x86_64-xlate.pl: fix jrcxz in nasm case.
[openssl.git] / crypto / perlasm / x86_64-xlate.pl
index 91017c69f313dadae2cc40b490704ec1d03732d0..bd165b152b833a157221af4df1e47554f80e4cec 100755 (executable)
@@ -62,12 +62,8 @@ 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: $!"
-       if ($stddev!=$outdev || $stdino!=$outino);
-}
+open STDOUT,">$output" || die "can't open $output: $!"
+       if (defined($output));
 
 my $gas=1;     $gas=0 if ($output =~ /\.asm$/);
 my $elf=1;     $elf=0 if (!$gas);
@@ -116,13 +112,15 @@ 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/) {
                $self->{sz} = "";
            } elsif ($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])$/) {
@@ -248,13 +246,19 @@ my %globals;
        $self->{index} =~ s/^[er](.?[0-9xpi])[d]?$/r\1/;
        $self->{base}  =~ s/^[er](.?[0-9xpi])[d]?$/r\1/;
 
+       # 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) {
-           # 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/([0-9]+\s*[\*\/\%]\s*[0-9]+)/eval($1)/eg;
-           $self->{label} =~ s/([0-9]+)/$1<<32>>32/eg;
            $self->{label} =~ s/^___imp_/__imp__/   if ($flavour eq "mingw64");
 
            if (defined($self->{index})) {
@@ -267,12 +271,14 @@ my %globals;
            }
        } else {
            %szmap = (  b=>"BYTE$PTR", w=>"WORD$PTR", l=>"DWORD$PTR",
-                       q=>"QWORD$PTR",o=>"OWORD$PTR",x=>"XMMWORD$PTR" );
+                       q=>"QWORD$PTR",o=>"OWORD$PTR",x=>"XMMWORD$PTR",
+                       y=>"" );
 
            $self->{label} =~ s/\./\$/g;
            $self->{label} =~ s/(?<![\w\$\.])0x([0-9a-f]+)/0$1h/ig;
            $self->{label} = "($self->{label})" if ($self->{label} =~ /[\*\+\-\/]/);
-           $sz="q" if ($self->{asterisk});
+           $sz="q" if ($self->{asterisk} || opcode->mnemonic() =~ /^v?movq$/);
+           $sz="l" if (opcode->mnemonic() =~ /^v?movd$/);
 
            if (defined($self->{index})) {
                sprintf "%s[%s%s*%d%s]",$szmap{$sz},
@@ -412,7 +418,7 @@ my %globals;
     }
     sub out {
        my $self = shift;
-       if ($nasm && opcode->mnemonic()=~m/^j/) {
+       if ($nasm && opcode->mnemonic()=~m/^j(?![re]cxz)/) {
            "NEAR ".$self->{value};
        } else {
            $self->{value};
@@ -515,6 +521,7 @@ my %globals;
                    elsif ($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;
@@ -564,7 +571,8 @@ my %globals;
                                            $v.=" READONLY";
                                            $v.=" ALIGN(".($1 eq "p" ? 4 : 8).")" if ($masm>=$masmref);
                                        } elsif ($line=~/\.CRT\$/i) {
-                                           $v.=" READONLY DWORD";
+                                           $v.=" READONLY ";
+                                           $v.=$masm>=$masmref ? "ALIGN(8)" : "DWORD";
                                        }
                                    }
                                    $current_segment = $line;
@@ -586,7 +594,7 @@ my %globals;
                                            $self->{value}="${decor}SEH_end_$current_function->{name}:";
                                            $self->{value}.=":\n" if($masm);
                                        }
-                                       $self->{value}.="$current_function->{name}\tENDP" if($masm);
+                                       $self->{value}.="$current_function->{name}\tENDP" if($masm && $current_function->{name});
                                        undef $current_function;
                                    }
                                    last;
@@ -625,10 +633,10 @@ my %globals;
                /\.comm/    && do { my @str=split(/,\s*/,$line);
                                    my $v=undef;
                                    if ($nasm) {
-                                       $v.="common     $prefix@str[0] @str[1]:near";
+                                       $v.="common     $prefix@str[0] @str[1]";
                                    } else {
                                        $v="$current_segment\tENDS\n" if ($current_segment);
-                                       $current_segment = ".data";
+                                       $current_segment = "_DATA";
                                        $v.="$current_segment\tSEGMENT\n";
                                        $v.="COMM       @str[0]:DWORD:".@str[1]/4;
                                    }
@@ -663,14 +671,14 @@ my %regrm = (     "%eax"=>0, "%ecx"=>1, "%edx"=>2, "%ebx"=>3,
 my $movq = sub {       # elderly gas can't handle inter-register movq
   my $arg = shift;
   my @opcode=(0x66);
-    if ($arg =~ /%xmm([0-9]+),%r(\w+)/) {
+    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 ($arg =~ /%r(\w+),%xmm([0-9]+)/) {
+    } elsif ($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);
@@ -683,7 +691,7 @@ my $movq = sub {    # elderly gas can't handle inter-register movq
 };
 
 my $pextrd = sub {
-    if (shift =~ /\$([0-9]+),%xmm([0-9]+),(%\w+)/) {
+    if (shift =~ /\$([0-9]+),\s*%xmm([0-9]+),\s*(%\w+)/) {
       my @opcode=(0x66);
        $imm=$1;
        $src=$2;
@@ -701,7 +709,7 @@ my $pextrd = sub {
 };
 
 my $pinsrd = sub {
-    if (shift =~ /\$([0-9]+),(%\w+),%xmm([0-9]+)/) {
+    if (shift =~ /\$([0-9]+),\s*(%\w+),\s*%xmm([0-9]+)/) {
       my @opcode=(0x66);
        $imm=$1;
        $src=$2;
@@ -719,7 +727,7 @@ my $pinsrd = sub {
 };
 
 my $pshufb = sub {
-    if (shift =~ /%xmm([0-9]+),%xmm([0-9]+)/) {
+    if (shift =~ /%xmm([0-9]+),\s*%xmm([0-9]+)/) {
       my @opcode=(0x66);
        rex(\@opcode,$2,$1);
        push @opcode,0x0f,0x38,0x00;
@@ -731,7 +739,7 @@ my $pshufb = sub {
 };
 
 my $palignr = sub {
-    if (shift =~ /\$([0-9]+),%xmm([0-9]+),%xmm([0-9]+)/) {
+    if (shift =~ /\$([0-9]+),\s*%xmm([0-9]+),\s*%xmm([0-9]+)/) {
       my @opcode=(0x66);
        rex(\@opcode,$3,$2);
        push @opcode,0x0f,0x3a,0x0f;
@@ -757,6 +765,58 @@ my $pclmulqdq = sub {
     }
 };
 
+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 {
+       ();
+    }
+};
+
+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 ($nasm) {
     print <<___;
 default        rel
@@ -822,6 +882,7 @@ while($line=<>) {
                    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 ($arg =~ /^mm[0-9]+$/)  { $insn.=$sz; $sz="q" if(!$sz); last; }
                }
                @args = reverse(@args);
@@ -1031,7 +1092,7 @@ close STDOUT;
 #      .rva    .LSEH_end_function
 #      .rva    function_unwind_info
 #
-# Reference to functon_unwind_info from .xdata segment is the anchor.
+# 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.