- Allow using negative indexes for MemObjectProxy. So you can do
'obj[-1]' to get the last child, (same as obj.c[-1]).
(Robert Xiao, #882356)
- We now only compile against Cython. I've finally hit some issues that
I don't want to work around. Namely sizeof(Class) doesn't work under
even pyrex 0.9.9. (John Arbash Meinel)
- Duplicate parent entries are filtered out. (eg the intern dict refers
to the same string 2x, you'll see 1 parent ref in the child, not 2.)
(John Arbash Meinel)
- We now default to limiting the maximum length of the parents list
(default 100). I had some dumps where a single object was referenced
50k times. Anything over about 10 is at the point where you won't
really walk them. This can be disabled with
load(max_parents=-1).
The main win is lowering memory consumption. A 50k parent list takes
200kB by itself (on 32-bit). (John Arbash Meinel)
- Fix a PyInt memory leak. We were calling __sizeof__ which returns an
PyInt and not DECREFing it. (John Arbash Meinel)
- Initial support for overriding
__sizeof__ definitions. It turns
out that object has a basic definition, so all new style classes
inherit it. We now provide meliae.scanner.add_special_size, which
takes a type name and some callbacks to determine the size of an
object. This lets you register them without importing the module.
- Small updates to fix the test suite under python2.7 and 64-bit
architectures. PyString changed size in 2.7, and we weren't
accounting for alignment effects on 64-bit machines.
(John Arbash Meinel, Jelmer Vernooij, #739310)
run_tests.py exits nonzero if a test fails.
(Jelmer Vernooij, #740109)- Add a MANIFEST.in file, to allow
python setup.py sdist to
properly see the .c and .h files it needs to include.
(Chris Adams, #735284)
The main update is the ability to do more queries on a subset of
the object graph. om.summarize(starting_at, excluding=[address])
lets you find out what is more directly "owned" by a given object.
- Add
__sizeof__ members to a lot of the core classes (IntSet,
etc.) (John Arbash Meinel)
ObjectManager.compute_total_size() now only computes the size of
a single object, rather than all objects. All objects took too long
to be useful anyway, better to make it easier to use the useful api.
(John Arbash Meinel)obj.iter_recursive_refs() can now be used to find all objects
referenced from this object (including obj). It can also take an
iterable of object addresses to exclude. Which makes it easy to ask,
"What objects are accessible from X that aren't accessible from Y?"
(John Arbash Meinel)ObjectManager.summarize() can now take an object and an exclusion
list, and summarize the referenced objects. This can be quite useful
when you want to look at just a subset of the graph. The syntax is
ObjectManager.summarize(obj, [not_address1, not_address2]).
(John Arbash Meinel)obj.all() and obj.compute_total_size() helpers. These let you
get the set of referenced objects matching the type (like
om.get_all()). But they also allow you to pass an exclusion
list, so you can only get things reachable from here and not
reachable from there. (John Arbash Meinel)
- When dumping a
PyFrame look at the function object for
co_name, so you don't have to wander through the references to
get their yourself. (Andrew Bennetts)
- Fixes for the simple regex parser (w/o simplejson). However,
simplejson is still recommended, because it is both faster and more
accurate (decodes unicode escapes, etc). (John Arbash Meinel)
loader.load() now defaults to computing parents and collapsing
instance dicts. It does mean that loading will be a bit slower, and
consume more memory, but it is almost always what you need to do
first anyway. (John Arbash Meinel)
- Avoid calling
PyType_Type.tp_traverse when the argument is not a
heap-class. There is an assert that gets tripped if you are running a
debug build (or something to do with how Fedora builds its python).
(John Arbash Meinel, #586122)
- Flush the file handle before exiting. We were doing it at the Python
layer, but that might not translate into the
FILE* object.
(John Arbash Meinel, #428165)
- Handle some issues when using Pyrex 0.9.8.4. It was treating
<unsigned long> as casting the object pointer, not as a Python
level "extract an integer". However, assignment to an cdef unsigned
long does the right thing. (John Arbash Meinel)
- Tweak some memory performance issues (Gary Poster, #581918)
A fairly major reworking of the internals, provides significant memory
savings and easier navigation of the object graph.
- The loader internals have been rewritten significantly. Instead of
storing a bunch of objects in a regular python dict, they are now
stored in a custom Pyrex collection, and proxy objects are created on
demand. This means significantly improved memory usage. (roughly
2:1). (John Arbash Meinel)
- The internals change also changes the interface a bit. When viewing
an object, a shorter display is given. One can use
obj[offset] to
get the object at that reference, rather than the reference itself.
(so fewer indirections through the ObjManager). (John Arbash Meinel)
- Now supports the
__sizeof__ interface introduced in python 2.6.
This allows a class (especially extension classes) to inform
Meliae/Python how many bytes it is consuming in memory.
(John Arbash Meinel)
- Memory objects now use a
.children and .parents notation,
rather than .ref_list and .referrers. For one, this makes it
clearer as 'referred to' and 'referred from' is a bit tricky. We also
now have .c and .p which return a list of Memory objects,
rather than just a list of addresses. This makes it very quick to
move around. Especially with .refs_as_dict and other such
niceties. (John Arbash Meinel)