And best of all if called, no registers need to be save. If inlined, it gets rescheduled
And it does the work in 5 instructions that used to take 15 (if inlined for 32 bit)
and if called (to memcmp), many dozen....

v6_equal:
        @ args = 0, pretend = 0, frame = 0
        @ frame_needed = 0, uses_anonymous_args = 0
        @ link register save eliminated.
        vld1.32 {d16-d17}, [r0] @ D.32765, MEM[(const __builtin_neon_si[4] *)p1_2(D)]
        vld1.32 {d18-d19}, [r1] @ D.32765, MEM[(const __builtin_neon_si[4] *)p2_3(D)]
        veor    q8, q8, q9      @ D.32765, D.32765, D.32765
        vorr    d16, d17, d16   @ D.32766, D.32765, D.32765
        vpmax.u32       d16, d16, d16   @, D.32766, D.32766, D.32766
        vmov.32 r0, d16[0]      @ D.32767, D.32766,
        clz     r0, r0  @, D.32767