release/src/router/nettle/arm/neon/sha3-permute.asm

   1 C nettle, low-level cryptographics library
   2 C
   3 C Copyright (C) 2013 Niels Möller
   4 C
   5 C The nettle library is free software; you can redistribute it and/or modify
   6 C it under the terms of the GNU Lesser General Public License as published by
   7 C the Free Software Foundation; either version 2.1 of the License, or (at your
   8 C option) any later version.
   9 C
  10 C The nettle library is distributed in the hope that it will be useful, but
  11 C WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY
  12 C or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU Lesser General Public
  13 C License for more details.
  14 C
  15 C You should have received a copy of the GNU Lesser General Public License
  16 C along with the nettle library; see the file COPYING.LIB.  If not, write to
  17 C the Free Software Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston,
  18 C MA 02111-1301, USA.
  19
  20         .file "sha3-permute.asm"
  21         .fpu    neon
  22
  23 define(<CTX>, <r0>)
  24 define(<COUNT>, <r1>)
  25 define(<RC>, <r2>)
  26 C First column
  27 define(<A0>, <d0>)
  28 define(<A5>, <d2>)
  29 define(<A10>, <d3>)
  30 define(<A15>, <d4>)
  31 define(<A20>, <d5>)
  32
  33 define(<A1>, <d6>)
  34 define(<A2>, <d7>)
  35 define(<A3>, <d8>)
  36 define(<A4>, <d9>)
  37
  38 define(<A6>, <d16>)
  39 define(<A7>, <d17>)
  40 define(<A8>, <d18>)
  41 define(<A9>, <d19>)
  42
  43 define(<A11>, <d20>)
  44 define(<A12>, <d21>)
  45 define(<A13>, <d22>)
  46 define(<A14>, <d23>)
  47
  48 define(<A16>, <d24>)
  49 define(<A17>, <d25>)
  50 define(<A18>, <d26>)
  51 define(<A19>, <d27>)
  52
  53 define(<A21>, <d28>)
  54 define(<A22>, <d29>)
  55 define(<A23>, <d30>)
  56 define(<A24>, <d31>)
  57
  58 define(<T0>, <d10>)
  59 define(<T1>, <d11>)
  60
  61 define(<C0>, <d1>)
  62 define(<C1>, <d12>)
  63 define(<C2>, <d13>)
  64 define(<C3>, <d14>)
  65 define(<C4>, <d15>)
  66
  67
  68 C ROL(DST, SRC, COUNT)
  69 C Must have SRC != DST
  70 define(<ROL>, <
  71         vshr.u64        $1, $2, #eval(64-$3)
  72         vsli.i64        $1, $2, #$3
  73         >)
  74 C sha3_permute(struct sha3_ctx *ctx)
  75
  76         .text
  77         .align  3
  78 .Lrc:
  79         .quad   0x0000000000000001
  80         .quad   0x0000000000008082
  81         .quad   0x800000000000808A
  82         .quad   0x8000000080008000
  83         .quad   0x000000000000808B
  84         .quad   0x0000000080000001
  85         .quad   0x8000000080008081
  86         .quad   0x8000000000008009
  87         .quad   0x000000000000008A
  88         .quad   0x0000000000000088
  89         .quad   0x0000000080008009
  90         .quad   0x000000008000000A
  91         .quad   0x000000008000808B
  92         .quad   0x800000000000008B
  93         .quad   0x8000000000008089
  94         .quad   0x8000000000008003
  95         .quad   0x8000000000008002
  96         .quad   0x8000000000000080
  97         .quad   0x000000000000800A
  98         .quad   0x800000008000000A
  99         .quad   0x8000000080008081
 100         .quad   0x8000000000008080
 101         .quad   0x0000000080000001
 102         .quad   0x8000000080008008
 103
 104 PROLOGUE(nettle_sha3_permute)
 105         vpush   {d8-d15}
 106
 107         vld1.64 {A0}, [CTX]!
 108         vldm    CTX!, {A1,A2,A3,A4}
 109         vld1.64 {A5}, [CTX]!
 110         vldm    CTX!, {A6,A7,A8,A9}
 111         vld1.64 {A10}, [CTX]!
 112         vldm    CTX!, {A11,A12,A13,A14}
 113         vld1.64 {A15}, [CTX]!
 114         vldm    CTX!, {A16,A17,A18,A19}
 115         vld1.64 {A20}, [CTX]!
 116         vldm    CTX, {A21,A22,A23,A24}
 117         sub     CTX, CTX, #168
 118
 119         mov     COUNT, #24
 120         adr     RC, .Lrc
 121
 122         .align 3
 123 .Loop:
 124         veor    QREG(T0), QREG(A5), QREG(A15)
 125         veor    C0, A0, T0
 126         veor    C0, C0, T1
 127         veor    QREG(C1), QREG(A1), QREG(A6)
 128         veor    QREG(C1), QREG(C1), QREG(A11)
 129         veor    QREG(C1), QREG(C1), QREG(A16)
 130         veor    QREG(C1), QREG(C1), QREG(A21)
 131
 132         veor    QREG(C3), QREG(A3), QREG(A8)
 133         veor    QREG(C3), QREG(C3), QREG(A13)
 134         veor    QREG(C3), QREG(C3), QREG(A18)
 135         veor    QREG(C3), QREG(C3), QREG(A23)
 136
 137         C       D0 = C4 ^ (C1 <<< 1)
 138         C       NOTE: Using ROL macro (and vsli) is slightly slower.
 139         vshl.i64        T0, C1, #1
 140         vshr.u64        T1, C1, #63
 141         veor    T0, T0, C4
 142         veor    T0, T0, T1
 143         vmov    T1, T0
 144         veor    A0, A0, T0
 145         veor    QREG(A5), QREG(A5), QREG(T0)
 146         veor    QREG(A15), QREG(A15), QREG(T0)
 147
 148         C       D1 = C0 ^ (C2 <<< 1)
 149         C       D2 = C1 ^ (C3 <<< 1)
 150         ROL(T0, C2, 1)
 151         ROL(T1, C3, 1)
 152         veor    T0, T0, C0
 153         veor    T1, T1, C1
 154         veor    QREG(A1), QREG(A1), QREG(T0)
 155         veor    QREG(A6), QREG(A6), QREG(T0)
 156         veor    QREG(A11), QREG(A11), QREG(T0)
 157         veor    QREG(A16), QREG(A16), QREG(T0)
 158         veor    QREG(A21), QREG(A21), QREG(T0)
 159
 160         C       D3 = C2 ^ (C4 <<< 1)
 161         C       D4 = C3 ^ (C0 <<< 1)
 162         ROL(T0, C4, 1)
 163         ROL(T1, C0, 1)
 164         veor    T0, T0, C2
 165         veor    T1, T1, C3
 166         veor    QREG(A3), QREG(A3), QREG(T0)
 167         veor    QREG(A8), QREG(A8), QREG(T0)
 168         veor    QREG(A13), QREG(A13), QREG(T0)
 169         veor    QREG(A18), QREG(A18), QREG(T0)
 170         veor    QREG(A23), QREG(A23), QREG(T0)
 171
 172         ROL( T0,  A1,  1)
 173         ROL( A1,  A6, 44)
 174         ROL( A6,  A9, 20)
 175         ROL( A9, A22, 61)
 176         ROL(A22, A14, 39)
 177         ROL(A14, A20, 18)
 178         ROL(A20,  A2, 62)
 179         ROL( A2, A12, 43)
 180         ROL(A12, A13, 25)
 181         ROL(A13, A19,  8)
 182         ROL(A19, A23, 56)
 183         ROL(A23, A15, 41)
 184         ROL(A15,  A4, 27)
 185         ROL( A4, A24, 14)
 186         ROL(A24, A21,  2)
 187         ROL(A21,  A8, 55)
 188         ROL( A8, A16, 45)
 189         ROL(A16,  A5, 36)
 190         ROL( A5,  A3, 28)
 191         ROL( A3, A18, 21)
 192         ROL(A18, A17, 15)
 193         ROL(A17, A11, 10)
 194         ROL(A11,  A7,  6)
 195         ROL( A7, A10,  3)
 196         C New A10 value left in T0
 197
 198         vbic    C0, A2, A1
 199         vbic    C1, A3, A2
 200         vbic    C2, A4, A3
 201         vbic    C3, A0, A4
 202         vbic    C4, A1, A0
 203
 204         veor    A0, A0, C0
 205         vld1.64 {C0}, [RC :64]!
 206         veor    QREG(A1), QREG(A1), QREG(C1)
 207         veor    QREG(A3), QREG(A3), QREG(C3)
 208         veor    A0, A0, C0
 209
 210         vbic    C0, A7, A6
 211         vbic    C1, A8, A7
 212         vbic    C2, A9, A8
 213         vbic    C3, A5, A9
 214         vbic    C4, A6, A5
 215
 216         veor    A5, A5, C0
 217         veor    QREG(A6), QREG(A6), QREG(C1)
 218         veor    QREG(A8), QREG(A8), QREG(C3)
 219
 220         vbic    C0, A12, A11
 221         vbic    C1, A13, A12
 222         vbic    C2, A14, A13
 223         vbic    C3, T0, A14
 224         vbic    C4, A11, T0
 225
 226         veor    A10, T0, C0
 227         veor    QREG(A11), QREG(A11), QREG(C1)
 228         veor    QREG(A13), QREG(A13), QREG(C3)
 229
 230         vbic    C0, A17, A16
 231         vbic    C1, A18, A17
 232         vbic    C2, A19, A18
 233         vbic    C3, A15, A19
 234         vbic    C4, A16, A15
 235
 236         veor    A15, A15, C0
 237         veor    QREG(A16), QREG(A16), QREG(C1)
 238         veor    QREG(A18), QREG(A18), QREG(C3)
 239
 240         vbic    C0, A22, A21
 241         vbic    C1, A23, A22
 242         vbic    C2, A24, A23
 243         vbic    C3, A20, A24
 244         vbic    C4, A21, A20
 245
 246         subs    COUNT, COUNT, #1
 247         veor    A20, A20, C0
 248         veor    QREG(A21), QREG(A21), QREG(C1)
 249         veor    QREG(A23), QREG(A23), QREG(C3)
 250
 251         bne     .Loop
 252
 253         vst1.64 {A0}, [CTX]!
 254         vstm    CTX!, {A1,A2,A3,A4}
 255         vst1.64 {A5}, [CTX]!
 256         vstm    CTX!, {A6,A7,A8,A9}
 257         vst1.64 {A10}, [CTX]!
 258         vstm    CTX!, {A11,A12,A13,A14}
 259         vst1.64 {A15}, [CTX]!
 260         vstm    CTX!, {A16,A17,A18,A19}
 261         vst1.64 {A20}, [CTX]!
 262         vstm    CTX, {A21,A22,A23,A24}
 263
 264         vpop    {d8-d15}
 265         bx      lr
 266 EPILOGUE(nettle_sha3_permute)