Appreciate the reference. Will look into it in the future. The Wikipedia summary made sense but the recompile part doesnt bother me as much as others. My preferrence is to make compiles fast as possible with good incremental compilation and caching. Ive even pushed for trying to FPGA-accelerate compiler passes if possible.
My solution allows incremental compilation. It is just a type system hack, and it does not even involve backtracking (my second favourite compilation technique), just a two pass typing instead of a single pass.
As for the compilation speed, the Nanpass-like approach is very parallelisable and there is a lot of space for nice optimisations (like pass fusion).
IIRC, you are into both tech like ML and hardware. I know of LISP machines & even a G-machine (Haskell) emulator on a Forth chip. However, given verification and non-Turing push, I was looking for a CPU to run algorithms written in Standard ML. Found one and thought you'd enjoy reading it too:
The older versions of Bigloo used to contain an ML compiler built on top of Scheme (they dropped it later for some reason), so the concept is not new.