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 for (@ARGV) {   $big_endian=1 if (/\-DB_ENDIAN/);
  71                 $big_endian=0 if (/\-DL_ENDIAN/);
  72                 $output=$_ if (/^\w[\w\-]*\.\w+$/);     }
  73 open STDOUT,">$output";
  74
  75 if (!defined($big_endian)) { $big_endian=(unpack('L',pack('N',1))==1); }
  76
  77 if ($output =~ /512/) {
  78         $label="512";
  79         $SZ=8;
  80         $LD="ld";               # load from memory
  81         $ST="sd";               # store to memory
  82         $SLL="dsll";            # shift left logical
  83         $SRL="dsrl";            # shift right logical
  84         $ADDU="daddu";
  85         @Sigma0=(28,34,39);
  86         @Sigma1=(14,18,41);
  87         @sigma0=( 7, 1, 8);     # right shift first
  88         @sigma1=( 6,19,61);     # right shift first
  89         $lastK=0x817;
  90         $rounds=80;
  91 } else {
  92         $label="256";
  93         $SZ=4;
  94         $LD="lw";               # load from memory
  95         $ST="sw";               # store to memory
  96         $SLL="sll";             # shift left logical
  97         $SRL="srl";             # shift right logical
  98         $ADDU="addu";
  99         @Sigma0=( 2,13,22);
 100         @Sigma1=( 6,11,25);
 101         @sigma0=( 3, 7,18);     # right shift first
 102         @sigma1=(10,17,19);     # right shift first
 103         $lastK=0x8f2;
 104         $rounds=64;
 105 }
 106
 107 @V=($A,$B,$C,$D,$E,$F,$G,$H)=map("\$$_",(1,2,3,7,24,25,30,31));
 108 @X=map("\$$_",(8..23));
 109
 110 $ctx=$a0;
 111 $inp=$a1;
 112 $len=$a2;       $Ktbl=$len;
 113
 114 sub BODY_00_15 {
 115 my ($i,$a,$b,$c,$d,$e,$f,$g,$h)=@_;
 116 my ($T1,$tmp0,$tmp1,$tmp2)=(@X[4],@X[5],@X[6],@X[7]);
 117
 118 $code.=<<___ if ($i<15);
 119         ${LD}l  @X[1],`($i+1)*$SZ+$MSB`($inp)
 120         ${LD}r  @X[1],`($i+1)*$SZ+$LSB`($inp)
 121 ___
 122 $code.=<<___    if (!$big_endian && $i<16);     # XXX no 64-bit byte swap yet
 123         srl     $tmp0,@X[0],24  # byte swap($i)
 124         srl     $tmp1,@X[0],8
 125         andi    $tmp2,@X[0],0xFF00
 126         sll     @X[0],@X[0],24
 127         andi    $tmp1,0xFF00
 128         sll     $tmp2,$tmp2,8
 129         or      @X[0],$tmp0
 130         or      $tmp1,$t2
 131         or      @X[0],$tmp1
 132 ___
 133 $code.=<<___;
 134         $ADDU   $T1,$X[0],$h                    # $i
 135         $SRL    $h,$e,@Sigma1[0]
 136         xor     $tmp2,$f,$g
 137         $SLL    $tmp1,$e,`$SZ*8-@Sigma1[2]`
 138         and     $tmp2,$e
 139         $SRL    $tmp0,$e,@Sigma1[1]
 140         xor     $h,$tmp1
 141         $SLL    $tmp1,$e,`$SZ*8-@Sigma1[1]`
 142         xor     $h,$tmp0
 143         $SRL    $tmp0,$e,@Sigma1[2]
 144         xor     $h,$tmp1
 145         $SLL    $tmp1,$e,`$SZ*8-@Sigma1[0]`
 146         xor     $h,$tmp0
 147         xor     $tmp2,$g                        # Ch(e,f,g)
 148         xor     $tmp0,$tmp1,$h                  # Sigma1(e)
 149
 150         $SRL    $h,$a,@Sigma0[0]
 151         $ADDU   $T1,$tmp2
 152         $LD     $tmp2,`$i*$SZ`($Ktbl)           # K[$i]
 153         $SLL    $tmp1,$a,`$SZ*8-@Sigma0[2]`
 154         $ADDU   $T1,$tmp0
 155         $SRL    $tmp0,$a,@Sigma0[1]
 156         xor     $h,$tmp1
 157         $SLL    $tmp1,$a,`$SZ*8-@Sigma0[1]`
 158         xor     $h,$tmp0
 159         $SRL    $tmp0,$a,@Sigma0[2]
 160         xor     $h,$tmp1
 161         $SLL    $tmp1,$a,`$SZ*8-@Sigma0[0]`
 162         xor     $h,$tmp0
 163         $ST     @X[0],`($i%16)*$SZ`($sp)
 164         xor     $h,$tmp1                        # Sigma0(a)
 165
 166         or      $tmp0,$a,$b
 167         and     $tmp1,$a,$b
 168         and     $tmp0,$c
 169         or      $tmp1,$tmp0                     # Maj(a,b,c)
 170         $ADDU   $T1,$tmp2                       # +=K[$i]
 171         $ADDU   $h,$tmp1
 172
 173         $ADDU   $d,$T1
 174         $ADDU   $h,$T1
 175 ___
 176 $code.=<<___ if ($i>=13);
 177         $LD     @X[3],`(($i+3)%16)*$SZ`($sp)    # prefetch
 178 ___
 179 }
 180
 181 sub BODY_16_XX {
 182 my $i=@_[0];
 183 my ($tmp0,$tmp1,$tmp2,$tmp3)=(@X[4],@X[5],@X[6],@X[7]);
 184
 185 $code.=<<___;
 186         $SRL    $tmp2,@X[1],@sigma0[0]          # Xupdate($i)
 187         $ADDU   @X[0],@X[9]                     # +=X[i+9]
 188         $SLL    $tmp1,@X[1],`$SZ*8-@sigma0[2]`
 189         $SRL    $tmp0,@X[1],@sigma0[1]
 190         xor     $tmp2,$tmp1
 191         $SLL    $tmp1,`@sigma0[2]-@sigma0[1]`
 192         xor     $tmp2,$tmp0
 193         $SRL    $tmp0,@X[1],@sigma0[2]
 194         xor     $tmp2,$tmp1
 195
 196         $SRL    $tmp3,@X[14],@sigma1[0]
 197         xor     $tmp2,$tmp0                     # sigma0(X[i+1])
 198         $SLL    $tmp1,@X[14],`$SZ*8-@sigma1[2]`
 199         $ADDU   @X[0],$tmp2
 200         $SRL    $tmp0,@X[14],@sigma1[1]
 201         xor     $tmp3,$tmp1
 202         $SLL    $tmp1,`@sigma1[2]-@sigma1[1]`
 203         xor     $tmp3,$tmp0
 204         $SRL    $tmp0,@X[14],@sigma1[2]
 205         xor     $tmp3,$tmp1
 206
 207         xor     $tmp3,$tmp0                     # sigma1(X[i+14])
 208         $ADDU   @X[0],$tmp3
 209
 210 ___
 211         &BODY_00_15(@_);
 212 }
 213
 214 $FRAMESIZE=16*$SZ+16*$SZREG;
 215 $SAVED_REGS_MASK = ($flavour =~ /nubi/i) ? 0xc0fff008 : 0xc0ff0000;
 216 $pf = ($flavour =~ /nubi/i) ? $t0 : $t2;
 217 $MSB = 0;
 218 $LSB = ($SZ-1)&~$MSB;
 219
 220 $code.=<<___;
 221 .text
 222 .set    noat
 223 .option pic2
 224
 225 .align  5
 226 .globl  sha${label}_block_data_order
 227 .ent    sha${label}_block_data_order
 228 sha${label}_block_data_order:
 229         .frame  $sp,$FRAMESIZE,$ra
 230         .mask   $SAVED_REGS_MASK,-$SZREG
 231         .set    noreorder
 232 ___
 233 $code.=<<___ if ($flavour =~ /o32/i);   # o32 PIC-ification
 234         .cpload $pf
 235 ___
 236 $code.=<<___;
 237         $PTR_SUB $sp,$FRAMESIZE
 238         $REG_S  $ra,$FRAMESIZE-1*$SZREG($sp)
 239         $REG_S  $fp,$FRAMESIZE-2*$SZREG($sp)
 240         $REG_S  $s11,$FRAMESIZE-3*$SZREG($sp)
 241         $REG_S  $s10,$FRAMESIZE-4*$SZREG($sp)
 242         $REG_S  $s9,$FRAMESIZE-5*$SZREG($sp)
 243         $REG_S  $s8,$FRAMESIZE-6*$SZREG($sp)
 244         $REG_S  $s7,$FRAMESIZE-7*$SZREG($sp)
 245         $REG_S  $s6,$FRAMESIZE-8*$SZREG($sp)
 246         $REG_S  $s5,$FRAMESIZE-9*$SZREG($sp)
 247         $REG_S  $s4,$FRAMESIZE-10*$SZREG($sp)
 248 ___
 249 $code.=<<___ if ($flavour =~ /nubi/i);  # optimize non-nubi prologue
 250         $REG_S  $s3,$FRAMESIZE-11*$SZREG($sp)
 251         $REG_S  $s2,$FRAMESIZE-12*$SZREG($sp)
 252         $REG_S  $s1,$FRAMESIZE-13*$SZREG($sp)
 253         $REG_S  $s0,$FRAMESIZE-14*$SZREG($sp)
 254         $REG_S  $gp,$FRAMESIZE-15*$SZREG($sp)
 255 ___
 256 $code.=<<___;
 257         $PTR_SLL @X[15],$len,`log(16*$SZ)/log(2)`
 258 ___
 259 $code.=<<___ if ($flavour !~ /o32/i);   # non-o32 PIC-ification
 260         .cplocal        $Ktbl
 261         .cpsetup        $pf,$zero,sha${label}_block_data_order
 262 ___
 263 $code.=<<___;
 264         .set    reorder
 265         la      $Ktbl,K${label}         # PIC-ified 'load address'
 266
 267         $LD     $A,0*$SZ($ctx)          # load context
 268         $LD     $B,1*$SZ($ctx)
 269         $LD     $C,2*$SZ($ctx)
 270         $LD     $D,3*$SZ($ctx)
 271         $LD     $E,4*$SZ($ctx)
 272         $LD     $F,5*$SZ($ctx)
 273         $LD     $G,6*$SZ($ctx)
 274         $LD     $H,7*$SZ($ctx)
 275
 276         $PTR_ADD @X[15],$inp            # pointer to the end of input
 277         $REG_S  @X[15],16*$SZ($sp)
 278         b       .Loop
 279
 280 .align  5
 281 .Loop:
 282         ${LD}l  @X[0],$MSB($inp)
 283         ${LD}r  @X[0],$LSB($inp)
 284 ___
 285 for ($i=0;$i<16;$i++)
 286 { &BODY_00_15($i,@V); unshift(@V,pop(@V)); push(@X,shift(@X)); }
 287 $code.=<<___;
 288         b       .L16_xx
 289 .align  4
 290 .L16_xx:
 291 ___
 292 for (;$i<32;$i++)
 293 { &BODY_16_XX($i,@V); unshift(@V,pop(@V)); push(@X,shift(@X)); }
 294 $code.=<<___;
 295         and     @X[6],0xfff
 296         li      @X[7],$lastK
 297         $PTR_ADD $Ktbl,16*$SZ           # Ktbl+=16
 298         bne     @X[6],@X[7],.L16_xx
 299
 300         $LD     @X[0],0*$SZ($ctx)
 301         $LD     @X[1],1*$SZ($ctx)
 302         $LD     @X[2],2*$SZ($ctx)
 303         $LD     @X[3],3*$SZ($ctx)
 304         $ADDU   $A,@X[0]
 305         $LD     @X[4],4*$SZ($ctx)
 306         $ADDU   $B,@X[1]
 307         $LD     @X[5],5*$SZ($ctx)
 308         $ADDU   $C,@X[2]
 309         $LD     @X[6],6*$SZ($ctx)
 310         $ADDU   $D,@X[3]
 311         $LD     @X[7],7*$SZ($ctx)
 312         $ADDU   $E,@X[4]
 313         $ST     $A,0*$SZ($ctx)
 314         $ADDU   $F,@X[5]
 315         $ST     $B,1*$SZ($ctx)
 316         $ADDU   $G,@X[6]
 317         $ST     $C,2*$SZ($ctx)
 318         $ADDU   $H,@X[7]
 319         $ST     $D,3*$SZ($ctx)
 320         $PTR_ADD $inp,16*$SZ
 321         $ST     $E,4*$SZ($ctx)
 322         $REG_L  @X[15],16*$SZ($sp)      # restore pointer to the end of input
 323         $ST     $F,5*$SZ($ctx)
 324         $ST     $G,6*$SZ($ctx)
 325         $ST     $H,7*$SZ($ctx)
 326
 327         $PTR_SUB $Ktbl,`($rounds-16)*$SZ`       # rewind $Ktbl
 328         bne     $inp,@X[15],.Loop
 329
 330         .set    noreorder
 331         $REG_L  $ra,$FRAMESIZE-1*$SZREG($sp)
 332         $REG_L  $fp,$FRAMESIZE-2*$SZREG($sp)
 333         $REG_L  $s11,$FRAMESIZE-3*$SZREG($sp)
 334         $REG_L  $s10,$FRAMESIZE-4*$SZREG($sp)
 335         $REG_L  $s9,$FRAMESIZE-5*$SZREG($sp)
 336         $REG_L  $s8,$FRAMESIZE-6*$SZREG($sp)
 337         $REG_L  $s7,$FRAMESIZE-7*$SZREG($sp)
 338         $REG_L  $s6,$FRAMESIZE-8*$SZREG($sp)
 339         $REG_L  $s5,$FRAMESIZE-9*$SZREG($sp)
 340         $REG_L  $s4,$FRAMESIZE-10*$SZREG($sp)
 341 ___
 342 $code.=<<___ if ($flavour =~ /nubi/i);
 343         $REG_L  $s3,$FRAMESIZE-11*$SZREG($sp)
 344         $REG_L  $s2,$FRAMESIZE-12*$SZREG($sp)
 345         $REG_L  $s1,$FRAMESIZE-13*$SZREG($sp)
 346         $REG_L  $s0,$FRAMESIZE-14*$SZREG($sp)
 347         $REG_L  $gp,$FRAMESIZE-15*$SZREG($sp)
 348 ___
 349 $code.=<<___;
 350         jr      $ra
 351         $PTR_ADD $sp,$FRAMESIZE
 352 .end    sha${label}_block_data_order
 353
 354 .rdata
 355 .align  5
 356 K${label}:
 357 ___
 358 if ($SZ==4) {
 359 $code.=<<___;
 360         .word   0x428a2f98, 0x71374491, 0xb5c0fbcf, 0xe9b5dba5
 361         .word   0x3956c25b, 0x59f111f1, 0x923f82a4, 0xab1c5ed5
 362         .word   0xd807aa98, 0x12835b01, 0x243185be, 0x550c7dc3
 363         .word   0x72be5d74, 0x80deb1fe, 0x9bdc06a7, 0xc19bf174
 364         .word   0xe49b69c1, 0xefbe4786, 0x0fc19dc6, 0x240ca1cc
 365         .word   0x2de92c6f, 0x4a7484aa, 0x5cb0a9dc, 0x76f988da
 366         .word   0x983e5152, 0xa831c66d, 0xb00327c8, 0xbf597fc7
 367         .word   0xc6e00bf3, 0xd5a79147, 0x06ca6351, 0x14292967
 368         .word   0x27b70a85, 0x2e1b2138, 0x4d2c6dfc, 0x53380d13
 369         .word   0x650a7354, 0x766a0abb, 0x81c2c92e, 0x92722c85
 370         .word   0xa2bfe8a1, 0xa81a664b, 0xc24b8b70, 0xc76c51a3
 371         .word   0xd192e819, 0xd6990624, 0xf40e3585, 0x106aa070
 372         .word   0x19a4c116, 0x1e376c08, 0x2748774c, 0x34b0bcb5
 373         .word   0x391c0cb3, 0x4ed8aa4a, 0x5b9cca4f, 0x682e6ff3
 374         .word   0x748f82ee, 0x78a5636f, 0x84c87814, 0x8cc70208
 375         .word   0x90befffa, 0xa4506ceb, 0xbef9a3f7, 0xc67178f2
 376 ___
 377 } else {
 378 $code.=<<___;
 379         .dword  0x428a2f98d728ae22, 0x7137449123ef65cd
 380         .dword  0xb5c0fbcfec4d3b2f, 0xe9b5dba58189dbbc
 381         .dword  0x3956c25bf348b538, 0x59f111f1b605d019
 382         .dword  0x923f82a4af194f9b, 0xab1c5ed5da6d8118
 383         .dword  0xd807aa98a3030242, 0x12835b0145706fbe
 384         .dword  0x243185be4ee4b28c, 0x550c7dc3d5ffb4e2
 385         .dword  0x72be5d74f27b896f, 0x80deb1fe3b1696b1
 386         .dword  0x9bdc06a725c71235, 0xc19bf174cf692694
 387         .dword  0xe49b69c19ef14ad2, 0xefbe4786384f25e3
 388         .dword  0x0fc19dc68b8cd5b5, 0x240ca1cc77ac9c65
 389         .dword  0x2de92c6f592b0275, 0x4a7484aa6ea6e483
 390         .dword  0x5cb0a9dcbd41fbd4, 0x76f988da831153b5
 391         .dword  0x983e5152ee66dfab, 0xa831c66d2db43210
 392         .dword  0xb00327c898fb213f, 0xbf597fc7beef0ee4
 393         .dword  0xc6e00bf33da88fc2, 0xd5a79147930aa725
 394         .dword  0x06ca6351e003826f, 0x142929670a0e6e70
 395         .dword  0x27b70a8546d22ffc, 0x2e1b21385c26c926
 396         .dword  0x4d2c6dfc5ac42aed, 0x53380d139d95b3df
 397         .dword  0x650a73548baf63de, 0x766a0abb3c77b2a8
 398         .dword  0x81c2c92e47edaee6, 0x92722c851482353b
 399         .dword  0xa2bfe8a14cf10364, 0xa81a664bbc423001
 400         .dword  0xc24b8b70d0f89791, 0xc76c51a30654be30
 401         .dword  0xd192e819d6ef5218, 0xd69906245565a910
 402         .dword  0xf40e35855771202a, 0x106aa07032bbd1b8
 403         .dword  0x19a4c116b8d2d0c8, 0x1e376c085141ab53
 404         .dword  0x2748774cdf8eeb99, 0x34b0bcb5e19b48a8
 405         .dword  0x391c0cb3c5c95a63, 0x4ed8aa4ae3418acb
 406         .dword  0x5b9cca4f7763e373, 0x682e6ff3d6b2b8a3
 407         .dword  0x748f82ee5defb2fc, 0x78a5636f43172f60
 408         .dword  0x84c87814a1f0ab72, 0x8cc702081a6439ec
 409         .dword  0x90befffa23631e28, 0xa4506cebde82bde9
 410         .dword  0xbef9a3f7b2c67915, 0xc67178f2e372532b
 411         .dword  0xca273eceea26619c, 0xd186b8c721c0c207
 412         .dword  0xeada7dd6cde0eb1e, 0xf57d4f7fee6ed178
 413         .dword  0x06f067aa72176fba, 0x0a637dc5a2c898a6
 414         .dword  0x113f9804bef90dae, 0x1b710b35131c471b
 415         .dword  0x28db77f523047d84, 0x32caab7b40c72493
 416         .dword  0x3c9ebe0a15c9bebc, 0x431d67c49c100d4c
 417         .dword  0x4cc5d4becb3e42b6, 0x597f299cfc657e2a
 418         .dword  0x5fcb6fab3ad6faec, 0x6c44198c4a475817
 419 ___
 420 }
 421 $code.=<<___;
 422 .asciiz "SHA${label} for MIPS, CRYPTOGAMS by <appro\@openssl.org>"
 423 .align  5
 424
 425 ___
 426
 427 $code =~ s/\`([^\`]*)\`/eval $1/gem;
 428 print $code;
 429 close STDOUT;