//------------------------------------------------------------------------
// LowerStoreLoc: Lower a store of a lclVar
//
// Arguments:
// storeLoc - the local store (GT_STORE_LCL_FLD or GT_STORE_LCL_VAR)
//
// Notes:
// This involves:
// - Widening operations of unsigneds.
//
void Lowering::LowerStoreLoc(GenTreeLclVarCommon* storeLoc)
{
// Try to widen the ops if they are going into a local var.
GenTree* op1 = storeLoc->gtGetOp1();
if ((storeLoc->gtOper == GT_STORE_LCL_VAR) && (op1->gtOper == GT_CNS_INT))
{
GenTreeIntCon* con = op1->AsIntCon();
ssize_t ival = con->gtIconVal;
unsigned varNum = storeLoc->gtLclNum;
LclVarDsc* varDsc = comp->lvaTable + varNum;
if (varDsc->lvIsSIMDType())
{
noway_assert(storeLoc->gtType != TYP_STRUCT);
}
unsigned size = genTypeSize(storeLoc);
// If we are storing a constant into a local variable
// we extend the size of the store here
if ((size < 4) && !varTypeIsStruct(varDsc))
{
if (!varTypeIsUnsigned(varDsc))
{
if (genTypeSize(storeLoc) == 1)
{
if ((ival & 0x7f) != ival)
{
ival = ival | 0xffffff00;
}
}
else
{
assert(genTypeSize(storeLoc) == 2);
if ((ival & 0x7fff) != ival)
{
ival = ival | 0xffff0000;
}
}
}
// A local stack slot is at least 4 bytes in size, regardless of
// what the local var is typed as, so auto-promote it here
// unless it is a field of a promoted struct
// TODO-CQ: if the field is promoted shouldn't we also be able to do this?
if (!varDsc->lvIsStructField)
{
storeLoc->gtType = TYP_INT;
con->SetIconValue(ival);
}
}
}
if (storeLoc->OperIs(GT_STORE_LCL_FLD))
{
// We should only encounter this for lclVars that are lvDoNotEnregister.
verifyLclFldDoNotEnregister(storeLoc->gtLclNum);
}
ContainCheckStoreLoc(storeLoc);
}
//----------------------------------------------------------------------------------------------
// Lowering::LowerHWIntrinsic: Perform containment analysis for a hardware intrinsic node.
//
// Arguments:
// node - The hardware intrinsic node.
//
void Lowering::LowerHWIntrinsic(GenTreeHWIntrinsic* node)
{
auto intrinsicID = node->gtHWIntrinsicId;
auto intrinsicInfo = HWIntrinsicInfo::lookup(node->gtHWIntrinsicId);
//
// Lower unsupported Unsigned Compare Zero intrinsics to their trivial transformations
//
// ARM64 does not support most forms of compare zero for Unsigned values
// This is because some are non-sensical, and the rest are trivial transformations of other operators
//
if ((intrinsicInfo.flags & HWIntrinsicInfo::LowerCmpUZero) && varTypeIsUnsigned(node->gtSIMDBaseType))
{
auto setAllVector = node->gtSIMDSize > 8 ? NI_ARM64_SIMD_SetAllVector128 : NI_ARM64_SIMD_SetAllVector64;
auto origOp1 = node->gtOp.gtOp1;
switch (intrinsicID)
{
case NI_ARM64_SIMD_GT_ZERO:
// Unsigned > 0 ==> !(Unsigned == 0)
node->gtOp.gtOp1 =
comp->gtNewSimdHWIntrinsicNode(node->TypeGet(), node->gtOp.gtOp1, NI_ARM64_SIMD_EQ_ZERO,
node->gtSIMDBaseType, node->gtSIMDSize);
node->gtHWIntrinsicId = NI_ARM64_SIMD_BitwiseNot;
BlockRange().InsertBefore(node, node->gtOp.gtOp1);
break;
case NI_ARM64_SIMD_LE_ZERO:
// Unsigned <= 0 ==> Unsigned == 0
node->gtHWIntrinsicId = NI_ARM64_SIMD_EQ_ZERO;
break;
case NI_ARM64_SIMD_GE_ZERO:
case NI_ARM64_SIMD_LT_ZERO:
// Unsigned >= 0 ==> Always true
// Unsigned < 0 ==> Always false
node->gtHWIntrinsicId = setAllVector;
node->gtOp.gtOp1 = comp->gtNewLconNode((intrinsicID == NI_ARM64_SIMD_GE_ZERO) ? ~0ULL : 0ULL);
BlockRange().InsertBefore(node, node->gtOp.gtOp1);
if ((origOp1->gtFlags & GTF_ALL_EFFECT) == 0)
{
BlockRange().Remove(origOp1, true);
}
else
{
origOp1->SetUnusedValue();
}
break;
default:
assert(!"Unhandled LowerCmpUZero case");
}
}
ContainCheckHWIntrinsic(node);
}
//------------------------------------------------------------------------
// DecomposeCast: Decompose GT_CAST.
//
// Arguments:
// use - the LIR::Use object for the def that needs to be decomposed.
//
// Return Value:
// The next node to process.
//
GenTree* DecomposeLongs::DecomposeCast(LIR::Use& use)
{
assert(use.IsInitialized());
assert(use.Def()->OperGet() == GT_CAST);
GenTree* cast = use.Def()->AsCast();
GenTree* loResult = nullptr;
GenTree* hiResult = nullptr;
var_types srcType = cast->CastFromType();
var_types dstType = cast->CastToType();
if ((cast->gtFlags & GTF_UNSIGNED) != 0)
{
srcType = genUnsignedType(srcType);
}
if (varTypeIsLong(srcType))
{
if (cast->gtOverflow() && (varTypeIsUnsigned(srcType) != varTypeIsUnsigned(dstType)))
{
GenTree* srcOp = cast->gtGetOp1();
noway_assert(srcOp->OperGet() == GT_LONG);
GenTree* loSrcOp = srcOp->gtGetOp1();
GenTree* hiSrcOp = srcOp->gtGetOp2();
//
// When casting between long types an overflow check is needed only if the types
// have different signedness. In both cases (long->ulong and ulong->long) we only
// need to check if the high part is negative or not. Use the existing cast node
// to perform a int->uint cast of the high part to take advantage of the overflow
// check provided by codegen.
//
loResult = loSrcOp;
hiResult = cast;
hiResult->gtType = TYP_INT;
hiResult->AsCast()->gtCastType = TYP_UINT;
hiResult->gtFlags &= ~GTF_UNSIGNED;
hiResult->gtOp.gtOp1 = hiSrcOp;
Range().Remove(cast);
Range().Remove(srcOp);
Range().InsertAfter(hiSrcOp, hiResult);
}
else
{
NYI("Unimplemented long->long no-op cast decomposition");
}
}
else if (varTypeIsIntegralOrI(srcType))
{
if (cast->gtOverflow() && !varTypeIsUnsigned(srcType) && varTypeIsUnsigned(dstType))
{
//
// An overflow check is needed only when casting from a signed type to ulong.
// Change the cast type to uint to take advantage of the overflow check provided
// by codegen and then zero extend the resulting uint to ulong.
//
loResult = cast;
loResult->AsCast()->gtCastType = TYP_UINT;
loResult->gtType = TYP_INT;
hiResult = m_compiler->gtNewZeroConNode(TYP_INT);
Range().InsertAfter(loResult, hiResult);
}
else
{
if (varTypeIsUnsigned(srcType))
{
loResult = cast->gtGetOp1();
hiResult = m_compiler->gtNewZeroConNode(TYP_INT);
Range().Remove(cast);
Range().InsertAfter(loResult, hiResult);
}
else
{
LIR::Use src(Range(), &(cast->gtOp.gtOp1), cast);
unsigned lclNum = src.ReplaceWithLclVar(m_compiler, m_blockWeight);
loResult = src.Def();
GenTree* loCopy = m_compiler->gtNewLclvNode(lclNum, TYP_INT);
GenTree* shiftBy = m_compiler->gtNewIconNode(31, TYP_INT);
hiResult = m_compiler->gtNewOperNode(GT_RSH, TYP_INT, loCopy, shiftBy);
Range().Remove(cast);
Range().InsertAfter(loResult, loCopy, shiftBy, hiResult);
m_compiler->lvaIncRefCnts(loCopy);
}
}
}
else
{
NYI("Unimplemented cast decomposition");
}
return FinalizeDecomposition(use, loResult, hiResult);
}
