.align 8
        .text

        .globl  _blitz_memcpy
        .align  8

_blitz_memcpy::
        fld.d  0(r16), f0
        fld.d 16(r16), f0

        fld.d  0(r17), f16
        fld.d  8(r17), f18
        fld.d 16(r17), f20
        fld.d 24(r17), f22
        fld.d 32(r17), f24
        fld.d 40(r17), f26
        fld.d 48(r17), f28
        fld.d 56(r17), f30


        fld.d 32(r16), f0
        fst.d f16,  0(r16)
        fst.d f18,  8(r16)
        fst.d f20, 16(r16)
        fst.d f22, 24(r16)
        fld.d 48(r16), f0
        fst.d f24, 32(r16)
        fst.d f26, 40(r16)
        fst.d f28, 48(r16)
        fst.d f30, 56(r16)

        adds  64, r16, r16
        adds  64, r17, r17

        adds -16, r18, r18
        xor r0, r18, r0
        bnc _blitz_memcpy
blitz_exit::
        bri r1
        nop