When I want to use valgrind to check for leaks in a Python program (or
test suite), I generally want pymalloc disabled.  When not running
valgrind I generally want it enabled.

Attached is a patch that automatically bypasses the pymalloc code when
running under valgrind but leaves it enabled overwise.

It is controlled by a WITH_VALGRIND #define, but I haven't updated the
configure script to allow turning it on.  I also haven't done much in
the way of profiling to see what the overhead is when not running under
valgrind.

A slightly cleaned up version of the previous patch.  I only needed to
include <valgrind/valgrind.h>.

Please provide a patch for Python 2.6 that includes a ./configure
--with-valgrind option. We may consider such a patch for 2.6 and 3.0 but
definitely not for 2.5.

Here's a patch with James changes to obmalloc and a --with-valgrind
option in configure.in.

Here's the updated version of my patch (the obmalloc.c bits applied
without conflicts to the newer source tree).

The configure changes are a bit different to Lauro's ones, in that they
check for the existence of the <valgrind/valgrind.h> header if Valgrind
support was requested.

An updated version of the patch.  The previous ones were missing the
valgrind check, resulting in the pymalloc code paths being executed
(which in turn cause unintialised read warnings from valgrind).

There are probably a few other performance optimisations that would be
good to turn off when running under valgrind.

A big one is the tuple cache: if there are tuple reference counting
issues, they won't necessarily be seen by valgrind if the tuple is kept
and reused rather than being freed.

James Henstridge schrieb:
> James Henstridge <james@jamesh.id.au> added the comment:
> 
> There are probably a few other performance optimisations that would be
> good to turn off when running under valgrind.
> 
> A big one is the tuple cache: if there are tuple reference counting
> issues, they won't necessarily be seen by valgrind if the tuple is kept
> and reused rather than being freed.

The tuple cache can't be disabled entirely. Some parts of Python and
some third party modules may depend on the fact that an empty tuple ()
is a singleton in Python.

The tuple cache could be reduced to empty tuple and some other caches
like the list and dict cache could be disabled entirely.

Christian