Re: [boost] Re: [shared_ptr] possible bug in sp_counted_base.hpp

From: Peter Dimov (pdimov_at_[hidden])
Date: 2005-04-14 15:41:17

Rene Rivera wrote:
>> No.. just one that tests all three of the functions. The timing test
>> you have only tests one of them.

No, it tests two of them. The push_back loop causes vector reallocations 
that copy and destroy the old contents. But increments are more frequent. If 
you put a v.resize( 0 ) after the loop, it will test an even number of 
increments and decrements but will also time 2^23 int destructions and this 
will skew the results. Your std::fill idea with an empty pointer isn't bad, 
though.

lock() should be tested in a weak_ptr_timing_test, but I haven't written one 
yet.

>> Well certainly an _mt timing test would be best. As it's not that
>> useful to only time the non-contention cases.

There are two _mt tests that are also timing tests, you even reported 
crashes with one of them. ;-)

The intent of shared_ptr_timing_test is precisely to time the uncontended 
case, because that's what I want to measure, the cost of MT safety when none 
is needed.