2 * A fast checksum routine using movem
3 * Copyright (c) 1998-2001 Axis Communications AB
5 * csum_partial(const unsigned char * buff, int len, unsigned int sum)
15 ;; check for breakeven length between movem and normal word looping versions
16 ;; we also do _NOT_ want to compute a checksum over more than the
17 ;; actual length when length < 40
23 ;; need to save the registers we use below in the movem loop
24 ;; this overhead is why we have a check above for breakeven length
25 ;; only r0 - r8 have to be saved, the other ones are clobber-able
26 ;; according to the ABI
31 ;; do a movem checksum
33 subq 10*4,$r11 ; update length for the first loop
35 _mloop: movem [$r10+],$r9 ; read 10 longwords
37 ;; perform dword checksumming on the 10 longwords
59 ;; fold the carry into the checksum, to avoid having to loop the carry
69 addq 10*4,$r11 ; compensate for last loop underflowing length
71 movem [$sp+],$r8 ; restore regs
74 ;; only fold if there is anything to fold.
79 ;; fold 32-bit checksum into a 16-bit checksum, to avoid carries below.
80 ;; r9 and r13 can be used as temporaries.
82 moveq -1,$r9 ; put 0xffff in r9, faster than move.d 0xffff,r9
86 lsrq 16,$r13 ; r13 = checksum >> 16
87 and.d $r9,$r12 ; checksum = checksum & 0xffff
88 add.d $r13,$r12 ; checksum += r13
95 ;; checksum the rest of the words
106 ;; see if we have one odd byte more
114 ;; copy and checksum the last byte