model-writtenrun 7018 - may be wrong, links go to the source messages. Commit-by-commit LLM analysis is in the Commits dialog.
- Tom Lane proposed making Datum exactly eight bytes on every platform to reduce maintenance costs while retaining 32-bit support, attaching a proof of concept that passed check-world on 32-bit FreeBSD in #1.
- Andres Freund reported GCC warnings from pointer conversions, (Datum) NULL casts, and varlena macros in #2; Tom proposed intermediate uintptr_t casts and consistent zero notation in #3, and Andres confirmed that the intermediate casts worked in #4.
- Nazir Bilal Yavuz measured essentially unchanged 32-bit pgbench performance in #6, which Tom accepted as sufficient evidence to proceed in #7.
- Tom proposed separate pointer and Datum variants of varlena helpers in #8; Peter Eisentraut linked related cleanup work in #9 and #10, while Tom favored concise _D variants and retaining existing zero notation in #11.
- Peter asked whether larger Datums should remain possible in #12, and Tom settled on exactly eight bytes because wider native registers were not a practical target and untested size flexibility would deteriorate anyway in #13.
- Tom committed the preparatory conversion mop-up as 665c3dbba4; although this thread motivated its null-cast and missing-conversion cleanup in #2 and #3, the supplied attribution for thread #52009 identifies its development there through Tom's announcement in #597529, superseding the earlier attribution to this thread.
- Tom identified a possible pg_upgrade hazard for extension SP-GiST indexes with small pass-by-value prefixes in #15, but found no affected core operator classes or extension examples; Joe Conway supported accepting the risk with reindexing as a fallback in #16, and Andrew Dunstan agreed in #18.
- Tom expanded the proposal into three patches in #17, retaining the core Datum change with GCC fixes and adding removal of obsolete Datum-width and pass-by-reference branches.
- Peter reviewed the expanded series in #19, requesting consistent int64 conversions and broader FLOAT8PASSBYVAL cleanup; the separate idea of deriving typbyval solely from type length encountered existing exceptions identified by Andres in #20 and extension-compatibility concerns from Tom in #21.
- Tom incorporated the concrete cleanup requests in #22, while retaining compatibility macros and control-file and ABI fields as vestigial and declining to split conversion edits into another patch.
- Peter approved the final attached series and requested remaining four-byte-Datum comment cleanup in #23, which Tom agreed to perform in #24.
- The final series in #22 landed as 2a600a93c7 with an added catalog-version bump, 6aebedc384 with additional comment cleanup, and ee54046601 unchanged.
- Tomas Vondra subsequently reported an alignment-sanitizer failure on a Raspberry Pi running 32-bit userspace in #25; Tom traced it to address arithmetic placing a Datum array after an insufficiently aligned SortItem array in #26.
- Tomas confirmed the proposed MAXALIGN corrections in #27, and Tom landed them with a defensive Size declaration for the allocation length in 09036dc71c, without backpatching.
- Ranier Vilela suggested two additional alignment fixes in #29, but Tom explained that those independent allocations lacked the problematic address arithmetic in #31, and Ranier accepted that explanation in #32, leaving the submitted series and reported regression resolved.