memset8_        proc    far
                jcxz thatsIt_8                        8 / 5

                push bx                               1
                mov ah,al                             1
                mov bx,ax                             1
                shl eax,16                            2
                mov ax,bx                             1
                pop bx                                4

                cld                                   2
                test di,1                             1
                jz aligned16_8                        3 / 1
                stosb                                 5
                dec cx                                1
                jz thatsIt_8                          3 / 1

aligned16_8:    test di,2                             1
                jz aligned32_8                        3 / 1
                stosb                                 5
                dec cx                                1
                jz thatsIt_8                          3 / 1
                stosb                                 5
                dec cx                                1

aligned32_8:    push cx                               1
                shr cx,2                              2
                rep stosd                             ?+5x
                pop cx                                4
                and cx,3                              1
                rep stosb                             ?+5x

thatsIt_8:      ret
memset8_        endp
                                             total:  60+5x / 49+5x / 33+5x

memset16_       proc far
                jcxz thatsIt_16                       8 / 5

                mov bx,ax                             1
                shl eax,16                            2
                mov ax,bx                             1

                mov dx,di                             1

                cld                                   2
                test di,1                             1
                jz aligned16_16                       3 / 1
                stosb                                 5
                ror eax,8                             2
                dec cx                                1
                jz lastByte_16                        3 / 1

aligned16_16:   test di,2                             1
                jz aligned32_16                       3 / 1
                stosw                                 5
                dec cx                                1

aligned32_16:   shr cx,1                              2
                rep stosd                             ?+5x
                jnc lastByte_16                       3 / 1
                stosw                                 5

lastByte_16:    test dx,1                             1
                jz thatsIt_16                         3 / 1
                stosb                                 5

thatsIt_16:     ret
memset16_       endp
                                                     57+5x / 42+5x / 29+5x


memset32_       proc far
                jcxz thatsIt_32                       8 / 5

                cld                                   2

                xchg dx,ax                            3
                shl eax,16                            2
                mov ax,dx                             1

                mov dx,di                             1

                test di,1                             1
                jz aligned16_32                       3 / 1
                stosb                                 5
                ror eax,8                             2
                dec cx                                1
                jz doRest_32                          3 / 1
                test di,2                             1
                jz mainCopy_32                        3 / 1
                stosw                                 5
                ror eax,16                            2
                jmp mainCopy_32                       3

aligned16_32:   stosw                                 5
                ror eax,16                            2
                dec cx                                1
                jz doRest_32                          3 / 1

mainCopy_32:    rep stosd                             ?+5x

doRest_32:      test dx,2                             1
                jz lastByte_32                        3 / 1
                stosw                                 5
                ror eax,16                            2
lastByte_32:    test dx,1                             1
                jz thatsIt_32                         3 / 1
                stosb                                 5

thatsIt_32:     ret
memset32_       endp
                                                     76+5x / 61+5x / 35+5x