ci: Set the -znostart-stop-gc flag for lld when testing

[torokati44]

Should fix CI failing with beta Rust (aside from the lints, for which there's #16679 and #16683).

Now that Rust beta also uses rust-lld by default on x86_64-unknown-linux-gnu, we can't use -Z as a workaround.

An alternative to #16448.
Also reverts #16390.
Referencing: dtolnay/linkme#94
For more info, see: https://lld.llvm.org/ELF/start-stop-gc

[torokati44]

This conservative behavior works for existing code which does not take GC into fair consideration, but unnecessarily increases sizes for modern metadata section usage which desires precise GC.

GNU ld 2.37 added -z start-stop-gc to restore the traditional behavior ld.lld 13.0.0 defaults to -z start-stop-gc and supports -z nostart-stop-gc to switch to the conservative behavior.

However, since this is only an issue for testing (we use linkme to collect the test cases IIRC correction: known stubs), binary size shouldn't be much of an issue.

[torokati44]

We may decide to wait for rustc to roll this change (switching to lld on linux by default) back from beta (dtolnay/linkme#88 (comment)) instead of merging this.
Or for linkme to fix this some other way with the change in rustc.

[danielhjacobs]

See rust-lang/rust#126282 and rust-lang/rust#126278

[lqd]

When using lld, this is likely to me to be the expected fix for linkme users, as I wrote in dtolnay/linkme#88 (comment) and the blog post announcing the change.

[torokati44]

I see, thanks!

For the record, while looking into it, I had some other alternative ideas too, but I have no idea whether any of them are actually feasible. Such as:

• Using the KEEP keyword in a linker script - linkme already has one for cortex-m, but I'm not sure it would be great to have one for x86_64 linux as well...
• Adding a retain attribute to rustc, similar to used, which would affect the linker symbol...
• Having linkme manually annotating the start/stop symbols to not be discarded...

[bjorn3]

Adding a retain attribute to rustc, similar to used, which would affect the linker symbol...

We have #[used(linker)] already. It is just unstable.

[torokati44]

Considering that the next beta is coming out almost 6 weeks from now, I'm proposing merging this now, just to clean up CI, and unblock merging PRs without repo admin force.
As seen above, this will likely not be a bad fix long term either.

[lqd]

Considering that the next beta is coming out almost 6 weeks from now

AFAICT the new beta is already out and lld is disabled there

[torokati44]

Oh, right, great! Anyway, I think leaving that option won't really hurt...