| Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1 | /* checksum.S: Sparc V9 optimized checksum code. |
| 2 | * |
| 3 | * Copyright(C) 1995 Linus Torvalds |
| 4 | * Copyright(C) 1995 Miguel de Icaza |
| 5 | * Copyright(C) 1996, 2000 David S. Miller |
| 6 | * Copyright(C) 1997 Jakub Jelinek |
| 7 | * |
| 8 | * derived from: |
| 9 | * Linux/Alpha checksum c-code |
| 10 | * Linux/ix86 inline checksum assembly |
| 11 | * RFC1071 Computing the Internet Checksum (esp. Jacobsons m68k code) |
| 12 | * David Mosberger-Tang for optimized reference c-code |
| 13 | * BSD4.4 portable checksum routine |
| 14 | */ |
| 15 | |
| Al Viro | d3867f04 | 2016-01-16 21:39:30 -0500 | [diff] [blame] | 16 | #include <asm/export.h> |
| Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 17 | .text |
| 18 | |
| 19 | csum_partial_fix_alignment: |
| 20 | /* We checked for zero length already, so there must be |
| 21 | * at least one byte. |
| 22 | */ |
| 23 | be,pt %icc, 1f |
| 24 | nop |
| 25 | ldub [%o0 + 0x00], %o4 |
| 26 | add %o0, 1, %o0 |
| 27 | sub %o1, 1, %o1 |
| 28 | 1: andcc %o0, 0x2, %g0 |
| 29 | be,pn %icc, csum_partial_post_align |
| 30 | cmp %o1, 2 |
| 31 | blu,pn %icc, csum_partial_end_cruft |
| 32 | nop |
| 33 | lduh [%o0 + 0x00], %o5 |
| 34 | add %o0, 2, %o0 |
| 35 | sub %o1, 2, %o1 |
| 36 | ba,pt %xcc, csum_partial_post_align |
| 37 | add %o5, %o4, %o4 |
| 38 | |
| 39 | .align 32 |
| 40 | .globl csum_partial |
| Al Viro | d3867f04 | 2016-01-16 21:39:30 -0500 | [diff] [blame] | 41 | EXPORT_SYMBOL(csum_partial) |
| Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 42 | csum_partial: /* %o0=buff, %o1=len, %o2=sum */ |
| 43 | prefetch [%o0 + 0x000], #n_reads |
| 44 | clr %o4 |
| 45 | prefetch [%o0 + 0x040], #n_reads |
| 46 | brz,pn %o1, csum_partial_finish |
| 47 | andcc %o0, 0x3, %g0 |
| 48 | |
| 49 | /* We "remember" whether the lowest bit in the address |
| 50 | * was set in %g7. Because if it is, we have to swap |
| 51 | * upper and lower 8 bit fields of the sum we calculate. |
| 52 | */ |
| 53 | bne,pn %icc, csum_partial_fix_alignment |
| 54 | andcc %o0, 0x1, %g7 |
| 55 | |
| 56 | csum_partial_post_align: |
| 57 | prefetch [%o0 + 0x080], #n_reads |
| 58 | andncc %o1, 0x3f, %o3 |
| 59 | |
| 60 | prefetch [%o0 + 0x0c0], #n_reads |
| 61 | sub %o1, %o3, %o1 |
| 62 | brz,pn %o3, 2f |
| 63 | prefetch [%o0 + 0x100], #n_reads |
| 64 | |
| 65 | /* So that we don't need to use the non-pairing |
| 66 | * add-with-carry instructions we accumulate 32-bit |
| 67 | * values into a 64-bit register. At the end of the |
| 68 | * loop we fold it down to 32-bits and so on. |
| 69 | */ |
| 70 | prefetch [%o0 + 0x140], #n_reads |
| 71 | 1: lduw [%o0 + 0x00], %o5 |
| 72 | lduw [%o0 + 0x04], %g1 |
| 73 | lduw [%o0 + 0x08], %g2 |
| 74 | add %o4, %o5, %o4 |
| 75 | lduw [%o0 + 0x0c], %g3 |
| 76 | add %o4, %g1, %o4 |
| 77 | lduw [%o0 + 0x10], %o5 |
| 78 | add %o4, %g2, %o4 |
| 79 | lduw [%o0 + 0x14], %g1 |
| 80 | add %o4, %g3, %o4 |
| 81 | lduw [%o0 + 0x18], %g2 |
| 82 | add %o4, %o5, %o4 |
| 83 | lduw [%o0 + 0x1c], %g3 |
| 84 | add %o4, %g1, %o4 |
| 85 | lduw [%o0 + 0x20], %o5 |
| 86 | add %o4, %g2, %o4 |
| 87 | lduw [%o0 + 0x24], %g1 |
| 88 | add %o4, %g3, %o4 |
| 89 | lduw [%o0 + 0x28], %g2 |
| 90 | add %o4, %o5, %o4 |
| 91 | lduw [%o0 + 0x2c], %g3 |
| 92 | add %o4, %g1, %o4 |
| 93 | lduw [%o0 + 0x30], %o5 |
| 94 | add %o4, %g2, %o4 |
| 95 | lduw [%o0 + 0x34], %g1 |
| 96 | add %o4, %g3, %o4 |
| 97 | lduw [%o0 + 0x38], %g2 |
| 98 | add %o4, %o5, %o4 |
| 99 | lduw [%o0 + 0x3c], %g3 |
| 100 | add %o4, %g1, %o4 |
| 101 | prefetch [%o0 + 0x180], #n_reads |
| 102 | add %o4, %g2, %o4 |
| 103 | subcc %o3, 0x40, %o3 |
| 104 | add %o0, 0x40, %o0 |
| 105 | bne,pt %icc, 1b |
| 106 | add %o4, %g3, %o4 |
| 107 | |
| 108 | 2: and %o1, 0x3c, %o3 |
| 109 | brz,pn %o3, 2f |
| 110 | sub %o1, %o3, %o1 |
| 111 | 1: lduw [%o0 + 0x00], %o5 |
| 112 | subcc %o3, 0x4, %o3 |
| 113 | add %o0, 0x4, %o0 |
| 114 | bne,pt %icc, 1b |
| 115 | add %o4, %o5, %o4 |
| 116 | |
| 117 | 2: |
| 118 | /* fold 64-->32 */ |
| 119 | srlx %o4, 32, %o5 |
| 120 | srl %o4, 0, %o4 |
| 121 | add %o4, %o5, %o4 |
| 122 | srlx %o4, 32, %o5 |
| 123 | srl %o4, 0, %o4 |
| 124 | add %o4, %o5, %o4 |
| 125 | |
| 126 | /* fold 32-->16 */ |
| 127 | sethi %hi(0xffff0000), %g1 |
| 128 | srl %o4, 16, %o5 |
| 129 | andn %o4, %g1, %g2 |
| 130 | add %o5, %g2, %o4 |
| 131 | srl %o4, 16, %o5 |
| 132 | andn %o4, %g1, %g2 |
| 133 | add %o5, %g2, %o4 |
| 134 | |
| 135 | csum_partial_end_cruft: |
| 136 | /* %o4 has the 16-bit sum we have calculated so-far. */ |
| 137 | cmp %o1, 2 |
| 138 | blu,pt %icc, 1f |
| 139 | nop |
| 140 | lduh [%o0 + 0x00], %o5 |
| 141 | sub %o1, 2, %o1 |
| 142 | add %o0, 2, %o0 |
| 143 | add %o4, %o5, %o4 |
| 144 | 1: brz,pt %o1, 1f |
| 145 | nop |
| 146 | ldub [%o0 + 0x00], %o5 |
| 147 | sub %o1, 1, %o1 |
| 148 | add %o0, 1, %o0 |
| 149 | sllx %o5, 8, %o5 |
| 150 | add %o4, %o5, %o4 |
| 151 | 1: |
| 152 | /* fold 32-->16 */ |
| 153 | sethi %hi(0xffff0000), %g1 |
| 154 | srl %o4, 16, %o5 |
| 155 | andn %o4, %g1, %g2 |
| 156 | add %o5, %g2, %o4 |
| 157 | srl %o4, 16, %o5 |
| 158 | andn %o4, %g1, %g2 |
| 159 | add %o5, %g2, %o4 |
| 160 | |
| 161 | 1: brz,pt %g7, 1f |
| 162 | nop |
| 163 | |
| 164 | /* We started with an odd byte, byte-swap the result. */ |
| 165 | srl %o4, 8, %o5 |
| 166 | and %o4, 0xff, %g1 |
| 167 | sll %g1, 8, %g1 |
| 168 | or %o5, %g1, %o4 |
| 169 | |
| David S. Miller | ae5de0f | 2006-06-04 21:32:01 -0700 | [diff] [blame] | 170 | 1: addcc %o2, %o4, %o2 |
| 171 | addc %g0, %o2, %o2 |
| Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 172 | |
| 173 | csum_partial_finish: |
| 174 | retl |
| David S. Miller | ae5de0f | 2006-06-04 21:32:01 -0700 | [diff] [blame] | 175 | srl %o2, 0, %o0 |