Alexandre Savard | 1b09e31 | 2012-08-07 20:33:29 -0400 | [diff] [blame] | 1 | // ==================================================================== |
| 2 | // Written by Andy Polyakov <appro@fy.chalmers.se> for the OpenSSL |
| 3 | // project. Rights for redistribution and usage in source and binary |
| 4 | // forms are granted according to the OpenSSL license. |
| 5 | // ==================================================================== |
| 6 | // |
| 7 | // What's wrong with compiler generated code? Compiler never uses |
| 8 | // variable 'shr' which is pairable with 'extr'/'dep' instructions. |
| 9 | // Then it uses 'zxt' which is an I-type, but can be replaced with |
| 10 | // 'and' which in turn can be assigned to M-port [there're double as |
| 11 | // much M-ports as there're I-ports on Itanium 2]. By sacrificing few |
| 12 | // registers for small constants (255, 24 and 16) to be used with |
| 13 | // 'shr' and 'and' instructions I can achieve better ILP, Intruction |
| 14 | // Level Parallelism, and performance. This code outperforms GCC 3.3 |
| 15 | // generated code by over factor of 2 (two), GCC 3.4 - by 70% and |
| 16 | // HP C - by 40%. Measured best-case scenario, i.e. aligned |
| 17 | // big-endian input, ECB timing on Itanium 2 is (18 + 13*rounds) |
| 18 | // ticks per block, or 9.25 CPU cycles per byte for 128 bit key. |
| 19 | |
| 20 | // Version 1.2 mitigates the hazard of cache-timing attacks by |
| 21 | // a) compressing S-boxes from 8KB to 2KB+256B, b) scheduling |
| 22 | // references to S-boxes for L2 cache latency, c) prefetching T[ed]4 |
| 23 | // prior last round. As result performance dropped to (26 + 15*rounds) |
| 24 | // ticks per block or 11 cycles per byte processed with 128-bit key. |
| 25 | // This is ~16% deterioration. For reference Itanium 2 L1 cache has |
| 26 | // 64 bytes line size and L2 - 128 bytes... |
| 27 | |
| 28 | .ident "aes-ia64.S, version 1.2" |
| 29 | .ident "IA-64 ISA artwork by Andy Polyakov <appro@fy.chalmers.se>" |
| 30 | .explicit |
| 31 | .text |
| 32 | |
| 33 | rk0=r8; rk1=r9; |
| 34 | |
| 35 | pfssave=r2; |
| 36 | lcsave=r10; |
| 37 | prsave=r3; |
| 38 | maskff=r11; |
| 39 | twenty4=r14; |
| 40 | sixteen=r15; |
| 41 | |
| 42 | te00=r16; te11=r17; te22=r18; te33=r19; |
| 43 | te01=r20; te12=r21; te23=r22; te30=r23; |
| 44 | te02=r24; te13=r25; te20=r26; te31=r27; |
| 45 | te03=r28; te10=r29; te21=r30; te32=r31; |
| 46 | |
| 47 | // these are rotating... |
| 48 | t0=r32; s0=r33; |
| 49 | t1=r34; s1=r35; |
| 50 | t2=r36; s2=r37; |
| 51 | t3=r38; s3=r39; |
| 52 | |
| 53 | te0=r40; te1=r41; te2=r42; te3=r43; |
| 54 | |
| 55 | #if defined(_HPUX_SOURCE) && !defined(_LP64) |
| 56 | # define ADDP addp4 |
| 57 | #else |
| 58 | # define ADDP add |
| 59 | #endif |
| 60 | |
| 61 | // Offsets from Te0 |
| 62 | #define TE0 0 |
| 63 | #define TE2 2 |
| 64 | #if defined(_HPUX_SOURCE) || defined(B_ENDIAN) |
| 65 | #define TE1 3 |
| 66 | #define TE3 1 |
| 67 | #else |
| 68 | #define TE1 1 |
| 69 | #define TE3 3 |
| 70 | #endif |
| 71 | |
| 72 | // This implies that AES_KEY comprises 32-bit key schedule elements |
| 73 | // even on LP64 platforms. |
| 74 | #ifndef KSZ |
| 75 | # define KSZ 4 |
| 76 | # define LDKEY ld4 |
| 77 | #endif |
| 78 | |
| 79 | .proc _ia64_AES_encrypt# |
| 80 | // Input: rk0-rk1 |
| 81 | // te0 |
| 82 | // te3 as AES_KEY->rounds!!! |
| 83 | // s0-s3 |
| 84 | // maskff,twenty4,sixteen |
| 85 | // Output: r16,r20,r24,r28 as s0-s3 |
| 86 | // Clobber: r16-r31,rk0-rk1,r32-r43 |
| 87 | .align 32 |
| 88 | _ia64_AES_encrypt: |
| 89 | .prologue |
| 90 | .altrp b6 |
| 91 | .body |
| 92 | { .mmi; alloc r16=ar.pfs,12,0,0,8 |
| 93 | LDKEY t0=[rk0],2*KSZ |
| 94 | mov pr.rot=1<<16 } |
| 95 | { .mmi; LDKEY t1=[rk1],2*KSZ |
| 96 | add te1=TE1,te0 |
| 97 | add te3=-3,te3 };; |
| 98 | { .mib; LDKEY t2=[rk0],2*KSZ |
| 99 | mov ar.ec=2 } |
| 100 | { .mib; LDKEY t3=[rk1],2*KSZ |
| 101 | add te2=TE2,te0 |
| 102 | brp.loop.imp .Le_top,.Le_end-16 };; |
| 103 | |
| 104 | { .mmi; xor s0=s0,t0 |
| 105 | xor s1=s1,t1 |
| 106 | mov ar.lc=te3 } |
| 107 | { .mmi; xor s2=s2,t2 |
| 108 | xor s3=s3,t3 |
| 109 | add te3=TE3,te0 };; |
| 110 | |
| 111 | .align 32 |
| 112 | .Le_top: |
| 113 | { .mmi; (p0) LDKEY t0=[rk0],2*KSZ // 0/0:rk[0] |
| 114 | (p0) and te33=s3,maskff // 0/0:s3&0xff |
| 115 | (p0) extr.u te22=s2,8,8 } // 0/0:s2>>8&0xff |
| 116 | { .mmi; (p0) LDKEY t1=[rk1],2*KSZ // 0/1:rk[1] |
| 117 | (p0) and te30=s0,maskff // 0/1:s0&0xff |
| 118 | (p0) shr.u te00=s0,twenty4 };; // 0/0:s0>>24 |
| 119 | { .mmi; (p0) LDKEY t2=[rk0],2*KSZ // 1/2:rk[2] |
| 120 | (p0) shladd te33=te33,3,te3 // 1/0:te0+s0>>24 |
| 121 | (p0) extr.u te23=s3,8,8 } // 1/1:s3>>8&0xff |
| 122 | { .mmi; (p0) LDKEY t3=[rk1],2*KSZ // 1/3:rk[3] |
| 123 | (p0) shladd te30=te30,3,te3 // 1/1:te3+s0 |
| 124 | (p0) shr.u te01=s1,twenty4 };; // 1/1:s1>>24 |
| 125 | { .mmi; (p0) ld4 te33=[te33] // 2/0:te3[s3&0xff] |
| 126 | (p0) shladd te22=te22,3,te2 // 2/0:te2+s2>>8&0xff |
| 127 | (p0) extr.u te20=s0,8,8 } // 2/2:s0>>8&0xff |
| 128 | { .mmi; (p0) ld4 te30=[te30] // 2/1:te3[s0] |
| 129 | (p0) shladd te23=te23,3,te2 // 2/1:te2+s3>>8 |
| 130 | (p0) shr.u te02=s2,twenty4 };; // 2/2:s2>>24 |
| 131 | { .mmi; (p0) ld4 te22=[te22] // 3/0:te2[s2>>8] |
| 132 | (p0) shladd te20=te20,3,te2 // 3/2:te2+s0>>8 |
| 133 | (p0) extr.u te21=s1,8,8 } // 3/3:s1>>8&0xff |
| 134 | { .mmi; (p0) ld4 te23=[te23] // 3/1:te2[s3>>8] |
| 135 | (p0) shladd te00=te00,3,te0 // 3/0:te0+s0>>24 |
| 136 | (p0) shr.u te03=s3,twenty4 };; // 3/3:s3>>24 |
| 137 | { .mmi; (p0) ld4 te20=[te20] // 4/2:te2[s0>>8] |
| 138 | (p0) shladd te21=te21,3,te2 // 4/3:te3+s2 |
| 139 | (p0) extr.u te11=s1,16,8 } // 4/0:s1>>16&0xff |
| 140 | { .mmi; (p0) ld4 te00=[te00] // 4/0:te0[s0>>24] |
| 141 | (p0) shladd te01=te01,3,te0 // 4/1:te0+s1>>24 |
| 142 | (p0) shr.u te13=s3,sixteen };; // 4/2:s3>>16 |
| 143 | { .mmi; (p0) ld4 te21=[te21] // 5/3:te2[s1>>8] |
| 144 | (p0) shladd te11=te11,3,te1 // 5/0:te1+s1>>16 |
| 145 | (p0) extr.u te12=s2,16,8 } // 5/1:s2>>16&0xff |
| 146 | { .mmi; (p0) ld4 te01=[te01] // 5/1:te0[s1>>24] |
| 147 | (p0) shladd te02=te02,3,te0 // 5/2:te0+s2>>24 |
| 148 | (p0) and te31=s1,maskff };; // 5/2:s1&0xff |
| 149 | { .mmi; (p0) ld4 te11=[te11] // 6/0:te1[s1>>16] |
| 150 | (p0) shladd te12=te12,3,te1 // 6/1:te1+s2>>16 |
| 151 | (p0) extr.u te10=s0,16,8 } // 6/3:s0>>16&0xff |
| 152 | { .mmi; (p0) ld4 te02=[te02] // 6/2:te0[s2>>24] |
| 153 | (p0) shladd te03=te03,3,te0 // 6/3:te1+s0>>16 |
| 154 | (p0) and te32=s2,maskff };; // 6/3:s2&0xff |
| 155 | |
| 156 | { .mmi; (p0) ld4 te12=[te12] // 7/1:te1[s2>>16] |
| 157 | (p0) shladd te31=te31,3,te3 // 7/2:te3+s1&0xff |
| 158 | (p0) and te13=te13,maskff} // 7/2:s3>>16&0xff |
| 159 | { .mmi; (p0) ld4 te03=[te03] // 7/3:te0[s3>>24] |
| 160 | (p0) shladd te32=te32,3,te3 // 7/3:te3+s2 |
| 161 | (p0) xor t0=t0,te33 };; // 7/0: |
| 162 | { .mmi; (p0) ld4 te31=[te31] // 8/2:te3[s1] |
| 163 | (p0) shladd te13=te13,3,te1 // 8/2:te1+s3>>16 |
| 164 | (p0) xor t0=t0,te22 } // 8/0: |
| 165 | { .mmi; (p0) ld4 te32=[te32] // 8/3:te3[s2] |
| 166 | (p0) shladd te10=te10,3,te1 // 8/3:te1+s0>>16 |
| 167 | (p0) xor t1=t1,te30 };; // 8/1: |
| 168 | { .mmi; (p0) ld4 te13=[te13] // 9/2:te1[s3>>16] |
| 169 | (p0) ld4 te10=[te10] // 9/3:te1[s0>>16] |
| 170 | (p0) xor t0=t0,te00 };; // 9/0: !L2 scheduling |
| 171 | { .mmi; (p0) xor t1=t1,te23 // 10[9]/1: |
| 172 | (p0) xor t2=t2,te20 // 10[9]/2: |
| 173 | (p0) xor t3=t3,te21 };; // 10[9]/3: |
| 174 | { .mmi; (p0) xor t0=t0,te11 // 11[10]/0:done! |
| 175 | (p0) xor t1=t1,te01 // 11[10]/1: |
| 176 | (p0) xor t2=t2,te02 };; // 11[10]/2: !L2 scheduling |
| 177 | { .mmi; (p0) xor t3=t3,te03 // 12[10]/3: |
| 178 | (p16) cmp.eq p0,p17=r0,r0 };; // 12[10]/clear (p17) |
| 179 | { .mmi; (p0) xor t1=t1,te12 // 13[11]/1:done! |
| 180 | (p0) xor t2=t2,te31 // 13[11]/2: |
| 181 | (p0) xor t3=t3,te32 } // 13[11]/3: |
| 182 | { .mmi; (p17) add te0=2048,te0 // 13[11]/ |
| 183 | (p17) add te1=2048+64-TE1,te1};; // 13[11]/ |
| 184 | { .mib; (p0) xor t2=t2,te13 // 14[12]/2:done! |
| 185 | (p17) add te2=2048+128-TE2,te2} // 14[12]/ |
| 186 | { .mib; (p0) xor t3=t3,te10 // 14[12]/3:done! |
| 187 | (p17) add te3=2048+192-TE3,te3 // 14[12]/ |
| 188 | br.ctop.sptk .Le_top };; |
| 189 | .Le_end: |
| 190 | |
| 191 | |
| 192 | { .mmi; ld8 te12=[te0] // prefetch Te4 |
| 193 | ld8 te31=[te1] } |
| 194 | { .mmi; ld8 te10=[te2] |
| 195 | ld8 te32=[te3] } |
| 196 | |
| 197 | { .mmi; LDKEY t0=[rk0],2*KSZ // 0/0:rk[0] |
| 198 | and te33=s3,maskff // 0/0:s3&0xff |
| 199 | extr.u te22=s2,8,8 } // 0/0:s2>>8&0xff |
| 200 | { .mmi; LDKEY t1=[rk1],2*KSZ // 0/1:rk[1] |
| 201 | and te30=s0,maskff // 0/1:s0&0xff |
| 202 | shr.u te00=s0,twenty4 };; // 0/0:s0>>24 |
| 203 | { .mmi; LDKEY t2=[rk0],2*KSZ // 1/2:rk[2] |
| 204 | add te33=te33,te0 // 1/0:te0+s0>>24 |
| 205 | extr.u te23=s3,8,8 } // 1/1:s3>>8&0xff |
| 206 | { .mmi; LDKEY t3=[rk1],2*KSZ // 1/3:rk[3] |
| 207 | add te30=te30,te0 // 1/1:te0+s0 |
| 208 | shr.u te01=s1,twenty4 };; // 1/1:s1>>24 |
| 209 | { .mmi; ld1 te33=[te33] // 2/0:te0[s3&0xff] |
| 210 | add te22=te22,te0 // 2/0:te0+s2>>8&0xff |
| 211 | extr.u te20=s0,8,8 } // 2/2:s0>>8&0xff |
| 212 | { .mmi; ld1 te30=[te30] // 2/1:te0[s0] |
| 213 | add te23=te23,te0 // 2/1:te0+s3>>8 |
| 214 | shr.u te02=s2,twenty4 };; // 2/2:s2>>24 |
| 215 | { .mmi; ld1 te22=[te22] // 3/0:te0[s2>>8] |
| 216 | add te20=te20,te0 // 3/2:te0+s0>>8 |
| 217 | extr.u te21=s1,8,8 } // 3/3:s1>>8&0xff |
| 218 | { .mmi; ld1 te23=[te23] // 3/1:te0[s3>>8] |
| 219 | add te00=te00,te0 // 3/0:te0+s0>>24 |
| 220 | shr.u te03=s3,twenty4 };; // 3/3:s3>>24 |
| 221 | { .mmi; ld1 te20=[te20] // 4/2:te0[s0>>8] |
| 222 | add te21=te21,te0 // 4/3:te0+s2 |
| 223 | extr.u te11=s1,16,8 } // 4/0:s1>>16&0xff |
| 224 | { .mmi; ld1 te00=[te00] // 4/0:te0[s0>>24] |
| 225 | add te01=te01,te0 // 4/1:te0+s1>>24 |
| 226 | shr.u te13=s3,sixteen };; // 4/2:s3>>16 |
| 227 | { .mmi; ld1 te21=[te21] // 5/3:te0[s1>>8] |
| 228 | add te11=te11,te0 // 5/0:te0+s1>>16 |
| 229 | extr.u te12=s2,16,8 } // 5/1:s2>>16&0xff |
| 230 | { .mmi; ld1 te01=[te01] // 5/1:te0[s1>>24] |
| 231 | add te02=te02,te0 // 5/2:te0+s2>>24 |
| 232 | and te31=s1,maskff };; // 5/2:s1&0xff |
| 233 | { .mmi; ld1 te11=[te11] // 6/0:te0[s1>>16] |
| 234 | add te12=te12,te0 // 6/1:te0+s2>>16 |
| 235 | extr.u te10=s0,16,8 } // 6/3:s0>>16&0xff |
| 236 | { .mmi; ld1 te02=[te02] // 6/2:te0[s2>>24] |
| 237 | add te03=te03,te0 // 6/3:te0+s0>>16 |
| 238 | and te32=s2,maskff };; // 6/3:s2&0xff |
| 239 | |
| 240 | { .mmi; ld1 te12=[te12] // 7/1:te0[s2>>16] |
| 241 | add te31=te31,te0 // 7/2:te0+s1&0xff |
| 242 | dep te33=te22,te33,8,8} // 7/0: |
| 243 | { .mmi; ld1 te03=[te03] // 7/3:te0[s3>>24] |
| 244 | add te32=te32,te0 // 7/3:te0+s2 |
| 245 | and te13=te13,maskff};; // 7/2:s3>>16&0xff |
| 246 | { .mmi; ld1 te31=[te31] // 8/2:te0[s1] |
| 247 | add te13=te13,te0 // 8/2:te0+s3>>16 |
| 248 | dep te30=te23,te30,8,8} // 8/1: |
| 249 | { .mmi; ld1 te32=[te32] // 8/3:te0[s2] |
| 250 | add te10=te10,te0 // 8/3:te0+s0>>16 |
| 251 | shl te00=te00,twenty4};; // 8/0: |
| 252 | { .mii; ld1 te13=[te13] // 9/2:te0[s3>>16] |
| 253 | dep te33=te11,te33,16,8 // 9/0: |
| 254 | shl te01=te01,twenty4};; // 9/1: |
| 255 | { .mii; ld1 te10=[te10] // 10/3:te0[s0>>16] |
| 256 | dep te31=te20,te31,8,8 // 10/2: |
| 257 | shl te02=te02,twenty4};; // 10/2: |
| 258 | { .mii; xor t0=t0,te33 // 11/0: |
| 259 | dep te32=te21,te32,8,8 // 11/3: |
| 260 | shl te12=te12,sixteen};; // 11/1: |
| 261 | { .mii; xor r16=t0,te00 // 12/0:done! |
| 262 | dep te31=te13,te31,16,8 // 12/2: |
| 263 | shl te03=te03,twenty4};; // 12/3: |
| 264 | { .mmi; xor t1=t1,te01 // 13/1: |
| 265 | xor t2=t2,te02 // 13/2: |
| 266 | dep te32=te10,te32,16,8};; // 13/3: |
| 267 | { .mmi; xor t1=t1,te30 // 14/1: |
| 268 | xor r24=t2,te31 // 14/2:done! |
| 269 | xor t3=t3,te32 };; // 14/3: |
| 270 | { .mib; xor r20=t1,te12 // 15/1:done! |
| 271 | xor r28=t3,te03 // 15/3:done! |
| 272 | br.ret.sptk b6 };; |
| 273 | .endp _ia64_AES_encrypt# |
| 274 | |
| 275 | // void AES_encrypt (const void *in,void *out,const AES_KEY *key); |
| 276 | .global AES_encrypt# |
| 277 | .proc AES_encrypt# |
| 278 | .align 32 |
| 279 | AES_encrypt: |
| 280 | .prologue |
| 281 | .save ar.pfs,pfssave |
| 282 | { .mmi; alloc pfssave=ar.pfs,3,1,12,0 |
| 283 | and out0=3,in0 |
| 284 | mov r3=ip } |
| 285 | { .mmi; ADDP in0=0,in0 |
| 286 | mov loc0=psr.um |
| 287 | ADDP out11=KSZ*60,in2 };; // &AES_KEY->rounds |
| 288 | |
| 289 | { .mmi; ld4 out11=[out11] // AES_KEY->rounds |
| 290 | add out8=(AES_Te#-AES_encrypt#),r3 // Te0 |
| 291 | .save pr,prsave |
| 292 | mov prsave=pr } |
| 293 | { .mmi; rum 1<<3 // clear um.ac |
| 294 | .save ar.lc,lcsave |
| 295 | mov lcsave=ar.lc };; |
| 296 | |
| 297 | .body |
| 298 | #if defined(_HPUX_SOURCE) // HPUX is big-endian, cut 15+15 cycles... |
| 299 | { .mib; cmp.ne p6,p0=out0,r0 |
| 300 | add out0=4,in0 |
| 301 | (p6) br.dpnt.many .Le_i_unaligned };; |
| 302 | |
| 303 | { .mmi; ld4 out1=[in0],8 // s0 |
| 304 | and out9=3,in1 |
| 305 | mov twenty4=24 } |
| 306 | { .mmi; ld4 out3=[out0],8 // s1 |
| 307 | ADDP rk0=0,in2 |
| 308 | mov sixteen=16 };; |
| 309 | { .mmi; ld4 out5=[in0] // s2 |
| 310 | cmp.ne p6,p0=out9,r0 |
| 311 | mov maskff=0xff } |
| 312 | { .mmb; ld4 out7=[out0] // s3 |
| 313 | ADDP rk1=KSZ,in2 |
| 314 | br.call.sptk.many b6=_ia64_AES_encrypt };; |
| 315 | |
| 316 | { .mib; ADDP in0=4,in1 |
| 317 | ADDP in1=0,in1 |
| 318 | (p6) br.spnt .Le_o_unaligned };; |
| 319 | |
| 320 | { .mii; mov psr.um=loc0 |
| 321 | mov ar.pfs=pfssave |
| 322 | mov ar.lc=lcsave };; |
| 323 | { .mmi; st4 [in1]=r16,8 // s0 |
| 324 | st4 [in0]=r20,8 // s1 |
| 325 | mov pr=prsave,0x1ffff };; |
| 326 | { .mmb; st4 [in1]=r24 // s2 |
| 327 | st4 [in0]=r28 // s3 |
| 328 | br.ret.sptk.many b0 };; |
| 329 | #endif |
| 330 | |
| 331 | .align 32 |
| 332 | .Le_i_unaligned: |
| 333 | { .mmi; add out0=1,in0 |
| 334 | add out2=2,in0 |
| 335 | add out4=3,in0 };; |
| 336 | { .mmi; ld1 r16=[in0],4 |
| 337 | ld1 r17=[out0],4 }//;; |
| 338 | { .mmi; ld1 r18=[out2],4 |
| 339 | ld1 out1=[out4],4 };; // s0 |
| 340 | { .mmi; ld1 r20=[in0],4 |
| 341 | ld1 r21=[out0],4 }//;; |
| 342 | { .mmi; ld1 r22=[out2],4 |
| 343 | ld1 out3=[out4],4 };; // s1 |
| 344 | { .mmi; ld1 r24=[in0],4 |
| 345 | ld1 r25=[out0],4 }//;; |
| 346 | { .mmi; ld1 r26=[out2],4 |
| 347 | ld1 out5=[out4],4 };; // s2 |
| 348 | { .mmi; ld1 r28=[in0] |
| 349 | ld1 r29=[out0] }//;; |
| 350 | { .mmi; ld1 r30=[out2] |
| 351 | ld1 out7=[out4] };; // s3 |
| 352 | |
| 353 | { .mii; |
| 354 | dep out1=r16,out1,24,8 //;; |
| 355 | dep out3=r20,out3,24,8 }//;; |
| 356 | { .mii; ADDP rk0=0,in2 |
| 357 | dep out5=r24,out5,24,8 //;; |
| 358 | dep out7=r28,out7,24,8 };; |
| 359 | { .mii; ADDP rk1=KSZ,in2 |
| 360 | dep out1=r17,out1,16,8 //;; |
| 361 | dep out3=r21,out3,16,8 }//;; |
| 362 | { .mii; mov twenty4=24 |
| 363 | dep out5=r25,out5,16,8 //;; |
| 364 | dep out7=r29,out7,16,8 };; |
| 365 | { .mii; mov sixteen=16 |
| 366 | dep out1=r18,out1,8,8 //;; |
| 367 | dep out3=r22,out3,8,8 }//;; |
| 368 | { .mii; mov maskff=0xff |
| 369 | dep out5=r26,out5,8,8 //;; |
| 370 | dep out7=r30,out7,8,8 };; |
| 371 | |
| 372 | { .mib; br.call.sptk.many b6=_ia64_AES_encrypt };; |
| 373 | |
| 374 | .Le_o_unaligned: |
| 375 | { .mii; ADDP out0=0,in1 |
| 376 | extr.u r17=r16,8,8 // s0 |
| 377 | shr.u r19=r16,twenty4 }//;; |
| 378 | { .mii; ADDP out1=1,in1 |
| 379 | extr.u r18=r16,16,8 |
| 380 | shr.u r23=r20,twenty4 }//;; // s1 |
| 381 | { .mii; ADDP out2=2,in1 |
| 382 | extr.u r21=r20,8,8 |
| 383 | shr.u r22=r20,sixteen }//;; |
| 384 | { .mii; ADDP out3=3,in1 |
| 385 | extr.u r25=r24,8,8 // s2 |
| 386 | shr.u r27=r24,twenty4 };; |
| 387 | { .mii; st1 [out3]=r16,4 |
| 388 | extr.u r26=r24,16,8 |
| 389 | shr.u r31=r28,twenty4 }//;; // s3 |
| 390 | { .mii; st1 [out2]=r17,4 |
| 391 | extr.u r29=r28,8,8 |
| 392 | shr.u r30=r28,sixteen }//;; |
| 393 | |
| 394 | { .mmi; st1 [out1]=r18,4 |
| 395 | st1 [out0]=r19,4 };; |
| 396 | { .mmi; st1 [out3]=r20,4 |
| 397 | st1 [out2]=r21,4 }//;; |
| 398 | { .mmi; st1 [out1]=r22,4 |
| 399 | st1 [out0]=r23,4 };; |
| 400 | { .mmi; st1 [out3]=r24,4 |
| 401 | st1 [out2]=r25,4 |
| 402 | mov pr=prsave,0x1ffff }//;; |
| 403 | { .mmi; st1 [out1]=r26,4 |
| 404 | st1 [out0]=r27,4 |
| 405 | mov ar.pfs=pfssave };; |
| 406 | { .mmi; st1 [out3]=r28 |
| 407 | st1 [out2]=r29 |
| 408 | mov ar.lc=lcsave }//;; |
| 409 | { .mmi; st1 [out1]=r30 |
| 410 | st1 [out0]=r31 } |
| 411 | { .mfb; mov psr.um=loc0 // restore user mask |
| 412 | br.ret.sptk.many b0 };; |
| 413 | .endp AES_encrypt# |
| 414 | |
| 415 | // *AES_decrypt are autogenerated by the following script: |
| 416 | #if 0 |
| 417 | #!/usr/bin/env perl |
| 418 | print "// *AES_decrypt are autogenerated by the following script:\n#if 0\n"; |
| 419 | open(PROG,'<'.$0); while(<PROG>) { print; } close(PROG); |
| 420 | print "#endif\n"; |
| 421 | while(<>) { |
| 422 | $process=1 if (/\.proc\s+_ia64_AES_encrypt/); |
| 423 | next if (!$process); |
| 424 | |
| 425 | #s/te00=s0/td00=s0/; s/te00/td00/g; |
| 426 | s/te11=s1/td13=s3/; s/te11/td13/g; |
| 427 | #s/te22=s2/td22=s2/; s/te22/td22/g; |
| 428 | s/te33=s3/td31=s1/; s/te33/td31/g; |
| 429 | |
| 430 | #s/te01=s1/td01=s1/; s/te01/td01/g; |
| 431 | s/te12=s2/td10=s0/; s/te12/td10/g; |
| 432 | #s/te23=s3/td23=s3/; s/te23/td23/g; |
| 433 | s/te30=s0/td32=s2/; s/te30/td32/g; |
| 434 | |
| 435 | #s/te02=s2/td02=s2/; s/te02/td02/g; |
| 436 | s/te13=s3/td11=s1/; s/te13/td11/g; |
| 437 | #s/te20=s0/td20=s0/; s/te20/td20/g; |
| 438 | s/te31=s1/td33=s3/; s/te31/td33/g; |
| 439 | |
| 440 | #s/te03=s3/td03=s3/; s/te03/td03/g; |
| 441 | s/te10=s0/td12=s2/; s/te10/td12/g; |
| 442 | #s/te21=s1/td21=s1/; s/te21/td21/g; |
| 443 | s/te32=s2/td30=s0/; s/te32/td30/g; |
| 444 | |
| 445 | s/td/te/g; |
| 446 | |
| 447 | s/AES_encrypt/AES_decrypt/g; |
| 448 | s/\.Le_/.Ld_/g; |
| 449 | s/AES_Te#/AES_Td#/g; |
| 450 | |
| 451 | print; |
| 452 | |
| 453 | exit if (/\.endp\s+AES_decrypt/); |
| 454 | } |
| 455 | #endif |
| 456 | .proc _ia64_AES_decrypt# |
| 457 | // Input: rk0-rk1 |
| 458 | // te0 |
| 459 | // te3 as AES_KEY->rounds!!! |
| 460 | // s0-s3 |
| 461 | // maskff,twenty4,sixteen |
| 462 | // Output: r16,r20,r24,r28 as s0-s3 |
| 463 | // Clobber: r16-r31,rk0-rk1,r32-r43 |
| 464 | .align 32 |
| 465 | _ia64_AES_decrypt: |
| 466 | .prologue |
| 467 | .altrp b6 |
| 468 | .body |
| 469 | { .mmi; alloc r16=ar.pfs,12,0,0,8 |
| 470 | LDKEY t0=[rk0],2*KSZ |
| 471 | mov pr.rot=1<<16 } |
| 472 | { .mmi; LDKEY t1=[rk1],2*KSZ |
| 473 | add te1=TE1,te0 |
| 474 | add te3=-3,te3 };; |
| 475 | { .mib; LDKEY t2=[rk0],2*KSZ |
| 476 | mov ar.ec=2 } |
| 477 | { .mib; LDKEY t3=[rk1],2*KSZ |
| 478 | add te2=TE2,te0 |
| 479 | brp.loop.imp .Ld_top,.Ld_end-16 };; |
| 480 | |
| 481 | { .mmi; xor s0=s0,t0 |
| 482 | xor s1=s1,t1 |
| 483 | mov ar.lc=te3 } |
| 484 | { .mmi; xor s2=s2,t2 |
| 485 | xor s3=s3,t3 |
| 486 | add te3=TE3,te0 };; |
| 487 | |
| 488 | .align 32 |
| 489 | .Ld_top: |
| 490 | { .mmi; (p0) LDKEY t0=[rk0],2*KSZ // 0/0:rk[0] |
| 491 | (p0) and te31=s1,maskff // 0/0:s3&0xff |
| 492 | (p0) extr.u te22=s2,8,8 } // 0/0:s2>>8&0xff |
| 493 | { .mmi; (p0) LDKEY t1=[rk1],2*KSZ // 0/1:rk[1] |
| 494 | (p0) and te32=s2,maskff // 0/1:s0&0xff |
| 495 | (p0) shr.u te00=s0,twenty4 };; // 0/0:s0>>24 |
| 496 | { .mmi; (p0) LDKEY t2=[rk0],2*KSZ // 1/2:rk[2] |
| 497 | (p0) shladd te31=te31,3,te3 // 1/0:te0+s0>>24 |
| 498 | (p0) extr.u te23=s3,8,8 } // 1/1:s3>>8&0xff |
| 499 | { .mmi; (p0) LDKEY t3=[rk1],2*KSZ // 1/3:rk[3] |
| 500 | (p0) shladd te32=te32,3,te3 // 1/1:te3+s0 |
| 501 | (p0) shr.u te01=s1,twenty4 };; // 1/1:s1>>24 |
| 502 | { .mmi; (p0) ld4 te31=[te31] // 2/0:te3[s3&0xff] |
| 503 | (p0) shladd te22=te22,3,te2 // 2/0:te2+s2>>8&0xff |
| 504 | (p0) extr.u te20=s0,8,8 } // 2/2:s0>>8&0xff |
| 505 | { .mmi; (p0) ld4 te32=[te32] // 2/1:te3[s0] |
| 506 | (p0) shladd te23=te23,3,te2 // 2/1:te2+s3>>8 |
| 507 | (p0) shr.u te02=s2,twenty4 };; // 2/2:s2>>24 |
| 508 | { .mmi; (p0) ld4 te22=[te22] // 3/0:te2[s2>>8] |
| 509 | (p0) shladd te20=te20,3,te2 // 3/2:te2+s0>>8 |
| 510 | (p0) extr.u te21=s1,8,8 } // 3/3:s1>>8&0xff |
| 511 | { .mmi; (p0) ld4 te23=[te23] // 3/1:te2[s3>>8] |
| 512 | (p0) shladd te00=te00,3,te0 // 3/0:te0+s0>>24 |
| 513 | (p0) shr.u te03=s3,twenty4 };; // 3/3:s3>>24 |
| 514 | { .mmi; (p0) ld4 te20=[te20] // 4/2:te2[s0>>8] |
| 515 | (p0) shladd te21=te21,3,te2 // 4/3:te3+s2 |
| 516 | (p0) extr.u te13=s3,16,8 } // 4/0:s1>>16&0xff |
| 517 | { .mmi; (p0) ld4 te00=[te00] // 4/0:te0[s0>>24] |
| 518 | (p0) shladd te01=te01,3,te0 // 4/1:te0+s1>>24 |
| 519 | (p0) shr.u te11=s1,sixteen };; // 4/2:s3>>16 |
| 520 | { .mmi; (p0) ld4 te21=[te21] // 5/3:te2[s1>>8] |
| 521 | (p0) shladd te13=te13,3,te1 // 5/0:te1+s1>>16 |
| 522 | (p0) extr.u te10=s0,16,8 } // 5/1:s2>>16&0xff |
| 523 | { .mmi; (p0) ld4 te01=[te01] // 5/1:te0[s1>>24] |
| 524 | (p0) shladd te02=te02,3,te0 // 5/2:te0+s2>>24 |
| 525 | (p0) and te33=s3,maskff };; // 5/2:s1&0xff |
| 526 | { .mmi; (p0) ld4 te13=[te13] // 6/0:te1[s1>>16] |
| 527 | (p0) shladd te10=te10,3,te1 // 6/1:te1+s2>>16 |
| 528 | (p0) extr.u te12=s2,16,8 } // 6/3:s0>>16&0xff |
| 529 | { .mmi; (p0) ld4 te02=[te02] // 6/2:te0[s2>>24] |
| 530 | (p0) shladd te03=te03,3,te0 // 6/3:te1+s0>>16 |
| 531 | (p0) and te30=s0,maskff };; // 6/3:s2&0xff |
| 532 | |
| 533 | { .mmi; (p0) ld4 te10=[te10] // 7/1:te1[s2>>16] |
| 534 | (p0) shladd te33=te33,3,te3 // 7/2:te3+s1&0xff |
| 535 | (p0) and te11=te11,maskff} // 7/2:s3>>16&0xff |
| 536 | { .mmi; (p0) ld4 te03=[te03] // 7/3:te0[s3>>24] |
| 537 | (p0) shladd te30=te30,3,te3 // 7/3:te3+s2 |
| 538 | (p0) xor t0=t0,te31 };; // 7/0: |
| 539 | { .mmi; (p0) ld4 te33=[te33] // 8/2:te3[s1] |
| 540 | (p0) shladd te11=te11,3,te1 // 8/2:te1+s3>>16 |
| 541 | (p0) xor t0=t0,te22 } // 8/0: |
| 542 | { .mmi; (p0) ld4 te30=[te30] // 8/3:te3[s2] |
| 543 | (p0) shladd te12=te12,3,te1 // 8/3:te1+s0>>16 |
| 544 | (p0) xor t1=t1,te32 };; // 8/1: |
| 545 | { .mmi; (p0) ld4 te11=[te11] // 9/2:te1[s3>>16] |
| 546 | (p0) ld4 te12=[te12] // 9/3:te1[s0>>16] |
| 547 | (p0) xor t0=t0,te00 };; // 9/0: !L2 scheduling |
| 548 | { .mmi; (p0) xor t1=t1,te23 // 10[9]/1: |
| 549 | (p0) xor t2=t2,te20 // 10[9]/2: |
| 550 | (p0) xor t3=t3,te21 };; // 10[9]/3: |
| 551 | { .mmi; (p0) xor t0=t0,te13 // 11[10]/0:done! |
| 552 | (p0) xor t1=t1,te01 // 11[10]/1: |
| 553 | (p0) xor t2=t2,te02 };; // 11[10]/2: !L2 scheduling |
| 554 | { .mmi; (p0) xor t3=t3,te03 // 12[10]/3: |
| 555 | (p16) cmp.eq p0,p17=r0,r0 };; // 12[10]/clear (p17) |
| 556 | { .mmi; (p0) xor t1=t1,te10 // 13[11]/1:done! |
| 557 | (p0) xor t2=t2,te33 // 13[11]/2: |
| 558 | (p0) xor t3=t3,te30 } // 13[11]/3: |
| 559 | { .mmi; (p17) add te0=2048,te0 // 13[11]/ |
| 560 | (p17) add te1=2048+64-TE1,te1};; // 13[11]/ |
| 561 | { .mib; (p0) xor t2=t2,te11 // 14[12]/2:done! |
| 562 | (p17) add te2=2048+128-TE2,te2} // 14[12]/ |
| 563 | { .mib; (p0) xor t3=t3,te12 // 14[12]/3:done! |
| 564 | (p17) add te3=2048+192-TE3,te3 // 14[12]/ |
| 565 | br.ctop.sptk .Ld_top };; |
| 566 | .Ld_end: |
| 567 | |
| 568 | |
| 569 | { .mmi; ld8 te10=[te0] // prefetch Td4 |
| 570 | ld8 te33=[te1] } |
| 571 | { .mmi; ld8 te12=[te2] |
| 572 | ld8 te30=[te3] } |
| 573 | |
| 574 | { .mmi; LDKEY t0=[rk0],2*KSZ // 0/0:rk[0] |
| 575 | and te31=s1,maskff // 0/0:s3&0xff |
| 576 | extr.u te22=s2,8,8 } // 0/0:s2>>8&0xff |
| 577 | { .mmi; LDKEY t1=[rk1],2*KSZ // 0/1:rk[1] |
| 578 | and te32=s2,maskff // 0/1:s0&0xff |
| 579 | shr.u te00=s0,twenty4 };; // 0/0:s0>>24 |
| 580 | { .mmi; LDKEY t2=[rk0],2*KSZ // 1/2:rk[2] |
| 581 | add te31=te31,te0 // 1/0:te0+s0>>24 |
| 582 | extr.u te23=s3,8,8 } // 1/1:s3>>8&0xff |
| 583 | { .mmi; LDKEY t3=[rk1],2*KSZ // 1/3:rk[3] |
| 584 | add te32=te32,te0 // 1/1:te0+s0 |
| 585 | shr.u te01=s1,twenty4 };; // 1/1:s1>>24 |
| 586 | { .mmi; ld1 te31=[te31] // 2/0:te0[s3&0xff] |
| 587 | add te22=te22,te0 // 2/0:te0+s2>>8&0xff |
| 588 | extr.u te20=s0,8,8 } // 2/2:s0>>8&0xff |
| 589 | { .mmi; ld1 te32=[te32] // 2/1:te0[s0] |
| 590 | add te23=te23,te0 // 2/1:te0+s3>>8 |
| 591 | shr.u te02=s2,twenty4 };; // 2/2:s2>>24 |
| 592 | { .mmi; ld1 te22=[te22] // 3/0:te0[s2>>8] |
| 593 | add te20=te20,te0 // 3/2:te0+s0>>8 |
| 594 | extr.u te21=s1,8,8 } // 3/3:s1>>8&0xff |
| 595 | { .mmi; ld1 te23=[te23] // 3/1:te0[s3>>8] |
| 596 | add te00=te00,te0 // 3/0:te0+s0>>24 |
| 597 | shr.u te03=s3,twenty4 };; // 3/3:s3>>24 |
| 598 | { .mmi; ld1 te20=[te20] // 4/2:te0[s0>>8] |
| 599 | add te21=te21,te0 // 4/3:te0+s2 |
| 600 | extr.u te13=s3,16,8 } // 4/0:s1>>16&0xff |
| 601 | { .mmi; ld1 te00=[te00] // 4/0:te0[s0>>24] |
| 602 | add te01=te01,te0 // 4/1:te0+s1>>24 |
| 603 | shr.u te11=s1,sixteen };; // 4/2:s3>>16 |
| 604 | { .mmi; ld1 te21=[te21] // 5/3:te0[s1>>8] |
| 605 | add te13=te13,te0 // 5/0:te0+s1>>16 |
| 606 | extr.u te10=s0,16,8 } // 5/1:s2>>16&0xff |
| 607 | { .mmi; ld1 te01=[te01] // 5/1:te0[s1>>24] |
| 608 | add te02=te02,te0 // 5/2:te0+s2>>24 |
| 609 | and te33=s3,maskff };; // 5/2:s1&0xff |
| 610 | { .mmi; ld1 te13=[te13] // 6/0:te0[s1>>16] |
| 611 | add te10=te10,te0 // 6/1:te0+s2>>16 |
| 612 | extr.u te12=s2,16,8 } // 6/3:s0>>16&0xff |
| 613 | { .mmi; ld1 te02=[te02] // 6/2:te0[s2>>24] |
| 614 | add te03=te03,te0 // 6/3:te0+s0>>16 |
| 615 | and te30=s0,maskff };; // 6/3:s2&0xff |
| 616 | |
| 617 | { .mmi; ld1 te10=[te10] // 7/1:te0[s2>>16] |
| 618 | add te33=te33,te0 // 7/2:te0+s1&0xff |
| 619 | dep te31=te22,te31,8,8} // 7/0: |
| 620 | { .mmi; ld1 te03=[te03] // 7/3:te0[s3>>24] |
| 621 | add te30=te30,te0 // 7/3:te0+s2 |
| 622 | and te11=te11,maskff};; // 7/2:s3>>16&0xff |
| 623 | { .mmi; ld1 te33=[te33] // 8/2:te0[s1] |
| 624 | add te11=te11,te0 // 8/2:te0+s3>>16 |
| 625 | dep te32=te23,te32,8,8} // 8/1: |
| 626 | { .mmi; ld1 te30=[te30] // 8/3:te0[s2] |
| 627 | add te12=te12,te0 // 8/3:te0+s0>>16 |
| 628 | shl te00=te00,twenty4};; // 8/0: |
| 629 | { .mii; ld1 te11=[te11] // 9/2:te0[s3>>16] |
| 630 | dep te31=te13,te31,16,8 // 9/0: |
| 631 | shl te01=te01,twenty4};; // 9/1: |
| 632 | { .mii; ld1 te12=[te12] // 10/3:te0[s0>>16] |
| 633 | dep te33=te20,te33,8,8 // 10/2: |
| 634 | shl te02=te02,twenty4};; // 10/2: |
| 635 | { .mii; xor t0=t0,te31 // 11/0: |
| 636 | dep te30=te21,te30,8,8 // 11/3: |
| 637 | shl te10=te10,sixteen};; // 11/1: |
| 638 | { .mii; xor r16=t0,te00 // 12/0:done! |
| 639 | dep te33=te11,te33,16,8 // 12/2: |
| 640 | shl te03=te03,twenty4};; // 12/3: |
| 641 | { .mmi; xor t1=t1,te01 // 13/1: |
| 642 | xor t2=t2,te02 // 13/2: |
| 643 | dep te30=te12,te30,16,8};; // 13/3: |
| 644 | { .mmi; xor t1=t1,te32 // 14/1: |
| 645 | xor r24=t2,te33 // 14/2:done! |
| 646 | xor t3=t3,te30 };; // 14/3: |
| 647 | { .mib; xor r20=t1,te10 // 15/1:done! |
| 648 | xor r28=t3,te03 // 15/3:done! |
| 649 | br.ret.sptk b6 };; |
| 650 | .endp _ia64_AES_decrypt# |
| 651 | |
| 652 | // void AES_decrypt (const void *in,void *out,const AES_KEY *key); |
| 653 | .global AES_decrypt# |
| 654 | .proc AES_decrypt# |
| 655 | .align 32 |
| 656 | AES_decrypt: |
| 657 | .prologue |
| 658 | .save ar.pfs,pfssave |
| 659 | { .mmi; alloc pfssave=ar.pfs,3,1,12,0 |
| 660 | and out0=3,in0 |
| 661 | mov r3=ip } |
| 662 | { .mmi; ADDP in0=0,in0 |
| 663 | mov loc0=psr.um |
| 664 | ADDP out11=KSZ*60,in2 };; // &AES_KEY->rounds |
| 665 | |
| 666 | { .mmi; ld4 out11=[out11] // AES_KEY->rounds |
| 667 | add out8=(AES_Td#-AES_decrypt#),r3 // Te0 |
| 668 | .save pr,prsave |
| 669 | mov prsave=pr } |
| 670 | { .mmi; rum 1<<3 // clear um.ac |
| 671 | .save ar.lc,lcsave |
| 672 | mov lcsave=ar.lc };; |
| 673 | |
| 674 | .body |
| 675 | #if defined(_HPUX_SOURCE) // HPUX is big-endian, cut 15+15 cycles... |
| 676 | { .mib; cmp.ne p6,p0=out0,r0 |
| 677 | add out0=4,in0 |
| 678 | (p6) br.dpnt.many .Ld_i_unaligned };; |
| 679 | |
| 680 | { .mmi; ld4 out1=[in0],8 // s0 |
| 681 | and out9=3,in1 |
| 682 | mov twenty4=24 } |
| 683 | { .mmi; ld4 out3=[out0],8 // s1 |
| 684 | ADDP rk0=0,in2 |
| 685 | mov sixteen=16 };; |
| 686 | { .mmi; ld4 out5=[in0] // s2 |
| 687 | cmp.ne p6,p0=out9,r0 |
| 688 | mov maskff=0xff } |
| 689 | { .mmb; ld4 out7=[out0] // s3 |
| 690 | ADDP rk1=KSZ,in2 |
| 691 | br.call.sptk.many b6=_ia64_AES_decrypt };; |
| 692 | |
| 693 | { .mib; ADDP in0=4,in1 |
| 694 | ADDP in1=0,in1 |
| 695 | (p6) br.spnt .Ld_o_unaligned };; |
| 696 | |
| 697 | { .mii; mov psr.um=loc0 |
| 698 | mov ar.pfs=pfssave |
| 699 | mov ar.lc=lcsave };; |
| 700 | { .mmi; st4 [in1]=r16,8 // s0 |
| 701 | st4 [in0]=r20,8 // s1 |
| 702 | mov pr=prsave,0x1ffff };; |
| 703 | { .mmb; st4 [in1]=r24 // s2 |
| 704 | st4 [in0]=r28 // s3 |
| 705 | br.ret.sptk.many b0 };; |
| 706 | #endif |
| 707 | |
| 708 | .align 32 |
| 709 | .Ld_i_unaligned: |
| 710 | { .mmi; add out0=1,in0 |
| 711 | add out2=2,in0 |
| 712 | add out4=3,in0 };; |
| 713 | { .mmi; ld1 r16=[in0],4 |
| 714 | ld1 r17=[out0],4 }//;; |
| 715 | { .mmi; ld1 r18=[out2],4 |
| 716 | ld1 out1=[out4],4 };; // s0 |
| 717 | { .mmi; ld1 r20=[in0],4 |
| 718 | ld1 r21=[out0],4 }//;; |
| 719 | { .mmi; ld1 r22=[out2],4 |
| 720 | ld1 out3=[out4],4 };; // s1 |
| 721 | { .mmi; ld1 r24=[in0],4 |
| 722 | ld1 r25=[out0],4 }//;; |
| 723 | { .mmi; ld1 r26=[out2],4 |
| 724 | ld1 out5=[out4],4 };; // s2 |
| 725 | { .mmi; ld1 r28=[in0] |
| 726 | ld1 r29=[out0] }//;; |
| 727 | { .mmi; ld1 r30=[out2] |
| 728 | ld1 out7=[out4] };; // s3 |
| 729 | |
| 730 | { .mii; |
| 731 | dep out1=r16,out1,24,8 //;; |
| 732 | dep out3=r20,out3,24,8 }//;; |
| 733 | { .mii; ADDP rk0=0,in2 |
| 734 | dep out5=r24,out5,24,8 //;; |
| 735 | dep out7=r28,out7,24,8 };; |
| 736 | { .mii; ADDP rk1=KSZ,in2 |
| 737 | dep out1=r17,out1,16,8 //;; |
| 738 | dep out3=r21,out3,16,8 }//;; |
| 739 | { .mii; mov twenty4=24 |
| 740 | dep out5=r25,out5,16,8 //;; |
| 741 | dep out7=r29,out7,16,8 };; |
| 742 | { .mii; mov sixteen=16 |
| 743 | dep out1=r18,out1,8,8 //;; |
| 744 | dep out3=r22,out3,8,8 }//;; |
| 745 | { .mii; mov maskff=0xff |
| 746 | dep out5=r26,out5,8,8 //;; |
| 747 | dep out7=r30,out7,8,8 };; |
| 748 | |
| 749 | { .mib; br.call.sptk.many b6=_ia64_AES_decrypt };; |
| 750 | |
| 751 | .Ld_o_unaligned: |
| 752 | { .mii; ADDP out0=0,in1 |
| 753 | extr.u r17=r16,8,8 // s0 |
| 754 | shr.u r19=r16,twenty4 }//;; |
| 755 | { .mii; ADDP out1=1,in1 |
| 756 | extr.u r18=r16,16,8 |
| 757 | shr.u r23=r20,twenty4 }//;; // s1 |
| 758 | { .mii; ADDP out2=2,in1 |
| 759 | extr.u r21=r20,8,8 |
| 760 | shr.u r22=r20,sixteen }//;; |
| 761 | { .mii; ADDP out3=3,in1 |
| 762 | extr.u r25=r24,8,8 // s2 |
| 763 | shr.u r27=r24,twenty4 };; |
| 764 | { .mii; st1 [out3]=r16,4 |
| 765 | extr.u r26=r24,16,8 |
| 766 | shr.u r31=r28,twenty4 }//;; // s3 |
| 767 | { .mii; st1 [out2]=r17,4 |
| 768 | extr.u r29=r28,8,8 |
| 769 | shr.u r30=r28,sixteen }//;; |
| 770 | |
| 771 | { .mmi; st1 [out1]=r18,4 |
| 772 | st1 [out0]=r19,4 };; |
| 773 | { .mmi; st1 [out3]=r20,4 |
| 774 | st1 [out2]=r21,4 }//;; |
| 775 | { .mmi; st1 [out1]=r22,4 |
| 776 | st1 [out0]=r23,4 };; |
| 777 | { .mmi; st1 [out3]=r24,4 |
| 778 | st1 [out2]=r25,4 |
| 779 | mov pr=prsave,0x1ffff }//;; |
| 780 | { .mmi; st1 [out1]=r26,4 |
| 781 | st1 [out0]=r27,4 |
| 782 | mov ar.pfs=pfssave };; |
| 783 | { .mmi; st1 [out3]=r28 |
| 784 | st1 [out2]=r29 |
| 785 | mov ar.lc=lcsave }//;; |
| 786 | { .mmi; st1 [out1]=r30 |
| 787 | st1 [out0]=r31 } |
| 788 | { .mfb; mov psr.um=loc0 // restore user mask |
| 789 | br.ret.sptk.many b0 };; |
| 790 | .endp AES_decrypt# |
| 791 | |
| 792 | // leave it in .text segment... |
| 793 | .align 64 |
| 794 | .global AES_Te# |
| 795 | .type AES_Te#,@object |
| 796 | AES_Te: data4 0xc66363a5,0xc66363a5, 0xf87c7c84,0xf87c7c84 |
| 797 | data4 0xee777799,0xee777799, 0xf67b7b8d,0xf67b7b8d |
| 798 | data4 0xfff2f20d,0xfff2f20d, 0xd66b6bbd,0xd66b6bbd |
| 799 | data4 0xde6f6fb1,0xde6f6fb1, 0x91c5c554,0x91c5c554 |
| 800 | data4 0x60303050,0x60303050, 0x02010103,0x02010103 |
| 801 | data4 0xce6767a9,0xce6767a9, 0x562b2b7d,0x562b2b7d |
| 802 | data4 0xe7fefe19,0xe7fefe19, 0xb5d7d762,0xb5d7d762 |
| 803 | data4 0x4dababe6,0x4dababe6, 0xec76769a,0xec76769a |
| 804 | data4 0x8fcaca45,0x8fcaca45, 0x1f82829d,0x1f82829d |
| 805 | data4 0x89c9c940,0x89c9c940, 0xfa7d7d87,0xfa7d7d87 |
| 806 | data4 0xeffafa15,0xeffafa15, 0xb25959eb,0xb25959eb |
| 807 | data4 0x8e4747c9,0x8e4747c9, 0xfbf0f00b,0xfbf0f00b |
| 808 | data4 0x41adadec,0x41adadec, 0xb3d4d467,0xb3d4d467 |
| 809 | data4 0x5fa2a2fd,0x5fa2a2fd, 0x45afafea,0x45afafea |
| 810 | data4 0x239c9cbf,0x239c9cbf, 0x53a4a4f7,0x53a4a4f7 |
| 811 | data4 0xe4727296,0xe4727296, 0x9bc0c05b,0x9bc0c05b |
| 812 | data4 0x75b7b7c2,0x75b7b7c2, 0xe1fdfd1c,0xe1fdfd1c |
| 813 | data4 0x3d9393ae,0x3d9393ae, 0x4c26266a,0x4c26266a |
| 814 | data4 0x6c36365a,0x6c36365a, 0x7e3f3f41,0x7e3f3f41 |
| 815 | data4 0xf5f7f702,0xf5f7f702, 0x83cccc4f,0x83cccc4f |
| 816 | data4 0x6834345c,0x6834345c, 0x51a5a5f4,0x51a5a5f4 |
| 817 | data4 0xd1e5e534,0xd1e5e534, 0xf9f1f108,0xf9f1f108 |
| 818 | data4 0xe2717193,0xe2717193, 0xabd8d873,0xabd8d873 |
| 819 | data4 0x62313153,0x62313153, 0x2a15153f,0x2a15153f |
| 820 | data4 0x0804040c,0x0804040c, 0x95c7c752,0x95c7c752 |
| 821 | data4 0x46232365,0x46232365, 0x9dc3c35e,0x9dc3c35e |
| 822 | data4 0x30181828,0x30181828, 0x379696a1,0x379696a1 |
| 823 | data4 0x0a05050f,0x0a05050f, 0x2f9a9ab5,0x2f9a9ab5 |
| 824 | data4 0x0e070709,0x0e070709, 0x24121236,0x24121236 |
| 825 | data4 0x1b80809b,0x1b80809b, 0xdfe2e23d,0xdfe2e23d |
| 826 | data4 0xcdebeb26,0xcdebeb26, 0x4e272769,0x4e272769 |
| 827 | data4 0x7fb2b2cd,0x7fb2b2cd, 0xea75759f,0xea75759f |
| 828 | data4 0x1209091b,0x1209091b, 0x1d83839e,0x1d83839e |
| 829 | data4 0x582c2c74,0x582c2c74, 0x341a1a2e,0x341a1a2e |
| 830 | data4 0x361b1b2d,0x361b1b2d, 0xdc6e6eb2,0xdc6e6eb2 |
| 831 | data4 0xb45a5aee,0xb45a5aee, 0x5ba0a0fb,0x5ba0a0fb |
| 832 | data4 0xa45252f6,0xa45252f6, 0x763b3b4d,0x763b3b4d |
| 833 | data4 0xb7d6d661,0xb7d6d661, 0x7db3b3ce,0x7db3b3ce |
| 834 | data4 0x5229297b,0x5229297b, 0xdde3e33e,0xdde3e33e |
| 835 | data4 0x5e2f2f71,0x5e2f2f71, 0x13848497,0x13848497 |
| 836 | data4 0xa65353f5,0xa65353f5, 0xb9d1d168,0xb9d1d168 |
| 837 | data4 0x00000000,0x00000000, 0xc1eded2c,0xc1eded2c |
| 838 | data4 0x40202060,0x40202060, 0xe3fcfc1f,0xe3fcfc1f |
| 839 | data4 0x79b1b1c8,0x79b1b1c8, 0xb65b5bed,0xb65b5bed |
| 840 | data4 0xd46a6abe,0xd46a6abe, 0x8dcbcb46,0x8dcbcb46 |
| 841 | data4 0x67bebed9,0x67bebed9, 0x7239394b,0x7239394b |
| 842 | data4 0x944a4ade,0x944a4ade, 0x984c4cd4,0x984c4cd4 |
| 843 | data4 0xb05858e8,0xb05858e8, 0x85cfcf4a,0x85cfcf4a |
| 844 | data4 0xbbd0d06b,0xbbd0d06b, 0xc5efef2a,0xc5efef2a |
| 845 | data4 0x4faaaae5,0x4faaaae5, 0xedfbfb16,0xedfbfb16 |
| 846 | data4 0x864343c5,0x864343c5, 0x9a4d4dd7,0x9a4d4dd7 |
| 847 | data4 0x66333355,0x66333355, 0x11858594,0x11858594 |
| 848 | data4 0x8a4545cf,0x8a4545cf, 0xe9f9f910,0xe9f9f910 |
| 849 | data4 0x04020206,0x04020206, 0xfe7f7f81,0xfe7f7f81 |
| 850 | data4 0xa05050f0,0xa05050f0, 0x783c3c44,0x783c3c44 |
| 851 | data4 0x259f9fba,0x259f9fba, 0x4ba8a8e3,0x4ba8a8e3 |
| 852 | data4 0xa25151f3,0xa25151f3, 0x5da3a3fe,0x5da3a3fe |
| 853 | data4 0x804040c0,0x804040c0, 0x058f8f8a,0x058f8f8a |
| 854 | data4 0x3f9292ad,0x3f9292ad, 0x219d9dbc,0x219d9dbc |
| 855 | data4 0x70383848,0x70383848, 0xf1f5f504,0xf1f5f504 |
| 856 | data4 0x63bcbcdf,0x63bcbcdf, 0x77b6b6c1,0x77b6b6c1 |
| 857 | data4 0xafdada75,0xafdada75, 0x42212163,0x42212163 |
| 858 | data4 0x20101030,0x20101030, 0xe5ffff1a,0xe5ffff1a |
| 859 | data4 0xfdf3f30e,0xfdf3f30e, 0xbfd2d26d,0xbfd2d26d |
| 860 | data4 0x81cdcd4c,0x81cdcd4c, 0x180c0c14,0x180c0c14 |
| 861 | data4 0x26131335,0x26131335, 0xc3ecec2f,0xc3ecec2f |
| 862 | data4 0xbe5f5fe1,0xbe5f5fe1, 0x359797a2,0x359797a2 |
| 863 | data4 0x884444cc,0x884444cc, 0x2e171739,0x2e171739 |
| 864 | data4 0x93c4c457,0x93c4c457, 0x55a7a7f2,0x55a7a7f2 |
| 865 | data4 0xfc7e7e82,0xfc7e7e82, 0x7a3d3d47,0x7a3d3d47 |
| 866 | data4 0xc86464ac,0xc86464ac, 0xba5d5de7,0xba5d5de7 |
| 867 | data4 0x3219192b,0x3219192b, 0xe6737395,0xe6737395 |
| 868 | data4 0xc06060a0,0xc06060a0, 0x19818198,0x19818198 |
| 869 | data4 0x9e4f4fd1,0x9e4f4fd1, 0xa3dcdc7f,0xa3dcdc7f |
| 870 | data4 0x44222266,0x44222266, 0x542a2a7e,0x542a2a7e |
| 871 | data4 0x3b9090ab,0x3b9090ab, 0x0b888883,0x0b888883 |
| 872 | data4 0x8c4646ca,0x8c4646ca, 0xc7eeee29,0xc7eeee29 |
| 873 | data4 0x6bb8b8d3,0x6bb8b8d3, 0x2814143c,0x2814143c |
| 874 | data4 0xa7dede79,0xa7dede79, 0xbc5e5ee2,0xbc5e5ee2 |
| 875 | data4 0x160b0b1d,0x160b0b1d, 0xaddbdb76,0xaddbdb76 |
| 876 | data4 0xdbe0e03b,0xdbe0e03b, 0x64323256,0x64323256 |
| 877 | data4 0x743a3a4e,0x743a3a4e, 0x140a0a1e,0x140a0a1e |
| 878 | data4 0x924949db,0x924949db, 0x0c06060a,0x0c06060a |
| 879 | data4 0x4824246c,0x4824246c, 0xb85c5ce4,0xb85c5ce4 |
| 880 | data4 0x9fc2c25d,0x9fc2c25d, 0xbdd3d36e,0xbdd3d36e |
| 881 | data4 0x43acacef,0x43acacef, 0xc46262a6,0xc46262a6 |
| 882 | data4 0x399191a8,0x399191a8, 0x319595a4,0x319595a4 |
| 883 | data4 0xd3e4e437,0xd3e4e437, 0xf279798b,0xf279798b |
| 884 | data4 0xd5e7e732,0xd5e7e732, 0x8bc8c843,0x8bc8c843 |
| 885 | data4 0x6e373759,0x6e373759, 0xda6d6db7,0xda6d6db7 |
| 886 | data4 0x018d8d8c,0x018d8d8c, 0xb1d5d564,0xb1d5d564 |
| 887 | data4 0x9c4e4ed2,0x9c4e4ed2, 0x49a9a9e0,0x49a9a9e0 |
| 888 | data4 0xd86c6cb4,0xd86c6cb4, 0xac5656fa,0xac5656fa |
| 889 | data4 0xf3f4f407,0xf3f4f407, 0xcfeaea25,0xcfeaea25 |
| 890 | data4 0xca6565af,0xca6565af, 0xf47a7a8e,0xf47a7a8e |
| 891 | data4 0x47aeaee9,0x47aeaee9, 0x10080818,0x10080818 |
| 892 | data4 0x6fbabad5,0x6fbabad5, 0xf0787888,0xf0787888 |
| 893 | data4 0x4a25256f,0x4a25256f, 0x5c2e2e72,0x5c2e2e72 |
| 894 | data4 0x381c1c24,0x381c1c24, 0x57a6a6f1,0x57a6a6f1 |
| 895 | data4 0x73b4b4c7,0x73b4b4c7, 0x97c6c651,0x97c6c651 |
| 896 | data4 0xcbe8e823,0xcbe8e823, 0xa1dddd7c,0xa1dddd7c |
| 897 | data4 0xe874749c,0xe874749c, 0x3e1f1f21,0x3e1f1f21 |
| 898 | data4 0x964b4bdd,0x964b4bdd, 0x61bdbddc,0x61bdbddc |
| 899 | data4 0x0d8b8b86,0x0d8b8b86, 0x0f8a8a85,0x0f8a8a85 |
| 900 | data4 0xe0707090,0xe0707090, 0x7c3e3e42,0x7c3e3e42 |
| 901 | data4 0x71b5b5c4,0x71b5b5c4, 0xcc6666aa,0xcc6666aa |
| 902 | data4 0x904848d8,0x904848d8, 0x06030305,0x06030305 |
| 903 | data4 0xf7f6f601,0xf7f6f601, 0x1c0e0e12,0x1c0e0e12 |
| 904 | data4 0xc26161a3,0xc26161a3, 0x6a35355f,0x6a35355f |
| 905 | data4 0xae5757f9,0xae5757f9, 0x69b9b9d0,0x69b9b9d0 |
| 906 | data4 0x17868691,0x17868691, 0x99c1c158,0x99c1c158 |
| 907 | data4 0x3a1d1d27,0x3a1d1d27, 0x279e9eb9,0x279e9eb9 |
| 908 | data4 0xd9e1e138,0xd9e1e138, 0xebf8f813,0xebf8f813 |
| 909 | data4 0x2b9898b3,0x2b9898b3, 0x22111133,0x22111133 |
| 910 | data4 0xd26969bb,0xd26969bb, 0xa9d9d970,0xa9d9d970 |
| 911 | data4 0x078e8e89,0x078e8e89, 0x339494a7,0x339494a7 |
| 912 | data4 0x2d9b9bb6,0x2d9b9bb6, 0x3c1e1e22,0x3c1e1e22 |
| 913 | data4 0x15878792,0x15878792, 0xc9e9e920,0xc9e9e920 |
| 914 | data4 0x87cece49,0x87cece49, 0xaa5555ff,0xaa5555ff |
| 915 | data4 0x50282878,0x50282878, 0xa5dfdf7a,0xa5dfdf7a |
| 916 | data4 0x038c8c8f,0x038c8c8f, 0x59a1a1f8,0x59a1a1f8 |
| 917 | data4 0x09898980,0x09898980, 0x1a0d0d17,0x1a0d0d17 |
| 918 | data4 0x65bfbfda,0x65bfbfda, 0xd7e6e631,0xd7e6e631 |
| 919 | data4 0x844242c6,0x844242c6, 0xd06868b8,0xd06868b8 |
| 920 | data4 0x824141c3,0x824141c3, 0x299999b0,0x299999b0 |
| 921 | data4 0x5a2d2d77,0x5a2d2d77, 0x1e0f0f11,0x1e0f0f11 |
| 922 | data4 0x7bb0b0cb,0x7bb0b0cb, 0xa85454fc,0xa85454fc |
| 923 | data4 0x6dbbbbd6,0x6dbbbbd6, 0x2c16163a,0x2c16163a |
| 924 | // Te4: |
| 925 | data1 0x63, 0x7c, 0x77, 0x7b, 0xf2, 0x6b, 0x6f, 0xc5 |
| 926 | data1 0x30, 0x01, 0x67, 0x2b, 0xfe, 0xd7, 0xab, 0x76 |
| 927 | data1 0xca, 0x82, 0xc9, 0x7d, 0xfa, 0x59, 0x47, 0xf0 |
| 928 | data1 0xad, 0xd4, 0xa2, 0xaf, 0x9c, 0xa4, 0x72, 0xc0 |
| 929 | data1 0xb7, 0xfd, 0x93, 0x26, 0x36, 0x3f, 0xf7, 0xcc |
| 930 | data1 0x34, 0xa5, 0xe5, 0xf1, 0x71, 0xd8, 0x31, 0x15 |
| 931 | data1 0x04, 0xc7, 0x23, 0xc3, 0x18, 0x96, 0x05, 0x9a |
| 932 | data1 0x07, 0x12, 0x80, 0xe2, 0xeb, 0x27, 0xb2, 0x75 |
| 933 | data1 0x09, 0x83, 0x2c, 0x1a, 0x1b, 0x6e, 0x5a, 0xa0 |
| 934 | data1 0x52, 0x3b, 0xd6, 0xb3, 0x29, 0xe3, 0x2f, 0x84 |
| 935 | data1 0x53, 0xd1, 0x00, 0xed, 0x20, 0xfc, 0xb1, 0x5b |
| 936 | data1 0x6a, 0xcb, 0xbe, 0x39, 0x4a, 0x4c, 0x58, 0xcf |
| 937 | data1 0xd0, 0xef, 0xaa, 0xfb, 0x43, 0x4d, 0x33, 0x85 |
| 938 | data1 0x45, 0xf9, 0x02, 0x7f, 0x50, 0x3c, 0x9f, 0xa8 |
| 939 | data1 0x51, 0xa3, 0x40, 0x8f, 0x92, 0x9d, 0x38, 0xf5 |
| 940 | data1 0xbc, 0xb6, 0xda, 0x21, 0x10, 0xff, 0xf3, 0xd2 |
| 941 | data1 0xcd, 0x0c, 0x13, 0xec, 0x5f, 0x97, 0x44, 0x17 |
| 942 | data1 0xc4, 0xa7, 0x7e, 0x3d, 0x64, 0x5d, 0x19, 0x73 |
| 943 | data1 0x60, 0x81, 0x4f, 0xdc, 0x22, 0x2a, 0x90, 0x88 |
| 944 | data1 0x46, 0xee, 0xb8, 0x14, 0xde, 0x5e, 0x0b, 0xdb |
| 945 | data1 0xe0, 0x32, 0x3a, 0x0a, 0x49, 0x06, 0x24, 0x5c |
| 946 | data1 0xc2, 0xd3, 0xac, 0x62, 0x91, 0x95, 0xe4, 0x79 |
| 947 | data1 0xe7, 0xc8, 0x37, 0x6d, 0x8d, 0xd5, 0x4e, 0xa9 |
| 948 | data1 0x6c, 0x56, 0xf4, 0xea, 0x65, 0x7a, 0xae, 0x08 |
| 949 | data1 0xba, 0x78, 0x25, 0x2e, 0x1c, 0xa6, 0xb4, 0xc6 |
| 950 | data1 0xe8, 0xdd, 0x74, 0x1f, 0x4b, 0xbd, 0x8b, 0x8a |
| 951 | data1 0x70, 0x3e, 0xb5, 0x66, 0x48, 0x03, 0xf6, 0x0e |
| 952 | data1 0x61, 0x35, 0x57, 0xb9, 0x86, 0xc1, 0x1d, 0x9e |
| 953 | data1 0xe1, 0xf8, 0x98, 0x11, 0x69, 0xd9, 0x8e, 0x94 |
| 954 | data1 0x9b, 0x1e, 0x87, 0xe9, 0xce, 0x55, 0x28, 0xdf |
| 955 | data1 0x8c, 0xa1, 0x89, 0x0d, 0xbf, 0xe6, 0x42, 0x68 |
| 956 | data1 0x41, 0x99, 0x2d, 0x0f, 0xb0, 0x54, 0xbb, 0x16 |
| 957 | .size AES_Te#,2048+256 // HP-UX assembler fails to ".-AES_Te#" |
| 958 | |
| 959 | .align 64 |
| 960 | .global AES_Td# |
| 961 | .type AES_Td#,@object |
| 962 | AES_Td: data4 0x51f4a750,0x51f4a750, 0x7e416553,0x7e416553 |
| 963 | data4 0x1a17a4c3,0x1a17a4c3, 0x3a275e96,0x3a275e96 |
| 964 | data4 0x3bab6bcb,0x3bab6bcb, 0x1f9d45f1,0x1f9d45f1 |
| 965 | data4 0xacfa58ab,0xacfa58ab, 0x4be30393,0x4be30393 |
| 966 | data4 0x2030fa55,0x2030fa55, 0xad766df6,0xad766df6 |
| 967 | data4 0x88cc7691,0x88cc7691, 0xf5024c25,0xf5024c25 |
| 968 | data4 0x4fe5d7fc,0x4fe5d7fc, 0xc52acbd7,0xc52acbd7 |
| 969 | data4 0x26354480,0x26354480, 0xb562a38f,0xb562a38f |
| 970 | data4 0xdeb15a49,0xdeb15a49, 0x25ba1b67,0x25ba1b67 |
| 971 | data4 0x45ea0e98,0x45ea0e98, 0x5dfec0e1,0x5dfec0e1 |
| 972 | data4 0xc32f7502,0xc32f7502, 0x814cf012,0x814cf012 |
| 973 | data4 0x8d4697a3,0x8d4697a3, 0x6bd3f9c6,0x6bd3f9c6 |
| 974 | data4 0x038f5fe7,0x038f5fe7, 0x15929c95,0x15929c95 |
| 975 | data4 0xbf6d7aeb,0xbf6d7aeb, 0x955259da,0x955259da |
| 976 | data4 0xd4be832d,0xd4be832d, 0x587421d3,0x587421d3 |
| 977 | data4 0x49e06929,0x49e06929, 0x8ec9c844,0x8ec9c844 |
| 978 | data4 0x75c2896a,0x75c2896a, 0xf48e7978,0xf48e7978 |
| 979 | data4 0x99583e6b,0x99583e6b, 0x27b971dd,0x27b971dd |
| 980 | data4 0xbee14fb6,0xbee14fb6, 0xf088ad17,0xf088ad17 |
| 981 | data4 0xc920ac66,0xc920ac66, 0x7dce3ab4,0x7dce3ab4 |
| 982 | data4 0x63df4a18,0x63df4a18, 0xe51a3182,0xe51a3182 |
| 983 | data4 0x97513360,0x97513360, 0x62537f45,0x62537f45 |
| 984 | data4 0xb16477e0,0xb16477e0, 0xbb6bae84,0xbb6bae84 |
| 985 | data4 0xfe81a01c,0xfe81a01c, 0xf9082b94,0xf9082b94 |
| 986 | data4 0x70486858,0x70486858, 0x8f45fd19,0x8f45fd19 |
| 987 | data4 0x94de6c87,0x94de6c87, 0x527bf8b7,0x527bf8b7 |
| 988 | data4 0xab73d323,0xab73d323, 0x724b02e2,0x724b02e2 |
| 989 | data4 0xe31f8f57,0xe31f8f57, 0x6655ab2a,0x6655ab2a |
| 990 | data4 0xb2eb2807,0xb2eb2807, 0x2fb5c203,0x2fb5c203 |
| 991 | data4 0x86c57b9a,0x86c57b9a, 0xd33708a5,0xd33708a5 |
| 992 | data4 0x302887f2,0x302887f2, 0x23bfa5b2,0x23bfa5b2 |
| 993 | data4 0x02036aba,0x02036aba, 0xed16825c,0xed16825c |
| 994 | data4 0x8acf1c2b,0x8acf1c2b, 0xa779b492,0xa779b492 |
| 995 | data4 0xf307f2f0,0xf307f2f0, 0x4e69e2a1,0x4e69e2a1 |
| 996 | data4 0x65daf4cd,0x65daf4cd, 0x0605bed5,0x0605bed5 |
| 997 | data4 0xd134621f,0xd134621f, 0xc4a6fe8a,0xc4a6fe8a |
| 998 | data4 0x342e539d,0x342e539d, 0xa2f355a0,0xa2f355a0 |
| 999 | data4 0x058ae132,0x058ae132, 0xa4f6eb75,0xa4f6eb75 |
| 1000 | data4 0x0b83ec39,0x0b83ec39, 0x4060efaa,0x4060efaa |
| 1001 | data4 0x5e719f06,0x5e719f06, 0xbd6e1051,0xbd6e1051 |
| 1002 | data4 0x3e218af9,0x3e218af9, 0x96dd063d,0x96dd063d |
| 1003 | data4 0xdd3e05ae,0xdd3e05ae, 0x4de6bd46,0x4de6bd46 |
| 1004 | data4 0x91548db5,0x91548db5, 0x71c45d05,0x71c45d05 |
| 1005 | data4 0x0406d46f,0x0406d46f, 0x605015ff,0x605015ff |
| 1006 | data4 0x1998fb24,0x1998fb24, 0xd6bde997,0xd6bde997 |
| 1007 | data4 0x894043cc,0x894043cc, 0x67d99e77,0x67d99e77 |
| 1008 | data4 0xb0e842bd,0xb0e842bd, 0x07898b88,0x07898b88 |
| 1009 | data4 0xe7195b38,0xe7195b38, 0x79c8eedb,0x79c8eedb |
| 1010 | data4 0xa17c0a47,0xa17c0a47, 0x7c420fe9,0x7c420fe9 |
| 1011 | data4 0xf8841ec9,0xf8841ec9, 0x00000000,0x00000000 |
| 1012 | data4 0x09808683,0x09808683, 0x322bed48,0x322bed48 |
| 1013 | data4 0x1e1170ac,0x1e1170ac, 0x6c5a724e,0x6c5a724e |
| 1014 | data4 0xfd0efffb,0xfd0efffb, 0x0f853856,0x0f853856 |
| 1015 | data4 0x3daed51e,0x3daed51e, 0x362d3927,0x362d3927 |
| 1016 | data4 0x0a0fd964,0x0a0fd964, 0x685ca621,0x685ca621 |
| 1017 | data4 0x9b5b54d1,0x9b5b54d1, 0x24362e3a,0x24362e3a |
| 1018 | data4 0x0c0a67b1,0x0c0a67b1, 0x9357e70f,0x9357e70f |
| 1019 | data4 0xb4ee96d2,0xb4ee96d2, 0x1b9b919e,0x1b9b919e |
| 1020 | data4 0x80c0c54f,0x80c0c54f, 0x61dc20a2,0x61dc20a2 |
| 1021 | data4 0x5a774b69,0x5a774b69, 0x1c121a16,0x1c121a16 |
| 1022 | data4 0xe293ba0a,0xe293ba0a, 0xc0a02ae5,0xc0a02ae5 |
| 1023 | data4 0x3c22e043,0x3c22e043, 0x121b171d,0x121b171d |
| 1024 | data4 0x0e090d0b,0x0e090d0b, 0xf28bc7ad,0xf28bc7ad |
| 1025 | data4 0x2db6a8b9,0x2db6a8b9, 0x141ea9c8,0x141ea9c8 |
| 1026 | data4 0x57f11985,0x57f11985, 0xaf75074c,0xaf75074c |
| 1027 | data4 0xee99ddbb,0xee99ddbb, 0xa37f60fd,0xa37f60fd |
| 1028 | data4 0xf701269f,0xf701269f, 0x5c72f5bc,0x5c72f5bc |
| 1029 | data4 0x44663bc5,0x44663bc5, 0x5bfb7e34,0x5bfb7e34 |
| 1030 | data4 0x8b432976,0x8b432976, 0xcb23c6dc,0xcb23c6dc |
| 1031 | data4 0xb6edfc68,0xb6edfc68, 0xb8e4f163,0xb8e4f163 |
| 1032 | data4 0xd731dcca,0xd731dcca, 0x42638510,0x42638510 |
| 1033 | data4 0x13972240,0x13972240, 0x84c61120,0x84c61120 |
| 1034 | data4 0x854a247d,0x854a247d, 0xd2bb3df8,0xd2bb3df8 |
| 1035 | data4 0xaef93211,0xaef93211, 0xc729a16d,0xc729a16d |
| 1036 | data4 0x1d9e2f4b,0x1d9e2f4b, 0xdcb230f3,0xdcb230f3 |
| 1037 | data4 0x0d8652ec,0x0d8652ec, 0x77c1e3d0,0x77c1e3d0 |
| 1038 | data4 0x2bb3166c,0x2bb3166c, 0xa970b999,0xa970b999 |
| 1039 | data4 0x119448fa,0x119448fa, 0x47e96422,0x47e96422 |
| 1040 | data4 0xa8fc8cc4,0xa8fc8cc4, 0xa0f03f1a,0xa0f03f1a |
| 1041 | data4 0x567d2cd8,0x567d2cd8, 0x223390ef,0x223390ef |
| 1042 | data4 0x87494ec7,0x87494ec7, 0xd938d1c1,0xd938d1c1 |
| 1043 | data4 0x8ccaa2fe,0x8ccaa2fe, 0x98d40b36,0x98d40b36 |
| 1044 | data4 0xa6f581cf,0xa6f581cf, 0xa57ade28,0xa57ade28 |
| 1045 | data4 0xdab78e26,0xdab78e26, 0x3fadbfa4,0x3fadbfa4 |
| 1046 | data4 0x2c3a9de4,0x2c3a9de4, 0x5078920d,0x5078920d |
| 1047 | data4 0x6a5fcc9b,0x6a5fcc9b, 0x547e4662,0x547e4662 |
| 1048 | data4 0xf68d13c2,0xf68d13c2, 0x90d8b8e8,0x90d8b8e8 |
| 1049 | data4 0x2e39f75e,0x2e39f75e, 0x82c3aff5,0x82c3aff5 |
| 1050 | data4 0x9f5d80be,0x9f5d80be, 0x69d0937c,0x69d0937c |
| 1051 | data4 0x6fd52da9,0x6fd52da9, 0xcf2512b3,0xcf2512b3 |
| 1052 | data4 0xc8ac993b,0xc8ac993b, 0x10187da7,0x10187da7 |
| 1053 | data4 0xe89c636e,0xe89c636e, 0xdb3bbb7b,0xdb3bbb7b |
| 1054 | data4 0xcd267809,0xcd267809, 0x6e5918f4,0x6e5918f4 |
| 1055 | data4 0xec9ab701,0xec9ab701, 0x834f9aa8,0x834f9aa8 |
| 1056 | data4 0xe6956e65,0xe6956e65, 0xaaffe67e,0xaaffe67e |
| 1057 | data4 0x21bccf08,0x21bccf08, 0xef15e8e6,0xef15e8e6 |
| 1058 | data4 0xbae79bd9,0xbae79bd9, 0x4a6f36ce,0x4a6f36ce |
| 1059 | data4 0xea9f09d4,0xea9f09d4, 0x29b07cd6,0x29b07cd6 |
| 1060 | data4 0x31a4b2af,0x31a4b2af, 0x2a3f2331,0x2a3f2331 |
| 1061 | data4 0xc6a59430,0xc6a59430, 0x35a266c0,0x35a266c0 |
| 1062 | data4 0x744ebc37,0x744ebc37, 0xfc82caa6,0xfc82caa6 |
| 1063 | data4 0xe090d0b0,0xe090d0b0, 0x33a7d815,0x33a7d815 |
| 1064 | data4 0xf104984a,0xf104984a, 0x41ecdaf7,0x41ecdaf7 |
| 1065 | data4 0x7fcd500e,0x7fcd500e, 0x1791f62f,0x1791f62f |
| 1066 | data4 0x764dd68d,0x764dd68d, 0x43efb04d,0x43efb04d |
| 1067 | data4 0xccaa4d54,0xccaa4d54, 0xe49604df,0xe49604df |
| 1068 | data4 0x9ed1b5e3,0x9ed1b5e3, 0x4c6a881b,0x4c6a881b |
| 1069 | data4 0xc12c1fb8,0xc12c1fb8, 0x4665517f,0x4665517f |
| 1070 | data4 0x9d5eea04,0x9d5eea04, 0x018c355d,0x018c355d |
| 1071 | data4 0xfa877473,0xfa877473, 0xfb0b412e,0xfb0b412e |
| 1072 | data4 0xb3671d5a,0xb3671d5a, 0x92dbd252,0x92dbd252 |
| 1073 | data4 0xe9105633,0xe9105633, 0x6dd64713,0x6dd64713 |
| 1074 | data4 0x9ad7618c,0x9ad7618c, 0x37a10c7a,0x37a10c7a |
| 1075 | data4 0x59f8148e,0x59f8148e, 0xeb133c89,0xeb133c89 |
| 1076 | data4 0xcea927ee,0xcea927ee, 0xb761c935,0xb761c935 |
| 1077 | data4 0xe11ce5ed,0xe11ce5ed, 0x7a47b13c,0x7a47b13c |
| 1078 | data4 0x9cd2df59,0x9cd2df59, 0x55f2733f,0x55f2733f |
| 1079 | data4 0x1814ce79,0x1814ce79, 0x73c737bf,0x73c737bf |
| 1080 | data4 0x53f7cdea,0x53f7cdea, 0x5ffdaa5b,0x5ffdaa5b |
| 1081 | data4 0xdf3d6f14,0xdf3d6f14, 0x7844db86,0x7844db86 |
| 1082 | data4 0xcaaff381,0xcaaff381, 0xb968c43e,0xb968c43e |
| 1083 | data4 0x3824342c,0x3824342c, 0xc2a3405f,0xc2a3405f |
| 1084 | data4 0x161dc372,0x161dc372, 0xbce2250c,0xbce2250c |
| 1085 | data4 0x283c498b,0x283c498b, 0xff0d9541,0xff0d9541 |
| 1086 | data4 0x39a80171,0x39a80171, 0x080cb3de,0x080cb3de |
| 1087 | data4 0xd8b4e49c,0xd8b4e49c, 0x6456c190,0x6456c190 |
| 1088 | data4 0x7bcb8461,0x7bcb8461, 0xd532b670,0xd532b670 |
| 1089 | data4 0x486c5c74,0x486c5c74, 0xd0b85742,0xd0b85742 |
| 1090 | // Td4: |
| 1091 | data1 0x52, 0x09, 0x6a, 0xd5, 0x30, 0x36, 0xa5, 0x38 |
| 1092 | data1 0xbf, 0x40, 0xa3, 0x9e, 0x81, 0xf3, 0xd7, 0xfb |
| 1093 | data1 0x7c, 0xe3, 0x39, 0x82, 0x9b, 0x2f, 0xff, 0x87 |
| 1094 | data1 0x34, 0x8e, 0x43, 0x44, 0xc4, 0xde, 0xe9, 0xcb |
| 1095 | data1 0x54, 0x7b, 0x94, 0x32, 0xa6, 0xc2, 0x23, 0x3d |
| 1096 | data1 0xee, 0x4c, 0x95, 0x0b, 0x42, 0xfa, 0xc3, 0x4e |
| 1097 | data1 0x08, 0x2e, 0xa1, 0x66, 0x28, 0xd9, 0x24, 0xb2 |
| 1098 | data1 0x76, 0x5b, 0xa2, 0x49, 0x6d, 0x8b, 0xd1, 0x25 |
| 1099 | data1 0x72, 0xf8, 0xf6, 0x64, 0x86, 0x68, 0x98, 0x16 |
| 1100 | data1 0xd4, 0xa4, 0x5c, 0xcc, 0x5d, 0x65, 0xb6, 0x92 |
| 1101 | data1 0x6c, 0x70, 0x48, 0x50, 0xfd, 0xed, 0xb9, 0xda |
| 1102 | data1 0x5e, 0x15, 0x46, 0x57, 0xa7, 0x8d, 0x9d, 0x84 |
| 1103 | data1 0x90, 0xd8, 0xab, 0x00, 0x8c, 0xbc, 0xd3, 0x0a |
| 1104 | data1 0xf7, 0xe4, 0x58, 0x05, 0xb8, 0xb3, 0x45, 0x06 |
| 1105 | data1 0xd0, 0x2c, 0x1e, 0x8f, 0xca, 0x3f, 0x0f, 0x02 |
| 1106 | data1 0xc1, 0xaf, 0xbd, 0x03, 0x01, 0x13, 0x8a, 0x6b |
| 1107 | data1 0x3a, 0x91, 0x11, 0x41, 0x4f, 0x67, 0xdc, 0xea |
| 1108 | data1 0x97, 0xf2, 0xcf, 0xce, 0xf0, 0xb4, 0xe6, 0x73 |
| 1109 | data1 0x96, 0xac, 0x74, 0x22, 0xe7, 0xad, 0x35, 0x85 |
| 1110 | data1 0xe2, 0xf9, 0x37, 0xe8, 0x1c, 0x75, 0xdf, 0x6e |
| 1111 | data1 0x47, 0xf1, 0x1a, 0x71, 0x1d, 0x29, 0xc5, 0x89 |
| 1112 | data1 0x6f, 0xb7, 0x62, 0x0e, 0xaa, 0x18, 0xbe, 0x1b |
| 1113 | data1 0xfc, 0x56, 0x3e, 0x4b, 0xc6, 0xd2, 0x79, 0x20 |
| 1114 | data1 0x9a, 0xdb, 0xc0, 0xfe, 0x78, 0xcd, 0x5a, 0xf4 |
| 1115 | data1 0x1f, 0xdd, 0xa8, 0x33, 0x88, 0x07, 0xc7, 0x31 |
| 1116 | data1 0xb1, 0x12, 0x10, 0x59, 0x27, 0x80, 0xec, 0x5f |
| 1117 | data1 0x60, 0x51, 0x7f, 0xa9, 0x19, 0xb5, 0x4a, 0x0d |
| 1118 | data1 0x2d, 0xe5, 0x7a, 0x9f, 0x93, 0xc9, 0x9c, 0xef |
| 1119 | data1 0xa0, 0xe0, 0x3b, 0x4d, 0xae, 0x2a, 0xf5, 0xb0 |
| 1120 | data1 0xc8, 0xeb, 0xbb, 0x3c, 0x83, 0x53, 0x99, 0x61 |
| 1121 | data1 0x17, 0x2b, 0x04, 0x7e, 0xba, 0x77, 0xd6, 0x26 |
| 1122 | data1 0xe1, 0x69, 0x14, 0x63, 0x55, 0x21, 0x0c, 0x7d |
| 1123 | .size AES_Td#,2048+256 // HP-UX assembler fails to ".-AES_Td#" |