model-writtenrun 8049 - may be wrong, links go to the source messages. Commit-by-commit LLM analysis is in the Commits dialog.
- John Naylor proposed replacing binary searches with perfect hashing for NFC and NFKC normalization quick checks, preceded by a change to the hash generator's candidate multipliers, and reported substantial speed improvements at a modest space cost in #1.
- Mark Dilger suggested caller-specific multipliers and identified a one-element increase in a generated keyword table in #2 and #4; John defended keeping multiplier selection internal and explained that the harmless increase avoided a table size divisible by a multiplier in #3 and #5.
- After a summer gap, John posted a revision that excluded the generated normalization header from pgindent instead of changing PerfectHash.pm's output formatting in #6.
- Mark supplied assembly comparisons for the multiplier change in #7 and tested whether the quick-check wrapper structure imposed runtime overhead in #8; the structure was retained for consistency, and he marked the work ready for a committer in #10.
- Michael Paquier inspected generated assembly and confirmed quick-check performance gains with gcc and MSVC in #11, then committed the unchanged multiplier prerequisite from #6 as 2a73164581.
- Michael merged the remaining NFC and NFKC changes into #12, adding explanatory comments; John questioned the separate pgindent exclusion section in #13, and Michael justified it by the different generating scripts in #14.
- The combined patch from #12 landed as 80f8eb79e2 with light naming and pgindent-related edits, and Michael confirmed further benchmarks while suggesting decomposition-table optimization in #16.
- Masahiko Sawada reported an undeclared htonl() warning on FreeBSD in #17; Michael posted a switch to pg_hton32() in #18, and Masahiko confirmed that the patch resolved the warning in #19.
- John supplied a separate decomposition-hashing experiment in #20, noting possible frontend binary growth and opportunities to improve recomposition; he proposed moving that work to a new thread, which Michael supported in #21.
- The tested portability fix from #18 landed unchanged as e578c17d81, with its application confirmed in #22.
- Peter Eisentraut reported generated-header whitespace errors in #23; Michael considered a .gitattributes exception in #24, and Tom Lane clarified the repository's custom whitespace rules in #26.
- John proposed right-padding the generated hash-table values in #27, and Michael retained the aligned columns while simplifying the generator and removing its final comma in #28.
- John accepted the formatting revision in #29, and it landed unchanged as 19ae53c92d; Michael's final confirmation in #30 leaves the quick-check optimization and both follow-up fixes committed, with decomposition optimization deferred to a separate thread.