On Mon, 18 Jun 2018 at 14:54, Ramana Radhakrishnan ramana.radhakrishnan@arm.com wrote:
I don't think the make bench discussion is totally relevant in this case. We'd like to see what the performance of this stuff is and it's always goign to be a battle to get things benchmarked perfectly on hardware and there is unlikely to be one implementation in the Arm world
+1, I don't think there's a risk in pushing these patches without a benchmark result. When implementations appear we can run benchmarks and tweak and choose accordingly. In the worst case if we find that *everyone* botched up their SVE implementations then we can just drop that code in some years :)
unlike other architectures. It's under an ifunc and gives folks something to reason with, if we are really that concerned about it's performance today lets put it under a tunable that allows us to turn it off easily ?
Existing tunables can do the job just fine:
- glibc.tune.cpu can be used to emulate a different processor if that's suitable for the core - glibc.tune.hwcap_mask can be used to mask out HWCAP_SVE completely
Siddhesh