model-writtenrun 1429 - may be wrong, links go to the source messages. Commit-by-commit LLM analysis is in the Commits dialog.
- Thomas Munro opened the thread with two LLVM 22 compatibility patches, hoping to reach the February minor releases: use LLVM's repaired SectionMemoryManager and replace invalid lifetime.end annotations with poison stores, with optimization validation still needed #1.
- Thomas discovered that LLVM's existing C interface did not enable ReserveAlloc, so merely dropping PostgreSQL's memory-manager replacement would lose the ARM memory-layout fix; he pursued an upstream API addition #2.
- Matheus Alcantara tested the lifetime change and compared assembly with LLVM 21 and with poison stores disabled, reporting apparent improvements from the proposed replacement #3 #5.
- After the upstream API work succeeded, Thomas posted a revised series using ReserveAlloc and added unsigned casts to address a new LLVM integer-constant assertion #6.
- Anthonin Bonnefoy's initial low-optimization tests found no difference with lifetime annotations or poison stores #7; Thomas then withdrew the integer-constant component from the LLVM 22 work because upstream had reverted the triggering change, deferring it to LLVM 23 #8.
- Andres Freund explained that inlining and optimization were necessary and supplied a reproducer showing that optimization-pass ordering defeated poison-based store elimination #9; Anthonin subsequently demonstrated that poison stores did not reproduce lifetime.end's behavior under full optimization #10.
- Devrim GÜNDÜZ reported successful builds and regression tests with the patches on Fedora's LLVM 22.1.0 #11, while Tom Lane later reproduced the unpatched lifetime.end failure on Fedora LLVM 22.1.1 and urged progress #12.
- Thomas chose to disable lifetime.end for LLVM 22 without replacement poison stores, deferring stack-allocation research #13; this reduced version of the lifetime component from #6 landed as de6b80e5ff, and Thomas announced completion in #14.
- When Tom questioned the breadth of backpatching, Thomas confirmed local failures on every supported branch #15 #16; Devrim subsequently reported building all supported releases on Fedora 44 #18.
- Andres reported that restoring useful optimization required broader expression changes and suggested a specialized calling convention for common scalar functions #19.
- The SectionMemoryManager component of #6 landed separately as be21341e13, preserving its code while refining explanatory text and retaining the existing license commentary.
- The thread ended with both LLVM 22 changes committed and the integer-constant fix deferred; Thomas continued exploring stack-based argument storage and expression cacheability, without another patch submission #20 #21.