xref: /freebsd/contrib/llvm-project/llvm/lib/Transforms/InstCombine/InstCombineMulDivRem.cpp (revision 0fca6ea1d4eea4c934cfff25ac9ee8ad6fe95583)
1 //===- InstCombineMulDivRem.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 visit functions for mul, fmul, sdiv, udiv, fdiv,
10 // srem, urem, frem.
11 //
12 //===----------------------------------------------------------------------===//
13 
14 #include "InstCombineInternal.h"
15 #include "llvm/ADT/APInt.h"
16 #include "llvm/ADT/SmallVector.h"
17 #include "llvm/Analysis/InstructionSimplify.h"
18 #include "llvm/Analysis/ValueTracking.h"
19 #include "llvm/IR/BasicBlock.h"
20 #include "llvm/IR/Constant.h"
21 #include "llvm/IR/Constants.h"
22 #include "llvm/IR/InstrTypes.h"
23 #include "llvm/IR/Instruction.h"
24 #include "llvm/IR/Instructions.h"
25 #include "llvm/IR/IntrinsicInst.h"
26 #include "llvm/IR/Intrinsics.h"
27 #include "llvm/IR/Operator.h"
28 #include "llvm/IR/PatternMatch.h"
29 #include "llvm/IR/Type.h"
30 #include "llvm/IR/Value.h"
31 #include "llvm/Support/Casting.h"
32 #include "llvm/Support/ErrorHandling.h"
33 #include "llvm/Transforms/InstCombine/InstCombiner.h"
34 #include "llvm/Transforms/Utils/BuildLibCalls.h"
35 #include <cassert>
36 
37 #define DEBUG_TYPE "instcombine"
38 #include "llvm/Transforms/Utils/InstructionWorklist.h"
39 
40 using namespace llvm;
41 using namespace PatternMatch;
42 
43 /// The specific integer value is used in a context where it is known to be
44 /// non-zero.  If this allows us to simplify the computation, do so and return
45 /// the new operand, otherwise return null.
simplifyValueKnownNonZero(Value * V,InstCombinerImpl & IC,Instruction & CxtI)46 static Value *simplifyValueKnownNonZero(Value *V, InstCombinerImpl &IC,
47                                         Instruction &CxtI) {
48   // If V has multiple uses, then we would have to do more analysis to determine
49   // if this is safe.  For example, the use could be in dynamically unreached
50   // code.
51   if (!V->hasOneUse()) return nullptr;
52 
53   bool MadeChange = false;
54 
55   // ((1 << A) >>u B) --> (1 << (A-B))
56   // Because V cannot be zero, we know that B is less than A.
57   Value *A = nullptr, *B = nullptr, *One = nullptr;
58   if (match(V, m_LShr(m_OneUse(m_Shl(m_Value(One), m_Value(A))), m_Value(B))) &&
59       match(One, m_One())) {
60     A = IC.Builder.CreateSub(A, B);
61     return IC.Builder.CreateShl(One, A);
62   }
63 
64   // (PowerOfTwo >>u B) --> isExact since shifting out the result would make it
65   // inexact.  Similarly for <<.
66   BinaryOperator *I = dyn_cast<BinaryOperator>(V);
67   if (I && I->isLogicalShift() &&
68       IC.isKnownToBeAPowerOfTwo(I->getOperand(0), false, 0, &CxtI)) {
69     // We know that this is an exact/nuw shift and that the input is a
70     // non-zero context as well.
71     if (Value *V2 = simplifyValueKnownNonZero(I->getOperand(0), IC, CxtI)) {
72       IC.replaceOperand(*I, 0, V2);
73       MadeChange = true;
74     }
75 
76     if (I->getOpcode() == Instruction::LShr && !I->isExact()) {
77       I->setIsExact();
78       MadeChange = true;
79     }
80 
81     if (I->getOpcode() == Instruction::Shl && !I->hasNoUnsignedWrap()) {
82       I->setHasNoUnsignedWrap();
83       MadeChange = true;
84     }
85   }
86 
87   // TODO: Lots more we could do here:
88   //    If V is a phi node, we can call this on each of its operands.
89   //    "select cond, X, 0" can simplify to "X".
90 
91   return MadeChange ? V : nullptr;
92 }
93 
94 // TODO: This is a specific form of a much more general pattern.
95 //       We could detect a select with any binop identity constant, or we
96 //       could use SimplifyBinOp to see if either arm of the select reduces.
97 //       But that needs to be done carefully and/or while removing potential
98 //       reverse canonicalizations as in InstCombiner::foldSelectIntoOp().
foldMulSelectToNegate(BinaryOperator & I,InstCombiner::BuilderTy & Builder)99 static Value *foldMulSelectToNegate(BinaryOperator &I,
100                                     InstCombiner::BuilderTy &Builder) {
101   Value *Cond, *OtherOp;
102 
103   // mul (select Cond, 1, -1), OtherOp --> select Cond, OtherOp, -OtherOp
104   // mul OtherOp, (select Cond, 1, -1) --> select Cond, OtherOp, -OtherOp
105   if (match(&I, m_c_Mul(m_OneUse(m_Select(m_Value(Cond), m_One(), m_AllOnes())),
106                         m_Value(OtherOp)))) {
107     bool HasAnyNoWrap = I.hasNoSignedWrap() || I.hasNoUnsignedWrap();
108     Value *Neg = Builder.CreateNeg(OtherOp, "", HasAnyNoWrap);
109     return Builder.CreateSelect(Cond, OtherOp, Neg);
110   }
111   // mul (select Cond, -1, 1), OtherOp --> select Cond, -OtherOp, OtherOp
112   // mul OtherOp, (select Cond, -1, 1) --> select Cond, -OtherOp, OtherOp
113   if (match(&I, m_c_Mul(m_OneUse(m_Select(m_Value(Cond), m_AllOnes(), m_One())),
114                         m_Value(OtherOp)))) {
115     bool HasAnyNoWrap = I.hasNoSignedWrap() || I.hasNoUnsignedWrap();
116     Value *Neg = Builder.CreateNeg(OtherOp, "", HasAnyNoWrap);
117     return Builder.CreateSelect(Cond, Neg, OtherOp);
118   }
119 
120   // fmul (select Cond, 1.0, -1.0), OtherOp --> select Cond, OtherOp, -OtherOp
121   // fmul OtherOp, (select Cond, 1.0, -1.0) --> select Cond, OtherOp, -OtherOp
122   if (match(&I, m_c_FMul(m_OneUse(m_Select(m_Value(Cond), m_SpecificFP(1.0),
123                                            m_SpecificFP(-1.0))),
124                          m_Value(OtherOp)))) {
125     IRBuilder<>::FastMathFlagGuard FMFGuard(Builder);
126     Builder.setFastMathFlags(I.getFastMathFlags());
127     return Builder.CreateSelect(Cond, OtherOp, Builder.CreateFNeg(OtherOp));
128   }
129 
130   // fmul (select Cond, -1.0, 1.0), OtherOp --> select Cond, -OtherOp, OtherOp
131   // fmul OtherOp, (select Cond, -1.0, 1.0) --> select Cond, -OtherOp, OtherOp
132   if (match(&I, m_c_FMul(m_OneUse(m_Select(m_Value(Cond), m_SpecificFP(-1.0),
133                                            m_SpecificFP(1.0))),
134                          m_Value(OtherOp)))) {
135     IRBuilder<>::FastMathFlagGuard FMFGuard(Builder);
136     Builder.setFastMathFlags(I.getFastMathFlags());
137     return Builder.CreateSelect(Cond, Builder.CreateFNeg(OtherOp), OtherOp);
138   }
139 
140   return nullptr;
141 }
142 
143 /// Reduce integer multiplication patterns that contain a (+/-1 << Z) factor.
144 /// Callers are expected to call this twice to handle commuted patterns.
foldMulShl1(BinaryOperator & Mul,bool CommuteOperands,InstCombiner::BuilderTy & Builder)145 static Value *foldMulShl1(BinaryOperator &Mul, bool CommuteOperands,
146                           InstCombiner::BuilderTy &Builder) {
147   Value *X = Mul.getOperand(0), *Y = Mul.getOperand(1);
148   if (CommuteOperands)
149     std::swap(X, Y);
150 
151   const bool HasNSW = Mul.hasNoSignedWrap();
152   const bool HasNUW = Mul.hasNoUnsignedWrap();
153 
154   // X * (1 << Z) --> X << Z
155   Value *Z;
156   if (match(Y, m_Shl(m_One(), m_Value(Z)))) {
157     bool PropagateNSW = HasNSW && cast<ShlOperator>(Y)->hasNoSignedWrap();
158     return Builder.CreateShl(X, Z, Mul.getName(), HasNUW, PropagateNSW);
159   }
160 
161   // Similar to above, but an increment of the shifted value becomes an add:
162   // X * ((1 << Z) + 1) --> (X * (1 << Z)) + X --> (X << Z) + X
163   // This increases uses of X, so it may require a freeze, but that is still
164   // expected to be an improvement because it removes the multiply.
165   BinaryOperator *Shift;
166   if (match(Y, m_OneUse(m_Add(m_BinOp(Shift), m_One()))) &&
167       match(Shift, m_OneUse(m_Shl(m_One(), m_Value(Z))))) {
168     bool PropagateNSW = HasNSW && Shift->hasNoSignedWrap();
169     Value *FrX = X;
170     if (!isGuaranteedNotToBeUndef(X))
171       FrX = Builder.CreateFreeze(X, X->getName() + ".fr");
172     Value *Shl = Builder.CreateShl(FrX, Z, "mulshl", HasNUW, PropagateNSW);
173     return Builder.CreateAdd(Shl, FrX, Mul.getName(), HasNUW, PropagateNSW);
174   }
175 
176   // Similar to above, but a decrement of the shifted value is disguised as
177   // 'not' and becomes a sub:
178   // X * (~(-1 << Z)) --> X * ((1 << Z) - 1) --> (X << Z) - X
179   // This increases uses of X, so it may require a freeze, but that is still
180   // expected to be an improvement because it removes the multiply.
181   if (match(Y, m_OneUse(m_Not(m_OneUse(m_Shl(m_AllOnes(), m_Value(Z))))))) {
182     Value *FrX = X;
183     if (!isGuaranteedNotToBeUndef(X))
184       FrX = Builder.CreateFreeze(X, X->getName() + ".fr");
185     Value *Shl = Builder.CreateShl(FrX, Z, "mulshl");
186     return Builder.CreateSub(Shl, FrX, Mul.getName());
187   }
188 
189   return nullptr;
190 }
191 
192 static Value *takeLog2(IRBuilderBase &Builder, Value *Op, unsigned Depth,
193                        bool AssumeNonZero, bool DoFold);
194 
visitMul(BinaryOperator & I)195 Instruction *InstCombinerImpl::visitMul(BinaryOperator &I) {
196   Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
197   if (Value *V =
198           simplifyMulInst(Op0, Op1, I.hasNoSignedWrap(), I.hasNoUnsignedWrap(),
199                           SQ.getWithInstruction(&I)))
200     return replaceInstUsesWith(I, V);
201 
202   if (SimplifyAssociativeOrCommutative(I))
203     return &I;
204 
205   if (Instruction *X = foldVectorBinop(I))
206     return X;
207 
208   if (Instruction *Phi = foldBinopWithPhiOperands(I))
209     return Phi;
210 
211   if (Value *V = foldUsingDistributiveLaws(I))
212     return replaceInstUsesWith(I, V);
213 
214   Type *Ty = I.getType();
215   const unsigned BitWidth = Ty->getScalarSizeInBits();
216   const bool HasNSW = I.hasNoSignedWrap();
217   const bool HasNUW = I.hasNoUnsignedWrap();
218 
219   // X * -1 --> 0 - X
220   if (match(Op1, m_AllOnes())) {
221     return HasNSW ? BinaryOperator::CreateNSWNeg(Op0)
222                   : BinaryOperator::CreateNeg(Op0);
223   }
224 
225   // Also allow combining multiply instructions on vectors.
226   {
227     Value *NewOp;
228     Constant *C1, *C2;
229     const APInt *IVal;
230     if (match(&I, m_Mul(m_Shl(m_Value(NewOp), m_ImmConstant(C2)),
231                         m_ImmConstant(C1))) &&
232         match(C1, m_APInt(IVal))) {
233       // ((X << C2)*C1) == (X * (C1 << C2))
234       Constant *Shl =
235           ConstantFoldBinaryOpOperands(Instruction::Shl, C1, C2, DL);
236       assert(Shl && "Constant folding of immediate constants failed");
237       BinaryOperator *Mul = cast<BinaryOperator>(I.getOperand(0));
238       BinaryOperator *BO = BinaryOperator::CreateMul(NewOp, Shl);
239       if (HasNUW && Mul->hasNoUnsignedWrap())
240         BO->setHasNoUnsignedWrap();
241       if (HasNSW && Mul->hasNoSignedWrap() && Shl->isNotMinSignedValue())
242         BO->setHasNoSignedWrap();
243       return BO;
244     }
245 
246     if (match(&I, m_Mul(m_Value(NewOp), m_Constant(C1)))) {
247       // Replace X*(2^C) with X << C, where C is either a scalar or a vector.
248       if (Constant *NewCst = ConstantExpr::getExactLogBase2(C1)) {
249         BinaryOperator *Shl = BinaryOperator::CreateShl(NewOp, NewCst);
250 
251         if (HasNUW)
252           Shl->setHasNoUnsignedWrap();
253         if (HasNSW) {
254           const APInt *V;
255           if (match(NewCst, m_APInt(V)) && *V != V->getBitWidth() - 1)
256             Shl->setHasNoSignedWrap();
257         }
258 
259         return Shl;
260       }
261     }
262   }
263 
264   if (Op0->hasOneUse() && match(Op1, m_NegatedPower2())) {
265     // Interpret  X * (-1<<C)  as  (-X) * (1<<C)  and try to sink the negation.
266     // The "* (1<<C)" thus becomes a potential shifting opportunity.
267     if (Value *NegOp0 =
268             Negator::Negate(/*IsNegation*/ true, HasNSW, Op0, *this)) {
269       auto *Op1C = cast<Constant>(Op1);
270       return replaceInstUsesWith(
271           I, Builder.CreateMul(NegOp0, ConstantExpr::getNeg(Op1C), "",
272                                /* HasNUW */ false,
273                                HasNSW && Op1C->isNotMinSignedValue()));
274     }
275 
276     // Try to convert multiply of extended operand to narrow negate and shift
277     // for better analysis.
278     // This is valid if the shift amount (trailing zeros in the multiplier
279     // constant) clears more high bits than the bitwidth difference between
280     // source and destination types:
281     // ({z/s}ext X) * (-1<<C) --> (zext (-X)) << C
282     const APInt *NegPow2C;
283     Value *X;
284     if (match(Op0, m_ZExtOrSExt(m_Value(X))) &&
285         match(Op1, m_APIntAllowPoison(NegPow2C))) {
286       unsigned SrcWidth = X->getType()->getScalarSizeInBits();
287       unsigned ShiftAmt = NegPow2C->countr_zero();
288       if (ShiftAmt >= BitWidth - SrcWidth) {
289         Value *N = Builder.CreateNeg(X, X->getName() + ".neg");
290         Value *Z = Builder.CreateZExt(N, Ty, N->getName() + ".z");
291         return BinaryOperator::CreateShl(Z, ConstantInt::get(Ty, ShiftAmt));
292       }
293     }
294   }
295 
296   if (Instruction *FoldedMul = foldBinOpIntoSelectOrPhi(I))
297     return FoldedMul;
298 
299   if (Value *FoldedMul = foldMulSelectToNegate(I, Builder))
300     return replaceInstUsesWith(I, FoldedMul);
301 
302   // Simplify mul instructions with a constant RHS.
303   Constant *MulC;
304   if (match(Op1, m_ImmConstant(MulC))) {
305     // Canonicalize (X+C1)*MulC -> X*MulC+C1*MulC.
306     // Canonicalize (X|C1)*MulC -> X*MulC+C1*MulC.
307     Value *X;
308     Constant *C1;
309     if (match(Op0, m_OneUse(m_AddLike(m_Value(X), m_ImmConstant(C1))))) {
310       // C1*MulC simplifies to a tidier constant.
311       Value *NewC = Builder.CreateMul(C1, MulC);
312       auto *BOp0 = cast<BinaryOperator>(Op0);
313       bool Op0NUW =
314           (BOp0->getOpcode() == Instruction::Or || BOp0->hasNoUnsignedWrap());
315       Value *NewMul = Builder.CreateMul(X, MulC);
316       auto *BO = BinaryOperator::CreateAdd(NewMul, NewC);
317       if (HasNUW && Op0NUW) {
318         // If NewMulBO is constant we also can set BO to nuw.
319         if (auto *NewMulBO = dyn_cast<BinaryOperator>(NewMul))
320           NewMulBO->setHasNoUnsignedWrap();
321         BO->setHasNoUnsignedWrap();
322       }
323       return BO;
324     }
325   }
326 
327   // abs(X) * abs(X) -> X * X
328   Value *X;
329   if (Op0 == Op1 && match(Op0, m_Intrinsic<Intrinsic::abs>(m_Value(X))))
330     return BinaryOperator::CreateMul(X, X);
331 
332   {
333     Value *Y;
334     // abs(X) * abs(Y) -> abs(X * Y)
335     if (I.hasNoSignedWrap() &&
336         match(Op0,
337               m_OneUse(m_Intrinsic<Intrinsic::abs>(m_Value(X), m_One()))) &&
338         match(Op1, m_OneUse(m_Intrinsic<Intrinsic::abs>(m_Value(Y), m_One()))))
339       return replaceInstUsesWith(
340           I, Builder.CreateBinaryIntrinsic(Intrinsic::abs,
341                                            Builder.CreateNSWMul(X, Y),
342                                            Builder.getTrue()));
343   }
344 
345   // -X * C --> X * -C
346   Value *Y;
347   Constant *Op1C;
348   if (match(Op0, m_Neg(m_Value(X))) && match(Op1, m_Constant(Op1C)))
349     return BinaryOperator::CreateMul(X, ConstantExpr::getNeg(Op1C));
350 
351   // -X * -Y --> X * Y
352   if (match(Op0, m_Neg(m_Value(X))) && match(Op1, m_Neg(m_Value(Y)))) {
353     auto *NewMul = BinaryOperator::CreateMul(X, Y);
354     if (HasNSW && cast<OverflowingBinaryOperator>(Op0)->hasNoSignedWrap() &&
355         cast<OverflowingBinaryOperator>(Op1)->hasNoSignedWrap())
356       NewMul->setHasNoSignedWrap();
357     return NewMul;
358   }
359 
360   // -X * Y --> -(X * Y)
361   // X * -Y --> -(X * Y)
362   if (match(&I, m_c_Mul(m_OneUse(m_Neg(m_Value(X))), m_Value(Y))))
363     return BinaryOperator::CreateNeg(Builder.CreateMul(X, Y));
364 
365   // (-X * Y) * -X --> (X * Y) * X
366   // (-X << Y) * -X --> (X << Y) * X
367   if (match(Op1, m_Neg(m_Value(X)))) {
368     if (Value *NegOp0 = Negator::Negate(false, /*IsNSW*/ false, Op0, *this))
369       return BinaryOperator::CreateMul(NegOp0, X);
370   }
371 
372   if (Op0->hasOneUse()) {
373     // (mul (div exact X, C0), C1)
374     //    -> (div exact X, C0 / C1)
375     // iff C0 % C1 == 0 and X / (C0 / C1) doesn't create UB.
376     const APInt *C1;
377     auto UDivCheck = [&C1](const APInt &C) { return C.urem(*C1).isZero(); };
378     auto SDivCheck = [&C1](const APInt &C) {
379       APInt Quot, Rem;
380       APInt::sdivrem(C, *C1, Quot, Rem);
381       return Rem.isZero() && !Quot.isAllOnes();
382     };
383     if (match(Op1, m_APInt(C1)) &&
384         (match(Op0, m_Exact(m_UDiv(m_Value(X), m_CheckedInt(UDivCheck)))) ||
385          match(Op0, m_Exact(m_SDiv(m_Value(X), m_CheckedInt(SDivCheck)))))) {
386       auto BOpc = cast<BinaryOperator>(Op0)->getOpcode();
387       return BinaryOperator::CreateExact(
388           BOpc, X,
389           Builder.CreateBinOp(BOpc, cast<BinaryOperator>(Op0)->getOperand(1),
390                               Op1));
391     }
392   }
393 
394   // (X / Y) *  Y = X - (X % Y)
395   // (X / Y) * -Y = (X % Y) - X
396   {
397     Value *Y = Op1;
398     BinaryOperator *Div = dyn_cast<BinaryOperator>(Op0);
399     if (!Div || (Div->getOpcode() != Instruction::UDiv &&
400                  Div->getOpcode() != Instruction::SDiv)) {
401       Y = Op0;
402       Div = dyn_cast<BinaryOperator>(Op1);
403     }
404     Value *Neg = dyn_castNegVal(Y);
405     if (Div && Div->hasOneUse() &&
406         (Div->getOperand(1) == Y || Div->getOperand(1) == Neg) &&
407         (Div->getOpcode() == Instruction::UDiv ||
408          Div->getOpcode() == Instruction::SDiv)) {
409       Value *X = Div->getOperand(0), *DivOp1 = Div->getOperand(1);
410 
411       // If the division is exact, X % Y is zero, so we end up with X or -X.
412       if (Div->isExact()) {
413         if (DivOp1 == Y)
414           return replaceInstUsesWith(I, X);
415         return BinaryOperator::CreateNeg(X);
416       }
417 
418       auto RemOpc = Div->getOpcode() == Instruction::UDiv ? Instruction::URem
419                                                           : Instruction::SRem;
420       // X must be frozen because we are increasing its number of uses.
421       Value *XFreeze = X;
422       if (!isGuaranteedNotToBeUndef(X))
423         XFreeze = Builder.CreateFreeze(X, X->getName() + ".fr");
424       Value *Rem = Builder.CreateBinOp(RemOpc, XFreeze, DivOp1);
425       if (DivOp1 == Y)
426         return BinaryOperator::CreateSub(XFreeze, Rem);
427       return BinaryOperator::CreateSub(Rem, XFreeze);
428     }
429   }
430 
431   // Fold the following two scenarios:
432   //   1) i1 mul -> i1 and.
433   //   2) X * Y --> X & Y, iff X, Y can be only {0,1}.
434   // Note: We could use known bits to generalize this and related patterns with
435   // shifts/truncs
436   if (Ty->isIntOrIntVectorTy(1) ||
437       (match(Op0, m_And(m_Value(), m_One())) &&
438        match(Op1, m_And(m_Value(), m_One()))))
439     return BinaryOperator::CreateAnd(Op0, Op1);
440 
441   if (Value *R = foldMulShl1(I, /* CommuteOperands */ false, Builder))
442     return replaceInstUsesWith(I, R);
443   if (Value *R = foldMulShl1(I, /* CommuteOperands */ true, Builder))
444     return replaceInstUsesWith(I, R);
445 
446   // (zext bool X) * (zext bool Y) --> zext (and X, Y)
447   // (sext bool X) * (sext bool Y) --> zext (and X, Y)
448   // Note: -1 * -1 == 1 * 1 == 1 (if the extends match, the result is the same)
449   if (((match(Op0, m_ZExt(m_Value(X))) && match(Op1, m_ZExt(m_Value(Y)))) ||
450        (match(Op0, m_SExt(m_Value(X))) && match(Op1, m_SExt(m_Value(Y))))) &&
451       X->getType()->isIntOrIntVectorTy(1) && X->getType() == Y->getType() &&
452       (Op0->hasOneUse() || Op1->hasOneUse() || X == Y)) {
453     Value *And = Builder.CreateAnd(X, Y, "mulbool");
454     return CastInst::Create(Instruction::ZExt, And, Ty);
455   }
456   // (sext bool X) * (zext bool Y) --> sext (and X, Y)
457   // (zext bool X) * (sext bool Y) --> sext (and X, Y)
458   // Note: -1 * 1 == 1 * -1  == -1
459   if (((match(Op0, m_SExt(m_Value(X))) && match(Op1, m_ZExt(m_Value(Y)))) ||
460        (match(Op0, m_ZExt(m_Value(X))) && match(Op1, m_SExt(m_Value(Y))))) &&
461       X->getType()->isIntOrIntVectorTy(1) && X->getType() == Y->getType() &&
462       (Op0->hasOneUse() || Op1->hasOneUse())) {
463     Value *And = Builder.CreateAnd(X, Y, "mulbool");
464     return CastInst::Create(Instruction::SExt, And, Ty);
465   }
466 
467   // (zext bool X) * Y --> X ? Y : 0
468   // Y * (zext bool X) --> X ? Y : 0
469   if (match(Op0, m_ZExt(m_Value(X))) && X->getType()->isIntOrIntVectorTy(1))
470     return SelectInst::Create(X, Op1, ConstantInt::getNullValue(Ty));
471   if (match(Op1, m_ZExt(m_Value(X))) && X->getType()->isIntOrIntVectorTy(1))
472     return SelectInst::Create(X, Op0, ConstantInt::getNullValue(Ty));
473 
474   // mul (sext X), Y -> select X, -Y, 0
475   // mul Y, (sext X) -> select X, -Y, 0
476   if (match(&I, m_c_Mul(m_OneUse(m_SExt(m_Value(X))), m_Value(Y))) &&
477       X->getType()->isIntOrIntVectorTy(1))
478     return SelectInst::Create(X, Builder.CreateNeg(Y, "", I.hasNoSignedWrap()),
479                               ConstantInt::getNullValue(Op0->getType()));
480 
481   Constant *ImmC;
482   if (match(Op1, m_ImmConstant(ImmC))) {
483     // (sext bool X) * C --> X ? -C : 0
484     if (match(Op0, m_SExt(m_Value(X))) && X->getType()->isIntOrIntVectorTy(1)) {
485       Constant *NegC = ConstantExpr::getNeg(ImmC);
486       return SelectInst::Create(X, NegC, ConstantInt::getNullValue(Ty));
487     }
488 
489     // (ashr i32 X, 31) * C --> (X < 0) ? -C : 0
490     const APInt *C;
491     if (match(Op0, m_OneUse(m_AShr(m_Value(X), m_APInt(C)))) &&
492         *C == C->getBitWidth() - 1) {
493       Constant *NegC = ConstantExpr::getNeg(ImmC);
494       Value *IsNeg = Builder.CreateIsNeg(X, "isneg");
495       return SelectInst::Create(IsNeg, NegC, ConstantInt::getNullValue(Ty));
496     }
497   }
498 
499   // (lshr X, 31) * Y --> (X < 0) ? Y : 0
500   // TODO: We are not checking one-use because the elimination of the multiply
501   //       is better for analysis?
502   const APInt *C;
503   if (match(&I, m_c_BinOp(m_LShr(m_Value(X), m_APInt(C)), m_Value(Y))) &&
504       *C == C->getBitWidth() - 1) {
505     Value *IsNeg = Builder.CreateIsNeg(X, "isneg");
506     return SelectInst::Create(IsNeg, Y, ConstantInt::getNullValue(Ty));
507   }
508 
509   // (and X, 1) * Y --> (trunc X) ? Y : 0
510   if (match(&I, m_c_BinOp(m_OneUse(m_And(m_Value(X), m_One())), m_Value(Y)))) {
511     Value *Tr = Builder.CreateTrunc(X, CmpInst::makeCmpResultType(Ty));
512     return SelectInst::Create(Tr, Y, ConstantInt::getNullValue(Ty));
513   }
514 
515   // ((ashr X, 31) | 1) * X --> abs(X)
516   // X * ((ashr X, 31) | 1) --> abs(X)
517   if (match(&I, m_c_BinOp(m_Or(m_AShr(m_Value(X),
518                                       m_SpecificIntAllowPoison(BitWidth - 1)),
519                                m_One()),
520                           m_Deferred(X)))) {
521     Value *Abs = Builder.CreateBinaryIntrinsic(
522         Intrinsic::abs, X, ConstantInt::getBool(I.getContext(), HasNSW));
523     Abs->takeName(&I);
524     return replaceInstUsesWith(I, Abs);
525   }
526 
527   if (Instruction *Ext = narrowMathIfNoOverflow(I))
528     return Ext;
529 
530   if (Instruction *Res = foldBinOpOfSelectAndCastOfSelectCondition(I))
531     return Res;
532 
533   // (mul Op0 Op1):
534   //    if Log2(Op0) folds away ->
535   //        (shl Op1, Log2(Op0))
536   //    if Log2(Op1) folds away ->
537   //        (shl Op0, Log2(Op1))
538   if (takeLog2(Builder, Op0, /*Depth*/ 0, /*AssumeNonZero*/ false,
539                /*DoFold*/ false)) {
540     Value *Res = takeLog2(Builder, Op0, /*Depth*/ 0, /*AssumeNonZero*/ false,
541                           /*DoFold*/ true);
542     BinaryOperator *Shl = BinaryOperator::CreateShl(Op1, Res);
543     // We can only propegate nuw flag.
544     Shl->setHasNoUnsignedWrap(HasNUW);
545     return Shl;
546   }
547   if (takeLog2(Builder, Op1, /*Depth*/ 0, /*AssumeNonZero*/ false,
548                /*DoFold*/ false)) {
549     Value *Res = takeLog2(Builder, Op1, /*Depth*/ 0, /*AssumeNonZero*/ false,
550                           /*DoFold*/ true);
551     BinaryOperator *Shl = BinaryOperator::CreateShl(Op0, Res);
552     // We can only propegate nuw flag.
553     Shl->setHasNoUnsignedWrap(HasNUW);
554     return Shl;
555   }
556 
557   bool Changed = false;
558   if (!HasNSW && willNotOverflowSignedMul(Op0, Op1, I)) {
559     Changed = true;
560     I.setHasNoSignedWrap(true);
561   }
562 
563   if (!HasNUW && willNotOverflowUnsignedMul(Op0, Op1, I, I.hasNoSignedWrap())) {
564     Changed = true;
565     I.setHasNoUnsignedWrap(true);
566   }
567 
568   return Changed ? &I : nullptr;
569 }
570 
foldFPSignBitOps(BinaryOperator & I)571 Instruction *InstCombinerImpl::foldFPSignBitOps(BinaryOperator &I) {
572   BinaryOperator::BinaryOps Opcode = I.getOpcode();
573   assert((Opcode == Instruction::FMul || Opcode == Instruction::FDiv) &&
574          "Expected fmul or fdiv");
575 
576   Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
577   Value *X, *Y;
578 
579   // -X * -Y --> X * Y
580   // -X / -Y --> X / Y
581   if (match(Op0, m_FNeg(m_Value(X))) && match(Op1, m_FNeg(m_Value(Y))))
582     return BinaryOperator::CreateWithCopiedFlags(Opcode, X, Y, &I);
583 
584   // fabs(X) * fabs(X) -> X * X
585   // fabs(X) / fabs(X) -> X / X
586   if (Op0 == Op1 && match(Op0, m_FAbs(m_Value(X))))
587     return BinaryOperator::CreateWithCopiedFlags(Opcode, X, X, &I);
588 
589   // fabs(X) * fabs(Y) --> fabs(X * Y)
590   // fabs(X) / fabs(Y) --> fabs(X / Y)
591   if (match(Op0, m_FAbs(m_Value(X))) && match(Op1, m_FAbs(m_Value(Y))) &&
592       (Op0->hasOneUse() || Op1->hasOneUse())) {
593     IRBuilder<>::FastMathFlagGuard FMFGuard(Builder);
594     Builder.setFastMathFlags(I.getFastMathFlags());
595     Value *XY = Builder.CreateBinOp(Opcode, X, Y);
596     Value *Fabs = Builder.CreateUnaryIntrinsic(Intrinsic::fabs, XY);
597     Fabs->takeName(&I);
598     return replaceInstUsesWith(I, Fabs);
599   }
600 
601   return nullptr;
602 }
603 
foldPowiReassoc(BinaryOperator & I)604 Instruction *InstCombinerImpl::foldPowiReassoc(BinaryOperator &I) {
605   auto createPowiExpr = [](BinaryOperator &I, InstCombinerImpl &IC, Value *X,
606                            Value *Y, Value *Z) {
607     InstCombiner::BuilderTy &Builder = IC.Builder;
608     Value *YZ = Builder.CreateAdd(Y, Z);
609     Instruction *NewPow = Builder.CreateIntrinsic(
610         Intrinsic::powi, {X->getType(), YZ->getType()}, {X, YZ}, &I);
611 
612     return NewPow;
613   };
614 
615   Value *X, *Y, *Z;
616   unsigned Opcode = I.getOpcode();
617   assert((Opcode == Instruction::FMul || Opcode == Instruction::FDiv) &&
618          "Unexpected opcode");
619 
620   // powi(X, Y) * X --> powi(X, Y+1)
621   // X * powi(X, Y) --> powi(X, Y+1)
622   if (match(&I, m_c_FMul(m_OneUse(m_AllowReassoc(m_Intrinsic<Intrinsic::powi>(
623                              m_Value(X), m_Value(Y)))),
624                          m_Deferred(X)))) {
625     Constant *One = ConstantInt::get(Y->getType(), 1);
626     if (willNotOverflowSignedAdd(Y, One, I)) {
627       Instruction *NewPow = createPowiExpr(I, *this, X, Y, One);
628       return replaceInstUsesWith(I, NewPow);
629     }
630   }
631 
632   // powi(x, y) * powi(x, z) -> powi(x, y + z)
633   Value *Op0 = I.getOperand(0);
634   Value *Op1 = I.getOperand(1);
635   if (Opcode == Instruction::FMul && I.isOnlyUserOfAnyOperand() &&
636       match(Op0, m_AllowReassoc(
637                      m_Intrinsic<Intrinsic::powi>(m_Value(X), m_Value(Y)))) &&
638       match(Op1, m_AllowReassoc(m_Intrinsic<Intrinsic::powi>(m_Specific(X),
639                                                              m_Value(Z)))) &&
640       Y->getType() == Z->getType()) {
641     Instruction *NewPow = createPowiExpr(I, *this, X, Y, Z);
642     return replaceInstUsesWith(I, NewPow);
643   }
644 
645   if (Opcode == Instruction::FDiv && I.hasAllowReassoc() && I.hasNoNaNs()) {
646     // powi(X, Y) / X --> powi(X, Y-1)
647     // This is legal when (Y - 1) can't wraparound, in which case reassoc and
648     // nnan are required.
649     // TODO: Multi-use may be also better off creating Powi(x,y-1)
650     if (match(Op0, m_OneUse(m_AllowReassoc(m_Intrinsic<Intrinsic::powi>(
651                        m_Specific(Op1), m_Value(Y))))) &&
652         willNotOverflowSignedSub(Y, ConstantInt::get(Y->getType(), 1), I)) {
653       Constant *NegOne = ConstantInt::getAllOnesValue(Y->getType());
654       Instruction *NewPow = createPowiExpr(I, *this, Op1, Y, NegOne);
655       return replaceInstUsesWith(I, NewPow);
656     }
657 
658     // powi(X, Y) / (X * Z) --> powi(X, Y-1) / Z
659     // This is legal when (Y - 1) can't wraparound, in which case reassoc and
660     // nnan are required.
661     // TODO: Multi-use may be also better off creating Powi(x,y-1)
662     if (match(Op0, m_OneUse(m_AllowReassoc(m_Intrinsic<Intrinsic::powi>(
663                        m_Value(X), m_Value(Y))))) &&
664         match(Op1, m_AllowReassoc(m_c_FMul(m_Specific(X), m_Value(Z)))) &&
665         willNotOverflowSignedSub(Y, ConstantInt::get(Y->getType(), 1), I)) {
666       Constant *NegOne = ConstantInt::getAllOnesValue(Y->getType());
667       auto *NewPow = createPowiExpr(I, *this, X, Y, NegOne);
668       return BinaryOperator::CreateFDivFMF(NewPow, Z, &I);
669     }
670   }
671 
672   return nullptr;
673 }
674 
foldFMulReassoc(BinaryOperator & I)675 Instruction *InstCombinerImpl::foldFMulReassoc(BinaryOperator &I) {
676   Value *Op0 = I.getOperand(0);
677   Value *Op1 = I.getOperand(1);
678   Value *X, *Y;
679   Constant *C;
680   BinaryOperator *Op0BinOp;
681 
682   // Reassociate constant RHS with another constant to form constant
683   // expression.
684   if (match(Op1, m_Constant(C)) && C->isFiniteNonZeroFP() &&
685       match(Op0, m_AllowReassoc(m_BinOp(Op0BinOp)))) {
686     // Everything in this scope folds I with Op0, intersecting their FMF.
687     FastMathFlags FMF = I.getFastMathFlags() & Op0BinOp->getFastMathFlags();
688     IRBuilder<>::FastMathFlagGuard FMFGuard(Builder);
689     Builder.setFastMathFlags(FMF);
690     Constant *C1;
691     if (match(Op0, m_OneUse(m_FDiv(m_Constant(C1), m_Value(X))))) {
692       // (C1 / X) * C --> (C * C1) / X
693       Constant *CC1 =
694           ConstantFoldBinaryOpOperands(Instruction::FMul, C, C1, DL);
695       if (CC1 && CC1->isNormalFP())
696         return BinaryOperator::CreateFDivFMF(CC1, X, FMF);
697     }
698     if (match(Op0, m_FDiv(m_Value(X), m_Constant(C1)))) {
699       // FIXME: This seems like it should also be checking for arcp
700       // (X / C1) * C --> X * (C / C1)
701       Constant *CDivC1 =
702           ConstantFoldBinaryOpOperands(Instruction::FDiv, C, C1, DL);
703       if (CDivC1 && CDivC1->isNormalFP())
704         return BinaryOperator::CreateFMulFMF(X, CDivC1, FMF);
705 
706       // If the constant was a denormal, try reassociating differently.
707       // (X / C1) * C --> X / (C1 / C)
708       Constant *C1DivC =
709           ConstantFoldBinaryOpOperands(Instruction::FDiv, C1, C, DL);
710       if (C1DivC && Op0->hasOneUse() && C1DivC->isNormalFP())
711         return BinaryOperator::CreateFDivFMF(X, C1DivC, FMF);
712     }
713 
714     // We do not need to match 'fadd C, X' and 'fsub X, C' because they are
715     // canonicalized to 'fadd X, C'. Distributing the multiply may allow
716     // further folds and (X * C) + C2 is 'fma'.
717     if (match(Op0, m_OneUse(m_FAdd(m_Value(X), m_Constant(C1))))) {
718       // (X + C1) * C --> (X * C) + (C * C1)
719       if (Constant *CC1 =
720               ConstantFoldBinaryOpOperands(Instruction::FMul, C, C1, DL)) {
721         Value *XC = Builder.CreateFMul(X, C);
722         return BinaryOperator::CreateFAddFMF(XC, CC1, FMF);
723       }
724     }
725     if (match(Op0, m_OneUse(m_FSub(m_Constant(C1), m_Value(X))))) {
726       // (C1 - X) * C --> (C * C1) - (X * C)
727       if (Constant *CC1 =
728               ConstantFoldBinaryOpOperands(Instruction::FMul, C, C1, DL)) {
729         Value *XC = Builder.CreateFMul(X, C);
730         return BinaryOperator::CreateFSubFMF(CC1, XC, FMF);
731       }
732     }
733   }
734 
735   Value *Z;
736   if (match(&I,
737             m_c_FMul(m_AllowReassoc(m_OneUse(m_FDiv(m_Value(X), m_Value(Y)))),
738                      m_Value(Z)))) {
739     BinaryOperator *DivOp = cast<BinaryOperator>(((Z == Op0) ? Op1 : Op0));
740     FastMathFlags FMF = I.getFastMathFlags() & DivOp->getFastMathFlags();
741     if (FMF.allowReassoc()) {
742       // Sink division: (X / Y) * Z --> (X * Z) / Y
743       IRBuilder<>::FastMathFlagGuard FMFGuard(Builder);
744       Builder.setFastMathFlags(FMF);
745       auto *NewFMul = Builder.CreateFMul(X, Z);
746       return BinaryOperator::CreateFDivFMF(NewFMul, Y, FMF);
747     }
748   }
749 
750   // sqrt(X) * sqrt(Y) -> sqrt(X * Y)
751   // nnan disallows the possibility of returning a number if both operands are
752   // negative (in that case, we should return NaN).
753   if (I.hasNoNaNs() && match(Op0, m_OneUse(m_Sqrt(m_Value(X)))) &&
754       match(Op1, m_OneUse(m_Sqrt(m_Value(Y))))) {
755     Value *XY = Builder.CreateFMulFMF(X, Y, &I);
756     Value *Sqrt = Builder.CreateUnaryIntrinsic(Intrinsic::sqrt, XY, &I);
757     return replaceInstUsesWith(I, Sqrt);
758   }
759 
760   // The following transforms are done irrespective of the number of uses
761   // for the expression "1.0/sqrt(X)".
762   //  1) 1.0/sqrt(X) * X -> X/sqrt(X)
763   //  2) X * 1.0/sqrt(X) -> X/sqrt(X)
764   // We always expect the backend to reduce X/sqrt(X) to sqrt(X), if it
765   // has the necessary (reassoc) fast-math-flags.
766   if (I.hasNoSignedZeros() &&
767       match(Op0, (m_FDiv(m_SpecificFP(1.0), m_Value(Y)))) &&
768       match(Y, m_Sqrt(m_Value(X))) && Op1 == X)
769     return BinaryOperator::CreateFDivFMF(X, Y, &I);
770   if (I.hasNoSignedZeros() &&
771       match(Op1, (m_FDiv(m_SpecificFP(1.0), m_Value(Y)))) &&
772       match(Y, m_Sqrt(m_Value(X))) && Op0 == X)
773     return BinaryOperator::CreateFDivFMF(X, Y, &I);
774 
775   // Like the similar transform in instsimplify, this requires 'nsz' because
776   // sqrt(-0.0) = -0.0, and -0.0 * -0.0 does not simplify to -0.0.
777   if (I.hasNoNaNs() && I.hasNoSignedZeros() && Op0 == Op1 && Op0->hasNUses(2)) {
778     // Peek through fdiv to find squaring of square root:
779     // (X / sqrt(Y)) * (X / sqrt(Y)) --> (X * X) / Y
780     if (match(Op0, m_FDiv(m_Value(X), m_Sqrt(m_Value(Y))))) {
781       Value *XX = Builder.CreateFMulFMF(X, X, &I);
782       return BinaryOperator::CreateFDivFMF(XX, Y, &I);
783     }
784     // (sqrt(Y) / X) * (sqrt(Y) / X) --> Y / (X * X)
785     if (match(Op0, m_FDiv(m_Sqrt(m_Value(Y)), m_Value(X)))) {
786       Value *XX = Builder.CreateFMulFMF(X, X, &I);
787       return BinaryOperator::CreateFDivFMF(Y, XX, &I);
788     }
789   }
790 
791   // pow(X, Y) * X --> pow(X, Y+1)
792   // X * pow(X, Y) --> pow(X, Y+1)
793   if (match(&I, m_c_FMul(m_OneUse(m_Intrinsic<Intrinsic::pow>(m_Value(X),
794                                                               m_Value(Y))),
795                          m_Deferred(X)))) {
796     Value *Y1 = Builder.CreateFAddFMF(Y, ConstantFP::get(I.getType(), 1.0), &I);
797     Value *Pow = Builder.CreateBinaryIntrinsic(Intrinsic::pow, X, Y1, &I);
798     return replaceInstUsesWith(I, Pow);
799   }
800 
801   if (Instruction *FoldedPowi = foldPowiReassoc(I))
802     return FoldedPowi;
803 
804   if (I.isOnlyUserOfAnyOperand()) {
805     // pow(X, Y) * pow(X, Z) -> pow(X, Y + Z)
806     if (match(Op0, m_Intrinsic<Intrinsic::pow>(m_Value(X), m_Value(Y))) &&
807         match(Op1, m_Intrinsic<Intrinsic::pow>(m_Specific(X), m_Value(Z)))) {
808       auto *YZ = Builder.CreateFAddFMF(Y, Z, &I);
809       auto *NewPow = Builder.CreateBinaryIntrinsic(Intrinsic::pow, X, YZ, &I);
810       return replaceInstUsesWith(I, NewPow);
811     }
812     // pow(X, Y) * pow(Z, Y) -> pow(X * Z, Y)
813     if (match(Op0, m_Intrinsic<Intrinsic::pow>(m_Value(X), m_Value(Y))) &&
814         match(Op1, m_Intrinsic<Intrinsic::pow>(m_Value(Z), m_Specific(Y)))) {
815       auto *XZ = Builder.CreateFMulFMF(X, Z, &I);
816       auto *NewPow = Builder.CreateBinaryIntrinsic(Intrinsic::pow, XZ, Y, &I);
817       return replaceInstUsesWith(I, NewPow);
818     }
819 
820     // exp(X) * exp(Y) -> exp(X + Y)
821     if (match(Op0, m_Intrinsic<Intrinsic::exp>(m_Value(X))) &&
822         match(Op1, m_Intrinsic<Intrinsic::exp>(m_Value(Y)))) {
823       Value *XY = Builder.CreateFAddFMF(X, Y, &I);
824       Value *Exp = Builder.CreateUnaryIntrinsic(Intrinsic::exp, XY, &I);
825       return replaceInstUsesWith(I, Exp);
826     }
827 
828     // exp2(X) * exp2(Y) -> exp2(X + Y)
829     if (match(Op0, m_Intrinsic<Intrinsic::exp2>(m_Value(X))) &&
830         match(Op1, m_Intrinsic<Intrinsic::exp2>(m_Value(Y)))) {
831       Value *XY = Builder.CreateFAddFMF(X, Y, &I);
832       Value *Exp2 = Builder.CreateUnaryIntrinsic(Intrinsic::exp2, XY, &I);
833       return replaceInstUsesWith(I, Exp2);
834     }
835   }
836 
837   // (X*Y) * X => (X*X) * Y where Y != X
838   //  The purpose is two-fold:
839   //   1) to form a power expression (of X).
840   //   2) potentially shorten the critical path: After transformation, the
841   //  latency of the instruction Y is amortized by the expression of X*X,
842   //  and therefore Y is in a "less critical" position compared to what it
843   //  was before the transformation.
844   if (match(Op0, m_OneUse(m_c_FMul(m_Specific(Op1), m_Value(Y)))) && Op1 != Y) {
845     Value *XX = Builder.CreateFMulFMF(Op1, Op1, &I);
846     return BinaryOperator::CreateFMulFMF(XX, Y, &I);
847   }
848   if (match(Op1, m_OneUse(m_c_FMul(m_Specific(Op0), m_Value(Y)))) && Op0 != Y) {
849     Value *XX = Builder.CreateFMulFMF(Op0, Op0, &I);
850     return BinaryOperator::CreateFMulFMF(XX, Y, &I);
851   }
852 
853   return nullptr;
854 }
855 
visitFMul(BinaryOperator & I)856 Instruction *InstCombinerImpl::visitFMul(BinaryOperator &I) {
857   if (Value *V = simplifyFMulInst(I.getOperand(0), I.getOperand(1),
858                                   I.getFastMathFlags(),
859                                   SQ.getWithInstruction(&I)))
860     return replaceInstUsesWith(I, V);
861 
862   if (SimplifyAssociativeOrCommutative(I))
863     return &I;
864 
865   if (Instruction *X = foldVectorBinop(I))
866     return X;
867 
868   if (Instruction *Phi = foldBinopWithPhiOperands(I))
869     return Phi;
870 
871   if (Instruction *FoldedMul = foldBinOpIntoSelectOrPhi(I))
872     return FoldedMul;
873 
874   if (Value *FoldedMul = foldMulSelectToNegate(I, Builder))
875     return replaceInstUsesWith(I, FoldedMul);
876 
877   if (Instruction *R = foldFPSignBitOps(I))
878     return R;
879 
880   if (Instruction *R = foldFBinOpOfIntCasts(I))
881     return R;
882 
883   // X * -1.0 --> -X
884   Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
885   if (match(Op1, m_SpecificFP(-1.0)))
886     return UnaryOperator::CreateFNegFMF(Op0, &I);
887 
888   // With no-nans/no-infs:
889   // X * 0.0 --> copysign(0.0, X)
890   // X * -0.0 --> copysign(0.0, -X)
891   const APFloat *FPC;
892   if (match(Op1, m_APFloatAllowPoison(FPC)) && FPC->isZero() &&
893       ((I.hasNoInfs() &&
894         isKnownNeverNaN(Op0, /*Depth=*/0, SQ.getWithInstruction(&I))) ||
895        isKnownNeverNaN(&I, /*Depth=*/0, SQ.getWithInstruction(&I)))) {
896     if (FPC->isNegative())
897       Op0 = Builder.CreateFNegFMF(Op0, &I);
898     CallInst *CopySign = Builder.CreateIntrinsic(Intrinsic::copysign,
899                                                  {I.getType()}, {Op1, Op0}, &I);
900     return replaceInstUsesWith(I, CopySign);
901   }
902 
903   // -X * C --> X * -C
904   Value *X, *Y;
905   Constant *C;
906   if (match(Op0, m_FNeg(m_Value(X))) && match(Op1, m_Constant(C)))
907     if (Constant *NegC = ConstantFoldUnaryOpOperand(Instruction::FNeg, C, DL))
908       return BinaryOperator::CreateFMulFMF(X, NegC, &I);
909 
910   if (I.hasNoNaNs() && I.hasNoSignedZeros()) {
911     // (uitofp bool X) * Y --> X ? Y : 0
912     // Y * (uitofp bool X) --> X ? Y : 0
913     // Note INF * 0 is NaN.
914     if (match(Op0, m_UIToFP(m_Value(X))) &&
915         X->getType()->isIntOrIntVectorTy(1)) {
916       auto *SI = SelectInst::Create(X, Op1, ConstantFP::get(I.getType(), 0.0));
917       SI->copyFastMathFlags(I.getFastMathFlags());
918       return SI;
919     }
920     if (match(Op1, m_UIToFP(m_Value(X))) &&
921         X->getType()->isIntOrIntVectorTy(1)) {
922       auto *SI = SelectInst::Create(X, Op0, ConstantFP::get(I.getType(), 0.0));
923       SI->copyFastMathFlags(I.getFastMathFlags());
924       return SI;
925     }
926   }
927 
928   // (select A, B, C) * (select A, D, E) --> select A, (B*D), (C*E)
929   if (Value *V = SimplifySelectsFeedingBinaryOp(I, Op0, Op1))
930     return replaceInstUsesWith(I, V);
931 
932   if (I.hasAllowReassoc())
933     if (Instruction *FoldedMul = foldFMulReassoc(I))
934       return FoldedMul;
935 
936   // log2(X * 0.5) * Y = log2(X) * Y - Y
937   if (I.isFast()) {
938     IntrinsicInst *Log2 = nullptr;
939     if (match(Op0, m_OneUse(m_Intrinsic<Intrinsic::log2>(
940             m_OneUse(m_FMul(m_Value(X), m_SpecificFP(0.5))))))) {
941       Log2 = cast<IntrinsicInst>(Op0);
942       Y = Op1;
943     }
944     if (match(Op1, m_OneUse(m_Intrinsic<Intrinsic::log2>(
945             m_OneUse(m_FMul(m_Value(X), m_SpecificFP(0.5))))))) {
946       Log2 = cast<IntrinsicInst>(Op1);
947       Y = Op0;
948     }
949     if (Log2) {
950       Value *Log2 = Builder.CreateUnaryIntrinsic(Intrinsic::log2, X, &I);
951       Value *LogXTimesY = Builder.CreateFMulFMF(Log2, Y, &I);
952       return BinaryOperator::CreateFSubFMF(LogXTimesY, Y, &I);
953     }
954   }
955 
956   // Simplify FMUL recurrences starting with 0.0 to 0.0 if nnan and nsz are set.
957   // Given a phi node with entry value as 0 and it used in fmul operation,
958   // we can replace fmul with 0 safely and eleminate loop operation.
959   PHINode *PN = nullptr;
960   Value *Start = nullptr, *Step = nullptr;
961   if (matchSimpleRecurrence(&I, PN, Start, Step) && I.hasNoNaNs() &&
962       I.hasNoSignedZeros() && match(Start, m_Zero()))
963     return replaceInstUsesWith(I, Start);
964 
965   // minimum(X, Y) * maximum(X, Y) => X * Y.
966   if (match(&I,
967             m_c_FMul(m_Intrinsic<Intrinsic::maximum>(m_Value(X), m_Value(Y)),
968                      m_c_Intrinsic<Intrinsic::minimum>(m_Deferred(X),
969                                                        m_Deferred(Y))))) {
970     BinaryOperator *Result = BinaryOperator::CreateFMulFMF(X, Y, &I);
971     // We cannot preserve ninf if nnan flag is not set.
972     // If X is NaN and Y is Inf then in original program we had NaN * NaN,
973     // while in optimized version NaN * Inf and this is a poison with ninf flag.
974     if (!Result->hasNoNaNs())
975       Result->setHasNoInfs(false);
976     return Result;
977   }
978 
979   return nullptr;
980 }
981 
982 /// Fold a divide or remainder with a select instruction divisor when one of the
983 /// select operands is zero. In that case, we can use the other select operand
984 /// because div/rem by zero is undefined.
simplifyDivRemOfSelectWithZeroOp(BinaryOperator & I)985 bool InstCombinerImpl::simplifyDivRemOfSelectWithZeroOp(BinaryOperator &I) {
986   SelectInst *SI = dyn_cast<SelectInst>(I.getOperand(1));
987   if (!SI)
988     return false;
989 
990   int NonNullOperand;
991   if (match(SI->getTrueValue(), m_Zero()))
992     // div/rem X, (Cond ? 0 : Y) -> div/rem X, Y
993     NonNullOperand = 2;
994   else if (match(SI->getFalseValue(), m_Zero()))
995     // div/rem X, (Cond ? Y : 0) -> div/rem X, Y
996     NonNullOperand = 1;
997   else
998     return false;
999 
1000   // Change the div/rem to use 'Y' instead of the select.
1001   replaceOperand(I, 1, SI->getOperand(NonNullOperand));
1002 
1003   // Okay, we know we replace the operand of the div/rem with 'Y' with no
1004   // problem.  However, the select, or the condition of the select may have
1005   // multiple uses.  Based on our knowledge that the operand must be non-zero,
1006   // propagate the known value for the select into other uses of it, and
1007   // propagate a known value of the condition into its other users.
1008 
1009   // If the select and condition only have a single use, don't bother with this,
1010   // early exit.
1011   Value *SelectCond = SI->getCondition();
1012   if (SI->use_empty() && SelectCond->hasOneUse())
1013     return true;
1014 
1015   // Scan the current block backward, looking for other uses of SI.
1016   BasicBlock::iterator BBI = I.getIterator(), BBFront = I.getParent()->begin();
1017   Type *CondTy = SelectCond->getType();
1018   while (BBI != BBFront) {
1019     --BBI;
1020     // If we found an instruction that we can't assume will return, so
1021     // information from below it cannot be propagated above it.
1022     if (!isGuaranteedToTransferExecutionToSuccessor(&*BBI))
1023       break;
1024 
1025     // Replace uses of the select or its condition with the known values.
1026     for (Use &Op : BBI->operands()) {
1027       if (Op == SI) {
1028         replaceUse(Op, SI->getOperand(NonNullOperand));
1029         Worklist.push(&*BBI);
1030       } else if (Op == SelectCond) {
1031         replaceUse(Op, NonNullOperand == 1 ? ConstantInt::getTrue(CondTy)
1032                                            : ConstantInt::getFalse(CondTy));
1033         Worklist.push(&*BBI);
1034       }
1035     }
1036 
1037     // If we past the instruction, quit looking for it.
1038     if (&*BBI == SI)
1039       SI = nullptr;
1040     if (&*BBI == SelectCond)
1041       SelectCond = nullptr;
1042 
1043     // If we ran out of things to eliminate, break out of the loop.
1044     if (!SelectCond && !SI)
1045       break;
1046 
1047   }
1048   return true;
1049 }
1050 
1051 /// True if the multiply can not be expressed in an int this size.
multiplyOverflows(const APInt & C1,const APInt & C2,APInt & Product,bool IsSigned)1052 static bool multiplyOverflows(const APInt &C1, const APInt &C2, APInt &Product,
1053                               bool IsSigned) {
1054   bool Overflow;
1055   Product = IsSigned ? C1.smul_ov(C2, Overflow) : C1.umul_ov(C2, Overflow);
1056   return Overflow;
1057 }
1058 
1059 /// True if C1 is a multiple of C2. Quotient contains C1/C2.
isMultiple(const APInt & C1,const APInt & C2,APInt & Quotient,bool IsSigned)1060 static bool isMultiple(const APInt &C1, const APInt &C2, APInt &Quotient,
1061                        bool IsSigned) {
1062   assert(C1.getBitWidth() == C2.getBitWidth() && "Constant widths not equal");
1063 
1064   // Bail if we will divide by zero.
1065   if (C2.isZero())
1066     return false;
1067 
1068   // Bail if we would divide INT_MIN by -1.
1069   if (IsSigned && C1.isMinSignedValue() && C2.isAllOnes())
1070     return false;
1071 
1072   APInt Remainder(C1.getBitWidth(), /*val=*/0ULL, IsSigned);
1073   if (IsSigned)
1074     APInt::sdivrem(C1, C2, Quotient, Remainder);
1075   else
1076     APInt::udivrem(C1, C2, Quotient, Remainder);
1077 
1078   return Remainder.isMinValue();
1079 }
1080 
foldIDivShl(BinaryOperator & I,InstCombiner::BuilderTy & Builder)1081 static Value *foldIDivShl(BinaryOperator &I, InstCombiner::BuilderTy &Builder) {
1082   assert((I.getOpcode() == Instruction::SDiv ||
1083           I.getOpcode() == Instruction::UDiv) &&
1084          "Expected integer divide");
1085 
1086   bool IsSigned = I.getOpcode() == Instruction::SDiv;
1087   Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
1088   Type *Ty = I.getType();
1089 
1090   Value *X, *Y, *Z;
1091 
1092   // With appropriate no-wrap constraints, remove a common factor in the
1093   // dividend and divisor that is disguised as a left-shifted value.
1094   if (match(Op1, m_Shl(m_Value(X), m_Value(Z))) &&
1095       match(Op0, m_c_Mul(m_Specific(X), m_Value(Y)))) {
1096     // Both operands must have the matching no-wrap for this kind of division.
1097     auto *Mul = cast<OverflowingBinaryOperator>(Op0);
1098     auto *Shl = cast<OverflowingBinaryOperator>(Op1);
1099     bool HasNUW = Mul->hasNoUnsignedWrap() && Shl->hasNoUnsignedWrap();
1100     bool HasNSW = Mul->hasNoSignedWrap() && Shl->hasNoSignedWrap();
1101 
1102     // (X * Y) u/ (X << Z) --> Y u>> Z
1103     if (!IsSigned && HasNUW)
1104       return Builder.CreateLShr(Y, Z, "", I.isExact());
1105 
1106     // (X * Y) s/ (X << Z) --> Y s/ (1 << Z)
1107     if (IsSigned && HasNSW && (Op0->hasOneUse() || Op1->hasOneUse())) {
1108       Value *Shl = Builder.CreateShl(ConstantInt::get(Ty, 1), Z);
1109       return Builder.CreateSDiv(Y, Shl, "", I.isExact());
1110     }
1111   }
1112 
1113   // With appropriate no-wrap constraints, remove a common factor in the
1114   // dividend and divisor that is disguised as a left-shift amount.
1115   if (match(Op0, m_Shl(m_Value(X), m_Value(Z))) &&
1116       match(Op1, m_Shl(m_Value(Y), m_Specific(Z)))) {
1117     auto *Shl0 = cast<OverflowingBinaryOperator>(Op0);
1118     auto *Shl1 = cast<OverflowingBinaryOperator>(Op1);
1119 
1120     // For unsigned div, we need 'nuw' on both shifts or
1121     // 'nsw' on both shifts + 'nuw' on the dividend.
1122     // (X << Z) / (Y << Z) --> X / Y
1123     if (!IsSigned &&
1124         ((Shl0->hasNoUnsignedWrap() && Shl1->hasNoUnsignedWrap()) ||
1125          (Shl0->hasNoUnsignedWrap() && Shl0->hasNoSignedWrap() &&
1126           Shl1->hasNoSignedWrap())))
1127       return Builder.CreateUDiv(X, Y, "", I.isExact());
1128 
1129     // For signed div, we need 'nsw' on both shifts + 'nuw' on the divisor.
1130     // (X << Z) / (Y << Z) --> X / Y
1131     if (IsSigned && Shl0->hasNoSignedWrap() && Shl1->hasNoSignedWrap() &&
1132         Shl1->hasNoUnsignedWrap())
1133       return Builder.CreateSDiv(X, Y, "", I.isExact());
1134   }
1135 
1136   // If X << Y and X << Z does not overflow, then:
1137   // (X << Y) / (X << Z) -> (1 << Y) / (1 << Z) -> 1 << Y >> Z
1138   if (match(Op0, m_Shl(m_Value(X), m_Value(Y))) &&
1139       match(Op1, m_Shl(m_Specific(X), m_Value(Z)))) {
1140     auto *Shl0 = cast<OverflowingBinaryOperator>(Op0);
1141     auto *Shl1 = cast<OverflowingBinaryOperator>(Op1);
1142 
1143     if (IsSigned ? (Shl0->hasNoSignedWrap() && Shl1->hasNoSignedWrap())
1144                  : (Shl0->hasNoUnsignedWrap() && Shl1->hasNoUnsignedWrap())) {
1145       Constant *One = ConstantInt::get(X->getType(), 1);
1146       // Only preserve the nsw flag if dividend has nsw
1147       // or divisor has nsw and operator is sdiv.
1148       Value *Dividend = Builder.CreateShl(
1149           One, Y, "shl.dividend",
1150           /*HasNUW*/ true,
1151           /*HasNSW*/
1152           IsSigned ? (Shl0->hasNoUnsignedWrap() || Shl1->hasNoUnsignedWrap())
1153                    : Shl0->hasNoSignedWrap());
1154       return Builder.CreateLShr(Dividend, Z, "", I.isExact());
1155     }
1156   }
1157 
1158   return nullptr;
1159 }
1160 
1161 /// This function implements the transforms common to both integer division
1162 /// instructions (udiv and sdiv). It is called by the visitors to those integer
1163 /// division instructions.
1164 /// Common integer divide transforms
commonIDivTransforms(BinaryOperator & I)1165 Instruction *InstCombinerImpl::commonIDivTransforms(BinaryOperator &I) {
1166   if (Instruction *Phi = foldBinopWithPhiOperands(I))
1167     return Phi;
1168 
1169   Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
1170   bool IsSigned = I.getOpcode() == Instruction::SDiv;
1171   Type *Ty = I.getType();
1172 
1173   // The RHS is known non-zero.
1174   if (Value *V = simplifyValueKnownNonZero(I.getOperand(1), *this, I))
1175     return replaceOperand(I, 1, V);
1176 
1177   // Handle cases involving: [su]div X, (select Cond, Y, Z)
1178   // This does not apply for fdiv.
1179   if (simplifyDivRemOfSelectWithZeroOp(I))
1180     return &I;
1181 
1182   // If the divisor is a select-of-constants, try to constant fold all div ops:
1183   // C / (select Cond, TrueC, FalseC) --> select Cond, (C / TrueC), (C / FalseC)
1184   // TODO: Adapt simplifyDivRemOfSelectWithZeroOp to allow this and other folds.
1185   if (match(Op0, m_ImmConstant()) &&
1186       match(Op1, m_Select(m_Value(), m_ImmConstant(), m_ImmConstant()))) {
1187     if (Instruction *R = FoldOpIntoSelect(I, cast<SelectInst>(Op1),
1188                                           /*FoldWithMultiUse*/ true))
1189       return R;
1190   }
1191 
1192   const APInt *C2;
1193   if (match(Op1, m_APInt(C2))) {
1194     Value *X;
1195     const APInt *C1;
1196 
1197     // (X / C1) / C2  -> X / (C1*C2)
1198     if ((IsSigned && match(Op0, m_SDiv(m_Value(X), m_APInt(C1)))) ||
1199         (!IsSigned && match(Op0, m_UDiv(m_Value(X), m_APInt(C1))))) {
1200       APInt Product(C1->getBitWidth(), /*val=*/0ULL, IsSigned);
1201       if (!multiplyOverflows(*C1, *C2, Product, IsSigned))
1202         return BinaryOperator::Create(I.getOpcode(), X,
1203                                       ConstantInt::get(Ty, Product));
1204     }
1205 
1206     APInt Quotient(C2->getBitWidth(), /*val=*/0ULL, IsSigned);
1207     if ((IsSigned && match(Op0, m_NSWMul(m_Value(X), m_APInt(C1)))) ||
1208         (!IsSigned && match(Op0, m_NUWMul(m_Value(X), m_APInt(C1))))) {
1209 
1210       // (X * C1) / C2 -> X / (C2 / C1) if C2 is a multiple of C1.
1211       if (isMultiple(*C2, *C1, Quotient, IsSigned)) {
1212         auto *NewDiv = BinaryOperator::Create(I.getOpcode(), X,
1213                                               ConstantInt::get(Ty, Quotient));
1214         NewDiv->setIsExact(I.isExact());
1215         return NewDiv;
1216       }
1217 
1218       // (X * C1) / C2 -> X * (C1 / C2) if C1 is a multiple of C2.
1219       if (isMultiple(*C1, *C2, Quotient, IsSigned)) {
1220         auto *Mul = BinaryOperator::Create(Instruction::Mul, X,
1221                                            ConstantInt::get(Ty, Quotient));
1222         auto *OBO = cast<OverflowingBinaryOperator>(Op0);
1223         Mul->setHasNoUnsignedWrap(!IsSigned && OBO->hasNoUnsignedWrap());
1224         Mul->setHasNoSignedWrap(OBO->hasNoSignedWrap());
1225         return Mul;
1226       }
1227     }
1228 
1229     if ((IsSigned && match(Op0, m_NSWShl(m_Value(X), m_APInt(C1))) &&
1230          C1->ult(C1->getBitWidth() - 1)) ||
1231         (!IsSigned && match(Op0, m_NUWShl(m_Value(X), m_APInt(C1))) &&
1232          C1->ult(C1->getBitWidth()))) {
1233       APInt C1Shifted = APInt::getOneBitSet(
1234           C1->getBitWidth(), static_cast<unsigned>(C1->getZExtValue()));
1235 
1236       // (X << C1) / C2 -> X / (C2 >> C1) if C2 is a multiple of 1 << C1.
1237       if (isMultiple(*C2, C1Shifted, Quotient, IsSigned)) {
1238         auto *BO = BinaryOperator::Create(I.getOpcode(), X,
1239                                           ConstantInt::get(Ty, Quotient));
1240         BO->setIsExact(I.isExact());
1241         return BO;
1242       }
1243 
1244       // (X << C1) / C2 -> X * ((1 << C1) / C2) if 1 << C1 is a multiple of C2.
1245       if (isMultiple(C1Shifted, *C2, Quotient, IsSigned)) {
1246         auto *Mul = BinaryOperator::Create(Instruction::Mul, X,
1247                                            ConstantInt::get(Ty, Quotient));
1248         auto *OBO = cast<OverflowingBinaryOperator>(Op0);
1249         Mul->setHasNoUnsignedWrap(!IsSigned && OBO->hasNoUnsignedWrap());
1250         Mul->setHasNoSignedWrap(OBO->hasNoSignedWrap());
1251         return Mul;
1252       }
1253     }
1254 
1255     // Distribute div over add to eliminate a matching div/mul pair:
1256     // ((X * C2) + C1) / C2 --> X + C1/C2
1257     // We need a multiple of the divisor for a signed add constant, but
1258     // unsigned is fine with any constant pair.
1259     if (IsSigned &&
1260         match(Op0, m_NSWAddLike(m_NSWMul(m_Value(X), m_SpecificInt(*C2)),
1261                                 m_APInt(C1))) &&
1262         isMultiple(*C1, *C2, Quotient, IsSigned)) {
1263       return BinaryOperator::CreateNSWAdd(X, ConstantInt::get(Ty, Quotient));
1264     }
1265     if (!IsSigned &&
1266         match(Op0, m_NUWAddLike(m_NUWMul(m_Value(X), m_SpecificInt(*C2)),
1267                                 m_APInt(C1)))) {
1268       return BinaryOperator::CreateNUWAdd(X,
1269                                           ConstantInt::get(Ty, C1->udiv(*C2)));
1270     }
1271 
1272     if (!C2->isZero()) // avoid X udiv 0
1273       if (Instruction *FoldedDiv = foldBinOpIntoSelectOrPhi(I))
1274         return FoldedDiv;
1275   }
1276 
1277   if (match(Op0, m_One())) {
1278     assert(!Ty->isIntOrIntVectorTy(1) && "i1 divide not removed?");
1279     if (IsSigned) {
1280       // 1 / 0 --> undef ; 1 / 1 --> 1 ; 1 / -1 --> -1 ; 1 / anything else --> 0
1281       // (Op1 + 1) u< 3 ? Op1 : 0
1282       // Op1 must be frozen because we are increasing its number of uses.
1283       Value *F1 = Op1;
1284       if (!isGuaranteedNotToBeUndef(Op1))
1285         F1 = Builder.CreateFreeze(Op1, Op1->getName() + ".fr");
1286       Value *Inc = Builder.CreateAdd(F1, Op0);
1287       Value *Cmp = Builder.CreateICmpULT(Inc, ConstantInt::get(Ty, 3));
1288       return SelectInst::Create(Cmp, F1, ConstantInt::get(Ty, 0));
1289     } else {
1290       // If Op1 is 0 then it's undefined behaviour. If Op1 is 1 then the
1291       // result is one, otherwise it's zero.
1292       return new ZExtInst(Builder.CreateICmpEQ(Op1, Op0), Ty);
1293     }
1294   }
1295 
1296   // See if we can fold away this div instruction.
1297   if (SimplifyDemandedInstructionBits(I))
1298     return &I;
1299 
1300   // (X - (X rem Y)) / Y -> X / Y; usually originates as ((X / Y) * Y) / Y
1301   Value *X, *Z;
1302   if (match(Op0, m_Sub(m_Value(X), m_Value(Z)))) // (X - Z) / Y; Y = Op1
1303     if ((IsSigned && match(Z, m_SRem(m_Specific(X), m_Specific(Op1)))) ||
1304         (!IsSigned && match(Z, m_URem(m_Specific(X), m_Specific(Op1)))))
1305       return BinaryOperator::Create(I.getOpcode(), X, Op1);
1306 
1307   // (X << Y) / X -> 1 << Y
1308   Value *Y;
1309   if (IsSigned && match(Op0, m_NSWShl(m_Specific(Op1), m_Value(Y))))
1310     return BinaryOperator::CreateNSWShl(ConstantInt::get(Ty, 1), Y);
1311   if (!IsSigned && match(Op0, m_NUWShl(m_Specific(Op1), m_Value(Y))))
1312     return BinaryOperator::CreateNUWShl(ConstantInt::get(Ty, 1), Y);
1313 
1314   // X / (X * Y) -> 1 / Y if the multiplication does not overflow.
1315   if (match(Op1, m_c_Mul(m_Specific(Op0), m_Value(Y)))) {
1316     bool HasNSW = cast<OverflowingBinaryOperator>(Op1)->hasNoSignedWrap();
1317     bool HasNUW = cast<OverflowingBinaryOperator>(Op1)->hasNoUnsignedWrap();
1318     if ((IsSigned && HasNSW) || (!IsSigned && HasNUW)) {
1319       replaceOperand(I, 0, ConstantInt::get(Ty, 1));
1320       replaceOperand(I, 1, Y);
1321       return &I;
1322     }
1323   }
1324 
1325   // (X << Z) / (X * Y) -> (1 << Z) / Y
1326   // TODO: Handle sdiv.
1327   if (!IsSigned && Op1->hasOneUse() &&
1328       match(Op0, m_NUWShl(m_Value(X), m_Value(Z))) &&
1329       match(Op1, m_c_Mul(m_Specific(X), m_Value(Y))))
1330     if (cast<OverflowingBinaryOperator>(Op1)->hasNoUnsignedWrap()) {
1331       Instruction *NewDiv = BinaryOperator::CreateUDiv(
1332           Builder.CreateShl(ConstantInt::get(Ty, 1), Z, "", /*NUW*/ true), Y);
1333       NewDiv->setIsExact(I.isExact());
1334       return NewDiv;
1335     }
1336 
1337   if (Value *R = foldIDivShl(I, Builder))
1338     return replaceInstUsesWith(I, R);
1339 
1340   // With the appropriate no-wrap constraint, remove a multiply by the divisor
1341   // after peeking through another divide:
1342   // ((Op1 * X) / Y) / Op1 --> X / Y
1343   if (match(Op0, m_BinOp(I.getOpcode(), m_c_Mul(m_Specific(Op1), m_Value(X)),
1344                          m_Value(Y)))) {
1345     auto *InnerDiv = cast<PossiblyExactOperator>(Op0);
1346     auto *Mul = cast<OverflowingBinaryOperator>(InnerDiv->getOperand(0));
1347     Instruction *NewDiv = nullptr;
1348     if (!IsSigned && Mul->hasNoUnsignedWrap())
1349       NewDiv = BinaryOperator::CreateUDiv(X, Y);
1350     else if (IsSigned && Mul->hasNoSignedWrap())
1351       NewDiv = BinaryOperator::CreateSDiv(X, Y);
1352 
1353     // Exact propagates only if both of the original divides are exact.
1354     if (NewDiv) {
1355       NewDiv->setIsExact(I.isExact() && InnerDiv->isExact());
1356       return NewDiv;
1357     }
1358   }
1359 
1360   // (X * Y) / (X * Z) --> Y / Z (and commuted variants)
1361   if (match(Op0, m_Mul(m_Value(X), m_Value(Y)))) {
1362     auto OB0HasNSW = cast<OverflowingBinaryOperator>(Op0)->hasNoSignedWrap();
1363     auto OB0HasNUW = cast<OverflowingBinaryOperator>(Op0)->hasNoUnsignedWrap();
1364 
1365     auto CreateDivOrNull = [&](Value *A, Value *B) -> Instruction * {
1366       auto OB1HasNSW = cast<OverflowingBinaryOperator>(Op1)->hasNoSignedWrap();
1367       auto OB1HasNUW =
1368           cast<OverflowingBinaryOperator>(Op1)->hasNoUnsignedWrap();
1369       const APInt *C1, *C2;
1370       if (IsSigned && OB0HasNSW) {
1371         if (OB1HasNSW && match(B, m_APInt(C1)) && !C1->isAllOnes())
1372           return BinaryOperator::CreateSDiv(A, B);
1373       }
1374       if (!IsSigned && OB0HasNUW) {
1375         if (OB1HasNUW)
1376           return BinaryOperator::CreateUDiv(A, B);
1377         if (match(A, m_APInt(C1)) && match(B, m_APInt(C2)) && C2->ule(*C1))
1378           return BinaryOperator::CreateUDiv(A, B);
1379       }
1380       return nullptr;
1381     };
1382 
1383     if (match(Op1, m_c_Mul(m_Specific(X), m_Value(Z)))) {
1384       if (auto *Val = CreateDivOrNull(Y, Z))
1385         return Val;
1386     }
1387     if (match(Op1, m_c_Mul(m_Specific(Y), m_Value(Z)))) {
1388       if (auto *Val = CreateDivOrNull(X, Z))
1389         return Val;
1390     }
1391   }
1392   return nullptr;
1393 }
1394 
1395 static const unsigned MaxDepth = 6;
1396 
1397 // Take the exact integer log2 of the value. If DoFold is true, create the
1398 // actual instructions, otherwise return a non-null dummy value. Return nullptr
1399 // on failure.
takeLog2(IRBuilderBase & Builder,Value * Op,unsigned Depth,bool AssumeNonZero,bool DoFold)1400 static Value *takeLog2(IRBuilderBase &Builder, Value *Op, unsigned Depth,
1401                        bool AssumeNonZero, bool DoFold) {
1402   auto IfFold = [DoFold](function_ref<Value *()> Fn) {
1403     if (!DoFold)
1404       return reinterpret_cast<Value *>(-1);
1405     return Fn();
1406   };
1407 
1408   // FIXME: assert that Op1 isn't/doesn't contain undef.
1409 
1410   // log2(2^C) -> C
1411   if (match(Op, m_Power2()))
1412     return IfFold([&]() {
1413       Constant *C = ConstantExpr::getExactLogBase2(cast<Constant>(Op));
1414       if (!C)
1415         llvm_unreachable("Failed to constant fold udiv -> logbase2");
1416       return C;
1417     });
1418 
1419   // The remaining tests are all recursive, so bail out if we hit the limit.
1420   if (Depth++ == MaxDepth)
1421     return nullptr;
1422 
1423   // log2(zext X) -> zext log2(X)
1424   // FIXME: Require one use?
1425   Value *X, *Y;
1426   if (match(Op, m_ZExt(m_Value(X))))
1427     if (Value *LogX = takeLog2(Builder, X, Depth, AssumeNonZero, DoFold))
1428       return IfFold([&]() { return Builder.CreateZExt(LogX, Op->getType()); });
1429 
1430   // log2(X << Y) -> log2(X) + Y
1431   // FIXME: Require one use unless X is 1?
1432   if (match(Op, m_Shl(m_Value(X), m_Value(Y)))) {
1433     auto *BO = cast<OverflowingBinaryOperator>(Op);
1434     // nuw will be set if the `shl` is trivially non-zero.
1435     if (AssumeNonZero || BO->hasNoUnsignedWrap() || BO->hasNoSignedWrap())
1436       if (Value *LogX = takeLog2(Builder, X, Depth, AssumeNonZero, DoFold))
1437         return IfFold([&]() { return Builder.CreateAdd(LogX, Y); });
1438   }
1439 
1440   // log2(Cond ? X : Y) -> Cond ? log2(X) : log2(Y)
1441   // FIXME: Require one use?
1442   if (SelectInst *SI = dyn_cast<SelectInst>(Op))
1443     if (Value *LogX = takeLog2(Builder, SI->getOperand(1), Depth,
1444                                AssumeNonZero, DoFold))
1445       if (Value *LogY = takeLog2(Builder, SI->getOperand(2), Depth,
1446                                  AssumeNonZero, DoFold))
1447         return IfFold([&]() {
1448           return Builder.CreateSelect(SI->getOperand(0), LogX, LogY);
1449         });
1450 
1451   // log2(umin(X, Y)) -> umin(log2(X), log2(Y))
1452   // log2(umax(X, Y)) -> umax(log2(X), log2(Y))
1453   auto *MinMax = dyn_cast<MinMaxIntrinsic>(Op);
1454   if (MinMax && MinMax->hasOneUse() && !MinMax->isSigned()) {
1455     // Use AssumeNonZero as false here. Otherwise we can hit case where
1456     // log2(umax(X, Y)) != umax(log2(X), log2(Y)) (because overflow).
1457     if (Value *LogX = takeLog2(Builder, MinMax->getLHS(), Depth,
1458                                /*AssumeNonZero*/ false, DoFold))
1459       if (Value *LogY = takeLog2(Builder, MinMax->getRHS(), Depth,
1460                                  /*AssumeNonZero*/ false, DoFold))
1461         return IfFold([&]() {
1462           return Builder.CreateBinaryIntrinsic(MinMax->getIntrinsicID(), LogX,
1463                                                LogY);
1464         });
1465   }
1466 
1467   return nullptr;
1468 }
1469 
1470 /// If we have zero-extended operands of an unsigned div or rem, we may be able
1471 /// to narrow the operation (sink the zext below the math).
narrowUDivURem(BinaryOperator & I,InstCombinerImpl & IC)1472 static Instruction *narrowUDivURem(BinaryOperator &I,
1473                                    InstCombinerImpl &IC) {
1474   Instruction::BinaryOps Opcode = I.getOpcode();
1475   Value *N = I.getOperand(0);
1476   Value *D = I.getOperand(1);
1477   Type *Ty = I.getType();
1478   Value *X, *Y;
1479   if (match(N, m_ZExt(m_Value(X))) && match(D, m_ZExt(m_Value(Y))) &&
1480       X->getType() == Y->getType() && (N->hasOneUse() || D->hasOneUse())) {
1481     // udiv (zext X), (zext Y) --> zext (udiv X, Y)
1482     // urem (zext X), (zext Y) --> zext (urem X, Y)
1483     Value *NarrowOp = IC.Builder.CreateBinOp(Opcode, X, Y);
1484     return new ZExtInst(NarrowOp, Ty);
1485   }
1486 
1487   Constant *C;
1488   if (isa<Instruction>(N) && match(N, m_OneUse(m_ZExt(m_Value(X)))) &&
1489       match(D, m_Constant(C))) {
1490     // If the constant is the same in the smaller type, use the narrow version.
1491     Constant *TruncC = IC.getLosslessUnsignedTrunc(C, X->getType());
1492     if (!TruncC)
1493       return nullptr;
1494 
1495     // udiv (zext X), C --> zext (udiv X, C')
1496     // urem (zext X), C --> zext (urem X, C')
1497     return new ZExtInst(IC.Builder.CreateBinOp(Opcode, X, TruncC), Ty);
1498   }
1499   if (isa<Instruction>(D) && match(D, m_OneUse(m_ZExt(m_Value(X)))) &&
1500       match(N, m_Constant(C))) {
1501     // If the constant is the same in the smaller type, use the narrow version.
1502     Constant *TruncC = IC.getLosslessUnsignedTrunc(C, X->getType());
1503     if (!TruncC)
1504       return nullptr;
1505 
1506     // udiv C, (zext X) --> zext (udiv C', X)
1507     // urem C, (zext X) --> zext (urem C', X)
1508     return new ZExtInst(IC.Builder.CreateBinOp(Opcode, TruncC, X), Ty);
1509   }
1510 
1511   return nullptr;
1512 }
1513 
visitUDiv(BinaryOperator & I)1514 Instruction *InstCombinerImpl::visitUDiv(BinaryOperator &I) {
1515   if (Value *V = simplifyUDivInst(I.getOperand(0), I.getOperand(1), I.isExact(),
1516                                   SQ.getWithInstruction(&I)))
1517     return replaceInstUsesWith(I, V);
1518 
1519   if (Instruction *X = foldVectorBinop(I))
1520     return X;
1521 
1522   // Handle the integer div common cases
1523   if (Instruction *Common = commonIDivTransforms(I))
1524     return Common;
1525 
1526   Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
1527   Value *X;
1528   const APInt *C1, *C2;
1529   if (match(Op0, m_LShr(m_Value(X), m_APInt(C1))) && match(Op1, m_APInt(C2))) {
1530     // (X lshr C1) udiv C2 --> X udiv (C2 << C1)
1531     bool Overflow;
1532     APInt C2ShlC1 = C2->ushl_ov(*C1, Overflow);
1533     if (!Overflow) {
1534       bool IsExact = I.isExact() && match(Op0, m_Exact(m_Value()));
1535       BinaryOperator *BO = BinaryOperator::CreateUDiv(
1536           X, ConstantInt::get(X->getType(), C2ShlC1));
1537       if (IsExact)
1538         BO->setIsExact();
1539       return BO;
1540     }
1541   }
1542 
1543   // Op0 / C where C is large (negative) --> zext (Op0 >= C)
1544   // TODO: Could use isKnownNegative() to handle non-constant values.
1545   Type *Ty = I.getType();
1546   if (match(Op1, m_Negative())) {
1547     Value *Cmp = Builder.CreateICmpUGE(Op0, Op1);
1548     return CastInst::CreateZExtOrBitCast(Cmp, Ty);
1549   }
1550   // Op0 / (sext i1 X) --> zext (Op0 == -1) (if X is 0, the div is undefined)
1551   if (match(Op1, m_SExt(m_Value(X))) && X->getType()->isIntOrIntVectorTy(1)) {
1552     Value *Cmp = Builder.CreateICmpEQ(Op0, ConstantInt::getAllOnesValue(Ty));
1553     return CastInst::CreateZExtOrBitCast(Cmp, Ty);
1554   }
1555 
1556   if (Instruction *NarrowDiv = narrowUDivURem(I, *this))
1557     return NarrowDiv;
1558 
1559   Value *A, *B;
1560 
1561   // Look through a right-shift to find the common factor:
1562   // ((Op1 *nuw A) >> B) / Op1 --> A >> B
1563   if (match(Op0, m_LShr(m_NUWMul(m_Specific(Op1), m_Value(A)), m_Value(B))) ||
1564       match(Op0, m_LShr(m_NUWMul(m_Value(A), m_Specific(Op1)), m_Value(B)))) {
1565     Instruction *Lshr = BinaryOperator::CreateLShr(A, B);
1566     if (I.isExact() && cast<PossiblyExactOperator>(Op0)->isExact())
1567       Lshr->setIsExact();
1568     return Lshr;
1569   }
1570 
1571   // Op1 udiv Op2 -> Op1 lshr log2(Op2), if log2() folds away.
1572   if (takeLog2(Builder, Op1, /*Depth*/ 0, /*AssumeNonZero*/ true,
1573                /*DoFold*/ false)) {
1574     Value *Res = takeLog2(Builder, Op1, /*Depth*/ 0,
1575                           /*AssumeNonZero*/ true, /*DoFold*/ true);
1576     return replaceInstUsesWith(
1577         I, Builder.CreateLShr(Op0, Res, I.getName(), I.isExact()));
1578   }
1579 
1580   return nullptr;
1581 }
1582 
visitSDiv(BinaryOperator & I)1583 Instruction *InstCombinerImpl::visitSDiv(BinaryOperator &I) {
1584   if (Value *V = simplifySDivInst(I.getOperand(0), I.getOperand(1), I.isExact(),
1585                                   SQ.getWithInstruction(&I)))
1586     return replaceInstUsesWith(I, V);
1587 
1588   if (Instruction *X = foldVectorBinop(I))
1589     return X;
1590 
1591   // Handle the integer div common cases
1592   if (Instruction *Common = commonIDivTransforms(I))
1593     return Common;
1594 
1595   Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
1596   Type *Ty = I.getType();
1597   Value *X;
1598   // sdiv Op0, -1 --> -Op0
1599   // sdiv Op0, (sext i1 X) --> -Op0 (because if X is 0, the op is undefined)
1600   if (match(Op1, m_AllOnes()) ||
1601       (match(Op1, m_SExt(m_Value(X))) && X->getType()->isIntOrIntVectorTy(1)))
1602     return BinaryOperator::CreateNSWNeg(Op0);
1603 
1604   // X / INT_MIN --> X == INT_MIN
1605   if (match(Op1, m_SignMask()))
1606     return new ZExtInst(Builder.CreateICmpEQ(Op0, Op1), Ty);
1607 
1608   if (I.isExact()) {
1609     // sdiv exact X, 1<<C --> ashr exact X, C   iff  1<<C  is non-negative
1610     if (match(Op1, m_Power2()) && match(Op1, m_NonNegative())) {
1611       Constant *C = ConstantExpr::getExactLogBase2(cast<Constant>(Op1));
1612       return BinaryOperator::CreateExactAShr(Op0, C);
1613     }
1614 
1615     // sdiv exact X, (1<<ShAmt) --> ashr exact X, ShAmt (if shl is non-negative)
1616     Value *ShAmt;
1617     if (match(Op1, m_NSWShl(m_One(), m_Value(ShAmt))))
1618       return BinaryOperator::CreateExactAShr(Op0, ShAmt);
1619 
1620     // sdiv exact X, -1<<C --> -(ashr exact X, C)
1621     if (match(Op1, m_NegatedPower2())) {
1622       Constant *NegPow2C = ConstantExpr::getNeg(cast<Constant>(Op1));
1623       Constant *C = ConstantExpr::getExactLogBase2(NegPow2C);
1624       Value *Ashr = Builder.CreateAShr(Op0, C, I.getName() + ".neg", true);
1625       return BinaryOperator::CreateNSWNeg(Ashr);
1626     }
1627   }
1628 
1629   const APInt *Op1C;
1630   if (match(Op1, m_APInt(Op1C))) {
1631     // If the dividend is sign-extended and the constant divisor is small enough
1632     // to fit in the source type, shrink the division to the narrower type:
1633     // (sext X) sdiv C --> sext (X sdiv C)
1634     Value *Op0Src;
1635     if (match(Op0, m_OneUse(m_SExt(m_Value(Op0Src)))) &&
1636         Op0Src->getType()->getScalarSizeInBits() >=
1637             Op1C->getSignificantBits()) {
1638 
1639       // In the general case, we need to make sure that the dividend is not the
1640       // minimum signed value because dividing that by -1 is UB. But here, we
1641       // know that the -1 divisor case is already handled above.
1642 
1643       Constant *NarrowDivisor =
1644           ConstantExpr::getTrunc(cast<Constant>(Op1), Op0Src->getType());
1645       Value *NarrowOp = Builder.CreateSDiv(Op0Src, NarrowDivisor);
1646       return new SExtInst(NarrowOp, Ty);
1647     }
1648 
1649     // -X / C --> X / -C (if the negation doesn't overflow).
1650     // TODO: This could be enhanced to handle arbitrary vector constants by
1651     //       checking if all elements are not the min-signed-val.
1652     if (!Op1C->isMinSignedValue() && match(Op0, m_NSWNeg(m_Value(X)))) {
1653       Constant *NegC = ConstantInt::get(Ty, -(*Op1C));
1654       Instruction *BO = BinaryOperator::CreateSDiv(X, NegC);
1655       BO->setIsExact(I.isExact());
1656       return BO;
1657     }
1658   }
1659 
1660   // -X / Y --> -(X / Y)
1661   Value *Y;
1662   if (match(&I, m_SDiv(m_OneUse(m_NSWNeg(m_Value(X))), m_Value(Y))))
1663     return BinaryOperator::CreateNSWNeg(
1664         Builder.CreateSDiv(X, Y, I.getName(), I.isExact()));
1665 
1666   // abs(X) / X --> X > -1 ? 1 : -1
1667   // X / abs(X) --> X > -1 ? 1 : -1
1668   if (match(&I, m_c_BinOp(
1669                     m_OneUse(m_Intrinsic<Intrinsic::abs>(m_Value(X), m_One())),
1670                     m_Deferred(X)))) {
1671     Value *Cond = Builder.CreateIsNotNeg(X);
1672     return SelectInst::Create(Cond, ConstantInt::get(Ty, 1),
1673                               ConstantInt::getAllOnesValue(Ty));
1674   }
1675 
1676   KnownBits KnownDividend = computeKnownBits(Op0, 0, &I);
1677   if (!I.isExact() &&
1678       (match(Op1, m_Power2(Op1C)) || match(Op1, m_NegatedPower2(Op1C))) &&
1679       KnownDividend.countMinTrailingZeros() >= Op1C->countr_zero()) {
1680     I.setIsExact();
1681     return &I;
1682   }
1683 
1684   if (KnownDividend.isNonNegative()) {
1685     // If both operands are unsigned, turn this into a udiv.
1686     if (isKnownNonNegative(Op1, SQ.getWithInstruction(&I))) {
1687       auto *BO = BinaryOperator::CreateUDiv(Op0, Op1, I.getName());
1688       BO->setIsExact(I.isExact());
1689       return BO;
1690     }
1691 
1692     if (match(Op1, m_NegatedPower2())) {
1693       // X sdiv (-(1 << C)) -> -(X sdiv (1 << C)) ->
1694       //                    -> -(X udiv (1 << C)) -> -(X u>> C)
1695       Constant *CNegLog2 = ConstantExpr::getExactLogBase2(
1696           ConstantExpr::getNeg(cast<Constant>(Op1)));
1697       Value *Shr = Builder.CreateLShr(Op0, CNegLog2, I.getName(), I.isExact());
1698       return BinaryOperator::CreateNeg(Shr);
1699     }
1700 
1701     if (isKnownToBeAPowerOfTwo(Op1, /*OrZero*/ true, 0, &I)) {
1702       // X sdiv (1 << Y) -> X udiv (1 << Y) ( -> X u>> Y)
1703       // Safe because the only negative value (1 << Y) can take on is
1704       // INT_MIN, and X sdiv INT_MIN == X udiv INT_MIN == 0 if X doesn't have
1705       // the sign bit set.
1706       auto *BO = BinaryOperator::CreateUDiv(Op0, Op1, I.getName());
1707       BO->setIsExact(I.isExact());
1708       return BO;
1709     }
1710   }
1711 
1712   // -X / X --> X == INT_MIN ? 1 : -1
1713   if (isKnownNegation(Op0, Op1)) {
1714     APInt MinVal = APInt::getSignedMinValue(Ty->getScalarSizeInBits());
1715     Value *Cond = Builder.CreateICmpEQ(Op0, ConstantInt::get(Ty, MinVal));
1716     return SelectInst::Create(Cond, ConstantInt::get(Ty, 1),
1717                               ConstantInt::getAllOnesValue(Ty));
1718   }
1719   return nullptr;
1720 }
1721 
1722 /// Remove negation and try to convert division into multiplication.
foldFDivConstantDivisor(BinaryOperator & I)1723 Instruction *InstCombinerImpl::foldFDivConstantDivisor(BinaryOperator &I) {
1724   Constant *C;
1725   if (!match(I.getOperand(1), m_Constant(C)))
1726     return nullptr;
1727 
1728   // -X / C --> X / -C
1729   Value *X;
1730   const DataLayout &DL = I.getDataLayout();
1731   if (match(I.getOperand(0), m_FNeg(m_Value(X))))
1732     if (Constant *NegC = ConstantFoldUnaryOpOperand(Instruction::FNeg, C, DL))
1733       return BinaryOperator::CreateFDivFMF(X, NegC, &I);
1734 
1735   // nnan X / +0.0 -> copysign(inf, X)
1736   // nnan nsz X / -0.0 -> copysign(inf, X)
1737   if (I.hasNoNaNs() &&
1738       (match(I.getOperand(1), m_PosZeroFP()) ||
1739        (I.hasNoSignedZeros() && match(I.getOperand(1), m_AnyZeroFP())))) {
1740     IRBuilder<> B(&I);
1741     CallInst *CopySign = B.CreateIntrinsic(
1742         Intrinsic::copysign, {C->getType()},
1743         {ConstantFP::getInfinity(I.getType()), I.getOperand(0)}, &I);
1744     CopySign->takeName(&I);
1745     return replaceInstUsesWith(I, CopySign);
1746   }
1747 
1748   // If the constant divisor has an exact inverse, this is always safe. If not,
1749   // then we can still create a reciprocal if fast-math-flags allow it and the
1750   // constant is a regular number (not zero, infinite, or denormal).
1751   if (!(C->hasExactInverseFP() || (I.hasAllowReciprocal() && C->isNormalFP())))
1752     return nullptr;
1753 
1754   // Disallow denormal constants because we don't know what would happen
1755   // on all targets.
1756   // TODO: Use Intrinsic::canonicalize or let function attributes tell us that
1757   // denorms are flushed?
1758   auto *RecipC = ConstantFoldBinaryOpOperands(
1759       Instruction::FDiv, ConstantFP::get(I.getType(), 1.0), C, DL);
1760   if (!RecipC || !RecipC->isNormalFP())
1761     return nullptr;
1762 
1763   // X / C --> X * (1 / C)
1764   return BinaryOperator::CreateFMulFMF(I.getOperand(0), RecipC, &I);
1765 }
1766 
1767 /// Remove negation and try to reassociate constant math.
foldFDivConstantDividend(BinaryOperator & I)1768 static Instruction *foldFDivConstantDividend(BinaryOperator &I) {
1769   Constant *C;
1770   if (!match(I.getOperand(0), m_Constant(C)))
1771     return nullptr;
1772 
1773   // C / -X --> -C / X
1774   Value *X;
1775   const DataLayout &DL = I.getDataLayout();
1776   if (match(I.getOperand(1), m_FNeg(m_Value(X))))
1777     if (Constant *NegC = ConstantFoldUnaryOpOperand(Instruction::FNeg, C, DL))
1778       return BinaryOperator::CreateFDivFMF(NegC, X, &I);
1779 
1780   if (!I.hasAllowReassoc() || !I.hasAllowReciprocal())
1781     return nullptr;
1782 
1783   // Try to reassociate C / X expressions where X includes another constant.
1784   Constant *C2, *NewC = nullptr;
1785   if (match(I.getOperand(1), m_FMul(m_Value(X), m_Constant(C2)))) {
1786     // C / (X * C2) --> (C / C2) / X
1787     NewC = ConstantFoldBinaryOpOperands(Instruction::FDiv, C, C2, DL);
1788   } else if (match(I.getOperand(1), m_FDiv(m_Value(X), m_Constant(C2)))) {
1789     // C / (X / C2) --> (C * C2) / X
1790     NewC = ConstantFoldBinaryOpOperands(Instruction::FMul, C, C2, DL);
1791   }
1792   // Disallow denormal constants because we don't know what would happen
1793   // on all targets.
1794   // TODO: Use Intrinsic::canonicalize or let function attributes tell us that
1795   // denorms are flushed?
1796   if (!NewC || !NewC->isNormalFP())
1797     return nullptr;
1798 
1799   return BinaryOperator::CreateFDivFMF(NewC, X, &I);
1800 }
1801 
1802 /// Negate the exponent of pow/exp to fold division-by-pow() into multiply.
foldFDivPowDivisor(BinaryOperator & I,InstCombiner::BuilderTy & Builder)1803 static Instruction *foldFDivPowDivisor(BinaryOperator &I,
1804                                        InstCombiner::BuilderTy &Builder) {
1805   Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
1806   auto *II = dyn_cast<IntrinsicInst>(Op1);
1807   if (!II || !II->hasOneUse() || !I.hasAllowReassoc() ||
1808       !I.hasAllowReciprocal())
1809     return nullptr;
1810 
1811   // Z / pow(X, Y) --> Z * pow(X, -Y)
1812   // Z / exp{2}(Y) --> Z * exp{2}(-Y)
1813   // In the general case, this creates an extra instruction, but fmul allows
1814   // for better canonicalization and optimization than fdiv.
1815   Intrinsic::ID IID = II->getIntrinsicID();
1816   SmallVector<Value *> Args;
1817   switch (IID) {
1818   case Intrinsic::pow:
1819     Args.push_back(II->getArgOperand(0));
1820     Args.push_back(Builder.CreateFNegFMF(II->getArgOperand(1), &I));
1821     break;
1822   case Intrinsic::powi: {
1823     // Require 'ninf' assuming that makes powi(X, -INT_MIN) acceptable.
1824     // That is, X ** (huge negative number) is 0.0, ~1.0, or INF and so
1825     // dividing by that is INF, ~1.0, or 0.0. Code that uses powi allows
1826     // non-standard results, so this corner case should be acceptable if the
1827     // code rules out INF values.
1828     if (!I.hasNoInfs())
1829       return nullptr;
1830     Args.push_back(II->getArgOperand(0));
1831     Args.push_back(Builder.CreateNeg(II->getArgOperand(1)));
1832     Type *Tys[] = {I.getType(), II->getArgOperand(1)->getType()};
1833     Value *Pow = Builder.CreateIntrinsic(IID, Tys, Args, &I);
1834     return BinaryOperator::CreateFMulFMF(Op0, Pow, &I);
1835   }
1836   case Intrinsic::exp:
1837   case Intrinsic::exp2:
1838     Args.push_back(Builder.CreateFNegFMF(II->getArgOperand(0), &I));
1839     break;
1840   default:
1841     return nullptr;
1842   }
1843   Value *Pow = Builder.CreateIntrinsic(IID, I.getType(), Args, &I);
1844   return BinaryOperator::CreateFMulFMF(Op0, Pow, &I);
1845 }
1846 
1847 /// Convert div to mul if we have an sqrt divisor iff sqrt's operand is a fdiv
1848 /// instruction.
foldFDivSqrtDivisor(BinaryOperator & I,InstCombiner::BuilderTy & Builder)1849 static Instruction *foldFDivSqrtDivisor(BinaryOperator &I,
1850                                         InstCombiner::BuilderTy &Builder) {
1851   // X / sqrt(Y / Z) -->  X * sqrt(Z / Y)
1852   if (!I.hasAllowReassoc() || !I.hasAllowReciprocal())
1853     return nullptr;
1854   Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
1855   auto *II = dyn_cast<IntrinsicInst>(Op1);
1856   if (!II || II->getIntrinsicID() != Intrinsic::sqrt || !II->hasOneUse() ||
1857       !II->hasAllowReassoc() || !II->hasAllowReciprocal())
1858     return nullptr;
1859 
1860   Value *Y, *Z;
1861   auto *DivOp = dyn_cast<Instruction>(II->getOperand(0));
1862   if (!DivOp)
1863     return nullptr;
1864   if (!match(DivOp, m_FDiv(m_Value(Y), m_Value(Z))))
1865     return nullptr;
1866   if (!DivOp->hasAllowReassoc() || !I.hasAllowReciprocal() ||
1867       !DivOp->hasOneUse())
1868     return nullptr;
1869   Value *SwapDiv = Builder.CreateFDivFMF(Z, Y, DivOp);
1870   Value *NewSqrt =
1871       Builder.CreateUnaryIntrinsic(II->getIntrinsicID(), SwapDiv, II);
1872   return BinaryOperator::CreateFMulFMF(Op0, NewSqrt, &I);
1873 }
1874 
visitFDiv(BinaryOperator & I)1875 Instruction *InstCombinerImpl::visitFDiv(BinaryOperator &I) {
1876   Module *M = I.getModule();
1877 
1878   if (Value *V = simplifyFDivInst(I.getOperand(0), I.getOperand(1),
1879                                   I.getFastMathFlags(),
1880                                   SQ.getWithInstruction(&I)))
1881     return replaceInstUsesWith(I, V);
1882 
1883   if (Instruction *X = foldVectorBinop(I))
1884     return X;
1885 
1886   if (Instruction *Phi = foldBinopWithPhiOperands(I))
1887     return Phi;
1888 
1889   if (Instruction *R = foldFDivConstantDivisor(I))
1890     return R;
1891 
1892   if (Instruction *R = foldFDivConstantDividend(I))
1893     return R;
1894 
1895   if (Instruction *R = foldFPSignBitOps(I))
1896     return R;
1897 
1898   Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
1899   if (isa<Constant>(Op0))
1900     if (SelectInst *SI = dyn_cast<SelectInst>(Op1))
1901       if (Instruction *R = FoldOpIntoSelect(I, SI))
1902         return R;
1903 
1904   if (isa<Constant>(Op1))
1905     if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
1906       if (Instruction *R = FoldOpIntoSelect(I, SI))
1907         return R;
1908 
1909   if (I.hasAllowReassoc() && I.hasAllowReciprocal()) {
1910     Value *X, *Y;
1911     if (match(Op0, m_OneUse(m_FDiv(m_Value(X), m_Value(Y)))) &&
1912         (!isa<Constant>(Y) || !isa<Constant>(Op1))) {
1913       // (X / Y) / Z => X / (Y * Z)
1914       Value *YZ = Builder.CreateFMulFMF(Y, Op1, &I);
1915       return BinaryOperator::CreateFDivFMF(X, YZ, &I);
1916     }
1917     if (match(Op1, m_OneUse(m_FDiv(m_Value(X), m_Value(Y)))) &&
1918         (!isa<Constant>(Y) || !isa<Constant>(Op0))) {
1919       // Z / (X / Y) => (Y * Z) / X
1920       Value *YZ = Builder.CreateFMulFMF(Y, Op0, &I);
1921       return BinaryOperator::CreateFDivFMF(YZ, X, &I);
1922     }
1923     // Z / (1.0 / Y) => (Y * Z)
1924     //
1925     // This is a special case of Z / (X / Y) => (Y * Z) / X, with X = 1.0. The
1926     // m_OneUse check is avoided because even in the case of the multiple uses
1927     // for 1.0/Y, the number of instructions remain the same and a division is
1928     // replaced by a multiplication.
1929     if (match(Op1, m_FDiv(m_SpecificFP(1.0), m_Value(Y))))
1930       return BinaryOperator::CreateFMulFMF(Y, Op0, &I);
1931   }
1932 
1933   if (I.hasAllowReassoc() && Op0->hasOneUse() && Op1->hasOneUse()) {
1934     // sin(X) / cos(X) -> tan(X)
1935     // cos(X) / sin(X) -> 1/tan(X) (cotangent)
1936     Value *X;
1937     bool IsTan = match(Op0, m_Intrinsic<Intrinsic::sin>(m_Value(X))) &&
1938                  match(Op1, m_Intrinsic<Intrinsic::cos>(m_Specific(X)));
1939     bool IsCot =
1940         !IsTan && match(Op0, m_Intrinsic<Intrinsic::cos>(m_Value(X))) &&
1941                   match(Op1, m_Intrinsic<Intrinsic::sin>(m_Specific(X)));
1942 
1943     if ((IsTan || IsCot) && hasFloatFn(M, &TLI, I.getType(), LibFunc_tan,
1944                                        LibFunc_tanf, LibFunc_tanl)) {
1945       IRBuilder<> B(&I);
1946       IRBuilder<>::FastMathFlagGuard FMFGuard(B);
1947       B.setFastMathFlags(I.getFastMathFlags());
1948       AttributeList Attrs =
1949           cast<CallBase>(Op0)->getCalledFunction()->getAttributes();
1950       Value *Res = emitUnaryFloatFnCall(X, &TLI, LibFunc_tan, LibFunc_tanf,
1951                                         LibFunc_tanl, B, Attrs);
1952       if (IsCot)
1953         Res = B.CreateFDiv(ConstantFP::get(I.getType(), 1.0), Res);
1954       return replaceInstUsesWith(I, Res);
1955     }
1956   }
1957 
1958   // X / (X * Y) --> 1.0 / Y
1959   // Reassociate to (X / X -> 1.0) is legal when NaNs are not allowed.
1960   // We can ignore the possibility that X is infinity because INF/INF is NaN.
1961   Value *X, *Y;
1962   if (I.hasNoNaNs() && I.hasAllowReassoc() &&
1963       match(Op1, m_c_FMul(m_Specific(Op0), m_Value(Y)))) {
1964     replaceOperand(I, 0, ConstantFP::get(I.getType(), 1.0));
1965     replaceOperand(I, 1, Y);
1966     return &I;
1967   }
1968 
1969   // X / fabs(X) -> copysign(1.0, X)
1970   // fabs(X) / X -> copysign(1.0, X)
1971   if (I.hasNoNaNs() && I.hasNoInfs() &&
1972       (match(&I, m_FDiv(m_Value(X), m_FAbs(m_Deferred(X)))) ||
1973        match(&I, m_FDiv(m_FAbs(m_Value(X)), m_Deferred(X))))) {
1974     Value *V = Builder.CreateBinaryIntrinsic(
1975         Intrinsic::copysign, ConstantFP::get(I.getType(), 1.0), X, &I);
1976     return replaceInstUsesWith(I, V);
1977   }
1978 
1979   if (Instruction *Mul = foldFDivPowDivisor(I, Builder))
1980     return Mul;
1981 
1982   if (Instruction *Mul = foldFDivSqrtDivisor(I, Builder))
1983     return Mul;
1984 
1985   // pow(X, Y) / X --> pow(X, Y-1)
1986   if (I.hasAllowReassoc() &&
1987       match(Op0, m_OneUse(m_Intrinsic<Intrinsic::pow>(m_Specific(Op1),
1988                                                       m_Value(Y))))) {
1989     Value *Y1 =
1990         Builder.CreateFAddFMF(Y, ConstantFP::get(I.getType(), -1.0), &I);
1991     Value *Pow = Builder.CreateBinaryIntrinsic(Intrinsic::pow, Op1, Y1, &I);
1992     return replaceInstUsesWith(I, Pow);
1993   }
1994 
1995   if (Instruction *FoldedPowi = foldPowiReassoc(I))
1996     return FoldedPowi;
1997 
1998   return nullptr;
1999 }
2000 
2001 // Variety of transform for:
2002 //  (urem/srem (mul X, Y), (mul X, Z))
2003 //  (urem/srem (shl X, Y), (shl X, Z))
2004 //  (urem/srem (shl Y, X), (shl Z, X))
2005 // NB: The shift cases are really just extensions of the mul case. We treat
2006 // shift as Val * (1 << Amt).
simplifyIRemMulShl(BinaryOperator & I,InstCombinerImpl & IC)2007 static Instruction *simplifyIRemMulShl(BinaryOperator &I,
2008                                        InstCombinerImpl &IC) {
2009   Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1), *X = nullptr;
2010   APInt Y, Z;
2011   bool ShiftByX = false;
2012 
2013   // If V is not nullptr, it will be matched using m_Specific.
2014   auto MatchShiftOrMulXC = [](Value *Op, Value *&V, APInt &C) -> bool {
2015     const APInt *Tmp = nullptr;
2016     if ((!V && match(Op, m_Mul(m_Value(V), m_APInt(Tmp)))) ||
2017         (V && match(Op, m_Mul(m_Specific(V), m_APInt(Tmp)))))
2018       C = *Tmp;
2019     else if ((!V && match(Op, m_Shl(m_Value(V), m_APInt(Tmp)))) ||
2020              (V && match(Op, m_Shl(m_Specific(V), m_APInt(Tmp)))))
2021       C = APInt(Tmp->getBitWidth(), 1) << *Tmp;
2022     if (Tmp != nullptr)
2023       return true;
2024 
2025     // Reset `V` so we don't start with specific value on next match attempt.
2026     V = nullptr;
2027     return false;
2028   };
2029 
2030   auto MatchShiftCX = [](Value *Op, APInt &C, Value *&V) -> bool {
2031     const APInt *Tmp = nullptr;
2032     if ((!V && match(Op, m_Shl(m_APInt(Tmp), m_Value(V)))) ||
2033         (V && match(Op, m_Shl(m_APInt(Tmp), m_Specific(V))))) {
2034       C = *Tmp;
2035       return true;
2036     }
2037 
2038     // Reset `V` so we don't start with specific value on next match attempt.
2039     V = nullptr;
2040     return false;
2041   };
2042 
2043   if (MatchShiftOrMulXC(Op0, X, Y) && MatchShiftOrMulXC(Op1, X, Z)) {
2044     // pass
2045   } else if (MatchShiftCX(Op0, Y, X) && MatchShiftCX(Op1, Z, X)) {
2046     ShiftByX = true;
2047   } else {
2048     return nullptr;
2049   }
2050 
2051   bool IsSRem = I.getOpcode() == Instruction::SRem;
2052 
2053   OverflowingBinaryOperator *BO0 = cast<OverflowingBinaryOperator>(Op0);
2054   // TODO: We may be able to deduce more about nsw/nuw of BO0/BO1 based on Y >=
2055   // Z or Z >= Y.
2056   bool BO0HasNSW = BO0->hasNoSignedWrap();
2057   bool BO0HasNUW = BO0->hasNoUnsignedWrap();
2058   bool BO0NoWrap = IsSRem ? BO0HasNSW : BO0HasNUW;
2059 
2060   APInt RemYZ = IsSRem ? Y.srem(Z) : Y.urem(Z);
2061   // (rem (mul nuw/nsw X, Y), (mul X, Z))
2062   //      if (rem Y, Z) == 0
2063   //          -> 0
2064   if (RemYZ.isZero() && BO0NoWrap)
2065     return IC.replaceInstUsesWith(I, ConstantInt::getNullValue(I.getType()));
2066 
2067   // Helper function to emit either (RemSimplificationC << X) or
2068   // (RemSimplificationC * X) depending on whether we matched Op0/Op1 as
2069   // (shl V, X) or (mul V, X) respectively.
2070   auto CreateMulOrShift =
2071       [&](const APInt &RemSimplificationC) -> BinaryOperator * {
2072     Value *RemSimplification =
2073         ConstantInt::get(I.getType(), RemSimplificationC);
2074     return ShiftByX ? BinaryOperator::CreateShl(RemSimplification, X)
2075                     : BinaryOperator::CreateMul(X, RemSimplification);
2076   };
2077 
2078   OverflowingBinaryOperator *BO1 = cast<OverflowingBinaryOperator>(Op1);
2079   bool BO1HasNSW = BO1->hasNoSignedWrap();
2080   bool BO1HasNUW = BO1->hasNoUnsignedWrap();
2081   bool BO1NoWrap = IsSRem ? BO1HasNSW : BO1HasNUW;
2082   // (rem (mul X, Y), (mul nuw/nsw X, Z))
2083   //      if (rem Y, Z) == Y
2084   //          -> (mul nuw/nsw X, Y)
2085   if (RemYZ == Y && BO1NoWrap) {
2086     BinaryOperator *BO = CreateMulOrShift(Y);
2087     // Copy any overflow flags from Op0.
2088     BO->setHasNoSignedWrap(IsSRem || BO0HasNSW);
2089     BO->setHasNoUnsignedWrap(!IsSRem || BO0HasNUW);
2090     return BO;
2091   }
2092 
2093   // (rem (mul nuw/nsw X, Y), (mul {nsw} X, Z))
2094   //      if Y >= Z
2095   //          -> (mul {nuw} nsw X, (rem Y, Z))
2096   if (Y.uge(Z) && (IsSRem ? (BO0HasNSW && BO1HasNSW) : BO0HasNUW)) {
2097     BinaryOperator *BO = CreateMulOrShift(RemYZ);
2098     BO->setHasNoSignedWrap();
2099     BO->setHasNoUnsignedWrap(BO0HasNUW);
2100     return BO;
2101   }
2102 
2103   return nullptr;
2104 }
2105 
2106 /// This function implements the transforms common to both integer remainder
2107 /// instructions (urem and srem). It is called by the visitors to those integer
2108 /// remainder instructions.
2109 /// Common integer remainder transforms
commonIRemTransforms(BinaryOperator & I)2110 Instruction *InstCombinerImpl::commonIRemTransforms(BinaryOperator &I) {
2111   if (Instruction *Phi = foldBinopWithPhiOperands(I))
2112     return Phi;
2113 
2114   Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2115 
2116   // The RHS is known non-zero.
2117   if (Value *V = simplifyValueKnownNonZero(I.getOperand(1), *this, I))
2118     return replaceOperand(I, 1, V);
2119 
2120   // Handle cases involving: rem X, (select Cond, Y, Z)
2121   if (simplifyDivRemOfSelectWithZeroOp(I))
2122     return &I;
2123 
2124   // If the divisor is a select-of-constants, try to constant fold all rem ops:
2125   // C % (select Cond, TrueC, FalseC) --> select Cond, (C % TrueC), (C % FalseC)
2126   // TODO: Adapt simplifyDivRemOfSelectWithZeroOp to allow this and other folds.
2127   if (match(Op0, m_ImmConstant()) &&
2128       match(Op1, m_Select(m_Value(), m_ImmConstant(), m_ImmConstant()))) {
2129     if (Instruction *R = FoldOpIntoSelect(I, cast<SelectInst>(Op1),
2130                                           /*FoldWithMultiUse*/ true))
2131       return R;
2132   }
2133 
2134   if (isa<Constant>(Op1)) {
2135     if (Instruction *Op0I = dyn_cast<Instruction>(Op0)) {
2136       if (SelectInst *SI = dyn_cast<SelectInst>(Op0I)) {
2137         if (Instruction *R = FoldOpIntoSelect(I, SI))
2138           return R;
2139       } else if (auto *PN = dyn_cast<PHINode>(Op0I)) {
2140         const APInt *Op1Int;
2141         if (match(Op1, m_APInt(Op1Int)) && !Op1Int->isMinValue() &&
2142             (I.getOpcode() == Instruction::URem ||
2143              !Op1Int->isMinSignedValue())) {
2144           // foldOpIntoPhi will speculate instructions to the end of the PHI's
2145           // predecessor blocks, so do this only if we know the srem or urem
2146           // will not fault.
2147           if (Instruction *NV = foldOpIntoPhi(I, PN))
2148             return NV;
2149         }
2150       }
2151 
2152       // See if we can fold away this rem instruction.
2153       if (SimplifyDemandedInstructionBits(I))
2154         return &I;
2155     }
2156   }
2157 
2158   if (Instruction *R = simplifyIRemMulShl(I, *this))
2159     return R;
2160 
2161   return nullptr;
2162 }
2163 
visitURem(BinaryOperator & I)2164 Instruction *InstCombinerImpl::visitURem(BinaryOperator &I) {
2165   if (Value *V = simplifyURemInst(I.getOperand(0), I.getOperand(1),
2166                                   SQ.getWithInstruction(&I)))
2167     return replaceInstUsesWith(I, V);
2168 
2169   if (Instruction *X = foldVectorBinop(I))
2170     return X;
2171 
2172   if (Instruction *common = commonIRemTransforms(I))
2173     return common;
2174 
2175   if (Instruction *NarrowRem = narrowUDivURem(I, *this))
2176     return NarrowRem;
2177 
2178   // X urem Y -> X and Y-1, where Y is a power of 2,
2179   Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2180   Type *Ty = I.getType();
2181   if (isKnownToBeAPowerOfTwo(Op1, /*OrZero*/ true, 0, &I)) {
2182     // This may increase instruction count, we don't enforce that Y is a
2183     // constant.
2184     Constant *N1 = Constant::getAllOnesValue(Ty);
2185     Value *Add = Builder.CreateAdd(Op1, N1);
2186     return BinaryOperator::CreateAnd(Op0, Add);
2187   }
2188 
2189   // 1 urem X -> zext(X != 1)
2190   if (match(Op0, m_One())) {
2191     Value *Cmp = Builder.CreateICmpNE(Op1, ConstantInt::get(Ty, 1));
2192     return CastInst::CreateZExtOrBitCast(Cmp, Ty);
2193   }
2194 
2195   // Op0 urem C -> Op0 < C ? Op0 : Op0 - C, where C >= signbit.
2196   // Op0 must be frozen because we are increasing its number of uses.
2197   if (match(Op1, m_Negative())) {
2198     Value *F0 = Op0;
2199     if (!isGuaranteedNotToBeUndef(Op0))
2200       F0 = Builder.CreateFreeze(Op0, Op0->getName() + ".fr");
2201     Value *Cmp = Builder.CreateICmpULT(F0, Op1);
2202     Value *Sub = Builder.CreateSub(F0, Op1);
2203     return SelectInst::Create(Cmp, F0, Sub);
2204   }
2205 
2206   // If the divisor is a sext of a boolean, then the divisor must be max
2207   // unsigned value (-1). Therefore, the remainder is Op0 unless Op0 is also
2208   // max unsigned value. In that case, the remainder is 0:
2209   // urem Op0, (sext i1 X) --> (Op0 == -1) ? 0 : Op0
2210   Value *X;
2211   if (match(Op1, m_SExt(m_Value(X))) && X->getType()->isIntOrIntVectorTy(1)) {
2212     Value *FrozenOp0 = Op0;
2213     if (!isGuaranteedNotToBeUndef(Op0))
2214       FrozenOp0 = Builder.CreateFreeze(Op0, Op0->getName() + ".frozen");
2215     Value *Cmp =
2216         Builder.CreateICmpEQ(FrozenOp0, ConstantInt::getAllOnesValue(Ty));
2217     return SelectInst::Create(Cmp, ConstantInt::getNullValue(Ty), FrozenOp0);
2218   }
2219 
2220   // For "(X + 1) % Op1" and if (X u< Op1) => (X + 1) == Op1 ? 0 : X + 1 .
2221   if (match(Op0, m_Add(m_Value(X), m_One()))) {
2222     Value *Val =
2223         simplifyICmpInst(ICmpInst::ICMP_ULT, X, Op1, SQ.getWithInstruction(&I));
2224     if (Val && match(Val, m_One())) {
2225       Value *FrozenOp0 = Op0;
2226       if (!isGuaranteedNotToBeUndef(Op0))
2227         FrozenOp0 = Builder.CreateFreeze(Op0, Op0->getName() + ".frozen");
2228       Value *Cmp = Builder.CreateICmpEQ(FrozenOp0, Op1);
2229       return SelectInst::Create(Cmp, ConstantInt::getNullValue(Ty), FrozenOp0);
2230     }
2231   }
2232 
2233   return nullptr;
2234 }
2235 
visitSRem(BinaryOperator & I)2236 Instruction *InstCombinerImpl::visitSRem(BinaryOperator &I) {
2237   if (Value *V = simplifySRemInst(I.getOperand(0), I.getOperand(1),
2238                                   SQ.getWithInstruction(&I)))
2239     return replaceInstUsesWith(I, V);
2240 
2241   if (Instruction *X = foldVectorBinop(I))
2242     return X;
2243 
2244   // Handle the integer rem common cases
2245   if (Instruction *Common = commonIRemTransforms(I))
2246     return Common;
2247 
2248   Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2249   {
2250     const APInt *Y;
2251     // X % -Y -> X % Y
2252     if (match(Op1, m_Negative(Y)) && !Y->isMinSignedValue())
2253       return replaceOperand(I, 1, ConstantInt::get(I.getType(), -*Y));
2254   }
2255 
2256   // -X srem Y --> -(X srem Y)
2257   Value *X, *Y;
2258   if (match(&I, m_SRem(m_OneUse(m_NSWNeg(m_Value(X))), m_Value(Y))))
2259     return BinaryOperator::CreateNSWNeg(Builder.CreateSRem(X, Y));
2260 
2261   // If the sign bits of both operands are zero (i.e. we can prove they are
2262   // unsigned inputs), turn this into a urem.
2263   APInt Mask(APInt::getSignMask(I.getType()->getScalarSizeInBits()));
2264   if (MaskedValueIsZero(Op1, Mask, 0, &I) &&
2265       MaskedValueIsZero(Op0, Mask, 0, &I)) {
2266     // X srem Y -> X urem Y, iff X and Y don't have sign bit set
2267     return BinaryOperator::CreateURem(Op0, Op1, I.getName());
2268   }
2269 
2270   // If it's a constant vector, flip any negative values positive.
2271   if (isa<ConstantVector>(Op1) || isa<ConstantDataVector>(Op1)) {
2272     Constant *C = cast<Constant>(Op1);
2273     unsigned VWidth = cast<FixedVectorType>(C->getType())->getNumElements();
2274 
2275     bool hasNegative = false;
2276     bool hasMissing = false;
2277     for (unsigned i = 0; i != VWidth; ++i) {
2278       Constant *Elt = C->getAggregateElement(i);
2279       if (!Elt) {
2280         hasMissing = true;
2281         break;
2282       }
2283 
2284       if (ConstantInt *RHS = dyn_cast<ConstantInt>(Elt))
2285         if (RHS->isNegative())
2286           hasNegative = true;
2287     }
2288 
2289     if (hasNegative && !hasMissing) {
2290       SmallVector<Constant *, 16> Elts(VWidth);
2291       for (unsigned i = 0; i != VWidth; ++i) {
2292         Elts[i] = C->getAggregateElement(i);  // Handle undef, etc.
2293         if (ConstantInt *RHS = dyn_cast<ConstantInt>(Elts[i])) {
2294           if (RHS->isNegative())
2295             Elts[i] = cast<ConstantInt>(ConstantExpr::getNeg(RHS));
2296         }
2297       }
2298 
2299       Constant *NewRHSV = ConstantVector::get(Elts);
2300       if (NewRHSV != C)  // Don't loop on -MININT
2301         return replaceOperand(I, 1, NewRHSV);
2302     }
2303   }
2304 
2305   return nullptr;
2306 }
2307 
visitFRem(BinaryOperator & I)2308 Instruction *InstCombinerImpl::visitFRem(BinaryOperator &I) {
2309   if (Value *V = simplifyFRemInst(I.getOperand(0), I.getOperand(1),
2310                                   I.getFastMathFlags(),
2311                                   SQ.getWithInstruction(&I)))
2312     return replaceInstUsesWith(I, V);
2313 
2314   if (Instruction *X = foldVectorBinop(I))
2315     return X;
2316 
2317   if (Instruction *Phi = foldBinopWithPhiOperands(I))
2318     return Phi;
2319 
2320   return nullptr;
2321 }
2322