Re: cpuflags (optimal compiler flags for building)

[David Brownlee <abs%NetBSD.org@localhost>]

On Fri, 5 Dec 2008, Thomas E. Spanjaard wrote:

> Thomas E. Spanjaard wrote:
> > It recommends '-march=pentium3' on a NetBSD/amd64 5.99.01 box (GCC
> > 4.1.3), with an Intel Core 2 Quad CPU at 3GHz (Kentsfield core). The
> > best fit for GCC 4.1.3 would be '-march=nocona', and less so
> > '-march=pentium-m'. Newer GCC versions support keywords for the Core 2,
> > but I haven't checked whether your script checks the GCC version and
> > keeps a list of supported options.
>
> I just read the source, and I see that you use '-march=pentium3' as
> fallback for '-march=core2' for older GCC versions; I'd suggest using
> '-march=nocona' instead, because it flags the target as supporting SSE2
> and SSE3. Instruction reordering by the compiler would theoretically
> benefit most from '-march=pentium-m' on older GCCs, but that lacks SSE3
> and x86-64. One snag could be that when using any of the Netburst types
> makes the compiler avoid things like ro{r,l}, but I have no clue as to
> how GCC really behaves with these flags.

        From cpuflags-1.22 it should automatically set -msse3 based
        on the features flags reported by cpuctl or from dmesg, so
        the fallback should be based on scheduling rather than SSE
        support?

        Based on that core2 should probably fall back to pentium-m first,
        and then pentium-m to pentium3 - I'll adjust the logic - thanks

--
                David/absolute       -- www.NetBSD.org: No hype required --