RFR: 8249893: AARCH64: optimize the construction of the value from the bits of the other two

[Andrew Haley]

Hi,

On 09/10/2020 20:01, Boris Ulasevich wrote:
> 
> Many thanks for looking at this again!
> 
> Benchmark link: [1]. Measurements on Cortex A73, Cortex A53 and Neoverse N1
> shows [2] 6-15% performance improvement for bench1 and 18-29% for bench2.

For me on ThunderX 2,

Before:

Benchmark           Mode  Cnt   Score   Error  Units
ConstructFF.bench1  avgt   10  15.170 ± 0.975  ns/op
ConstructFF.bench4  avgt   10  39.391 ± 2.617  ns/op

After:

ConstructFF.bench1  avgt   10  12.349 ± 2.535  ns/op
ConstructFF.bench4  avgt   10  24.353 ± 0.443  ns/op

So for this carefully-constructed benchmark, it looks like there's
a useful gain.

-- 
Andrew Haley  (he/him)
Java Platform Lead Engineer
Red Hat UK Ltd. <https://www.redhat.com>
https://keybase.io/andrewhaley
EAC8 43EB D3EF DB98 CC77 2FAD A5CD 6035 332F A671