model-writtenrun 7671 - may be wrong, links go to the source messages. Commit-by-commit LLM analysis is in the Commits dialog.
- Regina Obe reported that PostGIS would not compile against PostgreSQL 14 under MinGW-64 because of conflicting stat definitions #1; Tom Lane traced this to including libxml headers before postgres.h #4, and Regina confirmed that correcting the include order fixed compilation #6.
- The investigation then shifted to runtime crashes: Regina supplied PostGIS and PostgreSQL backtraces pointing through pg_re_throw into Windows unwinding code #9 #12, and Tom suspected broken longjmp behavior resembling an earlier MinGW report #13.
- Regina found that disabling assertions avoided the crashes and that PostgreSQL 13 passed with the same configuration #14 #16; her subsequent bisection identified an error-reporting change as the trigger, although its connection to setjmp remained unclear #20 #21.
- Juan José SantamarÃa Flecha reproduced the failure and supplied a patch using GCC's setjmp/longjmp builtins for MinGW-64, reporting that it fixed his reproduction #22.
- Tom questioned whether the native jmp_buf was compatible with those builtins #23, then posted a revision using an intptr_t[5] buffer and explained that the resulting ABI break ruled out backpatching #24.
- Regina confirmed that PostgreSQL and PostGIS passed all tests with the revised patch #25; Heikki Linnakangas connected the symptoms to another historical failure report, reinforcing the investigation into longstanding MinGW problems #26 #27.
- Tom investigated an old MinGW buffer-clearing bug #28, but Regina reproduced the crashes with a newer GCC 10.2 toolchain as well #29; after walleye's owners also confirmed the workaround, Tom announced that he would push it to HEAD for broader testing #30.
- Tom committed 146cb3889c from #24, retaining Juan's builtin-based workaround and the revised buffer while simplifying siglongjmp to a direct builtin alias; the ABI change remained confined to the development branch.
- After the commit, Andres Freund proposed a union buffer to preserve implementation flexibility #31, but Tom explained that mixing native and builtin setjmp/longjmp calls would still break compatibility #32; Andres accepted that objection and considered runtime dispatch too complicated, leaving the committed workaround in place without another patch #33.