------- Comment #10 from ubizjak at gmail dot com 2009-06-28 10:25 ------- (In reply to comment #9)
> This happens in veclower pass. type_for_widest_vector_mode () is forcing operations from v2sf to v4sf. This is not correct, since there will be garbage in the top two elements. For FP values, this will lead to all sorts of NaNs, subnormals, etc... -- http://gcc.gnu.org/bugzilla/show_bug.cgi?id=40550