model-writtenrun 7093 - may be wrong, links go to the source messages. Commit-by-commit LLM analysis is in the Commits dialog.
Feb - May 2022 · 2 entries
- Niyas Sait proposed native Windows ARM64 support in #1, extending MSVC build scripts and adding the spin-delay, CRC header, and symbol-export changes needed to build, with successful regression results on ARM64 and x86-64.
- Thomas Munro questioned ASLR's interaction with shared-memory mappings in #2, while Andres Freund explained existing retry protections in #6; Niyas subsequently demonstrated that the ARM64 linker rejects disabling ASLR in #18.
Jul - Sep 2022 · 4 entries
- Andres Freund's extensive successful ASLR testing in #23 encouraged Michael Paquier to propose enabling it for all MSVC builds in #26, which landed unchanged as 36389a060c.
- After that landing, Tom Lane reported reordered XML output on bowerbird in #30, and Andres Freund opposed hiding the unexplained failure in #36 while independently testing another Windows installation in #37.
- Michael Paquier and Tom Lane traced the instability to unordered XPath node sets in #40, #41, and #42, then rejected imposing sorting because of API and performance concerns discussed in #44 and #45.
- Tom Lane verified that avoiding node() preserved relevant code coverage in #46, adopted Michael Paquier's /e/n2 suggestion from #47, and committed the SQL and expected-output adjustment as 4ea07e7cf3; Michael subsequently reported bowerbird green in #49.
Nov 2022 - Jan 2023 · 4 entries
- Niyas clarified in #51 that only ASLR had landed and resubmitted the remaining ARM64 support, after which Andres identified missing Meson CRC support in #53.
- The revisions from #58 through #64 added Meson support and incorporated John Naylor's intrinsic review and Michael Paquier's corrections to architecture-specific CRC flags, library paths, and documentation.
- Andres argued against extending the retiring legacy MSVC build system in #66, Michael agreed in #71, and Niyas removed those changes while retaining Meson support in #72.
- Niyas retained runtime CRC detection because baseline ARMv8 made CRC optional, as explained in #70, then incorporated Andres's Meson header and control-flow suggestions from #74 into #76.
May - Sep 2023 · 2 entries
- Michael favored applying the simplified patch in #77, but Tom rejected adding a platform after feature freeze in #78, and Andrew Dunstan requested buildfarm coverage in #82.
- Niyas handed the work to Anthony Roberts after leaving Linaro in #87, but unavailable hardware and the requirement for a maintained buildfarm animal left progress stalled through #95, #96, and #107.
Jan - Feb 2024 · 5 entries
- Dave Cramer revived the work in #108 with a reduced documentation/export patch and a buildfarm offer, but acknowledged in #122 that his successful builds had produced x64 binaries.
- Selecting the ARM64 toolchain exposed missing spin-delay support, prompting Dave's experimental locking delta in #127 and its combined submission in #132; Andrew explained the missing definition in #133.
- Andrew returned to Niyas's earlier implementation in #135, obtaining a native build but reporting test access violations, and Dave later reproduced buildfarm failures and a VACUUM crash in #140 and #141.
- Andres questioned Windows SIGILL-based CRC probing in #142, and Thomas explained its incompatibility with Windows signals and proposed unconditional CRC instructions based on ARMv8.1 requirements in #155.
- Dave localized a failure to shared-memory code in #149, Andrew found that disabling optimization avoided it in #150, and Andres identified the inadequate compiler-only S_UNLOCK barrier in #154; Dave reported that a simple MemoryBarrier substitution did not resolve the failure in #158.
Sep - Oct 2024 · 3 entries
- After Dave resumed investigation in #159, Thomas supplied the independent C11 atomics experiment in #163 to help isolate memory-ordering problems, while noting that the known weak spinlocks might not explain the observed crash.
- Dave encountered conflicting atomic definitions in #165, received a pointer to architecture-standardization work in #166, and ended the visible thread in #167 still blocked by an atomic_uint lock-free check.
- The January 2026 ARM64 landing a516b3f00d is attributed to the later exact final patch #609330 in thread #52710, superseding the previous attribution to #135; it carries forward this thread's work with unconditional CRC selection, explicit intrinsic headers, and _InterlockedExchange-based unlocking, excludes the C11 experiment, and leaves further atomic, SIMD, and popcount optimizations for later.