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

[rguenth at gcc dot gnu.org]

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

--- Comment #31 from Richard Guenther <rguenth at gcc dot gnu.org> 2011-04-07 13:35:04 UTC ---
(In reply to comment #30)
> This is certainly a valid testcase:
> 
> struct S { long long x; int a; double b[10000]; int c; } s;
> double *p = &s.b[0];
> int
> foo (double *q)
> {
>   int i, j = 0;
>   for (i = 0; i < 10000; i++)
>     j += *q++;
>   return j;
> }
> int
> main (void)
> {
>   return foo (p);
> }
> 
> and without the target hook, vectorizer would think it can peel the loop and
> get it naturally aligned (assuming it isn't inlined at least etc. to see that q
> points to &s.b[0] initially).
> 
> I think the bug is in the ABIs that do that kind of thing, but it is too late
> to fix the ABIs up.

The question is what we should do in the middle-end.  If the above would
happen for strict-alignment targets then *p would simply trap.  The
only consistent way for the middle-end would be to simply assume all
pointer-to-doubles are only aligned to 4 bytes, thus use the minimal
alignment that can happen (without using GCC extensions).