Fix illumos .init_array section creation

[MarceColl]

Hey there! I've been debugging this issue for the some days and did a write up with more details that I link
to at the end of the PR.

I've tried to keep the disturbance to a minimum but sadly it seems that we need an unstable feature, and thus nightly
for illumos. I don't have a lot of rust experience so I may have done something wrong but I think all other platforms
should still work under stable.

Under illumos the #[used] tag emits an llvm.used flag, this gets converted into a SHF_SUNW_NODISCARD Solaris flag that illumos doesn't understand. The linker then doesn't merge the two .init_array sections since they have different flags.

Because we end up with two .init_array sections and the one pointed by default by DT_INIT_ARRAY is the first ""incorrect"" one, inventory doesn't work under illumos when using the illumos linker (the default).

This patch changes the default behaviour under illumos to not using #[used].

Some references:

• https://reviews.llvm.org/D107955#:~:text=One,them
• https://reviews.llvm.org/D97448
• My own writeup of investigating this issue: https://system-illumination.org/01-rustler.html

[dtolnay]

@jclulow, since you worked on illumos support for ctor in mmastrac/rust-ctor#135, would you be able to say whether this looks like the best solution to you?

[jclulow]

I feel like the best solution is probably to fix LLVM not to emit random flags on platforms that don't understand them, rather than just try to work around it?

[MarceColl]

@jclulow Yeah, that's probably a better idea. I'll keep the fork so I can unblock my own needs. And will look into an LLVM fix in the near future.

[dtolnay]

The rustup-distributed Rust compiler builds use a Rust-specific fork of LLVM (https://github.com/rust-lang/llvm-project). So as soon as you find a fix, it can be backported there and make it into the next nightly with a turnaround time of 1-2 days, instead of waiting months for the fix to end up in an LLVM release and major upgrade of Rust's LLVM dependency.

Since this workaround in inventory would be nightly-only anyway, I think I will not be merging it unless it turns out there is no reasonable LLVM fix somehow, or the backport is rejected. This workaround would only make a difference to someone who is building inventory for illumos while unable to upgrade from an old nightly.

[MarceColl]

Makes a lot of sense @dtolnay, will do that. Thanks for the tip, I didn't know. Feel free to close the PR! I can reopen it if it comes to that.

[bjorn3]

Coming here from your blog post. For reference I'm the one who changed #[used] to be equivalent to #[used(linker)] rather than #[used(compiler)] on ELF platforms in rust-lang/rust#140872 like it already was on other platforms.

Does Illumos have an equivalent to SHF_SUNW_NODISCARD/SHF_GNU_RETAIN?

[Noratrieb]

Let's move this to a Rust issue: rust-lang/rust#146169