model-writtenrun 3566 - may be wrong, links go to the source messages. Commit-by-commit LLM analysis is in the Commits dialog.
- Andres Freund proposed avoiding stack-frame setup in allocation hot paths by moving size and failure checks into allocator implementations and extracting non-inlined slow paths from AllocSetAlloc(), attaching a prototype and reporting preliminary performance gains in #1.
- David Rowley measured only a small improvement on his AMD system in #2; their follow-up profiling explored compiler and processor differences, with Andres identifying Ubuntu's enabled CET instrumentation in #7.
- After two years, Andres revived the work in #8, rebasing and polishing it into separate sibling-call and AllocSetAlloc() patches, with binary COPY benchmarks showing gains from both.
- David reported substantial allocation benchmark improvements in #9, while John Naylor's radix-tree tests in #10 showed reduced MemoryContextAlloc overhead from the sibling-call patch.
- David's detailed review in #11 requested clearer allocator API responsibilities, removal of a redundant Valgrind NULL check, and clarification of Slab size validation; he also discussed moving zeroing into allocators, which the submitted revisions did not implement.
- David took over preparation in #12, rebasing the sibling-call patch and supplying a separate documentation and cleanup fixup while postponing the AllocSetAlloc() rebase.
- Andres agreed to land the wrapper changes first in #13, explained that John's Slab-focused benchmark could not measure the AllocSet optimization, and suggested separately moving isReset updates into allocator slow paths.
- David restored both main patches in #14 and added an isReset experiment plus a benchmark helper, finding that Slab could avoid most flag writes, AllocSet could skip them on freelist allocations, and Generation offered no obvious opportunity.
- David committed the sibling-call component of #14 as 743112a2e9 with comment cleanup, then its AllocSetAlloc() component as 413c18401d with formatting, comment, and assertion refinements.
- After announcing those landings, David submitted #15, extending hot/cold separation to Generation and Slab and simplifying Generation block reuse by retaining an emptied current block and switching to an empty freeblock only when needed.
- David committed that extension with light cleanup as a0cd954480 and announced completion in #16; no eligible independent review of this extension appears before its commit.
- The thread ends with Andres's thanks in #17: all three main allocator changes landed, while the isReset experiment and benchmark helper from #14 are absent from these commits and have no stated final disposition.