rustbuilds rebuilds libstd even when it did not change [recent regression]

[RalfJung]

For just a few weeks now, I am seeing many cases where rustbuild rebuilds stage 0 libstd (and hence rebuilds everything) even though nothing changed there. I have not yet tried to reproduce this in a clean dir (that would take forever due to all the rebuilding), but the steps are roughly as follows:

• Start with a clean dir
• ./x.py --stage 0 build src/librustc
• ./x.py --stage 0 build src/librustc

The last step will rebuild everything even though nothing changed.

The timing coincides a little with libstd depending on crates io, but that could be mere coincidence.

Cc @alexcrichton

[RalfJung]

Hm, I tried in a clean dir and cannot reproduce. I could have sworn I saw this multiple times. I'll reopen if I can reproduce later.

[petrochenkov]

I see this too, it doesn't reproduce with clean build.
To reproduce you need to

• Run ./x.py test --stage 1 src/test/ui --test-args something
• Make some changes to libcore or liballoc (?) (not necessarily yourself, but for example by rebasing on master).
• Run ./x.py test --stage 1 src/test/ui --test-args something again.
• Run ./x.py test --stage 1 src/test/ui --test-args something again, which should be no-op, but it will actually start building almost everything starting with liballoc.

[petrochenkov]

I have confirmed that it's indeed can be reproduced by touching libcore.

[ehuss]

This has been bugging me, too. I bisected the regression to #56092. cc @alexcrichton. I don't fully understand why, but for some reason Cargo thinks that one of liballoc's dependencies has changed, but I don't see any of them being different. I'll keep poking to better understand why the fingerprint check is failing.

The following is a relatively quick repro:
touch src/libcore/lib.rs ; ./x.py build src/libstd --stage=0 ; ./x.py build src/libstd --stage=0
The second build should be fresh, but it isn't.

[ehuss]

I have posted a proposed fix for this.