model-writtenrun 9740 - may be wrong, links go to the source messages. Commit-by-commit LLM analysis is in the Commits dialog.
- Andres Freund proposed variable-length function-call information using adjacent Datum/nullness pairs, reducing a typical two-argument call structure from 936 to 64 bytes and preparing for better JIT optimization; the initial prototype left LLVM support incomplete #1.
- Serge Rielau shared an alternative using fixed and variable structures with argument pointers, but Andres preferred the flexible-array design to avoid extra indirection and allocations #2 #3; broader conversion of utility-function return values was deferred after Peter Eisentraut raised ABI and performance questions #5 #6.
- After four months, Andres posted a rebased revision with STACK_FCINFO_FOR_ARGS to hide stack-allocation boilerplate and a separate experimental patch replacing duplicated fmgr wrappers with macros #7.
- Andrew Gierth opposed backpatching compatibility accessors because extensions would still need to support older minor releases #9; Tom Lane agreed, found incorrect pointer-array padding, proposed LOCAL_FCINFO and a deliberate struct rename, requested documentation of the two aggregate null-check representations, and rejected the wrapper macros over runtime cost #10.
- Andres accepted the review and abandoned the wrapper-macro experiment #11; his next patch introduced FunctionCallInfoBaseData and LOCAL_FCINFO, corrected undersized allocations using FUNC_MAX_ARGS where necessary, completed LLVM support, and documented why tuple-slot and function-argument null checks still needed separate representations #12.
- Andres committed #12 as a9c35cf85c with light polishing, chiefly renaming NullableDatum.datum to value, rearranging declarations, and cleaning comments and initialization; he announced the landing in #13.
- After landing, Andres reported a buildfarm failure with an empty stage log #13; Tom Lane and Andrew Dunstan noted similar intermittent failures across machines and branches, leaving the apparent failure unconnected to this patch #14 #15.
- Greg Stark subsequently suggested a null bitmap followed by a Datum array #16, but Andres defended NullableDatum because bitmap access costs more instructions, cannot guarantee shared cache lines, and does not support existing bool-pointer interfaces #17.
- The thread ended with the main change landed and the wrapper-macro experiment abandoned; Andrew Dunstan traced the empty-log problem to a buildfarm-client logic bug, fixed it, and announced an upcoming client release #18.