model-writtenrun 9601 - may be wrong, links go to the source messages. Commit-by-commit LLM analysis is in the Commits dialog.
Dec 2018 - Jan 2019 · 19 entries
- John Naylor proposed replacing keyword-name pointers with offsets into a contiguous string to reduce storage and improve locality in #1; Tom Lane urged generated rather than hand-maintained PL/pgSQL tables and consistent conversion of all keyword consumers in #2 and #4.
- John's first implementation covered only PL/pgSQL unreserved keywords in #5, and Tom's review in #10 established a single generated header, a generator under src/tools, preservation of the existing keyword-list format, and distribution rules that would not require Perl for tarball builds.
- John suggested separating payload data in #11, and Tom developed that into an index-returning lookup API in #12; John implemented the resulting offset and payload arrays in #15.
- Andrew Gierth raised perfect hashing in #6, but tooling concerns initially kept the work focused on compact storage; Robert Haas proposed first-character dispatch in #16, which John added as an experimental second patch in #28.
- Tom questioned the maintainability of preferred initial search positions and reviewed generator details in #31, while Andrew Dunstan proposed Getopt::Long in #35; John incorporated the applicable feedback and converted all keyword consumers in #42, leaving the dispatch experiment out.
- The broader parser discussion produced no replacement project: Andres Freund advocated recursive descent in #22, Tom emphasized ambiguity detection and maintenance risks in #29, and Andrew Dunstan reported limitations of available generator alternatives in #30.
- Andres found frontend linkage and rebuild-dependency failures while testing the complete conversion in #43; John corrected the linkage, rebased the scanner changes, and fixed MSVC integration in #47.
- Joerg Sonnenberger supplied a compact perfect-hash example in #45 and explained its NetBSD implementation and license in #53, reopening hashing as a practical follow-up while the group chose to land the storage infrastructure first in #51.
- John supplied a delta separating token and category arrays in #52; Tom incorporated that idea, introduced ScanKeywordList, separated lookup code from keyword data, and removed backend token dependencies from src/common in #57, then repaired Windows build integration in #58.
- Tom's benchmark in #61 measured roughly a 4% raw-parsing improvement from compact storage and roughly 19% from an experimental hard-coded perfect hash, motivating generated hashing for every keyword set.
- Tom retained independent keyword sets and generated hash functions to avoid cross-component rebuild dependencies in #63, then committed the storage conversion from #58 as afb0d0712f, adding maximum-keyword-length rejection as reported in #64.
- Tom posted a complete Perl-generated perfect-hash implementation in #66, then revised its graph representation and signed table entries in #68 so unused entries could help reject non-keywords without a final modulo operation.
- Andres requested a reusable generator module in #67, and Tom extracted PerfectHash.pm and added automatic table-element sizing in #74.
- Andres identified the sparse builtin-function OID index as another possible hashing target in #73; Tom's experiment in #75 showed limited space savings and additional lookup work, leading him to prefer truncating the existing direct-index array.
- John reviewed module exports and case-sensitivity terminology in #76, while the OID experiment exposed weak multiplier choices; Tom's revised keyword patch in #80 added named options, fixed-length keys, and a broader search for workable hash parameters.
- Tom posted the simpler OID-array reduction in #81, Andres examined its external-bound lookup cost in #83, and the patch landed unchanged as 8ff5f824dc, saving nearly 8 KB without adopting OID hashing.
- John proposed a negatable case-sensitivity option in #89 and removal of the keyword-count macro in #90; Tom accepted the option direction but preferred case-fold terminology in #93 and retained the macro for possible future uses in #94.
- John also proposed warning about high function OIDs in #91, but Tom rejected the build-time warning in #92 because development patches deliberately use high OIDs and would generate excessive noise.
- John's final review in #95 identified confusing module comments and stale binary-search warnings; Tom accepted those changes in #96 and landed the hashing implementation from #80 with light naming and comment edits as c64d0cd5ce.
Mar 2019 · 2 entries
- After a two-month gap, Joel Jacobson proposed hashing several bytes per iteration in #97; Tom questioned the overall benefit and endianness implications in #98.
- Joel ended the thread with an experimental PerfectHash.pm patch, byte-swapping logic, microbenchmarks, and a tentative 3% overall parsing improvement in #99; the three supplied commits had landed, while this later optimization remained uncommitted in the available record.