On Mon, Sep 10, 2018 at 01:30:11PM -0700, Dave Hansen wrote:

On 09/08/2018 06:38 PM, owner-linux-mm@kvack.org wrote:

quoted

At last, here is the test result on my 4G virtual machine. I added printk
before and after sparse_memory_present_with_active_regions() and tested three
times with/without "likely".

               without      with
    Elapsed   0.000252     0.000250   -0.8%

The benefit seems to be too small on a 4G virtual machine or even this is not
stable. Not sure we can see some visible effect on a 32G machine.

I think it's highly unlikely you have found something significant here.
It's one system, in a VM and it's not being measured using a mechanism
that is suitable for benchmarking (the kernel dmesg timestamps).

Plus, if this is a really tight loop, the cpu's branch predictors will
be good at it.

Hi, Dave

Thanks for your reply.

I think you are right. This part is not significant and cpu may do its
job well.

Hmm... I am still willing to hear your opinion on my analysis of this
situation. In which case we would use likely/unlikely.

For example, in this case the possibility is (255/ 256) if the system
has 32G RAM. Do we have a threshold of the possibility to use
likely/unlikely. Or we'd prefer not to use this any more? Let compiler
and cpu do their job.

Look forward your insights.

-- 
Wei Yang
Help you, Help me