> One other strangeness, I ran all the benchmarks comparing the current > stable MLton and the (just checked in) RSSA version, both using the C > codegen. Weirdly enough, the RSSA version was noticeable faster in > most cases. I have no explanation. Here are the timings. You mentioned earlier that there is a new pseudo-register allocator. That might be enough for gcc to a little more (or less) with the code.