model-writtenrun 1407 - may be wrong, links go to the source messages. Commit-by-commit LLM analysis is in the Commits dialog.
- Anthonin Bonnefoy discovered that enabling JIT inlining without expensive optimization produced essentially unchanged bitcode on LLVM 21, and proposed restoring the missing inline pass for LLVM >= 17 in #1.
- Thomas Munro requested removal of a redundant condition, documentation of the regression's origin, and backpatching in #2; he and Andreas Karlsson questioned whether always-inline was needed for tuple deforming in #2 and #3.
- Anthonin incorporated the main-fix feedback and proposed a separate removal of the LLVM < 17 always-inline pass in #4, but Andres Freund objected because generated deform functions should be inlined even without broader function importing in #6.
- After verifying that default<O0> already includes always-inline, Anthonin withdrew the removal patch and added an explanatory comment in #7, preserving equivalent behavior across LLVM versions.
- Pierre Ducroquet reported a positive review in #10, Andres Freund approved the revision in #11, and Matheus Alcantara demonstrated successful inlining with LLVM 22 in #12.
- Thomas Munro announced that he would push the fix in the final message, #13, then committed #7 unchanged as e5d99b4d9e with backpatches through PostgreSQL 14; the proposed always-inline removal was abandoned and no patch remained pending.