http://gcc.gnu.org/bugzilla/show_bug.cgi?id=54733
Richard Guenther <rguenth at gcc dot gnu.org> changed: What |Removed |Added ---------------------------------------------------------------------------- Keywords| |missed-optimization Status|UNCONFIRMED |NEW Last reconfirmed| |2012-09-28 Ever Confirmed|0 |1 --- Comment #1 from Richard Guenther <rguenth at gcc dot gnu.org> 2012-09-28 10:47:44 UTC --- Confirmed. The original idea was to leverage the byte-swap pass for this. First you need to extend that to work on a memory source (which it currently does not handle), then you, in addition to detecting bswap, emit a word load and detect the "noop" pattern which then skips bswap. (the byte-swap pass was also supposed to utilize the vector shuffle engine for non-bswap shuffles).