model-writtenrun 5314 - may be wrong, links go to the source messages. Commit-by-commit LLM analysis is in the Commits dialog.
Jun - Oct 2021 · 7 entries
- Peter Eisentraut proposed generating copy, equal, output, input, and NodeTag support from annotated C structs in #1, preserving handwritten functions for exceptional cases and initially keeping generated coverage close to existing behavior.
- David Rowley suggested explicit array-length annotations and revisiting a metadata-driven implementation in #2, while Andres Freund advocated metadata and a proper C parser in #4; Tom Lane favored the Perl approach's modest tooling requirements and incremental delivery in #5.
- Four unchanged prerequisites from #1 landed first: the ExprState field rename as d9a38c52ce, the _outValue argument rename as 3d25b4ea6e, consistent support-function names as 31360381f0, and missing enum tags as 983bdc4fac.
- After Tom's explicit preparatory-patch review in #8, the remaining original cleanups landed unchanged: T_MemoryContext removal as 256909c6c1, SeqScan embedding as 2226b4189b, NestPath embedding as 18fea737b5, and centralized pointer-size checking as c1132aae33.
- Peter posted a second preparatory series in #9; Jacob Champion reviewed it and requested another Cardinality field in #10, while Noah Misch accepted the longer serialization labels in #12.
- The second series produced unchanged T_Expr removal 8539929197, array macros 308da179e7, WRITE_INDEX_ARRAY bdeb2c4ec2, and consistent output labels e581360696; Cardinality landed as 6fe0eb963d with Jacob's additional enrtuples substitution.
- The main generator returned in #14 after its prerequisites landed, now using explicit array_size annotations and supporting MSVC; Andrew Dunstan's comment-parser review in #17 led to whole-file comment stripping and generated include directives in #18.
Dec 2021 - Aug 2022 · 20 entries
- Tom's detailed review in #21 led to documented attributes, clearer generated filenames, preserved comments, and removal of debugging settings in #27; the metadata alternative remained future work after the debate in #22 through #26.
- David raised output-equivalence and comment-formatting concerns in #28, and Peter chose early PostgreSQL 16 development rather than the PostgreSQL 15 feature-freeze period to reduce conflicts and allow more testing in #29.
- Alvaro Herrera rebased the patch in #32 and uncovered JSON typedef obstacles; his enum-tag fix landed unchanged as d47a11da9e, while Tom and Peter's agreement to replace JsonPathSpec in #33 and #34 produced dc2be6ed47.
- Peter separated header comment cleanup into #37, which landed with extra RangeVar and Var formatting as 835d476fd2.
- Tom's Makefile review in #38 shaped #39, and his demand to expose special handling in node declarations in #40 led to struct-level attributes, explicit ignore markings, and rejection of unknown attributes in #42.
- Tom's combined review and patch in #44 separated copy and equality suppression, introduced explicit copy_as/read_as values, reduced unnecessary generated support, and identified Plan and Join as abstract; it also exposed a pgindent incompatibility with attribute placement.
- Additional preparation landed before the generator: 3140f08985 reformatted more comments for the expanded annotations, 251154bebe removed manual Plan and Join support following #44, and bf1f4a364d aligned the A_Expr serialization label with the same discussion.
- Peter folded Tom's changes into #46, but Tom preferred attributes near the top of each struct and supplied that layout plus zero-value and read_as corrections in #51, followed by a documentation correction in #52.
- The complete series in #51 landed as 964d01ae90 with only the final comment correction and catalog-version bump added; Peter reported completion after a brief VPATH buildfarm problem in #53.
- The first landing deliberately left broader utility-statement serialization and exceptional handwritten support outside its scope, and Peter planned immediate removal of the obsolete conditional sections in #46.
- Tom posted follow-up maintenance patches in #55; the expanded documentation became 3cd0ac9878, while accurate error locations, generated-file notices, and temporary-file cleanup landed together unchanged as 8eccaf6525.
- Andres questioned the input-header order in #56, and discussion of dependency ordering produced bf022d337e, with principal headers in dependency order and incidental headers alphabetized.
- Tom then warned in #63 that different build-system input orders could silently break extension ABI, while alphabetizing all tags enlarged the backend; Robert Haas argued for preserving automatic generation in #64.
- The ABI discussion settled on validating a canonical input list, and Tom posted explicit stable-branch numbering safeguards in #73; Andres accepted the canonical-list approach in #78 after discussing alternatives based on generated dependencies.
- Alvaro's separately posted XidList copying omission in #49 became 5ca0fe5c8a, expanded with equality and test-hook support after Tom asked about missing cases in #50.
- The ABI safeguard patch in #73 landed as eea9fa9b25 with canonical input-file validation added, allowing automatic NodeTags to remain rather than reverting to a manually maintained enum.
- Tom's parser-hardening patch in #81 became 7c0eb3c622, adding multiline and function-pointer parsing, rejecting unknown struct syntax, and gaining additional attribute-line and amcostestimate handling before commit.
- Amit Kapila reported stale generated files during Windows builds in #82 and traced the missing header-timestamp checks in #84, prompting Tom's MSVC repair in #85.
- Amit's test in #86 exposed empty entries in the extracted input list and supplied a working explicit-list alternative; Tom retained Makefile extraction, filtered empty entries, and improved diagnostics in #87.
- The final MSVC patch landed unchanged as 9a9f25e217 after cfbot testing reported in #88, and Amit confirmed the committed fix worked in the thread's last message, #89; no submitted fix remained pending.