Builder.CreateSelect(A, B, Zero));
2722 
2723     // select (select a, true, b), true, b -> select a, true, b
2724     if (match(CondVal, m_Select(m_Value(A), m_One(), m_Value(B))) &&
2725         match(TrueVal, m_One()) && match(FalseVal, m_Specific(B)))
2726       return replaceOperand(SI, 0, A);
2727     // select (select a, b, false), b, false -> select a, b, false
2728     if (match(CondVal, m_Select(m_Value(A), m_Value(B), m_Zero())) &&
2729         match(TrueVal, m_Specific(B)) && match(FalseVal, m_Zero()))
2730       return replaceOperand(SI, 0, A);
2731 
2732     if (!SelType->isVectorTy()) {
2733       if (Value *S = simplifyWithOpReplaced(TrueVal, CondVal, One, SQ,
2734                                             /* AllowRefinement */ true))
2735         return replaceOperand(SI, 1, S);
2736       if (Value *S = simplifyWithOpReplaced(FalseVal, CondVal, Zero, SQ,
2737                                             /* AllowRefinement */ true))
2738         return replaceOperand(SI, 2, S);
2739     }
2740 
2741     if (match(FalseVal, m_Zero()) || match(TrueVal, m_One())) {
2742       Use *Y = nullptr;
2743       bool IsAnd = match(FalseVal, m_Zero()) ? true : false;
2744       Value *Op1 = IsAnd ? TrueVal : FalseVal;
2745       if (isCheckForZeroAndMulWithOverflow(CondVal, Op1, IsAnd, Y)) {
2746         auto *FI = new FreezeInst(*Y, (*Y)->getName() + ".fr");
2747         InsertNewInstBefore(FI, *cast<Instruction>(Y->getUser()));
2748         replaceUse(*Y, FI);
2749         return replaceInstUsesWith(SI, Op1);
2750       }
2751 
2752       if (auto *Op1SI = dyn_cast<SelectInst>(Op1))
2753         if (auto *I = foldAndOrOfSelectUsingImpliedCond(CondVal, *Op1SI,
2754                                                         /* IsAnd */ IsAnd))
2755           return I;
2756 
2757       if (auto *ICmp0 = dyn_cast<ICmpInst>(CondVal))
2758         if (auto *ICmp1 = dyn_cast<ICmpInst>(Op1))
2759           if (auto *V = foldAndOrOfICmpsOfAndWithPow2(ICmp0, ICmp1, &SI, IsAnd,
2760                                                       /* IsLogical */ true))
2761             return replaceInstUsesWith(SI, V);
2762     }
2763 
2764     // select (select a, true, b), c, false -> select a, c, false
2765     // select c, (select a, true, b), false -> select c, a, false
2766     //   if c implies that b is false.
2767     if (match(CondVal, m_Select(m_Value(A), m_One(), m_Value(B))) &&
2768         match(FalseVal, m_Zero())) {
2769       Optional<bool> Res = isImpliedCondition(TrueVal, B, DL);
2770       if (Res && *Res == false)
2771         return replaceOperand(SI, 0, A);
2772     }
2773     if (match(TrueVal, m_Select(m_Value(A), m_One(), m_Value(B))) &&
2774         match(FalseVal, m_Zero())) {
2775       Optional<bool> Res = isImpliedCondition(CondVal, B, DL);
2776       if (Res && *Res == false)
2777         return replaceOperand(SI, 1, A);
2778     }
2779     // select c, true, (select a, b, false)  -> select c, true, a
2780     // select (select a, b, false), true, c  -> select a, true, c
2781     //   if c = false implies that b = true
2782     if (match(TrueVal, m_One()) &&
2783         match(FalseVal, m_Select(m_Value(A), m_Value(B), m_Zero()))) {
2784       Optional<bool> Res = isImpliedCondition(CondVal, B, DL, false);
2785       if (Res && *Res == true)
2786         return replaceOperand(SI, 2, A);
2787     }
2788     if (match(CondVal, m_Select(m_Value(A), m_Value(B), m_Zero())) &&
2789         match(TrueVal, m_One())) {
2790       Optional<bool> Res = isImpliedCondition(FalseVal, B, DL, false);
2791       if (Res && *Res == true)
2792         return replaceOperand(SI, 0, A);
2793     }
2794 
2795     // sel (sel c, a, false), true, (sel !c, b, false) -> sel c, a, b
2796     // sel (sel !c, a, false), true, (sel c, b, false) -> sel c, b, a
2797     Value *C1, *C2;
2798     if (match(CondVal, m_Select(m_Value(C1), m_Value(A), m_Zero())) &&
2799         match(TrueVal, m_One()) &&
2800         match(FalseVal, m_Select(m_Value(C2), m_Value(B), m_Zero()))) {
2801       if (match(C2, m_Not(m_Specific(C1)))) // first case
2802         return SelectInst::Create(C1, A, B);
2803       else if (match(C1, m_Not(m_Specific(C2)))) // second case
2804         return SelectInst::Create(C2, B, A);
2805     }
2806   }
2807 
2808   // Selecting between two integer or vector splat integer constants?
2809   //
2810   // Note that we don't handle a scalar select of vectors:
2811   // select i1 %c, <2 x i8> <1, 1>, <2 x i8> <0, 0>
2812   // because that may need 3 instructions to splat the condition value:
2813   // extend, insertelement, shufflevector.
2814   //
2815   // Do not handle i1 TrueVal and FalseVal otherwise would result in
2816   // zext/sext i1 to i1.
2817   if (SelType->isIntOrIntVectorTy() && !SelType->isIntOrIntVectorTy(1) &&
2818       CondVal->getType()->isVectorTy() == SelType->isVectorTy()) {
2819     // select C, 1, 0 -> zext C to int
2820     if (match(TrueVal, m_One()) && match(FalseVal, m_Zero()))
2821       return new ZExtInst(CondVal, SelType);
2822 
2823     // select C, -1, 0 -> sext C to int
2824     if (match(TrueVal, m_AllOnes()) && match(FalseVal, m_Zero()))
2825       return new SExtInst(CondVal, SelType);
2826 
2827     // select C, 0, 1 -> zext !C to int
2828     if (match(TrueVal, m_Zero()) && match(FalseVal, m_One())) {
2829       Value *NotCond = Builder.CreateNot(CondVal, "not." + CondVal->getName());
2830       return new ZExtInst(NotCond, SelType);
2831     }
2832 
2833     // select C, 0, -1 -> sext !C to int
2834     if (match(TrueVal, m_Zero()) && match(FalseVal, m_AllOnes())) {
2835       Value *NotCond = Builder.CreateNot(CondVal, "not." + CondVal->getName());
2836       return new SExtInst(NotCond, SelType);
2837     }
2838   }
2839 
2840   if (auto *FCmp = dyn_cast<FCmpInst>(CondVal)) {
2841     Value *Cmp0 = FCmp->getOperand(0), *Cmp1 = FCmp->getOperand(1);
2842     // Are we selecting a value based on a comparison of the two values?
2843     if ((Cmp0 == TrueVal && Cmp1 == FalseVal) ||
2844         (Cmp0 == FalseVal && Cmp1 == TrueVal)) {
2845       // Canonicalize to use ordered comparisons by swapping the select
2846       // operands.
2847       //
2848       // e.g.
2849       // (X ugt Y) ? X : Y -> (X ole Y) ? Y : X
2850       if (FCmp->hasOneUse() && FCmpInst::isUnordered(FCmp->getPredicate())) {
2851         FCmpInst::Predicate InvPred = FCmp->getInversePredicate();
2852         IRBuilder<>::FastMathFlagGuard FMFG(Builder);
2853         // FIXME: The FMF should propagate from the select, not the fcmp.
2854         Builder.setFastMathFlags(FCmp->getFastMathFlags());
2855         Value *NewCond = Builder.CreateFCmp(InvPred, Cmp0, Cmp1,
2856                                             FCmp->getName() + ".inv");
2857         Value *NewSel = Builder.CreateSelect(NewCond, FalseVal, TrueVal);
2858         return replaceInstUsesWith(SI, NewSel);
2859       }
2860 
2861       // NOTE: if we wanted to, this is where to detect MIN/MAX
2862     }
2863   }
2864 
2865   // Canonicalize select with fcmp to fabs(). -0.0 makes this tricky. We need
2866   // fast-math-flags (nsz) or fsub with +0.0 (not fneg) for this to work. We
2867   // also require nnan because we do not want to unintentionally change the
2868   // sign of a NaN value.
2869   // (X <= +/-0.0) ? (0.0 - X) : X --> fabs(X)
2870   Instruction *FSub;
2871   if (match(CondVal, m_FCmp(Pred, m_Specific(FalseVal), m_AnyZeroFP())) &&
2872       match(TrueVal, m_FSub(m_PosZeroFP(), m_Specific(FalseVal))) &&
2873       match(TrueVal, m_Instruction(FSub)) && FSub->hasNoNaNs() &&
2874       (Pred == FCmpInst::FCMP_OLE || Pred == FCmpInst::FCMP_ULE)) {
2875     Value *Fabs = Builder.CreateUnaryIntrinsic(Intrinsic::fabs, FalseVal, &SI);
2876     return replaceInstUsesWith(SI, Fabs);
2877   }
2878   // (X >  +/-0.0) ? X : (0.0 - X) --> fabs(X)
2879   if (match(CondVal, m_FCmp(Pred, m_Specific(TrueVal), m_AnyZeroFP())) &&
2880       match(FalseVal, m_FSub(m_PosZeroFP(), m_Specific(TrueVal))) &&
2881       match(FalseVal, m_Instruction(FSub)) && FSub->hasNoNaNs() &&
2882       (Pred == FCmpInst::FCMP_OGT || Pred == FCmpInst::FCMP_UGT)) {
2883     Value *Fabs = Builder.CreateUnaryIntrinsic(Intrinsic::fabs, TrueVal, &SI);
2884     return replaceInstUsesWith(SI, Fabs);
2885   }
2886   // With nnan and nsz:
2887   // (X <  +/-0.0) ? -X : X --> fabs(X)
2888   // (X <= +/-0.0) ? -X : X --> fabs(X)
2889   Instruction *FNeg;
2890   if (match(CondVal, m_FCmp(Pred, m_Specific(FalseVal), m_AnyZeroFP())) &&
2891       match(TrueVal, m_FNeg(m_Specific(FalseVal))) &&
2892       match(TrueVal, m_Instruction(FNeg)) && FNeg->hasNoNaNs() &&
2893       FNeg->hasNoSignedZeros() && SI.hasNoSignedZeros() &&
2894       (Pred == FCmpInst::FCMP_OLT || Pred == FCmpInst::FCMP_OLE ||
2895        Pred == FCmpInst::FCMP_ULT || Pred == FCmpInst::FCMP_ULE)) {
2896     Value *Fabs = Builder.CreateUnaryIntrinsic(Intrinsic::fabs, FalseVal, &SI);
2897     return replaceInstUsesWith(SI, Fabs);
2898   }
2899   // With nnan and nsz:
2900   // (X >  +/-0.0) ? X : -X --> fabs(X)
2901   // (X >= +/-0.0) ? X : -X --> fabs(X)
2902   if (match(CondVal, m_FCmp(Pred, m_Specific(TrueVal), m_AnyZeroFP())) &&
2903       match(FalseVal, m_FNeg(m_Specific(TrueVal))) &&
2904       match(FalseVal, m_Instruction(FNeg)) && FNeg->hasNoNaNs() &&
2905       FNeg->hasNoSignedZeros() && SI.hasNoSignedZeros() &&
2906       (Pred == FCmpInst::FCMP_OGT || Pred == FCmpInst::FCMP_OGE ||
2907        Pred == FCmpInst::FCMP_UGT || Pred == FCmpInst::FCMP_UGE)) {
2908     Value *Fabs = Builder.CreateUnaryIntrinsic(Intrinsic::fabs, TrueVal, &SI);
2909     return replaceInstUsesWith(SI, Fabs);
2910   }
2911 
2912   // See if we are selecting two values based on a comparison of the two values.
2913   if (ICmpInst *ICI = dyn_cast<ICmpInst>(CondVal))
2914     if (Instruction *Result = foldSelectInstWithICmp(SI, ICI))
2915       return Result;
2916 
2917   if (Instruction *Add = foldAddSubSelect(SI, Builder))
2918     return Add;
2919   if (Instruction *Add = foldOverflowingAddSubSelect(SI, Builder))
2920     return Add;
2921   if (Instruction *Or = foldSetClearBits(SI, Builder))
2922     return Or;
2923 
2924   // Turn (select C, (op X, Y), (op X, Z)) -> (op X, (select C, Y, Z))
2925   auto *TI = dyn_cast<Instruction>(TrueVal);
2926   auto *FI = dyn_cast<Instruction>(FalseVal);
2927   if (TI && FI && TI->getOpcode() == FI->getOpcode())
2928     if (Instruction *IV = foldSelectOpOp(SI, TI, FI))
2929       return IV;
2930 
2931   if (Instruction *I = foldSelectExtConst(SI))
2932     return I;
2933 
2934   // Fold (select C, (gep Ptr, Idx), Ptr) -> (gep Ptr, (select C, Idx, 0))
2935   // Fold (select C, Ptr, (gep Ptr, Idx)) -> (gep Ptr, (select C, 0, Idx))
2936   auto SelectGepWithBase = [&](GetElementPtrInst *Gep, Value *Base,
2937                                bool Swap) -> GetElementPtrInst * {
2938     Value *Ptr = Gep->getPointerOperand();
2939     if (Gep->getNumOperands() != 2 || Gep->getPointerOperand() != Base ||
2940         !Gep->hasOneUse())
2941       return nullptr;
2942     Type *ElementType = Gep->getResultElementType();
2943     Value *Idx = Gep->getOperand(1);
2944     Value *NewT = Idx;
2945     Value *NewF = Constant::getNullValue(Idx->getType());
2946     if (Swap)
2947       std::swap(NewT, NewF);
2948     Value *NewSI =
2949         Builder.CreateSelect(CondVal, NewT, NewF, SI.getName() + ".idx", &SI);
2950     return GetElementPtrInst::Create(ElementType, Ptr, {NewSI});
2951   };
2952   if (auto *TrueGep = dyn_cast<GetElementPtrInst>(TrueVal))
2953     if (auto *NewGep = SelectGepWithBase(TrueGep, FalseVal, false))
2954       return NewGep;
2955   if (auto *FalseGep = dyn_cast<GetElementPtrInst>(FalseVal))
2956     if (auto *NewGep = SelectGepWithBase(FalseGep, TrueVal, true))
2957       return NewGep;
2958 
2959   // See if we can fold the select into one of our operands.
2960   if (SelType->isIntOrIntVectorTy() || SelType->isFPOrFPVectorTy()) {
2961     if (Instruction *FoldI = foldSelectIntoOp(SI, TrueVal, FalseVal))
2962       return FoldI;
2963 
2964     Value *LHS, *RHS;
2965     Instruction::CastOps CastOp;
2966     SelectPatternResult SPR = matchSelectPattern(&SI, LHS, RHS, &CastOp);
2967     auto SPF = SPR.Flavor;
2968     if (SPF) {
2969       Value *LHS2, *RHS2;
2970       if (SelectPatternFlavor SPF2 = matchSelectPattern(LHS, LHS2, RHS2).Flavor)
2971         if (Instruction *R = foldSPFofSPF(cast<Instruction>(LHS), SPF2, LHS2,
2972                                           RHS2, SI, SPF, RHS))
2973           return R;
2974       if (SelectPatternFlavor SPF2 = matchSelectPattern(RHS, LHS2, RHS2).Flavor)
2975         if (Instruction *R = foldSPFofSPF(cast<Instruction>(RHS), SPF2, LHS2,
2976                                           RHS2, SI, SPF, LHS))
2977           return R;
2978       // TODO.
2979       // ABS(-X) -> ABS(X)
2980     }
2981 
2982     if (SelectPatternResult::isMinOrMax(SPF)) {
2983       // Canonicalize so that
2984       // - type casts are outside select patterns.
2985       // - float clamp is transformed to min/max pattern
2986 
2987       bool IsCastNeeded = LHS->getType() != SelType;
2988       Value *CmpLHS = cast<CmpInst>(CondVal)->getOperand(0);
2989       Value *CmpRHS = cast<CmpInst>(CondVal)->getOperand(1);
2990       if (IsCastNeeded ||
2991           (LHS->getType()->isFPOrFPVectorTy() &&
2992            ((CmpLHS != LHS && CmpLHS != RHS) ||
2993             (CmpRHS != LHS && CmpRHS != RHS)))) {
2994         CmpInst::Predicate MinMaxPred = getMinMaxPred(SPF, SPR.Ordered);
2995 
2996         Value *Cmp;
2997         if (CmpInst::isIntPredicate(MinMaxPred)) {
2998           Cmp = Builder.CreateICmp(MinMaxPred, LHS, RHS);
2999         } else {
3000           IRBuilder<>::FastMathFlagGuard FMFG(Builder);
3001           auto FMF =
3002               cast<FPMathOperator>(SI.getCondition())->getFastMathFlags();
3003           Builder.setFastMathFlags(FMF);
3004           Cmp = Builder.CreateFCmp(MinMaxPred, LHS, RHS);
3005         }
3006 
3007         Value *NewSI = Builder.CreateSelect(Cmp, LHS, RHS, SI.getName(), &SI);
3008         if (!IsCastNeeded)
3009           return replaceInstUsesWith(SI, NewSI);
3010 
3011         Value *NewCast = Builder.CreateCast(CastOp, NewSI, SelType);
3012         return replaceInstUsesWith(SI, NewCast);
3013       }
3014 
3015       // MAX(~a, ~b) -> ~MIN(a, b)
3016       // MAX(~a, C)  -> ~MIN(a, ~C)
3017       // MIN(~a, ~b) -> ~MAX(a, b)
3018       // MIN(~a, C)  -> ~MAX(a, ~C)
3019       auto moveNotAfterMinMax = [&](Value *X, Value *Y) -> Instruction * {
3020         Value *A;
3021         if (match(X, m_Not(m_Value(A))) && !X->hasNUsesOrMore(3) &&
3022             !isFreeToInvert(A, A->hasOneUse()) &&
3023             // Passing false to only consider m_Not and constants.
3024             isFreeToInvert(Y, false)) {
3025           Value *B = Builder.CreateNot(Y);
3026           Value *NewMinMax = createMinMax(Builder, getInverseMinMaxFlavor(SPF),
3027                                           A, B);
3028           // Copy the profile metadata.
3029           if (MDNode *MD = SI.getMetadata(LLVMContext::MD_prof)) {
3030             cast<SelectInst>(NewMinMax)->setMetadata(LLVMContext::MD_prof, MD);
3031             // Swap the metadata if the operands are swapped.
3032             if (X == SI.getFalseValue() && Y == SI.getTrueValue())
3033               cast<SelectInst>(NewMinMax)->swapProfMetadata();
3034           }
3035 
3036           return BinaryOperator::CreateNot(NewMinMax);
3037         }
3038 
3039         return nullptr;
3040       };
3041 
3042       if (Instruction *I = moveNotAfterMinMax(LHS, RHS))
3043         return I;
3044       if (Instruction *I = moveNotAfterMinMax(RHS, LHS))
3045         return I;
3046 
3047       if (Instruction *I = moveAddAfterMinMax(SPF, LHS, RHS, Builder))
3048         return I;
3049 
3050       if (Instruction *I = factorizeMinMaxTree(SPF, LHS, RHS, Builder))
3051         return I;
3052       if (Instruction *I = matchSAddSubSat(SI))
3053         return I;
3054     }
3055   }
3056 
3057   // Canonicalize select of FP values where NaN and -0.0 are not valid as
3058   // minnum/maxnum intrinsics.
3059   if (isa<FPMathOperator>(SI) && SI.hasNoNaNs() && SI.hasNoSignedZeros()) {
3060     Value *X, *Y;
3061     if (match(&SI, m_OrdFMax(m_Value(X), m_Value(Y))))
3062       return replaceInstUsesWith(
3063           SI, Builder.CreateBinaryIntrinsic(Intrinsic::maxnum, X, Y, &SI));
3064 
3065     if (match(&SI, m_OrdFMin(m_Value(X), m_Value(Y))))
3066       return replaceInstUsesWith(
3067           SI, Builder.CreateBinaryIntrinsic(Intrinsic::minnum, X, Y, &SI));
3068   }
3069 
3070   // See if we can fold the select into a phi node if the condition is a select.
3071   if (auto *PN = dyn_cast<PHINode>(SI.getCondition()))
3072     // The true/false values have to be live in the PHI predecessor's blocks.
3073     if (canSelectOperandBeMappingIntoPredBlock(TrueVal, SI) &&
3074         canSelectOperandBeMappingIntoPredBlock(FalseVal, SI))
3075       if (Instruction *NV = foldOpIntoPhi(SI, PN))
3076         return NV;
3077 
3078   if (SelectInst *TrueSI = dyn_cast<SelectInst>(TrueVal)) {
3079     if (TrueSI->getCondition()->getType() == CondVal->getType()) {
3080       // select(C, select(C, a, b), c) -> select(C, a, c)
3081       if (TrueSI->getCondition() == CondVal) {
3082         if (SI.getTrueValue() == TrueSI->getTrueValue())
3083           return nullptr;
3084         return replaceOperand(SI, 1, TrueSI->getTrueValue());
3085       }
3086       // select(C0, select(C1, a, b), b) -> select(C0&C1, a, b)
3087       // We choose this as normal form to enable folding on the And and
3088       // shortening paths for the values (this helps getUnderlyingObjects() for
3089       // example).
3090       if (TrueSI->getFalseValue() == FalseVal && TrueSI->hasOneUse()) {
3091         Value *And = Builder.CreateLogicalAnd(CondVal, TrueSI->getCondition());
3092         replaceOperand(SI, 0, And);
3093         replaceOperand(SI, 1, TrueSI->getTrueValue());
3094         return &SI;
3095       }
3096     }
3097   }
3098   if (SelectInst *FalseSI = dyn_cast<SelectInst>(FalseVal)) {
3099     if (FalseSI->getCondition()->getType() == CondVal->getType()) {
3100       // select(C, a, select(C, b, c)) -> select(C, a, c)
3101       if (FalseSI->getCondition() == CondVal) {
3102         if (SI.getFalseValue() == FalseSI->getFalseValue())
3103           return nullptr;
3104         return replaceOperand(SI, 2, FalseSI->getFalseValue());
3105       }
3106       // select(C0, a, select(C1, a, b)) -> select(C0|C1, a, b)
3107       if (FalseSI->getTrueValue() == TrueVal && FalseSI->hasOneUse()) {
3108         Value *Or = Builder.CreateLogicalOr(CondVal, FalseSI->getCondition());
3109         replaceOperand(SI, 0, Or);
3110         replaceOperand(SI, 2, FalseSI->getFalseValue());
3111         return &SI;
3112       }
3113     }
3114   }
3115 
3116   auto canMergeSelectThroughBinop = [](BinaryOperator *BO) {
3117     // The select might be preventing a division by 0.
3118     switch (BO->getOpcode()) {
3119     default:
3120       return true;
3121     case Instruction::SRem:
3122     case Instruction::URem:
3123     case Instruction::SDiv:
3124     case Instruction::UDiv:
3125       return false;
3126     }
3127   };
3128 
3129   // Try to simplify a binop sandwiched between 2 selects with the same
3130   // condition.
3131   // select(C, binop(select(C, X, Y), W), Z) -> select(C, binop(X, W), Z)
3132   BinaryOperator *TrueBO;
3133   if (match(TrueVal, m_OneUse(m_BinOp(TrueBO))) &&
3134       canMergeSelectThroughBinop(TrueBO)) {
3135     if (auto *TrueBOSI = dyn_cast<SelectInst>(TrueBO->getOperand(0))) {
3136       if (TrueBOSI->getCondition() == CondVal) {
3137         replaceOperand(*TrueBO, 0, TrueBOSI->getTrueValue());
3138         Worklist.push(TrueBO);
3139         return &SI;
3140       }
3141     }
3142     if (auto *TrueBOSI = dyn_cast<SelectInst>(TrueBO->getOperand(1))) {
3143       if (TrueBOSI->getCondition() == CondVal) {
3144         replaceOperand(*TrueBO, 1, TrueBOSI->getTrueValue());
3145         Worklist.push(TrueBO);
3146         return &SI;
3147       }
3148     }
3149   }
3150 
3151   // select(C, Z, binop(select(C, X, Y), W)) -> select(C, Z, binop(Y, W))
3152   BinaryOperator *FalseBO;
3153   if (match(FalseVal, m_OneUse(m_BinOp(FalseBO))) &&
3154       canMergeSelectThroughBinop(FalseBO)) {
3155     if (auto *FalseBOSI = dyn_cast<SelectInst>(FalseBO->getOperand(0))) {
3156       if (FalseBOSI->getCondition() == CondVal) {
3157         replaceOperand(*FalseBO, 0, FalseBOSI->getFalseValue());
3158         Worklist.push(FalseBO);
3159         return &SI;
3160       }
3161     }
3162     if (auto *FalseBOSI = dyn_cast<SelectInst>(FalseBO->getOperand(1))) {
3163       if (FalseBOSI->getCondition() == CondVal) {
3164         replaceOperand(*FalseBO, 1, FalseBOSI->getFalseValue());
3165         Worklist.push(FalseBO);
3166         return &SI;
3167       }
3168     }
3169   }
3170 
3171   Value *NotCond;
3172   if (match(CondVal, m_Not(m_Value(NotCond))) &&
3173       !InstCombiner::shouldAvoidAbsorbingNotIntoSelect(SI)) {
3174     replaceOperand(SI, 0, NotCond);
3175     SI.swapValues();
3176     SI.swapProfMetadata();
3177     return &SI;
3178   }
3179 
3180   if (Instruction *I = foldVectorSelect(SI))
3181     return I;
3182 
3183   // If we can compute the condition, there's no need for a select.
3184   // Like the above fold, we are attempting to reduce compile-time cost by
3185   // putting this fold here with limitations rather than in InstSimplify.
3186   // The motivation for this call into value tracking is to take advantage of
3187   // the assumption cache, so make sure that is populated.
3188   if (!CondVal->getType()->isVectorTy() && !AC.assumptions().empty()) {
3189     KnownBits Known(1);
3190     computeKnownBits(CondVal, Known, 0, &SI);
3191     if (Known.One.isOneValue())
3192       return replaceInstUsesWith(SI, TrueVal);
3193     if (Known.Zero.isOneValue())
3194       return replaceInstUsesWith(SI, FalseVal);
3195   }
3196 
3197   if (Instruction *BitCastSel = foldSelectCmpBitcasts(SI, Builder))
3198     return BitCastSel;
3199 
3200   // Simplify selects that test the returned flag of cmpxchg instructions.
3201   if (Value *V = foldSelectCmpXchg(SI))
3202     return replaceInstUsesWith(SI, V);
3203 
3204   if (Instruction *Select = foldSelectBinOpIdentity(SI, TLI, *this))
3205     return Select;
3206 
3207   if (Instruction *Funnel = foldSelectFunnelShift(SI, Builder))
3208     return Funnel;
3209 
3210   if (Instruction *Copysign = foldSelectToCopysign(SI, Builder))
3211     return Copysign;
3212 
3213   if (Instruction *PN = foldSelectToPhi(SI, DT, Builder))
3214     return replaceInstUsesWith(SI, PN);
3215 
3216   if (Value *Fr = foldSelectWithFrozenICmp(SI, Builder))
3217     return replaceInstUsesWith(SI, Fr);
3218 
3219   // select(mask, mload(,,mask,0), 0) -> mload(,,mask,0)
3220   // Load inst is intentionally not checked for hasOneUse()
3221   if (match(FalseVal, m_Zero()) &&
3222       match(TrueVal, m_MaskedLoad(m_Value(), m_Value(), m_Specific(CondVal),
3223                                   m_CombineOr(m_Undef(), m_Zero())))) {
3224     auto *MaskedLoad = cast<IntrinsicInst>(TrueVal);
3225     if (isa<UndefValue>(MaskedLoad->getArgOperand(3)))
3226       MaskedLoad->setArgOperand(3, FalseVal /* Zero */);
3227     return replaceInstUsesWith(SI, MaskedLoad);
3228   }
3229 
3230   Value *Mask;
3231   if (match(TrueVal, m_Zero()) &&
3232       match(FalseVal, m_MaskedLoad(m_Value(), m_Value(), m_Value(Mask),
3233                                    m_CombineOr(m_Undef(), m_Zero()))) &&
3234       (CondVal->getType() == Mask->getType())) {
3235     // We can remove the select by ensuring the load zeros all lanes the
3236     // select would have.  We determine this by proving there is no overlap
3237     // between the load and select masks.
3238     // (i.e (load_mask & select_mask) == 0 == no overlap)
3239     bool CanMergeSelectIntoLoad = false;
3240     if (Value *V = SimplifyAndInst(CondVal, Mask, SQ.getWithInstruction(&SI)))
3241       CanMergeSelectIntoLoad = match(V, m_Zero());
3242 
3243     if (CanMergeSelectIntoLoad) {
3244       auto *MaskedLoad = cast<IntrinsicInst>(FalseVal);
3245       if (isa<UndefValue>(MaskedLoad->getArgOperand(3)))
3246         MaskedLoad->setArgOperand(3, TrueVal /* Zero */);
3247       return replaceInstUsesWith(SI, MaskedLoad);
3248     }
3249   }
3250 
3251   return nullptr;
3252 }
3253