[patch, libgfortran] Fix PR20257 End of record occurs when writing large arrays

[Paul Thomas]

Jerry,

:REVIEWMAIL:

The patch is OK to commit, except:

> You will notice an absurd unit number being assigned to internal 
> units.  It turns out that the stderr unit number is zero.  To avoid 
> mistaking an internal unit as preconnected I set it to 135790.  This 
> also eliminates a case of accessing an uninitialized value pointed out 
> by valgrind.

In my experience, "absurd" numbers ALWAYS come back to bite you.  Use 
something determinable, like 2^31.  This needs to be documented and, 
preferably, an attempt to open unit = what_ever_you_chose should produce 
a warning that this is reserved internally.

> The final patch has been regression tested on i686, NIST tested, many 
> I/O tested and checked with valgrind on a few cases that valgrind 
> noted had problems.  All looks solid on this end.  (I even got a 
> little progress on tonto-1.0 with it)
>
Yes, it regtests OK on FC5 and I too got the little bit of progress on 
tonto.  It does not stop the segfault in FLUSH on non-trivial molecular 
tests, however.

> OK for trunk and then later 4.1.1?

Yup.

Paul

>