This is the mail archive of the
gcc-patches@gcc.gnu.org
mailing list for the GCC project.
Re: [wwwdocs] gcc-4.0/changes.html - Mention SMS as new optimization.
- From: "Giovanni Bajo" <giovannibajo at libero dot it>
- To: "Mostafa Hagog" <MUSTAFA at il dot ibm dot com>
- Cc: <gcc-patches at gcc dot gnu dot org>
- Date: Wed, 16 Feb 2005 01:19:25 +0100
- Subject: Re: [wwwdocs] gcc-4.0/changes.html - Mention SMS as new optimization.
- References: <OFA2178CCE.9176A349-ON42256FA9.00467441-42256FA9.00483F28@il.ibm.com>
Mostafa Hagog <MUSTAFA@il.ibm.com> wrote:
>> - why SMS makes the code better (what does "better scheduling" mean)
>
> I cannot explain this without technical information, like interleaving
> loop
> iterations and so on. We can just say that SMS is intended to
> schedule instructions in loops, rather than the traditional
> scheduling that look at basic blocks without giving special handling
> for loops.
IMO, that's exactly the kind of information I expect. I'm sure that the
abstract of the PDF paper will tell me more, if I'm interested.
>> - which are the targets which are best served by SMS? Will a
>> Pentium1 gain the same benefit as a Pentium 4 from this? And a PPC?
>> What about an embedded target?
> I will try doing this for a general classification of machines like
> in-order
> vs. out-of-order machines, super-scalar and so on. mentioning
> specific machines
> makes the description complicated, and I will end-up giving technical
> description
> again.
It does not mean to be super-accurate. It should just give an idea about which
users should most be interested in trying it.
>> - what are the figures on what SMS can achieve, on at least a
>> coupleof target? Is it 10% or 1% or 50% of run-time speed? You
>> mentioned it is disabled because it is slow, so you should give some
>> hints on why users should still try to use it.
>
> OK, I will look in the results that I have and make a summary, but in
> bottom line this is not (yet) a win (on SPEC).
Nice. Saying something like "at this point, SMS is still preliminar, so don't
expect big improvements in code generation, but stay tuned for 4.1" would be
great (assuming it's the truth).
Thanks for doing this!
Giovanni Bajo