Check for symbols with default visibility in symbol-check

[zmodem]

to ensure that compiler-builtins doesn't expose any symbols with default visibility. See discussion on rust-lang/rust#151998

[zmodem]

Lots of red on the CI, but that means the patch is working :-)

It's flagging the outline atomics targeted by the Rust-side PR. Also __aeabi_idivmod etc. which I'm assuming are similarly defined as naked_asm functions. Also __rust_probestack.

RSQRT_TAB is a variable though, so I guess that will need another patch to hide.

[tgross35]

Looking through the failures, all the naked functions fail as expected. But it seems like static RSQRT_TAB is also exposed? Any idea what's happening there?

[tgross35]

Oh sorry, you mentioned that in a comment already and I missed it :) Is that something that could also be fixed in rust-lang/rust#151998?

[zmodem]

Yes, I've updated that PR.

[tgross35]

I recently added some tests for symcheck, would you be able to add one to crates/symbol-check/tests/all.rs as well? It should look a bit like

compiler-builtins/crates/symbol-check/tests/all.rs

Lines 65 to 76 in a7b1d8e

	#[test]
	fn test_core_symbols() {
	let t = TestTarget::from_env();
	let dir = tempdir().unwrap();
	let lib_out = dir.path().join("libfoo.rlib");
	t.rustc_build(&input_dir().join("core_symbols.rs"), &lib_out, |cmd| cmd);
	let assert = t.symcheck_exe().arg(&lib_out).assert();
	assert
	.failure()
	.stderr_contains("found 1 undefined symbols from core")
	.stderr_contains("from_utf8");
	}

.

[zmodem]

I recently added some tests for symcheck,

Hmm, those tests don't play well with my visibility check :-)

The main issue is that the test is building things which are not compiler-builtins, such as executables and rust libraries, and those will have plenty of visible symbols.

I can avoid some of it by ignoring executables in my check, but for example the good function in good_lib.rs will be visible, and I can't fix that with -Zdefault-visibility=hidden because that's not available with stable rustc.

(I tried adding #![compiler_builtins] to good_lib.rs, but rustc didn't like that either.)

[rustbot]

This PR was rebased onto a different main commit. Here's a range-diff highlighting what actually changed.

Rebasing is a normal part of keeping PRs up to date, so no action is needed—this note is just to help reviewers.

[zmodem]

I added a --no-visibility flag which the tests that don't care about visibility checking can pass. Does that sound like a reasonable solution?