Re: [Lightning] Argument Management

[Paolo Bonzini]

Sandro Magi wrote:
> On Wed, Feb 20, 2008 at 1:50 AM, Paolo Bonzini <address@hidden> wrote:
> >  > Does this mean that I can't write a VM which performs interprocedural
> >  > register allocation using GNU Lightning? Or is this statement intended
> >  > to apply only when using the C call convention?
> >
> >  It's mostly because of the SPARC, which uses register windows.  If you
> >  don't care about it, you can pass parameters using GPR.
>
> I don't want to break compatibility unnecessarily. I'll read up on Sparc:
> http://www.sics.se/~psm/sparcstack.html
>
> Perhaps an additional macro to be defined by each arch is warranted.
> Something like: JIT_NEEDS_PROLOG. If false, I can use a different
> calling convention.

There is actually jit_leaf, which would be what you need.  But it's not 
implemented for the SPARC (it generates the same prolog) and I don't 
have access to SPARCs anymore.

> I've considered CPS, particularly in conjunction with the Cheney on
> the MTA-style GC. Unfortunately, CPS just kills the pipeline due to
> the indirect function calls. Given how deep pipelines are nowadays,
> and the fact that they're getting deeper, I'm reluctant to go this
> way.

Do you have so many function calls?

Paolo