Re: Linux-2.0.11

Philippe Strauss (root@urbanet.ch)
Thu, 8 Aug 1996 01:45:11 +0200


Robert L. Krawitz writes:
> Date: Tue, 6 Aug 1996 21:01:28 +0200
> From: Philippe Strauss <root@urbanet.ch>
>
> Anyway, since the i386, the instruction set has only been very
> slightly augmented on each new ix86 (only some stuff to make SMP
> boards without too much pain on 586). The number of register has growed,
> thats the reason of the -m486 flag of gcc. There is then no obvious
> reason to make specific extensions of gcc on each stepping of brand
> new intel, AMD, Cyrix or whatever.
>
> The number of user registers on each chip is the same (as opposed to
> the model specific registers of the Pentium, which aren't general
> purpose registers). There are very definite reasons to have different
> optimizers in gcc for each chip. The Pentium, for example, has much
> more complicated pipelines than earlier processors, and choosing
> instructions that pair in the two pipes can improve performance quite
> a bit.
>
> The reason why the kernel config ask you about your CPU is to try to
> turn on the best optimization for your CPU, but in no way gcc (at time
> of this writing) will generate code only for P6 or AMD kX.
> You can play with -malign-jumps=2 -malign-loops=2 -malign-functions=2
> which align jumps/loops/functions on 64bits boundaries, the
> performance improvement is most of the time psychologic :*)
>
> Actually, -malign-foobar=2 aligns (jumps|loops|functions) to 4-byte
> (2^2) boundaries.

I sincerely apologize for having answered without having enough
knowledge of the subject.

What about the ability of gcc to handle optimization of grouping
instruction for processor like Alpha 21164A, pentium, ultraSparc or
whatever chip which does't include a reordering stage before feeding
the pipeline?

> --
> Robert Krawitz <rlk@tiac.net> http://www.tiac.net/users/rlk/
>
> Member of the League for Programming Freedom -- mail lpf@uunet.uu.net
> Tall Clubs International -- tci-request@aptinc.com or 1-800-521-2512
>

-- 
Philippe Strauss, CH-1092 Belmont
Email:	philippe.strauss@urbanet.ch

******* Caution! don't use reply without changing "root" to "philippe.strauss" ******* in my email addr. (sendmail userdb not yet well configured ;*)