Re: [Qemu-devel] [PATCH v3 4/5] qtest: precompute hex nibs

[Eric Blake <eblake@redhat.com>]

[-- Attachment #1: Type: text/plain, Size: 1772 bytes --]

On 05/08/2015 10:22 AM, John Snow wrote:
>>> I'm a bit surprised - making a function call per byte generally executes
>>> more instructions than open-coding the conversion (albeit the branch
>>> prediction in the hardware probably does fairly well over long strings,
>>> since it is a tight and predictable loop).  Remember, sprintf() has to
>>> decode the format string on every call (unless the compiler is smart
>>> enough to open-code what sprintf would do).
>>
>> John's measurements show that the speed difference between snprintf()
>> and a local copy of formatting code gets thoroughly drowned in noise.
>>
>> The snprintf() version takes 18 lines less, according to diffstat.  Less
>> code, same measured performance, what's not to like?
>>
>> However, if you feel strongly about avoiding snprintf() here, I won't
>> argue further.  Except for the commit message: it needs to be fixed not
>> to claim avoiding "printf and friends" makes a speed difference.
>>
> 
> My reasoning was the same as Markus's: the difference was so negligible
> that I went with the "less home-rolled code" version.
> 
> I already staged this series without the nib functions and submitted the
> snprintf version as its own patch with a less disparaging (to printf and
> friends) commit message.
> 
> Any further micro-optimization is a waste of time to properly benchmark
> and split hairs. I already dropped the test from ~14s to ~4s. Good enough.

Okay, I'm convinced - keep the s[n]printf, on the grounds that it wasn't
the bottleneck.  You are right that premature optimization is not worth
the complexity if it gives no real speed gain.

-- 
Eric Blake   eblake redhat com    +1-919-301-3266
Libvirt virtualization library http://libvirt.org


[-- Attachment #2: OpenPGP digital signature --]
[-- Type: application/pgp-signature, Size: 604 bytes --]