fix: boxed string field

[xxchan]

fix #1159

Signed-off-by: xxchan [email protected]

[caspermeijn]

I think this test will pass without the other changes. CodeGenerator already makes the OneOf field a Box<T> if self.boxed() == true on line 667.

Can you explain why the other changes are needed?

[xxchan]

Hi @caspermeijn, thanks for the prompt review! Reverted unnecessary changes, and made CI Green.

I think this test will pass without the other changes.

I'm not sure what you mean by "other changes" here.

But anyway, here's how it works:

As mentioned in #1159, the type for the oneof field was correct. What was wrong is the generated merge function.

I noticed there are #[prost(boxed)] on normal messages fields, but not on the oneof field. After adding it, it works.

pub struct Foo {
    #[prost(message, optional, boxed, tag = "1")]
    pub bar: ::core::option::Option<::prost::alloc::boxed::Box<Bar>>,
    #[prost(oneof = "foo::OneofField", tags = "2, 3")]
    pub oneof_field: ::core::option::Option<foo::OneofField>,
}
/// Nested message and enum types in `Foo`.
    pub mod foo {
        pub enum OneofField {
            #[prost(string, tag = "2")]
            Baz(::prost::alloc::string::String),
// This PR added "boxed" here 
            #[prost(message, boxed, tag = "3")]  
            BoxQux(::prost::alloc::boxed::Box<super::Bar>),
        }
	}

[caspermeijn]

I think this test will pass without the other changes.

I'm not sure what you mean by "other changes" here.

When I only apply the changes in tests/, then the tests pass. See for example this commit: caspermeijn@06ee2ae

This means that either your tests are incomplete or that the code changes in code_generator are not necessary.

[caspermeijn]

You could simplify this to:

Suggested change

	cfg_if! {
	if #[cfg(feature = "edition-2015")] {
	use boxed_field::foo::OneofField;
	} else {
	use foo::OneofField;
	}
	}
	use self::foo::OneofField;

[xxchan]

Yes you are right. I found the problem is only for string oneof variants, but not message variant.

::prost::encoding::message::merge accepts M: Message, and we have impl<M> Message for Box<M>. But ::prost::encoding::string::merge only accepts &mut String. 🤔

[xxchan]

But it seems not useful at all to box a string.

[caspermeijn]

But it seems not useful at all to box a string.

I agree, a Box<String> doesn't make sense. Maybe the code generator should prevent a string from being boxed.

Perhaps it should just ignore boxed when the type is string. If it currently doesn't compile, then it is not a breaking change to change the field type.

To improve the tests, you could try to reuse TestAllTypesProto3 and add .boxed(".") to the build.rs. That way, we can verify that all field types actually compile.