model-writtenrun 5144 - may be wrong, links go to the source messages. Commit-by-commit LLM analysis is in the Commits dialog.
- Peter Eisentraut proposed replacing Datum conversion macros with inline functions to catch incorrect argument types, separating existing misuse fixes into a prerequisite patch and leaving some fmgr.h varlena macros unchanged because they already reached type-checking functions #1.
- Aleksander Alekseev approved the series and resent it with attribution and a discussion link #2; Tom Lane withdrew his initial ABI objection but opposed backpatching on risk grounds, while Robert Haas agreed that the conversion should initially remain development-only and both distinguished actual bug fixes from cosmetic cleanup #3 #5 #6 #8 #9.
- Aleksander added initializations in GiST, text search, and intarray code to silence compiler warnings exposed by the conversion, without changing its design #10 #11.
- Peter committed the prerequisite misuse fixes unchanged from the first patch of #11 as 6bcda4a721, leaving the inline conversion pending.
- Peter questioned whether PointerGetDatum's const argument misled the compiler about output parameters and proposed removing const, adding a nonconst variant, or retaining the initializations #12; Aleksander identified Clang 11 as the warning source and favored retaining the four initializations, which remained in the final code #13.
- After the prerequisite landed, Aleksander submitted the rebased conversion remainder #14, which Peter committed with only four comment edits as 595836e99b and announced in #15.
- Julien Rouhaud reported buildfarm failures immediately after the conversion landed #16; Peter reproduced problems on 32-bit platforms and reverted the conversion pending investigation #17.
- Two weeks later Peter requested help after becoming stuck diagnosing the failures #18, and Tom identified that the inline Fast conversion functions returned addresses of their own parameters on pass-by-reference platforms #19.
- Tom reported that restoring the two Fast macros passed check-world in a 32-bit VM and submitted the original conversion plus a repair using AssertVariableIsOfTypeMacro, preserving the caller's variable address while checking its type on both platform variants #19.
- Peter applied that complete repaired submission without code changes as c8b2ef05f4 and confirmed the landing, noting that the stricter assertions were suitable for these limited uses #20.
- The thread ended with the cleanup and corrected conversion landed; Tom noted that a UInt64GetDatumFast variant could be added if a caller needed it, but no current use required one #21.