/*
* A fast checksum routine using movem
* Copyright (c) 1998-2001 Axis Communications AB
*
* csum_partial(const unsigned char * buff, int len, unsigned int sum)
*/
.globl csum_partial
csum_partial:
;; r10 - src
;; r11 - length
;; r12 - checksum
;; check for breakeven length between movem and normal word looping versions
;; we also do _NOT_ want to compute a checksum over more than the
;; actual length when length < 40
cmpu.w 80,$r11
blo _word_loop
nop
;; need to save the registers we use below in the movem loop
;; this overhead is why we have a check above for breakeven length
;; only r0 - r8 have to be saved, the other ones are clobber-able
;; according to the ABI
subq 9*4,$sp
movem $r8,[$sp]
;; do a movem checksum
subq 10*4,$r11 ; update length for the first loop
_mloop: movem [$r10+],$r9 ; read 10 longwords
;; perform dword checksumming on the 10 longwords
add.d $r0,$r12
ax
add.d $r1,$r12
ax
add.d $r2,$r12
ax
add.d $r3,$r12
ax
add.d $r4,$r12
ax
add.d $r5,$r12
ax
add.d $r6,$r12
ax
add.d $r7,$r12
ax
add.d $r8,$r12
ax
add.d $r9,$r12
;; fold the carry into the checksum, to avoid having to loop the carry
;; back into the top
ax
addq 0,$r12
subq 10*4,$r11
bge _mloop
nop
addq 10*4,$r11 ; compensate for last loop underflowing length
movem [$sp+],$r8 ; restore regs
_word_loop:
;; only fold if there is anything to fold.
cmpq 0,$r12
beq _no_fold
;; fold 32-bit checksum into a 16-bit checksum, to avoid carries below.
;; r9 and r13 can be used as temporaries.
moveq -1,$r9 ; put 0xffff in r9, faster than move.d 0xffff,r9
lsrq 16,$r9
move.d $r12,$r13
lsrq 16,$r13 ; r13 = checksum >> 16
and.d $r9,$r12 ; checksum = checksum & 0xffff
add.d $r13,$r12 ; checksum += r13
_no_fold:
cmpq 2,$r11
blt _no_words
nop
;; checksum the rest of the words
subq 2,$r11
_wloop: subq 2,$r11
bge _wloop
addu.w [$r10+],$r12
addq 2,$r11
_no_words:
;; see if we have one odd byte more
cmpq 1,$r11
beq _do_byte
nop
ret
move.d $r12, $r10
_do_byte:
;; copy and checksum the last byte
addu.b [$r10],$r12
ret
move.d $r12, $r10
