Set source property on an evaluated expression

[Neil Jerram]

In 1.6.7, given these tests -

(with-test-prefix "set-source-property!"
  (read-enable 'positions)
  (let ((s (read (open-input-string "(display \"\")"))))

    (pass-if "set-source-property! before eval"
      (set-source-property! s 'test-sym 10)
      (eval s the-scm-module)
      #t)

    (pass-if "set-source-property! after eval"
      (hash-fold (lambda (x props acc)
                   (if (eq? (source-property x 'test-sym) 10)
                       (set-source-property! x 'test-sym 11)))
                 0
                 source-whash)
      #t)))

- the second test fails because set-source-property! throws a 
wrong-type-arg exception, which is because the car of x has been 
memoized to a gloc: in other words, x is now (address@hidden ""), not 
(display "").

A fix is to change the 3 occurrences of SCM_NCONSP in srcprop.c to 
SCM_NECONSP, as SCM_ECONSP allows both normal pairs and those in which 
the car is a gloc, whereas SCM_CONSP only allows normal pairs.  However, 
comments in the 1.6 code suggest that the design w.r.t. glocs is that 
they should never be visible on the Scheme level, so this may not be the 
best fix.  Instead, the real problem theoretically speaking may be that 
the source-whash (which is Scheme-accessible) contains glocs.

I suspect that it would be hard to remove glocs from the source-whash. 
Given that the source-whash is only of interest to people writing 
debugging tools (i.e. probably just me), I propose that the SCM_NECONSP 
fix is good enough in practice, and would like to release it into 1.6.x. 
 Any objections?

(In 1.8 glocs have been removed, so I presume there is no analogous 
problem there.)

(Curiously, the first quarter of this fix was made in 1999:

1999-09-12  Mikael Djurfeldt  <address@hidden>

        [...]

        * srcprop.c (scm_source_property): Bugfix: Use SCM_NECONSP instead
        of SCM_NCONSP.  (Thanks to Greg Badros.)
)

        Neil