/// Annotate the Shr in (X << IVOperand) >> C as exact using the
/// information from the IV's range. Returns true if anything changed, false
/// otherwise.
bool SimplifyIndvar::strengthenRightShift(BinaryOperator *BO,
Value *IVOperand) {
using namespace llvm::PatternMatch;
if (BO->getOpcode() == Instruction::Shl) {
bool Changed = false;
ConstantRange IVRange = SE->getUnsignedRange(SE->getSCEV(IVOperand));
for (auto *U : BO->users()) {
const APInt *C;
if (match(U,
m_AShr(m_Shl(m_Value(), m_Specific(IVOperand)), m_APInt(C))) ||
match(U,
m_LShr(m_Shl(m_Value(), m_Specific(IVOperand)), m_APInt(C)))) {
BinaryOperator *Shr = cast<BinaryOperator>(U);
if (!Shr->isExact() && IVRange.getUnsignedMin().uge(*C)) {
Shr->setIsExact(true);
Changed = true;
}
}
}
return Changed;
}
return false;
}
// X udiv 2^C -> X >> C
static Instruction *foldUDivPow2Cst(Value *Op0, Value *Op1,
const BinaryOperator &I, InstCombiner &IC) {
const APInt &C = cast<Constant>(Op1)->getUniqueInteger();
BinaryOperator *LShr = BinaryOperator::CreateLShr(
Op0, ConstantInt::get(Op0->getType(), C.logBase2()));
if (I.isExact()) LShr->setIsExact();
return LShr;
}
TEST(CloneInstruction, Exact) {
LLVMContext context;
Value *V = new Argument(Type::getInt32Ty(context));
BinaryOperator *SDiv = BinaryOperator::Create(Instruction::SDiv, V, V);
EXPECT_FALSE(cast<BinaryOperator>(SDiv->clone())->isExact());
SDiv->setIsExact(true);
EXPECT_TRUE(cast<BinaryOperator>(SDiv->clone())->isExact());
}
Value *AMDGPUCodeGenPrepare::copyFlags(
const BinaryOperator &I, Value *V) const {
assert(isa<BinaryOperator>(V) && "V must be binary operation");
BinaryOperator *BinOp = cast<BinaryOperator>(V);
if (isa<OverflowingBinaryOperator>(BinOp)) {
BinOp->setHasNoSignedWrap(I.hasNoSignedWrap());
BinOp->setHasNoUnsignedWrap(I.hasNoUnsignedWrap());
} else if (isa<PossiblyExactOperator>(BinOp))
BinOp->setIsExact(I.isExact());
return V;
}
// X udiv (C1 << N), where C1 is "1<<C2" --> X >> (N+C2)
static Instruction *foldUDivShl(Value *Op0, Value *Op1, const BinaryOperator &I,
InstCombiner &IC) {
Instruction *ShiftLeft = cast<Instruction>(Op1);
if (isa<ZExtInst>(ShiftLeft))
ShiftLeft = cast<Instruction>(ShiftLeft->getOperand(0));
const APInt &CI =
cast<Constant>(ShiftLeft->getOperand(0))->getUniqueInteger();
Value *N = ShiftLeft->getOperand(1);
if (CI != 1)
N = IC.Builder->CreateAdd(N, ConstantInt::get(N->getType(), CI.logBase2()));
if (ZExtInst *Z = dyn_cast<ZExtInst>(Op1))
N = IC.Builder->CreateZExt(N, Z->getDestTy());
BinaryOperator *LShr = BinaryOperator::CreateLShr(Op0, N);
if (I.isExact()) LShr->setIsExact();
return LShr;
}
/// FoldSelectIntoOp - Try fold the select into one of the operands to
/// facilitate further optimization.
Instruction *InstCombiner::FoldSelectIntoOp(SelectInst &SI, Value *TrueVal,
Value *FalseVal) {
// See the comment above GetSelectFoldableOperands for a description of the
// transformation we are doing here.
if (Instruction *TVI = dyn_cast<Instruction>(TrueVal)) {
if (TVI->hasOneUse() && TVI->getNumOperands() == 2 &&
!isa<Constant>(FalseVal)) {
if (unsigned SFO = GetSelectFoldableOperands(TVI)) {
unsigned OpToFold = 0;
if ((SFO & 1) && FalseVal == TVI->getOperand(0)) {
OpToFold = 1;
} else if ((SFO & 2) && FalseVal == TVI->getOperand(1)) {
OpToFold = 2;
}
if (OpToFold) {
Constant *C = GetSelectFoldableConstant(TVI);
Value *OOp = TVI->getOperand(2-OpToFold);
// Avoid creating select between 2 constants unless it's selecting
// between 0, 1 and -1.
if (!isa<Constant>(OOp) || isSelect01(C, cast<Constant>(OOp))) {
Value *NewSel = Builder->CreateSelect(SI.getCondition(), OOp, C);
NewSel->takeName(TVI);
BinaryOperator *TVI_BO = cast<BinaryOperator>(TVI);
BinaryOperator *BO = BinaryOperator::Create(TVI_BO->getOpcode(),
FalseVal, NewSel);
if (isa<PossiblyExactOperator>(BO))
BO->setIsExact(TVI_BO->isExact());
if (isa<OverflowingBinaryOperator>(BO)) {
BO->setHasNoUnsignedWrap(TVI_BO->hasNoUnsignedWrap());
BO->setHasNoSignedWrap(TVI_BO->hasNoSignedWrap());
}
return BO;
}
}
}
}
}
if (Instruction *FVI = dyn_cast<Instruction>(FalseVal)) {
if (FVI->hasOneUse() && FVI->getNumOperands() == 2 &&
!isa<Constant>(TrueVal)) {
if (unsigned SFO = GetSelectFoldableOperands(FVI)) {
unsigned OpToFold = 0;
if ((SFO & 1) && TrueVal == FVI->getOperand(0)) {
OpToFold = 1;
} else if ((SFO & 2) && TrueVal == FVI->getOperand(1)) {
OpToFold = 2;
}
if (OpToFold) {
Constant *C = GetSelectFoldableConstant(FVI);
Value *OOp = FVI->getOperand(2-OpToFold);
// Avoid creating select between 2 constants unless it's selecting
// between 0, 1 and -1.
if (!isa<Constant>(OOp) || isSelect01(C, cast<Constant>(OOp))) {
Value *NewSel = Builder->CreateSelect(SI.getCondition(), C, OOp);
NewSel->takeName(FVI);
BinaryOperator *FVI_BO = cast<BinaryOperator>(FVI);
BinaryOperator *BO = BinaryOperator::Create(FVI_BO->getOpcode(),
TrueVal, NewSel);
if (isa<PossiblyExactOperator>(BO))
BO->setIsExact(FVI_BO->isExact());
if (isa<OverflowingBinaryOperator>(BO)) {
BO->setHasNoUnsignedWrap(FVI_BO->hasNoUnsignedWrap());
BO->setHasNoSignedWrap(FVI_BO->hasNoSignedWrap());
}
return BO;
}
}
}
}
}
return nullptr;
}
/// Rebuild a new instruction just like 'I' but with the new operands given.
/// In the event of type mismatch, the type of the operands is correct.
static Value *BuildNew(Instruction *I, ArrayRef<Value*> NewOps) {
// We don't want to use the IRBuilder here because we want the replacement
// instructions to appear next to 'I', not the builder's insertion point.
switch (I->getOpcode()) {
case Instruction::Add:
case Instruction::FAdd:
case Instruction::Sub:
case Instruction::FSub:
case Instruction::Mul:
case Instruction::FMul:
case Instruction::UDiv:
case Instruction::SDiv:
case Instruction::FDiv:
case Instruction::URem:
case Instruction::SRem:
case Instruction::FRem:
case Instruction::Shl:
case Instruction::LShr:
case Instruction::AShr:
case Instruction::And:
case Instruction::Or:
case Instruction::Xor: {
BinaryOperator *BO = cast<BinaryOperator>(I);
assert(NewOps.size() == 2 && "binary operator with #ops != 2");
BinaryOperator *New =
BinaryOperator::Create(cast<BinaryOperator>(I)->getOpcode(),
NewOps[0], NewOps[1], "", BO);
if (isa<OverflowingBinaryOperator>(BO)) {
New->setHasNoUnsignedWrap(BO->hasNoUnsignedWrap());
New->setHasNoSignedWrap(BO->hasNoSignedWrap());
}
if (isa<PossiblyExactOperator>(BO)) {
New->setIsExact(BO->isExact());
}
return New;
}
case Instruction::ICmp:
assert(NewOps.size() == 2 && "icmp with #ops != 2");
return new ICmpInst(I, cast<ICmpInst>(I)->getPredicate(),
NewOps[0], NewOps[1]);
case Instruction::FCmp:
assert(NewOps.size() == 2 && "fcmp with #ops != 2");
return new FCmpInst(I, cast<FCmpInst>(I)->getPredicate(),
NewOps[0], NewOps[1]);
case Instruction::Trunc:
case Instruction::ZExt:
case Instruction::SExt:
case Instruction::FPToUI:
case Instruction::FPToSI:
case Instruction::UIToFP:
case Instruction::SIToFP:
case Instruction::FPTrunc:
case Instruction::FPExt: {
// It's possible that the mask has a different number of elements from
// the original cast. We recompute the destination type to match the mask.
Type *DestTy =
VectorType::get(I->getType()->getScalarType(),
NewOps[0]->getType()->getVectorNumElements());
assert(NewOps.size() == 1 && "cast with #ops != 1");
return CastInst::Create(cast<CastInst>(I)->getOpcode(), NewOps[0], DestTy,
"", I);
}
case Instruction::GetElementPtr: {
Value *Ptr = NewOps[0];
ArrayRef<Value*> Idx = NewOps.slice(1);
GetElementPtrInst *GEP = GetElementPtrInst::Create(Ptr, Idx, "", I);
GEP->setIsInBounds(cast<GetElementPtrInst>(I)->isInBounds());
return GEP;
}
}
llvm_unreachable("failed to rebuild vector instructions");
}
Instruction *InstCombiner::visitUDiv(BinaryOperator &I) {
Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
if (Value *V = SimplifyUDivInst(Op0, Op1, TD))
return ReplaceInstUsesWith(I, V);
// Handle the integer div common cases
if (Instruction *Common = commonIDivTransforms(I))
return Common;
{
// X udiv 2^C -> X >> C
// Check to see if this is an unsigned division with an exact power of 2,
// if so, convert to a right shift.
const APInt *C;
if (match(Op1, m_Power2(C))) {
BinaryOperator *LShr =
BinaryOperator::CreateLShr(Op0,
ConstantInt::get(Op0->getType(),
C->logBase2()));
if (I.isExact()) LShr->setIsExact();
return LShr;
}
}
if (ConstantInt *C = dyn_cast<ConstantInt>(Op1)) {
// X udiv C, where C >= signbit
if (C->getValue().isNegative()) {
Value *IC = Builder->CreateICmpULT(Op0, C);
return SelectInst::Create(IC, Constant::getNullValue(I.getType()),
ConstantInt::get(I.getType(), 1));
}
}
// X udiv (C1 << N), where C1 is "1<<C2" --> X >> (N+C2)
{ const APInt *CI; Value *N;
if (match(Op1, m_Shl(m_Power2(CI), m_Value(N)))) {
if (*CI != 1)
N = Builder->CreateAdd(N, ConstantInt::get(I.getType(),CI->logBase2()));
if (I.isExact())
return BinaryOperator::CreateExactLShr(Op0, N);
return BinaryOperator::CreateLShr(Op0, N);
}
}
// udiv X, (Select Cond, C1, C2) --> Select Cond, (shr X, C1), (shr X, C2)
// where C1&C2 are powers of two.
{ Value *Cond; const APInt *C1, *C2;
if (match(Op1, m_Select(m_Value(Cond), m_Power2(C1), m_Power2(C2)))) {
// Construct the "on true" case of the select
Value *TSI = Builder->CreateLShr(Op0, C1->logBase2(), Op1->getName()+".t",
I.isExact());
// Construct the "on false" case of the select
Value *FSI = Builder->CreateLShr(Op0, C2->logBase2(), Op1->getName()+".f",
I.isExact());
// construct the select instruction and return it.
return SelectInst::Create(Cond, TSI, FSI);
}
}
// (zext A) udiv (zext B) --> zext (A udiv B)
if (ZExtInst *ZOp0 = dyn_cast<ZExtInst>(Op0))
if (Value *ZOp1 = dyn_castZExtVal(Op1, ZOp0->getSrcTy()))
return new ZExtInst(Builder->CreateUDiv(ZOp0->getOperand(0), ZOp1, "div",
I.isExact()),
I.getType());
return 0;
}
static Value *genVal(int &Budget, unsigned Width, bool ConstOK, bool ArgOK) {
if (Budget > 0 && Choose(2)) {
if (Verbose)
errs() << "adding a phi, budget = " << Budget << "\n";
--Budget;
Value *V = Builder->CreatePHI(Type::getIntNTy(*C, Width), N);
Vals.push_back(V);
return V;
}
if (Budget > 0 && Budget != N && Choose(2)) {
if (Verbose)
errs() << "adding a branch, budget = " << Budget << "\n";
--Budget;
BranchInst *Br;
if (0 && Builder->GetInsertBlock()->size() > 0 && Choose(2)) {
Br = Builder->CreateBr(BBs[0]);
} else {
Value *C = genVal(Budget, 1, false, ArgOK);
Br = Builder->CreateCondBr(C, BBs[0], BBs[0]);
}
Branches.push_back(Br);
BasicBlock *BB = BasicBlock::Create(*C, "", F);
BBs.push_back(BB);
Builder->SetInsertPoint(BB);
return genVal(Budget, Width, ConstOK, ArgOK);
}
if (Budget > 0 && Width == W && Choose(2)) {
if (Verbose)
errs() << "adding a select with width = " << Width
<< " and budget = " << Budget << "\n";
--Budget;
Value *L, *R;
genLR(L, R, Budget, Width);
Value *C = genVal(Budget, 1, false);
Value *V = Builder->CreateSelect(C, L, R);
Vals.push_back(V);
return V;
}
if (Budget > 0 && Width == 1 && Choose(2)) {
if (Verbose)
errs() << "adding an icmp with width = " << Width
<< " and budget = " << Budget << "\n";
--Budget;
Value *L, *R;
genLR(L, R, Budget, Width);
CmpInst::Predicate P;
switch (OneICmp ? 0 : Choose(10)) {
case 0:
P = CmpInst::ICMP_EQ;
break;
case 1:
P = CmpInst::ICMP_NE;
break;
case 2:
P = CmpInst::ICMP_UGT;
break;
case 3:
P = CmpInst::ICMP_UGE;
break;
case 4:
P = CmpInst::ICMP_ULT;
break;
case 5:
P = CmpInst::ICMP_ULE;
break;
case 6:
P = CmpInst::ICMP_SGT;
break;
case 7:
P = CmpInst::ICMP_SGE;
break;
case 8:
P = CmpInst::ICMP_SLT;
break;
case 9:
P = CmpInst::ICMP_SLE;
break;
}
Value *V = Builder->CreateICmp(P, L, R);
Vals.push_back(V);
return V;
}
if (Budget > 0 && Width == W && Choose(2)) {
unsigned OldW = Width * 2;
if (Verbose)
errs() << "adding a trunc from " << OldW << " to " << Width
<< " and budget = " << Budget << "\n";
--Budget;
Value *V = Builder->CreateTrunc(genVal(Budget, OldW, false),
Type::getIntNTy(*C, Width));
Vals.push_back(V);
return V;
}
if (Budget > 0 && Width == W && Choose(2)) {
unsigned OldW = Width / 2;
if (OldW > 1 && Choose(2))
OldW = 1;
if (Verbose)
errs() << "adding a zext from " << OldW << " to " << Width
<< " and budget = " << Budget << "\n";
--Budget;
Value *V;
if (Choose(2))
V = Builder->CreateZExt(genVal(Budget, OldW, false),
Type::getIntNTy(*C, Width));
else
V = Builder->CreateSExt(genVal(Budget, OldW, false),
Type::getIntNTy(*C, Width));
Vals.push_back(V);
return V;
}
if (Budget > 0 && Width == W && Choose(2)) {
if (Verbose)
errs() << "adding a binop with width = " << Width
<< " and budget = " << Budget << "\n";
--Budget;
Instruction::BinaryOps Op;
switch (OneBinop ? 0 : Choose(10)) {
case 0:
Op = Instruction::Add;
break;
case 1:
Op = Instruction::Sub;
break;
case 2:
Op = Instruction::Mul;
break;
case 3:
Op = Instruction::SDiv;
break;
case 4:
Op = Instruction::UDiv;
break;
case 5:
Op = Instruction::SRem;
break;
case 6:
Op = Instruction::URem;
break;
case 7:
Op = Instruction::And;
break;
case 8:
Op = Instruction::Or;
break;
case 9:
Op = Instruction::Xor;
break;
}
Value *L, *R;
genLR(L, R, Budget, Width);
Value *V = Builder->CreateBinOp(Op, L, R);
if (!NoUB) {
if ((Op == Instruction::Add || Op == Instruction::Sub ||
Op == Instruction::Mul || Op == Instruction::Shl) &&
Choose(2)) {
BinaryOperator *B = cast<BinaryOperator>(V);
B->setHasNoSignedWrap(true);
}
if ((Op == Instruction::Add || Op == Instruction::Sub ||
Op == Instruction::Mul || Op == Instruction::Shl) &&
Choose(2)) {
BinaryOperator *B = cast<BinaryOperator>(V);
B->setHasNoUnsignedWrap(true);
}
if ((Op == Instruction::UDiv || Op == Instruction::SDiv ||
Op == Instruction::LShr || Op == Instruction::AShr) &&
Choose(2)) {
BinaryOperator *B = cast<BinaryOperator>(V);
B->setIsExact(true);
}
}
Vals.push_back(V);
return V;
}
if (ConstOK && Choose(2)) {
if (Verbose)
errs() << "adding a const with width = " << Width
<< " and budget = " << Budget << "\n";
if (OneConst) {
return ConstantInt::get(*C, APInt(Width, 1));
} else {
int n = Choose((1 << Width) + 1);
if (n == (1 << Width))
return UndefValue::get(Type::getIntNTy(*C, Width));
else
return ConstantInt::get(*C, APInt(Width, n));
}
}
if (Verbose)
errs() << "using existing val with width = " << Width
<< " and budget = " << Budget << " and ArgOK = " << ArgOK << "\n";
std::vector<Value *> Vs;
for (auto &it : Vals)
if (it->getType()->getPrimitiveSizeInBits() == Width)
Vs.push_back(it);
unsigned choices = Vs.size() + (ArgOK ? 1 : 0);
if (choices == 0)
exit(0);
unsigned which = Choose(choices);
if (which == Vs.size()) {
Value *V = 0;
for (auto it = F->arg_begin(); it != F->arg_end(); ++it) {
if (UsedArgs.find(it) == UsedArgs.end() &&
it->getType()->getPrimitiveSizeInBits() == Width) {
UsedArgs.insert(it);
V = it;
Vals.push_back(V);
break;
}
}
ensure(V);
return V;
} else {
return Vs.at(which);
}
}
Instruction *InstCombiner::FoldShiftByConstant(Value *Op0, Constant *Op1,
BinaryOperator &I) {
bool isLeftShift = I.getOpcode() == Instruction::Shl;
ConstantInt *COp1 = nullptr;
if (ConstantDataVector *CV = dyn_cast<ConstantDataVector>(Op1))
COp1 = dyn_cast_or_null<ConstantInt>(CV->getSplatValue());
else if (ConstantVector *CV = dyn_cast<ConstantVector>(Op1))
COp1 = dyn_cast_or_null<ConstantInt>(CV->getSplatValue());
else
COp1 = dyn_cast<ConstantInt>(Op1);
if (!COp1)
return nullptr;
// See if we can propagate this shift into the input, this covers the trivial
// cast of lshr(shl(x,c1),c2) as well as other more complex cases.
if (I.getOpcode() != Instruction::AShr &&
CanEvaluateShifted(Op0, COp1->getZExtValue(), isLeftShift, *this)) {
DEBUG(dbgs() << "ICE: GetShiftedValue propagating shift through expression"
" to eliminate shift:\n IN: " << *Op0 << "\n SH: " << I <<"\n");
return ReplaceInstUsesWith(I,
GetShiftedValue(Op0, COp1->getZExtValue(), isLeftShift, *this));
}
// See if we can simplify any instructions used by the instruction whose sole
// purpose is to compute bits we don't care about.
uint32_t TypeBits = Op0->getType()->getScalarSizeInBits();
assert(!COp1->uge(TypeBits) &&
"Shift over the type width should have been removed already");
// ((X*C1) << C2) == (X * (C1 << C2))
if (BinaryOperator *BO = dyn_cast<BinaryOperator>(Op0))
if (BO->getOpcode() == Instruction::Mul && isLeftShift)
if (Constant *BOOp = dyn_cast<Constant>(BO->getOperand(1)))
return BinaryOperator::CreateMul(BO->getOperand(0),
ConstantExpr::getShl(BOOp, Op1));
// Try to fold constant and into select arguments.
if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
if (Instruction *R = FoldOpIntoSelect(I, SI))
return R;
if (isa<PHINode>(Op0))
if (Instruction *NV = FoldOpIntoPhi(I))
return NV;
// Fold shift2(trunc(shift1(x,c1)), c2) -> trunc(shift2(shift1(x,c1),c2))
if (TruncInst *TI = dyn_cast<TruncInst>(Op0)) {
Instruction *TrOp = dyn_cast<Instruction>(TI->getOperand(0));
// If 'shift2' is an ashr, we would have to get the sign bit into a funny
// place. Don't try to do this transformation in this case. Also, we
// require that the input operand is a shift-by-constant so that we have
// confidence that the shifts will get folded together. We could do this
// xform in more cases, but it is unlikely to be profitable.
if (TrOp && I.isLogicalShift() && TrOp->isShift() &&
isa<ConstantInt>(TrOp->getOperand(1))) {
// Okay, we'll do this xform. Make the shift of shift.
Constant *ShAmt = ConstantExpr::getZExt(COp1, TrOp->getType());
// (shift2 (shift1 & 0x00FF), c2)
Value *NSh = Builder->CreateBinOp(I.getOpcode(), TrOp, ShAmt,I.getName());
// For logical shifts, the truncation has the effect of making the high
// part of the register be zeros. Emulate this by inserting an AND to
// clear the top bits as needed. This 'and' will usually be zapped by
// other xforms later if dead.
unsigned SrcSize = TrOp->getType()->getScalarSizeInBits();
unsigned DstSize = TI->getType()->getScalarSizeInBits();
APInt MaskV(APInt::getLowBitsSet(SrcSize, DstSize));
// The mask we constructed says what the trunc would do if occurring
// between the shifts. We want to know the effect *after* the second
// shift. We know that it is a logical shift by a constant, so adjust the
// mask as appropriate.
if (I.getOpcode() == Instruction::Shl)
MaskV <<= COp1->getZExtValue();
else {
assert(I.getOpcode() == Instruction::LShr && "Unknown logical shift");
MaskV = MaskV.lshr(COp1->getZExtValue());
}
// shift1 & 0x00FF
Value *And = Builder->CreateAnd(NSh,
ConstantInt::get(I.getContext(), MaskV),
TI->getName());
// Return the value truncated to the interesting size.
return new TruncInst(And, I.getType());
}
}
if (Op0->hasOneUse()) {
if (BinaryOperator *Op0BO = dyn_cast<BinaryOperator>(Op0)) {
// Turn ((X >> C) + Y) << C -> (X + (Y << C)) & (~0 << C)
Value *V1, *V2;
ConstantInt *CC;
switch (Op0BO->getOpcode()) {
default: break;
case Instruction::Add:
case Instruction::And:
case Instruction::Or:
case Instruction::Xor: {
// These operators commute.
// Turn (Y + (X >> C)) << C -> (X + (Y << C)) & (~0 << C)
if (isLeftShift && Op0BO->getOperand(1)->hasOneUse() &&
match(Op0BO->getOperand(1), m_Shr(m_Value(V1),
m_Specific(Op1)))) {
Value *YS = // (Y << C)
Builder->CreateShl(Op0BO->getOperand(0), Op1, Op0BO->getName());
// (X + (Y << C))
Value *X = Builder->CreateBinOp(Op0BO->getOpcode(), YS, V1,
Op0BO->getOperand(1)->getName());
uint32_t Op1Val = COp1->getLimitedValue(TypeBits);
APInt Bits = APInt::getHighBitsSet(TypeBits, TypeBits - Op1Val);
Constant *Mask = ConstantInt::get(I.getContext(), Bits);
if (VectorType *VT = dyn_cast<VectorType>(X->getType()))
Mask = ConstantVector::getSplat(VT->getNumElements(), Mask);
return BinaryOperator::CreateAnd(X, Mask);
}
// Turn (Y + ((X >> C) & CC)) << C -> ((X & (CC << C)) + (Y << C))
Value *Op0BOOp1 = Op0BO->getOperand(1);
if (isLeftShift && Op0BOOp1->hasOneUse() &&
match(Op0BOOp1,
m_And(m_OneUse(m_Shr(m_Value(V1), m_Specific(Op1))),
m_ConstantInt(CC)))) {
Value *YS = // (Y << C)
Builder->CreateShl(Op0BO->getOperand(0), Op1,
Op0BO->getName());
// X & (CC << C)
Value *XM = Builder->CreateAnd(V1, ConstantExpr::getShl(CC, Op1),
V1->getName()+".mask");
return BinaryOperator::Create(Op0BO->getOpcode(), YS, XM);
}
}
// FALL THROUGH.
case Instruction::Sub: {
// Turn ((X >> C) + Y) << C -> (X + (Y << C)) & (~0 << C)
if (isLeftShift && Op0BO->getOperand(0)->hasOneUse() &&
match(Op0BO->getOperand(0), m_Shr(m_Value(V1),
m_Specific(Op1)))) {
Value *YS = // (Y << C)
Builder->CreateShl(Op0BO->getOperand(1), Op1, Op0BO->getName());
// (X + (Y << C))
Value *X = Builder->CreateBinOp(Op0BO->getOpcode(), V1, YS,
Op0BO->getOperand(0)->getName());
uint32_t Op1Val = COp1->getLimitedValue(TypeBits);
APInt Bits = APInt::getHighBitsSet(TypeBits, TypeBits - Op1Val);
Constant *Mask = ConstantInt::get(I.getContext(), Bits);
if (VectorType *VT = dyn_cast<VectorType>(X->getType()))
Mask = ConstantVector::getSplat(VT->getNumElements(), Mask);
return BinaryOperator::CreateAnd(X, Mask);
}
// Turn (((X >> C)&CC) + Y) << C -> (X + (Y << C)) & (CC << C)
if (isLeftShift && Op0BO->getOperand(0)->hasOneUse() &&
match(Op0BO->getOperand(0),
m_And(m_OneUse(m_Shr(m_Value(V1), m_Value(V2))),
m_ConstantInt(CC))) && V2 == Op1) {
Value *YS = // (Y << C)
Builder->CreateShl(Op0BO->getOperand(1), Op1, Op0BO->getName());
// X & (CC << C)
Value *XM = Builder->CreateAnd(V1, ConstantExpr::getShl(CC, Op1),
V1->getName()+".mask");
return BinaryOperator::Create(Op0BO->getOpcode(), XM, YS);
}
break;
}
}
// If the operand is an bitwise operator with a constant RHS, and the
// shift is the only use, we can pull it out of the shift.
if (ConstantInt *Op0C = dyn_cast<ConstantInt>(Op0BO->getOperand(1))) {
bool isValid = true; // Valid only for And, Or, Xor
bool highBitSet = false; // Transform if high bit of constant set?
switch (Op0BO->getOpcode()) {
default: isValid = false; break; // Do not perform transform!
case Instruction::Add:
isValid = isLeftShift;
break;
case Instruction::Or:
case Instruction::Xor:
highBitSet = false;
break;
case Instruction::And:
highBitSet = true;
break;
}
// If this is a signed shift right, and the high bit is modified
// by the logical operation, do not perform the transformation.
// The highBitSet boolean indicates the value of the high bit of
// the constant which would cause it to be modified for this
// operation.
//
if (isValid && I.getOpcode() == Instruction::AShr)
isValid = Op0C->getValue()[TypeBits-1] == highBitSet;
if (isValid) {
Constant *NewRHS = ConstantExpr::get(I.getOpcode(), Op0C, Op1);
Value *NewShift =
Builder->CreateBinOp(I.getOpcode(), Op0BO->getOperand(0), Op1);
NewShift->takeName(Op0BO);
return BinaryOperator::Create(Op0BO->getOpcode(), NewShift,
NewRHS);
}
}
}
}
// Find out if this is a shift of a shift by a constant.
BinaryOperator *ShiftOp = dyn_cast<BinaryOperator>(Op0);
if (ShiftOp && !ShiftOp->isShift())
ShiftOp = nullptr;
if (ShiftOp && isa<ConstantInt>(ShiftOp->getOperand(1))) {
// This is a constant shift of a constant shift. Be careful about hiding
// shl instructions behind bit masks. They are used to represent multiplies
// by a constant, and it is important that simple arithmetic expressions
// are still recognizable by scalar evolution.
//
// The transforms applied to shl are very similar to the transforms applied
// to mul by constant. We can be more aggressive about optimizing right
// shifts.
//
// Combinations of right and left shifts will still be optimized in
// DAGCombine where scalar evolution no longer applies.
ConstantInt *ShiftAmt1C = cast<ConstantInt>(ShiftOp->getOperand(1));
uint32_t ShiftAmt1 = ShiftAmt1C->getLimitedValue(TypeBits);
uint32_t ShiftAmt2 = COp1->getLimitedValue(TypeBits);
assert(ShiftAmt2 != 0 && "Should have been simplified earlier");
if (ShiftAmt1 == 0) return nullptr; // Will be simplified in the future.
Value *X = ShiftOp->getOperand(0);
IntegerType *Ty = cast<IntegerType>(I.getType());
// Check for (X << c1) << c2 and (X >> c1) >> c2
if (I.getOpcode() == ShiftOp->getOpcode()) {
uint32_t AmtSum = ShiftAmt1+ShiftAmt2; // Fold into one big shift.
// If this is oversized composite shift, then unsigned shifts get 0, ashr
// saturates.
if (AmtSum >= TypeBits) {
if (I.getOpcode() != Instruction::AShr)
return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
AmtSum = TypeBits-1; // Saturate to 31 for i32 ashr.
}
return BinaryOperator::Create(I.getOpcode(), X,
ConstantInt::get(Ty, AmtSum));
}
if (ShiftAmt1 == ShiftAmt2) {
// If we have ((X << C) >>u C), turn this into X & (-1 >>u C).
if (I.getOpcode() == Instruction::LShr &&
ShiftOp->getOpcode() == Instruction::Shl) {
APInt Mask(APInt::getLowBitsSet(TypeBits, TypeBits - ShiftAmt1));
return BinaryOperator::CreateAnd(X,
ConstantInt::get(I.getContext(), Mask));
}
} else if (ShiftAmt1 < ShiftAmt2) {
uint32_t ShiftDiff = ShiftAmt2-ShiftAmt1;
// (X >>?,exact C1) << C2 --> X << (C2-C1)
// The inexact version is deferred to DAGCombine so we don't hide shl
// behind a bit mask.
if (I.getOpcode() == Instruction::Shl &&
ShiftOp->getOpcode() != Instruction::Shl &&
ShiftOp->isExact()) {
assert(ShiftOp->getOpcode() == Instruction::LShr ||
ShiftOp->getOpcode() == Instruction::AShr);
ConstantInt *ShiftDiffCst = ConstantInt::get(Ty, ShiftDiff);
BinaryOperator *NewShl = BinaryOperator::Create(Instruction::Shl,
X, ShiftDiffCst);
NewShl->setHasNoUnsignedWrap(I.hasNoUnsignedWrap());
NewShl->setHasNoSignedWrap(I.hasNoSignedWrap());
return NewShl;
}
// (X << C1) >>u C2 --> X >>u (C2-C1) & (-1 >> C2)
if (I.getOpcode() == Instruction::LShr &&
ShiftOp->getOpcode() == Instruction::Shl) {
ConstantInt *ShiftDiffCst = ConstantInt::get(Ty, ShiftDiff);
// (X <<nuw C1) >>u C2 --> X >>u (C2-C1)
if (ShiftOp->hasNoUnsignedWrap()) {
BinaryOperator *NewLShr = BinaryOperator::Create(Instruction::LShr,
X, ShiftDiffCst);
NewLShr->setIsExact(I.isExact());
return NewLShr;
}
Value *Shift = Builder->CreateLShr(X, ShiftDiffCst);
APInt Mask(APInt::getLowBitsSet(TypeBits, TypeBits - ShiftAmt2));
return BinaryOperator::CreateAnd(Shift,
ConstantInt::get(I.getContext(),Mask));
}
// We can't handle (X << C1) >>s C2, it shifts arbitrary bits in. However,
// we can handle (X <<nsw C1) >>s C2 since it only shifts in sign bits.
if (I.getOpcode() == Instruction::AShr &&
ShiftOp->getOpcode() == Instruction::Shl) {
if (ShiftOp->hasNoSignedWrap()) {
// (X <<nsw C1) >>s C2 --> X >>s (C2-C1)
ConstantInt *ShiftDiffCst = ConstantInt::get(Ty, ShiftDiff);
BinaryOperator *NewAShr = BinaryOperator::Create(Instruction::AShr,
X, ShiftDiffCst);
NewAShr->setIsExact(I.isExact());
return NewAShr;
}
}
} else {
assert(ShiftAmt2 < ShiftAmt1);
uint32_t ShiftDiff = ShiftAmt1-ShiftAmt2;
// (X >>?exact C1) << C2 --> X >>?exact (C1-C2)
// The inexact version is deferred to DAGCombine so we don't hide shl
// behind a bit mask.
if (I.getOpcode() == Instruction::Shl &&
ShiftOp->getOpcode() != Instruction::Shl &&
ShiftOp->isExact()) {
ConstantInt *ShiftDiffCst = ConstantInt::get(Ty, ShiftDiff);
BinaryOperator *NewShr = BinaryOperator::Create(ShiftOp->getOpcode(),
X, ShiftDiffCst);
NewShr->setIsExact(true);
return NewShr;
}
// (X << C1) >>u C2 --> X << (C1-C2) & (-1 >> C2)
if (I.getOpcode() == Instruction::LShr &&
ShiftOp->getOpcode() == Instruction::Shl) {
ConstantInt *ShiftDiffCst = ConstantInt::get(Ty, ShiftDiff);
if (ShiftOp->hasNoUnsignedWrap()) {
// (X <<nuw C1) >>u C2 --> X <<nuw (C1-C2)
BinaryOperator *NewShl = BinaryOperator::Create(Instruction::Shl,
X, ShiftDiffCst);
NewShl->setHasNoUnsignedWrap(true);
return NewShl;
}
Value *Shift = Builder->CreateShl(X, ShiftDiffCst);
APInt Mask(APInt::getLowBitsSet(TypeBits, TypeBits - ShiftAmt2));
return BinaryOperator::CreateAnd(Shift,
ConstantInt::get(I.getContext(),Mask));
}
// We can't handle (X << C1) >>s C2, it shifts arbitrary bits in. However,
// we can handle (X <<nsw C1) >>s C2 since it only shifts in sign bits.
if (I.getOpcode() == Instruction::AShr &&
ShiftOp->getOpcode() == Instruction::Shl) {
if (ShiftOp->hasNoSignedWrap()) {
// (X <<nsw C1) >>s C2 --> X <<nsw (C1-C2)
ConstantInt *ShiftDiffCst = ConstantInt::get(Ty, ShiftDiff);
BinaryOperator *NewShl = BinaryOperator::Create(Instruction::Shl,
X, ShiftDiffCst);
NewShl->setHasNoSignedWrap(true);
return NewShl;
}
}
}
}
return nullptr;
}
/// GetShiftedValue - When CanEvaluateShifted returned true for an expression,
/// this value inserts the new computation that produces the shifted value.
static Value *GetShiftedValue(Value *V, unsigned NumBits, bool isLeftShift,
InstCombiner &IC) {
// We can always evaluate constants shifted.
if (Constant *C = dyn_cast<Constant>(V)) {
if (isLeftShift)
V = IC.Builder->CreateShl(C, NumBits);
else
V = IC.Builder->CreateLShr(C, NumBits);
// If we got a constantexpr back, try to simplify it with TD info.
if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V))
V = ConstantFoldConstantExpression(CE, IC.getDataLayout(),
IC.getTargetLibraryInfo());
return V;
}
Instruction *I = cast<Instruction>(V);
IC.Worklist.Add(I);
switch (I->getOpcode()) {
default: llvm_unreachable("Inconsistency with CanEvaluateShifted");
case Instruction::And:
case Instruction::Or:
case Instruction::Xor:
// Bitwise operators can all arbitrarily be arbitrarily evaluated shifted.
I->setOperand(0, GetShiftedValue(I->getOperand(0), NumBits,isLeftShift,IC));
I->setOperand(1, GetShiftedValue(I->getOperand(1), NumBits,isLeftShift,IC));
return I;
case Instruction::Shl: {
BinaryOperator *BO = cast<BinaryOperator>(I);
unsigned TypeWidth = BO->getType()->getScalarSizeInBits();
// We only accept shifts-by-a-constant in CanEvaluateShifted.
ConstantInt *CI = cast<ConstantInt>(BO->getOperand(1));
// We can always fold shl(c1)+shl(c2) -> shl(c1+c2).
if (isLeftShift) {
// If this is oversized composite shift, then unsigned shifts get 0.
unsigned NewShAmt = NumBits+CI->getZExtValue();
if (NewShAmt >= TypeWidth)
return Constant::getNullValue(I->getType());
BO->setOperand(1, ConstantInt::get(BO->getType(), NewShAmt));
BO->setHasNoUnsignedWrap(false);
BO->setHasNoSignedWrap(false);
return I;
}
// We turn shl(c)+lshr(c) -> and(c2) if the input doesn't already have
// zeros.
if (CI->getValue() == NumBits) {
APInt Mask(APInt::getLowBitsSet(TypeWidth, TypeWidth - NumBits));
V = IC.Builder->CreateAnd(BO->getOperand(0),
ConstantInt::get(BO->getContext(), Mask));
if (Instruction *VI = dyn_cast<Instruction>(V)) {
VI->moveBefore(BO);
VI->takeName(BO);
}
return V;
}
// We turn shl(c1)+shr(c2) -> shl(c3)+and(c4), but only when we know that
// the and won't be needed.
assert(CI->getZExtValue() > NumBits);
BO->setOperand(1, ConstantInt::get(BO->getType(),
CI->getZExtValue() - NumBits));
BO->setHasNoUnsignedWrap(false);
BO->setHasNoSignedWrap(false);
return BO;
}
case Instruction::LShr: {
BinaryOperator *BO = cast<BinaryOperator>(I);
unsigned TypeWidth = BO->getType()->getScalarSizeInBits();
// We only accept shifts-by-a-constant in CanEvaluateShifted.
ConstantInt *CI = cast<ConstantInt>(BO->getOperand(1));
// We can always fold lshr(c1)+lshr(c2) -> lshr(c1+c2).
if (!isLeftShift) {
// If this is oversized composite shift, then unsigned shifts get 0.
unsigned NewShAmt = NumBits+CI->getZExtValue();
if (NewShAmt >= TypeWidth)
return Constant::getNullValue(BO->getType());
BO->setOperand(1, ConstantInt::get(BO->getType(), NewShAmt));
BO->setIsExact(false);
return I;
}
// We turn lshr(c)+shl(c) -> and(c2) if the input doesn't already have
// zeros.
if (CI->getValue() == NumBits) {
APInt Mask(APInt::getHighBitsSet(TypeWidth, TypeWidth - NumBits));
V = IC.Builder->CreateAnd(I->getOperand(0),
ConstantInt::get(BO->getContext(), Mask));
if (Instruction *VI = dyn_cast<Instruction>(V)) {
VI->moveBefore(I);
VI->takeName(I);
}
return V;
}
// We turn lshr(c1)+shl(c2) -> lshr(c3)+and(c4), but only when we know that
// the and won't be needed.
assert(CI->getZExtValue() > NumBits);
BO->setOperand(1, ConstantInt::get(BO->getType(),
CI->getZExtValue() - NumBits));
BO->setIsExact(false);
return BO;
}
case Instruction::Select:
I->setOperand(1, GetShiftedValue(I->getOperand(1), NumBits,isLeftShift,IC));
I->setOperand(2, GetShiftedValue(I->getOperand(2), NumBits,isLeftShift,IC));
return I;
case Instruction::PHI: {
// We can change a phi if we can change all operands. Note that we never
// get into trouble with cyclic PHIs here because we only consider
// instructions with a single use.
PHINode *PN = cast<PHINode>(I);
for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
PN->setIncomingValue(i, GetShiftedValue(PN->getIncomingValue(i),
NumBits, isLeftShift, IC));
return PN;
}
}
}
/// This function implements the transforms common to both integer division
/// instructions (udiv and sdiv). It is called by the visitors to those integer
/// division instructions.
/// @brief Common integer divide transforms
Instruction *InstCombiner::commonIDivTransforms(BinaryOperator &I) {
Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
// The RHS is known non-zero.
if (Value *V = simplifyValueKnownNonZero(I.getOperand(1), *this)) {
I.setOperand(1, V);
return &I;
}
// Handle cases involving: [su]div X, (select Cond, Y, Z)
// This does not apply for fdiv.
if (isa<SelectInst>(Op1) && SimplifyDivRemOfSelect(I))
return &I;
if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
if (Instruction *LHS = dyn_cast<Instruction>(Op0)) {
// (X / C1) / C2 -> X / (C1*C2)
if (Instruction::BinaryOps(LHS->getOpcode()) == I.getOpcode())
if (ConstantInt *LHSRHS = dyn_cast<ConstantInt>(LHS->getOperand(1))) {
if (MultiplyOverflows(RHS, LHSRHS,
I.getOpcode() == Instruction::SDiv))
return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
return BinaryOperator::Create(I.getOpcode(), LHS->getOperand(0),
ConstantExpr::getMul(RHS, LHSRHS));
}
Value *X;
const APInt *C1, *C2;
if (match(RHS, m_APInt(C2))) {
bool IsSigned = I.getOpcode() == Instruction::SDiv;
if ((IsSigned && match(LHS, m_NSWMul(m_Value(X), m_APInt(C1)))) ||
(!IsSigned && match(LHS, m_NUWMul(m_Value(X), m_APInt(C1))))) {
APInt Quotient(C1->getBitWidth(), /*Val=*/0ULL, IsSigned);
// (X * C1) / C2 -> X / (C2 / C1) if C2 is a multiple of C1.
if (IsMultiple(*C2, *C1, Quotient, IsSigned)) {
BinaryOperator *BO = BinaryOperator::Create(
I.getOpcode(), X, ConstantInt::get(X->getType(), Quotient));
BO->setIsExact(I.isExact());
return BO;
}
// (X * C1) / C2 -> X * (C1 / C2) if C1 is a multiple of C2.
if (IsMultiple(*C1, *C2, Quotient, IsSigned)) {
BinaryOperator *BO = BinaryOperator::Create(
Instruction::Mul, X, ConstantInt::get(X->getType(), Quotient));
BO->setHasNoUnsignedWrap(
!IsSigned &&
cast<OverflowingBinaryOperator>(LHS)->hasNoUnsignedWrap());
BO->setHasNoSignedWrap(
cast<OverflowingBinaryOperator>(LHS)->hasNoSignedWrap());
return BO;
}
}
if ((IsSigned && match(LHS, m_NSWShl(m_Value(X), m_APInt(C1)))) ||
(!IsSigned && match(LHS, m_NUWShl(m_Value(X), m_APInt(C1))))) {
APInt Quotient(C1->getBitWidth(), /*Val=*/0ULL, IsSigned);
APInt C1Shifted = APInt::getOneBitSet(
C1->getBitWidth(), static_cast<unsigned>(C1->getLimitedValue()));
// (X << C1) / C2 -> X / (C2 >> C1) if C2 is a multiple of C1.
if (IsMultiple(*C2, C1Shifted, Quotient, IsSigned)) {
BinaryOperator *BO = BinaryOperator::Create(
I.getOpcode(), X, ConstantInt::get(X->getType(), Quotient));
BO->setIsExact(I.isExact());
return BO;
}
// (X << C1) / C2 -> X * (C2 >> C1) if C1 is a multiple of C2.
if (IsMultiple(C1Shifted, *C2, Quotient, IsSigned)) {
BinaryOperator *BO = BinaryOperator::Create(
Instruction::Mul, X, ConstantInt::get(X->getType(), Quotient));
BO->setHasNoUnsignedWrap(
!IsSigned &&
cast<OverflowingBinaryOperator>(LHS)->hasNoUnsignedWrap());
BO->setHasNoSignedWrap(
cast<OverflowingBinaryOperator>(LHS)->hasNoSignedWrap());
return BO;
}
}
}
}
if (!RHS->isZero()) { // avoid X udiv 0
if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
if (Instruction *R = FoldOpIntoSelect(I, SI))
return R;
if (isa<PHINode>(Op0))
if (Instruction *NV = FoldOpIntoPhi(I))
return NV;
}
}
if (ConstantInt *One = dyn_cast<ConstantInt>(Op0)) {
if (One->isOne() && !I.getType()->isIntegerTy(1)) {
bool isSigned = I.getOpcode() == Instruction::SDiv;
if (isSigned) {
// If Op1 is 0 then it's undefined behaviour, if Op1 is 1 then the
// result is one, if Op1 is -1 then the result is minus one, otherwise
// it's zero.
Value *Inc = Builder->CreateAdd(Op1, One);
Value *Cmp = Builder->CreateICmpULT(
Inc, ConstantInt::get(I.getType(), 3));
return SelectInst::Create(Cmp, Op1, ConstantInt::get(I.getType(), 0));
} else {
// If Op1 is 0 then it's undefined behaviour. If Op1 is 1 then the
// result is one, otherwise it's zero.
return new ZExtInst(Builder->CreateICmpEQ(Op1, One), I.getType());
}
}
}
// See if we can fold away this div instruction.
if (SimplifyDemandedInstructionBits(I))
return &I;
// (X - (X rem Y)) / Y -> X / Y; usually originates as ((X / Y) * Y) / Y
Value *X = nullptr, *Z = nullptr;
if (match(Op0, m_Sub(m_Value(X), m_Value(Z)))) { // (X - Z) / Y; Y = Op1
bool isSigned = I.getOpcode() == Instruction::SDiv;
if ((isSigned && match(Z, m_SRem(m_Specific(X), m_Specific(Op1)))) ||
(!isSigned && match(Z, m_URem(m_Specific(X), m_Specific(Op1)))))
return BinaryOperator::Create(I.getOpcode(), X, Op1);
}
return nullptr;
}
