Replacing template alias that does not use its template parameter with one that does is broken about rust-bindgen HOT 6 CLOSED

I think it's not about being smarter about type parameters, but about being replacement-aware. That's partially why the replaces annotation sucks quite a bit (and it'd come a bit more handy to have a replaced_by annotation or something, but that's way more annoying).

Right now the amount of code that is replacement-aware is minimal. Probably we should make Context::resolve_item replacement-aware by default (it's not hard, just slightly annoying), and then have another function to bypass this in the appropriate places.

from rust-bindgen.

After digging in more, this is not what we thought this was. Code generation, which is when we get the named types we need to emit, happens after we have already processed replacements, so the issue is not that we are failing to be replacement aware.

In fact, it looks like we end up replacing the wrong items, and then end up ignoring the replacements or something.

I'm digging in further...

from rust-bindgen.

First, we parse the replaces_MaybeWrapped template alias, which has the use_instead_of (né replaces) annotation, and it ends up with an id of ItemId(13):

DEBUG:bindgen::ir::context: BindgenContext::add_item(Item { id: ItemId(13), local_id: Cell { value: None }, next_child_local_id: Cell { value: 1 }, canonical_name_cache: RefCell { value: None }, comment: Some("/// But the replacement type does use T!\n///\n/// <div rustbindgen replaces=\"MaybeWrapped\" />"), annotations: Annotations { opaque: false, hide: false, use_instead_of: Some("MaybeWrapped"), disallow_copy: false, private_fields: None, accessor_kind: None }, parent_id: ItemId(0), kind: Type(Type { name: None, layout: None, kind: TemplateAlias(ItemId(16), [ItemId(15)]), is_const: false }) }, declaration: Some(Cursor( kind: NoDeclFound, loc: builtin definitions, usr: None)), loc: Some(Cursor(replaces_MaybeWrapped kind: TypeAliasTemplateDecl, loc: tests/headers/replace_template_alias.hpp:23:1, usr: Some("c:@replaces_MaybeWrapped")))

Then, we define the replacement for MaybeWrapped (ie a call to BindgenContext::replace) as ItemId(18)(!!) and not the template alias above which is ItemId(13):

DEBUG:bindgen::ir::context: Defining replacement for MaybeWrapped as ItemId(18)

Finally, we add the item with id itemId(18), which is the named template type argument T:

DEBUG:bindgen::ir::context: BindgenContext::add_item(Item { id: ItemId(18), local_id: Cell { value: None }, next_child_local_id: Cell { value: 1 }, canonical_name_cache: RefCell { value: None }, comment: None, annotations: Annotations { opaque: false, hide: false, use_instead_of: None, disallow_copy: false, private_fields: None, accessor_kind: None }, parent_id: ItemId(13), kind: Type(Type { name: Some("T"), layout: None, kind: Named("T", None), is_const: false }) }, declaration: None, loc: None

So that seems to be the bug: we should be defining the replacement as ItemId(13) not ItemId(18).

Still figuring out why this is happening in the first place.

from rust-bindgen.

I bet we're inserting it twice in the hashmap?

from rust-bindgen.

I bet we're inserting it twice in the hashmap?

Great guess! ;)

DEBUG:bindgen::ir::context: Defining replacement for MaybeWrapped as ItemId(13)
DEBUG:bindgen::ir::context: Defining replacement for MaybeWrapped as ItemId(16)
DEBUG:bindgen::ir::context: Defining replacement for MaybeWrapped as ItemId(18)

from rust-bindgen.

I guess we can assume the first is the correct replacement and then warn for duplicate replacement definition attempts.

Fix incoming.

from rust-bindgen.