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