This is the mail archive of the
libc-alpha@sourceware.org
mailing list for the glibc project.
Re: [PATCH] PowerPC: removed branch prediction from rint implementation
On Sun, Mar 31, 2013 at 8:56 PM, Adhemerval Zanella
<azanella@linux.vnet.ibm.com> wrote:
> On 03/31/2013 10:24 PM, Ryan S. Arnold wrote:
>> On Sun, Mar 31, 2013 at 8:02 PM, Adhemerval Zanella
>> <azanella@linux.vnet.ibm.com> wrote:
>>> This patch removes the branch prediction instructions for PowerPC rint
>>> implementation, replacing them for normal instructions. These instructions
>>> are actually hurting performance, it is safer and generally better to let
>>> the hardware handle it.
>> This is true on Power7. I believe it was not the case on older hardware.
>
> Indeed, but assembly implementation is based on sysdeps/powerpc/fpu/s_rint.c, and the C
> implementation built with gcc 4.7.3 plus -O3 -mcpu=powerX does not generate any branch
> prediction instruction for neither power4, power5, power6, or power7.
>
> The assembly implementation make two assumptions: 1. 'fabs (x) < 2^52' is unlikely and
> 2. 'x > 0.0' is unlike (if 1. is true). So usually the algorithm expect path is
> 'x < 0.0' and 'abs(x) < 2^52'. Since it a general floating point function, expected input
> is not bounded and I think the restrictions for branch prediction does not make sense.
>
>>
>>> I added a rint benchmark test on benchtest framework. The results on a POWER7
>>> 64 bits are:
>>>
>>> MASTER: rint: ITERS:1e+09: TOTAL:12.0901s, MAX:48.908ns, MIN:5.312ns, 8.27124e+07 iter/s
>>> PATCH: rint: ITERS:1e+09: TOTAL:5.29302s, MAX:37.826ns, MIN:3.922ns, 1.88928e+08 iter/s
>>>
>>> Any tips, comments, advices?
>> We could conditionally exclude the branch hints on power7 and later
>> hardware as an alternative.
>
> Based on my previous comments, I don't think it is worth to add extra logic to handle specific
> platforms.
Alright, I stand corrected! This looks good to me.
Ryan