A while ago I sped up tree-sitter-haskell by 50x. I’m going to go through the process of discovering and fixing the worst bottleneck.

In the last post, we wrote a tiny C program, compiled it, disassembled it, then tweaked that assembly to make it six times faster. Now we’re going to beat it.

Sometimes humans can spot optimization opportunities that a compiler can’t doesn’t. In this post, we start with a loop generated from C code by clang, and tweak it in various ways, measuring the speedup.