[Bug c/43151] New: [4.5 Regression] Inconsistent diagnostic

[rguenth at gcc dot gnu dot org]

For

struct X { int i; int j[2]; };
const long x = ((struct X *)0)->j - (int *)0;
char a[((struct X *)0)->j - (int *)0];
char b[__builtin_offsetof(struct X, j)];
char c[x];

we now issue the following diagnostics:

> gcc-4.5 -S t.i 
t.i:3:17: warning: variably modified ‘a’ at file scope
t.i:5:6: error: variably modified ‘c’ at file scope

compared to

> gcc-4.4 -S t.i 
t.i:5: error: variably modified ‘c’ at file scope

which didn't warn about the declaration of a.  I would expect a
warning for x as well if we warn about a.  And of course I'd have
expected c to work anyway.

ICC accepts all of the above without a diagnostic including c.

I consider rejecting c errorneous and the inconsistent warning
about a and x at least annoying.  Only the warning is a regression.


-- 
           Summary: [4.5 Regression] Inconsistent diagnostic
           Product: gcc
           Version: 4.5.0
            Status: UNCONFIRMED
          Keywords: rejects-valid, diagnostic
          Severity: normal
          Priority: P3
         Component: c
        AssignedTo: unassigned at gcc dot gnu dot org
        ReportedBy: rguenth at gcc dot gnu dot org


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=43151