[aarch64-port-dev ] [16] RFR(S): 8251525: AARCH64: Faster Math.signum(fp)

[Dmitry Chuyko]

Sorry, I really should have used dev submit. Vladimir, thanks for 
immediate fix. It looks like the test was excluded in our CI for 16, 
will inspect that.

-Dmitry

On 9/3/20 10:56 PM, Vladimir Kozlov wrote:
> Please, run changes through jdk/submit before push. Current fix caused 
> tier1 failure:
>
> https://bugs.openjdk.java.net/browse/JDK-8252779
>
> Thanks,
> Vladimir K
>
> PS: currently jdk/submit under planned maintenance until Sep 7
>
> On 9/3/20 5:03 AM, Dmitry Chuyko wrote:
>> I pushed the change for signum, and opened an RFE [1] for fp 
>> re-materialization on aarch64. We've made some experiments already. 
>> On N1 simply enabling it makes workloads like Lucene geo search ~0.7% 
>> faster, will get back with more data.
>>
>> I haven't made RFE for "Constants as TEMPs" but will be glad to 
>> participate.
>>
>> -Dmitry
>>
>> [1] https://bugs.openjdk.java.net/browse/JDK-8252760
>>
>> On 9/3/20 11:34 AM, Andrew Haley wrote:
>>> On 02/09/2020 18:40, Vladimir Ivanov wrote:
>>>> C2 doesn't rematerialize FP constants on AArch64
>>> I think that's probably right, even though a few FP constants are
>>> cheap. Loads from L1 usually have a 4 (or 5) cycle latency, which can
>>> make spilling constants expensive around calls. We could be smarter
>>> about this.
>>>