https://gcc.gnu.org/bugzilla/show_bug.cgi?id=91117
--- Comment #1 from Uroš Bizjak <ubizjak at gmail dot com> ---
The generated code is created the way it is partially by design, and partially
by missing (!y,*x) alternative in *vec_extractv2di_0_sse. GCC lowers intrinsics
to generic operations, and this has its pluses (generic optimizations can be
used) and minuses. For the later, generic operations should avoid to allocate
MMX registers unless absolutely necessary, and the compiler discourages their
use by penalizing operations (moves) involving MMX registers.
gcc-10 emulates MMX operations using SSE instructions, so your code gets
compiled to:
pmullw %xmm0, %xmm0
movq %xmm0, %xmm0
ret
