crypto/sha/asm/sha512-mips.pl

   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 # SHA2 block procedures for MIPS.
  11
  12 # October 2010.
  13 #
  14 # SHA256 performance improvement on MIPS R5000 CPU is ~27% over gcc-
  15 # generated code in o32 build and ~55% in n32/64 build. SHA512 [which
  16 # for now can only be compiled for MIPS64 ISA] improvement is modest
  17 # ~17%, but it comes for free, because it's same instruction sequence.
  18 # Improvement coefficients are for aligned input.
  19
  20 ######################################################################
  21 # There is a number of MIPS ABI in use, O32 and N32/64 are most
  22 # widely used. Then there is a new contender: NUBI. It appears that if
  23 # one picks the latter, it's possible to arrange code in ABI neutral
  24 # manner. Therefore let's stick to NUBI register layout:
  25 #
  26 ($zero,$at,$t0,$t1,$t2)=map("\$$_",(0..2,24,25));
  27 ($a0,$a1,$a2,$a3,$a4,$a5,$a6,$a7)=map("\$$_",(4..11));
  28 ($s0,$s1,$s2,$s3,$s4,$s5,$s6,$s7,$s8,$s9,$s10,$s11)=map("\$$_",(12..23));
  29 ($gp,$tp,$sp,$fp,$ra)=map("\$$_",(3,28..31));
  30 #
  31 # The return value is placed in $a0. Following coding rules facilitate
  32 # interoperability:
  33 #
  34 # - never ever touch $tp, "thread pointer", former $gp;
  35 # - copy return value to $t0, former $v0 [or to $a0 if you're adapting
  36 #   old code];
  37 # - on O32 populate $a4-$a7 with 'lw $aN,4*N($sp)' if necessary;
  38 #
  39 # For reference here is register layout for N32/64 MIPS ABIs:
  40 #
  41 # ($zero,$at,$v0,$v1)=map("\$$_",(0..3));
  42 # ($a0,$a1,$a2,$a3,$a4,$a5,$a6,$a7)=map("\$$_",(4..11));
  43 # ($t0,$t1,$t2,$t3,$t8,$t9)=map("\$$_",(12..15,24,25));
  44 # ($s0,$s1,$s2,$s3,$s4,$s5,$s6,$s7)=map("\$$_",(16..23));
  45 # ($gp,$sp,$fp,$ra)=map("\$$_",(28..31));
  46 #
  47 $flavour = shift; # supported flavours are o32,n32,64,nubi32,nubi64
  48
  49 if ($flavour =~ /64|n32/i) {
  50         $PTR_ADD="dadd";        # incidentally works even on n32
  51         $PTR_SUB="dsub";        # incidentally works even on n32
  52         $REG_S="sd";
  53         $REG_L="ld";
  54         $PTR_SLL="dsll";        # incidentally works even on n32
  55         $SZREG=8;
  56 } else {
  57         $PTR_ADD="add";
  58         $PTR_SUB="sub";
  59         $REG_S="sw";
  60         $REG_L="lw";
  61         $PTR_SLL="sll";
  62         $SZREG=4;
  63 }
  64 #
  65 # <appro@openssl.org>
  66 #
  67 ######################################################################
  68
  69 $output=shift;
  70 open STDOUT,">$output";
  71
  72 if ($output =~ /512/) {
  73         $label="512";
  74         $SZ=8;
  75         $LD="ld";               # load from memory
  76         $ST="sd";               # store to memory
  77         $SLL="dsll";            # shift left logical
  78         $SRL="dsrl";            # shift right logical
  79         $ADDU="daddu";
  80         @Sigma0=(28,34,39);
  81         @Sigma1=(14,18,41);
  82         @sigma0=( 7, 1, 8);     # right shift first
  83         @sigma1=( 6,19,61);     # right shift first
  84         $lastK=0x817;
  85         $rounds=80;
  86 } else {
  87         $label="256";
  88         $SZ=4;
  89         $LD="lw";               # load from memory
  90         $ST="sw";               # store to memory
  91         $SLL="sll";             # shift left logical
  92         $SRL="srl";             # shift right logical
  93         $ADDU="addu";
  94         @Sigma0=( 2,13,22);
  95         @Sigma1=( 6,11,25);
  96         @sigma0=( 3, 7,18);     # right shift first
  97         @sigma1=(10,17,19);     # right shift first
  98         $lastK=0x8f2;
  99         $rounds=64;
 100 }
 101
 102 @V=($A,$B,$C,$D,$E,$F,$G,$H)=map("\$$_",(1,2,3,7,24,25,30,31));
 103 @X=map("\$$_",(8..23));
 104
 105 $ctx=$a0;
 106 $inp=$a1;
 107 $len=$a2;       $Ktbl=$len;
 108
 109 sub BODY_00_15 {
 110 my ($i,$a,$b,$c,$d,$e,$f,$g,$h)=@_;
 111 my ($T1,$tmp0,$tmp1,$tmp2)=(@X[4],@X[5],@X[6],@X[7]);
 112
 113 $code.=<<___ if ($i<15);
 114         ${LD}l  @X[1],`($i+1)*$SZ+$MSB`($inp)
 115         ${LD}r  @X[1],`($i+1)*$SZ+$LSB`($inp)
 116 ___
 117 $code.=<<___    if ($little_endian && $i<16);   # XXX no 64-bit byte swap yet
 118         srl     $tmp0,@X[0],24  # byte swap($i)
 119         srl     $tmp1,@X[0],8
 120         andi    $tmp2,@X[0],0xFF00
 121         sll     @X[0],@X[0],24
 122         andi    $tmp1,0xFF00
 123         sll     $tmp2,$tmp2,8
 124         or      @X[0],$tmp0
 125         or      $tmp1,$t2
 126         or      @X[0],$tmp1
 127 ___
 128 $code.=<<___;
 129         $ADDU   $T1,$X[0],$h                    # $i
 130         $SRL    $h,$e,@Sigma1[0]
 131         xor     $tmp2,$f,$g
 132         $SLL    $tmp1,$e,`$SZ*8-@Sigma1[2]`
 133         and     $tmp2,$e
 134         $SRL    $tmp0,$e,@Sigma1[1]
 135         xor     $h,$tmp1
 136         $SLL    $tmp1,$e,`$SZ*8-@Sigma1[1]`
 137         xor     $h,$tmp0
 138         $SRL    $tmp0,$e,@Sigma1[2]
 139         xor     $h,$tmp1
 140         $SLL    $tmp1,$e,`$SZ*8-@Sigma1[0]`
 141         xor     $h,$tmp0
 142         xor     $tmp2,$g                        # Ch(e,f,g)
 143         xor     $tmp0,$tmp1,$h                  # Sigma1(e)
 144
 145         $SRL    $h,$a,@Sigma0[0]
 146         $ADDU   $T1,$tmp2
 147         $LD     $tmp2,`$i*$SZ`($Ktbl)           # K[$i]
 148         $SLL    $tmp1,$a,`$SZ*8-@Sigma0[2]`
 149         $ADDU   $T1,$tmp0
 150         $SRL    $tmp0,$a,@Sigma0[1]
 151         xor     $h,$tmp1
 152         $SLL    $tmp1,$a,`$SZ*8-@Sigma0[1]`
 153         xor     $h,$tmp0
 154         $SRL    $tmp0,$a,@Sigma0[2]
 155         xor     $h,$tmp1
 156         $SLL    $tmp1,$a,`$SZ*8-@Sigma0[0]`
 157         xor     $h,$tmp0
 158         $ST     @X[0],`($i%16)*$SZ`($sp)
 159         xor     $h,$tmp1                        # Sigma0(a)
 160
 161         or      $tmp0,$a,$b
 162         and     $tmp1,$a,$b
 163         and     $tmp0,$c
 164         or      $tmp1,$tmp0                     # Maj(a,b,c)
 165         $ADDU   $T1,$tmp2                       # +=K[$i]
 166         $ADDU   $h,$tmp1
 167
 168         $ADDU   $d,$T1
 169         $ADDU   $h,$T1
 170 ___
 171 $code.=<<___ if ($i>=13);
 172         $LD     @X[3],`(($i+3)%16)*$SZ`($sp)    # prefetch
 173 ___
 174 }
 175
 176 sub BODY_16_XX {
 177 my $i=@_[0];
 178 my ($tmp0,$tmp1,$tmp2,$tmp3)=(@X[4],@X[5],@X[6],@X[7]);
 179
 180 $code.=<<___;
 181         $SRL    $tmp2,@X[1],@sigma0[0]          # Xupdate($i)
 182         $ADDU   @X[0],@X[9]                     # +=X[i+9]
 183         $SLL    $tmp1,@X[1],`$SZ*8-@sigma0[2]`
 184         $SRL    $tmp0,@X[1],@sigma0[1]
 185         xor     $tmp2,$tmp1
 186         $SLL    $tmp1,`@sigma0[2]-@sigma0[1]`
 187         xor     $tmp2,$tmp0
 188         $SRL    $tmp0,@X[1],@sigma0[2]
 189         xor     $tmp2,$tmp1
 190
 191         $SRL    $tmp3,@X[14],@sigma1[0]
 192         xor     $tmp2,$tmp0                     # sigma0(X[i+1])
 193         $SLL    $tmp1,@X[14],`$SZ*8-@sigma1[2]`
 194         $ADDU   @X[0],$tmp2
 195         $SRL    $tmp0,@X[14],@sigma1[1]
 196         xor     $tmp3,$tmp1
 197         $SLL    $tmp1,`@sigma1[2]-@sigma1[1]`
 198         xor     $tmp3,$tmp0
 199         $SRL    $tmp0,@X[14],@sigma1[2]
 200         xor     $tmp3,$tmp1
 201
 202         xor     $tmp3,$tmp0                     # sigma1(X[i+14])
 203         $ADDU   @X[0],$tmp3
 204
 205 ___
 206         &BODY_00_15(@_);
 207 }
 208
 209 $FRAMESIZE=16*$SZ+16*$SZREG;
 210 $SAVED_REGS_MASK = ($flavour =~ /nubi/i) ? 0xc0fff008 : 0xc0ff0000;
 211 $pf = ($flavour =~ /nubi/i) ? $t0 : $t2;
 212 $MSB = 0;
 213 $LSB = ($SZ-1)&~$MSB;
 214
 215 $code.=<<___;
 216 .text
 217 .set    noat
 218 .option pic2
 219
 220 .align  5
 221 .globl  sha${label}_block_data_order
 222 .ent    sha${label}_block_data_order
 223 sha${label}_block_data_order:
 224         .frame  $sp,$FRAMESIZE,$ra
 225         .mask   $SAVED_REGS_MASK,-$SZREG
 226         .set    noreorder
 227 ___
 228 $code.=<<___ if ($flavour =~ /o32/i);   # o32 PIC-ification
 229         .cpload $pf
 230 ___
 231 $code.=<<___;
 232         $PTR_SUB $sp,$FRAMESIZE
 233         $REG_S  $ra,$FRAMESIZE-1*$SZREG($sp)
 234         $REG_S  $fp,$FRAMESIZE-2*$SZREG($sp)
 235         $REG_S  $s11,$FRAMESIZE-3*$SZREG($sp)
 236         $REG_S  $s10,$FRAMESIZE-4*$SZREG($sp)
 237         $REG_S  $s9,$FRAMESIZE-5*$SZREG($sp)
 238         $REG_S  $s8,$FRAMESIZE-6*$SZREG($sp)
 239         $REG_S  $s7,$FRAMESIZE-7*$SZREG($sp)
 240         $REG_S  $s6,$FRAMESIZE-8*$SZREG($sp)
 241         $REG_S  $s5,$FRAMESIZE-9*$SZREG($sp)
 242         $REG_S  $s4,$FRAMESIZE-10*$SZREG($sp)
 243 ___
 244 $code.=<<___ if ($flavour =~ /nubi/i);  # optimize non-nubi prologue
 245         $REG_S  $s3,$FRAMESIZE-11*$SZREG($sp)
 246         $REG_S  $s2,$FRAMESIZE-12*$SZREG($sp)
 247         $REG_S  $s1,$FRAMESIZE-13*$SZREG($sp)
 248         $REG_S  $s0,$FRAMESIZE-14*$SZREG($sp)
 249         $REG_S  $gp,$FRAMESIZE-15*$SZREG($sp)
 250 ___
 251 $code.=<<___;
 252         $PTR_SLL @X[15],$len,`log(16*$SZ)/log(2)`
 253 ___
 254 $code.=<<___ if ($flavour !~ /o32/i);   # non-o32 PIC-ification
 255         .cplocal        $Ktbl
 256         .cpsetup        $pf,$zero,sha${label}_block_data_order
 257 ___
 258 $code.=<<___;
 259         .set    reorder
 260         la      $Ktbl,K${label}         # PIC-ified 'load address'
 261
 262         $LD     $A,0*$SZ($ctx)          # load context
 263         $LD     $B,1*$SZ($ctx)
 264         $LD     $C,2*$SZ($ctx)
 265         $LD     $D,3*$SZ($ctx)
 266         $LD     $E,4*$SZ($ctx)
 267         $LD     $F,5*$SZ($ctx)
 268         $LD     $G,6*$SZ($ctx)
 269         $LD     $H,7*$SZ($ctx)
 270
 271         $PTR_ADD @X[15],$inp            # pointer to the end of input
 272         $REG_S  @X[15],16*$SZ($sp)
 273         b       .Loop
 274
 275 .align  5
 276 .Loop:
 277         ${LD}l  @X[0],$MSB($inp)
 278         ${LD}r  @X[0],$LSB($inp)
 279 ___
 280 for ($i=0;$i<16;$i++)
 281 { &BODY_00_15($i,@V); unshift(@V,pop(@V)); push(@X,shift(@X)); }
 282 $code.=<<___;
 283         b       .L16_xx
 284 .align  4
 285 .L16_xx:
 286 ___
 287 for (;$i<32;$i++)
 288 { &BODY_16_XX($i,@V); unshift(@V,pop(@V)); push(@X,shift(@X)); }
 289 $code.=<<___;
 290         and     @X[6],0xfff
 291         li      @X[7],$lastK
 292         $PTR_ADD $Ktbl,16*$SZ           # Ktbl+=16
 293         bne     @X[6],@X[7],.L16_xx
 294
 295         $LD     @X[0],0*$SZ($ctx)
 296         $LD     @X[1],1*$SZ($ctx)
 297         $LD     @X[2],2*$SZ($ctx)
 298         $LD     @X[3],3*$SZ($ctx)
 299         $ADDU   $A,@X[0]
 300         $LD     @X[4],4*$SZ($ctx)
 301         $ADDU   $B,@X[1]
 302         $LD     @X[5],5*$SZ($ctx)
 303         $ADDU   $C,@X[2]
 304         $LD     @X[6],6*$SZ($ctx)
 305         $ADDU   $D,@X[3]
 306         $LD     @X[7],7*$SZ($ctx)
 307         $ADDU   $E,@X[4]
 308         $ST     $A,0*$SZ($ctx)
 309         $ADDU   $F,@X[5]
 310         $ST     $B,1*$SZ($ctx)
 311         $ADDU   $G,@X[6]
 312         $ST     $C,2*$SZ($ctx)
 313         $ADDU   $H,@X[7]
 314         $ST     $D,3*$SZ($ctx)
 315         $PTR_ADD $inp,16*$SZ
 316         $ST     $E,4*$SZ($ctx)
 317         $REG_L  @X[15],16*$SZ($sp)      # restore pointer to the end of input
 318         $ST     $F,5*$SZ($ctx)
 319         $ST     $G,6*$SZ($ctx)
 320         $ST     $H,7*$SZ($ctx)
 321
 322         $PTR_SUB $Ktbl,`($rounds-16)*$SZ`       # rewind $Ktbl
 323         bne     $inp,@X[15],.Loop
 324
 325         .set    noreorder
 326         $REG_L  $ra,$FRAMESIZE-1*$SZREG($sp)
 327         $REG_L  $fp,$FRAMESIZE-2*$SZREG($sp)
 328         $REG_L  $s11,$FRAMESIZE-3*$SZREG($sp)
 329         $REG_L  $s10,$FRAMESIZE-4*$SZREG($sp)
 330         $REG_L  $s9,$FRAMESIZE-5*$SZREG($sp)
 331         $REG_L  $s8,$FRAMESIZE-6*$SZREG($sp)
 332         $REG_L  $s7,$FRAMESIZE-7*$SZREG($sp)
 333         $REG_L  $s6,$FRAMESIZE-8*$SZREG($sp)
 334         $REG_L  $s5,$FRAMESIZE-9*$SZREG($sp)
 335         $REG_L  $s4,$FRAMESIZE-10*$SZREG($sp)
 336 ___
 337 $code.=<<___ if ($flavour =~ /nubi/i);
 338         $REG_L  $s3,$FRAMESIZE-11*$SZREG($sp)
 339         $REG_L  $s2,$FRAMESIZE-12*$SZREG($sp)
 340         $REG_L  $s1,$FRAMESIZE-13*$SZREG($sp)
 341         $REG_L  $s0,$FRAMESIZE-14*$SZREG($sp)
 342         $REG_L  $gp,$FRAMESIZE-15*$SZREG($sp)
 343 ___
 344 $code.=<<___;
 345         jr      $ra
 346         $PTR_ADD $sp,$FRAMESIZE
 347 .end    sha${label}_block_data_order
 348
 349 .rdata
 350 .align  5
 351 K${label}:
 352 ___
 353 if ($SZ==4) {
 354 $code.=<<___;
 355         .word   0x428a2f98, 0x71374491, 0xb5c0fbcf, 0xe9b5dba5
 356         .word   0x3956c25b, 0x59f111f1, 0x923f82a4, 0xab1c5ed5
 357         .word   0xd807aa98, 0x12835b01, 0x243185be, 0x550c7dc3
 358         .word   0x72be5d74, 0x80deb1fe, 0x9bdc06a7, 0xc19bf174
 359         .word   0xe49b69c1, 0xefbe4786, 0x0fc19dc6, 0x240ca1cc
 360         .word   0x2de92c6f, 0x4a7484aa, 0x5cb0a9dc, 0x76f988da
 361         .word   0x983e5152, 0xa831c66d, 0xb00327c8, 0xbf597fc7
 362         .word   0xc6e00bf3, 0xd5a79147, 0x06ca6351, 0x14292967
 363         .word   0x27b70a85, 0x2e1b2138, 0x4d2c6dfc, 0x53380d13
 364         .word   0x650a7354, 0x766a0abb, 0x81c2c92e, 0x92722c85
 365         .word   0xa2bfe8a1, 0xa81a664b, 0xc24b8b70, 0xc76c51a3
 366         .word   0xd192e819, 0xd6990624, 0xf40e3585, 0x106aa070
 367         .word   0x19a4c116, 0x1e376c08, 0x2748774c, 0x34b0bcb5
 368         .word   0x391c0cb3, 0x4ed8aa4a, 0x5b9cca4f, 0x682e6ff3
 369         .word   0x748f82ee, 0x78a5636f, 0x84c87814, 0x8cc70208
 370         .word   0x90befffa, 0xa4506ceb, 0xbef9a3f7, 0xc67178f2
 371 ___
 372 } else {
 373 $code.=<<___;
 374         .dword  0x428a2f98d728ae22, 0x7137449123ef65cd
 375         .dword  0xb5c0fbcfec4d3b2f, 0xe9b5dba58189dbbc
 376         .dword  0x3956c25bf348b538, 0x59f111f1b605d019
 377         .dword  0x923f82a4af194f9b, 0xab1c5ed5da6d8118
 378         .dword  0xd807aa98a3030242, 0x12835b0145706fbe
 379         .dword  0x243185be4ee4b28c, 0x550c7dc3d5ffb4e2
 380         .dword  0x72be5d74f27b896f, 0x80deb1fe3b1696b1
 381         .dword  0x9bdc06a725c71235, 0xc19bf174cf692694
 382         .dword  0xe49b69c19ef14ad2, 0xefbe4786384f25e3
 383         .dword  0x0fc19dc68b8cd5b5, 0x240ca1cc77ac9c65
 384         .dword  0x2de92c6f592b0275, 0x4a7484aa6ea6e483
 385         .dword  0x5cb0a9dcbd41fbd4, 0x76f988da831153b5
 386         .dword  0x983e5152ee66dfab, 0xa831c66d2db43210
 387         .dword  0xb00327c898fb213f, 0xbf597fc7beef0ee4
 388         .dword  0xc6e00bf33da88fc2, 0xd5a79147930aa725
 389         .dword  0x06ca6351e003826f, 0x142929670a0e6e70
 390         .dword  0x27b70a8546d22ffc, 0x2e1b21385c26c926
 391         .dword  0x4d2c6dfc5ac42aed, 0x53380d139d95b3df
 392         .dword  0x650a73548baf63de, 0x766a0abb3c77b2a8
 393         .dword  0x81c2c92e47edaee6, 0x92722c851482353b
 394         .dword  0xa2bfe8a14cf10364, 0xa81a664bbc423001
 395         .dword  0xc24b8b70d0f89791, 0xc76c51a30654be30
 396         .dword  0xd192e819d6ef5218, 0xd69906245565a910
 397         .dword  0xf40e35855771202a, 0x106aa07032bbd1b8
 398         .dword  0x19a4c116b8d2d0c8, 0x1e376c085141ab53
 399         .dword  0x2748774cdf8eeb99, 0x34b0bcb5e19b48a8
 400         .dword  0x391c0cb3c5c95a63, 0x4ed8aa4ae3418acb
 401         .dword  0x5b9cca4f7763e373, 0x682e6ff3d6b2b8a3
 402         .dword  0x748f82ee5defb2fc, 0x78a5636f43172f60
 403         .dword  0x84c87814a1f0ab72, 0x8cc702081a6439ec
 404         .dword  0x90befffa23631e28, 0xa4506cebde82bde9
 405         .dword  0xbef9a3f7b2c67915, 0xc67178f2e372532b
 406         .dword  0xca273eceea26619c, 0xd186b8c721c0c207
 407         .dword  0xeada7dd6cde0eb1e, 0xf57d4f7fee6ed178
 408         .dword  0x06f067aa72176fba, 0x0a637dc5a2c898a6
 409         .dword  0x113f9804bef90dae, 0x1b710b35131c471b
 410         .dword  0x28db77f523047d84, 0x32caab7b40c72493
 411         .dword  0x3c9ebe0a15c9bebc, 0x431d67c49c100d4c
 412         .dword  0x4cc5d4becb3e42b6, 0x597f299cfc657e2a
 413         .dword  0x5fcb6fab3ad6faec, 0x6c44198c4a475817
 414 ___
 415 }
 416 $code.=<<___;
 417 .asciiz "SHA${label} for MIPS, CRYPTOGAMS by <appro\@openssl.org>"
 418 .align  5
 419
 420 ___
 421
 422 $code =~ s/\`([^\`]*)\`/eval $1/gem;
 423 print $code;
 424 close STDOUT;