Hi Christian!
Armin Rigo wrote:
On Thu, Mar 16, 2006 at 01:26:28PM +0100, Armin Rigo wrote:
def rtype_destruct_object(hop):
v_any, = hop.inputargs(hop.args_r[0])
hop.genop('gc_unprotect', [v_any])
Updates (with feedback from Carl): the compute_result_annotation also
needs fixing. It should not say "annmodel.s_None", but just "None" --
which however must then be spelled in a lambda form, to avoid confusion
with "unspecified". Argh, this interface is not really polished. You
need:
..., compute_result_annotation=lambda *args_s: None, ...
Carl is hitting more bugs, which are not in backend_optimizations but in
the gctransformer which you run before -- he just added a checkgraph()
there that shows this. This needs fixing, indeed. (gctransform in
general needs to be cleaned up, too.)
It is actually the replace_gc_(un)protect code that you wrote. The
problem is that you replace an operation
v0 = gc_unprotect(obj)
with no operation. That means that afterwards v0 is undefined. For some
reason the checkgraph in the gctransformer was removed at one point (I
thought there was one there), so you found this problem only _much_
later in the backend optimizations. This is not entirely your fault, the
problem is indeed that gctransformer is a mess, you should not have to
care for generating a
v0 = same_as(Constant(None, Void))
yourself. The transformer definitively needs some refactoring.
Cheers,
Carl Friedrich
_______________________________________________
[email protected]
http://codespeak.net/mailman/listinfo/pypy-dev
