[slime-devel] Re: #+#. bug report (with fix, maybe)
heller at common-lisp.net
Sun Jan 27 10:24:11 UTC 2008
* Luis Oliveira [2008-01-22 01:33+0100] writes:
> SUPPRESS-SHARP-DOT sets things up so that #. returns a gensym instead of
> NIL in order to "avoid multiple entries for nil at toplevel in the
> source-map", according to the ChangeLog. Even if we ignore that problem,
> having #. return NIL in this case, while it would dodge the error, it
> wouldn't find or properly annotate the offending (/ 1 0) form since
> #+cl:nil usually suppresses forms.
> The attached patch gets rid of SUPPRESS-SHARP-DOT. That not only fixes
> my original problem but also makes the following example work better
> than it before, since the (/ 1 0) form is now annotated:
> #.'(defun foo () (/ 1 0))
True, but it seems a bit unusual to use #. for inserting big quoted
> I believe that was the case the ChangeLog entry I quoted was referring
> to. However, I have the feeling that there might be some good reason why
> #. was suppressed in the first place. Any ideas?
This thread is probably also related:
I think it boils down to the question whether *read-eval* should be
enabled or not.
What is important, is that we can read the entire toplevel form
without raising errors and that the "shape" of the form is more or
#. could have arbitrary side-effects. The most common case is
probably that #. can't read or evaluate the expression and signals an
error. Think of reading SBCL source code, which often refers to
packages or constants which no longer exist at runtime. Also, in most
cases the actual expression returned by #. is irrelevant because we
must skip over it anyway (#+#. is an exception). For the lookup in
the source-map it doesn't matter much if we intern symbols in the
wrong package (unnecessary interning is annoying, but not fatal). But
the proper package is usually important for #., since the expression
must be evaluated.
We could try both, first reading with *read-eval* enabled and if that
fails retry with *read-eval* disabled. (Let's assume we can rewind the
underlying stream.) But that's somewhat complicated and isn't a 100%
Your proposal, let *read-eval* be whatever it was, sounds better.
I will apply the patch.
> (BTW, it took me a while to figure out this bug was coming from within
> SLIME itself, and it might have taken even longer if someone hadn't
> noticed that compiling the software in question without SLIME worked
> without problem. Perhaps it would be helpful to somehow report internal
> errors as such?)
That sounds like a good idea. We have to think about how to do it
without cluttering the code to much.
More information about the slime-devel