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. Stick to explicit ip-relative addressing. If you have to use
  45 #    GOTPCREL addressing, stick to mov symbol@GOTPCREL(%rip),%r??.
  46 #    Both are recognized and translated to proper Win64 addressing
  47 #    modes. To support legacy code a synthetic directive, .picmeup,
  48 #    is implemented. It puts address of the *next* instruction into
  49 #    target register, e.g.:
  50 #
  51 #               .picmeup        %rax
  52 #               lea             .Label-.(%rax),%rax
  53 #
  54 # 8. In order to provide for structured exception handling unified
  55 #    Win64 prologue copies %rsp value to %rax. For further details
  56 #    see SEH paragraph at the end.
  57 # 9. .init segment is allowed to contain calls to functions only.
  58 \f
  59 my $flavour = shift;
  60 my $output  = shift;
  61 if ($flavour =~ /\./) { $output = $flavour; undef $flavour; }
  62
  63 { my ($stddev,$stdino,@junk)=stat(STDOUT);
  64   my ($outdev,$outino,@junk)=stat($output);
  65
  66     open STDOUT,">$output" || die "can't open $output: $!"
  67         if ($stddev!=$outdev || $stdino!=$outino);
  68 }
  69
  70 my $gas=1;      $gas=0 if ($output =~ /\.asm$/);
  71 my $elf=1;      $elf=0 if (!$gas);
  72 my $win64=0;
  73 my $prefix="";
  74 my $decor=".L";
  75
  76 my $masmref=8 + 50727*2**-32;   # 8.00.50727 shipped with VS2005
  77 my $masm=0;
  78 my $PTR=" PTR";
  79
  80 my $nasmref=2.03;
  81 my $nasm=0;
  82
  83 if    ($flavour eq "mingw64")   { $gas=1; $elf=0; $win64=1; $prefix="_"; }
  84 elsif ($flavour eq "macosx")    { $gas=1; $elf=0; $prefix="_"; $decor="L\$"; }
  85 elsif ($flavour eq "masm")      { $gas=0; $elf=0; $masm=$masmref; $win64=1; $decor="\$L\$"; }
  86 elsif ($flavour eq "nasm")      { $gas=0; $elf=0; $nasm=$nasmref; $win64=1; $decor="\$L\$"; $PTR=""; }
  87 elsif (!$gas)
  88 {   if ($ENV{ASM} =~ m/nasm/ && `nasm -v` =~ m/version ([0-9]+)\.([0-9]+)/i)
  89     {   $nasm = $1 + $2*0.01; $PTR="";  }
  90     elsif (`ml64 2>&1` =~ m/Version ([0-9]+)\.([0-9]+)(\.([0-9]+))?/)
  91     {   $masm = $1 + $2*2**-16 + $4*2**-32;   }
  92     die "no assembler found on %PATH" if (!($nasm || $masm));
  93     $win64=1;
  94     $elf=0;
  95     $decor="\$L\$";
  96 }
  97
  98 my $current_segment;
  99 my $current_function;
 100 my %globals;
 101
 102 { package opcode;       # pick up opcodes
 103     sub re {
 104         my      $self = shift;  # single instance in enough...
 105         local   *line = shift;
 106         undef   $ret;
 107
 108         if ($line =~ /^([a-z][a-z0-9]*)/i) {
 109             $self->{op} = $1;
 110             $ret = $self;
 111             $line = substr($line,@+[0]); $line =~ s/^\s+//;
 112
 113             undef $self->{sz};
 114             if ($self->{op} =~ /^(movz)b.*/) {  # movz is pain...
 115                 $self->{op} = $1;
 116                 $self->{sz} = "b";
 117             } elsif ($self->{op} =~ /call|jmp/) {
 118                 $self->{sz} = "";
 119             } elsif ($self->{op} =~ /^p/ && $' !~ /^(ush|op)/) { # SSEn
 120                 $self->{sz} = "";
 121             } elsif ($self->{op} =~ /([a-z]{3,})([qlwb])$/) {
 122                 $self->{op} = $1;
 123                 $self->{sz} = $2;
 124             }
 125         }
 126         $ret;
 127     }
 128     sub size {
 129         my $self = shift;
 130         my $sz   = shift;
 131         $self->{sz} = $sz if (defined($sz) && !defined($self->{sz}));
 132         $self->{sz};
 133     }
 134     sub out {
 135         my $self = shift;
 136         if ($gas) {
 137             if ($self->{op} eq "movz") {        # movz is pain...
 138                 sprintf "%s%s%s",$self->{op},$self->{sz},shift;
 139             } elsif ($self->{op} =~ /^set/) {
 140                 "$self->{op}";
 141             } elsif ($self->{op} eq "ret") {
 142                 my $epilogue = "";
 143                 if ($win64 && $current_function->{abi} eq "svr4") {
 144                     $epilogue = "movq   8(%rsp),%rdi\n\t" .
 145                                 "movq   16(%rsp),%rsi\n\t";
 146                 }
 147                 $epilogue . ".byte      0xf3,0xc3";
 148             } elsif ($self->{op} eq "call" && !$elf && $current_segment eq ".init") {
 149                 ".p2align\t3\n\t.quad";
 150             } else {
 151                 "$self->{op}$self->{sz}";
 152             }
 153         } else {
 154             $self->{op} =~ s/^movz/movzx/;
 155             if ($self->{op} eq "ret") {
 156                 $self->{op} = "";
 157                 if ($win64 && $current_function->{abi} eq "svr4") {
 158                     $self->{op} = "mov  rdi,QWORD${PTR}[8+rsp]\t;WIN64 epilogue\n\t".
 159                                   "mov  rsi,QWORD${PTR}[16+rsp]\n\t";
 160                 }
 161                 $self->{op} .= "DB\t0F3h,0C3h\t\t;repret";
 162             } elsif ($self->{op} =~ /^(pop|push)f/) {
 163                 $self->{op} .= $self->{sz};
 164             } elsif ($self->{op} eq "call" && $current_segment eq ".CRT\$XCU") {
 165                 $self->{op} = "ALIGN\t8\n\tDQ";
 166             }
 167             $self->{op};
 168         }
 169     }
 170     sub mnemonic {
 171         my $self=shift;
 172         my $op=shift;
 173         $self->{op}=$op if (defined($op));
 174         $self->{op};
 175     }
 176 }
 177 { package const;        # pick up constants, which start with $
 178     sub re {
 179         my      $self = shift;  # single instance in enough...
 180         local   *line = shift;
 181         undef   $ret;
 182
 183         if ($line =~ /^\$([^,]+)/) {
 184             $self->{value} = $1;
 185             $ret = $self;
 186             $line = substr($line,@+[0]); $line =~ s/^\s+//;
 187         }
 188         $ret;
 189     }
 190     sub out {
 191         my $self = shift;
 192
 193         if ($gas) {
 194             # Solaris /usr/ccs/bin/as can't handle multiplications
 195             # in $self->{value}
 196             $self->{value} =~ s/(?<![0-9a-f])(0[x0-9a-f]+)/oct($1)/egi;
 197             $self->{value} =~ s/([0-9]+\s*[\*\/\%]\s*[0-9]+)/eval($1)/eg;
 198             sprintf "\$%s",$self->{value};
 199         } else {
 200             $self->{value} =~ s/(0b[0-1]+)/oct($1)/eig;
 201             $self->{value} =~ s/0x([0-9a-f]+)/0$1h/ig if ($masm);
 202             sprintf "%s",$self->{value};
 203         }
 204     }
 205 }
 206 { package ea;           # pick up effective addresses: expr(%reg,%reg,scale)
 207     sub re {
 208         my      $self = shift;  # single instance in enough...
 209         local   *line = shift;
 210         undef   $ret;
 211
 212         # optional * ---vvv--- appears in indirect jmp/call
 213         if ($line =~ /^(\*?)([^\(,]*)\(([%\w,]+)\)/) {
 214             $self->{asterisk} = $1;
 215             $self->{label} = $2;
 216             ($self->{base},$self->{index},$self->{scale})=split(/,/,$3);
 217             $self->{scale} = 1 if (!defined($self->{scale}));
 218             $ret = $self;
 219             $line = substr($line,@+[0]); $line =~ s/^\s+//;
 220
 221             if ($win64 && $self->{label} =~ s/\@GOTPCREL//) {
 222                 die if (opcode->mnemonic() ne "mov");
 223                 opcode->mnemonic("lea");
 224             }
 225             $self->{base}  =~ s/^%//;
 226             $self->{index} =~ s/^%// if (defined($self->{index}));
 227         }
 228         $ret;
 229     }
 230     sub size {}
 231     sub out {
 232         my $self = shift;
 233         my $sz = shift;
 234
 235         $self->{label} =~ s/([_a-z][_a-z0-9]*)/$globals{$1} or $1/gei;
 236         $self->{label} =~ s/\.L/$decor/g;
 237
 238         # Silently convert all EAs to 64-bit. This is required for
 239         # elder GNU assembler and results in more compact code,
 240         # *but* most importantly AES module depends on this feature!
 241         $self->{index} =~ s/^[er](.?[0-9xpi])[d]?$/r\1/;
 242         $self->{base}  =~ s/^[er](.?[0-9xpi])[d]?$/r\1/;
 243
 244         if ($gas) {
 245             # Solaris /usr/ccs/bin/as can't handle multiplications
 246             # in $self->{label}, new gas requires sign extension...
 247             use integer;
 248             $self->{label} =~ s/(?<![0-9a-f])(0[x0-9a-f]+)/oct($1)/egi;
 249             $self->{label} =~ s/([0-9]+\s*[\*\/\%]\s*[0-9]+)/eval($1)/eg;
 250             $self->{label} =~ s/([0-9]+)/$1<<32>>32/eg;
 251             $self->{label} =~ s/^___imp_/__imp__/   if ($flavour eq "mingw64");
 252
 253             if (defined($self->{index})) {
 254                 sprintf "%s%s(%%%s,%%%s,%d)",$self->{asterisk},
 255                                         $self->{label},$self->{base},
 256                                         $self->{index},$self->{scale};
 257             } else {
 258                 sprintf "%s%s(%%%s)",   $self->{asterisk},$self->{label},$self->{base};
 259             }
 260         } else {
 261             %szmap = ( b=>"BYTE$PTR", w=>"WORD$PTR", l=>"DWORD$PTR", q=>"QWORD$PTR" );
 262
 263             $self->{label} =~ s/\./\$/g;
 264             $self->{label} =~ s/0x([0-9a-f]+)/0$1h/ig;
 265             $self->{label} = "($self->{label})" if ($self->{label} =~ /[\*\+\-\/]/);
 266             $sz="q" if ($self->{asterisk});
 267
 268             if (defined($self->{index})) {
 269                 sprintf "%s[%s%s*%d+%s]",$szmap{$sz},
 270                                         $self->{label}?"$self->{label}+":"",
 271                                         $self->{index},$self->{scale},
 272                                         $self->{base};
 273             } elsif ($self->{base} eq "rip") {
 274                 sprintf "%s[%s]",$szmap{$sz},$self->{label};
 275             } else {
 276                 sprintf "%s[%s%s]",$szmap{$sz},
 277                                         $self->{label}?"$self->{label}+":"",
 278                                         $self->{base};
 279             }
 280         }
 281     }
 282 }
 283 { package register;     # pick up registers, which start with %.
 284     sub re {
 285         my      $class = shift; # muliple instances...
 286         my      $self = {};
 287         local   *line = shift;
 288         undef   $ret;
 289
 290         # optional * ---vvv--- appears in indirect jmp/call
 291         if ($line =~ /^(\*?)%(\w+)/) {
 292             bless $self,$class;
 293             $self->{asterisk} = $1;
 294             $self->{value} = $2;
 295             $ret = $self;
 296             $line = substr($line,@+[0]); $line =~ s/^\s+//;
 297         }
 298         $ret;
 299     }
 300     sub size {
 301         my      $self = shift;
 302         undef   $ret;
 303
 304         if    ($self->{value} =~ /^r[\d]+b$/i)  { $ret="b"; }
 305         elsif ($self->{value} =~ /^r[\d]+w$/i)  { $ret="w"; }
 306         elsif ($self->{value} =~ /^r[\d]+d$/i)  { $ret="l"; }
 307         elsif ($self->{value} =~ /^r[\w]+$/i)   { $ret="q"; }
 308         elsif ($self->{value} =~ /^[a-d][hl]$/i){ $ret="b"; }
 309         elsif ($self->{value} =~ /^[\w]{2}l$/i) { $ret="b"; }
 310         elsif ($self->{value} =~ /^[\w]{2}$/i)  { $ret="w"; }
 311         elsif ($self->{value} =~ /^e[a-z]{2}$/i){ $ret="l"; }
 312
 313         $ret;
 314     }
 315     sub out {
 316         my $self = shift;
 317         if ($gas)       { sprintf "%s%%%s",$self->{asterisk},$self->{value}; }
 318         else            { $self->{value}; }
 319     }
 320 }
 321 { package label;        # pick up labels, which end with :
 322     sub re {
 323         my      $self = shift;  # single instance is enough...
 324         local   *line = shift;
 325         undef   $ret;
 326
 327         if ($line =~ /(^[\.\w]+)\:/) {
 328             $self->{value} = $1;
 329             $ret = $self;
 330             $line = substr($line,@+[0]); $line =~ s/^\s+//;
 331
 332             $self->{value} =~ s/^\.L/$decor/;
 333         }
 334         $ret;
 335     }
 336     sub out {
 337         my $self = shift;
 338
 339         if ($gas) {
 340             my $func = ($globals{$self->{value}} or $self->{value}) . ":";
 341             if ($win64  &&
 342                         $current_function->{name} eq $self->{value} &&
 343                         $current_function->{abi} eq "svr4") {
 344                 $func .= "\n";
 345                 $func .= "      movq    %rdi,8(%rsp)\n";
 346                 $func .= "      movq    %rsi,16(%rsp)\n";
 347                 $func .= "      movq    %rsp,%rax\n";
 348                 $func .= "${decor}SEH_begin_$current_function->{name}:\n";
 349                 my $narg = $current_function->{narg};
 350                 $narg=6 if (!defined($narg));
 351                 $func .= "      movq    %rcx,%rdi\n" if ($narg>0);
 352                 $func .= "      movq    %rdx,%rsi\n" if ($narg>1);
 353                 $func .= "      movq    %r8,%rdx\n"  if ($narg>2);
 354                 $func .= "      movq    %r9,%rcx\n"  if ($narg>3);
 355                 $func .= "      movq    40(%rsp),%r8\n" if ($narg>4);
 356                 $func .= "      movq    48(%rsp),%r9\n" if ($narg>5);
 357             }
 358             $func;
 359         } elsif ($self->{value} ne "$current_function->{name}") {
 360             $self->{value} .= ":" if ($masm && $ret!~m/^\$/);
 361             $self->{value} . ":";
 362         } elsif ($win64 && $current_function->{abi} eq "svr4") {
 363             my $func =  "$current_function->{name}" .
 364                         ($nasm ? ":" : "\tPROC $current_function->{scope}") .
 365                         "\n";
 366             $func .= "  mov     QWORD${PTR}[8+rsp],rdi\t;WIN64 prologue\n";
 367             $func .= "  mov     QWORD${PTR}[16+rsp],rsi\n";
 368             $func .= "  mov     rax,rsp\n";
 369             $func .= "${decor}SEH_begin_$current_function->{name}:";
 370             $func .= ":" if ($masm);
 371             $func .= "\n";
 372             my $narg = $current_function->{narg};
 373             $narg=6 if (!defined($narg));
 374             $func .= "  mov     rdi,rcx\n" if ($narg>0);
 375             $func .= "  mov     rsi,rdx\n" if ($narg>1);
 376             $func .= "  mov     rdx,r8\n"  if ($narg>2);
 377             $func .= "  mov     rcx,r9\n"  if ($narg>3);
 378             $func .= "  mov     r8,QWORD${PTR}[40+rsp]\n" if ($narg>4);
 379             $func .= "  mov     r9,QWORD${PTR}[48+rsp]\n" if ($narg>5);
 380             $func .= "\n";
 381         } else {
 382            "$current_function->{name}".
 383                         ($nasm ? ":" : "\tPROC $current_function->{scope}");
 384         }
 385     }
 386 }
 387 { package expr;         # pick up expressioins
 388     sub re {
 389         my      $self = shift;  # single instance is enough...
 390         local   *line = shift;
 391         undef   $ret;
 392
 393         if ($line =~ /(^[^,]+)/) {
 394             $self->{value} = $1;
 395             $ret = $self;
 396             $line = substr($line,@+[0]); $line =~ s/^\s+//;
 397
 398             $self->{value} =~ s/\@PLT// if (!$elf);
 399             $self->{value} =~ s/([_a-z][_a-z0-9]*)/$globals{$1} or $1/gei;
 400             $self->{value} =~ s/\.L/$decor/g;
 401         }
 402         $ret;
 403     }
 404     sub out {
 405         my $self = shift;
 406         if ($nasm && opcode->mnemonic()=~m/^j/) {
 407             "NEAR ".$self->{value};
 408         } else {
 409             $self->{value};
 410         }
 411     }
 412 }
 413 { package directive;    # pick up directives, which start with .
 414     sub re {
 415         my      $self = shift;  # single instance is enough...
 416         local   *line = shift;
 417         undef   $ret;
 418         my      $dir;
 419         my      %opcode =       # lea 2f-1f(%rip),%dst; 1: nop; 2:
 420                 (       "%rax"=>0x01058d48,     "%rcx"=>0x010d8d48,
 421                         "%rdx"=>0x01158d48,     "%rbx"=>0x011d8d48,
 422                         "%rsp"=>0x01258d48,     "%rbp"=>0x012d8d48,
 423                         "%rsi"=>0x01358d48,     "%rdi"=>0x013d8d48,
 424                         "%r8" =>0x01058d4c,     "%r9" =>0x010d8d4c,
 425                         "%r10"=>0x01158d4c,     "%r11"=>0x011d8d4c,
 426                         "%r12"=>0x01258d4c,     "%r13"=>0x012d8d4c,
 427                         "%r14"=>0x01358d4c,     "%r15"=>0x013d8d4c      );
 428
 429         if ($line =~ /^\s*(\.\w+)/) {
 430             $dir = $1;
 431             $ret = $self;
 432             undef $self->{value};
 433             $line = substr($line,@+[0]); $line =~ s/^\s+//;
 434
 435             SWITCH: for ($dir) {
 436                 /\.picmeup/ && do { if ($line =~ /(%r[\w]+)/i) {
 437                                         $dir="\t.long";
 438                                         $line=sprintf "0x%x,0x90000000",$opcode{$1};
 439                                     }
 440                                     last;
 441                                   };
 442                 /\.global|\.globl|\.extern/
 443                             && do { $globals{$line} = $prefix . $line;
 444                                     $line = $globals{$line} if ($prefix);
 445                                     last;
 446                                   };
 447                 /\.type/    && do { ($sym,$type,$narg) = split(',',$line);
 448                                     if ($type eq "\@function") {
 449                                         undef $current_function;
 450                                         $current_function->{name} = $sym;
 451                                         $current_function->{abi}  = "svr4";
 452                                         $current_function->{narg} = $narg;
 453                                         $current_function->{scope} = defined($globals{$sym})?"PUBLIC":"PRIVATE";
 454                                     } elsif ($type eq "\@abi-omnipotent") {
 455                                         undef $current_function;
 456                                         $current_function->{name} = $sym;
 457                                         $current_function->{scope} = defined($globals{$sym})?"PUBLIC":"PRIVATE";
 458                                     }
 459                                     $line =~ s/\@abi\-omnipotent/\@function/;
 460                                     $line =~ s/\@function.*/\@function/;
 461                                     last;
 462                                   };
 463                 /\.asciz/   && do { if ($line =~ /^"(.*)"$/) {
 464                                         $dir  = ".byte";
 465                                         $line = join(",",unpack("C*",$1),0);
 466                                     }
 467                                     last;
 468                                   };
 469                 /\.rva|\.long|\.quad/
 470                             && do { $line =~ s/([_a-z][_a-z0-9]*)/$globals{$1} or $1/gei;
 471                                     $line =~ s/\.L/$decor/g;
 472                                     last;
 473                                   };
 474             }
 475
 476             if ($gas) {
 477                 $self->{value} = $dir . "\t" . $line;
 478
 479                 if ($dir =~ /\.extern/) {
 480                     $self->{value} = ""; # swallow extern
 481                 } elsif (!$elf && $dir =~ /\.type/) {
 482                     $self->{value} = "";
 483                     $self->{value} = ".def\t" . ($globals{$1} or $1) . ";\t" .
 484                                 (defined($globals{$1})?".scl 2;":".scl 3;") .
 485                                 "\t.type 32;\t.endef"
 486                                 if ($win64 && $line =~ /([^,]+),\@function/);
 487                 } elsif (!$elf && $dir =~ /\.size/) {
 488                     $self->{value} = "";
 489                     if (defined($current_function)) {
 490                         $self->{value} .= "${decor}SEH_end_$current_function->{name}:"
 491                                 if ($win64 && $current_function->{abi} eq "svr4");
 492                         undef $current_function;
 493                     }
 494                 } elsif (!$elf && $dir =~ /\.align/) {
 495                     $self->{value} = ".p2align\t" . (log($line)/log(2));
 496                 } elsif ($dir eq ".section") {
 497                     $current_segment=$line;
 498                     if (!$elf && $current_segment eq ".init") {
 499                         if      ($flavour eq "macosx")  { $self->{value} = ".mod_init_func"; }
 500                         elsif   ($flavour eq "mingw64") { $self->{value} = ".section\t.ctors"; }
 501                     }
 502                 } elsif ($dir =~ /\.(text|data)/) {
 503                     $current_segment=".$1";
 504                 }
 505                 $line = "";
 506                 return $self;
 507             }
 508
 509             # non-gas case or nasm/masm
 510             SWITCH: for ($dir) {
 511                 /\.text/    && do { my $v=undef;
 512                                     if ($nasm) {
 513                                         $v="section     .text code align=64\n";
 514                                     } else {
 515                                         $v="$current_segment\tENDS\n" if ($current_segment);
 516                                         $current_segment = ".text\$";
 517                                         $v.="$current_segment\tSEGMENT ";
 518                                         $v.=$masm>=$masmref ? "ALIGN(64)" : "PAGE";
 519                                         $v.=" 'CODE'";
 520                                     }
 521                                     $self->{value} = $v;
 522                                     last;
 523                                   };
 524                 /\.data/    && do { my $v=undef;
 525                                     if ($nasm) {
 526                                         $v="section     .data data align=8\n";
 527                                     } else {
 528                                         $v="$current_segment\tENDS\n" if ($current_segment);
 529                                         $current_segment = "_DATA";
 530                                         $v.="$current_segment\tSEGMENT";
 531                                     }
 532                                     $self->{value} = $v;
 533                                     last;
 534                                   };
 535                 /\.section/ && do { my $v=undef;
 536                                     $line =~ s/([^,]*).*/$1/;
 537                                     $line = ".CRT\$XCU" if ($line eq ".init");
 538                                     if ($nasm) {
 539                                         $v="section     $line";
 540                                         if ($line=~/\.([px])data/) {
 541                                             $v.=" rdata align=";
 542                                             $v.=$1 eq "p"? 4 : 8;
 543                                         }
 544                                     } else {
 545                                         $v="$current_segment\tENDS\n" if ($current_segment);
 546                                         $v.="$line\tSEGMENT";
 547                                         if ($line=~/\.([px])data/) {
 548                                             $v.=" READONLY";
 549                                             $v.=" ALIGN(".($1 eq "p" ? 4 : 8).")" if ($masm>=$masmref);
 550                                         }
 551                                     }
 552                                     $current_segment = $line;
 553                                     $self->{value} = $v;
 554                                     last;
 555                                   };
 556                 /\.extern/  && do { $self->{value}  = "EXTERN\t".$line;
 557                                     $self->{value} .= ":NEAR" if ($masm);
 558                                     last;
 559                                   };
 560                 /\.globl|.global/
 561                             && do { $self->{value}  = $masm?"PUBLIC":"global";
 562                                     $self->{value} .= "\t".$line;
 563                                     last;
 564                                   };
 565                 /\.size/    && do { if (defined($current_function)) {
 566                                         undef $self->{value};
 567                                         if ($current_function->{abi} eq "svr4") {
 568                                             $self->{value}="${decor}SEH_end_$current_function->{name}:";
 569                                             $self->{value}.=":\n" if($masm);
 570                                         }
 571                                         $self->{value}.="$current_function->{name}\tENDP" if($masm);
 572                                         undef $current_function;
 573                                     }
 574                                     last;
 575                                   };
 576                 /\.align/   && do { $self->{value} = "ALIGN\t".$line; last; };
 577                 /\.(value|long|rva|quad)/
 578                             && do { my $sz  = substr($1,0,1);
 579                                     my @arr = split(/,\s*/,$line);
 580                                     my $last = pop(@arr);
 581                                     my $conv = sub  {   my $var=shift;
 582                                                         $var=~s/^(0b[0-1]+)/oct($1)/eig;
 583                                                         $var=~s/^0x([0-9a-f]+)/0$1h/ig if ($masm);
 584                                                         if ($sz eq "D" && ($current_segment=~/.[px]data/ || $dir eq ".rva"))
 585                                                         { $var=~s/([_a-z\$\@][_a-z0-9\$\@]*)/$nasm?"$1 wrt ..imagebase":"imagerel $1"/egi; }
 586                                                         $var;
 587                                                     };
 588
 589                                     $sz =~ tr/bvlrq/BWDDQ/;
 590                                     $self->{value} = "\tD$sz\t";
 591                                     for (@arr) { $self->{value} .= &$conv($_).","; }
 592                                     $self->{value} .= &$conv($last);
 593                                     last;
 594                                   };
 595                 /\.byte/    && do { my @str=split(/,\s*/,$line);
 596                                     map(s/(0b[0-1]+)/oct($1)/eig,@str);
 597                                     map(s/0x([0-9a-f]+)/0$1h/ig,@str) if ($masm);
 598                                     while ($#str>15) {
 599                                         $self->{value}.="DB\t"
 600                                                 .join(",",@str[0..15])."\n";
 601                                         foreach (0..15) { shift @str; }
 602                                     }
 603                                     $self->{value}.="DB\t"
 604                                                 .join(",",@str) if (@str);
 605                                     last;
 606                                   };
 607             }
 608             $line = "";
 609         }
 610
 611         $ret;
 612     }
 613     sub out {
 614         my $self = shift;
 615         $self->{value};
 616     }
 617 }
 618
 619 sub rex {
 620  local *opcode=shift;
 621  my ($dst,$src)=@_;
 622
 623    if ($dst>=8 || $src>=8) {
 624         $rex=0x40;
 625         $rex|=0x04 if($dst>=8);
 626         $rex|=0x01 if($src>=8);
 627         push @opcode,$rex;
 628    }
 629 }
 630
 631 # older gas doesn't handle SSE>2 instructions
 632 my %regrm = (   "%eax"=>0, "%ecx"=>1, "%edx"=>2, "%ebx"=>3,
 633                 "%esp"=>4, "%ebp"=>5, "%esi"=>6, "%edi"=>7      );
 634
 635 my $pextrd = sub {
 636   my ($imm,$src,$dst) = @_;
 637     if ("$imm:$src" =~ /\$([0-9]+):%xmm([0-9]+)/) {
 638       my @opcode=(0x66);
 639         $imm=$1;
 640         $src=$2;
 641         if ($dst =~ /%r([0-9]+)d/)      { $dst = $1; }
 642         elsif ($dst =~ /%e/)            { $dst = $regrm{$dst}; }
 643         rex(\@opcode,$src,$dst);
 644         push @opcode,0x0f,0x3a,0x16;
 645         push @opcode,0xc0|(($src&7)<<3)|($dst&7);       # ModR/M
 646         push @opcode,$imm;
 647         printf "\t.byte\t%s\n",join(',',@opcode);
 648     } else {
 649         printf "\tpextrd\t%s\n",join(',',@_);
 650     }
 651 } if ($gas);
 652
 653 my $pinsrd = sub {
 654   my ($imm,$src,$dst) = @_;
 655     if ("$imm:$dst" =~ /\$([0-9]+):%xmm([0-9]+)/) {
 656       my @opcode=(0x66);
 657         $imm=$1;
 658         $dst=$2;
 659         if ($src =~ /%r([0-9]+)d/)      { $src = $1; }
 660         elsif ($src =~ /%e/)            { $src = $regrm{$src}; }
 661         rex(\@opcode,$dst,$src);
 662         push @opcode,0x0f,0x3a,0x22;
 663         push @opcode,0xc0|(($dst&7)<<3)|($src&7);       # ModR/M
 664         push @opcode,$imm;
 665         printf "\t.byte\t%s\n",join(',',@opcode);
 666     } else {
 667         printf "\tpinsrd\t%s\n",join(',',@_);
 668     }
 669 } if ($gas);
 670
 671 my $pshufb = sub {
 672   my ($src,$dst) = @_;
 673     if ("$dst:$src" =~ /%xmm([0-9]+):%xmm([0-9]+)/) {
 674       my @opcode=(0x66);
 675         rex(\@opcode,$1,$2);
 676         push @opcode,0x0f,0x38,0x00;
 677         push @opcode,0xc0|($2&7)|(($1&7)<<3);   # ModR/M
 678         printf "\t.byte\t%s\n",join(',',@opcode);
 679     } else {
 680         printf "\tpshufb\t%s\n",join(',',@_);
 681     }
 682 } if ($gas);
 683
 684 if ($nasm) {
 685     print <<___;
 686 default rel
 687 ___
 688 } elsif ($masm) {
 689     print <<___;
 690 OPTION  DOTNAME
 691 ___
 692 }
 693 while($line=<>) {
 694
 695     chomp($line);
 696
 697     $line =~ s|[#!].*$||;       # get rid of asm-style comments...
 698     $line =~ s|/\*.*\*/||;      # ... and C-style comments...
 699     $line =~ s|^\s+||;          # ... and skip white spaces in beginning
 700
 701     undef $label;
 702     undef $opcode;
 703     undef $sz;
 704     undef @args;
 705
 706     if ($label=label->re(\$line))       { print $label->out(); }
 707
 708     if (directive->re(\$line)) {
 709         printf "%s",directive->out();
 710     } elsif ($opcode=opcode->re(\$line)) { ARGUMENT: while (1) {
 711         my $arg;
 712
 713         if ($arg=register->re(\$line))  { opcode->size($arg->size()); }
 714         elsif ($arg=const->re(\$line))  { }
 715         elsif ($arg=ea->re(\$line))     { }
 716         elsif ($arg=expr->re(\$line))   { }
 717         else                            { last ARGUMENT; }
 718
 719         push @args,$arg;
 720
 721         last ARGUMENT if ($line !~ /^,/);
 722
 723         $line =~ s/^,\s*//;
 724         } # ARGUMENT:
 725
 726         $sz=opcode->size();
 727
 728         if ($#args>=0) {
 729             my $insn;
 730             if ($gas) {
 731                 $insn = $opcode->out($#args>=1?$args[$#args]->size():$sz);
 732                 @args = map($_->out($sz),@args);
 733                 my $asm = eval("\$$insn");
 734                 if (ref($asm) eq 'CODE') { &$asm(@args); }
 735                 else { printf "\t%s\t%s",$insn,join(",",@args); }
 736             } else {
 737                 $insn = $opcode->out();
 738                 $insn .= $sz if (map($_->out() =~ /xmm|mmx/,@args));
 739                 @args = reverse(@args);
 740                 undef $sz if ($nasm && $opcode->mnemonic() eq "lea");
 741                 printf "\t%s\t%s",$insn,join(",",map($_->out($sz),@args));
 742             }
 743         } else {
 744             printf "\t%s",$opcode->out();
 745         }
 746     }
 747
 748     print $line,"\n";
 749 }
 750
 751 print "\n$current_segment\tENDS\n"      if ($current_segment && $masm);
 752 print "END\n"                           if ($masm);
 753
 754 close STDOUT;
 755
 756 \f#################################################
 757 # Cross-reference x86_64 ABI "card"
 758 #
 759 #               Unix            Win64
 760 # %rax          *               *
 761 # %rbx          -               -
 762 # %rcx          #4              #1
 763 # %rdx          #3              #2
 764 # %rsi          #2              -
 765 # %rdi          #1              -
 766 # %rbp          -               -
 767 # %rsp          -               -
 768 # %r8           #5              #3
 769 # %r9           #6              #4
 770 # %r10          *               *
 771 # %r11          *               *
 772 # %r12          -               -
 773 # %r13          -               -
 774 # %r14          -               -
 775 # %r15          -               -
 776 #
 777 # (*)   volatile register
 778 # (-)   preserved by callee
 779 # (#)   Nth argument, volatile
 780 #
 781 # In Unix terms top of stack is argument transfer area for arguments
 782 # which could not be accomodated in registers. Or in other words 7th
 783 # [integer] argument resides at 8(%rsp) upon function entry point.
 784 # 128 bytes above %rsp constitute a "red zone" which is not touched
 785 # by signal handlers and can be used as temporal storage without
 786 # allocating a frame.
 787 #
 788 # In Win64 terms N*8 bytes on top of stack is argument transfer area,
 789 # which belongs to/can be overwritten by callee. N is the number of
 790 # arguments passed to callee, *but* not less than 4! This means that
 791 # upon function entry point 5th argument resides at 40(%rsp), as well
 792 # as that 32 bytes from 8(%rsp) can always be used as temporal
 793 # storage [without allocating a frame]. One can actually argue that
 794 # one can assume a "red zone" above stack pointer under Win64 as well.
 795 # Point is that at apparently no occasion Windows kernel would alter
 796 # the area above user stack pointer in true asynchronous manner...
 797 #
 798 # All the above means that if assembler programmer adheres to Unix
 799 # register and stack layout, but disregards the "red zone" existense,
 800 # it's possible to use following prologue and epilogue to "gear" from
 801 # Unix to Win64 ABI in leaf functions with not more than 6 arguments.
 802 #
 803 # omnipotent_function:
 804 # ifdef WIN64
 805 #       movq    %rdi,8(%rsp)
 806 #       movq    %rsi,16(%rsp)
 807 #       movq    %rcx,%rdi       ; if 1st argument is actually present
 808 #       movq    %rdx,%rsi       ; if 2nd argument is actually ...
 809 #       movq    %r8,%rdx        ; if 3rd argument is ...
 810 #       movq    %r9,%rcx        ; if 4th argument ...
 811 #       movq    40(%rsp),%r8    ; if 5th ...
 812 #       movq    48(%rsp),%r9    ; if 6th ...
 813 # endif
 814 #       ...
 815 # ifdef WIN64
 816 #       movq    8(%rsp),%rdi
 817 #       movq    16(%rsp),%rsi
 818 # endif
 819 #       ret
 820 #
 821 \f#################################################
 822 # Win64 SEH, Structured Exception Handling.
 823 #
 824 # Unlike on Unix systems(*) lack of Win64 stack unwinding information
 825 # has undesired side-effect at run-time: if an exception is raised in
 826 # assembler subroutine such as those in question (basically we're
 827 # referring to segmentation violations caused by malformed input
 828 # parameters), the application is briskly terminated without invoking
 829 # any exception handlers, most notably without generating memory dump
 830 # or any user notification whatsoever. This poses a problem. It's
 831 # possible to address it by registering custom language-specific
 832 # handler that would restore processor context to the state at
 833 # subroutine entry point and return "exception is not handled, keep
 834 # unwinding" code. Writing such handler can be a challenge... But it's
 835 # doable, though requires certain coding convention. Consider following
 836 # snippet:
 837 #
 838 # .type function,@function
 839 # function:
 840 #       movq    %rsp,%rax       # copy rsp to volatile register
 841 #       pushq   %r15            # save non-volatile registers
 842 #       pushq   %rbx
 843 #       pushq   %rbp
 844 #       movq    %rsp,%r11
 845 #       subq    %rdi,%r11       # prepare [variable] stack frame
 846 #       andq    $-64,%r11
 847 #       movq    %rax,0(%r11)    # check for exceptions
 848 #       movq    %r11,%rsp       # allocate [variable] stack frame
 849 #       movq    %rax,0(%rsp)    # save original rsp value
 850 # magic_point:
 851 #       ...
 852 #       movq    0(%rsp),%rcx    # pull original rsp value
 853 #       movq    -24(%rcx),%rbp  # restore non-volatile registers
 854 #       movq    -16(%rcx),%rbx
 855 #       movq    -8(%rcx),%r15
 856 #       movq    %rcx,%rsp       # restore original rsp
 857 #       ret
 858 # .size function,.-function
 859 #
 860 # The key is that up to magic_point copy of original rsp value remains
 861 # in chosen volatile register and no non-volatile register, except for
 862 # rsp, is modified. While past magic_point rsp remains constant till
 863 # the very end of the function. In this case custom language-specific
 864 # exception handler would look like this:
 865 #
 866 # EXCEPTION_DISPOSITION handler (EXCEPTION_RECORD *rec,ULONG64 frame,
 867 #               CONTEXT *context,DISPATCHER_CONTEXT *disp)
 868 # {     ULONG64 *rsp = (ULONG64 *)context->Rax;
 869 #       if (context->Rip >= magic_point)
 870 #       {   rsp = ((ULONG64 **)context->Rsp)[0];
 871 #           context->Rbp = rsp[-3];
 872 #           context->Rbx = rsp[-2];
 873 #           context->R15 = rsp[-1];
 874 #       }
 875 #       context->Rsp = (ULONG64)rsp;
 876 #       context->Rdi = rsp[1];
 877 #       context->Rsi = rsp[2];
 878 #
 879 #       memcpy (disp->ContextRecord,context,sizeof(CONTEXT));
 880 #       RtlVirtualUnwind(UNW_FLAG_NHANDLER,disp->ImageBase,
 881 #               dips->ControlPc,disp->FunctionEntry,disp->ContextRecord,
 882 #               &disp->HandlerData,&disp->EstablisherFrame,NULL);
 883 #       return ExceptionContinueSearch;
 884 # }
 885 #
 886 # It's appropriate to implement this handler in assembler, directly in
 887 # function's module. In order to do that one has to know members'
 888 # offsets in CONTEXT and DISPATCHER_CONTEXT structures and some constant
 889 # values. Here they are:
 890 #
 891 #       CONTEXT.Rax                             120
 892 #       CONTEXT.Rcx                             128
 893 #       CONTEXT.Rdx                             136
 894 #       CONTEXT.Rbx                             144
 895 #       CONTEXT.Rsp                             152
 896 #       CONTEXT.Rbp                             160
 897 #       CONTEXT.Rsi                             168
 898 #       CONTEXT.Rdi                             176
 899 #       CONTEXT.R8                              184
 900 #       CONTEXT.R9                              192
 901 #       CONTEXT.R10                             200
 902 #       CONTEXT.R11                             208
 903 #       CONTEXT.R12                             216
 904 #       CONTEXT.R13                             224
 905 #       CONTEXT.R14                             232
 906 #       CONTEXT.R15                             240
 907 #       CONTEXT.Rip                             248
 908 #       CONTEXT.Xmm6                            512
 909 #       sizeof(CONTEXT)                         1232
 910 #       DISPATCHER_CONTEXT.ControlPc            0
 911 #       DISPATCHER_CONTEXT.ImageBase            8
 912 #       DISPATCHER_CONTEXT.FunctionEntry        16
 913 #       DISPATCHER_CONTEXT.EstablisherFrame     24
 914 #       DISPATCHER_CONTEXT.TargetIp             32
 915 #       DISPATCHER_CONTEXT.ContextRecord        40
 916 #       DISPATCHER_CONTEXT.LanguageHandler      48
 917 #       DISPATCHER_CONTEXT.HandlerData          56
 918 #       UNW_FLAG_NHANDLER                       0
 919 #       ExceptionContinueSearch                 1
 920 #
 921 # In order to tie the handler to the function one has to compose
 922 # couple of structures: one for .xdata segment and one for .pdata.
 923 #
 924 # UNWIND_INFO structure for .xdata segment would be
 925 #
 926 # function_unwind_info:
 927 #       .byte   9,0,0,0
 928 #       .rva    handler
 929 #
 930 # This structure designates exception handler for a function with
 931 # zero-length prologue, no stack frame or frame register.
 932 #
 933 # To facilitate composing of .pdata structures, auto-generated "gear"
 934 # prologue copies rsp value to rax and denotes next instruction with
 935 # .LSEH_begin_{function_name} label. This essentially defines the SEH
 936 # styling rule mentioned in the beginning. Position of this label is
 937 # chosen in such manner that possible exceptions raised in the "gear"
 938 # prologue would be accounted to caller and unwound from latter's frame.
 939 # End of function is marked with respective .LSEH_end_{function_name}
 940 # label. To summarize, .pdata segment would contain
 941 #
 942 #       .rva    .LSEH_begin_function
 943 #       .rva    .LSEH_end_function
 944 #       .rva    function_unwind_info
 945 #
 946 # Reference to functon_unwind_info from .xdata segment is the anchor.
 947 # In case you wonder why references are 32-bit .rvas and not 64-bit
 948 # .quads. References put into these two segments are required to be
 949 # *relative* to the base address of the current binary module, a.k.a.
 950 # image base. No Win64 module, be it .exe or .dll, can be larger than
 951 # 2GB and thus such relative references can be and are accommodated in
 952 # 32 bits.
 953 #
 954 # Having reviewed the example function code, one can argue that "movq
 955 # %rsp,%rax" above is redundant. It is not! Keep in mind that on Unix
 956 # rax would contain an undefined value. If this "offends" you, use
 957 # another register and refrain from modifying rax till magic_point is
 958 # reached, i.e. as if it was a non-volatile register. If more registers
 959 # are required prior [variable] frame setup is completed, note that
 960 # nobody says that you can have only one "magic point." You can
 961 # "liberate" non-volatile registers by denoting last stack off-load
 962 # instruction and reflecting it in finer grade unwind logic in handler.
 963 # After all, isn't it why it's called *language-specific* handler...
 964 #
 965 # Attentive reader can notice that exceptions would be mishandled in
 966 # auto-generated "gear" epilogue. Well, exception effectively can't
 967 # occur there, because if memory area used by it was subject to
 968 # segmentation violation, then it would be raised upon call to the
 969 # function (and as already mentioned be accounted to caller, which is
 970 # not a problem). If you're still not comfortable, then define tail
 971 # "magic point" just prior ret instruction and have handler treat it...
 972 #
 973 # (*)   Note that we're talking about run-time, not debug-time. Lack of
 974 #       unwind information makes debugging hard on both Windows and
 975 #       Unix. "Unlike" referes to the fact that on Unix signal handler
 976 #       will always be invoked, core dumped and appropriate exit code
 977 #       returned to parent (for user notification).