Re: Register Allocation

[Andrew MacLeod <amacleod@redhat.com>]

On Fri, 2005-11-18 at 10:53 +0100, Giovanni Bajo wrote:
> Andrew MacLeod <amacleod@redhat.com> wrote:

> 1) Do you believe there will be sub-parts of this project which could be
> carried on succesfully and efficiently by programmers without previous RTL
> experience? IIUC, the optimizers will be basically abstracted by RTL details,
> but I was thinking of something within the critical path.
> 

Some aspects require a little less intimate knowledge of RTL. I think
the insn annotation and operand summary could be done with just basic
RTL understanding, and then the live range code, if it were built on top
of that, would require very little RTL knowledge. 

The RTL library, insn selector, and insn rewriter will be the most
knowledge intensive.  The degree of knowledge required for the rest of
the critical path goes down a bit to where a more basic knowledge should
be all that is necessary. 

> 2) As for the new tables needed by the RTL library, I suppose they will be
> generated by some new gen* program. Did you consider using a scripting language
> as a fast prototype to munge .md files and generate those tables? I believe it
> would allow faster initial development and more flexibility in changes. Much
> later, it can be rewritten in C.
> 

Thats quite possible, I hadn't really thought about that too much.  I
was actually thinking about looking into commandeering one or more of
the gen programs and having them generate the tables at the same time as
what they currently generate. That might be a bad idea, as I have never
really looked at them.  There is a relationship between the assembler
table and the insn alternative table for instance, so I figured it might
be possible to enhance it for what I want instead of doing something
from scratch.

If working with an existing gen program turns out to not be a good idea,
perhaps a scripting language might be helpful, but I don't have much
experience with any of them. 

Andrew