Allow for 64-bit cache-line alignments in code segment.
[openssl.git] / crypto / perlasm / x86_64-xlate.pl
index 33ffc41..88a45b1 100755 (executable)
 #    stack frame allocation. If volatile storage is actually required
 #    that is. If not, just leave the stack as is.
 # 3. Functions tagged with ".type name,@function" get crafted with
-#    unified Windows prologue and epilogue automatically. If you want
+#    unified Win64 prologue and epilogue automatically. If you want
 #    to take care of ABI differences yourself, tag functions as
-#    ".type name,@abi-omnipotent."
-# 4. As minor optimization you can specify number of input arguments
-#    as ".type name,@function,N." Keep in mind that if N is larger
-#    than 6, then you *have to* write "abi-omnipotent" code, because
-#    it can't be resolved with unified prologue.
-# 5. Name local labels as .L*.
-# 6. Don't use repret, it's generated automatically.
+#    ".type name,@abi-omnipotent" instead.
+# 4. To optimize the Win64 prologue you can specify number of input
+#    arguments as ".type name,@function,N." Keep in mind that if N is
+#    larger than 6, then you *have to* write "abi-omnipotent" code,
+#    because >6 cases can't be addressed with unified prologue.
+# 5. Name local labels as .L*, do *not* use dynamic labels such as 1:
+#    (sorry about latter).
+# 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.
+#
+#    Example 1:
+#              .picmeup        %rax
+#              lea             .Label-.(%rax),%rax
+#    Example 2:
+#              .picmeup        %rcx
+#      .Lpic_point:
+#              ...
+#              lea             .Label-.Lpic_point(%rcx),%rbp
 
 my $output = shift;
 open STDOUT,">$output" || die "can't open $output: $!";
@@ -115,7 +131,13 @@ my $current_function;
     }
     sub out {
        my $self = shift;
-       sprintf $masm?"%s":"\$%s",$self->{value};
+
+       if (!$masm) {
+           sprintf "\$%s",$self->{value};
+       } else {
+           $self->{value} =~ s/0x([0-9a-f]+)/0$1h/ig;
+           sprintf "%s",$self->{value};
+       }
     }
 }
 { package ea;          # pick up effective addresses: expr(%reg,%reg,scale)
@@ -131,7 +153,6 @@ my $current_function;
            $ret = $self;
            $line = substr($line,@+[0]); $line =~ s/^\s+//;
 
-           $self->{label} =~ s/\.L/\$L/g;
            $self->{base}  =~ s/^%//;
            $self->{index} =~ s/^%// if (defined($self->{index}));
        }
@@ -143,23 +164,31 @@ my $current_function;
        my $sz = shift;
 
        if (!$masm) {
+           # elder GNU assembler insists on 64-bit EAs:-(
+           # on pros side, this results in more compact code:-)
+           $self->{index} =~ s/^[er](.?[0-9xp])[d]?$/r\1/;
+           $self->{base}  =~ s/^[er](.?[0-9xp])[d]?$/r\1/;
+
            if (defined($self->{index})) {
-               sprintf "%s(%%%s,%%%s,%d)",     $self->{label},$self->{base},
+               sprintf "%s(%%%s,%%%s,%d)",
+                                       $self->{label},$self->{base},
                                        $self->{index},$self->{scale};
-           }
-           else {
+           } else {
                sprintf "%s(%%%s)",     $self->{label},$self->{base};
            }
        } else {
            %szmap = ( b=>"BYTE", w=>"WORD", l=>"DWORD", q=>"QWORD" );
 
+           $self->{label} =~ s/\./\$/g;
+           $self->{label} =~ s/0x([0-9a-f]+)/0$1h/ig;
+           $self->{label} = "($self->{label})" if ($self->{label} =~ /[\*\+\-\/]/);
+
            if (defined($self->{index})) {
                sprintf "%s PTR %s[%s*%d+%s]",$szmap{$sz},
                                        $self->{label},
                                        $self->{index},$self->{scale},
                                        $self->{base};
-           }
-           else {
+           } else {
                sprintf "%s PTR %s[%s]",$szmap{$sz},
                                        $self->{label},$self->{base};
            }
@@ -267,13 +296,26 @@ my $current_function;
        local   *line = shift;
        undef   $ret;
        my      $dir;
+       my      %opcode =       # lea 2f-1f(%rip),%dst; 1: nop; 2:
+               (       "%rax"=>0x01058d48,     "%rcx"=>0x010d8d48,
+                       "%rdx"=>0x01158d48,     "%rbx"=>0x011d8d48,
+                       "%rsp"=>0x01258d48,     "%rbp"=>0x012d8d48,
+                       "%rsi"=>0x01358d48,     "%rdi"=>0x013d8d48,
+                       "%r8" =>0x01058d4c,     "%r9" =>0x010d8d4c,
+                       "%r10"=>0x01158d4c,     "%r11"=>0x011d8d4c,
+                       "%r12"=>0x01258d4c,     "%r13"=>0x012d8d4c,
+                       "%r14"=>0x01358d4c,     "%r15"=>0x013d8d4c      );
 
        if ($line =~ /^\s*(\.\w+)/) {
            if (!$masm) {
                $self->{value} = $1;
                $line =~ s/\@abi\-omnipotent/\@function/;
                $line =~ s/\@function.*/\@function/;
-               $self->{value} = $line;
+               if ($line =~ /\.picmeup\s+(%r[\w]+)/i) {
+                   $self->{value} = sprintf "\t.long\t0x%x,0x90000000",$opcode{$1};
+               } else {
+                   $self->{value} = $line;
+               }
                $line = "";
                return $self;
            }
@@ -283,31 +325,30 @@ my $current_function;
            undef $self->{value};
            $line = substr($line,@+[0]); $line =~ s/^\s+//;
            SWITCH: for ($dir) {
-               /\.(text|data)/
+               /\.(text)/
                            && do { my $v=undef;
                                    $v="$current_segment\tENDS\n" if ($current_segment);
-                                   $current_segment = "_$1";
+                                   $current_segment = "_$1\$";
                                    $current_segment =~ tr/[a-z]/[A-Z]/;
-                                   $v.="$current_segment\tSEGMENT PARA";
+                                   $v.="$current_segment\tSEGMENT ALIGN(64) 'CODE'";
                                    $self->{value} = $v;
                                    last;
                                  };
                /\.globl/   && do { $self->{value} = "PUBLIC\t".$line; last; };
                /\.type/    && do { ($sym,$type,$narg) = split(',',$line);
-                                   if ($type eq "\@function")
-                                   {   undef $current_function;
+                                   if ($type eq "\@function") {
+                                       undef $current_function;
                                        $current_function->{name} = $sym;
                                        $current_function->{abi}  = "svr4";
                                        $current_function->{narg} = $narg;
-                                   }
-                                   elsif ($type eq "\@abi-omnipotent")
-                                   {   undef $current_function;
+                                   } elsif ($type eq "\@abi-omnipotent") {
+                                       undef $current_function;
                                        $current_function->{name} = $sym;
                                    }
                                    last;
                                  };
-               /\.size/    && do { if (defined($current_function))
-                                   {   $self->{value}="$current_function->{name}\tENDP";
+               /\.size/    && do { if (defined($current_function)) {
+                                       $self->{value}="$current_function->{name}\tENDP";
                                        undef $current_function;
                                    }
                                    last;
@@ -324,6 +365,9 @@ my $current_function;
                                    $self->{value} .= sprintf"0%Xh",oct($last);
                                    last;
                                  };
+               /\.picmeup/ && do { $self->{value} = sprintf"\tDD\t 0%Xh,090000000h",$opcode{$line};
+                                   last;
+                                 };
            }
            $line = "";
        }
@@ -340,8 +384,9 @@ while($line=<>) {
 
     chomp($line);
 
-    $line =~ s/\[#!].*$//;     # get rid of comments...
-    $line =~ s/^\s+//;         # ... and skip white spaces
+    $line =~ s|[#!].*$||;      # get rid of asm-style comments...
+    $line =~ s|/\*.*\*/||;     # ... and C-style comments...
+    $line =~ s|^\s+||;         # ... and skip white spaces in beginning
 
     undef $label;
     undef $opcode;
@@ -376,13 +421,11 @@ while($line=<>) {
            if (!$masm) {
                printf "\t%s\t%s,%s",   $opcode->out($dst->size()),
                                        $src->out($sz),$dst->out($sz);
-           }
-           else {
+           } else {
                printf "\t%s\t%s,%s",   $opcode->out(),
                                        $dst->out($sz),$src->out($sz);
            }
-       }
-       elsif (defined($src)) {
+       } elsif (defined($src)) {
            printf "\t%s\t%s",$opcode->out(),$src->out($sz);
        } else {
            printf "\t%s",$opcode->out();