Christian: Thanks for that, I'll have to remember DEBUG_LEAK in the
future. :)

It looks like it may be acting correctly since there *is* now a
fastglobals object that needs collecting, and also a new tuple built
from co_names + "__builtins__". I don't know why it wouldn't collect
those in the shell, though. The shell does create a new stack frame for
every command (where a function's commands are all run in a single stack
frame), but I don't know why that would matter. I'll look further into it.

Chris: The build problems should be fixed in the next patch. Thanks for
spending so much time on benchmarking this.

Regarding PyEval_EvalCode creating a PyFastGlobalsObject: I'm not sure
whether it's right thing to do, but I think it is. PyEval_EvalCode gets
called for eval, exec, running an imported module's code, and everything
in the interactive prompt - basically, running any code that's not in a
function or generator. (I think that covers everything.) It seems like
the patch has PyEval_EvalCode doing the right thing in general, but it
may suffer a bit in benchmarks.