devel/perlasm/e_padlock-x86_64.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 # September 2011
  11 #
  12 # Assembler helpers for Padlock engine. See even e_padlock-x86.pl for
  13 # details.
  14
  15 $flavour = shift;
  16 $output  = shift;
  17 if ($flavour =~ /\./) { $output = $flavour; undef $flavour; }
  18
  19 $win64=0; $win64=1 if ($flavour =~ /[nm]asm|mingw64/ || $output =~ /\.asm$/);
  20
  21 $0 =~ m/(.*[\/\\])[^\/\\]+$/; $dir=$1;
  22 ( $xlate="${dir}x86_64-xlate.pl" and -f $xlate ) or
  23 ( $xlate="${dir}../../crypto/perlasm/x86_64-xlate.pl" and -f $xlate) or
  24 die "can't locate x86_64-xlate.pl";
  25
  26 open STDOUT,"| $^X $xlate $flavour $output";
  27
  28 $code=".text\n";
  29
  30 %PADLOCK_PREFETCH=(ecb=>128, cbc=>64, ctr32=>32);       # prefetch errata
  31 $PADLOCK_CHUNK=512;     # Must be a power of 2 between 32 and 2^20
  32
  33 $ctx="%rdx";
  34 $out="%rdi";
  35 $inp="%rsi";
  36 $len="%rcx";
  37 $chunk="%rbx";
  38
  39 ($arg1,$arg2,$arg3,$arg4)=$win64?("%rcx","%rdx","%r8", "%r9") : # Win64 order
  40                                  ("%rdi","%rsi","%rdx","%rcx"); # Unix order
  41
  42 $code.=<<___;
  43 .globl  padlock_capability
  44 .type   padlock_capability,\@abi-omnipotent
  45 .align  16
  46 padlock_capability:
  47         mov     %rbx,%r8
  48         xor     %eax,%eax
  49         cpuid
  50         xor     %eax,%eax
  51         cmp     \$`"0x".unpack("H*",'tneC')`,%ebx
  52         jne     .Lnoluck
  53         cmp     \$`"0x".unpack("H*",'Hrua')`,%edx
  54         jne     .Lnoluck
  55         cmp     \$`"0x".unpack("H*",'slua')`,%ecx
  56         jne     .Lnoluck
  57         mov     \$0xC0000000,%eax
  58         cpuid
  59         mov     %eax,%edx
  60         xor     %eax,%eax
  61         cmp     \$0xC0000001,%edx
  62         jb      .Lnoluck
  63         mov     \$0xC0000001,%eax
  64         cpuid
  65         mov     %edx,%eax
  66         and     \$0xffffffef,%eax
  67         or      \$0x10,%eax             # set Nano bit#4
  68 .Lnoluck:
  69         mov     %r8,%rbx
  70         ret
  71 .size   padlock_capability,.-padlock_capability
  72
  73 .globl  padlock_key_bswap
  74 .type   padlock_key_bswap,\@abi-omnipotent,0
  75 .align  16
  76 padlock_key_bswap:
  77         mov     240($arg1),%edx
  78 .Lbswap_loop:
  79         mov     ($arg1),%eax
  80         bswap   %eax
  81         mov     %eax,($arg1)
  82         lea     4($arg1),$arg1
  83         sub     \$1,%edx
  84         jnz     .Lbswap_loop
  85         ret
  86 .size   padlock_key_bswap,.-padlock_key_bswap
  87
  88 .globl  padlock_verify_context
  89 .type   padlock_verify_context,\@abi-omnipotent
  90 .align  16
  91 padlock_verify_context:
  92         mov     $arg1,$ctx
  93         pushf
  94         lea     .Lpadlock_saved_context(%rip),%rax
  95         call    _padlock_verify_ctx
  96         lea     8(%rsp),%rsp
  97         ret
  98 .size   padlock_verify_context,.-padlock_verify_context
  99
 100 .type   _padlock_verify_ctx,\@abi-omnipotent
 101 .align  16
 102 _padlock_verify_ctx:
 103         mov     8(%rsp),%r8
 104         bt      \$30,%r8
 105         jnc     .Lverified
 106         cmp     (%rax),$ctx
 107         je      .Lverified
 108         pushf
 109         popf
 110 .Lverified:
 111         mov     $ctx,(%rax)
 112         ret
 113 .size   _padlock_verify_ctx,.-_padlock_verify_ctx
 114
 115 .globl  padlock_reload_key
 116 .type   padlock_reload_key,\@abi-omnipotent
 117 .align  16
 118 padlock_reload_key:
 119         pushf
 120         popf
 121         ret
 122 .size   padlock_reload_key,.-padlock_reload_key
 123
 124 .globl  padlock_aes_block
 125 .type   padlock_aes_block,\@function,3
 126 .align  16
 127 padlock_aes_block:
 128         mov     %rbx,%r8
 129         mov     \$1,$len
 130         lea     32($ctx),%rbx           # key
 131         lea     16($ctx),$ctx           # control word
 132         .byte   0xf3,0x0f,0xa7,0xc8     # rep xcryptecb
 133         mov     %r8,%rbx
 134         ret
 135 .size   padlock_aes_block,.-padlock_aes_block
 136
 137 .globl  padlock_xstore
 138 .type   padlock_xstore,\@function,2
 139 .align  16
 140 padlock_xstore:
 141         mov     %esi,%edx
 142         .byte   0x0f,0xa7,0xc0          # xstore
 143         ret
 144 .size   padlock_xstore,.-padlock_xstore
 145
 146 .globl  padlock_sha1_oneshot
 147 .type   padlock_sha1_oneshot,\@function,3
 148 .align  16
 149 padlock_sha1_oneshot:
 150         mov     %rdx,%rcx
 151         mov     %rdi,%rdx               # put aside %rdi
 152         movups  (%rdi),%xmm0            # copy-in context
 153         sub     \$128+8,%rsp
 154         mov     16(%rdi),%eax
 155         movaps  %xmm0,(%rsp)
 156         mov     %rsp,%rdi
 157         mov     %eax,16(%rsp)
 158         xor     %rax,%rax
 159         .byte   0xf3,0x0f,0xa6,0xc8     # rep xsha1
 160         movaps  (%rsp),%xmm0
 161         mov     16(%rsp),%eax
 162         add     \$128+8,%rsp
 163         movups  %xmm0,(%rdx)            # copy-out context
 164         mov     %eax,16(%rdx)
 165         ret
 166 .size   padlock_sha1_oneshot,.-padlock_sha1_oneshot
 167
 168 .globl  padlock_sha1_blocks
 169 .type   padlock_sha1_blocks,\@function,3
 170 .align  16
 171 padlock_sha1_blocks:
 172         mov     %rdx,%rcx
 173         mov     %rdi,%rdx               # put aside %rdi
 174         movups  (%rdi),%xmm0            # copy-in context
 175         sub     \$128+8,%rsp
 176         mov     16(%rdi),%eax
 177         movaps  %xmm0,(%rsp)
 178         mov     %rsp,%rdi
 179         mov     %eax,16(%rsp)
 180         mov     \$-1,%rax
 181         .byte   0xf3,0x0f,0xa6,0xc8     # rep xsha1
 182         movaps  (%rsp),%xmm0
 183         mov     16(%rsp),%eax
 184         add     \$128+8,%rsp
 185         movups  %xmm0,(%rdx)            # copy-out context
 186         mov     %eax,16(%rdx)
 187         ret
 188 .size   padlock_sha1_blocks,.-padlock_sha1_blocks
 189
 190 .globl  padlock_sha256_oneshot
 191 .type   padlock_sha256_oneshot,\@function,3
 192 .align  16
 193 padlock_sha256_oneshot:
 194         mov     %rdx,%rcx
 195         mov     %rdi,%rdx               # put aside %rdi
 196         movups  (%rdi),%xmm0            # copy-in context
 197         sub     \$128+8,%rsp
 198         movups  16(%rdi),%xmm1
 199         movaps  %xmm0,(%rsp)
 200         mov     %rsp,%rdi
 201         movaps  %xmm1,16(%rsp)
 202         xor     %rax,%rax
 203         .byte   0xf3,0x0f,0xa6,0xd0     # rep xsha256
 204         movaps  (%rsp),%xmm0
 205         movaps  16(%rsp),%xmm1
 206         add     \$128+8,%rsp
 207         movups  %xmm0,(%rdx)            # copy-out context
 208         movups  %xmm1,16(%rdx)
 209         ret
 210 .size   padlock_sha256_oneshot,.-padlock_sha256_oneshot
 211
 212 .globl  padlock_sha256_blocks
 213 .type   padlock_sha256_blocks,\@function,3
 214 .align  16
 215 padlock_sha256_blocks:
 216         mov     %rdx,%rcx
 217         mov     %rdi,%rdx               # put aside %rdi
 218         movups  (%rdi),%xmm0            # copy-in context
 219         sub     \$128+8,%rsp
 220         movups  16(%rdi),%xmm1
 221         movaps  %xmm0,(%rsp)
 222         mov     %rsp,%rdi
 223         movaps  %xmm1,16(%rsp)
 224         mov     \$-1,%rax
 225         .byte   0xf3,0x0f,0xa6,0xd0     # rep xsha256
 226         movaps  (%rsp),%xmm0
 227         movaps  16(%rsp),%xmm1
 228         add     \$128+8,%rsp
 229         movups  %xmm0,(%rdx)            # copy-out context
 230         movups  %xmm1,16(%rdx)
 231         ret
 232 .size   padlock_sha256_blocks,.-padlock_sha256_blocks
 233
 234 .globl  padlock_sha512_blocks
 235 .type   padlock_sha512_blocks,\@function,3
 236 .align  16
 237 padlock_sha512_blocks:
 238         mov     %rdx,%rcx
 239         mov     %rdi,%rdx               # put aside %rdi
 240         movups  (%rdi),%xmm0            # copy-in context
 241         sub     \$128+8,%rsp
 242         movups  16(%rdi),%xmm1
 243         movups  32(%rdi),%xmm2
 244         movups  48(%rdi),%xmm3
 245         movaps  %xmm0,(%rsp)
 246         mov     %rsp,%rdi
 247         movaps  %xmm1,16(%rsp)
 248         movaps  %xmm2,32(%rsp)
 249         movaps  %xmm3,48(%rsp)
 250         .byte   0xf3,0x0f,0xa6,0xe0     # rep xha512
 251         movaps  (%rsp),%xmm0
 252         movaps  16(%rsp),%xmm1
 253         movaps  32(%rsp),%xmm2
 254         movaps  48(%rsp),%xmm3
 255         add     \$128+8,%rsp
 256         movups  %xmm0,(%rdx)            # copy-out context
 257         movups  %xmm1,16(%rdx)
 258         movups  %xmm2,32(%rdx)
 259         movups  %xmm3,48(%rdx)
 260         ret
 261 .size   padlock_sha512_blocks,.-padlock_sha512_blocks
 262 ___
 263
 264 sub generate_mode {
 265 my ($mode,$opcode) = @_;
 266 # int padlock_$mode_encrypt(void *out, const void *inp,
 267 #               struct padlock_cipher_data *ctx, size_t len);
 268 $code.=<<___;
 269 .globl  padlock_${mode}_encrypt
 270 .type   padlock_${mode}_encrypt,\@function,4
 271 .align  16
 272 padlock_${mode}_encrypt:
 273         push    %rbp
 274         push    %rbx
 275
 276         xor     %eax,%eax
 277         test    \$15,$ctx
 278         jnz     .L${mode}_abort
 279         test    \$15,$len
 280         jnz     .L${mode}_abort
 281         lea     .Lpadlock_saved_context(%rip),%rax
 282         pushf
 283         cld
 284         call    _padlock_verify_ctx
 285         lea     16($ctx),$ctx           # control word
 286         xor     %eax,%eax
 287         xor     %ebx,%ebx
 288         testl   \$`1<<5`,($ctx)         # align bit in control word
 289         jnz     .L${mode}_aligned
 290         test    \$0x0f,$out
 291         setz    %al                     # !out_misaligned
 292         test    \$0x0f,$inp
 293         setz    %bl                     # !inp_misaligned
 294         test    %ebx,%eax
 295         jnz     .L${mode}_aligned
 296         neg     %rax
 297         mov     \$$PADLOCK_CHUNK,$chunk
 298         not     %rax                    # out_misaligned?-1:0
 299         lea     (%rsp),%rbp
 300         cmp     $chunk,$len
 301         cmovc   $len,$chunk             # chunk=len>PADLOCK_CHUNK?PADLOCK_CHUNK:len
 302         and     $chunk,%rax             # out_misaligned?chunk:0
 303         mov     $len,$chunk
 304         neg     %rax
 305         and     \$$PADLOCK_CHUNK-1,$chunk       # chunk%=PADLOCK_CHUNK
 306         lea     (%rax,%rbp),%rsp
 307         mov     \$$PADLOCK_CHUNK,%rax
 308         cmovz   %rax,$chunk                     # chunk=chunk?:PADLOCK_CHUNK
 309 ___
 310 $code.=<<___                            if ($mode eq "ctr32");
 311 .L${mode}_reenter:
 312         mov     -4($ctx),%eax           # pull 32-bit counter
 313         bswap   %eax
 314         neg     %eax
 315         and     \$`$PADLOCK_CHUNK/16-1`,%eax
 316         mov     \$$PADLOCK_CHUNK,$chunk
 317         shl     \$4,%eax
 318         cmovz   $chunk,%rax
 319         cmp     %rax,$len
 320         cmova   %rax,$chunk             # don't let counter cross PADLOCK_CHUNK
 321         cmovbe  $len,$chunk
 322 ___
 323 $code.=<<___                            if ($PADLOCK_PREFETCH{$mode});
 324         cmp     $chunk,$len
 325         ja      .L${mode}_loop
 326         mov     $inp,%rax               # check if prefetch crosses page
 327         cmp     %rsp,%rbp
 328         cmove   $out,%rax
 329         add     $len,%rax
 330         neg     %rax
 331         and     \$0xfff,%rax            # distance to page boundary
 332         cmp     \$$PADLOCK_PREFETCH{$mode},%rax
 333         mov     \$-$PADLOCK_PREFETCH{$mode},%rax
 334         cmovae  $chunk,%rax             # mask=distance<prefetch?-prefetch:-1
 335         and     %rax,$chunk
 336         jz      .L${mode}_unaligned_tail
 337 ___
 338 $code.=<<___;
 339         jmp     .L${mode}_loop
 340 .align  16
 341 .L${mode}_loop:
 342         cmp     $len,$chunk             # ctr32 artefact
 343         cmova   $len,$chunk             # ctr32 artefact
 344         mov     $out,%r8                # save parameters
 345         mov     $inp,%r9
 346         mov     $len,%r10
 347         mov     $chunk,$len
 348         mov     $chunk,%r11
 349         test    \$0x0f,$out             # out_misaligned
 350         cmovnz  %rsp,$out
 351         test    \$0x0f,$inp             # inp_misaligned
 352         jz      .L${mode}_inp_aligned
 353         shr     \$3,$len
 354         .byte   0xf3,0x48,0xa5          # rep movsq
 355         sub     $chunk,$out
 356         mov     $chunk,$len
 357         mov     $out,$inp
 358 .L${mode}_inp_aligned:
 359         lea     -16($ctx),%rax          # ivp
 360         lea     16($ctx),%rbx           # key
 361         shr     \$4,$len
 362         .byte   0xf3,0x0f,0xa7,$opcode  # rep xcrypt*
 363 ___
 364 $code.=<<___                            if ($mode !~ /ecb|ctr/);
 365         movdqa  (%rax),%xmm0
 366         movdqa  %xmm0,-16($ctx)         # copy [or refresh] iv
 367 ___
 368 $code.=<<___                            if ($mode eq "ctr32");
 369         mov     -4($ctx),%eax           # pull 32-bit counter
 370         test    \$0xffff0000,%eax
 371         jnz     .L${mode}_no_carry
 372         bswap   %eax
 373         add     \$0x10000,%eax
 374         bswap   %eax
 375         mov     %eax,-4($ctx)
 376 .L${mode}_no_carry:
 377 ___
 378 $code.=<<___;
 379         mov     %r8,$out                # restore paramters
 380         mov     %r11,$chunk
 381         test    \$0x0f,$out
 382         jz      .L${mode}_out_aligned
 383         mov     $chunk,$len
 384         lea     (%rsp),$inp
 385         shr     \$3,$len
 386         .byte   0xf3,0x48,0xa5          # rep movsq
 387         sub     $chunk,$out
 388 .L${mode}_out_aligned:
 389         mov     %r9,$inp
 390         mov     %r10,$len
 391         add     $chunk,$out
 392         add     $chunk,$inp
 393         sub     $chunk,$len
 394         mov     \$$PADLOCK_CHUNK,$chunk
 395 ___
 396                                         if (!$PADLOCK_PREFETCH{$mode}) {
 397 $code.=<<___;
 398         jnz     .L${mode}_loop
 399 ___
 400                                         } else {
 401 $code.=<<___;
 402         jz      .L${mode}_break
 403         cmp     $chunk,$len
 404         jae     .L${mode}_loop
 405 ___
 406 $code.=<<___                            if ($mode eq "ctr32");
 407         mov     $len,$chunk
 408         mov     $inp,%rax               # check if prefetch crosses page
 409         cmp     %rsp,%rbp
 410         cmove   $out,%rax
 411         add     $len,%rax
 412         neg     %rax
 413         and     \$0xfff,%rax            # distance to page boundary
 414         cmp     \$$PADLOCK_PREFETCH{$mode},%rax
 415         mov     \$-$PADLOCK_PREFETCH{$mode},%rax
 416         cmovae  $chunk,%rax
 417         and     %rax,$chunk
 418         jnz     .L${mode}_loop
 419 ___
 420 $code.=<<___;
 421 .L${mode}_unaligned_tail:
 422         xor     %eax,%eax
 423         cmp     %rsp,%rbp
 424         cmove   $len,%rax
 425         mov     $out,%r8                # save parameters
 426         mov     $len,$chunk
 427         sub     %rax,%rsp               # alloca
 428         shr     \$3,$len
 429         lea     (%rsp),$out
 430         .byte   0xf3,0x48,0xa5          # rep movsq
 431         mov     %rsp,$inp
 432         mov     %r8, $out               # restore parameters
 433         mov     $chunk,$len
 434         jmp     .L${mode}_loop
 435 .align  16
 436 .L${mode}_break:
 437 ___
 438                                         }
 439 $code.=<<___;
 440         cmp     %rbp,%rsp
 441         je      .L${mode}_done
 442
 443         pxor    %xmm0,%xmm0
 444         lea     (%rsp),%rax
 445 .L${mode}_bzero:
 446         movaps  %xmm0,(%rax)
 447         lea     16(%rax),%rax
 448         cmp     %rax,%rbp
 449         ja      .L${mode}_bzero
 450
 451 .L${mode}_done:
 452         lea     (%rbp),%rsp
 453         jmp     .L${mode}_exit
 454
 455 .align  16
 456 .L${mode}_aligned:
 457 ___
 458 $code.=<<___                            if ($mode eq "ctr32");
 459         mov     -4($ctx),%eax           # pull 32-bit counter
 460         bswap   %eax
 461         neg     %eax
 462         and     \$0xffff,%eax
 463         mov     \$`16*0x10000`,$chunk
 464         shl     \$4,%eax
 465         cmovz   $chunk,%rax
 466         cmp     %rax,$len
 467         cmova   %rax,$chunk             # don't let counter cross 2^16
 468         cmovbe  $len,$chunk
 469         jbe     .L${mode}_aligned_skip
 470
 471 .L${mode}_aligned_loop:
 472         mov     $len,%r10               # save parameters
 473         mov     $chunk,$len
 474         mov     $chunk,%r11
 475
 476         lea     -16($ctx),%rax          # ivp
 477         lea     16($ctx),%rbx           # key
 478         shr     \$4,$len                # len/=AES_BLOCK_SIZE
 479         .byte   0xf3,0x0f,0xa7,$opcode  # rep xcrypt*
 480
 481         mov     -4($ctx),%eax           # pull 32-bit counter
 482         bswap   %eax
 483         add     \$0x10000,%eax
 484         bswap   %eax
 485         mov     %eax,-4($ctx)
 486
 487         mov     %r10,$len               # restore paramters
 488         sub     %r11,$len
 489         mov     \$`16*0x10000`,$chunk
 490         jz      .L${mode}_exit
 491         cmp     $chunk,$len
 492         jae     .L${mode}_aligned_loop
 493
 494 .L${mode}_aligned_skip:
 495 ___
 496 $code.=<<___                            if ($PADLOCK_PREFETCH{$mode});
 497         lea     ($inp,$len),%rbp
 498         neg     %rbp
 499         and     \$0xfff,%rbp            # distance to page boundary
 500         xor     %eax,%eax
 501         cmp     \$$PADLOCK_PREFETCH{$mode},%rbp
 502         mov     \$$PADLOCK_PREFETCH{$mode}-1,%rbp
 503         cmovae  %rax,%rbp
 504         and     $len,%rbp               # remainder
 505         sub     %rbp,$len
 506         jz      .L${mode}_aligned_tail
 507 ___
 508 $code.=<<___;
 509         lea     -16($ctx),%rax          # ivp
 510         lea     16($ctx),%rbx           # key
 511         shr     \$4,$len                # len/=AES_BLOCK_SIZE
 512         .byte   0xf3,0x0f,0xa7,$opcode  # rep xcrypt*
 513 ___
 514 $code.=<<___                            if ($mode !~ /ecb|ctr/);
 515         movdqa  (%rax),%xmm0
 516         movdqa  %xmm0,-16($ctx)         # copy [or refresh] iv
 517 ___
 518 $code.=<<___                            if ($PADLOCK_PREFETCH{$mode});
 519         test    %rbp,%rbp               # check remainder
 520         jz      .L${mode}_exit
 521
 522 .L${mode}_aligned_tail:
 523         mov     $out,%r8
 524         mov     %rbp,$chunk
 525         mov     %rbp,$len
 526         lea     (%rsp),%rbp
 527         sub     $len,%rsp
 528         shr     \$3,$len
 529         lea     (%rsp),$out
 530         .byte   0xf3,0x48,0xa5          # rep movsq
 531         lea     (%r8),$out
 532         lea     (%rsp),$inp
 533         mov     $chunk,$len
 534         jmp     .L${mode}_loop
 535 ___
 536 $code.=<<___;
 537 .L${mode}_exit:
 538         mov     \$1,%eax
 539         lea     8(%rsp),%rsp
 540 .L${mode}_abort:
 541         pop     %rbx
 542         pop     %rbp
 543         ret
 544 .size   padlock_${mode}_encrypt,.-padlock_${mode}_encrypt
 545 ___
 546 }
 547
 548 &generate_mode("ecb",0xc8);
 549 &generate_mode("cbc",0xd0);
 550 #&generate_mode("cfb",0xe0);
 551 #&generate_mode("ofb",0xe8);
 552 #&generate_mode("ctr32",0xd8);  # all 64-bit CPUs have working CTR...
 553
 554 $code.=<<___;
 555 .asciz  "VIA Padlock x86_64 module, CRYPTOGAMS by <appro\@openssl.org>"
 556 .align  16
 557 .data
 558 .align  8
 559 .Lpadlock_saved_context:
 560         .quad   0
 561 ___
 562 $code =~ s/\`([^\`]*)\`/eval($1)/gem;
 563
 564 print $code;
 565
 566 close STDOUT;