there were a lot of loss records.
The fix was:
- Avoid the O(m * n) looping over the chunks when creating the loss
records, by putting loss records into an OSet instead of a list, which
makes duplicate detection for each chunk an O(log n) operation instead of
an O(n) operation.
- Avoid the looping over loss records which was used to do a poor
man's sort, but was O(n^2). Instead copy pointers to the loss records
from the OSet into an array and sort it normally with VG_(ssort) (n log n,
usually) before printing.
This approach was similar to that used in the patch Philippe attached to the
bug report.
Other changes:
- Added Philippe's test programs in the new memcheck/perf directory. It
used to take 57s on my machine, now it takes 1.6s.
- Cleaned up massif/perf/Makefile.am to be consistent with other Makefiles.
- Improved some comments relating to VgHashTable and OSet.
- Avoided a redundant traversal of the hash table in VG_(HT_to_array), also
identified by Philippe..
- Made memcheck/tests/mempool's results independent of the pointer size, and
thus was able to remove its .stderr.exp64 file.
git-svn-id: svn://svn.valgrind.org/valgrind/trunk@9781
that are memory offsets) with PtrdiffT; OffT should only be used for file
sizes and offsets.
Change Off64T from a ULong to a Long, as it should be. Replace some uses
of ULong in the address space manager with Off64T to match.
Also add a comment explaining the meanings of the basic types like Addr,
OffT, SizeT, etc.
Also fix the prototype for VG_(pread) -- the last arg is an OffT, not an
Int.
git-svn-id: svn://svn.valgrind.org/valgrind/trunk@8959
relatively minor extensions to m_debuginfo, a major overhaul of
m_debuginfo/readdwarf3.c to get its space usage under control, and
changes throughout the system to enable heap-use profiling.
The majority of the merged changes were committed into
branches/PTRCHECK as the following revs: 8591 8595 8598 8599 8601 and
8161.
git-svn-id: svn://svn.valgrind.org/valgrind/trunk@8621
this does not always produce correct results. Instead use a slower
but correct method. Fixes#147545. (Nick Nethercote, Tom Hughes et
al)
git-svn-id: svn://svn.valgrind.org/valgrind/trunk@7283
interface and provides full power; and "OSetWord_", which is an
easier-to-use interface for if you just want to store words.
git-svn-id: svn://svn.valgrind.org/valgrind/trunk@6841
Inline stackPush and stackPop and placate gcc's resulting concerns
about uninitialised variables.
and also change ownership.
git-svn-id: svn://svn.valgrind.org/valgrind/trunk@6270
which is a sorted set with no duplicates. This is derived from
m_skiplist, which it will hopefully replace. The interface has the
following improvements:
- Avoided all mention of how the data structure is implemented in the
interface, so it could be replaced with another data structure without
changing external code.
- Two kinds of comparison: fast -- use the first word of each element
for comparison; slow -- use a custom function. The custom function
compares a key with an element, so non-overlapping interval lists can
be supported easily. m_skiplist only supports the slow variant, and it
makes things almost 2x faster.
- Users pass in malloc() and free() functions, so m_oset.c it doesn't
rely on any particular allocator.
- It has a Destroy() function which will deallocate all the nodes.
- It allows variable-sized nodes.
- No static constructor; I needed the flexibility of being able to
execute arbitrary code in the constructor. This also means no type
internals are exposed.
No part of Valgrind actually uses OSet yet, although I've privately
converted several data structures, and so I'm confident that the
interface is basically sound. Some functions may be added later.
The implementation uses AVL trees, and has the following
characteristics:
- Lookup is much faster than for skiplists -- around 3x. This is
because the inner lookup loop is much tighter.
- Insertion and removal is similar speed to skiplists, maybe a little
slower, but there's still some fat to be trimmed.
- The code is a bit longer and more complex than the skiplist code.
This was intended to replace the need for the VgHashTable type. But my
experiments have shown that VgHashTable is really fast, faster than both
AVL trees and skiplists in all but extreme cases (eg. if the hashtable
becomes way too full): insertion takes constant time, because you always
prepend to chains; lookup depends on chain length, but the inner loop
is so tight that you need about 20 elements per chain before it gets
worse than the AVL tree; removal is similar to lookup. And because
insertion uses prepending, any locality in accesses will help things. If
VgHashTable had its interface cleaned up to look like OSet's, and was made
to auto-resize when it got too full, it might be a better OSet (although
it's not sorted).
So, it's currently unclear exactly how the AVL tree OSet will be used.
The skiplist could be converted to the new interface (I have a 90%
complete version which I used in the comparison experiments). If
VgHashTable was converted to the same interface (or as close as
possible) it would make direct comparison of important places (eg.
Memcheck's malloc_lists) simple.
git-svn-id: svn://svn.valgrind.org/valgrind/trunk@4410
