crypto/perlasm/x86_64-xlate.pl

   1 #! /usr/bin/env perl
   2 # Copyright 2005-2016 The OpenSSL Project Authors. All Rights Reserved.
   3 #
   4 # Licensed under the OpenSSL license (the "License").  You may not use
   5 # this file except in compliance with the License.  You can obtain a copy
   6 # in the file LICENSE in the source distribution or at
   7 # https://www.openssl.org/source/license.html
   8
   9
  10 # Ascetic x86_64 AT&T to MASM/NASM assembler translator by <appro>.
  11 #
  12 # Why AT&T to MASM and not vice versa? Several reasons. Because AT&T
  13 # format is way easier to parse. Because it's simpler to "gear" from
  14 # Unix ABI to Windows one [see cross-reference "card" at the end of
  15 # file]. Because Linux targets were available first...
  16 #
  17 # In addition the script also "distills" code suitable for GNU
  18 # assembler, so that it can be compiled with more rigid assemblers,
  19 # such as Solaris /usr/ccs/bin/as.
  20 #
  21 # This translator is not designed to convert *arbitrary* assembler
  22 # code from AT&T format to MASM one. It's designed to convert just
  23 # enough to provide for dual-ABI OpenSSL modules development...
  24 # There *are* limitations and you might have to modify your assembler
  25 # code or this script to achieve the desired result...
  26 #
  27 # Currently recognized limitations:
  28 #
  29 # - can't use multiple ops per line;
  30 #
  31 # Dual-ABI styling rules.
  32 #
  33 # 1. Adhere to Unix register and stack layout [see cross-reference
  34 #    ABI "card" at the end for explanation].
  35 # 2. Forget about "red zone," stick to more traditional blended
  36 #    stack frame allocation. If volatile storage is actually required
  37 #    that is. If not, just leave the stack as is.
  38 # 3. Functions tagged with ".type name,@function" get crafted with
  39 #    unified Win64 prologue and epilogue automatically. If you want
  40 #    to take care of ABI differences yourself, tag functions as
  41 #    ".type name,@abi-omnipotent" instead.
  42 # 4. To optimize the Win64 prologue you can specify number of input
  43 #    arguments as ".type name,@function,N." Keep in mind that if N is
  44 #    larger than 6, then you *have to* write "abi-omnipotent" code,
  45 #    because >6 cases can't be addressed with unified prologue.
  46 # 5. Name local labels as .L*, do *not* use dynamic labels such as 1:
  47 #    (sorry about latter).
  48 # 6. Don't use [or hand-code with .byte] "rep ret." "ret" mnemonic is
  49 #    required to identify the spots, where to inject Win64 epilogue!
  50 #    But on the pros, it's then prefixed with rep automatically:-)
  51 # 7. Stick to explicit ip-relative addressing. If you have to use
  52 #    GOTPCREL addressing, stick to mov symbol@GOTPCREL(%rip),%r??.
  53 #    Both are recognized and translated to proper Win64 addressing
  54 #    modes.
  55 #
  56 # 8. In order to provide for structured exception handling unified
  57 #    Win64 prologue copies %rsp value to %rax. For further details
  58 #    see SEH paragraph at the end.
  59 # 9. .init segment is allowed to contain calls to functions only.
  60 # a. If function accepts more than 4 arguments *and* >4th argument
  61 #    is declared as non 64-bit value, do clear its upper part.
  62 \f
  63
  64 use strict;
  65
  66 my $flavour = shift;
  67 my $output  = shift;
  68 if ($flavour =~ /\./) { $output = $flavour; undef $flavour; }
  69
  70 open STDOUT,">$output" || die "can't open $output: $!"
  71         if (defined($output));
  72
  73 my $gas=1;      $gas=0 if ($output =~ /\.asm$/);
  74 my $elf=1;      $elf=0 if (!$gas);
  75 my $win64=0;
  76 my $prefix="";
  77 my $decor=".L";
  78
  79 my $masmref=8 + 50727*2**-32;   # 8.00.50727 shipped with VS2005
  80 my $masm=0;
  81 my $PTR=" PTR";
  82
  83 my $nasmref=2.03;
  84 my $nasm=0;
  85
  86 if    ($flavour eq "mingw64")   { $gas=1; $elf=0; $win64=1;
  87                                   $prefix=`echo __USER_LABEL_PREFIX__ | $ENV{CC} -E -P -`;
  88                                   $prefix =~ s|\R$||; # Better chomp
  89                                 }
  90 elsif ($flavour eq "macosx")    { $gas=1; $elf=0; $prefix="_"; $decor="L\$"; }
  91 elsif ($flavour eq "masm")      { $gas=0; $elf=0; $masm=$masmref; $win64=1; $decor="\$L\$"; }
  92 elsif ($flavour eq "nasm")      { $gas=0; $elf=0; $nasm=$nasmref; $win64=1; $decor="\$L\$"; $PTR=""; }
  93 elsif (!$gas)
  94 {   if ($ENV{ASM} =~ m/nasm/ && `nasm -v` =~ m/version ([0-9]+)\.([0-9]+)/i)
  95     {   $nasm = $1 + $2*0.01; $PTR="";  }
  96     elsif (`ml64 2>&1` =~ m/Version ([0-9]+)\.([0-9]+)(\.([0-9]+))?/)
  97     {   $masm = $1 + $2*2**-16 + $4*2**-32;   }
  98     die "no assembler found on %PATH%" if (!($nasm || $masm));
  99     $win64=1;
 100     $elf=0;
 101     $decor="\$L\$";
 102 }
 103
 104 my $current_segment;
 105 my $current_function;
 106 my %globals;
 107
 108 { package opcode;       # pick up opcodes
 109     sub re {
 110         my      ($class, $line) = @_;
 111         my      $self = {};
 112         my      $ret;
 113
 114         if ($$line =~ /^([a-z][a-z0-9]*)/i) {
 115             bless $self,$class;
 116             $self->{op} = $1;
 117             $ret = $self;
 118             $$line = substr($$line,@+[0]); $$line =~ s/^\s+//;
 119
 120             undef $self->{sz};
 121             if ($self->{op} =~ /^(movz)x?([bw]).*/) {   # movz is pain...
 122                 $self->{op} = $1;
 123                 $self->{sz} = $2;
 124             } elsif ($self->{op} =~ /call|jmp/) {
 125                 $self->{sz} = "";
 126             } elsif ($self->{op} =~ /^p/ && $' !~ /^(ush|op|insrw)/) { # SSEn
 127                 $self->{sz} = "";
 128             } elsif ($self->{op} =~ /^[vk]/) { # VEX or k* such as kmov
 129                 $self->{sz} = "";
 130             } elsif ($self->{op} =~ /mov[dq]/ && $$line =~ /%xmm/) {
 131                 $self->{sz} = "";
 132             } elsif ($self->{op} =~ /([a-z]{3,})([qlwb])$/) {
 133                 $self->{op} = $1;
 134                 $self->{sz} = $2;
 135             }
 136         }
 137         $ret;
 138     }
 139     sub size {
 140         my ($self, $sz) = @_;
 141         $self->{sz} = $sz if (defined($sz) && !defined($self->{sz}));
 142         $self->{sz};
 143     }
 144     sub out {
 145         my $self = shift;
 146         if ($gas) {
 147             if ($self->{op} eq "movz") {        # movz is pain...
 148                 sprintf "%s%s%s",$self->{op},$self->{sz},shift;
 149             } elsif ($self->{op} =~ /^set/) {
 150                 "$self->{op}";
 151             } elsif ($self->{op} eq "ret") {
 152                 my $epilogue = "";
 153                 if ($win64 && $current_function->{abi} eq "svr4") {
 154                     $epilogue = "movq   8(%rsp),%rdi\n\t" .
 155                                 "movq   16(%rsp),%rsi\n\t";
 156                 }
 157                 $epilogue . ".byte      0xf3,0xc3";
 158             } elsif ($self->{op} eq "call" && !$elf && $current_segment eq ".init") {
 159                 ".p2align\t3\n\t.quad";
 160             } else {
 161                 "$self->{op}$self->{sz}";
 162             }
 163         } else {
 164             $self->{op} =~ s/^movz/movzx/;
 165             if ($self->{op} eq "ret") {
 166                 $self->{op} = "";
 167                 if ($win64 && $current_function->{abi} eq "svr4") {
 168                     $self->{op} = "mov  rdi,QWORD$PTR\[8+rsp\]\t;WIN64 epilogue\n\t".
 169                                   "mov  rsi,QWORD$PTR\[16+rsp\]\n\t";
 170                 }
 171                 $self->{op} .= "DB\t0F3h,0C3h\t\t;repret";
 172             } elsif ($self->{op} =~ /^(pop|push)f/) {
 173                 $self->{op} .= $self->{sz};
 174             } elsif ($self->{op} eq "call" && $current_segment eq ".CRT\$XCU") {
 175                 $self->{op} = "\tDQ";
 176             }
 177             $self->{op};
 178         }
 179     }
 180     sub mnemonic {
 181         my ($self, $op) = @_;
 182         $self->{op}=$op if (defined($op));
 183         $self->{op};
 184     }
 185 }
 186 { package const;        # pick up constants, which start with $
 187     sub re {
 188         my      ($class, $line) = @_;
 189         my      $self = {};
 190         my      $ret;
 191
 192         if ($$line =~ /^\$([^,]+)/) {
 193             bless $self, $class;
 194             $self->{value} = $1;
 195             $ret = $self;
 196             $$line = substr($$line,@+[0]); $$line =~ s/^\s+//;
 197         }
 198         $ret;
 199     }
 200     sub out {
 201         my $self = shift;
 202
 203         $self->{value} =~ s/\b(0b[0-1]+)/oct($1)/eig;
 204         if ($gas) {
 205             # Solaris /usr/ccs/bin/as can't handle multiplications
 206             # in $self->{value}
 207             my $value = $self->{value};
 208             no warnings;    # oct might complain about overflow, ignore here...
 209             $value =~ s/(?<![\w\$\.])(0x?[0-9a-f]+)/oct($1)/egi;
 210             if ($value =~ s/([0-9]+\s*[\*\/\%]\s*[0-9]+)/eval($1)/eg) {
 211                 $self->{value} = $value;
 212             }
 213             sprintf "\$%s",$self->{value};
 214         } else {
 215             my $value = $self->{value};
 216             $value =~ s/0x([0-9a-f]+)/0$1h/ig if ($masm);
 217             sprintf "%s",$value;
 218         }
 219     }
 220 }
 221 { package ea;           # pick up effective addresses: expr(%reg,%reg,scale)
 222
 223     my %szmap = (       b=>"BYTE$PTR",    w=>"WORD$PTR",
 224                         l=>"DWORD$PTR",   d=>"DWORD$PTR",
 225                         q=>"QWORD$PTR",   o=>"OWORD$PTR",
 226                         x=>"XMMWORD$PTR", y=>"YMMWORD$PTR",
 227                         z=>"ZMMWORD$PTR" ) if (!$gas);
 228
 229     sub re {
 230         my      ($class, $line, $opcode) = @_;
 231         my      $self = {};
 232         my      $ret;
 233
 234         # optional * ----vvv--- appears in indirect jmp/call
 235         if ($$line =~ /^(\*?)([^\(,]*)\(([%\w,]+)\)((?:{[^}]+})*)/) {
 236             bless $self, $class;
 237             $self->{asterisk} = $1;
 238             $self->{label} = $2;
 239             ($self->{base},$self->{index},$self->{scale})=split(/,/,$3);
 240             $self->{scale} = 1 if (!defined($self->{scale}));
 241             $self->{opmask} = $4;
 242             $ret = $self;
 243             $$line = substr($$line,@+[0]); $$line =~ s/^\s+//;
 244
 245             if ($win64 && $self->{label} =~ s/\@GOTPCREL//) {
 246                 die if ($opcode->mnemonic() ne "mov");
 247                 $opcode->mnemonic("lea");
 248             }
 249             $self->{base}  =~ s/^%//;
 250             $self->{index} =~ s/^%// if (defined($self->{index}));
 251             $self->{opcode} = $opcode;
 252         }
 253         $ret;
 254     }
 255     sub size {}
 256     sub out {
 257         my ($self, $sz) = @_;
 258
 259         $self->{label} =~ s/([_a-z][_a-z0-9]*)/$globals{$1} or $1/gei;
 260         $self->{label} =~ s/\.L/$decor/g;
 261
 262         # Silently convert all EAs to 64-bit. This is required for
 263         # elder GNU assembler and results in more compact code,
 264         # *but* most importantly AES module depends on this feature!
 265         $self->{index} =~ s/^[er](.?[0-9xpi])[d]?$/r\1/;
 266         $self->{base}  =~ s/^[er](.?[0-9xpi])[d]?$/r\1/;
 267
 268         # Solaris /usr/ccs/bin/as can't handle multiplications
 269         # in $self->{label}...
 270         use integer;
 271         $self->{label} =~ s/(?<![\w\$\.])(0x?[0-9a-f]+)/oct($1)/egi;
 272         $self->{label} =~ s/\b([0-9]+\s*[\*\/\%]\s*[0-9]+)\b/eval($1)/eg;
 273
 274         # Some assemblers insist on signed presentation of 32-bit
 275         # offsets, but sign extension is a tricky business in perl...
 276         if ((1<<31)<<1) {
 277             $self->{label} =~ s/\b([0-9]+)\b/$1<<32>>32/eg;
 278         } else {
 279             $self->{label} =~ s/\b([0-9]+)\b/$1>>0/eg;
 280         }
 281
 282         # if base register is %rbp or %r13, see if it's possible to
 283         # flip base and index registers [for better performance]
 284         if (!$self->{label} && $self->{index} && $self->{scale}==1 &&
 285             $self->{base} =~ /(rbp|r13)/) {
 286                 $self->{base} = $self->{index}; $self->{index} = $1;
 287         }
 288
 289         if ($gas) {
 290             $self->{label} =~ s/^___imp_/__imp__/   if ($flavour eq "mingw64");
 291
 292             if (defined($self->{index})) {
 293                 sprintf "%s%s(%s,%%%s,%d)%s",
 294                                         $self->{asterisk},$self->{label},
 295                                         $self->{base}?"%$self->{base}":"",
 296                                         $self->{index},$self->{scale},
 297                                         $self->{opmask};
 298             } else {
 299                 sprintf "%s%s(%%%s)%s", $self->{asterisk},$self->{label},
 300                                         $self->{base},$self->{opmask};
 301             }
 302         } else {
 303             $self->{label} =~ s/\./\$/g;
 304             $self->{label} =~ s/(?<![\w\$\.])0x([0-9a-f]+)/0$1h/ig;
 305             $self->{label} = "($self->{label})" if ($self->{label} =~ /[\*\+\-\/]/);
 306
 307             my $mnemonic = $self->{opcode}->mnemonic();
 308             ($self->{asterisk})                         && ($sz="q") ||
 309             ($mnemonic =~ /^v?mov([qd])$/)              && ($sz=$1)  ||
 310             ($mnemonic =~ /^v?pinsr([qdwb])$/)          && ($sz=$1)  ||
 311             ($mnemonic =~ /^vpbroadcast([qdwb])$/)      && ($sz=$1)  ||
 312             ($mnemonic =~ /^v(?!perm)[a-z]+[fi]128$/)   && ($sz="x");
 313
 314             $self->{opmask}  =~ s/%(k[0-7])/$1/;
 315
 316             if (defined($self->{index})) {
 317                 sprintf "%s[%s%s*%d%s]%s",$szmap{$sz},
 318                                         $self->{label}?"$self->{label}+":"",
 319                                         $self->{index},$self->{scale},
 320                                         $self->{base}?"+$self->{base}":"",
 321                                         $self->{opmask};
 322             } elsif ($self->{base} eq "rip") {
 323                 sprintf "%s[%s]",$szmap{$sz},$self->{label};
 324             } else {
 325                 sprintf "%s[%s%s]%s",   $szmap{$sz},
 326                                         $self->{label}?"$self->{label}+":"",
 327                                         $self->{base},$self->{opmask};
 328             }
 329         }
 330     }
 331 }
 332 { package register;     # pick up registers, which start with %.
 333     sub re {
 334         my      ($class, $line, $opcode) = @_;
 335         my      $self = {};
 336         my      $ret;
 337
 338         # optional * ----vvv--- appears in indirect jmp/call
 339         if ($$line =~ /^(\*?)%(\w+)((?:{[^}]+})*)/) {
 340             bless $self,$class;
 341             $self->{asterisk} = $1;
 342             $self->{value} = $2;
 343             $self->{opmask} = $3;
 344             $opcode->size($self->size());
 345             $ret = $self;
 346             $$line = substr($$line,@+[0]); $$line =~ s/^\s+//;
 347         }
 348         $ret;
 349     }
 350     sub size {
 351         my      $self = shift;
 352         my      $ret;
 353
 354         if    ($self->{value} =~ /^r[\d]+b$/i)  { $ret="b"; }
 355         elsif ($self->{value} =~ /^r[\d]+w$/i)  { $ret="w"; }
 356         elsif ($self->{value} =~ /^r[\d]+d$/i)  { $ret="l"; }
 357         elsif ($self->{value} =~ /^r[\w]+$/i)   { $ret="q"; }
 358         elsif ($self->{value} =~ /^[a-d][hl]$/i){ $ret="b"; }
 359         elsif ($self->{value} =~ /^[\w]{2}l$/i) { $ret="b"; }
 360         elsif ($self->{value} =~ /^[\w]{2}$/i)  { $ret="w"; }
 361         elsif ($self->{value} =~ /^e[a-z]{2}$/i){ $ret="l"; }
 362
 363         $ret;
 364     }
 365     sub out {
 366         my $self = shift;
 367         if ($gas)       { sprintf "%s%%%s%s",   $self->{asterisk},
 368                                                 $self->{value},
 369                                                 $self->{opmask}; }
 370         else            { $self->{opmask} =~ s/%(k[0-7])/$1/;
 371                           $self->{value}.$self->{opmask}; }
 372     }
 373 }
 374 { package label;        # pick up labels, which end with :
 375     sub re {
 376         my      ($class, $line) = @_;
 377         my      $self = {};
 378         my      $ret;
 379
 380         if ($$line =~ /(^[\.\w]+)\:/) {
 381             bless $self,$class;
 382             $self->{value} = $1;
 383             $ret = $self;
 384             $$line = substr($$line,@+[0]); $$line =~ s/^\s+//;
 385
 386             $self->{value} =~ s/^\.L/$decor/;
 387         }
 388         $ret;
 389     }
 390     sub out {
 391         my $self = shift;
 392
 393         if ($gas) {
 394             my $func = ($globals{$self->{value}} or $self->{value}) . ":";
 395             if ($win64  && $current_function->{name} eq $self->{value}
 396                         && $current_function->{abi} eq "svr4") {
 397                 $func .= "\n";
 398                 $func .= "      movq    %rdi,8(%rsp)\n";
 399                 $func .= "      movq    %rsi,16(%rsp)\n";
 400                 $func .= "      movq    %rsp,%rax\n";
 401                 $func .= "${decor}SEH_begin_$current_function->{name}:\n";
 402                 my $narg = $current_function->{narg};
 403                 $narg=6 if (!defined($narg));
 404                 $func .= "      movq    %rcx,%rdi\n" if ($narg>0);
 405                 $func .= "      movq    %rdx,%rsi\n" if ($narg>1);
 406                 $func .= "      movq    %r8,%rdx\n"  if ($narg>2);
 407                 $func .= "      movq    %r9,%rcx\n"  if ($narg>3);
 408                 $func .= "      movq    40(%rsp),%r8\n" if ($narg>4);
 409                 $func .= "      movq    48(%rsp),%r9\n" if ($narg>5);
 410             }
 411             $func;
 412         } elsif ($self->{value} ne "$current_function->{name}") {
 413             # Make all labels in masm global.
 414             $self->{value} .= ":" if ($masm);
 415             $self->{value} . ":";
 416         } elsif ($win64 && $current_function->{abi} eq "svr4") {
 417             my $func =  "$current_function->{name}" .
 418                         ($nasm ? ":" : "\tPROC $current_function->{scope}") .
 419                         "\n";
 420             $func .= "  mov     QWORD$PTR\[8+rsp\],rdi\t;WIN64 prologue\n";
 421             $func .= "  mov     QWORD$PTR\[16+rsp\],rsi\n";
 422             $func .= "  mov     rax,rsp\n";
 423             $func .= "${decor}SEH_begin_$current_function->{name}:";
 424             $func .= ":" if ($masm);
 425             $func .= "\n";
 426             my $narg = $current_function->{narg};
 427             $narg=6 if (!defined($narg));
 428             $func .= "  mov     rdi,rcx\n" if ($narg>0);
 429             $func .= "  mov     rsi,rdx\n" if ($narg>1);
 430             $func .= "  mov     rdx,r8\n"  if ($narg>2);
 431             $func .= "  mov     rcx,r9\n"  if ($narg>3);
 432             $func .= "  mov     r8,QWORD$PTR\[40+rsp\]\n" if ($narg>4);
 433             $func .= "  mov     r9,QWORD$PTR\[48+rsp\]\n" if ($narg>5);
 434             $func .= "\n";
 435         } else {
 436            "$current_function->{name}".
 437                         ($nasm ? ":" : "\tPROC $current_function->{scope}");
 438         }
 439     }
 440 }
 441 { package expr;         # pick up expressions
 442     sub re {
 443         my      ($class, $line, $opcode) = @_;
 444         my      $self = {};
 445         my      $ret;
 446
 447         if ($$line =~ /(^[^,]+)/) {
 448             bless $self,$class;
 449             $self->{value} = $1;
 450             $ret = $self;
 451             $$line = substr($$line,@+[0]); $$line =~ s/^\s+//;
 452
 453             $self->{value} =~ s/\@PLT// if (!$elf);
 454             $self->{value} =~ s/([_a-z][_a-z0-9]*)/$globals{$1} or $1/gei;
 455             $self->{value} =~ s/\.L/$decor/g;
 456             $self->{opcode} = $opcode;
 457         }
 458         $ret;
 459     }
 460     sub out {
 461         my $self = shift;
 462         if ($nasm && $self->{opcode}->mnemonic()=~m/^j(?![re]cxz)/) {
 463             "NEAR ".$self->{value};
 464         } else {
 465             $self->{value};
 466         }
 467     }
 468 }
 469 { package cfi_directive;
 470     # CFI directives annotate instructions that are significant for
 471     # stack unwinding procedure compliant with DWARF specification,
 472     # see http://dwarfstd.org/. Besides naturally expected for this
 473     # script platform-specific filtering function, this module adds
 474     # three auxiliary synthetic directives not recognized by [GNU]
 475     # assembler:
 476     #
 477     # - .cfi_push to annotate push instructions in prologue, which
 478     #   translates to .cfi_adjust_cfa_offset (if needed) and
 479     #   .cfi_offset;
 480     # - .cfi_pop to annotate pop instructions in epilogue, which
 481     #   translates to .cfi_adjust_cfa_offset (if needed) and
 482     #   .cfi_restore;
 483     # - [and most notably] .cfi_cfa_expression which encodes
 484     #   DW_CFA_def_cfa_expression and passes it to .cfi_escape as
 485     #   byte vector;
 486     #
 487     # CFA expressions were introduced in DWARF specification version
 488     # 3 and describe how to deduce CFA, Canonical Frame Address. This
 489     # becomes handy if your stack frame is variable and you can't
 490     # spare register for [previous] frame pointer. Suggested directive
 491     # syntax is made-up mix of DWARF operator suffixes [subset of]
 492     # and references to registers with optional bias. Following example
 493     # describes offloaded *original* stack pointer at specific offset
 494     # from *current* stack pointer:
 495     #
 496     #   .cfi_cfa_expression     %rsp+40,deref,+8
 497     #
 498     # Final +8 has everything to do with the fact that CFA is defined
 499     # as reference to top of caller's stack, and on x86_64 call to
 500     # subroutine pushes 8-byte return address. In other words original
 501     # stack pointer upon entry to a subroutine is 8 bytes off from CFA.
 502
 503     # Below constants are taken from "DWARF Expressions" section of the
 504     # DWARF specification, section is numbered 7.7 in versions 3 and 4.
 505     my %DW_OP_simple = (        # no-arg operators, mapped directly
 506         deref   => 0x06,        dup     => 0x12,
 507         drop    => 0x13,        over    => 0x14,
 508         pick    => 0x15,        swap    => 0x16,
 509         rot     => 0x17,        xderef  => 0x18,
 510
 511         abs     => 0x19,        and     => 0x1a,
 512         div     => 0x1b,        minus   => 0x1c,
 513         mod     => 0x1d,        mul     => 0x1e,
 514         neg     => 0x1f,        not     => 0x20,
 515         or      => 0x21,        plus    => 0x22,
 516         shl     => 0x24,        shr     => 0x25,
 517         shra    => 0x26,        xor     => 0x27,
 518         );
 519
 520     my %DW_OP_complex = (       # used in specific subroutines
 521         constu          => 0x10,        # uleb128
 522         consts          => 0x11,        # sleb128
 523         plus_uconst     => 0x23,        # uleb128
 524         lit0            => 0x30,        # add 0-31 to opcode
 525         reg0            => 0x50,        # add 0-31 to opcode
 526         breg0           => 0x70,        # add 0-31 to opcole, sleb128
 527         regx            => 0x90,        # uleb28
 528         fbreg           => 0x91,        # sleb128
 529         bregx           => 0x92,        # uleb128, sleb128
 530         piece           => 0x93,        # uleb128
 531         );
 532
 533     # Following constants are defined in x86_64 ABI supplement, for
 534     # example available at https://www.uclibc.org/docs/psABI-x86_64.pdf,
 535     # see section 3.7 "Stack Unwind Algorithm".
 536     my %DW_reg_idx = (
 537         "%rax"=>0,  "%rdx"=>1,  "%rcx"=>2,  "%rbx"=>3,
 538         "%rsi"=>4,  "%rdi"=>5,  "%rbp"=>6,  "%rsp"=>7,
 539         "%r8" =>8,  "%r9" =>9,  "%r10"=>10, "%r11"=>11,
 540         "%r12"=>12, "%r13"=>13, "%r14"=>14, "%r15"=>15
 541         );
 542
 543     my ($cfa_reg, $cfa_rsp);
 544
 545     # [us]leb128 format is variable-length integer representation base
 546     # 2^128, with most significant bit of each byte being 0 denoting
 547     # *last* most significant digit. See "Variable Length Data" in the
 548     # DWARF specification, numbered 7.6 at least in versions 3 and 4.
 549     sub sleb128 {
 550         use integer;    # get right shift extend sign
 551
 552         my $val = shift;
 553         my $sign = ($val < 0) ? -1 : 0;
 554         my @ret = ();
 555
 556         while(1) {
 557             push @ret, $val&0x7f;
 558
 559             # see if remaining bits are same and equal to most
 560             # significant bit of the current digit, if so, it's
 561             # last digit...
 562             last if (($val>>6) == $sign);
 563
 564             @ret[-1] |= 0x80;
 565             $val >>= 7;
 566         }
 567
 568         return @ret;
 569     }
 570     sub uleb128 {
 571         my $val = shift;
 572         my @ret = ();
 573
 574         while(1) {
 575             push @ret, $val&0x7f;
 576
 577             # see if it's last significant digit...
 578             last if (($val >>= 7) == 0);
 579
 580             @ret[-1] |= 0x80;
 581         }
 582
 583         return @ret;
 584     }
 585     sub const {
 586         my $val = shift;
 587
 588         if ($val >= 0 && $val < 32) {
 589             return ($DW_OP_complex{lit0}+$val);
 590         }
 591         return ($DW_OP_complex{consts}, sleb128($val));
 592     }
 593     sub reg {
 594         my $val = shift;
 595
 596         return if ($val !~ m/^(%r\w+)(?:([\+\-])((?:0x)?[0-9a-f]+))?/);
 597
 598         my $reg = $DW_reg_idx{$1};
 599         my $off = eval ("0 $2 $3");
 600
 601         return (($DW_OP_complex{breg0} + $reg), sleb128($off));
 602         # Yes, we use DW_OP_bregX+0 to push register value and not
 603         # DW_OP_regX, because latter would require even DW_OP_piece,
 604         # which would be a waste under the circumstances. If you have
 605         # to use DWP_OP_reg, use "regx:N"...
 606     }
 607     sub cfa_expression {
 608         my $line = shift;
 609         my @ret;
 610
 611         foreach my $token (split(/,\s*/,$line)) {
 612             if ($token =~ /^%r/) {
 613                 push @ret,reg($token);
 614             } elsif ($token =~ /((?:0x)?[0-9a-f]+)\((%r\w+)\)/) {
 615                 push @ret,reg("$2+$1");
 616             } elsif ($token =~ /(\w+):(\-?(?:0x)?[0-9a-f]+)(U?)/i) {
 617                 my $i = 1*eval($2);
 618                 push @ret,$DW_OP_complex{$1}, ($3 ? uleb128($i) : sleb128($i));
 619             } elsif (my $i = 1*eval($token) or $token eq "0") {
 620                 if ($token =~ /^\+/) {
 621                     push @ret,$DW_OP_complex{plus_uconst},uleb128($i);
 622                 } else {
 623                     push @ret,const($i);
 624                 }
 625             } else {
 626                 push @ret,$DW_OP_simple{$token};
 627             }
 628         }
 629
 630         # Finally we return DW_CFA_def_cfa_expression, 15, followed by
 631         # length of the expression and of course the expression itself.
 632         return (15,scalar(@ret),@ret);
 633     }
 634     sub re {
 635         my      ($class, $line) = @_;
 636         my      $self = {};
 637         my      $ret;
 638
 639         if ($$line =~ s/^\s*\.cfi_(\w+)\s*//) {
 640             bless $self,$class;
 641             $ret = $self;
 642             undef $self->{value};
 643             my $dir = $1;
 644
 645             SWITCH: for ($dir) {
 646             # What is $cfa_rsp? Effectively it's difference between %rsp
 647             # value and current CFA, Canonical Frame Address, which is
 648             # why it starts with -8. Recall that CFA is top of caller's
 649             # stack...
 650             /startproc/ && do { ($cfa_reg, $cfa_rsp) = ("%rsp", -8); last; };
 651             /endproc/   && do { ($cfa_reg, $cfa_rsp) = ("%rsp",  0); last; };
 652             /def_cfa_register/
 653                         && do { $cfa_reg = $$line; last; };
 654             /def_cfa_offset/
 655                         && do { $cfa_rsp = -1*eval($$line) if ($cfa_reg eq "%rsp");
 656                                 last;
 657                               };
 658             /adjust_cfa_offset/
 659                         && do { $cfa_rsp -= 1*eval($$line) if ($cfa_reg eq "%rsp");
 660                                 last;
 661                               };
 662             /def_cfa/   && do { if ($$line =~ /(%r\w+)\s*,\s*(.+)/) {
 663                                     $cfa_reg = $1;
 664                                     $cfa_rsp = -1*eval($2) if ($cfa_reg eq "%rsp");
 665                                 }
 666                                 last;
 667                               };
 668             /push/      && do { $dir = undef;
 669                                 $cfa_rsp -= 8;
 670                                 if ($cfa_reg eq "%rsp") {
 671                                     $self->{value} = ".cfi_adjust_cfa_offset\t8\n";
 672                                 }
 673                                 $self->{value} .= ".cfi_offset\t$$line,$cfa_rsp";
 674                                 last;
 675                               };
 676             /pop/       && do { $dir = undef;
 677                                 $cfa_rsp += 8;
 678                                 if ($cfa_reg eq "%rsp") {
 679                                     $self->{value} = ".cfi_adjust_cfa_offset\t-8\n";
 680                                 }
 681                                 $self->{value} .= ".cfi_restore\t$$line";
 682                                 last;
 683                               };
 684             /cfa_expression/
 685                         && do { $dir = undef;
 686                                 $self->{value} = ".cfi_escape\t" .
 687                                         join(",", map(sprintf("0x%02x", $_),
 688                                                       cfa_expression($$line)));
 689                                 last;
 690                               };
 691             }
 692
 693             $self->{value} = ".cfi_$dir\t$$line" if ($dir);
 694
 695             $$line = "";
 696         }
 697
 698         return $ret;
 699     }
 700     sub out {
 701         my $self = shift;
 702         return ($elf ? $self->{value} : undef);
 703     }
 704 }
 705 { package directive;    # pick up directives, which start with .
 706     sub re {
 707         my      ($class, $line) = @_;
 708         my      $self = {};
 709         my      $ret;
 710         my      $dir;
 711
 712         # chain-call to cfi_directive
 713         $ret = cfi_directive->re($line) and return $ret;
 714
 715         if ($$line =~ /^\s*(\.\w+)/) {
 716             bless $self,$class;
 717             $dir = $1;
 718             $ret = $self;
 719             undef $self->{value};
 720             $$line = substr($$line,@+[0]); $$line =~ s/^\s+//;
 721
 722             SWITCH: for ($dir) {
 723                 /\.global|\.globl|\.extern/
 724                             && do { $globals{$$line} = $prefix . $$line;
 725                                     $$line = $globals{$$line} if ($prefix);
 726                                     last;
 727                                   };
 728                 /\.type/    && do { my ($sym,$type,$narg) = split(',',$$line);
 729                                     if ($type eq "\@function") {
 730                                         undef $current_function;
 731                                         $current_function->{name} = $sym;
 732                                         $current_function->{abi}  = "svr4";
 733                                         $current_function->{narg} = $narg;
 734                                         $current_function->{scope} = defined($globals{$sym})?"PUBLIC":"PRIVATE";
 735                                     } elsif ($type eq "\@abi-omnipotent") {
 736                                         undef $current_function;
 737                                         $current_function->{name} = $sym;
 738                                         $current_function->{scope} = defined($globals{$sym})?"PUBLIC":"PRIVATE";
 739                                     }
 740                                     $$line =~ s/\@abi\-omnipotent/\@function/;
 741                                     $$line =~ s/\@function.*/\@function/;
 742                                     last;
 743                                   };
 744                 /\.asciz/   && do { if ($$line =~ /^"(.*)"$/) {
 745                                         $dir  = ".byte";
 746                                         $$line = join(",",unpack("C*",$1),0);
 747                                     }
 748                                     last;
 749                                   };
 750                 /\.rva|\.long|\.quad/
 751                             && do { $$line =~ s/([_a-z][_a-z0-9]*)/$globals{$1} or $1/gei;
 752                                     $$line =~ s/\.L/$decor/g;
 753                                     last;
 754                                   };
 755             }
 756
 757             if ($gas) {
 758                 $self->{value} = $dir . "\t" . $$line;
 759
 760                 if ($dir =~ /\.extern/) {
 761                     $self->{value} = ""; # swallow extern
 762                 } elsif (!$elf && $dir =~ /\.type/) {
 763                     $self->{value} = "";
 764                     $self->{value} = ".def\t" . ($globals{$1} or $1) . ";\t" .
 765                                 (defined($globals{$1})?".scl 2;":".scl 3;") .
 766                                 "\t.type 32;\t.endef"
 767                                 if ($win64 && $$line =~ /([^,]+),\@function/);
 768                 } elsif (!$elf && $dir =~ /\.size/) {
 769                     $self->{value} = "";
 770                     if (defined($current_function)) {
 771                         $self->{value} .= "${decor}SEH_end_$current_function->{name}:"
 772                                 if ($win64 && $current_function->{abi} eq "svr4");
 773                         undef $current_function;
 774                     }
 775                 } elsif (!$elf && $dir =~ /\.align/) {
 776                     $self->{value} = ".p2align\t" . (log($$line)/log(2));
 777                 } elsif ($dir eq ".section") {
 778                     $current_segment=$$line;
 779                     if (!$elf && $current_segment eq ".init") {
 780                         if      ($flavour eq "macosx")  { $self->{value} = ".mod_init_func"; }
 781                         elsif   ($flavour eq "mingw64") { $self->{value} = ".section\t.ctors"; }
 782                     }
 783                 } elsif ($dir =~ /\.(text|data)/) {
 784                     $current_segment=".$1";
 785                 } elsif ($dir =~ /\.hidden/) {
 786                     if    ($flavour eq "macosx")  { $self->{value} = ".private_extern\t$prefix$$line"; }
 787                     elsif ($flavour eq "mingw64") { $self->{value} = ""; }
 788                 } elsif ($dir =~ /\.comm/) {
 789                     $self->{value} = "$dir\t$prefix$$line";
 790                     $self->{value} =~ s|,([0-9]+),([0-9]+)$|",$1,".log($2)/log(2)|e if ($flavour eq "macosx");
 791                 }
 792                 $$line = "";
 793                 return $self;
 794             }
 795
 796             # non-gas case or nasm/masm
 797             SWITCH: for ($dir) {
 798                 /\.text/    && do { my $v=undef;
 799                                     if ($nasm) {
 800                                         $v="section     .text code align=64\n";
 801                                     } else {
 802                                         $v="$current_segment\tENDS\n" if ($current_segment);
 803                                         $current_segment = ".text\$";
 804                                         $v.="$current_segment\tSEGMENT ";
 805                                         $v.=$masm>=$masmref ? "ALIGN(256)" : "PAGE";
 806                                         $v.=" 'CODE'";
 807                                     }
 808                                     $self->{value} = $v;
 809                                     last;
 810                                   };
 811                 /\.data/    && do { my $v=undef;
 812                                     if ($nasm) {
 813                                         $v="section     .data data align=8\n";
 814                                     } else {
 815                                         $v="$current_segment\tENDS\n" if ($current_segment);
 816                                         $current_segment = "_DATA";
 817                                         $v.="$current_segment\tSEGMENT";
 818                                     }
 819                                     $self->{value} = $v;
 820                                     last;
 821                                   };
 822                 /\.section/ && do { my $v=undef;
 823                                     $$line =~ s/([^,]*).*/$1/;
 824                                     $$line = ".CRT\$XCU" if ($$line eq ".init");
 825                                     if ($nasm) {
 826                                         $v="section     $$line";
 827                                         if ($$line=~/\.([px])data/) {
 828                                             $v.=" rdata align=";
 829                                             $v.=$1 eq "p"? 4 : 8;
 830                                         } elsif ($$line=~/\.CRT\$/i) {
 831                                             $v.=" rdata align=8";
 832                                         }
 833                                     } else {
 834                                         $v="$current_segment\tENDS\n" if ($current_segment);
 835                                         $v.="$$line\tSEGMENT";
 836                                         if ($$line=~/\.([px])data/) {
 837                                             $v.=" READONLY";
 838                                             $v.=" ALIGN(".($1 eq "p" ? 4 : 8).")" if ($masm>=$masmref);
 839                                         } elsif ($$line=~/\.CRT\$/i) {
 840                                             $v.=" READONLY ";
 841                                             $v.=$masm>=$masmref ? "ALIGN(8)" : "DWORD";
 842                                         }
 843                                     }
 844                                     $current_segment = $$line;
 845                                     $self->{value} = $v;
 846                                     last;
 847                                   };
 848                 /\.extern/  && do { $self->{value}  = "EXTERN\t".$$line;
 849                                     $self->{value} .= ":NEAR" if ($masm);
 850                                     last;
 851                                   };
 852                 /\.globl|.global/
 853                             && do { $self->{value}  = $masm?"PUBLIC":"global";
 854                                     $self->{value} .= "\t".$$line;
 855                                     last;
 856                                   };
 857                 /\.size/    && do { if (defined($current_function)) {
 858                                         undef $self->{value};
 859                                         if ($current_function->{abi} eq "svr4") {
 860                                             $self->{value}="${decor}SEH_end_$current_function->{name}:";
 861                                             $self->{value}.=":\n" if($masm);
 862                                         }
 863                                         $self->{value}.="$current_function->{name}\tENDP" if($masm && $current_function->{name});
 864                                         undef $current_function;
 865                                     }
 866                                     last;
 867                                   };
 868                 /\.align/   && do { my $max = ($masm && $masm>=$masmref) ? 256 : 4096;
 869                                     $self->{value} = "ALIGN\t".($$line>$max?$max:$$line);
 870                                     last;
 871                                   };
 872                 /\.(value|long|rva|quad)/
 873                             && do { my $sz  = substr($1,0,1);
 874                                     my @arr = split(/,\s*/,$$line);
 875                                     my $last = pop(@arr);
 876                                     my $conv = sub  {   my $var=shift;
 877                                                         $var=~s/^(0b[0-1]+)/oct($1)/eig;
 878                                                         $var=~s/^0x([0-9a-f]+)/0$1h/ig if ($masm);
 879                                                         if ($sz eq "D" && ($current_segment=~/.[px]data/ || $dir eq ".rva"))
 880                                                         { $var=~s/^([_a-z\$\@][_a-z0-9\$\@]*)/$nasm?"$1 wrt ..imagebase":"imagerel $1"/egi; }
 881                                                         $var;
 882                                                     };
 883
 884                                     $sz =~ tr/bvlrq/BWDDQ/;
 885                                     $self->{value} = "\tD$sz\t";
 886                                     for (@arr) { $self->{value} .= &$conv($_).","; }
 887                                     $self->{value} .= &$conv($last);
 888                                     last;
 889                                   };
 890                 /\.byte/    && do { my @str=split(/,\s*/,$$line);
 891                                     map(s/(0b[0-1]+)/oct($1)/eig,@str);
 892                                     map(s/0x([0-9a-f]+)/0$1h/ig,@str) if ($masm);
 893                                     while ($#str>15) {
 894                                         $self->{value}.="DB\t"
 895                                                 .join(",",@str[0..15])."\n";
 896                                         foreach (0..15) { shift @str; }
 897                                     }
 898                                     $self->{value}.="DB\t"
 899                                                 .join(",",@str) if (@str);
 900                                     last;
 901                                   };
 902                 /\.comm/    && do { my @str=split(/,\s*/,$$line);
 903                                     my $v=undef;
 904                                     if ($nasm) {
 905                                         $v.="common     $prefix@str[0] @str[1]";
 906                                     } else {
 907                                         $v="$current_segment\tENDS\n" if ($current_segment);
 908                                         $current_segment = "_DATA";
 909                                         $v.="$current_segment\tSEGMENT\n";
 910                                         $v.="COMM       @str[0]:DWORD:".@str[1]/4;
 911                                     }
 912                                     $self->{value} = $v;
 913                                     last;
 914                                   };
 915             }
 916             $$line = "";
 917         }
 918
 919         $ret;
 920     }
 921     sub out {
 922         my $self = shift;
 923         $self->{value};
 924     }
 925 }
 926
 927 # Upon initial x86_64 introduction SSE>2 extensions were not introduced
 928 # yet. In order not to be bothered by tracing exact assembler versions,
 929 # but at the same time to provide a bare security minimum of AES-NI, we
 930 # hard-code some instructions. Extensions past AES-NI on the other hand
 931 # are traced by examining assembler version in individual perlasm
 932 # modules...
 933
 934 my %regrm = (   "%eax"=>0, "%ecx"=>1, "%edx"=>2, "%ebx"=>3,
 935                 "%esp"=>4, "%ebp"=>5, "%esi"=>6, "%edi"=>7      );
 936
 937 sub rex {
 938  my $opcode=shift;
 939  my ($dst,$src,$rex)=@_;
 940
 941    $rex|=0x04 if($dst>=8);
 942    $rex|=0x01 if($src>=8);
 943    push @$opcode,($rex|0x40) if ($rex);
 944 }
 945
 946 my $movq = sub {        # elderly gas can't handle inter-register movq
 947   my $arg = shift;
 948   my @opcode=(0x66);
 949     if ($arg =~ /%xmm([0-9]+),\s*%r(\w+)/) {
 950         my ($src,$dst)=($1,$2);
 951         if ($dst !~ /[0-9]+/)   { $dst = $regrm{"%e$dst"}; }
 952         rex(\@opcode,$src,$dst,0x8);
 953         push @opcode,0x0f,0x7e;
 954         push @opcode,0xc0|(($src&7)<<3)|($dst&7);       # ModR/M
 955         @opcode;
 956     } elsif ($arg =~ /%r(\w+),\s*%xmm([0-9]+)/) {
 957         my ($src,$dst)=($2,$1);
 958         if ($dst !~ /[0-9]+/)   { $dst = $regrm{"%e$dst"}; }
 959         rex(\@opcode,$src,$dst,0x8);
 960         push @opcode,0x0f,0x6e;
 961         push @opcode,0xc0|(($src&7)<<3)|($dst&7);       # ModR/M
 962         @opcode;
 963     } else {
 964         ();
 965     }
 966 };
 967
 968 my $pextrd = sub {
 969     if (shift =~ /\$([0-9]+),\s*%xmm([0-9]+),\s*(%\w+)/) {
 970       my @opcode=(0x66);
 971         my $imm=$1;
 972         my $src=$2;
 973         my $dst=$3;
 974         if ($dst =~ /%r([0-9]+)d/)      { $dst = $1; }
 975         elsif ($dst =~ /%e/)            { $dst = $regrm{$dst}; }
 976         rex(\@opcode,$src,$dst);
 977         push @opcode,0x0f,0x3a,0x16;
 978         push @opcode,0xc0|(($src&7)<<3)|($dst&7);       # ModR/M
 979         push @opcode,$imm;
 980         @opcode;
 981     } else {
 982         ();
 983     }
 984 };
 985
 986 my $pinsrd = sub {
 987     if (shift =~ /\$([0-9]+),\s*(%\w+),\s*%xmm([0-9]+)/) {
 988       my @opcode=(0x66);
 989         my $imm=$1;
 990         my $src=$2;
 991         my $dst=$3;
 992         if ($src =~ /%r([0-9]+)/)       { $src = $1; }
 993         elsif ($src =~ /%e/)            { $src = $regrm{$src}; }
 994         rex(\@opcode,$dst,$src);
 995         push @opcode,0x0f,0x3a,0x22;
 996         push @opcode,0xc0|(($dst&7)<<3)|($src&7);       # ModR/M
 997         push @opcode,$imm;
 998         @opcode;
 999     } else {
1000         ();
1001     }
1002 };
1003
1004 my $pshufb = sub {
1005     if (shift =~ /%xmm([0-9]+),\s*%xmm([0-9]+)/) {
1006       my @opcode=(0x66);
1007         rex(\@opcode,$2,$1);
1008         push @opcode,0x0f,0x38,0x00;
1009         push @opcode,0xc0|($1&7)|(($2&7)<<3);           # ModR/M
1010         @opcode;
1011     } else {
1012         ();
1013     }
1014 };
1015
1016 my $palignr = sub {
1017     if (shift =~ /\$([0-9]+),\s*%xmm([0-9]+),\s*%xmm([0-9]+)/) {
1018       my @opcode=(0x66);
1019         rex(\@opcode,$3,$2);
1020         push @opcode,0x0f,0x3a,0x0f;
1021         push @opcode,0xc0|($2&7)|(($3&7)<<3);           # ModR/M
1022         push @opcode,$1;
1023         @opcode;
1024     } else {
1025         ();
1026     }
1027 };
1028
1029 my $pclmulqdq = sub {
1030     if (shift =~ /\$([x0-9a-f]+),\s*%xmm([0-9]+),\s*%xmm([0-9]+)/) {
1031       my @opcode=(0x66);
1032         rex(\@opcode,$3,$2);
1033         push @opcode,0x0f,0x3a,0x44;
1034         push @opcode,0xc0|($2&7)|(($3&7)<<3);           # ModR/M
1035         my $c=$1;
1036         push @opcode,$c=~/^0/?oct($c):$c;
1037         @opcode;
1038     } else {
1039         ();
1040     }
1041 };
1042
1043 my $rdrand = sub {
1044     if (shift =~ /%[er](\w+)/) {
1045       my @opcode=();
1046       my $dst=$1;
1047         if ($dst !~ /[0-9]+/) { $dst = $regrm{"%e$dst"}; }
1048         rex(\@opcode,0,$dst,8);
1049         push @opcode,0x0f,0xc7,0xf0|($dst&7);
1050         @opcode;
1051     } else {
1052         ();
1053     }
1054 };
1055
1056 my $rdseed = sub {
1057     if (shift =~ /%[er](\w+)/) {
1058       my @opcode=();
1059       my $dst=$1;
1060         if ($dst !~ /[0-9]+/) { $dst = $regrm{"%e$dst"}; }
1061         rex(\@opcode,0,$dst,8);
1062         push @opcode,0x0f,0xc7,0xf8|($dst&7);
1063         @opcode;
1064     } else {
1065         ();
1066     }
1067 };
1068
1069 # Not all AVX-capable assemblers recognize AMD XOP extension. Since we
1070 # are using only two instructions hand-code them in order to be excused
1071 # from chasing assembler versions...
1072
1073 sub rxb {
1074  my $opcode=shift;
1075  my ($dst,$src1,$src2,$rxb)=@_;
1076
1077    $rxb|=0x7<<5;
1078    $rxb&=~(0x04<<5) if($dst>=8);
1079    $rxb&=~(0x01<<5) if($src1>=8);
1080    $rxb&=~(0x02<<5) if($src2>=8);
1081    push @$opcode,$rxb;
1082 }
1083
1084 my $vprotd = sub {
1085     if (shift =~ /\$([x0-9a-f]+),\s*%xmm([0-9]+),\s*%xmm([0-9]+)/) {
1086       my @opcode=(0x8f);
1087         rxb(\@opcode,$3,$2,-1,0x08);
1088         push @opcode,0x78,0xc2;
1089         push @opcode,0xc0|($2&7)|(($3&7)<<3);           # ModR/M
1090         my $c=$1;
1091         push @opcode,$c=~/^0/?oct($c):$c;
1092         @opcode;
1093     } else {
1094         ();
1095     }
1096 };
1097
1098 my $vprotq = sub {
1099     if (shift =~ /\$([x0-9a-f]+),\s*%xmm([0-9]+),\s*%xmm([0-9]+)/) {
1100       my @opcode=(0x8f);
1101         rxb(\@opcode,$3,$2,-1,0x08);
1102         push @opcode,0x78,0xc3;
1103         push @opcode,0xc0|($2&7)|(($3&7)<<3);           # ModR/M
1104         my $c=$1;
1105         push @opcode,$c=~/^0/?oct($c):$c;
1106         @opcode;
1107     } else {
1108         ();
1109     }
1110 };
1111
1112 # Intel Control-flow Enforcement Technology extension. All functions and
1113 # indirect branch targets will have to start with this instruction...
1114
1115 my $endbranch = sub {
1116     (0xf3,0x0f,0x1e,0xfa);
1117 };
1118
1119 ########################################################################
1120
1121 if ($nasm) {
1122     print <<___;
1123 default rel
1124 %define XMMWORD
1125 %define YMMWORD
1126 %define ZMMWORD
1127 ___
1128 } elsif ($masm) {
1129     print <<___;
1130 OPTION  DOTNAME
1131 ___
1132 }
1133 while(defined(my $line=<>)) {
1134
1135     $line =~ s|\R$||;           # Better chomp
1136
1137     $line =~ s|[#!].*$||;       # get rid of asm-style comments...
1138     $line =~ s|/\*.*\*/||;      # ... and C-style comments...
1139     $line =~ s|^\s+||;          # ... and skip white spaces in beginning
1140     $line =~ s|\s+$||;          # ... and at the end
1141
1142     if (my $label=label->re(\$line))    { print $label->out(); }
1143
1144     if (my $directive=directive->re(\$line)) {
1145         printf "%s",$directive->out();
1146     } elsif (my $opcode=opcode->re(\$line)) {
1147         my $asm = eval("\$".$opcode->mnemonic());
1148
1149         if ((ref($asm) eq 'CODE') && scalar(my @bytes=&$asm($line))) {
1150             print $gas?".byte\t":"DB\t",join(',',@bytes),"\n";
1151             next;
1152         }
1153
1154         my @args;
1155         ARGUMENT: while (1) {
1156             my $arg;
1157
1158             ($arg=register->re(\$line, $opcode))||
1159             ($arg=const->re(\$line))            ||
1160             ($arg=ea->re(\$line, $opcode))      ||
1161             ($arg=expr->re(\$line, $opcode))    ||
1162             last ARGUMENT;
1163
1164             push @args,$arg;
1165
1166             last ARGUMENT if ($line !~ /^,/);
1167
1168             $line =~ s/^,\s*//;
1169         } # ARGUMENT:
1170
1171         if ($#args>=0) {
1172             my $insn;
1173             my $sz=$opcode->size();
1174
1175             if ($gas) {
1176                 $insn = $opcode->out($#args>=1?$args[$#args]->size():$sz);
1177                 @args = map($_->out($sz),@args);
1178                 printf "\t%s\t%s",$insn,join(",",@args);
1179             } else {
1180                 $insn = $opcode->out();
1181                 foreach (@args) {
1182                     my $arg = $_->out();
1183                     # $insn.=$sz compensates for movq, pinsrw, ...
1184                     if ($arg =~ /^xmm[0-9]+$/) { $insn.=$sz; $sz="x" if(!$sz); last; }
1185                     if ($arg =~ /^ymm[0-9]+$/) { $insn.=$sz; $sz="y" if(!$sz); last; }
1186                     if ($arg =~ /^zmm[0-9]+$/) { $insn.=$sz; $sz="z" if(!$sz); last; }
1187                     if ($arg =~ /^mm[0-9]+$/)  { $insn.=$sz; $sz="q" if(!$sz); last; }
1188                 }
1189                 @args = reverse(@args);
1190                 undef $sz if ($nasm && $opcode->mnemonic() eq "lea");
1191                 printf "\t%s\t%s",$insn,join(",",map($_->out($sz),@args));
1192             }
1193         } else {
1194             printf "\t%s",$opcode->out();
1195         }
1196     }
1197
1198     print $line,"\n";
1199 }
1200
1201 print "\n$current_segment\tENDS\n"      if ($current_segment && $masm);
1202 print "END\n"                           if ($masm);
1203
1204 close STDOUT;
1205
1206 \f#################################################
1207 # Cross-reference x86_64 ABI "card"
1208 #
1209 #               Unix            Win64
1210 # %rax          *               *
1211 # %rbx          -               -
1212 # %rcx          #4              #1
1213 # %rdx          #3              #2
1214 # %rsi          #2              -
1215 # %rdi          #1              -
1216 # %rbp          -               -
1217 # %rsp          -               -
1218 # %r8           #5              #3
1219 # %r9           #6              #4
1220 # %r10          *               *
1221 # %r11          *               *
1222 # %r12          -               -
1223 # %r13          -               -
1224 # %r14          -               -
1225 # %r15          -               -
1226 #
1227 # (*)   volatile register
1228 # (-)   preserved by callee
1229 # (#)   Nth argument, volatile
1230 #
1231 # In Unix terms top of stack is argument transfer area for arguments
1232 # which could not be accommodated in registers. Or in other words 7th
1233 # [integer] argument resides at 8(%rsp) upon function entry point.
1234 # 128 bytes above %rsp constitute a "red zone" which is not touched
1235 # by signal handlers and can be used as temporal storage without
1236 # allocating a frame.
1237 #
1238 # In Win64 terms N*8 bytes on top of stack is argument transfer area,
1239 # which belongs to/can be overwritten by callee. N is the number of
1240 # arguments passed to callee, *but* not less than 4! This means that
1241 # upon function entry point 5th argument resides at 40(%rsp), as well
1242 # as that 32 bytes from 8(%rsp) can always be used as temporal
1243 # storage [without allocating a frame]. One can actually argue that
1244 # one can assume a "red zone" above stack pointer under Win64 as well.
1245 # Point is that at apparently no occasion Windows kernel would alter
1246 # the area above user stack pointer in true asynchronous manner...
1247 #
1248 # All the above means that if assembler programmer adheres to Unix
1249 # register and stack layout, but disregards the "red zone" existence,
1250 # it's possible to use following prologue and epilogue to "gear" from
1251 # Unix to Win64 ABI in leaf functions with not more than 6 arguments.
1252 #
1253 # omnipotent_function:
1254 # ifdef WIN64
1255 #       movq    %rdi,8(%rsp)
1256 #       movq    %rsi,16(%rsp)
1257 #       movq    %rcx,%rdi       ; if 1st argument is actually present
1258 #       movq    %rdx,%rsi       ; if 2nd argument is actually ...
1259 #       movq    %r8,%rdx        ; if 3rd argument is ...
1260 #       movq    %r9,%rcx        ; if 4th argument ...
1261 #       movq    40(%rsp),%r8    ; if 5th ...
1262 #       movq    48(%rsp),%r9    ; if 6th ...
1263 # endif
1264 #       ...
1265 # ifdef WIN64
1266 #       movq    8(%rsp),%rdi
1267 #       movq    16(%rsp),%rsi
1268 # endif
1269 #       ret
1270 #
1271 \f#################################################
1272 # Win64 SEH, Structured Exception Handling.
1273 #
1274 # Unlike on Unix systems(*) lack of Win64 stack unwinding information
1275 # has undesired side-effect at run-time: if an exception is raised in
1276 # assembler subroutine such as those in question (basically we're
1277 # referring to segmentation violations caused by malformed input
1278 # parameters), the application is briskly terminated without invoking
1279 # any exception handlers, most notably without generating memory dump
1280 # or any user notification whatsoever. This poses a problem. It's
1281 # possible to address it by registering custom language-specific
1282 # handler that would restore processor context to the state at
1283 # subroutine entry point and return "exception is not handled, keep
1284 # unwinding" code. Writing such handler can be a challenge... But it's
1285 # doable, though requires certain coding convention. Consider following
1286 # snippet:
1287 #
1288 # .type function,@function
1289 # function:
1290 #       movq    %rsp,%rax       # copy rsp to volatile register
1291 #       pushq   %r15            # save non-volatile registers
1292 #       pushq   %rbx
1293 #       pushq   %rbp
1294 #       movq    %rsp,%r11
1295 #       subq    %rdi,%r11       # prepare [variable] stack frame
1296 #       andq    $-64,%r11
1297 #       movq    %rax,0(%r11)    # check for exceptions
1298 #       movq    %r11,%rsp       # allocate [variable] stack frame
1299 #       movq    %rax,0(%rsp)    # save original rsp value
1300 # magic_point:
1301 #       ...
1302 #       movq    0(%rsp),%rcx    # pull original rsp value
1303 #       movq    -24(%rcx),%rbp  # restore non-volatile registers
1304 #       movq    -16(%rcx),%rbx
1305 #       movq    -8(%rcx),%r15
1306 #       movq    %rcx,%rsp       # restore original rsp
1307 # magic_epilogue:
1308 #       ret
1309 # .size function,.-function
1310 #
1311 # The key is that up to magic_point copy of original rsp value remains
1312 # in chosen volatile register and no non-volatile register, except for
1313 # rsp, is modified. While past magic_point rsp remains constant till
1314 # the very end of the function. In this case custom language-specific
1315 # exception handler would look like this:
1316 #
1317 # EXCEPTION_DISPOSITION handler (EXCEPTION_RECORD *rec,ULONG64 frame,
1318 #               CONTEXT *context,DISPATCHER_CONTEXT *disp)
1319 # {     ULONG64 *rsp = (ULONG64 *)context->Rax;
1320 #       ULONG64  rip = context->Rip;
1321 #
1322 #       if (rip >= magic_point)
1323 #       {   rsp = (ULONG64 *)context->Rsp;
1324 #           if (rip < magic_epilogue)
1325 #           {   rsp = (ULONG64 *)rsp[0];
1326 #               context->Rbp = rsp[-3];
1327 #               context->Rbx = rsp[-2];
1328 #               context->R15 = rsp[-1];
1329 #           }
1330 #       }
1331 #       context->Rsp = (ULONG64)rsp;
1332 #       context->Rdi = rsp[1];
1333 #       context->Rsi = rsp[2];
1334 #
1335 #       memcpy (disp->ContextRecord,context,sizeof(CONTEXT));
1336 #       RtlVirtualUnwind(UNW_FLAG_NHANDLER,disp->ImageBase,
1337 #               dips->ControlPc,disp->FunctionEntry,disp->ContextRecord,
1338 #               &disp->HandlerData,&disp->EstablisherFrame,NULL);
1339 #       return ExceptionContinueSearch;
1340 # }
1341 #
1342 # It's appropriate to implement this handler in assembler, directly in
1343 # function's module. In order to do that one has to know members'
1344 # offsets in CONTEXT and DISPATCHER_CONTEXT structures and some constant
1345 # values. Here they are:
1346 #
1347 #       CONTEXT.Rax                             120
1348 #       CONTEXT.Rcx                             128
1349 #       CONTEXT.Rdx                             136
1350 #       CONTEXT.Rbx                             144
1351 #       CONTEXT.Rsp                             152
1352 #       CONTEXT.Rbp                             160
1353 #       CONTEXT.Rsi                             168
1354 #       CONTEXT.Rdi                             176
1355 #       CONTEXT.R8                              184
1356 #       CONTEXT.R9                              192
1357 #       CONTEXT.R10                             200
1358 #       CONTEXT.R11                             208
1359 #       CONTEXT.R12                             216
1360 #       CONTEXT.R13                             224
1361 #       CONTEXT.R14                             232
1362 #       CONTEXT.R15                             240
1363 #       CONTEXT.Rip                             248
1364 #       CONTEXT.Xmm6                            512
1365 #       sizeof(CONTEXT)                         1232
1366 #       DISPATCHER_CONTEXT.ControlPc            0
1367 #       DISPATCHER_CONTEXT.ImageBase            8
1368 #       DISPATCHER_CONTEXT.FunctionEntry        16
1369 #       DISPATCHER_CONTEXT.EstablisherFrame     24
1370 #       DISPATCHER_CONTEXT.TargetIp             32
1371 #       DISPATCHER_CONTEXT.ContextRecord        40
1372 #       DISPATCHER_CONTEXT.LanguageHandler      48
1373 #       DISPATCHER_CONTEXT.HandlerData          56
1374 #       UNW_FLAG_NHANDLER                       0
1375 #       ExceptionContinueSearch                 1
1376 #
1377 # In order to tie the handler to the function one has to compose
1378 # couple of structures: one for .xdata segment and one for .pdata.
1379 #
1380 # UNWIND_INFO structure for .xdata segment would be
1381 #
1382 # function_unwind_info:
1383 #       .byte   9,0,0,0
1384 #       .rva    handler
1385 #
1386 # This structure designates exception handler for a function with
1387 # zero-length prologue, no stack frame or frame register.
1388 #
1389 # To facilitate composing of .pdata structures, auto-generated "gear"
1390 # prologue copies rsp value to rax and denotes next instruction with
1391 # .LSEH_begin_{function_name} label. This essentially defines the SEH
1392 # styling rule mentioned in the beginning. Position of this label is
1393 # chosen in such manner that possible exceptions raised in the "gear"
1394 # prologue would be accounted to caller and unwound from latter's frame.
1395 # End of function is marked with respective .LSEH_end_{function_name}
1396 # label. To summarize, .pdata segment would contain
1397 #
1398 #       .rva    .LSEH_begin_function
1399 #       .rva    .LSEH_end_function
1400 #       .rva    function_unwind_info
1401 #
1402 # Reference to function_unwind_info from .xdata segment is the anchor.
1403 # In case you wonder why references are 32-bit .rvas and not 64-bit
1404 # .quads. References put into these two segments are required to be
1405 # *relative* to the base address of the current binary module, a.k.a.
1406 # image base. No Win64 module, be it .exe or .dll, can be larger than
1407 # 2GB and thus such relative references can be and are accommodated in
1408 # 32 bits.
1409 #
1410 # Having reviewed the example function code, one can argue that "movq
1411 # %rsp,%rax" above is redundant. It is not! Keep in mind that on Unix
1412 # rax would contain an undefined value. If this "offends" you, use
1413 # another register and refrain from modifying rax till magic_point is
1414 # reached, i.e. as if it was a non-volatile register. If more registers
1415 # are required prior [variable] frame setup is completed, note that
1416 # nobody says that you can have only one "magic point." You can
1417 # "liberate" non-volatile registers by denoting last stack off-load
1418 # instruction and reflecting it in finer grade unwind logic in handler.
1419 # After all, isn't it why it's called *language-specific* handler...
1420 #
1421 # SE handlers are also involved in unwinding stack when executable is
1422 # profiled or debugged. Profiling implies additional limitations that
1423 # are too subtle to discuss here. For now it's sufficient to say that
1424 # in order to simplify handlers one should either a) offload original
1425 # %rsp to stack (like discussed above); or b) if you have a register to
1426 # spare for frame pointer, choose volatile one.
1427 #
1428 # (*)   Note that we're talking about run-time, not debug-time. Lack of
1429 #       unwind information makes debugging hard on both Windows and
1430 #       Unix. "Unlike" refers to the fact that on Unix signal handler
1431 #       will always be invoked, core dumped and appropriate exit code
1432 #       returned to parent (for user notification).