void Rationalizer::RewriteAddress(LIR::Use& use)
{
assert(use.IsInitialized());
GenTreeUnOp* address = use.Def()->AsUnOp();
assert(address->OperGet() == GT_ADDR);
GenTree* location = address->gtGetOp1();
genTreeOps locationOp = location->OperGet();
if (location->IsLocal())
{
// We are changing the child from GT_LCL_VAR TO GT_LCL_VAR_ADDR.
// Therefore gtType of the child needs to be changed to a TYP_BYREF
#ifdef DEBUG
if (locationOp == GT_LCL_VAR)
{
JITDUMP("Rewriting GT_ADDR(GT_LCL_VAR) to GT_LCL_VAR_ADDR:\n");
}
else
{
assert(locationOp == GT_LCL_FLD);
JITDUMP("Rewriting GT_ADDR(GT_LCL_FLD) to GT_LCL_FLD_ADDR:\n");
}
#endif // DEBUG
location->SetOper(addrForm(locationOp));
location->gtType = TYP_BYREF;
copyFlags(location, address, GTF_ALL_EFFECT);
use.ReplaceWith(comp, location);
BlockRange().Remove(address);
}
else if (locationOp == GT_CLS_VAR)
{
location->SetOper(GT_CLS_VAR_ADDR);
location->gtType = TYP_BYREF;
copyFlags(location, address, GTF_ALL_EFFECT);
use.ReplaceWith(comp, location);
BlockRange().Remove(address);
JITDUMP("Rewriting GT_ADDR(GT_CLS_VAR) to GT_CLS_VAR_ADDR:\n");
}
else if (location->OperIsIndir())
{
use.ReplaceWith(comp, location->gtGetOp1());
BlockRange().Remove(location);
BlockRange().Remove(address);
JITDUMP("Rewriting GT_ADDR(GT_IND(X)) to X:\n");
}
DISPTREERANGE(BlockRange(), use.Def());
JITDUMP("\n");
}
void Rationalizer::RewriteAssignment(LIR::Use& use)
{
assert(use.IsInitialized());
GenTreeOp* assignment = use.Def()->AsOp();
assert(assignment->OperGet() == GT_ASG);
GenTree* location = assignment->gtGetOp1();
GenTree* value = assignment->gtGetOp2();
genTreeOps locationOp = location->OperGet();
switch (locationOp)
{
case GT_LCL_VAR:
case GT_LCL_FLD:
case GT_REG_VAR:
case GT_PHI_ARG:
RewriteAssignmentIntoStoreLclCore(assignment, location, value, locationOp);
BlockRange().Remove(location);
break;
case GT_IND:
{
GenTreeStoreInd* store =
new (comp, GT_STOREIND) GenTreeStoreInd(location->TypeGet(), location->gtGetOp1(), value);
copyFlags(store, assignment, GTF_ALL_EFFECT);
copyFlags(store, location, GTF_IND_FLAGS);
if (assignment->IsReverseOp())
{
store->gtFlags |= GTF_REVERSE_OPS;
}
// TODO: JIT dump
// Remove the GT_IND node and replace the assignment node with the store
BlockRange().Remove(location);
BlockRange().InsertBefore(assignment, store);
use.ReplaceWith(comp, store);
BlockRange().Remove(assignment);
}
break;
case GT_CLS_VAR:
{
location->SetOper(GT_CLS_VAR_ADDR);
location->gtType = TYP_BYREF;
assignment->SetOper(GT_STOREIND);
// TODO: JIT dump
}
break;
default:
unreached();
break;
}
}
// Rewrite a SIMD indirection as GT_IND(GT_LEA(obj.op1)), or as a simple
// lclVar if possible.
//
// Arguments:
// use - A use reference for a block node
// keepBlk - True if this should remain a block node if it is not a lclVar
//
// Return Value:
// None.
//
// TODO-1stClassStructs: These should be eliminated earlier, once we can handle
// lclVars in all the places that used to have GT_OBJ.
//
void Rationalizer::RewriteSIMDOperand(LIR::Use& use, bool keepBlk)
{
#ifdef FEATURE_SIMD
// No lowering is needed for non-SIMD nodes, so early out if featureSIMD is not enabled.
if (!comp->featureSIMD)
{
return;
}
GenTree* tree = use.Def();
if (!tree->OperIsIndir())
{
return;
}
var_types simdType = tree->TypeGet();
if (!varTypeIsSIMD(simdType))
{
return;
}
// If we have GT_IND(GT_LCL_VAR_ADDR) and the GT_LCL_VAR_ADDR is TYP_BYREF/TYP_I_IMPL,
// and the var is a SIMD type, replace the expression by GT_LCL_VAR.
GenTree* addr = tree->AsIndir()->Addr();
if (addr->OperIsLocalAddr() && comp->isAddrOfSIMDType(addr))
{
BlockRange().Remove(tree);
addr->SetOper(loadForm(addr->OperGet()));
addr->gtType = simdType;
use.ReplaceWith(comp, addr);
}
else if ((addr->OperGet() == GT_ADDR) && (addr->gtGetOp1()->OperGet() == GT_SIMD))
{
// if we have GT_IND(GT_ADDR(GT_SIMD)), remove the GT_IND(GT_ADDR()), leaving just the GT_SIMD.
BlockRange().Remove(tree);
BlockRange().Remove(addr);
use.ReplaceWith(comp, addr->gtGetOp1());
}
else if (!keepBlk)
{
tree->SetOper(GT_IND);
tree->gtType = simdType;
}
#endif // FEATURE_SIMD
}
//------------------------------------------------------------------------
// DecomposeLclVar: Decompose GT_LCL_VAR.
//
// Arguments:
// use - the LIR::Use object for the def that needs to be decomposed.
//
// Return Value:
// The next node to process.
//
GenTree* DecomposeLongs::DecomposeLclVar(LIR::Use& use)
{
assert(use.IsInitialized());
assert(use.Def()->OperGet() == GT_LCL_VAR);
GenTree* tree = use.Def();
unsigned varNum = tree->AsLclVarCommon()->gtLclNum;
LclVarDsc* varDsc = m_compiler->lvaTable + varNum;
m_compiler->lvaDecRefCnts(tree);
GenTree* loResult = tree;
loResult->gtType = TYP_INT;
GenTree* hiResult = m_compiler->gtNewLclLNode(varNum, TYP_INT);
hiResult->CopyCosts(loResult);
BlockRange().InsertAfter(loResult, hiResult);
if (varDsc->lvPromoted)
{
assert(varDsc->lvFieldCnt == 2);
unsigned loVarNum = varDsc->lvFieldLclStart;
unsigned hiVarNum = loVarNum + 1;
loResult->AsLclVarCommon()->SetLclNum(loVarNum);
hiResult->AsLclVarCommon()->SetLclNum(hiVarNum);
}
else
{
noway_assert(varDsc->lvLRACandidate == false);
loResult->SetOper(GT_LCL_FLD);
loResult->AsLclFld()->gtLclOffs = 0;
loResult->AsLclFld()->gtFieldSeq = FieldSeqStore::NotAField();
hiResult->SetOper(GT_LCL_FLD);
hiResult->AsLclFld()->gtLclOffs = 4;
hiResult->AsLclFld()->gtFieldSeq = FieldSeqStore::NotAField();
}
m_compiler->lvaIncRefCnts(loResult);
m_compiler->lvaIncRefCnts(hiResult);
return FinalizeDecomposition(use, loResult, hiResult);
}
//------------------------------------------------------------------------
// DecomposeLclVar: Decompose GT_LCL_VAR.
//
// Arguments:
// ppTree - the tree to decompose
// data - tree walk context
//
// Return Value:
// None.
//
void DecomposeLongs::DecomposeLclVar(GenTree** ppTree, Compiler::fgWalkData* data)
{
assert(ppTree != nullptr);
assert(*ppTree != nullptr);
assert(data != nullptr);
assert((*ppTree)->OperGet() == GT_LCL_VAR);
GenTree* tree = *ppTree;
unsigned varNum = tree->AsLclVarCommon()->gtLclNum;
LclVarDsc* varDsc = m_compiler->lvaTable + varNum;
m_compiler->lvaDecRefCnts(tree);
GenTree* loResult = tree;
loResult->gtType = TYP_INT;
GenTree* hiResult = m_compiler->gtNewLclLNode(varNum, TYP_INT);
if (varDsc->lvPromoted)
{
assert(varDsc->lvFieldCnt == 2);
unsigned loVarNum = varDsc->lvFieldLclStart;
unsigned hiVarNum = loVarNum + 1;
loResult->AsLclVarCommon()->SetLclNum(loVarNum);
hiResult->AsLclVarCommon()->SetLclNum(hiVarNum);
}
else
{
noway_assert(varDsc->lvLRACandidate == false);
loResult->SetOper(GT_LCL_FLD);
loResult->AsLclFld()->gtLclOffs = 0;
loResult->AsLclFld()->gtFieldSeq = FieldSeqStore::NotAField();
hiResult->SetOper(GT_LCL_FLD);
hiResult->AsLclFld()->gtLclOffs = 4;
hiResult->AsLclFld()->gtFieldSeq = FieldSeqStore::NotAField();
}
m_compiler->lvaIncRefCnts(loResult);
m_compiler->lvaIncRefCnts(hiResult);
FinalizeDecomposition(ppTree, data, loResult, hiResult);
}
// Rewrite a SIMD indirection as GT_IND(GT_LEA(obj.op1)), or as a simple
// lclVar if possible.
//
// Arguments:
// use - A use reference for a block node
// keepBlk - True if this should remain a block node if it is not a lclVar
//
// Return Value:
// None.
//
// TODO-1stClassStructs: These should be eliminated earlier, once we can handle
// lclVars in all the places that used to have GT_OBJ.
//
void Rationalizer::RewriteSIMDOperand(LIR::Use& use, bool keepBlk)
{
#ifdef FEATURE_SIMD
// No lowering is needed for non-SIMD nodes, so early out if featureSIMD is not enabled.
if (!comp->featureSIMD)
{
return;
}
GenTree* tree = use.Def();
if (!tree->OperIsIndir())
{
return;
}
var_types simdType = tree->TypeGet();
if (!varTypeIsSIMD(simdType))
{
return;
}
// If the operand of is a GT_ADDR(GT_LCL_VAR) and LclVar is known to be of simdType,
// replace obj by GT_LCL_VAR.
GenTree* addr = tree->AsIndir()->Addr();
if (addr->OperIsLocalAddr() && comp->isAddrOfSIMDType(addr))
{
BlockRange().Remove(tree);
addr->SetOper(loadForm(addr->OperGet()));
addr->gtType = simdType;
use.ReplaceWith(comp, addr);
}
else if (!keepBlk)
{
tree->SetOper(GT_IND);
tree->gtType = simdType;
}
#endif // FEATURE_SIMD
}
//------------------------------------------------------------------------
// 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);
}
// Rewrite GT_OBJ of SIMD Vector as GT_IND(GT_LEA(obj.op1)) of a SIMD type.
//
// Arguments:
// ppTree - A pointer-to-a-pointer for the GT_OBJ
// fgWalkData - A pointer to tree walk data providing the context
//
// Return Value:
// None.
//
// TODO-Cleanup: Once SIMD types are plumbed through the frontend, this will no longer
// be required.
//
void Rationalizer::RewriteObj(LIR::Use& use)
{
#ifdef FEATURE_SIMD
GenTreeObj* obj = use.Def()->AsObj();
// For UNIX struct passing, we can have Obj nodes for arguments.
// For other cases, we should never see a non-SIMD type here.
#ifdef FEATURE_UNIX_AMD64_STRUCT_PASSING
if (!varTypeIsSIMD(obj))
{
return;
}
#endif // FEATURE_UNIX_AMD64_STRUCT_PASSING
// Should come here only if featureSIMD is enabled
noway_assert(comp->featureSIMD);
// We should only call this with a SIMD type.
noway_assert(varTypeIsSIMD(obj));
var_types simdType = obj->TypeGet();
// If the operand of obj is a GT_ADDR(GT_LCL_VAR) and LclVar is known to be a SIMD type,
// replace obj by GT_LCL_VAR.
GenTree* srcAddr = obj->gtGetOp1();
if (srcAddr->OperIsLocalAddr() && comp->isAddrOfSIMDType(srcAddr))
{
BlockRange().Remove(obj);
srcAddr->SetOper(loadForm(srcAddr->OperGet()));
srcAddr->gtType = simdType;
use.ReplaceWith(comp, srcAddr);
}
else
{
obj->SetOper(GT_IND);
obj->gtType = simdType;
}
#else
// we should never reach without feature SIMD
assert(!"Unexpected obj during rationalization\n");
unreached();
#endif
}
//------------------------------------------------------------------------
// 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);
}
Compiler::fgWalkResult Rationalizer::RewriteNode(GenTree** useEdge, ArrayStack<GenTree*>& parentStack)
{
assert(useEdge != nullptr);
GenTree* node = *useEdge;
assert(node != nullptr);
#ifdef DEBUG
const bool isLateArg = (node->gtFlags & GTF_LATE_ARG) != 0;
#endif
// First, remove any preceeding list nodes, which are not otherwise visited by the tree walk.
//
// NOTE: GT_FIELD_LIST head nodes, and GT_LIST nodes used by phi nodes will in fact be visited.
for (GenTree* prev = node->gtPrev; prev != nullptr && prev->OperIsAnyList() && !(prev->OperIsFieldListHead());
prev = node->gtPrev)
{
BlockRange().Remove(prev);
}
// In addition, remove the current node if it is a GT_LIST node that is not an aggregate.
if (node->OperIsAnyList())
{
GenTreeArgList* list = node->AsArgList();
if (!list->OperIsFieldListHead())
{
BlockRange().Remove(list);
}
return Compiler::WALK_CONTINUE;
}
LIR::Use use;
if (parentStack.Height() < 2)
{
use = LIR::Use::GetDummyUse(BlockRange(), *useEdge);
}
else
{
use = LIR::Use(BlockRange(), useEdge, parentStack.Index(1));
}
assert(node == use.Def());
switch (node->OperGet())
{
case GT_ASG:
RewriteAssignment(use);
break;
case GT_BOX:
// GT_BOX at this level just passes through so get rid of it
use.ReplaceWith(comp, node->gtGetOp1());
BlockRange().Remove(node);
break;
case GT_ADDR:
RewriteAddress(use);
break;
case GT_IND:
// Clear the `GTF_IND_ASG_LHS` flag, which overlaps with `GTF_IND_REQ_ADDR_IN_REG`.
node->gtFlags &= ~GTF_IND_ASG_LHS;
if (varTypeIsSIMD(node))
{
RewriteSIMDOperand(use, false);
}
else
{
// Due to promotion of structs containing fields of type struct with a
// single scalar type field, we could potentially see IR nodes of the
// form GT_IND(GT_ADD(lclvarAddr, 0)) where 0 is an offset representing
// a field-seq. These get folded here.
//
// TODO: This code can be removed once JIT implements recursive struct
// promotion instead of lying about the type of struct field as the type
// of its single scalar field.
GenTree* addr = node->AsIndir()->Addr();
if (addr->OperGet() == GT_ADD && addr->gtGetOp1()->OperGet() == GT_LCL_VAR_ADDR &&
addr->gtGetOp2()->IsIntegralConst(0))
{
GenTreeLclVarCommon* lclVarNode = addr->gtGetOp1()->AsLclVarCommon();
unsigned lclNum = lclVarNode->GetLclNum();
LclVarDsc* varDsc = comp->lvaTable + lclNum;
if (node->TypeGet() == varDsc->TypeGet())
{
JITDUMP("Rewriting GT_IND(GT_ADD(LCL_VAR_ADDR,0)) to LCL_VAR\n");
lclVarNode->SetOper(GT_LCL_VAR);
lclVarNode->gtType = node->TypeGet();
use.ReplaceWith(comp, lclVarNode);
BlockRange().Remove(addr);
BlockRange().Remove(addr->gtGetOp2());
BlockRange().Remove(node);
}
}
}
break;
case GT_NOP:
// fgMorph sometimes inserts NOP nodes between defs and uses
// supposedly 'to prevent constant folding'. In this case, remove the
// NOP.
if (node->gtGetOp1() != nullptr)
{
use.ReplaceWith(comp, node->gtGetOp1());
BlockRange().Remove(node);
}
break;
case GT_COMMA:
{
GenTree* op1 = node->gtGetOp1();
if ((op1->gtFlags & GTF_ALL_EFFECT) == 0)
{
// The LHS has no side effects. Remove it.
bool isClosed = false;
unsigned sideEffects = 0;
LIR::ReadOnlyRange lhsRange = BlockRange().GetTreeRange(op1, &isClosed, &sideEffects);
// None of the transforms performed herein violate tree order, so these
// should always be true.
assert(isClosed);
assert((sideEffects & GTF_ALL_EFFECT) == 0);
BlockRange().Delete(comp, m_block, std::move(lhsRange));
}
GenTree* replacement = node->gtGetOp2();
if (!use.IsDummyUse())
{
use.ReplaceWith(comp, replacement);
}
else
{
// This is a top-level comma. If the RHS has no side effects we can remove
// it as well.
if ((replacement->gtFlags & GTF_ALL_EFFECT) == 0)
{
bool isClosed = false;
unsigned sideEffects = 0;
LIR::ReadOnlyRange rhsRange = BlockRange().GetTreeRange(replacement, &isClosed, &sideEffects);
// None of the transforms performed herein violate tree order, so these
// should always be true.
assert(isClosed);
assert((sideEffects & GTF_ALL_EFFECT) == 0);
BlockRange().Delete(comp, m_block, std::move(rhsRange));
}
}
BlockRange().Remove(node);
}
break;
case GT_ARGPLACE:
// Remove argplace and list nodes from the execution order.
//
// TODO: remove phi args and phi nodes as well?
BlockRange().Remove(node);
break;
#if defined(_TARGET_XARCH_) || defined(_TARGET_ARM_)
case GT_CLS_VAR:
{
// Class vars that are the target of an assignment will get rewritten into
// GT_STOREIND(GT_CLS_VAR_ADDR, val) by RewriteAssignment. This check is
// not strictly necessary--the GT_IND(GT_CLS_VAR_ADDR) pattern that would
// otherwise be generated would also be picked up by RewriteAssignment--but
// skipping the rewrite here saves an allocation and a bit of extra work.
const bool isLHSOfAssignment = (use.User()->OperGet() == GT_ASG) && (use.User()->gtGetOp1() == node);
if (!isLHSOfAssignment)
{
GenTree* ind = comp->gtNewOperNode(GT_IND, node->TypeGet(), node);
node->SetOper(GT_CLS_VAR_ADDR);
node->gtType = TYP_BYREF;
BlockRange().InsertAfter(node, ind);
use.ReplaceWith(comp, ind);
// TODO: JIT dump
}
}
break;
#endif // _TARGET_XARCH_
case GT_INTRINSIC:
// Non-target intrinsics should have already been rewritten back into user calls.
assert(Compiler::IsTargetIntrinsic(node->gtIntrinsic.gtIntrinsicId));
break;
#ifdef FEATURE_SIMD
case GT_BLK:
case GT_OBJ:
{
// TODO-1stClassStructs: These should have been transformed to GT_INDs, but in order
// to preserve existing behavior, we will keep this as a block node if this is the
// lhs of a block assignment, and either:
// - It is a "generic" TYP_STRUCT assignment, OR
// - It is an initblk, OR
// - Neither the lhs or rhs are known to be of SIMD type.
GenTree* parent = use.User();
bool keepBlk = false;
if ((parent->OperGet() == GT_ASG) && (node == parent->gtGetOp1()))
{
if ((node->TypeGet() == TYP_STRUCT) || parent->OperIsInitBlkOp())
{
keepBlk = true;
}
else if (!comp->isAddrOfSIMDType(node->AsBlk()->Addr()))
{
GenTree* dataSrc = parent->gtGetOp2();
if (!dataSrc->IsLocal() && (dataSrc->OperGet() != GT_SIMD))
{
noway_assert(dataSrc->OperIsIndir());
keepBlk = !comp->isAddrOfSIMDType(dataSrc->AsIndir()->Addr());
}
}
}
RewriteSIMDOperand(use, keepBlk);
}
break;
case GT_LCL_FLD:
case GT_STORE_LCL_FLD:
// TODO-1stClassStructs: Eliminate this.
FixupIfSIMDLocal(node->AsLclVarCommon());
break;
case GT_SIMD:
{
noway_assert(comp->featureSIMD);
GenTreeSIMD* simdNode = node->AsSIMD();
unsigned simdSize = simdNode->gtSIMDSize;
var_types simdType = comp->getSIMDTypeForSize(simdSize);
// TODO-1stClassStructs: This should be handled more generally for enregistered or promoted
// structs that are passed or returned in a different register type than their enregistered
// type(s).
if (simdNode->gtType == TYP_I_IMPL && simdNode->gtSIMDSize == TARGET_POINTER_SIZE)
{
// This happens when it is consumed by a GT_RET_EXPR.
// It can only be a Vector2f or Vector2i.
assert(genTypeSize(simdNode->gtSIMDBaseType) == 4);
simdNode->gtType = TYP_SIMD8;
}
// Certain SIMD trees require rationalizing.
if (simdNode->gtSIMD.gtSIMDIntrinsicID == SIMDIntrinsicInitArray)
{
// Rewrite this as an explicit load.
JITDUMP("Rewriting GT_SIMD array init as an explicit load:\n");
unsigned int baseTypeSize = genTypeSize(simdNode->gtSIMDBaseType);
GenTree* address = new (comp, GT_LEA) GenTreeAddrMode(TYP_BYREF, simdNode->gtOp1, simdNode->gtOp2,
baseTypeSize, offsetof(CORINFO_Array, u1Elems));
GenTree* ind = comp->gtNewOperNode(GT_IND, simdType, address);
BlockRange().InsertBefore(simdNode, address, ind);
use.ReplaceWith(comp, ind);
BlockRange().Remove(simdNode);
DISPTREERANGE(BlockRange(), use.Def());
JITDUMP("\n");
}
else
{
// This code depends on the fact that NONE of the SIMD intrinsics take vector operands
// of a different width. If that assumption changes, we will EITHER have to make these type
// transformations during importation, and plumb the types all the way through the JIT,
// OR add a lot of special handling here.
GenTree* op1 = simdNode->gtGetOp1();
if (op1 != nullptr && op1->gtType == TYP_STRUCT)
{
op1->gtType = simdType;
}
GenTree* op2 = simdNode->gtGetOp2IfPresent();
if (op2 != nullptr && op2->gtType == TYP_STRUCT)
{
op2->gtType = simdType;
}
}
}
break;
#endif // FEATURE_SIMD
default:
// JCC nodes should not be present in HIR.
assert(node->OperGet() != GT_JCC);
break;
}
// Do some extra processing on top-level nodes to remove unused local reads.
if (node->OperIsLocalRead())
{
if (use.IsDummyUse())
{
comp->lvaDecRefCnts(node);
BlockRange().Remove(node);
}
else
{
// Local reads are side-effect-free; clear any flags leftover from frontend transformations.
node->gtFlags &= ~GTF_ALL_EFFECT;
}
}
assert(isLateArg == ((use.Def()->gtFlags & GTF_LATE_ARG) != 0));
return Compiler::WALK_CONTINUE;
}
void Rationalizer::RewriteAssignment(LIR::Use& use)
{
assert(use.IsInitialized());
GenTreeOp* assignment = use.Def()->AsOp();
assert(assignment->OperGet() == GT_ASG);
GenTree* location = assignment->gtGetOp1();
GenTree* value = assignment->gtGetOp2();
genTreeOps locationOp = location->OperGet();
if (assignment->OperIsBlkOp())
{
#ifdef FEATURE_SIMD
if (varTypeIsSIMD(location) && assignment->OperIsInitBlkOp())
{
if (location->OperGet() == GT_LCL_VAR)
{
var_types simdType = location->TypeGet();
GenTree* initVal = assignment->gtOp.gtOp2;
var_types baseType = comp->getBaseTypeOfSIMDLocal(location);
if (baseType != TYP_UNKNOWN)
{
GenTreeSIMD* simdTree = new (comp, GT_SIMD)
GenTreeSIMD(simdType, initVal, SIMDIntrinsicInit, baseType, genTypeSize(simdType));
assignment->gtOp.gtOp2 = simdTree;
value = simdTree;
initVal->gtNext = simdTree;
simdTree->gtPrev = initVal;
simdTree->gtNext = location;
location->gtPrev = simdTree;
}
}
}
#endif // FEATURE_SIMD
if ((location->TypeGet() == TYP_STRUCT) && !assignment->IsPhiDefn() && !value->IsMultiRegCall())
{
if ((location->OperGet() == GT_LCL_VAR))
{
// We need to construct a block node for the location.
// Modify lcl to be the address form.
location->SetOper(addrForm(locationOp));
LclVarDsc* varDsc = &(comp->lvaTable[location->AsLclVarCommon()->gtLclNum]);
location->gtType = TYP_BYREF;
GenTreeBlk* storeBlk = nullptr;
unsigned int size = varDsc->lvExactSize;
if (varDsc->lvStructGcCount != 0)
{
CORINFO_CLASS_HANDLE structHnd = varDsc->lvVerTypeInfo.GetClassHandle();
GenTreeObj* objNode = comp->gtNewObjNode(structHnd, location)->AsObj();
unsigned int slots = (unsigned)(roundUp(size, TARGET_POINTER_SIZE) / TARGET_POINTER_SIZE);
objNode->SetGCInfo(varDsc->lvGcLayout, varDsc->lvStructGcCount, slots);
objNode->ChangeOper(GT_STORE_OBJ);
objNode->SetData(value);
comp->fgMorphUnsafeBlk(objNode);
storeBlk = objNode;
}
else
{
storeBlk = new (comp, GT_STORE_BLK) GenTreeBlk(GT_STORE_BLK, TYP_STRUCT, location, value, size);
}
storeBlk->gtFlags |= (GTF_REVERSE_OPS | GTF_ASG);
storeBlk->gtFlags |= ((location->gtFlags | value->gtFlags) & GTF_ALL_EFFECT);
GenTree* insertionPoint = location->gtNext;
BlockRange().InsertBefore(insertionPoint, storeBlk);
use.ReplaceWith(comp, storeBlk);
BlockRange().Remove(assignment);
JITDUMP("After transforming local struct assignment into a block op:\n");
DISPTREERANGE(BlockRange(), use.Def());
JITDUMP("\n");
return;
}
else
{
assert(location->OperIsBlk());
}
}
}
switch (locationOp)
{
case GT_LCL_VAR:
case GT_LCL_FLD:
case GT_REG_VAR:
case GT_PHI_ARG:
RewriteAssignmentIntoStoreLclCore(assignment, location, value, locationOp);
BlockRange().Remove(location);
break;
case GT_IND:
{
GenTreeStoreInd* store =
new (comp, GT_STOREIND) GenTreeStoreInd(location->TypeGet(), location->gtGetOp1(), value);
copyFlags(store, assignment, GTF_ALL_EFFECT);
copyFlags(store, location, GTF_IND_FLAGS);
if (assignment->IsReverseOp())
{
store->gtFlags |= GTF_REVERSE_OPS;
}
// TODO: JIT dump
// Remove the GT_IND node and replace the assignment node with the store
BlockRange().Remove(location);
BlockRange().InsertBefore(assignment, store);
use.ReplaceWith(comp, store);
BlockRange().Remove(assignment);
}
break;
case GT_CLS_VAR:
{
location->SetOper(GT_CLS_VAR_ADDR);
location->gtType = TYP_BYREF;
assignment->SetOper(GT_STOREIND);
// TODO: JIT dump
}
break;
case GT_BLK:
case GT_OBJ:
case GT_DYN_BLK:
{
assert(varTypeIsStruct(location));
GenTreeBlk* storeBlk = location->AsBlk();
genTreeOps storeOper;
switch (location->gtOper)
{
case GT_BLK:
storeOper = GT_STORE_BLK;
break;
case GT_OBJ:
storeOper = GT_STORE_OBJ;
break;
case GT_DYN_BLK:
storeOper = GT_STORE_DYN_BLK;
break;
default:
unreached();
}
JITDUMP("Rewriting GT_ASG(%s(X), Y) to %s(X,Y):\n", GenTree::NodeName(location->gtOper),
GenTree::NodeName(storeOper));
storeBlk->SetOperRaw(storeOper);
storeBlk->gtFlags &= ~GTF_DONT_CSE;
storeBlk->gtFlags |= (assignment->gtFlags & (GTF_ALL_EFFECT | GTF_REVERSE_OPS | GTF_BLK_VOLATILE |
GTF_BLK_UNALIGNED | GTF_DONT_CSE));
storeBlk->gtBlk.Data() = value;
// Replace the assignment node with the store
use.ReplaceWith(comp, storeBlk);
BlockRange().Remove(assignment);
DISPTREERANGE(BlockRange(), use.Def());
JITDUMP("\n");
}
break;
default:
unreached();
break;
}
}
//------------------------------------------------------------------------
// 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);
}
//------------------------------------------------------------------------
// DecomposeStoreLclVar: Decompose GT_STORE_LCL_VAR.
//
// Arguments:
// use - the LIR::Use object for the def that needs to be decomposed.
//
// Return Value:
// The next node to process.
//
GenTree* DecomposeLongs::DecomposeStoreLclVar(LIR::Use& use)
{
assert(use.IsInitialized());
assert(use.Def()->OperGet() == GT_STORE_LCL_VAR);
GenTree* tree = use.Def();
GenTree* rhs = tree->gtGetOp1();
if ((rhs->OperGet() == GT_PHI) || (rhs->OperGet() == GT_CALL) ||
((rhs->OperGet() == GT_MUL_LONG) && (rhs->gtFlags & GTF_MUL_64RSLT) != 0))
{
// GT_CALLs are not decomposed, so will not be converted to GT_LONG
// GT_STORE_LCL_VAR = GT_CALL are handled in genMultiRegCallStoreToLocal
// GT_MULs are not decomposed, so will not be converted to GT_LONG
return tree->gtNext;
}
noway_assert(rhs->OperGet() == GT_LONG);
unsigned varNum = tree->AsLclVarCommon()->gtLclNum;
LclVarDsc* varDsc = m_compiler->lvaTable + varNum;
m_compiler->lvaDecRefCnts(tree);
GenTree* loRhs = rhs->gtGetOp1();
GenTree* hiRhs = rhs->gtGetOp2();
GenTree* hiStore = m_compiler->gtNewLclLNode(varNum, TYP_INT);
if (varDsc->lvPromoted)
{
assert(varDsc->lvFieldCnt == 2);
unsigned loVarNum = varDsc->lvFieldLclStart;
unsigned hiVarNum = loVarNum + 1;
tree->AsLclVarCommon()->SetLclNum(loVarNum);
hiStore->SetOper(GT_STORE_LCL_VAR);
hiStore->AsLclVarCommon()->SetLclNum(hiVarNum);
}
else
{
noway_assert(varDsc->lvLRACandidate == false);
tree->SetOper(GT_STORE_LCL_FLD);
tree->AsLclFld()->gtLclOffs = 0;
tree->AsLclFld()->gtFieldSeq = FieldSeqStore::NotAField();
hiStore->SetOper(GT_STORE_LCL_FLD);
hiStore->AsLclFld()->gtLclOffs = 4;
hiStore->AsLclFld()->gtFieldSeq = FieldSeqStore::NotAField();
}
// 'tree' is going to steal the loRhs node for itself, so we need to remove the
// GT_LONG node from the threading.
Range().Remove(rhs);
tree->gtOp.gtOp1 = loRhs;
tree->gtType = TYP_INT;
hiStore->gtOp.gtOp1 = hiRhs;
hiStore->gtFlags |= GTF_VAR_DEF;
m_compiler->lvaIncRefCnts(tree);
m_compiler->lvaIncRefCnts(hiStore);
Range().InsertAfter(tree, hiStore);
return hiStore->gtNext;
}
Compiler::fgWalkResult Rationalizer::RewriteNode(GenTree** useEdge, ArrayStack<GenTree*>& parentStack)
{
assert(useEdge != nullptr);
GenTree* node = *useEdge;
assert(node != nullptr);
#ifdef DEBUG
const bool isLateArg = (node->gtFlags & GTF_LATE_ARG) != 0;
#endif
// First, remove any preceeding GT_LIST nodes, which are not otherwise visited by the tree walk.
//
// NOTE: GT_LIST nodes that are used by block ops and phi nodes will in fact be visited.
for (GenTree* prev = node->gtPrev; prev != nullptr && prev->OperGet() == GT_LIST; prev = node->gtPrev)
{
BlockRange().Remove(prev);
}
// In addition, remove the current node if it is a GT_LIST node.
if ((*useEdge)->OperGet() == GT_LIST)
{
BlockRange().Remove(*useEdge);
return Compiler::WALK_CONTINUE;
}
LIR::Use use;
if (parentStack.Height() < 2)
{
use = LIR::Use::GetDummyUse(BlockRange(), *useEdge);
}
else
{
use = LIR::Use(BlockRange(), useEdge, parentStack.Index(1));
}
assert(node == use.Def());
switch (node->OperGet())
{
case GT_ASG:
RewriteAssignment(use);
break;
case GT_BOX:
// GT_BOX at this level just passes through so get rid of it
use.ReplaceWith(comp, node->gtGetOp1());
BlockRange().Remove(node);
break;
case GT_ADDR:
RewriteAddress(use);
break;
case GT_NOP:
// fgMorph sometimes inserts NOP nodes between defs and uses
// supposedly 'to prevent constant folding'. In this case, remove the
// NOP.
if (node->gtGetOp1() != nullptr)
{
use.ReplaceWith(comp, node->gtGetOp1());
BlockRange().Remove(node);
}
break;
case GT_COMMA:
{
GenTree* op1 = node->gtGetOp1();
if ((op1->gtFlags & GTF_ALL_EFFECT) == 0)
{
// The LHS has no side effects. Remove it.
bool isClosed = false;
unsigned sideEffects = 0;
LIR::ReadOnlyRange lhsRange = BlockRange().GetTreeRange(op1, &isClosed, &sideEffects);
// None of the transforms performed herein violate tree order, so these
// should always be true.
assert(isClosed);
assert((sideEffects & GTF_ALL_EFFECT) == 0);
BlockRange().Delete(comp, m_block, std::move(lhsRange));
}
GenTree* replacement = node->gtGetOp2();
if (!use.IsDummyUse())
{
use.ReplaceWith(comp, replacement);
}
else
{
// This is a top-level comma. If the RHS has no side effects we can remove
// it as well.
if ((replacement->gtFlags & GTF_ALL_EFFECT) == 0)
{
bool isClosed = false;
unsigned sideEffects = 0;
LIR::ReadOnlyRange rhsRange = BlockRange().GetTreeRange(replacement, &isClosed, &sideEffects);
// None of the transforms performed herein violate tree order, so these
// should always be true.
assert(isClosed);
assert((sideEffects & GTF_ALL_EFFECT) == 0);
BlockRange().Delete(comp, m_block, std::move(rhsRange));
}
}
BlockRange().Remove(node);
}
break;
case GT_ARGPLACE:
// Remove argplace and list nodes from the execution order.
//
// TODO: remove phi args and phi nodes as well?
BlockRange().Remove(node);
break;
#ifdef _TARGET_XARCH_
case GT_CLS_VAR:
{
// Class vars that are the target of an assignment will get rewritten into
// GT_STOREIND(GT_CLS_VAR_ADDR, val) by RewriteAssignment. This check is
// not strictly necessary--the GT_IND(GT_CLS_VAR_ADDR) pattern that would
// otherwise be generated would also be picked up by RewriteAssignment--but
// skipping the rewrite here saves an allocation and a bit of extra work.
const bool isLHSOfAssignment = (use.User()->OperGet() == GT_ASG) && (use.User()->gtGetOp1() == node);
if (!isLHSOfAssignment)
{
GenTree* ind = comp->gtNewOperNode(GT_IND, node->TypeGet(), node);
node->SetOper(GT_CLS_VAR_ADDR);
node->gtType = TYP_BYREF;
BlockRange().InsertAfter(node, ind);
use.ReplaceWith(comp, ind);
// TODO: JIT dump
}
}
break;
#endif // _TARGET_XARCH_
case GT_INTRINSIC:
// Non-target intrinsics should have already been rewritten back into user calls.
assert(Compiler::IsTargetIntrinsic(node->gtIntrinsic.gtIntrinsicId));
break;
#ifdef FEATURE_SIMD
case GT_INITBLK:
RewriteInitBlk(use);
break;
case GT_COPYBLK:
RewriteCopyBlk(use);
break;
case GT_OBJ:
RewriteObj(use);
break;
case GT_LCL_FLD:
case GT_STORE_LCL_FLD:
// TODO-1stClassStructs: Eliminate this.
FixupIfSIMDLocal(node->AsLclVarCommon());
break;
case GT_STOREIND:
case GT_IND:
if (node->gtType == TYP_STRUCT)
{
GenTree* addr = node->AsIndir()->Addr();
assert(addr->TypeGet() == TYP_BYREF);
if (addr->OperIsLocal())
{
LclVarDsc* varDsc = &(comp->lvaTable[addr->AsLclVarCommon()->gtLclNum]);
assert(varDsc->lvSIMDType);
unsigned simdSize = (unsigned int)roundUp(varDsc->lvExactSize, TARGET_POINTER_SIZE);
node->gtType = comp->getSIMDTypeForSize(simdSize);
}
#if DEBUG
else
{
// If the address is not a local var, assert that the user of this IND is an ADDR node.
assert((use.User()->OperGet() == GT_ADDR) || use.User()->OperIsLocalAddr());
}
#endif
}
break;
case GT_SIMD:
{
noway_assert(comp->featureSIMD);
GenTreeSIMD* simdNode = node->AsSIMD();
unsigned simdSize = simdNode->gtSIMDSize;
var_types simdType = comp->getSIMDTypeForSize(simdSize);
// TODO-1stClassStructs: This should be handled more generally for enregistered or promoted
// structs that are passed or returned in a different register type than their enregistered
// type(s).
if (simdNode->gtType == TYP_I_IMPL && simdNode->gtSIMDSize == TARGET_POINTER_SIZE)
{
// This happens when it is consumed by a GT_RET_EXPR.
// It can only be a Vector2f or Vector2i.
assert(genTypeSize(simdNode->gtSIMDBaseType) == 4);
simdNode->gtType = TYP_SIMD8;
}
else if (simdNode->gtType == TYP_STRUCT || varTypeIsSIMD(simdNode))
{
node->gtType = simdType;
}
// Certain SIMD trees require rationalizing.
if (simdNode->gtSIMD.gtSIMDIntrinsicID == SIMDIntrinsicInitArray)
{
// Rewrite this as an explicit load.
JITDUMP("Rewriting GT_SIMD array init as an explicit load:\n");
unsigned int baseTypeSize = genTypeSize(simdNode->gtSIMDBaseType);
GenTree* address = new (comp, GT_LEA) GenTreeAddrMode(TYP_BYREF, simdNode->gtOp1, simdNode->gtOp2,
baseTypeSize, offsetof(CORINFO_Array, u1Elems));
GenTree* ind = comp->gtNewOperNode(GT_IND, simdType, address);
BlockRange().InsertBefore(simdNode, address, ind);
use.ReplaceWith(comp, ind);
BlockRange().Remove(simdNode);
DISPTREERANGE(BlockRange(), use.Def());
JITDUMP("\n");
}
else
{
// This code depends on the fact that NONE of the SIMD intrinsics take vector operands
// of a different width. If that assumption changes, we will EITHER have to make these type
// transformations during importation, and plumb the types all the way through the JIT,
// OR add a lot of special handling here.
GenTree* op1 = simdNode->gtGetOp1();
if (op1 != nullptr && op1->gtType == TYP_STRUCT)
{
op1->gtType = simdType;
}
GenTree* op2 = simdNode->gtGetOp2();
if (op2 != nullptr && op2->gtType == TYP_STRUCT)
{
op2->gtType = simdType;
}
}
}
break;
#endif // FEATURE_SIMD
default:
break;
}
// Do some extra processing on top-level nodes to remove unused local reads.
if (use.IsDummyUse() && node->OperIsLocalRead())
{
assert((node->gtFlags & GTF_ALL_EFFECT) == 0);
comp->lvaDecRefCnts(node);
BlockRange().Remove(node);
}
assert(isLateArg == ((node->gtFlags & GTF_LATE_ARG) != 0));
return Compiler::WALK_CONTINUE;
}
// Transform CopyBlk involving SIMD vectors into stlclvar or stind of a SIMD type.
// Transformation is done if either src or dst are known to be SIMD vectors.
//
// Arguments:
// ppTree - A pointer-to-a-pointer for the GT_COPYBLK
// fgWalkData - A pointer to tree walk data providing the context
//
// Return Value:
// None.
//
// If either the source or the dst are known to be SIMD (a lclVar or SIMD intrinsic),
// get the simdType (TYP_DOUBLE or a SIMD type for SSE2) from the size of the SIMD node.
//
// For the source:
// - If it is a SIMD intrinsic or a lvSIMDType lclVar, change the node type to simdType.
// - Otherwise, add a GT_IND of simdType.
// For the dst:
// - If it is a lclVar of a SIMD type, chanage the node type to simdType.
// - Otherwise, change it to a GT_STORE_IND of simdType
//
// TODO-Cleanup: Once SIMD types are plumbed through the frontend, this will no longer
// be required.
//
void Rationalizer::RewriteCopyBlk(LIR::Use& use)
{
#ifdef FEATURE_SIMD
// No need to transofrm non-SIMD nodes, if featureSIMD is not enabled.
if (!comp->featureSIMD)
{
return;
}
// See if this is a SIMD copyBlk
GenTreeCpBlk* cpBlk = use.Def()->AsCpBlk();
GenTreePtr dstAddr = cpBlk->Dest();
GenTree* srcAddr = cpBlk->Source();
const bool srcIsSIMDAddr = comp->isAddrOfSIMDType(srcAddr);
const bool dstIsSIMDAddr = comp->isAddrOfSIMDType(dstAddr);
// Do not transform if neither src or dst is known to be a SIMD type.
// If src tree type is something we cannot reason but if dst is known to be of a SIMD type
// we will treat src tree as a SIMD type and vice versa.
if (!srcIsSIMDAddr && !dstIsSIMDAddr)
{
return;
}
// At this point it is known to be a copyblk of SIMD vectors and we can
// start transforming the original tree. Prior to this point do not perform
// any modifications to the original tree.
JITDUMP("\nRewriting SIMD CopyBlk\n");
DISPTREERANGE(BlockRange(), cpBlk);
// There are currently only three sizes supported: 8 bytes, 12 bytes, 16 bytes or the vector register length.
GenTreeIntConCommon* sizeNode = cpBlk->Size()->AsIntConCommon();
var_types simdType = comp->getSIMDTypeForSize((unsigned int)sizeNode->IconValue());
// Remove 'size' from execution order
BlockRange().Remove(sizeNode);
// Is destination a lclVar which is not an arg?
// If yes then we can turn it to a stlcl.var, otherwise turn into stind.
GenTree* simdDst = nullptr;
genTreeOps oper = GT_NONE;
if (dstIsSIMDAddr && dstAddr->OperIsLocalAddr())
{
simdDst = dstAddr;
simdDst->gtType = simdType;
oper = GT_STORE_LCL_VAR;
// For structs that are padded (e.g. Vector3f, Vector3i), the morpher will have marked them
// as GTF_VAR_USEASG. Unmark them.
simdDst->gtFlags &= ~(GTF_VAR_USEASG);
}
else
{
// Address of a non-local var
simdDst = dstAddr;
oper = GT_STOREIND;
}
GenTree* simdSrc = nullptr;
if ((srcAddr->OperGet() == GT_ADDR) && varTypeIsSIMD(srcAddr->gtGetOp1()))
{
// Get rid of parent node of GT_ADDR(..) if its child happens to be of a SIMD type.
BlockRange().Remove(srcAddr);
simdSrc = srcAddr->gtGetOp1();
}
else if (srcIsSIMDAddr && srcAddr->OperIsLocalAddr())
{
// If the source has been rewritten into a local addr node, rewrite it back into a
// local var node.
simdSrc = srcAddr;
simdSrc->SetOper(loadForm(srcAddr->OperGet()));
}
else
{
// Since destination is known to be a SIMD type, src must be a SIMD type too
// though we cannot figure it out easily enough. Transform src into
// GT_IND(src) of simdType.
GenTree* indir = comp->gtNewOperNode(GT_IND, simdType, srcAddr);
BlockRange().InsertAfter(srcAddr, indir);
cpBlk->gtGetOp1()->gtOp.gtOp2 = indir;
simdSrc = indir;
}
simdSrc->gtType = simdType;
// Change cpblk to either a st.lclvar or st.ind.
// At this point we are manipulating cpblk node with the knowledge of
// its internals (i.e. op1 is the size node, and the src & dst are in a GT_LIST on op2).
// This logic might need to be changed if we ever restructure cpblk node.
assert(simdDst != nullptr);
assert(simdSrc != nullptr);
GenTree* newNode = nullptr;
if (oper == GT_STORE_LCL_VAR)
{
newNode = simdDst;
newNode->SetOper(oper);
GenTreeLclVar* store = newNode->AsLclVar();
store->gtOp1 = simdSrc;
store->gtType = simdType;
store->gtFlags |= ((simdSrc->gtFlags & GTF_ALL_EFFECT) | GTF_ASG);
BlockRange().Remove(simdDst);
BlockRange().InsertAfter(simdSrc, store);
}
else
{
assert(oper == GT_STOREIND);
newNode = cpBlk->gtGetOp1();
newNode->SetOper(oper);
GenTreeStoreInd* storeInd = newNode->AsStoreInd();
storeInd->gtType = simdType;
storeInd->gtFlags |= ((simdSrc->gtFlags & GTF_ALL_EFFECT) | GTF_ASG);
storeInd->gtOp1 = simdDst;
storeInd->gtOp2 = simdSrc;
BlockRange().InsertBefore(cpBlk, storeInd);
}
use.ReplaceWith(comp, newNode);
BlockRange().Remove(cpBlk);
JITDUMP("After rewriting SIMD CopyBlk:\n");
DISPTREERANGE(BlockRange(), use.Def());
JITDUMP("\n");
#endif // FEATURE_SIMD
}
// Rewrite InitBlk involving SIMD vector into stlcl.var of a SIMD type.
//
// Arguments:
// ppTree - A pointer-to-a-pointer for the GT_INITBLK
// fgWalkData - A pointer to tree walk data providing the context
//
// Return Value:
// None.
//
// TODO-Cleanup: Once SIMD types are plumbed through the frontend, this will no longer
// be required.
//
void Rationalizer::RewriteInitBlk(LIR::Use& use)
{
#ifdef FEATURE_SIMD
// No lowering is needed for non-SIMD nodes, so early out if featureSIMD is not enabled.
if (!comp->featureSIMD)
{
return;
}
// See if this is a SIMD initBlk that needs to be changed to a simple st.lclVar.
GenTreeInitBlk* initBlk = use.Def()->AsInitBlk();
// Is the dstAddr is addr of a SIMD type lclVar?
GenTree* dstAddr = initBlk->Dest();
if (!comp->isAddrOfSIMDType(dstAddr) || !dstAddr->OperIsLocalAddr())
{
return;
}
unsigned lclNum = dstAddr->AsLclVarCommon()->gtLclNum;
if (!comp->lvaTable[lclNum].lvSIMDType)
{
return;
}
var_types baseType = comp->lvaTable[lclNum].lvBaseType;
CORINFO_CLASS_HANDLE typeHnd = comp->lvaTable[lclNum].lvVerTypeInfo.GetClassHandle();
unsigned simdLocalSize = comp->getSIMDTypeSizeInBytes(typeHnd);
JITDUMP("Rewriting SIMD InitBlk\n");
DISPTREERANGE(BlockRange(), initBlk);
assert((dstAddr->gtFlags & GTF_VAR_USEASG) == 0);
// There are currently only three sizes supported: 8 bytes, 16 bytes or the vector register length.
GenTreeIntConCommon* sizeNode = initBlk->Size()->AsIntConCommon();
unsigned int size = (unsigned int)roundUp(sizeNode->IconValue(), TARGET_POINTER_SIZE);
var_types simdType = comp->getSIMDTypeForSize(size);
assert(roundUp(simdLocalSize, TARGET_POINTER_SIZE) == size);
GenTree* initVal = initBlk->InitVal();
GenTreeSIMD* simdNode = new (comp, GT_SIMD)
GenTreeSIMD(simdType, initVal, SIMDIntrinsicInit, baseType, (unsigned)sizeNode->IconValue());
dstAddr->SetOper(GT_STORE_LCL_VAR);
GenTreeLclVar* store = dstAddr->AsLclVar();
store->gtType = simdType;
store->gtOp.gtOp1 = simdNode;
store->gtFlags |= ((simdNode->gtFlags & GTF_ALL_EFFECT) | GTF_ASG);
BlockRange().Remove(store);
// Insert the new nodes into the block
BlockRange().InsertAfter(initVal, simdNode, store);
use.ReplaceWith(comp, store);
// Remove the old size and GT_INITBLK nodes.
BlockRange().Remove(sizeNode);
BlockRange().Remove(initBlk);
JITDUMP("After rewriting SIMD InitBlk:\n");
DISPTREERANGE(BlockRange(), use.Def());
JITDUMP("\n");
#endif // FEATURE_SIMD
}
//------------------------------------------------------------------------
// DecomposeStoreLclVar: Decompose GT_STORE_LCL_VAR.
//
// Arguments:
// ppTree - the tree to decompose
// data - tree walk context
//
// Return Value:
// None.
//
void DecomposeLongs::DecomposeStoreLclVar(GenTree** ppTree, Compiler::fgWalkData* data)
{
assert(ppTree != nullptr);
assert(*ppTree != nullptr);
assert(data != nullptr);
assert((*ppTree)->OperGet() == GT_STORE_LCL_VAR);
assert(m_compiler->compCurStmt != nullptr);
GenTreeStmt* curStmt = m_compiler->compCurStmt->AsStmt();
GenTree* tree = *ppTree;
GenTree* nextTree = tree->gtNext;
GenTree* rhs = tree->gtGetOp1();
if ((rhs->OperGet() == GT_PHI) || (rhs->OperGet() == GT_CALL))
{
// GT_CALLs are not decomposed, so will not be converted to GT_LONG
// GT_STORE_LCL_VAR = GT_CALL are handled in genMultiRegCallStoreToLocal
return;
}
noway_assert(rhs->OperGet() == GT_LONG);
unsigned varNum = tree->AsLclVarCommon()->gtLclNum;
LclVarDsc* varDsc = m_compiler->lvaTable + varNum;
m_compiler->lvaDecRefCnts(tree);
GenTree* loRhs = rhs->gtGetOp1();
GenTree* hiRhs = rhs->gtGetOp2();
GenTree* hiStore = m_compiler->gtNewLclLNode(varNum, TYP_INT);
if (varDsc->lvPromoted)
{
assert(varDsc->lvFieldCnt == 2);
unsigned loVarNum = varDsc->lvFieldLclStart;
unsigned hiVarNum = loVarNum + 1;
tree->AsLclVarCommon()->SetLclNum(loVarNum);
hiStore->SetOper(GT_STORE_LCL_VAR);
hiStore->AsLclVarCommon()->SetLclNum(hiVarNum);
}
else
{
noway_assert(varDsc->lvLRACandidate == false);
tree->SetOper(GT_STORE_LCL_FLD);
tree->AsLclFld()->gtLclOffs = 0;
tree->AsLclFld()->gtFieldSeq = FieldSeqStore::NotAField();
hiStore->SetOper(GT_STORE_LCL_FLD);
hiStore->AsLclFld()->gtLclOffs = 4;
hiStore->AsLclFld()->gtFieldSeq = FieldSeqStore::NotAField();
}
tree->gtOp.gtOp1 = loRhs;
tree->gtType = TYP_INT;
loRhs->gtNext = tree;
tree->gtPrev = loRhs;
hiStore->gtOp.gtOp1 = hiRhs;
hiStore->CopyCosts(tree);
hiStore->gtFlags |= GTF_VAR_DEF;
m_compiler->lvaIncRefCnts(tree);
m_compiler->lvaIncRefCnts(hiStore);
tree->gtNext = hiRhs;
hiRhs->gtPrev = tree;
hiRhs->gtNext = hiStore;
hiStore->gtPrev = hiRhs;
hiStore->gtNext = nextTree;
if (nextTree != nullptr)
{
nextTree->gtPrev = hiStore;
}
nextTree = hiRhs;
bool isEmbeddedStmt = !curStmt->gtStmtIsTopLevel();
if (!isEmbeddedStmt)
{
tree->gtNext = nullptr;
hiRhs->gtPrev = nullptr;
}
InsertNodeAsStmt(hiStore);
}
void Rationalizer::RewriteAssignment(LIR::Use& use)
{
assert(use.IsInitialized());
GenTreeOp* assignment = use.Def()->AsOp();
assert(assignment->OperGet() == GT_ASG);
GenTree* location = assignment->gtGetOp1();
GenTree* value = assignment->gtGetOp2();
genTreeOps locationOp = location->OperGet();
#ifdef FEATURE_SIMD
if (varTypeIsSIMD(location) && assignment->OperIsInitBlkOp())
{
if (location->OperGet() == GT_LCL_VAR)
{
var_types simdType = location->TypeGet();
GenTree* initVal = assignment->gtOp.gtOp2;
var_types baseType = comp->getBaseTypeOfSIMDLocal(location);
if (baseType != TYP_UNKNOWN)
{
GenTreeSIMD* simdTree = new (comp, GT_SIMD)
GenTreeSIMD(simdType, initVal, SIMDIntrinsicInit, baseType, genTypeSize(simdType));
assignment->gtOp.gtOp2 = simdTree;
value = simdTree;
initVal->gtNext = simdTree;
simdTree->gtPrev = initVal;
simdTree->gtNext = location;
location->gtPrev = simdTree;
}
}
else
{
assert(location->OperIsBlk());
}
}
#endif // FEATURE_SIMD
switch (locationOp)
{
case GT_LCL_VAR:
case GT_LCL_FLD:
case GT_REG_VAR:
case GT_PHI_ARG:
RewriteAssignmentIntoStoreLclCore(assignment, location, value, locationOp);
BlockRange().Remove(location);
break;
case GT_IND:
{
GenTreeStoreInd* store =
new (comp, GT_STOREIND) GenTreeStoreInd(location->TypeGet(), location->gtGetOp1(), value);
copyFlags(store, assignment, GTF_ALL_EFFECT);
copyFlags(store, location, GTF_IND_FLAGS);
if (assignment->IsReverseOp())
{
store->gtFlags |= GTF_REVERSE_OPS;
}
// TODO: JIT dump
// Remove the GT_IND node and replace the assignment node with the store
BlockRange().Remove(location);
BlockRange().InsertBefore(assignment, store);
use.ReplaceWith(comp, store);
BlockRange().Remove(assignment);
}
break;
case GT_CLS_VAR:
{
location->SetOper(GT_CLS_VAR_ADDR);
location->gtType = TYP_BYREF;
assignment->SetOper(GT_STOREIND);
// TODO: JIT dump
}
break;
case GT_BLK:
case GT_OBJ:
case GT_DYN_BLK:
{
assert(varTypeIsStruct(location));
GenTreeBlk* storeBlk = location->AsBlk();
genTreeOps storeOper;
switch (location->gtOper)
{
case GT_BLK:
storeOper = GT_STORE_BLK;
break;
case GT_OBJ:
storeOper = GT_STORE_OBJ;
break;
case GT_DYN_BLK:
storeOper = GT_STORE_DYN_BLK;
break;
default:
unreached();
}
JITDUMP("Rewriting GT_ASG(%s(X), Y) to %s(X,Y):\n", GenTree::NodeName(location->gtOper),
GenTree::NodeName(storeOper));
storeBlk->SetOperRaw(storeOper);
storeBlk->gtFlags &= ~GTF_DONT_CSE;
storeBlk->gtFlags |= (assignment->gtFlags & (GTF_ALL_EFFECT | GTF_REVERSE_OPS | GTF_BLK_VOLATILE |
GTF_BLK_UNALIGNED | GTF_BLK_INIT | GTF_DONT_CSE));
storeBlk->gtBlk.Data() = value;
// Replace the assignment node with the store
use.ReplaceWith(comp, storeBlk);
BlockRange().Remove(assignment);
DISPTREERANGE(BlockRange(), use.Def());
JITDUMP("\n");
}
break;
default:
unreached();
break;
}
}