crypto/perlasm/x86_64-xlate.pl

   1 #!/usr/bin/env perl
   2
   3 # Ascetic x86_64 AT&T to MASM/NASM assembler translator by <appro>.
   4 #
   5 # Why AT&T to MASM and not vice versa? Several reasons. Because AT&T
   6 # format is way easier to parse. Because it's simpler to "gear" from
   7 # Unix ABI to Windows one [see cross-reference "card" at the end of
   8 # file]. Because Linux targets were available first...
   9 #
  10 # In addition the script also "distills" code suitable for GNU
  11 # assembler, so that it can be compiled with more rigid assemblers,
  12 # such as Solaris /usr/ccs/bin/as.
  13 #
  14 # This translator is not designed to convert *arbitrary* assembler
  15 # code from AT&T format to MASM one. It's designed to convert just
  16 # enough to provide for dual-ABI OpenSSL modules development...
  17 # There *are* limitations and you might have to modify your assembler
  18 # code or this script to achieve the desired result...
  19 #
  20 # Currently recognized limitations:
  21 #
  22 # - can't use multiple ops per line;
  23 #
  24 # Dual-ABI styling rules.
  25 #
  26 # 1. Adhere to Unix register and stack layout [see cross-reference
  27 #    ABI "card" at the end for explanation].
  28 # 2. Forget about "red zone," stick to more traditional blended
  29 #    stack frame allocation. If volatile storage is actually required
  30 #    that is. If not, just leave the stack as is.
  31 # 3. Functions tagged with ".type name,@function" get crafted with
  32 #    unified Win64 prologue and epilogue automatically. If you want
  33 #    to take care of ABI differences yourself, tag functions as
  34 #    ".type name,@abi-omnipotent" instead.
  35 # 4. To optimize the Win64 prologue you can specify number of input
  36 #    arguments as ".type name,@function,N." Keep in mind that if N is
  37 #    larger than 6, then you *have to* write "abi-omnipotent" code,
  38 #    because >6 cases can't be addressed with unified prologue.
  39 # 5. Name local labels as .L*, do *not* use dynamic labels such as 1:
  40 #    (sorry about latter).
  41 # 6. Don't use [or hand-code with .byte] "rep ret." "ret" mnemonic is
  42 #    required to identify the spots, where to inject Win64 epilogue!
  43 #    But on the pros, it's then prefixed with rep automatically:-)
  44 # 7. Due to MASM limitations [and certain general counter-intuitivity
  45 #    of ip-relative addressing] generation of position-independent
  46 #    code is assisted by synthetic directive, .picmeup, which puts
  47 #    address of the *next* instruction into target register.
  48 #
  49 #    Example 1:
  50 #               .picmeup        %rax
  51 #               lea             .Label-.(%rax),%rax
  52 #    Example 2:
  53 #               .picmeup        %rcx
  54 #       .Lpic_point:
  55 #               ...
  56 #               lea             .Label-.Lpic_point(%rcx),%rbp
  57 #
  58 # 8. In order to provide for structured exception handling unified
  59 #    Win64 prologue copies %rsp value to %rax. For further details
  60 #    see SEH paragraph at the end.
  61 \f
  62 my $output = shift;
  63
  64 { my ($stddev,$stdino,@junk)=stat(STDOUT);
  65   my ($outdev,$outino,@junk)=stat($output);
  66
  67     open STDOUT,">$output" || die "can't open $output: $!"
  68         if ($stddev!=$outdev || $stdino!=$outino);
  69 }
  70
  71 my $win64=1 if ($output =~ /\.asm/);
  72
  73 my $masmref=8 + 50727*2**-32;   # 8.00.50727 shipped with VS2005
  74 my $masm=0;
  75 my $PTR=" PTR";
  76
  77 my $nasmref=2.03;
  78 my $nasm=0;
  79
  80 if ($win64)
  81 {   if ($ENV{ASM} =~ m/nasm/ && `nasm -v` =~ m/version ([0-9]+)\.([0-9]+)/i)
  82     {   $nasm = $1 + $2*0.01; $PTR="";  }
  83     elsif (`ml64 2>&1` =~ m/Version ([0-9]+)\.([0-9]+)(\.([0-9]+))?/)
  84     {   $masm = $1 + $2*2**-16 + $4*2**-32;   }
  85     die "no assembler found on %PATH" if (!($nasm || $masm));
  86 }
  87
  88 my $current_segment;
  89 my $current_function;
  90 my %globals;
  91
  92 { package opcode;       # pick up opcodes
  93     sub re {
  94         my      $self = shift;  # single instance in enough...
  95         local   *line = shift;
  96         undef   $ret;
  97
  98         if ($line =~ /^([a-z][a-z0-9]*)/i) {
  99             $self->{op} = $1;
 100             $ret = $self;
 101             $line = substr($line,@+[0]); $line =~ s/^\s+//;
 102
 103             undef $self->{sz};
 104             if ($self->{op} =~ /^(movz)b.*/) {  # movz is pain...
 105                 $self->{op} = $1;
 106                 $self->{sz} = "b";
 107             } elsif ($self->{op} =~ /call|jmp/) {
 108                 $self->{sz} = ""
 109             } elsif ($self->{op} =~ /([a-z]{3,})([qlwb])$/) {
 110                 $self->{op} = $1;
 111                 $self->{sz} = $2;
 112             }
 113         }
 114         $ret;
 115     }
 116     sub size {
 117         my $self = shift;
 118         my $sz   = shift;
 119         $self->{sz} = $sz if (defined($sz) && !defined($self->{sz}));
 120         $self->{sz};
 121     }
 122     sub out {
 123         my $self = shift;
 124         if (!$win64) {
 125             if ($self->{op} eq "movz") {        # movz is pain...
 126                 sprintf "%s%s%s",$self->{op},$self->{sz},shift;
 127             } elsif ($self->{op} =~ /^set/) {
 128                 "$self->{op}";
 129             } elsif ($self->{op} eq "ret") {
 130                 ".byte  0xf3,0xc3";
 131             } else {
 132                 "$self->{op}$self->{sz}";
 133             }
 134         } else {
 135             $self->{op} =~ s/^movz/movzx/;
 136             if ($self->{op} eq "ret") {
 137                 $self->{op} = "";
 138                 if ($current_function->{abi} eq "svr4") {
 139                     $self->{op} = "mov  rdi,QWORD${PTR}[8+rsp]\t;WIN64 epilogue\n\t".
 140                                   "mov  rsi,QWORD${PTR}[16+rsp]\n\t";
 141                 }
 142                 $self->{op} .= "DB\t0F3h,0C3h\t\t;repret";
 143             }
 144             elsif ($self->{op} =~ /^(pop|push)f/) {
 145                 $self->{op} .= $self->{sz};
 146             }
 147             $self->{op};
 148         }
 149     }
 150     sub mnemonic { shift->{op}; }
 151 }
 152 { package const;        # pick up constants, which start with $
 153     sub re {
 154         my      $self = shift;  # single instance in enough...
 155         local   *line = shift;
 156         undef   $ret;
 157
 158         if ($line =~ /^\$([^,]+)/) {
 159             $self->{value} = $1;
 160             $ret = $self;
 161             $line = substr($line,@+[0]); $line =~ s/^\s+//;
 162         }
 163         $ret;
 164     }
 165     sub out {
 166         my $self = shift;
 167
 168         if (!$win64) {
 169             # Solaris /usr/ccs/bin/as can't handle multiplications
 170             # in $self->{value}
 171             $self->{value} =~ s/(?<![0-9a-f])(0[x0-9a-f]+)/oct($1)/egi;
 172             $self->{value} =~ s/([0-9]+\s*[\*\/\%]\s*[0-9]+)/eval($1)/eg;
 173             sprintf "\$%s",$self->{value};
 174         } else {
 175             $self->{value} =~ s/0x([0-9a-f]+)/0$1h/ig;
 176             sprintf "%s",$self->{value};
 177         }
 178     }
 179 }
 180 { package ea;           # pick up effective addresses: expr(%reg,%reg,scale)
 181     sub re {
 182         my      $self = shift;  # single instance in enough...
 183         local   *line = shift;
 184         undef   $ret;
 185
 186         # optional * ---vvv--- appears in indirect jmp/call
 187         if ($line =~ /^(\*?)([^\(,]*)\(([%\w,]+)\)/) {
 188             $self->{asterisk} = $1;
 189             $self->{label} = $2;
 190             ($self->{base},$self->{index},$self->{scale})=split(/,/,$3);
 191             $self->{scale} = 1 if (!defined($self->{scale}));
 192             $ret = $self;
 193             $line = substr($line,@+[0]); $line =~ s/^\s+//;
 194
 195             $self->{base}  =~ s/^%//;
 196             $self->{index} =~ s/^%// if (defined($self->{index}));
 197         }
 198         $ret;
 199     }
 200     sub size {}
 201     sub out {
 202         my $self = shift;
 203         my $sz = shift;
 204
 205         # Silently convert all EAs to 64-bit. This is required for
 206         # elder GNU assembler and results in more compact code,
 207         # *but* most importantly AES module depends on this feature!
 208         $self->{index} =~ s/^[er](.?[0-9xpi])[d]?$/r\1/;
 209         $self->{base}  =~ s/^[er](.?[0-9xpi])[d]?$/r\1/;
 210
 211         if (!$win64) {
 212             # Solaris /usr/ccs/bin/as can't handle multiplications
 213             # in $self->{label}
 214             $self->{label} =~ s/(?<![0-9a-f])(0[x0-9a-f]+)/oct($1)/egi;
 215             $self->{label} =~ s/([0-9]+\s*[\*\/\%]\s*[0-9]+)/eval($1)/eg;
 216
 217             if (defined($self->{index})) {
 218                 sprintf "%s%s(%%%s,%%%s,%d)",$self->{asterisk},
 219                                         $self->{label},$self->{base},
 220                                         $self->{index},$self->{scale};
 221             } else {
 222                 sprintf "%s%s(%%%s)",   $self->{asterisk},$self->{label},$self->{base};
 223             }
 224         } else {
 225             %szmap = ( b=>"BYTE$PTR", w=>"WORD$PTR", l=>"DWORD$PTR", q=>"QWORD$PTR" );
 226
 227             $self->{label} =~ s/\.L/\$L\$/g;
 228             $self->{label} =~ s/\./\$/g;
 229             $self->{label} =~ s/0x([0-9a-f]+)/0$1h/ig;
 230             $self->{label} = "($self->{label})" if ($self->{label} =~ /[\*\+\-\/]/);
 231             $sz="q" if ($self->{asterisk});
 232
 233             if (defined($self->{index})) {
 234                 sprintf "%s[%s%s*%d+%s]",$szmap{$sz},
 235                                         $self->{label}?"$self->{label}+":"",
 236                                         $self->{index},$self->{scale},
 237                                         $self->{base};
 238             } elsif ($self->{base} eq "rip") {
 239                 sprintf "%s[%s]",$szmap{$sz},$self->{label};
 240             } else {
 241                 sprintf "%s[%s%s]",$szmap{$sz},
 242                                         $self->{label}?"$self->{label}+":"",
 243                                         $self->{base};
 244             }
 245         }
 246     }
 247 }
 248 { package register;     # pick up registers, which start with %.
 249     sub re {
 250         my      $class = shift; # muliple instances...
 251         my      $self = {};
 252         local   *line = shift;
 253         undef   $ret;
 254
 255         # optional * ---vvv--- appears in indirect jmp/call
 256         if ($line =~ /^(\*?)%(\w+)/) {
 257             bless $self,$class;
 258             $self->{asterisk} = $1;
 259             $self->{value} = $2;
 260             $ret = $self;
 261             $line = substr($line,@+[0]); $line =~ s/^\s+//;
 262         }
 263         $ret;
 264     }
 265     sub size {
 266         my      $self = shift;
 267         undef   $ret;
 268
 269         if    ($self->{value} =~ /^r[\d]+b$/i)  { $ret="b"; }
 270         elsif ($self->{value} =~ /^r[\d]+w$/i)  { $ret="w"; }
 271         elsif ($self->{value} =~ /^r[\d]+d$/i)  { $ret="l"; }
 272         elsif ($self->{value} =~ /^r[\w]+$/i)   { $ret="q"; }
 273         elsif ($self->{value} =~ /^[a-d][hl]$/i){ $ret="b"; }
 274         elsif ($self->{value} =~ /^[\w]{2}l$/i) { $ret="b"; }
 275         elsif ($self->{value} =~ /^[\w]{2}$/i)  { $ret="w"; }
 276         elsif ($self->{value} =~ /^e[a-z]{2}$/i){ $ret="l"; }
 277
 278         $ret;
 279     }
 280     sub out {
 281         my $self = shift;
 282         if (!$win64)    { sprintf "%s%%%s",$self->{asterisk},$self->{value}; }
 283         else            { $self->{value}; }
 284     }
 285 }
 286 { package label;        # pick up labels, which end with :
 287     sub re {
 288         my      $self = shift;  # single instance is enough...
 289         local   *line = shift;
 290         undef   $ret;
 291
 292         if ($line =~ /(^[\.\w]+\:)/) {
 293             $self->{value} = $1;
 294             $ret = $self;
 295             $line = substr($line,@+[0]); $line =~ s/^\s+//;
 296
 297             $self->{value} =~ s/\.L/\$L\$/ if ($win64);
 298         }
 299         $ret;
 300     }
 301     sub out {
 302         my $self = shift;
 303
 304         if (!$win64) {
 305             $self->{value};
 306         } elsif ($self->{value} ne "$current_function->{name}:") {
 307             $self->{value} .= ":" if ($masm && $ret!~m/^\$/);
 308             $self->{value};
 309         } elsif ($current_function->{abi} eq "svr4") {
 310             my $func =  "$current_function->{name}" .
 311                         ($nasm ? ":" : "\tPROC $current_function->{scope}") .
 312                         "\n";
 313             $func .= "  mov     QWORD${PTR}[8+rsp],rdi\t;WIN64 prologue\n";
 314             $func .= "  mov     QWORD${PTR}[16+rsp],rsi\n";
 315             $func .= "  mov     rax,rsp\n";
 316             $func .= "\$L\$SEH_begin_$current_function->{name}:";
 317             $func .= ":" if ($masm);
 318             $func .= "\n";
 319             my $narg = $current_function->{narg};
 320             $narg=6 if (!defined($narg));
 321             $func .= "  mov     rdi,rcx\n" if ($narg>0);
 322             $func .= "  mov     rsi,rdx\n" if ($narg>1);
 323             $func .= "  mov     rdx,r8\n"  if ($narg>2);
 324             $func .= "  mov     rcx,r9\n"  if ($narg>3);
 325             $func .= "  mov     r8,QWORD${PTR}[40+rsp]\n" if ($narg>4);
 326             $func .= "  mov     r9,QWORD${PTR}[48+rsp]\n" if ($narg>5);
 327             $func .= "\n";
 328         } else {
 329            "$current_function->{name}".
 330                         ($nasm ? ":" : "\tPROC $current_function->{scope}");
 331         }
 332     }
 333 }
 334 { package expr;         # pick up expressioins
 335     sub re {
 336         my      $self = shift;  # single instance is enough...
 337         local   *line = shift;
 338         undef   $ret;
 339
 340         if ($line =~ /(^[^,]+)/) {
 341             $self->{value} = $1;
 342             $ret = $self;
 343             $line = substr($line,@+[0]); $line =~ s/^\s+//;
 344
 345             $self->{value} =~ s/\.L/\$L\$/g if ($win64);
 346         }
 347         $ret;
 348     }
 349     sub out {
 350         my $self = shift;
 351         if ($nasm && opcode->mnemonic()=~m/^j/) {
 352             "NEAR ".$self->{value};
 353         } else {
 354             $self->{value};
 355         }
 356     }
 357 }
 358 { package directive;    # pick up directives, which start with .
 359     sub re {
 360         my      $self = shift;  # single instance is enough...
 361         local   *line = shift;
 362         undef   $ret;
 363         my      $dir;
 364         my      %opcode =       # lea 2f-1f(%rip),%dst; 1: nop; 2:
 365                 (       "%rax"=>0x01058d48,     "%rcx"=>0x010d8d48,
 366                         "%rdx"=>0x01158d48,     "%rbx"=>0x011d8d48,
 367                         "%rsp"=>0x01258d48,     "%rbp"=>0x012d8d48,
 368                         "%rsi"=>0x01358d48,     "%rdi"=>0x013d8d48,
 369                         "%r8" =>0x01058d4c,     "%r9" =>0x010d8d4c,
 370                         "%r10"=>0x01158d4c,     "%r11"=>0x011d8d4c,
 371                         "%r12"=>0x01258d4c,     "%r13"=>0x012d8d4c,
 372                         "%r14"=>0x01358d4c,     "%r15"=>0x013d8d4c      );
 373
 374         if ($line =~ /^\s*(\.\w+)/) {
 375             if (!$win64) {
 376                 $self->{value} = $1;
 377                 $line =~ s/\@abi\-omnipotent/\@function/;
 378                 $line =~ s/\@function.*/\@function/;
 379                 if ($line =~ /\.picmeup\s+(%r[\w]+)/i) {
 380                     $self->{value} = sprintf "\t.long\t0x%x,0x90000000",$opcode{$1};
 381                 } elsif ($line =~ /\.asciz\s+"(.*)"$/) {
 382                     $self->{value} = ".byte\t".join(",",unpack("C*",$1),0);
 383                 } elsif ($line =~ /\.extern/) {
 384                     $self->{value} = ""; # swallow extern
 385                 } else {
 386                     $self->{value} = $line;
 387                 }
 388                 $line = "";
 389                 return $self;
 390             }
 391
 392             $dir = $1;
 393             $ret = $self;
 394             undef $self->{value};
 395             $line = substr($line,@+[0]); $line =~ s/^\s+//;
 396             SWITCH: for ($dir) {
 397                 /\.text/    && do { my $v=undef;
 398                                     if ($nasm) {
 399                                         $v="section     .text code align=64\n";
 400                                     } else {
 401                                         $v="$current_segment\tENDS\n" if ($current_segment);
 402                                         $current_segment = ".text\$";
 403                                         $v.="$current_segment\tSEGMENT ";
 404                                         $v.=$masm>=$masmref ? "ALIGN(64)" : "PAGE";
 405                                         $v.=" 'CODE'";
 406                                     }
 407                                     $self->{value} = $v;
 408                                     last;
 409                                   };
 410                 /\.data/    && do { my $v=undef;
 411                                     if ($nasm) {
 412                                         $v="section     .data data align=8\n";
 413                                     } else {
 414                                         $v="$current_segment\tENDS\n" if ($current_segment);
 415                                         $current_segment = "_DATA";
 416                                         $v.="$current_segment\tSEGMENT";
 417                                     }
 418                                     $self->{value} = $v;
 419                                     last;
 420                                   };
 421                 /\.section/ && do { my $v=undef;
 422                                     if ($nasm) {
 423                                         $v="section     $line";
 424                                         if ($line=~/\.([px])data/) {
 425                                             $v.=" rdata align=";
 426                                             $v.=$1 eq "p"? 4 : 8;
 427                                         }
 428                                     } else {
 429                                         $v="$current_segment\tENDS\n" if ($current_segment);
 430                                         $v.="$line\tSEGMENT";
 431                                         if ($line=~/\.([px])data/) {
 432                                             $v.=" READONLY";
 433                                             $v.=" ALIGN(".($1 eq "p" ? 4 : 8).")" if ($masm>=$masmref);
 434                                         }
 435                                     }
 436                                     $current_segment = $line;
 437                                     $self->{value} = $v;
 438                                     last;
 439                                   };
 440                 /\.extern/  && do { $self->{value}  = "EXTERN\t".$line;
 441                                     $self->{value} .= ":NEAR" if ($masm);
 442                                     last;
 443                                   };
 444                 /\.globl|.global/
 445                             && do { $self->{value}  = $masm?"PUBLIC":"global";
 446                                     $self->{value} .= "\t".$line;
 447                                     $globals{$line} = $line;
 448                                     last;
 449                                   };
 450                 /\.type/    && do { ($sym,$type,$narg) = split(',',$line);
 451                                     if ($type eq "\@function") {
 452                                         undef $current_function;
 453                                         $current_function->{name} = $sym;
 454                                         $current_function->{abi}  = "svr4";
 455                                         $current_function->{narg} = $narg;
 456                                         $current_function->{scope} = defined($globals{$sym})?"PUBLIC":"PRIVATE";
 457                                     } elsif ($type eq "\@abi-omnipotent") {
 458                                         undef $current_function;
 459                                         $current_function->{name} = $sym;
 460                                         $current_function->{scope} = defined($globals{$sym})?"PUBLIC":"PRIVATE";
 461                                     }
 462                                     last;
 463                                   };
 464                 /\.size/    && do { if (defined($current_function)) {
 465                                         $self->{value}="\$L\$SEH_end_$current_function->{name}:";
 466                                         $self->{value}.=":\n$current_function->{name}\tENDP" if($masm);
 467                                         undef $current_function;
 468                                     }
 469                                     last;
 470                                   };
 471                 /\.align/   && do { $self->{value} = "ALIGN\t".$line; last; };
 472                 /\.(byte|value|long|quad)/
 473                             && do { my $sz  = substr($1,0,1);
 474                                     my @arr = split(',',$line);
 475                                     my $last = pop(@arr);
 476                                     my $conv = sub  {   my $var=shift;
 477                                                         $var=~s/0x([0-9a-f]+)/0$1h/ig;
 478                                                         $var=~s/\.L/\$L\$/g;
 479                                                         if ($current_segment=~/.[px]data/)
 480                                                         { $var=~s/([_a-z\$\@][_a-z0-9\$\@]*)/$nasm?"$1 wrt ..imagebase":"imagerel $1"/egi; }
 481                                                         $var;
 482                                                     };
 483
 484                                     $sz =~ tr/bvlq/BWDQ/;
 485                                     $self->{value} = "\tD$sz\t";
 486                                     for (@arr) { $self->{value} .= &$conv($_).","; }
 487                                     $self->{value} .= &$conv($last);
 488                                     last;
 489                                   };
 490                 /\.picmeup/ && do { $self->{value} = sprintf"\tDD\t0%Xh,090000000h",$opcode{$line};
 491                                     last;
 492                                   };
 493                 /\.asciz/   && do { if ($line =~ /^"(.*)"$/) {
 494                                         my @str=unpack("C*",$1);
 495                                         push @str,0;
 496                                         while ($#str>15) {
 497                                             $self->{value}.="DB\t"
 498                                                 .join(",",@str[0..15])."\n";
 499                                             foreach (0..15) { shift @str; }
 500                                         }
 501                                         $self->{value}.="DB\t"
 502                                                 .join(",",@str) if (@str);
 503                                     }
 504                                     last;
 505                                   };
 506             }
 507             $line = "";
 508         }
 509
 510         $ret;
 511     }
 512     sub out {
 513         my $self = shift;
 514         $self->{value};
 515     }
 516 }
 517
 518 if ($nasm) {
 519     print <<___;
 520 default rel
 521 ___
 522 } elsif ($masm) {
 523     print <<___;
 524 OPTION  DOTNAME
 525 ___
 526 }
 527 while($line=<>) {
 528
 529     chomp($line);
 530
 531     $line =~ s|[#!].*$||;       # get rid of asm-style comments...
 532     $line =~ s|/\*.*\*/||;      # ... and C-style comments...
 533     $line =~ s|^\s+||;          # ... and skip white spaces in beginning
 534
 535     undef $label;
 536     undef $opcode;
 537     undef $dst;
 538     undef $src;
 539     undef $sz;
 540
 541     if ($label=label->re(\$line))       { print $label->out(); }
 542
 543     if (directive->re(\$line)) {
 544         printf "%s",directive->out();
 545     } elsif ($opcode=opcode->re(\$line)) { ARGUMENT: {
 546
 547         if ($src=register->re(\$line))  { opcode->size($src->size()); }
 548         elsif ($src=const->re(\$line))  { }
 549         elsif ($src=ea->re(\$line))     { }
 550         elsif ($src=expr->re(\$line))   { }
 551
 552         last ARGUMENT if ($line !~ /^,/);
 553
 554         $line = substr($line,1); $line =~ s/^\s+//;
 555
 556         if ($dst=register->re(\$line))  { opcode->size($dst->size()); }
 557         elsif ($dst=const->re(\$line))  { }
 558         elsif ($dst=ea->re(\$line))     { }
 559
 560         } # ARGUMENT:
 561
 562         $sz=opcode->size();
 563
 564         if (defined($dst)) {
 565             if (!$win64) {
 566                 printf "\t%s\t%s,%s",   $opcode->out($dst->size()),
 567                                         $src->out($sz),$dst->out($sz);
 568             } else {
 569                 undef $sz if ($nasm && $opcode->mnemonic() eq "lea");
 570                 printf "\t%s\t%s,%s",   $opcode->out(),
 571                                         $dst->out($sz),$src->out($sz);
 572             }
 573         } elsif (defined($src)) {
 574             printf "\t%s\t%s",$opcode->out(),$src->out($sz);
 575         } else {
 576             printf "\t%s",$opcode->out();
 577         }
 578     }
 579
 580     print $line,"\n";
 581 }
 582
 583 print "\n$current_segment\tENDS\nEND\n" if ($current_segment && $masm);
 584
 585 close STDOUT;
 586
 587 \f#################################################
 588 # Cross-reference x86_64 ABI "card"
 589 #
 590 #               Unix            Win64
 591 # %rax          *               *
 592 # %rbx          -               -
 593 # %rcx          #4              #1
 594 # %rdx          #3              #2
 595 # %rsi          #2              -
 596 # %rdi          #1              -
 597 # %rbp          -               -
 598 # %rsp          -               -
 599 # %r8           #5              #3
 600 # %r9           #6              #4
 601 # %r10          *               *
 602 # %r11          *               *
 603 # %r12          -               -
 604 # %r13          -               -
 605 # %r14          -               -
 606 # %r15          -               -
 607 #
 608 # (*)   volatile register
 609 # (-)   preserved by callee
 610 # (#)   Nth argument, volatile
 611 #
 612 # In Unix terms top of stack is argument transfer area for arguments
 613 # which could not be accomodated in registers. Or in other words 7th
 614 # [integer] argument resides at 8(%rsp) upon function entry point.
 615 # 128 bytes above %rsp constitute a "red zone" which is not touched
 616 # by signal handlers and can be used as temporal storage without
 617 # allocating a frame.
 618 #
 619 # In Win64 terms N*8 bytes on top of stack is argument transfer area,
 620 # which belongs to/can be overwritten by callee. N is the number of
 621 # arguments passed to callee, *but* not less than 4! This means that
 622 # upon function entry point 5th argument resides at 40(%rsp), as well
 623 # as that 32 bytes from 8(%rsp) can always be used as temporal
 624 # storage [without allocating a frame]. One can actually argue that
 625 # one can assume a "red zone" above stack pointer under Win64 as well.
 626 # Point is that at apparently no occasion Windows kernel would alter
 627 # the area above user stack pointer in true asynchronous manner...
 628 #
 629 # All the above means that if assembler programmer adheres to Unix
 630 # register and stack layout, but disregards the "red zone" existense,
 631 # it's possible to use following prologue and epilogue to "gear" from
 632 # Unix to Win64 ABI in leaf functions with not more than 6 arguments.
 633 #
 634 # omnipotent_function:
 635 # ifdef WIN64
 636 #       movq    %rdi,8(%rsp)
 637 #       movq    %rsi,16(%rsp)
 638 #       movq    %rcx,%rdi       ; if 1st argument is actually present
 639 #       movq    %rdx,%rsi       ; if 2nd argument is actually ...
 640 #       movq    %r8,%rdx        ; if 3rd argument is ...
 641 #       movq    %r9,%rcx        ; if 4th argument ...
 642 #       movq    40(%rsp),%r8    ; if 5th ...
 643 #       movq    48(%rsp),%r9    ; if 6th ...
 644 # endif
 645 #       ...
 646 # ifdef WIN64
 647 #       movq    8(%rsp),%rdi
 648 #       movq    16(%rsp),%rsi
 649 # endif
 650 #       ret
 651 #
 652 \f#################################################
 653 # Win64 SEH, Structured Exception Handling.
 654 #
 655 # Unlike on Unix systems(*) lack of Win64 stack unwinding information
 656 # has undesired side-effect at run-time: if an exception is raised in
 657 # assembler subroutine such as those in question (basically we're
 658 # referring to segmentation violations caused by malformed input
 659 # parameters), the application is briskly terminated without invoking
 660 # any exception handlers, most notably without generating memory dump
 661 # or any user notification whatsoever. This poses a problem. It's
 662 # possible to address it by registering custom language-specific
 663 # handler that would restore processor context to the state at
 664 # subroutine entry point and return "exception is not handled, keep
 665 # unwinding" code. Writing such handler can be a challenge... But it's
 666 # doable, though requires certain coding convention. Consider following
 667 # snippet:
 668 #
 669 # .type function,@function
 670 # function:
 671 #       movq    %rsp,%rax       # copy rsp to volatile register
 672 #       pushq   %r15            # save non-volatile registers
 673 #       pushq   %rbx
 674 #       pushq   %rbp
 675 #       movq    %rsp,%r11
 676 #       subq    %rdi,%r11       # prepare [variable] stack frame
 677 #       andq    $-64,%r11
 678 #       movq    %rax,0(%r11)    # check for exceptions
 679 #       movq    %r11,%rsp       # allocate [variable] stack frame
 680 #       movq    %rax,0(%rsp)    # save original rsp value
 681 # magic_point:
 682 #       ...
 683 #       movq    0(%rsp),%rcx    # pull original rsp value
 684 #       movq    -24(%rcx),%rbp  # restore non-volatile registers
 685 #       movq    -16(%rcx),%rbx
 686 #       movq    -8(%rcx),%r15
 687 #       movq    %rcx,%rsp       # restore original rsp
 688 #       ret
 689 # .size function,.-function
 690 #
 691 # The key is that up to magic_point copy of original rsp value remains
 692 # in chosen volatile register and no non-volatile register, except for
 693 # rsp, is modified. While past magic_point rsp remains constant till
 694 # the very end of the function. In this case custom language-specific
 695 # exception handler would look like this:
 696 #
 697 # EXCEPTION_DISPOSITION handler (EXCEPTION_RECORD *rec,ULONG64 frame,
 698 #               CONTEXT *context,DISPATCHER_CONTEXT *disp)
 699 # {     ULONG64 *rsp = (ULONG64 *)context->Rax;
 700 #       if (context->Rip >= magic_point)
 701 #       {   rsp = ((ULONG64 **)context->Rsp)[0];
 702 #           context->Rbp = rsp[-3];
 703 #           context->Rbx = rsp[-2];
 704 #           context->R15 = rsp[-1];
 705 #       }
 706 #       context->Rsp = (ULONG64)rsp;
 707 #       context->Rdi = rsp[1];
 708 #       context->Rsi = rsp[2];
 709 #
 710 #       memcpy (disp->ContextRecord,context,sizeof(CONTEXT));
 711 #       RtlVirtualUnwind(UNW_FLAG_NHANDLER,disp->ImageBase,
 712 #               dips->ControlPc,disp->FunctionEntry,disp->ContextRecord,
 713 #               &disp->HandlerData,&disp->EstablisherFrame,NULL);
 714 #       return ExceptionContinueSearch;
 715 # }
 716 #
 717 # It's appropriate to implement this handler in assembler, directly in
 718 # function's module. In order to do that one has to know members'
 719 # offsets in CONTEXT and DISPATCHER_CONTEXT structures and some constant
 720 # values. Here they are:
 721 #
 722 #       CONTEXT.Rax                             120
 723 #       CONTEXT.Rcx                             128
 724 #       CONTEXT.Rdx                             136
 725 #       CONTEXT.Rbx                             144
 726 #       CONTEXT.Rsp                             152
 727 #       CONTEXT.Rbp                             160
 728 #       CONTEXT.Rsi                             168
 729 #       CONTEXT.Rdi                             176
 730 #       CONTEXT.R8                              184
 731 #       CONTEXT.R9                              192
 732 #       CONTEXT.R10                             200
 733 #       CONTEXT.R11                             208
 734 #       CONTEXT.R12                             216
 735 #       CONTEXT.R13                             224
 736 #       CONTEXT.R14                             232
 737 #       CONTEXT.R15                             240
 738 #       CONTEXT.Rip                             248
 739 #       sizeof(CONTEXT)                         1232
 740 #       DISPATCHER_CONTEXT.ControlPc            0
 741 #       DISPATCHER_CONTEXT.ImageBase            8
 742 #       DISPATCHER_CONTEXT.FunctionEntry        16
 743 #       DISPATCHER_CONTEXT.EstablisherFrame     24
 744 #       DISPATCHER_CONTEXT.TargetIp             32
 745 #       DISPATCHER_CONTEXT.ContextRecord        40
 746 #       DISPATCHER_CONTEXT.LanguageHandler      48
 747 #       DISPATCHER_CONTEXT.HandlerData          56
 748 #       UNW_FLAG_NHANDLER                       0
 749 #       ExceptionContinueSearch                 1
 750 #
 751 # In order to tie the handler to the function one has to compose
 752 # couple of structures: one for .xdata segment and one for .pdata.
 753 #
 754 # UNWIND_INFO structure for .xdata segment would be
 755 #
 756 # function_unwind_info:
 757 #       .byte   9,0,0,0
 758 #       .long   handler
 759 #
 760 # This structure designates exception handler for a function with
 761 # zero-length prologue, no stack frame or frame register.
 762 #
 763 # To facilitate composing of .pdata structures, auto-generated "gear"
 764 # prologue copies rsp value to rax and denotes next instruction with
 765 # $L$SEH_begin_{function_name} label. This essentially defines the SEH
 766 # styling rule mentioned in the beginning. Position of this label is
 767 # chosen in such manner that possible exceptions raised in the "gear"
 768 # prologue would be accounted to caller and unwound from latter's frame.
 769 # End of function is marked with respective $L$SEH_end_{function_name}
 770 # label. To summarize, .pdata segment would contain
 771 #
 772 #       .long   $L$SEH_begin_function
 773 #       .long   $L$SEH_end_function
 774 #       .long   function_unwind_info
 775 #
 776 # Reference to functon_unwind_info from .xdata segment is the anchor.
 777 # In case you wonder why references are 32-bit .longs and not 64-bit
 778 # .quads. References put into these two segments are required to be
 779 # *relative* to the base address of the current binary module, a.k.a.
 780 # image base. No Win64 module, be it .exe or .dll, can be larger than
 781 # 2GB and thus such relative references can be and are accommodated in
 782 # 32 bits.
 783 #
 784 # Having reviewed the example function code, one can argue that "movq
 785 # %rsp,%rax" above is redundant. It is not! Keep in mind that on Unix
 786 # rax would contain an undefined value. If this "offends" you, use
 787 # another register and refrain from modifying rax till magic_point is
 788 # reached, i.e. as if it was a non-volatile register. If more registers
 789 # are required prior [variable] frame setup is completed, note that
 790 # nobody says that you can have only one "magic point." You can
 791 # "liberate" non-volatile registers by denoting last stack off-load
 792 # instruction and reflecting it in finer grade unwind logic in handler.
 793 # After all, isn't it why it's called *language-specific* handler...
 794 #
 795 # Attentive reader can notice that exceptions would be mishandled in
 796 # auto-generated "gear" epilogue. Well, exception effectively can't
 797 # occur there, because if memory area used by it was subject to
 798 # segmentation violation, then it would be raised upon call to the
 799 # function (and as already mentioned be accounted to caller, which is
 800 # not a problem). If you're still not comfortable, then define tail
 801 # "magic point" just prior ret instruction and have handler treat it...
 802 #
 803 # (*)   Note that we're talking about run-time, not debug-time. Lack of
 804 #       unwind information makes debugging hard on both Windows and
 805 #       Unix. "Unlike" referes to the fact that on Unix signal handler
 806 #       will always be invoked, core dumped and appropriate exit code
 807 #       returned to parent (for user notification).