arch/cris/arch-v32/lib/checksumcopy.S

   1 /*
   2  * A fast checksum+copy routine using movem
   3  * Copyright (c) 1998, 2001, 2003 Axis Communications AB
   4  *
   5  * Authors:     Bjorn Wesen
   6  *
   7  * csum_partial_copy_nocheck(const char *src, char *dst,
   8  *                           int len, unsigned int sum)
   9  */
  10
  11         .globl  csum_partial_copy_nocheck
  12 csum_partial_copy_nocheck:
  13
  14         ;; r10 - src
  15         ;; r11 - dst
  16         ;; r12 - length
  17         ;; r13 - checksum
  18
  19         ;; check for breakeven length between movem and normal word looping versions
  20         ;; we also do _NOT_ want to compute a checksum over more than the
  21         ;; actual length when length < 40
  22
  23         cmpu.w  80,$r12
  24         blo     _word_loop
  25         nop
  26
  27         ;; need to save the registers we use below in the movem loop
  28         ;; this overhead is why we have a check above for breakeven length
  29         ;; only r0 - r8 have to be saved, the other ones are clobber-able
  30         ;; according to the ABI
  31
  32         subq    9*4,$sp
  33         subq    10*4,$r12       ; update length for the first loop
  34         movem   $r8,[$sp]
  35
  36         ;; do a movem copy and checksum
  37
  38 1:      ;; A failing userspace access (the read) will have this as PC.
  39 _mloop: movem   [$r10+],$r9     ; read 10 longwords
  40         movem   $r9,[$r11+]     ; write 10 longwords
  41
  42         ;; perform dword checksumming on the 10 longwords
  43
  44         add.d   $r0,$r13
  45         addc    $r1,$r13
  46         addc    $r2,$r13
  47         addc    $r3,$r13
  48         addc    $r4,$r13
  49         addc    $r5,$r13
  50         addc    $r6,$r13
  51         addc    $r7,$r13
  52         addc    $r8,$r13
  53         addc    $r9,$r13
  54
  55         ;; fold the carry into the checksum, to avoid having to loop the carry
  56         ;; back into the top
  57
  58         addc    0,$r13
  59         addc    0,$r13          ; do it again, since we might have generated a carry
  60
  61         subq    10*4,$r12
  62         bge     _mloop
  63         nop
  64
  65         addq    10*4,$r12       ; compensate for last loop underflowing length
  66
  67         movem   [$sp+],$r8      ; restore regs
  68
  69 _word_loop:
  70         ;; only fold if there is anything to fold.
  71
  72         cmpq    0,$r13
  73         beq     _no_fold
  74
  75         ;; fold 32-bit checksum into a 16-bit checksum, to avoid carries below
  76         ;; r9 can be used as temporary.
  77
  78         move.d  $r13,$r9
  79         lsrq    16,$r9          ; r0 = checksum >> 16
  80         and.d   0xffff,$r13     ; checksum = checksum & 0xffff
  81         add.d   $r9,$r13        ; checksum += r0
  82         move.d  $r13,$r9        ; do the same again, maybe we got a carry last add
  83         lsrq    16,$r9
  84         and.d   0xffff,$r13
  85         add.d   $r9,$r13
  86
  87 _no_fold:
  88         cmpq    2,$r12
  89         blt     _no_words
  90         nop
  91
  92         ;; copy and checksum the rest of the words
  93
  94         subq    2,$r12
  95
  96 2:      ;; A failing userspace access for the read below will have this as PC.
  97 _wloop: move.w  [$r10+],$r9
  98         addu.w  $r9,$r13
  99         subq    2,$r12
 100         bge     _wloop
 101         move.w  $r9,[$r11+]
 102
 103         addq    2,$r12
 104
 105 _no_words:
 106         ;; see if we have one odd byte more
 107         cmpq    1,$r12
 108         beq     _do_byte
 109         nop
 110         ret
 111         move.d  $r13,$r10
 112
 113 _do_byte:
 114         ;; copy and checksum the last byte
 115 3:      ;; A failing userspace access for the read below will have this as PC.
 116         move.b  [$r10],$r9
 117         addu.b  $r9,$r13
 118         move.b  $r9,[$r11]
 119         ret
 120         move.d  $r13,$r10