model-writtenrun 844 - may be wrong, links go to the source messages. Commit-by-commit LLM analysis is in the Commits dialog.
- Jakob Egger reported universal-build CPUID failures and an Intel/Rosetta checksum compilation failure in #1; after Daniel Gustafsson and Tom Lane reported working Intel builds, Jakob narrowed the Intel failure to Rosetta with Xcode 26.2 and confirmed that Command Line Tools 26.4 worked in #8.
- Tom Lane reproduced the universal failure and explained that a single configure run cannot reliably detect architecture-specific features for both targets in #9; he favored restoring basic buildability without redesigning configure, and Nathan Bossart agreed in #10.
- Tom proposed making pg_cpuid() return zeroes when CPUID support is unavailable in #11, Lukas Fittl explained why TSC would safely fall back in #12, and John Naylor supplied that implementation in #13.
- Tobias Bussmann confirmed successful universal builds and passing checks on both architectures in #14, after which John committed the patch unchanged as 901ed9b352 and announced it in #15.
- John's configure-output analysis in #15 showed that universal builds lacked optimizations on both architectures, and Tom explained in #16 that the slower CRC fallback was longstanding rather than a new performance regression.
- Tobias reopened the build problem in #17 with comparisons across six configurations, showing that building ARM or universal binaries from an x86 environment incorrectly enabled AVX2 and failed in checksum.c; he also relayed confirmation from a native Intel host.
- Tom reproduced the failure and supplied a temporary checksum.c AVX2 guard plus a pg_cpu.h guard adjustment in #20, recommending the workaround for EDB's beta build while broader review continued.
- Tobias suggested combining separately optimized builds with lipo in #19, but Tom considered that investment disproportionate to the remaining lifetime of Intel macOS support in #21, leaving optimized universal-build support unresolved.
- Tom corrected his diagnosis in #22: the earlier pg_cpu.h adjustment was a no-op, while configure accepted an AVX2 target attribute that ARM compilation merely warned about; his replacement patch made x86_feature_available() return false on non-x86 platforms but still allowed compiler warnings.
- At the thread's last message, Tom replaced those approaches with global suppression of AVX2 and AVX512 configuration macros on non-x86_64 targets in src/include/c.h, explaining in #24 that this would protect future consumers and accommodate pending related work.
- The final proposal from #24 landed the following day as e88bd2736f with only an #endif comment correction; the temporary checksum guard and feature-query stub were superseded, and the commit required no backpatch because the build regression was new in v19.