Running `./x test mir-opt` rebuilds std every time

[Zalathar]

Using profile = compiler, running ./x test mir-opt causes bootstrap to rebuild all of std before running the tests, even when the build cache ought to be warm from a previous ./x test mir-opt.

Needless to say, this is not very fun when actually trying to run one or more mir-opt tests.

[Zalathar]

cc @jieyouxu

[Zalathar]

I will note that --keep-stage-std=1 does seem to avoid the rebuild when re-running the tests.

[ChrisDenton]

Is there some way for CI to error on spurious rebuilds? I don't know what that would look like but this has happened in a few different places for different reasons so it'd be nice if there were some way to catch it early.

[jieyouxu]

Usually the spurious rebuilds happen if you try to run the same command, twice. I don't think that's very tractable for CI, but am happy to be proven wrong. Maybe we can try to devise some kind of mechanism in bootstrap to make some invalidations impossible, but I don't know how such a mechanism would work in general.

[onur-ozkan]

One mechanism for that would be to keep the rustflags logic in a single sealed module, so they can't be set externally from the individual build steps, which means they can't invalidate the build cache. Next to that approach we can add test coverage to rustflags logic.

[jieyouxu]

I opened a short-term immediate fix in #131442, but the long term solution to prevent resurrgence of RUSTFLAGS-induced invalidation is a proper prevention mechanism as described above.

[jieyouxu]

Also this one actually builds same stage std twice for host if host == target (i.e. not cross-compile) because std was ensure'd twice with differing RUSTFLAGS. The double std build was the dead giveaway that fortunately makes it very easy to tell what the rebuild was caused by.