`#[used(linker)]` is broken on illumos

[Noratrieb]

On illumos, #[used(linker)] generates the section flag 0x100000, which is known as SHF_SUNW_NODISCARD on Solaris, but unknown to illumos. This causes incorrect behavior when used in combination with .init_array, for example by inventory.

This regressed in #140872 which made #[used] mean #[used(linker)] on all ELF targets, including illumos.

See

• Original investigation by @MarceColl on this issue: https://system-illumination.org/01-rustler.html
  • discussed on lobste.rs: https://lobste.rs/s/nr8kkd/debugging_rustler_on_illumos
  • Attempt to work-around in the inventory crate: Fix illumos .init_array section creation dtolnay/inventory#84
• LLVM change to start using SHF_SUNW_NODISCARD on Solaris: https://reviews.llvm.org/D107955#:~:text=One,them
• LLVM change related to retained section behavior: https://reviews.llvm.org/D97448

[Noratrieb]

We tell LLVM that our illumos is Solaris, so it is not surprising that LLVM applies Solaris-specific attributes:

rust/compiler/rustc_target/src/spec/targets/x86_64_unknown_illumos.rs

Line 14 in fd75a9c

llvm_target: "x86_64-pc-solaris".into(),

@jclulow @pfmooney what do you think about this? Ideally this would be fixed in LLVM by teaching it about illumos and making it not apply SHF_SUNW_NODISCARD for it (and possibly adding similar functionality to illumos so #[used(linker)] can be supported properly :D). As a work-around, we could make #[used] mean #[used(compiler)] on illumos again I suppose.

[jclulow]

Yeah I think we need to go back to #[used(compiler)] for now. We can probably look at adding support for the flag in our linker in the future, but even if we had it today it won't help on existing (e.g., LTS) systems for quite some time. It has been suggested to me that if you're not using -z ignore, our linker doesn't do the same level of proactive garbage collection of unreferenced sections that the Oracle Solaris linker will do.

As far as getting a first class illumos triple in LLVM itself, I agree that this is a positive long term direction -- but obviously is a huge amount of work, for the community at large (for whom clang, etc, would start working differently) as well as just in the LLVM contribution process, so I wouldn't want to put that on any critical path.

[bjorn3]

We do use -z ignore...

rust/compiler/rustc_codegen_ssa/src/back/linker.rs

Line 914 in fd75a9c

self.link_args(&["-z", "ignore"]);

rust/compiler/rustc_codegen_ssa/src/back/link.rs

Lines 2289 to 2293 in fd75a9c

	// On Solaris-like systems, `-z ignore` acts as both `--as-needed` and `--gc-sections`
	// so it will ignore unreferenced ELF sections from relocatable objects.
	// For that reason, we put this flag after metadata objects as they would otherwise be removed.
	// FIXME: Support more fine-grained dead code removal on Solaris/illumos
	// and move this option back to the top.

[Noratrieb]

Using #[used(compiler)] again would mean essentially reverting #140872 for illumos only, which should be a simple change.

[bjorn3]

Not quite. With the attribute representation changes, #[used] already gets interpreted as #[used(linker)] when parsing the attribute. You did have to add a third variant to the UsedBy enum and use it at

rust/compiler/rustc_attr_parsing/src/attributes/codegen_attrs.rs

Line 317 in fd75a9c

ArgParser::NoArgs => UsedBy::Linker,

and then pick the appropriate flag to set for the target at

rust/compiler/rustc_codegen_ssa/src/codegen_attrs.rs

Lines 263 to 266 in fd75a9c

	AttributeKind::Used { used_by, .. } => match used_by {
	UsedBy::Compiler => codegen_fn_attrs.flags |= CodegenFnAttrFlags::USED_COMPILER,
	UsedBy::Linker => codegen_fn_attrs.flags |= CodegenFnAttrFlags::USED_LINKER,
	},

when this variant is used.

[MarceColl]

👋 thanks for looking into this. Maybe I should have tried contacting and opening an issue here instead of trying to work around the issue downstream.

Let me know if you need me to run any tests or help debug on the Illumos side. Even though @jclulow would probably be better for this, but if I can take some load off his shoulders I'm happy to do so.