model-writtenrun 3330 - may be wrong, links go to the source messages. Commit-by-commit LLM analysis is in the Commits dialog.
- Pavel Stehule reported regression-test crashes after Fedora 40 upgraded to LLVM 18.1 in #1, confirmed that fresh master still failed in #3, and Thomas Munro reproduced the failure on FreeBSD and traced it to LLVM's inliner in #5 and #6.
- Dmitry Dolgov isolated the failure to generated deform functions returning void in #7, identified a triggering LLVM change and filed an upstream issue in #8, then attached a dummy-return workaround in #10.
- Thomas initially preferred waiting for the proposed upstream LLVM fix because its release could precede PostgreSQL's next release, avoiding the workaround, as explained in #11.
- Thomas identified PostgreSQL's copying of return-value attributes from a Datum-returning template onto void-returning deform functions and supplied an inline guard in #12; Dmitry confirmed that it fixed the crash in #13, superseding the dummy-return approach.
- Dmitry demonstrated that LLVM verification would have caught the invalid noundef attribute in #14, then proposed a jit_verify_bitcode GUC in #15, which also exposed an incorrectly mangled lifetime intrinsic name.
- Thomas combined the attribute fix with a separate verifier patch in #16, replacing the proposed GUC with automatic verification in assertion builds and correcting the intrinsic name.
- Dmitry measured increased optimization time but considered it potentially acceptable for assertion builds, and proposed using LLVMPassBuilderOptionsSetVerifyEach instead of a pipeline prefix in #17.
- The attribute patch from #16 landed with unchanged logic as development-branch commit 53c8d6c9f1 and its separately listed stable-branch counterpart bf1cfe77ed; Thomas announced the fix in #18, and Pavel confirmed all Fedora 40 tests passed in #19.
- Verification remained pending while Thomas questioned CI overhead and compatibility of the intrinsic rename with older LLVM versions in #18; Dmitry reported only a few percent slowdown in local installcheck in #20, and Andres Freund supported verification while warning that mangling rules had changed historically in #21.
- After Robert Haas closed the original bug's commitfest entry in #22, Dmitry reposted the remaining verifier patch in #23, retaining Thomas's authorship, incorporating the verification API suggestion, and specifying master-only installation without backpatching.
- Thomas committed the remaining patch unchanged as 8583b1f993, enabling assertion-build verification for LLVM 17 and later and fixing the intrinsic name, then announced completion and closure in #24.