"The regression" is actually due to that 2.2.1 had a bug that prevented
the regression (which isn't a regression at all) from showing up. "The
regression" is actually a glitch in cyclic gc that's been there forever.
As the generation being collected is analyzed, objects that can't be
collected (because, e.g., we find they're externally referenced, or
are in an unreachable cycle but have a __del__ method) are moved out
of the list of candidates. A tricksy scheme uses negative values of
gc_refs to mark such objects as being moved. However, the exact
negative value set at the start may become "more negative" over time
for objects not in the generation being collected, and the scheme was
checking for an exact match on the negative value originally assigned.
As a result, objects in generations older than the one being collected
could get scanned too, and yanked back into a younger generation. Doing
so doesn't lead to an error, but doesn't do any good, and can burn an
unbounded amount of time doing useless work.
A test case is simple (thanks to Kevin Jacobs for finding it!):
x = []
for i in xrange(200000):
x.append((1,))
Without the patch, this ends up scanning all of x on every gen0 collection,
scans all of x twice on every gen1 collection, and x gets yanked back into
gen1 on every gen0 collection. With the patch, once x gets to gen2, it's
never scanned again until another gen2 collection, and stays in gen2.
Bugfix candidate, although the code has changed enough that I think I'll
need to port it by hand. 2.2.1 also has a different bug that causes
bound method objects not to get tracked at all (so the test case doesn't
burn absurd amounts of time in 2.2.1, but *should* <wink>).
generations is now an array. This cleans up some code and makes it easy
to change the number of generations. Also, implemented a
gc_list_is_empty() function. This makes the logic a little clearer in
places. The performance impact of these changes should be negligible.
One functional change is that allocation/collection counters are always
zeroed at the start of a collection. This should fix SF bug #551915.
This change is too big for back-porting but the minimal patch on SF
looks good for a bugfix release.
compatibility function.
Make PyObject_GC_Track and PyObject_GC_UnTrack functions instead of
trivial macros wrapping functions. Provide binary compatibility
functions.
The fix makes it possible to call PyObject_GC_UnTrack() more than once
on the same object, and then move the PyObject_GC_UnTrack() call to
*before* the trashcan code is invoked.
BUGFIX CANDIDATE!
objects to save in gc.garbage. This should be the last change needed to
fix SF bug 477059: "__del__ on new classes vs. GC".
Note that this change slightly changes the behavior of the collector.
Before, if a cycle was found that contained instances with __del__
methods then all instance objects in that cycle were saved in
gc.garbage. Now, only objects with __del__ methods are saved in
gc.garbage.
When moving objects with a __del__ attribute to a special list, look
for __del__ on new-style classes with the HEAPTYPE flag set as well.
(HEAPTYPE means the class was created by a class statement.)
The platform requires 8-byte alignment for doubles, but the GC header
was 12 bytes and that threw off the natural alignment of the double
members of a subtype of complex. The fix puts the GC header into a
union with a double as the other member, to force no-looser-than
double alignment of GC headers. On boxes that require 8-byte alignment
for doubles, this may add pad bytes to the GC header accordingly; ditto
for platforms that *prefer* 8-byte alignment for doubles. On platforms
that don't care, it shouldn't change the memory layout (because the
size of the old GC header is certainly greater than the size of a double
on all platforms, so unioning with a double shouldn't change size or
alignment on such boxes).
This simplifies the rounding in _PyObject_VAR_SIZE, allows to restore the
pre-rounding calling sequence, and allows some nice little simplifications
in its callers. I'm still making it return a size_t, though.
As Guido suggested, this makes the new subclassing code substantially
simpler. But the mechanics of doing it w/ C macro semantics are a mess,
and _PyObject_VAR_SIZE has a new calling sequence now.
Question: The PyObject_NEW_VAR macro appears to be part of the public API.
Regardless of what it expands to, the notion that it has to round up the
memory it allocates is new, and extensions containing the old
PyObject_NEW_VAR macro expansion (which was embedded in the
PyObject_NEW_VAR expansion) won't do this rounding. But the rounding
isn't actually *needed* except for new-style instances with dict pointers
after a variable-length blob of embedded data. So my guess is that we do
not need to bump the API version for this (as the rounding isn't needed
for anything an extension can do unless it's recompiled anyway). What's
your guess?
pad memory to properly align the __dict__ pointer in all cases.
gcmodule.c/objimpl.h, _PyObject_GC_Malloc:
+ Added a "padding" argument so that this flavor of malloc can allocate
enough bytes for alignment padding (it can't know this is needed, but
its callers do).
typeobject.c, PyType_GenericAlloc:
+ Allocated enough bytes to align the __dict__ pointer.
+ Sped and simplified the round-up-to-PTRSIZE logic.
+ Added blank lines so I could parse the if/else blocks <0.7 wink>.
visit_finalizer_reachable since it's the same as visit_reachable.
Rename visit_reachable to visit_move. Objects can now have the GC type
flag set, reachable by tp_traverse and not be in a GC linked list. This
should make the collector more robust and easier to use by extension
module writers. Add memory management functions for container objects
(new, del, resize).
collector will be saved in gc.garbage. This is useful for debugging a
program that creates reference cycles.
- Fix else statements in gcmodule.c to conform to Python coding standards.
Tim Peters
Neil Schemenauer
Martin v. Löwis
Jeremy Hylton
Guido van Rossum
Neal Norwitz
Vladimir Marangozov
Barry Warsaw
Fred Drake
Peter Schneider-Kamp