release/src/router/nettle/sparc64/arcfour-crypt.asm

   1 C -*- mode: asm; asm-comment-char: ?C; -*-
   2 C nettle, low-level cryptographics library
   3 C
   4 C Copyright (C) 2002, 2005 Niels Möller
   5 C
   6 C The nettle library is free software; you can redistribute it and/or modify
   7 C it under the terms of the GNU Lesser General Public License as published by
   8 C the Free Software Foundation; either version 2.1 of the License, or (at your
   9 C option) any later version.
  10 C
  11 C The nettle library is distributed in the hope that it will be useful, but
  12 C WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY
  13 C or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU Lesser General Public
  14 C License for more details.
  15 C
  16 C You should have received a copy of the GNU Lesser General Public License
  17 C along with the nettle library; see the file COPYING.LIB.  If not, write to
  18 C the Free Software Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston,
  19 C MA 02111-1301, USA.
  20
  21 C       Define to YES, to enable the complex code to special case SRC
  22 C       and DST with compatible alignment.
  23
  24 define(<WITH_ALIGN>, <YES>)
  25
  26 C       Registers
  27
  28 define(<CTX>,   <%i0>)
  29 define(<LENGTH>,<%i1>)
  30 define(<DST>,   <%i2>)
  31 define(<SRC>,   <%i3>)
  32
  33 define(<I1>,    <%i4>)
  34 define(<I2>,    <%i5>)
  35 define(<J>,     <%g1>)
  36 define(<SI>,    <%g2>)
  37 define(<SJ>,    <%g3>)
  38 define(<TMP>,   <%o0>)
  39 define(<TMP2>,  <%o1>)
  40 define(<N>,     <%o2>)
  41 define(<DATA>,  <%o3>)
  42
  43 C       Computes the next byte of the key stream. As input, i must
  44 C       already point to the index for the current access, the index
  45 C       for the next access is stored in ni. The resulting key byte is
  46 C       stored in res.
  47 C       ARCFOUR_BYTE(i, ni, res)
  48 define(<ARCFOUR_BYTE>, <
  49         ldub    [CTX + $1], SI
  50         add     $1, 1, $2
  51         add     J, SI, J
  52         and     J, 0xff, J
  53         ldub    [CTX + J], SJ
  54         and     $2, 0xff, $2
  55         stb     SI, [CTX + J]
  56         add     SI, SJ, SI
  57         and     SI, 0xff, SI
  58         stb     SJ, [CTX + $1]
  59         ldub    [CTX + SI], $3
  60 >)dnl
  61
  62 define(<FRAME_SIZE>, 192)
  63
  64         .file "arcfour-crypt.asm"
  65
  66         C arcfour_crypt(struct arcfour_ctx *ctx,
  67         C               unsigned length, uint8_t *dst,
  68         C               const uint8_t *src)
  69
  70         .section        ".text"
  71         .align 16
  72         .proc   020
  73
  74 PROLOGUE(nettle_arcfour_crypt)
  75
  76         save    %sp, -FRAME_SIZE, %sp
  77         cmp     LENGTH, 0
  78         be      .Lend
  79         nop
  80
  81         C       Load both I and J
  82         lduh    [CTX + ARCFOUR_I], I1
  83         and     I1, 0xff, J
  84         srl     I1, 8, I1
  85
  86         C       We want an even address for DST
  87         andcc   DST, 1, %g0
  88         add     I1, 1 ,I1
  89         beq     .Laligned2
  90         and     I1, 0xff, I1
  91
  92         mov     I1, I2
  93         ldub    [SRC], DATA
  94         ARCFOUR_BYTE(I2, I1, TMP)
  95         subcc   LENGTH, 1, LENGTH
  96         add     SRC, 1, SRC
  97         xor     DATA, TMP, DATA
  98         stb     DATA, [DST]
  99         beq     .Ldone
 100         add     DST, 1, DST
 101
 102 .Laligned2:
 103
 104         cmp     LENGTH, 2
 105         blu     .Lfinal1
 106         C       Harmless delay slot instruction
 107         andcc   DST, 2, %g0
 108         beq     .Laligned4
 109         nop
 110
 111         ldub    [SRC], DATA
 112         ARCFOUR_BYTE(I1, I2, TMP)
 113         ldub    [SRC + 1], TMP2
 114         add     SRC, 2, SRC
 115         xor     DATA, TMP, DATA
 116         sll     DATA, 8, DATA
 117
 118         ARCFOUR_BYTE(I2, I1, TMP)
 119         xor     TMP2, TMP, TMP
 120         subcc   LENGTH, 2, LENGTH
 121         or      DATA, TMP, DATA
 122
 123         sth     DATA, [DST]
 124         beq     .Ldone
 125         add     DST, 2, DST
 126
 127 .Laligned4:
 128         cmp     LENGTH, 4
 129         blu     .Lfinal2
 130         C       Harmless delay slot instruction
 131         srl     LENGTH, 2, N
 132
 133 .Loop:
 134         C       Main loop, with aligned writes
 135
 136         C       FIXME: Could check if SRC is aligned, and
 137         C       use 32-bit reads in that case.
 138
 139         ldub    [SRC], DATA
 140         ARCFOUR_BYTE(I1, I2, TMP)
 141         ldub    [SRC + 1], TMP2
 142         xor     TMP, DATA, DATA
 143         sll     DATA, 8, DATA
 144
 145         ARCFOUR_BYTE(I2, I1, TMP)
 146         xor     TMP2, TMP, TMP
 147         ldub    [SRC + 2], TMP2
 148         or      TMP, DATA, DATA
 149         sll     DATA, 8, DATA
 150
 151         ARCFOUR_BYTE(I1, I2, TMP)
 152         xor     TMP2, TMP, TMP
 153         ldub    [SRC + 3], TMP2
 154         or      TMP, DATA, DATA
 155         sll     DATA, 8, DATA
 156
 157         ARCFOUR_BYTE(I2, I1, TMP)
 158         xor     TMP2, TMP, TMP
 159         or      TMP, DATA, DATA
 160         subcc   N, 1, N
 161         add     SRC, 4, SRC
 162         st      DATA, [DST]
 163         bne     .Loop
 164         add     DST, 4, DST
 165
 166         andcc   LENGTH, 3, LENGTH
 167         beq     .Ldone
 168         nop
 169
 170 .Lfinal2:
 171         C       DST address must be 2-aligned
 172         cmp     LENGTH, 2
 173         blu     .Lfinal1
 174         nop
 175
 176         ldub    [SRC], DATA
 177         ARCFOUR_BYTE(I1, I2, TMP)
 178         ldub    [SRC + 1], TMP2
 179         add     SRC, 2, SRC
 180         xor     DATA, TMP, DATA
 181         sll     DATA, 8, DATA
 182
 183         ARCFOUR_BYTE(I2, I1, TMP)
 184         xor     TMP2, TMP, TMP
 185         or      DATA, TMP, DATA
 186
 187         sth     DATA, [DST]
 188         beq     .Ldone
 189         add     DST, 2, DST
 190
 191 .Lfinal1:
 192         mov     I1, I2
 193         ldub    [SRC], DATA
 194         ARCFOUR_BYTE(I2, I1, TMP)
 195         xor     DATA, TMP, DATA
 196         stb     DATA, [DST]
 197
 198 .Ldone:
 199         C       Save back I and J
 200         sll     I2, 8, I2
 201         or      I2, J, I2
 202         stuh    I2, [CTX + ARCFOUR_I]
 203
 204 .Lend:
 205         ret
 206         restore
 207
 208 EPILOGUE(nettle_arcfour_crypt)
 209
 210 C       Stats for AES 128 on sellafield.lysator.liu.se (UE450, 296 MHz)
 211
 212 C 1:    nettle-1.13 C-code
 213 C 2:    New assembler code (basically the same as for sparc32)
 214
 215 C       MB/s    cycles/byte
 216 C 1:    3.6     77.7
 217 C 2:    21.8    13.0