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 isKnownToBeAPowerOfTwo(I->getOperand(0), IC.getDataLayout(), false, 0,
51 &IC.getAssumptionCache(), &CxtI,
52 &IC.getDominatorTree())) {
53 // We know that this is an exact/nuw shift and that the input is a
54 // non-zero context as well.
55 if (Value *V2 = simplifyValueKnownNonZero(I->getOperand(0), IC, CxtI)) {
60 if (I->getOpcode() == Instruction::LShr && !I->isExact()) {
65 if (I->getOpcode() == Instruction::Shl && !I->hasNoUnsignedWrap()) {
66 I->setHasNoUnsignedWrap();
71 // TODO: Lots more we could do here:
72 // If V is a phi node, we can call this on each of its operands.
73 // "select cond, X, 0" can simplify to "X".
75 return MadeChange ? V : nullptr;
79 /// True if the multiply can not be expressed in an int this size.
80 static bool MultiplyOverflows(const APInt &C1, const APInt &C2, APInt &Product,
84 Product = C1.smul_ov(C2, Overflow);
86 Product = C1.umul_ov(C2, Overflow);
91 /// \brief True if C2 is a multiple of C1. Quotient contains C2/C1.
92 static bool IsMultiple(const APInt &C1, const APInt &C2, APInt &Quotient,
94 assert(C1.getBitWidth() == C2.getBitWidth() &&
95 "Inconsistent width of constants!");
97 // Bail if we will divide by zero.
101 // Bail if we would divide INT_MIN by -1.
102 if (IsSigned && C1.isMinSignedValue() && C2.isAllOnesValue())
105 APInt Remainder(C1.getBitWidth(), /*Val=*/0ULL, IsSigned);
107 APInt::sdivrem(C1, C2, Quotient, Remainder);
109 APInt::udivrem(C1, C2, Quotient, Remainder);
111 return Remainder.isMinValue();
114 /// \brief A helper routine of InstCombiner::visitMul().
116 /// If C is a vector of known powers of 2, then this function returns
117 /// a new vector obtained from C replacing each element with its logBase2.
118 /// Return a null pointer otherwise.
119 static Constant *getLogBase2Vector(ConstantDataVector *CV) {
121 SmallVector<Constant *, 4> Elts;
123 for (unsigned I = 0, E = CV->getNumElements(); I != E; ++I) {
124 Constant *Elt = CV->getElementAsConstant(I);
125 if (!match(Elt, m_APInt(IVal)) || !IVal->isPowerOf2())
127 Elts.push_back(ConstantInt::get(Elt->getType(), IVal->logBase2()));
130 return ConstantVector::get(Elts);
133 /// \brief Return true if we can prove that:
134 /// (mul LHS, RHS) === (mul nsw LHS, RHS)
135 bool InstCombiner::WillNotOverflowSignedMul(Value *LHS, Value *RHS,
137 // Multiplying n * m significant bits yields a result of n + m significant
138 // bits. If the total number of significant bits does not exceed the
139 // result bit width (minus 1), there is no overflow.
140 // This means if we have enough leading sign bits in the operands
141 // we can guarantee that the result does not overflow.
142 // Ref: "Hacker's Delight" by Henry Warren
143 unsigned BitWidth = LHS->getType()->getScalarSizeInBits();
145 // Note that underestimating the number of sign bits gives a more
146 // conservative answer.
148 ComputeNumSignBits(LHS, 0, &CxtI) + ComputeNumSignBits(RHS, 0, &CxtI);
150 // First handle the easy case: if we have enough sign bits there's
151 // definitely no overflow.
152 if (SignBits > BitWidth + 1)
155 // There are two ambiguous cases where there can be no overflow:
156 // SignBits == BitWidth + 1 and
157 // SignBits == BitWidth
158 // The second case is difficult to check, therefore we only handle the
160 if (SignBits == BitWidth + 1) {
161 // It overflows only when both arguments are negative and the true
162 // product is exactly the minimum negative number.
163 // E.g. mul i16 with 17 sign bits: 0xff00 * 0xff80 = 0x8000
164 // For simplicity we just check if at least one side is not negative.
165 bool LHSNonNegative, LHSNegative;
166 bool RHSNonNegative, RHSNegative;
167 ComputeSignBit(LHS, LHSNonNegative, LHSNegative, /*Depth=*/0, &CxtI);
168 ComputeSignBit(RHS, RHSNonNegative, RHSNegative, /*Depth=*/0, &CxtI);
169 if (LHSNonNegative || RHSNonNegative)
175 Instruction *InstCombiner::visitMul(BinaryOperator &I) {
176 bool Changed = SimplifyAssociativeOrCommutative(I);
177 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
179 if (Value *V = SimplifyVectorOp(I))
180 return replaceInstUsesWith(I, V);
182 if (Value *V = SimplifyMulInst(Op0, Op1, DL, &TLI, &DT, &AC))
183 return replaceInstUsesWith(I, V);
185 if (Value *V = SimplifyUsingDistributiveLaws(I))
186 return replaceInstUsesWith(I, V);
189 if (match(Op1, m_AllOnes())) {
190 BinaryOperator *BO = BinaryOperator::CreateNeg(Op0, I.getName());
191 if (I.hasNoSignedWrap())
192 BO->setHasNoSignedWrap();
196 // Also allow combining multiply instructions on vectors.
201 if (match(&I, m_Mul(m_Shl(m_Value(NewOp), m_Constant(C2)),
203 match(C1, m_APInt(IVal))) {
204 // ((X << C2)*C1) == (X * (C1 << C2))
205 Constant *Shl = ConstantExpr::getShl(C1, C2);
206 BinaryOperator *Mul = cast<BinaryOperator>(I.getOperand(0));
207 BinaryOperator *BO = BinaryOperator::CreateMul(NewOp, Shl);
208 if (I.hasNoUnsignedWrap() && Mul->hasNoUnsignedWrap())
209 BO->setHasNoUnsignedWrap();
210 if (I.hasNoSignedWrap() && Mul->hasNoSignedWrap() &&
211 Shl->isNotMinSignedValue())
212 BO->setHasNoSignedWrap();
216 if (match(&I, m_Mul(m_Value(NewOp), m_Constant(C1)))) {
217 Constant *NewCst = nullptr;
218 if (match(C1, m_APInt(IVal)) && IVal->isPowerOf2())
219 // Replace X*(2^C) with X << C, where C is either a scalar or a splat.
220 NewCst = ConstantInt::get(NewOp->getType(), IVal->logBase2());
221 else if (ConstantDataVector *CV = dyn_cast<ConstantDataVector>(C1))
222 // Replace X*(2^C) with X << C, where C is a vector of known
223 // constant powers of 2.
224 NewCst = getLogBase2Vector(CV);
227 unsigned Width = NewCst->getType()->getPrimitiveSizeInBits();
228 BinaryOperator *Shl = BinaryOperator::CreateShl(NewOp, NewCst);
230 if (I.hasNoUnsignedWrap())
231 Shl->setHasNoUnsignedWrap();
232 if (I.hasNoSignedWrap()) {
234 if (match(NewCst, m_ConstantInt(V)) && V != Width - 1)
235 Shl->setHasNoSignedWrap();
243 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
244 // (Y - X) * (-(2**n)) -> (X - Y) * (2**n), for positive nonzero n
245 // (Y + const) * (-(2**n)) -> (-constY) * (2**n), for positive nonzero n
246 // The "* (2**n)" thus becomes a potential shifting opportunity.
248 const APInt & Val = CI->getValue();
249 const APInt &PosVal = Val.abs();
250 if (Val.isNegative() && PosVal.isPowerOf2()) {
251 Value *X = nullptr, *Y = nullptr;
252 if (Op0->hasOneUse()) {
254 Value *Sub = nullptr;
255 if (match(Op0, m_Sub(m_Value(Y), m_Value(X))))
256 Sub = Builder->CreateSub(X, Y, "suba");
257 else if (match(Op0, m_Add(m_Value(Y), m_ConstantInt(C1))))
258 Sub = Builder->CreateSub(Builder->CreateNeg(C1), Y, "subc");
261 BinaryOperator::CreateMul(Sub,
262 ConstantInt::get(Y->getType(), PosVal));
268 // Simplify mul instructions with a constant RHS.
269 if (isa<Constant>(Op1)) {
270 if (Instruction *FoldedMul = foldOpWithConstantIntoOperand(I))
273 // Canonicalize (X+C1)*CI -> X*CI+C1*CI.
277 if (match(Op0, m_OneUse(m_Add(m_Value(X), m_Constant(C1))))) {
278 Value *Mul = Builder->CreateMul(C1, Op1);
279 // Only go forward with the transform if C1*CI simplifies to a tidier
281 if (!match(Mul, m_Mul(m_Value(), m_Value())))
282 return BinaryOperator::CreateAdd(Builder->CreateMul(X, Op1), Mul);
287 if (Value *Op0v = dyn_castNegVal(Op0)) { // -X * -Y = X*Y
288 if (Value *Op1v = dyn_castNegVal(Op1)) {
289 BinaryOperator *BO = BinaryOperator::CreateMul(Op0v, Op1v);
290 if (I.hasNoSignedWrap() &&
291 match(Op0, m_NSWSub(m_Value(), m_Value())) &&
292 match(Op1, m_NSWSub(m_Value(), m_Value())))
293 BO->setHasNoSignedWrap();
298 // (X / Y) * Y = X - (X % Y)
299 // (X / Y) * -Y = (X % Y) - X
302 BinaryOperator *BO = dyn_cast<BinaryOperator>(Op0);
304 (BO->getOpcode() != Instruction::UDiv &&
305 BO->getOpcode() != Instruction::SDiv)) {
307 BO = dyn_cast<BinaryOperator>(Op1);
309 Value *Neg = dyn_castNegVal(Op1C);
310 if (BO && BO->hasOneUse() &&
311 (BO->getOperand(1) == Op1C || BO->getOperand(1) == Neg) &&
312 (BO->getOpcode() == Instruction::UDiv ||
313 BO->getOpcode() == Instruction::SDiv)) {
314 Value *Op0BO = BO->getOperand(0), *Op1BO = BO->getOperand(1);
316 // If the division is exact, X % Y is zero, so we end up with X or -X.
317 if (PossiblyExactOperator *SDiv = dyn_cast<PossiblyExactOperator>(BO))
318 if (SDiv->isExact()) {
320 return replaceInstUsesWith(I, Op0BO);
321 return BinaryOperator::CreateNeg(Op0BO);
325 if (BO->getOpcode() == Instruction::UDiv)
326 Rem = Builder->CreateURem(Op0BO, Op1BO);
328 Rem = Builder->CreateSRem(Op0BO, Op1BO);
332 return BinaryOperator::CreateSub(Op0BO, Rem);
333 return BinaryOperator::CreateSub(Rem, Op0BO);
337 /// i1 mul -> i1 and.
338 if (I.getType()->getScalarType()->isIntegerTy(1))
339 return BinaryOperator::CreateAnd(Op0, Op1);
341 // X*(1 << Y) --> X << Y
342 // (1 << Y)*X --> X << Y
345 BinaryOperator *BO = nullptr;
347 if (match(Op0, m_Shl(m_One(), m_Value(Y)))) {
348 BO = BinaryOperator::CreateShl(Op1, Y);
349 ShlNSW = cast<ShlOperator>(Op0)->hasNoSignedWrap();
350 } else if (match(Op1, m_Shl(m_One(), m_Value(Y)))) {
351 BO = BinaryOperator::CreateShl(Op0, Y);
352 ShlNSW = cast<ShlOperator>(Op1)->hasNoSignedWrap();
355 if (I.hasNoUnsignedWrap())
356 BO->setHasNoUnsignedWrap();
357 if (I.hasNoSignedWrap() && ShlNSW)
358 BO->setHasNoSignedWrap();
363 // If one of the operands of the multiply is a cast from a boolean value, then
364 // we know the bool is either zero or one, so this is a 'masking' multiply.
365 // X * Y (where Y is 0 or 1) -> X & (0-Y)
366 if (!I.getType()->isVectorTy()) {
367 // -2 is "-1 << 1" so it is all bits set except the low one.
368 APInt Negative2(I.getType()->getPrimitiveSizeInBits(), (uint64_t)-2, true);
370 Value *BoolCast = nullptr, *OtherOp = nullptr;
371 if (MaskedValueIsZero(Op0, Negative2, 0, &I)) {
374 } else if (MaskedValueIsZero(Op1, Negative2, 0, &I)) {
380 Value *V = Builder->CreateSub(Constant::getNullValue(I.getType()),
382 return BinaryOperator::CreateAnd(V, OtherOp);
386 // Check for (mul (sext x), y), see if we can merge this into an
387 // integer mul followed by a sext.
388 if (SExtInst *Op0Conv = dyn_cast<SExtInst>(Op0)) {
389 // (mul (sext x), cst) --> (sext (mul x, cst'))
390 if (ConstantInt *Op1C = dyn_cast<ConstantInt>(Op1)) {
391 if (Op0Conv->hasOneUse()) {
393 ConstantExpr::getTrunc(Op1C, Op0Conv->getOperand(0)->getType());
394 if (ConstantExpr::getSExt(CI, I.getType()) == Op1C &&
395 WillNotOverflowSignedMul(Op0Conv->getOperand(0), CI, I)) {
396 // Insert the new, smaller mul.
398 Builder->CreateNSWMul(Op0Conv->getOperand(0), CI, "mulconv");
399 return new SExtInst(NewMul, I.getType());
404 // (mul (sext x), (sext y)) --> (sext (mul int x, y))
405 if (SExtInst *Op1Conv = dyn_cast<SExtInst>(Op1)) {
406 // Only do this if x/y have the same type, if at last one of them has a
407 // single use (so we don't increase the number of sexts), and if the
408 // integer mul will not overflow.
409 if (Op0Conv->getOperand(0)->getType() ==
410 Op1Conv->getOperand(0)->getType() &&
411 (Op0Conv->hasOneUse() || Op1Conv->hasOneUse()) &&
412 WillNotOverflowSignedMul(Op0Conv->getOperand(0),
413 Op1Conv->getOperand(0), I)) {
414 // Insert the new integer mul.
415 Value *NewMul = Builder->CreateNSWMul(
416 Op0Conv->getOperand(0), Op1Conv->getOperand(0), "mulconv");
417 return new SExtInst(NewMul, I.getType());
422 // Check for (mul (zext x), y), see if we can merge this into an
423 // integer mul followed by a zext.
424 if (auto *Op0Conv = dyn_cast<ZExtInst>(Op0)) {
425 // (mul (zext x), cst) --> (zext (mul x, cst'))
426 if (ConstantInt *Op1C = dyn_cast<ConstantInt>(Op1)) {
427 if (Op0Conv->hasOneUse()) {
429 ConstantExpr::getTrunc(Op1C, Op0Conv->getOperand(0)->getType());
430 if (ConstantExpr::getZExt(CI, I.getType()) == Op1C &&
431 computeOverflowForUnsignedMul(Op0Conv->getOperand(0), CI, &I) ==
432 OverflowResult::NeverOverflows) {
433 // Insert the new, smaller mul.
435 Builder->CreateNUWMul(Op0Conv->getOperand(0), CI, "mulconv");
436 return new ZExtInst(NewMul, I.getType());
441 // (mul (zext x), (zext y)) --> (zext (mul int x, y))
442 if (auto *Op1Conv = dyn_cast<ZExtInst>(Op1)) {
443 // Only do this if x/y have the same type, if at last one of them has a
444 // single use (so we don't increase the number of zexts), and if the
445 // integer mul will not overflow.
446 if (Op0Conv->getOperand(0)->getType() ==
447 Op1Conv->getOperand(0)->getType() &&
448 (Op0Conv->hasOneUse() || Op1Conv->hasOneUse()) &&
449 computeOverflowForUnsignedMul(Op0Conv->getOperand(0),
450 Op1Conv->getOperand(0),
451 &I) == OverflowResult::NeverOverflows) {
452 // Insert the new integer mul.
453 Value *NewMul = Builder->CreateNUWMul(
454 Op0Conv->getOperand(0), Op1Conv->getOperand(0), "mulconv");
455 return new ZExtInst(NewMul, I.getType());
460 if (!I.hasNoSignedWrap() && WillNotOverflowSignedMul(Op0, Op1, I)) {
462 I.setHasNoSignedWrap(true);
465 if (!I.hasNoUnsignedWrap() &&
466 computeOverflowForUnsignedMul(Op0, Op1, &I) ==
467 OverflowResult::NeverOverflows) {
469 I.setHasNoUnsignedWrap(true);
472 return Changed ? &I : nullptr;
475 /// Detect pattern log2(Y * 0.5) with corresponding fast math flags.
476 static void detectLog2OfHalf(Value *&Op, Value *&Y, IntrinsicInst *&Log2) {
477 if (!Op->hasOneUse())
480 IntrinsicInst *II = dyn_cast<IntrinsicInst>(Op);
483 if (II->getIntrinsicID() != Intrinsic::log2 || !II->hasUnsafeAlgebra())
487 Value *OpLog2Of = II->getArgOperand(0);
488 if (!OpLog2Of->hasOneUse())
491 Instruction *I = dyn_cast<Instruction>(OpLog2Of);
494 if (I->getOpcode() != Instruction::FMul || !I->hasUnsafeAlgebra())
497 if (match(I->getOperand(0), m_SpecificFP(0.5)))
498 Y = I->getOperand(1);
499 else if (match(I->getOperand(1), m_SpecificFP(0.5)))
500 Y = I->getOperand(0);
503 static bool isFiniteNonZeroFp(Constant *C) {
504 if (C->getType()->isVectorTy()) {
505 for (unsigned I = 0, E = C->getType()->getVectorNumElements(); I != E;
507 ConstantFP *CFP = dyn_cast_or_null<ConstantFP>(C->getAggregateElement(I));
508 if (!CFP || !CFP->getValueAPF().isFiniteNonZero())
514 return isa<ConstantFP>(C) &&
515 cast<ConstantFP>(C)->getValueAPF().isFiniteNonZero();
518 static bool isNormalFp(Constant *C) {
519 if (C->getType()->isVectorTy()) {
520 for (unsigned I = 0, E = C->getType()->getVectorNumElements(); I != E;
522 ConstantFP *CFP = dyn_cast_or_null<ConstantFP>(C->getAggregateElement(I));
523 if (!CFP || !CFP->getValueAPF().isNormal())
529 return isa<ConstantFP>(C) && cast<ConstantFP>(C)->getValueAPF().isNormal();
532 /// Helper function of InstCombiner::visitFMul(BinaryOperator(). It returns
533 /// true iff the given value is FMul or FDiv with one and only one operand
534 /// being a normal constant (i.e. not Zero/NaN/Infinity).
535 static bool isFMulOrFDivWithConstant(Value *V) {
536 Instruction *I = dyn_cast<Instruction>(V);
537 if (!I || (I->getOpcode() != Instruction::FMul &&
538 I->getOpcode() != Instruction::FDiv))
541 Constant *C0 = dyn_cast<Constant>(I->getOperand(0));
542 Constant *C1 = dyn_cast<Constant>(I->getOperand(1));
547 return (C0 && isFiniteNonZeroFp(C0)) || (C1 && isFiniteNonZeroFp(C1));
550 /// foldFMulConst() is a helper routine of InstCombiner::visitFMul().
551 /// The input \p FMulOrDiv is a FMul/FDiv with one and only one operand
552 /// being a constant (i.e. isFMulOrFDivWithConstant(FMulOrDiv) == true).
553 /// This function is to simplify "FMulOrDiv * C" and returns the
554 /// resulting expression. Note that this function could return NULL in
555 /// case the constants cannot be folded into a normal floating-point.
557 Value *InstCombiner::foldFMulConst(Instruction *FMulOrDiv, Constant *C,
558 Instruction *InsertBefore) {
559 assert(isFMulOrFDivWithConstant(FMulOrDiv) && "V is invalid");
561 Value *Opnd0 = FMulOrDiv->getOperand(0);
562 Value *Opnd1 = FMulOrDiv->getOperand(1);
564 Constant *C0 = dyn_cast<Constant>(Opnd0);
565 Constant *C1 = dyn_cast<Constant>(Opnd1);
567 BinaryOperator *R = nullptr;
569 // (X * C0) * C => X * (C0*C)
570 if (FMulOrDiv->getOpcode() == Instruction::FMul) {
571 Constant *F = ConstantExpr::getFMul(C1 ? C1 : C0, C);
573 R = BinaryOperator::CreateFMul(C1 ? Opnd0 : Opnd1, F);
576 // (C0 / X) * C => (C0 * C) / X
577 if (FMulOrDiv->hasOneUse()) {
578 // It would otherwise introduce another div.
579 Constant *F = ConstantExpr::getFMul(C0, C);
581 R = BinaryOperator::CreateFDiv(F, Opnd1);
584 // (X / C1) * C => X * (C/C1) if C/C1 is not a denormal
585 Constant *F = ConstantExpr::getFDiv(C, C1);
587 R = BinaryOperator::CreateFMul(Opnd0, F);
589 // (X / C1) * C => X / (C1/C)
590 Constant *F = ConstantExpr::getFDiv(C1, C);
592 R = BinaryOperator::CreateFDiv(Opnd0, F);
598 R->setHasUnsafeAlgebra(true);
599 InsertNewInstWith(R, *InsertBefore);
605 Instruction *InstCombiner::visitFMul(BinaryOperator &I) {
606 bool Changed = SimplifyAssociativeOrCommutative(I);
607 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
609 if (Value *V = SimplifyVectorOp(I))
610 return replaceInstUsesWith(I, V);
612 if (isa<Constant>(Op0))
616 SimplifyFMulInst(Op0, Op1, I.getFastMathFlags(), DL, &TLI, &DT, &AC))
617 return replaceInstUsesWith(I, V);
619 bool AllowReassociate = I.hasUnsafeAlgebra();
621 // Simplify mul instructions with a constant RHS.
622 if (isa<Constant>(Op1)) {
623 if (Instruction *FoldedMul = foldOpWithConstantIntoOperand(I))
626 // (fmul X, -1.0) --> (fsub -0.0, X)
627 if (match(Op1, m_SpecificFP(-1.0))) {
628 Constant *NegZero = ConstantFP::getNegativeZero(Op1->getType());
629 Instruction *RI = BinaryOperator::CreateFSub(NegZero, Op0);
630 RI->copyFastMathFlags(&I);
634 Constant *C = cast<Constant>(Op1);
635 if (AllowReassociate && isFiniteNonZeroFp(C)) {
636 // Let MDC denote an expression in one of these forms:
637 // X * C, C/X, X/C, where C is a constant.
639 // Try to simplify "MDC * Constant"
640 if (isFMulOrFDivWithConstant(Op0))
641 if (Value *V = foldFMulConst(cast<Instruction>(Op0), C, &I))
642 return replaceInstUsesWith(I, V);
644 // (MDC +/- C1) * C => (MDC * C) +/- (C1 * C)
645 Instruction *FAddSub = dyn_cast<Instruction>(Op0);
647 (FAddSub->getOpcode() == Instruction::FAdd ||
648 FAddSub->getOpcode() == Instruction::FSub)) {
649 Value *Opnd0 = FAddSub->getOperand(0);
650 Value *Opnd1 = FAddSub->getOperand(1);
651 Constant *C0 = dyn_cast<Constant>(Opnd0);
652 Constant *C1 = dyn_cast<Constant>(Opnd1);
656 std::swap(Opnd0, Opnd1);
660 if (C1 && isFiniteNonZeroFp(C1) && isFMulOrFDivWithConstant(Opnd0)) {
661 Value *M1 = ConstantExpr::getFMul(C1, C);
662 Value *M0 = isNormalFp(cast<Constant>(M1)) ?
663 foldFMulConst(cast<Instruction>(Opnd0), C, &I) :
666 if (Swap && FAddSub->getOpcode() == Instruction::FSub)
669 Instruction *RI = (FAddSub->getOpcode() == Instruction::FAdd)
670 ? BinaryOperator::CreateFAdd(M0, M1)
671 : BinaryOperator::CreateFSub(M0, M1);
672 RI->copyFastMathFlags(&I);
681 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(Op0)) {
682 // sqrt(X) * sqrt(X) -> X
683 if (AllowReassociate && II->getIntrinsicID() == Intrinsic::sqrt)
684 return replaceInstUsesWith(I, II->getOperand(0));
686 // fabs(X) * fabs(X) -> X * X
687 if (II->getIntrinsicID() == Intrinsic::fabs) {
688 Instruction *FMulVal = BinaryOperator::CreateFMul(II->getOperand(0),
691 FMulVal->copyFastMathFlags(&I);
697 // Under unsafe algebra do:
698 // X * log2(0.5*Y) = X*log2(Y) - X
699 if (AllowReassociate) {
700 Value *OpX = nullptr;
701 Value *OpY = nullptr;
703 detectLog2OfHalf(Op0, OpY, Log2);
707 detectLog2OfHalf(Op1, OpY, Log2);
712 // if pattern detected emit alternate sequence
714 BuilderTy::FastMathFlagGuard Guard(*Builder);
715 Builder->setFastMathFlags(Log2->getFastMathFlags());
716 Log2->setArgOperand(0, OpY);
717 Value *FMulVal = Builder->CreateFMul(OpX, Log2);
718 Value *FSub = Builder->CreateFSub(FMulVal, OpX);
720 return replaceInstUsesWith(I, FSub);
724 // Handle symmetric situation in a 2-iteration loop
727 for (int i = 0; i < 2; i++) {
728 bool IgnoreZeroSign = I.hasNoSignedZeros();
729 if (BinaryOperator::isFNeg(Opnd0, IgnoreZeroSign)) {
730 BuilderTy::FastMathFlagGuard Guard(*Builder);
731 Builder->setFastMathFlags(I.getFastMathFlags());
733 Value *N0 = dyn_castFNegVal(Opnd0, IgnoreZeroSign);
734 Value *N1 = dyn_castFNegVal(Opnd1, IgnoreZeroSign);
738 Value *FMul = Builder->CreateFMul(N0, N1);
740 return replaceInstUsesWith(I, FMul);
743 if (Opnd0->hasOneUse()) {
744 // -X * Y => -(X*Y) (Promote negation as high as possible)
745 Value *T = Builder->CreateFMul(N0, Opnd1);
746 Value *Neg = Builder->CreateFNeg(T);
748 return replaceInstUsesWith(I, Neg);
752 // (X*Y) * X => (X*X) * Y where Y != X
753 // The purpose is two-fold:
754 // 1) to form a power expression (of X).
755 // 2) potentially shorten the critical path: After transformation, the
756 // latency of the instruction Y is amortized by the expression of X*X,
757 // and therefore Y is in a "less critical" position compared to what it
758 // was before the transformation.
760 if (AllowReassociate) {
761 Value *Opnd0_0, *Opnd0_1;
762 if (Opnd0->hasOneUse() &&
763 match(Opnd0, m_FMul(m_Value(Opnd0_0), m_Value(Opnd0_1)))) {
765 if (Opnd0_0 == Opnd1 && Opnd0_1 != Opnd1)
767 else if (Opnd0_1 == Opnd1 && Opnd0_0 != Opnd1)
771 BuilderTy::FastMathFlagGuard Guard(*Builder);
772 Builder->setFastMathFlags(I.getFastMathFlags());
773 Value *T = Builder->CreateFMul(Opnd1, Opnd1);
774 Value *R = Builder->CreateFMul(T, Y);
776 return replaceInstUsesWith(I, R);
781 if (!isa<Constant>(Op1))
782 std::swap(Opnd0, Opnd1);
787 return Changed ? &I : nullptr;
790 /// Try to fold a divide or remainder of a select instruction.
791 bool InstCombiner::SimplifyDivRemOfSelect(BinaryOperator &I) {
792 SelectInst *SI = cast<SelectInst>(I.getOperand(1));
794 // div/rem X, (Cond ? 0 : Y) -> div/rem X, Y
795 int NonNullOperand = -1;
796 if (Constant *ST = dyn_cast<Constant>(SI->getOperand(1)))
797 if (ST->isNullValue())
799 // div/rem X, (Cond ? Y : 0) -> div/rem X, Y
800 if (Constant *ST = dyn_cast<Constant>(SI->getOperand(2)))
801 if (ST->isNullValue())
804 if (NonNullOperand == -1)
807 Value *SelectCond = SI->getOperand(0);
809 // Change the div/rem to use 'Y' instead of the select.
810 I.setOperand(1, SI->getOperand(NonNullOperand));
812 // Okay, we know we replace the operand of the div/rem with 'Y' with no
813 // problem. However, the select, or the condition of the select may have
814 // multiple uses. Based on our knowledge that the operand must be non-zero,
815 // propagate the known value for the select into other uses of it, and
816 // propagate a known value of the condition into its other users.
818 // If the select and condition only have a single use, don't bother with this,
820 if (SI->use_empty() && SelectCond->hasOneUse())
823 // Scan the current block backward, looking for other uses of SI.
824 BasicBlock::iterator BBI = I.getIterator(), BBFront = I.getParent()->begin();
826 while (BBI != BBFront) {
828 // If we found a call to a function, we can't assume it will return, so
829 // information from below it cannot be propagated above it.
830 if (isa<CallInst>(BBI) && !isa<IntrinsicInst>(BBI))
833 // Replace uses of the select or its condition with the known values.
834 for (Instruction::op_iterator I = BBI->op_begin(), E = BBI->op_end();
837 *I = SI->getOperand(NonNullOperand);
839 } else if (*I == SelectCond) {
840 *I = Builder->getInt1(NonNullOperand == 1);
845 // If we past the instruction, quit looking for it.
848 if (&*BBI == SelectCond)
849 SelectCond = nullptr;
851 // If we ran out of things to eliminate, break out of the loop.
852 if (!SelectCond && !SI)
860 /// This function implements the transforms common to both integer division
861 /// instructions (udiv and sdiv). It is called by the visitors to those integer
862 /// division instructions.
863 /// @brief Common integer divide transforms
864 Instruction *InstCombiner::commonIDivTransforms(BinaryOperator &I) {
865 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
867 // The RHS is known non-zero.
868 if (Value *V = simplifyValueKnownNonZero(I.getOperand(1), *this, I)) {
873 // Handle cases involving: [su]div X, (select Cond, Y, Z)
874 // This does not apply for fdiv.
875 if (isa<SelectInst>(Op1) && SimplifyDivRemOfSelect(I))
878 if (Instruction *LHS = dyn_cast<Instruction>(Op0)) {
880 if (match(Op1, m_APInt(C2))) {
883 bool IsSigned = I.getOpcode() == Instruction::SDiv;
885 // (X / C1) / C2 -> X / (C1*C2)
886 if ((IsSigned && match(LHS, m_SDiv(m_Value(X), m_APInt(C1)))) ||
887 (!IsSigned && match(LHS, m_UDiv(m_Value(X), m_APInt(C1))))) {
888 APInt Product(C1->getBitWidth(), /*Val=*/0ULL, IsSigned);
889 if (!MultiplyOverflows(*C1, *C2, Product, IsSigned))
890 return BinaryOperator::Create(I.getOpcode(), X,
891 ConstantInt::get(I.getType(), Product));
894 if ((IsSigned && match(LHS, m_NSWMul(m_Value(X), m_APInt(C1)))) ||
895 (!IsSigned && match(LHS, m_NUWMul(m_Value(X), m_APInt(C1))))) {
896 APInt Quotient(C1->getBitWidth(), /*Val=*/0ULL, IsSigned);
898 // (X * C1) / C2 -> X / (C2 / C1) if C2 is a multiple of C1.
899 if (IsMultiple(*C2, *C1, Quotient, IsSigned)) {
900 BinaryOperator *BO = BinaryOperator::Create(
901 I.getOpcode(), X, ConstantInt::get(X->getType(), Quotient));
902 BO->setIsExact(I.isExact());
906 // (X * C1) / C2 -> X * (C1 / C2) if C1 is a multiple of C2.
907 if (IsMultiple(*C1, *C2, Quotient, IsSigned)) {
908 BinaryOperator *BO = BinaryOperator::Create(
909 Instruction::Mul, X, ConstantInt::get(X->getType(), Quotient));
910 BO->setHasNoUnsignedWrap(
912 cast<OverflowingBinaryOperator>(LHS)->hasNoUnsignedWrap());
913 BO->setHasNoSignedWrap(
914 cast<OverflowingBinaryOperator>(LHS)->hasNoSignedWrap());
919 if ((IsSigned && match(LHS, m_NSWShl(m_Value(X), m_APInt(C1))) &&
920 *C1 != C1->getBitWidth() - 1) ||
921 (!IsSigned && match(LHS, m_NUWShl(m_Value(X), m_APInt(C1))))) {
922 APInt Quotient(C1->getBitWidth(), /*Val=*/0ULL, IsSigned);
923 APInt C1Shifted = APInt::getOneBitSet(
924 C1->getBitWidth(), static_cast<unsigned>(C1->getLimitedValue()));
926 // (X << C1) / C2 -> X / (C2 >> C1) if C2 is a multiple of C1.
927 if (IsMultiple(*C2, C1Shifted, Quotient, IsSigned)) {
928 BinaryOperator *BO = BinaryOperator::Create(
929 I.getOpcode(), X, ConstantInt::get(X->getType(), Quotient));
930 BO->setIsExact(I.isExact());
934 // (X << C1) / C2 -> X * (C2 >> C1) if C1 is a multiple of C2.
935 if (IsMultiple(C1Shifted, *C2, Quotient, IsSigned)) {
936 BinaryOperator *BO = BinaryOperator::Create(
937 Instruction::Mul, X, ConstantInt::get(X->getType(), Quotient));
938 BO->setHasNoUnsignedWrap(
940 cast<OverflowingBinaryOperator>(LHS)->hasNoUnsignedWrap());
941 BO->setHasNoSignedWrap(
942 cast<OverflowingBinaryOperator>(LHS)->hasNoSignedWrap());
947 if (*C2 != 0) // avoid X udiv 0
948 if (Instruction *FoldedDiv = foldOpWithConstantIntoOperand(I))
953 if (ConstantInt *One = dyn_cast<ConstantInt>(Op0)) {
954 if (One->isOne() && !I.getType()->isIntegerTy(1)) {
955 bool isSigned = I.getOpcode() == Instruction::SDiv;
957 // If Op1 is 0 then it's undefined behaviour, if Op1 is 1 then the
958 // result is one, if Op1 is -1 then the result is minus one, otherwise
960 Value *Inc = Builder->CreateAdd(Op1, One);
961 Value *Cmp = Builder->CreateICmpULT(
962 Inc, ConstantInt::get(I.getType(), 3));
963 return SelectInst::Create(Cmp, Op1, ConstantInt::get(I.getType(), 0));
965 // If Op1 is 0 then it's undefined behaviour. If Op1 is 1 then the
966 // result is one, otherwise it's zero.
967 return new ZExtInst(Builder->CreateICmpEQ(Op1, One), I.getType());
972 // See if we can fold away this div instruction.
973 if (SimplifyDemandedInstructionBits(I))
976 // (X - (X rem Y)) / Y -> X / Y; usually originates as ((X / Y) * Y) / Y
977 Value *X = nullptr, *Z = nullptr;
978 if (match(Op0, m_Sub(m_Value(X), m_Value(Z)))) { // (X - Z) / Y; Y = Op1
979 bool isSigned = I.getOpcode() == Instruction::SDiv;
980 if ((isSigned && match(Z, m_SRem(m_Specific(X), m_Specific(Op1)))) ||
981 (!isSigned && match(Z, m_URem(m_Specific(X), m_Specific(Op1)))))
982 return BinaryOperator::Create(I.getOpcode(), X, Op1);
988 /// dyn_castZExtVal - Checks if V is a zext or constant that can
989 /// be truncated to Ty without losing bits.
990 static Value *dyn_castZExtVal(Value *V, Type *Ty) {
991 if (ZExtInst *Z = dyn_cast<ZExtInst>(V)) {
992 if (Z->getSrcTy() == Ty)
993 return Z->getOperand(0);
994 } else if (ConstantInt *C = dyn_cast<ConstantInt>(V)) {
995 if (C->getValue().getActiveBits() <= cast<IntegerType>(Ty)->getBitWidth())
996 return ConstantExpr::getTrunc(C, Ty);
1002 const unsigned MaxDepth = 6;
1003 typedef Instruction *(*FoldUDivOperandCb)(Value *Op0, Value *Op1,
1004 const BinaryOperator &I,
1007 /// \brief Used to maintain state for visitUDivOperand().
1008 struct UDivFoldAction {
1009 FoldUDivOperandCb FoldAction; ///< Informs visitUDiv() how to fold this
1010 ///< operand. This can be zero if this action
1011 ///< joins two actions together.
1013 Value *OperandToFold; ///< Which operand to fold.
1015 Instruction *FoldResult; ///< The instruction returned when FoldAction is
1018 size_t SelectLHSIdx; ///< Stores the LHS action index if this action
1019 ///< joins two actions together.
1022 UDivFoldAction(FoldUDivOperandCb FA, Value *InputOperand)
1023 : FoldAction(FA), OperandToFold(InputOperand), FoldResult(nullptr) {}
1024 UDivFoldAction(FoldUDivOperandCb FA, Value *InputOperand, size_t SLHS)
1025 : FoldAction(FA), OperandToFold(InputOperand), SelectLHSIdx(SLHS) {}
1029 // X udiv 2^C -> X >> C
1030 static Instruction *foldUDivPow2Cst(Value *Op0, Value *Op1,
1031 const BinaryOperator &I, InstCombiner &IC) {
1032 const APInt &C = cast<Constant>(Op1)->getUniqueInteger();
1033 BinaryOperator *LShr = BinaryOperator::CreateLShr(
1034 Op0, ConstantInt::get(Op0->getType(), C.logBase2()));
1040 // X udiv C, where C >= signbit
1041 static Instruction *foldUDivNegCst(Value *Op0, Value *Op1,
1042 const BinaryOperator &I, InstCombiner &IC) {
1043 Value *ICI = IC.Builder->CreateICmpULT(Op0, cast<ConstantInt>(Op1));
1045 return SelectInst::Create(ICI, Constant::getNullValue(I.getType()),
1046 ConstantInt::get(I.getType(), 1));
1049 // X udiv (C1 << N), where C1 is "1<<C2" --> X >> (N+C2)
1050 // X udiv (zext (C1 << N)), where C1 is "1<<C2" --> X >> (N+C2)
1051 static Instruction *foldUDivShl(Value *Op0, Value *Op1, const BinaryOperator &I,
1054 if (!match(Op1, m_ZExt(m_Value(ShiftLeft))))
1059 if (!match(ShiftLeft, m_Shl(m_APInt(CI), m_Value(N))))
1060 llvm_unreachable("match should never fail here!");
1062 N = IC.Builder->CreateAdd(N,
1063 ConstantInt::get(N->getType(), CI->logBase2()));
1064 if (Op1 != ShiftLeft)
1065 N = IC.Builder->CreateZExt(N, Op1->getType());
1066 BinaryOperator *LShr = BinaryOperator::CreateLShr(Op0, N);
1072 // \brief Recursively visits the possible right hand operands of a udiv
1073 // instruction, seeing through select instructions, to determine if we can
1074 // replace the udiv with something simpler. If we find that an operand is not
1075 // able to simplify the udiv, we abort the entire transformation.
1076 static size_t visitUDivOperand(Value *Op0, Value *Op1, const BinaryOperator &I,
1077 SmallVectorImpl<UDivFoldAction> &Actions,
1078 unsigned Depth = 0) {
1079 // Check to see if this is an unsigned division with an exact power of 2,
1080 // if so, convert to a right shift.
1081 if (match(Op1, m_Power2())) {
1082 Actions.push_back(UDivFoldAction(foldUDivPow2Cst, Op1));
1083 return Actions.size();
1086 if (ConstantInt *C = dyn_cast<ConstantInt>(Op1))
1087 // X udiv C, where C >= signbit
1088 if (C->getValue().isNegative()) {
1089 Actions.push_back(UDivFoldAction(foldUDivNegCst, C));
1090 return Actions.size();
1093 // X udiv (C1 << N), where C1 is "1<<C2" --> X >> (N+C2)
1094 if (match(Op1, m_Shl(m_Power2(), m_Value())) ||
1095 match(Op1, m_ZExt(m_Shl(m_Power2(), m_Value())))) {
1096 Actions.push_back(UDivFoldAction(foldUDivShl, Op1));
1097 return Actions.size();
1100 // The remaining tests are all recursive, so bail out if we hit the limit.
1101 if (Depth++ == MaxDepth)
1104 if (SelectInst *SI = dyn_cast<SelectInst>(Op1))
1106 visitUDivOperand(Op0, SI->getOperand(1), I, Actions, Depth))
1107 if (visitUDivOperand(Op0, SI->getOperand(2), I, Actions, Depth)) {
1108 Actions.push_back(UDivFoldAction(nullptr, Op1, LHSIdx - 1));
1109 return Actions.size();
1115 Instruction *InstCombiner::visitUDiv(BinaryOperator &I) {
1116 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
1118 if (Value *V = SimplifyVectorOp(I))
1119 return replaceInstUsesWith(I, V);
1121 if (Value *V = SimplifyUDivInst(Op0, Op1, DL, &TLI, &DT, &AC))
1122 return replaceInstUsesWith(I, V);
1124 // Handle the integer div common cases
1125 if (Instruction *Common = commonIDivTransforms(I))
1128 // (x lshr C1) udiv C2 --> x udiv (C2 << C1)
1131 const APInt *C1, *C2;
1132 if (match(Op0, m_LShr(m_Value(X), m_APInt(C1))) &&
1133 match(Op1, m_APInt(C2))) {
1135 APInt C2ShlC1 = C2->ushl_ov(*C1, Overflow);
1137 bool IsExact = I.isExact() && match(Op0, m_Exact(m_Value()));
1138 BinaryOperator *BO = BinaryOperator::CreateUDiv(
1139 X, ConstantInt::get(X->getType(), C2ShlC1));
1147 // (zext A) udiv (zext B) --> zext (A udiv B)
1148 if (ZExtInst *ZOp0 = dyn_cast<ZExtInst>(Op0))
1149 if (Value *ZOp1 = dyn_castZExtVal(Op1, ZOp0->getSrcTy()))
1150 return new ZExtInst(
1151 Builder->CreateUDiv(ZOp0->getOperand(0), ZOp1, "div", I.isExact()),
1154 // (LHS udiv (select (select (...)))) -> (LHS >> (select (select (...))))
1155 SmallVector<UDivFoldAction, 6> UDivActions;
1156 if (visitUDivOperand(Op0, Op1, I, UDivActions))
1157 for (unsigned i = 0, e = UDivActions.size(); i != e; ++i) {
1158 FoldUDivOperandCb Action = UDivActions[i].FoldAction;
1159 Value *ActionOp1 = UDivActions[i].OperandToFold;
1162 Inst = Action(Op0, ActionOp1, I, *this);
1164 // This action joins two actions together. The RHS of this action is
1165 // simply the last action we processed, we saved the LHS action index in
1166 // the joining action.
1167 size_t SelectRHSIdx = i - 1;
1168 Value *SelectRHS = UDivActions[SelectRHSIdx].FoldResult;
1169 size_t SelectLHSIdx = UDivActions[i].SelectLHSIdx;
1170 Value *SelectLHS = UDivActions[SelectLHSIdx].FoldResult;
1171 Inst = SelectInst::Create(cast<SelectInst>(ActionOp1)->getCondition(),
1172 SelectLHS, SelectRHS);
1175 // If this is the last action to process, return it to the InstCombiner.
1176 // Otherwise, we insert it before the UDiv and record it so that we may
1177 // use it as part of a joining action (i.e., a SelectInst).
1179 Inst->insertBefore(&I);
1180 UDivActions[i].FoldResult = Inst;
1188 Instruction *InstCombiner::visitSDiv(BinaryOperator &I) {
1189 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
1191 if (Value *V = SimplifyVectorOp(I))
1192 return replaceInstUsesWith(I, V);
1194 if (Value *V = SimplifySDivInst(Op0, Op1, DL, &TLI, &DT, &AC))
1195 return replaceInstUsesWith(I, V);
1197 // Handle the integer div common cases
1198 if (Instruction *Common = commonIDivTransforms(I))
1202 if (match(Op1, m_APInt(Op1C))) {
1204 if (Op1C->isAllOnesValue())
1205 return BinaryOperator::CreateNeg(Op0);
1207 // sdiv exact X, C --> ashr exact X, log2(C)
1208 if (I.isExact() && Op1C->isNonNegative() && Op1C->isPowerOf2()) {
1209 Value *ShAmt = ConstantInt::get(Op1->getType(), Op1C->exactLogBase2());
1210 return BinaryOperator::CreateExactAShr(Op0, ShAmt, I.getName());
1213 // If the dividend is sign-extended and the constant divisor is small enough
1214 // to fit in the source type, shrink the division to the narrower type:
1215 // (sext X) sdiv C --> sext (X sdiv C)
1217 if (match(Op0, m_OneUse(m_SExt(m_Value(Op0Src)))) &&
1218 Op0Src->getType()->getScalarSizeInBits() >= Op1C->getMinSignedBits()) {
1220 // In the general case, we need to make sure that the dividend is not the
1221 // minimum signed value because dividing that by -1 is UB. But here, we
1222 // know that the -1 divisor case is already handled above.
1224 Constant *NarrowDivisor =
1225 ConstantExpr::getTrunc(cast<Constant>(Op1), Op0Src->getType());
1226 Value *NarrowOp = Builder->CreateSDiv(Op0Src, NarrowDivisor);
1227 return new SExtInst(NarrowOp, Op0->getType());
1231 if (Constant *RHS = dyn_cast<Constant>(Op1)) {
1232 // X/INT_MIN -> X == INT_MIN
1233 if (RHS->isMinSignedValue())
1234 return new ZExtInst(Builder->CreateICmpEQ(Op0, Op1), I.getType());
1236 // -X/C --> X/-C provided the negation doesn't overflow.
1238 if (match(Op0, m_NSWSub(m_Zero(), m_Value(X)))) {
1239 auto *BO = BinaryOperator::CreateSDiv(X, ConstantExpr::getNeg(RHS));
1240 BO->setIsExact(I.isExact());
1245 // If the sign bits of both operands are zero (i.e. we can prove they are
1246 // unsigned inputs), turn this into a udiv.
1247 if (I.getType()->isIntegerTy()) {
1248 APInt Mask(APInt::getSignBit(I.getType()->getPrimitiveSizeInBits()));
1249 if (MaskedValueIsZero(Op0, Mask, 0, &I)) {
1250 if (MaskedValueIsZero(Op1, Mask, 0, &I)) {
1251 // X sdiv Y -> X udiv Y, iff X and Y don't have sign bit set
1252 auto *BO = BinaryOperator::CreateUDiv(Op0, Op1, I.getName());
1253 BO->setIsExact(I.isExact());
1257 if (isKnownToBeAPowerOfTwo(Op1, DL, /*OrZero*/ true, 0, &AC, &I, &DT)) {
1258 // X sdiv (1 << Y) -> X udiv (1 << Y) ( -> X u>> Y)
1259 // Safe because the only negative value (1 << Y) can take on is
1260 // INT_MIN, and X sdiv INT_MIN == X udiv INT_MIN == 0 if X doesn't have
1261 // the sign bit set.
1262 auto *BO = BinaryOperator::CreateUDiv(Op0, Op1, I.getName());
1263 BO->setIsExact(I.isExact());
1272 /// CvtFDivConstToReciprocal tries to convert X/C into X*1/C if C not a special
1274 /// 1) 1/C is exact, or
1275 /// 2) reciprocal is allowed.
1276 /// If the conversion was successful, the simplified expression "X * 1/C" is
1277 /// returned; otherwise, NULL is returned.
1279 static Instruction *CvtFDivConstToReciprocal(Value *Dividend, Constant *Divisor,
1280 bool AllowReciprocal) {
1281 if (!isa<ConstantFP>(Divisor)) // TODO: handle vectors.
1284 const APFloat &FpVal = cast<ConstantFP>(Divisor)->getValueAPF();
1285 APFloat Reciprocal(FpVal.getSemantics());
1286 bool Cvt = FpVal.getExactInverse(&Reciprocal);
1288 if (!Cvt && AllowReciprocal && FpVal.isFiniteNonZero()) {
1289 Reciprocal = APFloat(FpVal.getSemantics(), 1.0f);
1290 (void)Reciprocal.divide(FpVal, APFloat::rmNearestTiesToEven);
1291 Cvt = !Reciprocal.isDenormal();
1298 R = ConstantFP::get(Dividend->getType()->getContext(), Reciprocal);
1299 return BinaryOperator::CreateFMul(Dividend, R);
1302 Instruction *InstCombiner::visitFDiv(BinaryOperator &I) {
1303 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
1305 if (Value *V = SimplifyVectorOp(I))
1306 return replaceInstUsesWith(I, V);
1308 if (Value *V = SimplifyFDivInst(Op0, Op1, I.getFastMathFlags(),
1309 DL, &TLI, &DT, &AC))
1310 return replaceInstUsesWith(I, V);
1312 if (isa<Constant>(Op0))
1313 if (SelectInst *SI = dyn_cast<SelectInst>(Op1))
1314 if (Instruction *R = FoldOpIntoSelect(I, SI))
1317 bool AllowReassociate = I.hasUnsafeAlgebra();
1318 bool AllowReciprocal = I.hasAllowReciprocal();
1320 if (Constant *Op1C = dyn_cast<Constant>(Op1)) {
1321 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
1322 if (Instruction *R = FoldOpIntoSelect(I, SI))
1325 if (AllowReassociate) {
1326 Constant *C1 = nullptr;
1327 Constant *C2 = Op1C;
1329 Instruction *Res = nullptr;
1331 if (match(Op0, m_FMul(m_Value(X), m_Constant(C1)))) {
1332 // (X*C1)/C2 => X * (C1/C2)
1334 Constant *C = ConstantExpr::getFDiv(C1, C2);
1336 Res = BinaryOperator::CreateFMul(X, C);
1337 } else if (match(Op0, m_FDiv(m_Value(X), m_Constant(C1)))) {
1338 // (X/C1)/C2 => X /(C2*C1) [=> X * 1/(C2*C1) if reciprocal is allowed]
1340 Constant *C = ConstantExpr::getFMul(C1, C2);
1341 if (isNormalFp(C)) {
1342 Res = CvtFDivConstToReciprocal(X, C, AllowReciprocal);
1344 Res = BinaryOperator::CreateFDiv(X, C);
1349 Res->setFastMathFlags(I.getFastMathFlags());
1355 if (Instruction *T = CvtFDivConstToReciprocal(Op0, Op1C, AllowReciprocal)) {
1356 T->copyFastMathFlags(&I);
1363 if (AllowReassociate && isa<Constant>(Op0)) {
1364 Constant *C1 = cast<Constant>(Op0), *C2;
1365 Constant *Fold = nullptr;
1367 bool CreateDiv = true;
1369 // C1 / (X*C2) => (C1/C2) / X
1370 if (match(Op1, m_FMul(m_Value(X), m_Constant(C2))))
1371 Fold = ConstantExpr::getFDiv(C1, C2);
1372 else if (match(Op1, m_FDiv(m_Value(X), m_Constant(C2)))) {
1373 // C1 / (X/C2) => (C1*C2) / X
1374 Fold = ConstantExpr::getFMul(C1, C2);
1375 } else if (match(Op1, m_FDiv(m_Constant(C2), m_Value(X)))) {
1376 // C1 / (C2/X) => (C1/C2) * X
1377 Fold = ConstantExpr::getFDiv(C1, C2);
1381 if (Fold && isNormalFp(Fold)) {
1382 Instruction *R = CreateDiv ? BinaryOperator::CreateFDiv(Fold, X)
1383 : BinaryOperator::CreateFMul(X, Fold);
1384 R->setFastMathFlags(I.getFastMathFlags());
1390 if (AllowReassociate) {
1392 Value *NewInst = nullptr;
1393 Instruction *SimpR = nullptr;
1395 if (Op0->hasOneUse() && match(Op0, m_FDiv(m_Value(X), m_Value(Y)))) {
1396 // (X/Y) / Z => X / (Y*Z)
1398 if (!isa<Constant>(Y) || !isa<Constant>(Op1)) {
1399 NewInst = Builder->CreateFMul(Y, Op1);
1400 if (Instruction *RI = dyn_cast<Instruction>(NewInst)) {
1401 FastMathFlags Flags = I.getFastMathFlags();
1402 Flags &= cast<Instruction>(Op0)->getFastMathFlags();
1403 RI->setFastMathFlags(Flags);
1405 SimpR = BinaryOperator::CreateFDiv(X, NewInst);
1407 } else if (Op1->hasOneUse() && match(Op1, m_FDiv(m_Value(X), m_Value(Y)))) {
1408 // Z / (X/Y) => Z*Y / X
1410 if (!isa<Constant>(Y) || !isa<Constant>(Op0)) {
1411 NewInst = Builder->CreateFMul(Op0, Y);
1412 if (Instruction *RI = dyn_cast<Instruction>(NewInst)) {
1413 FastMathFlags Flags = I.getFastMathFlags();
1414 Flags &= cast<Instruction>(Op1)->getFastMathFlags();
1415 RI->setFastMathFlags(Flags);
1417 SimpR = BinaryOperator::CreateFDiv(NewInst, X);
1422 if (Instruction *T = dyn_cast<Instruction>(NewInst))
1423 T->setDebugLoc(I.getDebugLoc());
1424 SimpR->setFastMathFlags(I.getFastMathFlags());
1433 if (match(Op0, m_FNeg(m_Value(LHS))) && match(Op1, m_FNeg(m_Value(RHS)))) {
1434 I.setOperand(0, LHS);
1435 I.setOperand(1, RHS);
1442 /// This function implements the transforms common to both integer remainder
1443 /// instructions (urem and srem). It is called by the visitors to those integer
1444 /// remainder instructions.
1445 /// @brief Common integer remainder transforms
1446 Instruction *InstCombiner::commonIRemTransforms(BinaryOperator &I) {
1447 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
1449 // The RHS is known non-zero.
1450 if (Value *V = simplifyValueKnownNonZero(I.getOperand(1), *this, I)) {
1455 // Handle cases involving: rem X, (select Cond, Y, Z)
1456 if (isa<SelectInst>(Op1) && SimplifyDivRemOfSelect(I))
1459 if (isa<Constant>(Op1)) {
1460 if (Instruction *Op0I = dyn_cast<Instruction>(Op0)) {
1461 if (SelectInst *SI = dyn_cast<SelectInst>(Op0I)) {
1462 if (Instruction *R = FoldOpIntoSelect(I, SI))
1464 } else if (isa<PHINode>(Op0I)) {
1465 using namespace llvm::PatternMatch;
1466 const APInt *Op1Int;
1467 if (match(Op1, m_APInt(Op1Int)) && !Op1Int->isMinValue() &&
1468 (I.getOpcode() == Instruction::URem ||
1469 !Op1Int->isMinSignedValue())) {
1470 // FoldOpIntoPhi will speculate instructions to the end of the PHI's
1471 // predecessor blocks, so do this only if we know the srem or urem
1473 if (Instruction *NV = FoldOpIntoPhi(I))
1478 // See if we can fold away this rem instruction.
1479 if (SimplifyDemandedInstructionBits(I))
1487 Instruction *InstCombiner::visitURem(BinaryOperator &I) {
1488 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
1490 if (Value *V = SimplifyVectorOp(I))
1491 return replaceInstUsesWith(I, V);
1493 if (Value *V = SimplifyURemInst(Op0, Op1, DL, &TLI, &DT, &AC))
1494 return replaceInstUsesWith(I, V);
1496 if (Instruction *common = commonIRemTransforms(I))
1499 // (zext A) urem (zext B) --> zext (A urem B)
1500 if (ZExtInst *ZOp0 = dyn_cast<ZExtInst>(Op0))
1501 if (Value *ZOp1 = dyn_castZExtVal(Op1, ZOp0->getSrcTy()))
1502 return new ZExtInst(Builder->CreateURem(ZOp0->getOperand(0), ZOp1),
1505 // X urem Y -> X and Y-1, where Y is a power of 2,
1506 if (isKnownToBeAPowerOfTwo(Op1, DL, /*OrZero*/ true, 0, &AC, &I, &DT)) {
1507 Constant *N1 = Constant::getAllOnesValue(I.getType());
1508 Value *Add = Builder->CreateAdd(Op1, N1);
1509 return BinaryOperator::CreateAnd(Op0, Add);
1512 // 1 urem X -> zext(X != 1)
1513 if (match(Op0, m_One())) {
1514 Value *Cmp = Builder->CreateICmpNE(Op1, Op0);
1515 Value *Ext = Builder->CreateZExt(Cmp, I.getType());
1516 return replaceInstUsesWith(I, Ext);
1519 // X urem C -> X < C ? X : X - C, where C >= signbit.
1520 const APInt *DivisorC;
1521 if (match(Op1, m_APInt(DivisorC)) && DivisorC->isNegative()) {
1522 Value *Cmp = Builder->CreateICmpULT(Op0, Op1);
1523 Value *Sub = Builder->CreateSub(Op0, Op1);
1524 return SelectInst::Create(Cmp, Op0, Sub);
1530 Instruction *InstCombiner::visitSRem(BinaryOperator &I) {
1531 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
1533 if (Value *V = SimplifyVectorOp(I))
1534 return replaceInstUsesWith(I, V);
1536 if (Value *V = SimplifySRemInst(Op0, Op1, DL, &TLI, &DT, &AC))
1537 return replaceInstUsesWith(I, V);
1539 // Handle the integer rem common cases
1540 if (Instruction *Common = commonIRemTransforms(I))
1546 if (match(Op1, m_APInt(Y)) && Y->isNegative() && !Y->isMinSignedValue()) {
1547 Worklist.AddValue(I.getOperand(1));
1548 I.setOperand(1, ConstantInt::get(I.getType(), -*Y));
1553 // If the sign bits of both operands are zero (i.e. we can prove they are
1554 // unsigned inputs), turn this into a urem.
1555 if (I.getType()->isIntegerTy()) {
1556 APInt Mask(APInt::getSignBit(I.getType()->getPrimitiveSizeInBits()));
1557 if (MaskedValueIsZero(Op1, Mask, 0, &I) &&
1558 MaskedValueIsZero(Op0, Mask, 0, &I)) {
1559 // X srem Y -> X urem Y, iff X and Y don't have sign bit set
1560 return BinaryOperator::CreateURem(Op0, Op1, I.getName());
1564 // If it's a constant vector, flip any negative values positive.
1565 if (isa<ConstantVector>(Op1) || isa<ConstantDataVector>(Op1)) {
1566 Constant *C = cast<Constant>(Op1);
1567 unsigned VWidth = C->getType()->getVectorNumElements();
1569 bool hasNegative = false;
1570 bool hasMissing = false;
1571 for (unsigned i = 0; i != VWidth; ++i) {
1572 Constant *Elt = C->getAggregateElement(i);
1578 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Elt))
1579 if (RHS->isNegative())
1583 if (hasNegative && !hasMissing) {
1584 SmallVector<Constant *, 16> Elts(VWidth);
1585 for (unsigned i = 0; i != VWidth; ++i) {
1586 Elts[i] = C->getAggregateElement(i); // Handle undef, etc.
1587 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Elts[i])) {
1588 if (RHS->isNegative())
1589 Elts[i] = cast<ConstantInt>(ConstantExpr::getNeg(RHS));
1593 Constant *NewRHSV = ConstantVector::get(Elts);
1594 if (NewRHSV != C) { // Don't loop on -MININT
1595 Worklist.AddValue(I.getOperand(1));
1596 I.setOperand(1, NewRHSV);
1605 Instruction *InstCombiner::visitFRem(BinaryOperator &I) {
1606 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
1608 if (Value *V = SimplifyVectorOp(I))
1609 return replaceInstUsesWith(I, V);
1611 if (Value *V = SimplifyFRemInst(Op0, Op1, I.getFastMathFlags(),
1612 DL, &TLI, &DT, &AC))
1613 return replaceInstUsesWith(I, V);
1615 // Handle cases involving: rem X, (select Cond, Y, Z)
1616 if (isa<SelectInst>(Op1) && SimplifyDivRemOfSelect(I))