proto_tree_add_item was valid *before* we short-circuited based on a NULL tree.
This was good in that it removed a common source of really-long-loop bugs. It
was less good in that it cost us about 8% in speed when doing a tree-less
dissection, but we decided the tradeoff was worth it.
After Anders' recent mail to -dev about performance, I started thinking about
how to optimize this. It occurred to me that the vast majority of the logic
involved in the check was dealing with the length value - fetching the actual
length if it was a counted string, calculating the length if it was -1, adding
the length to the offset in a way that was free from overflows, etc.
All of this is (theoretically) unnecessary - simply checking the offset without
worrying about the length will still catch the very-long-loops, since it is the
offset that increases in each iteration, not the length.
All that to justify:
- implement tvb_ensure_offset_exists which throws an exception if the offset is
not within the tvb
- use it instead of all the complicated other logic in the pre-short-circuit
step of proto_tree_add_item and friends
This gives us back about 3/4 of the performance we lost in the original patch.
We're still ~2% slower than without any check, but this is the best I can think
of right now.
svn path=/trunk/; revision=52578
dissecting without tree, they are costly because they now happen for every
proto_tree_add_item call even if tree is NULL.
svn path=/trunk/; revision=52575
problem/solution.
"Long" since solved by copying the solution from Makefile.nmake.
Small update to the status of cmake builds on windows
svn path=/trunk/; revision=52574
tree_data and reference it directly when allocating/freeing tree items. This
lets us keep multiple around when we need them, and still lets us use
wmem_free_all for a major speedup. It also, coincidentally, lets us get rid of
the annoying fi_tmp hack that was needed before, since that element gets swept
up in the free_all with everything else.
Keep one pool cached to avoid creating/destroying a pool for each packet,
another minor performance win.
The various changes in approach seem to balance out pretty much exactly, this
still gives ~11% over pre-52569.
svn path=/trunk/; revision=52573
field_info separately. We still have to walk the tree in order to free certain
fvalues, but that's not a big deal. Another ~11% speed-up running "tshark -nVr"
on a large capture.
svn path=/trunk/; revision=52569
Even without making use of free_all (which should be possible) this still
results in ~8% speedup running "tshark -nVr" on a large file in my tests.
svn path=/trunk/; revision=52568
Add a new strutil function format_text_chr that replaces unprintable characters
with a single passed-in char (eg a '.' or a '-') instead of a C-style escape.
This is useful for displaying binary data that frequently but not always
contains text; otherwise the number of C escape codes makes it unreadable.
svn path=/trunk/; revision=52563
a typedef of a *pointer* to the struct, not the struct itself, which are
different sizes.
This doesn't show up under valgrind because the length isn't checked in that
case, everything gets subsumed in valgrind's malloc/free hooks.
Fixes https://bugs.wireshark.org/bugzilla/show_bug.cgi?id=9264
svn path=/trunk/; revision=52560
List Windows Server 2012 (and a couple of Home Server versions) as supported OS.
While at it, remove the list of the various "NT-based" Windows versions from
a couple of other places (this effectively removes statements that we still
support NT4.0 and 2000).
svn path=/trunk/; revision=52549
Move the declaration of set_last_open_dir() to ui/util.h. It still
has separate GTK+ and Qt implemenations. We might want to move it to
ui/util.c at some point. Remove a lot of unnecessary GTK+ includes.
Remove most of the references to the global cfile while we're at it.
svn path=/trunk/; revision=52542