Test case uses locale-dependent output

[crlf0710]

I was running ui-tests, and this test case fail on my windows machine:

---- [ui] tests\ui\unpretty\staged-api-invalid-path-108697.rs stdout ----
$DIR\lol
$DIR\staged-api-invalid-path-108697.rs
Saved the actual stderr to "L:\\ThirdParty\\rust\\build\\x86_64-pc-windows-msvc\\test\\ui\\unpretty\\staged-api-invalid-path-108697\\staged-api-invalid-path-108697.stderr"
diff of stderr:

-       error: couldn't read `$DIR/lol`: No such file or directory (os error 2)
+       error: couldn't read `$DIR/lol`: 系统找不到指定的文件。 (os error 2)
2         --> $DIR/staged-api-invalid-path-108697.rs:8:1
3          |
4       LL | mod foo;

Maybe we should not rely on such locale-dependent output, or maybe we could configure Rust standard library/compiler to use English as UI language.

[ChrisDenton]

Maybe we should not rely on such locale-dependent output

If that's the only test that fails then I think we don't normally rely on locale? It may just be this test needs tweaking.

EDIT: wait that's not a linker erorr, huh. Ignore this next bit.

linker stuff

or maybe we could configure Rust standard library/compiler to use English as UI language.

We do configure Rust to use English for the linker

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

Lines 30 to 41 in c3fe9e7

	/// Disables non-English messages from localized linkers.
	/// Such messages may cause issues with text encoding on Windows (#35785)
	/// and prevent inspection of linker output in case of errors, which we occasionally do.
	/// This should be acceptable because other messages from rustc are in English anyway,
	/// and may also be desirable to improve searchability of the linker diagnostics.
	pub(crate) fn disable_localization(linker: &mut Command) {
	// No harm in setting both env vars simultaneously.
	// Unix-style linkers.
	linker.env("LC_ALL", "C");
	// MSVC's `link.exe`.
	linker.env("VSLANG", "1033");
	}

However it relies on the English language pack being installed via the Visual Studio Installer.

[crlf0710]

Currently the std::io::Error's debug output relies on winapi FormatMessage behavior, which relies on thread language (locale).

unsafe {
    SetThreadUILanguage(MAKELANGID(LANG_ENGLISH, SUBLANG_ENGLISH_US));
}

If above code is called from rustc, i think the output will be deterministic.

[ChrisDenton]

I think that would need to be an option exposed by libtest. It manages running test threads.

[Noratrieb]

This error comes from rustc, not a libtest thread. rustc itself would need to set its language to english for this

[ChrisDenton]

In which case we could set it in rustc main (early, before anything can error) but I think that would need a Major Change Proposal (MCP).

[jieyouxu]

If that's the only test that fails then I think we don't normally rely on locale? It may just be this test needs tweaking.

This is a general problem where the test infra doesn't try to account for locale (because rustc will forward system error messages in different locale IIRC).

I don't think the test infra can really properly account for this (e.g. snapshot tests) without setting the locale for rustc somehow.