Fix sysroot handling

[RalfJung]

No description provided.

[RalfJung]

Turns out I screwed up in my last sysroot PR, I hope this PR (updated it) now fixes that.

I also changed it so that the panic from compile_time_sysroot surfaces even if MIRI_SYSROOT is set; this means we actually test that that panic dos not happen as part of the test suite.

[RalfJung]

CI looks good. Still, this needs a new review I think.

[oli-obk]

lgtm, but I'm confused how this happened. Is it because I created the sysroot hacks in our test suite and thus changes in the miri binary were ignored by our tests since the tests passed their own --sysroot?

[RalfJung]

What happened is that the test suite in rustc always passes a --sysroot, yes. Same for our test suite here though, and same for cargo miri. So I think it has been the case since forever that the fallback code (with no MIRI_SYSROOT/--sysroot) panicked, it's just that we never ran into that code.

But now I added this sanity check in cargo miri that does exactly that.

[RalfJung]

Oh also we don't test cargo miri at all on the rustc side. Would be great if we could somehow make that happen -- and maybe also recompile libstd there instead of relying on setting --cfg miri in bootstrap, which I feel uneasy about.

[oli-obk]

maybe also recompile libstd there instead of relying on setting --cfg miri in bootstrap

You mean via cargo miri setup or just straight up adding an additional libstd-miri thing in bootstrap that Miri depends on?

[RalfJung]

Well ideally the same way users do it, so via cargo miri setup.