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/APFloat.h"
16 #include "llvm/ADT/APInt.h"
17 #include "llvm/ADT/SmallVector.h"
18 #include "llvm/Analysis/InstructionSimplify.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/Support/KnownBits.h"
34 #include "llvm/Transforms/InstCombine/InstCombineWorklist.h"
35 #include "llvm/Transforms/Utils/BuildLibCalls.h"
36 #include <cassert>
37 #include <cstddef>
38 #include <cstdint>
39 #include <utility>
40 
41 using namespace llvm;
42 using namespace PatternMatch;
43 
44 #define DEBUG_TYPE "instcombine"
45 
46 /// The specific integer value is used in a context where it is known to be
47 /// non-zero.  If this allows us to simplify the computation, do so and return
48 /// the new operand, otherwise return null.
49 static Value *simplifyValueKnownNonZero(Value *V, InstCombiner &IC,
50                                         Instruction &CxtI) {
51   // If V has multiple uses, then we would have to do more analysis to determine
52   // if this is safe.  For example, the use could be in dynamically unreached
53   // code.
54   if (!V->hasOneUse()) return nullptr;
55 
56   bool MadeChange = false;
57 
58   // ((1 << A) >>u B) --> (1 << (A-B))
59   // Because V cannot be zero, we know that B is less than A.
60   Value *A = nullptr, *B = nullptr, *One = nullptr;
61   if (match(V, m_LShr(m_OneUse(m_Shl(m_Value(One), m_Value(A))), m_Value(B))) &&
62       match(One, m_One())) {
63     A = IC.Builder.CreateSub(A, B);
64     return IC.Builder.CreateShl(One, A);
65   }
66 
67   // (PowerOfTwo >>u B) --> isExact since shifting out the result would make it
68   // inexact.  Similarly for <<.
69   BinaryOperator *I = dyn_cast<BinaryOperator>(V);
70   if (I && I->isLogicalShift() &&
71       IC.isKnownToBeAPowerOfTwo(I->getOperand(0), false, 0, &CxtI)) {
72     // We know that this is an exact/nuw shift and that the input is a
73     // non-zero context as well.
74     if (Value *V2 = simplifyValueKnownNonZero(I->getOperand(0), IC, CxtI)) {
75       I->setOperand(0, V2);
76       MadeChange = true;
77     }
78 
79     if (I->getOpcode() == Instruction::LShr && !I->isExact()) {
80       I->setIsExact();
81       MadeChange = true;
82     }
83 
84     if (I->getOpcode() == Instruction::Shl && !I->hasNoUnsignedWrap()) {
85       I->setHasNoUnsignedWrap();
86       MadeChange = true;
87     }
88   }
89 
90   // TODO: Lots more we could do here:
91   //    If V is a phi node, we can call this on each of its operands.
92   //    "select cond, X, 0" can simplify to "X".
93 
94   return MadeChange ? V : nullptr;
95 }
96 
97 /// A helper routine of InstCombiner::visitMul().
98 ///
99 /// If C is a scalar/vector of known powers of 2, then this function returns
100 /// a new scalar/vector obtained from logBase2 of C.
101 /// Return a null pointer otherwise.
102 static Constant *getLogBase2(Type *Ty, Constant *C) {
103   const APInt *IVal;
104   if (match(C, m_APInt(IVal)) && IVal->isPowerOf2())
105     return ConstantInt::get(Ty, IVal->logBase2());
106 
107   if (!Ty->isVectorTy())
108     return nullptr;
109 
110   SmallVector<Constant *, 4> Elts;
111   for (unsigned I = 0, E = Ty->getVectorNumElements(); I != E; ++I) {
112     Constant *Elt = C->getAggregateElement(I);
113     if (!Elt)
114       return nullptr;
115     if (isa<UndefValue>(Elt)) {
116       Elts.push_back(UndefValue::get(Ty->getScalarType()));
117       continue;
118     }
119     if (!match(Elt, m_APInt(IVal)) || !IVal->isPowerOf2())
120       return nullptr;
121     Elts.push_back(ConstantInt::get(Ty->getScalarType(), IVal->logBase2()));
122   }
123 
124   return ConstantVector::get(Elts);
125 }
126 
127 // TODO: This is a specific form of a much more general pattern.
128 //       We could detect a select with any binop identity constant, or we
129 //       could use SimplifyBinOp to see if either arm of the select reduces.
130 //       But that needs to be done carefully and/or while removing potential
131 //       reverse canonicalizations as in InstCombiner::foldSelectIntoOp().
132 static Value *foldMulSelectToNegate(BinaryOperator &I,
133                                     InstCombiner::BuilderTy &Builder) {
134   Value *Cond, *OtherOp;
135 
136   // mul (select Cond, 1, -1), OtherOp --> select Cond, OtherOp, -OtherOp
137   // mul OtherOp, (select Cond, 1, -1) --> select Cond, OtherOp, -OtherOp
138   if (match(&I, m_c_Mul(m_OneUse(m_Select(m_Value(Cond), m_One(), m_AllOnes())),
139                         m_Value(OtherOp))))
140     return Builder.CreateSelect(Cond, OtherOp, Builder.CreateNeg(OtherOp));
141 
142   // mul (select Cond, -1, 1), OtherOp --> select Cond, -OtherOp, OtherOp
143   // mul OtherOp, (select Cond, -1, 1) --> select Cond, -OtherOp, OtherOp
144   if (match(&I, m_c_Mul(m_OneUse(m_Select(m_Value(Cond), m_AllOnes(), m_One())),
145                         m_Value(OtherOp))))
146     return Builder.CreateSelect(Cond, Builder.CreateNeg(OtherOp), OtherOp);
147 
148   // fmul (select Cond, 1.0, -1.0), OtherOp --> select Cond, OtherOp, -OtherOp
149   // fmul OtherOp, (select Cond, 1.0, -1.0) --> select Cond, OtherOp, -OtherOp
150   if (match(&I, m_c_FMul(m_OneUse(m_Select(m_Value(Cond), m_SpecificFP(1.0),
151                                            m_SpecificFP(-1.0))),
152                          m_Value(OtherOp)))) {
153     IRBuilder<>::FastMathFlagGuard FMFGuard(Builder);
154     Builder.setFastMathFlags(I.getFastMathFlags());
155     return Builder.CreateSelect(Cond, OtherOp, Builder.CreateFNeg(OtherOp));
156   }
157 
158   // fmul (select Cond, -1.0, 1.0), OtherOp --> select Cond, -OtherOp, OtherOp
159   // fmul OtherOp, (select Cond, -1.0, 1.0) --> select Cond, -OtherOp, OtherOp
160   if (match(&I, m_c_FMul(m_OneUse(m_Select(m_Value(Cond), m_SpecificFP(-1.0),
161                                            m_SpecificFP(1.0))),
162                          m_Value(OtherOp)))) {
163     IRBuilder<>::FastMathFlagGuard FMFGuard(Builder);
164     Builder.setFastMathFlags(I.getFastMathFlags());
165     return Builder.CreateSelect(Cond, Builder.CreateFNeg(OtherOp), OtherOp);
166   }
167 
168   return nullptr;
169 }
170 
171 Instruction *InstCombiner::visitMul(BinaryOperator &I) {
172   if (Value *V = SimplifyMulInst(I.getOperand(0), I.getOperand(1),
173                                  SQ.getWithInstruction(&I)))
174     return replaceInstUsesWith(I, V);
175 
176   if (SimplifyAssociativeOrCommutative(I))
177     return &I;
178 
179   if (Instruction *X = foldVectorBinop(I))
180     return X;
181 
182   if (Value *V = SimplifyUsingDistributiveLaws(I))
183     return replaceInstUsesWith(I, V);
184 
185   // X * -1 == 0 - X
186   Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
187   if (match(Op1, m_AllOnes())) {
188     BinaryOperator *BO = BinaryOperator::CreateNeg(Op0, I.getName());
189     if (I.hasNoSignedWrap())
190       BO->setHasNoSignedWrap();
191     return BO;
192   }
193 
194   // Also allow combining multiply instructions on vectors.
195   {
196     Value *NewOp;
197     Constant *C1, *C2;
198     const APInt *IVal;
199     if (match(&I, m_Mul(m_Shl(m_Value(NewOp), m_Constant(C2)),
200                         m_Constant(C1))) &&
201         match(C1, m_APInt(IVal))) {
202       // ((X << C2)*C1) == (X * (C1 << C2))
203       Constant *Shl = ConstantExpr::getShl(C1, C2);
204       BinaryOperator *Mul = cast<BinaryOperator>(I.getOperand(0));
205       BinaryOperator *BO = BinaryOperator::CreateMul(NewOp, Shl);
206       if (I.hasNoUnsignedWrap() && Mul->hasNoUnsignedWrap())
207         BO->setHasNoUnsignedWrap();
208       if (I.hasNoSignedWrap() && Mul->hasNoSignedWrap() &&
209           Shl->isNotMinSignedValue())
210         BO->setHasNoSignedWrap();
211       return BO;
212     }
213 
214     if (match(&I, m_Mul(m_Value(NewOp), m_Constant(C1)))) {
215       // Replace X*(2^C) with X << C, where C is either a scalar or a vector.
216       if (Constant *NewCst = getLogBase2(NewOp->getType(), C1)) {
217         BinaryOperator *Shl = BinaryOperator::CreateShl(NewOp, NewCst);
218 
219         if (I.hasNoUnsignedWrap())
220           Shl->setHasNoUnsignedWrap();
221         if (I.hasNoSignedWrap()) {
222           const APInt *V;
223           if (match(NewCst, m_APInt(V)) && *V != V->getBitWidth() - 1)
224             Shl->setHasNoSignedWrap();
225         }
226 
227         return Shl;
228       }
229     }
230   }
231 
232   if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
233     // (Y - X) * (-(2**n)) -> (X - Y) * (2**n), for positive nonzero n
234     // (Y + const) * (-(2**n)) -> (-constY) * (2**n), for positive nonzero n
235     // The "* (2**n)" thus becomes a potential shifting opportunity.
236     {
237       const APInt &   Val = CI->getValue();
238       const APInt &PosVal = Val.abs();
239       if (Val.isNegative() && PosVal.isPowerOf2()) {
240         Value *X = nullptr, *Y = nullptr;
241         if (Op0->hasOneUse()) {
242           ConstantInt *C1;
243           Value *Sub = nullptr;
244           if (match(Op0, m_Sub(m_Value(Y), m_Value(X))))
245             Sub = Builder.CreateSub(X, Y, "suba");
246           else if (match(Op0, m_Add(m_Value(Y), m_ConstantInt(C1))))
247             Sub = Builder.CreateSub(Builder.CreateNeg(C1), Y, "subc");
248           if (Sub)
249             return
250               BinaryOperator::CreateMul(Sub,
251                                         ConstantInt::get(Y->getType(), PosVal));
252         }
253       }
254     }
255   }
256 
257   if (Instruction *FoldedMul = foldBinOpIntoSelectOrPhi(I))
258     return FoldedMul;
259 
260   if (Value *FoldedMul = foldMulSelectToNegate(I, Builder))
261     return replaceInstUsesWith(I, FoldedMul);
262 
263   // Simplify mul instructions with a constant RHS.
264   if (isa<Constant>(Op1)) {
265     // Canonicalize (X+C1)*CI -> X*CI+C1*CI.
266     Value *X;
267     Constant *C1;
268     if (match(Op0, m_OneUse(m_Add(m_Value(X), m_Constant(C1))))) {
269       Value *Mul = Builder.CreateMul(C1, Op1);
270       // Only go forward with the transform if C1*CI simplifies to a tidier
271       // constant.
272       if (!match(Mul, m_Mul(m_Value(), m_Value())))
273         return BinaryOperator::CreateAdd(Builder.CreateMul(X, Op1), Mul);
274     }
275   }
276 
277   // -X * C --> X * -C
278   Value *X, *Y;
279   Constant *Op1C;
280   if (match(Op0, m_Neg(m_Value(X))) && match(Op1, m_Constant(Op1C)))
281     return BinaryOperator::CreateMul(X, ConstantExpr::getNeg(Op1C));
282 
283   // -X * -Y --> X * Y
284   if (match(Op0, m_Neg(m_Value(X))) && match(Op1, m_Neg(m_Value(Y)))) {
285     auto *NewMul = BinaryOperator::CreateMul(X, Y);
286     if (I.hasNoSignedWrap() &&
287         cast<OverflowingBinaryOperator>(Op0)->hasNoSignedWrap() &&
288         cast<OverflowingBinaryOperator>(Op1)->hasNoSignedWrap())
289       NewMul->setHasNoSignedWrap();
290     return NewMul;
291   }
292 
293   // -X * Y --> -(X * Y)
294   // X * -Y --> -(X * Y)
295   if (match(&I, m_c_Mul(m_OneUse(m_Neg(m_Value(X))), m_Value(Y))))
296     return BinaryOperator::CreateNeg(Builder.CreateMul(X, Y));
297 
298   // (X / Y) *  Y = X - (X % Y)
299   // (X / Y) * -Y = (X % Y) - X
300   {
301     Value *Y = Op1;
302     BinaryOperator *Div = dyn_cast<BinaryOperator>(Op0);
303     if (!Div || (Div->getOpcode() != Instruction::UDiv &&
304                  Div->getOpcode() != Instruction::SDiv)) {
305       Y = Op0;
306       Div = dyn_cast<BinaryOperator>(Op1);
307     }
308     Value *Neg = dyn_castNegVal(Y);
309     if (Div && Div->hasOneUse() &&
310         (Div->getOperand(1) == Y || Div->getOperand(1) == Neg) &&
311         (Div->getOpcode() == Instruction::UDiv ||
312          Div->getOpcode() == Instruction::SDiv)) {
313       Value *X = Div->getOperand(0), *DivOp1 = Div->getOperand(1);
314 
315       // If the division is exact, X % Y is zero, so we end up with X or -X.
316       if (Div->isExact()) {
317         if (DivOp1 == Y)
318           return replaceInstUsesWith(I, X);
319         return BinaryOperator::CreateNeg(X);
320       }
321 
322       auto RemOpc = Div->getOpcode() == Instruction::UDiv ? Instruction::URem
323                                                           : Instruction::SRem;
324       Value *Rem = Builder.CreateBinOp(RemOpc, X, DivOp1);
325       if (DivOp1 == Y)
326         return BinaryOperator::CreateSub(X, Rem);
327       return BinaryOperator::CreateSub(Rem, X);
328     }
329   }
330 
331   /// i1 mul -> i1 and.
332   if (I.getType()->isIntOrIntVectorTy(1))
333     return BinaryOperator::CreateAnd(Op0, Op1);
334 
335   // X*(1 << Y) --> X << Y
336   // (1 << Y)*X --> X << Y
337   {
338     Value *Y;
339     BinaryOperator *BO = nullptr;
340     bool ShlNSW = false;
341     if (match(Op0, m_Shl(m_One(), m_Value(Y)))) {
342       BO = BinaryOperator::CreateShl(Op1, Y);
343       ShlNSW = cast<ShlOperator>(Op0)->hasNoSignedWrap();
344     } else if (match(Op1, m_Shl(m_One(), m_Value(Y)))) {
345       BO = BinaryOperator::CreateShl(Op0, Y);
346       ShlNSW = cast<ShlOperator>(Op1)->hasNoSignedWrap();
347     }
348     if (BO) {
349       if (I.hasNoUnsignedWrap())
350         BO->setHasNoUnsignedWrap();
351       if (I.hasNoSignedWrap() && ShlNSW)
352         BO->setHasNoSignedWrap();
353       return BO;
354     }
355   }
356 
357   // (bool X) * Y --> X ? Y : 0
358   // Y * (bool X) --> X ? Y : 0
359   if (match(Op0, m_ZExt(m_Value(X))) && X->getType()->isIntOrIntVectorTy(1))
360     return SelectInst::Create(X, Op1, ConstantInt::get(I.getType(), 0));
361   if (match(Op1, m_ZExt(m_Value(X))) && X->getType()->isIntOrIntVectorTy(1))
362     return SelectInst::Create(X, Op0, ConstantInt::get(I.getType(), 0));
363 
364   // (lshr X, 31) * Y --> (ashr X, 31) & Y
365   // Y * (lshr X, 31) --> (ashr X, 31) & Y
366   // TODO: We are not checking one-use because the elimination of the multiply
367   //       is better for analysis?
368   // TODO: Should we canonicalize to '(X < 0) ? Y : 0' instead? That would be
369   //       more similar to what we're doing above.
370   const APInt *C;
371   if (match(Op0, m_LShr(m_Value(X), m_APInt(C))) && *C == C->getBitWidth() - 1)
372     return BinaryOperator::CreateAnd(Builder.CreateAShr(X, *C), Op1);
373   if (match(Op1, m_LShr(m_Value(X), m_APInt(C))) && *C == C->getBitWidth() - 1)
374     return BinaryOperator::CreateAnd(Builder.CreateAShr(X, *C), Op0);
375 
376   if (Instruction *Ext = narrowMathIfNoOverflow(I))
377     return Ext;
378 
379   bool Changed = false;
380   if (!I.hasNoSignedWrap() && willNotOverflowSignedMul(Op0, Op1, I)) {
381     Changed = true;
382     I.setHasNoSignedWrap(true);
383   }
384 
385   if (!I.hasNoUnsignedWrap() && willNotOverflowUnsignedMul(Op0, Op1, I)) {
386     Changed = true;
387     I.setHasNoUnsignedWrap(true);
388   }
389 
390   return Changed ? &I : nullptr;
391 }
392 
393 Instruction *InstCombiner::visitFMul(BinaryOperator &I) {
394   if (Value *V = SimplifyFMulInst(I.getOperand(0), I.getOperand(1),
395                                   I.getFastMathFlags(),
396                                   SQ.getWithInstruction(&I)))
397     return replaceInstUsesWith(I, V);
398 
399   if (SimplifyAssociativeOrCommutative(I))
400     return &I;
401 
402   if (Instruction *X = foldVectorBinop(I))
403     return X;
404 
405   if (Instruction *FoldedMul = foldBinOpIntoSelectOrPhi(I))
406     return FoldedMul;
407 
408   if (Value *FoldedMul = foldMulSelectToNegate(I, Builder))
409     return replaceInstUsesWith(I, FoldedMul);
410 
411   // X * -1.0 --> -X
412   Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
413   if (match(Op1, m_SpecificFP(-1.0)))
414     return BinaryOperator::CreateFNegFMF(Op0, &I);
415 
416   // -X * -Y --> X * Y
417   Value *X, *Y;
418   if (match(Op0, m_FNeg(m_Value(X))) && match(Op1, m_FNeg(m_Value(Y))))
419     return BinaryOperator::CreateFMulFMF(X, Y, &I);
420 
421   // -X * C --> X * -C
422   Constant *C;
423   if (match(Op0, m_FNeg(m_Value(X))) && match(Op1, m_Constant(C)))
424     return BinaryOperator::CreateFMulFMF(X, ConstantExpr::getFNeg(C), &I);
425 
426   // fabs(X) * fabs(X) -> X * X
427   if (Op0 == Op1 && match(Op0, m_Intrinsic<Intrinsic::fabs>(m_Value(X))))
428     return BinaryOperator::CreateFMulFMF(X, X, &I);
429 
430   // (select A, B, C) * (select A, D, E) --> select A, (B*D), (C*E)
431   if (Value *V = SimplifySelectsFeedingBinaryOp(I, Op0, Op1))
432     return replaceInstUsesWith(I, V);
433 
434   if (I.hasAllowReassoc()) {
435     // Reassociate constant RHS with another constant to form constant
436     // expression.
437     if (match(Op1, m_Constant(C)) && C->isFiniteNonZeroFP()) {
438       Constant *C1;
439       if (match(Op0, m_OneUse(m_FDiv(m_Constant(C1), m_Value(X))))) {
440         // (C1 / X) * C --> (C * C1) / X
441         Constant *CC1 = ConstantExpr::getFMul(C, C1);
442         if (CC1->isNormalFP())
443           return BinaryOperator::CreateFDivFMF(CC1, X, &I);
444       }
445       if (match(Op0, m_FDiv(m_Value(X), m_Constant(C1)))) {
446         // (X / C1) * C --> X * (C / C1)
447         Constant *CDivC1 = ConstantExpr::getFDiv(C, C1);
448         if (CDivC1->isNormalFP())
449           return BinaryOperator::CreateFMulFMF(X, CDivC1, &I);
450 
451         // If the constant was a denormal, try reassociating differently.
452         // (X / C1) * C --> X / (C1 / C)
453         Constant *C1DivC = ConstantExpr::getFDiv(C1, C);
454         if (Op0->hasOneUse() && C1DivC->isNormalFP())
455           return BinaryOperator::CreateFDivFMF(X, C1DivC, &I);
456       }
457 
458       // We do not need to match 'fadd C, X' and 'fsub X, C' because they are
459       // canonicalized to 'fadd X, C'. Distributing the multiply may allow
460       // further folds and (X * C) + C2 is 'fma'.
461       if (match(Op0, m_OneUse(m_FAdd(m_Value(X), m_Constant(C1))))) {
462         // (X + C1) * C --> (X * C) + (C * C1)
463         Constant *CC1 = ConstantExpr::getFMul(C, C1);
464         Value *XC = Builder.CreateFMulFMF(X, C, &I);
465         return BinaryOperator::CreateFAddFMF(XC, CC1, &I);
466       }
467       if (match(Op0, m_OneUse(m_FSub(m_Constant(C1), m_Value(X))))) {
468         // (C1 - X) * C --> (C * C1) - (X * C)
469         Constant *CC1 = ConstantExpr::getFMul(C, C1);
470         Value *XC = Builder.CreateFMulFMF(X, C, &I);
471         return BinaryOperator::CreateFSubFMF(CC1, XC, &I);
472       }
473     }
474 
475     Value *Z;
476     if (match(&I, m_c_FMul(m_OneUse(m_FDiv(m_Value(X), m_Value(Y))),
477                            m_Value(Z)))) {
478       // Sink division: (X / Y) * Z --> (X * Z) / Y
479       Value *NewFMul = Builder.CreateFMulFMF(X, Z, &I);
480       return BinaryOperator::CreateFDivFMF(NewFMul, Y, &I);
481     }
482 
483     // sqrt(X) * sqrt(Y) -> sqrt(X * Y)
484     // nnan disallows the possibility of returning a number if both operands are
485     // negative (in that case, we should return NaN).
486     if (I.hasNoNaNs() &&
487         match(Op0, m_OneUse(m_Intrinsic<Intrinsic::sqrt>(m_Value(X)))) &&
488         match(Op1, m_OneUse(m_Intrinsic<Intrinsic::sqrt>(m_Value(Y))))) {
489       Value *XY = Builder.CreateFMulFMF(X, Y, &I);
490       Value *Sqrt = Builder.CreateUnaryIntrinsic(Intrinsic::sqrt, XY, &I);
491       return replaceInstUsesWith(I, Sqrt);
492     }
493 
494     // Like the similar transform in instsimplify, this requires 'nsz' because
495     // sqrt(-0.0) = -0.0, and -0.0 * -0.0 does not simplify to -0.0.
496     if (I.hasNoNaNs() && I.hasNoSignedZeros() && Op0 == Op1 &&
497         Op0->hasNUses(2)) {
498       // Peek through fdiv to find squaring of square root:
499       // (X / sqrt(Y)) * (X / sqrt(Y)) --> (X * X) / Y
500       if (match(Op0, m_FDiv(m_Value(X),
501                             m_Intrinsic<Intrinsic::sqrt>(m_Value(Y))))) {
502         Value *XX = Builder.CreateFMulFMF(X, X, &I);
503         return BinaryOperator::CreateFDivFMF(XX, Y, &I);
504       }
505       // (sqrt(Y) / X) * (sqrt(Y) / X) --> Y / (X * X)
506       if (match(Op0, m_FDiv(m_Intrinsic<Intrinsic::sqrt>(m_Value(Y)),
507                             m_Value(X)))) {
508         Value *XX = Builder.CreateFMulFMF(X, X, &I);
509         return BinaryOperator::CreateFDivFMF(Y, XX, &I);
510       }
511     }
512 
513     // exp(X) * exp(Y) -> exp(X + Y)
514     // Match as long as at least one of exp has only one use.
515     if (match(Op0, m_Intrinsic<Intrinsic::exp>(m_Value(X))) &&
516         match(Op1, m_Intrinsic<Intrinsic::exp>(m_Value(Y))) &&
517         (Op0->hasOneUse() || Op1->hasOneUse())) {
518       Value *XY = Builder.CreateFAddFMF(X, Y, &I);
519       Value *Exp = Builder.CreateUnaryIntrinsic(Intrinsic::exp, XY, &I);
520       return replaceInstUsesWith(I, Exp);
521     }
522 
523     // exp2(X) * exp2(Y) -> exp2(X + Y)
524     // Match as long as at least one of exp2 has only one use.
525     if (match(Op0, m_Intrinsic<Intrinsic::exp2>(m_Value(X))) &&
526         match(Op1, m_Intrinsic<Intrinsic::exp2>(m_Value(Y))) &&
527         (Op0->hasOneUse() || Op1->hasOneUse())) {
528       Value *XY = Builder.CreateFAddFMF(X, Y, &I);
529       Value *Exp2 = Builder.CreateUnaryIntrinsic(Intrinsic::exp2, XY, &I);
530       return replaceInstUsesWith(I, Exp2);
531     }
532 
533     // (X*Y) * X => (X*X) * Y where Y != X
534     //  The purpose is two-fold:
535     //   1) to form a power expression (of X).
536     //   2) potentially shorten the critical path: After transformation, the
537     //  latency of the instruction Y is amortized by the expression of X*X,
538     //  and therefore Y is in a "less critical" position compared to what it
539     //  was before the transformation.
540     if (match(Op0, m_OneUse(m_c_FMul(m_Specific(Op1), m_Value(Y)))) &&
541         Op1 != Y) {
542       Value *XX = Builder.CreateFMulFMF(Op1, Op1, &I);
543       return BinaryOperator::CreateFMulFMF(XX, Y, &I);
544     }
545     if (match(Op1, m_OneUse(m_c_FMul(m_Specific(Op0), m_Value(Y)))) &&
546         Op0 != Y) {
547       Value *XX = Builder.CreateFMulFMF(Op0, Op0, &I);
548       return BinaryOperator::CreateFMulFMF(XX, Y, &I);
549     }
550   }
551 
552   // log2(X * 0.5) * Y = log2(X) * Y - Y
553   if (I.isFast()) {
554     IntrinsicInst *Log2 = nullptr;
555     if (match(Op0, m_OneUse(m_Intrinsic<Intrinsic::log2>(
556             m_OneUse(m_FMul(m_Value(X), m_SpecificFP(0.5))))))) {
557       Log2 = cast<IntrinsicInst>(Op0);
558       Y = Op1;
559     }
560     if (match(Op1, m_OneUse(m_Intrinsic<Intrinsic::log2>(
561             m_OneUse(m_FMul(m_Value(X), m_SpecificFP(0.5))))))) {
562       Log2 = cast<IntrinsicInst>(Op1);
563       Y = Op0;
564     }
565     if (Log2) {
566       Log2->setArgOperand(0, X);
567       Log2->copyFastMathFlags(&I);
568       Value *LogXTimesY = Builder.CreateFMulFMF(Log2, Y, &I);
569       return BinaryOperator::CreateFSubFMF(LogXTimesY, Y, &I);
570     }
571   }
572 
573   return nullptr;
574 }
575 
576 /// Fold a divide or remainder with a select instruction divisor when one of the
577 /// select operands is zero. In that case, we can use the other select operand
578 /// because div/rem by zero is undefined.
579 bool InstCombiner::simplifyDivRemOfSelectWithZeroOp(BinaryOperator &I) {
580   SelectInst *SI = dyn_cast<SelectInst>(I.getOperand(1));
581   if (!SI)
582     return false;
583 
584   int NonNullOperand;
585   if (match(SI->getTrueValue(), m_Zero()))
586     // div/rem X, (Cond ? 0 : Y) -> div/rem X, Y
587     NonNullOperand = 2;
588   else if (match(SI->getFalseValue(), m_Zero()))
589     // div/rem X, (Cond ? Y : 0) -> div/rem X, Y
590     NonNullOperand = 1;
591   else
592     return false;
593 
594   // Change the div/rem to use 'Y' instead of the select.
595   I.setOperand(1, SI->getOperand(NonNullOperand));
596 
597   // Okay, we know we replace the operand of the div/rem with 'Y' with no
598   // problem.  However, the select, or the condition of the select may have
599   // multiple uses.  Based on our knowledge that the operand must be non-zero,
600   // propagate the known value for the select into other uses of it, and
601   // propagate a known value of the condition into its other users.
602 
603   // If the select and condition only have a single use, don't bother with this,
604   // early exit.
605   Value *SelectCond = SI->getCondition();
606   if (SI->use_empty() && SelectCond->hasOneUse())
607     return true;
608 
609   // Scan the current block backward, looking for other uses of SI.
610   BasicBlock::iterator BBI = I.getIterator(), BBFront = I.getParent()->begin();
611   Type *CondTy = SelectCond->getType();
612   while (BBI != BBFront) {
613     --BBI;
614     // If we found an instruction that we can't assume will return, so
615     // information from below it cannot be propagated above it.
616     if (!isGuaranteedToTransferExecutionToSuccessor(&*BBI))
617       break;
618 
619     // Replace uses of the select or its condition with the known values.
620     for (Instruction::op_iterator I = BBI->op_begin(), E = BBI->op_end();
621          I != E; ++I) {
622       if (*I == SI) {
623         *I = SI->getOperand(NonNullOperand);
624         Worklist.Add(&*BBI);
625       } else if (*I == SelectCond) {
626         *I = NonNullOperand == 1 ? ConstantInt::getTrue(CondTy)
627                                  : ConstantInt::getFalse(CondTy);
628         Worklist.Add(&*BBI);
629       }
630     }
631 
632     // If we past the instruction, quit looking for it.
633     if (&*BBI == SI)
634       SI = nullptr;
635     if (&*BBI == SelectCond)
636       SelectCond = nullptr;
637 
638     // If we ran out of things to eliminate, break out of the loop.
639     if (!SelectCond && !SI)
640       break;
641 
642   }
643   return true;
644 }
645 
646 /// True if the multiply can not be expressed in an int this size.
647 static bool multiplyOverflows(const APInt &C1, const APInt &C2, APInt &Product,
648                               bool IsSigned) {
649   bool Overflow;
650   Product = IsSigned ? C1.smul_ov(C2, Overflow) : C1.umul_ov(C2, Overflow);
651   return Overflow;
652 }
653 
654 /// True if C1 is a multiple of C2. Quotient contains C1/C2.
655 static bool isMultiple(const APInt &C1, const APInt &C2, APInt &Quotient,
656                        bool IsSigned) {
657   assert(C1.getBitWidth() == C2.getBitWidth() && "Constant widths not equal");
658 
659   // Bail if we will divide by zero.
660   if (C2.isNullValue())
661     return false;
662 
663   // Bail if we would divide INT_MIN by -1.
664   if (IsSigned && C1.isMinSignedValue() && C2.isAllOnesValue())
665     return false;
666 
667   APInt Remainder(C1.getBitWidth(), /*val=*/0ULL, IsSigned);
668   if (IsSigned)
669     APInt::sdivrem(C1, C2, Quotient, Remainder);
670   else
671     APInt::udivrem(C1, C2, Quotient, Remainder);
672 
673   return Remainder.isMinValue();
674 }
675 
676 /// This function implements the transforms common to both integer division
677 /// instructions (udiv and sdiv). It is called by the visitors to those integer
678 /// division instructions.
679 /// Common integer divide transforms
680 Instruction *InstCombiner::commonIDivTransforms(BinaryOperator &I) {
681   Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
682   bool IsSigned = I.getOpcode() == Instruction::SDiv;
683   Type *Ty = I.getType();
684 
685   // The RHS is known non-zero.
686   if (Value *V = simplifyValueKnownNonZero(I.getOperand(1), *this, I)) {
687     I.setOperand(1, V);
688     return &I;
689   }
690 
691   // Handle cases involving: [su]div X, (select Cond, Y, Z)
692   // This does not apply for fdiv.
693   if (simplifyDivRemOfSelectWithZeroOp(I))
694     return &I;
695 
696   const APInt *C2;
697   if (match(Op1, m_APInt(C2))) {
698     Value *X;
699     const APInt *C1;
700 
701     // (X / C1) / C2  -> X / (C1*C2)
702     if ((IsSigned && match(Op0, m_SDiv(m_Value(X), m_APInt(C1)))) ||
703         (!IsSigned && match(Op0, m_UDiv(m_Value(X), m_APInt(C1))))) {
704       APInt Product(C1->getBitWidth(), /*val=*/0ULL, IsSigned);
705       if (!multiplyOverflows(*C1, *C2, Product, IsSigned))
706         return BinaryOperator::Create(I.getOpcode(), X,
707                                       ConstantInt::get(Ty, Product));
708     }
709 
710     if ((IsSigned && match(Op0, m_NSWMul(m_Value(X), m_APInt(C1)))) ||
711         (!IsSigned && match(Op0, m_NUWMul(m_Value(X), m_APInt(C1))))) {
712       APInt Quotient(C1->getBitWidth(), /*val=*/0ULL, IsSigned);
713 
714       // (X * C1) / C2 -> X / (C2 / C1) if C2 is a multiple of C1.
715       if (isMultiple(*C2, *C1, Quotient, IsSigned)) {
716         auto *NewDiv = BinaryOperator::Create(I.getOpcode(), X,
717                                               ConstantInt::get(Ty, Quotient));
718         NewDiv->setIsExact(I.isExact());
719         return NewDiv;
720       }
721 
722       // (X * C1) / C2 -> X * (C1 / C2) if C1 is a multiple of C2.
723       if (isMultiple(*C1, *C2, Quotient, IsSigned)) {
724         auto *Mul = BinaryOperator::Create(Instruction::Mul, X,
725                                            ConstantInt::get(Ty, Quotient));
726         auto *OBO = cast<OverflowingBinaryOperator>(Op0);
727         Mul->setHasNoUnsignedWrap(!IsSigned && OBO->hasNoUnsignedWrap());
728         Mul->setHasNoSignedWrap(OBO->hasNoSignedWrap());
729         return Mul;
730       }
731     }
732 
733     if ((IsSigned && match(Op0, m_NSWShl(m_Value(X), m_APInt(C1))) &&
734          *C1 != C1->getBitWidth() - 1) ||
735         (!IsSigned && match(Op0, m_NUWShl(m_Value(X), m_APInt(C1))))) {
736       APInt Quotient(C1->getBitWidth(), /*val=*/0ULL, IsSigned);
737       APInt C1Shifted = APInt::getOneBitSet(
738           C1->getBitWidth(), static_cast<unsigned>(C1->getLimitedValue()));
739 
740       // (X << C1) / C2 -> X / (C2 >> C1) if C2 is a multiple of 1 << C1.
741       if (isMultiple(*C2, C1Shifted, Quotient, IsSigned)) {
742         auto *BO = BinaryOperator::Create(I.getOpcode(), X,
743                                           ConstantInt::get(Ty, Quotient));
744         BO->setIsExact(I.isExact());
745         return BO;
746       }
747 
748       // (X << C1) / C2 -> X * ((1 << C1) / C2) if 1 << C1 is a multiple of C2.
749       if (isMultiple(C1Shifted, *C2, Quotient, IsSigned)) {
750         auto *Mul = BinaryOperator::Create(Instruction::Mul, X,
751                                            ConstantInt::get(Ty, Quotient));
752         auto *OBO = cast<OverflowingBinaryOperator>(Op0);
753         Mul->setHasNoUnsignedWrap(!IsSigned && OBO->hasNoUnsignedWrap());
754         Mul->setHasNoSignedWrap(OBO->hasNoSignedWrap());
755         return Mul;
756       }
757     }
758 
759     if (!C2->isNullValue()) // avoid X udiv 0
760       if (Instruction *FoldedDiv = foldBinOpIntoSelectOrPhi(I))
761         return FoldedDiv;
762   }
763 
764   if (match(Op0, m_One())) {
765     assert(!Ty->isIntOrIntVectorTy(1) && "i1 divide not removed?");
766     if (IsSigned) {
767       // If Op1 is 0 then it's undefined behaviour, if Op1 is 1 then the
768       // result is one, if Op1 is -1 then the result is minus one, otherwise
769       // it's zero.
770       Value *Inc = Builder.CreateAdd(Op1, Op0);
771       Value *Cmp = Builder.CreateICmpULT(Inc, ConstantInt::get(Ty, 3));
772       return SelectInst::Create(Cmp, Op1, ConstantInt::get(Ty, 0));
773     } else {
774       // If Op1 is 0 then it's undefined behaviour. If Op1 is 1 then the
775       // result is one, otherwise it's zero.
776       return new ZExtInst(Builder.CreateICmpEQ(Op1, Op0), Ty);
777     }
778   }
779 
780   // See if we can fold away this div instruction.
781   if (SimplifyDemandedInstructionBits(I))
782     return &I;
783 
784   // (X - (X rem Y)) / Y -> X / Y; usually originates as ((X / Y) * Y) / Y
785   Value *X, *Z;
786   if (match(Op0, m_Sub(m_Value(X), m_Value(Z)))) // (X - Z) / Y; Y = Op1
787     if ((IsSigned && match(Z, m_SRem(m_Specific(X), m_Specific(Op1)))) ||
788         (!IsSigned && match(Z, m_URem(m_Specific(X), m_Specific(Op1)))))
789       return BinaryOperator::Create(I.getOpcode(), X, Op1);
790 
791   // (X << Y) / X -> 1 << Y
792   Value *Y;
793   if (IsSigned && match(Op0, m_NSWShl(m_Specific(Op1), m_Value(Y))))
794     return BinaryOperator::CreateNSWShl(ConstantInt::get(Ty, 1), Y);
795   if (!IsSigned && match(Op0, m_NUWShl(m_Specific(Op1), m_Value(Y))))
796     return BinaryOperator::CreateNUWShl(ConstantInt::get(Ty, 1), Y);
797 
798   // X / (X * Y) -> 1 / Y if the multiplication does not overflow.
799   if (match(Op1, m_c_Mul(m_Specific(Op0), m_Value(Y)))) {
800     bool HasNSW = cast<OverflowingBinaryOperator>(Op1)->hasNoSignedWrap();
801     bool HasNUW = cast<OverflowingBinaryOperator>(Op1)->hasNoUnsignedWrap();
802     if ((IsSigned && HasNSW) || (!IsSigned && HasNUW)) {
803       I.setOperand(0, ConstantInt::get(Ty, 1));
804       I.setOperand(1, Y);
805       return &I;
806     }
807   }
808 
809   return nullptr;
810 }
811 
812 static const unsigned MaxDepth = 6;
813 
814 namespace {
815 
816 using FoldUDivOperandCb = Instruction *(*)(Value *Op0, Value *Op1,
817                                            const BinaryOperator &I,
818                                            InstCombiner &IC);
819 
820 /// Used to maintain state for visitUDivOperand().
821 struct UDivFoldAction {
822   /// Informs visitUDiv() how to fold this operand.  This can be zero if this
823   /// action joins two actions together.
824   FoldUDivOperandCb FoldAction;
825 
826   /// Which operand to fold.
827   Value *OperandToFold;
828 
829   union {
830     /// The instruction returned when FoldAction is invoked.
831     Instruction *FoldResult;
832 
833     /// Stores the LHS action index if this action joins two actions together.
834     size_t SelectLHSIdx;
835   };
836 
837   UDivFoldAction(FoldUDivOperandCb FA, Value *InputOperand)
838       : FoldAction(FA), OperandToFold(InputOperand), FoldResult(nullptr) {}
839   UDivFoldAction(FoldUDivOperandCb FA, Value *InputOperand, size_t SLHS)
840       : FoldAction(FA), OperandToFold(InputOperand), SelectLHSIdx(SLHS) {}
841 };
842 
843 } // end anonymous namespace
844 
845 // X udiv 2^C -> X >> C
846 static Instruction *foldUDivPow2Cst(Value *Op0, Value *Op1,
847                                     const BinaryOperator &I, InstCombiner &IC) {
848   Constant *C1 = getLogBase2(Op0->getType(), cast<Constant>(Op1));
849   if (!C1)
850     llvm_unreachable("Failed to constant fold udiv -> logbase2");
851   BinaryOperator *LShr = BinaryOperator::CreateLShr(Op0, C1);
852   if (I.isExact())
853     LShr->setIsExact();
854   return LShr;
855 }
856 
857 // X udiv (C1 << N), where C1 is "1<<C2"  -->  X >> (N+C2)
858 // X udiv (zext (C1 << N)), where C1 is "1<<C2"  -->  X >> (N+C2)
859 static Instruction *foldUDivShl(Value *Op0, Value *Op1, const BinaryOperator &I,
860                                 InstCombiner &IC) {
861   Value *ShiftLeft;
862   if (!match(Op1, m_ZExt(m_Value(ShiftLeft))))
863     ShiftLeft = Op1;
864 
865   Constant *CI;
866   Value *N;
867   if (!match(ShiftLeft, m_Shl(m_Constant(CI), m_Value(N))))
868     llvm_unreachable("match should never fail here!");
869   Constant *Log2Base = getLogBase2(N->getType(), CI);
870   if (!Log2Base)
871     llvm_unreachable("getLogBase2 should never fail here!");
872   N = IC.Builder.CreateAdd(N, Log2Base);
873   if (Op1 != ShiftLeft)
874     N = IC.Builder.CreateZExt(N, Op1->getType());
875   BinaryOperator *LShr = BinaryOperator::CreateLShr(Op0, N);
876   if (I.isExact())
877     LShr->setIsExact();
878   return LShr;
879 }
880 
881 // Recursively visits the possible right hand operands of a udiv
882 // instruction, seeing through select instructions, to determine if we can
883 // replace the udiv with something simpler.  If we find that an operand is not
884 // able to simplify the udiv, we abort the entire transformation.
885 static size_t visitUDivOperand(Value *Op0, Value *Op1, const BinaryOperator &I,
886                                SmallVectorImpl<UDivFoldAction> &Actions,
887                                unsigned Depth = 0) {
888   // Check to see if this is an unsigned division with an exact power of 2,
889   // if so, convert to a right shift.
890   if (match(Op1, m_Power2())) {
891     Actions.push_back(UDivFoldAction(foldUDivPow2Cst, Op1));
892     return Actions.size();
893   }
894 
895   // X udiv (C1 << N), where C1 is "1<<C2"  -->  X >> (N+C2)
896   if (match(Op1, m_Shl(m_Power2(), m_Value())) ||
897       match(Op1, m_ZExt(m_Shl(m_Power2(), m_Value())))) {
898     Actions.push_back(UDivFoldAction(foldUDivShl, Op1));
899     return Actions.size();
900   }
901 
902   // The remaining tests are all recursive, so bail out if we hit the limit.
903   if (Depth++ == MaxDepth)
904     return 0;
905 
906   if (SelectInst *SI = dyn_cast<SelectInst>(Op1))
907     if (size_t LHSIdx =
908             visitUDivOperand(Op0, SI->getOperand(1), I, Actions, Depth))
909       if (visitUDivOperand(Op0, SI->getOperand(2), I, Actions, Depth)) {
910         Actions.push_back(UDivFoldAction(nullptr, Op1, LHSIdx - 1));
911         return Actions.size();
912       }
913 
914   return 0;
915 }
916 
917 /// If we have zero-extended operands of an unsigned div or rem, we may be able
918 /// to narrow the operation (sink the zext below the math).
919 static Instruction *narrowUDivURem(BinaryOperator &I,
920                                    InstCombiner::BuilderTy &Builder) {
921   Instruction::BinaryOps Opcode = I.getOpcode();
922   Value *N = I.getOperand(0);
923   Value *D = I.getOperand(1);
924   Type *Ty = I.getType();
925   Value *X, *Y;
926   if (match(N, m_ZExt(m_Value(X))) && match(D, m_ZExt(m_Value(Y))) &&
927       X->getType() == Y->getType() && (N->hasOneUse() || D->hasOneUse())) {
928     // udiv (zext X), (zext Y) --> zext (udiv X, Y)
929     // urem (zext X), (zext Y) --> zext (urem X, Y)
930     Value *NarrowOp = Builder.CreateBinOp(Opcode, X, Y);
931     return new ZExtInst(NarrowOp, Ty);
932   }
933 
934   Constant *C;
935   if ((match(N, m_OneUse(m_ZExt(m_Value(X)))) && match(D, m_Constant(C))) ||
936       (match(D, m_OneUse(m_ZExt(m_Value(X)))) && match(N, m_Constant(C)))) {
937     // If the constant is the same in the smaller type, use the narrow version.
938     Constant *TruncC = ConstantExpr::getTrunc(C, X->getType());
939     if (ConstantExpr::getZExt(TruncC, Ty) != C)
940       return nullptr;
941 
942     // udiv (zext X), C --> zext (udiv X, C')
943     // urem (zext X), C --> zext (urem X, C')
944     // udiv C, (zext X) --> zext (udiv C', X)
945     // urem C, (zext X) --> zext (urem C', X)
946     Value *NarrowOp = isa<Constant>(D) ? Builder.CreateBinOp(Opcode, X, TruncC)
947                                        : Builder.CreateBinOp(Opcode, TruncC, X);
948     return new ZExtInst(NarrowOp, Ty);
949   }
950 
951   return nullptr;
952 }
953 
954 Instruction *InstCombiner::visitUDiv(BinaryOperator &I) {
955   if (Value *V = SimplifyUDivInst(I.getOperand(0), I.getOperand(1),
956                                   SQ.getWithInstruction(&I)))
957     return replaceInstUsesWith(I, V);
958 
959   if (Instruction *X = foldVectorBinop(I))
960     return X;
961 
962   // Handle the integer div common cases
963   if (Instruction *Common = commonIDivTransforms(I))
964     return Common;
965 
966   Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
967   Value *X;
968   const APInt *C1, *C2;
969   if (match(Op0, m_LShr(m_Value(X), m_APInt(C1))) && match(Op1, m_APInt(C2))) {
970     // (X lshr C1) udiv C2 --> X udiv (C2 << C1)
971     bool Overflow;
972     APInt C2ShlC1 = C2->ushl_ov(*C1, Overflow);
973     if (!Overflow) {
974       bool IsExact = I.isExact() && match(Op0, m_Exact(m_Value()));
975       BinaryOperator *BO = BinaryOperator::CreateUDiv(
976           X, ConstantInt::get(X->getType(), C2ShlC1));
977       if (IsExact)
978         BO->setIsExact();
979       return BO;
980     }
981   }
982 
983   // Op0 / C where C is large (negative) --> zext (Op0 >= C)
984   // TODO: Could use isKnownNegative() to handle non-constant values.
985   Type *Ty = I.getType();
986   if (match(Op1, m_Negative())) {
987     Value *Cmp = Builder.CreateICmpUGE(Op0, Op1);
988     return CastInst::CreateZExtOrBitCast(Cmp, Ty);
989   }
990   // Op0 / (sext i1 X) --> zext (Op0 == -1) (if X is 0, the div is undefined)
991   if (match(Op1, m_SExt(m_Value(X))) && X->getType()->isIntOrIntVectorTy(1)) {
992     Value *Cmp = Builder.CreateICmpEQ(Op0, ConstantInt::getAllOnesValue(Ty));
993     return CastInst::CreateZExtOrBitCast(Cmp, Ty);
994   }
995 
996   if (Instruction *NarrowDiv = narrowUDivURem(I, Builder))
997     return NarrowDiv;
998 
999   // If the udiv operands are non-overflowing multiplies with a common operand,
1000   // then eliminate the common factor:
1001   // (A * B) / (A * X) --> B / X (and commuted variants)
1002   // TODO: The code would be reduced if we had m_c_NUWMul pattern matching.
1003   // TODO: If -reassociation handled this generally, we could remove this.
1004   Value *A, *B;
1005   if (match(Op0, m_NUWMul(m_Value(A), m_Value(B)))) {
1006     if (match(Op1, m_NUWMul(m_Specific(A), m_Value(X))) ||
1007         match(Op1, m_NUWMul(m_Value(X), m_Specific(A))))
1008       return BinaryOperator::CreateUDiv(B, X);
1009     if (match(Op1, m_NUWMul(m_Specific(B), m_Value(X))) ||
1010         match(Op1, m_NUWMul(m_Value(X), m_Specific(B))))
1011       return BinaryOperator::CreateUDiv(A, X);
1012   }
1013 
1014   // (LHS udiv (select (select (...)))) -> (LHS >> (select (select (...))))
1015   SmallVector<UDivFoldAction, 6> UDivActions;
1016   if (visitUDivOperand(Op0, Op1, I, UDivActions))
1017     for (unsigned i = 0, e = UDivActions.size(); i != e; ++i) {
1018       FoldUDivOperandCb Action = UDivActions[i].FoldAction;
1019       Value *ActionOp1 = UDivActions[i].OperandToFold;
1020       Instruction *Inst;
1021       if (Action)
1022         Inst = Action(Op0, ActionOp1, I, *this);
1023       else {
1024         // This action joins two actions together.  The RHS of this action is
1025         // simply the last action we processed, we saved the LHS action index in
1026         // the joining action.
1027         size_t SelectRHSIdx = i - 1;
1028         Value *SelectRHS = UDivActions[SelectRHSIdx].FoldResult;
1029         size_t SelectLHSIdx = UDivActions[i].SelectLHSIdx;
1030         Value *SelectLHS = UDivActions[SelectLHSIdx].FoldResult;
1031         Inst = SelectInst::Create(cast<SelectInst>(ActionOp1)->getCondition(),
1032                                   SelectLHS, SelectRHS);
1033       }
1034 
1035       // If this is the last action to process, return it to the InstCombiner.
1036       // Otherwise, we insert it before the UDiv and record it so that we may
1037       // use it as part of a joining action (i.e., a SelectInst).
1038       if (e - i != 1) {
1039         Inst->insertBefore(&I);
1040         UDivActions[i].FoldResult = Inst;
1041       } else
1042         return Inst;
1043     }
1044 
1045   return nullptr;
1046 }
1047 
1048 Instruction *InstCombiner::visitSDiv(BinaryOperator &I) {
1049   if (Value *V = SimplifySDivInst(I.getOperand(0), I.getOperand(1),
1050                                   SQ.getWithInstruction(&I)))
1051     return replaceInstUsesWith(I, V);
1052 
1053   if (Instruction *X = foldVectorBinop(I))
1054     return X;
1055 
1056   // Handle the integer div common cases
1057   if (Instruction *Common = commonIDivTransforms(I))
1058     return Common;
1059 
1060   Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
1061   Value *X;
1062   // sdiv Op0, -1 --> -Op0
1063   // sdiv Op0, (sext i1 X) --> -Op0 (because if X is 0, the op is undefined)
1064   if (match(Op1, m_AllOnes()) ||
1065       (match(Op1, m_SExt(m_Value(X))) && X->getType()->isIntOrIntVectorTy(1)))
1066     return BinaryOperator::CreateNeg(Op0);
1067 
1068   // X / INT_MIN --> X == INT_MIN
1069   if (match(Op1, m_SignMask()))
1070     return new ZExtInst(Builder.CreateICmpEQ(Op0, Op1), I.getType());
1071 
1072   const APInt *Op1C;
1073   if (match(Op1, m_APInt(Op1C))) {
1074     // sdiv exact X, C  -->  ashr exact X, log2(C)
1075     if (I.isExact() && Op1C->isNonNegative() && Op1C->isPowerOf2()) {
1076       Value *ShAmt = ConstantInt::get(Op1->getType(), Op1C->exactLogBase2());
1077       return BinaryOperator::CreateExactAShr(Op0, ShAmt, I.getName());
1078     }
1079 
1080     // If the dividend is sign-extended and the constant divisor is small enough
1081     // to fit in the source type, shrink the division to the narrower type:
1082     // (sext X) sdiv C --> sext (X sdiv C)
1083     Value *Op0Src;
1084     if (match(Op0, m_OneUse(m_SExt(m_Value(Op0Src)))) &&
1085         Op0Src->getType()->getScalarSizeInBits() >= Op1C->getMinSignedBits()) {
1086 
1087       // In the general case, we need to make sure that the dividend is not the
1088       // minimum signed value because dividing that by -1 is UB. But here, we
1089       // know that the -1 divisor case is already handled above.
1090 
1091       Constant *NarrowDivisor =
1092           ConstantExpr::getTrunc(cast<Constant>(Op1), Op0Src->getType());
1093       Value *NarrowOp = Builder.CreateSDiv(Op0Src, NarrowDivisor);
1094       return new SExtInst(NarrowOp, Op0->getType());
1095     }
1096 
1097     // -X / C --> X / -C (if the negation doesn't overflow).
1098     // TODO: This could be enhanced to handle arbitrary vector constants by
1099     //       checking if all elements are not the min-signed-val.
1100     if (!Op1C->isMinSignedValue() &&
1101         match(Op0, m_NSWSub(m_Zero(), m_Value(X)))) {
1102       Constant *NegC = ConstantInt::get(I.getType(), -(*Op1C));
1103       Instruction *BO = BinaryOperator::CreateSDiv(X, NegC);
1104       BO->setIsExact(I.isExact());
1105       return BO;
1106     }
1107   }
1108 
1109   // -X / Y --> -(X / Y)
1110   Value *Y;
1111   if (match(&I, m_SDiv(m_OneUse(m_NSWSub(m_Zero(), m_Value(X))), m_Value(Y))))
1112     return BinaryOperator::CreateNSWNeg(
1113         Builder.CreateSDiv(X, Y, I.getName(), I.isExact()));
1114 
1115   // If the sign bits of both operands are zero (i.e. we can prove they are
1116   // unsigned inputs), turn this into a udiv.
1117   APInt Mask(APInt::getSignMask(I.getType()->getScalarSizeInBits()));
1118   if (MaskedValueIsZero(Op0, Mask, 0, &I)) {
1119     if (MaskedValueIsZero(Op1, Mask, 0, &I)) {
1120       // X sdiv Y -> X udiv Y, iff X and Y don't have sign bit set
1121       auto *BO = BinaryOperator::CreateUDiv(Op0, Op1, I.getName());
1122       BO->setIsExact(I.isExact());
1123       return BO;
1124     }
1125 
1126     if (isKnownToBeAPowerOfTwo(Op1, /*OrZero*/ true, 0, &I)) {
1127       // X sdiv (1 << Y) -> X udiv (1 << Y) ( -> X u>> Y)
1128       // Safe because the only negative value (1 << Y) can take on is
1129       // INT_MIN, and X sdiv INT_MIN == X udiv INT_MIN == 0 if X doesn't have
1130       // the sign bit set.
1131       auto *BO = BinaryOperator::CreateUDiv(Op0, Op1, I.getName());
1132       BO->setIsExact(I.isExact());
1133       return BO;
1134     }
1135   }
1136 
1137   return nullptr;
1138 }
1139 
1140 /// Remove negation and try to convert division into multiplication.
1141 static Instruction *foldFDivConstantDivisor(BinaryOperator &I) {
1142   Constant *C;
1143   if (!match(I.getOperand(1), m_Constant(C)))
1144     return nullptr;
1145 
1146   // -X / C --> X / -C
1147   Value *X;
1148   if (match(I.getOperand(0), m_FNeg(m_Value(X))))
1149     return BinaryOperator::CreateFDivFMF(X, ConstantExpr::getFNeg(C), &I);
1150 
1151   // If the constant divisor has an exact inverse, this is always safe. If not,
1152   // then we can still create a reciprocal if fast-math-flags allow it and the
1153   // constant is a regular number (not zero, infinite, or denormal).
1154   if (!(C->hasExactInverseFP() || (I.hasAllowReciprocal() && C->isNormalFP())))
1155     return nullptr;
1156 
1157   // Disallow denormal constants because we don't know what would happen
1158   // on all targets.
1159   // TODO: Use Intrinsic::canonicalize or let function attributes tell us that
1160   // denorms are flushed?
1161   auto *RecipC = ConstantExpr::getFDiv(ConstantFP::get(I.getType(), 1.0), C);
1162   if (!RecipC->isNormalFP())
1163     return nullptr;
1164 
1165   // X / C --> X * (1 / C)
1166   return BinaryOperator::CreateFMulFMF(I.getOperand(0), RecipC, &I);
1167 }
1168 
1169 /// Remove negation and try to reassociate constant math.
1170 static Instruction *foldFDivConstantDividend(BinaryOperator &I) {
1171   Constant *C;
1172   if (!match(I.getOperand(0), m_Constant(C)))
1173     return nullptr;
1174 
1175   // C / -X --> -C / X
1176   Value *X;
1177   if (match(I.getOperand(1), m_FNeg(m_Value(X))))
1178     return BinaryOperator::CreateFDivFMF(ConstantExpr::getFNeg(C), X, &I);
1179 
1180   if (!I.hasAllowReassoc() || !I.hasAllowReciprocal())
1181     return nullptr;
1182 
1183   // Try to reassociate C / X expressions where X includes another constant.
1184   Constant *C2, *NewC = nullptr;
1185   if (match(I.getOperand(1), m_FMul(m_Value(X), m_Constant(C2)))) {
1186     // C / (X * C2) --> (C / C2) / X
1187     NewC = ConstantExpr::getFDiv(C, C2);
1188   } else if (match(I.getOperand(1), m_FDiv(m_Value(X), m_Constant(C2)))) {
1189     // C / (X / C2) --> (C * C2) / X
1190     NewC = ConstantExpr::getFMul(C, C2);
1191   }
1192   // Disallow denormal constants because we don't know what would happen
1193   // on all targets.
1194   // TODO: Use Intrinsic::canonicalize or let function attributes tell us that
1195   // denorms are flushed?
1196   if (!NewC || !NewC->isNormalFP())
1197     return nullptr;
1198 
1199   return BinaryOperator::CreateFDivFMF(NewC, X, &I);
1200 }
1201 
1202 Instruction *InstCombiner::visitFDiv(BinaryOperator &I) {
1203   if (Value *V = SimplifyFDivInst(I.getOperand(0), I.getOperand(1),
1204                                   I.getFastMathFlags(),
1205                                   SQ.getWithInstruction(&I)))
1206     return replaceInstUsesWith(I, V);
1207 
1208   if (Instruction *X = foldVectorBinop(I))
1209     return X;
1210 
1211   if (Instruction *R = foldFDivConstantDivisor(I))
1212     return R;
1213 
1214   if (Instruction *R = foldFDivConstantDividend(I))
1215     return R;
1216 
1217   Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
1218   if (isa<Constant>(Op0))
1219     if (SelectInst *SI = dyn_cast<SelectInst>(Op1))
1220       if (Instruction *R = FoldOpIntoSelect(I, SI))
1221         return R;
1222 
1223   if (isa<Constant>(Op1))
1224     if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
1225       if (Instruction *R = FoldOpIntoSelect(I, SI))
1226         return R;
1227 
1228   if (I.hasAllowReassoc() && I.hasAllowReciprocal()) {
1229     Value *X, *Y;
1230     if (match(Op0, m_OneUse(m_FDiv(m_Value(X), m_Value(Y)))) &&
1231         (!isa<Constant>(Y) || !isa<Constant>(Op1))) {
1232       // (X / Y) / Z => X / (Y * Z)
1233       Value *YZ = Builder.CreateFMulFMF(Y, Op1, &I);
1234       return BinaryOperator::CreateFDivFMF(X, YZ, &I);
1235     }
1236     if (match(Op1, m_OneUse(m_FDiv(m_Value(X), m_Value(Y)))) &&
1237         (!isa<Constant>(Y) || !isa<Constant>(Op0))) {
1238       // Z / (X / Y) => (Y * Z) / X
1239       Value *YZ = Builder.CreateFMulFMF(Y, Op0, &I);
1240       return BinaryOperator::CreateFDivFMF(YZ, X, &I);
1241     }
1242     // Z / (1.0 / Y) => (Y * Z)
1243     //
1244     // This is a special case of Z / (X / Y) => (Y * Z) / X, with X = 1.0. The
1245     // m_OneUse check is avoided because even in the case of the multiple uses
1246     // for 1.0/Y, the number of instructions remain the same and a division is
1247     // replaced by a multiplication.
1248     if (match(Op1, m_FDiv(m_SpecificFP(1.0), m_Value(Y))))
1249       return BinaryOperator::CreateFMulFMF(Y, Op0, &I);
1250   }
1251 
1252   if (I.hasAllowReassoc() && Op0->hasOneUse() && Op1->hasOneUse()) {
1253     // sin(X) / cos(X) -> tan(X)
1254     // cos(X) / sin(X) -> 1/tan(X) (cotangent)
1255     Value *X;
1256     bool IsTan = match(Op0, m_Intrinsic<Intrinsic::sin>(m_Value(X))) &&
1257                  match(Op1, m_Intrinsic<Intrinsic::cos>(m_Specific(X)));
1258     bool IsCot =
1259         !IsTan && match(Op0, m_Intrinsic<Intrinsic::cos>(m_Value(X))) &&
1260                   match(Op1, m_Intrinsic<Intrinsic::sin>(m_Specific(X)));
1261 
1262     if ((IsTan || IsCot) &&
1263         hasFloatFn(&TLI, I.getType(), LibFunc_tan, LibFunc_tanf, LibFunc_tanl)) {
1264       IRBuilder<> B(&I);
1265       IRBuilder<>::FastMathFlagGuard FMFGuard(B);
1266       B.setFastMathFlags(I.getFastMathFlags());
1267       AttributeList Attrs =
1268           cast<CallBase>(Op0)->getCalledFunction()->getAttributes();
1269       Value *Res = emitUnaryFloatFnCall(X, &TLI, LibFunc_tan, LibFunc_tanf,
1270                                         LibFunc_tanl, B, Attrs);
1271       if (IsCot)
1272         Res = B.CreateFDiv(ConstantFP::get(I.getType(), 1.0), Res);
1273       return replaceInstUsesWith(I, Res);
1274     }
1275   }
1276 
1277   // -X / -Y -> X / Y
1278   Value *X, *Y;
1279   if (match(Op0, m_FNeg(m_Value(X))) && match(Op1, m_FNeg(m_Value(Y)))) {
1280     I.setOperand(0, X);
1281     I.setOperand(1, Y);
1282     return &I;
1283   }
1284 
1285   // X / (X * Y) --> 1.0 / Y
1286   // Reassociate to (X / X -> 1.0) is legal when NaNs are not allowed.
1287   // We can ignore the possibility that X is infinity because INF/INF is NaN.
1288   if (I.hasNoNaNs() && I.hasAllowReassoc() &&
1289       match(Op1, m_c_FMul(m_Specific(Op0), m_Value(Y)))) {
1290     I.setOperand(0, ConstantFP::get(I.getType(), 1.0));
1291     I.setOperand(1, Y);
1292     return &I;
1293   }
1294 
1295   // X / fabs(X) -> copysign(1.0, X)
1296   // fabs(X) / X -> copysign(1.0, X)
1297   if (I.hasNoNaNs() && I.hasNoInfs() &&
1298       (match(&I,
1299              m_FDiv(m_Value(X), m_Intrinsic<Intrinsic::fabs>(m_Deferred(X)))) ||
1300        match(&I, m_FDiv(m_Intrinsic<Intrinsic::fabs>(m_Value(X)),
1301                         m_Deferred(X))))) {
1302     Value *V = Builder.CreateBinaryIntrinsic(
1303         Intrinsic::copysign, ConstantFP::get(I.getType(), 1.0), X, &I);
1304     return replaceInstUsesWith(I, V);
1305   }
1306   return nullptr;
1307 }
1308 
1309 /// This function implements the transforms common to both integer remainder
1310 /// instructions (urem and srem). It is called by the visitors to those integer
1311 /// remainder instructions.
1312 /// Common integer remainder transforms
1313 Instruction *InstCombiner::commonIRemTransforms(BinaryOperator &I) {
1314   Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
1315 
1316   // The RHS is known non-zero.
1317   if (Value *V = simplifyValueKnownNonZero(I.getOperand(1), *this, I)) {
1318     I.setOperand(1, V);
1319     return &I;
1320   }
1321 
1322   // Handle cases involving: rem X, (select Cond, Y, Z)
1323   if (simplifyDivRemOfSelectWithZeroOp(I))
1324     return &I;
1325 
1326   if (isa<Constant>(Op1)) {
1327     if (Instruction *Op0I = dyn_cast<Instruction>(Op0)) {
1328       if (SelectInst *SI = dyn_cast<SelectInst>(Op0I)) {
1329         if (Instruction *R = FoldOpIntoSelect(I, SI))
1330           return R;
1331       } else if (auto *PN = dyn_cast<PHINode>(Op0I)) {
1332         const APInt *Op1Int;
1333         if (match(Op1, m_APInt(Op1Int)) && !Op1Int->isMinValue() &&
1334             (I.getOpcode() == Instruction::URem ||
1335              !Op1Int->isMinSignedValue())) {
1336           // foldOpIntoPhi will speculate instructions to the end of the PHI's
1337           // predecessor blocks, so do this only if we know the srem or urem
1338           // will not fault.
1339           if (Instruction *NV = foldOpIntoPhi(I, PN))
1340             return NV;
1341         }
1342       }
1343 
1344       // See if we can fold away this rem instruction.
1345       if (SimplifyDemandedInstructionBits(I))
1346         return &I;
1347     }
1348   }
1349 
1350   return nullptr;
1351 }
1352 
1353 Instruction *InstCombiner::visitURem(BinaryOperator &I) {
1354   if (Value *V = SimplifyURemInst(I.getOperand(0), I.getOperand(1),
1355                                   SQ.getWithInstruction(&I)))
1356     return replaceInstUsesWith(I, V);
1357 
1358   if (Instruction *X = foldVectorBinop(I))
1359     return X;
1360 
1361   if (Instruction *common = commonIRemTransforms(I))
1362     return common;
1363 
1364   if (Instruction *NarrowRem = narrowUDivURem(I, Builder))
1365     return NarrowRem;
1366 
1367   // X urem Y -> X and Y-1, where Y is a power of 2,
1368   Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
1369   Type *Ty = I.getType();
1370   if (isKnownToBeAPowerOfTwo(Op1, /*OrZero*/ true, 0, &I)) {
1371     // This may increase instruction count, we don't enforce that Y is a
1372     // constant.
1373     Constant *N1 = Constant::getAllOnesValue(Ty);
1374     Value *Add = Builder.CreateAdd(Op1, N1);
1375     return BinaryOperator::CreateAnd(Op0, Add);
1376   }
1377 
1378   // 1 urem X -> zext(X != 1)
1379   if (match(Op0, m_One())) {
1380     Value *Cmp = Builder.CreateICmpNE(Op1, ConstantInt::get(Ty, 1));
1381     return CastInst::CreateZExtOrBitCast(Cmp, Ty);
1382   }
1383 
1384   // X urem C -> X < C ? X : X - C, where C >= signbit.
1385   if (match(Op1, m_Negative())) {
1386     Value *Cmp = Builder.CreateICmpULT(Op0, Op1);
1387     Value *Sub = Builder.CreateSub(Op0, Op1);
1388     return SelectInst::Create(Cmp, Op0, Sub);
1389   }
1390 
1391   // If the divisor is a sext of a boolean, then the divisor must be max
1392   // unsigned value (-1). Therefore, the remainder is Op0 unless Op0 is also
1393   // max unsigned value. In that case, the remainder is 0:
1394   // urem Op0, (sext i1 X) --> (Op0 == -1) ? 0 : Op0
1395   Value *X;
1396   if (match(Op1, m_SExt(m_Value(X))) && X->getType()->isIntOrIntVectorTy(1)) {
1397     Value *Cmp = Builder.CreateICmpEQ(Op0, ConstantInt::getAllOnesValue(Ty));
1398     return SelectInst::Create(Cmp, ConstantInt::getNullValue(Ty), Op0);
1399   }
1400 
1401   return nullptr;
1402 }
1403 
1404 Instruction *InstCombiner::visitSRem(BinaryOperator &I) {
1405   if (Value *V = SimplifySRemInst(I.getOperand(0), I.getOperand(1),
1406                                   SQ.getWithInstruction(&I)))
1407     return replaceInstUsesWith(I, V);
1408 
1409   if (Instruction *X = foldVectorBinop(I))
1410     return X;
1411 
1412   // Handle the integer rem common cases
1413   if (Instruction *Common = commonIRemTransforms(I))
1414     return Common;
1415 
1416   Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
1417   {
1418     const APInt *Y;
1419     // X % -Y -> X % Y
1420     if (match(Op1, m_Negative(Y)) && !Y->isMinSignedValue()) {
1421       Worklist.AddValue(I.getOperand(1));
1422       I.setOperand(1, ConstantInt::get(I.getType(), -*Y));
1423       return &I;
1424     }
1425   }
1426 
1427   // -X srem Y --> -(X srem Y)
1428   Value *X, *Y;
1429   if (match(&I, m_SRem(m_OneUse(m_NSWSub(m_Zero(), m_Value(X))), m_Value(Y))))
1430     return BinaryOperator::CreateNSWNeg(Builder.CreateSRem(X, Y));
1431 
1432   // If the sign bits of both operands are zero (i.e. we can prove they are
1433   // unsigned inputs), turn this into a urem.
1434   APInt Mask(APInt::getSignMask(I.getType()->getScalarSizeInBits()));
1435   if (MaskedValueIsZero(Op1, Mask, 0, &I) &&
1436       MaskedValueIsZero(Op0, Mask, 0, &I)) {
1437     // X srem Y -> X urem Y, iff X and Y don't have sign bit set
1438     return BinaryOperator::CreateURem(Op0, Op1, I.getName());
1439   }
1440 
1441   // If it's a constant vector, flip any negative values positive.
1442   if (isa<ConstantVector>(Op1) || isa<ConstantDataVector>(Op1)) {
1443     Constant *C = cast<Constant>(Op1);
1444     unsigned VWidth = C->getType()->getVectorNumElements();
1445 
1446     bool hasNegative = false;
1447     bool hasMissing = false;
1448     for (unsigned i = 0; i != VWidth; ++i) {
1449       Constant *Elt = C->getAggregateElement(i);
1450       if (!Elt) {
1451         hasMissing = true;
1452         break;
1453       }
1454 
1455       if (ConstantInt *RHS = dyn_cast<ConstantInt>(Elt))
1456         if (RHS->isNegative())
1457           hasNegative = true;
1458     }
1459 
1460     if (hasNegative && !hasMissing) {
1461       SmallVector<Constant *, 16> Elts(VWidth);
1462       for (unsigned i = 0; i != VWidth; ++i) {
1463         Elts[i] = C->getAggregateElement(i);  // Handle undef, etc.
1464         if (ConstantInt *RHS = dyn_cast<ConstantInt>(Elts[i])) {
1465           if (RHS->isNegative())
1466             Elts[i] = cast<ConstantInt>(ConstantExpr::getNeg(RHS));
1467         }
1468       }
1469 
1470       Constant *NewRHSV = ConstantVector::get(Elts);
1471       if (NewRHSV != C) {  // Don't loop on -MININT
1472         Worklist.AddValue(I.getOperand(1));
1473         I.setOperand(1, NewRHSV);
1474         return &I;
1475       }
1476     }
1477   }
1478 
1479   return nullptr;
1480 }
1481 
1482 Instruction *InstCombiner::visitFRem(BinaryOperator &I) {
1483   if (Value *V = SimplifyFRemInst(I.getOperand(0), I.getOperand(1),
1484                                   I.getFastMathFlags(),
1485                                   SQ.getWithInstruction(&I)))
1486     return replaceInstUsesWith(I, V);
1487 
1488   if (Instruction *X = foldVectorBinop(I))
1489     return X;
1490 
1491   return nullptr;
1492 }
1493