First discovered in bytecodealliance/wasmtime-go#241 which is one of the main users of MinGW builds (AFAIK Go doesn't have MSVC integration). I've bisected through Github Actions (couldn't do so locally) and discovered the following:

• Wasmtime is currently built with the github actions windows-2022 runner which uses gcc 12.2.0 from MinGW/MSYS2/something.
• When these precompiled artifacts are used in wasmtime-go in tests test can fail with exception 0xc0000028 which stands for STATUS_BAD_STACK (aka resurrection of STATUS_BAD_STACK on Windows on trap #1967)
• Bisection showed the following properties:
  • The faulty PR is Test out -D__USE_MINGW_SETJMP_NON_SEH #9929 (which in retrospect isn't too surprising)
  • If the "faulty PR" is built with windows-2025, which uses gcc 14.2, then the bug does not happen.

So it looks like our workaround of -D__USE_MINGW_SETJMP_NON_SEH, used to fix another bug in MinGW gcc is perhaps triggering some other bug in MinGW which is seemingly fixed with more recent GCC versions.

Overall I personally wish we could just drop this target and not worry about it, but I'm unaware of how to integrate with Go on Windows any other way...