The bug is at this line:

Breakpoint 6, xmlparser_gc_clear (self=0x7ffff7e28c08) at /home/haypo/prog/python/master/Modules/_elementtree.c:3414


static int
xmlparser_gc_clear(XMLParserObject *self)
{
    EXPAT(ParserFree)(self->parser);   // <--- HERE
    ...
}

This function calls XML_ParserFree() twice on the same parser object. The first call is fine and frees the memory. Since we now use Python memory allocators, XML_ParserFree() fills the freed memory with 0xDB byte pattern (when Python is in debug mode).

The second XML_ParserFree() call uses freed memory (filled with 0xDB in debug mode).

Call 1: a GC collection

Breakpoint 6, xmlparser_gc_clear (self=0x7ffff7e28c08) at /home/haypo/prog/python/master/Modules/_elementtree.c:3414
(gdb) up
#1  0x0000000000446636 in delete_garbage (collectable=0x7fffffffd9a0, old=0x9b8f90 <_PyRuntime+432>) at Modules/gcmodule.c:759
(gdb) up
#2  0x0000000000446ade in collect (generation=2, n_collected=0x7fffffffda30, n_uncollectable=0x7fffffffda28, nofail=0) at Modules/gcmodule.c:911
(gdb) cont
Continuing.


Call 2: xmlparser_dealloc()

Breakpoint 6, xmlparser_gc_clear (self=0x7ffff7e28c08) at /home/haypo/prog/python/master/Modules/_elementtree.c:3414
(gdb) up
#1  0x00007ffff0038cb8 in xmlparser_dealloc (self=0x7ffff7e28c08) at /home/haypo/prog/python/master/Modules/_elementtree.c:3435


IMHO it's an obvious bug in Python. The question is more why/how the code didn't crash before? :-)