run-make-fulldeps/linker-output-non-utf8 failure

[nikomatsakis]

when running x.py test locally I get errors in the test run-make-fulldeps/linker-output-non-utf8:

---- [run-make] run-make-fulldeps/linker-output-non-utf8 stdout ----

error: make failed
status: exit code: 2
command: "make"
stdout:
------------------------------------------
LD_LIBRARY_PATH="/home/nmatsakis/versioned/rust-6/build/x86_64-unknown-linux-gnu/test/run-make-fulldeps/linker-output-non-utf8/linker-output-non-utf8:/home/nmatsakis/versioned/rust-6/build/x86_64-unknown-linux-gnu/stage2/lib:/home/nmatsakis/versioned/rust-6/build/x86_64\
-unknown-linux-gnu/stage0-bootstrap-tools/x86_64-unknown-linux-gnu/release/deps:/home/nmatsakis/versioned/rust-6/build/x86_64-unknown-linux-gnu/stage0/lib" '/home/nmatsakis/versioned/rust-6/build/x86_64-unknown-linux-gnu/stage2/bin/rustc' --out-dir /home/nmatsakis/versi\
oned/rust-6/build/x86_64-unknown-linux-gnu/test/run-make-fulldeps/linker-output-non-utf8/linker-output-non-utf8 -L /home/nmatsakis/versioned/rust-6/build/x86_64-unknown-linux-gnu/test/run-make-fulldeps/linker-output-non-utf8/linker-output-non-utf8  library.rs
mkdir /home/nmatsakis/versioned/rust-6/build/x86_64-unknown-linux-gnu/test/run-make-fulldeps/linker-output-non-utf8/linker-output-non-utf8/zzz$'\xff'
mv /home/nmatsakis/versioned/rust-6/build/x86_64-unknown-linux-gnu/test/run-make-fulldeps/linker-output-non-utf8/linker-output-non-utf8/liblibrary.a /home/nmatsakis/versioned/rust-6/build/x86_64-unknown-linux-gnu/test/run-make-fulldeps/linker-output-non-utf8/linker-outp\
ut-non-utf8/zzz$'\xff'
LD_LIBRARY_PATH="/home/nmatsakis/versioned/rust-6/build/x86_64-unknown-linux-gnu/test/run-make-fulldeps/linker-output-non-utf8/linker-output-non-utf8:/home/nmatsakis/versioned/rust-6/build/x86_64-unknown-linux-gnu/stage2/lib:/home/nmatsakis/versioned/rust-6/build/x86_64\
-unknown-linux-gnu/stage0-bootstrap-tools/x86_64-unknown-linux-gnu/release/deps:/home/nmatsakis/versioned/rust-6/build/x86_64-unknown-linux-gnu/stage0/lib" '/home/nmatsakis/versioned/rust-6/build/x86_64-unknown-linux-gnu/stage2/bin/rustc' --out-dir /home/nmatsakis/versi\
oned/rust-6/build/x86_64-unknown-linux-gnu/test/run-make-fulldeps/linker-output-non-utf8/linker-output-non-utf8 -L /home/nmatsakis/versioned/rust-6/build/x86_64-unknown-linux-gnu/test/run-make-fulldeps/linker-output-non-utf8/linker-output-non-utf8  -L /home/nmatsakis/ve\
rsioned/rust-6/build/x86_64-unknown-linux-gnu/test/run-make-fulldeps/linker-output-non-utf8/linker-output-non-utf8/zzz$'\xff' exec.rs 2>&1 | "/home/nmatsakis/versioned/rust-6/src/etc/cat-and-grep.sh" this_symbol_not_defined
[[[ begin stdout ]]]
^[[90merror: Argument 6 is not valid Unicode: "/home/nmatsakis/versioned/rust-6/build/x86_64-unknown-linux-gnu/test/run-make-fulldeps/linker-output-non-utf8/linker-output-non-utf8/zzz\xFF"

^[[0m
[[[ end stdout ]]]
^[[1;31mError: cannot match: this_symbol_not_defined^[[0m

------------------------------------------
stderr:
------------------------------------------
make: *** [Makefile:23: all] Error 1

------------------------------------------

The most notable part being this:

error: Argument 6 is not valid Unicode: "/home/nmatsakis/versioned/rust-6/build/x86_64-unknown-linux-gnu/test/run-make-fulldeps/linker-output-non-utf8/linker-output-non-utf8/zzz\xFF"

perhaps related to the linker on my system?

[nikomatsakis]

Compiler team pre-triage: marking as P-medium, this is probably something wacky in my local setup. We'll see if others complain about it.

[cuviper]

I have also seen this in Fedora and RHEL builds, but I didn't worry enough to chase it down yet.

This error is actually from rustc itself:

rust/src/librustc_driver/lib.rs

Lines 1189 to 1194 in 1cdcea9

	let args = env::args_os().enumerate()
	.map(|(i, arg)| arg.into_string().unwrap_or_else(|arg| {
	early_error(ErrorOutputType::default(),
	&format!("Argument {} is not valid Unicode: {:?}", i, arg))
	}))
	.collect::<Vec<_>>();

Arguments 5 and 6 from the rustc command in your report are:

-L /home/nmatsakis/versioned/rust-6/build/x86_64-unknown-linux-gnu/test/run-make-fulldeps/linker-output-non-utf8/linker-output-non-utf8/zzz$'\xff'

That line is deliberately added in the linker-output-non-utf8 test:

rust/src/test/run-make-fulldeps/linker-output-non-utf8/Makefile

Lines 17 to 23 in 1cdcea9

	bad_dir := $(TMPDIR)/zzz$$'\xff'
	all:
	$(RUSTC) library.rs
	mkdir $(bad_dir)
	mv $(TMPDIR)/liblibrary.a $(bad_dir)
	$(RUSTC) -L $(bad_dir) exec.rs 2>&1 | $(CGREP) this_symbol_not_defined

I have no idea why this isn't an issue in CI though...

[cuviper]

I have no idea why this isn't an issue in CI though...

Oh, I think I do know why. That bad_dir in the Makefile is assuming shell expansion of that $$'\xff' (with doubled $$ to escape from make itself). Fedora and RHEL use sh -> bash which does expand it, but Debian and Ubuntu (used in CI) have sh -> dash, which is a more limited shell.

bash:

$ echo -n $'\xff' | xxd
00000000: ff                                       .

dash:

$ echo -n $'\xff' | xxd
00000000: 245c 7866 66                             $\xff

So I think this test hasn't really been working as desired on dash-using platforms at all, since the bad_dir isn't really "bad" when not escaped. And it can't work even then because rustc won't accept non-utf8 arguments in the first place, never mind reaching the linker.

[cuviper]

Note, the dist-i686-linux and dist-x86_64-linux builders do use CentOS 5 as the base image, which would have this issue too, but we don't actually run tests there. It would be nice to add some diversity to the testing images, either with Fedora or perhaps the redistributable RHEL8 UBI.

[mati865]

Any idea what to do with this test?
Having to use --exclude src/test/run-make-fulldeps for over two weeks isn't very practical.

[Mark-Simulacrum]

We can probably just remove that test, it doesn't seem worth the pain of having it if it's causing people problems. It sounds like we don't have a test case that would work universally even on *nix platforms, right? Or that's my interpretation of #63520 (comment) at least.

It's also pretty unclear to me that ICE-ing on non-utf8 output from linkers is something we need to test for. That sounds both incredibly uncommon, and also pretty easily fixable if we need to.

One thought for testing this is to override the linker rustc uses with a shell script or even a rust program that runs the linker internally but just does something like println or eprintln with some non-utf8 output.

[nikomatsakis]

This test continues to be super annoying. I'm happy to just remove it, per @Mark-Simulacrum's suggestion.

[Mark-Simulacrum]

Okay I've filed #65321 to remove the test -- seems like everyone is in agreement it does not pull its weight.