charlie cummings

now that i implemented nun boxing, and values fit in one register, i was able to also implement copy&patch compilation. with some additional optimizations im now finally faster than luajit's interpreter-only mode on a few benchmarks as well 🥳

| Command | Mean [s] | Min [s] | Max [s] | Relative | |:---|---:|---:|---:|---:| | lua5.1 bench.lua -- lua_benchmarking/benchmarks/life/bench 5000 \|\| true | 1.967 ± 0.021 | 1.947 | 2.008 | 2.86 ± 0.04 | | luajit -joff bench.lua -- lua_benchmarking/benchmarks/life/bench 5000 | 1.114 ± 0.012 | 1.103 | 1.139 | 1.62 ± 0.02 | | luajit bench.lua -- lua_benchmarking/benchmarks/life/bench 5000 | 0.687 ± 0.004 | 0.682 | 0.694 | 1.00 | | ./target/interpreter/release/bench life.bin 5000 | 2.689 ± 0.036 | 2.659 | 2.783 | 3.91 ± 0.06 | | ./target/release/bench life.bin 5000 | 1.068 ± 0.003 | 1.063 | 1.073 | 1.56 ± 0.01 | | ./target/unsafe/bench life.bin 5000 | 0.939 ± 0.005 | 0.933 | 0.946 | 1.37 ± 0.01 |

its so messed up how fast luajit is man. im beating luau and luajit's interpreter-only mode but still getting absolutely crushed by the real jit

@bradlarsen@infosec.exchange caveat emptor im just doing this for fun as a side project so it has a bunch of bugs, doesn't properly implement a bunch of things including metatables, etc

@bradlarsen@infosec.exchange i have a repo up at https://github.com/chc4/lunacy and a previous blog post at https://redvice.org/2026/lunacy/ - the most recent set of commits are kinda claude-slop quality that i have to clean up, and i haven't pushed the copy&patch part yet for the same reason though.

@bradlarsen@infosec.exchange im very happy with the approach i took still, the generator-based lbbv approach and simple template jit over that has made it extremely easy to implement and extend

i gotta write a new blog post outlining all the improvements since the last one. i have a cute graph with benchmark times per commit which is fun to look at, and also nicely illustrates which optimizations were actually an important feature vs minor twiddling

https://redvice.org/2026/coverage-the-wrong-way/ wrote a blogpost about using kvm to collect code coverage for blackbox binaries.