Re: Register VM WIP

[David Kastrup <dak@gnu.org>]

Mark H Weaver <mhw@netris.org> writes:

> I certainly agree that we should have a generous number of registers,
> but I suspect that the sweet spot for a VM is 256, because it enables
> more compact dispatching code in the VM, and yet is more than enough to
> allow a decent register allocator to generate good code.
>
> That's my educated guess anyway.  Feel free to prove me wrong :)

The counterproof will usually be done by benchmarking, and will even
differ between different processors sharing the same instruction set.

I see two ways out:
a) pick the register size individually for each function, as small as
possible without spillage.  Which makes the whole indistinguishable from
a stack-based VM.
b) don't generate the final bytecode until the code is actually being run.

That means that _if_ code is precompiled, it will be precompiled into
either stack-based VM or some other representation better suited to
compile into code for a certain amount of registers.  Of course, the
threshold to picking actual registers of the available processor and
compiling native code is then not all too large.

-- 
David Kastrup