- Valgrind runs much faster when the --smc-check=all option is given.
- Cachegrind has a new processing script, cg_diff, which finds the
difference between two profiles. It's very useful for evaluating the
performance effects of a change in a program.
Related to this change, the meaning of cg_annotate's (rarely-used)
--threshold option has changed; this is unlikely to affect many people, if
you do use it please see the user manual for details.
- Callgrind now can do branch prediction simulation, similar to Cachegrind.
In addition, it optionally can count the number of executed global bus events.
Both can be used for a better approximation of a "Cycle Estimation" as
derived event (you need to update the event formula in KCachegrind yourself).
- Cachegrind and Callgrind now refer to the LL (last-level) cache rather
than the L2 cache. This is to accommodate machines with three levels of
caches -- if Cachegrind/Callgrind auto-detects the cache configuration of
such a machine it will run the simulation as if the L2 cache isn't
present. This means the results are less likely to match the true result
for the machine, but Cachegrind/Callgrind's results are already only
approximate, and should not be considered authoritative. The results are
still useful for giving a general idea about a program's locality.
- Massif has a new option, --pages-as-heap, which is disabled by default.
When enabled, instead of tracking allocations at the level of heap blocks
(as allocated with malloc/new/new[]), it instead tracks memory allocations
at the level of memory pages (as mapped by mmap, brk, etc). Each mapped
page is treated as its own block. Interpreting the page-level output is
harder than the heap-level output, but this option is useful if you want
to account for every byte of memory used by a program.
- Added new memcheck command-line option --show-possibly-lost.
OBS-URL: https://build.opensuse.org/package/show/devel:tools/valgrind?expand=0&rev=17
Dirk Mueller
a3155bc151