projects / linux-block.git / blame
| Commit | Line | Data |
|---|---|---|
| b2441318 | 1 | /* SPDX-License-Identifier: GPL-2.0 */ |
| 9bf79539 | 2 | /* |
| 1da177e4 LT |
3 | * A fast checksum routine using movem |
| 4 | * Copyright (c) 1998-2001 Axis Communications AB | |
| 5 | * | |
| 6 | * csum_partial(const unsigned char * buff, int len, unsigned int sum) | |
| 7 | */ | |
| 8 | ||
| 9 | .globl csum_partial | |
| 10 | csum_partial: | |
| 11 | ||
| 12 | ;; r10 - src | |
| 13 | ;; r11 - length | |
| 14 | ;; r12 - checksum | |
| 15 | ||
| 16 | ;; check for breakeven length between movem and normal word looping versions | |
| 17 | ;; we also do _NOT_ want to compute a checksum over more than the | |
| 18 | ;; actual length when length < 40 | |
| 19 | ||
| 20 | cmpu.w 80,$r11 | |
| 21 | blo _word_loop | |
| 22 | nop | |
| 23 | ||
| 24 | ;; need to save the registers we use below in the movem loop | |
| 25 | ;; this overhead is why we have a check above for breakeven length | |
| 26 | ;; only r0 - r8 have to be saved, the other ones are clobber-able | |
| 27 | ;; according to the ABI | |
| 28 | ||
| 29 | subq 9*4,$sp | |
| 30 | movem $r8,[$sp] | |
| 31 | ||
| 32 | ;; do a movem checksum | |
| 33 | ||
| 34 | subq 10*4,$r11 ; update length for the first loop | |
| 35 | ||
| 36 | _mloop: movem [$r10+],$r9 ; read 10 longwords | |
| 37 | ||
| 38 | ;; perform dword checksumming on the 10 longwords | |
| 39 | ||
| 40 | add.d $r0,$r12 | |
| 41 | ax | |
| 42 | add.d $r1,$r12 | |
| 43 | ax | |
| 44 | add.d $r2,$r12 | |
| 45 | ax | |
| 46 | add.d $r3,$r12 | |
| 47 | ax | |
| 48 | add.d $r4,$r12 | |
| 49 | ax | |
| 50 | add.d $r5,$r12 | |
| 51 | ax | |
| 52 | add.d $r6,$r12 | |
| 53 | ax | |
| 54 | add.d $r7,$r12 | |
| 55 | ax | |
| 56 | add.d $r8,$r12 | |
| 57 | ax | |
| 58 | add.d $r9,$r12 | |
| 59 | ||
| 60 | ;; fold the carry into the checksum, to avoid having to loop the carry | |
| 61 | ;; back into the top | |
| 62 | ||
| 63 | ax | |
| 64 | addq 0,$r12 | |
| 1da177e4 LT |
65 | |
| 66 | subq 10*4,$r11 | |
| 67 | bge _mloop | |
| 68 | nop | |
| 69 | ||
| 70 | addq 10*4,$r11 ; compensate for last loop underflowing length | |
| 71 | ||
| 72 | movem [$sp+],$r8 ; restore regs | |
| 73 | ||
| 74 | _word_loop: | |
| 75 | ;; only fold if there is anything to fold. | |
| 76 | ||
| 77 | cmpq 0,$r12 | |
| 78 | beq _no_fold | |
| 79 | ||
| 80 | ;; fold 32-bit checksum into a 16-bit checksum, to avoid carries below. | |
| 81 | ;; r9 and r13 can be used as temporaries. | |
| 82 | ||
| 83 | moveq -1,$r9 ; put 0xffff in r9, faster than move.d 0xffff,r9 | |
| 84 | lsrq 16,$r9 | |
| 85 | ||
| 86 | move.d $r12,$r13 | |
| 87 | lsrq 16,$r13 ; r13 = checksum >> 16 | |
| 88 | and.d $r9,$r12 ; checksum = checksum & 0xffff | |
| 89 | add.d $r13,$r12 ; checksum += r13 | |
| 1da177e4 LT |
90 | |
| 91 | _no_fold: | |
| 92 | cmpq 2,$r11 | |
| 93 | blt _no_words | |
| 94 | nop | |
| 95 | ||
| 96 | ;; checksum the rest of the words | |
| 97 | ||
| 98 | subq 2,$r11 | |
| 99 | ||
| 100 | _wloop: subq 2,$r11 | |
| 101 | bge _wloop | |
| 102 | addu.w [$r10+],$r12 | |
| 103 | ||
| 104 | addq 2,$r11 | |
| 105 | ||
| 106 | _no_words: | |
| 107 | ;; see if we have one odd byte more | |
| 108 | cmpq 1,$r11 | |
| 109 | beq _do_byte | |
| 110 | nop | |
| 111 | ret | |
| 112 | move.d $r12, $r10 | |
| 113 | ||
| 114 | _do_byte: | |
| 115 | ;; copy and checksum the last byte | |
| 116 | addu.b [$r10],$r12 | |
| 117 | ret | |
| 118 | move.d $r12, $r10 | |
| 119 |