Dependence on `fmt::Debug`

[kornelski]

I'm proposing an option to disable fmt::Debug, mainly to reduce executable sizes, and to thoroughly strip symbol names from executables.

When testing it, it quickly became apparent that thiserror uses fmt::Debug, and this is a common pattern in proc macros.

impl ToTokens for Trait {
    fn to_tokens(&self, tokens: &mut TokenStream) {
        let trait_name = format_ident!("{}", format!("{:?}", self));
        tokens.extend(quote!(::core::fmt::#trait_name));
    }
}

I wanted to check how you feel about an option "breaking" fmt::Debug. Shall fmt::Debug be guaranteed to work in proc-macros? Would you be okay changing the implementation to use something else to stringify enums? (strum's Display, or manual match … => "…"). I'm also proposing to add #[cfg(debug_fmt_detail = "full")] to let crates use fmt::Debug when it's on, and fall back to something else when it's a no-op.

[dtolnay]

As an opt-in compiler flag, I think it's fine that it causes some proc macros to not work. I don't think special consideration needs to be built into the option to behave differently in proc macros.

That said, motivation number 1 in rust-lang/rust#123940 (comment) does not apply to proc macros. If motivation 1 is most of the sincere motivation for this option, and motivation 2 is much less valuable to anyone in practice, then it might be nice if you could make {:?} behave normally in proc macros.

That sounds hard though, at least when Cargo's "cross compile mode" is not in effect (i.e. when --target is not passed). For crates that are a dependency of both macro and non-macro code, it's not necessarily worth making them get compiled twice just to have different {:?} behavior. (As I understand it, in "cross compile mode" they're built twice already anyway, even if host and target are identical targets.) So if you do add a carve-out for proc macros but it only applies when --target is used, I don't see that being valuable because it won't substantially reduce the need for proc macros to adjust their code accommodating this new option.

[kornelski]

Thanks for the response and the code update.