http://gcc.gnu.org/bugzilla/show_bug.cgi?id=53101
--- Comment #4 from Marc Glisse <marc.glisse at normalesup dot org> 2012-05-03 19:19:00 UTC --- (define_peephole2 [(set (mem:VI8F_256 (match_operand 2)) (match_operand:VI8F_256 1 "register_operand")) (set (match_operand:<ssehalfvecmode> 0 "register_operand") (mem:<ssehalfvecmode> (match_dup 2)))] "TARGET_AVX" [(set (match_dup 0) (vec_select:<ssehalfvecmode> (match_dup 1) (parallel [(const_int 0) (const_int 1)])))] ) (and similar for VI4F_256) is much less hackish than the XEXP stuff. I was quite sure I'd tested exactly this and it didn't work, but now it looks like it does :-/ Except that following http://gcc.gnu.org/ml/gcc-patches/2012-05/msg00197.html , this is not the right place to try and add such logic. That's a good thing because it is way too fragile, another instruction can easily squeeze between the two sets and disable the peephole.