Re: profile mode: statistics for vector and map

[Gedare Bloom <gedare at rtems dot org>]

On Tue, May 7, 2013 at 6:16 PM, Gedare Bloom <gedare@rtems.org> wrote:
> On Tue, May 7, 2013 at 4:56 PM, François Dumont <frs.dumont@gmail.com> wrote:
>>     IMHO we should adopt the same kind of design I have introduced for the
>> unordered containers that is to say a base class containing in its
>> constructor the call to the instance registration and in the destructor the
>> unregistration. This way special functions could be defaulted which will
>> simplify maintenance and we would never forget the registration like in the
>> vector constructor from initializer_list<value_type> one.
>>
> That's fine. If I have time I'll study your approach in more detail. I
> had a look at the patch you sent, and your approach seems more robust.
> The mechanisms here with macros and hand-inserted function call-outs
> to wrap the "real" operations is easy to mess up.
>
If someone can rework the existing vector code for the
profiler_vector_size.h/profiler_container_size.h files using the
approach suggested, then I can probably generalize the approach to the
container statistics that I implemented without too much trouble. I
used the vector size as the basis for my design and implementation.

-Gedare

> I did this work in a rush last year and just recently had time to
> package it and submit it. Either the patch bit-rots, someone fixes it,
> or I can get around to it myself eventually. I do not have the
> time/money for it right now. Someone requested the code recently, and
> I provided it for 4.7.1 and thought I should push it upstream lest you
> all think no one uses profile mode. :)
>
>  Side note: Profile mode improvements might make good GSoC projects.
>
> -Gedare
>
>> François
>>
>>
>>
>> On 05/07/2013 08:14 PM, Gedare Bloom wrote:
>>>
>>> Hi,
>>>
>>> The attached patch offers a start at the Statistics aspect of
>>> libstdc++ profile mode. So far it implements vector and map statistics
>>> for size (redundant/deprecates existing vector size) and common
>>> operation counts. There are macros for turning on/off the vector and
>>> map statistics individually, e.g.
>>> _GLIBCXX_PROFILE_NO_VECTOR_STATISTICS and _NO_MAP_STATISTICS.
>>>
>>> Some improvements could be made still or remain for the future, such as:
>>> * Use the _GLIBCXX_PROFILE_STATISTICS flag to turn on/off all statistics
>>> * Cost heuristics or high resolution timers to get the costs of operations
>>> * There were some problems with iterators, though I forget the specifics.
>>> * More containers supported and tested.
>>>
>>> I have not done extensive testing on this version of the code. This
>>> work was done previously on 4.7.1 with x86_64. I made some effort to
>>> rebase this patch from Francois' unordered map patch. Hopefully there
>>> are no conflicts remaining.
>>>
>>> 2013-05-07  Gedare Bloom <gedare@rtems.org>
>>>
>>>      * include/profile/impl/profiler_container_statistics.h: New.
>>>      * include/profile/impl/profiler_vector_statistics.h: New.
>>>      * include/profile/impl/profiler_map_statistics.h: New.
>>>      * include/Makefile.am: Add new profile headers.
>>>      * include/Makefile.in: Regenerate.
>>>      * include/profile/impl/profiler.h: New macros for
>>>      __profcxx_vector_statistics and __profcxx_map_statistics.
>>>      * include/profile/impl/profiler_trace.h: Class and function
>>> declarations
>>>      for std::profile::__trace_vector_statistics and
>>>      std::profile::__trace_map_statistics.
>>>      * include/profile/map.h: Use __profcxx_map_statistics macros.
>>>      * include/profile/vector: Use __profcxx_vector_statistics macros.
>>
>>