gcc 4.0 compiles kernel with altivec

Segher Boessenkool segher at kernel.crashing.org
Fri Feb 18 23:59:18 EST 2005

>> Last time I benchmarked it (on 970), it was best to test per group of
>> 4 to 8 registers (bigger was worse, smaller was worse).  This was an
>> artificial benchmark though, real usage patterns might skew the 
>> results
>> a bit.  Anyway, between 4 to 8 the graph was quite flat.
> Ah good !
> I'm curious how much registers are used by common gcc constructs using
> altivec implicitely (and possibly altivec-optimized memcpy & co that
> I'll put in the vdso eventually). That will probably be the key
> threshold.

Yes, real usage is the key.  Note though that the way GCC does
its register allocation, you tend to end up with a clump of
used vector registers at both ends of the register file, i.e.,
VRSAVE will always be something like  0xe800003f or such, so saving
per group of, say, 4 registers, doesn't really make you save much
too much.


More information about the Linuxppc64-dev mailing list