> Dyalog APL's source code is about 500KLOC according to Dyalog's staff. GNU APL is over 100KLOC per its homepage. April's .lisp files currently stand at 6350LOC.
A ways into the video it says the core Dyalog interpreter is about 100KLOC, with the rest being GUI code and interop such as file handling and formats, cross-platform support, .NET interop, and so on. If the 100k figure surprises you then what you are missing is just how deep array optimization is. For example, searching through an array will use an SSSE3 table lookup for 1-byte elements, a different table for 2-byte elements, and several different kinds of hash tables for other types. I tried to give a sense of the scale of this sort of thing in the first slide of my Dyalog '19 talk on Reductions (https://www.youtube.com/watch?v=TqmpSP8Knvg). You can write pages and pages of code and still there will be a way to speed up a common operation substantially.
By working on multidimensional arrays, APL expands the range of computations that are accessible to optimization. I think the writer of a scalar compiler like C would generally be happy making a single loop as fast as possible. APL immediately makes the implementer think about what happens if there are many axes (each of which is equivalent to a nested loop) even if the later axes (inner loops) are small.
-April implements the language and just punts to CL for all the I/O and system level stuff. Like it doesn't have to implement a REPL, or a line editor or APL workspace files/array stores.
Thanks. I'm guessing (based on my knowledge of J and all its optimizations/special forms) that Dyalog APL would still be more performant too? I didn't see it mentioned on the GH README.
By working on multidimensional arrays, APL expands the range of computations that are accessible to optimization. I think the writer of a scalar compiler like C would generally be happy making a single loop as fast as possible. APL immediately makes the implementer think about what happens if there are many axes (each of which is equivalent to a nested loop) even if the later axes (inner loops) are small.