engines/asm/e_padlock-x86.pl

   1 #! /usr/bin/env perl
   2 # Copyright 2011-2016 The OpenSSL Project Authors. All Rights Reserved.
   3 #
   4 # Licensed under the OpenSSL license (the "License").  You may not use
   5 # this file except in compliance with the License.  You can obtain a copy
   6 # in the file LICENSE in the source distribution or at
   7 # https://www.openssl.org/source/license.html
   8
   9
  10 # ====================================================================
  11 # Written by Andy Polyakov <appro@openssl.org> for the OpenSSL
  12 # project. The module is, however, dual licensed under OpenSSL and
  13 # CRYPTOGAMS licenses depending on where you obtain it. For further
  14 # details see http://www.openssl.org/~appro/cryptogams/.
  15 # ====================================================================
  16
  17 # September 2011
  18 #
  19 # Assembler helpers for Padlock engine. Compared to original engine
  20 # version relying on inline assembler and compiled with gcc 3.4.6 it
  21 # was measured to provide ~100% improvement on misaligned data in ECB
  22 # mode and ~75% in CBC mode. For aligned data improvement can be
  23 # observed for short inputs only, e.g. 45% for 64-byte messages in
  24 # ECB mode, 20% in CBC. Difference in performance for aligned vs.
  25 # misaligned data depends on misalignment and is either ~1.8x or 2.9x.
  26 # These are approximately same factors as for hardware support, so
  27 # there is little reason to rely on the latter. On the contrary, it
  28 # might actually hurt performance in mixture of aligned and misaligned
  29 # buffers, because a) if you choose to flip 'align' flag in control
  30 # word on per-buffer basis, then you'd have to reload key context,
  31 # which incurs penalty; b) if you choose to set 'align' flag
  32 # permanently, it limits performance even for aligned data to ~1/2.
  33 # All above mentioned results were collected on 1.5GHz C7. Nano on the
  34 # other hand handles unaligned data more gracefully. Depending on
  35 # algorithm and how unaligned data is, hardware can be up to 70% more
  36 # efficient than below software alignment procedures, nor does 'align'
  37 # flag have affect on aligned performance [if has any meaning at all].
  38 # Therefore suggestion is to unconditionally set 'align' flag on Nano
  39 # for optimal performance.
  40
  41 $0 =~ m/(.*[\/\\])[^\/\\]+$/; $dir=$1;
  42 push(@INC,"${dir}","${dir}../../crypto/perlasm");
  43 require "x86asm.pl";
  44
  45 $output=pop;
  46 open STDOUT,">$output";
  47
  48 &asm_init($ARGV[0],$0);
  49
  50 %PADLOCK_PREFETCH=(ecb=>128, cbc=>64);  # prefetch errata
  51 $PADLOCK_CHUNK=512;     # Must be a power of 2 larger than 16
  52
  53 $ctx="edx";
  54 $out="edi";
  55 $inp="esi";
  56 $len="ecx";
  57 $chunk="ebx";
  58
  59 &function_begin_B("padlock_capability");
  60         &push   ("ebx");
  61         &pushf  ();
  62         &pop    ("eax");
  63         &mov    ("ecx","eax");
  64         &xor    ("eax",1<<21);
  65         &push   ("eax");
  66         &popf   ();
  67         &pushf  ();
  68         &pop    ("eax");
  69         &xor    ("ecx","eax");
  70         &xor    ("eax","eax");
  71         &bt     ("ecx",21);
  72         &jnc    (&label("noluck"));
  73         &cpuid  ();
  74         &xor    ("eax","eax");
  75         &cmp    ("ebx","0x".unpack("H*",'tneC'));
  76         &jne    (&label("noluck"));
  77         &cmp    ("edx","0x".unpack("H*",'Hrua'));
  78         &jne    (&label("noluck"));
  79         &cmp    ("ecx","0x".unpack("H*",'slua'));
  80         &jne    (&label("noluck"));
  81         &mov    ("eax",0xC0000000);
  82         &cpuid  ();
  83         &mov    ("edx","eax");
  84         &xor    ("eax","eax");
  85         &cmp    ("edx",0xC0000001);
  86         &jb     (&label("noluck"));
  87         &mov    ("eax",1);
  88         &cpuid  ();
  89         &or     ("eax",0x0f);
  90         &xor    ("ebx","ebx");
  91         &and    ("eax",0x0fff);
  92         &cmp    ("eax",0x06ff);         # check for Nano
  93         &sete   ("bl");
  94         &mov    ("eax",0xC0000001);
  95         &push   ("ebx");
  96         &cpuid  ();
  97         &pop    ("ebx");
  98         &mov    ("eax","edx");
  99         &shl    ("ebx",4);              # bit#4 denotes Nano
 100         &and    ("eax",0xffffffef);
 101         &or     ("eax","ebx")
 102 &set_label("noluck");
 103         &pop    ("ebx");
 104         &ret    ();
 105 &function_end_B("padlock_capability")
 106
 107 &function_begin_B("padlock_key_bswap");
 108         &mov    ("edx",&wparam(0));
 109         &mov    ("ecx",&DWP(240,"edx"));
 110 &set_label("bswap_loop");
 111         &mov    ("eax",&DWP(0,"edx"));
 112         &bswap  ("eax");
 113         &mov    (&DWP(0,"edx"),"eax");
 114         &lea    ("edx",&DWP(4,"edx"));
 115         &sub    ("ecx",1);
 116         &jnz    (&label("bswap_loop"));
 117         &ret    ();
 118 &function_end_B("padlock_key_bswap");
 119
 120 # This is heuristic key context tracing. At first one
 121 # believes that one should use atomic swap instructions,
 122 # but it's not actually necessary. Point is that if
 123 # padlock_saved_context was changed by another thread
 124 # after we've read it and before we compare it with ctx,
 125 # our key *shall* be reloaded upon thread context switch
 126 # and we are therefore set in either case...
 127 &static_label("padlock_saved_context");
 128
 129 &function_begin_B("padlock_verify_context");
 130         &mov    ($ctx,&wparam(0));
 131         &lea    ("eax",($::win32 or $::coff) ? &DWP(&label("padlock_saved_context")) :
 132                        &DWP(&label("padlock_saved_context")."-".&label("verify_pic_point")));
 133         &pushf  ();
 134         &call   ("_padlock_verify_ctx");
 135 &set_label("verify_pic_point");
 136         &lea    ("esp",&DWP(4,"esp"));
 137         &ret    ();
 138 &function_end_B("padlock_verify_context");
 139
 140 &function_begin_B("_padlock_verify_ctx");
 141         &add    ("eax",&DWP(0,"esp")) if(!($::win32 or $::coff));# &padlock_saved_context
 142         &bt     (&DWP(4,"esp"),30);             # eflags
 143         &jnc    (&label("verified"));
 144         &cmp    ($ctx,&DWP(0,"eax"));
 145         &je     (&label("verified"));
 146         &pushf  ();
 147         &popf   ();
 148 &set_label("verified");
 149         &mov    (&DWP(0,"eax"),$ctx);
 150         &ret    ();
 151 &function_end_B("_padlock_verify_ctx");
 152
 153 &function_begin_B("padlock_reload_key");
 154         &pushf  ();
 155         &popf   ();
 156         &ret    ();
 157 &function_end_B("padlock_reload_key");
 158
 159 &function_begin_B("padlock_aes_block");
 160         &push   ("edi");
 161         &push   ("esi");
 162         &push   ("ebx");
 163         &mov    ($out,&wparam(0));              # must be 16-byte aligned
 164         &mov    ($inp,&wparam(1));              # must be 16-byte aligned
 165         &mov    ($ctx,&wparam(2));
 166         &mov    ($len,1);
 167         &lea    ("ebx",&DWP(32,$ctx));          # key
 168         &lea    ($ctx,&DWP(16,$ctx));           # control word
 169         &data_byte(0xf3,0x0f,0xa7,0xc8);        # rep xcryptecb
 170         &pop    ("ebx");
 171         &pop    ("esi");
 172         &pop    ("edi");
 173         &ret    ();
 174 &function_end_B("padlock_aes_block");
 175
 176 sub generate_mode {
 177 my ($mode,$opcode) = @_;
 178 # int padlock_$mode_encrypt(void *out, const void *inp,
 179 #               struct padlock_cipher_data *ctx, size_t len);
 180 &function_begin("padlock_${mode}_encrypt");
 181         &mov    ($out,&wparam(0));
 182         &mov    ($inp,&wparam(1));
 183         &mov    ($ctx,&wparam(2));
 184         &mov    ($len,&wparam(3));
 185         &test   ($ctx,15);
 186         &jnz    (&label("${mode}_abort"));
 187         &test   ($len,15);
 188         &jnz    (&label("${mode}_abort"));
 189         &lea    ("eax",($::win32 or $::coff) ? &DWP(&label("padlock_saved_context")) :
 190                        &DWP(&label("padlock_saved_context")."-".&label("${mode}_pic_point")));
 191         &pushf  ();
 192         &cld    ();
 193         &call   ("_padlock_verify_ctx");
 194 &set_label("${mode}_pic_point");
 195         &lea    ($ctx,&DWP(16,$ctx));   # control word
 196         &xor    ("eax","eax");
 197                                         if ($mode eq "ctr32") {
 198         &movq   ("mm0",&QWP(-16,$ctx)); # load [upper part of] counter
 199                                         } else {
 200         &xor    ("ebx","ebx");
 201         &test   (&DWP(0,$ctx),1<<5);    # align bit in control word
 202         &jnz    (&label("${mode}_aligned"));
 203         &test   ($out,0x0f);
 204         &setz   ("al");                 # !out_misaligned
 205         &test   ($inp,0x0f);
 206         &setz   ("bl");                 # !inp_misaligned
 207         &test   ("eax","ebx");
 208         &jnz    (&label("${mode}_aligned"));
 209         &neg    ("eax");
 210                                         }
 211         &mov    ($chunk,$PADLOCK_CHUNK);
 212         &not    ("eax");                # out_misaligned?-1:0
 213         &lea    ("ebp",&DWP(-24,"esp"));
 214         &cmp    ($len,$chunk);
 215         &cmovc  ($chunk,$len);          # chunk=len>PADLOCK_CHUNK?PADLOCK_CHUNK:len
 216         &and    ("eax",$chunk);         # out_misaligned?chunk:0
 217         &mov    ($chunk,$len);
 218         &neg    ("eax");
 219         &and    ($chunk,$PADLOCK_CHUNK-1);      # chunk=len%PADLOCK_CHUNK
 220         &lea    ("esp",&DWP(0,"eax","ebp"));    # alloca
 221         &mov    ("eax",$PADLOCK_CHUNK);
 222         &cmovz  ($chunk,"eax");                 # chunk=chunk?:PADLOCK_CHUNK
 223         &mov    ("eax","ebp");
 224         &and    ("ebp",-16);
 225         &and    ("esp",-16);
 226         &mov    (&DWP(16,"ebp"),"eax");
 227     if ($PADLOCK_PREFETCH{$mode}) {
 228         &cmp    ($len,$chunk);
 229         &ja     (&label("${mode}_loop"));
 230         &mov    ("eax",$inp);           # check if prefetch crosses page
 231         &cmp    ("ebp","esp");
 232         &cmove  ("eax",$out);
 233         &add    ("eax",$len);
 234         &neg    ("eax");
 235         &and    ("eax",0xfff);          # distance to page boundary
 236         &cmp    ("eax",$PADLOCK_PREFETCH{$mode});
 237         &mov    ("eax",-$PADLOCK_PREFETCH{$mode});
 238         &cmovae ("eax",$chunk);         # mask=distance<prefetch?-prefetch:-1
 239         &and    ($chunk,"eax");
 240         &jz     (&label("${mode}_unaligned_tail"));
 241     }
 242         &jmp    (&label("${mode}_loop"));
 243
 244 &set_label("${mode}_loop",16);
 245         &mov    (&DWP(0,"ebp"),$out);           # save parameters
 246         &mov    (&DWP(4,"ebp"),$inp);
 247         &mov    (&DWP(8,"ebp"),$len);
 248         &mov    ($len,$chunk);
 249         &mov    (&DWP(12,"ebp"),$chunk);        # chunk
 250                                                 if ($mode eq "ctr32") {
 251         &mov    ("ecx",&DWP(-4,$ctx));
 252         &xor    ($out,$out);
 253         &mov    ("eax",&DWP(-8,$ctx));          # borrow $len
 254 &set_label("${mode}_prepare");
 255         &mov    (&DWP(12,"esp",$out),"ecx");
 256         &bswap  ("ecx");
 257         &movq   (&QWP(0,"esp",$out),"mm0");
 258         &inc    ("ecx");
 259         &mov    (&DWP(8,"esp",$out),"eax");
 260         &bswap  ("ecx");
 261         &lea    ($out,&DWP(16,$out));
 262         &cmp    ($out,$chunk);
 263         &jb     (&label("${mode}_prepare"));
 264
 265         &mov    (&DWP(-4,$ctx),"ecx");
 266         &lea    ($inp,&DWP(0,"esp"));
 267         &lea    ($out,&DWP(0,"esp"));
 268         &mov    ($len,$chunk);
 269                                                 } else {
 270         &test   ($out,0x0f);                    # out_misaligned
 271         &cmovnz ($out,"esp");
 272         &test   ($inp,0x0f);                    # inp_misaligned
 273         &jz     (&label("${mode}_inp_aligned"));
 274         &shr    ($len,2);
 275         &data_byte(0xf3,0xa5);                  # rep movsl
 276         &sub    ($out,$chunk);
 277         &mov    ($len,$chunk);
 278         &mov    ($inp,$out);
 279 &set_label("${mode}_inp_aligned");
 280                                                 }
 281         &lea    ("eax",&DWP(-16,$ctx));         # ivp
 282         &lea    ("ebx",&DWP(16,$ctx));          # key
 283         &shr    ($len,4);                       # len/=AES_BLOCK_SIZE
 284         &data_byte(0xf3,0x0f,0xa7,$opcode);     # rep xcrypt*
 285                                                 if ($mode !~ /ecb|ctr/) {
 286         &movaps ("xmm0",&QWP(0,"eax"));
 287         &movaps (&QWP(-16,$ctx),"xmm0");        # copy [or refresh] iv
 288                                                 }
 289         &mov    ($out,&DWP(0,"ebp"));           # restore parameters
 290         &mov    ($chunk,&DWP(12,"ebp"));
 291                                                 if ($mode eq "ctr32") {
 292         &mov    ($inp,&DWP(4,"ebp"));
 293         &xor    ($len,$len);
 294 &set_label("${mode}_xor");
 295         &movups ("xmm1",&QWP(0,$inp,$len));
 296         &lea    ($len,&DWP(16,$len));
 297         &pxor   ("xmm1",&QWP(-16,"esp",$len));
 298         &movups (&QWP(-16,$out,$len),"xmm1");
 299         &cmp    ($len,$chunk);
 300         &jb     (&label("${mode}_xor"));
 301                                                 } else {
 302         &test   ($out,0x0f);
 303         &jz     (&label("${mode}_out_aligned"));
 304         &mov    ($len,$chunk);
 305         &lea    ($inp,&DWP(0,"esp"));
 306         &shr    ($len,2);
 307         &data_byte(0xf3,0xa5);                  # rep movsl
 308         &sub    ($out,$chunk);
 309 &set_label("${mode}_out_aligned");
 310         &mov    ($inp,&DWP(4,"ebp"));
 311                                                 }
 312         &mov    ($len,&DWP(8,"ebp"));
 313         &add    ($out,$chunk);
 314         &add    ($inp,$chunk);
 315         &sub    ($len,$chunk);
 316         &mov    ($chunk,$PADLOCK_CHUNK);
 317     if (!$PADLOCK_PREFETCH{$mode}) {
 318         &jnz    (&label("${mode}_loop"));
 319     } else {
 320         &jz     (&label("${mode}_break"));
 321         &cmp    ($len,$chunk);
 322         &jae    (&label("${mode}_loop"));
 323
 324 &set_label("${mode}_unaligned_tail");
 325         &xor    ("eax","eax");
 326         &cmp    ("esp","ebp");
 327         &cmove  ("eax",$len);
 328         &sub    ("esp","eax");                  # alloca
 329         &mov    ("eax", $out);                  # save parameters
 330         &mov    ($chunk,$len);
 331         &shr    ($len,2);
 332         &lea    ($out,&DWP(0,"esp"));
 333         &data_byte(0xf3,0xa5);                  # rep movsl
 334         &mov    ($inp,"esp");
 335         &mov    ($out,"eax");                   # restore parameters
 336         &mov    ($len,$chunk);
 337         &jmp    (&label("${mode}_loop"));
 338
 339 &set_label("${mode}_break",16);
 340     }
 341                                                 if ($mode ne "ctr32") {
 342         &cmp    ("esp","ebp");
 343         &je     (&label("${mode}_done"));
 344                                                 }
 345         &pxor   ("xmm0","xmm0");
 346         &lea    ("eax",&DWP(0,"esp"));
 347 &set_label("${mode}_bzero");
 348         &movaps (&QWP(0,"eax"),"xmm0");
 349         &lea    ("eax",&DWP(16,"eax"));
 350         &cmp    ("ebp","eax");
 351         &ja     (&label("${mode}_bzero"));
 352
 353 &set_label("${mode}_done");
 354         &mov    ("ebp",&DWP(16,"ebp"));
 355         &lea    ("esp",&DWP(24,"ebp"));
 356                                                 if ($mode ne "ctr32") {
 357         &jmp    (&label("${mode}_exit"));
 358
 359 &set_label("${mode}_aligned",16);
 360     if ($PADLOCK_PREFETCH{$mode}) {
 361         &lea    ("ebp",&DWP(0,$inp,$len));
 362         &neg    ("ebp");
 363         &and    ("ebp",0xfff);                  # distance to page boundary
 364         &xor    ("eax","eax");
 365         &cmp    ("ebp",$PADLOCK_PREFETCH{$mode});
 366         &mov    ("ebp",$PADLOCK_PREFETCH{$mode}-1);
 367         &cmovae ("ebp","eax");
 368         &and    ("ebp",$len);                   # remainder
 369         &sub    ($len,"ebp");
 370         &jz     (&label("${mode}_aligned_tail"));
 371     }
 372         &lea    ("eax",&DWP(-16,$ctx));         # ivp
 373         &lea    ("ebx",&DWP(16,$ctx));          # key
 374         &shr    ($len,4);                       # len/=AES_BLOCK_SIZE
 375         &data_byte(0xf3,0x0f,0xa7,$opcode);     # rep xcrypt*
 376                                                 if ($mode ne "ecb") {
 377         &movaps ("xmm0",&QWP(0,"eax"));
 378         &movaps (&QWP(-16,$ctx),"xmm0");        # copy [or refresh] iv
 379                                                 }
 380     if ($PADLOCK_PREFETCH{$mode}) {
 381         &test   ("ebp","ebp");
 382         &jz     (&label("${mode}_exit"));
 383
 384 &set_label("${mode}_aligned_tail");
 385         &mov    ($len,"ebp");
 386         &lea    ("ebp",&DWP(-24,"esp"));
 387         &mov    ("esp","ebp");
 388         &mov    ("eax","ebp");
 389         &sub    ("esp",$len);
 390         &and    ("ebp",-16);
 391         &and    ("esp",-16);
 392         &mov    (&DWP(16,"ebp"),"eax");
 393         &mov    ("eax", $out);                  # save parameters
 394         &mov    ($chunk,$len);
 395         &shr    ($len,2);
 396         &lea    ($out,&DWP(0,"esp"));
 397         &data_byte(0xf3,0xa5);                  # rep movsl
 398         &mov    ($inp,"esp");
 399         &mov    ($out,"eax");                   # restore parameters
 400         &mov    ($len,$chunk);
 401         &jmp    (&label("${mode}_loop"));
 402     }
 403 &set_label("${mode}_exit");                     }
 404         &mov    ("eax",1);
 405         &lea    ("esp",&DWP(4,"esp"));          # popf
 406         &emms   ()                              if ($mode eq "ctr32");
 407 &set_label("${mode}_abort");
 408 &function_end("padlock_${mode}_encrypt");
 409 }
 410
 411 &generate_mode("ecb",0xc8);
 412 &generate_mode("cbc",0xd0);
 413 &generate_mode("cfb",0xe0);
 414 &generate_mode("ofb",0xe8);
 415 &generate_mode("ctr32",0xc8);   # yes, it implements own CTR with ECB opcode,
 416                                 # because hardware CTR was introduced later
 417                                 # and even has errata on certain C7 stepping.
 418                                 # own implementation *always* works, though
 419                                 # ~15% slower than dedicated hardware...
 420
 421 &function_begin_B("padlock_xstore");
 422         &push   ("edi");
 423         &mov    ("edi",&wparam(0));
 424         &mov    ("edx",&wparam(1));
 425         &data_byte(0x0f,0xa7,0xc0);             # xstore
 426         &pop    ("edi");
 427         &ret    ();
 428 &function_end_B("padlock_xstore");
 429
 430 &function_begin_B("_win32_segv_handler");
 431         &mov    ("eax",1);                      # ExceptionContinueSearch
 432         &mov    ("edx",&wparam(0));             # *ExceptionRecord
 433         &mov    ("ecx",&wparam(2));             # *ContextRecord
 434         &cmp    (&DWP(0,"edx"),0xC0000005)      # ExceptionRecord->ExceptionCode == STATUS_ACCESS_VIOLATION
 435         &jne    (&label("ret"));
 436         &add    (&DWP(184,"ecx"),4);            # skip over rep sha*
 437         &mov    ("eax",0);                      # ExceptionContinueExecution
 438 &set_label("ret");
 439         &ret    ();
 440 &function_end_B("_win32_segv_handler");
 441 &safeseh("_win32_segv_handler")                 if ($::win32);
 442
 443 &function_begin_B("padlock_sha1_oneshot");
 444         &push   ("edi");
 445         &push   ("esi");
 446         &xor    ("eax","eax");
 447         &mov    ("edi",&wparam(0));
 448         &mov    ("esi",&wparam(1));
 449         &mov    ("ecx",&wparam(2));
 450     if ($::win32 or $::coff) {
 451         &push   (&::islabel("_win32_segv_handler"));
 452         &data_byte(0x64,0xff,0x30);             # push  %fs:(%eax)
 453         &data_byte(0x64,0x89,0x20);             # mov   %esp,%fs:(%eax)
 454     }
 455         &mov    ("edx","esp");                  # put aside %esp
 456         &add    ("esp",-128);                   # 32 is enough but spec says 128
 457         &movups ("xmm0",&QWP(0,"edi"));         # copy-in context
 458         &and    ("esp",-16);
 459         &mov    ("eax",&DWP(16,"edi"));
 460         &movaps (&QWP(0,"esp"),"xmm0");
 461         &mov    ("edi","esp");
 462         &mov    (&DWP(16,"esp"),"eax");
 463         &xor    ("eax","eax");
 464         &data_byte(0xf3,0x0f,0xa6,0xc8);        # rep xsha1
 465         &movaps ("xmm0",&QWP(0,"esp"));
 466         &mov    ("eax",&DWP(16,"esp"));
 467         &mov    ("esp","edx");                  # restore %esp
 468     if ($::win32 or $::coff) {
 469         &data_byte(0x64,0x8f,0x05,0,0,0,0);     # pop   %fs:0
 470         &lea    ("esp",&DWP(4,"esp"));
 471     }
 472         &mov    ("edi",&wparam(0));
 473         &movups (&QWP(0,"edi"),"xmm0");         # copy-out context
 474         &mov    (&DWP(16,"edi"),"eax");
 475         &pop    ("esi");
 476         &pop    ("edi");
 477         &ret    ();
 478 &function_end_B("padlock_sha1_oneshot");
 479
 480 &function_begin_B("padlock_sha1_blocks");
 481         &push   ("edi");
 482         &push   ("esi");
 483         &mov    ("edi",&wparam(0));
 484         &mov    ("esi",&wparam(1));
 485         &mov    ("edx","esp");                  # put aside %esp
 486         &mov    ("ecx",&wparam(2));
 487         &add    ("esp",-128);
 488         &movups ("xmm0",&QWP(0,"edi"));         # copy-in context
 489         &and    ("esp",-16);
 490         &mov    ("eax",&DWP(16,"edi"));
 491         &movaps (&QWP(0,"esp"),"xmm0");
 492         &mov    ("edi","esp");
 493         &mov    (&DWP(16,"esp"),"eax");
 494         &mov    ("eax",-1);
 495         &data_byte(0xf3,0x0f,0xa6,0xc8);        # rep xsha1
 496         &movaps ("xmm0",&QWP(0,"esp"));
 497         &mov    ("eax",&DWP(16,"esp"));
 498         &mov    ("esp","edx");                  # restore %esp
 499         &mov    ("edi",&wparam(0));
 500         &movups (&QWP(0,"edi"),"xmm0");         # copy-out context
 501         &mov    (&DWP(16,"edi"),"eax");
 502         &pop    ("esi");
 503         &pop    ("edi");
 504         &ret    ();
 505 &function_end_B("padlock_sha1_blocks");
 506
 507 &function_begin_B("padlock_sha256_oneshot");
 508         &push   ("edi");
 509         &push   ("esi");
 510         &xor    ("eax","eax");
 511         &mov    ("edi",&wparam(0));
 512         &mov    ("esi",&wparam(1));
 513         &mov    ("ecx",&wparam(2));
 514     if ($::win32 or $::coff) {
 515         &push   (&::islabel("_win32_segv_handler"));
 516         &data_byte(0x64,0xff,0x30);             # push  %fs:(%eax)
 517         &data_byte(0x64,0x89,0x20);             # mov   %esp,%fs:(%eax)
 518     }
 519         &mov    ("edx","esp");                  # put aside %esp
 520         &add    ("esp",-128);
 521         &movups ("xmm0",&QWP(0,"edi"));         # copy-in context
 522         &and    ("esp",-16);
 523         &movups ("xmm1",&QWP(16,"edi"));
 524         &movaps (&QWP(0,"esp"),"xmm0");
 525         &mov    ("edi","esp");
 526         &movaps (&QWP(16,"esp"),"xmm1");
 527         &xor    ("eax","eax");
 528         &data_byte(0xf3,0x0f,0xa6,0xd0);        # rep xsha256
 529         &movaps ("xmm0",&QWP(0,"esp"));
 530         &movaps ("xmm1",&QWP(16,"esp"));
 531         &mov    ("esp","edx");                  # restore %esp
 532     if ($::win32 or $::coff) {
 533         &data_byte(0x64,0x8f,0x05,0,0,0,0);     # pop   %fs:0
 534         &lea    ("esp",&DWP(4,"esp"));
 535     }
 536         &mov    ("edi",&wparam(0));
 537         &movups (&QWP(0,"edi"),"xmm0");         # copy-out context
 538         &movups (&QWP(16,"edi"),"xmm1");
 539         &pop    ("esi");
 540         &pop    ("edi");
 541         &ret    ();
 542 &function_end_B("padlock_sha256_oneshot");
 543
 544 &function_begin_B("padlock_sha256_blocks");
 545         &push   ("edi");
 546         &push   ("esi");
 547         &mov    ("edi",&wparam(0));
 548         &mov    ("esi",&wparam(1));
 549         &mov    ("ecx",&wparam(2));
 550         &mov    ("edx","esp");                  # put aside %esp
 551         &add    ("esp",-128);
 552         &movups ("xmm0",&QWP(0,"edi"));         # copy-in context
 553         &and    ("esp",-16);
 554         &movups ("xmm1",&QWP(16,"edi"));
 555         &movaps (&QWP(0,"esp"),"xmm0");
 556         &mov    ("edi","esp");
 557         &movaps (&QWP(16,"esp"),"xmm1");
 558         &mov    ("eax",-1);
 559         &data_byte(0xf3,0x0f,0xa6,0xd0);        # rep xsha256
 560         &movaps ("xmm0",&QWP(0,"esp"));
 561         &movaps ("xmm1",&QWP(16,"esp"));
 562         &mov    ("esp","edx");                  # restore %esp
 563         &mov    ("edi",&wparam(0));
 564         &movups (&QWP(0,"edi"),"xmm0");         # copy-out context
 565         &movups (&QWP(16,"edi"),"xmm1");
 566         &pop    ("esi");
 567         &pop    ("edi");
 568         &ret    ();
 569 &function_end_B("padlock_sha256_blocks");
 570
 571 &function_begin_B("padlock_sha512_blocks");
 572         &push   ("edi");
 573         &push   ("esi");
 574         &mov    ("edi",&wparam(0));
 575         &mov    ("esi",&wparam(1));
 576         &mov    ("ecx",&wparam(2));
 577         &mov    ("edx","esp");                  # put aside %esp
 578         &add    ("esp",-128);
 579         &movups ("xmm0",&QWP(0,"edi"));         # copy-in context
 580         &and    ("esp",-16);
 581         &movups ("xmm1",&QWP(16,"edi"));
 582         &movups ("xmm2",&QWP(32,"edi"));
 583         &movups ("xmm3",&QWP(48,"edi"));
 584         &movaps (&QWP(0,"esp"),"xmm0");
 585         &mov    ("edi","esp");
 586         &movaps (&QWP(16,"esp"),"xmm1");
 587         &movaps (&QWP(32,"esp"),"xmm2");
 588         &movaps (&QWP(48,"esp"),"xmm3");
 589         &data_byte(0xf3,0x0f,0xa6,0xe0);        # rep xsha512
 590         &movaps ("xmm0",&QWP(0,"esp"));
 591         &movaps ("xmm1",&QWP(16,"esp"));
 592         &movaps ("xmm2",&QWP(32,"esp"));
 593         &movaps ("xmm3",&QWP(48,"esp"));
 594         &mov    ("esp","edx");                  # restore %esp
 595         &mov    ("edi",&wparam(0));
 596         &movups (&QWP(0,"edi"),"xmm0");         # copy-out context
 597         &movups (&QWP(16,"edi"),"xmm1");
 598         &movups (&QWP(32,"edi"),"xmm2");
 599         &movups (&QWP(48,"edi"),"xmm3");
 600         &pop    ("esi");
 601         &pop    ("edi");
 602         &ret    ();
 603 &function_end_B("padlock_sha512_blocks");
 604
 605 &asciz  ("VIA Padlock x86 module, CRYPTOGAMS by <appro\@openssl.org>");
 606 &align  (16);
 607
 608 &dataseg();
 609 # Essentially this variable belongs in thread local storage.
 610 # Having this variable global on the other hand can only cause
 611 # few bogus key reloads [if any at all on signle-CPU system],
 612 # so we accept the penalty...
 613 &set_label("padlock_saved_context",4);
 614 &data_word(0);
 615
 616 &asm_finish();
 617
 618 close STDOUT;