1 //===- InstCombineMulDivRem.cpp -------------------------------------------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This file implements the visit functions for mul, fmul, sdiv, udiv, fdiv,
13 //===----------------------------------------------------------------------===//
15 #include "InstCombineInternal.h"
16 #include "llvm/Analysis/InstructionSimplify.h"
17 #include "llvm/IR/IntrinsicInst.h"
18 #include "llvm/IR/PatternMatch.h"
20 using namespace PatternMatch;
22 #define DEBUG_TYPE "instcombine"
25 /// The specific integer value is used in a context where it is known to be
26 /// non-zero. If this allows us to simplify the computation, do so and return
27 /// the new operand, otherwise return null.
28 static Value *simplifyValueKnownNonZero(Value *V, InstCombiner &IC,
30 // If V has multiple uses, then we would have to do more analysis to determine
31 // if this is safe. For example, the use could be in dynamically unreached
33 if (!V->hasOneUse()) return nullptr;
35 bool MadeChange = false;
37 // ((1 << A) >>u B) --> (1 << (A-B))
38 // Because V cannot be zero, we know that B is less than A.
39 Value *A = nullptr, *B = nullptr, *One = nullptr;
40 if (match(V, m_LShr(m_OneUse(m_Shl(m_Value(One), m_Value(A))), m_Value(B))) &&
41 match(One, m_One())) {
42 A = IC.Builder->CreateSub(A, B);
43 return IC.Builder->CreateShl(One, A);
46 // (PowerOfTwo >>u B) --> isExact since shifting out the result would make it
47 // inexact. Similarly for <<.
48 BinaryOperator *I = dyn_cast<BinaryOperator>(V);
49 if (I && I->isLogicalShift() &&
50 IC.isKnownToBeAPowerOfTwo(I->getOperand(0), false, 0, &CxtI)) {
51 // We know that this is an exact/nuw shift and that the input is a
52 // non-zero context as well.
53 if (Value *V2 = simplifyValueKnownNonZero(I->getOperand(0), IC, CxtI)) {
58 if (I->getOpcode() == Instruction::LShr && !I->isExact()) {
63 if (I->getOpcode() == Instruction::Shl && !I->hasNoUnsignedWrap()) {
64 I->setHasNoUnsignedWrap();
69 // TODO: Lots more we could do here:
70 // If V is a phi node, we can call this on each of its operands.
71 // "select cond, X, 0" can simplify to "X".
73 return MadeChange ? V : nullptr;
77 /// True if the multiply can not be expressed in an int this size.
78 static bool MultiplyOverflows(const APInt &C1, const APInt &C2, APInt &Product,
82 Product = C1.smul_ov(C2, Overflow);
84 Product = C1.umul_ov(C2, Overflow);
89 /// \brief True if C2 is a multiple of C1. Quotient contains C2/C1.
90 static bool IsMultiple(const APInt &C1, const APInt &C2, APInt &Quotient,
92 assert(C1.getBitWidth() == C2.getBitWidth() &&
93 "Inconsistent width of constants!");
95 // Bail if we will divide by zero.
99 // Bail if we would divide INT_MIN by -1.
100 if (IsSigned && C1.isMinSignedValue() && C2.isAllOnesValue())
103 APInt Remainder(C1.getBitWidth(), /*Val=*/0ULL, IsSigned);
105 APInt::sdivrem(C1, C2, Quotient, Remainder);
107 APInt::udivrem(C1, C2, Quotient, Remainder);
109 return Remainder.isMinValue();
112 /// \brief A helper routine of InstCombiner::visitMul().
114 /// If C is a vector of known powers of 2, then this function returns
115 /// a new vector obtained from C replacing each element with its logBase2.
116 /// Return a null pointer otherwise.
117 static Constant *getLogBase2Vector(ConstantDataVector *CV) {
119 SmallVector<Constant *, 4> Elts;
121 for (unsigned I = 0, E = CV->getNumElements(); I != E; ++I) {
122 Constant *Elt = CV->getElementAsConstant(I);
123 if (!match(Elt, m_APInt(IVal)) || !IVal->isPowerOf2())
125 Elts.push_back(ConstantInt::get(Elt->getType(), IVal->logBase2()));
128 return ConstantVector::get(Elts);
131 /// \brief Return true if we can prove that:
132 /// (mul LHS, RHS) === (mul nsw LHS, RHS)
133 bool InstCombiner::willNotOverflowSignedMul(const Value *LHS,
135 const Instruction &CxtI) const {
136 // Multiplying n * m significant bits yields a result of n + m significant
137 // bits. If the total number of significant bits does not exceed the
138 // result bit width (minus 1), there is no overflow.
139 // This means if we have enough leading sign bits in the operands
140 // we can guarantee that the result does not overflow.
141 // Ref: "Hacker's Delight" by Henry Warren
142 unsigned BitWidth = LHS->getType()->getScalarSizeInBits();
144 // Note that underestimating the number of sign bits gives a more
145 // conservative answer.
147 ComputeNumSignBits(LHS, 0, &CxtI) + ComputeNumSignBits(RHS, 0, &CxtI);
149 // First handle the easy case: if we have enough sign bits there's
150 // definitely no overflow.
151 if (SignBits > BitWidth + 1)
154 // There are two ambiguous cases where there can be no overflow:
155 // SignBits == BitWidth + 1 and
156 // SignBits == BitWidth
157 // The second case is difficult to check, therefore we only handle the
159 if (SignBits == BitWidth + 1) {
160 // It overflows only when both arguments are negative and the true
161 // product is exactly the minimum negative number.
162 // E.g. mul i16 with 17 sign bits: 0xff00 * 0xff80 = 0x8000
163 // For simplicity we just check if at least one side is not negative.
164 KnownBits LHSKnown = computeKnownBits(LHS, /*Depth=*/0, &CxtI);
165 KnownBits RHSKnown = computeKnownBits(RHS, /*Depth=*/0, &CxtI);
166 if (LHSKnown.isNonNegative() || RHSKnown.isNonNegative())
172 Instruction *InstCombiner::visitMul(BinaryOperator &I) {
173 bool Changed = SimplifyAssociativeOrCommutative(I);
174 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
176 if (Value *V = SimplifyVectorOp(I))
177 return replaceInstUsesWith(I, V);
179 if (Value *V = SimplifyMulInst(Op0, Op1, SQ))
180 return replaceInstUsesWith(I, V);
182 if (Value *V = SimplifyUsingDistributiveLaws(I))
183 return replaceInstUsesWith(I, V);
186 if (match(Op1, m_AllOnes())) {
187 BinaryOperator *BO = BinaryOperator::CreateNeg(Op0, I.getName());
188 if (I.hasNoSignedWrap())
189 BO->setHasNoSignedWrap();
193 // Also allow combining multiply instructions on vectors.
198 if (match(&I, m_Mul(m_Shl(m_Value(NewOp), m_Constant(C2)),
200 match(C1, m_APInt(IVal))) {
201 // ((X << C2)*C1) == (X * (C1 << C2))
202 Constant *Shl = ConstantExpr::getShl(C1, C2);
203 BinaryOperator *Mul = cast<BinaryOperator>(I.getOperand(0));
204 BinaryOperator *BO = BinaryOperator::CreateMul(NewOp, Shl);
205 if (I.hasNoUnsignedWrap() && Mul->hasNoUnsignedWrap())
206 BO->setHasNoUnsignedWrap();
207 if (I.hasNoSignedWrap() && Mul->hasNoSignedWrap() &&
208 Shl->isNotMinSignedValue())
209 BO->setHasNoSignedWrap();
213 if (match(&I, m_Mul(m_Value(NewOp), m_Constant(C1)))) {
214 Constant *NewCst = nullptr;
215 if (match(C1, m_APInt(IVal)) && IVal->isPowerOf2())
216 // Replace X*(2^C) with X << C, where C is either a scalar or a splat.
217 NewCst = ConstantInt::get(NewOp->getType(), IVal->logBase2());
218 else if (ConstantDataVector *CV = dyn_cast<ConstantDataVector>(C1))
219 // Replace X*(2^C) with X << C, where C is a vector of known
220 // constant powers of 2.
221 NewCst = getLogBase2Vector(CV);
224 unsigned Width = NewCst->getType()->getPrimitiveSizeInBits();
225 BinaryOperator *Shl = BinaryOperator::CreateShl(NewOp, NewCst);
227 if (I.hasNoUnsignedWrap())
228 Shl->setHasNoUnsignedWrap();
229 if (I.hasNoSignedWrap()) {
231 if (match(NewCst, m_ConstantInt(V)) && V != Width - 1)
232 Shl->setHasNoSignedWrap();
240 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
241 // (Y - X) * (-(2**n)) -> (X - Y) * (2**n), for positive nonzero n
242 // (Y + const) * (-(2**n)) -> (-constY) * (2**n), for positive nonzero n
243 // The "* (2**n)" thus becomes a potential shifting opportunity.
245 const APInt & Val = CI->getValue();
246 const APInt &PosVal = Val.abs();
247 if (Val.isNegative() && PosVal.isPowerOf2()) {
248 Value *X = nullptr, *Y = nullptr;
249 if (Op0->hasOneUse()) {
251 Value *Sub = nullptr;
252 if (match(Op0, m_Sub(m_Value(Y), m_Value(X))))
253 Sub = Builder->CreateSub(X, Y, "suba");
254 else if (match(Op0, m_Add(m_Value(Y), m_ConstantInt(C1))))
255 Sub = Builder->CreateSub(Builder->CreateNeg(C1), Y, "subc");
258 BinaryOperator::CreateMul(Sub,
259 ConstantInt::get(Y->getType(), PosVal));
265 // Simplify mul instructions with a constant RHS.
266 if (isa<Constant>(Op1)) {
267 if (Instruction *FoldedMul = foldOpWithConstantIntoOperand(I))
270 // Canonicalize (X+C1)*CI -> X*CI+C1*CI.
274 if (match(Op0, m_OneUse(m_Add(m_Value(X), m_Constant(C1))))) {
275 Value *Mul = Builder->CreateMul(C1, Op1);
276 // Only go forward with the transform if C1*CI simplifies to a tidier
278 if (!match(Mul, m_Mul(m_Value(), m_Value())))
279 return BinaryOperator::CreateAdd(Builder->CreateMul(X, Op1), Mul);
284 if (Value *Op0v = dyn_castNegVal(Op0)) { // -X * -Y = X*Y
285 if (Value *Op1v = dyn_castNegVal(Op1)) {
286 BinaryOperator *BO = BinaryOperator::CreateMul(Op0v, Op1v);
287 if (I.hasNoSignedWrap() &&
288 match(Op0, m_NSWSub(m_Value(), m_Value())) &&
289 match(Op1, m_NSWSub(m_Value(), m_Value())))
290 BO->setHasNoSignedWrap();
295 // (X / Y) * Y = X - (X % Y)
296 // (X / Y) * -Y = (X % Y) - X
299 BinaryOperator *Div = dyn_cast<BinaryOperator>(Op0);
300 if (!Div || (Div->getOpcode() != Instruction::UDiv &&
301 Div->getOpcode() != Instruction::SDiv)) {
303 Div = dyn_cast<BinaryOperator>(Op1);
305 Value *Neg = dyn_castNegVal(Y);
306 if (Div && Div->hasOneUse() &&
307 (Div->getOperand(1) == Y || Div->getOperand(1) == Neg) &&
308 (Div->getOpcode() == Instruction::UDiv ||
309 Div->getOpcode() == Instruction::SDiv)) {
310 Value *X = Div->getOperand(0), *DivOp1 = Div->getOperand(1);
312 // If the division is exact, X % Y is zero, so we end up with X or -X.
313 if (Div->isExact()) {
315 return replaceInstUsesWith(I, X);
316 return BinaryOperator::CreateNeg(X);
319 auto RemOpc = Div->getOpcode() == Instruction::UDiv ? Instruction::URem
321 Value *Rem = Builder->CreateBinOp(RemOpc, X, DivOp1);
323 return BinaryOperator::CreateSub(X, Rem);
324 return BinaryOperator::CreateSub(Rem, X);
328 /// i1 mul -> i1 and.
329 if (I.getType()->getScalarType()->isIntegerTy(1))
330 return BinaryOperator::CreateAnd(Op0, Op1);
332 // X*(1 << Y) --> X << Y
333 // (1 << Y)*X --> X << Y
336 BinaryOperator *BO = nullptr;
338 if (match(Op0, m_Shl(m_One(), m_Value(Y)))) {
339 BO = BinaryOperator::CreateShl(Op1, Y);
340 ShlNSW = cast<ShlOperator>(Op0)->hasNoSignedWrap();
341 } else if (match(Op1, m_Shl(m_One(), m_Value(Y)))) {
342 BO = BinaryOperator::CreateShl(Op0, Y);
343 ShlNSW = cast<ShlOperator>(Op1)->hasNoSignedWrap();
346 if (I.hasNoUnsignedWrap())
347 BO->setHasNoUnsignedWrap();
348 if (I.hasNoSignedWrap() && ShlNSW)
349 BO->setHasNoSignedWrap();
354 // If one of the operands of the multiply is a cast from a boolean value, then
355 // we know the bool is either zero or one, so this is a 'masking' multiply.
356 // X * Y (where Y is 0 or 1) -> X & (0-Y)
357 if (!I.getType()->isVectorTy()) {
358 // -2 is "-1 << 1" so it is all bits set except the low one.
359 APInt Negative2(I.getType()->getPrimitiveSizeInBits(), (uint64_t)-2, true);
361 Value *BoolCast = nullptr, *OtherOp = nullptr;
362 if (MaskedValueIsZero(Op0, Negative2, 0, &I)) {
365 } else if (MaskedValueIsZero(Op1, Negative2, 0, &I)) {
371 Value *V = Builder->CreateSub(Constant::getNullValue(I.getType()),
373 return BinaryOperator::CreateAnd(V, OtherOp);
377 // Check for (mul (sext x), y), see if we can merge this into an
378 // integer mul followed by a sext.
379 if (SExtInst *Op0Conv = dyn_cast<SExtInst>(Op0)) {
380 // (mul (sext x), cst) --> (sext (mul x, cst'))
381 if (ConstantInt *Op1C = dyn_cast<ConstantInt>(Op1)) {
382 if (Op0Conv->hasOneUse()) {
384 ConstantExpr::getTrunc(Op1C, Op0Conv->getOperand(0)->getType());
385 if (ConstantExpr::getSExt(CI, I.getType()) == Op1C &&
386 willNotOverflowSignedMul(Op0Conv->getOperand(0), CI, I)) {
387 // Insert the new, smaller mul.
389 Builder->CreateNSWMul(Op0Conv->getOperand(0), CI, "mulconv");
390 return new SExtInst(NewMul, I.getType());
395 // (mul (sext x), (sext y)) --> (sext (mul int x, y))
396 if (SExtInst *Op1Conv = dyn_cast<SExtInst>(Op1)) {
397 // Only do this if x/y have the same type, if at last one of them has a
398 // single use (so we don't increase the number of sexts), and if the
399 // integer mul will not overflow.
400 if (Op0Conv->getOperand(0)->getType() ==
401 Op1Conv->getOperand(0)->getType() &&
402 (Op0Conv->hasOneUse() || Op1Conv->hasOneUse()) &&
403 willNotOverflowSignedMul(Op0Conv->getOperand(0),
404 Op1Conv->getOperand(0), I)) {
405 // Insert the new integer mul.
406 Value *NewMul = Builder->CreateNSWMul(
407 Op0Conv->getOperand(0), Op1Conv->getOperand(0), "mulconv");
408 return new SExtInst(NewMul, I.getType());
413 // Check for (mul (zext x), y), see if we can merge this into an
414 // integer mul followed by a zext.
415 if (auto *Op0Conv = dyn_cast<ZExtInst>(Op0)) {
416 // (mul (zext x), cst) --> (zext (mul x, cst'))
417 if (ConstantInt *Op1C = dyn_cast<ConstantInt>(Op1)) {
418 if (Op0Conv->hasOneUse()) {
420 ConstantExpr::getTrunc(Op1C, Op0Conv->getOperand(0)->getType());
421 if (ConstantExpr::getZExt(CI, I.getType()) == Op1C &&
422 willNotOverflowUnsignedMul(Op0Conv->getOperand(0), CI, I)) {
423 // Insert the new, smaller mul.
425 Builder->CreateNUWMul(Op0Conv->getOperand(0), CI, "mulconv");
426 return new ZExtInst(NewMul, I.getType());
431 // (mul (zext x), (zext y)) --> (zext (mul int x, y))
432 if (auto *Op1Conv = dyn_cast<ZExtInst>(Op1)) {
433 // Only do this if x/y have the same type, if at last one of them has a
434 // single use (so we don't increase the number of zexts), and if the
435 // integer mul will not overflow.
436 if (Op0Conv->getOperand(0)->getType() ==
437 Op1Conv->getOperand(0)->getType() &&
438 (Op0Conv->hasOneUse() || Op1Conv->hasOneUse()) &&
439 willNotOverflowUnsignedMul(Op0Conv->getOperand(0),
440 Op1Conv->getOperand(0), I)) {
441 // Insert the new integer mul.
442 Value *NewMul = Builder->CreateNUWMul(
443 Op0Conv->getOperand(0), Op1Conv->getOperand(0), "mulconv");
444 return new ZExtInst(NewMul, I.getType());
449 if (!I.hasNoSignedWrap() && willNotOverflowSignedMul(Op0, Op1, I)) {
451 I.setHasNoSignedWrap(true);
454 if (!I.hasNoUnsignedWrap() && willNotOverflowUnsignedMul(Op0, Op1, I)) {
456 I.setHasNoUnsignedWrap(true);
459 return Changed ? &I : nullptr;
462 /// Detect pattern log2(Y * 0.5) with corresponding fast math flags.
463 static void detectLog2OfHalf(Value *&Op, Value *&Y, IntrinsicInst *&Log2) {
464 if (!Op->hasOneUse())
467 IntrinsicInst *II = dyn_cast<IntrinsicInst>(Op);
470 if (II->getIntrinsicID() != Intrinsic::log2 || !II->hasUnsafeAlgebra())
474 Value *OpLog2Of = II->getArgOperand(0);
475 if (!OpLog2Of->hasOneUse())
478 Instruction *I = dyn_cast<Instruction>(OpLog2Of);
481 if (I->getOpcode() != Instruction::FMul || !I->hasUnsafeAlgebra())
484 if (match(I->getOperand(0), m_SpecificFP(0.5)))
485 Y = I->getOperand(1);
486 else if (match(I->getOperand(1), m_SpecificFP(0.5)))
487 Y = I->getOperand(0);
490 static bool isFiniteNonZeroFp(Constant *C) {
491 if (C->getType()->isVectorTy()) {
492 for (unsigned I = 0, E = C->getType()->getVectorNumElements(); I != E;
494 ConstantFP *CFP = dyn_cast_or_null<ConstantFP>(C->getAggregateElement(I));
495 if (!CFP || !CFP->getValueAPF().isFiniteNonZero())
501 return isa<ConstantFP>(C) &&
502 cast<ConstantFP>(C)->getValueAPF().isFiniteNonZero();
505 static bool isNormalFp(Constant *C) {
506 if (C->getType()->isVectorTy()) {
507 for (unsigned I = 0, E = C->getType()->getVectorNumElements(); I != E;
509 ConstantFP *CFP = dyn_cast_or_null<ConstantFP>(C->getAggregateElement(I));
510 if (!CFP || !CFP->getValueAPF().isNormal())
516 return isa<ConstantFP>(C) && cast<ConstantFP>(C)->getValueAPF().isNormal();
519 /// Helper function of InstCombiner::visitFMul(BinaryOperator(). It returns
520 /// true iff the given value is FMul or FDiv with one and only one operand
521 /// being a normal constant (i.e. not Zero/NaN/Infinity).
522 static bool isFMulOrFDivWithConstant(Value *V) {
523 Instruction *I = dyn_cast<Instruction>(V);
524 if (!I || (I->getOpcode() != Instruction::FMul &&
525 I->getOpcode() != Instruction::FDiv))
528 Constant *C0 = dyn_cast<Constant>(I->getOperand(0));
529 Constant *C1 = dyn_cast<Constant>(I->getOperand(1));
534 return (C0 && isFiniteNonZeroFp(C0)) || (C1 && isFiniteNonZeroFp(C1));
537 /// foldFMulConst() is a helper routine of InstCombiner::visitFMul().
538 /// The input \p FMulOrDiv is a FMul/FDiv with one and only one operand
539 /// being a constant (i.e. isFMulOrFDivWithConstant(FMulOrDiv) == true).
540 /// This function is to simplify "FMulOrDiv * C" and returns the
541 /// resulting expression. Note that this function could return NULL in
542 /// case the constants cannot be folded into a normal floating-point.
544 Value *InstCombiner::foldFMulConst(Instruction *FMulOrDiv, Constant *C,
545 Instruction *InsertBefore) {
546 assert(isFMulOrFDivWithConstant(FMulOrDiv) && "V is invalid");
548 Value *Opnd0 = FMulOrDiv->getOperand(0);
549 Value *Opnd1 = FMulOrDiv->getOperand(1);
551 Constant *C0 = dyn_cast<Constant>(Opnd0);
552 Constant *C1 = dyn_cast<Constant>(Opnd1);
554 BinaryOperator *R = nullptr;
556 // (X * C0) * C => X * (C0*C)
557 if (FMulOrDiv->getOpcode() == Instruction::FMul) {
558 Constant *F = ConstantExpr::getFMul(C1 ? C1 : C0, C);
560 R = BinaryOperator::CreateFMul(C1 ? Opnd0 : Opnd1, F);
563 // (C0 / X) * C => (C0 * C) / X
564 if (FMulOrDiv->hasOneUse()) {
565 // It would otherwise introduce another div.
566 Constant *F = ConstantExpr::getFMul(C0, C);
568 R = BinaryOperator::CreateFDiv(F, Opnd1);
571 // (X / C1) * C => X * (C/C1) if C/C1 is not a denormal
572 Constant *F = ConstantExpr::getFDiv(C, C1);
574 R = BinaryOperator::CreateFMul(Opnd0, F);
576 // (X / C1) * C => X / (C1/C)
577 Constant *F = ConstantExpr::getFDiv(C1, C);
579 R = BinaryOperator::CreateFDiv(Opnd0, F);
585 R->setHasUnsafeAlgebra(true);
586 InsertNewInstWith(R, *InsertBefore);
592 Instruction *InstCombiner::visitFMul(BinaryOperator &I) {
593 bool Changed = SimplifyAssociativeOrCommutative(I);
594 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
596 if (Value *V = SimplifyVectorOp(I))
597 return replaceInstUsesWith(I, V);
599 if (isa<Constant>(Op0))
602 if (Value *V = SimplifyFMulInst(Op0, Op1, I.getFastMathFlags(), SQ))
603 return replaceInstUsesWith(I, V);
605 bool AllowReassociate = I.hasUnsafeAlgebra();
607 // Simplify mul instructions with a constant RHS.
608 if (isa<Constant>(Op1)) {
609 if (Instruction *FoldedMul = foldOpWithConstantIntoOperand(I))
612 // (fmul X, -1.0) --> (fsub -0.0, X)
613 if (match(Op1, m_SpecificFP(-1.0))) {
614 Constant *NegZero = ConstantFP::getNegativeZero(Op1->getType());
615 Instruction *RI = BinaryOperator::CreateFSub(NegZero, Op0);
616 RI->copyFastMathFlags(&I);
620 Constant *C = cast<Constant>(Op1);
621 if (AllowReassociate && isFiniteNonZeroFp(C)) {
622 // Let MDC denote an expression in one of these forms:
623 // X * C, C/X, X/C, where C is a constant.
625 // Try to simplify "MDC * Constant"
626 if (isFMulOrFDivWithConstant(Op0))
627 if (Value *V = foldFMulConst(cast<Instruction>(Op0), C, &I))
628 return replaceInstUsesWith(I, V);
630 // (MDC +/- C1) * C => (MDC * C) +/- (C1 * C)
631 Instruction *FAddSub = dyn_cast<Instruction>(Op0);
633 (FAddSub->getOpcode() == Instruction::FAdd ||
634 FAddSub->getOpcode() == Instruction::FSub)) {
635 Value *Opnd0 = FAddSub->getOperand(0);
636 Value *Opnd1 = FAddSub->getOperand(1);
637 Constant *C0 = dyn_cast<Constant>(Opnd0);
638 Constant *C1 = dyn_cast<Constant>(Opnd1);
642 std::swap(Opnd0, Opnd1);
646 if (C1 && isFiniteNonZeroFp(C1) && isFMulOrFDivWithConstant(Opnd0)) {
647 Value *M1 = ConstantExpr::getFMul(C1, C);
648 Value *M0 = isNormalFp(cast<Constant>(M1)) ?
649 foldFMulConst(cast<Instruction>(Opnd0), C, &I) :
652 if (Swap && FAddSub->getOpcode() == Instruction::FSub)
655 Instruction *RI = (FAddSub->getOpcode() == Instruction::FAdd)
656 ? BinaryOperator::CreateFAdd(M0, M1)
657 : BinaryOperator::CreateFSub(M0, M1);
658 RI->copyFastMathFlags(&I);
667 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(Op0)) {
668 // sqrt(X) * sqrt(X) -> X
669 if (AllowReassociate && II->getIntrinsicID() == Intrinsic::sqrt)
670 return replaceInstUsesWith(I, II->getOperand(0));
672 // fabs(X) * fabs(X) -> X * X
673 if (II->getIntrinsicID() == Intrinsic::fabs) {
674 Instruction *FMulVal = BinaryOperator::CreateFMul(II->getOperand(0),
677 FMulVal->copyFastMathFlags(&I);
683 // Under unsafe algebra do:
684 // X * log2(0.5*Y) = X*log2(Y) - X
685 if (AllowReassociate) {
686 Value *OpX = nullptr;
687 Value *OpY = nullptr;
689 detectLog2OfHalf(Op0, OpY, Log2);
693 detectLog2OfHalf(Op1, OpY, Log2);
698 // if pattern detected emit alternate sequence
700 BuilderTy::FastMathFlagGuard Guard(*Builder);
701 Builder->setFastMathFlags(Log2->getFastMathFlags());
702 Log2->setArgOperand(0, OpY);
703 Value *FMulVal = Builder->CreateFMul(OpX, Log2);
704 Value *FSub = Builder->CreateFSub(FMulVal, OpX);
706 return replaceInstUsesWith(I, FSub);
710 // Handle symmetric situation in a 2-iteration loop
713 for (int i = 0; i < 2; i++) {
714 bool IgnoreZeroSign = I.hasNoSignedZeros();
715 if (BinaryOperator::isFNeg(Opnd0, IgnoreZeroSign)) {
716 BuilderTy::FastMathFlagGuard Guard(*Builder);
717 Builder->setFastMathFlags(I.getFastMathFlags());
719 Value *N0 = dyn_castFNegVal(Opnd0, IgnoreZeroSign);
720 Value *N1 = dyn_castFNegVal(Opnd1, IgnoreZeroSign);
724 Value *FMul = Builder->CreateFMul(N0, N1);
726 return replaceInstUsesWith(I, FMul);
729 if (Opnd0->hasOneUse()) {
730 // -X * Y => -(X*Y) (Promote negation as high as possible)
731 Value *T = Builder->CreateFMul(N0, Opnd1);
732 Value *Neg = Builder->CreateFNeg(T);
734 return replaceInstUsesWith(I, Neg);
738 // (X*Y) * X => (X*X) * Y where Y != X
739 // The purpose is two-fold:
740 // 1) to form a power expression (of X).
741 // 2) potentially shorten the critical path: After transformation, the
742 // latency of the instruction Y is amortized by the expression of X*X,
743 // and therefore Y is in a "less critical" position compared to what it
744 // was before the transformation.
746 if (AllowReassociate) {
747 Value *Opnd0_0, *Opnd0_1;
748 if (Opnd0->hasOneUse() &&
749 match(Opnd0, m_FMul(m_Value(Opnd0_0), m_Value(Opnd0_1)))) {
751 if (Opnd0_0 == Opnd1 && Opnd0_1 != Opnd1)
753 else if (Opnd0_1 == Opnd1 && Opnd0_0 != Opnd1)
757 BuilderTy::FastMathFlagGuard Guard(*Builder);
758 Builder->setFastMathFlags(I.getFastMathFlags());
759 Value *T = Builder->CreateFMul(Opnd1, Opnd1);
760 Value *R = Builder->CreateFMul(T, Y);
762 return replaceInstUsesWith(I, R);
767 if (!isa<Constant>(Op1))
768 std::swap(Opnd0, Opnd1);
773 return Changed ? &I : nullptr;
776 /// Try to fold a divide or remainder of a select instruction.
777 bool InstCombiner::SimplifyDivRemOfSelect(BinaryOperator &I) {
778 SelectInst *SI = cast<SelectInst>(I.getOperand(1));
780 // div/rem X, (Cond ? 0 : Y) -> div/rem X, Y
781 int NonNullOperand = -1;
782 if (Constant *ST = dyn_cast<Constant>(SI->getOperand(1)))
783 if (ST->isNullValue())
785 // div/rem X, (Cond ? Y : 0) -> div/rem X, Y
786 if (Constant *ST = dyn_cast<Constant>(SI->getOperand(2)))
787 if (ST->isNullValue())
790 if (NonNullOperand == -1)
793 Value *SelectCond = SI->getOperand(0);
795 // Change the div/rem to use 'Y' instead of the select.
796 I.setOperand(1, SI->getOperand(NonNullOperand));
798 // Okay, we know we replace the operand of the div/rem with 'Y' with no
799 // problem. However, the select, or the condition of the select may have
800 // multiple uses. Based on our knowledge that the operand must be non-zero,
801 // propagate the known value for the select into other uses of it, and
802 // propagate a known value of the condition into its other users.
804 // If the select and condition only have a single use, don't bother with this,
806 if (SI->use_empty() && SelectCond->hasOneUse())
809 // Scan the current block backward, looking for other uses of SI.
810 BasicBlock::iterator BBI = I.getIterator(), BBFront = I.getParent()->begin();
812 while (BBI != BBFront) {
814 // If we found a call to a function, we can't assume it will return, so
815 // information from below it cannot be propagated above it.
816 if (isa<CallInst>(BBI) && !isa<IntrinsicInst>(BBI))
819 // Replace uses of the select or its condition with the known values.
820 for (Instruction::op_iterator I = BBI->op_begin(), E = BBI->op_end();
823 *I = SI->getOperand(NonNullOperand);
825 } else if (*I == SelectCond) {
826 *I = Builder->getInt1(NonNullOperand == 1);
831 // If we past the instruction, quit looking for it.
834 if (&*BBI == SelectCond)
835 SelectCond = nullptr;
837 // If we ran out of things to eliminate, break out of the loop.
838 if (!SelectCond && !SI)
846 /// This function implements the transforms common to both integer division
847 /// instructions (udiv and sdiv). It is called by the visitors to those integer
848 /// division instructions.
849 /// @brief Common integer divide transforms
850 Instruction *InstCombiner::commonIDivTransforms(BinaryOperator &I) {
851 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
853 // The RHS is known non-zero.
854 if (Value *V = simplifyValueKnownNonZero(I.getOperand(1), *this, I)) {
859 // Handle cases involving: [su]div X, (select Cond, Y, Z)
860 // This does not apply for fdiv.
861 if (isa<SelectInst>(Op1) && SimplifyDivRemOfSelect(I))
864 if (Instruction *LHS = dyn_cast<Instruction>(Op0)) {
866 if (match(Op1, m_APInt(C2))) {
869 bool IsSigned = I.getOpcode() == Instruction::SDiv;
871 // (X / C1) / C2 -> X / (C1*C2)
872 if ((IsSigned && match(LHS, m_SDiv(m_Value(X), m_APInt(C1)))) ||
873 (!IsSigned && match(LHS, m_UDiv(m_Value(X), m_APInt(C1))))) {
874 APInt Product(C1->getBitWidth(), /*Val=*/0ULL, IsSigned);
875 if (!MultiplyOverflows(*C1, *C2, Product, IsSigned))
876 return BinaryOperator::Create(I.getOpcode(), X,
877 ConstantInt::get(I.getType(), Product));
880 if ((IsSigned && match(LHS, m_NSWMul(m_Value(X), m_APInt(C1)))) ||
881 (!IsSigned && match(LHS, m_NUWMul(m_Value(X), m_APInt(C1))))) {
882 APInt Quotient(C1->getBitWidth(), /*Val=*/0ULL, IsSigned);
884 // (X * C1) / C2 -> X / (C2 / C1) if C2 is a multiple of C1.
885 if (IsMultiple(*C2, *C1, Quotient, IsSigned)) {
886 BinaryOperator *BO = BinaryOperator::Create(
887 I.getOpcode(), X, ConstantInt::get(X->getType(), Quotient));
888 BO->setIsExact(I.isExact());
892 // (X * C1) / C2 -> X * (C1 / C2) if C1 is a multiple of C2.
893 if (IsMultiple(*C1, *C2, Quotient, IsSigned)) {
894 BinaryOperator *BO = BinaryOperator::Create(
895 Instruction::Mul, X, ConstantInt::get(X->getType(), Quotient));
896 BO->setHasNoUnsignedWrap(
898 cast<OverflowingBinaryOperator>(LHS)->hasNoUnsignedWrap());
899 BO->setHasNoSignedWrap(
900 cast<OverflowingBinaryOperator>(LHS)->hasNoSignedWrap());
905 if ((IsSigned && match(LHS, m_NSWShl(m_Value(X), m_APInt(C1))) &&
906 *C1 != C1->getBitWidth() - 1) ||
907 (!IsSigned && match(LHS, m_NUWShl(m_Value(X), m_APInt(C1))))) {
908 APInt Quotient(C1->getBitWidth(), /*Val=*/0ULL, IsSigned);
909 APInt C1Shifted = APInt::getOneBitSet(
910 C1->getBitWidth(), static_cast<unsigned>(C1->getLimitedValue()));
912 // (X << C1) / C2 -> X / (C2 >> C1) if C2 is a multiple of C1.
913 if (IsMultiple(*C2, C1Shifted, Quotient, IsSigned)) {
914 BinaryOperator *BO = BinaryOperator::Create(
915 I.getOpcode(), X, ConstantInt::get(X->getType(), Quotient));
916 BO->setIsExact(I.isExact());
920 // (X << C1) / C2 -> X * (C2 >> C1) if C1 is a multiple of C2.
921 if (IsMultiple(C1Shifted, *C2, Quotient, IsSigned)) {
922 BinaryOperator *BO = BinaryOperator::Create(
923 Instruction::Mul, X, ConstantInt::get(X->getType(), Quotient));
924 BO->setHasNoUnsignedWrap(
926 cast<OverflowingBinaryOperator>(LHS)->hasNoUnsignedWrap());
927 BO->setHasNoSignedWrap(
928 cast<OverflowingBinaryOperator>(LHS)->hasNoSignedWrap());
933 if (*C2 != 0) // avoid X udiv 0
934 if (Instruction *FoldedDiv = foldOpWithConstantIntoOperand(I))
939 if (match(Op0, m_One())) {
940 assert(!I.getType()->getScalarType()->isIntegerTy(1) &&
941 "i1 divide not removed?");
942 if (I.getOpcode() == Instruction::SDiv) {
943 // If Op1 is 0 then it's undefined behaviour, if Op1 is 1 then the
944 // result is one, if Op1 is -1 then the result is minus one, otherwise
946 Value *Inc = Builder->CreateAdd(Op1, Op0);
947 Value *Cmp = Builder->CreateICmpULT(
948 Inc, ConstantInt::get(I.getType(), 3));
949 return SelectInst::Create(Cmp, Op1, ConstantInt::get(I.getType(), 0));
951 // If Op1 is 0 then it's undefined behaviour. If Op1 is 1 then the
952 // result is one, otherwise it's zero.
953 return new ZExtInst(Builder->CreateICmpEQ(Op1, Op0), I.getType());
957 // See if we can fold away this div instruction.
958 if (SimplifyDemandedInstructionBits(I))
961 // (X - (X rem Y)) / Y -> X / Y; usually originates as ((X / Y) * Y) / Y
962 Value *X = nullptr, *Z = nullptr;
963 if (match(Op0, m_Sub(m_Value(X), m_Value(Z)))) { // (X - Z) / Y; Y = Op1
964 bool isSigned = I.getOpcode() == Instruction::SDiv;
965 if ((isSigned && match(Z, m_SRem(m_Specific(X), m_Specific(Op1)))) ||
966 (!isSigned && match(Z, m_URem(m_Specific(X), m_Specific(Op1)))))
967 return BinaryOperator::Create(I.getOpcode(), X, Op1);
973 /// dyn_castZExtVal - Checks if V is a zext or constant that can
974 /// be truncated to Ty without losing bits.
975 static Value *dyn_castZExtVal(Value *V, Type *Ty) {
976 if (ZExtInst *Z = dyn_cast<ZExtInst>(V)) {
977 if (Z->getSrcTy() == Ty)
978 return Z->getOperand(0);
979 } else if (ConstantInt *C = dyn_cast<ConstantInt>(V)) {
980 if (C->getValue().getActiveBits() <= cast<IntegerType>(Ty)->getBitWidth())
981 return ConstantExpr::getTrunc(C, Ty);
987 const unsigned MaxDepth = 6;
988 typedef Instruction *(*FoldUDivOperandCb)(Value *Op0, Value *Op1,
989 const BinaryOperator &I,
992 /// \brief Used to maintain state for visitUDivOperand().
993 struct UDivFoldAction {
994 FoldUDivOperandCb FoldAction; ///< Informs visitUDiv() how to fold this
995 ///< operand. This can be zero if this action
996 ///< joins two actions together.
998 Value *OperandToFold; ///< Which operand to fold.
1000 Instruction *FoldResult; ///< The instruction returned when FoldAction is
1003 size_t SelectLHSIdx; ///< Stores the LHS action index if this action
1004 ///< joins two actions together.
1007 UDivFoldAction(FoldUDivOperandCb FA, Value *InputOperand)
1008 : FoldAction(FA), OperandToFold(InputOperand), FoldResult(nullptr) {}
1009 UDivFoldAction(FoldUDivOperandCb FA, Value *InputOperand, size_t SLHS)
1010 : FoldAction(FA), OperandToFold(InputOperand), SelectLHSIdx(SLHS) {}
1014 // X udiv 2^C -> X >> C
1015 static Instruction *foldUDivPow2Cst(Value *Op0, Value *Op1,
1016 const BinaryOperator &I, InstCombiner &IC) {
1017 const APInt &C = cast<Constant>(Op1)->getUniqueInteger();
1018 BinaryOperator *LShr = BinaryOperator::CreateLShr(
1019 Op0, ConstantInt::get(Op0->getType(), C.logBase2()));
1025 // X udiv C, where C >= signbit
1026 static Instruction *foldUDivNegCst(Value *Op0, Value *Op1,
1027 const BinaryOperator &I, InstCombiner &IC) {
1028 Value *ICI = IC.Builder->CreateICmpULT(Op0, cast<ConstantInt>(Op1));
1030 return SelectInst::Create(ICI, Constant::getNullValue(I.getType()),
1031 ConstantInt::get(I.getType(), 1));
1034 // X udiv (C1 << N), where C1 is "1<<C2" --> X >> (N+C2)
1035 // X udiv (zext (C1 << N)), where C1 is "1<<C2" --> X >> (N+C2)
1036 static Instruction *foldUDivShl(Value *Op0, Value *Op1, const BinaryOperator &I,
1039 if (!match(Op1, m_ZExt(m_Value(ShiftLeft))))
1044 if (!match(ShiftLeft, m_Shl(m_APInt(CI), m_Value(N))))
1045 llvm_unreachable("match should never fail here!");
1047 N = IC.Builder->CreateAdd(N,
1048 ConstantInt::get(N->getType(), CI->logBase2()));
1049 if (Op1 != ShiftLeft)
1050 N = IC.Builder->CreateZExt(N, Op1->getType());
1051 BinaryOperator *LShr = BinaryOperator::CreateLShr(Op0, N);
1057 // \brief Recursively visits the possible right hand operands of a udiv
1058 // instruction, seeing through select instructions, to determine if we can
1059 // replace the udiv with something simpler. If we find that an operand is not
1060 // able to simplify the udiv, we abort the entire transformation.
1061 static size_t visitUDivOperand(Value *Op0, Value *Op1, const BinaryOperator &I,
1062 SmallVectorImpl<UDivFoldAction> &Actions,
1063 unsigned Depth = 0) {
1064 // Check to see if this is an unsigned division with an exact power of 2,
1065 // if so, convert to a right shift.
1066 if (match(Op1, m_Power2())) {
1067 Actions.push_back(UDivFoldAction(foldUDivPow2Cst, Op1));
1068 return Actions.size();
1071 if (ConstantInt *C = dyn_cast<ConstantInt>(Op1))
1072 // X udiv C, where C >= signbit
1073 if (C->getValue().isNegative()) {
1074 Actions.push_back(UDivFoldAction(foldUDivNegCst, C));
1075 return Actions.size();
1078 // X udiv (C1 << N), where C1 is "1<<C2" --> X >> (N+C2)
1079 if (match(Op1, m_Shl(m_Power2(), m_Value())) ||
1080 match(Op1, m_ZExt(m_Shl(m_Power2(), m_Value())))) {
1081 Actions.push_back(UDivFoldAction(foldUDivShl, Op1));
1082 return Actions.size();
1085 // The remaining tests are all recursive, so bail out if we hit the limit.
1086 if (Depth++ == MaxDepth)
1089 if (SelectInst *SI = dyn_cast<SelectInst>(Op1))
1091 visitUDivOperand(Op0, SI->getOperand(1), I, Actions, Depth))
1092 if (visitUDivOperand(Op0, SI->getOperand(2), I, Actions, Depth)) {
1093 Actions.push_back(UDivFoldAction(nullptr, Op1, LHSIdx - 1));
1094 return Actions.size();
1100 Instruction *InstCombiner::visitUDiv(BinaryOperator &I) {
1101 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
1103 if (Value *V = SimplifyVectorOp(I))
1104 return replaceInstUsesWith(I, V);
1106 if (Value *V = SimplifyUDivInst(Op0, Op1, SQ))
1107 return replaceInstUsesWith(I, V);
1109 // Handle the integer div common cases
1110 if (Instruction *Common = commonIDivTransforms(I))
1113 // (x lshr C1) udiv C2 --> x udiv (C2 << C1)
1116 const APInt *C1, *C2;
1117 if (match(Op0, m_LShr(m_Value(X), m_APInt(C1))) &&
1118 match(Op1, m_APInt(C2))) {
1120 APInt C2ShlC1 = C2->ushl_ov(*C1, Overflow);
1122 bool IsExact = I.isExact() && match(Op0, m_Exact(m_Value()));
1123 BinaryOperator *BO = BinaryOperator::CreateUDiv(
1124 X, ConstantInt::get(X->getType(), C2ShlC1));
1132 // (zext A) udiv (zext B) --> zext (A udiv B)
1133 if (ZExtInst *ZOp0 = dyn_cast<ZExtInst>(Op0))
1134 if (Value *ZOp1 = dyn_castZExtVal(Op1, ZOp0->getSrcTy()))
1135 return new ZExtInst(
1136 Builder->CreateUDiv(ZOp0->getOperand(0), ZOp1, "div", I.isExact()),
1139 // (LHS udiv (select (select (...)))) -> (LHS >> (select (select (...))))
1140 SmallVector<UDivFoldAction, 6> UDivActions;
1141 if (visitUDivOperand(Op0, Op1, I, UDivActions))
1142 for (unsigned i = 0, e = UDivActions.size(); i != e; ++i) {
1143 FoldUDivOperandCb Action = UDivActions[i].FoldAction;
1144 Value *ActionOp1 = UDivActions[i].OperandToFold;
1147 Inst = Action(Op0, ActionOp1, I, *this);
1149 // This action joins two actions together. The RHS of this action is
1150 // simply the last action we processed, we saved the LHS action index in
1151 // the joining action.
1152 size_t SelectRHSIdx = i - 1;
1153 Value *SelectRHS = UDivActions[SelectRHSIdx].FoldResult;
1154 size_t SelectLHSIdx = UDivActions[i].SelectLHSIdx;
1155 Value *SelectLHS = UDivActions[SelectLHSIdx].FoldResult;
1156 Inst = SelectInst::Create(cast<SelectInst>(ActionOp1)->getCondition(),
1157 SelectLHS, SelectRHS);
1160 // If this is the last action to process, return it to the InstCombiner.
1161 // Otherwise, we insert it before the UDiv and record it so that we may
1162 // use it as part of a joining action (i.e., a SelectInst).
1164 Inst->insertBefore(&I);
1165 UDivActions[i].FoldResult = Inst;
1173 Instruction *InstCombiner::visitSDiv(BinaryOperator &I) {
1174 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
1176 if (Value *V = SimplifyVectorOp(I))
1177 return replaceInstUsesWith(I, V);
1179 if (Value *V = SimplifySDivInst(Op0, Op1, SQ))
1180 return replaceInstUsesWith(I, V);
1182 // Handle the integer div common cases
1183 if (Instruction *Common = commonIDivTransforms(I))
1187 if (match(Op1, m_APInt(Op1C))) {
1189 if (Op1C->isAllOnesValue())
1190 return BinaryOperator::CreateNeg(Op0);
1192 // sdiv exact X, C --> ashr exact X, log2(C)
1193 if (I.isExact() && Op1C->isNonNegative() && Op1C->isPowerOf2()) {
1194 Value *ShAmt = ConstantInt::get(Op1->getType(), Op1C->exactLogBase2());
1195 return BinaryOperator::CreateExactAShr(Op0, ShAmt, I.getName());
1198 // If the dividend is sign-extended and the constant divisor is small enough
1199 // to fit in the source type, shrink the division to the narrower type:
1200 // (sext X) sdiv C --> sext (X sdiv C)
1202 if (match(Op0, m_OneUse(m_SExt(m_Value(Op0Src)))) &&
1203 Op0Src->getType()->getScalarSizeInBits() >= Op1C->getMinSignedBits()) {
1205 // In the general case, we need to make sure that the dividend is not the
1206 // minimum signed value because dividing that by -1 is UB. But here, we
1207 // know that the -1 divisor case is already handled above.
1209 Constant *NarrowDivisor =
1210 ConstantExpr::getTrunc(cast<Constant>(Op1), Op0Src->getType());
1211 Value *NarrowOp = Builder->CreateSDiv(Op0Src, NarrowDivisor);
1212 return new SExtInst(NarrowOp, Op0->getType());
1216 if (Constant *RHS = dyn_cast<Constant>(Op1)) {
1217 // X/INT_MIN -> X == INT_MIN
1218 if (RHS->isMinSignedValue())
1219 return new ZExtInst(Builder->CreateICmpEQ(Op0, Op1), I.getType());
1221 // -X/C --> X/-C provided the negation doesn't overflow.
1223 if (match(Op0, m_NSWSub(m_Zero(), m_Value(X)))) {
1224 auto *BO = BinaryOperator::CreateSDiv(X, ConstantExpr::getNeg(RHS));
1225 BO->setIsExact(I.isExact());
1230 // If the sign bits of both operands are zero (i.e. we can prove they are
1231 // unsigned inputs), turn this into a udiv.
1232 APInt Mask(APInt::getSignMask(I.getType()->getScalarSizeInBits()));
1233 if (MaskedValueIsZero(Op0, Mask, 0, &I)) {
1234 if (MaskedValueIsZero(Op1, Mask, 0, &I)) {
1235 // X sdiv Y -> X udiv Y, iff X and Y don't have sign bit set
1236 auto *BO = BinaryOperator::CreateUDiv(Op0, Op1, I.getName());
1237 BO->setIsExact(I.isExact());
1241 if (isKnownToBeAPowerOfTwo(Op1, /*OrZero*/ true, 0, &I)) {
1242 // X sdiv (1 << Y) -> X udiv (1 << Y) ( -> X u>> Y)
1243 // Safe because the only negative value (1 << Y) can take on is
1244 // INT_MIN, and X sdiv INT_MIN == X udiv INT_MIN == 0 if X doesn't have
1245 // the sign bit set.
1246 auto *BO = BinaryOperator::CreateUDiv(Op0, Op1, I.getName());
1247 BO->setIsExact(I.isExact());
1255 /// CvtFDivConstToReciprocal tries to convert X/C into X*1/C if C not a special
1257 /// 1) 1/C is exact, or
1258 /// 2) reciprocal is allowed.
1259 /// If the conversion was successful, the simplified expression "X * 1/C" is
1260 /// returned; otherwise, NULL is returned.
1262 static Instruction *CvtFDivConstToReciprocal(Value *Dividend, Constant *Divisor,
1263 bool AllowReciprocal) {
1264 if (!isa<ConstantFP>(Divisor)) // TODO: handle vectors.
1267 const APFloat &FpVal = cast<ConstantFP>(Divisor)->getValueAPF();
1268 APFloat Reciprocal(FpVal.getSemantics());
1269 bool Cvt = FpVal.getExactInverse(&Reciprocal);
1271 if (!Cvt && AllowReciprocal && FpVal.isFiniteNonZero()) {
1272 Reciprocal = APFloat(FpVal.getSemantics(), 1.0f);
1273 (void)Reciprocal.divide(FpVal, APFloat::rmNearestTiesToEven);
1274 Cvt = !Reciprocal.isDenormal();
1281 R = ConstantFP::get(Dividend->getType()->getContext(), Reciprocal);
1282 return BinaryOperator::CreateFMul(Dividend, R);
1285 Instruction *InstCombiner::visitFDiv(BinaryOperator &I) {
1286 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
1288 if (Value *V = SimplifyVectorOp(I))
1289 return replaceInstUsesWith(I, V);
1291 if (Value *V = SimplifyFDivInst(Op0, Op1, I.getFastMathFlags(), SQ))
1292 return replaceInstUsesWith(I, V);
1294 if (isa<Constant>(Op0))
1295 if (SelectInst *SI = dyn_cast<SelectInst>(Op1))
1296 if (Instruction *R = FoldOpIntoSelect(I, SI))
1299 bool AllowReassociate = I.hasUnsafeAlgebra();
1300 bool AllowReciprocal = I.hasAllowReciprocal();
1302 if (Constant *Op1C = dyn_cast<Constant>(Op1)) {
1303 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
1304 if (Instruction *R = FoldOpIntoSelect(I, SI))
1307 if (AllowReassociate) {
1308 Constant *C1 = nullptr;
1309 Constant *C2 = Op1C;
1311 Instruction *Res = nullptr;
1313 if (match(Op0, m_FMul(m_Value(X), m_Constant(C1)))) {
1314 // (X*C1)/C2 => X * (C1/C2)
1316 Constant *C = ConstantExpr::getFDiv(C1, C2);
1318 Res = BinaryOperator::CreateFMul(X, C);
1319 } else if (match(Op0, m_FDiv(m_Value(X), m_Constant(C1)))) {
1320 // (X/C1)/C2 => X /(C2*C1) [=> X * 1/(C2*C1) if reciprocal is allowed]
1322 Constant *C = ConstantExpr::getFMul(C1, C2);
1323 if (isNormalFp(C)) {
1324 Res = CvtFDivConstToReciprocal(X, C, AllowReciprocal);
1326 Res = BinaryOperator::CreateFDiv(X, C);
1331 Res->setFastMathFlags(I.getFastMathFlags());
1337 if (Instruction *T = CvtFDivConstToReciprocal(Op0, Op1C, AllowReciprocal)) {
1338 T->copyFastMathFlags(&I);
1345 if (AllowReassociate && isa<Constant>(Op0)) {
1346 Constant *C1 = cast<Constant>(Op0), *C2;
1347 Constant *Fold = nullptr;
1349 bool CreateDiv = true;
1351 // C1 / (X*C2) => (C1/C2) / X
1352 if (match(Op1, m_FMul(m_Value(X), m_Constant(C2))))
1353 Fold = ConstantExpr::getFDiv(C1, C2);
1354 else if (match(Op1, m_FDiv(m_Value(X), m_Constant(C2)))) {
1355 // C1 / (X/C2) => (C1*C2) / X
1356 Fold = ConstantExpr::getFMul(C1, C2);
1357 } else if (match(Op1, m_FDiv(m_Constant(C2), m_Value(X)))) {
1358 // C1 / (C2/X) => (C1/C2) * X
1359 Fold = ConstantExpr::getFDiv(C1, C2);
1363 if (Fold && isNormalFp(Fold)) {
1364 Instruction *R = CreateDiv ? BinaryOperator::CreateFDiv(Fold, X)
1365 : BinaryOperator::CreateFMul(X, Fold);
1366 R->setFastMathFlags(I.getFastMathFlags());
1372 if (AllowReassociate) {
1374 Value *NewInst = nullptr;
1375 Instruction *SimpR = nullptr;
1377 if (Op0->hasOneUse() && match(Op0, m_FDiv(m_Value(X), m_Value(Y)))) {
1378 // (X/Y) / Z => X / (Y*Z)
1380 if (!isa<Constant>(Y) || !isa<Constant>(Op1)) {
1381 NewInst = Builder->CreateFMul(Y, Op1);
1382 if (Instruction *RI = dyn_cast<Instruction>(NewInst)) {
1383 FastMathFlags Flags = I.getFastMathFlags();
1384 Flags &= cast<Instruction>(Op0)->getFastMathFlags();
1385 RI->setFastMathFlags(Flags);
1387 SimpR = BinaryOperator::CreateFDiv(X, NewInst);
1389 } else if (Op1->hasOneUse() && match(Op1, m_FDiv(m_Value(X), m_Value(Y)))) {
1390 // Z / (X/Y) => Z*Y / X
1392 if (!isa<Constant>(Y) || !isa<Constant>(Op0)) {
1393 NewInst = Builder->CreateFMul(Op0, Y);
1394 if (Instruction *RI = dyn_cast<Instruction>(NewInst)) {
1395 FastMathFlags Flags = I.getFastMathFlags();
1396 Flags &= cast<Instruction>(Op1)->getFastMathFlags();
1397 RI->setFastMathFlags(Flags);
1399 SimpR = BinaryOperator::CreateFDiv(NewInst, X);
1404 if (Instruction *T = dyn_cast<Instruction>(NewInst))
1405 T->setDebugLoc(I.getDebugLoc());
1406 SimpR->setFastMathFlags(I.getFastMathFlags());
1415 if (match(Op0, m_FNeg(m_Value(LHS))) && match(Op1, m_FNeg(m_Value(RHS)))) {
1416 I.setOperand(0, LHS);
1417 I.setOperand(1, RHS);
1424 /// This function implements the transforms common to both integer remainder
1425 /// instructions (urem and srem). It is called by the visitors to those integer
1426 /// remainder instructions.
1427 /// @brief Common integer remainder transforms
1428 Instruction *InstCombiner::commonIRemTransforms(BinaryOperator &I) {
1429 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
1431 // The RHS is known non-zero.
1432 if (Value *V = simplifyValueKnownNonZero(I.getOperand(1), *this, I)) {
1437 // Handle cases involving: rem X, (select Cond, Y, Z)
1438 if (isa<SelectInst>(Op1) && SimplifyDivRemOfSelect(I))
1441 if (isa<Constant>(Op1)) {
1442 if (Instruction *Op0I = dyn_cast<Instruction>(Op0)) {
1443 if (SelectInst *SI = dyn_cast<SelectInst>(Op0I)) {
1444 if (Instruction *R = FoldOpIntoSelect(I, SI))
1446 } else if (auto *PN = dyn_cast<PHINode>(Op0I)) {
1447 using namespace llvm::PatternMatch;
1448 const APInt *Op1Int;
1449 if (match(Op1, m_APInt(Op1Int)) && !Op1Int->isMinValue() &&
1450 (I.getOpcode() == Instruction::URem ||
1451 !Op1Int->isMinSignedValue())) {
1452 // foldOpIntoPhi will speculate instructions to the end of the PHI's
1453 // predecessor blocks, so do this only if we know the srem or urem
1455 if (Instruction *NV = foldOpIntoPhi(I, PN))
1460 // See if we can fold away this rem instruction.
1461 if (SimplifyDemandedInstructionBits(I))
1469 Instruction *InstCombiner::visitURem(BinaryOperator &I) {
1470 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
1472 if (Value *V = SimplifyVectorOp(I))
1473 return replaceInstUsesWith(I, V);
1475 if (Value *V = SimplifyURemInst(Op0, Op1, SQ))
1476 return replaceInstUsesWith(I, V);
1478 if (Instruction *common = commonIRemTransforms(I))
1481 // (zext A) urem (zext B) --> zext (A urem B)
1482 if (ZExtInst *ZOp0 = dyn_cast<ZExtInst>(Op0))
1483 if (Value *ZOp1 = dyn_castZExtVal(Op1, ZOp0->getSrcTy()))
1484 return new ZExtInst(Builder->CreateURem(ZOp0->getOperand(0), ZOp1),
1487 // X urem Y -> X and Y-1, where Y is a power of 2,
1488 if (isKnownToBeAPowerOfTwo(Op1, /*OrZero*/ true, 0, &I)) {
1489 Constant *N1 = Constant::getAllOnesValue(I.getType());
1490 Value *Add = Builder->CreateAdd(Op1, N1);
1491 return BinaryOperator::CreateAnd(Op0, Add);
1494 // 1 urem X -> zext(X != 1)
1495 if (match(Op0, m_One())) {
1496 Value *Cmp = Builder->CreateICmpNE(Op1, Op0);
1497 Value *Ext = Builder->CreateZExt(Cmp, I.getType());
1498 return replaceInstUsesWith(I, Ext);
1501 // X urem C -> X < C ? X : X - C, where C >= signbit.
1502 const APInt *DivisorC;
1503 if (match(Op1, m_APInt(DivisorC)) && DivisorC->isNegative()) {
1504 Value *Cmp = Builder->CreateICmpULT(Op0, Op1);
1505 Value *Sub = Builder->CreateSub(Op0, Op1);
1506 return SelectInst::Create(Cmp, Op0, Sub);
1512 Instruction *InstCombiner::visitSRem(BinaryOperator &I) {
1513 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
1515 if (Value *V = SimplifyVectorOp(I))
1516 return replaceInstUsesWith(I, V);
1518 if (Value *V = SimplifySRemInst(Op0, Op1, SQ))
1519 return replaceInstUsesWith(I, V);
1521 // Handle the integer rem common cases
1522 if (Instruction *Common = commonIRemTransforms(I))
1528 if (match(Op1, m_APInt(Y)) && Y->isNegative() && !Y->isMinSignedValue()) {
1529 Worklist.AddValue(I.getOperand(1));
1530 I.setOperand(1, ConstantInt::get(I.getType(), -*Y));
1535 // If the sign bits of both operands are zero (i.e. we can prove they are
1536 // unsigned inputs), turn this into a urem.
1537 APInt Mask(APInt::getSignMask(I.getType()->getScalarSizeInBits()));
1538 if (MaskedValueIsZero(Op1, Mask, 0, &I) &&
1539 MaskedValueIsZero(Op0, Mask, 0, &I)) {
1540 // X srem Y -> X urem Y, iff X and Y don't have sign bit set
1541 return BinaryOperator::CreateURem(Op0, Op1, I.getName());
1544 // If it's a constant vector, flip any negative values positive.
1545 if (isa<ConstantVector>(Op1) || isa<ConstantDataVector>(Op1)) {
1546 Constant *C = cast<Constant>(Op1);
1547 unsigned VWidth = C->getType()->getVectorNumElements();
1549 bool hasNegative = false;
1550 bool hasMissing = false;
1551 for (unsigned i = 0; i != VWidth; ++i) {
1552 Constant *Elt = C->getAggregateElement(i);
1558 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Elt))
1559 if (RHS->isNegative())
1563 if (hasNegative && !hasMissing) {
1564 SmallVector<Constant *, 16> Elts(VWidth);
1565 for (unsigned i = 0; i != VWidth; ++i) {
1566 Elts[i] = C->getAggregateElement(i); // Handle undef, etc.
1567 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Elts[i])) {
1568 if (RHS->isNegative())
1569 Elts[i] = cast<ConstantInt>(ConstantExpr::getNeg(RHS));
1573 Constant *NewRHSV = ConstantVector::get(Elts);
1574 if (NewRHSV != C) { // Don't loop on -MININT
1575 Worklist.AddValue(I.getOperand(1));
1576 I.setOperand(1, NewRHSV);
1585 Instruction *InstCombiner::visitFRem(BinaryOperator &I) {
1586 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
1588 if (Value *V = SimplifyVectorOp(I))
1589 return replaceInstUsesWith(I, V);
1591 if (Value *V = SimplifyFRemInst(Op0, Op1, I.getFastMathFlags(), SQ))
1592 return replaceInstUsesWith(I, V);
1594 // Handle cases involving: rem X, (select Cond, Y, Z)
1595 if (isa<SelectInst>(Op1) && SimplifyDivRemOfSelect(I))