[aarch64-port-dev ] aarch64 port "review"

[Stuart Monteith]

Hello,
   Ed's given me the task of measuring the effects of prefetching,
using the following options:

-XX:PrefetchCopyIntervalInBytes, -XX:PrefetchScanIntervalInBytes,
-XX:AllocatePrefetchStyle, -XX:AllocatePrefetchDistance,
-XX:AllocatePrefetchStepSize

This isn't quite what you are discussing, but I'll post the results
once I have something in case they are informative.

Regards,
    Stuart


On 9 May 2016 at 18:32, Jasty, Ananth <Ananth.Jasty at cavium.com> wrote:
> Gcc does prefetching to great benefit, largely on loops. It also does profile based code opt if that's what you mean but my results have been less impressive on aarch64.
>
> I tried adding prefetching support to macro and template assembler, with what seemed like very solid results but it'll require a more rigorous test. Ed has my poc patch, I'll try to write a more limited, focused patch this week.
>
> Ananth
>
>> On May 9, 2016, at 5:52 AM, Roland Westrelin <rwestrel at redhat.com> wrote:
>>
>>
>> Hi Ananth,
>>
>> Thanks for your comments.
>>
>>> So far the codegen of the aarch64 JIT looks very strong, no major
>>> pitfalls to be seen. However, given the nature of aarch64 cores, we
>>> are seeing a greater need for predictive pre-fetching, both in
>>> compiled code, and in the runtime itself (GC is particularly
>>> sensitive).
>>
>> C2 has very little support for prefetching. I suppose it's been tried on
>> other platforms but I don't know what the history is. Do you know if
>> other compilers (gcc for instance) successfully take advantage of
>> profiling on arm?
>>
>> Roland.