Add a compiler option for a 'pseudo-cwd' for rustc errors

[alexcrichton]

Cargo's behavior of what the cwd is when rustc is invoked changed in rust-lang/cargo#4788 for a number of reasons, but this has caused a number of regressions in tooling because the paths printed in error messages are now relative to a different location.

In talking with @nikomatsakis an idea that came up was that we could add an option to the compiler to say "print diagnostic paths relative to this location". Cargo could pass the current cwd of the cargo process itself when invoking rustc and the compiler would otherwise know to not encode this value into the output artifact at all.

cc @estebank, you may be interested in this as well!

[nikomatsakis]

cc @rust-lang/compiler

[estebank]

This should be useful for any tooling that deals with the compiler output, even beyond cargo.

[eddyb]

Should this be tied/related to the debuginfo path renaming options? (cc @michaelwoerister)

[alexcrichton]

In this context @eddyb I don't think so. That was actually the bugfix in Cargo because Cargo was incorrectly caching artifacts when it otherwise would have changed because the debuginfo would have been different. Instead Cargo now produces artifacts with deterministic debuginfo (regardless of the cwd of who invoked Cargo to build the crate), but for tooling it seems like it may be best to continue to produce diagnostics relative to the cwd

[nikomatsakis]

Question is: who has time to hack this up =)

[ehuss]

Would/should this option also affect the file! macro?

I parse the output from tests to help the user jump to any failed tests, so it needs to resolve those relative paths as well.

[alexcrichton]

@ehuss since that affects the actual output binary itself i'd imagine that this flag wouldn't affect that (as then Cargo would go back to recompiling whenever cwd changed!)

[oli-obk]

How about making this an env var, so tooling could use this without caring how to pass down the path through arbitrary layers of scripts an tools?

This way no part of the compiler has to know about it except for the diagnostic emitters

[nikomatsakis]

I agree with @alexcrichton that this should purely affect compiler output.

I don't know how I feel about using an environment variable. I guess it might be ok; in general, though, environment variables always feel sort of like "action at a distance" to me.

[nikomatsakis]

triage: P-high

If we're going to do something here, it should be soon!

What is the status in terms of other tools (e.g., IDEs) getting mitigations?