[tathougies]

Compiling to machine code is not a panacea for optimization. A optimized JIT compiler is going to blow an AOT compiler out of the water. Being smart about the machine code generated is significantly more important than generating machine code. In particular, PyPy makes several optimizations over python code that a more direct implementation of CPython at the machine level probably wouldn't. For example, PyPy erases dictionary lookups for object member access if the object shape is statically known. Given how prevalent this kind of lookup is in Python code, it's possible that even an interpreter that made this optimization would be faster than a machine code version that used an actual hash table.

I think this compiler also makes this particular optimization, but this is just one of many many optimizations PyPy does. I imagine that with sufficient work, this compiler could be brought up to speed with PyPy, but as it stands right now, PyPy simply benefits from having years of optimization work that a new project doesn't.