[Bug target/53907] gcc uses unaligned load when aligned load was requested

[rguenth at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=53907

--- Comment #2 from Richard Guenther <rguenth at gcc dot gnu.org> 2012-07-10 
09:16:29 UTC ---
  s.0_2 = (sizetype) s_1(D);
  D.4303_3 = s.0_2 & 15;
  D.4304_4 = -D.4303_3;
  s_5 = s_1(D) + D.4304_4;
  sz_6 = MEM[(const __m128i * {ref-all})s_5];
  return sz_6;

it's hard to see for GCC that s_5 is aligned because of the way it is computed.
If I change the source to use

#include <emmintrin.h>
__m128i x(char *s){
    __m128i sz,z,mvec;
    s=(char *)(((unsigned long) s)& -16);
    sz=_mm_load_si128(s);
    return sz;
}

then GCC sees that s is aligned.  The code generated for the re-alignment
is also simpler, thus we should do this transform somewhere.

I'll prepare a patch.