`tests/run-make/dump-ice-to-disk` is flakey on i686-mingw

[jieyouxu]

I've seen multiple instances of dump-ice-to-disk failing on completely unrelated PRs, but it's completely non-obvious to me how it can fail non-deterministically.

• Enable zstd for debug compression. #125642 (comment)
  • i686-mingw: assertion left == right failed (left: 59, right: 60)

  • rust/tests/run-make/dump-ice-to-disk/rmake.rs

    Line 32 in 730d5d4

    assert_eq!(default, default_set);

• trans her gender #128293 (comment)
  • i686-mingw: assertion left == right failed (left: 61, right: 62)

  • rust/tests/run-make/dump-ice-to-disk/rmake.rs

    Line 55 in 2b78d92

    assert_eq!(short, default_set);

• Link std statically in rustc_driver #122362 (comment) failed twice
  • i686-mingw: assertion left == right failed (left: 60, right: 59)

  • rust/tests/run-make/dump-ice-to-disk/rmake.rs

    Line 76 in 730d5d4

    assert_eq!(full, baseline);

  • i686-mingw: assertion left == right failed (left: 60, right: 59)

  • rust/tests/run-make/dump-ice-to-disk/rmake.rs

    Line 63 in c9bd03c

    assert_eq!(short, baseline);

---

Past investigations

Update: see https://rust-lang.zulipchat.com/#narrow/stream/182449-t-compiler.2Fhelp/topic/getting.20.60Box.3Cdyn.20Any.3E.60.20as.20ICE.20message posted by @camelid:

I'm getting the following ICE message during development of a PR:

thread 'rustc' panicked at compiler/rustc_middle/src/ty/context.rs:3013:21:
Box<dyn Any>
stack backtrace:

this line in my PR is

bug!("No bound vars found for {}", self.hir().node_to_string(id))

it turns out it displays the message correctly to stderr but not to the rustc-ice file

Is it possible that the test isn't flakey, but that our ICE dump actually regressed? cc @estebank and @nnethercote for diagnostic and ICE dumping in case you guys have any ideas.

Update 2: #128956 has the Box<dyn Any> symptom as well, it might not be limited to ICE dumps to files but also to stderr? Unclear at this moment.

thread 'rustc' panicked at compiler\rustc_codegen_llvm\src\context.rs:1138:21:
Box<dyn Any>
stack backtrace:
   0:     0x7fffe0c051b0 - std::backtrace_rs::backtrace::dbghelp64::trace::h0891390128512157

Update 3: Box<dyn Any> seems to be a separate thing.

[jieyouxu]

I will make it dump the ICE message later today to aid debugging.

[jieyouxu]

Possibly related: https://rust-lang.zulipchat.com/#narrow/stream/182449-t-compiler.2Fhelp/topic/getting.20.60Box.3Cdyn.20Any.3E.60.20as.20ICE.20message with Box<dyn Any> in ICE dumps, we'll see if #128916 gets merged and we still see assertion failures in random unrelated PRs.

I am beginning to suspect that this test is in fact not flakey, but it's actually correctly detecting that ICE dumps regressed.

[jieyouxu]

Hmm, on further investigation, could we have two separate bug here? That i686-mingw somehow has ICEs that slightly differ, and that ICE dumps may have regressed to Box<dyn Any>? Not sure.