model-writtenrun 2581 - may be wrong, links go to the source messages. Commit-by-commit LLM analysis is in the Commits dialog.
- Tristan Partin proposed requiring uselocale() on non-Windows systems to simplify thread-safe locale work in #1, but Tom Lane identified NetBSD as a counterexample in #7, preventing that blanket requirement.
- Dagfinn Ilmari Mannsåker highlighted platform bugs discovered by Perl in #4, while Thomas Munro identified ECPG's setlocale() fallback as unsafe for concurrent use in #8.
- Thomas Munro proposed explicit locale conversion functions in #10, and Tom Lane favored using available _l() functions rather than implementing a new floating-point parser in #11.
- Thomas Munro's first patch in #12 exported conversion wrappers from pgtypeslib, while acknowledging that independent pgtypeslib use and Windows DLL variable sharing complicated initialization.
- Tom Lane tested that implementation on NetBSD and supplied a printf-attribute fix in #13, then realized ECPG already used PostgreSQL's snprintf.c and suggested changing that implementation instead in #14.
- The discussion converged on making snprintf.c consistently use the C locale in #18; Tom Lane noted that the existing Ryu implementation did not readily satisfy printf's requested digit counts in #19, and Ryu conversion remained outside the patch.
- After a long gap, Thomas Munro returned with a redesigned shared locale constructor and direct strtod_l() calls in #20, then retained the backend's existing locale behavior and added constant-locale support and allocation checks through #21 and #23.
- Tristan Partin approved the revised direction and supplied a strtod() poisoning add-on in #24; Thomas welcomed it in #25, but subsequent submissions and the eventual commit did not incorporate it, with no rejection reason recorded.
- Peter Eisentraut reviewed the design in #26, requested separate xlocale.h cleanup and consistent header placement, and questioned the initialization coupling between ecpglib and pgtypeslib.
- Peter extracted the xlocale.h portion of #23 into 9c2a6c5a5f, correcting the configure probe and include placement as reported in #27.
- Thomas resumed the remaining ECPG work in #28 and applied Peter's Windows strtod_l mapping feedback in #29, while acknowledging that allocation-failure handling for standalone pgtypeslib callers remained imperfect.
- Heikki Linnakangas endorsed the ECPG patch while questioning initialization coverage and unnecessary floating-point parsing in #30; Peter separately requested staged Windows runtime cleanup in #31.
- Thomas supplied the three Windows cleanup patches in #32, Peter endorsed them in #34, and Thomas revised their explanations and raised obsolete MinGW documentation in #35.
- Michael Paquier supported simplifying the old MinGW instructions in #36, and Peter supplied replacement MSYS2 and Flex/Bison guidance in #38, which Thomas accepted in #39.
- The first and third patches of #35 landed with unchanged code as f1da075d9a and a62d90f2e5, removing the obsolete configure test and reverting the workaround for MinGW's dummy _configthreadlocale().
- The runtime cleanup from #35 and documentation from #38 were combined into 1758d42446, formalizing reliance on ucrt and removing obsolete MinGW locale restrictions.
- Peter revived the outstanding ECPG patch in #40, rebasing Thomas's work and adding compiler-warning and ordering fixups, then committed the combined submission as 8e993bff53 and announced it in #41.
- Masahiko Sawada reported buildfarm failures and reproduced a test_json_parser linker error in #42; Peter decided to revert in #43 and backed out the ECPG change with 3c8e463b0d.
- After the revert, Masahiko isolated a Make-versus-Meson pthread-linking difference and reported a locally successful Makefile adjustment in #45, but no replacement landing appears in this thread.
- Tom Lane later documented in #46 that the already-landed xlocale.h cleanup also avoided a newer macOS build problem affecting older historical revisions, providing a workaround for bisecting those revisions.
- At the thread's end, the header and Windows cleanup remained landed while the ECPG redesign remained reverted: Alvaro Herrera revived the issue in #47, Peter proposed live buildfarm debugging in #48, and Alvaro pointed to missing pthread flags in test_json_parser in #49.