The attached patch exposes the PyImport_FrozenModules
global variable. It can be written through the new
imp.set_frozenmodules() function. No read functionality
has been provided since that would be harder to do
<wink>, and would IMO not be all that useful.

Rationale:
It allows py2exe/McMillan-Installer-like tools to be
written WITHOUT __import__ hooks, while still being
able to bootstrap with pure Python code.

Use case:
I'm in the process of writing a freeze-like tool for
MacOSX. Applications on this platform are (usually)
simply a directory containing some meta files and
arbitrary stuff. I can (and will) make it simply add
.pyc files to this directory, but it will be a huge win
to be able to put a single file in there containing all
needed Python modules. I would prefer to do this
without resorting to __import__ hooks, as they add
overhead and brittleness. I use a vanilla python
executable and bootstrap in Python. Look ma, no C!

Implementation:
PyImport_FrozenModules is an array of structs
containing a pointer to a C string (the name), a
pointer to the marshalled code object data and an int
specifying the size of the code data (negated if the
module is a package). None of these are Python objects.
imp.set_frozenmodules() takes a sequence of tuples as
its argument: (name, codedata, ispkg). Internally this
sequence is converted to a tuple, and a reference is
kept in an additional static variable in import.c. The
elements of the newly allocated PyImport_FrozenModules
array simply point to the actual string data, as
referenced by the tuple.

The patch doesn't include documentation yet, but I will
add that as soon as the patch is otherwise accepted.