[comparebyte.patch](/uploads/d99fe5cbb17612f72748edfe125b473d/comparebyte.patch) 1. Check ```nasm .L16x_Tail: cmp %r9, %r10 ; r10 = end of buf1, r9 = end of full XMMs in buf1 je .LNothing ; ...handle 1 to 15 tail bytes ``` determines if there are tail bytes, but is required only after 16× loop; jump directly to `.L16x_Tail` is performed for `len` ∈ [4; 15]. But instead of just moving (or in addition to moving) `.L16x_Tail` label two instructions down, I moved the entire 16× loop out, tuning the <16b case further. 2. ```nasm movzbl (%rcx,%rax), %ecx movzbl (%rdx,%rax), %edx mov %rcx, %rax sub %rdx, %rax ret ``` can be carefully modified to get rid of `mov`.