http://gcc.gnu.org/bugzilla/show_bug.cgi?id=55461
--- Comment #3 from Uros Bizjak <ubizjak at gmail dot com> 2012-11-25 18:48:25 UTC --- (In reply to comment #2) > That's a design not matching the hardware then -- read up on reformatting > penalties. > > Also the way the rule in the md file is written it more looks like a bug to > me. > It seems to try to figure out the right type, and it works most of the time, > just sometimes it fails. You are right, there is a bug in the way TARGET_SSE_TYPELESS_STORES due to recent load/store split, but it doesn't change the fact that -Os forces movups. Do not use -Os if you care about performance.