oqed

oqed is an LLM-guided superoptimizer for x86-64. You point it at a C or C++ project and it does the rest: - builds the project and runs its workload; - profiles it to find the hot functions; - has Meta's Muse models write faster assembly for them; - keeps only what is correct and beats clang -O3 and gcc -O3. Then you build with CC=oqed-cc and your source doesn't change. oqed-cc re-checks every replacement against the current source on each build and falls back to the compiler's output if anything changed. Most of what the models write is wrong, so most of oqed is the checker. Every candidate has to: - assemble; - follow the calling convention; - pass differential testing and the project's own tests; - beat both compilers in an interleaved, paired benchmark. Small functions also get a symbolic-execution + Z3 proof that they're equivalent for every possible input. Results on real open-source projects, whole program, output identical on every run: - protobuf-c 1.27x (clang) / 1.32x (gcc) - zlib 1.17x / 1.15x - cJSON 1.07x - LevelDB 1.05x (clang) Single functions are proven equivalent and run up to 6.5x faster (18 instructions → one lzcnt). There's a live UI with every real run, including search replays, failure reasons and proofs: ignaciosearles.github.io/oqed