skip_serializing does not inhibit generic bounds

[arcnmx]

The following will fail because the derive will automatically add a T: Serialize/Deserialize bound despite the attributes. It would be nice if the codegen were smart enough to figure out the bound isn't necessary.

Also note that default is not quite correct in this test case, as it doesn't imply that the type should never be deserialized. A skip_deserializing attribute should also exist (see also #183).

Also related: the : Default bound should not be necessary, and automatically added to the derived impl Deserialize in the same manner.

#![feature(plugin, custom_derive, custom_attribute)]
#![plugin(serde_macros)]

extern crate serde;

#[derive(Serialize, Deserialize)]
struct A<T: Default> {
    #[serde(skip_serializing, default)]
    t: T,
}

#[derive(Default)]
struct B;

fn main() {
    fn assert<T: serde::Serialize + serde::Deserialize>(_: T) { }

    assert(A { t: B })
}

[erickt]

@arcnmx: #[serde(skip_deserializing)] doesn't make sense since it'd be impossible to create the structure. It would be neat if we could skip adding bounds if we know need them.

[arcnmx]

@erickt the difference between default and skip_deserializing is that default still implies that type will be deserialized if possible. skip_deserializing would mean to never even try, just unconditionally initialize it from a given or default value and don't look for the key at all, etc.

[arcnmx]

(and would therefore mean a T: Deserialize bound is unneeded and wrong etc.)

[erickt]

Sure, that makes sense. Bumping this off the 0.7.0 release since it could be implemented backwards compatibly.

[dtolnay]

Using the changes in #285, the code above now compiles if you replace #[serde(skip_serializing, default)] with #[serde(skip_serializing, skip_deserializing)] (which is I think how @arcnmx intended for it to work).

[arcnmx]

@dtolnay awesome work, looks like that will indeed kill this issue off for good! I'd add that perhaps it's worth discussing whether #[serde(skip_deserializing, default)] should be required for consistency(ish) reasons, but I don't have a personal opinion on that anyway.

[dtolnay]

Thanks for bringing that up. My opinion is no - skip_deserializing should imply default unless you specify a different default with default=....