When building into frameworks, the generated code doesn't always have direct
access to the proto internals. Instead of opening up the access, just use the
public method to fetch the correct oneof.
Fixes https://github.com/google/protobuf/issues/1789
Currently some public API methods are defined in GenreatedMessage.java
and they have a generric return type:
class GeneratedMessage {
class Builder<BuilderType extends Builder<BuilderType>> {
public BuilderType setField(...);
public BuilderType setExtension(...);
}
}
With these definitions, the compiled byte code of a callsite will have
a direct reference to GeneratedMessage. For example:
fooBuilder.setField(...);
becomes:
##: invokevirtual // Method Builder.setField:(...)LGeneratedMessage.Builder
##: checkcast // class Builder
This will prevent us from updating generated classes to subclass a
different versioned GeneratedMessageV3 class in the future (we can't do
it in a binary compatible way).
This change addresses the problem by overriding these methods directly
in the generated class:
class Foo {
class Builder extends GeneratedMessage.Builder<Builder> {
public Builder setField(...) {
return super.setField(...);
}
}
}
After this, fooBuilder.setField(...) will be compiled to:
##: invokevirtual // Method Builder.setField:(...)LFoo.Builder
The callsites will no longer reference GeneratedMessage directly and we
can change Foo to subclass GeneratedMessageV3 without breaking binary
compatiblity.
The downside of this change is:
1. It increases generated code size (though it saves some instructions
on the callsites).
2. We can never stop generating these overrides because doing that
will break binary compatibility.
Change-Id: I879afbbc1325a66324a51565e017143489b06e97
We now just perform the optimization within AddRange itself.
This is a breaking change in terms of "drop in the DLL", but is
source compatible, which should be fine.
As bazel folks are looking at getting auto generation of module maps going and
the importing of sources files causes issues there. We were only do it to
hack around some of the apple linker behaviors around objc classes and
categories, but even that isn't complete and CocoaPods was already doing -ObjC,
and developers not using pods could have still needed it to ensure everything
was linked anyways; so drop the hack of importing sources.
A couple of small tweaks were needed to fix test failures that were
inadvertently introduced in the last down integration. I also had to add
the map.js file which was missing.
