model-writtenrun 1191 - may be wrong, links go to the source messages. Commit-by-commit LLM analysis is in the Commits dialog.
- Bertrand Drouvot proposed rejecting -ffast-math during Meson configuration to match autoconf, and Nazir Bilal Yavuz applied the patch and confirmed it worked #1 #2.
- Andres Freund argued that configuration checks could miss flags inherited from dependencies, while Peter Eisentraut explained that the configure check addressed historical Linux distributions and proposed relying on the existing source-level guards instead #3 #4.
- Tom Lane questioned whether guards in datetime code were obsolete after floating-point timestamps disappeared; after Bertrand reported regression failures, Tom independently reproduced widespread failures involving isnan(), isinf(), and near-overflow handling and proposed moving the defense to floating-point code #5 #6 #8.
- Andres Freund recommended putting the guard in float.c rather than a shared header so extensions could intentionally enable -ffast-math #9.
- Bertrand traced his uncaught failures to -ffast-math remaining in link flags, and Andres challenged the build setup; Bertrand then isolated a two-step Meson configuration that replaced compilation flags while retaining the problematic linking behavior and renewed his argument for a Meson check #10 #12 #13 #14.
- Bertrand and Tom accepted placing the guard in float.c to accommodate extensions, and Bertrand submitted a replacement patch that also removed the configure check and obsolete datetime guards, abandoning the proposed Meson check #15 #16 #17.
- Tom committed the replacement patch with expanded comments explaining fast-math hazards and extension implications 82ff54377e; his final message confirmed the push, leaving the revised defense landed without a separate fix for the reported Meson link-flag scenario #18.