https://gcc.gnu.org/bugzilla/show_bug.cgi?id=79880
--- Comment #3 from Jakub Jelinek <jakub at gcc dot gnu.org> --- Perhaps something to report to llvm that they accept invalid, that is a bug. Also, why do you use Intel syntax when you don't know how should it be used? In AT&T syntax you wouldn't need to specify this extra stuff, just vpgatherdd %xmm4, (%r13 + %xmm3), %xmm4. Also note that for inline asm you don't specify clobbers etc. (but that might be just from simplification of the testcase).
