On Wed, 2011-05-04 at 18:38 -0400, Chris Tyler wrote: > On Wed, 2011-05-04 at 14:46 -0400, Jon Masters wrote: > > I think we should build for ARM (as opposed to Thumb2) but we should > > support interworking with Thumb2 code through the toolchain options. We > > should then later consider implementing some Thumb2 optimization. It's > > more armv7thl, but the (t) is implied since it's ARMv7 anyway. > > The thumb2 situation needs a bit more investigation. I spoke with some folks this evening about this, and other optimization criteria. Apparently, we can expect up to a 30% improvement from Thumb2 but the reality is more like 18-24% for high cache-hit/non-data heavy cases where we could really get an improvement. However, it depends on the micro-architecture in use. We'll possibly see great improvements in Thumb2 performance come A15 but may not notice them on A8 and A9MP. I'm leaning toward... > I think some comprehensive testing is needed. And I agree. Do you have someone looking into Thumb2 some more? As to VFPv3, I'm coming around to the idea of settling on VFPv3-D16 since it seems to be the lowest common denominator on v7. We could do with some comparisons in terms of performance between the two AAPCS variants, if someone is willing to do the investigation. And of course it would really vary from application to application somewhat. My current revised thought, until proven otherwise: *). Little endian (obviously, but worth stating) (l) *). Cortex-A8 or higher fully compliant core(s) *). ARM VFP3 hardware floating point (h) (D16 though) *). Not required, but some optimized libs for ARM NEON Architecture *). Thumb2 interworking support but not built for Thumb2 Jon. _______________________________________________ arm mailing list arm@xxxxxxxxxxxxxxxxxxxxxxx https://admin.fedoraproject.org/mailman/listinfo/arm
