libtest output is not colored

[RalfJung]

cargo miri test output is not colored the same way cargo test is: the output that is generated inside Miri (the libtest runner) uses no colors. I think this is probably because of the hack where we remove the TERM environment variable even with isolation disabled -- that makes libtest think that the terminal does not support color so it falls back to no colors. (EDIT: looks like I am wrong, see below.)

We remove the TERM variable because having it makes libtest startup super slow (that was #1702). Maybe we need another work-around for this problem, like adding a cfg(miri) somewhere to skip the slow code -- or actually making Miri less slow on that code.^^

@saethlin this is the bug you mentioned in one of your PRs; it has nothing to do with --color=always (other than, even with --color=always, libtest still uses no color unless TERM indicates that the terminal supports color).

[RalfJung]

With #2293 we can use MIRIFLAGS="-Zmiri-disable-isolation -Zmiri-env-forward=TERM" cargo miri test to test what happens when TERM is properly forwarded... and strangely there are still no colors. Hm...

[RalfJung]

Ah, got it -- in addition to forwarding TERM, we also need to adjust our isatty shim, which currently just says nothing is a TTY.

I am not sure what the best solution is here, given the multi-second wait triggered by forwarding TERM.

[saethlin]

Is it sufficient to just lie about whether we are a TTY? I was honestly thinking of hacking up something like that for myself to induce colors to appear.

[RalfJung]

As I said, we need both TERM and the isatty change.

[saethlin]

Argh I was reading your phrasing in the most desperately optimistic way possible.

terminfo files are a mess, and I really wonder how much portability this is actually providing for libtest. Are people actually running cargo test or cargo miri test on a terminal that only supports 8-bit colors? That would be shocking. I wonder if whoever is in charge of this code would be okay with us just bypassing the parsing with a #[cfg(miri)] and selecting some terminfo that we think is reasonable everywhere that Cargo/Miri can run.

[RalfJung]

Yeah they might be. However, is always forcing isatty even what we want? Playground was not happy last time we did that...

[saethlin]

I was thinking of a flag so the user could pick, but it would be better to have the shim work correctly.

[RalfJung]

If the host is also Unix (or maybe only for Linux), we could forward to libc. That will help at least in some cases?

[saethlin]

Everyone goes through this crate https://crates.io/crates/atty

[RalfJung]

Hm, not sure how I feel about adding external dependencies. We are quite conservative about that here. OTOH that sounds like exactly the API we want and it is even cross-platform...

But anyway it seems like there are two orthogonal steps here:

• make whatever code in libtest parses the TERM stuff just skip all that under cfg(miri) and assume a reasonably modern terminal, but still honor isatty
• do something more clever with Miri's isatty shim

[RalfJung]

Everyone goes through this crate https://crates.io/crates/atty

Alternatively we could copy this file which is tiny.