http://gcc.gnu.org/bugzilla/show_bug.cgi?id=55573
--- Comment #5 from Marc Glisse <glisse at gcc dot gnu.org> 2012-12-03 13:38:19 UTC --- Thanks Jakub, each time I introduce a regression, you fix it before I even have time to read the PR :-) I haven't looked into it yet, but would changing the assert to scalarish_type_p fail? I am a little concerned that converting a vector means VIEW_CONVERT_EXPR, while converting the elements one by one is a true conversion, but since the comment mentions NOP_EXPR, this path is likely only about cases where it doesn't matter.