//------------------------------------------------------------------------
// 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* tree = use.Def();
GenTree* loResult = nullptr;
GenTree* hiResult = nullptr;
assert(tree->gtPrev == tree->gtGetOp1());
NYI_IF(tree->gtOverflow(), "TYP_LONG cast with overflow");
switch (tree->AsCast()->CastFromType())
{
case TYP_INT:
if (tree->gtFlags & GTF_UNSIGNED)
{
loResult = tree->gtGetOp1();
Range().Remove(tree);
hiResult = new (m_compiler, GT_CNS_INT) GenTreeIntCon(TYP_INT, 0);
Range().InsertAfter(loResult, hiResult);
}
else
{
NYI("Lowering of signed cast TYP_INT->TYP_LONG");
}
break;
default:
NYI("Unimplemented type for Lowering of cast to TYP_LONG");
break;
}
return FinalizeDecomposition(use, loResult, hiResult);
}
//------------------------------------------------------------------------
// DecomposeArith: Decompose GT_ADD, GT_SUB, GT_OR, GT_XOR, GT_AND.
//
// Arguments:
// use - the LIR::Use object for the def that needs to be decomposed.
//
// Return Value:
// The next node to process.
//
GenTree* DecomposeLongs::DecomposeArith(LIR::Use& use)
{
assert(use.IsInitialized());
GenTree* tree = use.Def();
genTreeOps oper = tree->OperGet();
assert((oper == GT_ADD) || (oper == GT_SUB) || (oper == GT_OR) || (oper == GT_XOR) || (oper == GT_AND));
GenTree* op1 = tree->gtGetOp1();
GenTree* op2 = tree->gtGetOp2();
// Both operands must have already been decomposed into GT_LONG operators.
noway_assert((op1->OperGet() == GT_LONG) && (op2->OperGet() == GT_LONG));
// Capture the lo and hi halves of op1 and op2.
GenTree* loOp1 = op1->gtGetOp1();
GenTree* hiOp1 = op1->gtGetOp2();
GenTree* loOp2 = op2->gtGetOp1();
GenTree* hiOp2 = op2->gtGetOp2();
// Now, remove op1 and op2 from the node list.
Range().Remove(op1);
Range().Remove(op2);
// We will reuse "tree" for the loResult, which will now be of TYP_INT, and its operands
// will be the lo halves of op1 from above.
GenTree* loResult = tree;
loResult->SetOper(GetLoOper(oper));
loResult->gtType = TYP_INT;
loResult->gtOp.gtOp1 = loOp1;
loResult->gtOp.gtOp2 = loOp2;
GenTree* hiResult = new (m_compiler, oper) GenTreeOp(GetHiOper(oper), TYP_INT, hiOp1, hiOp2);
Range().InsertAfter(loResult, hiResult);
if ((oper == GT_ADD) || (oper == GT_SUB))
{
if (loResult->gtOverflow())
{
hiResult->gtFlags |= GTF_OVERFLOW;
loResult->gtFlags &= ~GTF_OVERFLOW;
}
if (loResult->gtFlags & GTF_UNSIGNED)
{
hiResult->gtFlags |= GTF_UNSIGNED;
}
}
return FinalizeDecomposition(use, loResult, hiResult);
}
//------------------------------------------------------------------------
// DecomposeCast: Decompose GT_CAST.
//
// Arguments:
// ppTree - the tree to decompose
// data - tree walk context
//
// Return Value:
// None.
//
void DecomposeLongs::DecomposeCast(GenTree** ppTree, Compiler::fgWalkData* data)
{
assert(ppTree != nullptr);
assert(*ppTree != nullptr);
assert(data != nullptr);
assert((*ppTree)->OperGet() == GT_CAST);
assert(m_compiler->compCurStmt != nullptr);
GenTree* tree = *ppTree;
GenTree* loResult = nullptr;
GenTree* hiResult = nullptr;
GenTreeStmt* curStmt = m_compiler->compCurStmt->AsStmt();
assert(tree->gtPrev == tree->gtGetOp1());
NYI_IF(tree->gtOverflow(), "TYP_LONG cast with overflow");
switch (tree->AsCast()->CastFromType())
{
case TYP_INT:
if (tree->gtFlags & GTF_UNSIGNED)
{
loResult = tree->gtGetOp1();
hiResult = new (m_compiler, GT_CNS_INT) GenTreeIntCon(TYP_INT, 0);
m_compiler->fgSnipNode(curStmt, tree);
}
else
{
NYI("Lowering of signed cast TYP_INT->TYP_LONG");
}
break;
default:
NYI("Unimplemented type for Lowering of cast to TYP_LONG");
break;
}
FinalizeDecomposition(ppTree, data, loResult, hiResult);
}
//------------------------------------------------------------------------
// DecomposeArith: Decompose GT_ADD, GT_SUB, GT_OR, GT_XOR, GT_AND.
//
// Arguments:
// ppTree - the tree to decompose
// data - tree walk context
//
// Return Value:
// None.
//
void DecomposeLongs::DecomposeArith(GenTree** ppTree, Compiler::fgWalkData* data)
{
assert(ppTree != nullptr);
assert(*ppTree != nullptr);
assert(data != nullptr);
assert(m_compiler->compCurStmt != nullptr);
GenTreeStmt* curStmt = m_compiler->compCurStmt->AsStmt();
GenTree* tree = *ppTree;
genTreeOps oper = tree->OperGet();
assert((oper == GT_ADD) ||
(oper == GT_SUB) ||
(oper == GT_OR) ||
(oper == GT_XOR) ||
(oper == GT_AND));
NYI_IF((tree->gtFlags & GTF_REVERSE_OPS) != 0, "Binary operator with GTF_REVERSE_OPS");
GenTree* op1 = tree->gtGetOp1();
GenTree* op2 = tree->gtGetOp2();
// Both operands must have already been decomposed into GT_LONG operators.
noway_assert((op1->OperGet() == GT_LONG) && (op2->OperGet() == GT_LONG));
// Capture the lo and hi halves of op1 and op2.
GenTree* loOp1 = op1->gtGetOp1();
GenTree* hiOp1 = op1->gtGetOp2();
GenTree* loOp2 = op2->gtGetOp1();
GenTree* hiOp2 = op2->gtGetOp2();
// We don't have support to decompose a TYP_LONG node that already has a child that has
// been decomposed into parts, where the high part depends on the value generated by the
// low part (via the flags register). For example, if we have:
// +(gt_long(+(lo3, lo4), +Hi(hi3, hi4)), gt_long(lo2, hi2))
// We would decompose it here to:
// gt_long(+(+(lo3, lo4), lo2), +Hi(+Hi(hi3, hi4), hi2))
// But this would generate incorrect code, because the "+Hi(hi3, hi4)" code generation
// needs to immediately follow the "+(lo3, lo4)" part. Also, if this node is one that
// requires a unique high operator, and the child nodes are not simple locals (e.g.,
// they are decomposed nodes), then we also can't decompose the node, as we aren't
// guaranteed the high and low parts will be executed immediately after each other.
NYI_IF(hiOp1->OperIsHigh() ||
hiOp2->OperIsHigh() ||
(GenTree::OperIsHigh(GetHiOper(oper)) &&
(!loOp1->OperIsLeaf() ||
!hiOp1->OperIsLeaf() ||
!loOp1->OperIsLeaf() ||
!hiOp2->OperIsLeaf())),
"Can't decompose expression tree TYP_LONG node");
// Now, remove op1 and op2 from the node list.
m_compiler->fgSnipNode(curStmt, op1);
m_compiler->fgSnipNode(curStmt, op2);
// We will reuse "tree" for the loResult, which will now be of TYP_INT, and its operands
// will be the lo halves of op1 from above.
GenTree* loResult = tree;
loResult->SetOper(GetLoOper(loResult->OperGet()));
loResult->gtType = TYP_INT;
loResult->gtOp.gtOp1 = loOp1;
loResult->gtOp.gtOp2 = loOp2;
// The various halves will be correctly threaded internally. We simply need to
// relink them into the proper order, i.e. loOp1 is followed by loOp2, and then
// the loResult node.
// (This rethreading, and that below, are where we need to address the reverse ops case).
// The current order is (after snipping op1 and op2):
// ... loOp1-> ... hiOp1->loOp2First ... loOp2->hiOp2First ... hiOp2
// The order we want is:
// ... loOp1->loOp2First ... loOp2->loResult
// ... hiOp1->hiOp2First ... hiOp2->hiResult
// i.e. we swap hiOp1 and loOp2, and create (for now) separate loResult and hiResult trees
GenTree* loOp2First = hiOp1->gtNext;
GenTree* hiOp2First = loOp2->gtNext;
// First, we will NYI if both hiOp1 and loOp2 have side effects.
NYI_IF(((loOp2->gtFlags & GTF_ALL_EFFECT) != 0) && ((hiOp1->gtFlags & GTF_ALL_EFFECT) != 0),
"Binary long operator with non-reorderable sub expressions");
// Now, we reorder the loOps and the loResult.
loOp1->gtNext = loOp2First;
loOp2First->gtPrev = loOp1;
loOp2->gtNext = loResult;
loResult->gtPrev = loOp2;
// Next, reorder the hiOps and the hiResult.
GenTree* hiResult = new (m_compiler, oper) GenTreeOp(GetHiOper(oper), TYP_INT, hiOp1, hiOp2);
hiOp1->gtNext = hiOp2First;
hiOp2First->gtPrev = hiOp1;
hiOp2->gtNext = hiResult;
hiResult->gtPrev = hiOp2;
if ((oper == GT_ADD) || (oper == GT_SUB))
{
if (loResult->gtOverflow())
{
hiResult->gtFlags |= GTF_OVERFLOW;
loResult->gtFlags &= ~GTF_OVERFLOW;
}
if (loResult->gtFlags & GTF_UNSIGNED)
{
hiResult->gtFlags |= GTF_UNSIGNED;
}
}
FinalizeDecomposition(ppTree, data, loResult, hiResult);
}
//------------------------------------------------------------------------
// 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);
}
//------------------------------------------------------------------------
// DecomposeArith: Decompose GT_ADD, GT_SUB, GT_OR, GT_XOR, GT_AND.
//
// Arguments:
// use - the LIR::Use object for the def that needs to be decomposed.
//
// Return Value:
// The next node to process.
//
GenTree* DecomposeLongs::DecomposeArith(LIR::Use& use)
{
assert(use.IsInitialized());
GenTree* tree = use.Def();
genTreeOps oper = tree->OperGet();
assert((oper == GT_ADD) || (oper == GT_SUB) || (oper == GT_OR) || (oper == GT_XOR) || (oper == GT_AND));
GenTree* op1 = tree->gtGetOp1();
GenTree* op2 = tree->gtGetOp2();
// Both operands must have already been decomposed into GT_LONG operators.
noway_assert((op1->OperGet() == GT_LONG) && (op2->OperGet() == GT_LONG));
// Capture the lo and hi halves of op1 and op2.
GenTree* loOp1 = op1->gtGetOp1();
GenTree* hiOp1 = op1->gtGetOp2();
GenTree* loOp2 = op2->gtGetOp1();
GenTree* hiOp2 = op2->gtGetOp2();
// We don't have support to decompose a TYP_LONG node that already has a child that has
// been decomposed into parts, where the high part depends on the value generated by the
// low part (via the flags register). For example, if we have:
// +(gt_long(+(lo3, lo4), +Hi(hi3, hi4)), gt_long(lo2, hi2))
// We would decompose it here to:
// gt_long(+(+(lo3, lo4), lo2), +Hi(+Hi(hi3, hi4), hi2))
// But this would generate incorrect code, because the "+Hi(hi3, hi4)" code generation
// needs to immediately follow the "+(lo3, lo4)" part. Also, if this node is one that
// requires a unique high operator, and the child nodes are not simple locals (e.g.,
// they are decomposed nodes), then we also can't decompose the node, as we aren't
// guaranteed the high and low parts will be executed immediately after each other.
NYI_IF(hiOp1->OperIsHigh() || hiOp2->OperIsHigh() ||
(GenTree::OperIsHigh(GetHiOper(oper)) &&
(!loOp1->OperIsLeaf() || !hiOp1->OperIsLeaf() || !loOp1->OperIsLeaf() || !hiOp2->OperIsLeaf())),
"Can't decompose expression tree TYP_LONG node");
// Now, remove op1 and op2 from the node list.
BlockRange().Remove(op1);
BlockRange().Remove(op2);
// We will reuse "tree" for the loResult, which will now be of TYP_INT, and its operands
// will be the lo halves of op1 from above.
GenTree* loResult = tree;
loResult->SetOper(GetLoOper(loResult->OperGet()));
loResult->gtType = TYP_INT;
loResult->gtOp.gtOp1 = loOp1;
loResult->gtOp.gtOp2 = loOp2;
GenTree* hiResult = new (m_compiler, oper) GenTreeOp(GetHiOper(oper), TYP_INT, hiOp1, hiOp2);
hiResult->CopyCosts(loResult);
BlockRange().InsertAfter(loResult, hiResult);
if ((oper == GT_ADD) || (oper == GT_SUB))
{
if (loResult->gtOverflow())
{
hiResult->gtFlags |= GTF_OVERFLOW;
loResult->gtFlags &= ~GTF_OVERFLOW;
}
if (loResult->gtFlags & GTF_UNSIGNED)
{
hiResult->gtFlags |= GTF_UNSIGNED;
}
}
return FinalizeDecomposition(use, loResult, hiResult);
}