Even better news, however, is that I've finally reached a point where Basic Block Versioning (BBV), the compiler optimization technique I've been implementing in Higgs as part of my PhD thesis, finally delivers measurable speedups on benchmark programs. Many optimizations are required in order to make dynamic languages fast, and the quality of the code generated by Higgs was just too poor: removing 50% of type tests or more didn't make a noticeable performance impact. I was extremely surprised to find out that on loop microbenchmarks, the naive register allocator produced code that ran more than 10 times faster than similar code loading and storing values on the stack.