Really cool work. The upshot of it is that SIMD optimizations can work within traced numpy calls. This is preferable to calling out to an optimized subroutine,

Really cool work. The upshot of it is that SIMD optimizations can work within traced numpy calls. This is preferable to calling out to an optimized subroutine, because the first lets you keep your registers intact (and your trace) and the last does not. I wonder what issues prevent them from tracing the happy-case list yet.