Brrt (Bart Wiegmans) is untangling the intricacies of floating point
expressions so close to the metal in his [MoarVM JIT Compiler
provides some of his insights [in his recent blog
post](http://brrt-to-the-future.blogspot.com/2019/01/a-short-posts-about-types-and.html). Here is his brief report (edited slightly by me):
> Last month, I've been busy with floating point support. Getting that ready involved two things:
> * Getting the type system for the expression template precompiler ready. This works now, I'm happy with the result.
> * Fixing the register allocator to support it. This works as well, but isn't fully tested.
> * Ensured that we convert the register to x86-compatible form in the register allocator rather than in the emitter, which means that we can often select better registers.
> The next thing on my list is irregular instruction handling.