perlasm/x86gas.pl: limit special OPENSSL_ia32cap_P treatment to ELF.
[openssl.git] / crypto / perlasm / x86gas.pl
index 5347bf80a03c5501a781684c7645c39a2c5fbe6c..63b2301fd1f096a13e00576af8238a81d124634d 100644 (file)
@@ -25,33 +25,28 @@ sub opsize()
 # expand opcode with size suffix;
 # prefix numeric constants with $;
 sub ::generic
-{ my($opcode,$dst,$src)=@_;
-  my($tmp,$suffix,@arg);
-
-    if (defined($src))
-    {  $src =~ s/^(e?[a-dsixphl]{2})$/%$1/o;
-       $src =~ s/^(x?mm[0-7])$/%$1/o;
-       $src =~ s/^(\-?[0-9]+)$/\$$1/o;
-       $src =~ s/^(\-?0x[0-9a-f]+)$/\$$1/o;
-       push(@arg,$src);
-    }
-    if (defined($dst))
-    {  $dst =~ s/^(\*?)(e?[a-dsixphl]{2})$/$1%$2/o;
-       $dst =~ s/^(x?mm[0-7])$/%$1/o;
-       $dst =~ s/^(\-?[0-9]+)$/\$$1/o          if(!defined($src));
-       $dst =~ s/^(\-?0x[0-9a-f]+)$/\$$1/o     if(!defined($src));
-       push(@arg,$dst);
+{ my($opcode,@arg)=@_;
+  my($suffix,$dst,$src);
+
+    @arg=reverse(@arg);
+
+    for (@arg)
+    {  s/^(\*?)(e?[a-dsixphl]{2})$/$1%$2/o;    # gp registers
+       s/^([xy]?mm[0-7])$/%$1/o;               # xmm/mmx registers
+       s/^(\-?[0-9]+)$/\$$1/o;                 # constants
+       s/^(\-?0x[0-9a-f]+)$/\$$1/o;            # constants
     }
 
+    $dst = $arg[$#arg]         if ($#arg>=0);
+    $src = $arg[$#arg-1]       if ($#arg>=1);
     if    ($dst =~ m/^%/o)     { $suffix=&opsize($dst); }
     elsif ($src =~ m/^%/o)     { $suffix=&opsize($src); }
     else                       { $suffix="l";           }
     undef $suffix if ($dst =~ m/^%[xm]/o || $src =~ m/^%[xm]/o);
 
     if ($#_==0)                                { &::emit($opcode);             }
-    elsif ($opcode =~ m/^j/o && $#_==1)        { &::emit($opcode,@arg);        }
-    elsif ($opcode eq "call" && $#_==1)        { &::emit($opcode,@arg);        }
-    elsif ($opcode =~ m/^set/&& $#_==1)        { &::emit($opcode,@arg);        }
+    elsif ($#_==1 && $opcode =~ m/^(call|clflush|j|loop|set)/o)
+                                       { &::emit($opcode,@arg);        }
     else                               { &::emit($opcode.$suffix,@arg);}
 
   1;
@@ -71,23 +66,12 @@ sub ::jmp_ptr       { &::generic("jmp","*$_[0]");   }
 
 *::bswap = sub { &::emit("bswap","%$_[0]");    } if (!$::i386);
 
-*::pshufw = sub
-{ my($dst,$src,$magic)=@_;
-    &::emit("pshufw","\$$magic","%$src","%$dst");
-};
-*::shld = sub
-{ my($dst,$src,$bits)=@_;
-    &::emit("shldl",$bit eq "cl"?"%cl":"\$$bits","%$src","%$dst");
-};
-*::shrd = sub
-{ my($dst,$src,$bits)=@_;
-    &::emit("shrdl",$bit eq "cl"?"%cl":"\$$bits","%$src","%$dst");
-};
-
 sub ::DWP
 { my($addr,$reg1,$reg2,$idx)=@_;
   my $ret="";
 
+    if (!defined($idx) && 1*$reg2) { $idx=$reg2; $reg2=$reg1; undef $reg1; }
+
     $addr =~ s/^\s+//;
     # prepend global references with optional underscore
     $addr =~ s/^([^\+\-0-9][^\+\-]*)/&::islabel($1) or "$nmdecor$1"/ige;
@@ -108,6 +92,7 @@ sub ::DWP
 }
 sub ::QWP      { &::DWP(@_);   }
 sub ::BP       { &::DWP(@_);   }
+sub ::WP       { &::DWP(@_);   }
 sub ::BC       { @_;           }
 sub ::DWC      { @_;           }
 
@@ -124,7 +109,7 @@ sub ::function_begin_B
 
     push(@out,".globl\t$func\n")       if ($global);
     if ($::coff)
-    {  push(@out,".def\t$func;\t.scl\t2;\t.type\t32;\t.endef\n"); }
+    {  push(@out,".def\t$func;\t.scl\t".(3-$global).";\t.type\t32;\t.endef\n"); }
     elsif (($::aout and !$::pic) or $::macosx)
     { }
     else
@@ -166,29 +151,30 @@ sub ::public_label
 {   push(@out,".globl\t".&::LABEL($_[0],$nmdecor.$_[0])."\n");   }
 
 sub ::file_end
-{   if (grep {/\b${nmdecor}OPENSSL_ia32cap_P\b/i} @out) {
-       my $tmp=".comm\t${nmdecor}OPENSSL_ia32cap_P,4";
-       if ($::elf)     { push (@out,"$tmp,4\n"); }
-       else            { push (@out,"$tmp\n"); }
-    }
-    if ($::macosx)
+{   if ($::macosx)
     {  if (%non_lazy_ptr)
        {   push(@out,".section __IMPORT,__pointers,non_lazy_symbol_pointers\n");
            foreach $i (keys %non_lazy_ptr)
            {   push(@out,"$non_lazy_ptr{$i}:\n.indirect_symbol\t$i\n.long\t0\n");   }
        }
     }
+    if (grep {/\b${nmdecor}OPENSSL_ia32cap_P\b/i} @out) {
+       my $tmp=".comm\t${nmdecor}OPENSSL_ia32cap_P,16";
+       if ($::macosx)  { push (@out,"$tmp,2\n"); }
+       elsif ($::elf)  { push (@out,"$tmp,4\n"); }
+       else            { push (@out,"$tmp\n"); }
+    }
     push(@out,$initseg) if ($initseg);
 }
 
 sub ::data_byte        {   push(@out,".byte\t".join(',',@_)."\n");   }
+sub ::data_short{   push(@out,".value\t".join(',',@_)."\n");  }
 sub ::data_word {   push(@out,".long\t".join(',',@_)."\n");   }
 
 sub ::align
-{ my $val=$_[0],$p2,$i;
+{ my $val=$_[0];
     if ($::aout)
-    {  for ($p2=0;$val!=0;$val>>=1) { $p2++; }
-       $val=$p2-1;
+    {  $val=int(log($val)/log(2));
        $val.=",0x90";
     }
     push(@out,".align\t$val\n");
@@ -197,7 +183,7 @@ sub ::align
 sub ::picmeup
 { my($dst,$sym,$base,$reflabel)=@_;
 
-    if ($::pic && ($::elf || $::aout))
+    if (($::pic && ($::elf || $::aout)) || $::macosx)
     {  if (!defined($base))
        {   &::call(&::label("PIC_me_up"));
            &::set_label("PIC_me_up");
@@ -210,6 +196,8 @@ sub ::picmeup
            &::mov($dst,&::DWP("$indirect-$reflabel",$base));
            $non_lazy_ptr{"$nmdecor$sym"}=$indirect;
        }
+       elsif ($sym eq "OPENSSL_ia32cap_P" && $::elf>0)
+       {   &::lea($dst,&::DWP("$sym-$reflabel",$base));   }
        else
        {   &::lea($dst,&::DWP("_GLOBAL_OFFSET_TABLE_+[.-$reflabel]",
                            $base));
@@ -223,13 +211,17 @@ sub ::picmeup
 sub ::initseg
 { my $f=$nmdecor.shift;
 
-    if ($::elf)
+    if ($::android)
+    {  $initseg.=<<___;
+.section       .init_array
+.align 4
+.long  $f
+___
+    }
+    elsif ($::elf)
     {  $initseg.=<<___;
 .section       .init
        call    $f
-       jmp     .Linitalign
-.align $align
-.Linitalign:
 ___
     }
     elsif ($::coff)
@@ -258,4 +250,9 @@ ___
     }
 }
 
+sub ::dataseg
+{   push(@out,".data\n");   }
+
+*::hidden = sub { push(@out,".hidden\t$nmdecor$_[0]\n"); } if ($::elf);
+
 1;