[aarch64-port-dev ] RFC: JEP - One AArch64 Port, Not Two

Ningsheng Jian ningsheng.jian at linaro.org
Mon Aug 27 08:49:34 UTC 2018


Hi,

On 27 August 2018 at 16:13, Andrew Haley <aph at redhat.com> wrote:
> On 08/27/2018 07:30 AM, Magnus Ihse Bursie wrote:
>> One question: I remember there were some guys from Linaro who compared
>> aarch64 vs arm64 in microbenchmarks, and found that while aarch64 had
>> the superior performance most of the time, there were some benchmarks
>> where arm64 was fastest.
>>
>> Has these been analyzed, and the performance on aarch64 brought in line
>> with arm64?
>
> I doubt that's possible. As we know, JIT heuristics can sometimes go one
> way and then the other, and it can be just a matter of luck. I haven't
> seen any optimizations in the arm64 port that were absent in the aarch64
> port. If anyone can find anything we'll port it over.
>
> Looking here
>
> https://docs.google.com/spreadsheets/d/18iklOrbaL67i46XHsPTrdTObWqUJJK8EO4-k0eqdLKM/edit#gid=909743165
>
> we see that tradebeans is faster on one implementation, slower on another.
> sunflow is spectacularly better on arm64, but it failed to converge most
> of the time on one implementation, so I don't believe the result.
>

Yes, some benchmark data is unstable. We did analyze the arm64 and
aarch64 codegen difference with microbenchmarks. AArch64 backend
generates better code on most cases, and I believe that we have
addressed those minor worse cases we found. E.g. JDK-8183533,
JDK-8185786, JDK-8187601.

Thanks,
Ningsheng


More information about the aarch64-port-dev mailing list