Spurious (?) error in "compile-fail\rfc-2126-extern-in-paths\single-segment.rs" etc

[kennytm]

First introduced by #47828 (comment).

Symptom: The following 3 test cases involving error E0432 will fail.

failures:
    [compile-fail] compile-fail\rfc-2126-extern-in-paths\single-segment.rs
    [compile-fail] compile-fail\use-keyword.rs
    [compile-fail] compile-fail\use-mod-2.rs

Mainly affects Windows and macOS machines (maybe because they are tested first).

Current instances:

PR	Image
#47828 (comment)	check i686-pc-windows-gnu
#47828 (comment)	check x86_64-apple-darwin
#47828 (comment)	check i686-pc-windows-gnu
#47657 (comment)	check i686-pc-windows-gnu
#48092 (comment)	check i686-pc-windows-gnu
#48127 (comment)	check x86_64-apple-darwin
#47614 (comment)	check x86_64-apple-darwin
#47752 (comment)	check x86_64-windows-gnu
#47804 (comment)	check x86_64-apple-darwin
#47804 (comment)	check x86_64-apple-darwin
#48158 (comment)	check x86_64-apple-darwin
#47906 (comment)	check x86_64-apple-darwin

[kennytm]

The second victim (#47657) contains nothing useful in the log https://ci.appveyor.com/project/rust-lang/rust/build/1.0.6293/job/k1wbokf5c59o4ghj.

cc @alexcrichton @eddyb

[eddyb]

Does compiletest really hide ICEs now? How did this happen 😞?

[alexcrichton]

There's some more information as well at the top of #47828 (comment), notably that this has been seen at least once on x86_64-apple-darwin.

The fact that it's spurious yet deterministic about these three tests is pretty disturbing, and sort of points to maybe this being a nondeterministic miscompilation in librustc_resolve itself. I think it's pretty certain this was introduced via the LLVM 6 upgrade, but given the lack of ability to reproduce or anything else my preferred course of action here is:

• Leave LLVM 6 in master as-is. We want this for a whole slew of other reasons. That, and it's not like we have a perfect test suite otherwise.
• We're very close to the beta branching point. If it becomes a problem we should revert the LLVM upgrade on beta immediately (shouldn't be hard to do with just a few submodule updates).
• Hopefully we'll get some more testing of this via nightly users on the nightly branch. I'm sort of hoping that someone else runs into a segfault/bug which allows us to actually debug this...

@kennytm if this starts bouncing every other PR though we can certainly reconsider!

[kennytm]

@eddyb I've verified that the ICE is swallowed by compile-test 😢. However, even if it is not swallowed, the ICE is useless without RUST_BACKTRACE=1 😓

Edit: Wow, even worse, the test case for #48000 passed successfully even though it is still ICE-ing. check_no_compiler_crash is not effective anymore?

Edit²: I think this (ICE being ignored) is caused by #47634

Edit³: Nevermind, the ICE detection is functioning properly with #48048. The ICE in #48000 never failed with --error-format json. But it specifically catches the words "error: internal compiler error".

Edit⁴: The "hide ICE" thing should be fixed in #48127.