Alexandre Savard | 1b09e31 | 2012-08-07 20:33:29 -0400 | [diff] [blame] | 1 | #!/usr/bin/env perl |
| 2 | # |
| 3 | # ==================================================================== |
| 4 | # Written by Andy Polyakov <appro@fy.chalmers.se> for the OpenSSL |
| 5 | # project. The module is, however, dual licensed under OpenSSL and |
| 6 | # CRYPTOGAMS licenses depending on where you obtain it. For further |
| 7 | # details see http://www.openssl.org/~appro/cryptogams/. |
| 8 | # ==================================================================== |
| 9 | # |
| 10 | # SHA256/512_Transform for Itanium. |
| 11 | # |
| 12 | # sha512_block runs in 1003 cycles on Itanium 2, which is almost 50% |
| 13 | # faster than gcc and >60%(!) faster than code generated by HP-UX |
| 14 | # compiler (yes, HP-UX is generating slower code, because unlike gcc, |
| 15 | # it failed to deploy "shift right pair," 'shrp' instruction, which |
| 16 | # substitutes for 64-bit rotate). |
| 17 | # |
| 18 | # 924 cycles long sha256_block outperforms gcc by over factor of 2(!) |
| 19 | # and HP-UX compiler - by >40% (yes, gcc won sha512_block, but lost |
| 20 | # this one big time). Note that "formally" 924 is about 100 cycles |
| 21 | # too much. I mean it's 64 32-bit rounds vs. 80 virtually identical |
| 22 | # 64-bit ones and 1003*64/80 gives 802. Extra cycles, 2 per round, |
| 23 | # are spent on extra work to provide for 32-bit rotations. 32-bit |
| 24 | # rotations are still handled by 'shrp' instruction and for this |
| 25 | # reason lower 32 bits are deposited to upper half of 64-bit register |
| 26 | # prior 'shrp' issue. And in order to minimize the amount of such |
| 27 | # operations, X[16] values are *maintained* with copies of lower |
| 28 | # halves in upper halves, which is why you'll spot such instructions |
| 29 | # as custom 'mux2', "parallel 32-bit add," 'padd4' and "parallel |
| 30 | # 32-bit unsigned right shift," 'pshr4.u' instructions here. |
| 31 | # |
| 32 | # Rules of engagement. |
| 33 | # |
| 34 | # There is only one integer shifter meaning that if I have two rotate, |
| 35 | # deposit or extract instructions in adjacent bundles, they shall |
| 36 | # split [at run-time if they have to]. But note that variable and |
| 37 | # parallel shifts are performed by multi-media ALU and *are* pairable |
| 38 | # with rotates [and alike]. On the backside MMALU is rather slow: it |
| 39 | # takes 2 extra cycles before the result of integer operation is |
| 40 | # available *to* MMALU and 2(*) extra cycles before the result of MM |
| 41 | # operation is available "back" *to* integer ALU, not to mention that |
| 42 | # MMALU itself has 2 cycles latency. However! I explicitly scheduled |
| 43 | # these MM instructions to avoid MM stalls, so that all these extra |
| 44 | # latencies get "hidden" in instruction-level parallelism. |
| 45 | # |
| 46 | # (*) 2 cycles on Itanium 1 and 1 cycle on Itanium 2. But I schedule |
| 47 | # for 2 in order to provide for best *overall* performance, |
| 48 | # because on Itanium 1 stall on MM result is accompanied by |
| 49 | # pipeline flush, which takes 6 cycles:-( |
| 50 | # |
| 51 | # Resulting performance numbers for 900MHz Itanium 2 system: |
| 52 | # |
| 53 | # The 'numbers' are in 1000s of bytes per second processed. |
| 54 | # type 16 bytes 64 bytes 256 bytes 1024 bytes 8192 bytes |
| 55 | # sha1(*) 6210.14k 20376.30k 52447.83k 85870.05k 105478.12k |
| 56 | # sha256 7476.45k 20572.05k 41538.34k 56062.29k 62093.18k |
| 57 | # sha512 4996.56k 20026.28k 47597.20k 85278.79k 111501.31k |
| 58 | # |
| 59 | # (*) SHA1 numbers are for HP-UX compiler and are presented purely |
| 60 | # for reference purposes. I bet it can improved too... |
| 61 | # |
| 62 | # To generate code, pass the file name with either 256 or 512 in its |
| 63 | # name and compiler flags. |
| 64 | |
| 65 | $output=shift; |
| 66 | |
| 67 | if ($output =~ /512.*\.[s|asm]/) { |
| 68 | $SZ=8; |
| 69 | $BITS=8*$SZ; |
| 70 | $LDW="ld8"; |
| 71 | $STW="st8"; |
| 72 | $ADD="add"; |
| 73 | $SHRU="shr.u"; |
| 74 | $TABLE="K512"; |
| 75 | $func="sha512_block_data_order"; |
| 76 | @Sigma0=(28,34,39); |
| 77 | @Sigma1=(14,18,41); |
| 78 | @sigma0=(1, 8, 7); |
| 79 | @sigma1=(19,61, 6); |
| 80 | $rounds=80; |
| 81 | } elsif ($output =~ /256.*\.[s|asm]/) { |
| 82 | $SZ=4; |
| 83 | $BITS=8*$SZ; |
| 84 | $LDW="ld4"; |
| 85 | $STW="st4"; |
| 86 | $ADD="padd4"; |
| 87 | $SHRU="pshr4.u"; |
| 88 | $TABLE="K256"; |
| 89 | $func="sha256_block_data_order"; |
| 90 | @Sigma0=( 2,13,22); |
| 91 | @Sigma1=( 6,11,25); |
| 92 | @sigma0=( 7,18, 3); |
| 93 | @sigma1=(17,19,10); |
| 94 | $rounds=64; |
| 95 | } else { die "nonsense $output"; } |
| 96 | |
| 97 | open STDOUT,">$output" || die "can't open $output: $!"; |
| 98 | |
| 99 | if ($^O eq "hpux") { |
| 100 | $ADDP="addp4"; |
| 101 | for (@ARGV) { $ADDP="add" if (/[\+DD|\-mlp]64/); } |
| 102 | } else { $ADDP="add"; } |
| 103 | for (@ARGV) { $big_endian=1 if (/\-DB_ENDIAN/); |
| 104 | $big_endian=0 if (/\-DL_ENDIAN/); } |
| 105 | if (!defined($big_endian)) |
| 106 | { $big_endian=(unpack('L',pack('N',1))==1); } |
| 107 | |
| 108 | $code=<<___; |
| 109 | .ident \"$output, version 1.1\" |
| 110 | .ident \"IA-64 ISA artwork by Andy Polyakov <appro\@fy.chalmers.se>\" |
| 111 | .explicit |
| 112 | .text |
| 113 | |
| 114 | pfssave=r2; |
| 115 | lcsave=r3; |
| 116 | prsave=r14; |
| 117 | K=r15; |
| 118 | A=r16; B=r17; C=r18; D=r19; |
| 119 | E=r20; F=r21; G=r22; H=r23; |
| 120 | T1=r24; T2=r25; |
| 121 | s0=r26; s1=r27; t0=r28; t1=r29; |
| 122 | Ktbl=r30; |
| 123 | ctx=r31; // 1st arg |
| 124 | input=r48; // 2nd arg |
| 125 | num=r49; // 3rd arg |
| 126 | sgm0=r50; sgm1=r51; // small constants |
| 127 | A_=r54; B_=r55; C_=r56; D_=r57; |
| 128 | E_=r58; F_=r59; G_=r60; H_=r61; |
| 129 | |
| 130 | // void $func (SHA_CTX *ctx, const void *in,size_t num[,int host]) |
| 131 | .global $func# |
| 132 | .proc $func# |
| 133 | .align 32 |
| 134 | $func: |
| 135 | .prologue |
| 136 | .save ar.pfs,pfssave |
| 137 | { .mmi; alloc pfssave=ar.pfs,3,27,0,16 |
| 138 | $ADDP ctx=0,r32 // 1st arg |
| 139 | .save ar.lc,lcsave |
| 140 | mov lcsave=ar.lc } |
| 141 | { .mmi; $ADDP input=0,r33 // 2nd arg |
| 142 | mov num=r34 // 3rd arg |
| 143 | .save pr,prsave |
| 144 | mov prsave=pr };; |
| 145 | |
| 146 | .body |
| 147 | { .mib; add r8=0*$SZ,ctx |
| 148 | add r9=1*$SZ,ctx |
| 149 | brp.loop.imp .L_first16,.L_first16_end-16 } |
| 150 | { .mib; add r10=2*$SZ,ctx |
| 151 | add r11=3*$SZ,ctx |
| 152 | brp.loop.imp .L_rest,.L_rest_end-16 };; |
| 153 | |
| 154 | // load A-H |
| 155 | .Lpic_point: |
| 156 | { .mmi; $LDW A_=[r8],4*$SZ |
| 157 | $LDW B_=[r9],4*$SZ |
| 158 | mov Ktbl=ip } |
| 159 | { .mmi; $LDW C_=[r10],4*$SZ |
| 160 | $LDW D_=[r11],4*$SZ |
| 161 | mov sgm0=$sigma0[2] };; |
| 162 | { .mmi; $LDW E_=[r8] |
| 163 | $LDW F_=[r9] |
| 164 | add Ktbl=($TABLE#-.Lpic_point),Ktbl } |
| 165 | { .mmi; $LDW G_=[r10] |
| 166 | $LDW H_=[r11] |
| 167 | cmp.ne p0,p16=0,r0 };; // used in sha256_block |
| 168 | ___ |
| 169 | $code.=<<___ if ($BITS==64); |
| 170 | { .mii; and r8=7,input |
| 171 | and input=~7,input;; |
| 172 | cmp.eq p9,p0=1,r8 } |
| 173 | { .mmi; cmp.eq p10,p0=2,r8 |
| 174 | cmp.eq p11,p0=3,r8 |
| 175 | cmp.eq p12,p0=4,r8 } |
| 176 | { .mmi; cmp.eq p13,p0=5,r8 |
| 177 | cmp.eq p14,p0=6,r8 |
| 178 | cmp.eq p15,p0=7,r8 };; |
| 179 | ___ |
| 180 | $code.=<<___; |
| 181 | .L_outer: |
| 182 | .rotr X[16] |
| 183 | { .mmi; mov A=A_ |
| 184 | mov B=B_ |
| 185 | mov ar.lc=14 } |
| 186 | { .mmi; mov C=C_ |
| 187 | mov D=D_ |
| 188 | mov E=E_ } |
| 189 | { .mmi; mov F=F_ |
| 190 | mov G=G_ |
| 191 | mov ar.ec=2 } |
| 192 | { .mmi; ld1 X[15]=[input],$SZ // eliminated in 64-bit |
| 193 | mov H=H_ |
| 194 | mov sgm1=$sigma1[2] };; |
| 195 | |
| 196 | ___ |
| 197 | $t0="t0", $t1="t1", $code.=<<___ if ($BITS==32); |
| 198 | .align 32 |
| 199 | .L_first16: |
| 200 | { .mmi; add r9=1-$SZ,input |
| 201 | add r10=2-$SZ,input |
| 202 | add r11=3-$SZ,input };; |
| 203 | { .mmi; ld1 r9=[r9] |
| 204 | ld1 r10=[r10] |
| 205 | dep.z $t1=E,32,32 } |
| 206 | { .mmi; $LDW K=[Ktbl],$SZ |
| 207 | ld1 r11=[r11] |
| 208 | zxt4 E=E };; |
| 209 | { .mii; or $t1=$t1,E |
| 210 | dep X[15]=X[15],r9,8,8 |
| 211 | dep r11=r10,r11,8,8 };; |
| 212 | { .mmi; and T1=F,E |
| 213 | and T2=A,B |
| 214 | dep X[15]=X[15],r11,16,16 } |
| 215 | { .mmi; andcm r8=G,E |
| 216 | and r9=A,C |
| 217 | mux2 $t0=A,0x44 };; // copy lower half to upper |
| 218 | { .mmi; (p16) ld1 X[15-1]=[input],$SZ // prefetch |
| 219 | xor T1=T1,r8 // T1=((e & f) ^ (~e & g)) |
| 220 | _rotr r11=$t1,$Sigma1[0] } // ROTR(e,14) |
| 221 | { .mib; and r10=B,C |
| 222 | xor T2=T2,r9 };; |
| 223 | ___ |
| 224 | $t0="A", $t1="E", $code.=<<___ if ($BITS==64); |
| 225 | // in 64-bit mode I load whole X[16] at once and take care of alignment... |
| 226 | { .mmi; add r8=1*$SZ,input |
| 227 | add r9=2*$SZ,input |
| 228 | add r10=3*$SZ,input };; |
| 229 | { .mmb; $LDW X[15]=[input],4*$SZ |
| 230 | $LDW X[14]=[r8],4*$SZ |
| 231 | (p9) br.cond.dpnt.many .L1byte };; |
| 232 | { .mmb; $LDW X[13]=[r9],4*$SZ |
| 233 | $LDW X[12]=[r10],4*$SZ |
| 234 | (p10) br.cond.dpnt.many .L2byte };; |
| 235 | { .mmb; $LDW X[11]=[input],4*$SZ |
| 236 | $LDW X[10]=[r8],4*$SZ |
| 237 | (p11) br.cond.dpnt.many .L3byte };; |
| 238 | { .mmb; $LDW X[ 9]=[r9],4*$SZ |
| 239 | $LDW X[ 8]=[r10],4*$SZ |
| 240 | (p12) br.cond.dpnt.many .L4byte };; |
| 241 | { .mmb; $LDW X[ 7]=[input],4*$SZ |
| 242 | $LDW X[ 6]=[r8],4*$SZ |
| 243 | (p13) br.cond.dpnt.many .L5byte };; |
| 244 | { .mmb; $LDW X[ 5]=[r9],4*$SZ |
| 245 | $LDW X[ 4]=[r10],4*$SZ |
| 246 | (p14) br.cond.dpnt.many .L6byte };; |
| 247 | { .mmb; $LDW X[ 3]=[input],4*$SZ |
| 248 | $LDW X[ 2]=[r8],4*$SZ |
| 249 | (p15) br.cond.dpnt.many .L7byte };; |
| 250 | { .mmb; $LDW X[ 1]=[r9],4*$SZ |
| 251 | $LDW X[ 0]=[r10],4*$SZ |
| 252 | br.many .L_first16 };; |
| 253 | .L1byte: |
| 254 | { .mmi; $LDW X[13]=[r9],4*$SZ |
| 255 | $LDW X[12]=[r10],4*$SZ |
| 256 | shrp X[15]=X[15],X[14],56 };; |
| 257 | { .mmi; $LDW X[11]=[input],4*$SZ |
| 258 | $LDW X[10]=[r8],4*$SZ |
| 259 | shrp X[14]=X[14],X[13],56 } |
| 260 | { .mmi; $LDW X[ 9]=[r9],4*$SZ |
| 261 | $LDW X[ 8]=[r10],4*$SZ |
| 262 | shrp X[13]=X[13],X[12],56 };; |
| 263 | { .mmi; $LDW X[ 7]=[input],4*$SZ |
| 264 | $LDW X[ 6]=[r8],4*$SZ |
| 265 | shrp X[12]=X[12],X[11],56 } |
| 266 | { .mmi; $LDW X[ 5]=[r9],4*$SZ |
| 267 | $LDW X[ 4]=[r10],4*$SZ |
| 268 | shrp X[11]=X[11],X[10],56 };; |
| 269 | { .mmi; $LDW X[ 3]=[input],4*$SZ |
| 270 | $LDW X[ 2]=[r8],4*$SZ |
| 271 | shrp X[10]=X[10],X[ 9],56 } |
| 272 | { .mmi; $LDW X[ 1]=[r9],4*$SZ |
| 273 | $LDW X[ 0]=[r10],4*$SZ |
| 274 | shrp X[ 9]=X[ 9],X[ 8],56 };; |
| 275 | { .mii; $LDW T1=[input] |
| 276 | shrp X[ 8]=X[ 8],X[ 7],56 |
| 277 | shrp X[ 7]=X[ 7],X[ 6],56 } |
| 278 | { .mii; shrp X[ 6]=X[ 6],X[ 5],56 |
| 279 | shrp X[ 5]=X[ 5],X[ 4],56 };; |
| 280 | { .mii; shrp X[ 4]=X[ 4],X[ 3],56 |
| 281 | shrp X[ 3]=X[ 3],X[ 2],56 } |
| 282 | { .mii; shrp X[ 2]=X[ 2],X[ 1],56 |
| 283 | shrp X[ 1]=X[ 1],X[ 0],56 } |
| 284 | { .mib; shrp X[ 0]=X[ 0],T1,56 |
| 285 | br.many .L_first16 };; |
| 286 | .L2byte: |
| 287 | { .mmi; $LDW X[11]=[input],4*$SZ |
| 288 | $LDW X[10]=[r8],4*$SZ |
| 289 | shrp X[15]=X[15],X[14],48 } |
| 290 | { .mmi; $LDW X[ 9]=[r9],4*$SZ |
| 291 | $LDW X[ 8]=[r10],4*$SZ |
| 292 | shrp X[14]=X[14],X[13],48 };; |
| 293 | { .mmi; $LDW X[ 7]=[input],4*$SZ |
| 294 | $LDW X[ 6]=[r8],4*$SZ |
| 295 | shrp X[13]=X[13],X[12],48 } |
| 296 | { .mmi; $LDW X[ 5]=[r9],4*$SZ |
| 297 | $LDW X[ 4]=[r10],4*$SZ |
| 298 | shrp X[12]=X[12],X[11],48 };; |
| 299 | { .mmi; $LDW X[ 3]=[input],4*$SZ |
| 300 | $LDW X[ 2]=[r8],4*$SZ |
| 301 | shrp X[11]=X[11],X[10],48 } |
| 302 | { .mmi; $LDW X[ 1]=[r9],4*$SZ |
| 303 | $LDW X[ 0]=[r10],4*$SZ |
| 304 | shrp X[10]=X[10],X[ 9],48 };; |
| 305 | { .mii; $LDW T1=[input] |
| 306 | shrp X[ 9]=X[ 9],X[ 8],48 |
| 307 | shrp X[ 8]=X[ 8],X[ 7],48 } |
| 308 | { .mii; shrp X[ 7]=X[ 7],X[ 6],48 |
| 309 | shrp X[ 6]=X[ 6],X[ 5],48 };; |
| 310 | { .mii; shrp X[ 5]=X[ 5],X[ 4],48 |
| 311 | shrp X[ 4]=X[ 4],X[ 3],48 } |
| 312 | { .mii; shrp X[ 3]=X[ 3],X[ 2],48 |
| 313 | shrp X[ 2]=X[ 2],X[ 1],48 } |
| 314 | { .mii; shrp X[ 1]=X[ 1],X[ 0],48 |
| 315 | shrp X[ 0]=X[ 0],T1,48 } |
| 316 | { .mfb; br.many .L_first16 };; |
| 317 | .L3byte: |
| 318 | { .mmi; $LDW X[ 9]=[r9],4*$SZ |
| 319 | $LDW X[ 8]=[r10],4*$SZ |
| 320 | shrp X[15]=X[15],X[14],40 };; |
| 321 | { .mmi; $LDW X[ 7]=[input],4*$SZ |
| 322 | $LDW X[ 6]=[r8],4*$SZ |
| 323 | shrp X[14]=X[14],X[13],40 } |
| 324 | { .mmi; $LDW X[ 5]=[r9],4*$SZ |
| 325 | $LDW X[ 4]=[r10],4*$SZ |
| 326 | shrp X[13]=X[13],X[12],40 };; |
| 327 | { .mmi; $LDW X[ 3]=[input],4*$SZ |
| 328 | $LDW X[ 2]=[r8],4*$SZ |
| 329 | shrp X[12]=X[12],X[11],40 } |
| 330 | { .mmi; $LDW X[ 1]=[r9],4*$SZ |
| 331 | $LDW X[ 0]=[r10],4*$SZ |
| 332 | shrp X[11]=X[11],X[10],40 };; |
| 333 | { .mii; $LDW T1=[input] |
| 334 | shrp X[10]=X[10],X[ 9],40 |
| 335 | shrp X[ 9]=X[ 9],X[ 8],40 } |
| 336 | { .mii; shrp X[ 8]=X[ 8],X[ 7],40 |
| 337 | shrp X[ 7]=X[ 7],X[ 6],40 };; |
| 338 | { .mii; shrp X[ 6]=X[ 6],X[ 5],40 |
| 339 | shrp X[ 5]=X[ 5],X[ 4],40 } |
| 340 | { .mii; shrp X[ 4]=X[ 4],X[ 3],40 |
| 341 | shrp X[ 3]=X[ 3],X[ 2],40 } |
| 342 | { .mii; shrp X[ 2]=X[ 2],X[ 1],40 |
| 343 | shrp X[ 1]=X[ 1],X[ 0],40 } |
| 344 | { .mib; shrp X[ 0]=X[ 0],T1,40 |
| 345 | br.many .L_first16 };; |
| 346 | .L4byte: |
| 347 | { .mmi; $LDW X[ 7]=[input],4*$SZ |
| 348 | $LDW X[ 6]=[r8],4*$SZ |
| 349 | shrp X[15]=X[15],X[14],32 } |
| 350 | { .mmi; $LDW X[ 5]=[r9],4*$SZ |
| 351 | $LDW X[ 4]=[r10],4*$SZ |
| 352 | shrp X[14]=X[14],X[13],32 };; |
| 353 | { .mmi; $LDW X[ 3]=[input],4*$SZ |
| 354 | $LDW X[ 2]=[r8],4*$SZ |
| 355 | shrp X[13]=X[13],X[12],32 } |
| 356 | { .mmi; $LDW X[ 1]=[r9],4*$SZ |
| 357 | $LDW X[ 0]=[r10],4*$SZ |
| 358 | shrp X[12]=X[12],X[11],32 };; |
| 359 | { .mii; $LDW T1=[input] |
| 360 | shrp X[11]=X[11],X[10],32 |
| 361 | shrp X[10]=X[10],X[ 9],32 } |
| 362 | { .mii; shrp X[ 9]=X[ 9],X[ 8],32 |
| 363 | shrp X[ 8]=X[ 8],X[ 7],32 };; |
| 364 | { .mii; shrp X[ 7]=X[ 7],X[ 6],32 |
| 365 | shrp X[ 6]=X[ 6],X[ 5],32 } |
| 366 | { .mii; shrp X[ 5]=X[ 5],X[ 4],32 |
| 367 | shrp X[ 4]=X[ 4],X[ 3],32 } |
| 368 | { .mii; shrp X[ 3]=X[ 3],X[ 2],32 |
| 369 | shrp X[ 2]=X[ 2],X[ 1],32 } |
| 370 | { .mii; shrp X[ 1]=X[ 1],X[ 0],32 |
| 371 | shrp X[ 0]=X[ 0],T1,32 } |
| 372 | { .mfb; br.many .L_first16 };; |
| 373 | .L5byte: |
| 374 | { .mmi; $LDW X[ 5]=[r9],4*$SZ |
| 375 | $LDW X[ 4]=[r10],4*$SZ |
| 376 | shrp X[15]=X[15],X[14],24 };; |
| 377 | { .mmi; $LDW X[ 3]=[input],4*$SZ |
| 378 | $LDW X[ 2]=[r8],4*$SZ |
| 379 | shrp X[14]=X[14],X[13],24 } |
| 380 | { .mmi; $LDW X[ 1]=[r9],4*$SZ |
| 381 | $LDW X[ 0]=[r10],4*$SZ |
| 382 | shrp X[13]=X[13],X[12],24 };; |
| 383 | { .mii; $LDW T1=[input] |
| 384 | shrp X[12]=X[12],X[11],24 |
| 385 | shrp X[11]=X[11],X[10],24 } |
| 386 | { .mii; shrp X[10]=X[10],X[ 9],24 |
| 387 | shrp X[ 9]=X[ 9],X[ 8],24 };; |
| 388 | { .mii; shrp X[ 8]=X[ 8],X[ 7],24 |
| 389 | shrp X[ 7]=X[ 7],X[ 6],24 } |
| 390 | { .mii; shrp X[ 6]=X[ 6],X[ 5],24 |
| 391 | shrp X[ 5]=X[ 5],X[ 4],24 } |
| 392 | { .mii; shrp X[ 4]=X[ 4],X[ 3],24 |
| 393 | shrp X[ 3]=X[ 3],X[ 2],24 } |
| 394 | { .mii; shrp X[ 2]=X[ 2],X[ 1],24 |
| 395 | shrp X[ 1]=X[ 1],X[ 0],24 } |
| 396 | { .mib; shrp X[ 0]=X[ 0],T1,24 |
| 397 | br.many .L_first16 };; |
| 398 | .L6byte: |
| 399 | { .mmi; $LDW X[ 3]=[input],4*$SZ |
| 400 | $LDW X[ 2]=[r8],4*$SZ |
| 401 | shrp X[15]=X[15],X[14],16 } |
| 402 | { .mmi; $LDW X[ 1]=[r9],4*$SZ |
| 403 | $LDW X[ 0]=[r10],4*$SZ |
| 404 | shrp X[14]=X[14],X[13],16 };; |
| 405 | { .mii; $LDW T1=[input] |
| 406 | shrp X[13]=X[13],X[12],16 |
| 407 | shrp X[12]=X[12],X[11],16 } |
| 408 | { .mii; shrp X[11]=X[11],X[10],16 |
| 409 | shrp X[10]=X[10],X[ 9],16 };; |
| 410 | { .mii; shrp X[ 9]=X[ 9],X[ 8],16 |
| 411 | shrp X[ 8]=X[ 8],X[ 7],16 } |
| 412 | { .mii; shrp X[ 7]=X[ 7],X[ 6],16 |
| 413 | shrp X[ 6]=X[ 6],X[ 5],16 } |
| 414 | { .mii; shrp X[ 5]=X[ 5],X[ 4],16 |
| 415 | shrp X[ 4]=X[ 4],X[ 3],16 } |
| 416 | { .mii; shrp X[ 3]=X[ 3],X[ 2],16 |
| 417 | shrp X[ 2]=X[ 2],X[ 1],16 } |
| 418 | { .mii; shrp X[ 1]=X[ 1],X[ 0],16 |
| 419 | shrp X[ 0]=X[ 0],T1,16 } |
| 420 | { .mfb; br.many .L_first16 };; |
| 421 | .L7byte: |
| 422 | { .mmi; $LDW X[ 1]=[r9],4*$SZ |
| 423 | $LDW X[ 0]=[r10],4*$SZ |
| 424 | shrp X[15]=X[15],X[14],8 };; |
| 425 | { .mii; $LDW T1=[input] |
| 426 | shrp X[14]=X[14],X[13],8 |
| 427 | shrp X[13]=X[13],X[12],8 } |
| 428 | { .mii; shrp X[12]=X[12],X[11],8 |
| 429 | shrp X[11]=X[11],X[10],8 };; |
| 430 | { .mii; shrp X[10]=X[10],X[ 9],8 |
| 431 | shrp X[ 9]=X[ 9],X[ 8],8 } |
| 432 | { .mii; shrp X[ 8]=X[ 8],X[ 7],8 |
| 433 | shrp X[ 7]=X[ 7],X[ 6],8 } |
| 434 | { .mii; shrp X[ 6]=X[ 6],X[ 5],8 |
| 435 | shrp X[ 5]=X[ 5],X[ 4],8 } |
| 436 | { .mii; shrp X[ 4]=X[ 4],X[ 3],8 |
| 437 | shrp X[ 3]=X[ 3],X[ 2],8 } |
| 438 | { .mii; shrp X[ 2]=X[ 2],X[ 1],8 |
| 439 | shrp X[ 1]=X[ 1],X[ 0],8 } |
| 440 | { .mib; shrp X[ 0]=X[ 0],T1,8 |
| 441 | br.many .L_first16 };; |
| 442 | |
| 443 | .align 32 |
| 444 | .L_first16: |
| 445 | { .mmi; $LDW K=[Ktbl],$SZ |
| 446 | and T1=F,E |
| 447 | and T2=A,B } |
| 448 | { .mmi; //$LDW X[15]=[input],$SZ // X[i]=*input++ |
| 449 | andcm r8=G,E |
| 450 | and r9=A,C };; |
| 451 | { .mmi; xor T1=T1,r8 //T1=((e & f) ^ (~e & g)) |
| 452 | and r10=B,C |
| 453 | _rotr r11=$t1,$Sigma1[0] } // ROTR(e,14) |
| 454 | { .mmi; xor T2=T2,r9 |
| 455 | mux1 X[15]=X[15],\@rev };; // eliminated in big-endian |
| 456 | ___ |
| 457 | $code.=<<___; |
| 458 | { .mib; add T1=T1,H // T1=Ch(e,f,g)+h |
| 459 | _rotr r8=$t1,$Sigma1[1] } // ROTR(e,18) |
| 460 | { .mib; xor T2=T2,r10 // T2=((a & b) ^ (a & c) ^ (b & c)) |
| 461 | mov H=G };; |
| 462 | { .mib; xor r11=r8,r11 |
| 463 | _rotr r9=$t1,$Sigma1[2] } // ROTR(e,41) |
| 464 | { .mib; mov G=F |
| 465 | mov F=E };; |
| 466 | { .mib; xor r9=r9,r11 // r9=Sigma1(e) |
| 467 | _rotr r10=$t0,$Sigma0[0] } // ROTR(a,28) |
| 468 | { .mib; add T1=T1,K // T1=Ch(e,f,g)+h+K512[i] |
| 469 | mov E=D };; |
| 470 | { .mib; add T1=T1,r9 // T1+=Sigma1(e) |
| 471 | _rotr r11=$t0,$Sigma0[1] } // ROTR(a,34) |
| 472 | { .mib; mov D=C |
| 473 | mov C=B };; |
| 474 | { .mib; add T1=T1,X[15] // T1+=X[i] |
| 475 | _rotr r8=$t0,$Sigma0[2] } // ROTR(a,39) |
| 476 | { .mib; xor r10=r10,r11 |
| 477 | mux2 X[15]=X[15],0x44 };; // eliminated in 64-bit |
| 478 | { .mmi; xor r10=r8,r10 // r10=Sigma0(a) |
| 479 | mov B=A |
| 480 | add A=T1,T2 };; |
| 481 | { .mib; add E=E,T1 |
| 482 | add A=A,r10 // T2=Maj(a,b,c)+Sigma0(a) |
| 483 | br.ctop.sptk .L_first16 };; |
| 484 | .L_first16_end: |
| 485 | |
| 486 | { .mii; mov ar.lc=$rounds-17 |
| 487 | mov ar.ec=1 };; |
| 488 | |
| 489 | .align 32 |
| 490 | .L_rest: |
| 491 | .rotr X[16] |
| 492 | { .mib; $LDW K=[Ktbl],$SZ |
| 493 | _rotr r8=X[15-1],$sigma0[0] } // ROTR(s0,1) |
| 494 | { .mib; $ADD X[15]=X[15],X[15-9] // X[i&0xF]+=X[(i+9)&0xF] |
| 495 | $SHRU s0=X[15-1],sgm0 };; // s0=X[(i+1)&0xF]>>7 |
| 496 | { .mib; and T1=F,E |
| 497 | _rotr r9=X[15-1],$sigma0[1] } // ROTR(s0,8) |
| 498 | { .mib; andcm r10=G,E |
| 499 | $SHRU s1=X[15-14],sgm1 };; // s1=X[(i+14)&0xF]>>6 |
| 500 | { .mmi; xor T1=T1,r10 // T1=((e & f) ^ (~e & g)) |
| 501 | xor r9=r8,r9 |
| 502 | _rotr r10=X[15-14],$sigma1[0] };;// ROTR(s1,19) |
| 503 | { .mib; and T2=A,B |
| 504 | _rotr r11=X[15-14],$sigma1[1] }// ROTR(s1,61) |
| 505 | { .mib; and r8=A,C };; |
| 506 | ___ |
| 507 | $t0="t0", $t1="t1", $code.=<<___ if ($BITS==32); |
| 508 | // I adhere to mmi; in order to hold Itanium 1 back and avoid 6 cycle |
| 509 | // pipeline flush in last bundle. Note that even on Itanium2 the |
| 510 | // latter stalls for one clock cycle... |
| 511 | { .mmi; xor s0=s0,r9 // s0=sigma0(X[(i+1)&0xF]) |
| 512 | dep.z $t1=E,32,32 } |
| 513 | { .mmi; xor r10=r11,r10 |
| 514 | zxt4 E=E };; |
| 515 | { .mmi; or $t1=$t1,E |
| 516 | xor s1=s1,r10 // s1=sigma1(X[(i+14)&0xF]) |
| 517 | mux2 $t0=A,0x44 };; // copy lower half to upper |
| 518 | { .mmi; xor T2=T2,r8 |
| 519 | _rotr r9=$t1,$Sigma1[0] } // ROTR(e,14) |
| 520 | { .mmi; and r10=B,C |
| 521 | add T1=T1,H // T1=Ch(e,f,g)+h |
| 522 | $ADD X[15]=X[15],s0 };; // X[i&0xF]+=sigma0(X[(i+1)&0xF]) |
| 523 | ___ |
| 524 | $t0="A", $t1="E", $code.=<<___ if ($BITS==64); |
| 525 | { .mib; xor s0=s0,r9 // s0=sigma0(X[(i+1)&0xF]) |
| 526 | _rotr r9=$t1,$Sigma1[0] } // ROTR(e,14) |
| 527 | { .mib; xor r10=r11,r10 |
| 528 | xor T2=T2,r8 };; |
| 529 | { .mib; xor s1=s1,r10 // s1=sigma1(X[(i+14)&0xF]) |
| 530 | add T1=T1,H } |
| 531 | { .mib; and r10=B,C |
| 532 | $ADD X[15]=X[15],s0 };; // X[i&0xF]+=sigma0(X[(i+1)&0xF]) |
| 533 | ___ |
| 534 | $code.=<<___; |
| 535 | { .mmi; xor T2=T2,r10 // T2=((a & b) ^ (a & c) ^ (b & c)) |
| 536 | mov H=G |
| 537 | _rotr r8=$t1,$Sigma1[1] };; // ROTR(e,18) |
| 538 | { .mmi; xor r11=r8,r9 |
| 539 | $ADD X[15]=X[15],s1 // X[i&0xF]+=sigma1(X[(i+14)&0xF]) |
| 540 | _rotr r9=$t1,$Sigma1[2] } // ROTR(e,41) |
| 541 | { .mmi; mov G=F |
| 542 | mov F=E };; |
| 543 | { .mib; xor r9=r9,r11 // r9=Sigma1(e) |
| 544 | _rotr r10=$t0,$Sigma0[0] } // ROTR(a,28) |
| 545 | { .mib; add T1=T1,K // T1=Ch(e,f,g)+h+K512[i] |
| 546 | mov E=D };; |
| 547 | { .mib; add T1=T1,r9 // T1+=Sigma1(e) |
| 548 | _rotr r11=$t0,$Sigma0[1] } // ROTR(a,34) |
| 549 | { .mib; mov D=C |
| 550 | mov C=B };; |
| 551 | { .mmi; add T1=T1,X[15] // T1+=X[i] |
| 552 | xor r10=r10,r11 |
| 553 | _rotr r8=$t0,$Sigma0[2] };; // ROTR(a,39) |
| 554 | { .mmi; xor r10=r8,r10 // r10=Sigma0(a) |
| 555 | mov B=A |
| 556 | add A=T1,T2 };; |
| 557 | { .mib; add E=E,T1 |
| 558 | add A=A,r10 // T2=Maj(a,b,c)+Sigma0(a) |
| 559 | br.ctop.sptk .L_rest };; |
| 560 | .L_rest_end: |
| 561 | |
| 562 | { .mmi; add A_=A_,A |
| 563 | add B_=B_,B |
| 564 | add C_=C_,C } |
| 565 | { .mmi; add D_=D_,D |
| 566 | add E_=E_,E |
| 567 | cmp.ltu p16,p0=1,num };; |
| 568 | { .mmi; add F_=F_,F |
| 569 | add G_=G_,G |
| 570 | add H_=H_,H } |
| 571 | { .mmb; add Ktbl=-$SZ*$rounds,Ktbl |
| 572 | (p16) add num=-1,num |
| 573 | (p16) br.dptk.many .L_outer };; |
| 574 | |
| 575 | { .mib; add r8=0*$SZ,ctx |
| 576 | add r9=1*$SZ,ctx } |
| 577 | { .mib; add r10=2*$SZ,ctx |
| 578 | add r11=3*$SZ,ctx };; |
| 579 | { .mmi; $STW [r8]=A_,4*$SZ |
| 580 | $STW [r9]=B_,4*$SZ |
| 581 | mov ar.lc=lcsave } |
| 582 | { .mmi; $STW [r10]=C_,4*$SZ |
| 583 | $STW [r11]=D_,4*$SZ |
| 584 | mov pr=prsave,0x1ffff };; |
| 585 | { .mmb; $STW [r8]=E_ |
| 586 | $STW [r9]=F_ } |
| 587 | { .mmb; $STW [r10]=G_ |
| 588 | $STW [r11]=H_ |
| 589 | br.ret.sptk.many b0 };; |
| 590 | .endp $func# |
| 591 | ___ |
| 592 | |
| 593 | $code =~ s/\`([^\`]*)\`/eval $1/gem; |
| 594 | $code =~ s/_rotr(\s+)([^=]+)=([^,]+),([0-9]+)/shrp$1$2=$3,$3,$4/gm; |
| 595 | if ($BITS==64) { |
| 596 | $code =~ s/mux2(\s+)\S+/nop.i$1 0x0/gm; |
| 597 | $code =~ s/mux1(\s+)\S+/nop.i$1 0x0/gm if ($big_endian); |
| 598 | $code =~ s/(shrp\s+X\[[^=]+)=([^,]+),([^,]+),([1-9]+)/$1=$3,$2,64-$4/gm |
| 599 | if (!$big_endian); |
| 600 | $code =~ s/ld1(\s+)X\[\S+/nop.m$1 0x0/gm; |
| 601 | } |
| 602 | |
| 603 | print $code; |
| 604 | |
| 605 | print<<___ if ($BITS==32); |
| 606 | .align 64 |
| 607 | .type K256#,\@object |
| 608 | K256: data4 0x428a2f98,0x71374491,0xb5c0fbcf,0xe9b5dba5 |
| 609 | data4 0x3956c25b,0x59f111f1,0x923f82a4,0xab1c5ed5 |
| 610 | data4 0xd807aa98,0x12835b01,0x243185be,0x550c7dc3 |
| 611 | data4 0x72be5d74,0x80deb1fe,0x9bdc06a7,0xc19bf174 |
| 612 | data4 0xe49b69c1,0xefbe4786,0x0fc19dc6,0x240ca1cc |
| 613 | data4 0x2de92c6f,0x4a7484aa,0x5cb0a9dc,0x76f988da |
| 614 | data4 0x983e5152,0xa831c66d,0xb00327c8,0xbf597fc7 |
| 615 | data4 0xc6e00bf3,0xd5a79147,0x06ca6351,0x14292967 |
| 616 | data4 0x27b70a85,0x2e1b2138,0x4d2c6dfc,0x53380d13 |
| 617 | data4 0x650a7354,0x766a0abb,0x81c2c92e,0x92722c85 |
| 618 | data4 0xa2bfe8a1,0xa81a664b,0xc24b8b70,0xc76c51a3 |
| 619 | data4 0xd192e819,0xd6990624,0xf40e3585,0x106aa070 |
| 620 | data4 0x19a4c116,0x1e376c08,0x2748774c,0x34b0bcb5 |
| 621 | data4 0x391c0cb3,0x4ed8aa4a,0x5b9cca4f,0x682e6ff3 |
| 622 | data4 0x748f82ee,0x78a5636f,0x84c87814,0x8cc70208 |
| 623 | data4 0x90befffa,0xa4506ceb,0xbef9a3f7,0xc67178f2 |
| 624 | .size K256#,$SZ*$rounds |
| 625 | stringz "SHA256 block transform for IA64, CRYPTOGAMS by <appro\@openssl.org>" |
| 626 | ___ |
| 627 | print<<___ if ($BITS==64); |
| 628 | .align 64 |
| 629 | .type K512#,\@object |
| 630 | K512: data8 0x428a2f98d728ae22,0x7137449123ef65cd |
| 631 | data8 0xb5c0fbcfec4d3b2f,0xe9b5dba58189dbbc |
| 632 | data8 0x3956c25bf348b538,0x59f111f1b605d019 |
| 633 | data8 0x923f82a4af194f9b,0xab1c5ed5da6d8118 |
| 634 | data8 0xd807aa98a3030242,0x12835b0145706fbe |
| 635 | data8 0x243185be4ee4b28c,0x550c7dc3d5ffb4e2 |
| 636 | data8 0x72be5d74f27b896f,0x80deb1fe3b1696b1 |
| 637 | data8 0x9bdc06a725c71235,0xc19bf174cf692694 |
| 638 | data8 0xe49b69c19ef14ad2,0xefbe4786384f25e3 |
| 639 | data8 0x0fc19dc68b8cd5b5,0x240ca1cc77ac9c65 |
| 640 | data8 0x2de92c6f592b0275,0x4a7484aa6ea6e483 |
| 641 | data8 0x5cb0a9dcbd41fbd4,0x76f988da831153b5 |
| 642 | data8 0x983e5152ee66dfab,0xa831c66d2db43210 |
| 643 | data8 0xb00327c898fb213f,0xbf597fc7beef0ee4 |
| 644 | data8 0xc6e00bf33da88fc2,0xd5a79147930aa725 |
| 645 | data8 0x06ca6351e003826f,0x142929670a0e6e70 |
| 646 | data8 0x27b70a8546d22ffc,0x2e1b21385c26c926 |
| 647 | data8 0x4d2c6dfc5ac42aed,0x53380d139d95b3df |
| 648 | data8 0x650a73548baf63de,0x766a0abb3c77b2a8 |
| 649 | data8 0x81c2c92e47edaee6,0x92722c851482353b |
| 650 | data8 0xa2bfe8a14cf10364,0xa81a664bbc423001 |
| 651 | data8 0xc24b8b70d0f89791,0xc76c51a30654be30 |
| 652 | data8 0xd192e819d6ef5218,0xd69906245565a910 |
| 653 | data8 0xf40e35855771202a,0x106aa07032bbd1b8 |
| 654 | data8 0x19a4c116b8d2d0c8,0x1e376c085141ab53 |
| 655 | data8 0x2748774cdf8eeb99,0x34b0bcb5e19b48a8 |
| 656 | data8 0x391c0cb3c5c95a63,0x4ed8aa4ae3418acb |
| 657 | data8 0x5b9cca4f7763e373,0x682e6ff3d6b2b8a3 |
| 658 | data8 0x748f82ee5defb2fc,0x78a5636f43172f60 |
| 659 | data8 0x84c87814a1f0ab72,0x8cc702081a6439ec |
| 660 | data8 0x90befffa23631e28,0xa4506cebde82bde9 |
| 661 | data8 0xbef9a3f7b2c67915,0xc67178f2e372532b |
| 662 | data8 0xca273eceea26619c,0xd186b8c721c0c207 |
| 663 | data8 0xeada7dd6cde0eb1e,0xf57d4f7fee6ed178 |
| 664 | data8 0x06f067aa72176fba,0x0a637dc5a2c898a6 |
| 665 | data8 0x113f9804bef90dae,0x1b710b35131c471b |
| 666 | data8 0x28db77f523047d84,0x32caab7b40c72493 |
| 667 | data8 0x3c9ebe0a15c9bebc,0x431d67c49c100d4c |
| 668 | data8 0x4cc5d4becb3e42b6,0x597f299cfc657e2a |
| 669 | data8 0x5fcb6fab3ad6faec,0x6c44198c4a475817 |
| 670 | .size K512#,$SZ*$rounds |
| 671 | stringz "SHA512 block transform for IA64, CRYPTOGAMS by <appro\@openssl.org>" |
| 672 | ___ |