https://gcc.gnu.org/bugzilla/show_bug.cgi?id=104151
--- Comment #2 from Hongtao.liu <crazylht at gmail dot com> ---
with -fno-tree-vectorize, gcc also produce optimal code.
mov rax, rsi
mov rdx, rdi
bswap rax
bswap rdx
ret
Guess it's related to vectorizer cost model.
