| Adrian Bunk | 88278ca | 2008-05-19 16:53:02 -0700 | [diff] [blame] | 1 | /* |
| Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2 | * memscan.S: Optimized memscan for the Sparc. |
| 3 | * |
| 4 | * Copyright (C) 1996 David S. Miller (davem@caip.rutgers.edu) |
| 5 | */ |
| 6 | |
| 7 | /* In essence, this is just a fancy strlen. */ |
| 8 | |
| 9 | #define LO_MAGIC 0x01010101 |
| 10 | #define HI_MAGIC 0x80808080 |
| 11 | |
| 12 | .text |
| 13 | .align 4 |
| 14 | .globl __memscan_zero, __memscan_generic |
| 15 | .globl memscan |
| 16 | __memscan_zero: |
| 17 | /* %o0 = addr, %o1 = size */ |
| 18 | cmp %o1, 0 |
| 19 | bne,a 1f |
| 20 | andcc %o0, 3, %g0 |
| 21 | |
| 22 | retl |
| 23 | nop |
| 24 | |
| 25 | 1: |
| 26 | be mzero_scan_word |
| 27 | sethi %hi(HI_MAGIC), %g2 |
| 28 | |
| 29 | ldsb [%o0], %g3 |
| 30 | mzero_still_not_word_aligned: |
| 31 | cmp %g3, 0 |
| 32 | bne 1f |
| 33 | add %o0, 1, %o0 |
| 34 | |
| 35 | retl |
| 36 | sub %o0, 1, %o0 |
| 37 | |
| 38 | 1: |
| 39 | subcc %o1, 1, %o1 |
| 40 | bne,a 1f |
| 41 | andcc %o0, 3, %g0 |
| 42 | |
| 43 | retl |
| 44 | nop |
| 45 | |
| 46 | 1: |
| 47 | bne,a mzero_still_not_word_aligned |
| 48 | ldsb [%o0], %g3 |
| 49 | |
| 50 | sethi %hi(HI_MAGIC), %g2 |
| 51 | mzero_scan_word: |
| 52 | or %g2, %lo(HI_MAGIC), %o3 |
| 53 | sethi %hi(LO_MAGIC), %g3 |
| 54 | or %g3, %lo(LO_MAGIC), %o2 |
| 55 | mzero_next_word: |
| 56 | ld [%o0], %g2 |
| 57 | mzero_next_word_preloaded: |
| 58 | sub %g2, %o2, %g2 |
| 59 | mzero_next_word_preloaded_next: |
| 60 | andcc %g2, %o3, %g0 |
| 61 | bne mzero_byte_zero |
| 62 | add %o0, 4, %o0 |
| 63 | |
| 64 | mzero_check_out_of_fuel: |
| 65 | subcc %o1, 4, %o1 |
| 66 | bg,a 1f |
| 67 | ld [%o0], %g2 |
| 68 | |
| 69 | retl |
| 70 | nop |
| 71 | |
| 72 | 1: |
| 73 | b mzero_next_word_preloaded_next |
| 74 | sub %g2, %o2, %g2 |
| 75 | |
| 76 | /* Check every byte. */ |
| 77 | mzero_byte_zero: |
| 78 | ldsb [%o0 - 4], %g2 |
| 79 | cmp %g2, 0 |
| 80 | bne mzero_byte_one |
| 81 | sub %o0, 4, %g3 |
| 82 | |
| 83 | retl |
| 84 | mov %g3, %o0 |
| 85 | |
| 86 | mzero_byte_one: |
| 87 | ldsb [%o0 - 3], %g2 |
| 88 | cmp %g2, 0 |
| 89 | bne,a mzero_byte_two_and_three |
| 90 | ldsb [%o0 - 2], %g2 |
| 91 | |
| 92 | retl |
| 93 | sub %o0, 3, %o0 |
| 94 | |
| 95 | mzero_byte_two_and_three: |
| 96 | cmp %g2, 0 |
| 97 | bne,a 1f |
| 98 | ldsb [%o0 - 1], %g2 |
| 99 | |
| 100 | retl |
| 101 | sub %o0, 2, %o0 |
| 102 | |
| 103 | 1: |
| 104 | cmp %g2, 0 |
| 105 | bne,a mzero_next_word_preloaded |
| 106 | ld [%o0], %g2 |
| 107 | |
| 108 | retl |
| 109 | sub %o0, 1, %o0 |
| 110 | |
| 111 | mzero_found_it: |
| 112 | retl |
| 113 | sub %o0, 2, %o0 |
| 114 | |
| 115 | memscan: |
| 116 | __memscan_generic: |
| 117 | /* %o0 = addr, %o1 = c, %o2 = size */ |
| 118 | cmp %o2, 0 |
| 119 | bne,a 0f |
| 120 | ldub [%o0], %g2 |
| 121 | |
| 122 | b,a 2f |
| 123 | 1: |
| 124 | ldub [%o0], %g2 |
| 125 | 0: |
| 126 | cmp %g2, %o1 |
| 127 | be 2f |
| 128 | addcc %o2, -1, %o2 |
| 129 | bne 1b |
| 130 | add %o0, 1, %o0 |
| 131 | 2: |
| 132 | retl |
| 133 | nop |