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. |
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, P: m_LShr(L: m_OneUse(SubPattern: m_Shl(L: m_Value(V&: One), R: m_Value(V&: A))), R: m_Value(V&: B))) && |
59 | match(V: One, P: m_One())) { |
60 | A = IC.Builder.CreateSub(LHS: A, RHS: B); |
61 | return IC.Builder.CreateShl(LHS: One, RHS: 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>(Val: V); |
67 | if (I && I->isLogicalShift() && |
68 | IC.isKnownToBeAPowerOfTwo(V: I->getOperand(i_nocapture: 0), OrZero: false, Depth: 0, CxtI: &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(V: I->getOperand(i_nocapture: 0), IC, CxtI)) { |
72 | IC.replaceOperand(I&: *I, OpNum: 0, V: 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(). |
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(V: &I, P: m_c_Mul(L: m_OneUse(SubPattern: m_Select(C: m_Value(V&: Cond), L: m_One(), R: m_AllOnes())), |
106 | R: m_Value(V&: OtherOp)))) { |
107 | bool HasAnyNoWrap = I.hasNoSignedWrap() || I.hasNoUnsignedWrap(); |
108 | Value *Neg = Builder.CreateNeg(V: OtherOp, Name: "" , HasNSW: HasAnyNoWrap); |
109 | return Builder.CreateSelect(C: Cond, True: OtherOp, False: 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(V: &I, P: m_c_Mul(L: m_OneUse(SubPattern: m_Select(C: m_Value(V&: Cond), L: m_AllOnes(), R: m_One())), |
114 | R: m_Value(V&: OtherOp)))) { |
115 | bool HasAnyNoWrap = I.hasNoSignedWrap() || I.hasNoUnsignedWrap(); |
116 | Value *Neg = Builder.CreateNeg(V: OtherOp, Name: "" , HasNSW: HasAnyNoWrap); |
117 | return Builder.CreateSelect(C: Cond, True: Neg, False: 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(V: &I, P: m_c_FMul(L: m_OneUse(SubPattern: m_Select(C: m_Value(V&: Cond), L: m_SpecificFP(V: 1.0), |
123 | R: m_SpecificFP(V: -1.0))), |
124 | R: m_Value(V&: OtherOp)))) { |
125 | IRBuilder<>::FastMathFlagGuard FMFGuard(Builder); |
126 | Builder.setFastMathFlags(I.getFastMathFlags()); |
127 | return Builder.CreateSelect(C: Cond, True: OtherOp, False: Builder.CreateFNeg(V: 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(V: &I, P: m_c_FMul(L: m_OneUse(SubPattern: m_Select(C: m_Value(V&: Cond), L: m_SpecificFP(V: -1.0), |
133 | R: m_SpecificFP(V: 1.0))), |
134 | R: m_Value(V&: OtherOp)))) { |
135 | IRBuilder<>::FastMathFlagGuard FMFGuard(Builder); |
136 | Builder.setFastMathFlags(I.getFastMathFlags()); |
137 | return Builder.CreateSelect(C: Cond, True: Builder.CreateFNeg(V: OtherOp), False: 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. |
145 | static Value *foldMulShl1(BinaryOperator &Mul, bool CommuteOperands, |
146 | InstCombiner::BuilderTy &Builder) { |
147 | Value *X = Mul.getOperand(i_nocapture: 0), *Y = Mul.getOperand(i_nocapture: 1); |
148 | if (CommuteOperands) |
149 | std::swap(a&: X, b&: 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(V: Y, P: m_Shl(L: m_One(), R: m_Value(V&: Z)))) { |
157 | bool PropagateNSW = HasNSW && cast<ShlOperator>(Val: Y)->hasNoSignedWrap(); |
158 | return Builder.CreateShl(LHS: X, RHS: Z, Name: Mul.getName(), HasNUW, HasNSW: 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(V: Y, P: m_OneUse(SubPattern: m_Add(L: m_BinOp(I&: Shift), R: m_One()))) && |
167 | match(V: Shift, P: m_OneUse(SubPattern: m_Shl(L: m_One(), R: m_Value(V&: Z))))) { |
168 | bool PropagateNSW = HasNSW && Shift->hasNoSignedWrap(); |
169 | Value *FrX = Builder.CreateFreeze(V: X, Name: X->getName() + ".fr" ); |
170 | Value *Shl = Builder.CreateShl(LHS: FrX, RHS: Z, Name: "mulshl" , HasNUW, HasNSW: PropagateNSW); |
171 | return Builder.CreateAdd(LHS: Shl, RHS: FrX, Name: Mul.getName(), HasNUW, HasNSW: PropagateNSW); |
172 | } |
173 | |
174 | // Similar to above, but a decrement of the shifted value is disguised as |
175 | // 'not' and becomes a sub: |
176 | // X * (~(-1 << Z)) --> X * ((1 << Z) - 1) --> (X << Z) - X |
177 | // This increases uses of X, so it may require a freeze, but that is still |
178 | // expected to be an improvement because it removes the multiply. |
179 | if (match(V: Y, P: m_OneUse(SubPattern: m_Not(V: m_OneUse(SubPattern: m_Shl(L: m_AllOnes(), R: m_Value(V&: Z))))))) { |
180 | Value *FrX = Builder.CreateFreeze(V: X, Name: X->getName() + ".fr" ); |
181 | Value *Shl = Builder.CreateShl(LHS: FrX, RHS: Z, Name: "mulshl" ); |
182 | return Builder.CreateSub(LHS: Shl, RHS: FrX, Name: Mul.getName()); |
183 | } |
184 | |
185 | return nullptr; |
186 | } |
187 | |
188 | static Value *takeLog2(IRBuilderBase &Builder, Value *Op, unsigned Depth, |
189 | bool AssumeNonZero, bool DoFold); |
190 | |
191 | Instruction *InstCombinerImpl::visitMul(BinaryOperator &I) { |
192 | Value *Op0 = I.getOperand(i_nocapture: 0), *Op1 = I.getOperand(i_nocapture: 1); |
193 | if (Value *V = |
194 | simplifyMulInst(LHS: Op0, RHS: Op1, IsNSW: I.hasNoSignedWrap(), IsNUW: I.hasNoUnsignedWrap(), |
195 | Q: SQ.getWithInstruction(I: &I))) |
196 | return replaceInstUsesWith(I, V); |
197 | |
198 | if (SimplifyAssociativeOrCommutative(I)) |
199 | return &I; |
200 | |
201 | if (Instruction *X = foldVectorBinop(Inst&: I)) |
202 | return X; |
203 | |
204 | if (Instruction *Phi = foldBinopWithPhiOperands(BO&: I)) |
205 | return Phi; |
206 | |
207 | if (Value *V = foldUsingDistributiveLaws(I)) |
208 | return replaceInstUsesWith(I, V); |
209 | |
210 | Type *Ty = I.getType(); |
211 | const unsigned BitWidth = Ty->getScalarSizeInBits(); |
212 | const bool HasNSW = I.hasNoSignedWrap(); |
213 | const bool HasNUW = I.hasNoUnsignedWrap(); |
214 | |
215 | // X * -1 --> 0 - X |
216 | if (match(V: Op1, P: m_AllOnes())) { |
217 | return HasNSW ? BinaryOperator::CreateNSWNeg(Op: Op0) |
218 | : BinaryOperator::CreateNeg(Op: Op0); |
219 | } |
220 | |
221 | // Also allow combining multiply instructions on vectors. |
222 | { |
223 | Value *NewOp; |
224 | Constant *C1, *C2; |
225 | const APInt *IVal; |
226 | if (match(V: &I, P: m_Mul(L: m_Shl(L: m_Value(V&: NewOp), R: m_Constant(C&: C2)), |
227 | R: m_Constant(C&: C1))) && |
228 | match(V: C1, P: m_APInt(Res&: IVal))) { |
229 | // ((X << C2)*C1) == (X * (C1 << C2)) |
230 | Constant *Shl = ConstantExpr::getShl(C1, C2); |
231 | BinaryOperator *Mul = cast<BinaryOperator>(Val: I.getOperand(i_nocapture: 0)); |
232 | BinaryOperator *BO = BinaryOperator::CreateMul(V1: NewOp, V2: Shl); |
233 | if (HasNUW && Mul->hasNoUnsignedWrap()) |
234 | BO->setHasNoUnsignedWrap(); |
235 | if (HasNSW && Mul->hasNoSignedWrap() && Shl->isNotMinSignedValue()) |
236 | BO->setHasNoSignedWrap(); |
237 | return BO; |
238 | } |
239 | |
240 | if (match(V: &I, P: m_Mul(L: m_Value(V&: NewOp), R: m_Constant(C&: C1)))) { |
241 | // Replace X*(2^C) with X << C, where C is either a scalar or a vector. |
242 | if (Constant *NewCst = ConstantExpr::getExactLogBase2(C: C1)) { |
243 | BinaryOperator *Shl = BinaryOperator::CreateShl(V1: NewOp, V2: NewCst); |
244 | |
245 | if (HasNUW) |
246 | Shl->setHasNoUnsignedWrap(); |
247 | if (HasNSW) { |
248 | const APInt *V; |
249 | if (match(V: NewCst, P: m_APInt(Res&: V)) && *V != V->getBitWidth() - 1) |
250 | Shl->setHasNoSignedWrap(); |
251 | } |
252 | |
253 | return Shl; |
254 | } |
255 | } |
256 | } |
257 | |
258 | if (Op0->hasOneUse() && match(V: Op1, P: m_NegatedPower2())) { |
259 | // Interpret X * (-1<<C) as (-X) * (1<<C) and try to sink the negation. |
260 | // The "* (1<<C)" thus becomes a potential shifting opportunity. |
261 | if (Value *NegOp0 = |
262 | Negator::Negate(/*IsNegation*/ LHSIsZero: true, IsNSW: HasNSW, Root: Op0, IC&: *this)) { |
263 | auto *Op1C = cast<Constant>(Val: Op1); |
264 | return replaceInstUsesWith( |
265 | I, V: Builder.CreateMul(LHS: NegOp0, RHS: ConstantExpr::getNeg(C: Op1C), Name: "" , |
266 | /* HasNUW */ false, |
267 | HasNSW: HasNSW && Op1C->isNotMinSignedValue())); |
268 | } |
269 | |
270 | // Try to convert multiply of extended operand to narrow negate and shift |
271 | // for better analysis. |
272 | // This is valid if the shift amount (trailing zeros in the multiplier |
273 | // constant) clears more high bits than the bitwidth difference between |
274 | // source and destination types: |
275 | // ({z/s}ext X) * (-1<<C) --> (zext (-X)) << C |
276 | const APInt *NegPow2C; |
277 | Value *X; |
278 | if (match(V: Op0, P: m_ZExtOrSExt(Op: m_Value(V&: X))) && |
279 | match(V: Op1, P: m_APIntAllowPoison(Res&: NegPow2C))) { |
280 | unsigned SrcWidth = X->getType()->getScalarSizeInBits(); |
281 | unsigned ShiftAmt = NegPow2C->countr_zero(); |
282 | if (ShiftAmt >= BitWidth - SrcWidth) { |
283 | Value *N = Builder.CreateNeg(V: X, Name: X->getName() + ".neg" ); |
284 | Value *Z = Builder.CreateZExt(V: N, DestTy: Ty, Name: N->getName() + ".z" ); |
285 | return BinaryOperator::CreateShl(V1: Z, V2: ConstantInt::get(Ty, V: ShiftAmt)); |
286 | } |
287 | } |
288 | } |
289 | |
290 | if (Instruction *FoldedMul = foldBinOpIntoSelectOrPhi(I)) |
291 | return FoldedMul; |
292 | |
293 | if (Value *FoldedMul = foldMulSelectToNegate(I, Builder)) |
294 | return replaceInstUsesWith(I, V: FoldedMul); |
295 | |
296 | // Simplify mul instructions with a constant RHS. |
297 | Constant *MulC; |
298 | if (match(V: Op1, P: m_ImmConstant(C&: MulC))) { |
299 | // Canonicalize (X+C1)*MulC -> X*MulC+C1*MulC. |
300 | // Canonicalize (X|C1)*MulC -> X*MulC+C1*MulC. |
301 | Value *X; |
302 | Constant *C1; |
303 | if (match(V: Op0, P: m_OneUse(SubPattern: m_AddLike(L: m_Value(V&: X), R: m_ImmConstant(C&: C1))))) { |
304 | // C1*MulC simplifies to a tidier constant. |
305 | Value *NewC = Builder.CreateMul(LHS: C1, RHS: MulC); |
306 | auto *BOp0 = cast<BinaryOperator>(Val: Op0); |
307 | bool Op0NUW = |
308 | (BOp0->getOpcode() == Instruction::Or || BOp0->hasNoUnsignedWrap()); |
309 | Value *NewMul = Builder.CreateMul(LHS: X, RHS: MulC); |
310 | auto *BO = BinaryOperator::CreateAdd(V1: NewMul, V2: NewC); |
311 | if (HasNUW && Op0NUW) { |
312 | // If NewMulBO is constant we also can set BO to nuw. |
313 | if (auto *NewMulBO = dyn_cast<BinaryOperator>(Val: NewMul)) |
314 | NewMulBO->setHasNoUnsignedWrap(); |
315 | BO->setHasNoUnsignedWrap(); |
316 | } |
317 | return BO; |
318 | } |
319 | } |
320 | |
321 | // abs(X) * abs(X) -> X * X |
322 | Value *X; |
323 | if (Op0 == Op1 && match(Op0, m_Intrinsic<Intrinsic::abs>(m_Value(V&: X)))) |
324 | return BinaryOperator::CreateMul(V1: X, V2: X); |
325 | |
326 | { |
327 | Value *Y; |
328 | // abs(X) * abs(Y) -> abs(X * Y) |
329 | if (I.hasNoSignedWrap() && |
330 | match(Op0, |
331 | m_OneUse(m_Intrinsic<Intrinsic::abs>(m_Value(V&: X), m_One()))) && |
332 | match(Op1, m_OneUse(m_Intrinsic<Intrinsic::abs>(m_Value(V&: Y), m_One())))) |
333 | return replaceInstUsesWith( |
334 | I, V: Builder.CreateBinaryIntrinsic(Intrinsic::ID: abs, |
335 | LHS: Builder.CreateNSWMul(LHS: X, RHS: Y), |
336 | RHS: Builder.getTrue())); |
337 | } |
338 | |
339 | // -X * C --> X * -C |
340 | Value *Y; |
341 | Constant *Op1C; |
342 | if (match(V: Op0, P: m_Neg(V: m_Value(V&: X))) && match(V: Op1, P: m_Constant(C&: Op1C))) |
343 | return BinaryOperator::CreateMul(V1: X, V2: ConstantExpr::getNeg(C: Op1C)); |
344 | |
345 | // -X * -Y --> X * Y |
346 | if (match(V: Op0, P: m_Neg(V: m_Value(V&: X))) && match(V: Op1, P: m_Neg(V: m_Value(V&: Y)))) { |
347 | auto *NewMul = BinaryOperator::CreateMul(V1: X, V2: Y); |
348 | if (HasNSW && cast<OverflowingBinaryOperator>(Val: Op0)->hasNoSignedWrap() && |
349 | cast<OverflowingBinaryOperator>(Val: Op1)->hasNoSignedWrap()) |
350 | NewMul->setHasNoSignedWrap(); |
351 | return NewMul; |
352 | } |
353 | |
354 | // -X * Y --> -(X * Y) |
355 | // X * -Y --> -(X * Y) |
356 | if (match(V: &I, P: m_c_Mul(L: m_OneUse(SubPattern: m_Neg(V: m_Value(V&: X))), R: m_Value(V&: Y)))) |
357 | return BinaryOperator::CreateNeg(Op: Builder.CreateMul(LHS: X, RHS: Y)); |
358 | |
359 | // (-X * Y) * -X --> (X * Y) * X |
360 | // (-X << Y) * -X --> (X << Y) * X |
361 | if (match(V: Op1, P: m_Neg(V: m_Value(V&: X)))) { |
362 | if (Value *NegOp0 = Negator::Negate(LHSIsZero: false, /*IsNSW*/ false, Root: Op0, IC&: *this)) |
363 | return BinaryOperator::CreateMul(V1: NegOp0, V2: X); |
364 | } |
365 | |
366 | // (X / Y) * Y = X - (X % Y) |
367 | // (X / Y) * -Y = (X % Y) - X |
368 | { |
369 | Value *Y = Op1; |
370 | BinaryOperator *Div = dyn_cast<BinaryOperator>(Val: Op0); |
371 | if (!Div || (Div->getOpcode() != Instruction::UDiv && |
372 | Div->getOpcode() != Instruction::SDiv)) { |
373 | Y = Op0; |
374 | Div = dyn_cast<BinaryOperator>(Val: Op1); |
375 | } |
376 | Value *Neg = dyn_castNegVal(V: Y); |
377 | if (Div && Div->hasOneUse() && |
378 | (Div->getOperand(i_nocapture: 1) == Y || Div->getOperand(i_nocapture: 1) == Neg) && |
379 | (Div->getOpcode() == Instruction::UDiv || |
380 | Div->getOpcode() == Instruction::SDiv)) { |
381 | Value *X = Div->getOperand(i_nocapture: 0), *DivOp1 = Div->getOperand(i_nocapture: 1); |
382 | |
383 | // If the division is exact, X % Y is zero, so we end up with X or -X. |
384 | if (Div->isExact()) { |
385 | if (DivOp1 == Y) |
386 | return replaceInstUsesWith(I, V: X); |
387 | return BinaryOperator::CreateNeg(Op: X); |
388 | } |
389 | |
390 | auto RemOpc = Div->getOpcode() == Instruction::UDiv ? Instruction::URem |
391 | : Instruction::SRem; |
392 | // X must be frozen because we are increasing its number of uses. |
393 | Value *XFreeze = Builder.CreateFreeze(V: X, Name: X->getName() + ".fr" ); |
394 | Value *Rem = Builder.CreateBinOp(Opc: RemOpc, LHS: XFreeze, RHS: DivOp1); |
395 | if (DivOp1 == Y) |
396 | return BinaryOperator::CreateSub(V1: XFreeze, V2: Rem); |
397 | return BinaryOperator::CreateSub(V1: Rem, V2: XFreeze); |
398 | } |
399 | } |
400 | |
401 | // Fold the following two scenarios: |
402 | // 1) i1 mul -> i1 and. |
403 | // 2) X * Y --> X & Y, iff X, Y can be only {0,1}. |
404 | // Note: We could use known bits to generalize this and related patterns with |
405 | // shifts/truncs |
406 | if (Ty->isIntOrIntVectorTy(BitWidth: 1) || |
407 | (match(V: Op0, P: m_And(L: m_Value(), R: m_One())) && |
408 | match(V: Op1, P: m_And(L: m_Value(), R: m_One())))) |
409 | return BinaryOperator::CreateAnd(V1: Op0, V2: Op1); |
410 | |
411 | if (Value *R = foldMulShl1(Mul&: I, /* CommuteOperands */ false, Builder)) |
412 | return replaceInstUsesWith(I, V: R); |
413 | if (Value *R = foldMulShl1(Mul&: I, /* CommuteOperands */ true, Builder)) |
414 | return replaceInstUsesWith(I, V: R); |
415 | |
416 | // (zext bool X) * (zext bool Y) --> zext (and X, Y) |
417 | // (sext bool X) * (sext bool Y) --> zext (and X, Y) |
418 | // Note: -1 * -1 == 1 * 1 == 1 (if the extends match, the result is the same) |
419 | if (((match(V: Op0, P: m_ZExt(Op: m_Value(V&: X))) && match(V: Op1, P: m_ZExt(Op: m_Value(V&: Y)))) || |
420 | (match(V: Op0, P: m_SExt(Op: m_Value(V&: X))) && match(V: Op1, P: m_SExt(Op: m_Value(V&: Y))))) && |
421 | X->getType()->isIntOrIntVectorTy(BitWidth: 1) && X->getType() == Y->getType() && |
422 | (Op0->hasOneUse() || Op1->hasOneUse() || X == Y)) { |
423 | Value *And = Builder.CreateAnd(LHS: X, RHS: Y, Name: "mulbool" ); |
424 | return CastInst::Create(Instruction::ZExt, S: And, Ty); |
425 | } |
426 | // (sext bool X) * (zext bool Y) --> sext (and X, Y) |
427 | // (zext bool X) * (sext bool Y) --> sext (and X, Y) |
428 | // Note: -1 * 1 == 1 * -1 == -1 |
429 | if (((match(V: Op0, P: m_SExt(Op: m_Value(V&: X))) && match(V: Op1, P: m_ZExt(Op: m_Value(V&: Y)))) || |
430 | (match(V: Op0, P: m_ZExt(Op: m_Value(V&: X))) && match(V: Op1, P: m_SExt(Op: m_Value(V&: Y))))) && |
431 | X->getType()->isIntOrIntVectorTy(BitWidth: 1) && X->getType() == Y->getType() && |
432 | (Op0->hasOneUse() || Op1->hasOneUse())) { |
433 | Value *And = Builder.CreateAnd(LHS: X, RHS: Y, Name: "mulbool" ); |
434 | return CastInst::Create(Instruction::SExt, S: And, Ty); |
435 | } |
436 | |
437 | // (zext bool X) * Y --> X ? Y : 0 |
438 | // Y * (zext bool X) --> X ? Y : 0 |
439 | if (match(V: Op0, P: m_ZExt(Op: m_Value(V&: X))) && X->getType()->isIntOrIntVectorTy(BitWidth: 1)) |
440 | return SelectInst::Create(C: X, S1: Op1, S2: ConstantInt::getNullValue(Ty)); |
441 | if (match(V: Op1, P: m_ZExt(Op: m_Value(V&: X))) && X->getType()->isIntOrIntVectorTy(BitWidth: 1)) |
442 | return SelectInst::Create(C: X, S1: Op0, S2: ConstantInt::getNullValue(Ty)); |
443 | |
444 | // mul (sext X), Y -> select X, -Y, 0 |
445 | // mul Y, (sext X) -> select X, -Y, 0 |
446 | if (match(V: &I, P: m_c_Mul(L: m_OneUse(SubPattern: m_SExt(Op: m_Value(V&: X))), R: m_Value(V&: Y))) && |
447 | X->getType()->isIntOrIntVectorTy(BitWidth: 1)) |
448 | return SelectInst::Create(C: X, S1: Builder.CreateNeg(V: Y, Name: "" , HasNSW: I.hasNoSignedWrap()), |
449 | S2: ConstantInt::getNullValue(Ty: Op0->getType())); |
450 | |
451 | Constant *ImmC; |
452 | if (match(V: Op1, P: m_ImmConstant(C&: ImmC))) { |
453 | // (sext bool X) * C --> X ? -C : 0 |
454 | if (match(V: Op0, P: m_SExt(Op: m_Value(V&: X))) && X->getType()->isIntOrIntVectorTy(BitWidth: 1)) { |
455 | Constant *NegC = ConstantExpr::getNeg(C: ImmC); |
456 | return SelectInst::Create(C: X, S1: NegC, S2: ConstantInt::getNullValue(Ty)); |
457 | } |
458 | |
459 | // (ashr i32 X, 31) * C --> (X < 0) ? -C : 0 |
460 | const APInt *C; |
461 | if (match(V: Op0, P: m_OneUse(SubPattern: m_AShr(L: m_Value(V&: X), R: m_APInt(Res&: C)))) && |
462 | *C == C->getBitWidth() - 1) { |
463 | Constant *NegC = ConstantExpr::getNeg(C: ImmC); |
464 | Value *IsNeg = Builder.CreateIsNeg(Arg: X, Name: "isneg" ); |
465 | return SelectInst::Create(C: IsNeg, S1: NegC, S2: ConstantInt::getNullValue(Ty)); |
466 | } |
467 | } |
468 | |
469 | // (lshr X, 31) * Y --> (X < 0) ? Y : 0 |
470 | // TODO: We are not checking one-use because the elimination of the multiply |
471 | // is better for analysis? |
472 | const APInt *C; |
473 | if (match(V: &I, P: m_c_BinOp(L: m_LShr(L: m_Value(V&: X), R: m_APInt(Res&: C)), R: m_Value(V&: Y))) && |
474 | *C == C->getBitWidth() - 1) { |
475 | Value *IsNeg = Builder.CreateIsNeg(Arg: X, Name: "isneg" ); |
476 | return SelectInst::Create(C: IsNeg, S1: Y, S2: ConstantInt::getNullValue(Ty)); |
477 | } |
478 | |
479 | // (and X, 1) * Y --> (trunc X) ? Y : 0 |
480 | if (match(V: &I, P: m_c_BinOp(L: m_OneUse(SubPattern: m_And(L: m_Value(V&: X), R: m_One())), R: m_Value(V&: Y)))) { |
481 | Value *Tr = Builder.CreateTrunc(V: X, DestTy: CmpInst::makeCmpResultType(opnd_type: Ty)); |
482 | return SelectInst::Create(C: Tr, S1: Y, S2: ConstantInt::getNullValue(Ty)); |
483 | } |
484 | |
485 | // ((ashr X, 31) | 1) * X --> abs(X) |
486 | // X * ((ashr X, 31) | 1) --> abs(X) |
487 | if (match(V: &I, P: m_c_BinOp(L: m_Or(L: m_AShr(L: m_Value(V&: X), |
488 | R: m_SpecificIntAllowPoison(V: BitWidth - 1)), |
489 | R: m_One()), |
490 | R: m_Deferred(V: X)))) { |
491 | Value *Abs = Builder.CreateBinaryIntrinsic( |
492 | Intrinsic::ID: abs, LHS: X, RHS: ConstantInt::getBool(Context&: I.getContext(), V: HasNSW)); |
493 | Abs->takeName(V: &I); |
494 | return replaceInstUsesWith(I, V: Abs); |
495 | } |
496 | |
497 | if (Instruction *Ext = narrowMathIfNoOverflow(I)) |
498 | return Ext; |
499 | |
500 | if (Instruction *Res = foldBinOpOfSelectAndCastOfSelectCondition(I)) |
501 | return Res; |
502 | |
503 | // (mul Op0 Op1): |
504 | // if Log2(Op0) folds away -> |
505 | // (shl Op1, Log2(Op0)) |
506 | // if Log2(Op1) folds away -> |
507 | // (shl Op0, Log2(Op1)) |
508 | if (takeLog2(Builder, Op: Op0, /*Depth*/ 0, /*AssumeNonZero*/ false, |
509 | /*DoFold*/ false)) { |
510 | Value *Res = takeLog2(Builder, Op: Op0, /*Depth*/ 0, /*AssumeNonZero*/ false, |
511 | /*DoFold*/ true); |
512 | BinaryOperator *Shl = BinaryOperator::CreateShl(V1: Op1, V2: Res); |
513 | // We can only propegate nuw flag. |
514 | Shl->setHasNoUnsignedWrap(HasNUW); |
515 | return Shl; |
516 | } |
517 | if (takeLog2(Builder, Op: Op1, /*Depth*/ 0, /*AssumeNonZero*/ false, |
518 | /*DoFold*/ false)) { |
519 | Value *Res = takeLog2(Builder, Op: Op1, /*Depth*/ 0, /*AssumeNonZero*/ false, |
520 | /*DoFold*/ true); |
521 | BinaryOperator *Shl = BinaryOperator::CreateShl(V1: Op0, V2: Res); |
522 | // We can only propegate nuw flag. |
523 | Shl->setHasNoUnsignedWrap(HasNUW); |
524 | return Shl; |
525 | } |
526 | |
527 | bool Changed = false; |
528 | if (!HasNSW && willNotOverflowSignedMul(LHS: Op0, RHS: Op1, CxtI: I)) { |
529 | Changed = true; |
530 | I.setHasNoSignedWrap(true); |
531 | } |
532 | |
533 | if (!HasNUW && willNotOverflowUnsignedMul(LHS: Op0, RHS: Op1, CxtI: I)) { |
534 | Changed = true; |
535 | I.setHasNoUnsignedWrap(true); |
536 | } |
537 | |
538 | return Changed ? &I : nullptr; |
539 | } |
540 | |
541 | Instruction *InstCombinerImpl::foldFPSignBitOps(BinaryOperator &I) { |
542 | BinaryOperator::BinaryOps Opcode = I.getOpcode(); |
543 | assert((Opcode == Instruction::FMul || Opcode == Instruction::FDiv) && |
544 | "Expected fmul or fdiv" ); |
545 | |
546 | Value *Op0 = I.getOperand(i_nocapture: 0), *Op1 = I.getOperand(i_nocapture: 1); |
547 | Value *X, *Y; |
548 | |
549 | // -X * -Y --> X * Y |
550 | // -X / -Y --> X / Y |
551 | if (match(V: Op0, P: m_FNeg(X: m_Value(V&: X))) && match(V: Op1, P: m_FNeg(X: m_Value(V&: Y)))) |
552 | return BinaryOperator::CreateWithCopiedFlags(Opc: Opcode, V1: X, V2: Y, CopyO: &I); |
553 | |
554 | // fabs(X) * fabs(X) -> X * X |
555 | // fabs(X) / fabs(X) -> X / X |
556 | if (Op0 == Op1 && match(V: Op0, P: m_FAbs(Op0: m_Value(V&: X)))) |
557 | return BinaryOperator::CreateWithCopiedFlags(Opc: Opcode, V1: X, V2: X, CopyO: &I); |
558 | |
559 | // fabs(X) * fabs(Y) --> fabs(X * Y) |
560 | // fabs(X) / fabs(Y) --> fabs(X / Y) |
561 | if (match(V: Op0, P: m_FAbs(Op0: m_Value(V&: X))) && match(V: Op1, P: m_FAbs(Op0: m_Value(V&: Y))) && |
562 | (Op0->hasOneUse() || Op1->hasOneUse())) { |
563 | IRBuilder<>::FastMathFlagGuard FMFGuard(Builder); |
564 | Builder.setFastMathFlags(I.getFastMathFlags()); |
565 | Value *XY = Builder.CreateBinOp(Opc: Opcode, LHS: X, RHS: Y); |
566 | Value *Fabs = Builder.CreateUnaryIntrinsic(Intrinsic::ID: fabs, V: XY); |
567 | Fabs->takeName(V: &I); |
568 | return replaceInstUsesWith(I, V: Fabs); |
569 | } |
570 | |
571 | return nullptr; |
572 | } |
573 | |
574 | Instruction *InstCombinerImpl::foldPowiReassoc(BinaryOperator &I) { |
575 | auto createPowiExpr = [](BinaryOperator &I, InstCombinerImpl &IC, Value *X, |
576 | Value *Y, Value *Z) { |
577 | InstCombiner::BuilderTy &Builder = IC.Builder; |
578 | Value *YZ = Builder.CreateAdd(LHS: Y, RHS: Z); |
579 | Instruction *NewPow = Builder.CreateIntrinsic( |
580 | Intrinsic::powi, {X->getType(), YZ->getType()}, {X, YZ}, &I); |
581 | |
582 | return NewPow; |
583 | }; |
584 | |
585 | Value *X, *Y, *Z; |
586 | unsigned Opcode = I.getOpcode(); |
587 | assert((Opcode == Instruction::FMul || Opcode == Instruction::FDiv) && |
588 | "Unexpected opcode" ); |
589 | |
590 | // powi(X, Y) * X --> powi(X, Y+1) |
591 | // X * powi(X, Y) --> powi(X, Y+1) |
592 | if (match(&I, m_c_FMul(m_OneUse(m_AllowReassoc(m_Intrinsic<Intrinsic::powi>( |
593 | m_Value(X), m_Value(Y)))), |
594 | m_Deferred(X)))) { |
595 | Constant *One = ConstantInt::get(Ty: Y->getType(), V: 1); |
596 | if (willNotOverflowSignedAdd(LHS: Y, RHS: One, CxtI: I)) { |
597 | Instruction *NewPow = createPowiExpr(I, *this, X, Y, One); |
598 | return replaceInstUsesWith(I, V: NewPow); |
599 | } |
600 | } |
601 | |
602 | // powi(x, y) * powi(x, z) -> powi(x, y + z) |
603 | Value *Op0 = I.getOperand(i_nocapture: 0); |
604 | Value *Op1 = I.getOperand(i_nocapture: 1); |
605 | if (Opcode == Instruction::FMul && I.isOnlyUserOfAnyOperand() && |
606 | match(Op0, m_AllowReassoc( |
607 | m_Intrinsic<Intrinsic::powi>(m_Value(X), m_Value(Y)))) && |
608 | match(Op1, m_AllowReassoc(m_Intrinsic<Intrinsic::powi>(m_Specific(X), |
609 | m_Value(Z)))) && |
610 | Y->getType() == Z->getType()) { |
611 | Instruction *NewPow = createPowiExpr(I, *this, X, Y, Z); |
612 | return replaceInstUsesWith(I, V: NewPow); |
613 | } |
614 | |
615 | if (Opcode == Instruction::FDiv && I.hasAllowReassoc() && I.hasNoNaNs()) { |
616 | // powi(X, Y) / X --> powi(X, Y-1) |
617 | // This is legal when (Y - 1) can't wraparound, in which case reassoc and |
618 | // nnan are required. |
619 | // TODO: Multi-use may be also better off creating Powi(x,y-1) |
620 | if (match(Op0, m_OneUse(m_AllowReassoc(m_Intrinsic<Intrinsic::powi>( |
621 | m_Specific(Op1), m_Value(Y))))) && |
622 | willNotOverflowSignedSub(Y, ConstantInt::get(Y->getType(), 1), I)) { |
623 | Constant *NegOne = ConstantInt::getAllOnesValue(Ty: Y->getType()); |
624 | Instruction *NewPow = createPowiExpr(I, *this, Op1, Y, NegOne); |
625 | return replaceInstUsesWith(I, V: NewPow); |
626 | } |
627 | |
628 | // powi(X, Y) / (X * Z) --> powi(X, Y-1) / Z |
629 | // This is legal when (Y - 1) can't wraparound, in which case reassoc and |
630 | // nnan are required. |
631 | // TODO: Multi-use may be also better off creating Powi(x,y-1) |
632 | if (match(Op0, m_OneUse(m_AllowReassoc(m_Intrinsic<Intrinsic::powi>( |
633 | m_Value(X), m_Value(Y))))) && |
634 | match(Op1, m_AllowReassoc(m_c_FMul(m_Specific(X), m_Value(Z)))) && |
635 | willNotOverflowSignedSub(Y, ConstantInt::get(Y->getType(), 1), I)) { |
636 | Constant *NegOne = ConstantInt::getAllOnesValue(Ty: Y->getType()); |
637 | auto *NewPow = createPowiExpr(I, *this, X, Y, NegOne); |
638 | return BinaryOperator::CreateFDivFMF(NewPow, Z, &I); |
639 | } |
640 | } |
641 | |
642 | return nullptr; |
643 | } |
644 | |
645 | Instruction *InstCombinerImpl::foldFMulReassoc(BinaryOperator &I) { |
646 | Value *Op0 = I.getOperand(i_nocapture: 0); |
647 | Value *Op1 = I.getOperand(i_nocapture: 1); |
648 | Value *X, *Y; |
649 | Constant *C; |
650 | BinaryOperator *Op0BinOp; |
651 | |
652 | // Reassociate constant RHS with another constant to form constant |
653 | // expression. |
654 | if (match(V: Op1, P: m_Constant(C)) && C->isFiniteNonZeroFP() && |
655 | match(V: Op0, P: m_AllowReassoc(SubPattern: m_BinOp(I&: Op0BinOp)))) { |
656 | // Everything in this scope folds I with Op0, intersecting their FMF. |
657 | FastMathFlags FMF = I.getFastMathFlags() & Op0BinOp->getFastMathFlags(); |
658 | IRBuilder<>::FastMathFlagGuard FMFGuard(Builder); |
659 | Builder.setFastMathFlags(FMF); |
660 | Constant *C1; |
661 | if (match(V: Op0, P: m_OneUse(SubPattern: m_FDiv(L: m_Constant(C&: C1), R: m_Value(V&: X))))) { |
662 | // (C1 / X) * C --> (C * C1) / X |
663 | Constant *CC1 = |
664 | ConstantFoldBinaryOpOperands(Opcode: Instruction::FMul, LHS: C, RHS: C1, DL); |
665 | if (CC1 && CC1->isNormalFP()) |
666 | return BinaryOperator::CreateFDivFMF(V1: CC1, V2: X, FMF); |
667 | } |
668 | if (match(V: Op0, P: m_FDiv(L: m_Value(V&: X), R: m_Constant(C&: C1)))) { |
669 | // FIXME: This seems like it should also be checking for arcp |
670 | // (X / C1) * C --> X * (C / C1) |
671 | Constant *CDivC1 = |
672 | ConstantFoldBinaryOpOperands(Opcode: Instruction::FDiv, LHS: C, RHS: C1, DL); |
673 | if (CDivC1 && CDivC1->isNormalFP()) |
674 | return BinaryOperator::CreateFMulFMF(V1: X, V2: CDivC1, FMF); |
675 | |
676 | // If the constant was a denormal, try reassociating differently. |
677 | // (X / C1) * C --> X / (C1 / C) |
678 | Constant *C1DivC = |
679 | ConstantFoldBinaryOpOperands(Opcode: Instruction::FDiv, LHS: C1, RHS: C, DL); |
680 | if (C1DivC && Op0->hasOneUse() && C1DivC->isNormalFP()) |
681 | return BinaryOperator::CreateFDivFMF(V1: X, V2: C1DivC, FMF); |
682 | } |
683 | |
684 | // We do not need to match 'fadd C, X' and 'fsub X, C' because they are |
685 | // canonicalized to 'fadd X, C'. Distributing the multiply may allow |
686 | // further folds and (X * C) + C2 is 'fma'. |
687 | if (match(V: Op0, P: m_OneUse(SubPattern: m_FAdd(L: m_Value(V&: X), R: m_Constant(C&: C1))))) { |
688 | // (X + C1) * C --> (X * C) + (C * C1) |
689 | if (Constant *CC1 = |
690 | ConstantFoldBinaryOpOperands(Opcode: Instruction::FMul, LHS: C, RHS: C1, DL)) { |
691 | Value *XC = Builder.CreateFMul(L: X, R: C); |
692 | return BinaryOperator::CreateFAddFMF(V1: XC, V2: CC1, FMF); |
693 | } |
694 | } |
695 | if (match(V: Op0, P: m_OneUse(SubPattern: m_FSub(L: m_Constant(C&: C1), R: m_Value(V&: X))))) { |
696 | // (C1 - X) * C --> (C * C1) - (X * C) |
697 | if (Constant *CC1 = |
698 | ConstantFoldBinaryOpOperands(Opcode: Instruction::FMul, LHS: C, RHS: C1, DL)) { |
699 | Value *XC = Builder.CreateFMul(L: X, R: C); |
700 | return BinaryOperator::CreateFSubFMF(V1: CC1, V2: XC, FMF); |
701 | } |
702 | } |
703 | } |
704 | |
705 | Value *Z; |
706 | if (match(V: &I, |
707 | P: m_c_FMul(L: m_AllowReassoc(SubPattern: m_OneUse(SubPattern: m_FDiv(L: m_Value(V&: X), R: m_Value(V&: Y)))), |
708 | R: m_Value(V&: Z)))) { |
709 | BinaryOperator *DivOp = cast<BinaryOperator>(Val: ((Z == Op0) ? Op1 : Op0)); |
710 | FastMathFlags FMF = I.getFastMathFlags() & DivOp->getFastMathFlags(); |
711 | if (FMF.allowReassoc()) { |
712 | // Sink division: (X / Y) * Z --> (X * Z) / Y |
713 | IRBuilder<>::FastMathFlagGuard FMFGuard(Builder); |
714 | Builder.setFastMathFlags(FMF); |
715 | auto *NewFMul = Builder.CreateFMul(L: X, R: Z); |
716 | return BinaryOperator::CreateFDivFMF(V1: NewFMul, V2: Y, FMF); |
717 | } |
718 | } |
719 | |
720 | // sqrt(X) * sqrt(Y) -> sqrt(X * Y) |
721 | // nnan disallows the possibility of returning a number if both operands are |
722 | // negative (in that case, we should return NaN). |
723 | if (I.hasNoNaNs() && match(V: Op0, P: m_OneUse(SubPattern: m_Sqrt(Op0: m_Value(V&: X)))) && |
724 | match(V: Op1, P: m_OneUse(SubPattern: m_Sqrt(Op0: m_Value(V&: Y))))) { |
725 | Value *XY = Builder.CreateFMulFMF(L: X, R: Y, FMFSource: &I); |
726 | Value *Sqrt = Builder.CreateUnaryIntrinsic(Intrinsic::ID: sqrt, V: XY, FMFSource: &I); |
727 | return replaceInstUsesWith(I, V: Sqrt); |
728 | } |
729 | |
730 | // The following transforms are done irrespective of the number of uses |
731 | // for the expression "1.0/sqrt(X)". |
732 | // 1) 1.0/sqrt(X) * X -> X/sqrt(X) |
733 | // 2) X * 1.0/sqrt(X) -> X/sqrt(X) |
734 | // We always expect the backend to reduce X/sqrt(X) to sqrt(X), if it |
735 | // has the necessary (reassoc) fast-math-flags. |
736 | if (I.hasNoSignedZeros() && |
737 | match(V: Op0, P: (m_FDiv(L: m_SpecificFP(V: 1.0), R: m_Value(V&: Y)))) && |
738 | match(V: Y, P: m_Sqrt(Op0: m_Value(V&: X))) && Op1 == X) |
739 | return BinaryOperator::CreateFDivFMF(V1: X, V2: Y, FMFSource: &I); |
740 | if (I.hasNoSignedZeros() && |
741 | match(V: Op1, P: (m_FDiv(L: m_SpecificFP(V: 1.0), R: m_Value(V&: Y)))) && |
742 | match(V: Y, P: m_Sqrt(Op0: m_Value(V&: X))) && Op0 == X) |
743 | return BinaryOperator::CreateFDivFMF(V1: X, V2: Y, FMFSource: &I); |
744 | |
745 | // Like the similar transform in instsimplify, this requires 'nsz' because |
746 | // sqrt(-0.0) = -0.0, and -0.0 * -0.0 does not simplify to -0.0. |
747 | if (I.hasNoNaNs() && I.hasNoSignedZeros() && Op0 == Op1 && Op0->hasNUses(N: 2)) { |
748 | // Peek through fdiv to find squaring of square root: |
749 | // (X / sqrt(Y)) * (X / sqrt(Y)) --> (X * X) / Y |
750 | if (match(V: Op0, P: m_FDiv(L: m_Value(V&: X), R: m_Sqrt(Op0: m_Value(V&: Y))))) { |
751 | Value *XX = Builder.CreateFMulFMF(L: X, R: X, FMFSource: &I); |
752 | return BinaryOperator::CreateFDivFMF(V1: XX, V2: Y, FMFSource: &I); |
753 | } |
754 | // (sqrt(Y) / X) * (sqrt(Y) / X) --> Y / (X * X) |
755 | if (match(V: Op0, P: m_FDiv(L: m_Sqrt(Op0: m_Value(V&: Y)), R: m_Value(V&: X)))) { |
756 | Value *XX = Builder.CreateFMulFMF(L: X, R: X, FMFSource: &I); |
757 | return BinaryOperator::CreateFDivFMF(V1: Y, V2: XX, FMFSource: &I); |
758 | } |
759 | } |
760 | |
761 | // pow(X, Y) * X --> pow(X, Y+1) |
762 | // X * pow(X, Y) --> pow(X, Y+1) |
763 | if (match(&I, m_c_FMul(m_OneUse(m_Intrinsic<Intrinsic::pow>(m_Value(X), |
764 | m_Value(Y))), |
765 | m_Deferred(X)))) { |
766 | Value *Y1 = Builder.CreateFAddFMF(L: Y, R: ConstantFP::get(Ty: I.getType(), V: 1.0), FMFSource: &I); |
767 | Value *Pow = Builder.CreateBinaryIntrinsic(Intrinsic::ID: pow, LHS: X, RHS: Y1, FMFSource: &I); |
768 | return replaceInstUsesWith(I, V: Pow); |
769 | } |
770 | |
771 | if (Instruction *FoldedPowi = foldPowiReassoc(I)) |
772 | return FoldedPowi; |
773 | |
774 | if (I.isOnlyUserOfAnyOperand()) { |
775 | // pow(X, Y) * pow(X, Z) -> pow(X, Y + Z) |
776 | if (match(Op0, m_Intrinsic<Intrinsic::pow>(m_Value(X), m_Value(Y))) && |
777 | match(Op1, m_Intrinsic<Intrinsic::pow>(m_Specific(X), m_Value(Z)))) { |
778 | auto *YZ = Builder.CreateFAddFMF(L: Y, R: Z, FMFSource: &I); |
779 | auto *NewPow = Builder.CreateBinaryIntrinsic(Intrinsic::pow, X, YZ, &I); |
780 | return replaceInstUsesWith(I, V: NewPow); |
781 | } |
782 | // pow(X, Y) * pow(Z, Y) -> pow(X * Z, Y) |
783 | if (match(Op0, m_Intrinsic<Intrinsic::pow>(m_Value(X), m_Value(Y))) && |
784 | match(Op1, m_Intrinsic<Intrinsic::pow>(m_Value(Z), m_Specific(Y)))) { |
785 | auto *XZ = Builder.CreateFMulFMF(L: X, R: Z, FMFSource: &I); |
786 | auto *NewPow = Builder.CreateBinaryIntrinsic(Intrinsic::pow, XZ, Y, &I); |
787 | return replaceInstUsesWith(I, V: NewPow); |
788 | } |
789 | |
790 | // exp(X) * exp(Y) -> exp(X + Y) |
791 | if (match(Op0, m_Intrinsic<Intrinsic::exp>(m_Value(X))) && |
792 | match(Op1, m_Intrinsic<Intrinsic::exp>(m_Value(Y)))) { |
793 | Value *XY = Builder.CreateFAddFMF(L: X, R: Y, FMFSource: &I); |
794 | Value *Exp = Builder.CreateUnaryIntrinsic(Intrinsic::ID: exp, V: XY, FMFSource: &I); |
795 | return replaceInstUsesWith(I, V: Exp); |
796 | } |
797 | |
798 | // exp2(X) * exp2(Y) -> exp2(X + Y) |
799 | if (match(Op0, m_Intrinsic<Intrinsic::exp2>(m_Value(X))) && |
800 | match(Op1, m_Intrinsic<Intrinsic::exp2>(m_Value(Y)))) { |
801 | Value *XY = Builder.CreateFAddFMF(L: X, R: Y, FMFSource: &I); |
802 | Value *Exp2 = Builder.CreateUnaryIntrinsic(Intrinsic::ID: exp2, V: XY, FMFSource: &I); |
803 | return replaceInstUsesWith(I, V: Exp2); |
804 | } |
805 | } |
806 | |
807 | // (X*Y) * X => (X*X) * Y where Y != X |
808 | // The purpose is two-fold: |
809 | // 1) to form a power expression (of X). |
810 | // 2) potentially shorten the critical path: After transformation, the |
811 | // latency of the instruction Y is amortized by the expression of X*X, |
812 | // and therefore Y is in a "less critical" position compared to what it |
813 | // was before the transformation. |
814 | if (match(V: Op0, P: m_OneUse(SubPattern: m_c_FMul(L: m_Specific(V: Op1), R: m_Value(V&: Y)))) && Op1 != Y) { |
815 | Value *XX = Builder.CreateFMulFMF(L: Op1, R: Op1, FMFSource: &I); |
816 | return BinaryOperator::CreateFMulFMF(V1: XX, V2: Y, FMFSource: &I); |
817 | } |
818 | if (match(V: Op1, P: m_OneUse(SubPattern: m_c_FMul(L: m_Specific(V: Op0), R: m_Value(V&: Y)))) && Op0 != Y) { |
819 | Value *XX = Builder.CreateFMulFMF(L: Op0, R: Op0, FMFSource: &I); |
820 | return BinaryOperator::CreateFMulFMF(V1: XX, V2: Y, FMFSource: &I); |
821 | } |
822 | |
823 | return nullptr; |
824 | } |
825 | |
826 | Instruction *InstCombinerImpl::visitFMul(BinaryOperator &I) { |
827 | if (Value *V = simplifyFMulInst(LHS: I.getOperand(i_nocapture: 0), RHS: I.getOperand(i_nocapture: 1), |
828 | FMF: I.getFastMathFlags(), |
829 | Q: SQ.getWithInstruction(I: &I))) |
830 | return replaceInstUsesWith(I, V); |
831 | |
832 | if (SimplifyAssociativeOrCommutative(I)) |
833 | return &I; |
834 | |
835 | if (Instruction *X = foldVectorBinop(Inst&: I)) |
836 | return X; |
837 | |
838 | if (Instruction *Phi = foldBinopWithPhiOperands(BO&: I)) |
839 | return Phi; |
840 | |
841 | if (Instruction *FoldedMul = foldBinOpIntoSelectOrPhi(I)) |
842 | return FoldedMul; |
843 | |
844 | if (Value *FoldedMul = foldMulSelectToNegate(I, Builder)) |
845 | return replaceInstUsesWith(I, V: FoldedMul); |
846 | |
847 | if (Instruction *R = foldFPSignBitOps(I)) |
848 | return R; |
849 | |
850 | if (Instruction *R = foldFBinOpOfIntCasts(I)) |
851 | return R; |
852 | |
853 | // X * -1.0 --> -X |
854 | Value *Op0 = I.getOperand(i_nocapture: 0), *Op1 = I.getOperand(i_nocapture: 1); |
855 | if (match(V: Op1, P: m_SpecificFP(V: -1.0))) |
856 | return UnaryOperator::CreateFNegFMF(Op: Op0, FMFSource: &I); |
857 | |
858 | // With no-nans/no-infs: |
859 | // X * 0.0 --> copysign(0.0, X) |
860 | // X * -0.0 --> copysign(0.0, -X) |
861 | const APFloat *FPC; |
862 | if (match(V: Op1, P: m_APFloatAllowPoison(Res&: FPC)) && FPC->isZero() && |
863 | ((I.hasNoInfs() && |
864 | isKnownNeverNaN(V: Op0, /*Depth=*/0, SQ: SQ.getWithInstruction(I: &I))) || |
865 | isKnownNeverNaN(V: &I, /*Depth=*/0, SQ: SQ.getWithInstruction(I: &I)))) { |
866 | if (FPC->isNegative()) |
867 | Op0 = Builder.CreateFNegFMF(V: Op0, FMFSource: &I); |
868 | CallInst *CopySign = Builder.CreateIntrinsic(Intrinsic::copysign, |
869 | {I.getType()}, {Op1, Op0}, &I); |
870 | return replaceInstUsesWith(I, V: CopySign); |
871 | } |
872 | |
873 | // -X * C --> X * -C |
874 | Value *X, *Y; |
875 | Constant *C; |
876 | if (match(V: Op0, P: m_FNeg(X: m_Value(V&: X))) && match(V: Op1, P: m_Constant(C))) |
877 | if (Constant *NegC = ConstantFoldUnaryOpOperand(Opcode: Instruction::FNeg, Op: C, DL)) |
878 | return BinaryOperator::CreateFMulFMF(V1: X, V2: NegC, FMFSource: &I); |
879 | |
880 | // (select A, B, C) * (select A, D, E) --> select A, (B*D), (C*E) |
881 | if (Value *V = SimplifySelectsFeedingBinaryOp(I, LHS: Op0, RHS: Op1)) |
882 | return replaceInstUsesWith(I, V); |
883 | |
884 | if (I.hasAllowReassoc()) |
885 | if (Instruction *FoldedMul = foldFMulReassoc(I)) |
886 | return FoldedMul; |
887 | |
888 | // log2(X * 0.5) * Y = log2(X) * Y - Y |
889 | if (I.isFast()) { |
890 | IntrinsicInst *Log2 = nullptr; |
891 | if (match(Op0, m_OneUse(m_Intrinsic<Intrinsic::log2>( |
892 | m_OneUse(SubPattern: m_FMul(L: m_Value(V&: X), R: m_SpecificFP(V: 0.5))))))) { |
893 | Log2 = cast<IntrinsicInst>(Val: Op0); |
894 | Y = Op1; |
895 | } |
896 | if (match(Op1, m_OneUse(m_Intrinsic<Intrinsic::log2>( |
897 | m_OneUse(SubPattern: m_FMul(L: m_Value(V&: X), R: m_SpecificFP(V: 0.5))))))) { |
898 | Log2 = cast<IntrinsicInst>(Val: Op1); |
899 | Y = Op0; |
900 | } |
901 | if (Log2) { |
902 | Value *Log2 = Builder.CreateUnaryIntrinsic(Intrinsic::ID: log2, V: X, FMFSource: &I); |
903 | Value *LogXTimesY = Builder.CreateFMulFMF(L: Log2, R: Y, FMFSource: &I); |
904 | return BinaryOperator::CreateFSubFMF(V1: LogXTimesY, V2: Y, FMFSource: &I); |
905 | } |
906 | } |
907 | |
908 | // Simplify FMUL recurrences starting with 0.0 to 0.0 if nnan and nsz are set. |
909 | // Given a phi node with entry value as 0 and it used in fmul operation, |
910 | // we can replace fmul with 0 safely and eleminate loop operation. |
911 | PHINode *PN = nullptr; |
912 | Value *Start = nullptr, *Step = nullptr; |
913 | if (matchSimpleRecurrence(I: &I, P&: PN, Start, Step) && I.hasNoNaNs() && |
914 | I.hasNoSignedZeros() && match(V: Start, P: m_Zero())) |
915 | return replaceInstUsesWith(I, V: Start); |
916 | |
917 | // minimum(X, Y) * maximum(X, Y) => X * Y. |
918 | if (match(&I, |
919 | m_c_FMul(m_Intrinsic<Intrinsic::maximum>(m_Value(V&: X), m_Value(V&: Y)), |
920 | m_c_Intrinsic<Intrinsic::minimum>(m_Deferred(V: X), |
921 | m_Deferred(V: Y))))) { |
922 | BinaryOperator *Result = BinaryOperator::CreateFMulFMF(V1: X, V2: Y, FMFSource: &I); |
923 | // We cannot preserve ninf if nnan flag is not set. |
924 | // If X is NaN and Y is Inf then in original program we had NaN * NaN, |
925 | // while in optimized version NaN * Inf and this is a poison with ninf flag. |
926 | if (!Result->hasNoNaNs()) |
927 | Result->setHasNoInfs(false); |
928 | return Result; |
929 | } |
930 | |
931 | return nullptr; |
932 | } |
933 | |
934 | /// Fold a divide or remainder with a select instruction divisor when one of the |
935 | /// select operands is zero. In that case, we can use the other select operand |
936 | /// because div/rem by zero is undefined. |
937 | bool InstCombinerImpl::simplifyDivRemOfSelectWithZeroOp(BinaryOperator &I) { |
938 | SelectInst *SI = dyn_cast<SelectInst>(Val: I.getOperand(i_nocapture: 1)); |
939 | if (!SI) |
940 | return false; |
941 | |
942 | int NonNullOperand; |
943 | if (match(V: SI->getTrueValue(), P: m_Zero())) |
944 | // div/rem X, (Cond ? 0 : Y) -> div/rem X, Y |
945 | NonNullOperand = 2; |
946 | else if (match(V: SI->getFalseValue(), P: m_Zero())) |
947 | // div/rem X, (Cond ? Y : 0) -> div/rem X, Y |
948 | NonNullOperand = 1; |
949 | else |
950 | return false; |
951 | |
952 | // Change the div/rem to use 'Y' instead of the select. |
953 | replaceOperand(I, OpNum: 1, V: SI->getOperand(i_nocapture: NonNullOperand)); |
954 | |
955 | // Okay, we know we replace the operand of the div/rem with 'Y' with no |
956 | // problem. However, the select, or the condition of the select may have |
957 | // multiple uses. Based on our knowledge that the operand must be non-zero, |
958 | // propagate the known value for the select into other uses of it, and |
959 | // propagate a known value of the condition into its other users. |
960 | |
961 | // If the select and condition only have a single use, don't bother with this, |
962 | // early exit. |
963 | Value *SelectCond = SI->getCondition(); |
964 | if (SI->use_empty() && SelectCond->hasOneUse()) |
965 | return true; |
966 | |
967 | // Scan the current block backward, looking for other uses of SI. |
968 | BasicBlock::iterator BBI = I.getIterator(), BBFront = I.getParent()->begin(); |
969 | Type *CondTy = SelectCond->getType(); |
970 | while (BBI != BBFront) { |
971 | --BBI; |
972 | // If we found an instruction that we can't assume will return, so |
973 | // information from below it cannot be propagated above it. |
974 | if (!isGuaranteedToTransferExecutionToSuccessor(I: &*BBI)) |
975 | break; |
976 | |
977 | // Replace uses of the select or its condition with the known values. |
978 | for (Use &Op : BBI->operands()) { |
979 | if (Op == SI) { |
980 | replaceUse(U&: Op, NewValue: SI->getOperand(i_nocapture: NonNullOperand)); |
981 | Worklist.push(I: &*BBI); |
982 | } else if (Op == SelectCond) { |
983 | replaceUse(U&: Op, NewValue: NonNullOperand == 1 ? ConstantInt::getTrue(Ty: CondTy) |
984 | : ConstantInt::getFalse(Ty: CondTy)); |
985 | Worklist.push(I: &*BBI); |
986 | } |
987 | } |
988 | |
989 | // If we past the instruction, quit looking for it. |
990 | if (&*BBI == SI) |
991 | SI = nullptr; |
992 | if (&*BBI == SelectCond) |
993 | SelectCond = nullptr; |
994 | |
995 | // If we ran out of things to eliminate, break out of the loop. |
996 | if (!SelectCond && !SI) |
997 | break; |
998 | |
999 | } |
1000 | return true; |
1001 | } |
1002 | |
1003 | /// True if the multiply can not be expressed in an int this size. |
1004 | static bool multiplyOverflows(const APInt &C1, const APInt &C2, APInt &Product, |
1005 | bool IsSigned) { |
1006 | bool Overflow; |
1007 | Product = IsSigned ? C1.smul_ov(RHS: C2, Overflow) : C1.umul_ov(RHS: C2, Overflow); |
1008 | return Overflow; |
1009 | } |
1010 | |
1011 | /// True if C1 is a multiple of C2. Quotient contains C1/C2. |
1012 | static bool isMultiple(const APInt &C1, const APInt &C2, APInt &Quotient, |
1013 | bool IsSigned) { |
1014 | assert(C1.getBitWidth() == C2.getBitWidth() && "Constant widths not equal" ); |
1015 | |
1016 | // Bail if we will divide by zero. |
1017 | if (C2.isZero()) |
1018 | return false; |
1019 | |
1020 | // Bail if we would divide INT_MIN by -1. |
1021 | if (IsSigned && C1.isMinSignedValue() && C2.isAllOnes()) |
1022 | return false; |
1023 | |
1024 | APInt Remainder(C1.getBitWidth(), /*val=*/0ULL, IsSigned); |
1025 | if (IsSigned) |
1026 | APInt::sdivrem(LHS: C1, RHS: C2, Quotient, Remainder); |
1027 | else |
1028 | APInt::udivrem(LHS: C1, RHS: C2, Quotient, Remainder); |
1029 | |
1030 | return Remainder.isMinValue(); |
1031 | } |
1032 | |
1033 | static Value *foldIDivShl(BinaryOperator &I, InstCombiner::BuilderTy &Builder) { |
1034 | assert((I.getOpcode() == Instruction::SDiv || |
1035 | I.getOpcode() == Instruction::UDiv) && |
1036 | "Expected integer divide" ); |
1037 | |
1038 | bool IsSigned = I.getOpcode() == Instruction::SDiv; |
1039 | Value *Op0 = I.getOperand(i_nocapture: 0), *Op1 = I.getOperand(i_nocapture: 1); |
1040 | Type *Ty = I.getType(); |
1041 | |
1042 | Value *X, *Y, *Z; |
1043 | |
1044 | // With appropriate no-wrap constraints, remove a common factor in the |
1045 | // dividend and divisor that is disguised as a left-shifted value. |
1046 | if (match(V: Op1, P: m_Shl(L: m_Value(V&: X), R: m_Value(V&: Z))) && |
1047 | match(V: Op0, P: m_c_Mul(L: m_Specific(V: X), R: m_Value(V&: Y)))) { |
1048 | // Both operands must have the matching no-wrap for this kind of division. |
1049 | auto *Mul = cast<OverflowingBinaryOperator>(Val: Op0); |
1050 | auto *Shl = cast<OverflowingBinaryOperator>(Val: Op1); |
1051 | bool HasNUW = Mul->hasNoUnsignedWrap() && Shl->hasNoUnsignedWrap(); |
1052 | bool HasNSW = Mul->hasNoSignedWrap() && Shl->hasNoSignedWrap(); |
1053 | |
1054 | // (X * Y) u/ (X << Z) --> Y u>> Z |
1055 | if (!IsSigned && HasNUW) |
1056 | return Builder.CreateLShr(LHS: Y, RHS: Z, Name: "" , isExact: I.isExact()); |
1057 | |
1058 | // (X * Y) s/ (X << Z) --> Y s/ (1 << Z) |
1059 | if (IsSigned && HasNSW && (Op0->hasOneUse() || Op1->hasOneUse())) { |
1060 | Value *Shl = Builder.CreateShl(LHS: ConstantInt::get(Ty, V: 1), RHS: Z); |
1061 | return Builder.CreateSDiv(LHS: Y, RHS: Shl, Name: "" , isExact: I.isExact()); |
1062 | } |
1063 | } |
1064 | |
1065 | // With appropriate no-wrap constraints, remove a common factor in the |
1066 | // dividend and divisor that is disguised as a left-shift amount. |
1067 | if (match(V: Op0, P: m_Shl(L: m_Value(V&: X), R: m_Value(V&: Z))) && |
1068 | match(V: Op1, P: m_Shl(L: m_Value(V&: Y), R: m_Specific(V: Z)))) { |
1069 | auto *Shl0 = cast<OverflowingBinaryOperator>(Val: Op0); |
1070 | auto *Shl1 = cast<OverflowingBinaryOperator>(Val: Op1); |
1071 | |
1072 | // For unsigned div, we need 'nuw' on both shifts or |
1073 | // 'nsw' on both shifts + 'nuw' on the dividend. |
1074 | // (X << Z) / (Y << Z) --> X / Y |
1075 | if (!IsSigned && |
1076 | ((Shl0->hasNoUnsignedWrap() && Shl1->hasNoUnsignedWrap()) || |
1077 | (Shl0->hasNoUnsignedWrap() && Shl0->hasNoSignedWrap() && |
1078 | Shl1->hasNoSignedWrap()))) |
1079 | return Builder.CreateUDiv(LHS: X, RHS: Y, Name: "" , isExact: I.isExact()); |
1080 | |
1081 | // For signed div, we need 'nsw' on both shifts + 'nuw' on the divisor. |
1082 | // (X << Z) / (Y << Z) --> X / Y |
1083 | if (IsSigned && Shl0->hasNoSignedWrap() && Shl1->hasNoSignedWrap() && |
1084 | Shl1->hasNoUnsignedWrap()) |
1085 | return Builder.CreateSDiv(LHS: X, RHS: Y, Name: "" , isExact: I.isExact()); |
1086 | } |
1087 | |
1088 | // If X << Y and X << Z does not overflow, then: |
1089 | // (X << Y) / (X << Z) -> (1 << Y) / (1 << Z) -> 1 << Y >> Z |
1090 | if (match(V: Op0, P: m_Shl(L: m_Value(V&: X), R: m_Value(V&: Y))) && |
1091 | match(V: Op1, P: m_Shl(L: m_Specific(V: X), R: m_Value(V&: Z)))) { |
1092 | auto *Shl0 = cast<OverflowingBinaryOperator>(Val: Op0); |
1093 | auto *Shl1 = cast<OverflowingBinaryOperator>(Val: Op1); |
1094 | |
1095 | if (IsSigned ? (Shl0->hasNoSignedWrap() && Shl1->hasNoSignedWrap()) |
1096 | : (Shl0->hasNoUnsignedWrap() && Shl1->hasNoUnsignedWrap())) { |
1097 | Constant *One = ConstantInt::get(Ty: X->getType(), V: 1); |
1098 | // Only preserve the nsw flag if dividend has nsw |
1099 | // or divisor has nsw and operator is sdiv. |
1100 | Value *Dividend = Builder.CreateShl( |
1101 | LHS: One, RHS: Y, Name: "shl.dividend" , |
1102 | /*HasNUW*/ true, |
1103 | /*HasNSW*/ |
1104 | IsSigned ? (Shl0->hasNoUnsignedWrap() || Shl1->hasNoUnsignedWrap()) |
1105 | : Shl0->hasNoSignedWrap()); |
1106 | return Builder.CreateLShr(LHS: Dividend, RHS: Z, Name: "" , isExact: I.isExact()); |
1107 | } |
1108 | } |
1109 | |
1110 | return nullptr; |
1111 | } |
1112 | |
1113 | /// This function implements the transforms common to both integer division |
1114 | /// instructions (udiv and sdiv). It is called by the visitors to those integer |
1115 | /// division instructions. |
1116 | /// Common integer divide transforms |
1117 | Instruction *InstCombinerImpl::commonIDivTransforms(BinaryOperator &I) { |
1118 | if (Instruction *Phi = foldBinopWithPhiOperands(BO&: I)) |
1119 | return Phi; |
1120 | |
1121 | Value *Op0 = I.getOperand(i_nocapture: 0), *Op1 = I.getOperand(i_nocapture: 1); |
1122 | bool IsSigned = I.getOpcode() == Instruction::SDiv; |
1123 | Type *Ty = I.getType(); |
1124 | |
1125 | // The RHS is known non-zero. |
1126 | if (Value *V = simplifyValueKnownNonZero(V: I.getOperand(i_nocapture: 1), IC&: *this, CxtI&: I)) |
1127 | return replaceOperand(I, OpNum: 1, V); |
1128 | |
1129 | // Handle cases involving: [su]div X, (select Cond, Y, Z) |
1130 | // This does not apply for fdiv. |
1131 | if (simplifyDivRemOfSelectWithZeroOp(I)) |
1132 | return &I; |
1133 | |
1134 | // If the divisor is a select-of-constants, try to constant fold all div ops: |
1135 | // C / (select Cond, TrueC, FalseC) --> select Cond, (C / TrueC), (C / FalseC) |
1136 | // TODO: Adapt simplifyDivRemOfSelectWithZeroOp to allow this and other folds. |
1137 | if (match(V: Op0, P: m_ImmConstant()) && |
1138 | match(V: Op1, P: m_Select(C: m_Value(), L: m_ImmConstant(), R: m_ImmConstant()))) { |
1139 | if (Instruction *R = FoldOpIntoSelect(Op&: I, SI: cast<SelectInst>(Val: Op1), |
1140 | /*FoldWithMultiUse*/ true)) |
1141 | return R; |
1142 | } |
1143 | |
1144 | const APInt *C2; |
1145 | if (match(V: Op1, P: m_APInt(Res&: C2))) { |
1146 | Value *X; |
1147 | const APInt *C1; |
1148 | |
1149 | // (X / C1) / C2 -> X / (C1*C2) |
1150 | if ((IsSigned && match(V: Op0, P: m_SDiv(L: m_Value(V&: X), R: m_APInt(Res&: C1)))) || |
1151 | (!IsSigned && match(V: Op0, P: m_UDiv(L: m_Value(V&: X), R: m_APInt(Res&: C1))))) { |
1152 | APInt Product(C1->getBitWidth(), /*val=*/0ULL, IsSigned); |
1153 | if (!multiplyOverflows(C1: *C1, C2: *C2, Product, IsSigned)) |
1154 | return BinaryOperator::Create(Op: I.getOpcode(), S1: X, |
1155 | S2: ConstantInt::get(Ty, V: Product)); |
1156 | } |
1157 | |
1158 | APInt Quotient(C2->getBitWidth(), /*val=*/0ULL, IsSigned); |
1159 | if ((IsSigned && match(V: Op0, P: m_NSWMul(L: m_Value(V&: X), R: m_APInt(Res&: C1)))) || |
1160 | (!IsSigned && match(V: Op0, P: m_NUWMul(L: m_Value(V&: X), R: m_APInt(Res&: C1))))) { |
1161 | |
1162 | // (X * C1) / C2 -> X / (C2 / C1) if C2 is a multiple of C1. |
1163 | if (isMultiple(C1: *C2, C2: *C1, Quotient, IsSigned)) { |
1164 | auto *NewDiv = BinaryOperator::Create(Op: I.getOpcode(), S1: X, |
1165 | S2: ConstantInt::get(Ty, V: Quotient)); |
1166 | NewDiv->setIsExact(I.isExact()); |
1167 | return NewDiv; |
1168 | } |
1169 | |
1170 | // (X * C1) / C2 -> X * (C1 / C2) if C1 is a multiple of C2. |
1171 | if (isMultiple(C1: *C1, C2: *C2, Quotient, IsSigned)) { |
1172 | auto *Mul = BinaryOperator::Create(Op: Instruction::Mul, S1: X, |
1173 | S2: ConstantInt::get(Ty, V: Quotient)); |
1174 | auto *OBO = cast<OverflowingBinaryOperator>(Val: Op0); |
1175 | Mul->setHasNoUnsignedWrap(!IsSigned && OBO->hasNoUnsignedWrap()); |
1176 | Mul->setHasNoSignedWrap(OBO->hasNoSignedWrap()); |
1177 | return Mul; |
1178 | } |
1179 | } |
1180 | |
1181 | if ((IsSigned && match(V: Op0, P: m_NSWShl(L: m_Value(V&: X), R: m_APInt(Res&: C1))) && |
1182 | C1->ult(RHS: C1->getBitWidth() - 1)) || |
1183 | (!IsSigned && match(V: Op0, P: m_NUWShl(L: m_Value(V&: X), R: m_APInt(Res&: C1))) && |
1184 | C1->ult(RHS: C1->getBitWidth()))) { |
1185 | APInt C1Shifted = APInt::getOneBitSet( |
1186 | numBits: C1->getBitWidth(), BitNo: static_cast<unsigned>(C1->getZExtValue())); |
1187 | |
1188 | // (X << C1) / C2 -> X / (C2 >> C1) if C2 is a multiple of 1 << C1. |
1189 | if (isMultiple(C1: *C2, C2: C1Shifted, Quotient, IsSigned)) { |
1190 | auto *BO = BinaryOperator::Create(Op: I.getOpcode(), S1: X, |
1191 | S2: ConstantInt::get(Ty, V: Quotient)); |
1192 | BO->setIsExact(I.isExact()); |
1193 | return BO; |
1194 | } |
1195 | |
1196 | // (X << C1) / C2 -> X * ((1 << C1) / C2) if 1 << C1 is a multiple of C2. |
1197 | if (isMultiple(C1: C1Shifted, C2: *C2, Quotient, IsSigned)) { |
1198 | auto *Mul = BinaryOperator::Create(Op: Instruction::Mul, S1: X, |
1199 | S2: ConstantInt::get(Ty, V: Quotient)); |
1200 | auto *OBO = cast<OverflowingBinaryOperator>(Val: Op0); |
1201 | Mul->setHasNoUnsignedWrap(!IsSigned && OBO->hasNoUnsignedWrap()); |
1202 | Mul->setHasNoSignedWrap(OBO->hasNoSignedWrap()); |
1203 | return Mul; |
1204 | } |
1205 | } |
1206 | |
1207 | // Distribute div over add to eliminate a matching div/mul pair: |
1208 | // ((X * C2) + C1) / C2 --> X + C1/C2 |
1209 | // We need a multiple of the divisor for a signed add constant, but |
1210 | // unsigned is fine with any constant pair. |
1211 | if (IsSigned && |
1212 | match(V: Op0, P: m_NSWAddLike(L: m_NSWMul(L: m_Value(V&: X), R: m_SpecificInt(V: *C2)), |
1213 | R: m_APInt(Res&: C1))) && |
1214 | isMultiple(C1: *C1, C2: *C2, Quotient, IsSigned)) { |
1215 | return BinaryOperator::CreateNSWAdd(V1: X, V2: ConstantInt::get(Ty, V: Quotient)); |
1216 | } |
1217 | if (!IsSigned && |
1218 | match(V: Op0, P: m_NUWAddLike(L: m_NUWMul(L: m_Value(V&: X), R: m_SpecificInt(V: *C2)), |
1219 | R: m_APInt(Res&: C1)))) { |
1220 | return BinaryOperator::CreateNUWAdd(V1: X, |
1221 | V2: ConstantInt::get(Ty, V: C1->udiv(RHS: *C2))); |
1222 | } |
1223 | |
1224 | if (!C2->isZero()) // avoid X udiv 0 |
1225 | if (Instruction *FoldedDiv = foldBinOpIntoSelectOrPhi(I)) |
1226 | return FoldedDiv; |
1227 | } |
1228 | |
1229 | if (match(V: Op0, P: m_One())) { |
1230 | assert(!Ty->isIntOrIntVectorTy(1) && "i1 divide not removed?" ); |
1231 | if (IsSigned) { |
1232 | // 1 / 0 --> undef ; 1 / 1 --> 1 ; 1 / -1 --> -1 ; 1 / anything else --> 0 |
1233 | // (Op1 + 1) u< 3 ? Op1 : 0 |
1234 | // Op1 must be frozen because we are increasing its number of uses. |
1235 | Value *F1 = Builder.CreateFreeze(V: Op1, Name: Op1->getName() + ".fr" ); |
1236 | Value *Inc = Builder.CreateAdd(LHS: F1, RHS: Op0); |
1237 | Value *Cmp = Builder.CreateICmpULT(LHS: Inc, RHS: ConstantInt::get(Ty, V: 3)); |
1238 | return SelectInst::Create(C: Cmp, S1: F1, S2: ConstantInt::get(Ty, V: 0)); |
1239 | } else { |
1240 | // If Op1 is 0 then it's undefined behaviour. If Op1 is 1 then the |
1241 | // result is one, otherwise it's zero. |
1242 | return new ZExtInst(Builder.CreateICmpEQ(LHS: Op1, RHS: Op0), Ty); |
1243 | } |
1244 | } |
1245 | |
1246 | // See if we can fold away this div instruction. |
1247 | if (SimplifyDemandedInstructionBits(Inst&: I)) |
1248 | return &I; |
1249 | |
1250 | // (X - (X rem Y)) / Y -> X / Y; usually originates as ((X / Y) * Y) / Y |
1251 | Value *X, *Z; |
1252 | if (match(V: Op0, P: m_Sub(L: m_Value(V&: X), R: m_Value(V&: Z)))) // (X - Z) / Y; Y = Op1 |
1253 | if ((IsSigned && match(V: Z, P: m_SRem(L: m_Specific(V: X), R: m_Specific(V: Op1)))) || |
1254 | (!IsSigned && match(V: Z, P: m_URem(L: m_Specific(V: X), R: m_Specific(V: Op1))))) |
1255 | return BinaryOperator::Create(Op: I.getOpcode(), S1: X, S2: Op1); |
1256 | |
1257 | // (X << Y) / X -> 1 << Y |
1258 | Value *Y; |
1259 | if (IsSigned && match(V: Op0, P: m_NSWShl(L: m_Specific(V: Op1), R: m_Value(V&: Y)))) |
1260 | return BinaryOperator::CreateNSWShl(V1: ConstantInt::get(Ty, V: 1), V2: Y); |
1261 | if (!IsSigned && match(V: Op0, P: m_NUWShl(L: m_Specific(V: Op1), R: m_Value(V&: Y)))) |
1262 | return BinaryOperator::CreateNUWShl(V1: ConstantInt::get(Ty, V: 1), V2: Y); |
1263 | |
1264 | // X / (X * Y) -> 1 / Y if the multiplication does not overflow. |
1265 | if (match(V: Op1, P: m_c_Mul(L: m_Specific(V: Op0), R: m_Value(V&: Y)))) { |
1266 | bool HasNSW = cast<OverflowingBinaryOperator>(Val: Op1)->hasNoSignedWrap(); |
1267 | bool HasNUW = cast<OverflowingBinaryOperator>(Val: Op1)->hasNoUnsignedWrap(); |
1268 | if ((IsSigned && HasNSW) || (!IsSigned && HasNUW)) { |
1269 | replaceOperand(I, OpNum: 0, V: ConstantInt::get(Ty, V: 1)); |
1270 | replaceOperand(I, OpNum: 1, V: Y); |
1271 | return &I; |
1272 | } |
1273 | } |
1274 | |
1275 | // (X << Z) / (X * Y) -> (1 << Z) / Y |
1276 | // TODO: Handle sdiv. |
1277 | if (!IsSigned && Op1->hasOneUse() && |
1278 | match(V: Op0, P: m_NUWShl(L: m_Value(V&: X), R: m_Value(V&: Z))) && |
1279 | match(V: Op1, P: m_c_Mul(L: m_Specific(V: X), R: m_Value(V&: Y)))) |
1280 | if (cast<OverflowingBinaryOperator>(Val: Op1)->hasNoUnsignedWrap()) { |
1281 | Instruction *NewDiv = BinaryOperator::CreateUDiv( |
1282 | V1: Builder.CreateShl(LHS: ConstantInt::get(Ty, V: 1), RHS: Z, Name: "" , /*NUW*/ HasNUW: true), V2: Y); |
1283 | NewDiv->setIsExact(I.isExact()); |
1284 | return NewDiv; |
1285 | } |
1286 | |
1287 | if (Value *R = foldIDivShl(I, Builder)) |
1288 | return replaceInstUsesWith(I, V: R); |
1289 | |
1290 | // With the appropriate no-wrap constraint, remove a multiply by the divisor |
1291 | // after peeking through another divide: |
1292 | // ((Op1 * X) / Y) / Op1 --> X / Y |
1293 | if (match(V: Op0, P: m_BinOp(Opcode: I.getOpcode(), L: m_c_Mul(L: m_Specific(V: Op1), R: m_Value(V&: X)), |
1294 | R: m_Value(V&: Y)))) { |
1295 | auto *InnerDiv = cast<PossiblyExactOperator>(Val: Op0); |
1296 | auto *Mul = cast<OverflowingBinaryOperator>(Val: InnerDiv->getOperand(i_nocapture: 0)); |
1297 | Instruction *NewDiv = nullptr; |
1298 | if (!IsSigned && Mul->hasNoUnsignedWrap()) |
1299 | NewDiv = BinaryOperator::CreateUDiv(V1: X, V2: Y); |
1300 | else if (IsSigned && Mul->hasNoSignedWrap()) |
1301 | NewDiv = BinaryOperator::CreateSDiv(V1: X, V2: Y); |
1302 | |
1303 | // Exact propagates only if both of the original divides are exact. |
1304 | if (NewDiv) { |
1305 | NewDiv->setIsExact(I.isExact() && InnerDiv->isExact()); |
1306 | return NewDiv; |
1307 | } |
1308 | } |
1309 | |
1310 | // (X * Y) / (X * Z) --> Y / Z (and commuted variants) |
1311 | if (match(V: Op0, P: m_Mul(L: m_Value(V&: X), R: m_Value(V&: Y)))) { |
1312 | auto OB0HasNSW = cast<OverflowingBinaryOperator>(Val: Op0)->hasNoSignedWrap(); |
1313 | auto OB0HasNUW = cast<OverflowingBinaryOperator>(Val: Op0)->hasNoUnsignedWrap(); |
1314 | |
1315 | auto CreateDivOrNull = [&](Value *A, Value *B) -> Instruction * { |
1316 | auto OB1HasNSW = cast<OverflowingBinaryOperator>(Val: Op1)->hasNoSignedWrap(); |
1317 | auto OB1HasNUW = |
1318 | cast<OverflowingBinaryOperator>(Val: Op1)->hasNoUnsignedWrap(); |
1319 | const APInt *C1, *C2; |
1320 | if (IsSigned && OB0HasNSW) { |
1321 | if (OB1HasNSW && match(V: B, P: m_APInt(Res&: C1)) && !C1->isAllOnes()) |
1322 | return BinaryOperator::CreateSDiv(V1: A, V2: B); |
1323 | } |
1324 | if (!IsSigned && OB0HasNUW) { |
1325 | if (OB1HasNUW) |
1326 | return BinaryOperator::CreateUDiv(V1: A, V2: B); |
1327 | if (match(V: A, P: m_APInt(Res&: C1)) && match(V: B, P: m_APInt(Res&: C2)) && C2->ule(RHS: *C1)) |
1328 | return BinaryOperator::CreateUDiv(V1: A, V2: B); |
1329 | } |
1330 | return nullptr; |
1331 | }; |
1332 | |
1333 | if (match(V: Op1, P: m_c_Mul(L: m_Specific(V: X), R: m_Value(V&: Z)))) { |
1334 | if (auto *Val = CreateDivOrNull(Y, Z)) |
1335 | return Val; |
1336 | } |
1337 | if (match(V: Op1, P: m_c_Mul(L: m_Specific(V: Y), R: m_Value(V&: Z)))) { |
1338 | if (auto *Val = CreateDivOrNull(X, Z)) |
1339 | return Val; |
1340 | } |
1341 | } |
1342 | return nullptr; |
1343 | } |
1344 | |
1345 | static const unsigned MaxDepth = 6; |
1346 | |
1347 | // Take the exact integer log2 of the value. If DoFold is true, create the |
1348 | // actual instructions, otherwise return a non-null dummy value. Return nullptr |
1349 | // on failure. |
1350 | static Value *takeLog2(IRBuilderBase &Builder, Value *Op, unsigned Depth, |
1351 | bool AssumeNonZero, bool DoFold) { |
1352 | auto IfFold = [DoFold](function_ref<Value *()> Fn) { |
1353 | if (!DoFold) |
1354 | return reinterpret_cast<Value *>(-1); |
1355 | return Fn(); |
1356 | }; |
1357 | |
1358 | // FIXME: assert that Op1 isn't/doesn't contain undef. |
1359 | |
1360 | // log2(2^C) -> C |
1361 | if (match(V: Op, P: m_Power2())) |
1362 | return IfFold([&]() { |
1363 | Constant *C = ConstantExpr::getExactLogBase2(C: cast<Constant>(Val: Op)); |
1364 | if (!C) |
1365 | llvm_unreachable("Failed to constant fold udiv -> logbase2" ); |
1366 | return C; |
1367 | }); |
1368 | |
1369 | // The remaining tests are all recursive, so bail out if we hit the limit. |
1370 | if (Depth++ == MaxDepth) |
1371 | return nullptr; |
1372 | |
1373 | // log2(zext X) -> zext log2(X) |
1374 | // FIXME: Require one use? |
1375 | Value *X, *Y; |
1376 | if (match(V: Op, P: m_ZExt(Op: m_Value(V&: X)))) |
1377 | if (Value *LogX = takeLog2(Builder, Op: X, Depth, AssumeNonZero, DoFold)) |
1378 | return IfFold([&]() { return Builder.CreateZExt(V: LogX, DestTy: Op->getType()); }); |
1379 | |
1380 | // log2(X << Y) -> log2(X) + Y |
1381 | // FIXME: Require one use unless X is 1? |
1382 | if (match(V: Op, P: m_Shl(L: m_Value(V&: X), R: m_Value(V&: Y)))) { |
1383 | auto *BO = cast<OverflowingBinaryOperator>(Val: Op); |
1384 | // nuw will be set if the `shl` is trivially non-zero. |
1385 | if (AssumeNonZero || BO->hasNoUnsignedWrap() || BO->hasNoSignedWrap()) |
1386 | if (Value *LogX = takeLog2(Builder, Op: X, Depth, AssumeNonZero, DoFold)) |
1387 | return IfFold([&]() { return Builder.CreateAdd(LHS: LogX, RHS: Y); }); |
1388 | } |
1389 | |
1390 | // log2(Cond ? X : Y) -> Cond ? log2(X) : log2(Y) |
1391 | // FIXME: Require one use? |
1392 | if (SelectInst *SI = dyn_cast<SelectInst>(Val: Op)) |
1393 | if (Value *LogX = takeLog2(Builder, Op: SI->getOperand(i_nocapture: 1), Depth, |
1394 | AssumeNonZero, DoFold)) |
1395 | if (Value *LogY = takeLog2(Builder, Op: SI->getOperand(i_nocapture: 2), Depth, |
1396 | AssumeNonZero, DoFold)) |
1397 | return IfFold([&]() { |
1398 | return Builder.CreateSelect(C: SI->getOperand(i_nocapture: 0), True: LogX, False: LogY); |
1399 | }); |
1400 | |
1401 | // log2(umin(X, Y)) -> umin(log2(X), log2(Y)) |
1402 | // log2(umax(X, Y)) -> umax(log2(X), log2(Y)) |
1403 | auto *MinMax = dyn_cast<MinMaxIntrinsic>(Val: Op); |
1404 | if (MinMax && MinMax->hasOneUse() && !MinMax->isSigned()) { |
1405 | // Use AssumeNonZero as false here. Otherwise we can hit case where |
1406 | // log2(umax(X, Y)) != umax(log2(X), log2(Y)) (because overflow). |
1407 | if (Value *LogX = takeLog2(Builder, Op: MinMax->getLHS(), Depth, |
1408 | /*AssumeNonZero*/ false, DoFold)) |
1409 | if (Value *LogY = takeLog2(Builder, Op: MinMax->getRHS(), Depth, |
1410 | /*AssumeNonZero*/ false, DoFold)) |
1411 | return IfFold([&]() { |
1412 | return Builder.CreateBinaryIntrinsic(ID: MinMax->getIntrinsicID(), LHS: LogX, |
1413 | RHS: LogY); |
1414 | }); |
1415 | } |
1416 | |
1417 | return nullptr; |
1418 | } |
1419 | |
1420 | /// If we have zero-extended operands of an unsigned div or rem, we may be able |
1421 | /// to narrow the operation (sink the zext below the math). |
1422 | static Instruction *narrowUDivURem(BinaryOperator &I, |
1423 | InstCombinerImpl &IC) { |
1424 | Instruction::BinaryOps Opcode = I.getOpcode(); |
1425 | Value *N = I.getOperand(i_nocapture: 0); |
1426 | Value *D = I.getOperand(i_nocapture: 1); |
1427 | Type *Ty = I.getType(); |
1428 | Value *X, *Y; |
1429 | if (match(V: N, P: m_ZExt(Op: m_Value(V&: X))) && match(V: D, P: m_ZExt(Op: m_Value(V&: Y))) && |
1430 | X->getType() == Y->getType() && (N->hasOneUse() || D->hasOneUse())) { |
1431 | // udiv (zext X), (zext Y) --> zext (udiv X, Y) |
1432 | // urem (zext X), (zext Y) --> zext (urem X, Y) |
1433 | Value *NarrowOp = IC.Builder.CreateBinOp(Opc: Opcode, LHS: X, RHS: Y); |
1434 | return new ZExtInst(NarrowOp, Ty); |
1435 | } |
1436 | |
1437 | Constant *C; |
1438 | if (isa<Instruction>(Val: N) && match(V: N, P: m_OneUse(SubPattern: m_ZExt(Op: m_Value(V&: X)))) && |
1439 | match(V: D, P: m_Constant(C))) { |
1440 | // If the constant is the same in the smaller type, use the narrow version. |
1441 | Constant *TruncC = IC.getLosslessUnsignedTrunc(C, TruncTy: X->getType()); |
1442 | if (!TruncC) |
1443 | return nullptr; |
1444 | |
1445 | // udiv (zext X), C --> zext (udiv X, C') |
1446 | // urem (zext X), C --> zext (urem X, C') |
1447 | return new ZExtInst(IC.Builder.CreateBinOp(Opc: Opcode, LHS: X, RHS: TruncC), Ty); |
1448 | } |
1449 | if (isa<Instruction>(Val: D) && match(V: D, P: m_OneUse(SubPattern: m_ZExt(Op: m_Value(V&: X)))) && |
1450 | match(V: N, P: m_Constant(C))) { |
1451 | // If the constant is the same in the smaller type, use the narrow version. |
1452 | Constant *TruncC = IC.getLosslessUnsignedTrunc(C, TruncTy: X->getType()); |
1453 | if (!TruncC) |
1454 | return nullptr; |
1455 | |
1456 | // udiv C, (zext X) --> zext (udiv C', X) |
1457 | // urem C, (zext X) --> zext (urem C', X) |
1458 | return new ZExtInst(IC.Builder.CreateBinOp(Opc: Opcode, LHS: TruncC, RHS: X), Ty); |
1459 | } |
1460 | |
1461 | return nullptr; |
1462 | } |
1463 | |
1464 | Instruction *InstCombinerImpl::visitUDiv(BinaryOperator &I) { |
1465 | if (Value *V = simplifyUDivInst(LHS: I.getOperand(i_nocapture: 0), RHS: I.getOperand(i_nocapture: 1), IsExact: I.isExact(), |
1466 | Q: SQ.getWithInstruction(I: &I))) |
1467 | return replaceInstUsesWith(I, V); |
1468 | |
1469 | if (Instruction *X = foldVectorBinop(Inst&: I)) |
1470 | return X; |
1471 | |
1472 | // Handle the integer div common cases |
1473 | if (Instruction *Common = commonIDivTransforms(I)) |
1474 | return Common; |
1475 | |
1476 | Value *Op0 = I.getOperand(i_nocapture: 0), *Op1 = I.getOperand(i_nocapture: 1); |
1477 | Value *X; |
1478 | const APInt *C1, *C2; |
1479 | if (match(V: Op0, P: m_LShr(L: m_Value(V&: X), R: m_APInt(Res&: C1))) && match(V: Op1, P: m_APInt(Res&: C2))) { |
1480 | // (X lshr C1) udiv C2 --> X udiv (C2 << C1) |
1481 | bool Overflow; |
1482 | APInt C2ShlC1 = C2->ushl_ov(Amt: *C1, Overflow); |
1483 | if (!Overflow) { |
1484 | bool IsExact = I.isExact() && match(V: Op0, P: m_Exact(SubPattern: m_Value())); |
1485 | BinaryOperator *BO = BinaryOperator::CreateUDiv( |
1486 | V1: X, V2: ConstantInt::get(Ty: X->getType(), V: C2ShlC1)); |
1487 | if (IsExact) |
1488 | BO->setIsExact(); |
1489 | return BO; |
1490 | } |
1491 | } |
1492 | |
1493 | // Op0 / C where C is large (negative) --> zext (Op0 >= C) |
1494 | // TODO: Could use isKnownNegative() to handle non-constant values. |
1495 | Type *Ty = I.getType(); |
1496 | if (match(V: Op1, P: m_Negative())) { |
1497 | Value *Cmp = Builder.CreateICmpUGE(LHS: Op0, RHS: Op1); |
1498 | return CastInst::CreateZExtOrBitCast(S: Cmp, Ty); |
1499 | } |
1500 | // Op0 / (sext i1 X) --> zext (Op0 == -1) (if X is 0, the div is undefined) |
1501 | if (match(V: Op1, P: m_SExt(Op: m_Value(V&: X))) && X->getType()->isIntOrIntVectorTy(BitWidth: 1)) { |
1502 | Value *Cmp = Builder.CreateICmpEQ(LHS: Op0, RHS: ConstantInt::getAllOnesValue(Ty)); |
1503 | return CastInst::CreateZExtOrBitCast(S: Cmp, Ty); |
1504 | } |
1505 | |
1506 | if (Instruction *NarrowDiv = narrowUDivURem(I, IC&: *this)) |
1507 | return NarrowDiv; |
1508 | |
1509 | Value *A, *B; |
1510 | |
1511 | // Look through a right-shift to find the common factor: |
1512 | // ((Op1 *nuw A) >> B) / Op1 --> A >> B |
1513 | if (match(V: Op0, P: m_LShr(L: m_NUWMul(L: m_Specific(V: Op1), R: m_Value(V&: A)), R: m_Value(V&: B))) || |
1514 | match(V: Op0, P: m_LShr(L: m_NUWMul(L: m_Value(V&: A), R: m_Specific(V: Op1)), R: m_Value(V&: B)))) { |
1515 | Instruction *Lshr = BinaryOperator::CreateLShr(V1: A, V2: B); |
1516 | if (I.isExact() && cast<PossiblyExactOperator>(Val: Op0)->isExact()) |
1517 | Lshr->setIsExact(); |
1518 | return Lshr; |
1519 | } |
1520 | |
1521 | // Op1 udiv Op2 -> Op1 lshr log2(Op2), if log2() folds away. |
1522 | if (takeLog2(Builder, Op: Op1, /*Depth*/ 0, /*AssumeNonZero*/ true, |
1523 | /*DoFold*/ false)) { |
1524 | Value *Res = takeLog2(Builder, Op: Op1, /*Depth*/ 0, |
1525 | /*AssumeNonZero*/ true, /*DoFold*/ true); |
1526 | return replaceInstUsesWith( |
1527 | I, V: Builder.CreateLShr(LHS: Op0, RHS: Res, Name: I.getName(), isExact: I.isExact())); |
1528 | } |
1529 | |
1530 | return nullptr; |
1531 | } |
1532 | |
1533 | Instruction *InstCombinerImpl::visitSDiv(BinaryOperator &I) { |
1534 | if (Value *V = simplifySDivInst(LHS: I.getOperand(i_nocapture: 0), RHS: I.getOperand(i_nocapture: 1), IsExact: I.isExact(), |
1535 | Q: SQ.getWithInstruction(I: &I))) |
1536 | return replaceInstUsesWith(I, V); |
1537 | |
1538 | if (Instruction *X = foldVectorBinop(Inst&: I)) |
1539 | return X; |
1540 | |
1541 | // Handle the integer div common cases |
1542 | if (Instruction *Common = commonIDivTransforms(I)) |
1543 | return Common; |
1544 | |
1545 | Value *Op0 = I.getOperand(i_nocapture: 0), *Op1 = I.getOperand(i_nocapture: 1); |
1546 | Type *Ty = I.getType(); |
1547 | Value *X; |
1548 | // sdiv Op0, -1 --> -Op0 |
1549 | // sdiv Op0, (sext i1 X) --> -Op0 (because if X is 0, the op is undefined) |
1550 | if (match(V: Op1, P: m_AllOnes()) || |
1551 | (match(V: Op1, P: m_SExt(Op: m_Value(V&: X))) && X->getType()->isIntOrIntVectorTy(BitWidth: 1))) |
1552 | return BinaryOperator::CreateNSWNeg(Op: Op0); |
1553 | |
1554 | // X / INT_MIN --> X == INT_MIN |
1555 | if (match(V: Op1, P: m_SignMask())) |
1556 | return new ZExtInst(Builder.CreateICmpEQ(LHS: Op0, RHS: Op1), Ty); |
1557 | |
1558 | if (I.isExact()) { |
1559 | // sdiv exact X, 1<<C --> ashr exact X, C iff 1<<C is non-negative |
1560 | if (match(V: Op1, P: m_Power2()) && match(V: Op1, P: m_NonNegative())) { |
1561 | Constant *C = ConstantExpr::getExactLogBase2(C: cast<Constant>(Val: Op1)); |
1562 | return BinaryOperator::CreateExactAShr(V1: Op0, V2: C); |
1563 | } |
1564 | |
1565 | // sdiv exact X, (1<<ShAmt) --> ashr exact X, ShAmt (if shl is non-negative) |
1566 | Value *ShAmt; |
1567 | if (match(V: Op1, P: m_NSWShl(L: m_One(), R: m_Value(V&: ShAmt)))) |
1568 | return BinaryOperator::CreateExactAShr(V1: Op0, V2: ShAmt); |
1569 | |
1570 | // sdiv exact X, -1<<C --> -(ashr exact X, C) |
1571 | if (match(V: Op1, P: m_NegatedPower2())) { |
1572 | Constant *NegPow2C = ConstantExpr::getNeg(C: cast<Constant>(Val: Op1)); |
1573 | Constant *C = ConstantExpr::getExactLogBase2(C: NegPow2C); |
1574 | Value *Ashr = Builder.CreateAShr(LHS: Op0, RHS: C, Name: I.getName() + ".neg" , isExact: true); |
1575 | return BinaryOperator::CreateNSWNeg(Op: Ashr); |
1576 | } |
1577 | } |
1578 | |
1579 | const APInt *Op1C; |
1580 | if (match(V: Op1, P: m_APInt(Res&: Op1C))) { |
1581 | // If the dividend is sign-extended and the constant divisor is small enough |
1582 | // to fit in the source type, shrink the division to the narrower type: |
1583 | // (sext X) sdiv C --> sext (X sdiv C) |
1584 | Value *Op0Src; |
1585 | if (match(V: Op0, P: m_OneUse(SubPattern: m_SExt(Op: m_Value(V&: Op0Src)))) && |
1586 | Op0Src->getType()->getScalarSizeInBits() >= |
1587 | Op1C->getSignificantBits()) { |
1588 | |
1589 | // In the general case, we need to make sure that the dividend is not the |
1590 | // minimum signed value because dividing that by -1 is UB. But here, we |
1591 | // know that the -1 divisor case is already handled above. |
1592 | |
1593 | Constant *NarrowDivisor = |
1594 | ConstantExpr::getTrunc(C: cast<Constant>(Val: Op1), Ty: Op0Src->getType()); |
1595 | Value *NarrowOp = Builder.CreateSDiv(LHS: Op0Src, RHS: NarrowDivisor); |
1596 | return new SExtInst(NarrowOp, Ty); |
1597 | } |
1598 | |
1599 | // -X / C --> X / -C (if the negation doesn't overflow). |
1600 | // TODO: This could be enhanced to handle arbitrary vector constants by |
1601 | // checking if all elements are not the min-signed-val. |
1602 | if (!Op1C->isMinSignedValue() && match(V: Op0, P: m_NSWNeg(V: m_Value(V&: X)))) { |
1603 | Constant *NegC = ConstantInt::get(Ty, V: -(*Op1C)); |
1604 | Instruction *BO = BinaryOperator::CreateSDiv(V1: X, V2: NegC); |
1605 | BO->setIsExact(I.isExact()); |
1606 | return BO; |
1607 | } |
1608 | } |
1609 | |
1610 | // -X / Y --> -(X / Y) |
1611 | Value *Y; |
1612 | if (match(V: &I, P: m_SDiv(L: m_OneUse(SubPattern: m_NSWNeg(V: m_Value(V&: X))), R: m_Value(V&: Y)))) |
1613 | return BinaryOperator::CreateNSWNeg( |
1614 | Op: Builder.CreateSDiv(LHS: X, RHS: Y, Name: I.getName(), isExact: I.isExact())); |
1615 | |
1616 | // abs(X) / X --> X > -1 ? 1 : -1 |
1617 | // X / abs(X) --> X > -1 ? 1 : -1 |
1618 | if (match(&I, m_c_BinOp( |
1619 | m_OneUse(m_Intrinsic<Intrinsic::abs>(m_Value(V&: X), m_One())), |
1620 | m_Deferred(V: X)))) { |
1621 | Value *Cond = Builder.CreateIsNotNeg(Arg: X); |
1622 | return SelectInst::Create(C: Cond, S1: ConstantInt::get(Ty, V: 1), |
1623 | S2: ConstantInt::getAllOnesValue(Ty)); |
1624 | } |
1625 | |
1626 | KnownBits KnownDividend = computeKnownBits(V: Op0, Depth: 0, CxtI: &I); |
1627 | if (!I.isExact() && |
1628 | (match(V: Op1, P: m_Power2(V&: Op1C)) || match(V: Op1, P: m_NegatedPower2(V&: Op1C))) && |
1629 | KnownDividend.countMinTrailingZeros() >= Op1C->countr_zero()) { |
1630 | I.setIsExact(); |
1631 | return &I; |
1632 | } |
1633 | |
1634 | if (KnownDividend.isNonNegative()) { |
1635 | // If both operands are unsigned, turn this into a udiv. |
1636 | if (isKnownNonNegative(V: Op1, SQ: SQ.getWithInstruction(I: &I))) { |
1637 | auto *BO = BinaryOperator::CreateUDiv(V1: Op0, V2: Op1, Name: I.getName()); |
1638 | BO->setIsExact(I.isExact()); |
1639 | return BO; |
1640 | } |
1641 | |
1642 | if (match(V: Op1, P: m_NegatedPower2())) { |
1643 | // X sdiv (-(1 << C)) -> -(X sdiv (1 << C)) -> |
1644 | // -> -(X udiv (1 << C)) -> -(X u>> C) |
1645 | Constant *CNegLog2 = ConstantExpr::getExactLogBase2( |
1646 | C: ConstantExpr::getNeg(C: cast<Constant>(Val: Op1))); |
1647 | Value *Shr = Builder.CreateLShr(LHS: Op0, RHS: CNegLog2, Name: I.getName(), isExact: I.isExact()); |
1648 | return BinaryOperator::CreateNeg(Op: Shr); |
1649 | } |
1650 | |
1651 | if (isKnownToBeAPowerOfTwo(V: Op1, /*OrZero*/ true, Depth: 0, CxtI: &I)) { |
1652 | // X sdiv (1 << Y) -> X udiv (1 << Y) ( -> X u>> Y) |
1653 | // Safe because the only negative value (1 << Y) can take on is |
1654 | // INT_MIN, and X sdiv INT_MIN == X udiv INT_MIN == 0 if X doesn't have |
1655 | // the sign bit set. |
1656 | auto *BO = BinaryOperator::CreateUDiv(V1: Op0, V2: Op1, Name: I.getName()); |
1657 | BO->setIsExact(I.isExact()); |
1658 | return BO; |
1659 | } |
1660 | } |
1661 | |
1662 | // -X / X --> X == INT_MIN ? 1 : -1 |
1663 | if (isKnownNegation(X: Op0, Y: Op1)) { |
1664 | APInt MinVal = APInt::getSignedMinValue(numBits: Ty->getScalarSizeInBits()); |
1665 | Value *Cond = Builder.CreateICmpEQ(LHS: Op0, RHS: ConstantInt::get(Ty, V: MinVal)); |
1666 | return SelectInst::Create(C: Cond, S1: ConstantInt::get(Ty, V: 1), |
1667 | S2: ConstantInt::getAllOnesValue(Ty)); |
1668 | } |
1669 | return nullptr; |
1670 | } |
1671 | |
1672 | /// Remove negation and try to convert division into multiplication. |
1673 | Instruction *InstCombinerImpl::foldFDivConstantDivisor(BinaryOperator &I) { |
1674 | Constant *C; |
1675 | if (!match(V: I.getOperand(i_nocapture: 1), P: m_Constant(C))) |
1676 | return nullptr; |
1677 | |
1678 | // -X / C --> X / -C |
1679 | Value *X; |
1680 | const DataLayout &DL = I.getModule()->getDataLayout(); |
1681 | if (match(V: I.getOperand(i_nocapture: 0), P: m_FNeg(X: m_Value(V&: X)))) |
1682 | if (Constant *NegC = ConstantFoldUnaryOpOperand(Opcode: Instruction::FNeg, Op: C, DL)) |
1683 | return BinaryOperator::CreateFDivFMF(V1: X, V2: NegC, FMFSource: &I); |
1684 | |
1685 | // nnan X / +0.0 -> copysign(inf, X) |
1686 | // nnan nsz X / -0.0 -> copysign(inf, X) |
1687 | if (I.hasNoNaNs() && |
1688 | (match(V: I.getOperand(i_nocapture: 1), P: m_PosZeroFP()) || |
1689 | (I.hasNoSignedZeros() && match(V: I.getOperand(i_nocapture: 1), P: m_AnyZeroFP())))) { |
1690 | IRBuilder<> B(&I); |
1691 | CallInst *CopySign = B.CreateIntrinsic( |
1692 | Intrinsic::copysign, {C->getType()}, |
1693 | {ConstantFP::getInfinity(Ty: I.getType()), I.getOperand(i_nocapture: 0)}, &I); |
1694 | CopySign->takeName(V: &I); |
1695 | return replaceInstUsesWith(I, V: CopySign); |
1696 | } |
1697 | |
1698 | // If the constant divisor has an exact inverse, this is always safe. If not, |
1699 | // then we can still create a reciprocal if fast-math-flags allow it and the |
1700 | // constant is a regular number (not zero, infinite, or denormal). |
1701 | if (!(C->hasExactInverseFP() || (I.hasAllowReciprocal() && C->isNormalFP()))) |
1702 | return nullptr; |
1703 | |
1704 | // Disallow denormal constants because we don't know what would happen |
1705 | // on all targets. |
1706 | // TODO: Use Intrinsic::canonicalize or let function attributes tell us that |
1707 | // denorms are flushed? |
1708 | auto *RecipC = ConstantFoldBinaryOpOperands( |
1709 | Opcode: Instruction::FDiv, LHS: ConstantFP::get(Ty: I.getType(), V: 1.0), RHS: C, DL); |
1710 | if (!RecipC || !RecipC->isNormalFP()) |
1711 | return nullptr; |
1712 | |
1713 | // X / C --> X * (1 / C) |
1714 | return BinaryOperator::CreateFMulFMF(V1: I.getOperand(i_nocapture: 0), V2: RecipC, FMFSource: &I); |
1715 | } |
1716 | |
1717 | /// Remove negation and try to reassociate constant math. |
1718 | static Instruction *foldFDivConstantDividend(BinaryOperator &I) { |
1719 | Constant *C; |
1720 | if (!match(V: I.getOperand(i_nocapture: 0), P: m_Constant(C))) |
1721 | return nullptr; |
1722 | |
1723 | // C / -X --> -C / X |
1724 | Value *X; |
1725 | const DataLayout &DL = I.getModule()->getDataLayout(); |
1726 | if (match(V: I.getOperand(i_nocapture: 1), P: m_FNeg(X: m_Value(V&: X)))) |
1727 | if (Constant *NegC = ConstantFoldUnaryOpOperand(Opcode: Instruction::FNeg, Op: C, DL)) |
1728 | return BinaryOperator::CreateFDivFMF(V1: NegC, V2: X, FMFSource: &I); |
1729 | |
1730 | if (!I.hasAllowReassoc() || !I.hasAllowReciprocal()) |
1731 | return nullptr; |
1732 | |
1733 | // Try to reassociate C / X expressions where X includes another constant. |
1734 | Constant *C2, *NewC = nullptr; |
1735 | if (match(V: I.getOperand(i_nocapture: 1), P: m_FMul(L: m_Value(V&: X), R: m_Constant(C&: C2)))) { |
1736 | // C / (X * C2) --> (C / C2) / X |
1737 | NewC = ConstantFoldBinaryOpOperands(Opcode: Instruction::FDiv, LHS: C, RHS: C2, DL); |
1738 | } else if (match(V: I.getOperand(i_nocapture: 1), P: m_FDiv(L: m_Value(V&: X), R: m_Constant(C&: C2)))) { |
1739 | // C / (X / C2) --> (C * C2) / X |
1740 | NewC = ConstantFoldBinaryOpOperands(Opcode: Instruction::FMul, LHS: C, RHS: C2, DL); |
1741 | } |
1742 | // Disallow denormal constants because we don't know what would happen |
1743 | // on all targets. |
1744 | // TODO: Use Intrinsic::canonicalize or let function attributes tell us that |
1745 | // denorms are flushed? |
1746 | if (!NewC || !NewC->isNormalFP()) |
1747 | return nullptr; |
1748 | |
1749 | return BinaryOperator::CreateFDivFMF(V1: NewC, V2: X, FMFSource: &I); |
1750 | } |
1751 | |
1752 | /// Negate the exponent of pow/exp to fold division-by-pow() into multiply. |
1753 | static Instruction *foldFDivPowDivisor(BinaryOperator &I, |
1754 | InstCombiner::BuilderTy &Builder) { |
1755 | Value *Op0 = I.getOperand(i_nocapture: 0), *Op1 = I.getOperand(i_nocapture: 1); |
1756 | auto *II = dyn_cast<IntrinsicInst>(Val: Op1); |
1757 | if (!II || !II->hasOneUse() || !I.hasAllowReassoc() || |
1758 | !I.hasAllowReciprocal()) |
1759 | return nullptr; |
1760 | |
1761 | // Z / pow(X, Y) --> Z * pow(X, -Y) |
1762 | // Z / exp{2}(Y) --> Z * exp{2}(-Y) |
1763 | // In the general case, this creates an extra instruction, but fmul allows |
1764 | // for better canonicalization and optimization than fdiv. |
1765 | Intrinsic::ID IID = II->getIntrinsicID(); |
1766 | SmallVector<Value *> Args; |
1767 | switch (IID) { |
1768 | case Intrinsic::pow: |
1769 | Args.push_back(Elt: II->getArgOperand(i: 0)); |
1770 | Args.push_back(Elt: Builder.CreateFNegFMF(V: II->getArgOperand(i: 1), FMFSource: &I)); |
1771 | break; |
1772 | case Intrinsic::powi: { |
1773 | // Require 'ninf' assuming that makes powi(X, -INT_MIN) acceptable. |
1774 | // That is, X ** (huge negative number) is 0.0, ~1.0, or INF and so |
1775 | // dividing by that is INF, ~1.0, or 0.0. Code that uses powi allows |
1776 | // non-standard results, so this corner case should be acceptable if the |
1777 | // code rules out INF values. |
1778 | if (!I.hasNoInfs()) |
1779 | return nullptr; |
1780 | Args.push_back(Elt: II->getArgOperand(i: 0)); |
1781 | Args.push_back(Elt: Builder.CreateNeg(V: II->getArgOperand(i: 1))); |
1782 | Type *Tys[] = {I.getType(), II->getArgOperand(i: 1)->getType()}; |
1783 | Value *Pow = Builder.CreateIntrinsic(ID: IID, Types: Tys, Args, FMFSource: &I); |
1784 | return BinaryOperator::CreateFMulFMF(V1: Op0, V2: Pow, FMFSource: &I); |
1785 | } |
1786 | case Intrinsic::exp: |
1787 | case Intrinsic::exp2: |
1788 | Args.push_back(Elt: Builder.CreateFNegFMF(V: II->getArgOperand(i: 0), FMFSource: &I)); |
1789 | break; |
1790 | default: |
1791 | return nullptr; |
1792 | } |
1793 | Value *Pow = Builder.CreateIntrinsic(ID: IID, Types: I.getType(), Args, FMFSource: &I); |
1794 | return BinaryOperator::CreateFMulFMF(V1: Op0, V2: Pow, FMFSource: &I); |
1795 | } |
1796 | |
1797 | /// Convert div to mul if we have an sqrt divisor iff sqrt's operand is a fdiv |
1798 | /// instruction. |
1799 | static Instruction *foldFDivSqrtDivisor(BinaryOperator &I, |
1800 | InstCombiner::BuilderTy &Builder) { |
1801 | // X / sqrt(Y / Z) --> X * sqrt(Z / Y) |
1802 | if (!I.hasAllowReassoc() || !I.hasAllowReciprocal()) |
1803 | return nullptr; |
1804 | Value *Op0 = I.getOperand(i_nocapture: 0), *Op1 = I.getOperand(i_nocapture: 1); |
1805 | auto *II = dyn_cast<IntrinsicInst>(Val: Op1); |
1806 | if (!II || II->getIntrinsicID() != Intrinsic::sqrt || !II->hasOneUse() || |
1807 | !II->hasAllowReassoc() || !II->hasAllowReciprocal()) |
1808 | return nullptr; |
1809 | |
1810 | Value *Y, *Z; |
1811 | auto *DivOp = dyn_cast<Instruction>(Val: II->getOperand(i_nocapture: 0)); |
1812 | if (!DivOp) |
1813 | return nullptr; |
1814 | if (!match(V: DivOp, P: m_FDiv(L: m_Value(V&: Y), R: m_Value(V&: Z)))) |
1815 | return nullptr; |
1816 | if (!DivOp->hasAllowReassoc() || !I.hasAllowReciprocal() || |
1817 | !DivOp->hasOneUse()) |
1818 | return nullptr; |
1819 | Value *SwapDiv = Builder.CreateFDivFMF(L: Z, R: Y, FMFSource: DivOp); |
1820 | Value *NewSqrt = |
1821 | Builder.CreateUnaryIntrinsic(ID: II->getIntrinsicID(), V: SwapDiv, FMFSource: II); |
1822 | return BinaryOperator::CreateFMulFMF(V1: Op0, V2: NewSqrt, FMFSource: &I); |
1823 | } |
1824 | |
1825 | Instruction *InstCombinerImpl::visitFDiv(BinaryOperator &I) { |
1826 | Module *M = I.getModule(); |
1827 | |
1828 | if (Value *V = simplifyFDivInst(LHS: I.getOperand(i_nocapture: 0), RHS: I.getOperand(i_nocapture: 1), |
1829 | FMF: I.getFastMathFlags(), |
1830 | Q: SQ.getWithInstruction(I: &I))) |
1831 | return replaceInstUsesWith(I, V); |
1832 | |
1833 | if (Instruction *X = foldVectorBinop(Inst&: I)) |
1834 | return X; |
1835 | |
1836 | if (Instruction *Phi = foldBinopWithPhiOperands(BO&: I)) |
1837 | return Phi; |
1838 | |
1839 | if (Instruction *R = foldFDivConstantDivisor(I)) |
1840 | return R; |
1841 | |
1842 | if (Instruction *R = foldFDivConstantDividend(I)) |
1843 | return R; |
1844 | |
1845 | if (Instruction *R = foldFPSignBitOps(I)) |
1846 | return R; |
1847 | |
1848 | Value *Op0 = I.getOperand(i_nocapture: 0), *Op1 = I.getOperand(i_nocapture: 1); |
1849 | if (isa<Constant>(Val: Op0)) |
1850 | if (SelectInst *SI = dyn_cast<SelectInst>(Val: Op1)) |
1851 | if (Instruction *R = FoldOpIntoSelect(Op&: I, SI)) |
1852 | return R; |
1853 | |
1854 | if (isa<Constant>(Val: Op1)) |
1855 | if (SelectInst *SI = dyn_cast<SelectInst>(Val: Op0)) |
1856 | if (Instruction *R = FoldOpIntoSelect(Op&: I, SI)) |
1857 | return R; |
1858 | |
1859 | if (I.hasAllowReassoc() && I.hasAllowReciprocal()) { |
1860 | Value *X, *Y; |
1861 | if (match(V: Op0, P: m_OneUse(SubPattern: m_FDiv(L: m_Value(V&: X), R: m_Value(V&: Y)))) && |
1862 | (!isa<Constant>(Val: Y) || !isa<Constant>(Val: Op1))) { |
1863 | // (X / Y) / Z => X / (Y * Z) |
1864 | Value *YZ = Builder.CreateFMulFMF(L: Y, R: Op1, FMFSource: &I); |
1865 | return BinaryOperator::CreateFDivFMF(V1: X, V2: YZ, FMFSource: &I); |
1866 | } |
1867 | if (match(V: Op1, P: m_OneUse(SubPattern: m_FDiv(L: m_Value(V&: X), R: m_Value(V&: Y)))) && |
1868 | (!isa<Constant>(Val: Y) || !isa<Constant>(Val: Op0))) { |
1869 | // Z / (X / Y) => (Y * Z) / X |
1870 | Value *YZ = Builder.CreateFMulFMF(L: Y, R: Op0, FMFSource: &I); |
1871 | return BinaryOperator::CreateFDivFMF(V1: YZ, V2: X, FMFSource: &I); |
1872 | } |
1873 | // Z / (1.0 / Y) => (Y * Z) |
1874 | // |
1875 | // This is a special case of Z / (X / Y) => (Y * Z) / X, with X = 1.0. The |
1876 | // m_OneUse check is avoided because even in the case of the multiple uses |
1877 | // for 1.0/Y, the number of instructions remain the same and a division is |
1878 | // replaced by a multiplication. |
1879 | if (match(V: Op1, P: m_FDiv(L: m_SpecificFP(V: 1.0), R: m_Value(V&: Y)))) |
1880 | return BinaryOperator::CreateFMulFMF(V1: Y, V2: Op0, FMFSource: &I); |
1881 | } |
1882 | |
1883 | if (I.hasAllowReassoc() && Op0->hasOneUse() && Op1->hasOneUse()) { |
1884 | // sin(X) / cos(X) -> tan(X) |
1885 | // cos(X) / sin(X) -> 1/tan(X) (cotangent) |
1886 | Value *X; |
1887 | bool IsTan = match(Op0, m_Intrinsic<Intrinsic::sin>(m_Value(X))) && |
1888 | match(Op1, m_Intrinsic<Intrinsic::cos>(m_Specific(X))); |
1889 | bool IsCot = |
1890 | !IsTan && match(Op0, m_Intrinsic<Intrinsic::cos>(m_Value(X))) && |
1891 | match(Op1, m_Intrinsic<Intrinsic::sin>(m_Specific(X))); |
1892 | |
1893 | if ((IsTan || IsCot) && hasFloatFn(M, TLI: &TLI, Ty: I.getType(), DoubleFn: LibFunc_tan, |
1894 | FloatFn: LibFunc_tanf, LongDoubleFn: LibFunc_tanl)) { |
1895 | IRBuilder<> B(&I); |
1896 | IRBuilder<>::FastMathFlagGuard FMFGuard(B); |
1897 | B.setFastMathFlags(I.getFastMathFlags()); |
1898 | AttributeList Attrs = |
1899 | cast<CallBase>(Val: Op0)->getCalledFunction()->getAttributes(); |
1900 | Value *Res = emitUnaryFloatFnCall(Op: X, TLI: &TLI, DoubleFn: LibFunc_tan, FloatFn: LibFunc_tanf, |
1901 | LongDoubleFn: LibFunc_tanl, B, Attrs); |
1902 | if (IsCot) |
1903 | Res = B.CreateFDiv(L: ConstantFP::get(Ty: I.getType(), V: 1.0), R: Res); |
1904 | return replaceInstUsesWith(I, V: Res); |
1905 | } |
1906 | } |
1907 | |
1908 | // X / (X * Y) --> 1.0 / Y |
1909 | // Reassociate to (X / X -> 1.0) is legal when NaNs are not allowed. |
1910 | // We can ignore the possibility that X is infinity because INF/INF is NaN. |
1911 | Value *X, *Y; |
1912 | if (I.hasNoNaNs() && I.hasAllowReassoc() && |
1913 | match(V: Op1, P: m_c_FMul(L: m_Specific(V: Op0), R: m_Value(V&: Y)))) { |
1914 | replaceOperand(I, OpNum: 0, V: ConstantFP::get(Ty: I.getType(), V: 1.0)); |
1915 | replaceOperand(I, OpNum: 1, V: Y); |
1916 | return &I; |
1917 | } |
1918 | |
1919 | // X / fabs(X) -> copysign(1.0, X) |
1920 | // fabs(X) / X -> copysign(1.0, X) |
1921 | if (I.hasNoNaNs() && I.hasNoInfs() && |
1922 | (match(V: &I, P: m_FDiv(L: m_Value(V&: X), R: m_FAbs(Op0: m_Deferred(V: X)))) || |
1923 | match(V: &I, P: m_FDiv(L: m_FAbs(Op0: m_Value(V&: X)), R: m_Deferred(V: X))))) { |
1924 | Value *V = Builder.CreateBinaryIntrinsic( |
1925 | Intrinsic::copysign, ConstantFP::get(I.getType(), 1.0), X, &I); |
1926 | return replaceInstUsesWith(I, V); |
1927 | } |
1928 | |
1929 | if (Instruction *Mul = foldFDivPowDivisor(I, Builder)) |
1930 | return Mul; |
1931 | |
1932 | if (Instruction *Mul = foldFDivSqrtDivisor(I, Builder)) |
1933 | return Mul; |
1934 | |
1935 | // pow(X, Y) / X --> pow(X, Y-1) |
1936 | if (I.hasAllowReassoc() && |
1937 | match(Op0, m_OneUse(m_Intrinsic<Intrinsic::pow>(m_Specific(Op1), |
1938 | m_Value(Y))))) { |
1939 | Value *Y1 = |
1940 | Builder.CreateFAddFMF(L: Y, R: ConstantFP::get(Ty: I.getType(), V: -1.0), FMFSource: &I); |
1941 | Value *Pow = Builder.CreateBinaryIntrinsic(Intrinsic::pow, Op1, Y1, &I); |
1942 | return replaceInstUsesWith(I, V: Pow); |
1943 | } |
1944 | |
1945 | if (Instruction *FoldedPowi = foldPowiReassoc(I)) |
1946 | return FoldedPowi; |
1947 | |
1948 | return nullptr; |
1949 | } |
1950 | |
1951 | // Variety of transform for: |
1952 | // (urem/srem (mul X, Y), (mul X, Z)) |
1953 | // (urem/srem (shl X, Y), (shl X, Z)) |
1954 | // (urem/srem (shl Y, X), (shl Z, X)) |
1955 | // NB: The shift cases are really just extensions of the mul case. We treat |
1956 | // shift as Val * (1 << Amt). |
1957 | static Instruction *simplifyIRemMulShl(BinaryOperator &I, |
1958 | InstCombinerImpl &IC) { |
1959 | Value *Op0 = I.getOperand(i_nocapture: 0), *Op1 = I.getOperand(i_nocapture: 1), *X = nullptr; |
1960 | APInt Y, Z; |
1961 | bool ShiftByX = false; |
1962 | |
1963 | // If V is not nullptr, it will be matched using m_Specific. |
1964 | auto MatchShiftOrMulXC = [](Value *Op, Value *&V, APInt &C) -> bool { |
1965 | const APInt *Tmp = nullptr; |
1966 | if ((!V && match(V: Op, P: m_Mul(L: m_Value(V), R: m_APInt(Res&: Tmp)))) || |
1967 | (V && match(V: Op, P: m_Mul(L: m_Specific(V), R: m_APInt(Res&: Tmp))))) |
1968 | C = *Tmp; |
1969 | else if ((!V && match(V: Op, P: m_Shl(L: m_Value(V), R: m_APInt(Res&: Tmp)))) || |
1970 | (V && match(V: Op, P: m_Shl(L: m_Specific(V), R: m_APInt(Res&: Tmp))))) |
1971 | C = APInt(Tmp->getBitWidth(), 1) << *Tmp; |
1972 | if (Tmp != nullptr) |
1973 | return true; |
1974 | |
1975 | // Reset `V` so we don't start with specific value on next match attempt. |
1976 | V = nullptr; |
1977 | return false; |
1978 | }; |
1979 | |
1980 | auto MatchShiftCX = [](Value *Op, APInt &C, Value *&V) -> bool { |
1981 | const APInt *Tmp = nullptr; |
1982 | if ((!V && match(V: Op, P: m_Shl(L: m_APInt(Res&: Tmp), R: m_Value(V)))) || |
1983 | (V && match(V: Op, P: m_Shl(L: m_APInt(Res&: Tmp), R: m_Specific(V))))) { |
1984 | C = *Tmp; |
1985 | return true; |
1986 | } |
1987 | |
1988 | // Reset `V` so we don't start with specific value on next match attempt. |
1989 | V = nullptr; |
1990 | return false; |
1991 | }; |
1992 | |
1993 | if (MatchShiftOrMulXC(Op0, X, Y) && MatchShiftOrMulXC(Op1, X, Z)) { |
1994 | // pass |
1995 | } else if (MatchShiftCX(Op0, Y, X) && MatchShiftCX(Op1, Z, X)) { |
1996 | ShiftByX = true; |
1997 | } else { |
1998 | return nullptr; |
1999 | } |
2000 | |
2001 | bool IsSRem = I.getOpcode() == Instruction::SRem; |
2002 | |
2003 | OverflowingBinaryOperator *BO0 = cast<OverflowingBinaryOperator>(Val: Op0); |
2004 | // TODO: We may be able to deduce more about nsw/nuw of BO0/BO1 based on Y >= |
2005 | // Z or Z >= Y. |
2006 | bool BO0HasNSW = BO0->hasNoSignedWrap(); |
2007 | bool BO0HasNUW = BO0->hasNoUnsignedWrap(); |
2008 | bool BO0NoWrap = IsSRem ? BO0HasNSW : BO0HasNUW; |
2009 | |
2010 | APInt RemYZ = IsSRem ? Y.srem(RHS: Z) : Y.urem(RHS: Z); |
2011 | // (rem (mul nuw/nsw X, Y), (mul X, Z)) |
2012 | // if (rem Y, Z) == 0 |
2013 | // -> 0 |
2014 | if (RemYZ.isZero() && BO0NoWrap) |
2015 | return IC.replaceInstUsesWith(I, V: ConstantInt::getNullValue(Ty: I.getType())); |
2016 | |
2017 | // Helper function to emit either (RemSimplificationC << X) or |
2018 | // (RemSimplificationC * X) depending on whether we matched Op0/Op1 as |
2019 | // (shl V, X) or (mul V, X) respectively. |
2020 | auto CreateMulOrShift = |
2021 | [&](const APInt &RemSimplificationC) -> BinaryOperator * { |
2022 | Value *RemSimplification = |
2023 | ConstantInt::get(Ty: I.getType(), V: RemSimplificationC); |
2024 | return ShiftByX ? BinaryOperator::CreateShl(V1: RemSimplification, V2: X) |
2025 | : BinaryOperator::CreateMul(V1: X, V2: RemSimplification); |
2026 | }; |
2027 | |
2028 | OverflowingBinaryOperator *BO1 = cast<OverflowingBinaryOperator>(Val: Op1); |
2029 | bool BO1HasNSW = BO1->hasNoSignedWrap(); |
2030 | bool BO1HasNUW = BO1->hasNoUnsignedWrap(); |
2031 | bool BO1NoWrap = IsSRem ? BO1HasNSW : BO1HasNUW; |
2032 | // (rem (mul X, Y), (mul nuw/nsw X, Z)) |
2033 | // if (rem Y, Z) == Y |
2034 | // -> (mul nuw/nsw X, Y) |
2035 | if (RemYZ == Y && BO1NoWrap) { |
2036 | BinaryOperator *BO = CreateMulOrShift(Y); |
2037 | // Copy any overflow flags from Op0. |
2038 | BO->setHasNoSignedWrap(IsSRem || BO0HasNSW); |
2039 | BO->setHasNoUnsignedWrap(!IsSRem || BO0HasNUW); |
2040 | return BO; |
2041 | } |
2042 | |
2043 | // (rem (mul nuw/nsw X, Y), (mul {nsw} X, Z)) |
2044 | // if Y >= Z |
2045 | // -> (mul {nuw} nsw X, (rem Y, Z)) |
2046 | if (Y.uge(RHS: Z) && (IsSRem ? (BO0HasNSW && BO1HasNSW) : BO0HasNUW)) { |
2047 | BinaryOperator *BO = CreateMulOrShift(RemYZ); |
2048 | BO->setHasNoSignedWrap(); |
2049 | BO->setHasNoUnsignedWrap(BO0HasNUW); |
2050 | return BO; |
2051 | } |
2052 | |
2053 | return nullptr; |
2054 | } |
2055 | |
2056 | /// This function implements the transforms common to both integer remainder |
2057 | /// instructions (urem and srem). It is called by the visitors to those integer |
2058 | /// remainder instructions. |
2059 | /// Common integer remainder transforms |
2060 | Instruction *InstCombinerImpl::commonIRemTransforms(BinaryOperator &I) { |
2061 | if (Instruction *Phi = foldBinopWithPhiOperands(BO&: I)) |
2062 | return Phi; |
2063 | |
2064 | Value *Op0 = I.getOperand(i_nocapture: 0), *Op1 = I.getOperand(i_nocapture: 1); |
2065 | |
2066 | // The RHS is known non-zero. |
2067 | if (Value *V = simplifyValueKnownNonZero(V: I.getOperand(i_nocapture: 1), IC&: *this, CxtI&: I)) |
2068 | return replaceOperand(I, OpNum: 1, V); |
2069 | |
2070 | // Handle cases involving: rem X, (select Cond, Y, Z) |
2071 | if (simplifyDivRemOfSelectWithZeroOp(I)) |
2072 | return &I; |
2073 | |
2074 | // If the divisor is a select-of-constants, try to constant fold all rem ops: |
2075 | // C % (select Cond, TrueC, FalseC) --> select Cond, (C % TrueC), (C % FalseC) |
2076 | // TODO: Adapt simplifyDivRemOfSelectWithZeroOp to allow this and other folds. |
2077 | if (match(V: Op0, P: m_ImmConstant()) && |
2078 | match(V: Op1, P: m_Select(C: m_Value(), L: m_ImmConstant(), R: m_ImmConstant()))) { |
2079 | if (Instruction *R = FoldOpIntoSelect(Op&: I, SI: cast<SelectInst>(Val: Op1), |
2080 | /*FoldWithMultiUse*/ true)) |
2081 | return R; |
2082 | } |
2083 | |
2084 | if (isa<Constant>(Val: Op1)) { |
2085 | if (Instruction *Op0I = dyn_cast<Instruction>(Val: Op0)) { |
2086 | if (SelectInst *SI = dyn_cast<SelectInst>(Val: Op0I)) { |
2087 | if (Instruction *R = FoldOpIntoSelect(Op&: I, SI)) |
2088 | return R; |
2089 | } else if (auto *PN = dyn_cast<PHINode>(Val: Op0I)) { |
2090 | const APInt *Op1Int; |
2091 | if (match(V: Op1, P: m_APInt(Res&: Op1Int)) && !Op1Int->isMinValue() && |
2092 | (I.getOpcode() == Instruction::URem || |
2093 | !Op1Int->isMinSignedValue())) { |
2094 | // foldOpIntoPhi will speculate instructions to the end of the PHI's |
2095 | // predecessor blocks, so do this only if we know the srem or urem |
2096 | // will not fault. |
2097 | if (Instruction *NV = foldOpIntoPhi(I, PN)) |
2098 | return NV; |
2099 | } |
2100 | } |
2101 | |
2102 | // See if we can fold away this rem instruction. |
2103 | if (SimplifyDemandedInstructionBits(Inst&: I)) |
2104 | return &I; |
2105 | } |
2106 | } |
2107 | |
2108 | if (Instruction *R = simplifyIRemMulShl(I, IC&: *this)) |
2109 | return R; |
2110 | |
2111 | return nullptr; |
2112 | } |
2113 | |
2114 | Instruction *InstCombinerImpl::visitURem(BinaryOperator &I) { |
2115 | if (Value *V = simplifyURemInst(LHS: I.getOperand(i_nocapture: 0), RHS: I.getOperand(i_nocapture: 1), |
2116 | Q: SQ.getWithInstruction(I: &I))) |
2117 | return replaceInstUsesWith(I, V); |
2118 | |
2119 | if (Instruction *X = foldVectorBinop(Inst&: I)) |
2120 | return X; |
2121 | |
2122 | if (Instruction *common = commonIRemTransforms(I)) |
2123 | return common; |
2124 | |
2125 | if (Instruction *NarrowRem = narrowUDivURem(I, IC&: *this)) |
2126 | return NarrowRem; |
2127 | |
2128 | // X urem Y -> X and Y-1, where Y is a power of 2, |
2129 | Value *Op0 = I.getOperand(i_nocapture: 0), *Op1 = I.getOperand(i_nocapture: 1); |
2130 | Type *Ty = I.getType(); |
2131 | if (isKnownToBeAPowerOfTwo(V: Op1, /*OrZero*/ true, Depth: 0, CxtI: &I)) { |
2132 | // This may increase instruction count, we don't enforce that Y is a |
2133 | // constant. |
2134 | Constant *N1 = Constant::getAllOnesValue(Ty); |
2135 | Value *Add = Builder.CreateAdd(LHS: Op1, RHS: N1); |
2136 | return BinaryOperator::CreateAnd(V1: Op0, V2: Add); |
2137 | } |
2138 | |
2139 | // 1 urem X -> zext(X != 1) |
2140 | if (match(V: Op0, P: m_One())) { |
2141 | Value *Cmp = Builder.CreateICmpNE(LHS: Op1, RHS: ConstantInt::get(Ty, V: 1)); |
2142 | return CastInst::CreateZExtOrBitCast(S: Cmp, Ty); |
2143 | } |
2144 | |
2145 | // Op0 urem C -> Op0 < C ? Op0 : Op0 - C, where C >= signbit. |
2146 | // Op0 must be frozen because we are increasing its number of uses. |
2147 | if (match(V: Op1, P: m_Negative())) { |
2148 | Value *F0 = Builder.CreateFreeze(V: Op0, Name: Op0->getName() + ".fr" ); |
2149 | Value *Cmp = Builder.CreateICmpULT(LHS: F0, RHS: Op1); |
2150 | Value *Sub = Builder.CreateSub(LHS: F0, RHS: Op1); |
2151 | return SelectInst::Create(C: Cmp, S1: F0, S2: Sub); |
2152 | } |
2153 | |
2154 | // If the divisor is a sext of a boolean, then the divisor must be max |
2155 | // unsigned value (-1). Therefore, the remainder is Op0 unless Op0 is also |
2156 | // max unsigned value. In that case, the remainder is 0: |
2157 | // urem Op0, (sext i1 X) --> (Op0 == -1) ? 0 : Op0 |
2158 | Value *X; |
2159 | if (match(V: Op1, P: m_SExt(Op: m_Value(V&: X))) && X->getType()->isIntOrIntVectorTy(BitWidth: 1)) { |
2160 | Value *FrozenOp0 = Builder.CreateFreeze(V: Op0, Name: Op0->getName() + ".frozen" ); |
2161 | Value *Cmp = |
2162 | Builder.CreateICmpEQ(LHS: FrozenOp0, RHS: ConstantInt::getAllOnesValue(Ty)); |
2163 | return SelectInst::Create(C: Cmp, S1: ConstantInt::getNullValue(Ty), S2: FrozenOp0); |
2164 | } |
2165 | |
2166 | // For "(X + 1) % Op1" and if (X u< Op1) => (X + 1) == Op1 ? 0 : X + 1 . |
2167 | if (match(V: Op0, P: m_Add(L: m_Value(V&: X), R: m_One()))) { |
2168 | Value *Val = |
2169 | simplifyICmpInst(Predicate: ICmpInst::ICMP_ULT, LHS: X, RHS: Op1, Q: SQ.getWithInstruction(I: &I)); |
2170 | if (Val && match(V: Val, P: m_One())) { |
2171 | Value *FrozenOp0 = Builder.CreateFreeze(V: Op0, Name: Op0->getName() + ".frozen" ); |
2172 | Value *Cmp = Builder.CreateICmpEQ(LHS: FrozenOp0, RHS: Op1); |
2173 | return SelectInst::Create(C: Cmp, S1: ConstantInt::getNullValue(Ty), S2: FrozenOp0); |
2174 | } |
2175 | } |
2176 | |
2177 | return nullptr; |
2178 | } |
2179 | |
2180 | Instruction *InstCombinerImpl::visitSRem(BinaryOperator &I) { |
2181 | if (Value *V = simplifySRemInst(LHS: I.getOperand(i_nocapture: 0), RHS: I.getOperand(i_nocapture: 1), |
2182 | Q: SQ.getWithInstruction(I: &I))) |
2183 | return replaceInstUsesWith(I, V); |
2184 | |
2185 | if (Instruction *X = foldVectorBinop(Inst&: I)) |
2186 | return X; |
2187 | |
2188 | // Handle the integer rem common cases |
2189 | if (Instruction *Common = commonIRemTransforms(I)) |
2190 | return Common; |
2191 | |
2192 | Value *Op0 = I.getOperand(i_nocapture: 0), *Op1 = I.getOperand(i_nocapture: 1); |
2193 | { |
2194 | const APInt *Y; |
2195 | // X % -Y -> X % Y |
2196 | if (match(V: Op1, P: m_Negative(V&: Y)) && !Y->isMinSignedValue()) |
2197 | return replaceOperand(I, OpNum: 1, V: ConstantInt::get(Ty: I.getType(), V: -*Y)); |
2198 | } |
2199 | |
2200 | // -X srem Y --> -(X srem Y) |
2201 | Value *X, *Y; |
2202 | if (match(V: &I, P: m_SRem(L: m_OneUse(SubPattern: m_NSWNeg(V: m_Value(V&: X))), R: m_Value(V&: Y)))) |
2203 | return BinaryOperator::CreateNSWNeg(Op: Builder.CreateSRem(LHS: X, RHS: Y)); |
2204 | |
2205 | // If the sign bits of both operands are zero (i.e. we can prove they are |
2206 | // unsigned inputs), turn this into a urem. |
2207 | APInt Mask(APInt::getSignMask(BitWidth: I.getType()->getScalarSizeInBits())); |
2208 | if (MaskedValueIsZero(V: Op1, Mask, Depth: 0, CxtI: &I) && |
2209 | MaskedValueIsZero(V: Op0, Mask, Depth: 0, CxtI: &I)) { |
2210 | // X srem Y -> X urem Y, iff X and Y don't have sign bit set |
2211 | return BinaryOperator::CreateURem(V1: Op0, V2: Op1, Name: I.getName()); |
2212 | } |
2213 | |
2214 | // If it's a constant vector, flip any negative values positive. |
2215 | if (isa<ConstantVector>(Val: Op1) || isa<ConstantDataVector>(Val: Op1)) { |
2216 | Constant *C = cast<Constant>(Val: Op1); |
2217 | unsigned VWidth = cast<FixedVectorType>(Val: C->getType())->getNumElements(); |
2218 | |
2219 | bool hasNegative = false; |
2220 | bool hasMissing = false; |
2221 | for (unsigned i = 0; i != VWidth; ++i) { |
2222 | Constant *Elt = C->getAggregateElement(Elt: i); |
2223 | if (!Elt) { |
2224 | hasMissing = true; |
2225 | break; |
2226 | } |
2227 | |
2228 | if (ConstantInt *RHS = dyn_cast<ConstantInt>(Val: Elt)) |
2229 | if (RHS->isNegative()) |
2230 | hasNegative = true; |
2231 | } |
2232 | |
2233 | if (hasNegative && !hasMissing) { |
2234 | SmallVector<Constant *, 16> Elts(VWidth); |
2235 | for (unsigned i = 0; i != VWidth; ++i) { |
2236 | Elts[i] = C->getAggregateElement(Elt: i); // Handle undef, etc. |
2237 | if (ConstantInt *RHS = dyn_cast<ConstantInt>(Val: Elts[i])) { |
2238 | if (RHS->isNegative()) |
2239 | Elts[i] = cast<ConstantInt>(Val: ConstantExpr::getNeg(C: RHS)); |
2240 | } |
2241 | } |
2242 | |
2243 | Constant *NewRHSV = ConstantVector::get(V: Elts); |
2244 | if (NewRHSV != C) // Don't loop on -MININT |
2245 | return replaceOperand(I, OpNum: 1, V: NewRHSV); |
2246 | } |
2247 | } |
2248 | |
2249 | return nullptr; |
2250 | } |
2251 | |
2252 | Instruction *InstCombinerImpl::visitFRem(BinaryOperator &I) { |
2253 | if (Value *V = simplifyFRemInst(LHS: I.getOperand(i_nocapture: 0), RHS: I.getOperand(i_nocapture: 1), |
2254 | FMF: I.getFastMathFlags(), |
2255 | Q: SQ.getWithInstruction(I: &I))) |
2256 | return replaceInstUsesWith(I, V); |
2257 | |
2258 | if (Instruction *X = foldVectorBinop(Inst&: I)) |
2259 | return X; |
2260 | |
2261 | if (Instruction *Phi = foldBinopWithPhiOperands(BO&: I)) |
2262 | return Phi; |
2263 | |
2264 | return nullptr; |
2265 | } |
2266 | |