[openib-general] Re: [PATCH 13/13] [RFC] ipath Kconfig and Makefile
Adrian Bunk
bunk at stusta.de
Sat Dec 17 15:55:54 PST 2005
On Sat, Dec 17, 2005 at 02:54:44PM -0800, Robert Walsh wrote:
> > The driver shouldn't use assembler code and therefore no longer depend
> > on X86_64.
>
> Agreed about the assembler, but one way or the other, x86_64 is the only
> arch we support.
>...
There's a difference between "technically supported by the driver" and
"officially supported for our costumers":
It's fine if you tell the costumers buying your hardware "anything else
than 64bit x86_64 kernels is completely unsupported", but for getting
your driver included into the kernel it should be 32bit clean [1] and
should also work for people using 32bit kernels on an Opteron.
> > -O3 doesn't make much sense since the fight for producing the fastest
> > code is between -O2 and -Os.
>
> Makes many nanoseconds of difference to us for our latency numbers. At
> the low latency numbers we measuring (1.29us), this is a very important
> difference to our customers.
>...
There's no doubt that this is important for your customers.
What surprises me is that -O3 turned out to be the fastest flag for you.
Can you send numbers comparing -Os/-O2/-O3 (without -g3, preferable with
gcc 4.0) including a description what and how you are measuring?
> Regards,
> Robert.
cu
Adrian
[1] not long ago, it used to be the other way round that drivers weren't
64bit clean...
--
"Is there not promise of rain?" Ling Tan asked suddenly out
of the darkness. There had been need of rain for many days.
"Only a promise," Lao Er said.
Pearl S. Buck - Dragon Seed
More information about the general
mailing list