The Zig 0.14.0 release happened and did not address the single most important issue for the pkmn project which motivated me to take matters into my own hands. Breaking changes to the Zig language will soon make it untenable to keep supporting 0.11.0 to preserve performance, and a 20%+ regression is too large to ignore, so I wanted to figure out if I could come up with a minimal patch to the Zig compiler that hopefully will be easier to maintain. As it turns out, a simple change to less than 10 lines of the compiler’s code not only claws back the performance lost since 0.11.0, but surpasses it due to the other improvements made by LLVM in the intervening period. The effort to reward ratio here is huge, and while needing to patch the compiler is a pretty extreme step it feels more than justified1.
The other big change that happened this month was Pokémon Showdown overhauling its Generation I implementation of binding moves. This will entail major changes to the engine, hopefully all strictly for the better in terms of simplification, though more realistically there may be some new weirdness and bugs to workaround. It’s positive to see this issue finally being addressed, regardless of the amount of work it will be to match in the engine. This work, coupled with resolving the division-by-zero glitch issues which constantly plague the fuzz tests are on deck to be handled next month (and perhaps the momentum of resolving those will also help motivate a push towards working on Haze, the final mechanic remaining to finish -Dchance and -Dcalc suppport).
The engine’s calculator/solver demo is somewhat blocked on not having a clear vision of what I feel a calculator should be. Perhaps a good goal for this coming month will be to work on (finish?) the “Damage Calculators From First Principles” post that I’ve got stashed away as a draft…
— pre
-
“Why can’t Zig just merge this patch upstream?” – the patch reverts logic which is crucial for ensuring undefined bits of sub-byte integers don’t get read. Given that the engine makes heavy use of sub-byte integers to maximize cache line space, additional operations around each and every load greatly slows down the code. The engine’s extensive test suite allows us to be confident that it doesn’t actually ever read the undefined bits the loading logic is meant to guard against which means its safe for the engine to ignore, even if in general it is required to ensure correctness of Zig programs in other domains.↩︎