model-writtenrun 7715 - may be wrong, links go to the source messages. Commit-by-commit LLM analysis is in the Commits dialog.
- Tom Lane opened the thread with #1, explaining that back-references incorrectly inherited constraints such as anchors from their referenced subexpressions, and supplied a patch to remove those constraints while preserving character matching.
- The second patch in #1 replaced approximate NFA checks with exact string comparisons whenever the captured string was already available, avoiding expensive failed matches in common right-deep concatenation trees; this optimization depended on the proposed semantics correction.
- Joel Jacobson tested both patches in #2: a query he had canceled after 23 minutes finished in roughly 20–25 milliseconds, while all 3,360,068 corpus cases retained their results and overall corpus timings were similar.
- In #3, Tom reported a possible overall improvement of about 1% and abandoned an attempted optimization for left-deep trees because it changed ambiguous-match results; at the thread's last message, he was waiting for objections to the semantics change.
- The semantics patch from #1 subsequently landed unchanged as 4aea704a5b, with Joel credited as reviewer and backpatching explicitly ruled out because the behavior change would break compatibility.
- The performance patch from #1 then landed unchanged as 0c3405cf11, also crediting Joel, completing the submitted series while leaving the left-deep-tree limitation in place.