RFR: 8258932: AArch64: Enhance floating-point Min/MaxReductionV with fminp/fmaxp [v3]

[Andrew Haley]

On 1/11/21 11:41 AM, Dong Bo wrote:
> On Mon, 11 Jan 2021 10:38:41 GMT, Andrew Haley <aph at openjdk.org> wrote:

>> Did you try math.abs() for doubles?
> 
> The `Math.abs(doublesA[i] - doublesB[i])` has `~36%` improvements.
OK, so we are seeing some improvement in the performance of plausible
Java benchmarks, if not real programs.

Is it correct that there have been no improvements to anything you have
tried using single-precision floating-point operands? If not, is there
any possibility that this optimization might do anything useful with
single-precision floats?

-- 
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