/*

 A fast CRC16 algorithm using 16 word table in RAM

*/

#include <avr/io.h>

//

CRC_POLYNOM = 0xA001
CRC_INIT    = 0xFFFF

// register name assignment

zero       = 1
tmp        = 18
next       = 19 
crc_l      = 20
crc_h      = 21
data_cnt   = 22         ; 22-25, mixed order!
crc_ptr    = 26         ; X
data_ptr   = 30         ; Z

.macro crc_nibble
    mov     tmp, crc_l                  ; save crc_l, low nibble is crc table index
                                        ; shift in next nibble into CRC, right shift
    ror     next
    ror     crc_h
    ror     crc_l

    ror     next
    ror     crc_h
    ror     crc_l

    ror     next
    ror     crc_h
    ror     crc_l

    ror     next
    ror     crc_h
    ror     crc_l

    andi    tmp, 0x0F
    lsl     tmp                        ; 16 bit word table
    ldi     crc_ptr, lo8(crc16RAM)     ; CRC16 table pointer
;    ldi     crc_ptr+1, hi8(crc16RAM)  ; RAM is aligned to 32 Bytes, only LSB must be reloaded
    add     crc_ptr, tmp
;    adc     crc_ptr+1, zero
    ld      tmp, x+
    eor     crc_l, tmp
    ld      tmp, x
    eor     crc_h, tmp
.endm

/*
int16_t calculate_crc(uint32_t app_end);
registers              r23 r22 r25 r24
*/

.text
.global calculate_crc
calculate_crc:
    ; copy CRC table into RAM for faster access via X-Pointer
    ldi     tmp, 32
    ldi     crc_ptr, lo8(crc16RAM)
    ldi     crc_ptr, hi8(crc16RAM)
    ldi     data_ptr, lo8(crc16ROM)
    ldi     data_ptr, hi8(crc16ROM)
copy_loop:
    lpm     next, Z+
    st      X+, next
    subi    tmp, 1
    brne    copy_loop

    ldi     crc_l, lo8(CRC_INIT)
    ldi     crc_h, hi8(CRC_INIT)
    sts     RAMPZ, zero
    ldi     data_ptr, 0                 ; data pointer
    ldi     data_ptr+1, 0
    ldi     crc_ptr+1, hi8(crc16RAM)
crc_loop:
    elpm    next, Z+                    ; 3 cycles
    crc_nibble                          ; 23 cycles
    crc_nibble                          ; 23 cycles
    subi    r24, 1                      ; -1, 32 bit
    sbc     r25, zero
    sbc     r22, zero
    sbc     r23, zero
    brne    crc_loop                    ; 55 cycles / Byte

    mov     r24, crc_l                  ; return CRC
    mov     r25, crc_h
    ret

crc16ROM:
   .word  0xFFFF, 0xFFFF, 0xFFFF, 0xFFFF, 0xFFFF, 0xFFFF, 0xFFFF, 0xFFFF
   .word  0xFFFF, 0xFFFF, 0xFFFF, 0xFFFF, 0xFFFF, 0xFFFF, 0xFFFF, 0xFFFF

.data
.global crc16RAM
.balign 32
crc16RAM:
.space  32