[Clang][RFC] Resugar attributed type alias

[zyn0217]

#142608 reflects a case where the type sugar on the attributed type alias is lost.

It might be possible to add the sugar back at the cost of duplicating the type alias decl, along with a modified type source info.

Though I'm not sure if the solution is feasible, and since there's no test regression I'll just elaborate the difference through a diagnostic.

[mizvekov]

Suggested change

	auto *ET = cast<TypedefType>(Old);
	QualType Underlying = wrap(C, ET->desugar(), I);
	TypedefNameDecl *Typedef;
	if (auto *TD = dyn_cast<TypedefDecl>(ET->getDecl())) {
	Typedef = TypedefDecl::Create(C, TD->getDeclContext(),
	TD->getBeginLoc(), TD->getLocation(),
	TD->getIdentifier(), nullptr);
	Typedef->setModedTypeSourceInfo(TD->getTypeSourceInfo(), Underlying);
	} else {
	auto *Alias = cast<TypeAliasDecl>(ET->getDecl());
	Typedef = TypedefDecl::Create(
	C, Alias->getDeclContext(), Alias->getBeginLoc(),
	Alias->getLocation(), Alias->getIdentifier(), nullptr);
	Typedef->setModedTypeSourceInfo(Alias->getTypeSourceInfo(),
	Underlying);
	}
	Typedef->setPreviousDecl(ET->getDecl());
	return C.getTypedefType(Typedef, Underlying);
	return C.getTypedefType(cast<TypedefType>(Old)->getDecl(), Underlying);

[mizvekov]

Otherwise, passing that second 'Underlying' argument would be unnecessary.

The reason we have that second parameter is to support a feature called 'divergent typedefs', which was added a couple of years ago in order to make resugaring more efficient.

This is currently used by the getCommonSugaredType infrastructure as well.

But yeah, this allows you to invent a different underlying type for a TypedefType, without having to create a fake TypedefDecl for it.

[zyn0217]

But yeah, this allows you to invent a different underlying type for a TypedefType, without having to create a fake TypedefDecl for it.

No? There's an assert in that function:

assert(hasSameType(Decl->getUnderlyingType(), Underlying));

[mizvekov]

Yes, that's has same type, not identical types. Are we not just modifying the type sugar here?

[zyn0217]

The context of this resugar is to find the FunctionType and add an attribute (by creating a new function type) on it.

The resugarer tries to add back anything it gets rid of, like the TypedefType in this case.

Or do you know if there's other way around to avoid copying the function type?

[mizvekov]

Yeah, but the new function type is still the same as the original one, it's just not identical (different sugar).

If that's the case, that's still under contract with getTypedefType.

If that's not the case, I think it would be possible to remove that assert, you just need to investigate and deal with the consequences elsewhere.

[zyn0217]

Thanks, I'll take a look

[mizvekov]

In both cases, either divergent typedef or creating a new decl, if the types are not the same, you have to deal with these consequences.

For example, getCommonSugaredType expects all TypedefTypes of the same typedef declaration to be the same type.