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 "InstCombine.h"
16 #include "llvm/Analysis/InstructionSimplify.h"
17 #include "llvm/IR/IntrinsicInst.h"
18 #include "llvm/Support/PatternMatch.h"
20 using namespace PatternMatch;
23 /// simplifyValueKnownNonZero - The specific integer value is used in a context
24 /// where it is known to be non-zero. If this allows us to simplify the
25 /// computation, do so and return the new operand, otherwise return null.
26 static Value *simplifyValueKnownNonZero(Value *V, InstCombiner &IC) {
27 // If V has multiple uses, then we would have to do more analysis to determine
28 // if this is safe. For example, the use could be in dynamically unreached
30 if (!V->hasOneUse()) return 0;
32 bool MadeChange = false;
34 // ((1 << A) >>u B) --> (1 << (A-B))
35 // Because V cannot be zero, we know that B is less than A.
36 Value *A = 0, *B = 0, *PowerOf2 = 0;
37 if (match(V, m_LShr(m_OneUse(m_Shl(m_Value(PowerOf2), m_Value(A))),
39 // The "1" can be any value known to be a power of 2.
40 isKnownToBeAPowerOfTwo(PowerOf2)) {
41 A = IC.Builder->CreateSub(A, B);
42 return IC.Builder->CreateShl(PowerOf2, A);
45 // (PowerOfTwo >>u B) --> isExact since shifting out the result would make it
46 // inexact. Similarly for <<.
47 if (BinaryOperator *I = dyn_cast<BinaryOperator>(V))
48 if (I->isLogicalShift() && isKnownToBeAPowerOfTwo(I->getOperand(0))) {
49 // We know that this is an exact/nuw shift and that the input is a
50 // non-zero context as well.
51 if (Value *V2 = simplifyValueKnownNonZero(I->getOperand(0), IC)) {
56 if (I->getOpcode() == Instruction::LShr && !I->isExact()) {
61 if (I->getOpcode() == Instruction::Shl && !I->hasNoUnsignedWrap()) {
62 I->setHasNoUnsignedWrap();
67 // TODO: Lots more we could do here:
68 // If V is a phi node, we can call this on each of its operands.
69 // "select cond, X, 0" can simplify to "X".
71 return MadeChange ? V : 0;
75 /// MultiplyOverflows - True if the multiply can not be expressed in an int
77 static bool MultiplyOverflows(ConstantInt *C1, ConstantInt *C2, bool sign) {
78 uint32_t W = C1->getBitWidth();
79 APInt LHSExt = C1->getValue(), RHSExt = C2->getValue();
81 LHSExt = LHSExt.sext(W * 2);
82 RHSExt = RHSExt.sext(W * 2);
84 LHSExt = LHSExt.zext(W * 2);
85 RHSExt = RHSExt.zext(W * 2);
88 APInt MulExt = LHSExt * RHSExt;
91 return MulExt.ugt(APInt::getLowBitsSet(W * 2, W));
93 APInt Min = APInt::getSignedMinValue(W).sext(W * 2);
94 APInt Max = APInt::getSignedMaxValue(W).sext(W * 2);
95 return MulExt.slt(Min) || MulExt.sgt(Max);
98 /// \brief A helper routine of InstCombiner::visitMul().
100 /// If C is a vector of known powers of 2, then this function returns
101 /// a new vector obtained from C replacing each element with its logBase2.
102 /// Return a null pointer otherwise.
103 static Constant *getLogBase2Vector(ConstantDataVector *CV) {
105 SmallVector<Constant *, 4> Elts;
107 for (unsigned I = 0, E = CV->getNumElements(); I != E; ++I) {
108 Constant *Elt = CV->getElementAsConstant(I);
109 if (!match(Elt, m_APInt(IVal)) || !IVal->isPowerOf2())
111 Elts.push_back(ConstantInt::get(Elt->getType(), IVal->logBase2()));
114 return ConstantVector::get(Elts);
117 Instruction *InstCombiner::visitMul(BinaryOperator &I) {
118 bool Changed = SimplifyAssociativeOrCommutative(I);
119 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
121 if (Value *V = SimplifyMulInst(Op0, Op1, TD))
122 return ReplaceInstUsesWith(I, V);
124 if (Value *V = SimplifyUsingDistributiveLaws(I))
125 return ReplaceInstUsesWith(I, V);
127 if (match(Op1, m_AllOnes())) // X * -1 == 0 - X
128 return BinaryOperator::CreateNeg(Op0, I.getName());
130 // Also allow combining multiply instructions on vectors.
135 if (match(&I, m_Mul(m_Shl(m_Value(NewOp), m_Constant(C2)),
137 match(C1, m_APInt(IVal)))
138 // ((X << C1)*C2) == (X * (C2 << C1))
139 return BinaryOperator::CreateMul(NewOp, ConstantExpr::getShl(C1, C2));
141 if (match(&I, m_Mul(m_Value(NewOp), m_Constant(C1)))) {
142 Constant *NewCst = 0;
143 if (match(C1, m_APInt(IVal)) && IVal->isPowerOf2())
144 // Replace X*(2^C) with X << C, where C is either a scalar or a splat.
145 NewCst = ConstantInt::get(NewOp->getType(), IVal->logBase2());
146 else if (ConstantDataVector *CV = dyn_cast<ConstantDataVector>(C1))
147 // Replace X*(2^C) with X << C, where C is a vector of known
148 // constant powers of 2.
149 NewCst = getLogBase2Vector(CV);
152 BinaryOperator *Shl = BinaryOperator::CreateShl(NewOp, NewCst);
153 if (I.hasNoSignedWrap()) Shl->setHasNoSignedWrap();
154 if (I.hasNoUnsignedWrap()) Shl->setHasNoUnsignedWrap();
160 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
161 // Canonicalize (X+C1)*CI -> X*CI+C1*CI.
162 { Value *X; ConstantInt *C1;
163 if (Op0->hasOneUse() &&
164 match(Op0, m_Add(m_Value(X), m_ConstantInt(C1)))) {
165 Value *Add = Builder->CreateMul(X, CI);
166 return BinaryOperator::CreateAdd(Add, Builder->CreateMul(C1, CI));
170 // (Y - X) * (-(2**n)) -> (X - Y) * (2**n), for positive nonzero n
171 // (Y + const) * (-(2**n)) -> (-constY) * (2**n), for positive nonzero n
172 // The "* (2**n)" thus becomes a potential shifting opportunity.
174 const APInt & Val = CI->getValue();
175 const APInt &PosVal = Val.abs();
176 if (Val.isNegative() && PosVal.isPowerOf2()) {
177 Value *X = 0, *Y = 0;
178 if (Op0->hasOneUse()) {
181 if (match(Op0, m_Sub(m_Value(Y), m_Value(X))))
182 Sub = Builder->CreateSub(X, Y, "suba");
183 else if (match(Op0, m_Add(m_Value(Y), m_ConstantInt(C1))))
184 Sub = Builder->CreateSub(Builder->CreateNeg(C1), Y, "subc");
187 BinaryOperator::CreateMul(Sub,
188 ConstantInt::get(Y->getType(), PosVal));
194 // Simplify mul instructions with a constant RHS.
195 if (isa<Constant>(Op1)) {
196 // Try to fold constant mul into select arguments.
197 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
198 if (Instruction *R = FoldOpIntoSelect(I, SI))
201 if (isa<PHINode>(Op0))
202 if (Instruction *NV = FoldOpIntoPhi(I))
206 if (Value *Op0v = dyn_castNegVal(Op0)) // -X * -Y = X*Y
207 if (Value *Op1v = dyn_castNegVal(Op1))
208 return BinaryOperator::CreateMul(Op0v, Op1v);
210 // (X / Y) * Y = X - (X % Y)
211 // (X / Y) * -Y = (X % Y) - X
214 BinaryOperator *BO = dyn_cast<BinaryOperator>(Op0);
216 (BO->getOpcode() != Instruction::UDiv &&
217 BO->getOpcode() != Instruction::SDiv)) {
219 BO = dyn_cast<BinaryOperator>(Op1);
221 Value *Neg = dyn_castNegVal(Op1C);
222 if (BO && BO->hasOneUse() &&
223 (BO->getOperand(1) == Op1C || BO->getOperand(1) == Neg) &&
224 (BO->getOpcode() == Instruction::UDiv ||
225 BO->getOpcode() == Instruction::SDiv)) {
226 Value *Op0BO = BO->getOperand(0), *Op1BO = BO->getOperand(1);
228 // If the division is exact, X % Y is zero, so we end up with X or -X.
229 if (PossiblyExactOperator *SDiv = dyn_cast<PossiblyExactOperator>(BO))
230 if (SDiv->isExact()) {
232 return ReplaceInstUsesWith(I, Op0BO);
233 return BinaryOperator::CreateNeg(Op0BO);
237 if (BO->getOpcode() == Instruction::UDiv)
238 Rem = Builder->CreateURem(Op0BO, Op1BO);
240 Rem = Builder->CreateSRem(Op0BO, Op1BO);
244 return BinaryOperator::CreateSub(Op0BO, Rem);
245 return BinaryOperator::CreateSub(Rem, Op0BO);
249 /// i1 mul -> i1 and.
250 if (I.getType()->isIntegerTy(1))
251 return BinaryOperator::CreateAnd(Op0, Op1);
253 // X*(1 << Y) --> X << Y
254 // (1 << Y)*X --> X << Y
257 if (match(Op0, m_Shl(m_One(), m_Value(Y))))
258 return BinaryOperator::CreateShl(Op1, Y);
259 if (match(Op1, m_Shl(m_One(), m_Value(Y))))
260 return BinaryOperator::CreateShl(Op0, Y);
263 // If one of the operands of the multiply is a cast from a boolean value, then
264 // we know the bool is either zero or one, so this is a 'masking' multiply.
265 // X * Y (where Y is 0 or 1) -> X & (0-Y)
266 if (!I.getType()->isVectorTy()) {
267 // -2 is "-1 << 1" so it is all bits set except the low one.
268 APInt Negative2(I.getType()->getPrimitiveSizeInBits(), (uint64_t)-2, true);
270 Value *BoolCast = 0, *OtherOp = 0;
271 if (MaskedValueIsZero(Op0, Negative2))
272 BoolCast = Op0, OtherOp = Op1;
273 else if (MaskedValueIsZero(Op1, Negative2))
274 BoolCast = Op1, OtherOp = Op0;
277 Value *V = Builder->CreateSub(Constant::getNullValue(I.getType()),
279 return BinaryOperator::CreateAnd(V, OtherOp);
283 return Changed ? &I : 0;
291 // And check for corresponding fast math flags
294 static void detectLog2OfHalf(Value *&Op, Value *&Y, IntrinsicInst *&Log2) {
296 if (!Op->hasOneUse())
299 IntrinsicInst *II = dyn_cast<IntrinsicInst>(Op);
302 if (II->getIntrinsicID() != Intrinsic::log2 || !II->hasUnsafeAlgebra())
306 Value *OpLog2Of = II->getArgOperand(0);
307 if (!OpLog2Of->hasOneUse())
310 Instruction *I = dyn_cast<Instruction>(OpLog2Of);
313 if (I->getOpcode() != Instruction::FMul || !I->hasUnsafeAlgebra())
316 ConstantFP *CFP = dyn_cast<ConstantFP>(I->getOperand(0));
317 if (CFP && CFP->isExactlyValue(0.5)) {
318 Y = I->getOperand(1);
321 CFP = dyn_cast<ConstantFP>(I->getOperand(1));
322 if (CFP && CFP->isExactlyValue(0.5))
323 Y = I->getOperand(0);
326 /// Helper function of InstCombiner::visitFMul(BinaryOperator(). It returns
327 /// true iff the given value is FMul or FDiv with one and only one operand
328 /// being a normal constant (i.e. not Zero/NaN/Infinity).
329 static bool isFMulOrFDivWithConstant(Value *V) {
330 Instruction *I = dyn_cast<Instruction>(V);
331 if (!I || (I->getOpcode() != Instruction::FMul &&
332 I->getOpcode() != Instruction::FDiv))
335 ConstantFP *C0 = dyn_cast<ConstantFP>(I->getOperand(0));
336 ConstantFP *C1 = dyn_cast<ConstantFP>(I->getOperand(1));
341 return (C0 && C0->getValueAPF().isFiniteNonZero()) ||
342 (C1 && C1->getValueAPF().isFiniteNonZero());
345 static bool isNormalFp(const ConstantFP *C) {
346 const APFloat &Flt = C->getValueAPF();
347 return Flt.isNormal();
350 /// foldFMulConst() is a helper routine of InstCombiner::visitFMul().
351 /// The input \p FMulOrDiv is a FMul/FDiv with one and only one operand
352 /// being a constant (i.e. isFMulOrFDivWithConstant(FMulOrDiv) == true).
353 /// This function is to simplify "FMulOrDiv * C" and returns the
354 /// resulting expression. Note that this function could return NULL in
355 /// case the constants cannot be folded into a normal floating-point.
357 Value *InstCombiner::foldFMulConst(Instruction *FMulOrDiv, ConstantFP *C,
358 Instruction *InsertBefore) {
359 assert(isFMulOrFDivWithConstant(FMulOrDiv) && "V is invalid");
361 Value *Opnd0 = FMulOrDiv->getOperand(0);
362 Value *Opnd1 = FMulOrDiv->getOperand(1);
364 ConstantFP *C0 = dyn_cast<ConstantFP>(Opnd0);
365 ConstantFP *C1 = dyn_cast<ConstantFP>(Opnd1);
367 BinaryOperator *R = 0;
369 // (X * C0) * C => X * (C0*C)
370 if (FMulOrDiv->getOpcode() == Instruction::FMul) {
371 Constant *F = ConstantExpr::getFMul(C1 ? C1 : C0, C);
372 if (isNormalFp(cast<ConstantFP>(F)))
373 R = BinaryOperator::CreateFMul(C1 ? Opnd0 : Opnd1, F);
376 // (C0 / X) * C => (C0 * C) / X
377 if (FMulOrDiv->hasOneUse()) {
378 // It would otherwise introduce another div.
379 ConstantFP *F = cast<ConstantFP>(ConstantExpr::getFMul(C0, C));
381 R = BinaryOperator::CreateFDiv(F, Opnd1);
384 // (X / C1) * C => X * (C/C1) if C/C1 is not a denormal
385 ConstantFP *F = cast<ConstantFP>(ConstantExpr::getFDiv(C, C1));
387 R = BinaryOperator::CreateFMul(Opnd0, F);
389 // (X / C1) * C => X / (C1/C)
390 Constant *F = ConstantExpr::getFDiv(C1, C);
391 if (isNormalFp(cast<ConstantFP>(F)))
392 R = BinaryOperator::CreateFDiv(Opnd0, F);
398 R->setHasUnsafeAlgebra(true);
399 InsertNewInstWith(R, *InsertBefore);
405 Instruction *InstCombiner::visitFMul(BinaryOperator &I) {
406 bool Changed = SimplifyAssociativeOrCommutative(I);
407 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
409 if (isa<Constant>(Op0))
412 if (Value *V = SimplifyFMulInst(Op0, Op1, I.getFastMathFlags(), TD))
413 return ReplaceInstUsesWith(I, V);
415 bool AllowReassociate = I.hasUnsafeAlgebra();
417 // Simplify mul instructions with a constant RHS.
418 if (isa<Constant>(Op1)) {
419 // Try to fold constant mul into select arguments.
420 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
421 if (Instruction *R = FoldOpIntoSelect(I, SI))
424 if (isa<PHINode>(Op0))
425 if (Instruction *NV = FoldOpIntoPhi(I))
428 ConstantFP *C = dyn_cast<ConstantFP>(Op1);
429 if (C && AllowReassociate && C->getValueAPF().isFiniteNonZero()) {
430 // Let MDC denote an expression in one of these forms:
431 // X * C, C/X, X/C, where C is a constant.
433 // Try to simplify "MDC * Constant"
434 if (isFMulOrFDivWithConstant(Op0)) {
435 Value *V = foldFMulConst(cast<Instruction>(Op0), C, &I);
437 return ReplaceInstUsesWith(I, V);
440 // (MDC +/- C1) * C => (MDC * C) +/- (C1 * C)
441 Instruction *FAddSub = dyn_cast<Instruction>(Op0);
443 (FAddSub->getOpcode() == Instruction::FAdd ||
444 FAddSub->getOpcode() == Instruction::FSub)) {
445 Value *Opnd0 = FAddSub->getOperand(0);
446 Value *Opnd1 = FAddSub->getOperand(1);
447 ConstantFP *C0 = dyn_cast<ConstantFP>(Opnd0);
448 ConstantFP *C1 = dyn_cast<ConstantFP>(Opnd1);
452 std::swap(Opnd0, Opnd1);
456 if (C1 && C1->getValueAPF().isFiniteNonZero() &&
457 isFMulOrFDivWithConstant(Opnd0)) {
458 Value *M1 = ConstantExpr::getFMul(C1, C);
459 Value *M0 = isNormalFp(cast<ConstantFP>(M1)) ?
460 foldFMulConst(cast<Instruction>(Opnd0), C, &I) :
463 if (Swap && FAddSub->getOpcode() == Instruction::FSub)
466 Value *R = (FAddSub->getOpcode() == Instruction::FAdd) ?
467 BinaryOperator::CreateFAdd(M0, M1) :
468 BinaryOperator::CreateFSub(M0, M1);
469 Instruction *RI = cast<Instruction>(R);
470 RI->copyFastMathFlags(&I);
479 // Under unsafe algebra do:
480 // X * log2(0.5*Y) = X*log2(Y) - X
481 if (I.hasUnsafeAlgebra()) {
485 detectLog2OfHalf(Op0, OpY, Log2);
489 detectLog2OfHalf(Op1, OpY, Log2);
494 // if pattern detected emit alternate sequence
496 Log2->setArgOperand(0, OpY);
497 Value *FMulVal = Builder->CreateFMul(OpX, Log2);
498 Instruction *FMul = cast<Instruction>(FMulVal);
499 FMul->copyFastMathFlags(Log2);
500 Instruction *FSub = BinaryOperator::CreateFSub(FMulVal, OpX);
501 FSub->copyFastMathFlags(Log2);
506 // Handle symmetric situation in a 2-iteration loop
509 for (int i = 0; i < 2; i++) {
510 bool IgnoreZeroSign = I.hasNoSignedZeros();
511 if (BinaryOperator::isFNeg(Opnd0, IgnoreZeroSign)) {
512 Value *N0 = dyn_castFNegVal(Opnd0, IgnoreZeroSign);
513 Value *N1 = dyn_castFNegVal(Opnd1, IgnoreZeroSign);
517 return BinaryOperator::CreateFMul(N0, N1);
519 if (Opnd0->hasOneUse()) {
520 // -X * Y => -(X*Y) (Promote negation as high as possible)
521 Value *T = Builder->CreateFMul(N0, Opnd1);
522 Instruction *Neg = BinaryOperator::CreateFNeg(T);
523 if (I.getFastMathFlags().any()) {
524 if (Instruction *TI = dyn_cast<Instruction>(T))
525 TI->copyFastMathFlags(&I);
526 Neg->copyFastMathFlags(&I);
532 // (X*Y) * X => (X*X) * Y where Y != X
533 // The purpose is two-fold:
534 // 1) to form a power expression (of X).
535 // 2) potentially shorten the critical path: After transformation, the
536 // latency of the instruction Y is amortized by the expression of X*X,
537 // and therefore Y is in a "less critical" position compared to what it
538 // was before the transformation.
540 if (AllowReassociate) {
541 Value *Opnd0_0, *Opnd0_1;
542 if (Opnd0->hasOneUse() &&
543 match(Opnd0, m_FMul(m_Value(Opnd0_0), m_Value(Opnd0_1)))) {
545 if (Opnd0_0 == Opnd1 && Opnd0_1 != Opnd1)
547 else if (Opnd0_1 == Opnd1 && Opnd0_0 != Opnd1)
551 Instruction *T = cast<Instruction>(Builder->CreateFMul(Opnd1, Opnd1));
552 T->copyFastMathFlags(&I);
553 T->setDebugLoc(I.getDebugLoc());
555 Instruction *R = BinaryOperator::CreateFMul(T, Y);
556 R->copyFastMathFlags(&I);
562 // B * (uitofp i1 C) -> select C, B, 0
563 if (I.hasNoNaNs() && I.hasNoInfs() && I.hasNoSignedZeros()) {
564 Value *LHS = Op0, *RHS = Op1;
566 if (!match(RHS, m_UIToFP(m_Value(C))))
569 if (match(RHS, m_UIToFP(m_Value(C))) && C->getType()->isIntegerTy(1)) {
571 Value *Zero = ConstantFP::getNegativeZero(B->getType());
572 return SelectInst::Create(C, B, Zero);
576 // A * (1 - uitofp i1 C) -> select C, 0, A
577 if (I.hasNoNaNs() && I.hasNoInfs() && I.hasNoSignedZeros()) {
578 Value *LHS = Op0, *RHS = Op1;
580 if (!match(RHS, m_FSub(m_FPOne(), m_UIToFP(m_Value(C)))))
583 if (match(RHS, m_FSub(m_FPOne(), m_UIToFP(m_Value(C)))) &&
584 C->getType()->isIntegerTy(1)) {
586 Value *Zero = ConstantFP::getNegativeZero(A->getType());
587 return SelectInst::Create(C, Zero, A);
591 if (!isa<Constant>(Op1))
592 std::swap(Opnd0, Opnd1);
597 return Changed ? &I : 0;
600 /// SimplifyDivRemOfSelect - Try to fold a divide or remainder of a select
602 bool InstCombiner::SimplifyDivRemOfSelect(BinaryOperator &I) {
603 SelectInst *SI = cast<SelectInst>(I.getOperand(1));
605 // div/rem X, (Cond ? 0 : Y) -> div/rem X, Y
606 int NonNullOperand = -1;
607 if (Constant *ST = dyn_cast<Constant>(SI->getOperand(1)))
608 if (ST->isNullValue())
610 // div/rem X, (Cond ? Y : 0) -> div/rem X, Y
611 if (Constant *ST = dyn_cast<Constant>(SI->getOperand(2)))
612 if (ST->isNullValue())
615 if (NonNullOperand == -1)
618 Value *SelectCond = SI->getOperand(0);
620 // Change the div/rem to use 'Y' instead of the select.
621 I.setOperand(1, SI->getOperand(NonNullOperand));
623 // Okay, we know we replace the operand of the div/rem with 'Y' with no
624 // problem. However, the select, or the condition of the select may have
625 // multiple uses. Based on our knowledge that the operand must be non-zero,
626 // propagate the known value for the select into other uses of it, and
627 // propagate a known value of the condition into its other users.
629 // If the select and condition only have a single use, don't bother with this,
631 if (SI->use_empty() && SelectCond->hasOneUse())
634 // Scan the current block backward, looking for other uses of SI.
635 BasicBlock::iterator BBI = &I, BBFront = I.getParent()->begin();
637 while (BBI != BBFront) {
639 // If we found a call to a function, we can't assume it will return, so
640 // information from below it cannot be propagated above it.
641 if (isa<CallInst>(BBI) && !isa<IntrinsicInst>(BBI))
644 // Replace uses of the select or its condition with the known values.
645 for (Instruction::op_iterator I = BBI->op_begin(), E = BBI->op_end();
648 *I = SI->getOperand(NonNullOperand);
650 } else if (*I == SelectCond) {
651 *I = Builder->getInt1(NonNullOperand == 1);
656 // If we past the instruction, quit looking for it.
659 if (&*BBI == SelectCond)
662 // If we ran out of things to eliminate, break out of the loop.
663 if (SelectCond == 0 && SI == 0)
671 /// This function implements the transforms common to both integer division
672 /// instructions (udiv and sdiv). It is called by the visitors to those integer
673 /// division instructions.
674 /// @brief Common integer divide transforms
675 Instruction *InstCombiner::commonIDivTransforms(BinaryOperator &I) {
676 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
678 // The RHS is known non-zero.
679 if (Value *V = simplifyValueKnownNonZero(I.getOperand(1), *this)) {
684 // Handle cases involving: [su]div X, (select Cond, Y, Z)
685 // This does not apply for fdiv.
686 if (isa<SelectInst>(Op1) && SimplifyDivRemOfSelect(I))
689 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
690 // (X / C1) / C2 -> X / (C1*C2)
691 if (Instruction *LHS = dyn_cast<Instruction>(Op0))
692 if (Instruction::BinaryOps(LHS->getOpcode()) == I.getOpcode())
693 if (ConstantInt *LHSRHS = dyn_cast<ConstantInt>(LHS->getOperand(1))) {
694 if (MultiplyOverflows(RHS, LHSRHS,
695 I.getOpcode()==Instruction::SDiv))
696 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
697 return BinaryOperator::Create(I.getOpcode(), LHS->getOperand(0),
698 ConstantExpr::getMul(RHS, LHSRHS));
701 if (!RHS->isZero()) { // avoid X udiv 0
702 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
703 if (Instruction *R = FoldOpIntoSelect(I, SI))
705 if (isa<PHINode>(Op0))
706 if (Instruction *NV = FoldOpIntoPhi(I))
711 // See if we can fold away this div instruction.
712 if (SimplifyDemandedInstructionBits(I))
715 // (X - (X rem Y)) / Y -> X / Y; usually originates as ((X / Y) * Y) / Y
716 Value *X = 0, *Z = 0;
717 if (match(Op0, m_Sub(m_Value(X), m_Value(Z)))) { // (X - Z) / Y; Y = Op1
718 bool isSigned = I.getOpcode() == Instruction::SDiv;
719 if ((isSigned && match(Z, m_SRem(m_Specific(X), m_Specific(Op1)))) ||
720 (!isSigned && match(Z, m_URem(m_Specific(X), m_Specific(Op1)))))
721 return BinaryOperator::Create(I.getOpcode(), X, Op1);
727 /// dyn_castZExtVal - Checks if V is a zext or constant that can
728 /// be truncated to Ty without losing bits.
729 static Value *dyn_castZExtVal(Value *V, Type *Ty) {
730 if (ZExtInst *Z = dyn_cast<ZExtInst>(V)) {
731 if (Z->getSrcTy() == Ty)
732 return Z->getOperand(0);
733 } else if (ConstantInt *C = dyn_cast<ConstantInt>(V)) {
734 if (C->getValue().getActiveBits() <= cast<IntegerType>(Ty)->getBitWidth())
735 return ConstantExpr::getTrunc(C, Ty);
741 const unsigned MaxDepth = 6;
742 typedef Instruction *(*FoldUDivOperandCb)(Value *Op0, Value *Op1,
743 const BinaryOperator &I,
746 /// \brief Used to maintain state for visitUDivOperand().
747 struct UDivFoldAction {
748 FoldUDivOperandCb FoldAction; ///< Informs visitUDiv() how to fold this
749 ///< operand. This can be zero if this action
750 ///< joins two actions together.
752 Value *OperandToFold; ///< Which operand to fold.
754 Instruction *FoldResult; ///< The instruction returned when FoldAction is
757 size_t SelectLHSIdx; ///< Stores the LHS action index if this action
758 ///< joins two actions together.
761 UDivFoldAction(FoldUDivOperandCb FA, Value *InputOperand)
762 : FoldAction(FA), OperandToFold(InputOperand), FoldResult(0) {}
763 UDivFoldAction(FoldUDivOperandCb FA, Value *InputOperand, size_t SLHS)
764 : FoldAction(FA), OperandToFold(InputOperand), SelectLHSIdx(SLHS) {}
768 // X udiv 2^C -> X >> C
769 static Instruction *foldUDivPow2Cst(Value *Op0, Value *Op1,
770 const BinaryOperator &I, InstCombiner &IC) {
771 const APInt &C = cast<Constant>(Op1)->getUniqueInteger();
772 BinaryOperator *LShr = BinaryOperator::CreateLShr(
773 Op0, ConstantInt::get(Op0->getType(), C.logBase2()));
774 if (I.isExact()) LShr->setIsExact();
778 // X udiv C, where C >= signbit
779 static Instruction *foldUDivNegCst(Value *Op0, Value *Op1,
780 const BinaryOperator &I, InstCombiner &IC) {
781 Value *ICI = IC.Builder->CreateICmpULT(Op0, cast<ConstantInt>(Op1));
783 return SelectInst::Create(ICI, Constant::getNullValue(I.getType()),
784 ConstantInt::get(I.getType(), 1));
787 // X udiv (C1 << N), where C1 is "1<<C2" --> X >> (N+C2)
788 static Instruction *foldUDivShl(Value *Op0, Value *Op1, const BinaryOperator &I,
790 Instruction *ShiftLeft = cast<Instruction>(Op1);
791 if (isa<ZExtInst>(ShiftLeft))
792 ShiftLeft = cast<Instruction>(ShiftLeft->getOperand(0));
795 cast<Constant>(ShiftLeft->getOperand(0))->getUniqueInteger();
796 Value *N = ShiftLeft->getOperand(1);
798 N = IC.Builder->CreateAdd(N, ConstantInt::get(N->getType(), CI.logBase2()));
799 if (ZExtInst *Z = dyn_cast<ZExtInst>(Op1))
800 N = IC.Builder->CreateZExt(N, Z->getDestTy());
801 BinaryOperator *LShr = BinaryOperator::CreateLShr(Op0, N);
802 if (I.isExact()) LShr->setIsExact();
806 // \brief Recursively visits the possible right hand operands of a udiv
807 // instruction, seeing through select instructions, to determine if we can
808 // replace the udiv with something simpler. If we find that an operand is not
809 // able to simplify the udiv, we abort the entire transformation.
810 static size_t visitUDivOperand(Value *Op0, Value *Op1, const BinaryOperator &I,
811 SmallVectorImpl<UDivFoldAction> &Actions,
812 unsigned Depth = 0) {
813 // Check to see if this is an unsigned division with an exact power of 2,
814 // if so, convert to a right shift.
815 if (match(Op1, m_Power2())) {
816 Actions.push_back(UDivFoldAction(foldUDivPow2Cst, Op1));
817 return Actions.size();
820 if (ConstantInt *C = dyn_cast<ConstantInt>(Op1))
821 // X udiv C, where C >= signbit
822 if (C->getValue().isNegative()) {
823 Actions.push_back(UDivFoldAction(foldUDivNegCst, C));
824 return Actions.size();
827 // X udiv (C1 << N), where C1 is "1<<C2" --> X >> (N+C2)
828 if (match(Op1, m_Shl(m_Power2(), m_Value())) ||
829 match(Op1, m_ZExt(m_Shl(m_Power2(), m_Value())))) {
830 Actions.push_back(UDivFoldAction(foldUDivShl, Op1));
831 return Actions.size();
834 // The remaining tests are all recursive, so bail out if we hit the limit.
835 if (Depth++ == MaxDepth)
838 if (SelectInst *SI = dyn_cast<SelectInst>(Op1))
839 if (size_t LHSIdx = visitUDivOperand(Op0, SI->getOperand(1), I, Actions))
840 if (visitUDivOperand(Op0, SI->getOperand(2), I, Actions)) {
841 Actions.push_back(UDivFoldAction((FoldUDivOperandCb)0, Op1, LHSIdx-1));
842 return Actions.size();
848 Instruction *InstCombiner::visitUDiv(BinaryOperator &I) {
849 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
851 if (Value *V = SimplifyUDivInst(Op0, Op1, TD))
852 return ReplaceInstUsesWith(I, V);
854 // Handle the integer div common cases
855 if (Instruction *Common = commonIDivTransforms(I))
858 // (x lshr C1) udiv C2 --> x udiv (C2 << C1)
859 if (ConstantInt *C2 = dyn_cast<ConstantInt>(Op1)) {
862 if (match(Op0, m_LShr(m_Value(X), m_ConstantInt(C1)))) {
863 APInt NC = C2->getValue().shl(C1->getLimitedValue(C1->getBitWidth()-1));
864 return BinaryOperator::CreateUDiv(X, Builder->getInt(NC));
868 // (zext A) udiv (zext B) --> zext (A udiv B)
869 if (ZExtInst *ZOp0 = dyn_cast<ZExtInst>(Op0))
870 if (Value *ZOp1 = dyn_castZExtVal(Op1, ZOp0->getSrcTy()))
871 return new ZExtInst(Builder->CreateUDiv(ZOp0->getOperand(0), ZOp1, "div",
875 // (LHS udiv (select (select (...)))) -> (LHS >> (select (select (...))))
876 SmallVector<UDivFoldAction, 6> UDivActions;
877 if (visitUDivOperand(Op0, Op1, I, UDivActions))
878 for (unsigned i = 0, e = UDivActions.size(); i != e; ++i) {
879 FoldUDivOperandCb Action = UDivActions[i].FoldAction;
880 Value *ActionOp1 = UDivActions[i].OperandToFold;
883 Inst = Action(Op0, ActionOp1, I, *this);
885 // This action joins two actions together. The RHS of this action is
886 // simply the last action we processed, we saved the LHS action index in
887 // the joining action.
888 size_t SelectRHSIdx = i - 1;
889 Value *SelectRHS = UDivActions[SelectRHSIdx].FoldResult;
890 size_t SelectLHSIdx = UDivActions[i].SelectLHSIdx;
891 Value *SelectLHS = UDivActions[SelectLHSIdx].FoldResult;
892 Inst = SelectInst::Create(cast<SelectInst>(ActionOp1)->getCondition(),
893 SelectLHS, SelectRHS);
896 // If this is the last action to process, return it to the InstCombiner.
897 // Otherwise, we insert it before the UDiv and record it so that we may
898 // use it as part of a joining action (i.e., a SelectInst).
900 Inst->insertBefore(&I);
901 UDivActions[i].FoldResult = Inst;
909 Instruction *InstCombiner::visitSDiv(BinaryOperator &I) {
910 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
912 if (Value *V = SimplifySDivInst(Op0, Op1, TD))
913 return ReplaceInstUsesWith(I, V);
915 // Handle the integer div common cases
916 if (Instruction *Common = commonIDivTransforms(I))
919 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
921 if (RHS->isAllOnesValue())
922 return BinaryOperator::CreateNeg(Op0);
924 // sdiv X, C --> ashr exact X, log2(C)
925 if (I.isExact() && RHS->getValue().isNonNegative() &&
926 RHS->getValue().isPowerOf2()) {
927 Value *ShAmt = llvm::ConstantInt::get(RHS->getType(),
928 RHS->getValue().exactLogBase2());
929 return BinaryOperator::CreateExactAShr(Op0, ShAmt, I.getName());
932 // -X/C --> X/-C provided the negation doesn't overflow.
933 if (SubOperator *Sub = dyn_cast<SubOperator>(Op0))
934 if (match(Sub->getOperand(0), m_Zero()) && Sub->hasNoSignedWrap())
935 return BinaryOperator::CreateSDiv(Sub->getOperand(1),
936 ConstantExpr::getNeg(RHS));
939 // If the sign bits of both operands are zero (i.e. we can prove they are
940 // unsigned inputs), turn this into a udiv.
941 if (I.getType()->isIntegerTy()) {
942 APInt Mask(APInt::getSignBit(I.getType()->getPrimitiveSizeInBits()));
943 if (MaskedValueIsZero(Op0, Mask)) {
944 if (MaskedValueIsZero(Op1, Mask)) {
945 // X sdiv Y -> X udiv Y, iff X and Y don't have sign bit set
946 return BinaryOperator::CreateUDiv(Op0, Op1, I.getName());
949 if (match(Op1, m_Shl(m_Power2(), m_Value()))) {
950 // X sdiv (1 << Y) -> X udiv (1 << Y) ( -> X u>> Y)
951 // Safe because the only negative value (1 << Y) can take on is
952 // INT_MIN, and X sdiv INT_MIN == X udiv INT_MIN == 0 if X doesn't have
954 return BinaryOperator::CreateUDiv(Op0, Op1, I.getName());
962 /// CvtFDivConstToReciprocal tries to convert X/C into X*1/C if C not a special
964 /// 1) 1/C is exact, or
965 /// 2) reciprocal is allowed.
966 /// If the conversion was successful, the simplified expression "X * 1/C" is
967 /// returned; otherwise, NULL is returned.
969 static Instruction *CvtFDivConstToReciprocal(Value *Dividend,
971 bool AllowReciprocal) {
972 const APFloat &FpVal = Divisor->getValueAPF();
973 APFloat Reciprocal(FpVal.getSemantics());
974 bool Cvt = FpVal.getExactInverse(&Reciprocal);
976 if (!Cvt && AllowReciprocal && FpVal.isFiniteNonZero()) {
977 Reciprocal = APFloat(FpVal.getSemantics(), 1.0f);
978 (void)Reciprocal.divide(FpVal, APFloat::rmNearestTiesToEven);
979 Cvt = !Reciprocal.isDenormal();
986 R = ConstantFP::get(Dividend->getType()->getContext(), Reciprocal);
987 return BinaryOperator::CreateFMul(Dividend, R);
990 Instruction *InstCombiner::visitFDiv(BinaryOperator &I) {
991 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
993 if (Value *V = SimplifyFDivInst(Op0, Op1, TD))
994 return ReplaceInstUsesWith(I, V);
996 if (isa<Constant>(Op0))
997 if (SelectInst *SI = dyn_cast<SelectInst>(Op1))
998 if (Instruction *R = FoldOpIntoSelect(I, SI))
1001 bool AllowReassociate = I.hasUnsafeAlgebra();
1002 bool AllowReciprocal = I.hasAllowReciprocal();
1004 if (ConstantFP *Op1C = dyn_cast<ConstantFP>(Op1)) {
1005 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
1006 if (Instruction *R = FoldOpIntoSelect(I, SI))
1009 if (AllowReassociate) {
1011 ConstantFP *C2 = Op1C;
1013 Instruction *Res = 0;
1015 if (match(Op0, m_FMul(m_Value(X), m_ConstantFP(C1)))) {
1016 // (X*C1)/C2 => X * (C1/C2)
1018 Constant *C = ConstantExpr::getFDiv(C1, C2);
1019 const APFloat &F = cast<ConstantFP>(C)->getValueAPF();
1021 Res = BinaryOperator::CreateFMul(X, C);
1022 } else if (match(Op0, m_FDiv(m_Value(X), m_ConstantFP(C1)))) {
1023 // (X/C1)/C2 => X /(C2*C1) [=> X * 1/(C2*C1) if reciprocal is allowed]
1025 Constant *C = ConstantExpr::getFMul(C1, C2);
1026 const APFloat &F = cast<ConstantFP>(C)->getValueAPF();
1028 Res = CvtFDivConstToReciprocal(X, cast<ConstantFP>(C),
1031 Res = BinaryOperator::CreateFDiv(X, C);
1036 Res->setFastMathFlags(I.getFastMathFlags());
1042 if (Instruction *T = CvtFDivConstToReciprocal(Op0, Op1C, AllowReciprocal))
1048 if (AllowReassociate && isa<ConstantFP>(Op0)) {
1049 ConstantFP *C1 = cast<ConstantFP>(Op0), *C2;
1052 bool CreateDiv = true;
1054 // C1 / (X*C2) => (C1/C2) / X
1055 if (match(Op1, m_FMul(m_Value(X), m_ConstantFP(C2))))
1056 Fold = ConstantExpr::getFDiv(C1, C2);
1057 else if (match(Op1, m_FDiv(m_Value(X), m_ConstantFP(C2)))) {
1058 // C1 / (X/C2) => (C1*C2) / X
1059 Fold = ConstantExpr::getFMul(C1, C2);
1060 } else if (match(Op1, m_FDiv(m_ConstantFP(C2), m_Value(X)))) {
1061 // C1 / (C2/X) => (C1/C2) * X
1062 Fold = ConstantExpr::getFDiv(C1, C2);
1067 const APFloat &FoldC = cast<ConstantFP>(Fold)->getValueAPF();
1068 if (FoldC.isNormal()) {
1069 Instruction *R = CreateDiv ?
1070 BinaryOperator::CreateFDiv(Fold, X) :
1071 BinaryOperator::CreateFMul(X, Fold);
1072 R->setFastMathFlags(I.getFastMathFlags());
1079 if (AllowReassociate) {
1082 Instruction *SimpR = 0;
1084 if (Op0->hasOneUse() && match(Op0, m_FDiv(m_Value(X), m_Value(Y)))) {
1085 // (X/Y) / Z => X / (Y*Z)
1087 if (!isa<ConstantFP>(Y) || !isa<ConstantFP>(Op1)) {
1088 NewInst = Builder->CreateFMul(Y, Op1);
1089 SimpR = BinaryOperator::CreateFDiv(X, NewInst);
1091 } else if (Op1->hasOneUse() && match(Op1, m_FDiv(m_Value(X), m_Value(Y)))) {
1092 // Z / (X/Y) => Z*Y / X
1094 if (!isa<ConstantFP>(Y) || !isa<ConstantFP>(Op0)) {
1095 NewInst = Builder->CreateFMul(Op0, Y);
1096 SimpR = BinaryOperator::CreateFDiv(NewInst, X);
1101 if (Instruction *T = dyn_cast<Instruction>(NewInst))
1102 T->setDebugLoc(I.getDebugLoc());
1103 SimpR->setFastMathFlags(I.getFastMathFlags());
1111 /// This function implements the transforms common to both integer remainder
1112 /// instructions (urem and srem). It is called by the visitors to those integer
1113 /// remainder instructions.
1114 /// @brief Common integer remainder transforms
1115 Instruction *InstCombiner::commonIRemTransforms(BinaryOperator &I) {
1116 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
1118 // The RHS is known non-zero.
1119 if (Value *V = simplifyValueKnownNonZero(I.getOperand(1), *this)) {
1124 // Handle cases involving: rem X, (select Cond, Y, Z)
1125 if (isa<SelectInst>(Op1) && SimplifyDivRemOfSelect(I))
1128 if (isa<ConstantInt>(Op1)) {
1129 if (Instruction *Op0I = dyn_cast<Instruction>(Op0)) {
1130 if (SelectInst *SI = dyn_cast<SelectInst>(Op0I)) {
1131 if (Instruction *R = FoldOpIntoSelect(I, SI))
1133 } else if (isa<PHINode>(Op0I)) {
1134 if (Instruction *NV = FoldOpIntoPhi(I))
1138 // See if we can fold away this rem instruction.
1139 if (SimplifyDemandedInstructionBits(I))
1147 Instruction *InstCombiner::visitURem(BinaryOperator &I) {
1148 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
1150 if (Value *V = SimplifyURemInst(Op0, Op1, TD))
1151 return ReplaceInstUsesWith(I, V);
1153 if (Instruction *common = commonIRemTransforms(I))
1156 // (zext A) urem (zext B) --> zext (A urem B)
1157 if (ZExtInst *ZOp0 = dyn_cast<ZExtInst>(Op0))
1158 if (Value *ZOp1 = dyn_castZExtVal(Op1, ZOp0->getSrcTy()))
1159 return new ZExtInst(Builder->CreateURem(ZOp0->getOperand(0), ZOp1),
1162 // X urem Y -> X and Y-1, where Y is a power of 2,
1163 if (isKnownToBeAPowerOfTwo(Op1, /*OrZero*/true)) {
1164 Constant *N1 = Constant::getAllOnesValue(I.getType());
1165 Value *Add = Builder->CreateAdd(Op1, N1);
1166 return BinaryOperator::CreateAnd(Op0, Add);
1169 // 1 urem X -> zext(X != 1)
1170 if (match(Op0, m_One())) {
1171 Value *Cmp = Builder->CreateICmpNE(Op1, Op0);
1172 Value *Ext = Builder->CreateZExt(Cmp, I.getType());
1173 return ReplaceInstUsesWith(I, Ext);
1179 Instruction *InstCombiner::visitSRem(BinaryOperator &I) {
1180 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
1182 if (Value *V = SimplifySRemInst(Op0, Op1, TD))
1183 return ReplaceInstUsesWith(I, V);
1185 // Handle the integer rem common cases
1186 if (Instruction *Common = commonIRemTransforms(I))
1189 if (Value *RHSNeg = dyn_castNegVal(Op1))
1190 if (!isa<Constant>(RHSNeg) ||
1191 (isa<ConstantInt>(RHSNeg) &&
1192 cast<ConstantInt>(RHSNeg)->getValue().isStrictlyPositive())) {
1194 Worklist.AddValue(I.getOperand(1));
1195 I.setOperand(1, RHSNeg);
1199 // If the sign bits of both operands are zero (i.e. we can prove they are
1200 // unsigned inputs), turn this into a urem.
1201 if (I.getType()->isIntegerTy()) {
1202 APInt Mask(APInt::getSignBit(I.getType()->getPrimitiveSizeInBits()));
1203 if (MaskedValueIsZero(Op1, Mask) && MaskedValueIsZero(Op0, Mask)) {
1204 // X srem Y -> X urem Y, iff X and Y don't have sign bit set
1205 return BinaryOperator::CreateURem(Op0, Op1, I.getName());
1209 // If it's a constant vector, flip any negative values positive.
1210 if (isa<ConstantVector>(Op1) || isa<ConstantDataVector>(Op1)) {
1211 Constant *C = cast<Constant>(Op1);
1212 unsigned VWidth = C->getType()->getVectorNumElements();
1214 bool hasNegative = false;
1215 bool hasMissing = false;
1216 for (unsigned i = 0; i != VWidth; ++i) {
1217 Constant *Elt = C->getAggregateElement(i);
1223 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Elt))
1224 if (RHS->isNegative())
1228 if (hasNegative && !hasMissing) {
1229 SmallVector<Constant *, 16> Elts(VWidth);
1230 for (unsigned i = 0; i != VWidth; ++i) {
1231 Elts[i] = C->getAggregateElement(i); // Handle undef, etc.
1232 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Elts[i])) {
1233 if (RHS->isNegative())
1234 Elts[i] = cast<ConstantInt>(ConstantExpr::getNeg(RHS));
1238 Constant *NewRHSV = ConstantVector::get(Elts);
1239 if (NewRHSV != C) { // Don't loop on -MININT
1240 Worklist.AddValue(I.getOperand(1));
1241 I.setOperand(1, NewRHSV);
1250 Instruction *InstCombiner::visitFRem(BinaryOperator &I) {
1251 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
1253 if (Value *V = SimplifyFRemInst(Op0, Op1, TD))
1254 return ReplaceInstUsesWith(I, V);
1256 // Handle cases involving: rem X, (select Cond, Y, Z)
1257 if (isa<SelectInst>(Op1) && SimplifyDivRemOfSelect(I))