All three bugs found in both 2.2.3 and 2.3b2. It's one
patch
instead of three because the fixes share a refactoring of
Python/compile.c:symtable_build.

Bug 1: Wrong file name

    import symtable
    symtable.symtable('def foo(a): global a', 'spam', 'exec')

The error message says '<string>' instead of 'spam'.
(Cause: PyNode_CompileSymtable doesn't set
st_filename.)

Bug 2: Memory leak

    while True:
        try:
            compile('def foo(a): global a', 'spam', 'exec')
        except SyntaxError:
            pass

(Cause: symtable_build doesn't free c->c_symtable on
error.)

The patch is missing a test case for this one; I don't see
how to write it.

Bug 3: Exception clobbered

    def foo(a):
        global a  # SyntaxError

    def bar():
        b = 1
        global b  # SyntaxWarning

Running this as a script, the SyntaxWarning is issued,
and then
the interpreter dies silently (that is, without printing a
traceback reporting the SyntaxError). Compiling it with
compile()
causes a SystemError. (Cause: see below.)

What to do about bugs 1 and 2 is obvious, but bug 3
(which was
actually reported on c.l.py) is not so clear to me. Here's
how
the problem occurs:

symtable_global() sees the global statement in foo(),
raises a
SyntaxError, increments st_errors, and returns.
Processing
continues. Later, symtable_global() sees the global
statement in
bar() and issues a warning by (indirectly) calling
PyErr_WarnExplicit(). This call clears the SyntaxError,
which is
still pending at that time. (It's actually cleared during the
attempt to import the warnings module, in
PyImport_Import, which
seems to think the exception was raised in
PyEval_GetGlobals.)
But st_errors is still > 0, as it should be, so
symtable_build()
returns -1 (resp. PyNode_CompileSymtable, NULL), its
callers
return their error values, etc., until eventually PyErr_Print
tries to print the exception that isn't there.

What the patch implements is this:

Do not issue SyntaxWarnings if an exception is pending;
fail
instead and let that exception propagate. Also, as a
defensive
measure against other bugs of this type (present and
future), when
returning with error from symtable_build(), verify that
there's an
exception pending (and raise SystemError if not). Finally,
refactor symtable_build() so PyNode_CompileSymtable
can use it and
thereby benefit from that defensive measure.

Alternatives (and why I don't like them):

1. Do not try to continue processing after a SyntaxError is
   raised. (Seems like the Right Thing to me, but also
   seems to be
   contrary to the intent of the existing code. There are
   oodles of
   places in compile.c which call symtable_node without
   checking
   st_errors immediately afterwards.)

2. Put the check for a pending exception in
   PyErr_WarnExplicit()
   instead of in the helper function in compile.c. (Doesn't
   seem
   like a common enough coding error to merit a check
   there. In
   symtable_node etc we deliberately let SyntaxErrors, er,
   "pend"
   while we do a bit more compiling, so *there* it's worth
   checking.
   Note that jcompile() already has a check for something
   similar,
   though not for the symbol-table-building phase.)

3. Squirrel the pending exception away, issue the
   warning, then
   restore the exception. (Not worth the bother, IMO. And
   if the
   warning gets strengthened into an exception, should that
   exception
   or the squirrelled one propagate? Ick.)

Logged In: YES
user_id=666873

Just noticed another alternative: leave the checking of
whether a global statement declares a function parameter
global to the later pass. There's already a check for this
condition in symtable_load_symbols (with a different error
message, which is a subject for another patch).

The problem with this idea is that it breaks
symtable.symtable(), which only uses symtable_build. Thus this
function would no longer raise SyntaxErrors for this case.