RFR: aarch32: Another version of SoftFP implementation

yangyongyong yangyongyong at huawei.com
Thu Oct 27 12:45:51 UTC 2016


Hi, Ed and Sergey,

I am very glad to see this patch. I have already built on my local machine and run it on our arm platform. This version and our softfp version are comparable at performance with minor difference in average. And I think it is acceptable. It is a great thing if we can pull the code from openjdk community, and build it directly and run it well. Next, we will push some bugfix and performance patches.

Meanwhile, I share our softfp implementation at http://cr.openjdk.java.net/~fyang/161024-softfp.hg.diff , in case someone is interested in a static way to implement softfp.

Best regards.

-----Original Message-----
From: aarch32-port-dev [mailto:aarch32-port-dev-bounces at openjdk.java.net] On Behalf Of Edward Nevill
Sent: Wednesday, October 26, 2016 11:37 PM
To: Sergey Nazarkin; aarch32-port-dev
Subject: Re: RFR: aarch32: Another version of SoftFP implementation

Hi Sergey,

I have tested this out on an armv7l platform and it seems to work fine.

The fact that it tests to see whether FPU is available is nice because it means we can build a single softfp image which will work on hardfp or softfp with minimal performance loss on hardfp.

It would be nice to have a -XX:+UseSoftFP for testing purposes as at the moment I have to rebuild forcing hasFPU() to false in order to test SoftFP (I don't have any non FPU hardware).

Looks ok to push.

yangyongyong: Do you have any comments?

All the best,
Ed.

On Mon, 2016-10-24 at 17:12 +0000, Sergey Nazarkin wrote:
> Hi!
> 
> Please review alternative version of SoftFP implementation 
> (Interpreter and C1) with VFP support in the soft-float binary[1]. In 
> contrast to [2] we keep assembler unchanged because we believe this 
> improve code readability.
> 
> Another feature is java calling convention that is defined in runtime 
> and depends  on FPU availability. It may improve performance on 
> FPU/SoftFPABI platforms in comparison with traditional C-calling 
> convention.
> 
> webrev: http://cr.openjdk.java.net/~snazarki/softfp/
> 
> [1] http://mail.openjdk.java.net/pipermail/aarch32-port-dev/2016-Octo
> ber/000554.html
> [2] http://mail.openjdk.java.net/pipermail/aarch32-port-dev/2016-Octo
> ber/000546.html
> 
> Sergey Nazarkin
> 
> 
> 
> 


More information about the aarch32-port-dev mailing list