There are several changes here:
1) When wrapping a callback interface object for JS, just extract the
underlying JSObject from inside it and hand that object out.
2) Flag callback interface descriptors as "not concrete" (only matters
for cases when they have constants on the interface object) and not
wrappercached (will catch bugs if someone tries to treat them as a
Gecko object).
3) Fix a preexisting bug in sequence wrapping where we'd try to
JS_DefineElement twice if we were wrapping a null value for a
sequence of nullable interface objects.
The new setup is:
* Consequential interfaces with no explicit descriptor are still skipped.
* If no information is provided about an interface, an empty descriptor is
assumed.
* If a list is provided and the only entry is a worker descriptor, an empty
main-thread descriptor is assumed.
The valid annotations are [Throws], [GetterThrows], and [SetterThrows], which can all either take no value or
be set to MainThread or Workers if the throwing behavior is only happening on main thread or in workers
In this new setup, there are three new extended attributes: Infallible,
GetterInfallible, SetterInfallible. The first one applies to both methods and
attributes, while the last two apply only to attributes. Each one can be
either set with no value specified (in which case it applies both on main
thread and in workers) or can be set to MainThread or Workers to limit where it
applies.
If Infallible is set on an attribute, then GetterInfallible and
SetterInfallible will be ignored. So if you want to specify, for example, an
attribute that's infallible on the main thread and has an infallible getter in
workers, the right way to do that is [GetterInfallible,
SetterInfallible=MainThread].
Another implementation option would be to put all the dictionaries in a single
file and have a static global set of ids which works across all dictionaries
and is initialized once at startup or so. That would also handle cross-file
dictionary inheritance better.
One problem that remains is the fake descriptor business. At the moment this
does not allow interface types inside dictionaries. We could probably work
around this by either refactoring code to make it possible to get the declType
independently of the actual conversion template (whether because it lives in a
separate function or because the conversion template generator knows to just
return an empty string when the fake descriptor provirder is passed) or by
figuring out a way to pass an actual descriptor provider to dictionary codegen.
Boris Zbarsky
Peter Van der Beken