model-writtenrun 7140 - may be wrong, links go to the source messages. Commit-by-commit LLM analysis is in the Commits dialog.
- Tom Lane proposed removing the remaining HPPA support after retiring its last buildfarm machine, arguing that untested architecture-specific spinlocks no longer served their original testing purpose #1.
- Michael Paquier approved the C changes and asked about configuration-script references, which Tom explained should remain because those scripts are imported from upstream #2 #3.
- Noah Misch initially opposed removing apparently low-maintenance support for NetBSD/OpenBSD HPPA, while Tom argued that nominal package availability did not establish testing or practical use and that supported architectures needed buildfarm coverage #4 #5 #7.
- Andres Freund identified a concrete benefit: HPPA required spinlock-based atomic emulation that prevented some signal-handler uses; the discussion converged on requiring native 32-bit atomics while retaining 64-bit emulation and considering removal of semaphore-based spinlocks #8 #10 #11 #12.
- Noah accepted broader atomics cleanup as a sound justification and later explicitly supported HPPA removal; Tom then committed the original patch with routine rebasing as edadeb0710 and announced completion #13 #19 #20.
- Thomas Munro followed with experimental patches removing --disable-spinlocks and --disable-atomics and optionally implementing atomics through stdatomic.h, explaining that native 32-bit atomics would support his planned SIGALRM-handler operations #21.
- Tom supported removing the two disabling options but questioned surrendering control of synchronization semantics and performance to stdatomic.h, so Thomas shelved that experiment and added compiler- and memory-barrier fallback removal instead #22 #23.
- The barrier-removal proposals originated in the LTO discussion thread #49720, but Thomas supplied their matching implementations here in #23, making this thread the source of the committed patches.
- Heikki Linnakangas approved the revised series and confirmed through a compilation experiment that MSVC was missing the x86-specific code paths, including an optimized CRC path, because __x86_64__ was not defined #24.
- Thomas supplied a separate MSVC fix using explicit Windows architecture macros, preferring that approach over defining a GCC-style macro that external code might interpret as permission to use GCC assembly; backpatch scope remained under discussion #25.
- The four patches from #23 landed unchanged as spinlock-fallback removal e25626677f, 32-bit atomics requirements 8138526136, compiler-barrier requirements a011dc399c, and memory-barrier requirements 83aadbeb96, with Thomas announcing the push and expected buildfarm configuration fallout #26.
- Thomas next proposed implementing spinlocks through pg_atomic_flag, preserving relaxed-load spinning optimizations while deleting duplicated assembly and the unused SpinLockFree interface #27.
- Heikki supported the spinlock conversion but highlighted its acquire/release semantics and requested additional confidence in the 32-bit x86 behavior; Thomas cited Intel guidance and acknowledged the need to document and investigate the ordering implications #28 #29.
- Andres requested a way to detect repeated SpinLockRelease calls, retired the obsolete fallback-testing buildfarm configurations from master, and supplied contention benchmarks illustrating the cost of stronger release barriers #30 #31 #32.
- Thomas revised the conversion to delete s_lock.h entirely and added debug-only initialization markers and release assertions, but Andres rejected the debug-only structure member because it changed the ABI between assertion and non-assertion builds #33 #34.
- Thomas replaced the ABI-changing marker with nonzero states inside the x86 atomic-flag implementation and separated the double-release assertion, leaving this revised spinlock series awaiting further review #35.
- The final message revisited the MSVC detection fix as a broader draft defining centralized PG_ARCH_* macros; HPPA removal and the four fallback-removal commits had landed, while architecture detection and the spinlock conversion remained proposals and stdatomic.h remained deferred #36.