resolve: Bind primitive types to items in libcore

[petrochenkov]

This is a more radical alternative to #32131
A new attribute #[primitive_type] is introduced. If path used in type position resolves to a #[primitive_type] item, the definition of a primitive type with the same name is used instead of this item.
A path cannot resolve to a primitive type in any other way. Several different #[primitive_type] items can be resolved to one primitive type.
Most of primitive types are bound to modules (core::u8, core::str etc.) for backward compatibility.
All the primitive type items are added to the prelude.

Pros:
This makes primitive types appear to be less special and more similar to lang items - they are defined in libcore (and also libcollections, librustc_unicode and libstd), shadowed by other items (including modules) and need to be imported before use (but prelude does it for you).

Cons:
It doesn't makes name resolution simpler internally, primitive types are still a hardcoded limited set of Def::PrimTys known to the compiler.
It puts the burden of defining/importing primitive types on #[no_core] and #[no_implicit_prelude] code. The alternative PR presents them as the second small prelude which is hardcoded and has lower priority, so the primitive types are always available.

[breaking-change]
Needs crater before proceeding.

cc #32131 (comment)
cc @jseyfried
r? @eddyb

[eddyb]

Why not #[lang = "..."]?

[petrochenkov]

Lang items have to be unique, no?
str, for example, has to be bound to both core::str and collections::str for backward compatibility. The same is true for char, f32, f64.

[eddyb]

(and also libcollections, librustc_unicode and libstd)

I would still allow the impls to be in different crates (if I'm understanding this right) but move the types in libcore.

[petrochenkov]

The problem is not in impls, but in shadowing.
Neither use core::str; nor use std::str; (they are two different modules) should make primitive str unavailable.
If we mark only core::str with #[primitive_type], then use std::str; will break our code using str.
If we mark only std::str with #[primitive_type], then we won't be able to use str in libcore at all.

[eddyb]

@petrochenkov What's in collections::str that can't be in libcore? I guess this is the other thing aside from impls - modules that should've been replaced the moment UFCS was implemented.
Sadly we couldn't do it for constants at the time so we ended up leaving functions in modules too.

[petrochenkov]

What's in collections::str that can't be in libcore?

Hm, only EncodeUtf16 stabilized in not yet released 1.8.0. There's still chance to move it.
std::f32 and std::f64 contain only reexports from the libcore modules.
rustc_unicode::char contains two stable structs though.

[eddyb]

@petrochenkov Even if stabilized, couldn't move everything but char's two structures? And char is only an issue because it depends on machinery in librustc_unicode, right?

[petrochenkov]

EncodeUtf16 can't be moved to libcore too if stabilized, it depends on librustc_unicode as well.
(If I correctly understood your question)

[eddyb]

@petrochenkov Why do all of these things depend on an unicode library that we don't want to have in libcore for bloat reasons? 😭

[jseyfried]

I'd prefer to check for Def::Err explicitly instead of introducing the method opt_def_id.

[petrochenkov]

Def::Err and Def::SelfTy, but anyway, def kind doesn't matter here, only the presence of def_id.

[jseyfried]

Good point.

[jseyfried]

Reviewed, looks good to me.

[petrochenkov]

(The failure caught by Travis happens only during docs generation for libc, it doesn't prevent running crater, I'll fix it later.)

[DemiMarie]

Could libstd et al pub use the libcore export of the primitive type?

[eddyb]

@drbo I believe the issue is that libcore doesn't have everything about all of those types.

[petrochenkov]

I know there are plans to make a snapshot in the near future, it would be ideal to land this PR before that snapshot.

[alexcrichton]

eh we're still in "make snapshots whenever" mode so it's not like we can't have one for another 6 months :)

I went ahead and kicked one off as @eddyb requested it, but if this lands we can just make a new one.

[bors]

☔️ The latest upstream changes (presumably #32302) made this pull request unmergeable. Please resolve the merge conflicts.

[petrochenkov]

Rebased.

[bors]

☔️ The latest upstream changes (presumably #32054) made this pull request unmergeable. Please resolve the merge conflicts.

[petrochenkov]

Rebased.

[petrochenkov]

Closing as per #32131 (comment)