crypto/sha/asm/sha512-c64xplus.pl

   1 #!/usr/bin/env perl
   2 #
   3 # ====================================================================
   4 # Written by Andy Polyakov <appro@openssl.org> 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 # SHA512 for C64x+.
  11 #
  12 # January 2012
  13 #
  14 # Performance is 19 cycles per processed byte. Compared to block
  15 # transform function from sha512.c compiled with cl6x with -mv6400+
  16 # -o2 -DOPENSSL_SMALL_FOOTPRINT it's almost 7x faster and 2x smaller.
  17 # Loop unroll won't make it, this implementation, any faster, because
  18 # it's effectively dominated by SHRU||SHL pairs and you can't schedule
  19 # more of them.
  20 #
  21 # !!! Note that this module uses AMR, which means that all interrupt
  22 # service routines are expected to preserve it and for own well-being
  23 # zero it upon entry.
  24
  25 while (($output=shift) && ($output!~/\w[\w\-]*\.\w+$/)) {}
  26 open STDOUT,">$output";
  27
  28 ($CTXA,$INP,$NUM) = ("A4","B4","A6");            # arguments
  29  $K512="A3";
  30
  31 ($Ahi,$Actxhi,$Bhi,$Bctxhi,$Chi,$Cctxhi,$Dhi,$Dctxhi,
  32  $Ehi,$Ectxhi,$Fhi,$Fctxhi,$Ghi,$Gctxhi,$Hhi,$Hctxhi)=map("A$_",(16..31));
  33 ($Alo,$Actxlo,$Blo,$Bctxlo,$Clo,$Cctxlo,$Dlo,$Dctxlo,
  34  $Elo,$Ectxlo,$Flo,$Fctxlo,$Glo,$Gctxlo,$Hlo,$Hctxlo)=map("B$_",(16..31));
  35
  36 ($S1hi,$CHhi,$S0hi,$t0hi)=map("A$_",(10..13));
  37 ($S1lo,$CHlo,$S0lo,$t0lo)=map("B$_",(10..13));
  38 ($T1hi,         $T2hi)=         ("A6","A7");
  39 ($T1lo,$T1carry,$T2lo,$T2carry)=("B6","B7","B8","B9");
  40 ($Khi,$Klo)=("A9","A8");
  41 ($MAJhi,$MAJlo)=($T2hi,$T2lo);
  42 ($t1hi,$t1lo)=($Khi,"B2");
  43  $CTXB=$t1lo;
  44
  45 ($Xihi,$Xilo)=("A5","B5");                      # circular/ring buffer
  46
  47 $code.=<<___;
  48         .text
  49         .if     __TI_EABI__
  50         .nocmp
  51         .asg    sha512_block_data_order,_sha512_block_data_order
  52         .endif
  53
  54         .asg    B3,RA
  55         .asg    A15,FP
  56         .asg    B15,SP
  57
  58         .if     .BIG_ENDIAN
  59         .asg    $Khi,KHI
  60         .asg    $Klo,KLO
  61         .else
  62         .asg    $Khi,KLO
  63         .asg    $Klo,KHI
  64         .endif
  65
  66         .global _sha512_block_data_order
  67 _sha512_block_data_order:
  68         .asmfunc stack_usage(40+128)
  69         MV      $NUM,A0                         ; reassign $NUM
  70 ||      MVK     -128,B0
  71   [!A0] BNOP    RA                              ; if ($NUM==0) return;
  72 || [A0] STW     FP,*SP--(40)                    ; save frame pointer
  73 || [A0] MV      SP,FP
  74    [A0] STDW    B13:B12,*SP[4]
  75 || [A0] MVK     0x00404,B1
  76    [A0] STDW    B11:B10,*SP[3]
  77 || [A0] STDW    A13:A12,*FP[-3]
  78 || [A0] MVKH    0x60000,B1
  79    [A0] STDW    A11:A10,*SP[1]
  80 || [A0] MVC     B1,AMR                          ; setup circular addressing
  81 || [A0] ADD     B0,SP,SP                        ; alloca(128)
  82         .if     __TI_EABI__
  83    [A0] AND     B0,SP,SP                        ; align stack at 128 bytes
  84 || [A0] ADDKPC  _sha512_block_data_order,B1
  85 || [A0] MVKL    \$PCR_OFFSET(K512,_sha512_block_data_order),$K512
  86    [A0] MVKH    \$PCR_OFFSET(K512,_sha512_block_data_order),$K512
  87 || [A0] SUBAW   SP,2,SP                         ; reserve two words above buffer
  88         .else
  89    [A0] AND     B0,SP,SP                        ; align stack at 128 bytes
  90 || [A0] ADDKPC  _sha512_block_data_order,B1
  91 || [A0] MVKL    (K512-_sha512_block_data_order),$K512
  92    [A0] MVKH    (K512-_sha512_block_data_order),$K512
  93 || [A0] SUBAW   SP,2,SP                         ; reserve two words above buffer
  94         .endif
  95         ADDAW   SP,3,$Xilo
  96         ADDAW   SP,2,$Xihi
  97
  98 ||      MV      $CTXA,$CTXB
  99         LDW     *${CTXA}[0^.LITTLE_ENDIAN],$Ahi ; load ctx
 100 ||      LDW     *${CTXB}[1^.LITTLE_ENDIAN],$Alo
 101 ||      ADD     B1,$K512,$K512
 102         LDW     *${CTXA}[2^.LITTLE_ENDIAN],$Bhi
 103 ||      LDW     *${CTXB}[3^.LITTLE_ENDIAN],$Blo
 104         LDW     *${CTXA}[4^.LITTLE_ENDIAN],$Chi
 105 ||      LDW     *${CTXB}[5^.LITTLE_ENDIAN],$Clo
 106         LDW     *${CTXA}[6^.LITTLE_ENDIAN],$Dhi
 107 ||      LDW     *${CTXB}[7^.LITTLE_ENDIAN],$Dlo
 108         LDW     *${CTXA}[8^.LITTLE_ENDIAN],$Ehi
 109 ||      LDW     *${CTXB}[9^.LITTLE_ENDIAN],$Elo
 110         LDW     *${CTXA}[10^.LITTLE_ENDIAN],$Fhi
 111 ||      LDW     *${CTXB}[11^.LITTLE_ENDIAN],$Flo
 112         LDW     *${CTXA}[12^.LITTLE_ENDIAN],$Ghi
 113 ||      LDW     *${CTXB}[13^.LITTLE_ENDIAN],$Glo
 114         LDW     *${CTXA}[14^.LITTLE_ENDIAN],$Hhi
 115 ||      LDW     *${CTXB}[15^.LITTLE_ENDIAN],$Hlo
 116
 117         LDNDW   *$INP++,B11:B10                 ; pre-fetch input
 118         LDDW    *$K512++,$Khi:$Klo              ; pre-fetch K512[0]
 119 outerloop?:
 120         MVK     15,B0                           ; loop counters
 121 ||      MVK     64,B1
 122 ||      SUB     A0,1,A0
 123         MV      $Ahi,$Actxhi
 124 ||      MV      $Alo,$Actxlo
 125 ||      MV      $Bhi,$Bctxhi
 126 ||      MV      $Blo,$Bctxlo
 127 ||      MV      $Chi,$Cctxhi
 128 ||      MV      $Clo,$Cctxlo
 129 ||      MVD     $Dhi,$Dctxhi
 130 ||      MVD     $Dlo,$Dctxlo
 131         MV      $Ehi,$Ectxhi
 132 ||      MV      $Elo,$Ectxlo
 133 ||      MV      $Fhi,$Fctxhi
 134 ||      MV      $Flo,$Fctxlo
 135 ||      MV      $Ghi,$Gctxhi
 136 ||      MV      $Glo,$Gctxlo
 137 ||      MVD     $Hhi,$Hctxhi
 138 ||      MVD     $Hlo,$Hctxlo
 139 loop0_15?:
 140         .if     .BIG_ENDIAN
 141         MV      B11,$T1hi
 142 ||      MV      B10,$T1lo
 143         .else
 144         SWAP4   B10,$T1hi
 145 ||      SWAP4   B11,$T1lo
 146         SWAP2   $T1hi,$T1hi
 147 ||      SWAP2   $T1lo,$T1lo
 148         .endif
 149 loop16_79?:
 150         STW     $T1hi,*$Xihi++[2]
 151 ||      STW     $T1lo,*$Xilo++[2]                       ; X[i] = T1
 152 ||      ADD     $Hhi,$T1hi,$T1hi
 153 ||      ADDU    $Hlo,$T1lo,$T1carry:$T1lo               ; T1 += h
 154 ||      SHRU    $Ehi,14,$S1hi
 155 ||      SHL     $Ehi,32-14,$S1lo
 156         XOR     $Fhi,$Ghi,$CHhi
 157 ||      XOR     $Flo,$Glo,$CHlo
 158 ||      ADD     KHI,$T1hi,$T1hi
 159 ||      ADDU    KLO,$T1carry:$T1lo,$T1carry:$T1lo       ; T1 += K512[i]
 160 ||      SHRU    $Elo,14,$t0lo
 161 ||      SHL     $Elo,32-14,$t0hi
 162         XOR     $t0hi,$S1hi,$S1hi
 163 ||      XOR     $t0lo,$S1lo,$S1lo
 164 ||      AND     $Ehi,$CHhi,$CHhi
 165 ||      AND     $Elo,$CHlo,$CHlo
 166 ||      ROTL    $Ghi,0,$Hhi
 167 ||      ROTL    $Glo,0,$Hlo                             ; h = g
 168 ||      SHRU    $Ehi,18,$t0hi
 169 ||      SHL     $Ehi,32-18,$t0lo
 170         XOR     $t0hi,$S1hi,$S1hi
 171 ||      XOR     $t0lo,$S1lo,$S1lo
 172 ||      XOR     $Ghi,$CHhi,$CHhi
 173 ||      XOR     $Glo,$CHlo,$CHlo                        ; Ch(e,f,g) = ((f^g)&e)^g
 174 ||      ROTL    $Fhi,0,$Ghi
 175 ||      ROTL    $Flo,0,$Glo                             ; g = f
 176 ||      SHRU    $Elo,18,$t0lo
 177 ||      SHL     $Elo,32-18,$t0hi
 178         XOR     $t0hi,$S1hi,$S1hi
 179 ||      XOR     $t0lo,$S1lo,$S1lo
 180 ||      OR      $Ahi,$Bhi,$MAJhi
 181 ||      OR      $Alo,$Blo,$MAJlo
 182 ||      ROTL    $Ehi,0,$Fhi
 183 ||      ROTL    $Elo,0,$Flo                             ; f = e
 184 ||      SHRU    $Ehi,41-32,$t0lo
 185 ||      SHL     $Ehi,64-41,$t0hi
 186         XOR     $t0hi,$S1hi,$S1hi
 187 ||      XOR     $t0lo,$S1lo,$S1lo
 188 ||      AND     $Chi,$MAJhi,$MAJhi
 189 ||      AND     $Clo,$MAJlo,$MAJlo
 190 ||      ROTL    $Dhi,0,$Ehi
 191 ||      ROTL    $Dlo,0,$Elo                             ; e = d
 192 ||      SHRU    $Elo,41-32,$t0hi
 193 ||      SHL     $Elo,64-41,$t0lo
 194         XOR     $t0hi,$S1hi,$S1hi
 195 ||      XOR     $t0lo,$S1lo,$S1lo                       ; Sigma1(e)
 196 ||      AND     $Ahi,$Bhi,$t1hi
 197 ||      AND     $Alo,$Blo,$t1lo
 198 ||      ROTL    $Chi,0,$Dhi
 199 ||      ROTL    $Clo,0,$Dlo                             ; d = c
 200 ||      SHRU    $Ahi,28,$S0hi
 201 ||      SHL     $Ahi,32-28,$S0lo
 202         OR      $t1hi,$MAJhi,$MAJhi
 203 ||      OR      $t1lo,$MAJlo,$MAJlo                     ; Maj(a,b,c) = ((a|b)&c)|(a&b)
 204 ||      ADD     $CHhi,$T1hi,$T1hi
 205 ||      ADDU    $CHlo,$T1carry:$T1lo,$T1carry:$T1lo     ; T1 += Ch(e,f,g)
 206 ||      ROTL    $Bhi,0,$Chi
 207 ||      ROTL    $Blo,0,$Clo                             ; c = b
 208 ||      SHRU    $Alo,28,$t0lo
 209 ||      SHL     $Alo,32-28,$t0hi
 210         XOR     $t0hi,$S0hi,$S0hi
 211 ||      XOR     $t0lo,$S0lo,$S0lo
 212 ||      ADD     $S1hi,$T1hi,$T1hi
 213 ||      ADDU    $S1lo,$T1carry:$T1lo,$T1carry:$T1lo     ; T1 += Sigma1(e)
 214 ||      ROTL    $Ahi,0,$Bhi
 215 ||      ROTL    $Alo,0,$Blo                             ; b = a
 216 ||      SHRU    $Ahi,34-32,$t0lo
 217 ||      SHL     $Ahi,64-34,$t0hi
 218         XOR     $t0hi,$S0hi,$S0hi
 219 ||      XOR     $t0lo,$S0lo,$S0lo
 220 ||      ADD     $MAJhi,$T1hi,$T2hi
 221 ||      ADDU    $MAJlo,$T1carry:$T1lo,$T2carry:$T2lo    ; T2 = T1+Maj(a,b,c)
 222 ||      SHRU    $Alo,34-32,$t0hi
 223 ||      SHL     $Alo,64-34,$t0lo
 224         XOR     $t0hi,$S0hi,$S0hi
 225 ||      XOR     $t0lo,$S0lo,$S0lo
 226 ||      ADD     $Ehi,$T1hi,$T1hi
 227 ||      ADDU    $Elo,$T1carry:$T1lo,$T1carry:$T1lo      ; T1 += e
 228 || [B0] BNOP    loop0_15?
 229 ||      SHRU    $Ahi,39-32,$t0lo
 230 ||      SHL     $Ahi,64-39,$t0hi
 231         XOR     $t0hi,$S0hi,$S0hi
 232 ||      XOR     $t0lo,$S0lo,$S0lo
 233 || [B0] LDNDW   *$INP++,B11:B10                         ; pre-fetch input
 234 ||[!B1] BNOP    break?
 235 ||      SHRU    $Alo,39-32,$t0hi
 236 ||      SHL     $Alo,64-39,$t0lo
 237         XOR     $t0hi,$S0hi,$S0hi
 238 ||      XOR     $t0lo,$S0lo,$S0lo                       ; Sigma0(a)
 239 ||      ADD     $T1carry,$T1hi,$Ehi
 240 ||      MV      $T1lo,$Elo                              ; e = T1
 241 ||[!B0] LDW     *${Xihi}[28],$T1hi
 242 ||[!B0] LDW     *${Xilo}[28],$T1lo                      ; X[i+14]
 243         ADD     $S0hi,$T2hi,$T2hi
 244 ||      ADDU    $S0lo,$T2carry:$T2lo,$T2carry:$T2lo     ; T2 += Sigma0(a)
 245 || [B1] LDDW    *$K512++,$Khi:$Klo                      ; pre-fetch K512[i]
 246         NOP                                             ; avoid cross-path stall
 247         ADD     $T2carry,$T2hi,$Ahi
 248 ||      MV      $T2lo,$Alo                              ; a = T2
 249 || [B0] SUB     B0,1,B0
 250 ;;===== branch to loop00_15? is taken here
 251         NOP
 252 ;;===== branch to break? is taken here
 253         LDW     *${Xihi}[2],$T2hi
 254 ||      LDW     *${Xilo}[2],$T2lo                       ; X[i+1]
 255 ||      SHRU    $T1hi,19,$S1hi
 256 ||      SHL     $T1hi,32-19,$S1lo
 257         SHRU    $T1lo,19,$t0lo
 258 ||      SHL     $T1lo,32-19,$t0hi
 259         XOR     $t0hi,$S1hi,$S1hi
 260 ||      XOR     $t0lo,$S1lo,$S1lo
 261 ||      SHRU    $T1hi,61-32,$t0lo
 262 ||      SHL     $T1hi,64-61,$t0hi
 263         XOR     $t0hi,$S1hi,$S1hi
 264 ||      XOR     $t0lo,$S1lo,$S1lo
 265 ||      SHRU    $T1lo,61-32,$t0hi
 266 ||      SHL     $T1lo,64-61,$t0lo
 267         XOR     $t0hi,$S1hi,$S1hi
 268 ||      XOR     $t0lo,$S1lo,$S1lo
 269 ||      SHRU    $T1hi,6,$t0hi
 270 ||      SHL     $T1hi,32-6,$t0lo
 271         XOR     $t0hi,$S1hi,$S1hi
 272 ||      XOR     $t0lo,$S1lo,$S1lo
 273 ||      SHRU    $T1lo,6,$t0lo
 274 ||      LDW     *${Xihi}[18],$T1hi
 275 ||      LDW     *${Xilo}[18],$T1lo                      ; X[i+9]
 276         XOR     $t0lo,$S1lo,$S1lo                       ; sigma1(Xi[i+14])
 277
 278 ||      LDW     *${Xihi}[0],$CHhi
 279 ||      LDW     *${Xilo}[0],$CHlo                       ; X[i]
 280 ||      SHRU    $T2hi,1,$S0hi
 281 ||      SHL     $T2hi,32-1,$S0lo
 282         SHRU    $T2lo,1,$t0lo
 283 ||      SHL     $T2lo,32-1,$t0hi
 284         XOR     $t0hi,$S0hi,$S0hi
 285 ||      XOR     $t0lo,$S0lo,$S0lo
 286 ||      SHRU    $T2hi,8,$t0hi
 287 ||      SHL     $T2hi,32-8,$t0lo
 288         XOR     $t0hi,$S0hi,$S0hi
 289 ||      XOR     $t0lo,$S0lo,$S0lo
 290 ||      SHRU    $T2lo,8,$t0lo
 291 ||      SHL     $T2lo,32-8,$t0hi
 292         XOR     $t0hi,$S0hi,$S0hi
 293 ||      XOR     $t0lo,$S0lo,$S0lo
 294 ||      ADD     $S1hi,$T1hi,$T1hi
 295 ||      ADDU    $S1lo,$T1lo,$T1carry:$T1lo              ; T1 = X[i+9]+sigma1()
 296 || [B1] BNOP    loop16_79?
 297 ||      SHRU    $T2hi,7,$t0hi
 298 ||      SHL     $T2hi,32-7,$t0lo
 299         XOR     $t0hi,$S0hi,$S0hi
 300 ||      XOR     $t0lo,$S0lo,$S0lo
 301 ||      ADD     $CHhi,$T1hi,$T1hi
 302 ||      ADDU    $CHlo,$T1carry:$T1lo,$T1carry:$T1lo     ; T1 += X[i]
 303 ||      SHRU    $T2lo,7,$t0lo
 304         XOR     $t0lo,$S0lo,$S0lo                       ; sigma0(Xi[i+1]
 305
 306         ADD     $S0hi,$T1hi,$T1hi
 307 ||      ADDU    $S0lo,$T1carry:$T1lo,$T1carry:$T1lo     ; T1 += sigma0()
 308 || [B1] SUB     B1,1,B1
 309         NOP                                             ; avoid cross-path stall
 310         ADD     $T1carry,$T1hi,$T1hi
 311 ;;===== branch to loop16_79? is taken here
 312
 313 break?:
 314         ADD     $Ahi,$Actxhi,$Ahi               ; accumulate ctx
 315 ||      ADDU    $Alo,$Actxlo,$Actxlo:$Alo
 316 || [A0] LDNDW   *$INP++,B11:B10                 ; pre-fetch input
 317 || [A0] ADDK    -640,$K512                      ; rewind pointer to K512
 318         ADD     $Bhi,$Bctxhi,$Bhi
 319 ||      ADDU    $Blo,$Bctxlo,$Bctxlo:$Blo
 320 || [A0] LDDW    *$K512++,$Khi:$Klo              ; pre-fetch K512[0]
 321         ADD     $Chi,$Cctxhi,$Chi
 322 ||      ADDU    $Clo,$Cctxlo,$Cctxlo:$Clo
 323 ||      ADD     $Actxlo,$Ahi,$Ahi
 324 ||[!A0] MV      $CTXA,$CTXB
 325         ADD     $Dhi,$Dctxhi,$Dhi
 326 ||      ADDU    $Dlo,$Dctxlo,$Dctxlo:$Dlo
 327 ||      ADD     $Bctxlo,$Bhi,$Bhi
 328 ||[!A0] STW     $Ahi,*${CTXA}[0^.LITTLE_ENDIAN] ; save ctx
 329 ||[!A0] STW     $Alo,*${CTXB}[1^.LITTLE_ENDIAN]
 330         ADD     $Ehi,$Ectxhi,$Ehi
 331 ||      ADDU    $Elo,$Ectxlo,$Ectxlo:$Elo
 332 ||      ADD     $Cctxlo,$Chi,$Chi
 333 || [A0] BNOP    outerloop?
 334 ||[!A0] STW     $Bhi,*${CTXA}[2^.LITTLE_ENDIAN]
 335 ||[!A0] STW     $Blo,*${CTXB}[3^.LITTLE_ENDIAN]
 336         ADD     $Fhi,$Fctxhi,$Fhi
 337 ||      ADDU    $Flo,$Fctxlo,$Fctxlo:$Flo
 338 ||      ADD     $Dctxlo,$Dhi,$Dhi
 339 ||[!A0] STW     $Chi,*${CTXA}[4^.LITTLE_ENDIAN]
 340 ||[!A0] STW     $Clo,*${CTXB}[5^.LITTLE_ENDIAN]
 341         ADD     $Ghi,$Gctxhi,$Ghi
 342 ||      ADDU    $Glo,$Gctxlo,$Gctxlo:$Glo
 343 ||      ADD     $Ectxlo,$Ehi,$Ehi
 344 ||[!A0] STW     $Dhi,*${CTXA}[6^.LITTLE_ENDIAN]
 345 ||[!A0] STW     $Dlo,*${CTXB}[7^.LITTLE_ENDIAN]
 346         ADD     $Hhi,$Hctxhi,$Hhi
 347 ||      ADDU    $Hlo,$Hctxlo,$Hctxlo:$Hlo
 348 ||      ADD     $Fctxlo,$Fhi,$Fhi
 349 ||[!A0] STW     $Ehi,*${CTXA}[8^.LITTLE_ENDIAN]
 350 ||[!A0] STW     $Elo,*${CTXB}[9^.LITTLE_ENDIAN]
 351         ADD     $Gctxlo,$Ghi,$Ghi
 352 ||[!A0] STW     $Fhi,*${CTXA}[10^.LITTLE_ENDIAN]
 353 ||[!A0] STW     $Flo,*${CTXB}[11^.LITTLE_ENDIAN]
 354         ADD     $Hctxlo,$Hhi,$Hhi
 355 ||[!A0] STW     $Ghi,*${CTXA}[12^.LITTLE_ENDIAN]
 356 ||[!A0] STW     $Glo,*${CTXB}[13^.LITTLE_ENDIAN]
 357 ;;===== branch to outerloop? is taken here
 358
 359         STW     $Hhi,*${CTXA}[14^.LITTLE_ENDIAN]
 360 ||      STW     $Hlo,*${CTXB}[15^.LITTLE_ENDIAN]
 361 ||      MVK     -40,B0
 362         ADD     FP,B0,SP                        ; destroy circular buffer
 363 ||      LDDW    *FP[-4],A11:A10
 364         LDDW    *SP[2],A13:A12
 365 ||      LDDW    *FP[-2],B11:B10
 366         LDDW    *SP[4],B13:B12
 367 ||      BNOP    RA
 368         LDW     *++SP(40),FP                    ; restore frame pointer
 369         MVK     0,B0
 370         MVC     B0,AMR                          ; clear AMR
 371         NOP     2                               ; wait till FP is committed
 372         .endasmfunc
 373
 374         .if     __TI_EABI__
 375         .sect   ".text:sha_asm.const"
 376         .else
 377         .sect   ".const:sha_asm"
 378         .endif
 379         .align  128
 380 K512:
 381         .uword  0x428a2f98,0xd728ae22, 0x71374491,0x23ef65cd
 382         .uword  0xb5c0fbcf,0xec4d3b2f, 0xe9b5dba5,0x8189dbbc
 383         .uword  0x3956c25b,0xf348b538, 0x59f111f1,0xb605d019
 384         .uword  0x923f82a4,0xaf194f9b, 0xab1c5ed5,0xda6d8118
 385         .uword  0xd807aa98,0xa3030242, 0x12835b01,0x45706fbe
 386         .uword  0x243185be,0x4ee4b28c, 0x550c7dc3,0xd5ffb4e2
 387         .uword  0x72be5d74,0xf27b896f, 0x80deb1fe,0x3b1696b1
 388         .uword  0x9bdc06a7,0x25c71235, 0xc19bf174,0xcf692694
 389         .uword  0xe49b69c1,0x9ef14ad2, 0xefbe4786,0x384f25e3
 390         .uword  0x0fc19dc6,0x8b8cd5b5, 0x240ca1cc,0x77ac9c65
 391         .uword  0x2de92c6f,0x592b0275, 0x4a7484aa,0x6ea6e483
 392         .uword  0x5cb0a9dc,0xbd41fbd4, 0x76f988da,0x831153b5
 393         .uword  0x983e5152,0xee66dfab, 0xa831c66d,0x2db43210
 394         .uword  0xb00327c8,0x98fb213f, 0xbf597fc7,0xbeef0ee4
 395         .uword  0xc6e00bf3,0x3da88fc2, 0xd5a79147,0x930aa725
 396         .uword  0x06ca6351,0xe003826f, 0x14292967,0x0a0e6e70
 397         .uword  0x27b70a85,0x46d22ffc, 0x2e1b2138,0x5c26c926
 398         .uword  0x4d2c6dfc,0x5ac42aed, 0x53380d13,0x9d95b3df
 399         .uword  0x650a7354,0x8baf63de, 0x766a0abb,0x3c77b2a8
 400         .uword  0x81c2c92e,0x47edaee6, 0x92722c85,0x1482353b
 401         .uword  0xa2bfe8a1,0x4cf10364, 0xa81a664b,0xbc423001
 402         .uword  0xc24b8b70,0xd0f89791, 0xc76c51a3,0x0654be30
 403         .uword  0xd192e819,0xd6ef5218, 0xd6990624,0x5565a910
 404         .uword  0xf40e3585,0x5771202a, 0x106aa070,0x32bbd1b8
 405         .uword  0x19a4c116,0xb8d2d0c8, 0x1e376c08,0x5141ab53
 406         .uword  0x2748774c,0xdf8eeb99, 0x34b0bcb5,0xe19b48a8
 407         .uword  0x391c0cb3,0xc5c95a63, 0x4ed8aa4a,0xe3418acb
 408         .uword  0x5b9cca4f,0x7763e373, 0x682e6ff3,0xd6b2b8a3
 409         .uword  0x748f82ee,0x5defb2fc, 0x78a5636f,0x43172f60
 410         .uword  0x84c87814,0xa1f0ab72, 0x8cc70208,0x1a6439ec
 411         .uword  0x90befffa,0x23631e28, 0xa4506ceb,0xde82bde9
 412         .uword  0xbef9a3f7,0xb2c67915, 0xc67178f2,0xe372532b
 413         .uword  0xca273ece,0xea26619c, 0xd186b8c7,0x21c0c207
 414         .uword  0xeada7dd6,0xcde0eb1e, 0xf57d4f7f,0xee6ed178
 415         .uword  0x06f067aa,0x72176fba, 0x0a637dc5,0xa2c898a6
 416         .uword  0x113f9804,0xbef90dae, 0x1b710b35,0x131c471b
 417         .uword  0x28db77f5,0x23047d84, 0x32caab7b,0x40c72493
 418         .uword  0x3c9ebe0a,0x15c9bebc, 0x431d67c4,0x9c100d4c
 419         .uword  0x4cc5d4be,0xcb3e42b6, 0x597f299c,0xfc657e2a
 420         .uword  0x5fcb6fab,0x3ad6faec, 0x6c44198c,0x4a475817
 421         .cstring "SHA512 block transform for C64x+, CRYPTOGAMS by <appro\@openssl.org>"
 422         .align  4
 423 ___
 424
 425 print $code;
 426 close STDOUT;