model-writtenrun 74 - may be wrong, links go to the source messages. Commit-by-commit LLM analysis is in the Commits dialog.
- Thomas Munro opened the thread with experimental standard C/C++ implementations of port/atomics.h and spinlocks, plus header and build-system cleanups, emphasizing that portability, generated code, and performance still needed investigation #1.
- Heikki Linnakangas endorsed committing the limits.h prerequisite independently and corrected include placement; he and Tom Lane also questioned removal of the Solaris template, which Thomas acknowledged was an accidental rebase error #2 #3 #4.
- Thomas committed the limits.h prerequisite from #1 with Heikki's include-order corrections as 017249b828, then posted the remaining work with Solaris restored and Visual Studio 2022 support enabled #5.
- Peter Eisentraut flagged older GCC buildfarm compilers, Aleksander Alekseev reported successful Linux x64 testing with and without Valgrind, and Thomas identified Visual Studio requirements and cross-platform code generation as outstanding adoption questions #6 #7 #8 #9.
- Greg Burd supplied a Windows ARM64 fix on top of Thomas's series and reported passing tests; Thomas then questioned the choice of YIELD versus ISB for spin delay and called for architecture-specific performance research #11 #15.
- Tom Lane, Alvaro Herrera, and Thomas clarified that Solaris/illumos support was intended to continue and that its template removal had already been corrected #12 #13 #14.
- After a four-month gap, Greg took over the series, reported testing across Linux, FreeBSD, RISC-V, and Windows ARM64, and added an ISB-based Windows ARM64 spin delay while retaining Thomas's overall conversion #17.
- Peter warned that raising the MSVC baseline was too late for the release cycle, and Tom requested a fallback-capable alternative implementation rather than immediate deletion of the existing code; Greg accepted the direction and deferred the work to the next cycle #18 #20 #21.
- Greg returned with a major redesign making stdatomic optional and default-off, retaining traditional atomics, and attaching a benchmark harness and a separate future-removal patch that were not proposed for immediate merge #22.
- Greg reported RISC-V parallel-hash-join failures with relaxed atomic reads and an ARM performance regression with stronger writes, motivating sequentially consistent reads, relaxed writes, and combined compiler/thread fences; subsequent refreshes clarified those choices and adjusted which auxiliary patches CI applied #22 #23 #24.
- Peter questioned the value of permanently maintaining two implementations, and Greg confirmed that complete replacement remained the goal, proposing an initial period with both paths available for testing #25 #26.
- After another rebase, Greg corrected implementation mistakes including full-barrier operations, spinlock release ordering, and the explanation of narrow atomics, while removing unused 8-bit and 16-bit atomic types #27 #28.
- Shihao Zhong tested the patch on macOS ARM64 and found a C++23 linkage failure and inefficient flag code generation; Greg fixed the include linkage and adopted atomic exchange for flag acquisition #29 #30.
- Shihao next discovered that the supposedly optional patch had become mandatory in both build systems, prompting Greg to restore default-off selection and correct configuration help; Shihao's subsequent testing found no further problems, and Greg added review credit while correcting stale commit descriptions #31 #32 #33 #34.
- Greg finally offered a companion single-patch version combining the implementation with deletion of traditional atomics and spinlocks, leaving the conservative opt-in version available and the choice between them to a committer #35.
- At the thread's end, only the limits.h prerequisite had landed as 017249b828; Greg paused further submissions pending interest, with the GCC 13.3 RISC-V failure still unexplained, stronger-read costs insufficiently benchmarked, and POWER, s390x, and Solaris/illumos testing outstanding #35.