Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Jump to bottom

Substitute GT_RET_EXPR in inline candidate arguments #69117

Merged
merged 6 commits into from
May 13, 2022
Merged

Substitute GT_RET_EXPR in inline candidate arguments #69117

Show file tree
Hide file tree
Changes from 5 commits
Commits
Show all changes
6 commits
Select commit Hold shift + click to select a range
ea46012
Substitute GT_RET_EXPR in inline candidate arguments before inlining
jakobbotsch May 10, 2022
857c50d
Clean up usages of gtRetExprVal
jakobbotsch May 10, 2022
28b116f
Small cleanup
jakobbotsch May 10, 2022
f837f19
Fix a comment
jakobbotsch May 10, 2022
41e0af2
Do late devirt on args too
jakobbotsch May 10, 2022
ea9e136
Rename GTF_CALL_M_LATE_DEVIRT -> GTF_CALL_M_HAS_LATE_DEVIRT_INFO
jakobbotsch May 11, 2022
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
95 changes: 49 additions & 46 deletions src/coreclr/jit/fginline.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -121,6 +121,22 @@ PhaseStatus Compiler::fgInline()
// candidate stage. So scan the tree looking for those early failures.
fgWalkTreePre(stmt->GetRootNodePointer(), fgFindNonInlineCandidate, stmt);
#endif
// See if we need to replace some return value place holders.
// Also, see if this replacement enables further devirtualization.
//
// Note we have both preorder and postorder callbacks here.
//
// The preorder callback is responsible for replacing GT_RET_EXPRs
// with the appropriate expansion (call or inline result).
// Replacement may introduce subtrees with GT_RET_EXPR and so
// we rely on the preorder to recursively process those as well.
//
// On the way back up, the postorder callback then re-examines nodes for
// possible further optimization, as the (now complete) GT_RET_EXPR
// replacement may have enabled optimizations by providing more
// specific types for trees or variables.
fgWalkTree(stmt->GetRootNodePointer(), fgUpdateInlineReturnExpressionPlaceHolder, fgLateDevirtualization,
&madeChanges);

GenTree* expr = stmt->GetRootNode();

Expand Down Expand Up @@ -162,23 +178,6 @@ PhaseStatus Compiler::fgInline()
}
}

// See if we need to replace some return value place holders.
// Also, see if this replacement enables further devirtualization.
//
// Note we have both preorder and postorder callbacks here.
//
// The preorder callback is responsible for replacing GT_RET_EXPRs
// with the appropriate expansion (call or inline result).
// Replacement may introduce subtrees with GT_RET_EXPR and so
// we rely on the preorder to recursively process those as well.
//
// On the way back up, the postorder callback then re-examines nodes for
// possible further optimization, as the (now complete) GT_RET_EXPR
// replacement may have enabled optimizations by providing more
// specific types for trees or variables.
fgWalkTree(stmt->GetRootNodePointer(), fgUpdateInlineReturnExpressionPlaceHolder, fgLateDevirtualization,
(void*)&madeChanges);

// See if stmt is of the form GT_COMMA(call, nop)
// If yes, we can get rid of GT_COMMA.
if (expr->OperGet() == GT_COMMA && expr->AsOp()->gtOp1->OperGet() == GT_CALL &&
Expand Down Expand Up @@ -492,7 +491,7 @@ Compiler::fgWalkResult Compiler::fgUpdateInlineReturnExpressionPlaceHolder(GenTr
Compiler* comp = data->compiler;
CORINFO_CLASS_HANDLE retClsHnd = NO_CLASS_HANDLE;

while (tree->OperGet() == GT_RET_EXPR)
while (tree->OperIs(GT_RET_EXPR))
{
// We are going to copy the tree from the inlinee,
// so record the handle now.
Expand All @@ -505,15 +504,21 @@ Compiler::fgWalkResult Compiler::fgUpdateInlineReturnExpressionPlaceHolder(GenTr
// Skip through chains of GT_RET_EXPRs (say from nested inlines)
// to the actual tree to use.
//
// Also we might as well try and fold the return value.
// Eg returns of constant bools will have CASTS.
// This folding may uncover more GT_RET_EXPRs, so we loop around
// until we've got something distinct.
BasicBlockFlags bbFlags;
GenTree* inlineCandidate = tree;
do
{
GenTreeRetExpr* retExpr = inlineCandidate->AsRetExpr();
inlineCandidate = retExpr->gtInlineCandidate;
bbFlags = retExpr->bbFlags;
} while (inlineCandidate->OperIs(GT_RET_EXPR));

// We might as well try and fold the return value. Eg returns of
// constant bools will have CASTS. This folding may uncover more
// GT_RET_EXPRs, so we loop around until we've got something distinct.
//
BasicBlockFlags bbFlags = BBF_EMPTY;
GenTree* inlineCandidate = tree->gtRetExprVal(&bbFlags);
inlineCandidate = comp->gtFoldExpr(inlineCandidate);
var_types retType = tree->TypeGet();
inlineCandidate = comp->gtFoldExpr(inlineCandidate);
var_types retType = tree->TypeGet();

#ifdef DEBUG
if (comp->verbose)
Expand Down Expand Up @@ -718,8 +723,12 @@ Compiler::fgWalkResult Compiler::fgLateDevirtualization(GenTree** pTree, fgWalkD

if ((call->gtCallMoreFlags & GTF_CALL_M_LATE_DEVIRT) != 0)
{
context = call->gtLateDevirtualizationInfo->exactContextHnd;
call->gtLateDevirtualizationInfo = nullptr;
context = call->gtLateDevirtualizationInfo->exactContextHnd;
// Note: we might call this multiple times for the same trees.
// If the devirtualization below succeeds, the call becomes
// non-virtual and we won't get here again. If it does not
// succeed we might get here again so we keep the late devirt
// info.
}

comp->impDevirtualizeCall(call, nullptr, &method, &methodFlags, &context, nullptr, isLateDevirtualization,
Expand Down Expand Up @@ -1537,8 +1546,7 @@ Statement* Compiler::fgInlinePrependStatements(InlineInfo* inlineInfo)
bool argHasPutArg = !varTypeIsStruct(arg->GetSignatureType()) &&
(genTypeSize(argNode) != genTypeSize(arg->GetSignatureType()));

BasicBlockFlags bbFlags = BBF_EMPTY;
argNode = argNode->gtRetExprVal(&bbFlags);
assert(!argNode->OperIs(GT_RET_EXPR));

if (argInfo.argHasTmp)
{
Expand Down Expand Up @@ -1601,7 +1609,6 @@ Statement* Compiler::fgInlinePrependStatements(InlineInfo* inlineInfo)
}
#endif // DEBUG
}
block->bbFlags |= (bbFlags & BBF_SPLIT_GAINED);
}
else if (argInfo.argIsByRefToStructLocal)
{
Expand Down Expand Up @@ -1649,39 +1656,37 @@ Statement* Compiler::fgInlinePrependStatements(InlineInfo* inlineInfo)
// This helper call is marked as a "Special DCE" helper during
// importation, over in fgGetStaticsCCtorHelper.
//
// (2) NYI. If, after tunneling through GT_RET_VALs, we find that
// the actual arg expression has no side effects, we can skip
// appending all together. This will help jit TP a bit.
// (2) NYI. If we find that the actual arg expression
// has no side effects, we can skip appending all
// together. This will help jit TP a bit.
//
// Chase through any GT_RET_EXPRs to find the actual argument
// expression.
GenTree* actualArgNode = argNode->gtRetExprVal(&bbFlags);
assert(!argNode->OperIs(GT_RET_EXPR));

// For case (1)
//
// Look for the following tree shapes
// prejit: (IND (ADD (CONST, CALL(special dce helper...))))
// jit : (COMMA (CALL(special dce helper...), (FIELD ...)))
if (actualArgNode->gtOper == GT_COMMA)
if (argNode->gtOper == GT_COMMA)
{
// Look for (COMMA (CALL(special dce helper...), (FIELD ...)))
GenTree* op1 = actualArgNode->AsOp()->gtOp1;
GenTree* op2 = actualArgNode->AsOp()->gtOp2;
GenTree* op1 = argNode->AsOp()->gtOp1;
GenTree* op2 = argNode->AsOp()->gtOp2;
if (op1->IsCall() &&
((op1->AsCall()->gtCallMoreFlags & GTF_CALL_M_HELPER_SPECIAL_DCE) != 0) &&
(op2->gtOper == GT_FIELD) && ((op2->gtFlags & GTF_EXCEPT) == 0))
{
JITDUMP("\nPerforming special dce on unused arg [%06u]:"
" actual arg [%06u] helper call [%06u]\n",
argNode->gtTreeID, actualArgNode->gtTreeID, op1->gtTreeID);
argNode->gtTreeID, argNode->gtTreeID, op1->gtTreeID);
// Drop the whole tree
append = false;
}
}
else if (actualArgNode->gtOper == GT_IND)
else if (argNode->gtOper == GT_IND)
{
// Look for (IND (ADD (CONST, CALL(special dce helper...))))
GenTree* addr = actualArgNode->AsOp()->gtOp1;
GenTree* addr = argNode->AsOp()->gtOp1;

if (addr->gtOper == GT_ADD)
{
Expand All @@ -1694,7 +1699,7 @@ Statement* Compiler::fgInlinePrependStatements(InlineInfo* inlineInfo)
// Drop the whole tree
JITDUMP("\nPerforming special dce on unused arg [%06u]:"
" actual arg [%06u] helper call [%06u]\n",
argNode->gtTreeID, actualArgNode->gtTreeID, op1->gtTreeID);
argNode->gtTreeID, argNode->gtTreeID, op1->gtTreeID);
append = false;
}
}
Expand Down Expand Up @@ -1731,8 +1736,6 @@ Statement* Compiler::fgInlinePrependStatements(InlineInfo* inlineInfo)
// since the box itself will be ignored.
gtTryRemoveBoxUpstreamEffects(argNode);
}

block->bbFlags |= (bbFlags & BBF_SPLIT_GAINED);
}
}
}
Expand Down
36 changes: 0 additions & 36 deletions src/coreclr/jit/gentree.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -6161,42 +6161,6 @@ GenTree* GenTree::gtGetParent(GenTree*** pUse)
return user;
}

//-------------------------------------------------------------------------
// gtRetExprVal - walk back through GT_RET_EXPRs
//
// Arguments:
// pbbFlags - out-parameter that is set to the flags of the basic block
// containing the inlinee return value. The value is 0
// for unsuccessful inlines.
//
// Returns:
// tree representing return value from a successful inline,
// or original call for failed or yet to be determined inline.
//
// Notes:
// Multi-level inlines can form chains of GT_RET_EXPRs.
// This method walks back to the root of the chain.
//
GenTree* GenTree::gtRetExprVal(BasicBlockFlags* pbbFlags /* = nullptr */)
{
GenTree* retExprVal = this;
BasicBlockFlags bbFlags = BBF_EMPTY;

while (retExprVal->OperIs(GT_RET_EXPR))
{
const GenTreeRetExpr* retExpr = retExprVal->AsRetExpr();
bbFlags = retExpr->bbFlags;
retExprVal = retExpr->gtInlineCandidate;
}

if (pbbFlags != nullptr)
{
*pbbFlags = bbFlags;
}

return retExprVal;
}

//------------------------------------------------------------------------------
// OperRequiresAsgFlag : Check whether the operation requires GTF_ASG flag regardless
// of the children's flags.
Expand Down
3 changes: 0 additions & 3 deletions src/coreclr/jit/gentree.h
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -1762,9 +1762,6 @@ struct GenTree

inline GenTree* gtCommaAssignVal();

// Tunnel through any GT_RET_EXPRs
GenTree* gtRetExprVal(BasicBlockFlags* pbbFlags = nullptr);

// Return the child of this node if it is a GT_RELOAD or GT_COPY; otherwise simply return the node itself
inline GenTree* gtSkipReloadOrCopy();

Expand Down
6 changes: 4 additions & 2 deletions src/coreclr/jit/importer.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -19572,7 +19572,7 @@ void Compiler::impInlineRecordArgInfo(InlineInfo* pInlineInfo,
inlCurArgInfo->arg = arg;
GenTree* curArgVal = arg->GetNode();

curArgVal = curArgVal->gtRetExprVal();
assert(!curArgVal->OperIs(GT_RET_EXPR));

if (curArgVal->gtOper == GT_MKREFANY)
{
Expand Down Expand Up @@ -20168,7 +20168,8 @@ GenTree* Compiler::impInlineFetchArg(unsigned lclNum, InlArgInfo* inlArgInfo, In
const var_types lclTyp = lclInfo.lclTypeInfo;
GenTree* op1 = nullptr;

GenTree* argNode = argInfo.arg->GetNode()->gtRetExprVal();
GenTree* argNode = argInfo.arg->GetNode();
assert(!argNode->OperIs(GT_RET_EXPR));

if (argInfo.argIsInvariant && !argCanBeModified)
{
Expand Down Expand Up @@ -21299,6 +21300,7 @@ void Compiler::impDevirtualizeCall(GenTreeCall* call,
// it's a union field used for other things by virtual
// stubs)
call->gtInlineCandidateInfo = nullptr;
call->gtCallMoreFlags &= ~GTF_CALL_M_LATE_DEVIRT;
Copy link
Member

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Maybe we should rename this flag? GTF_CALL_M_HAS_LATE_DEVIRT_INFO perhaps? I introduced it but the name could better describe what it means.

Copy link
Member Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Done.


#if defined(DEBUG)
if (verbose)
Expand Down