crypto/whrlpool/asm/wp-x86_64.pl

   1 #!/usr/bin/env perl
   2 #
   3 # ====================================================================
   4 # Written by Andy Polyakov <appro@fy.chalmers.se> for the OpenSSL
   5 # project. Rights for redistribution and usage in source and binary
   6 # forms are granted according to the OpenSSL license.
   7 # ====================================================================
   8 #
   9 # whirlpool_block for x86_64.
  10 #
  11 # 2500 cycles per 64-byte input block on AMD64, which is *identical*
  12 # to 32-bit MMX version executed on same CPU. So why did I bother?
  13 # Well, it's faster than gcc 3.3.2 generated code by over 50%, and
  14 # over 80% faster than PathScale 1.4, an "ambitious" commercial
  15 # compiler. Furthermore it surpasses gcc 3.4.3 by 170% and Sun Studio
  16 # 10 - by 360%[!]... What is it with x86_64 compilers? It's not the
  17 # first example when they fail to generate more optimal code, when
  18 # I believe they had *all* chances to...
  19 #
  20 # Note that register and stack frame layout are virtually identical
  21 # to 32-bit MMX version, except that %r8-15 are used instead of
  22 # %mm0-8. You can even notice that K[i] and S[i] are loaded to
  23 # %eax:%ebx as pair of 32-bit values and not as single 64-bit one.
  24 # This is done in order to avoid 64-bit shift penalties on Intel
  25 # EM64T core. Speaking of which! I bet it's possible to improve
  26 # Opteron performance by compressing the table to 2KB and replacing
  27 # unaligned references with complementary rotations [which would
  28 # incidentally replace lea instructions], but it would definitely
  29 # just "kill" EM64T, because it has only 1 shifter/rotator [against
  30 # 3 on Opteron] and which is *unacceptably* slow with 64-bit
  31 # operand.
  32
  33 $flavour = shift;
  34 $output  = shift;
  35 if ($flavour =~ /\./) { $output = $flavour; undef $flavour; }
  36
  37 $win64=0; $win64=1 if ($flavour =~ /[nm]asm|mingw64/ || $output =~ /\.asm$/);
  38
  39 $0 =~ m/(.*[\/\\])[^\/\\]+$/; my $dir=$1; my $xlate;
  40 ( $xlate="${dir}x86_64-xlate.pl" and -f $xlate ) or
  41 ( $xlate="${dir}../../perlasm/x86_64-xlate.pl" and -f $xlate) or
  42 die "can't locate x86_64-xlate.pl";
  43
  44 open STDOUT,"| \"$^X\" $xlate $flavour $output";
  45
  46 sub L() { $code.=".byte ".join(',',@_)."\n"; }
  47 sub LL(){ $code.=".byte ".join(',',@_).",".join(',',@_)."\n"; }
  48
  49 @mm=("%r8","%r9","%r10","%r11","%r12","%r13","%r14","%r15");
  50
  51 $func="whirlpool_block";
  52 $table=".Ltable";
  53
  54 $code=<<___;
  55 .text
  56
  57 .globl  $func
  58 .type   $func,\@function,3
  59 .align  16
  60 $func:
  61         push    %rbx
  62         push    %rbp
  63         push    %r12
  64         push    %r13
  65         push    %r14
  66         push    %r15
  67
  68         mov     %rsp,%r11
  69         sub     \$128+40,%rsp
  70         and     \$-64,%rsp
  71
  72         lea     128(%rsp),%r10
  73         mov     %rdi,0(%r10)            # save parameter block
  74         mov     %rsi,8(%r10)
  75         mov     %rdx,16(%r10)
  76         mov     %r11,32(%r10)           # saved stack pointer
  77 .Lprologue:
  78
  79         mov     %r10,%rbx
  80         lea     $table(%rip),%rbp
  81
  82         xor     %rcx,%rcx
  83         xor     %rdx,%rdx
  84 ___
  85 for($i=0;$i<8;$i++) { $code.="mov $i*8(%rdi),@mm[$i]\n"; }      # L=H
  86 $code.=".Louterloop:\n";
  87 for($i=0;$i<8;$i++) { $code.="mov @mm[$i],$i*8(%rsp)\n"; }      # K=L
  88 for($i=0;$i<8;$i++) { $code.="xor $i*8(%rsi),@mm[$i]\n"; }      # L^=inp
  89 for($i=0;$i<8;$i++) { $code.="mov @mm[$i],64+$i*8(%rsp)\n"; }   # S=L
  90 $code.=<<___;
  91         xor     %rsi,%rsi
  92         mov     %rsi,24(%rbx)           # zero round counter
  93         jmp     .Lround
  94 .align  16
  95 .Lround:
  96         mov     4096(%rbp,%rsi,8),@mm[0]        # rc[r]
  97         mov     0(%rsp),%eax
  98         mov     4(%rsp),%ebx
  99         movz    %al,%ecx
 100         movz    %ah,%edx
 101 ___
 102 for($i=0;$i<8;$i++) {
 103     my $func = ($i==0)? "mov" : "xor";
 104     $code.=<<___;
 105         shr     \$16,%eax
 106         lea     (%rcx,%rcx),%rsi
 107         movz    %al,%ecx
 108         lea     (%rdx,%rdx),%rdi
 109         movz    %ah,%edx
 110         xor     0(%rbp,%rsi,8),@mm[0]
 111         $func   7(%rbp,%rdi,8),@mm[1]
 112         mov     $i*8+8(%rsp),%eax               # ($i+1)*8
 113         lea     (%rcx,%rcx),%rsi
 114         movz    %bl,%ecx
 115         lea     (%rdx,%rdx),%rdi
 116         movz    %bh,%edx
 117         $func   6(%rbp,%rsi,8),@mm[2]
 118         $func   5(%rbp,%rdi,8),@mm[3]
 119         shr     \$16,%ebx
 120         lea     (%rcx,%rcx),%rsi
 121         movz    %bl,%ecx
 122         lea     (%rdx,%rdx),%rdi
 123         movz    %bh,%edx
 124         $func   4(%rbp,%rsi,8),@mm[4]
 125         $func   3(%rbp,%rdi,8),@mm[5]
 126         mov     $i*8+8+4(%rsp),%ebx             # ($i+1)*8+4
 127         lea     (%rcx,%rcx),%rsi
 128         movz    %al,%ecx
 129         lea     (%rdx,%rdx),%rdi
 130         movz    %ah,%edx
 131         $func   2(%rbp,%rsi,8),@mm[6]
 132         $func   1(%rbp,%rdi,8),@mm[7]
 133 ___
 134     push(@mm,shift(@mm));
 135 }
 136 for($i=0;$i<8;$i++) { $code.="mov @mm[$i],$i*8(%rsp)\n"; }      # K=L
 137 for($i=0;$i<8;$i++) {
 138     $code.=<<___;
 139         shr     \$16,%eax
 140         lea     (%rcx,%rcx),%rsi
 141         movz    %al,%ecx
 142         lea     (%rdx,%rdx),%rdi
 143         movz    %ah,%edx
 144         xor     0(%rbp,%rsi,8),@mm[0]
 145         xor     7(%rbp,%rdi,8),@mm[1]
 146         `"mov   64+$i*8+8(%rsp),%eax"   if($i<7);`      # 64+($i+1)*8
 147         lea     (%rcx,%rcx),%rsi
 148         movz    %bl,%ecx
 149         lea     (%rdx,%rdx),%rdi
 150         movz    %bh,%edx
 151         xor     6(%rbp,%rsi,8),@mm[2]
 152         xor     5(%rbp,%rdi,8),@mm[3]
 153         shr     \$16,%ebx
 154         lea     (%rcx,%rcx),%rsi
 155         movz    %bl,%ecx
 156         lea     (%rdx,%rdx),%rdi
 157         movz    %bh,%edx
 158         xor     4(%rbp,%rsi,8),@mm[4]
 159         xor     3(%rbp,%rdi,8),@mm[5]
 160         `"mov   64+$i*8+8+4(%rsp),%ebx" if($i<7);`      # 64+($i+1)*8+4
 161         lea     (%rcx,%rcx),%rsi
 162         movz    %al,%ecx
 163         lea     (%rdx,%rdx),%rdi
 164         movz    %ah,%edx
 165         xor     2(%rbp,%rsi,8),@mm[6]
 166         xor     1(%rbp,%rdi,8),@mm[7]
 167 ___
 168     push(@mm,shift(@mm));
 169 }
 170 $code.=<<___;
 171         lea     128(%rsp),%rbx
 172         mov     24(%rbx),%rsi           # pull round counter
 173         add     \$1,%rsi
 174         cmp     \$10,%rsi
 175         je      .Lroundsdone
 176
 177         mov     %rsi,24(%rbx)           # update round counter
 178 ___
 179 for($i=0;$i<8;$i++) { $code.="mov @mm[$i],64+$i*8(%rsp)\n"; }   # S=L
 180 $code.=<<___;
 181         jmp     .Lround
 182 .align  16
 183 .Lroundsdone:
 184         mov     0(%rbx),%rdi            # reload argument block
 185         mov     8(%rbx),%rsi
 186         mov     16(%rbx),%rax
 187 ___
 188 for($i=0;$i<8;$i++) { $code.="xor $i*8(%rsi),@mm[$i]\n"; }      # L^=inp
 189 for($i=0;$i<8;$i++) { $code.="xor $i*8(%rdi),@mm[$i]\n"; }      # L^=H
 190 for($i=0;$i<8;$i++) { $code.="mov @mm[$i],$i*8(%rdi)\n"; }      # H=L
 191 $code.=<<___;
 192         lea     64(%rsi),%rsi           # inp+=64
 193         sub     \$1,%rax                # num--
 194         jz      .Lalldone
 195         mov     %rsi,8(%rbx)            # update parameter block
 196         mov     %rax,16(%rbx)
 197         jmp     .Louterloop
 198 .Lalldone:
 199         mov     32(%rbx),%rsi           # restore saved pointer
 200         mov     (%rsi),%r15
 201         mov     8(%rsi),%r14
 202         mov     16(%rsi),%r13
 203         mov     24(%rsi),%r12
 204         mov     32(%rsi),%rbp
 205         mov     40(%rsi),%rbx
 206         lea     48(%rsi),%rsp
 207 .Lepilogue:
 208         ret
 209 .size   $func,.-$func
 210
 211 .align  64
 212 .type   $table,\@object
 213 $table:
 214 ___
 215         &LL(0x18,0x18,0x60,0x18,0xc0,0x78,0x30,0xd8);
 216         &LL(0x23,0x23,0x8c,0x23,0x05,0xaf,0x46,0x26);
 217         &LL(0xc6,0xc6,0x3f,0xc6,0x7e,0xf9,0x91,0xb8);
 218         &LL(0xe8,0xe8,0x87,0xe8,0x13,0x6f,0xcd,0xfb);
 219         &LL(0x87,0x87,0x26,0x87,0x4c,0xa1,0x13,0xcb);
 220         &LL(0xb8,0xb8,0xda,0xb8,0xa9,0x62,0x6d,0x11);
 221         &LL(0x01,0x01,0x04,0x01,0x08,0x05,0x02,0x09);
 222         &LL(0x4f,0x4f,0x21,0x4f,0x42,0x6e,0x9e,0x0d);
 223         &LL(0x36,0x36,0xd8,0x36,0xad,0xee,0x6c,0x9b);
 224         &LL(0xa6,0xa6,0xa2,0xa6,0x59,0x04,0x51,0xff);
 225         &LL(0xd2,0xd2,0x6f,0xd2,0xde,0xbd,0xb9,0x0c);
 226         &LL(0xf5,0xf5,0xf3,0xf5,0xfb,0x06,0xf7,0x0e);
 227         &LL(0x79,0x79,0xf9,0x79,0xef,0x80,0xf2,0x96);
 228         &LL(0x6f,0x6f,0xa1,0x6f,0x5f,0xce,0xde,0x30);
 229         &LL(0x91,0x91,0x7e,0x91,0xfc,0xef,0x3f,0x6d);
 230         &LL(0x52,0x52,0x55,0x52,0xaa,0x07,0xa4,0xf8);
 231         &LL(0x60,0x60,0x9d,0x60,0x27,0xfd,0xc0,0x47);
 232         &LL(0xbc,0xbc,0xca,0xbc,0x89,0x76,0x65,0x35);
 233         &LL(0x9b,0x9b,0x56,0x9b,0xac,0xcd,0x2b,0x37);
 234         &LL(0x8e,0x8e,0x02,0x8e,0x04,0x8c,0x01,0x8a);
 235         &LL(0xa3,0xa3,0xb6,0xa3,0x71,0x15,0x5b,0xd2);
 236         &LL(0x0c,0x0c,0x30,0x0c,0x60,0x3c,0x18,0x6c);
 237         &LL(0x7b,0x7b,0xf1,0x7b,0xff,0x8a,0xf6,0x84);
 238         &LL(0x35,0x35,0xd4,0x35,0xb5,0xe1,0x6a,0x80);
 239         &LL(0x1d,0x1d,0x74,0x1d,0xe8,0x69,0x3a,0xf5);
 240         &LL(0xe0,0xe0,0xa7,0xe0,0x53,0x47,0xdd,0xb3);
 241         &LL(0xd7,0xd7,0x7b,0xd7,0xf6,0xac,0xb3,0x21);
 242         &LL(0xc2,0xc2,0x2f,0xc2,0x5e,0xed,0x99,0x9c);
 243         &LL(0x2e,0x2e,0xb8,0x2e,0x6d,0x96,0x5c,0x43);
 244         &LL(0x4b,0x4b,0x31,0x4b,0x62,0x7a,0x96,0x29);
 245         &LL(0xfe,0xfe,0xdf,0xfe,0xa3,0x21,0xe1,0x5d);
 246         &LL(0x57,0x57,0x41,0x57,0x82,0x16,0xae,0xd5);
 247         &LL(0x15,0x15,0x54,0x15,0xa8,0x41,0x2a,0xbd);
 248         &LL(0x77,0x77,0xc1,0x77,0x9f,0xb6,0xee,0xe8);
 249         &LL(0x37,0x37,0xdc,0x37,0xa5,0xeb,0x6e,0x92);
 250         &LL(0xe5,0xe5,0xb3,0xe5,0x7b,0x56,0xd7,0x9e);
 251         &LL(0x9f,0x9f,0x46,0x9f,0x8c,0xd9,0x23,0x13);
 252         &LL(0xf0,0xf0,0xe7,0xf0,0xd3,0x17,0xfd,0x23);
 253         &LL(0x4a,0x4a,0x35,0x4a,0x6a,0x7f,0x94,0x20);
 254         &LL(0xda,0xda,0x4f,0xda,0x9e,0x95,0xa9,0x44);
 255         &LL(0x58,0x58,0x7d,0x58,0xfa,0x25,0xb0,0xa2);
 256         &LL(0xc9,0xc9,0x03,0xc9,0x06,0xca,0x8f,0xcf);
 257         &LL(0x29,0x29,0xa4,0x29,0x55,0x8d,0x52,0x7c);
 258         &LL(0x0a,0x0a,0x28,0x0a,0x50,0x22,0x14,0x5a);
 259         &LL(0xb1,0xb1,0xfe,0xb1,0xe1,0x4f,0x7f,0x50);
 260         &LL(0xa0,0xa0,0xba,0xa0,0x69,0x1a,0x5d,0xc9);
 261         &LL(0x6b,0x6b,0xb1,0x6b,0x7f,0xda,0xd6,0x14);
 262         &LL(0x85,0x85,0x2e,0x85,0x5c,0xab,0x17,0xd9);
 263         &LL(0xbd,0xbd,0xce,0xbd,0x81,0x73,0x67,0x3c);
 264         &LL(0x5d,0x5d,0x69,0x5d,0xd2,0x34,0xba,0x8f);
 265         &LL(0x10,0x10,0x40,0x10,0x80,0x50,0x20,0x90);
 266         &LL(0xf4,0xf4,0xf7,0xf4,0xf3,0x03,0xf5,0x07);
 267         &LL(0xcb,0xcb,0x0b,0xcb,0x16,0xc0,0x8b,0xdd);
 268         &LL(0x3e,0x3e,0xf8,0x3e,0xed,0xc6,0x7c,0xd3);
 269         &LL(0x05,0x05,0x14,0x05,0x28,0x11,0x0a,0x2d);
 270         &LL(0x67,0x67,0x81,0x67,0x1f,0xe6,0xce,0x78);
 271         &LL(0xe4,0xe4,0xb7,0xe4,0x73,0x53,0xd5,0x97);
 272         &LL(0x27,0x27,0x9c,0x27,0x25,0xbb,0x4e,0x02);
 273         &LL(0x41,0x41,0x19,0x41,0x32,0x58,0x82,0x73);
 274         &LL(0x8b,0x8b,0x16,0x8b,0x2c,0x9d,0x0b,0xa7);
 275         &LL(0xa7,0xa7,0xa6,0xa7,0x51,0x01,0x53,0xf6);
 276         &LL(0x7d,0x7d,0xe9,0x7d,0xcf,0x94,0xfa,0xb2);
 277         &LL(0x95,0x95,0x6e,0x95,0xdc,0xfb,0x37,0x49);
 278         &LL(0xd8,0xd8,0x47,0xd8,0x8e,0x9f,0xad,0x56);
 279         &LL(0xfb,0xfb,0xcb,0xfb,0x8b,0x30,0xeb,0x70);
 280         &LL(0xee,0xee,0x9f,0xee,0x23,0x71,0xc1,0xcd);
 281         &LL(0x7c,0x7c,0xed,0x7c,0xc7,0x91,0xf8,0xbb);
 282         &LL(0x66,0x66,0x85,0x66,0x17,0xe3,0xcc,0x71);
 283         &LL(0xdd,0xdd,0x53,0xdd,0xa6,0x8e,0xa7,0x7b);
 284         &LL(0x17,0x17,0x5c,0x17,0xb8,0x4b,0x2e,0xaf);
 285         &LL(0x47,0x47,0x01,0x47,0x02,0x46,0x8e,0x45);
 286         &LL(0x9e,0x9e,0x42,0x9e,0x84,0xdc,0x21,0x1a);
 287         &LL(0xca,0xca,0x0f,0xca,0x1e,0xc5,0x89,0xd4);
 288         &LL(0x2d,0x2d,0xb4,0x2d,0x75,0x99,0x5a,0x58);
 289         &LL(0xbf,0xbf,0xc6,0xbf,0x91,0x79,0x63,0x2e);
 290         &LL(0x07,0x07,0x1c,0x07,0x38,0x1b,0x0e,0x3f);
 291         &LL(0xad,0xad,0x8e,0xad,0x01,0x23,0x47,0xac);
 292         &LL(0x5a,0x5a,0x75,0x5a,0xea,0x2f,0xb4,0xb0);
 293         &LL(0x83,0x83,0x36,0x83,0x6c,0xb5,0x1b,0xef);
 294         &LL(0x33,0x33,0xcc,0x33,0x85,0xff,0x66,0xb6);
 295         &LL(0x63,0x63,0x91,0x63,0x3f,0xf2,0xc6,0x5c);
 296         &LL(0x02,0x02,0x08,0x02,0x10,0x0a,0x04,0x12);
 297         &LL(0xaa,0xaa,0x92,0xaa,0x39,0x38,0x49,0x93);
 298         &LL(0x71,0x71,0xd9,0x71,0xaf,0xa8,0xe2,0xde);
 299         &LL(0xc8,0xc8,0x07,0xc8,0x0e,0xcf,0x8d,0xc6);
 300         &LL(0x19,0x19,0x64,0x19,0xc8,0x7d,0x32,0xd1);
 301         &LL(0x49,0x49,0x39,0x49,0x72,0x70,0x92,0x3b);
 302         &LL(0xd9,0xd9,0x43,0xd9,0x86,0x9a,0xaf,0x5f);
 303         &LL(0xf2,0xf2,0xef,0xf2,0xc3,0x1d,0xf9,0x31);
 304         &LL(0xe3,0xe3,0xab,0xe3,0x4b,0x48,0xdb,0xa8);
 305         &LL(0x5b,0x5b,0x71,0x5b,0xe2,0x2a,0xb6,0xb9);
 306         &LL(0x88,0x88,0x1a,0x88,0x34,0x92,0x0d,0xbc);
 307         &LL(0x9a,0x9a,0x52,0x9a,0xa4,0xc8,0x29,0x3e);
 308         &LL(0x26,0x26,0x98,0x26,0x2d,0xbe,0x4c,0x0b);
 309         &LL(0x32,0x32,0xc8,0x32,0x8d,0xfa,0x64,0xbf);
 310         &LL(0xb0,0xb0,0xfa,0xb0,0xe9,0x4a,0x7d,0x59);
 311         &LL(0xe9,0xe9,0x83,0xe9,0x1b,0x6a,0xcf,0xf2);
 312         &LL(0x0f,0x0f,0x3c,0x0f,0x78,0x33,0x1e,0x77);
 313         &LL(0xd5,0xd5,0x73,0xd5,0xe6,0xa6,0xb7,0x33);
 314         &LL(0x80,0x80,0x3a,0x80,0x74,0xba,0x1d,0xf4);
 315         &LL(0xbe,0xbe,0xc2,0xbe,0x99,0x7c,0x61,0x27);
 316         &LL(0xcd,0xcd,0x13,0xcd,0x26,0xde,0x87,0xeb);
 317         &LL(0x34,0x34,0xd0,0x34,0xbd,0xe4,0x68,0x89);
 318         &LL(0x48,0x48,0x3d,0x48,0x7a,0x75,0x90,0x32);
 319         &LL(0xff,0xff,0xdb,0xff,0xab,0x24,0xe3,0x54);
 320         &LL(0x7a,0x7a,0xf5,0x7a,0xf7,0x8f,0xf4,0x8d);
 321         &LL(0x90,0x90,0x7a,0x90,0xf4,0xea,0x3d,0x64);
 322         &LL(0x5f,0x5f,0x61,0x5f,0xc2,0x3e,0xbe,0x9d);
 323         &LL(0x20,0x20,0x80,0x20,0x1d,0xa0,0x40,0x3d);
 324         &LL(0x68,0x68,0xbd,0x68,0x67,0xd5,0xd0,0x0f);
 325         &LL(0x1a,0x1a,0x68,0x1a,0xd0,0x72,0x34,0xca);
 326         &LL(0xae,0xae,0x82,0xae,0x19,0x2c,0x41,0xb7);
 327         &LL(0xb4,0xb4,0xea,0xb4,0xc9,0x5e,0x75,0x7d);
 328         &LL(0x54,0x54,0x4d,0x54,0x9a,0x19,0xa8,0xce);
 329         &LL(0x93,0x93,0x76,0x93,0xec,0xe5,0x3b,0x7f);
 330         &LL(0x22,0x22,0x88,0x22,0x0d,0xaa,0x44,0x2f);
 331         &LL(0x64,0x64,0x8d,0x64,0x07,0xe9,0xc8,0x63);
 332         &LL(0xf1,0xf1,0xe3,0xf1,0xdb,0x12,0xff,0x2a);
 333         &LL(0x73,0x73,0xd1,0x73,0xbf,0xa2,0xe6,0xcc);
 334         &LL(0x12,0x12,0x48,0x12,0x90,0x5a,0x24,0x82);
 335         &LL(0x40,0x40,0x1d,0x40,0x3a,0x5d,0x80,0x7a);
 336         &LL(0x08,0x08,0x20,0x08,0x40,0x28,0x10,0x48);
 337         &LL(0xc3,0xc3,0x2b,0xc3,0x56,0xe8,0x9b,0x95);
 338         &LL(0xec,0xec,0x97,0xec,0x33,0x7b,0xc5,0xdf);
 339         &LL(0xdb,0xdb,0x4b,0xdb,0x96,0x90,0xab,0x4d);
 340         &LL(0xa1,0xa1,0xbe,0xa1,0x61,0x1f,0x5f,0xc0);
 341         &LL(0x8d,0x8d,0x0e,0x8d,0x1c,0x83,0x07,0x91);
 342         &LL(0x3d,0x3d,0xf4,0x3d,0xf5,0xc9,0x7a,0xc8);
 343         &LL(0x97,0x97,0x66,0x97,0xcc,0xf1,0x33,0x5b);
 344         &LL(0x00,0x00,0x00,0x00,0x00,0x00,0x00,0x00);
 345         &LL(0xcf,0xcf,0x1b,0xcf,0x36,0xd4,0x83,0xf9);
 346         &LL(0x2b,0x2b,0xac,0x2b,0x45,0x87,0x56,0x6e);
 347         &LL(0x76,0x76,0xc5,0x76,0x97,0xb3,0xec,0xe1);
 348         &LL(0x82,0x82,0x32,0x82,0x64,0xb0,0x19,0xe6);
 349         &LL(0xd6,0xd6,0x7f,0xd6,0xfe,0xa9,0xb1,0x28);
 350         &LL(0x1b,0x1b,0x6c,0x1b,0xd8,0x77,0x36,0xc3);
 351         &LL(0xb5,0xb5,0xee,0xb5,0xc1,0x5b,0x77,0x74);
 352         &LL(0xaf,0xaf,0x86,0xaf,0x11,0x29,0x43,0xbe);
 353         &LL(0x6a,0x6a,0xb5,0x6a,0x77,0xdf,0xd4,0x1d);
 354         &LL(0x50,0x50,0x5d,0x50,0xba,0x0d,0xa0,0xea);
 355         &LL(0x45,0x45,0x09,0x45,0x12,0x4c,0x8a,0x57);
 356         &LL(0xf3,0xf3,0xeb,0xf3,0xcb,0x18,0xfb,0x38);
 357         &LL(0x30,0x30,0xc0,0x30,0x9d,0xf0,0x60,0xad);
 358         &LL(0xef,0xef,0x9b,0xef,0x2b,0x74,0xc3,0xc4);
 359         &LL(0x3f,0x3f,0xfc,0x3f,0xe5,0xc3,0x7e,0xda);
 360         &LL(0x55,0x55,0x49,0x55,0x92,0x1c,0xaa,0xc7);
 361         &LL(0xa2,0xa2,0xb2,0xa2,0x79,0x10,0x59,0xdb);
 362         &LL(0xea,0xea,0x8f,0xea,0x03,0x65,0xc9,0xe9);
 363         &LL(0x65,0x65,0x89,0x65,0x0f,0xec,0xca,0x6a);
 364         &LL(0xba,0xba,0xd2,0xba,0xb9,0x68,0x69,0x03);
 365         &LL(0x2f,0x2f,0xbc,0x2f,0x65,0x93,0x5e,0x4a);
 366         &LL(0xc0,0xc0,0x27,0xc0,0x4e,0xe7,0x9d,0x8e);
 367         &LL(0xde,0xde,0x5f,0xde,0xbe,0x81,0xa1,0x60);
 368         &LL(0x1c,0x1c,0x70,0x1c,0xe0,0x6c,0x38,0xfc);
 369         &LL(0xfd,0xfd,0xd3,0xfd,0xbb,0x2e,0xe7,0x46);
 370         &LL(0x4d,0x4d,0x29,0x4d,0x52,0x64,0x9a,0x1f);
 371         &LL(0x92,0x92,0x72,0x92,0xe4,0xe0,0x39,0x76);
 372         &LL(0x75,0x75,0xc9,0x75,0x8f,0xbc,0xea,0xfa);
 373         &LL(0x06,0x06,0x18,0x06,0x30,0x1e,0x0c,0x36);
 374         &LL(0x8a,0x8a,0x12,0x8a,0x24,0x98,0x09,0xae);
 375         &LL(0xb2,0xb2,0xf2,0xb2,0xf9,0x40,0x79,0x4b);
 376         &LL(0xe6,0xe6,0xbf,0xe6,0x63,0x59,0xd1,0x85);
 377         &LL(0x0e,0x0e,0x38,0x0e,0x70,0x36,0x1c,0x7e);
 378         &LL(0x1f,0x1f,0x7c,0x1f,0xf8,0x63,0x3e,0xe7);
 379         &LL(0x62,0x62,0x95,0x62,0x37,0xf7,0xc4,0x55);
 380         &LL(0xd4,0xd4,0x77,0xd4,0xee,0xa3,0xb5,0x3a);
 381         &LL(0xa8,0xa8,0x9a,0xa8,0x29,0x32,0x4d,0x81);
 382         &LL(0x96,0x96,0x62,0x96,0xc4,0xf4,0x31,0x52);
 383         &LL(0xf9,0xf9,0xc3,0xf9,0x9b,0x3a,0xef,0x62);
 384         &LL(0xc5,0xc5,0x33,0xc5,0x66,0xf6,0x97,0xa3);
 385         &LL(0x25,0x25,0x94,0x25,0x35,0xb1,0x4a,0x10);
 386         &LL(0x59,0x59,0x79,0x59,0xf2,0x20,0xb2,0xab);
 387         &LL(0x84,0x84,0x2a,0x84,0x54,0xae,0x15,0xd0);
 388         &LL(0x72,0x72,0xd5,0x72,0xb7,0xa7,0xe4,0xc5);
 389         &LL(0x39,0x39,0xe4,0x39,0xd5,0xdd,0x72,0xec);
 390         &LL(0x4c,0x4c,0x2d,0x4c,0x5a,0x61,0x98,0x16);
 391         &LL(0x5e,0x5e,0x65,0x5e,0xca,0x3b,0xbc,0x94);
 392         &LL(0x78,0x78,0xfd,0x78,0xe7,0x85,0xf0,0x9f);
 393         &LL(0x38,0x38,0xe0,0x38,0xdd,0xd8,0x70,0xe5);
 394         &LL(0x8c,0x8c,0x0a,0x8c,0x14,0x86,0x05,0x98);
 395         &LL(0xd1,0xd1,0x63,0xd1,0xc6,0xb2,0xbf,0x17);
 396         &LL(0xa5,0xa5,0xae,0xa5,0x41,0x0b,0x57,0xe4);
 397         &LL(0xe2,0xe2,0xaf,0xe2,0x43,0x4d,0xd9,0xa1);
 398         &LL(0x61,0x61,0x99,0x61,0x2f,0xf8,0xc2,0x4e);
 399         &LL(0xb3,0xb3,0xf6,0xb3,0xf1,0x45,0x7b,0x42);
 400         &LL(0x21,0x21,0x84,0x21,0x15,0xa5,0x42,0x34);
 401         &LL(0x9c,0x9c,0x4a,0x9c,0x94,0xd6,0x25,0x08);
 402         &LL(0x1e,0x1e,0x78,0x1e,0xf0,0x66,0x3c,0xee);
 403         &LL(0x43,0x43,0x11,0x43,0x22,0x52,0x86,0x61);
 404         &LL(0xc7,0xc7,0x3b,0xc7,0x76,0xfc,0x93,0xb1);
 405         &LL(0xfc,0xfc,0xd7,0xfc,0xb3,0x2b,0xe5,0x4f);
 406         &LL(0x04,0x04,0x10,0x04,0x20,0x14,0x08,0x24);
 407         &LL(0x51,0x51,0x59,0x51,0xb2,0x08,0xa2,0xe3);
 408         &LL(0x99,0x99,0x5e,0x99,0xbc,0xc7,0x2f,0x25);
 409         &LL(0x6d,0x6d,0xa9,0x6d,0x4f,0xc4,0xda,0x22);
 410         &LL(0x0d,0x0d,0x34,0x0d,0x68,0x39,0x1a,0x65);
 411         &LL(0xfa,0xfa,0xcf,0xfa,0x83,0x35,0xe9,0x79);
 412         &LL(0xdf,0xdf,0x5b,0xdf,0xb6,0x84,0xa3,0x69);
 413         &LL(0x7e,0x7e,0xe5,0x7e,0xd7,0x9b,0xfc,0xa9);
 414         &LL(0x24,0x24,0x90,0x24,0x3d,0xb4,0x48,0x19);
 415         &LL(0x3b,0x3b,0xec,0x3b,0xc5,0xd7,0x76,0xfe);
 416         &LL(0xab,0xab,0x96,0xab,0x31,0x3d,0x4b,0x9a);
 417         &LL(0xce,0xce,0x1f,0xce,0x3e,0xd1,0x81,0xf0);
 418         &LL(0x11,0x11,0x44,0x11,0x88,0x55,0x22,0x99);
 419         &LL(0x8f,0x8f,0x06,0x8f,0x0c,0x89,0x03,0x83);
 420         &LL(0x4e,0x4e,0x25,0x4e,0x4a,0x6b,0x9c,0x04);
 421         &LL(0xb7,0xb7,0xe6,0xb7,0xd1,0x51,0x73,0x66);
 422         &LL(0xeb,0xeb,0x8b,0xeb,0x0b,0x60,0xcb,0xe0);
 423         &LL(0x3c,0x3c,0xf0,0x3c,0xfd,0xcc,0x78,0xc1);
 424         &LL(0x81,0x81,0x3e,0x81,0x7c,0xbf,0x1f,0xfd);
 425         &LL(0x94,0x94,0x6a,0x94,0xd4,0xfe,0x35,0x40);
 426         &LL(0xf7,0xf7,0xfb,0xf7,0xeb,0x0c,0xf3,0x1c);
 427         &LL(0xb9,0xb9,0xde,0xb9,0xa1,0x67,0x6f,0x18);
 428         &LL(0x13,0x13,0x4c,0x13,0x98,0x5f,0x26,0x8b);
 429         &LL(0x2c,0x2c,0xb0,0x2c,0x7d,0x9c,0x58,0x51);
 430         &LL(0xd3,0xd3,0x6b,0xd3,0xd6,0xb8,0xbb,0x05);
 431         &LL(0xe7,0xe7,0xbb,0xe7,0x6b,0x5c,0xd3,0x8c);
 432         &LL(0x6e,0x6e,0xa5,0x6e,0x57,0xcb,0xdc,0x39);
 433         &LL(0xc4,0xc4,0x37,0xc4,0x6e,0xf3,0x95,0xaa);
 434         &LL(0x03,0x03,0x0c,0x03,0x18,0x0f,0x06,0x1b);
 435         &LL(0x56,0x56,0x45,0x56,0x8a,0x13,0xac,0xdc);
 436         &LL(0x44,0x44,0x0d,0x44,0x1a,0x49,0x88,0x5e);
 437         &LL(0x7f,0x7f,0xe1,0x7f,0xdf,0x9e,0xfe,0xa0);
 438         &LL(0xa9,0xa9,0x9e,0xa9,0x21,0x37,0x4f,0x88);
 439         &LL(0x2a,0x2a,0xa8,0x2a,0x4d,0x82,0x54,0x67);
 440         &LL(0xbb,0xbb,0xd6,0xbb,0xb1,0x6d,0x6b,0x0a);
 441         &LL(0xc1,0xc1,0x23,0xc1,0x46,0xe2,0x9f,0x87);
 442         &LL(0x53,0x53,0x51,0x53,0xa2,0x02,0xa6,0xf1);
 443         &LL(0xdc,0xdc,0x57,0xdc,0xae,0x8b,0xa5,0x72);
 444         &LL(0x0b,0x0b,0x2c,0x0b,0x58,0x27,0x16,0x53);
 445         &LL(0x9d,0x9d,0x4e,0x9d,0x9c,0xd3,0x27,0x01);
 446         &LL(0x6c,0x6c,0xad,0x6c,0x47,0xc1,0xd8,0x2b);
 447         &LL(0x31,0x31,0xc4,0x31,0x95,0xf5,0x62,0xa4);
 448         &LL(0x74,0x74,0xcd,0x74,0x87,0xb9,0xe8,0xf3);
 449         &LL(0xf6,0xf6,0xff,0xf6,0xe3,0x09,0xf1,0x15);
 450         &LL(0x46,0x46,0x05,0x46,0x0a,0x43,0x8c,0x4c);
 451         &LL(0xac,0xac,0x8a,0xac,0x09,0x26,0x45,0xa5);
 452         &LL(0x89,0x89,0x1e,0x89,0x3c,0x97,0x0f,0xb5);
 453         &LL(0x14,0x14,0x50,0x14,0xa0,0x44,0x28,0xb4);
 454         &LL(0xe1,0xe1,0xa3,0xe1,0x5b,0x42,0xdf,0xba);
 455         &LL(0x16,0x16,0x58,0x16,0xb0,0x4e,0x2c,0xa6);
 456         &LL(0x3a,0x3a,0xe8,0x3a,0xcd,0xd2,0x74,0xf7);
 457         &LL(0x69,0x69,0xb9,0x69,0x6f,0xd0,0xd2,0x06);
 458         &LL(0x09,0x09,0x24,0x09,0x48,0x2d,0x12,0x41);
 459         &LL(0x70,0x70,0xdd,0x70,0xa7,0xad,0xe0,0xd7);
 460         &LL(0xb6,0xb6,0xe2,0xb6,0xd9,0x54,0x71,0x6f);
 461         &LL(0xd0,0xd0,0x67,0xd0,0xce,0xb7,0xbd,0x1e);
 462         &LL(0xed,0xed,0x93,0xed,0x3b,0x7e,0xc7,0xd6);
 463         &LL(0xcc,0xcc,0x17,0xcc,0x2e,0xdb,0x85,0xe2);
 464         &LL(0x42,0x42,0x15,0x42,0x2a,0x57,0x84,0x68);
 465         &LL(0x98,0x98,0x5a,0x98,0xb4,0xc2,0x2d,0x2c);
 466         &LL(0xa4,0xa4,0xaa,0xa4,0x49,0x0e,0x55,0xed);
 467         &LL(0x28,0x28,0xa0,0x28,0x5d,0x88,0x50,0x75);
 468         &LL(0x5c,0x5c,0x6d,0x5c,0xda,0x31,0xb8,0x86);
 469         &LL(0xf8,0xf8,0xc7,0xf8,0x93,0x3f,0xed,0x6b);
 470         &LL(0x86,0x86,0x22,0x86,0x44,0xa4,0x11,0xc2);
 471
 472         &L(0x18,0x23,0xc6,0xe8,0x87,0xb8,0x01,0x4f);    # rc[ROUNDS]
 473         &L(0x36,0xa6,0xd2,0xf5,0x79,0x6f,0x91,0x52);
 474         &L(0x60,0xbc,0x9b,0x8e,0xa3,0x0c,0x7b,0x35);
 475         &L(0x1d,0xe0,0xd7,0xc2,0x2e,0x4b,0xfe,0x57);
 476         &L(0x15,0x77,0x37,0xe5,0x9f,0xf0,0x4a,0xda);
 477         &L(0x58,0xc9,0x29,0x0a,0xb1,0xa0,0x6b,0x85);
 478         &L(0xbd,0x5d,0x10,0xf4,0xcb,0x3e,0x05,0x67);
 479         &L(0xe4,0x27,0x41,0x8b,0xa7,0x7d,0x95,0xd8);
 480         &L(0xfb,0xee,0x7c,0x66,0xdd,0x17,0x47,0x9e);
 481         &L(0xca,0x2d,0xbf,0x07,0xad,0x5a,0x83,0x33);
 482
 483 # EXCEPTION_DISPOSITION handler (EXCEPTION_RECORD *rec,ULONG64 frame,
 484 #               CONTEXT *context,DISPATCHER_CONTEXT *disp)
 485 if ($win64) {
 486 $rec="%rcx";
 487 $frame="%rdx";
 488 $context="%r8";
 489 $disp="%r9";
 490
 491 $code.=<<___;
 492 .extern __imp_RtlVirtualUnwind
 493 .type   se_handler,\@abi-omnipotent
 494 .align  16
 495 se_handler:
 496         push    %rsi
 497         push    %rdi
 498         push    %rbx
 499         push    %rbp
 500         push    %r12
 501         push    %r13
 502         push    %r14
 503         push    %r15
 504         pushfq
 505         sub     \$64,%rsp
 506
 507         mov     120($context),%rax      # pull context->Rax
 508         mov     248($context),%rbx      # pull context->Rip
 509
 510         lea     .Lprologue(%rip),%r10
 511         cmp     %r10,%rbx               # context->Rip<.Lprologue
 512         jb      .Lin_prologue
 513
 514         mov     152($context),%rax      # pull context->Rsp
 515
 516         lea     .Lepilogue(%rip),%r10
 517         cmp     %r10,%rbx               # context->Rip>=.Lepilogue
 518         jae     .Lin_prologue
 519
 520         mov     128+32(%rax),%rax       # pull saved stack pointer
 521         lea     48(%rax),%rax
 522
 523         mov     -8(%rax),%rbx
 524         mov     -16(%rax),%rbp
 525         mov     -24(%rax),%r12
 526         mov     -32(%rax),%r13
 527         mov     -40(%rax),%r14
 528         mov     -48(%rax),%r15
 529         mov     %rbx,144($context)      # restore context->Rbx
 530         mov     %rbp,160($context)      # restore context->Rbp
 531         mov     %r12,216($context)      # restore context->R12
 532         mov     %r13,224($context)      # restore context->R13
 533         mov     %r14,232($context)      # restore context->R14
 534         mov     %r15,240($context)      # restore context->R15
 535
 536 .Lin_prologue:
 537         mov     8(%rax),%rdi
 538         mov     16(%rax),%rsi
 539         mov     %rax,152($context)      # restore context->Rsp
 540         mov     %rsi,168($context)      # restore context->Rsi
 541         mov     %rdi,176($context)      # restore context->Rdi
 542
 543         mov     40($disp),%rdi          # disp->ContextRecord
 544         mov     $context,%rsi           # context
 545         mov     \$154,%ecx              # sizeof(CONTEXT)
 546         .long   0xa548f3fc              # cld; rep movsq
 547
 548         mov     $disp,%rsi
 549         xor     %rcx,%rcx               # arg1, UNW_FLAG_NHANDLER
 550         mov     8(%rsi),%rdx            # arg2, disp->ImageBase
 551         mov     0(%rsi),%r8             # arg3, disp->ControlPc
 552         mov     16(%rsi),%r9            # arg4, disp->FunctionEntry
 553         mov     40(%rsi),%r10           # disp->ContextRecord
 554         lea     56(%rsi),%r11           # &disp->HandlerData
 555         lea     24(%rsi),%r12           # &disp->EstablisherFrame
 556         mov     %r10,32(%rsp)           # arg5
 557         mov     %r11,40(%rsp)           # arg6
 558         mov     %r12,48(%rsp)           # arg7
 559         mov     %rcx,56(%rsp)           # arg8, (NULL)
 560         call    *__imp_RtlVirtualUnwind(%rip)
 561
 562         mov     \$1,%eax                # ExceptionContinueSearch
 563         add     \$64,%rsp
 564         popfq
 565         pop     %r15
 566         pop     %r14
 567         pop     %r13
 568         pop     %r12
 569         pop     %rbp
 570         pop     %rbx
 571         pop     %rdi
 572         pop     %rsi
 573         ret
 574 .size   se_handler,.-se_handler
 575
 576 .section        .pdata
 577 .align  4
 578         .rva    .LSEH_begin_$func
 579         .rva    .LSEH_end_$func
 580         .rva    .LSEH_info_$func
 581
 582 .section        .xdata
 583 .align  8
 584 .LSEH_info_$func:
 585         .byte   9,0,0,0
 586         .rva    se_handler
 587 ___
 588 }
 589
 590 $code =~ s/\`([^\`]*)\`/eval $1/gem;
 591 print $code;
 592 close STDOUT;