Re: [Valgrind-developers] huge performance regressions between yesterday and today ?

[Florian K. <fl...@ei...>]

On 25.07.2014 21:37, Philippe Waroquiers wrote:
> On Fri, 2014-07-25 at 08:07 +0200, Philippe Waroquiers wrote:
>> Intuitively, the biggest change (the VexendNess change) is
>> the first candidate to look at. 
> And as extremely often with performance related subjects,
> intuition gives reliable guesses, as reliable as
>   interpreting the way the birds fly
>   or as a crystal ball,
>   or interpret coffee grounds
>   and other similar scientific ways to predict perf impact.
> 
> So, it is not at all the VexendNess change.
> 
> Rather from (manual) bisect, the regression is caused by r14186,
> which is given below.
> 

Thanks for investigating this.
I'm working on a patch which attempts to eliminate these large buffers
that are meant to hold function names (and file names). Those have
caused usability issues in the past e.g. BZ 155125.  That patch should
eliminate many of these strncpy calls and therefore help with the
perfprmance regression. But let's see..

   Florian