Introduce something about rmeta and metadata in general

[crlf0710]

[mark-i-m]

@crlf0710 I don't actually know anything about rmeta or metadata. Who even is the expert? Is there at least a code snippet you can point us to?

[crlf0710]

@mark-i-m Sure, there's the crate src/librustc_metadata in tree. For the experts part, i guess @eddyb @petrochenkov can say a lot on this topic.

[eddyb]

Btw, this is one of those areas I'd prefer refactoring into irrelevance - see rust-lang/rust#65407.

[crlf0710]

@eddyb i believe metadata is the key to rust-lang/rust#65427, do you think this issue is worth resolving before the refactoring you mentioned, or are they orthogonal?

[eddyb]

I think they're unrelated, talk to Cargo people (e.g. @alexcrichton and @ehuss) about how pipelining works for cargo build and how it relies on .rmetas for almost the entire build (except bin/staticlib/cdylib crates, which need linking and therefore .rlibs).

But also, even without reusing the .rmetas, you could make Cargo use the same incremental directory between cargo check and cargo build and that "should just work".

EDIT: oh, I found rust-lang/cargo#3501 (comment) - is the "metadata is different" part of that what you meant? It's technically true, but not for metadata reasons, but "does this crate need codegen and therefore information about codegen'd monomorphizations?" reasons.

[crlf0710]

@eddyb ok thanks!

Is it possible when you have time to give some brief introduction (in text, audio or whatever form) about the current status (or if the current status is not ideal, a sketch of new design) of crate metadata storage on disk and the metadata embedded into final rlib/dylib files. I'm eager to learn more about this, and it will be good include it as part of rustc-guide.

[ehuss]

I went ahead and wrote something up in #1044 on this. I realize things change over time, but I think a little information is useful. If anyone following this issue can give it a read, please let me know if anything is missing you'd like to see (or if you notice anything wrong).