diff --git a/sparc64/aes.asm b/sparc64/aes.asm new file mode 100644 index 0000000000000000000000000000000000000000..248455bd7d1800660cfcd7558ac97e43b1ea8182 --- /dev/null +++ b/sparc64/aes.asm @@ -0,0 +1,369 @@ +! -*- mode: asm; asm-comment-char: ?!; -*- +! nettle, low-level cryptographics library +! +! Copyright (C) 2002 Niels Möller +! +! The nettle library is free software; you can redistribute it and/or modify +! it under the terms of the GNU Lesser General Public License as published by +! the Free Software Foundation; either version 2.1 of the License, or (at your +! option) any later version. +! +! The nettle library is distributed in the hope that it will be useful, but +! WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY +! or FITNESS FOR A PARTICULAR PURPOSE. See the GNU Lesser General Public +! License for more details. +! +! You should have received a copy of the GNU Lesser General Public License +! along with the nettle library; see the file COPYING.LIB. If not, write to +! the Free Software Foundation, Inc., 59 Temple Place - Suite 330, Boston, +! MA 02111-1307, USA. + +! The only difference between this code and the sparc32 code is the +! frame offsets, and the magic BIAS when accessing the stack. + +! FIXME: For improved ultra sparc performance, we should avoid ALU +! instructions that use the result of an immediately preceeding ALU +! instruction. It is also a good idea to have a greater distance than +! one instruction between a load and use of its value, as that reduces +! the penalty for cache misses. Such instruction sequences are marked +! with !U comments. + +! NOTE: Some of the %g registers are reserved for operating system etc +! (see gcc/config/sparc.h). The only %g registers that seems safe to +! use are %g1-%g3. + + ! Used registers: %l0,1,2,3,4,5,6,7 + ! %i0,1,2,3,4 (%i6=%fp, %i7 = return) + ! %o0,1,2,3,4 (%o6=%sp) + ! + + .file "aes.asm" + +! Arguments +define(ctx, %i0) +define(T, %i1) +define(length, %i2) +define(dst, %i3) +define(src, %i4) + +! Loop invariants +define(wtxt, %l0) +define(tmp, %l1) +define(diff, %l2) +define(nrounds, %l3) + +! Further loop invariants +define(T0, %l4) +define(T1, %l5) +define(T2, %l6) +define(T3, %l7) + +! Teporaries +define(t0, %o0) +define(t1, %o1) +define(t2, %o2) + +! Loop variables +define(round, %o3) +define(key, %o4) + +C IDX1 cointains the permutation values * 4 + 2 +define(IDX1, <T + AES_SIDX1 >) +C IDX3 cointains the permutation values * 4 +define(IDX3, <T + AES_SIDX3 >) + + +C AES_LOAD(i) +C Get one word of input, XOR with first subkey, store in wtxt +define(<AES_LOAD>, < + ldub [src+$1], t0 + ldub [src+$1+1], t1 + ldub [src+$1+2], t2 + sll t1, 8, t1 + + or t0, t1, t0 ! U + ldub [src+$1+3], t1 + sll t2, 16, t2 + or t0, t2, t0 + + sll t1, 24, t1 + ! Get subkey + ld [ctx + $1], t2 + or t0, t1, t0 + xor t0, t2, t0 + + st t0, [wtxt+$1]>)dnl + +C AES_ROUND(i) +C Compute one word in the round function. +C Input in wtxt, output stored in tmp + i. +C +C The comments mark which j in T->table[j][ Bj(wtxt[IDXi(i)]) ] +C the instruction is a part of. +define(<AES_ROUND>, < + ld [IDX1+$1], t1 ! 1 + ldub [wtxt+$1+3], t0 ! 0 + ldub [wtxt+t1], t1 ! 1 + sll t0, 2, t0 ! 0 + + ld [T0+t0], t0 ! 0 + sll t1, 2, t1 ! 1 + ld [T1+t1], t1 ! 1 !U + ld [IDX3+$1], t2 ! 3 + + xor t0, t1, t0 ! 0, 1 + ! IDX2(j) = j XOR 2 + ldub [wtxt+eval($1 ^ 8)+1], t1 ! 2 + ldub [wtxt+t2], t2 ! 3 + sll t1, 2, t1 ! 2 + + ld [T2+t1], t1 ! 2 !U + sll t2, 2, t2 ! 3 + ld [T3+t2], t2 ! 3 !U + xor t0, t1, t0 ! 0, 1, 2 + + ! Fetch roundkey + ld [key + $1], t1 + xor t0, t2, t0 ! 0, 1, 2, 3 + xor t0, t1, t0 !U + st t0, [tmp + $1]>)dnl + +C AES_FINAL_ROUND(i) +C Compute one word in the final round function. +C Input in wtxt, output converted to an octet string and stored at dst. +C +C The comments mark which j in T->table[j][ Bj(wtxt[IDXi(i)]) ] +C the instruction is a part of. +define(<AES_FINAL_ROUND>, < + ld [IDX1+$1], t1 ! 1 + ldub [wtxt+$1+3], t0 ! 0 + ldub [wtxt+t1], t1 ! 1 + ldub [T+t0], t0 ! 0 + + ldub [T+t1], t1 ! 1 + ld [IDX3 + $1], t2 ! 3 + sll t1, 8, t1 ! 1 + or t0, t1, t0 ! 0, 1 !U + + ! IDX2(j) = j XOR 2 + ldub [wtxt+eval($1 ^ 8)+1], t1 ! 2 + ldub [wtxt+t2], t2 ! 3 + ldub [T+t1], t1 ! 2 + ldub [T+t2], t2 ! 3 + + sll t1, 16, t1 ! 2 + or t0, t1, t0 ! 0, 1, 2 !U + sll t2, 24, t2 ! 3 + ld [key + $1], t1 + + or t0, t2, t0 ! 0, 1, 2, 3 + xor t0, t1, t0 !U + srl t0, 24, t1 !U + stb t1, [dst+$1+3] !U + + srl t0, 16, t1 + stb t1, [dst+$1+2] !U + srl t0, 8, t1 + stb t1, [dst+$1+1] !U + + stb t0, [dst+$1]>)dnl + +C The stack frame looks like +C +C %fp - 8: OS-dependent link field +C %fp - 16: OS-dependent link field +C %fp - 32: tmp, uint32_t[4] +C %fp - 48: wtxt, uint32_t[4] +C %fp - 224: OS register save area. (22*8 == 176 bytes) +define(<FRAME_SIZE>, 224) +define(<BIAS>, 2047) C Magic stack bias for the Sparc64 ABI + + .section ".text" + .align 16 + .global _nettle_aes_crypt + .type _nettle_aes_crypt,#function + .proc 020 + +_nettle_aes_crypt: + save %sp, -FRAME_SIZE, %sp + cmp length, 0 + be .Lend + ! wtxt + add %fp, BIAS-32, wtxt + + add %fp, BIAS-40, tmp + ld [ctx + AES_NROUNDS], nrounds + ! Compute xor, so that we can swap efficiently. + xor wtxt, tmp, diff + ! The loop variable will be multiplied by 16. + ! More loop invariants + add T, AES_TABLE0, T0 + + add T, AES_TABLE1, T1 + add T, AES_TABLE2, T2 + add T, AES_TABLE3, T3 + nop + +.Lblock_loop: + C Read src, and add initial subkey + AES_LOAD(0) ! i = 0 + AES_LOAD(4) ! i = 1 + AES_LOAD(8) ! i = 2 + AES_LOAD(12) ! i = 3 + add src, 16, src + + sub nrounds, 1, round + add ctx, 16, key + nop +.Lround_loop: + + AES_ROUND(0) ! i = 0 + AES_ROUND(4) ! i = 1 + AES_ROUND(8) ! i = 2 + AES_ROUND(12) ! i = 3 + + ! switch roles for tmp and wtxt + xor wtxt, diff, wtxt + xor tmp, diff, tmp + subcc round, 1, round + bne .Lround_loop + + add key, 16, key + + C Final round, and storage of the output + + AES_FINAL_ROUND(0) ! i = 0 + AES_FINAL_ROUND(4) ! i = 1 + AES_FINAL_ROUND(8) ! i = 2 + AES_FINAL_ROUND(12) ! i = 3 + + addcc length, -16, length + bne .Lblock_loop + add dst, 16, dst + +.Lend: + ret + restore +.Leord: + .size _nettle_aes_crypt,.Leord-_nettle_aes_crypt + + ! Benchmarks on my slow sparcstation: + ! Original C code + ! aes128 (ECB encrypt): 14.36s, 0.696MB/s + ! aes128 (ECB decrypt): 17.19s, 0.582MB/s + ! aes128 (CBC encrypt): 16.08s, 0.622MB/s + ! aes128 ((CBC decrypt)): 18.79s, 0.532MB/s + ! + ! aes192 (ECB encrypt): 16.85s, 0.593MB/s + ! aes192 (ECB decrypt): 19.64s, 0.509MB/s + ! aes192 (CBC encrypt): 18.43s, 0.543MB/s + ! aes192 (CBC decrypt): 20.76s, 0.482MB/s + ! + ! aes256 (ECB encrypt): 19.12s, 0.523MB/s + ! aes256 (ECB decrypt): 22.57s, 0.443MB/s + ! aes256 (CBC encrypt): 20.92s, 0.478MB/s + ! aes256 (CBC decrypt): 23.22s, 0.431MB/s + + ! After unrolling key_addition32, and getting rid of + ! some sll x, 2, x, encryption speed is 0.760 MB/s. + + ! Next, the C code was optimized to use larger tables and + ! no rotates. New timings: + ! aes128 (ECB encrypt): 13.10s, 0.763MB/s + ! aes128 (ECB decrypt): 11.51s, 0.869MB/s + ! aes128 (CBC encrypt): 15.15s, 0.660MB/s + ! aes128 (CBC decrypt): 13.10s, 0.763MB/s + ! + ! aes192 (ECB encrypt): 15.68s, 0.638MB/s + ! aes192 (ECB decrypt): 13.59s, 0.736MB/s + ! aes192 (CBC encrypt): 17.65s, 0.567MB/s + ! aes192 (CBC decrypt): 15.31s, 0.653MB/s + ! + ! aes256 (ECB encrypt): 17.95s, 0.557MB/s + ! aes256 (ECB decrypt): 15.90s, 0.629MB/s + ! aes256 (CBC encrypt): 20.16s, 0.496MB/s + ! aes256 (CBC decrypt): 17.47s, 0.572MB/s + + ! After optimization using pre-shifted indices + ! (AES_SIDX[1-3]): + ! aes128 (ECB encrypt): 12.46s, 0.803MB/s + ! aes128 (ECB decrypt): 10.74s, 0.931MB/s + ! aes128 (CBC encrypt): 17.74s, 0.564MB/s + ! aes128 (CBC decrypt): 12.43s, 0.805MB/s + ! + ! aes192 (ECB encrypt): 14.59s, 0.685MB/s + ! aes192 (ECB decrypt): 12.76s, 0.784MB/s + ! aes192 (CBC encrypt): 19.97s, 0.501MB/s + ! aes192 (CBC decrypt): 14.46s, 0.692MB/s + ! + ! aes256 (ECB encrypt): 17.00s, 0.588MB/s + ! aes256 (ECB decrypt): 14.81s, 0.675MB/s + ! aes256 (CBC encrypt): 22.65s, 0.442MB/s + ! aes256 (CBC decrypt): 16.46s, 0.608MB/s + + ! After implementing double buffering + ! aes128 (ECB encrypt): 12.59s, 0.794MB/s + ! aes128 (ECB decrypt): 10.56s, 0.947MB/s + ! aes128 (CBC encrypt): 17.91s, 0.558MB/s + ! aes128 (CBC decrypt): 12.30s, 0.813MB/s + ! + ! aes192 (ECB encrypt): 15.03s, 0.665MB/s + ! aes192 (ECB decrypt): 12.56s, 0.796MB/s + ! aes192 (CBC encrypt): 20.30s, 0.493MB/s + ! aes192 (CBC decrypt): 14.26s, 0.701MB/s + ! + ! aes256 (ECB encrypt): 17.30s, 0.578MB/s + ! aes256 (ECB decrypt): 14.51s, 0.689MB/s + ! aes256 (CBC encrypt): 22.75s, 0.440MB/s + ! aes256 (CBC decrypt): 16.35s, 0.612MB/s + + ! After reordering aes-encrypt.c and aes-decypt.c + ! (the order probably causes strange cache-effects): + ! aes128 (ECB encrypt): 9.21s, 1.086MB/s + ! aes128 (ECB decrypt): 11.13s, 0.898MB/s + ! aes128 (CBC encrypt): 14.12s, 0.708MB/s + ! aes128 (CBC decrypt): 13.77s, 0.726MB/s + ! + ! aes192 (ECB encrypt): 10.86s, 0.921MB/s + ! aes192 (ECB decrypt): 13.17s, 0.759MB/s + ! aes192 (CBC encrypt): 15.74s, 0.635MB/s + ! aes192 (CBC decrypt): 15.91s, 0.629MB/s + ! + ! aes256 (ECB encrypt): 12.71s, 0.787MB/s + ! aes256 (ECB decrypt): 15.38s, 0.650MB/s + ! aes256 (CBC encrypt): 17.49s, 0.572MB/s + ! aes256 (CBC decrypt): 17.87s, 0.560MB/s + + ! After further optimizations of the initial and final loops, + ! source_loop and final_loop. + ! aes128 (ECB encrypt): 8.07s, 1.239MB/s + ! aes128 (ECB decrypt): 9.48s, 1.055MB/s + ! aes128 (CBC encrypt): 12.76s, 0.784MB/s + ! aes128 (CBC decrypt): 12.15s, 0.823MB/s + ! + ! aes192 (ECB encrypt): 9.43s, 1.060MB/s + ! aes192 (ECB decrypt): 11.20s, 0.893MB/s + ! aes192 (CBC encrypt): 14.19s, 0.705MB/s + ! aes192 (CBC decrypt): 13.97s, 0.716MB/s + ! + ! aes256 (ECB encrypt): 10.81s, 0.925MB/s + ! aes256 (ECB decrypt): 12.92s, 0.774MB/s + ! aes256 (CBC encrypt): 15.59s, 0.641MB/s + ! aes256 (CBC decrypt): 15.76s, 0.635MB/s + + ! After unrolling loops, and other optimizations suggested by + ! Marcus: + ! aes128 (ECB encrypt): 6.40s, 1.562MB/s + ! aes128 (ECB decrypt): 8.17s, 1.224MB/s + ! aes128 (CBC encrypt): 13.11s, 0.763MB/s + ! aes128 (CBC decrypt): 10.05s, 0.995MB/s + ! + ! aes192 (ECB encrypt): 7.43s, 1.346MB/s + ! aes192 (ECB decrypt): 9.51s, 1.052MB/s + ! aes192 (CBC encrypt): 14.09s, 0.710MB/s + ! aes192 (CBC decrypt): 11.58s, 0.864MB/s + ! + ! aes256 (ECB encrypt): 8.57s, 1.167MB/s + ! aes256 (ECB decrypt): 11.13s, 0.898MB/s + ! aes256 (CBC encrypt): 15.30s, 0.654MB/s + ! aes256 (CBC decrypt): 12.93s, 0.773MB/s diff --git a/sparc64/machine.m4 b/sparc64/machine.m4 new file mode 100644 index 0000000000000000000000000000000000000000..e69de29bb2d1d6434b8b29ae775ad8c2e48c5391