Skip to content

[Flang] Propogate fir.declare attributes through cg-rewrite #137207

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

Closed
wants to merge 5 commits into from
Closed

[Flang] Propogate fir.declare attributes through cg-rewrite #137207

Show file tree
Hide file tree
Changes from 2 commits
Commits
Show all changes
5 commits
Select commit Hold shift + click to select a range
3301f3e
propogate attributes through cg-rewrite
SusanTan Apr 24, 2025
353e86a
format
SusanTan Apr 24, 2025
a26f871
Merge remote-tracking branch 'origin/main' into susant/cg-rewrite
SusanTan Apr 28, 2025
aa366ae
remove propogation when fir.declare is removed
SusanTan Apr 28, 2025
196dfeb
update comment
SusanTan Apr 28, 2025
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
21 changes: 20 additions & 1 deletion flang/lib/Optimizer/CodeGen/PreCGRewrite.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -281,6 +281,20 @@ class DeclareOpConversion : public mlir::OpRewritePattern<fir::DeclareOp> {
matchAndRewrite(fir::DeclareOp declareOp,
mlir::PatternRewriter &rewriter) const override {
if (!preserveDeclare) {
auto memrefOp = declareOp.getMemref().getDefiningOp();
Copy link
Contributor

@jeanPerier jeanPerier Apr 25, 2025

Choose a reason for hiding this comment

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

I find it a bit weird to propagate the attributes on the declare memref input producing op. First of all, you may end-up in weird situations where a single alloca maps to several declare (looking at you equivalences), so that could conflict.

Why do you need to propagate attribute if the cg declare is not emitted?
It looks like to me that if one wants to preserve language level info, the declare should also be preserved.

Copy link
Contributor Author

@SusanTan SusanTan Apr 28, 2025

Choose a reason for hiding this comment

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

that's a good catch, thanks Jean! I will remove this part and sought solutions downstream. thanks!

if (!memrefOp) {
rewriter.replaceOp(declareOp, declareOp.getMemref());
Copy link
Contributor

@razvanlupusoru razvanlupusoru Apr 24, 2025

Choose a reason for hiding this comment

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

Instead of duplicating this code - I would simply guard the attachment of attributes with if (declareOp.getMemref().getDefiningOp())

return mlir::success();
}

// attach metadatas from the fir.declare to its memref (if it's an
// operation)
mlir::NamedAttrList elidedAttrs =
mlir::NamedAttrList{memrefOp->getAttrs()};
for (const mlir::NamedAttribute &attr : declareOp->getAttrs())
if (!elidedAttrs.get(attr.getName()))
memrefOp->setAttr(attr.getName(), attr.getValue());

rewriter.replaceOp(declareOp, declareOp.getMemref());
return mlir::success();
}
Expand All @@ -299,11 +313,16 @@ class DeclareOpConversion : public mlir::OpRewritePattern<fir::DeclareOp> {
else
return mlir::failure();
}
// FIXME: Add FortranAttrs and CudaAttrs

auto xDeclOp = rewriter.create<fir::cg::XDeclareOp>(
loc, declareOp.getType(), declareOp.getMemref(), shapeOpers, shiftOpers,
declareOp.getTypeparams(), declareOp.getDummyScope(),
declareOp.getUniqName());

// attach metadatas from fir.declare to fircg.ext_declare
Copy link
Contributor

@razvanlupusoru razvanlupusoru Apr 28, 2025

Choose a reason for hiding this comment

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

I would update this comment to say something like:
"// Propagate all attributes to fircg.ext_declare to avoid losing metadata including CUDA and OpenACC attributes."

for (const mlir::NamedAttribute &attr : declareOp->getAttrs())
xDeclOp->setAttr(attr.getName(), attr.getValue());
Copy link
Contributor

@jeanPerier jeanPerier Apr 30, 2025

Choose a reason for hiding this comment

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

How is this not overriding the operandSegmentSizes attribute of the AttrSizedOperandSegments interface of the newly created xDeclOp with the one from the declareOp?

I would suggest checking if the attributes names are not already used in the xDeclop to avoid any issues.


LLVM_DEBUG(llvm::dbgs()
<< "rewriting " << declareOp << " to " << xDeclOp << '\n');
rewriter.replaceOp(declareOp, xDeclOp.getOperation()->getResults());
Expand Down
22 changes: 19 additions & 3 deletions flang/test/Fir/declare-codegen.fir
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -23,7 +23,7 @@ func.func private @bar(%arg0: !fir.ref<!fir.array<12x23xi32>>)

// DECL-LABEL: func.func @test(
// DECL-SAME: %[[arg0:.*]]: !fir.ref<!fir.array<12x23xi32>>) {
// DECL: fircg.ext_declare
// DECL: fircg.ext_declare{{.*}}{uniq_name = "_QFarray_numeric_lboundsEx"}


func.func @useless_shape_with_duplicate_extent_operand(%arg0: !fir.ref<!fir.array<3x3xf32>>) {
Expand All @@ -37,7 +37,7 @@ func.func @useless_shape_with_duplicate_extent_operand(%arg0: !fir.ref<!fir.arra
// NODECL-NEXT: return

// DECL-LABEL: func.func @useless_shape_with_duplicate_extent_operand(
// DECL: fircg.ext_declare
// DECL: fircg.ext_declare{{.*}}{uniq_name = "u"}

// Test DCE does not crash because of unreachable code.
func.func @unreachable_code(%arg0: !fir.ref<!fir.char<1,10>>) {
Expand All @@ -51,4 +51,20 @@ func.func @unreachable_code(%arg0: !fir.ref<!fir.char<1,10>>) {
// NODECL-LABEL: func.func @unreachable_code(
// NODECL-NOT: uniq_name = "live_code"
// DECL-LABEL: func.func @unreachable_code(
// DECL: uniq_name = "live_code"
// DECL: fircg.ext_declare{{.*}}{uniq_name = "live_code"}
// DECL: fir.declare{{.*}}{uniq_name = "dead_code"}


// Test that attributes get attached to the memref operation when fir.declare is not preserved
func.func @test_local_memref() {
%0 = fir.alloca !fir.array<10xi32> {uniq_name = "local_array"}
%1 = fir.declare %0 {uniq_name = "local_array_declare", acc.declare = #acc.declare<dataClause = acc_create>} : (!fir.ref<!fir.array<10xi32>>) -> !fir.ref<!fir.array<10xi32>>
return
}
// NODECL-LABEL: func.func @test_local_memref() {

// Note: the uniq_name is not attached of fir.declare is not preserved because it conflicts with the uniq_name of the memref
// NODECL: fir.alloca{{.*}}{acc.declare = #acc.declare<dataClause = acc_create>, uniq_name = "local_array"}

// DECL-LABEL: func.func @test_local_memref() {
// DECL: fircg.ext_declare{{.*}}{acc.declare = #acc.declare<dataClause = acc_create>, uniq_name = "local_array_declare"}
Loading