pentium?

pentium?

[Neal Becker]

Any plans to incorporate pentium optimizations into egcs?

Re: pentium?

[Jeffrey A Law]

  In message < u9zpouvri1.fsf@neal.ctd.comsat.com >you write:
  > Any plans to incorporate pentium optimizations into egcs?
Are you referring to the pgcc stuff?

Yes, but it's going to take time -- most of the stuff intel wrote
that became pgcc was poorly written, sometimes broken for the x86
and often wrong for other machines.

Much/most of it will need to be rewritten.

I'm casually looking at some of the loop stuff in my copious
spare time :-)  Some of their loop changes are _much_ more
complicated than they need to be, others have been made either
mostly or completely obsolete by changes already in gcc.  Still
others will have to be rewritten.

We also have a pretty good idea how to replace all that ugly riscify
code, but nobody's stepped up to do the work.  If anyone is interested,
they can contact me and I'll describe a cleaner solution which re-uses
gcc's instruction splitting code.

A lot of their reload improvements are becomming obsolete as
reload_cse gets better.  Once we pick up the latest reload_cse stuff
from gcc2 I think the only serious thing we'll be missing is hoisting
of spills out of loops -- and that may be made obsolete by the
live range splitting code Cygnus is working on.

So, the summary is I think we're slowly addressing the kinds of
things pgcc was trying to optimize, but potentially in different
ways -- hopefully in ways that will benefit more ports than just
the pentium.

Jeff

Re: pentium?

[Neal Becker]

On a related subject - we use also use i960 here.  I have to use the
intel version, because they have added a number of features.
Unfortunately, their version is pretty out-of-date, and I'm not sure
they're maintaining it.  Any plans to incorporate any of intel's i960
improvements?

Re: pentium?

[Joel Sherrill]

On Wed, 1 Oct 1997, Neal Becker wrote:

> On a related subject - we use also use i960 here.  I have to use the
> intel version, because they have added a number of features.
> Unfortunately, their version is pretty out-of-date, and I'm not sure
> they're maintaining it.  Any plans to incorporate any of intel's i960
> improvements?

Specifically what features are missing in egcs?

--joel

Re: pentium?

[Neal Becker]

>>>>> "Joel" == Joel Sherrill <joel@OARcorp.com> writes:

    Joel> On Wed, 1 Oct 1997, Neal Becker wrote:

    >> On a related subject - we use also use i960 here.  I have to use the
    >> intel version, because they have added a number of features.
    >> Unfortunately, their version is pretty out-of-date, and I'm not sure
    >> they're maintaining it.  Any plans to incorporate any of intel's i960
    >> improvements?

    Joel> Specifically what features are missing in egcs?

-mlong-calls
#pragma interrupt

various optimizations, judgeing from intel's descriptions - including
a number of i960 model dependent optimizations.

Re: intel 960 support Was: pentium?

[Robert Lipe]

Joel Sherrill wrote:
> On Wed, 1 Oct 1997, Neal Becker wrote:
> 
> > On a related subject - we use also use i960 here.  I have to use the
> > intel version, because they have added a number of features.
> > Unfortunately, their version is pretty out-of-date, and I'm not sure
> > they're maintaining it.  Any plans to incorporate any of intel's i960
> > improvements?

Intel released version a major overhaul of the tools about a year or 
so ago.  By that time, my involvement with i960 was long over.

I recall in late '95, perhaps early '96 doing a big shootout between the
Cygnus progressive compiler of that era and the Intel tools (which I recently
took great delight in removing, so I can't tell you the version number, but
it was < 4.0).    It was about a draw.   The Intel version contained some
chip-specific things for the JK and JA and could therefore dhrystone about
2-3% better.  However, the Cygnus version would actually work given any
non-trivial input. :-)

> Specifically what features are missing in egcs?

Gosh, I can't recall.   There were (obviously) chip-specific 
optimizations present that the Cygnus/FSF version didn't have.
It knew more about the pipelines.   The linker knew how to do 
jump to a jump sorts of optimizations.    Nothing earthshattering
that I recall.

The thing I do recall was the that the Intel version wasn't very
reliable and was always having to look over its shoulder.

RJL

Re: intel 960 support Was: pentium?

[Jim Wilson]

Yes, this is certainly one of the items on our list of things to look at
when we find the time.

However, the intel i960 support has many of the same issues as the pentium
support.  There are interesting ideas there, but many of them have been
implemented poorly, or implemented in such a way that they will work for the
i960, but will break other targets, and hence can't be used as is.  There is
also the issue of copyright assignments.  We can't use Intel's code without a
copyright assignment from Intel.  I am not sure how hard it will be to get one;
it might be easier to just write new code from scratch in some cases.

Jim