[Bug middle-end/48377] [4.6/4.7 regression] miscompilation at -O3

[jakub at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=48377

Jakub Jelinek <jakub at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
             Status|RESOLVED                    |REOPENED
                 CC|                            |irar at gcc dot gnu.org
         Resolution|INVALID                     |
            Summary|[4.6 regression]            |[4.6/4.7 regression]
                   |miscompilation at -O3       |miscompilation at -O3

--- Comment #12 from Jakub Jelinek <jakub at gcc dot gnu.org> 2011-04-07 
07:44:18 UTC ---
I'd say that is a vectorization bug then.  Either it shouldn't vectorize it at
all, or needs to ensure unaligned vector loads (resp. stores) are used
everywhere.

Short testcase:

typedef unsigned int U __attribute__((__aligned__ (1), __may_alias__));

__attribute__((noinline, noclone)) unsigned int
foo (const char *s, int len)
{
  const U *p = (const U *) s;
  unsigned int f = len / sizeof (unsigned int), hash = len, i;

  for (i = 0; i < f; ++i)
    hash += *p++;
  return hash;
}

char buf[64] __attribute__((aligned (32)));

int
main (void)
{
  return foo (buf + 1, 26);
}