Undo a90081576c94f9f54de1755188a00ccc1760549a
[openssl.git] / crypto / perlasm / x86_64-xlate.pl
index 10b8629e61472ca7def09ef782cef6f59063741d..84745f600a4363e95bd89503c06fe2e0570003e6 100755 (executable)
 # 6. Don't use [or hand-code with .byte] "rep ret." "ret" mnemonic is
 #    required to identify the spots, where to inject Win64 epilogue!
 #    But on the pros, it's then prefixed with rep automatically:-)
-# 7. Due to MASM limitations [and certain general counter-intuitivity
-#    of ip-relative addressing] generation of position-independent
-#    code is assisted by synthetic directive, .picmeup, which puts
-#    address of the *next* instruction into target register.
+# 7. Stick to explicit ip-relative addressing. If you have to use
+#    GOTPCREL addressing, stick to mov symbol@GOTPCREL(%rip),%r??.
+#    Both are recognized and translated to proper Win64 addressing
+#    modes. To support legacy code a synthetic directive, .picmeup,
+#    is implemented. It puts address of the *next* instruction into
+#    target register, e.g.:
 #
-#    Example 1:
 #              .picmeup        %rax
 #              lea             .Label-.(%rax),%rax
-#    Example 2:
-#              .picmeup        %rcx
-#      .Lpic_point:
-#              ...
-#              lea             .Label-.Lpic_point(%rcx),%rbp
 #
 # 8. In order to provide for structured exception handling unified
 #    Win64 prologue copies %rsp value to %rax. For further details
 #    see SEH paragraph at the end.
+# 9. .init segment is allowed to contain calls to functions only.
+# a. If function accepts more than 4 arguments *and* >4th argument
+#    is declared as non 64-bit value, do clear its upper part.
 \f
-my $output = shift;
+my $flavour = shift;
+my $output  = shift;
+if ($flavour =~ /\./) { $output = $flavour; undef $flavour; }
 
-{ my ($stddev,$stdino,@junk)=stat(STDOUT);
-  my ($outdev,$outino,@junk)=stat($output);
+open STDOUT,">$output" || die "can't open $output: $!"
+       if (defined($output));
 
-    open STDOUT,">$output" || die "can't open $output: $!"
-       if ($stddev!=$outdev || $stdino!=$outino);
-}
-
-my $win64=1 if ($output =~ /\.asm/);
+my $gas=1;     $gas=0 if ($output =~ /\.asm$/);
+my $elf=1;     $elf=0 if (!$gas);
+my $win64=0;
+my $prefix="";
+my $decor=".L";
 
 my $masmref=8 + 50727*2**-32;  # 8.00.50727 shipped with VS2005
 my $masm=0;
@@ -77,12 +78,22 @@ my $PTR=" PTR";
 my $nasmref=2.03;
 my $nasm=0;
 
-if ($win64)
+if    ($flavour eq "mingw64")  { $gas=1; $elf=0; $win64=1;
+                                 $prefix=`echo __USER_LABEL_PREFIX__ | $ENV{CC} -E -P -`;
+                                 chomp($prefix);
+                               }
+elsif ($flavour eq "macosx")   { $gas=1; $elf=0; $prefix="_"; $decor="L\$"; }
+elsif ($flavour eq "masm")     { $gas=0; $elf=0; $masm=$masmref; $win64=1; $decor="\$L\$"; }
+elsif ($flavour eq "nasm")     { $gas=0; $elf=0; $nasm=$nasmref; $win64=1; $decor="\$L\$"; $PTR=""; }
+elsif (!$gas)
 {   if ($ENV{ASM} =~ m/nasm/ && `nasm -v` =~ m/version ([0-9]+)\.([0-9]+)/i)
     {  $nasm = $1 + $2*0.01; $PTR="";  }
     elsif (`ml64 2>&1` =~ m/Version ([0-9]+)\.([0-9]+)(\.([0-9]+))?/)
     {  $masm = $1 + $2*2**-16 + $4*2**-32;   }
     die "no assembler found on %PATH" if (!($nasm || $masm));
+    $win64=1;
+    $elf=0;
+    $decor="\$L\$";
 }
 
 my $current_segment;
@@ -101,11 +112,17 @@ my %globals;
            $line = substr($line,@+[0]); $line =~ s/^\s+//;
 
            undef $self->{sz};
-           if ($self->{op} =~ /^(movz)b.*/) {  # movz is pain...
+           if ($self->{op} =~ /^(movz)x?([bw]).*/) {   # movz is pain...
                $self->{op} = $1;
-               $self->{sz} = "b";
+               $self->{sz} = $2;
            } elsif ($self->{op} =~ /call|jmp/) {
-               $self->{sz} = ""
+               $self->{sz} = "";
+           } elsif ($self->{op} =~ /^p/ && $' !~ /^(ush|op|insrw)/) { # SSEn
+               $self->{sz} = "";
+           } elsif ($self->{op} =~ /^v/) { # VEX
+               $self->{sz} = "";
+           } elsif ($self->{op} =~ /movq/ && $line =~ /%xmm/) {
+               $self->{sz} = "";
            } elsif ($self->{op} =~ /([a-z]{3,})([qlwb])$/) {
                $self->{op} = $1;
                $self->{sz} = $2;
@@ -121,13 +138,20 @@ my %globals;
     }
     sub out {
        my $self = shift;
-       if (!$win64) {
+       if ($gas) {
            if ($self->{op} eq "movz") {        # movz is pain...
                sprintf "%s%s%s",$self->{op},$self->{sz},shift;
            } elsif ($self->{op} =~ /^set/) { 
                "$self->{op}";
            } elsif ($self->{op} eq "ret") {
-               ".byte  0xf3,0xc3";
+               my $epilogue = "";
+               if ($win64 && $current_function->{abi} eq "svr4") {
+                   $epilogue = "movq   8(%rsp),%rdi\n\t" .
+                               "movq   16(%rsp),%rsi\n\t";
+               }
+               $epilogue . ".byte      0xf3,0xc3";
+           } elsif ($self->{op} eq "call" && !$elf && $current_segment eq ".init") {
+               ".p2align\t3\n\t.quad";
            } else {
                "$self->{op}$self->{sz}";
            }
@@ -135,19 +159,25 @@ my %globals;
            $self->{op} =~ s/^movz/movzx/;
            if ($self->{op} eq "ret") {
                $self->{op} = "";
-               if ($current_function->{abi} eq "svr4") {
+               if ($win64 && $current_function->{abi} eq "svr4") {
                    $self->{op} = "mov  rdi,QWORD${PTR}[8+rsp]\t;WIN64 epilogue\n\t".
                                  "mov  rsi,QWORD${PTR}[16+rsp]\n\t";
                }
                $self->{op} .= "DB\t0F3h,0C3h\t\t;repret";
-           }
-           elsif ($self->{op} =~ /^(pop|push)f/) {
+           } elsif ($self->{op} =~ /^(pop|push)f/) {
                $self->{op} .= $self->{sz};
-           }
+           } elsif ($self->{op} eq "call" && $current_segment eq ".CRT\$XCU") {
+               $self->{op} = "\tDQ";
+           } 
            $self->{op};
        }
     }
-    sub mnemonic { shift->{op}; }
+    sub mnemonic {
+       my $self=shift;
+       my $op=shift;
+       $self->{op}=$op if (defined($op));
+       $self->{op};
+    }
 }
 { package const;       # pick up constants, which start with $
     sub re {
@@ -165,14 +195,15 @@ my %globals;
     sub out {
        my $self = shift;
 
-       if (!$win64) {
+       if ($gas) {
            # Solaris /usr/ccs/bin/as can't handle multiplications
            # in $self->{value}
-           $self->{value} =~ s/(?<![0-9a-f])(0[x0-9a-f]+)/oct($1)/egi;
+           $self->{value} =~ s/(?<![\w\$\.])(0x?[0-9a-f]+)/oct($1)/egi;
            $self->{value} =~ s/([0-9]+\s*[\*\/\%]\s*[0-9]+)/eval($1)/eg;
            sprintf "\$%s",$self->{value};
        } else {
-           $self->{value} =~ s/0x([0-9a-f]+)/0$1h/ig;
+           $self->{value} =~ s/(0b[0-1]+)/oct($1)/eig;
+           $self->{value} =~ s/0x([0-9a-f]+)/0$1h/ig if ($masm);
            sprintf "%s",$self->{value};
        }
     }
@@ -192,6 +223,10 @@ my %globals;
            $ret = $self;
            $line = substr($line,@+[0]); $line =~ s/^\s+//;
 
+           if ($win64 && $self->{label} =~ s/\@GOTPCREL//) {
+               die if (opcode->mnemonic() ne "mov");
+               opcode->mnemonic("lea");
+           }
            $self->{base}  =~ s/^%//;
            $self->{index} =~ s/^%// if (defined($self->{index}));
        }
@@ -202,39 +237,59 @@ my %globals;
        my $self = shift;
        my $sz = shift;
 
+       $self->{label} =~ s/([_a-z][_a-z0-9]*)/$globals{$1} or $1/gei;
+       $self->{label} =~ s/\.L/$decor/g;
+
        # Silently convert all EAs to 64-bit. This is required for
        # elder GNU assembler and results in more compact code,
        # *but* most importantly AES module depends on this feature!
        $self->{index} =~ s/^[er](.?[0-9xpi])[d]?$/r\1/;
        $self->{base}  =~ s/^[er](.?[0-9xpi])[d]?$/r\1/;
 
-       if (!$win64) {
-           # Solaris /usr/ccs/bin/as can't handle multiplications
-           # in $self->{label}
-           $self->{label} =~ s/(?<![0-9a-f])(0[x0-9a-f]+)/oct($1)/egi;
-           $self->{label} =~ s/([0-9]+\s*[\*\/\%]\s*[0-9]+)/eval($1)/eg;
+       # Solaris /usr/ccs/bin/as can't handle multiplications
+       # in $self->{label}, new gas requires sign extension...
+       use integer;
+       $self->{label} =~ s/(?<![\w\$\.])(0x?[0-9a-f]+)/oct($1)/egi;
+       $self->{label} =~ s/\b([0-9]+\s*[\*\/\%]\s*[0-9]+)\b/eval($1)/eg;
+       $self->{label} =~ s/\b([0-9]+)\b/$1<<32>>32/eg;
+
+       if (!$self->{label} && $self->{index} && $self->{scale}==1 &&
+           $self->{base} =~ /(rbp|r13)/) {
+               $self->{base} = $self->{index}; $self->{index} = $1;
+       }
+
+       if ($gas) {
+           $self->{label} =~ s/^___imp_/__imp__/   if ($flavour eq "mingw64");
 
            if (defined($self->{index})) {
-               sprintf "%s%s(%%%s,%%%s,%d)",$self->{asterisk},
-                                       $self->{label},$self->{base},
+               sprintf "%s%s(%s,%%%s,%d)",$self->{asterisk},
+                                       $self->{label},
+                                       $self->{base}?"%$self->{base}":"",
                                        $self->{index},$self->{scale};
            } else {
                sprintf "%s%s(%%%s)",   $self->{asterisk},$self->{label},$self->{base};
            }
        } else {
-           %szmap = ( b=>"BYTE$PTR", w=>"WORD$PTR", l=>"DWORD$PTR", q=>"QWORD$PTR" );
+           %szmap = (  b=>"BYTE$PTR",  w=>"WORD$PTR",
+                       l=>"DWORD$PTR", d=>"DWORD$PTR",
+                       q=>"QWORD$PTR", o=>"OWORD$PTR",
+                       x=>"XMMWORD$PTR", y=>"YMMWORD$PTR", z=>"ZMMWORD$PTR" );
 
-           $self->{label} =~ s/\.L/\$L\$/g;
            $self->{label} =~ s/\./\$/g;
-           $self->{label} =~ s/0x([0-9a-f]+)/0$1h/ig;
+           $self->{label} =~ s/(?<![\w\$\.])0x([0-9a-f]+)/0$1h/ig;
            $self->{label} = "($self->{label})" if ($self->{label} =~ /[\*\+\-\/]/);
-           $sz="q" if ($self->{asterisk});
+
+           ($self->{asterisk})                                 && ($sz="q") ||
+           (opcode->mnemonic() =~ /^v?mov([qd])$/)             && ($sz=$1)  ||
+           (opcode->mnemonic() =~ /^v?pinsr([qdwb])$/)         && ($sz=$1)  ||
+           (opcode->mnemonic() =~ /^vpbroadcast([qdwb])$/)     && ($sz=$1)  ||
+           (opcode->mnemonic() =~ /^vinsert[fi]128$/)          && ($sz="x");
 
            if (defined($self->{index})) {
-               sprintf "%s[%s%s*%d+%s]",$szmap{$sz},
+               sprintf "%s[%s%s*%d%s]",$szmap{$sz},
                                        $self->{label}?"$self->{label}+":"",
                                        $self->{index},$self->{scale},
-                                       $self->{base};
+                                       $self->{base}?"+$self->{base}":"";
            } elsif ($self->{base} eq "rip") {
                sprintf "%s[%s]",$szmap{$sz},$self->{label};
            } else {
@@ -279,7 +334,7 @@ my %globals;
     }
     sub out {
        my $self = shift;
-       if (!$win64)    { sprintf "%s%%%s",$self->{asterisk},$self->{value}; }
+       if ($gas)       { sprintf "%s%%%s",$self->{asterisk},$self->{value}; }
        else            { $self->{value}; }
     }
 }
@@ -289,31 +344,49 @@ my %globals;
        local   *line = shift;
        undef   $ret;
 
-       if ($line =~ /(^[\.\w]+\:)/) {
+       if ($line =~ /(^[\.\w]+)\:/) {
            $self->{value} = $1;
            $ret = $self;
            $line = substr($line,@+[0]); $line =~ s/^\s+//;
 
-           $self->{value} =~ s/\.L/\$L\$/ if ($win64);
+           $self->{value} =~ s/^\.L/$decor/;
        }
        $ret;
     }
     sub out {
        my $self = shift;
 
-       if (!$win64) {
-           $self->{value};
-       } elsif ($self->{value} ne "$current_function->{name}:") {
+       if ($gas) {
+           my $func = ($globals{$self->{value}} or $self->{value}) . ":";
+           if ($win64  &&
+                       $current_function->{name} eq $self->{value} &&
+                       $current_function->{abi} eq "svr4") {
+               $func .= "\n";
+               $func .= "      movq    %rdi,8(%rsp)\n";
+               $func .= "      movq    %rsi,16(%rsp)\n";
+               $func .= "      movq    %rsp,%rax\n";
+               $func .= "${decor}SEH_begin_$current_function->{name}:\n";
+               my $narg = $current_function->{narg};
+               $narg=6 if (!defined($narg));
+               $func .= "      movq    %rcx,%rdi\n" if ($narg>0);
+               $func .= "      movq    %rdx,%rsi\n" if ($narg>1);
+               $func .= "      movq    %r8,%rdx\n"  if ($narg>2);
+               $func .= "      movq    %r9,%rcx\n"  if ($narg>3);
+               $func .= "      movq    40(%rsp),%r8\n" if ($narg>4);
+               $func .= "      movq    48(%rsp),%r9\n" if ($narg>5);
+           }
+           $func;
+       } elsif ($self->{value} ne "$current_function->{name}") {
            $self->{value} .= ":" if ($masm && $ret!~m/^\$/);
-           $self->{value};
-       } elsif ($current_function->{abi} eq "svr4") {
+           $self->{value} . ":";
+       } elsif ($win64 && $current_function->{abi} eq "svr4") {
            my $func =  "$current_function->{name}" .
                        ($nasm ? ":" : "\tPROC $current_function->{scope}") .
                        "\n";
            $func .= "  mov     QWORD${PTR}[8+rsp],rdi\t;WIN64 prologue\n";
            $func .= "  mov     QWORD${PTR}[16+rsp],rsi\n";
            $func .= "  mov     rax,rsp\n";
-           $func .= "\$L\$SEH_begin_$current_function->{name}:";
+           $func .= "${decor}SEH_begin_$current_function->{name}:";
            $func .= ":" if ($masm);
            $func .= "\n";
            my $narg = $current_function->{narg};
@@ -342,13 +415,15 @@ my %globals;
            $ret = $self;
            $line = substr($line,@+[0]); $line =~ s/^\s+//;
 
-           $self->{value} =~ s/\.L/\$L\$/g if ($win64);
+           $self->{value} =~ s/\@PLT// if (!$elf);
+           $self->{value} =~ s/([_a-z][_a-z0-9]*)/$globals{$1} or $1/gei;
+           $self->{value} =~ s/\.L/$decor/g;
        }
        $ret;
     }
     sub out {
        my $self = shift;
-       if ($nasm && opcode->mnemonic()=~m/^j/) {
+       if ($nasm && opcode->mnemonic()=~m/^j(?![re]cxz)/) {
            "NEAR ".$self->{value};
        } else {
            $self->{value};
@@ -372,27 +447,92 @@ my %globals;
                        "%r14"=>0x01358d4c,     "%r15"=>0x013d8d4c      );
 
        if ($line =~ /^\s*(\.\w+)/) {
-           if (!$win64) {
-               $self->{value} = $1;
-               $line =~ s/\@abi\-omnipotent/\@function/;
-               $line =~ s/\@function.*/\@function/;
-               if ($line =~ /\.picmeup\s+(%r[\w]+)/i) {
-                   $self->{value} = sprintf "\t.long\t0x%x,0x90000000",$opcode{$1};
-               } elsif ($line =~ /\.asciz\s+"(.*)"$/) {
-                   $self->{value} = ".byte\t".join(",",unpack("C*",$1),0);
-               } elsif ($line =~ /\.extern/) {
+           $dir = $1;
+           $ret = $self;
+           undef $self->{value};
+           $line = substr($line,@+[0]); $line =~ s/^\s+//;
+
+           SWITCH: for ($dir) {
+               /\.picmeup/ && do { if ($line =~ /(%r[\w]+)/i) {
+                                       $dir="\t.long";
+                                       $line=sprintf "0x%x,0x90000000",$opcode{$1};
+                                   }
+                                   last;
+                                 };
+               /\.global|\.globl|\.extern/
+                           && do { $globals{$line} = $prefix . $line;
+                                   $line = $globals{$line} if ($prefix);
+                                   last;
+                                 };
+               /\.type/    && do { ($sym,$type,$narg) = split(',',$line);
+                                   if ($type eq "\@function") {
+                                       undef $current_function;
+                                       $current_function->{name} = $sym;
+                                       $current_function->{abi}  = "svr4";
+                                       $current_function->{narg} = $narg;
+                                       $current_function->{scope} = defined($globals{$sym})?"PUBLIC":"PRIVATE";
+                                   } elsif ($type eq "\@abi-omnipotent") {
+                                       undef $current_function;
+                                       $current_function->{name} = $sym;
+                                       $current_function->{scope} = defined($globals{$sym})?"PUBLIC":"PRIVATE";
+                                   }
+                                   $line =~ s/\@abi\-omnipotent/\@function/;
+                                   $line =~ s/\@function.*/\@function/;
+                                   last;
+                                 };
+               /\.asciz/   && do { if ($line =~ /^"(.*)"$/) {
+                                       $dir  = ".byte";
+                                       $line = join(",",unpack("C*",$1),0);
+                                   }
+                                   last;
+                                 };
+               /\.rva|\.long|\.quad/
+                           && do { $line =~ s/([_a-z][_a-z0-9]*)/$globals{$1} or $1/gei;
+                                   $line =~ s/\.L/$decor/g;
+                                   last;
+                                 };
+           }
+
+           if ($gas) {
+               $self->{value} = $dir . "\t" . $line;
+
+               if ($dir =~ /\.extern/) {
                    $self->{value} = ""; # swallow extern
-               } else {
-                   $self->{value} = $line;
+               } elsif (!$elf && $dir =~ /\.type/) {
+                   $self->{value} = "";
+                   $self->{value} = ".def\t" . ($globals{$1} or $1) . ";\t" .
+                               (defined($globals{$1})?".scl 2;":".scl 3;") .
+                               "\t.type 32;\t.endef"
+                               if ($win64 && $line =~ /([^,]+),\@function/);
+               } elsif (!$elf && $dir =~ /\.size/) {
+                   $self->{value} = "";
+                   if (defined($current_function)) {
+                       $self->{value} .= "${decor}SEH_end_$current_function->{name}:"
+                               if ($win64 && $current_function->{abi} eq "svr4");
+                       undef $current_function;
+                   }
+               } elsif (!$elf && $dir =~ /\.align/) {
+                   $self->{value} = ".p2align\t" . (log($line)/log(2));
+               } elsif ($dir eq ".section") {
+                   $current_segment=$line;
+                   if (!$elf && $current_segment eq ".init") {
+                       if      ($flavour eq "macosx")  { $self->{value} = ".mod_init_func"; }
+                       elsif   ($flavour eq "mingw64") { $self->{value} = ".section\t.ctors"; }
+                   }
+               } elsif ($dir =~ /\.(text|data)/) {
+                   $current_segment=".$1";
+               } elsif ($dir =~ /\.hidden/) {
+                   if    ($flavour eq "macosx")  { $self->{value} = ".private_extern\t$prefix$line"; }
+                   elsif ($flavour eq "mingw64") { $self->{value} = ""; }
+               } elsif ($dir =~ /\.comm/) {
+                   $self->{value} = "$dir\t$prefix$line";
+                   $self->{value} =~ s|,([0-9]+),([0-9]+)$|",$1,".log($2)/log(2)|e if ($flavour eq "macosx");
                }
                $line = "";
                return $self;
            }
 
-           $dir = $1;
-           $ret = $self;
-           undef $self->{value};
-           $line = substr($line,@+[0]); $line =~ s/^\s+//;
+           # non-gas case or nasm/masm
            SWITCH: for ($dir) {
                /\.text/    && do { my $v=undef;
                                    if ($nasm) {
@@ -401,7 +541,7 @@ my %globals;
                                        $v="$current_segment\tENDS\n" if ($current_segment);
                                        $current_segment = ".text\$";
                                        $v.="$current_segment\tSEGMENT ";
-                                       $v.=$masm>=$masmref ? "ALIGN(64)" : "PAGE";
+                                       $v.=$masm>=$masmref ? "ALIGN(256)" : "PAGE";
                                        $v.=" 'CODE'";
                                    }
                                    $self->{value} = $v;
@@ -419,11 +559,15 @@ my %globals;
                                    last;
                                  };
                /\.section/ && do { my $v=undef;
+                                   $line =~ s/([^,]*).*/$1/;
+                                   $line = ".CRT\$XCU" if ($line eq ".init");
                                    if ($nasm) {
                                        $v="section     $line";
                                        if ($line=~/\.([px])data/) {
                                            $v.=" rdata align=";
                                            $v.=$1 eq "p"? 4 : 8;
+                                       } elsif ($line=~/\.CRT\$/i) {
+                                           $v.=" rdata align=8";
                                        }
                                    } else {
                                        $v="$current_segment\tENDS\n" if ($current_segment);
@@ -431,6 +575,9 @@ my %globals;
                                        if ($line=~/\.([px])data/) {
                                            $v.=" READONLY";
                                            $v.=" ALIGN(".($1 eq "p" ? 4 : 8).")" if ($masm>=$masmref);
+                                       } elsif ($line=~/\.CRT\$/i) {
+                                           $v.=" READONLY ";
+                                           $v.=$masm>=$masmref ? "ALIGN(8)" : "DWORD";
                                        }
                                    }
                                    $current_segment = $line;
@@ -444,63 +591,61 @@ my %globals;
                /\.globl|.global/
                            && do { $self->{value}  = $masm?"PUBLIC":"global";
                                    $self->{value} .= "\t".$line;
-                                   $globals{$line} = $line;
-                                   last;
-                                 };
-               /\.type/    && do { ($sym,$type,$narg) = split(',',$line);
-                                   if ($type eq "\@function") {
-                                       undef $current_function;
-                                       $current_function->{name} = $sym;
-                                       $current_function->{abi}  = "svr4";
-                                       $current_function->{narg} = $narg;
-                                       $current_function->{scope} = defined($globals{$sym})?"PUBLIC":"PRIVATE";
-                                   } elsif ($type eq "\@abi-omnipotent") {
-                                       undef $current_function;
-                                       $current_function->{name} = $sym;
-                                       $current_function->{scope} = defined($globals{$sym})?"PUBLIC":"PRIVATE";
-                                   }
                                    last;
                                  };
                /\.size/    && do { if (defined($current_function)) {
-                                       $self->{value}="\$L\$SEH_end_$current_function->{name}:";
-                                       $self->{value}.=":\n$current_function->{name}\tENDP" if($masm);
+                                       undef $self->{value};
+                                       if ($current_function->{abi} eq "svr4") {
+                                           $self->{value}="${decor}SEH_end_$current_function->{name}:";
+                                           $self->{value}.=":\n" if($masm);
+                                       }
+                                       $self->{value}.="$current_function->{name}\tENDP" if($masm && $current_function->{name});
                                        undef $current_function;
                                    }
                                    last;
                                  };
                /\.align/   && do { $self->{value} = "ALIGN\t".$line; last; };
-               /\.(byte|value|long|quad)/
+               /\.(value|long|rva|quad)/
                            && do { my $sz  = substr($1,0,1);
-                                   my @arr = split(',',$line);
+                                   my @arr = split(/,\s*/,$line);
                                    my $last = pop(@arr);
                                    my $conv = sub  {   my $var=shift;
-                                                       $var=~s/0x([0-9a-f]+)/0$1h/ig;
-                                                       $var=~s/\.L/\$L\$/g;
-                                                       if ($current_segment=~/.[px]data/)
+                                                       $var=~s/^(0b[0-1]+)/oct($1)/eig;
+                                                       $var=~s/^0x([0-9a-f]+)/0$1h/ig if ($masm);
+                                                       if ($sz eq "D" && ($current_segment=~/.[px]data/ || $dir eq ".rva"))
                                                        { $var=~s/([_a-z\$\@][_a-z0-9\$\@]*)/$nasm?"$1 wrt ..imagebase":"imagerel $1"/egi; }
                                                        $var;
                                                    };  
 
-                                   $sz =~ tr/bvlq/BWDQ/;
+                                   $sz =~ tr/bvlrq/BWDDQ/;
                                    $self->{value} = "\tD$sz\t";
                                    for (@arr) { $self->{value} .= &$conv($_).","; }
                                    $self->{value} .= &$conv($last);
                                    last;
                                  };
-               /\.picmeup/ && do { $self->{value} = sprintf"\tDD\t0%Xh,090000000h",$opcode{$line};
-                                   last;
-                                 };
-               /\.asciz/   && do { if ($line =~ /^"(.*)"$/) {
-                                       my @str=unpack("C*",$1);
-                                       push @str,0;
-                                       while ($#str>15) {
-                                           $self->{value}.="DB\t"
-                                               .join(",",@str[0..15])."\n";
-                                           foreach (0..15) { shift @str; }
-                                       }
+               /\.byte/    && do { my @str=split(/,\s*/,$line);
+                                   map(s/(0b[0-1]+)/oct($1)/eig,@str);
+                                   map(s/0x([0-9a-f]+)/0$1h/ig,@str) if ($masm);       
+                                   while ($#str>15) {
                                        $self->{value}.="DB\t"
+                                               .join(",",@str[0..15])."\n";
+                                       foreach (0..15) { shift @str; }
+                                   }
+                                   $self->{value}.="DB\t"
                                                .join(",",@str) if (@str);
+                                   last;
+                                 };
+               /\.comm/    && do { my @str=split(/,\s*/,$line);
+                                   my $v=undef;
+                                   if ($nasm) {
+                                       $v.="common     $prefix@str[0] @str[1]";
+                                   } else {
+                                       $v="$current_segment\tENDS\n" if ($current_segment);
+                                       $current_segment = "_DATA";
+                                       $v.="$current_segment\tSEGMENT\n";
+                                       $v.="COMM       @str[0]:DWORD:".@str[1]/4;
                                    }
+                                   $self->{value} = $v;
                                    last;
                                  };
            }
@@ -515,9 +660,187 @@ my %globals;
     }
 }
 
+sub rex {
+ local *opcode=shift;
+ my ($dst,$src,$rex)=@_;
+
+   $rex|=0x04 if($dst>=8);
+   $rex|=0x01 if($src>=8);
+   push @opcode,($rex|0x40) if ($rex);
+}
+
+# older gas and ml64 don't handle SSE>2 instructions
+my %regrm = (  "%eax"=>0, "%ecx"=>1, "%edx"=>2, "%ebx"=>3,
+               "%esp"=>4, "%ebp"=>5, "%esi"=>6, "%edi"=>7      );
+
+my $movq = sub {       # elderly gas can't handle inter-register movq
+  my $arg = shift;
+  my @opcode=(0x66);
+    if ($arg =~ /%xmm([0-9]+),\s*%r(\w+)/) {
+       my ($src,$dst)=($1,$2);
+       if ($dst !~ /[0-9]+/)   { $dst = $regrm{"%e$dst"}; }
+       rex(\@opcode,$src,$dst,0x8);
+       push @opcode,0x0f,0x7e;
+       push @opcode,0xc0|(($src&7)<<3)|($dst&7);       # ModR/M
+       @opcode;
+    } elsif ($arg =~ /%r(\w+),\s*%xmm([0-9]+)/) {
+       my ($src,$dst)=($2,$1);
+       if ($dst !~ /[0-9]+/)   { $dst = $regrm{"%e$dst"}; }
+       rex(\@opcode,$src,$dst,0x8);
+       push @opcode,0x0f,0x6e;
+       push @opcode,0xc0|(($src&7)<<3)|($dst&7);       # ModR/M
+       @opcode;
+    } else {
+       ();
+    }
+};
+
+my $pextrd = sub {
+    if (shift =~ /\$([0-9]+),\s*%xmm([0-9]+),\s*(%\w+)/) {
+      my @opcode=(0x66);
+       $imm=$1;
+       $src=$2;
+       $dst=$3;
+       if ($dst =~ /%r([0-9]+)d/)      { $dst = $1; }
+       elsif ($dst =~ /%e/)            { $dst = $regrm{$dst}; }
+       rex(\@opcode,$src,$dst);
+       push @opcode,0x0f,0x3a,0x16;
+       push @opcode,0xc0|(($src&7)<<3)|($dst&7);       # ModR/M
+       push @opcode,$imm;
+       @opcode;
+    } else {
+       ();
+    }
+};
+
+my $pinsrd = sub {
+    if (shift =~ /\$([0-9]+),\s*(%\w+),\s*%xmm([0-9]+)/) {
+      my @opcode=(0x66);
+       $imm=$1;
+       $src=$2;
+       $dst=$3;
+       if ($src =~ /%r([0-9]+)/)       { $src = $1; }
+       elsif ($src =~ /%e/)            { $src = $regrm{$src}; }
+       rex(\@opcode,$dst,$src);
+       push @opcode,0x0f,0x3a,0x22;
+       push @opcode,0xc0|(($dst&7)<<3)|($src&7);       # ModR/M
+       push @opcode,$imm;
+       @opcode;
+    } else {
+       ();
+    }
+};
+
+my $pshufb = sub {
+    if (shift =~ /%xmm([0-9]+),\s*%xmm([0-9]+)/) {
+      my @opcode=(0x66);
+       rex(\@opcode,$2,$1);
+       push @opcode,0x0f,0x38,0x00;
+       push @opcode,0xc0|($1&7)|(($2&7)<<3);           # ModR/M
+       @opcode;
+    } else {
+       ();
+    }
+};
+
+my $palignr = sub {
+    if (shift =~ /\$([0-9]+),\s*%xmm([0-9]+),\s*%xmm([0-9]+)/) {
+      my @opcode=(0x66);
+       rex(\@opcode,$3,$2);
+       push @opcode,0x0f,0x3a,0x0f;
+       push @opcode,0xc0|($2&7)|(($3&7)<<3);           # ModR/M
+       push @opcode,$1;
+       @opcode;
+    } else {
+       ();
+    }
+};
+
+my $pclmulqdq = sub {
+    if (shift =~ /\$([x0-9a-f]+),\s*%xmm([0-9]+),\s*%xmm([0-9]+)/) {
+      my @opcode=(0x66);
+       rex(\@opcode,$3,$2);
+       push @opcode,0x0f,0x3a,0x44;
+       push @opcode,0xc0|($2&7)|(($3&7)<<3);           # ModR/M
+       my $c=$1;
+       push @opcode,$c=~/^0/?oct($c):$c;
+       @opcode;
+    } else {
+       ();
+    }
+};
+
+my $rdrand = sub {
+    if (shift =~ /%[er](\w+)/) {
+      my @opcode=();
+      my $dst=$1;
+       if ($dst !~ /[0-9]+/) { $dst = $regrm{"%e$dst"}; }
+       rex(\@opcode,0,$1,8);
+       push @opcode,0x0f,0xc7,0xf0|($dst&7);
+       @opcode;
+    } else {
+       ();
+    }
+};
+
+my $rdseed = sub {
+    if (shift =~ /%[er](\w+)/) {
+      my @opcode=();
+      my $dst=$1;
+       if ($dst !~ /[0-9]+/) { $dst = $regrm{"%e$dst"}; }
+       rex(\@opcode,0,$1,8);
+       push @opcode,0x0f,0xc7,0xf8|($dst&7);
+       @opcode;
+    } else {
+       ();
+    }
+};
+
+sub rxb {
+ local *opcode=shift;
+ my ($dst,$src1,$src2,$rxb)=@_;
+
+   $rxb|=0x7<<5;
+   $rxb&=~(0x04<<5) if($dst>=8);
+   $rxb&=~(0x01<<5) if($src1>=8);
+   $rxb&=~(0x02<<5) if($src2>=8);
+   push @opcode,$rxb;
+}
+
+my $vprotd = sub {
+    if (shift =~ /\$([x0-9a-f]+),\s*%xmm([0-9]+),\s*%xmm([0-9]+)/) {
+      my @opcode=(0x8f);
+       rxb(\@opcode,$3,$2,-1,0x08);
+       push @opcode,0x78,0xc2;
+       push @opcode,0xc0|($2&7)|(($3&7)<<3);           # ModR/M
+       my $c=$1;
+       push @opcode,$c=~/^0/?oct($c):$c;
+       @opcode;
+    } else {
+       ();
+    }
+};
+
+my $vprotq = sub {
+    if (shift =~ /\$([x0-9a-f]+),\s*%xmm([0-9]+),\s*%xmm([0-9]+)/) {
+      my @opcode=(0x8f);
+       rxb(\@opcode,$3,$2,-1,0x08);
+       push @opcode,0x78,0xc3;
+       push @opcode,0xc0|($2&7)|(($3&7)<<3);           # ModR/M
+       my $c=$1;
+       push @opcode,$c=~/^0/?oct($c):$c;
+       @opcode;
+    } else {
+       ();
+    }
+};
+
 if ($nasm) {
     print <<___;
 default        rel
+%define XMMWORD
+%define YMMWORD
+%define ZMMWORD
 ___
 } elsif ($masm) {
     print <<___;
@@ -531,47 +854,63 @@ while($line=<>) {
     $line =~ s|[#!].*$||;      # get rid of asm-style comments...
     $line =~ s|/\*.*\*/||;     # ... and C-style comments...
     $line =~ s|^\s+||;         # ... and skip white spaces in beginning
+    $line =~ s|\s+$||;         # ... and at the end
 
     undef $label;
     undef $opcode;
-    undef $dst;
-    undef $src;
-    undef $sz;
+    undef @args;
 
     if ($label=label->re(\$line))      { print $label->out(); }
 
     if (directive->re(\$line)) {
        printf "%s",directive->out();
-    } elsif ($opcode=opcode->re(\$line)) { ARGUMENT: {
+    } elsif ($opcode=opcode->re(\$line)) {
+       my $asm = eval("\$".$opcode->mnemonic());
+       undef @bytes;
+       
+       if ((ref($asm) eq 'CODE') && scalar(@bytes=&$asm($line))) {
+           print $gas?".byte\t":"DB\t",join(',',@bytes),"\n";
+           next;
+       }
 
-       if ($src=register->re(\$line))  { opcode->size($src->size()); }
-       elsif ($src=const->re(\$line))  { }
-       elsif ($src=ea->re(\$line))     { }
-       elsif ($src=expr->re(\$line))   { }
+       ARGUMENT: while (1) {
+       my $arg;
 
-       last ARGUMENT if ($line !~ /^,/);
+       if ($arg=register->re(\$line))  { opcode->size($arg->size()); }
+       elsif ($arg=const->re(\$line))  { }
+       elsif ($arg=ea->re(\$line))     { }
+       elsif ($arg=expr->re(\$line))   { }
+       else                            { last ARGUMENT; }
 
-       $line = substr($line,1); $line =~ s/^\s+//;
+       push @args,$arg;
 
-       if ($dst=register->re(\$line))  { opcode->size($dst->size()); }
-       elsif ($dst=const->re(\$line))  { }
-       elsif ($dst=ea->re(\$line))     { }
+       last ARGUMENT if ($line !~ /^,/);
 
+       $line =~ s/^,\s*//;
        } # ARGUMENT:
 
-       $sz=opcode->size();
+       if ($#args>=0) {
+           my $insn;
+           my $sz=opcode->size();
 
-       if (defined($dst)) {
-           if (!$win64) {
-               printf "\t%s\t%s,%s",   $opcode->out($dst->size()),
-                                       $src->out($sz),$dst->out($sz);
+           if ($gas) {
+               $insn = $opcode->out($#args>=1?$args[$#args]->size():$sz);
+               @args = map($_->out($sz),@args);
+               printf "\t%s\t%s",$insn,join(",",@args);
            } else {
+               $insn = $opcode->out();
+               foreach (@args) {
+                   my $arg = $_->out();
+                   # $insn.=$sz compensates for movq, pinsrw, ...
+                   if ($arg =~ /^xmm[0-9]+$/) { $insn.=$sz; $sz="x" if(!$sz); last; }
+                   if ($arg =~ /^ymm[0-9]+$/) { $insn.=$sz; $sz="y" if(!$sz); last; }
+                   if ($arg =~ /^zmm[0-9]+$/) { $insn.=$sz; $sz="z" if(!$sz); last; }
+                   if ($arg =~ /^mm[0-9]+$/)  { $insn.=$sz; $sz="q" if(!$sz); last; }
+               }
+               @args = reverse(@args);
                undef $sz if ($nasm && $opcode->mnemonic() eq "lea");
-               printf "\t%s\t%s,%s",   $opcode->out(),
-                                       $dst->out($sz),$src->out($sz);
+               printf "\t%s\t%s",$insn,join(",",map($_->out($sz),@args));
            }
-       } elsif (defined($src)) {
-           printf "\t%s\t%s",$opcode->out(),$src->out($sz);
        } else {
            printf "\t%s",$opcode->out();
        }
@@ -580,7 +919,8 @@ while($line=<>) {
     print $line,"\n";
 }
 
-print "\n$current_segment\tENDS\nEND\n" if ($current_segment && $masm);
+print "\n$current_segment\tENDS\n"     if ($current_segment && $masm);
+print "END\n"                          if ($masm);
 
 close STDOUT;
 
@@ -736,6 +1076,7 @@ close STDOUT;
 #      CONTEXT.R14                             232
 #      CONTEXT.R15                             240
 #      CONTEXT.Rip                             248
+#      CONTEXT.Xmm6                            512
 #      sizeof(CONTEXT)                         1232
 #      DISPATCHER_CONTEXT.ControlPc            0
 #      DISPATCHER_CONTEXT.ImageBase            8
@@ -755,26 +1096,26 @@ close STDOUT;
 #
 # function_unwind_info:
 #      .byte   9,0,0,0
-#      .long   handler
+#      .rva    handler
 #
 # This structure designates exception handler for a function with
 # zero-length prologue, no stack frame or frame register.
 #
 # To facilitate composing of .pdata structures, auto-generated "gear"
 # prologue copies rsp value to rax and denotes next instruction with
-# $L$SEH_begin_{function_name} label. This essentially defines the SEH
+# .LSEH_begin_{function_name} label. This essentially defines the SEH
 # styling rule mentioned in the beginning. Position of this label is
 # chosen in such manner that possible exceptions raised in the "gear"
 # prologue would be accounted to caller and unwound from latter's frame.
-# End of function is marked with respective $L$SEH_end_{function_name}
+# End of function is marked with respective .LSEH_end_{function_name}
 # label. To summarize, .pdata segment would contain
 #
-#      .long   $L$SEH_begin_function
-#      .long   $L$SEH_end_function
-#      .long   function_unwind_info
+#      .rva    .LSEH_begin_function
+#      .rva    .LSEH_end_function
+#      .rva    function_unwind_info
 #
-# Reference to functon_unwind_info from .xdata segment is the anchor.
-# In case you wonder why references are 32-bit .longs and not 64-bit
+# Reference to function_unwind_info from .xdata segment is the anchor.
+# In case you wonder why references are 32-bit .rvas and not 64-bit
 # .quads. References put into these two segments are required to be
 # *relative* to the base address of the current binary module, a.k.a.
 # image base. No Win64 module, be it .exe or .dll, can be larger than