model-writtenrun 7590 - may be wrong, links go to the source messages. Commit-by-commit LLM analysis is in the Commits dialog.
Jul - Oct 2018 · 5 entries
- Fabien COELHO proposed a seeded permutation function for pgbench in #1, intended to scatter non-uniformly selected keys without collisions or unwanted locality, using inexpensive rounds of XOR and prime-based transformations.
- Hironobu SUZUKI demonstrated that multiplication overflow destroyed the permutation property for large sizes and supplied modular multiplication in #2; Fabien incorporated the fix, smaller primes, shortcuts, and large-input tests in #4, explicitly recognizing Hironobu as a co-author.
- Builtin optimization and portability discussion with Thomas Munro led from popcount support in #5 to leading-zero counting in #9, followed by a test-conflict rebase in #13.
- Hironobu and Fabien refined arithmetic handling across #14, #15, #16, and #17, retaining intentional unsigned LCG wraparound, removing an unnecessary addition guard, and adding 63-bit tests.
- Alvaro Herrera's formatting feedback in #20 and #22 produced the indentation and protected algorithm comments in #21 and #24.
Feb - Mar 2019 · 1 entry
- After several months, Hironobu refreshed the patch and expanded its mathematical explanations in #26; Andres Freund then requested reuse of existing bit utilities in #27, which Fabien implemented in #29.
May 2019 · 1 entry
- Fabien added an int128 multiplication shortcut in #32 and incorporated Thomas Munro's wording corrections from #33 into #34, while retaining the manual fallback.
Jan - Feb 2020 · 1 entry
- Peter Eisentraut and Alvaro Herrera questioned whether readers could understand the benchmark benefit in #37 and #38; Fabien explained the locality bias in #39 and added a warning and concrete permutation examples in #40.
Apr 2020 · 1 entry
- Build problems and doubts about consensus threatened the patch's continuation in #41, but Alvaro argued to retain it in #43; the refreshed patch in #46 removed an unfinished multiplication experiment and adapted the documentation, with an XML correction in #47.
Oct 2020 · 1 entry
- Muhammad Usama reported successful feature and installcheck-world testing in #48, but the patch remained uncommitted until renewed discussion in March 2021 brought Dean Rasheed into the review through #52.
Mar - Apr 2021 · 9 entries
- Dean questioned the bespoke algorithm and proposed an indexed random sequence in #53, while David Bowen suggested a conventional shuffle in #54; constant-time permutation requirements and memory costs ruled out those alternatives in #55 and #56.
- Thomas Munro and Alvaro supported simplifying the arithmetic in #57 and #58, leading Fabien to remove the manual multiplication fallback in #59 while preserving smaller-input operation on platforms without int128.
- Alvaro's detailed review in #60 led to the permute() name, removal of unused code, terminology cleanup, and platform-specific test handling in #61 and #62; broader expression-evaluation refactoring was deferred.
- Dean's statistical experiments in #65 and #66 exposed non-uniform permutation choices, particularly for power-of-two sizes and even seeds, prompting his replacement algorithm in #66 with random odd multipliers, XOR, rotations, and offsets that also eliminated the int128 requirement.
- Fabien identified a path that could avoid the half transformations in #67, and Dean removed the intervening full-range offset and simplified rotation and mask calculations in #71.
- Fabien objected that a single 48-bit random state ignored part of the supplied seed in #70, so Dean introduced two random states using all seed bits in #73.
- Fabien proposed integer-only random generation in #75, but Dean identified biased bit construction, added cost, and overflow that could break bijectivity in #76; Fabien supplied a revised experimental helper and bounded offset in #77.
- Dean committed 6b258e3d68 with the final patch structure from #77, replacing its experimental helper with getrand(), increasing the rounds to six based on testing described in #76, and updating tests and comments.
- Dean confirmed the landing and explained the getrand() choice in #78, and Fabien accepted that compromise in #79; the feature was complete, with broader random-generator modernization left for separate work.