[matplotlib-devel] vanishing memory leaks

[John H. <jdh...@ac...>]

Last night hayden (jo...@on...) sent me a patch in which he
identified and fixed two leaks in _backend_agg.cpp, one in the
write_png method and one elsewhere.  Spurred on by his successes, I
tracked down and fixed some additional leaks in the _transforms and
ft2font modules.

Using units/memleak_hawaii3.py in CVS for test script, on typical runs
creating hundreds of random, complex figures, I get anywhere from -5
up to 40 bytes per figure.  Since this range is sometimes negative, it
may reflect statistical fluctuation.  This number is down from 400
bytes per figure in 0.60.2 (and 24K in 0.54.2) and is getting into the
"manageably small" range.  Of course, there is always the possibility
of a leak in an untested area, but I'm encouraged....

Lessons for developers (and me):

  * in pycxx extension code, you almost always want to return
    Py::asObject rather than Py::Object.  One exception is returning
    None as Py::Object().  This caused several leaks in the ft2font
    module.

  * Never incref where you meant to decref!  This was the cause of the
    leak in the _transforms module, which had escaped my earlier unit
    tests because I wasn't unit testing the leak of binary operations
    of lazy values.  In the BinOp class, I was incref-ing in the
    destructor.  Very bad...

  * Use freetype 2.1.9 since this fixes some leaks in earlier
    versions.

JDH