Re: Register Pressure in Instruction Level Parallelism (fwd)

[Michael Matz <matzmich at cs dot tu-berlin dot de>]

Hi,

On Sun, 14 Jul 2002, Pop Sébastian wrote:

> Yes, but a major part of Sid's thesis is to provide efficient heuristics
> for these hard to solve problems:

What an interesting thread can result from a statement I made without
actually reading the paper, which I explicitely mentioned ;)  Yes, I was
referring to the slowness of integer programming when applied to larger
problems.  If you only apply it to small regions you again end up with
suboptimal solutions, which the exact algorithms were supposed to solve.
Then you anyway can also use other heuristics which don't need a complete
IP solver, which we would need to incorporate in gcc.  Remember we're
talking about register allocation, which _is_ global.  Applying IP to
other things (scheduling e.g.) might be usefull, I don't know.

But I not only had the slowness in mind.  We would need to write or
incorporate an IP solver, preferrably a good one, as simple-minded IP
solving _is_ slow as hell.  And we would need to write the IP expressions
which should be solved, probably slightly other for each platform, i.e.
much infrastructure change.  I.e. I think IP would currently be a big
task.


Ciao,
Michael.