Two different versions of a crate interacting leads to unhelpful error messages

[huonw]

https://github.com/huonw/bad-error-messages-a contains a crate a. https://github.com/huonw/bad-error-messages-bc contains two crates, b and c.

• a contains a trait Foo and a function test that requires that trait.
• b depends on a specific revision (8b532f5, not the HEAD) of a and contains a type Bar that implements a::Foo.
• c depends on a's HEAD (happens to be 84cfe230) and b, and tries to use b::Bar with a::test.

c fails to compile despite Bar "clearly" implementing Foo... it's even displayed in the docs!

Cloning https://github.com/huonw/bad-error-messages-bc and running cargo build in the root (which is c) gives

   Compiling a v0.0.1 (https://github.com/huonw/bad-error-messages-a#84cfe230)
   Compiling a v0.0.1 (https://github.com/huonw/bad-error-messages-a?rev=8b532f5#8b532f51)
   Compiling b v0.0.1 (file:///home/huon/projects/test-rust/error-messages/c)
   Compiling c v0.0.1 (file:///home/huon/projects/test-rust/error-messages/c)
src/lib.rs:5:5: 5:12 error: the trait `a::Foo` is not implemented for the type `b::Bar` [E0277]
src/lib.rs:5     a::test(b::Bar);
                 ^~~~~~~
error: aborting due to previous error

There's no indication of the fundamental problem from either rustdoc or rustc: that there's two different versions of a being used, meaning a#84cfe230::Foo and a#8b532f5::Foo are different. Bar only implements the latter, but in c, the name a refers to a#84cfe230 so a::test(...) requires the former. (Using name#rev to represent the crate with that version.)

There is an indication of the difference in the example above, since a is compiled twice, but that comes from cargo, not rustc. This issue is focusing on the fact that rustc itself does not give helpful errors and/or allow tools like cargo to control those errors.

It would be nice if rustc:

• disambiguated ambiguous names, probably by filepath/name by default
• allowed cargo to specify semantically relevant info for the disambiguation e.g. pass in the version/revision/..., so that rustc can print foo#0.2.3 instead of the filepath/name
• detected errors that may be caused by ambiguous names and noted it explicitly, e.g. when there is a error involving an ambiguous crate print (once) some extra lines like "note: there's multiple versions of crate ... in use": along with a listing of the possibilities.

(NB. I've used cargo to construct this example because it is the easiest way I can get two different versions of a crate. This appears in rust-lang/rust's makefiles too, where the std test runner is called "std" internally, but links against another crate called "std", i.e. the conventional standard library.)

[Hoverbear]

This bit me today with uuid! It took me a few minutes before I even understood that it wasn't "my" code that was giving this off.

[Kimundi]

Related, its also annoying if only crate local paths get emitted, like in backtraces.

We need a way for crates to identify themself uniquely, and not just by an identifier. Maybe a scheme like <crate foo #ef12f212>::bar::baz?

[golddranks]

I was bitten by this too.

Kimundi: Agreed for the need to unique identification. But then again, in some cases the local path tells more directly where the problem is. I wonder if there are cases where path is better and cases where a hash is better. Or maybe some kind of hybrid: If version information from cargo is available, use that. If not, if version control information is available, use commit SHA, and if no other information exists, use a path...?

[huonw]

Nominating for high priority. I think this is one of the main pain points people have when using cargo, and one of the major reasons for people complaining about cargo allowing multiple versions of a crate. It's usually easy to identify once you've seen it a few times, but it is a horribly confusing and annoying experience.

(Tagged with both compiler and tools since cargo presumably plays into this, but this is probably most relevant to the compiler subteam.)

@golddranks: the question of connecting semantically relevant info to crates is discussed a little in the issue. I think the best plan would external tools to pass in arbitrary strings connected to each crate, defaulting to the path, meaning rustc doesn't have to actually understand anything, just print strings (the compiler already has an --extern NAME=PATH argument that tools like cargo use, it could be extended to also allow --extern NAME#"ID"=PATH or something).

[steveklabnik]

I've also seen an increasing number of IRC questions about this.

On Sep 7, 2015, at 20:58, Huon Wilson notifications@github.com wrote:

Nominating for high priority. I think this is one of the main pain points people have when using cargo, and one of the major reasons for people complaining about cargo allowing multiple versions of a crate. It's usually easy to identify once you've seen it a few times, but it is a horribly confusing and annoying experience.

(Tagged with both compiler and tools since cargo presumably plays into this, but this is probably most relevant to the compiler subteam.)

@golddranks: the question of connecting semantically relevant info to crates is discussed a little in the issue. I think the best plan would external tools to pass in arbitrary strings connected to each crate, defaulting to the path, meaning rustc doesn't have to actually understand anything, just print strings (the compiler already has an --extern NAME=PATH argument that tools like cargo use, it could be extended to also allow --extern NAME#"ID"=PATH or something).

―
Reply to this email directly or view it on GitHub.

[Manishearth]

I'll take a crack at this

[Manishearth]

I think at the very least, for mismatched type and unimplemented trait errors, if we detect a crate mismatch, we should mention it. Any other such errors?