| Mikael Starvik | 51533b6 | 2005-07-27 11:44:44 -0700 | [diff] [blame] | 1 | /* |
| 2 | * A fast checksum routine using movem |
| Jesper Nilsson | 41f9412 | 2008-01-25 17:54:14 +0100 | [diff] [blame^] | 3 | * Copyright (c) 1998-2007 Axis Communications AB |
| Mikael Starvik | 51533b6 | 2005-07-27 11:44:44 -0700 | [diff] [blame] | 4 | * |
| 5 | * csum_partial(const unsigned char * buff, int len, unsigned int sum) |
| 6 | */ |
| 7 | |
| 8 | .globl csum_partial |
| 9 | csum_partial: |
| 10 | |
| 11 | ;; r10 - src |
| 12 | ;; r11 - length |
| 13 | ;; r12 - checksum |
| 14 | |
| Jesper Nilsson | 41f9412 | 2008-01-25 17:54:14 +0100 | [diff] [blame^] | 15 | ;; Optimized for large packets |
| 16 | subq 10*4, $r11 |
| 17 | blt _word_loop |
| 18 | move.d $r11, $acr |
| Mikael Starvik | 51533b6 | 2005-07-27 11:44:44 -0700 | [diff] [blame] | 19 | |
| 20 | subq 9*4,$sp |
| Jesper Nilsson | 41f9412 | 2008-01-25 17:54:14 +0100 | [diff] [blame^] | 21 | clearf c |
| Mikael Starvik | 51533b6 | 2005-07-27 11:44:44 -0700 | [diff] [blame] | 22 | movem $r8,[$sp] |
| 23 | |
| 24 | ;; do a movem checksum |
| 25 | |
| 26 | _mloop: movem [$r10+],$r9 ; read 10 longwords |
| Jesper Nilsson | 41f9412 | 2008-01-25 17:54:14 +0100 | [diff] [blame^] | 27 | ;; Loop count without touching the c flag. |
| 28 | addoq -10*4, $acr, $acr |
| Mikael Starvik | 51533b6 | 2005-07-27 11:44:44 -0700 | [diff] [blame] | 29 | ;; perform dword checksumming on the 10 longwords |
| 30 | |
| Jesper Nilsson | 41f9412 | 2008-01-25 17:54:14 +0100 | [diff] [blame^] | 31 | addc $r0,$r12 |
| Mikael Starvik | 51533b6 | 2005-07-27 11:44:44 -0700 | [diff] [blame] | 32 | addc $r1,$r12 |
| 33 | addc $r2,$r12 |
| 34 | addc $r3,$r12 |
| 35 | addc $r4,$r12 |
| 36 | addc $r5,$r12 |
| 37 | addc $r6,$r12 |
| 38 | addc $r7,$r12 |
| 39 | addc $r8,$r12 |
| 40 | addc $r9,$r12 |
| 41 | |
| Jesper Nilsson | 41f9412 | 2008-01-25 17:54:14 +0100 | [diff] [blame^] | 42 | ;; test $acr without trashing carry. |
| 43 | move.d $acr, $acr |
| 44 | bpl _mloop |
| 45 | ;; r11 <= acr is not really needed in the mloop, just using the dslot |
| 46 | ;; to prepare for what is needed after mloop. |
| 47 | move.d $acr, $r11 |
| Mikael Starvik | 51533b6 | 2005-07-27 11:44:44 -0700 | [diff] [blame] | 48 | |
| Jesper Nilsson | 41f9412 | 2008-01-25 17:54:14 +0100 | [diff] [blame^] | 49 | ;; fold the last carry into r13 |
| 50 | addc 0, $r12 |
| Mikael Starvik | 51533b6 | 2005-07-27 11:44:44 -0700 | [diff] [blame] | 51 | movem [$sp+],$r8 ; restore regs |
| 52 | |
| 53 | _word_loop: |
| Jesper Nilsson | 41f9412 | 2008-01-25 17:54:14 +0100 | [diff] [blame^] | 54 | addq 10*4,$r11 ; compensate for last loop underflowing length |
| Mikael Starvik | 51533b6 | 2005-07-27 11:44:44 -0700 | [diff] [blame] | 55 | |
| 56 | moveq -1,$r9 ; put 0xffff in r9, faster than move.d 0xffff,r9 |
| 57 | lsrq 16,$r9 |
| 58 | |
| 59 | move.d $r12,$r13 |
| 60 | lsrq 16,$r13 ; r13 = checksum >> 16 |
| Jesper Nilsson | 41f9412 | 2008-01-25 17:54:14 +0100 | [diff] [blame^] | 61 | and.d $r9,$r12 ; checksum = checksum & 0xffff |
| Mikael Starvik | 51533b6 | 2005-07-27 11:44:44 -0700 | [diff] [blame] | 62 | |
| 63 | _no_fold: |
| Jesper Nilsson | 41f9412 | 2008-01-25 17:54:14 +0100 | [diff] [blame^] | 64 | subq 2,$r11 |
| Mikael Starvik | 51533b6 | 2005-07-27 11:44:44 -0700 | [diff] [blame] | 65 | blt _no_words |
| Jesper Nilsson | 41f9412 | 2008-01-25 17:54:14 +0100 | [diff] [blame^] | 66 | add.d $r13,$r12 ; checksum += r13 |
| Mikael Starvik | 51533b6 | 2005-07-27 11:44:44 -0700 | [diff] [blame] | 67 | |
| 68 | ;; checksum the rest of the words |
| Mikael Starvik | 51533b6 | 2005-07-27 11:44:44 -0700 | [diff] [blame] | 69 | _wloop: subq 2,$r11 |
| 70 | bge _wloop |
| 71 | addu.w [$r10+],$r12 |
| 72 | |
| Mikael Starvik | 51533b6 | 2005-07-27 11:44:44 -0700 | [diff] [blame] | 73 | _no_words: |
| Jesper Nilsson | 41f9412 | 2008-01-25 17:54:14 +0100 | [diff] [blame^] | 74 | addq 2,$r11 |
| Mikael Starvik | 51533b6 | 2005-07-27 11:44:44 -0700 | [diff] [blame] | 75 | ;; see if we have one odd byte more |
| Jesper Nilsson | 41f9412 | 2008-01-25 17:54:14 +0100 | [diff] [blame^] | 76 | bne _do_byte |
| Mikael Starvik | 51533b6 | 2005-07-27 11:44:44 -0700 | [diff] [blame] | 77 | nop |
| 78 | ret |
| 79 | move.d $r12,$r10 |
| 80 | |
| 81 | _do_byte: |
| 82 | ;; copy and checksum the last byte |
| 83 | addu.b [$r10],$r12 |
| 84 | ret |
| 85 | move.d $r12,$r10 |