Created attachment 36805 [details]
memcheck patch to implement "delta leak search" client requests

At work, before using valgrind, we used wdb (hp-ux gdb version) leak search.
wdb has a nice feature that allows to output only "new leaks".
We are using this functionality to search for leaks between "test phases/actions" of a bigger test. This allows to see more precisely which test phase/action is leaking (or allocating, if show-reachable=yes) additional memory.

The attached patch implements two new client requests:
     VALGRIND_DO_ADDED_LEAK_CHECK : only shows the loss record with increased
       nr of bytes or blocks
     VALGRIND_DO_CHANGED_LEAK_CHECK : shows loss records with increase or
          decreased bytes or blocks.
Using such a request gives errors such as:
==21888== 10 (+10) bytes in 1 (+1) blocks are definitely lost in loss record 2 of 3
==21888==    at 0x400591C: malloc (vg_replace_malloc.c:195)
==21888==    by 0x8048541: f (leak-delta.c:14)
==21888==    by 0x8048C37: main (leak-delta.c:61)
==21888== 

So, instead of showing all loss records with absolute nr of bytes/blocks losts,
these new requests will only show the loss records for which there is an increase (or a change) of the nr of bytes/blocks lost (or allocated).
The "unchanged" loss records are not output.
This allows a much easier link between a leak (or memory allocation) and
the test phase producing this leak/memory allocation change.

The regression tests of memcheck have been run with the patch (no new problem).
I will also attach a test program (which must be run as:
   valgrind --leak-check=full --show-reachable=yes leak-delta

If this will be integrated in valgrind, I can then provide patches for
the NEWS file + user manual (basically, would be the above text).