use #[naked] for __rust_probestack

[folkertdev]

This should work now. #[naked] takes care of adding the directives that define_rust_probestack added.

I deliberately did not format/indent the assembly code to make review easier. I could do that in a separate commit it that is desirable.

Let's see if CI is OK with this.

[folkertdev]

I'm assuming that exporting this as a C symbol is important, but maybe it's not?

[bjorn3]

Yes, rustc hard codes the name of this function. It should probably be changed to use #[rustc_std_internal_symbol] mangling though, but that would require coordination between rustc and compiler-builtins.

[bjorn3]

Actually using #[no_mangle] would cause __rust_probestack to be exported from cdylibs, so I think it will have to be changed to #[rustc_std_internal_symbol] in this PR and changed on the rustc side when bumping the compiler-builtins dependency. It is currently not exported from cdylibs as rustc isn't aware that the global_asm!() exports __rust_probestack.

[folkertdev]

hmm,

error[E0658]: this is an internal attribute that will never be stable
  --> compiler-builtins/src/probestack.rs:64:1
   |
64 | #[rustc_std_internal_symbol]
   | ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
   |
   = help: add `#![feature(rustc_attrs)]` to the crate attributes to enable
   = note: this compiler was built on 2025-04-21; consider upgrading it if it is out of date

error[E0736]: attribute incompatible with `#[unsafe(naked)]`
  --> compiler-builtins/src/probestack.rs:64:1
   |
63 | #[unsafe(naked)]
   | ---------------- function marked with `#[unsafe(naked)]` here
64 | #[rustc_std_internal_symbol]
   | ^^^^^^^^^^^^^^^^^^^^^^^^^^^^ the `rustc_std_internal_symbol` attribute is incompatible with `#[unsafe(naked)]`

I guess compiler-builtins can just enable #![feature(rustc_attrs)], but that second error needs a fix in rustc first.

[tgross35]

Could you mention in the doc comment that the custom ABI is the reason this is unsafe?

[folkertdev]

Do you mean in the module doc comment? We can add some specific details (about what registers are used etc) if we add a doc comment to the function itself.

[tgross35]

I was referring to the comments above each of the functions, I guess they're not actually doc comments though they could be. Something like (hence the `unsafe`, since the ABI does not actually match `extern "C"`) in the paragraph describing ABI is sufficient, unless you want to add more. Just to have a note that it isn't supposed to be extern "C".

[tgross35]

That reminded me of a few other cases we have where naked functions are extern "C" but that doesn't really make sense. Seems like a somewhat common issue, I wrote up rust-lang/rust#140566 as a possible way to make this more clear (not for this pr of course).

[tgross35]

Thanks for putting this together, this is a nice bit of simplification. After the rustc_std_internal_symbol change I guess it's probably easiest to wait until the beta bump to merge this, since that's in only ~8 days and avoids any bootstrap problems.

I deliberately did not format/indent the assembly code to make review easier. I could do that in a separate commit it that is desirable.

Feel free to do any cleanup you see fit. Split NFC commits are great, GH's "ignore whitespace" option on the review UI works for everything else 🙂

[folkertdev]

Hmm, formatting does not really help much, because formatting in asm is undefined. So let's leave it as it is.

I did merge the definition for sdx into the normal x86_64 version (one day that can use #[cfg()] in the asm block, for now a macro will do). That removes a bunch of duplication.

[bjorn3]

After the rustc_std_internal_symbol change I guess it's probably easiest to wait until the beta bump to merge this, since that's in only ~8 days and avoids any bootstrap problems.

There will need to be a change on the rustc side anyway when updating compiler-builtins. #[rustc_std_internal_symbol] changes the exact symbol name, so the __rust_probestack reference inside cg_llvm needs to be wrapped in a mangle_internal_symbol call to match what compiler-builtins will define when it adds #[rustc_std_internal_symbol].

[folkertdev]

Well currently the bootstrap compiler rejects #[unsafe(naked)], so we're using #[cfg(bootstrap)] to make that work in some other places in compiler-builtins. Similar, bootstrap would still reject #[rustc_std_internal_symbol] in combination with #[naked].

So yeah waiting for the new beta is probably easiest (not sure how rustweek impacts the schedule there, but we'll see)