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 # a. If function accepts more than 4 arguments *and* >4th argument
  59 #    is declared as non 64-bit value, do clear its upper part.
  60 \f
  61 my $flavour = shift;
  62 my $output  = shift;
  63 if ($flavour =~ /\./) { $output = $flavour; undef $flavour; }
  64
  65 { my ($stddev,$stdino,@junk)=stat(STDOUT);
  66   my ($outdev,$outino,@junk)=stat($output);
  67
  68     open STDOUT,">$output" || die "can't open $output: $!"
  69         if ($stddev!=$outdev || $stdino!=$outino);
  70 }
  71
  72 my $gas=1;      $gas=0 if ($output =~ /\.asm$/);
  73 my $elf=1;      $elf=0 if (!$gas);
  74 my $win64=0;
  75 my $prefix="";
  76 my $decor=".L";
  77
  78 my $masmref=8 + 50727*2**-32;   # 8.00.50727 shipped with VS2005
  79 my $masm=0;
  80 my $PTR=" PTR";
  81
  82 my $nasmref=2.03;
  83 my $nasm=0;
  84
  85 if    ($flavour eq "mingw64")   { $gas=1; $elf=0; $win64=1;
  86                                   $prefix=`echo __USER_LABEL_PREFIX__ | $ENV{CC} -E -P -`;
  87                                   chomp($prefix);
  88                                 }
  89 elsif ($flavour eq "macosx")    { $gas=1; $elf=0; $prefix="_"; $decor="L\$"; }
  90 elsif ($flavour eq "masm")      { $gas=0; $elf=0; $masm=$masmref; $win64=1; $decor="\$L\$"; }
  91 elsif ($flavour eq "nasm")      { $gas=0; $elf=0; $nasm=$nasmref; $win64=1; $decor="\$L\$"; $PTR=""; }
  92 elsif (!$gas)
  93 {   if ($ENV{ASM} =~ m/nasm/ && `nasm -v` =~ m/version ([0-9]+)\.([0-9]+)/i)
  94     {   $nasm = $1 + $2*0.01; $PTR="";  }
  95     elsif (`ml64 2>&1` =~ m/Version ([0-9]+)\.([0-9]+)(\.([0-9]+))?/)
  96     {   $masm = $1 + $2*2**-16 + $4*2**-32;   }
  97     die "no assembler found on %PATH" if (!($nasm || $masm));
  98     $win64=1;
  99     $elf=0;
 100     $decor="\$L\$";
 101 }
 102
 103 my $current_segment;
 104 my $current_function;
 105 my %globals;
 106
 107 { package opcode;       # pick up opcodes
 108     sub re {
 109         my      $self = shift;  # single instance in enough...
 110         local   *line = shift;
 111         undef   $ret;
 112
 113         if ($line =~ /^([a-z][a-z0-9]*)/i) {
 114             $self->{op} = $1;
 115             $ret = $self;
 116             $line = substr($line,@+[0]); $line =~ s/^\s+//;
 117
 118             undef $self->{sz};
 119             if ($self->{op} =~ /^(movz)b.*/) {  # movz is pain...
 120                 $self->{op} = $1;
 121                 $self->{sz} = "b";
 122             } elsif ($self->{op} =~ /call|jmp/) {
 123                 $self->{sz} = "";
 124             } elsif ($self->{op} =~ /^p/ && $' !~ /^(ush|op|insrw)/) { # SSEn
 125                 $self->{sz} = "";
 126             } elsif ($self->{op} =~ /([a-z]{3,})([qlwb])$/) {
 127                 $self->{op} = $1;
 128                 $self->{sz} = $2;
 129             }
 130         }
 131         $ret;
 132     }
 133     sub size {
 134         my $self = shift;
 135         my $sz   = shift;
 136         $self->{sz} = $sz if (defined($sz) && !defined($self->{sz}));
 137         $self->{sz};
 138     }
 139     sub out {
 140         my $self = shift;
 141         if ($gas) {
 142             if ($self->{op} eq "movz") {        # movz is pain...
 143                 sprintf "%s%s%s",$self->{op},$self->{sz},shift;
 144             } elsif ($self->{op} =~ /^set/) {
 145                 "$self->{op}";
 146             } elsif ($self->{op} eq "ret") {
 147                 my $epilogue = "";
 148                 if ($win64 && $current_function->{abi} eq "svr4") {
 149                     $epilogue = "movq   8(%rsp),%rdi\n\t" .
 150                                 "movq   16(%rsp),%rsi\n\t";
 151                 }
 152                 $epilogue . ".byte      0xf3,0xc3";
 153             } elsif ($self->{op} eq "call" && !$elf && $current_segment eq ".init") {
 154                 ".p2align\t3\n\t.quad";
 155             } else {
 156                 "$self->{op}$self->{sz}";
 157             }
 158         } else {
 159             $self->{op} =~ s/^movz/movzx/;
 160             if ($self->{op} eq "ret") {
 161                 $self->{op} = "";
 162                 if ($win64 && $current_function->{abi} eq "svr4") {
 163                     $self->{op} = "mov  rdi,QWORD${PTR}[8+rsp]\t;WIN64 epilogue\n\t".
 164                                   "mov  rsi,QWORD${PTR}[16+rsp]\n\t";
 165                 }
 166                 $self->{op} .= "DB\t0F3h,0C3h\t\t;repret";
 167             } elsif ($self->{op} =~ /^(pop|push)f/) {
 168                 $self->{op} .= $self->{sz};
 169             } elsif ($self->{op} eq "call" && $current_segment eq ".CRT\$XCU") {
 170                 $self->{op} = "\tDQ";
 171             }
 172             $self->{op};
 173         }
 174     }
 175     sub mnemonic {
 176         my $self=shift;
 177         my $op=shift;
 178         $self->{op}=$op if (defined($op));
 179         $self->{op};
 180     }
 181 }
 182 { package const;        # pick up constants, which start with $
 183     sub re {
 184         my      $self = shift;  # single instance in enough...
 185         local   *line = shift;
 186         undef   $ret;
 187
 188         if ($line =~ /^\$([^,]+)/) {
 189             $self->{value} = $1;
 190             $ret = $self;
 191             $line = substr($line,@+[0]); $line =~ s/^\s+//;
 192         }
 193         $ret;
 194     }
 195     sub out {
 196         my $self = shift;
 197
 198         if ($gas) {
 199             # Solaris /usr/ccs/bin/as can't handle multiplications
 200             # in $self->{value}
 201             $self->{value} =~ s/(?<![\w\$\.])(0x?[0-9a-f]+)/oct($1)/egi;
 202             $self->{value} =~ s/([0-9]+\s*[\*\/\%]\s*[0-9]+)/eval($1)/eg;
 203             sprintf "\$%s",$self->{value};
 204         } else {
 205             $self->{value} =~ s/(0b[0-1]+)/oct($1)/eig;
 206             $self->{value} =~ s/0x([0-9a-f]+)/0$1h/ig if ($masm);
 207             sprintf "%s",$self->{value};
 208         }
 209     }
 210 }
 211 { package ea;           # pick up effective addresses: expr(%reg,%reg,scale)
 212     sub re {
 213         my      $self = shift;  # single instance in enough...
 214         local   *line = shift;
 215         undef   $ret;
 216
 217         # optional * ---vvv--- appears in indirect jmp/call
 218         if ($line =~ /^(\*?)([^\(,]*)\(([%\w,]+)\)/) {
 219             $self->{asterisk} = $1;
 220             $self->{label} = $2;
 221             ($self->{base},$self->{index},$self->{scale})=split(/,/,$3);
 222             $self->{scale} = 1 if (!defined($self->{scale}));
 223             $ret = $self;
 224             $line = substr($line,@+[0]); $line =~ s/^\s+//;
 225
 226             if ($win64 && $self->{label} =~ s/\@GOTPCREL//) {
 227                 die if (opcode->mnemonic() ne "mov");
 228                 opcode->mnemonic("lea");
 229             }
 230             $self->{base}  =~ s/^%//;
 231             $self->{index} =~ s/^%// if (defined($self->{index}));
 232         }
 233         $ret;
 234     }
 235     sub size {}
 236     sub out {
 237         my $self = shift;
 238         my $sz = shift;
 239
 240         $self->{label} =~ s/([_a-z][_a-z0-9]*)/$globals{$1} or $1/gei;
 241         $self->{label} =~ s/\.L/$decor/g;
 242
 243         # Silently convert all EAs to 64-bit. This is required for
 244         # elder GNU assembler and results in more compact code,
 245         # *but* most importantly AES module depends on this feature!
 246         $self->{index} =~ s/^[er](.?[0-9xpi])[d]?$/r\1/;
 247         $self->{base}  =~ s/^[er](.?[0-9xpi])[d]?$/r\1/;
 248
 249         if ($gas) {
 250             # Solaris /usr/ccs/bin/as can't handle multiplications
 251             # in $self->{label}, new gas requires sign extension...
 252             use integer;
 253             $self->{label} =~ s/(?<![\w\$\.])(0x?[0-9a-f]+)/oct($1)/egi;
 254             $self->{label} =~ s/([0-9]+\s*[\*\/\%]\s*[0-9]+)/eval($1)/eg;
 255             $self->{label} =~ s/([0-9]+)/$1<<32>>32/eg;
 256             $self->{label} =~ s/^___imp_/__imp__/   if ($flavour eq "mingw64");
 257
 258             if (defined($self->{index})) {
 259                 sprintf "%s%s(%%%s,%%%s,%d)",$self->{asterisk},
 260                                         $self->{label},$self->{base},
 261                                         $self->{index},$self->{scale};
 262             } else {
 263                 sprintf "%s%s(%%%s)",   $self->{asterisk},$self->{label},$self->{base};
 264             }
 265         } else {
 266             %szmap = (  b=>"BYTE$PTR", w=>"WORD$PTR", l=>"DWORD$PTR",
 267                         q=>"QWORD$PTR",o=>"OWORD$PTR",x=>"XMMWORD$PTR" );
 268
 269             $self->{label} =~ s/\./\$/g;
 270             $self->{label} =~ s/(?<![\w\$\.])0x([0-9a-f]+)/0$1h/ig;
 271             $self->{label} = "($self->{label})" if ($self->{label} =~ /[\*\+\-\/]/);
 272             $sz="q" if ($self->{asterisk});
 273
 274             if (defined($self->{index})) {
 275                 sprintf "%s[%s%s*%d+%s]",$szmap{$sz},
 276                                         $self->{label}?"$self->{label}+":"",
 277                                         $self->{index},$self->{scale},
 278                                         $self->{base};
 279             } elsif ($self->{base} eq "rip") {
 280                 sprintf "%s[%s]",$szmap{$sz},$self->{label};
 281             } else {
 282                 sprintf "%s[%s%s]",$szmap{$sz},
 283                                         $self->{label}?"$self->{label}+":"",
 284                                         $self->{base};
 285             }
 286         }
 287     }
 288 }
 289 { package register;     # pick up registers, which start with %.
 290     sub re {
 291         my      $class = shift; # muliple instances...
 292         my      $self = {};
 293         local   *line = shift;
 294         undef   $ret;
 295
 296         # optional * ---vvv--- appears in indirect jmp/call
 297         if ($line =~ /^(\*?)%(\w+)/) {
 298             bless $self,$class;
 299             $self->{asterisk} = $1;
 300             $self->{value} = $2;
 301             $ret = $self;
 302             $line = substr($line,@+[0]); $line =~ s/^\s+//;
 303         }
 304         $ret;
 305     }
 306     sub size {
 307         my      $self = shift;
 308         undef   $ret;
 309
 310         if    ($self->{value} =~ /^r[\d]+b$/i)  { $ret="b"; }
 311         elsif ($self->{value} =~ /^r[\d]+w$/i)  { $ret="w"; }
 312         elsif ($self->{value} =~ /^r[\d]+d$/i)  { $ret="l"; }
 313         elsif ($self->{value} =~ /^r[\w]+$/i)   { $ret="q"; }
 314         elsif ($self->{value} =~ /^[a-d][hl]$/i){ $ret="b"; }
 315         elsif ($self->{value} =~ /^[\w]{2}l$/i) { $ret="b"; }
 316         elsif ($self->{value} =~ /^[\w]{2}$/i)  { $ret="w"; }
 317         elsif ($self->{value} =~ /^e[a-z]{2}$/i){ $ret="l"; }
 318
 319         $ret;
 320     }
 321     sub out {
 322         my $self = shift;
 323         if ($gas)       { sprintf "%s%%%s",$self->{asterisk},$self->{value}; }
 324         else            { $self->{value}; }
 325     }
 326 }
 327 { package label;        # pick up labels, which end with :
 328     sub re {
 329         my      $self = shift;  # single instance is enough...
 330         local   *line = shift;
 331         undef   $ret;
 332
 333         if ($line =~ /(^[\.\w]+)\:/) {
 334             $self->{value} = $1;
 335             $ret = $self;
 336             $line = substr($line,@+[0]); $line =~ s/^\s+//;
 337
 338             $self->{value} =~ s/^\.L/$decor/;
 339         }
 340         $ret;
 341     }
 342     sub out {
 343         my $self = shift;
 344
 345         if ($gas) {
 346             my $func = ($globals{$self->{value}} or $self->{value}) . ":";
 347             if ($win64  &&
 348                         $current_function->{name} eq $self->{value} &&
 349                         $current_function->{abi} eq "svr4") {
 350                 $func .= "\n";
 351                 $func .= "      movq    %rdi,8(%rsp)\n";
 352                 $func .= "      movq    %rsi,16(%rsp)\n";
 353                 $func .= "      movq    %rsp,%rax\n";
 354                 $func .= "${decor}SEH_begin_$current_function->{name}:\n";
 355                 my $narg = $current_function->{narg};
 356                 $narg=6 if (!defined($narg));
 357                 $func .= "      movq    %rcx,%rdi\n" if ($narg>0);
 358                 $func .= "      movq    %rdx,%rsi\n" if ($narg>1);
 359                 $func .= "      movq    %r8,%rdx\n"  if ($narg>2);
 360                 $func .= "      movq    %r9,%rcx\n"  if ($narg>3);
 361                 $func .= "      movq    40(%rsp),%r8\n" if ($narg>4);
 362                 $func .= "      movq    48(%rsp),%r9\n" if ($narg>5);
 363             }
 364             $func;
 365         } elsif ($self->{value} ne "$current_function->{name}") {
 366             $self->{value} .= ":" if ($masm && $ret!~m/^\$/);
 367             $self->{value} . ":";
 368         } elsif ($win64 && $current_function->{abi} eq "svr4") {
 369             my $func =  "$current_function->{name}" .
 370                         ($nasm ? ":" : "\tPROC $current_function->{scope}") .
 371                         "\n";
 372             $func .= "  mov     QWORD${PTR}[8+rsp],rdi\t;WIN64 prologue\n";
 373             $func .= "  mov     QWORD${PTR}[16+rsp],rsi\n";
 374             $func .= "  mov     rax,rsp\n";
 375             $func .= "${decor}SEH_begin_$current_function->{name}:";
 376             $func .= ":" if ($masm);
 377             $func .= "\n";
 378             my $narg = $current_function->{narg};
 379             $narg=6 if (!defined($narg));
 380             $func .= "  mov     rdi,rcx\n" if ($narg>0);
 381             $func .= "  mov     rsi,rdx\n" if ($narg>1);
 382             $func .= "  mov     rdx,r8\n"  if ($narg>2);
 383             $func .= "  mov     rcx,r9\n"  if ($narg>3);
 384             $func .= "  mov     r8,QWORD${PTR}[40+rsp]\n" if ($narg>4);
 385             $func .= "  mov     r9,QWORD${PTR}[48+rsp]\n" if ($narg>5);
 386             $func .= "\n";
 387         } else {
 388            "$current_function->{name}".
 389                         ($nasm ? ":" : "\tPROC $current_function->{scope}");
 390         }
 391     }
 392 }
 393 { package expr;         # pick up expressioins
 394     sub re {
 395         my      $self = shift;  # single instance is enough...
 396         local   *line = shift;
 397         undef   $ret;
 398
 399         if ($line =~ /(^[^,]+)/) {
 400             $self->{value} = $1;
 401             $ret = $self;
 402             $line = substr($line,@+[0]); $line =~ s/^\s+//;
 403
 404             $self->{value} =~ s/\@PLT// if (!$elf);
 405             $self->{value} =~ s/([_a-z][_a-z0-9]*)/$globals{$1} or $1/gei;
 406             $self->{value} =~ s/\.L/$decor/g;
 407         }
 408         $ret;
 409     }
 410     sub out {
 411         my $self = shift;
 412         if ($nasm && opcode->mnemonic()=~m/^j/) {
 413             "NEAR ".$self->{value};
 414         } else {
 415             $self->{value};
 416         }
 417     }
 418 }
 419 { package directive;    # pick up directives, which start with .
 420     sub re {
 421         my      $self = shift;  # single instance is enough...
 422         local   *line = shift;
 423         undef   $ret;
 424         my      $dir;
 425         my      %opcode =       # lea 2f-1f(%rip),%dst; 1: nop; 2:
 426                 (       "%rax"=>0x01058d48,     "%rcx"=>0x010d8d48,
 427                         "%rdx"=>0x01158d48,     "%rbx"=>0x011d8d48,
 428                         "%rsp"=>0x01258d48,     "%rbp"=>0x012d8d48,
 429                         "%rsi"=>0x01358d48,     "%rdi"=>0x013d8d48,
 430                         "%r8" =>0x01058d4c,     "%r9" =>0x010d8d4c,
 431                         "%r10"=>0x01158d4c,     "%r11"=>0x011d8d4c,
 432                         "%r12"=>0x01258d4c,     "%r13"=>0x012d8d4c,
 433                         "%r14"=>0x01358d4c,     "%r15"=>0x013d8d4c      );
 434
 435         if ($line =~ /^\s*(\.\w+)/) {
 436             $dir = $1;
 437             $ret = $self;
 438             undef $self->{value};
 439             $line = substr($line,@+[0]); $line =~ s/^\s+//;
 440
 441             SWITCH: for ($dir) {
 442                 /\.picmeup/ && do { if ($line =~ /(%r[\w]+)/i) {
 443                                         $dir="\t.long";
 444                                         $line=sprintf "0x%x,0x90000000",$opcode{$1};
 445                                     }
 446                                     last;
 447                                   };
 448                 /\.global|\.globl|\.extern/
 449                             && do { $globals{$line} = $prefix . $line;
 450                                     $line = $globals{$line} if ($prefix);
 451                                     last;
 452                                   };
 453                 /\.type/    && do { ($sym,$type,$narg) = split(',',$line);
 454                                     if ($type eq "\@function") {
 455                                         undef $current_function;
 456                                         $current_function->{name} = $sym;
 457                                         $current_function->{abi}  = "svr4";
 458                                         $current_function->{narg} = $narg;
 459                                         $current_function->{scope} = defined($globals{$sym})?"PUBLIC":"PRIVATE";
 460                                     } elsif ($type eq "\@abi-omnipotent") {
 461                                         undef $current_function;
 462                                         $current_function->{name} = $sym;
 463                                         $current_function->{scope} = defined($globals{$sym})?"PUBLIC":"PRIVATE";
 464                                     }
 465                                     $line =~ s/\@abi\-omnipotent/\@function/;
 466                                     $line =~ s/\@function.*/\@function/;
 467                                     last;
 468                                   };
 469                 /\.asciz/   && do { if ($line =~ /^"(.*)"$/) {
 470                                         $dir  = ".byte";
 471                                         $line = join(",",unpack("C*",$1),0);
 472                                     }
 473                                     last;
 474                                   };
 475                 /\.rva|\.long|\.quad/
 476                             && do { $line =~ s/([_a-z][_a-z0-9]*)/$globals{$1} or $1/gei;
 477                                     $line =~ s/\.L/$decor/g;
 478                                     last;
 479                                   };
 480             }
 481
 482             if ($gas) {
 483                 $self->{value} = $dir . "\t" . $line;
 484
 485                 if ($dir =~ /\.extern/) {
 486                     $self->{value} = ""; # swallow extern
 487                 } elsif (!$elf && $dir =~ /\.type/) {
 488                     $self->{value} = "";
 489                     $self->{value} = ".def\t" . ($globals{$1} or $1) . ";\t" .
 490                                 (defined($globals{$1})?".scl 2;":".scl 3;") .
 491                                 "\t.type 32;\t.endef"
 492                                 if ($win64 && $line =~ /([^,]+),\@function/);
 493                 } elsif (!$elf && $dir =~ /\.size/) {
 494                     $self->{value} = "";
 495                     if (defined($current_function)) {
 496                         $self->{value} .= "${decor}SEH_end_$current_function->{name}:"
 497                                 if ($win64 && $current_function->{abi} eq "svr4");
 498                         undef $current_function;
 499                     }
 500                 } elsif (!$elf && $dir =~ /\.align/) {
 501                     $self->{value} = ".p2align\t" . (log($line)/log(2));
 502                 } elsif ($dir eq ".section") {
 503                     $current_segment=$line;
 504                     if (!$elf && $current_segment eq ".init") {
 505                         if      ($flavour eq "macosx")  { $self->{value} = ".mod_init_func"; }
 506                         elsif   ($flavour eq "mingw64") { $self->{value} = ".section\t.ctors"; }
 507                     }
 508                 } elsif ($dir =~ /\.(text|data)/) {
 509                     $current_segment=".$1";
 510                 }
 511                 $line = "";
 512                 return $self;
 513             }
 514
 515             # non-gas case or nasm/masm
 516             SWITCH: for ($dir) {
 517                 /\.text/    && do { my $v=undef;
 518                                     if ($nasm) {
 519                                         $v="section     .text code align=64\n";
 520                                     } else {
 521                                         $v="$current_segment\tENDS\n" if ($current_segment);
 522                                         $current_segment = ".text\$";
 523                                         $v.="$current_segment\tSEGMENT ";
 524                                         $v.=$masm>=$masmref ? "ALIGN(64)" : "PAGE";
 525                                         $v.=" 'CODE'";
 526                                     }
 527                                     $self->{value} = $v;
 528                                     last;
 529                                   };
 530                 /\.data/    && do { my $v=undef;
 531                                     if ($nasm) {
 532                                         $v="section     .data data align=8\n";
 533                                     } else {
 534                                         $v="$current_segment\tENDS\n" if ($current_segment);
 535                                         $current_segment = "_DATA";
 536                                         $v.="$current_segment\tSEGMENT";
 537                                     }
 538                                     $self->{value} = $v;
 539                                     last;
 540                                   };
 541                 /\.section/ && do { my $v=undef;
 542                                     $line =~ s/([^,]*).*/$1/;
 543                                     $line = ".CRT\$XCU" if ($line eq ".init");
 544                                     if ($nasm) {
 545                                         $v="section     $line";
 546                                         if ($line=~/\.([px])data/) {
 547                                             $v.=" rdata align=";
 548                                             $v.=$1 eq "p"? 4 : 8;
 549                                         } elsif ($line=~/\.CRT\$/i) {
 550                                             $v.=" rdata align=8";
 551                                         }
 552                                     } else {
 553                                         $v="$current_segment\tENDS\n" if ($current_segment);
 554                                         $v.="$line\tSEGMENT";
 555                                         if ($line=~/\.([px])data/) {
 556                                             $v.=" READONLY";
 557                                             $v.=" ALIGN(".($1 eq "p" ? 4 : 8).")" if ($masm>=$masmref);
 558                                         } elsif ($line=~/\.CRT\$/i) {
 559                                             $v.=" READONLY DWORD";
 560                                         }
 561                                     }
 562                                     $current_segment = $line;
 563                                     $self->{value} = $v;
 564                                     last;
 565                                   };
 566                 /\.extern/  && do { $self->{value}  = "EXTERN\t".$line;
 567                                     $self->{value} .= ":NEAR" if ($masm);
 568                                     last;
 569                                   };
 570                 /\.globl|.global/
 571                             && do { $self->{value}  = $masm?"PUBLIC":"global";
 572                                     $self->{value} .= "\t".$line;
 573                                     last;
 574                                   };
 575                 /\.size/    && do { if (defined($current_function)) {
 576                                         undef $self->{value};
 577                                         if ($current_function->{abi} eq "svr4") {
 578                                             $self->{value}="${decor}SEH_end_$current_function->{name}:";
 579                                             $self->{value}.=":\n" if($masm);
 580                                         }
 581                                         $self->{value}.="$current_function->{name}\tENDP" if($masm);
 582                                         undef $current_function;
 583                                     }
 584                                     last;
 585                                   };
 586                 /\.align/   && do { $self->{value} = "ALIGN\t".$line; last; };
 587                 /\.(value|long|rva|quad)/
 588                             && do { my $sz  = substr($1,0,1);
 589                                     my @arr = split(/,\s*/,$line);
 590                                     my $last = pop(@arr);
 591                                     my $conv = sub  {   my $var=shift;
 592                                                         $var=~s/^(0b[0-1]+)/oct($1)/eig;
 593                                                         $var=~s/^0x([0-9a-f]+)/0$1h/ig if ($masm);
 594                                                         if ($sz eq "D" && ($current_segment=~/.[px]data/ || $dir eq ".rva"))
 595                                                         { $var=~s/([_a-z\$\@][_a-z0-9\$\@]*)/$nasm?"$1 wrt ..imagebase":"imagerel $1"/egi; }
 596                                                         $var;
 597                                                     };
 598
 599                                     $sz =~ tr/bvlrq/BWDDQ/;
 600                                     $self->{value} = "\tD$sz\t";
 601                                     for (@arr) { $self->{value} .= &$conv($_).","; }
 602                                     $self->{value} .= &$conv($last);
 603                                     last;
 604                                   };
 605                 /\.byte/    && do { my @str=split(/,\s*/,$line);
 606                                     map(s/(0b[0-1]+)/oct($1)/eig,@str);
 607                                     map(s/0x([0-9a-f]+)/0$1h/ig,@str) if ($masm);
 608                                     while ($#str>15) {
 609                                         $self->{value}.="DB\t"
 610                                                 .join(",",@str[0..15])."\n";
 611                                         foreach (0..15) { shift @str; }
 612                                     }
 613                                     $self->{value}.="DB\t"
 614                                                 .join(",",@str) if (@str);
 615                                     last;
 616                                   };
 617             }
 618             $line = "";
 619         }
 620
 621         $ret;
 622     }
 623     sub out {
 624         my $self = shift;
 625         $self->{value};
 626     }
 627 }
 628
 629 sub rex {
 630  local *opcode=shift;
 631  my ($dst,$src)=@_;
 632
 633    if ($dst>=8 || $src>=8) {
 634         $rex=0x40;
 635         $rex|=0x04 if($dst>=8);
 636         $rex|=0x01 if($src>=8);
 637         push @opcode,$rex;
 638    }
 639 }
 640
 641 # older gas and ml64 don't handle SSE>2 instructions
 642 my %regrm = (   "%eax"=>0, "%ecx"=>1, "%edx"=>2, "%ebx"=>3,
 643                 "%esp"=>4, "%ebp"=>5, "%esi"=>6, "%edi"=>7      );
 644
 645 my $pextrd = sub {
 646     if (shift =~ /\$([0-9]+),%xmm([0-9]+),(%\w+)/) {
 647       my @opcode=(0x66);
 648         $imm=$1;
 649         $src=$2;
 650         $dst=$3;
 651         if ($dst =~ /%r([0-9]+)d/)      { $dst = $1; }
 652         elsif ($dst =~ /%e/)            { $dst = $regrm{$dst}; }
 653         rex(\@opcode,$src,$dst);
 654         push @opcode,0x0f,0x3a,0x16;
 655         push @opcode,0xc0|(($src&7)<<3)|($dst&7);       # ModR/M
 656         push @opcode,$imm;
 657         @opcode;
 658     } else {
 659         ();
 660     }
 661 };
 662
 663 my $pinsrd = sub {
 664     if (shift =~ /\$([0-9]+),(%\w+),%xmm([0-9]+)/) {
 665       my @opcode=(0x66);
 666         $imm=$1;
 667         $src=$2;
 668         $dst=$3;
 669         if ($src =~ /%r([0-9]+)d/)      { $src = $1; }
 670         elsif ($src =~ /%e/)            { $src = $regrm{$src}; }
 671         rex(\@opcode,$dst,$src);
 672         push @opcode,0x0f,0x3a,0x22;
 673         push @opcode,0xc0|(($dst&7)<<3)|($src&7);       # ModR/M
 674         push @opcode,$imm;
 675         @opcode;
 676     } else {
 677         ();
 678     }
 679 };
 680
 681 my $pshufb = sub {
 682     if (shift =~ /%xmm([0-9]+),%xmm([0-9]+)/) {
 683       my @opcode=(0x66);
 684         rex(\@opcode,$2,$1);
 685         push @opcode,0x0f,0x38,0x00;
 686         push @opcode,0xc0|($1&7)|(($2&7)<<3);           # ModR/M
 687         @opcode;
 688     } else {
 689         ();
 690     }
 691 };
 692
 693 if ($nasm) {
 694     print <<___;
 695 default rel
 696 %define XMMWORD
 697 ___
 698 } elsif ($masm) {
 699     print <<___;
 700 OPTION  DOTNAME
 701 ___
 702 }
 703 while($line=<>) {
 704
 705     chomp($line);
 706
 707     $line =~ s|[#!].*$||;       # get rid of asm-style comments...
 708     $line =~ s|/\*.*\*/||;      # ... and C-style comments...
 709     $line =~ s|^\s+||;          # ... and skip white spaces in beginning
 710
 711     undef $label;
 712     undef $opcode;
 713     undef @args;
 714
 715     if ($label=label->re(\$line))       { print $label->out(); }
 716
 717     if (directive->re(\$line)) {
 718         printf "%s",directive->out();
 719     } elsif ($opcode=opcode->re(\$line)) {
 720         my $asm = eval("\$".$opcode->mnemonic());
 721         undef @bytes;
 722
 723         if ((ref($asm) eq 'CODE') && scalar(@bytes=&$asm($line))) {
 724             print $gas?".byte\t":"DB\t",join(',',@bytes),"\n";
 725             next;
 726         }
 727
 728         ARGUMENT: while (1) {
 729         my $arg;
 730
 731         if ($arg=register->re(\$line))  { opcode->size($arg->size()); }
 732         elsif ($arg=const->re(\$line))  { }
 733         elsif ($arg=ea->re(\$line))     { }
 734         elsif ($arg=expr->re(\$line))   { }
 735         else                            { last ARGUMENT; }
 736
 737         push @args,$arg;
 738
 739         last ARGUMENT if ($line !~ /^,/);
 740
 741         $line =~ s/^,\s*//;
 742         } # ARGUMENT:
 743
 744         if ($#args>=0) {
 745             my $insn;
 746             my $sz=opcode->size();
 747
 748             if ($gas) {
 749                 $insn = $opcode->out($#args>=1?$args[$#args]->size():$sz);
 750                 @args = map($_->out($sz),@args);
 751                 printf "\t%s\t%s",$insn,join(",",@args);
 752             } else {
 753                 $insn = $opcode->out();
 754                 foreach (@args) {
 755                     my $arg = $_->out();
 756                     # $insn.=$sz compensates for movq, pinsrw, ...
 757                     if ($arg =~ /^xmm[0-9]+$/) { $insn.=$sz; $sz="x" if(!$sz); last; }
 758                     if ($arg =~ /^mm[0-9]+$/)  { $insn.=$sz; $sz="q" if(!$sz); last; }
 759                 }
 760                 @args = reverse(@args);
 761                 undef $sz if ($nasm && $opcode->mnemonic() eq "lea");
 762                 printf "\t%s\t%s",$insn,join(",",map($_->out($sz),@args));
 763             }
 764         } else {
 765             printf "\t%s",$opcode->out();
 766         }
 767     }
 768
 769     print $line,"\n";
 770 }
 771
 772 print "\n$current_segment\tENDS\n"      if ($current_segment && $masm);
 773 print "END\n"                           if ($masm);
 774
 775 close STDOUT;
 776
 777 \f#################################################
 778 # Cross-reference x86_64 ABI "card"
 779 #
 780 #               Unix            Win64
 781 # %rax          *               *
 782 # %rbx          -               -
 783 # %rcx          #4              #1
 784 # %rdx          #3              #2
 785 # %rsi          #2              -
 786 # %rdi          #1              -
 787 # %rbp          -               -
 788 # %rsp          -               -
 789 # %r8           #5              #3
 790 # %r9           #6              #4
 791 # %r10          *               *
 792 # %r11          *               *
 793 # %r12          -               -
 794 # %r13          -               -
 795 # %r14          -               -
 796 # %r15          -               -
 797 #
 798 # (*)   volatile register
 799 # (-)   preserved by callee
 800 # (#)   Nth argument, volatile
 801 #
 802 # In Unix terms top of stack is argument transfer area for arguments
 803 # which could not be accomodated in registers. Or in other words 7th
 804 # [integer] argument resides at 8(%rsp) upon function entry point.
 805 # 128 bytes above %rsp constitute a "red zone" which is not touched
 806 # by signal handlers and can be used as temporal storage without
 807 # allocating a frame.
 808 #
 809 # In Win64 terms N*8 bytes on top of stack is argument transfer area,
 810 # which belongs to/can be overwritten by callee. N is the number of
 811 # arguments passed to callee, *but* not less than 4! This means that
 812 # upon function entry point 5th argument resides at 40(%rsp), as well
 813 # as that 32 bytes from 8(%rsp) can always be used as temporal
 814 # storage [without allocating a frame]. One can actually argue that
 815 # one can assume a "red zone" above stack pointer under Win64 as well.
 816 # Point is that at apparently no occasion Windows kernel would alter
 817 # the area above user stack pointer in true asynchronous manner...
 818 #
 819 # All the above means that if assembler programmer adheres to Unix
 820 # register and stack layout, but disregards the "red zone" existense,
 821 # it's possible to use following prologue and epilogue to "gear" from
 822 # Unix to Win64 ABI in leaf functions with not more than 6 arguments.
 823 #
 824 # omnipotent_function:
 825 # ifdef WIN64
 826 #       movq    %rdi,8(%rsp)
 827 #       movq    %rsi,16(%rsp)
 828 #       movq    %rcx,%rdi       ; if 1st argument is actually present
 829 #       movq    %rdx,%rsi       ; if 2nd argument is actually ...
 830 #       movq    %r8,%rdx        ; if 3rd argument is ...
 831 #       movq    %r9,%rcx        ; if 4th argument ...
 832 #       movq    40(%rsp),%r8    ; if 5th ...
 833 #       movq    48(%rsp),%r9    ; if 6th ...
 834 # endif
 835 #       ...
 836 # ifdef WIN64
 837 #       movq    8(%rsp),%rdi
 838 #       movq    16(%rsp),%rsi
 839 # endif
 840 #       ret
 841 #
 842 \f#################################################
 843 # Win64 SEH, Structured Exception Handling.
 844 #
 845 # Unlike on Unix systems(*) lack of Win64 stack unwinding information
 846 # has undesired side-effect at run-time: if an exception is raised in
 847 # assembler subroutine such as those in question (basically we're
 848 # referring to segmentation violations caused by malformed input
 849 # parameters), the application is briskly terminated without invoking
 850 # any exception handlers, most notably without generating memory dump
 851 # or any user notification whatsoever. This poses a problem. It's
 852 # possible to address it by registering custom language-specific
 853 # handler that would restore processor context to the state at
 854 # subroutine entry point and return "exception is not handled, keep
 855 # unwinding" code. Writing such handler can be a challenge... But it's
 856 # doable, though requires certain coding convention. Consider following
 857 # snippet:
 858 #
 859 # .type function,@function
 860 # function:
 861 #       movq    %rsp,%rax       # copy rsp to volatile register
 862 #       pushq   %r15            # save non-volatile registers
 863 #       pushq   %rbx
 864 #       pushq   %rbp
 865 #       movq    %rsp,%r11
 866 #       subq    %rdi,%r11       # prepare [variable] stack frame
 867 #       andq    $-64,%r11
 868 #       movq    %rax,0(%r11)    # check for exceptions
 869 #       movq    %r11,%rsp       # allocate [variable] stack frame
 870 #       movq    %rax,0(%rsp)    # save original rsp value
 871 # magic_point:
 872 #       ...
 873 #       movq    0(%rsp),%rcx    # pull original rsp value
 874 #       movq    -24(%rcx),%rbp  # restore non-volatile registers
 875 #       movq    -16(%rcx),%rbx
 876 #       movq    -8(%rcx),%r15
 877 #       movq    %rcx,%rsp       # restore original rsp
 878 #       ret
 879 # .size function,.-function
 880 #
 881 # The key is that up to magic_point copy of original rsp value remains
 882 # in chosen volatile register and no non-volatile register, except for
 883 # rsp, is modified. While past magic_point rsp remains constant till
 884 # the very end of the function. In this case custom language-specific
 885 # exception handler would look like this:
 886 #
 887 # EXCEPTION_DISPOSITION handler (EXCEPTION_RECORD *rec,ULONG64 frame,
 888 #               CONTEXT *context,DISPATCHER_CONTEXT *disp)
 889 # {     ULONG64 *rsp = (ULONG64 *)context->Rax;
 890 #       if (context->Rip >= magic_point)
 891 #       {   rsp = ((ULONG64 **)context->Rsp)[0];
 892 #           context->Rbp = rsp[-3];
 893 #           context->Rbx = rsp[-2];
 894 #           context->R15 = rsp[-1];
 895 #       }
 896 #       context->Rsp = (ULONG64)rsp;
 897 #       context->Rdi = rsp[1];
 898 #       context->Rsi = rsp[2];
 899 #
 900 #       memcpy (disp->ContextRecord,context,sizeof(CONTEXT));
 901 #       RtlVirtualUnwind(UNW_FLAG_NHANDLER,disp->ImageBase,
 902 #               dips->ControlPc,disp->FunctionEntry,disp->ContextRecord,
 903 #               &disp->HandlerData,&disp->EstablisherFrame,NULL);
 904 #       return ExceptionContinueSearch;
 905 # }
 906 #
 907 # It's appropriate to implement this handler in assembler, directly in
 908 # function's module. In order to do that one has to know members'
 909 # offsets in CONTEXT and DISPATCHER_CONTEXT structures and some constant
 910 # values. Here they are:
 911 #
 912 #       CONTEXT.Rax                             120
 913 #       CONTEXT.Rcx                             128
 914 #       CONTEXT.Rdx                             136
 915 #       CONTEXT.Rbx                             144
 916 #       CONTEXT.Rsp                             152
 917 #       CONTEXT.Rbp                             160
 918 #       CONTEXT.Rsi                             168
 919 #       CONTEXT.Rdi                             176
 920 #       CONTEXT.R8                              184
 921 #       CONTEXT.R9                              192
 922 #       CONTEXT.R10                             200
 923 #       CONTEXT.R11                             208
 924 #       CONTEXT.R12                             216
 925 #       CONTEXT.R13                             224
 926 #       CONTEXT.R14                             232
 927 #       CONTEXT.R15                             240
 928 #       CONTEXT.Rip                             248
 929 #       CONTEXT.Xmm6                            512
 930 #       sizeof(CONTEXT)                         1232
 931 #       DISPATCHER_CONTEXT.ControlPc            0
 932 #       DISPATCHER_CONTEXT.ImageBase            8
 933 #       DISPATCHER_CONTEXT.FunctionEntry        16
 934 #       DISPATCHER_CONTEXT.EstablisherFrame     24
 935 #       DISPATCHER_CONTEXT.TargetIp             32
 936 #       DISPATCHER_CONTEXT.ContextRecord        40
 937 #       DISPATCHER_CONTEXT.LanguageHandler      48
 938 #       DISPATCHER_CONTEXT.HandlerData          56
 939 #       UNW_FLAG_NHANDLER                       0
 940 #       ExceptionContinueSearch                 1
 941 #
 942 # In order to tie the handler to the function one has to compose
 943 # couple of structures: one for .xdata segment and one for .pdata.
 944 #
 945 # UNWIND_INFO structure for .xdata segment would be
 946 #
 947 # function_unwind_info:
 948 #       .byte   9,0,0,0
 949 #       .rva    handler
 950 #
 951 # This structure designates exception handler for a function with
 952 # zero-length prologue, no stack frame or frame register.
 953 #
 954 # To facilitate composing of .pdata structures, auto-generated "gear"
 955 # prologue copies rsp value to rax and denotes next instruction with
 956 # .LSEH_begin_{function_name} label. This essentially defines the SEH
 957 # styling rule mentioned in the beginning. Position of this label is
 958 # chosen in such manner that possible exceptions raised in the "gear"
 959 # prologue would be accounted to caller and unwound from latter's frame.
 960 # End of function is marked with respective .LSEH_end_{function_name}
 961 # label. To summarize, .pdata segment would contain
 962 #
 963 #       .rva    .LSEH_begin_function
 964 #       .rva    .LSEH_end_function
 965 #       .rva    function_unwind_info
 966 #
 967 # Reference to functon_unwind_info from .xdata segment is the anchor.
 968 # In case you wonder why references are 32-bit .rvas and not 64-bit
 969 # .quads. References put into these two segments are required to be
 970 # *relative* to the base address of the current binary module, a.k.a.
 971 # image base. No Win64 module, be it .exe or .dll, can be larger than
 972 # 2GB and thus such relative references can be and are accommodated in
 973 # 32 bits.
 974 #
 975 # Having reviewed the example function code, one can argue that "movq
 976 # %rsp,%rax" above is redundant. It is not! Keep in mind that on Unix
 977 # rax would contain an undefined value. If this "offends" you, use
 978 # another register and refrain from modifying rax till magic_point is
 979 # reached, i.e. as if it was a non-volatile register. If more registers
 980 # are required prior [variable] frame setup is completed, note that
 981 # nobody says that you can have only one "magic point." You can
 982 # "liberate" non-volatile registers by denoting last stack off-load
 983 # instruction and reflecting it in finer grade unwind logic in handler.
 984 # After all, isn't it why it's called *language-specific* handler...
 985 #
 986 # Attentive reader can notice that exceptions would be mishandled in
 987 # auto-generated "gear" epilogue. Well, exception effectively can't
 988 # occur there, because if memory area used by it was subject to
 989 # segmentation violation, then it would be raised upon call to the
 990 # function (and as already mentioned be accounted to caller, which is
 991 # not a problem). If you're still not comfortable, then define tail
 992 # "magic point" just prior ret instruction and have handler treat it...
 993 #
 994 # (*)   Note that we're talking about run-time, not debug-time. Lack of
 995 #       unwind information makes debugging hard on both Windows and
 996 #       Unix. "Unlike" referes to the fact that on Unix signal handler
 997 #       will always be invoked, core dumped and appropriate exit code
 998 #       returned to parent (for user notification).