Really weird (bogus?) regressions

[Paolo Carlini <pcarlini at unitus dot it>]

Hi,

I'm really puzzled. Recently, running make check at my site reveal many 
regression, which, judiging from the posted testresults should be 
completely bogus. I'm afraid something is broken in my setup :(
I'm on a very standard i686-pc-linux-gnu, and the problem happens both 
for 3.4 and 3.3.

The bogus failures belong to gcc.dg/cpp, gcc.dg/noncompile, and few 
others. They are *all* of the same form.
Consider, f.i., gcc.dg/noncompile/920721-2.c:
void f(int n)
{
int s;
for(s=0;s<n;s++)
 s==5?1 n=1;        /* { dg-error "parse error" } */
}

Whereas it's expected:
920721-2.c: In function `f':
920721-2.c:5: parse error before "n"

I get, instead:
920721-2.c: In function `f':
920721-2.c:5: error: syntax error before "n"

Thanks in advance for any help,
Paolo.