Crash in bytecode->ir translation while accessing co_llvm

[GoogleCodeExporter]

The following snippet crashes while accessing co_llvm:

def y():
    raise Exception

for i in xrange(50001):
    try:
        y()
    except:
        pass
print "---"
print y.__code__.co_llvm

Original issue reported on code.google.com by joerg...@gmail.com on 6 Jan 2010 at 11:39

[GoogleCodeExporter]

Original comment by collinw on 6 Jan 2010 at 11:40

• Changed state: Accepted
• Added labels: Priority-Critical, Release-2009Q4, Type-Defect
• Removed labels: Priority-Medium, Type-Enhancement

[GoogleCodeExporter]

I think happens like this:
First y() becomes hot and is compiled.  The LOAD_GLOBAL_fast optimization is 
used for
the lookup of "Exception".  The next iteration of the loop modifies the global 
"i"
and invalidates the compiled code by setting co_assumed_globals to NULL.  It 
leaves
the CO_FDO_GLOBALS flag set so that no attempt will be made to execute the 
code.  All
the remaining calls go through the interpreter which is unfortunate but not 
this bug.

When the IR has to be regenerated it tries to use the LOAD_GLOBAL_fast path 
again but
now we don't have a co_assumed_globals anymore so trying to do a lookup crashes
inside PyDict_GetItem().

Original comment by abbeyj on 12 Jan 2010 at 1:37

[GoogleCodeExporter]

Original comment by collinw on 10 Mar 2010 at 2:38

• Changed state: Started

[GoogleCodeExporter]

This was fixed as part of the watcher redesign in r1120. I added a regression 
test for 
this segfault in r1130.

Original comment by collinw on 10 Mar 2010 at 7:27

• Changed state: Fixed