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 if (match(I->getOperand(0), m_SpecificFP(0.5)))
317 Y = I->getOperand(1);
318 else if (match(I->getOperand(1), m_SpecificFP(0.5)))
319 Y = I->getOperand(0);
322 static bool isFiniteNonZeroFp(Constant *C) {
323 if (C->getType()->isVectorTy()) {
324 for (unsigned I = 0, E = C->getType()->getVectorNumElements(); I != E;
326 ConstantFP *CFP = dyn_cast<ConstantFP>(C->getAggregateElement(I));
327 if (!CFP || !CFP->getValueAPF().isFiniteNonZero())
333 return isa<ConstantFP>(C) &&
334 cast<ConstantFP>(C)->getValueAPF().isFiniteNonZero();
337 static bool isNormalFp(Constant *C) {
338 if (C->getType()->isVectorTy()) {
339 for (unsigned I = 0, E = C->getType()->getVectorNumElements(); I != E;
341 ConstantFP *CFP = dyn_cast<ConstantFP>(C->getAggregateElement(I));
342 if (!CFP || !CFP->getValueAPF().isNormal())
348 return isa<ConstantFP>(C) && cast<ConstantFP>(C)->getValueAPF().isNormal();
351 /// Helper function of InstCombiner::visitFMul(BinaryOperator(). It returns
352 /// true iff the given value is FMul or FDiv with one and only one operand
353 /// being a normal constant (i.e. not Zero/NaN/Infinity).
354 static bool isFMulOrFDivWithConstant(Value *V) {
355 Instruction *I = dyn_cast<Instruction>(V);
356 if (!I || (I->getOpcode() != Instruction::FMul &&
357 I->getOpcode() != Instruction::FDiv))
360 Constant *C0 = dyn_cast<Constant>(I->getOperand(0));
361 Constant *C1 = dyn_cast<Constant>(I->getOperand(1));
366 return (C0 && isFiniteNonZeroFp(C0)) || (C1 && isFiniteNonZeroFp(C1));
369 /// foldFMulConst() is a helper routine of InstCombiner::visitFMul().
370 /// The input \p FMulOrDiv is a FMul/FDiv with one and only one operand
371 /// being a constant (i.e. isFMulOrFDivWithConstant(FMulOrDiv) == true).
372 /// This function is to simplify "FMulOrDiv * C" and returns the
373 /// resulting expression. Note that this function could return NULL in
374 /// case the constants cannot be folded into a normal floating-point.
376 Value *InstCombiner::foldFMulConst(Instruction *FMulOrDiv, Constant *C,
377 Instruction *InsertBefore) {
378 assert(isFMulOrFDivWithConstant(FMulOrDiv) && "V is invalid");
380 Value *Opnd0 = FMulOrDiv->getOperand(0);
381 Value *Opnd1 = FMulOrDiv->getOperand(1);
383 Constant *C0 = dyn_cast<Constant>(Opnd0);
384 Constant *C1 = dyn_cast<Constant>(Opnd1);
386 BinaryOperator *R = 0;
388 // (X * C0) * C => X * (C0*C)
389 if (FMulOrDiv->getOpcode() == Instruction::FMul) {
390 Constant *F = ConstantExpr::getFMul(C1 ? C1 : C0, C);
392 R = BinaryOperator::CreateFMul(C1 ? Opnd0 : Opnd1, F);
395 // (C0 / X) * C => (C0 * C) / X
396 if (FMulOrDiv->hasOneUse()) {
397 // It would otherwise introduce another div.
398 Constant *F = ConstantExpr::getFMul(C0, C);
400 R = BinaryOperator::CreateFDiv(F, Opnd1);
403 // (X / C1) * C => X * (C/C1) if C/C1 is not a denormal
404 Constant *F = ConstantExpr::getFDiv(C, C1);
406 R = BinaryOperator::CreateFMul(Opnd0, F);
408 // (X / C1) * C => X / (C1/C)
409 Constant *F = ConstantExpr::getFDiv(C1, C);
411 R = BinaryOperator::CreateFDiv(Opnd0, F);
417 R->setHasUnsafeAlgebra(true);
418 InsertNewInstWith(R, *InsertBefore);
424 Instruction *InstCombiner::visitFMul(BinaryOperator &I) {
425 bool Changed = SimplifyAssociativeOrCommutative(I);
426 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
428 if (isa<Constant>(Op0))
431 if (Value *V = SimplifyFMulInst(Op0, Op1, I.getFastMathFlags(), TD))
432 return ReplaceInstUsesWith(I, V);
434 bool AllowReassociate = I.hasUnsafeAlgebra();
436 // Simplify mul instructions with a constant RHS.
437 if (isa<Constant>(Op1)) {
438 // Try to fold constant mul into select arguments.
439 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
440 if (Instruction *R = FoldOpIntoSelect(I, SI))
443 if (isa<PHINode>(Op0))
444 if (Instruction *NV = FoldOpIntoPhi(I))
447 // (fmul X, -1.0) --> (fsub -0.0, X)
448 if (match(Op1, m_SpecificFP(-1.0))) {
449 Constant *NegZero = ConstantFP::getNegativeZero(Op1->getType());
450 Instruction *RI = BinaryOperator::CreateFSub(NegZero, Op0);
451 RI->copyFastMathFlags(&I);
455 Constant *C = cast<Constant>(Op1);
456 if (AllowReassociate && isFiniteNonZeroFp(C)) {
457 // Let MDC denote an expression in one of these forms:
458 // X * C, C/X, X/C, where C is a constant.
460 // Try to simplify "MDC * Constant"
461 if (isFMulOrFDivWithConstant(Op0))
462 if (Value *V = foldFMulConst(cast<Instruction>(Op0), C, &I))
463 return ReplaceInstUsesWith(I, V);
465 // (MDC +/- C1) * C => (MDC * C) +/- (C1 * C)
466 Instruction *FAddSub = dyn_cast<Instruction>(Op0);
468 (FAddSub->getOpcode() == Instruction::FAdd ||
469 FAddSub->getOpcode() == Instruction::FSub)) {
470 Value *Opnd0 = FAddSub->getOperand(0);
471 Value *Opnd1 = FAddSub->getOperand(1);
472 Constant *C0 = dyn_cast<Constant>(Opnd0);
473 Constant *C1 = dyn_cast<Constant>(Opnd1);
477 std::swap(Opnd0, Opnd1);
481 if (C1 && isFiniteNonZeroFp(C1) && isFMulOrFDivWithConstant(Opnd0)) {
482 Value *M1 = ConstantExpr::getFMul(C1, C);
483 Value *M0 = isNormalFp(cast<Constant>(M1)) ?
484 foldFMulConst(cast<Instruction>(Opnd0), C, &I) :
487 if (Swap && FAddSub->getOpcode() == Instruction::FSub)
490 Instruction *RI = (FAddSub->getOpcode() == Instruction::FAdd)
491 ? BinaryOperator::CreateFAdd(M0, M1)
492 : BinaryOperator::CreateFSub(M0, M1);
493 RI->copyFastMathFlags(&I);
502 // Under unsafe algebra do:
503 // X * log2(0.5*Y) = X*log2(Y) - X
504 if (I.hasUnsafeAlgebra()) {
508 detectLog2OfHalf(Op0, OpY, Log2);
512 detectLog2OfHalf(Op1, OpY, Log2);
517 // if pattern detected emit alternate sequence
519 BuilderTy::FastMathFlagGuard Guard(*Builder);
520 Builder->SetFastMathFlags(Log2->getFastMathFlags());
521 Log2->setArgOperand(0, OpY);
522 Value *FMulVal = Builder->CreateFMul(OpX, Log2);
523 Value *FSub = Builder->CreateFSub(FMulVal, OpX);
525 return ReplaceInstUsesWith(I, FSub);
529 // Handle symmetric situation in a 2-iteration loop
532 for (int i = 0; i < 2; i++) {
533 bool IgnoreZeroSign = I.hasNoSignedZeros();
534 if (BinaryOperator::isFNeg(Opnd0, IgnoreZeroSign)) {
535 BuilderTy::FastMathFlagGuard Guard(*Builder);
536 Builder->SetFastMathFlags(I.getFastMathFlags());
538 Value *N0 = dyn_castFNegVal(Opnd0, IgnoreZeroSign);
539 Value *N1 = dyn_castFNegVal(Opnd1, IgnoreZeroSign);
543 Value *FMul = Builder->CreateFMul(N0, N1);
545 return ReplaceInstUsesWith(I, FMul);
548 if (Opnd0->hasOneUse()) {
549 // -X * Y => -(X*Y) (Promote negation as high as possible)
550 Value *T = Builder->CreateFMul(N0, Opnd1);
551 Value *Neg = Builder->CreateFNeg(T);
553 return ReplaceInstUsesWith(I, Neg);
557 // (X*Y) * X => (X*X) * Y where Y != X
558 // The purpose is two-fold:
559 // 1) to form a power expression (of X).
560 // 2) potentially shorten the critical path: After transformation, the
561 // latency of the instruction Y is amortized by the expression of X*X,
562 // and therefore Y is in a "less critical" position compared to what it
563 // was before the transformation.
565 if (AllowReassociate) {
566 Value *Opnd0_0, *Opnd0_1;
567 if (Opnd0->hasOneUse() &&
568 match(Opnd0, m_FMul(m_Value(Opnd0_0), m_Value(Opnd0_1)))) {
570 if (Opnd0_0 == Opnd1 && Opnd0_1 != Opnd1)
572 else if (Opnd0_1 == Opnd1 && Opnd0_0 != Opnd1)
576 BuilderTy::FastMathFlagGuard Guard(*Builder);
577 Builder->SetFastMathFlags(I.getFastMathFlags());
578 Value *T = Builder->CreateFMul(Opnd1, Opnd1);
580 Value *R = Builder->CreateFMul(T, Y);
582 return ReplaceInstUsesWith(I, R);
587 // B * (uitofp i1 C) -> select C, B, 0
588 if (I.hasNoNaNs() && I.hasNoInfs() && I.hasNoSignedZeros()) {
589 Value *LHS = Op0, *RHS = Op1;
591 if (!match(RHS, m_UIToFP(m_Value(C))))
594 if (match(RHS, m_UIToFP(m_Value(C))) &&
595 C->getType()->getScalarType()->isIntegerTy(1)) {
597 Value *Zero = ConstantFP::getNegativeZero(B->getType());
598 return SelectInst::Create(C, B, Zero);
602 // A * (1 - uitofp i1 C) -> select C, 0, A
603 if (I.hasNoNaNs() && I.hasNoInfs() && I.hasNoSignedZeros()) {
604 Value *LHS = Op0, *RHS = Op1;
606 if (!match(RHS, m_FSub(m_FPOne(), m_UIToFP(m_Value(C)))))
609 if (match(RHS, m_FSub(m_FPOne(), m_UIToFP(m_Value(C)))) &&
610 C->getType()->getScalarType()->isIntegerTy(1)) {
612 Value *Zero = ConstantFP::getNegativeZero(A->getType());
613 return SelectInst::Create(C, Zero, A);
617 if (!isa<Constant>(Op1))
618 std::swap(Opnd0, Opnd1);
623 return Changed ? &I : 0;
626 /// SimplifyDivRemOfSelect - Try to fold a divide or remainder of a select
628 bool InstCombiner::SimplifyDivRemOfSelect(BinaryOperator &I) {
629 SelectInst *SI = cast<SelectInst>(I.getOperand(1));
631 // div/rem X, (Cond ? 0 : Y) -> div/rem X, Y
632 int NonNullOperand = -1;
633 if (Constant *ST = dyn_cast<Constant>(SI->getOperand(1)))
634 if (ST->isNullValue())
636 // div/rem X, (Cond ? Y : 0) -> div/rem X, Y
637 if (Constant *ST = dyn_cast<Constant>(SI->getOperand(2)))
638 if (ST->isNullValue())
641 if (NonNullOperand == -1)
644 Value *SelectCond = SI->getOperand(0);
646 // Change the div/rem to use 'Y' instead of the select.
647 I.setOperand(1, SI->getOperand(NonNullOperand));
649 // Okay, we know we replace the operand of the div/rem with 'Y' with no
650 // problem. However, the select, or the condition of the select may have
651 // multiple uses. Based on our knowledge that the operand must be non-zero,
652 // propagate the known value for the select into other uses of it, and
653 // propagate a known value of the condition into its other users.
655 // If the select and condition only have a single use, don't bother with this,
657 if (SI->use_empty() && SelectCond->hasOneUse())
660 // Scan the current block backward, looking for other uses of SI.
661 BasicBlock::iterator BBI = &I, BBFront = I.getParent()->begin();
663 while (BBI != BBFront) {
665 // If we found a call to a function, we can't assume it will return, so
666 // information from below it cannot be propagated above it.
667 if (isa<CallInst>(BBI) && !isa<IntrinsicInst>(BBI))
670 // Replace uses of the select or its condition with the known values.
671 for (Instruction::op_iterator I = BBI->op_begin(), E = BBI->op_end();
674 *I = SI->getOperand(NonNullOperand);
676 } else if (*I == SelectCond) {
677 *I = Builder->getInt1(NonNullOperand == 1);
682 // If we past the instruction, quit looking for it.
685 if (&*BBI == SelectCond)
688 // If we ran out of things to eliminate, break out of the loop.
689 if (SelectCond == 0 && SI == 0)
697 /// This function implements the transforms common to both integer division
698 /// instructions (udiv and sdiv). It is called by the visitors to those integer
699 /// division instructions.
700 /// @brief Common integer divide transforms
701 Instruction *InstCombiner::commonIDivTransforms(BinaryOperator &I) {
702 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
704 // The RHS is known non-zero.
705 if (Value *V = simplifyValueKnownNonZero(I.getOperand(1), *this)) {
710 // Handle cases involving: [su]div X, (select Cond, Y, Z)
711 // This does not apply for fdiv.
712 if (isa<SelectInst>(Op1) && SimplifyDivRemOfSelect(I))
715 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
716 // (X / C1) / C2 -> X / (C1*C2)
717 if (Instruction *LHS = dyn_cast<Instruction>(Op0))
718 if (Instruction::BinaryOps(LHS->getOpcode()) == I.getOpcode())
719 if (ConstantInt *LHSRHS = dyn_cast<ConstantInt>(LHS->getOperand(1))) {
720 if (MultiplyOverflows(RHS, LHSRHS,
721 I.getOpcode()==Instruction::SDiv))
722 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
723 return BinaryOperator::Create(I.getOpcode(), LHS->getOperand(0),
724 ConstantExpr::getMul(RHS, LHSRHS));
727 if (!RHS->isZero()) { // avoid X udiv 0
728 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
729 if (Instruction *R = FoldOpIntoSelect(I, SI))
731 if (isa<PHINode>(Op0))
732 if (Instruction *NV = FoldOpIntoPhi(I))
737 // See if we can fold away this div instruction.
738 if (SimplifyDemandedInstructionBits(I))
741 // (X - (X rem Y)) / Y -> X / Y; usually originates as ((X / Y) * Y) / Y
742 Value *X = 0, *Z = 0;
743 if (match(Op0, m_Sub(m_Value(X), m_Value(Z)))) { // (X - Z) / Y; Y = Op1
744 bool isSigned = I.getOpcode() == Instruction::SDiv;
745 if ((isSigned && match(Z, m_SRem(m_Specific(X), m_Specific(Op1)))) ||
746 (!isSigned && match(Z, m_URem(m_Specific(X), m_Specific(Op1)))))
747 return BinaryOperator::Create(I.getOpcode(), X, Op1);
753 /// dyn_castZExtVal - Checks if V is a zext or constant that can
754 /// be truncated to Ty without losing bits.
755 static Value *dyn_castZExtVal(Value *V, Type *Ty) {
756 if (ZExtInst *Z = dyn_cast<ZExtInst>(V)) {
757 if (Z->getSrcTy() == Ty)
758 return Z->getOperand(0);
759 } else if (ConstantInt *C = dyn_cast<ConstantInt>(V)) {
760 if (C->getValue().getActiveBits() <= cast<IntegerType>(Ty)->getBitWidth())
761 return ConstantExpr::getTrunc(C, Ty);
767 const unsigned MaxDepth = 6;
768 typedef Instruction *(*FoldUDivOperandCb)(Value *Op0, Value *Op1,
769 const BinaryOperator &I,
772 /// \brief Used to maintain state for visitUDivOperand().
773 struct UDivFoldAction {
774 FoldUDivOperandCb FoldAction; ///< Informs visitUDiv() how to fold this
775 ///< operand. This can be zero if this action
776 ///< joins two actions together.
778 Value *OperandToFold; ///< Which operand to fold.
780 Instruction *FoldResult; ///< The instruction returned when FoldAction is
783 size_t SelectLHSIdx; ///< Stores the LHS action index if this action
784 ///< joins two actions together.
787 UDivFoldAction(FoldUDivOperandCb FA, Value *InputOperand)
788 : FoldAction(FA), OperandToFold(InputOperand), FoldResult(0) {}
789 UDivFoldAction(FoldUDivOperandCb FA, Value *InputOperand, size_t SLHS)
790 : FoldAction(FA), OperandToFold(InputOperand), SelectLHSIdx(SLHS) {}
794 // X udiv 2^C -> X >> C
795 static Instruction *foldUDivPow2Cst(Value *Op0, Value *Op1,
796 const BinaryOperator &I, InstCombiner &IC) {
797 const APInt &C = cast<Constant>(Op1)->getUniqueInteger();
798 BinaryOperator *LShr = BinaryOperator::CreateLShr(
799 Op0, ConstantInt::get(Op0->getType(), C.logBase2()));
800 if (I.isExact()) LShr->setIsExact();
804 // X udiv C, where C >= signbit
805 static Instruction *foldUDivNegCst(Value *Op0, Value *Op1,
806 const BinaryOperator &I, InstCombiner &IC) {
807 Value *ICI = IC.Builder->CreateICmpULT(Op0, cast<ConstantInt>(Op1));
809 return SelectInst::Create(ICI, Constant::getNullValue(I.getType()),
810 ConstantInt::get(I.getType(), 1));
813 // X udiv (C1 << N), where C1 is "1<<C2" --> X >> (N+C2)
814 static Instruction *foldUDivShl(Value *Op0, Value *Op1, const BinaryOperator &I,
816 Instruction *ShiftLeft = cast<Instruction>(Op1);
817 if (isa<ZExtInst>(ShiftLeft))
818 ShiftLeft = cast<Instruction>(ShiftLeft->getOperand(0));
821 cast<Constant>(ShiftLeft->getOperand(0))->getUniqueInteger();
822 Value *N = ShiftLeft->getOperand(1);
824 N = IC.Builder->CreateAdd(N, ConstantInt::get(N->getType(), CI.logBase2()));
825 if (ZExtInst *Z = dyn_cast<ZExtInst>(Op1))
826 N = IC.Builder->CreateZExt(N, Z->getDestTy());
827 BinaryOperator *LShr = BinaryOperator::CreateLShr(Op0, N);
828 if (I.isExact()) LShr->setIsExact();
832 // \brief Recursively visits the possible right hand operands of a udiv
833 // instruction, seeing through select instructions, to determine if we can
834 // replace the udiv with something simpler. If we find that an operand is not
835 // able to simplify the udiv, we abort the entire transformation.
836 static size_t visitUDivOperand(Value *Op0, Value *Op1, const BinaryOperator &I,
837 SmallVectorImpl<UDivFoldAction> &Actions,
838 unsigned Depth = 0) {
839 // Check to see if this is an unsigned division with an exact power of 2,
840 // if so, convert to a right shift.
841 if (match(Op1, m_Power2())) {
842 Actions.push_back(UDivFoldAction(foldUDivPow2Cst, Op1));
843 return Actions.size();
846 if (ConstantInt *C = dyn_cast<ConstantInt>(Op1))
847 // X udiv C, where C >= signbit
848 if (C->getValue().isNegative()) {
849 Actions.push_back(UDivFoldAction(foldUDivNegCst, C));
850 return Actions.size();
853 // X udiv (C1 << N), where C1 is "1<<C2" --> X >> (N+C2)
854 if (match(Op1, m_Shl(m_Power2(), m_Value())) ||
855 match(Op1, m_ZExt(m_Shl(m_Power2(), m_Value())))) {
856 Actions.push_back(UDivFoldAction(foldUDivShl, Op1));
857 return Actions.size();
860 // The remaining tests are all recursive, so bail out if we hit the limit.
861 if (Depth++ == MaxDepth)
864 if (SelectInst *SI = dyn_cast<SelectInst>(Op1))
865 if (size_t LHSIdx = visitUDivOperand(Op0, SI->getOperand(1), I, Actions))
866 if (visitUDivOperand(Op0, SI->getOperand(2), I, Actions)) {
867 Actions.push_back(UDivFoldAction((FoldUDivOperandCb)0, Op1, LHSIdx-1));
868 return Actions.size();
874 Instruction *InstCombiner::visitUDiv(BinaryOperator &I) {
875 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
877 if (Value *V = SimplifyUDivInst(Op0, Op1, TD))
878 return ReplaceInstUsesWith(I, V);
880 // Handle the integer div common cases
881 if (Instruction *Common = commonIDivTransforms(I))
884 // (x lshr C1) udiv C2 --> x udiv (C2 << C1)
885 if (ConstantInt *C2 = dyn_cast<ConstantInt>(Op1)) {
888 if (match(Op0, m_LShr(m_Value(X), m_ConstantInt(C1)))) {
889 APInt NC = C2->getValue().shl(C1->getLimitedValue(C1->getBitWidth()-1));
890 return BinaryOperator::CreateUDiv(X, Builder->getInt(NC));
894 // (zext A) udiv (zext B) --> zext (A udiv B)
895 if (ZExtInst *ZOp0 = dyn_cast<ZExtInst>(Op0))
896 if (Value *ZOp1 = dyn_castZExtVal(Op1, ZOp0->getSrcTy()))
897 return new ZExtInst(Builder->CreateUDiv(ZOp0->getOperand(0), ZOp1, "div",
901 // (LHS udiv (select (select (...)))) -> (LHS >> (select (select (...))))
902 SmallVector<UDivFoldAction, 6> UDivActions;
903 if (visitUDivOperand(Op0, Op1, I, UDivActions))
904 for (unsigned i = 0, e = UDivActions.size(); i != e; ++i) {
905 FoldUDivOperandCb Action = UDivActions[i].FoldAction;
906 Value *ActionOp1 = UDivActions[i].OperandToFold;
909 Inst = Action(Op0, ActionOp1, I, *this);
911 // This action joins two actions together. The RHS of this action is
912 // simply the last action we processed, we saved the LHS action index in
913 // the joining action.
914 size_t SelectRHSIdx = i - 1;
915 Value *SelectRHS = UDivActions[SelectRHSIdx].FoldResult;
916 size_t SelectLHSIdx = UDivActions[i].SelectLHSIdx;
917 Value *SelectLHS = UDivActions[SelectLHSIdx].FoldResult;
918 Inst = SelectInst::Create(cast<SelectInst>(ActionOp1)->getCondition(),
919 SelectLHS, SelectRHS);
922 // If this is the last action to process, return it to the InstCombiner.
923 // Otherwise, we insert it before the UDiv and record it so that we may
924 // use it as part of a joining action (i.e., a SelectInst).
926 Inst->insertBefore(&I);
927 UDivActions[i].FoldResult = Inst;
935 Instruction *InstCombiner::visitSDiv(BinaryOperator &I) {
936 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
938 if (Value *V = SimplifySDivInst(Op0, Op1, TD))
939 return ReplaceInstUsesWith(I, V);
941 // Handle the integer div common cases
942 if (Instruction *Common = commonIDivTransforms(I))
945 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
947 if (RHS->isAllOnesValue())
948 return BinaryOperator::CreateNeg(Op0);
950 // sdiv X, C --> ashr exact X, log2(C)
951 if (I.isExact() && RHS->getValue().isNonNegative() &&
952 RHS->getValue().isPowerOf2()) {
953 Value *ShAmt = llvm::ConstantInt::get(RHS->getType(),
954 RHS->getValue().exactLogBase2());
955 return BinaryOperator::CreateExactAShr(Op0, ShAmt, I.getName());
958 // -X/C --> X/-C provided the negation doesn't overflow.
959 if (SubOperator *Sub = dyn_cast<SubOperator>(Op0))
960 if (match(Sub->getOperand(0), m_Zero()) && Sub->hasNoSignedWrap())
961 return BinaryOperator::CreateSDiv(Sub->getOperand(1),
962 ConstantExpr::getNeg(RHS));
965 // If the sign bits of both operands are zero (i.e. we can prove they are
966 // unsigned inputs), turn this into a udiv.
967 if (I.getType()->isIntegerTy()) {
968 APInt Mask(APInt::getSignBit(I.getType()->getPrimitiveSizeInBits()));
969 if (MaskedValueIsZero(Op0, Mask)) {
970 if (MaskedValueIsZero(Op1, Mask)) {
971 // X sdiv Y -> X udiv Y, iff X and Y don't have sign bit set
972 return BinaryOperator::CreateUDiv(Op0, Op1, I.getName());
975 if (match(Op1, m_Shl(m_Power2(), m_Value()))) {
976 // X sdiv (1 << Y) -> X udiv (1 << Y) ( -> X u>> Y)
977 // Safe because the only negative value (1 << Y) can take on is
978 // INT_MIN, and X sdiv INT_MIN == X udiv INT_MIN == 0 if X doesn't have
980 return BinaryOperator::CreateUDiv(Op0, Op1, I.getName());
988 /// CvtFDivConstToReciprocal tries to convert X/C into X*1/C if C not a special
990 /// 1) 1/C is exact, or
991 /// 2) reciprocal is allowed.
992 /// If the conversion was successful, the simplified expression "X * 1/C" is
993 /// returned; otherwise, NULL is returned.
995 static Instruction *CvtFDivConstToReciprocal(Value *Dividend,
997 bool AllowReciprocal) {
998 if (!isa<ConstantFP>(Divisor)) // TODO: handle vectors.
1001 const APFloat &FpVal = cast<ConstantFP>(Divisor)->getValueAPF();
1002 APFloat Reciprocal(FpVal.getSemantics());
1003 bool Cvt = FpVal.getExactInverse(&Reciprocal);
1005 if (!Cvt && AllowReciprocal && FpVal.isFiniteNonZero()) {
1006 Reciprocal = APFloat(FpVal.getSemantics(), 1.0f);
1007 (void)Reciprocal.divide(FpVal, APFloat::rmNearestTiesToEven);
1008 Cvt = !Reciprocal.isDenormal();
1015 R = ConstantFP::get(Dividend->getType()->getContext(), Reciprocal);
1016 return BinaryOperator::CreateFMul(Dividend, R);
1019 Instruction *InstCombiner::visitFDiv(BinaryOperator &I) {
1020 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
1022 if (Value *V = SimplifyFDivInst(Op0, Op1, TD))
1023 return ReplaceInstUsesWith(I, V);
1025 if (isa<Constant>(Op0))
1026 if (SelectInst *SI = dyn_cast<SelectInst>(Op1))
1027 if (Instruction *R = FoldOpIntoSelect(I, SI))
1030 bool AllowReassociate = I.hasUnsafeAlgebra();
1031 bool AllowReciprocal = I.hasAllowReciprocal();
1033 if (Constant *Op1C = dyn_cast<Constant>(Op1)) {
1034 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
1035 if (Instruction *R = FoldOpIntoSelect(I, SI))
1038 if (AllowReassociate) {
1040 Constant *C2 = Op1C;
1042 Instruction *Res = 0;
1044 if (match(Op0, m_FMul(m_Value(X), m_Constant(C1)))) {
1045 // (X*C1)/C2 => X * (C1/C2)
1047 Constant *C = ConstantExpr::getFDiv(C1, C2);
1049 Res = BinaryOperator::CreateFMul(X, C);
1050 } else if (match(Op0, m_FDiv(m_Value(X), m_Constant(C1)))) {
1051 // (X/C1)/C2 => X /(C2*C1) [=> X * 1/(C2*C1) if reciprocal is allowed]
1053 Constant *C = ConstantExpr::getFMul(C1, C2);
1054 if (isNormalFp(C)) {
1055 Res = CvtFDivConstToReciprocal(X, C, AllowReciprocal);
1057 Res = BinaryOperator::CreateFDiv(X, C);
1062 Res->setFastMathFlags(I.getFastMathFlags());
1068 if (Instruction *T = CvtFDivConstToReciprocal(Op0, Op1C, AllowReciprocal)) {
1069 T->copyFastMathFlags(&I);
1076 if (AllowReassociate && isa<Constant>(Op0)) {
1077 Constant *C1 = cast<Constant>(Op0), *C2;
1080 bool CreateDiv = true;
1082 // C1 / (X*C2) => (C1/C2) / X
1083 if (match(Op1, m_FMul(m_Value(X), m_Constant(C2))))
1084 Fold = ConstantExpr::getFDiv(C1, C2);
1085 else if (match(Op1, m_FDiv(m_Value(X), m_Constant(C2)))) {
1086 // C1 / (X/C2) => (C1*C2) / X
1087 Fold = ConstantExpr::getFMul(C1, C2);
1088 } else if (match(Op1, m_FDiv(m_Constant(C2), m_Value(X)))) {
1089 // C1 / (C2/X) => (C1/C2) * X
1090 Fold = ConstantExpr::getFDiv(C1, C2);
1094 if (Fold && isNormalFp(Fold)) {
1095 Instruction *R = CreateDiv ? BinaryOperator::CreateFDiv(Fold, X)
1096 : BinaryOperator::CreateFMul(X, Fold);
1097 R->setFastMathFlags(I.getFastMathFlags());
1103 if (AllowReassociate) {
1106 Instruction *SimpR = 0;
1108 if (Op0->hasOneUse() && match(Op0, m_FDiv(m_Value(X), m_Value(Y)))) {
1109 // (X/Y) / Z => X / (Y*Z)
1111 if (!isa<Constant>(Y) || !isa<Constant>(Op1)) {
1112 NewInst = Builder->CreateFMul(Y, Op1);
1113 SimpR = BinaryOperator::CreateFDiv(X, NewInst);
1115 } else if (Op1->hasOneUse() && match(Op1, m_FDiv(m_Value(X), m_Value(Y)))) {
1116 // Z / (X/Y) => Z*Y / X
1118 if (!isa<Constant>(Y) || !isa<Constant>(Op0)) {
1119 NewInst = Builder->CreateFMul(Op0, Y);
1120 SimpR = BinaryOperator::CreateFDiv(NewInst, X);
1125 if (Instruction *T = dyn_cast<Instruction>(NewInst))
1126 T->setDebugLoc(I.getDebugLoc());
1127 SimpR->setFastMathFlags(I.getFastMathFlags());
1135 /// This function implements the transforms common to both integer remainder
1136 /// instructions (urem and srem). It is called by the visitors to those integer
1137 /// remainder instructions.
1138 /// @brief Common integer remainder transforms
1139 Instruction *InstCombiner::commonIRemTransforms(BinaryOperator &I) {
1140 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
1142 // The RHS is known non-zero.
1143 if (Value *V = simplifyValueKnownNonZero(I.getOperand(1), *this)) {
1148 // Handle cases involving: rem X, (select Cond, Y, Z)
1149 if (isa<SelectInst>(Op1) && SimplifyDivRemOfSelect(I))
1152 if (isa<ConstantInt>(Op1)) {
1153 if (Instruction *Op0I = dyn_cast<Instruction>(Op0)) {
1154 if (SelectInst *SI = dyn_cast<SelectInst>(Op0I)) {
1155 if (Instruction *R = FoldOpIntoSelect(I, SI))
1157 } else if (isa<PHINode>(Op0I)) {
1158 if (Instruction *NV = FoldOpIntoPhi(I))
1162 // See if we can fold away this rem instruction.
1163 if (SimplifyDemandedInstructionBits(I))
1171 Instruction *InstCombiner::visitURem(BinaryOperator &I) {
1172 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
1174 if (Value *V = SimplifyURemInst(Op0, Op1, TD))
1175 return ReplaceInstUsesWith(I, V);
1177 if (Instruction *common = commonIRemTransforms(I))
1180 // (zext A) urem (zext B) --> zext (A urem B)
1181 if (ZExtInst *ZOp0 = dyn_cast<ZExtInst>(Op0))
1182 if (Value *ZOp1 = dyn_castZExtVal(Op1, ZOp0->getSrcTy()))
1183 return new ZExtInst(Builder->CreateURem(ZOp0->getOperand(0), ZOp1),
1186 // X urem Y -> X and Y-1, where Y is a power of 2,
1187 if (isKnownToBeAPowerOfTwo(Op1, /*OrZero*/true)) {
1188 Constant *N1 = Constant::getAllOnesValue(I.getType());
1189 Value *Add = Builder->CreateAdd(Op1, N1);
1190 return BinaryOperator::CreateAnd(Op0, Add);
1193 // 1 urem X -> zext(X != 1)
1194 if (match(Op0, m_One())) {
1195 Value *Cmp = Builder->CreateICmpNE(Op1, Op0);
1196 Value *Ext = Builder->CreateZExt(Cmp, I.getType());
1197 return ReplaceInstUsesWith(I, Ext);
1203 Instruction *InstCombiner::visitSRem(BinaryOperator &I) {
1204 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
1206 if (Value *V = SimplifySRemInst(Op0, Op1, TD))
1207 return ReplaceInstUsesWith(I, V);
1209 // Handle the integer rem common cases
1210 if (Instruction *Common = commonIRemTransforms(I))
1213 if (Value *RHSNeg = dyn_castNegVal(Op1))
1214 if (!isa<Constant>(RHSNeg) ||
1215 (isa<ConstantInt>(RHSNeg) &&
1216 cast<ConstantInt>(RHSNeg)->getValue().isStrictlyPositive())) {
1218 Worklist.AddValue(I.getOperand(1));
1219 I.setOperand(1, RHSNeg);
1223 // If the sign bits of both operands are zero (i.e. we can prove they are
1224 // unsigned inputs), turn this into a urem.
1225 if (I.getType()->isIntegerTy()) {
1226 APInt Mask(APInt::getSignBit(I.getType()->getPrimitiveSizeInBits()));
1227 if (MaskedValueIsZero(Op1, Mask) && MaskedValueIsZero(Op0, Mask)) {
1228 // X srem Y -> X urem Y, iff X and Y don't have sign bit set
1229 return BinaryOperator::CreateURem(Op0, Op1, I.getName());
1233 // If it's a constant vector, flip any negative values positive.
1234 if (isa<ConstantVector>(Op1) || isa<ConstantDataVector>(Op1)) {
1235 Constant *C = cast<Constant>(Op1);
1236 unsigned VWidth = C->getType()->getVectorNumElements();
1238 bool hasNegative = false;
1239 bool hasMissing = false;
1240 for (unsigned i = 0; i != VWidth; ++i) {
1241 Constant *Elt = C->getAggregateElement(i);
1247 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Elt))
1248 if (RHS->isNegative())
1252 if (hasNegative && !hasMissing) {
1253 SmallVector<Constant *, 16> Elts(VWidth);
1254 for (unsigned i = 0; i != VWidth; ++i) {
1255 Elts[i] = C->getAggregateElement(i); // Handle undef, etc.
1256 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Elts[i])) {
1257 if (RHS->isNegative())
1258 Elts[i] = cast<ConstantInt>(ConstantExpr::getNeg(RHS));
1262 Constant *NewRHSV = ConstantVector::get(Elts);
1263 if (NewRHSV != C) { // Don't loop on -MININT
1264 Worklist.AddValue(I.getOperand(1));
1265 I.setOperand(1, NewRHSV);
1274 Instruction *InstCombiner::visitFRem(BinaryOperator &I) {
1275 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
1277 if (Value *V = SimplifyFRemInst(Op0, Op1, TD))
1278 return ReplaceInstUsesWith(I, V);
1280 // Handle cases involving: rem X, (select Cond, Y, Z)
1281 if (isa<SelectInst>(Op1) && SimplifyDivRemOfSelect(I))