mdw@git.distorted.org.uk Git - catacomb/blob - symm/chacha-x86-sse2.S

   1 /// -*- mode: asm; asm-comment-char: ?/ -*-
   2 ///
   3 /// Fancy SIMD implementation of ChaCha
   4 ///
   5 /// (c) 2015 Straylight/Edgeware
   6 ///
   7
   8 ///----- Licensing notice ---------------------------------------------------
   9 ///
  10 /// This file is part of Catacomb.
  11 ///
  12 /// Catacomb is free software; you can redistribute it and/or modify
  13 /// it under the terms of the GNU Library General Public License as
  14 /// published by the Free Software Foundation; either version 2 of the
  15 /// License, or (at your option) any later version.
  16 ///
  17 /// Catacomb is distributed in the hope that it will be useful,
  18 /// but WITHOUT ANY WARRANTY; without even the implied warranty of
  19 /// MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  20 /// GNU Library General Public License for more details.
  21 ///
  22 /// You should have received a copy of the GNU Library General Public
  23 /// License along with Catacomb; if not, write to the Free
  24 /// Software Foundation, Inc., 59 Temple Place - Suite 330, Boston,
  25 /// MA 02111-1307, USA.
  26
  27 ///--------------------------------------------------------------------------
  28 /// External definitions.
  29
  30 #include "config.h"
  31 #include "asm-common.h"
  32
  33 ///--------------------------------------------------------------------------
  34 /// Main code.
  35
  36         .arch pentium4
  37         .section .text
  38
  39 FUNC(chacha_core_x86_sse2)
  40
  41         // Initial state.  We have three arguments:
  42         // [ebp +  8] is the number of rounds to do
  43         // [ebp + 12] points to the input matrix
  44         // [ebp + 16] points to the output matrix
  45         push    ebp
  46         mov     ebp, esp
  47         sub     esp, 16
  48         mov     edx, [ebp + 12]
  49         and     esp, ~15
  50
  51         // First job is to slurp the matrix into XMM registers.  Be careful:
  52         // the input matrix isn't likely to be properly aligned.
  53         //
  54         //      [ 0  1  2  3] (a, xmm0)
  55         //      [ 4  5  6  7] (b, xmm1)
  56         //      [ 8  9 10 11] (c, xmm2)
  57         //      [12 13 14 15] (d, xmm3)
  58         movdqu  xmm0, [edx +  0]
  59         movdqu  xmm1, [edx + 16]
  60         movdqu  xmm2, [edx + 32]
  61         movdqu  xmm3, [edx + 48]
  62
  63         // Prepare for the main loop.
  64         mov     ecx, [ebp + 8]
  65
  66         // Take a copy for later.  This one is aligned properly, by
  67         // construction.
  68         movdqa  [esp], xmm0
  69         movdqa  xmm5, xmm1
  70         movdqa  xmm6, xmm2
  71         movdqa  xmm7, xmm3
  72
  73 loop:
  74         // Apply a column quarterround to each of the columns simultaneously.
  75         // Alas, there doesn't seem to be a packed doubleword rotate, so we
  76         // have to synthesize it.
  77
  78         // a += b; d ^= a; d <<<= 16
  79         paddd   xmm0, xmm1
  80         pxor    xmm3, xmm0
  81         movdqa  xmm4, xmm3
  82         pslld   xmm3, 16
  83         psrld   xmm4, 16
  84         por     xmm3, xmm4
  85
  86         // c += d; b ^= c; b <<<= 12
  87         paddd   xmm2, xmm3
  88         pxor    xmm1, xmm2
  89         movdqa  xmm4, xmm1
  90         pslld   xmm1, 12
  91         psrld   xmm4, 20
  92         por     xmm1, xmm4
  93
  94         // a += b; d ^= a; d <<<=  8
  95         paddd   xmm0, xmm1
  96         pxor    xmm3, xmm0
  97         movdqa  xmm4, xmm3
  98         pslld   xmm3, 8
  99         psrld   xmm4, 24
 100         por     xmm3, xmm4
 101
 102         // c += d; b ^= c; b <<<=  7
 103         paddd   xmm2, xmm3
 104         pshufd  xmm3, xmm3, 0x93
 105         pxor    xmm1, xmm2
 106         pshufd  xmm2, xmm2, 0x4e
 107         movdqa  xmm4, xmm1
 108         pslld   xmm1, 7
 109         psrld   xmm4, 25
 110         por     xmm1, xmm4
 111
 112         // The not-quite-transpose conveniently only involves reordering
 113         // elements of individual rows, which can be done quite easily.  It
 114         // doesn't involve any movement of elements between rows, or even
 115         // renaming of the rows.
 116         //
 117         //      [ 0  1  2  3]           [ 0  1  2  3] (a, xmm0)
 118         //      [ 4  5  6  7]    -->    [ 5  6  7  4] (b, xmm1)
 119         //      [ 8  9 10 11]           [10 11  8  9] (c, xmm2)
 120         //      [12 13 14 15]           [15 12 13 14] (d, xmm3)
 121         //
 122         // The shuffles have quite high latency, so they've mostly been
 123         // pushed upwards.  The remaining one can't be moved, though.
 124         pshufd  xmm1, xmm1, 0x39
 125
 126         // Apply the diagonal quarterround to each of the columns
 127         // simultaneously.
 128
 129         // a += b; d ^= a; d <<<= 16
 130         paddd   xmm0, xmm1
 131         pxor    xmm3, xmm0
 132         movdqa  xmm4, xmm3
 133         pslld   xmm3, 16
 134         psrld   xmm4, 16
 135         por     xmm3, xmm4
 136
 137         // c += d; b ^= c; b <<<= 12
 138         paddd   xmm2, xmm3
 139         pxor    xmm1, xmm2
 140         movdqa  xmm4, xmm1
 141         pslld   xmm1, 12
 142         psrld   xmm4, 20
 143         por     xmm1, xmm4
 144
 145         // a += b; d ^= a; d <<<=  8
 146         paddd   xmm0, xmm1
 147         pxor    xmm3, xmm0
 148         movdqa  xmm4, xmm3
 149         pslld   xmm3, 8
 150         psrld   xmm4, 24
 151         por     xmm3, xmm4
 152
 153         // c += d; b ^= c; b <<<=  7
 154         paddd   xmm2, xmm3
 155         pshufd  xmm3, xmm3, 0x39
 156         pxor    xmm1, xmm2
 157         pshufd  xmm2, xmm2, 0x4e
 158         movdqa  xmm4, xmm1
 159         pslld   xmm1, 7
 160         psrld   xmm4, 25
 161         por     xmm1, xmm4
 162
 163         // Finally, finish off undoing the transpose, and we're done for this
 164         // doubleround.  Again, most of this was done above so we don't have
 165         // to wait for the shuffles.
 166         pshufd  xmm1, xmm1, 0x93
 167
 168         // Decrement the loop counter and see if we should go round again.
 169         sub     ecx, 2
 170         ja      loop
 171
 172         // Almost there.  Firstly, the feedforward addition.
 173         mov     edx, [ebp + 16]
 174         paddd   xmm0, [esp]
 175         paddd   xmm1, xmm5
 176         paddd   xmm2, xmm6
 177         paddd   xmm3, xmm7
 178
 179         // And now we write out the result.  This one won't be aligned
 180         // either.
 181         movdqu  [edx +  0], xmm0
 182         movdqu  [edx + 16], xmm1
 183         movdqu  [edx + 32], xmm2
 184         movdqu  [edx + 48], xmm3
 185
 186         // Tidy things up.
 187         mov     esp, ebp
 188         pop     ebp
 189
 190         // And with that, we're done.
 191         ret
 192
 193 ENDFUNC
 194
 195 ///----- That's all, folks --------------------------------------------------