Re: [Sbcl-devel] "Failed AVER": A bug in SBCL?

[Nikodemus S. <nik...@ra...>]

On 24 June 2013 18:03, Teemu Likonen <tli...@ik...> wrote:

>>     failed AVER: (NOT (EQL (FUNCTIONAL-KIND FUNCTIONAL) DELETED)) This
>>     is probably a bug in SBCL itself.
>
>> I managed to shrink the relevant code quite a bit. As a shrunken
>> version it doesn't make sense but it should still compile. I'm using
>> SBCL 1.1.8.30-51bc001.
>>
>>
>>     (defun test (string)
>>       (cond ((every #'digit-char-p string)
>>              nil)
>>             ((some (lambda (c)
>>                      (digit-char-p c))
>>                    string))))
>
> More info: It seems that the AVER error is cause by this:
>
>     (declaim (optimize (space 0)))
>
> With (space 1) it compiles.

Many, MANY thanks for the followup, and the original wonderfully
compact test case.

I managed to reproduce this. I did not look at the details yet (nor am
I likely to in the near future), but it seems that our MAYBE-INLINE is
broken.

We delete the optional dispatch while there is still a REF to a
DEFINED-FUN that has a references (without a REF) the optional
dispatch. This seems wrong -- but checking for that breaks elsewhere.

The optional dispatch also gets used in two components. This seems
iffy. Checking for that fixes the symptom, but I didn't check if it
breaks anything else...

diff --git a/src/compiler/ir1opt.lisp b/src/compiler/ir1opt.lisp
index 7f44b64..ecd913c 100644
--- a/src/compiler/ir1opt.lisp
+++ b/src/compiler/ir1opt.lisp
@@ -1081,7 +1081,8 @@
                  (change-ref-leaf ref res))))
         (let ((fun (defined-fun-functional leaf)))
           (if (or (not fun)
-                  (and (eq inlinep :inline) (functional-kind fun)))
+                  (and (eq inlinep :inline) (functional-kind fun))
+                  (not (member (node-component call)
(functional-components fun))))
               ;; Convert.
               (if ir1-converting-not-optimizing-p
                   (frob)

I'd need to page back more to convince myself this is the right thing,
though. I kind of suspect it is not.

Frankly, it seems to me that we're have a lot of special casing for
IMO dubious benefit due MAYBE-INLINE. Maybe we should remove it?

I'm leaving this here for someone else to ponder -- I'm unlikely to
return to the subject shortly unless I'm inspired to go on
MAYBE-INLINE killing spree some time next week.

Cheers,

 -- nikodemus