model-writtenrun 8677 - may be wrong, links go to the source messages. Commit-by-commit LLM analysis is in the Commits dialog.
- Peter Eisentraut proposed extending the SASLprep normalization infrastructure to all four Unicode forms and exposing SQL-standard NORMALIZE and IS NORMALIZED expressions, with quick checking and protocol-level normalization identified as possible further work #1.
- Daniel Verite found matching results against ICU but severe NFC/NFKC performance costs; Peter traced those costs to sequential recomposition-table scans and suggested a future reverse mapping #2 #3.
- Peter added Unicode quick-check property tables to accelerate IS NORMALIZED, then corrected an inappropriate UTF-16-style codepoint skip after Daniel demonstrated that it could over-read the input #4 #5 #6.
- Andreas Karlsson reviewed table size, conversion overhead, and code style; Peter accepted the style changes but questioned whether partial normalization after a MAYBE result was valid #7 #9.
- Daniel again verified the revised implementation against ICU and documented its remaining performance gap; Peter fixed a separately reported pre-existing comment typo and declined switching to ICU because of libpq dependencies and the burden of maintaining multiple implementations #10 #11 #12 #13.
- Andreas considered the performance sufficient for useful deployment, and Daniel marked the series ready after finding no further issues #15 #16.
- Peter committed the unchanged API portion of #6 as d40d564c5a, while retaining the SQL portion for further work on its table and binary size #17.
- Peter resubmitted the SQL remainder with backend-only quick-check tables, 32-bit bitfield entries, and no NFD/NFKD quick-check tables, accepting full normalization checks for those cheaper decomposed forms #18.
- John Naylor supplied a perfect-hash alternative and reported passing regression tests; his subsequent benchmark improved the NFC check from 419 ms to 310 ms while increasing added binary size from 33 kB to 57 kB #19 #20.
- Peter committed the binary-search version from #18 as 2991ac5fc9, with only catalog-version and expected-output differences, and deferred John's hash changes because questions about shared hash parameters and future Unicode updates were too late for the release cycle #21 #22.
- The thread ended with both original features landed and the hash optimization deferred: John explained the parameter behavior and proposed a future split between array restructuring and hash lookup, with hash functions generated during the build #23.