Source code of a function

[cscherrer]

@dfdx

The tape can then be transformed and recompiled back to a normal function, so interpretation only happens once.

That’s really interesting. It seems like this with your Loop construct could be great for probabilistic programming. It would probably take some extension though. In PPL, for any loop that includes sampling, we’d need (or prefer anyway, I think) to be able to pick up at an arbitrary iteration. This could be something like a sequence of loop states stored as a StructArray.

@dfdx

IIUC, invokelatest(f, args...) tells Julia not to link to a specific version of f() during compilation, but instead go to the method table and get the latest version during runtime. Basically, it’s just a dynamic dispatch, and if you call it only once for the top-level function, its cost is relatively small.

That’s fine for some applications. But for me, for example rand needs to build code to sample from a model. This is usually in an inner loop, so it’s important that it only be built once for each type-distinct set of arguments. I think there are lots of other subtleties beyond this. Basically things got a lot better when I started avoiding eval when possible.