[Sbcl-devel] Re: [Sbcl-help] Debugging questions

[Alexey D. <ade...@co...>]

Dave Roberts <ld...@dr...> writes:

> (defun decode-mx-rdata (buffer size index)
>   (let ((length (get16 buffer size index))
> 	(preference (get16 buffer size (+ 2 index))))
>     (multiple-value-bind (exchange end-idx) (decode-name buffer size
> 							 (+ 4 index))
>       (assert nil) ;(= (+ index length 2) end-idx))
>       (values (list preference exchange) end-idx))))
> 
> Basically, when this broke, I was trying to find out the value of index
> and length. Index is a function argument, and I could find that with
> either the LIST-LOCALS command or SB-DEBUG:ARG. I could not figure out
> how to determine the value of LENGTH, however.

Obstacles to dumping debug information:

1. If a variable is not referenced (e.g. LENGTH above), or all
   references are optimized away, it is discarded.

2. If <var1> is never set and its definition is (LET ((var1 var2))
   ...), it is substituted with <var2>.

3. If a variable is never set and is referenced exactly once, the
   reference is substituted with the variable initial value.

I am not sure about case 1: unused variable seems to be only useful
for debugging, and in this case it can be replaced with a special.

Regarding cases 2 and 3, I think these optimizations should be
disabled according to the general policy (debug>=2 for interned and =3
for uninterned symbols), except for IR1-TRANSFORMed lambdas.

Case 2 has an interesting consequence: variables, introduced with
MULTIPLE-VALUE-BIND, are not recorded with DEBUG=3: with this policy,
a variable is considered live everywhere in its scope, but M-V-BIND
is translated to M-V-CALL of (lambda (&optional v1 v2 ...) ...), its
body is in the scope of variables from the varargs entry, but
actually uses variables from the last &optional entry, which are
considered to be dead :-). We can try to prolong their scope, but it
looks easier to just disable let-substitution.

BTW, what about compiling open-coded lambdas with (DEBUG 0)? We
already had to disable CATCH insertion, and now there are two new
optimizations to be enabled.

-- 
Regards,
Alexey Dejneka

"Alas, the spheres of truth are less transparent than those of
illusion." -- L.E.J. Brouwer