IIRC numexpr got about x2-x5 faster when they switched from numpy's implementation (which naively mostly works as you describe, one op at a time, refcounted, th

IIRC numexpr got about x2-x5 faster when they switched from numpy's implementation (which naively mostly works as you describe, one op at a time, refcounted, though no good reuse) to a VM that reuses temporary result memory and works in 8KB pages (thus getting way, way better cache use). But I'm not familiar with an APL/J/K implementation that does.