8 files changed, 671 insertions, 0 deletions
diff --git a/openssl/crypto/bn/asm/x86/add.pl b/openssl/crypto/bn/asm/x86/add.pl
new file mode 100644
index 000000000..0b5cf583e
--- /dev/null
+++ b/openssl/crypto/bn/asm/x86/add.pl
@@ -0,0 +1,76 @@
+#!/usr/local/bin/perl
+# x86 assember
+
+sub bn_add_words
+	{
+	local($name)=@_;
+
+	&function_begin($name,"");
+
+	&comment("");
+	$a="esi";
+	$b="edi";
+	$c="eax";
+	$r="ebx";
+	$tmp1="ecx";
+	$tmp2="edx";
+	$num="ebp";
+
+	&mov($r,&wparam(0));	# get r
+	 &mov($a,&wparam(1));	# get a
+	&mov($b,&wparam(2));	# get b
+	 &mov($num,&wparam(3));	# get num
+	&xor($c,$c);		# clear carry
+	 &and($num,0xfffffff8);	# num / 8
+
+	&jz(&label("aw_finish"));
+
+	&set_label("aw_loop",0);
+	for ($i=0; $i<8; $i++)
+		{
+		&comment("Round $i");
+
+		&mov($tmp1,&DWP($i*4,$a,"",0)); 	# *a
+		 &mov($tmp2,&DWP($i*4,$b,"",0)); 	# *b
+		&add($tmp1,$c);
+		 &mov($c,0);
+		&adc($c,$c);
+		 &add($tmp1,$tmp2);
+		&adc($c,0);
+		 &mov(&DWP($i*4,$r,"",0),$tmp1); 	# *r
+		}
+
+	&comment("");
+	&add($a,32);
+	 &add($b,32);
+	&add($r,32);
+	 &sub($num,8);
+	&jnz(&label("aw_loop"));
+
+	&set_label("aw_finish",0);
+	&mov($num,&wparam(3));	# get num
+	&and($num,7);
+	 &jz(&label("aw_end"));
+
+	for ($i=0; $i<7; $i++)
+		{
+		&comment("Tail Round $i");
+		&mov($tmp1,&DWP($i*4,$a,"",0));	# *a
+		 &mov($tmp2,&DWP($i*4,$b,"",0));# *b
+		&add($tmp1,$c);
+		 &mov($c,0);
+		&adc($c,$c);
+		 &add($tmp1,$tmp2);
+		&adc($c,0);
+		 &dec($num) if ($i != 6);
+		&mov(&DWP($i*4,$r,"",0),$tmp1);	# *a
+		 &jz(&label("aw_end")) if ($i != 6);
+		}
+	&set_label("aw_end",0);
+
+#	&mov("eax",$c);		# $c is "eax"
+
+	&function_end($name);
+	}
+
+1;
diff --git a/openssl/crypto/bn/asm/x86/comba.pl b/openssl/crypto/bn/asm/x86/comba.pl
new file mode 100644
index 000000000..229125362
--- /dev/null
+++ b/openssl/crypto/bn/asm/x86/comba.pl
@@ -0,0 +1,277 @@
+#!/usr/local/bin/perl
+# x86 assember
+
+sub mul_add_c
+	{
+	local($a,$ai,$b,$bi,$c0,$c1,$c2,$pos,$i,$na,$nb)=@_;
+
+	# pos == -1 if eax and edx are pre-loaded, 0 to load from next
+	# words, and 1 if load return value
+
+	&comment("mul a[$ai]*b[$bi]");
+
+	# "eax" and "edx" will always be pre-loaded.
+	# &mov("eax",&DWP($ai*4,$a,"",0)) ;
+	# &mov("edx",&DWP($bi*4,$b,"",0));
+
+	&mul("edx");
+	&add($c0,"eax");
+	 &mov("eax",&DWP(($na)*4,$a,"",0)) if $pos == 0;	# laod next a
+	 &mov("eax",&wparam(0)) if $pos > 0;			# load r[]
+	 ###
+	&adc($c1,"edx");
+	 &mov("edx",&DWP(($nb)*4,$b,"",0)) if $pos == 0;	# laod next b
+	 &mov("edx",&DWP(($nb)*4,$b,"",0)) if $pos == 1;	# laod next b
+	 ###
+	&adc($c2,0);
+	 # is pos > 1, it means it is the last loop 
+	 &mov(&DWP($i*4,"eax","",0),$c0) if $pos > 0;		# save r[];
+	&mov("eax",&DWP(($na)*4,$a,"",0)) if $pos == 1;		# laod next a
+	}
+
+sub sqr_add_c
+	{
+	local($r,$a,$ai,$bi,$c0,$c1,$c2,$pos,$i,$na,$nb)=@_;
+
+	# pos == -1 if eax and edx are pre-loaded, 0 to load from next
+	# words, and 1 if load return value
+
+	&comment("sqr a[$ai]*a[$bi]");
+
+	# "eax" and "edx" will always be pre-loaded.
+	# &mov("eax",&DWP($ai*4,$a,"",0)) ;
+	# &mov("edx",&DWP($bi*4,$b,"",0));
+
+	if ($ai == $bi)
+		{ &mul("eax");}
+	else
+		{ &mul("edx");}
+	&add($c0,"eax");
+	 &mov("eax",&DWP(($na)*4,$a,"",0)) if $pos == 0;	# load next a
+	 ###
+	&adc($c1,"edx");
+	 &mov("edx",&DWP(($nb)*4,$a,"",0)) if ($pos == 1) && ($na != $nb);
+	 ###
+	&adc($c2,0);
+	 # is pos > 1, it means it is the last loop 
+	 &mov(&DWP($i*4,$r,"",0),$c0) if $pos > 0;		# save r[];
+	&mov("eax",&DWP(($na)*4,$a,"",0)) if $pos == 1;		# load next b
+	}
+
+sub sqr_add_c2
+	{
+	local($r,$a,$ai,$bi,$c0,$c1,$c2,$pos,$i,$na,$nb)=@_;
+
+	# pos == -1 if eax and edx are pre-loaded, 0 to load from next
+	# words, and 1 if load return value
+
+	&comment("sqr a[$ai]*a[$bi]");
+
+	# "eax" and "edx" will always be pre-loaded.
+	# &mov("eax",&DWP($ai*4,$a,"",0)) ;
+	# &mov("edx",&DWP($bi*4,$a,"",0));
+
+	if ($ai == $bi)
+		{ &mul("eax");}
+	else
+		{ &mul("edx");}
+	&add("eax","eax");
+	 ###
+	&adc("edx","edx");
+	 ###
+	&adc($c2,0);
+	 &add($c0,"eax");
+	&adc($c1,"edx");
+	 &mov("eax",&DWP(($na)*4,$a,"",0)) if $pos == 0;	# load next a
+	 &mov("eax",&DWP(($na)*4,$a,"",0)) if $pos == 1;	# load next b
+	&adc($c2,0);
+	&mov(&DWP($i*4,$r,"",0),$c0) if $pos > 0;		# save r[];
+	 &mov("edx",&DWP(($nb)*4,$a,"",0)) if ($pos <= 1) && ($na != $nb);
+	 ###
+	}
+
+sub bn_mul_comba
+	{
+	local($name,$num)=@_;
+	local($a,$b,$c0,$c1,$c2);
+	local($i,$as,$ae,$bs,$be,$ai,$bi);
+	local($tot,$end);
+
+	&function_begin_B($name,"");
+
+	$c0="ebx";
+	$c1="ecx";
+	$c2="ebp";
+	$a="esi";
+	$b="edi";
+	
+	$as=0;
+	$ae=0;
+	$bs=0;
+	$be=0;
+	$tot=$num+$num-1;
+
+	&push("esi");
+	 &mov($a,&wparam(1));
+	&push("edi");
+	 &mov($b,&wparam(2));
+	&push("ebp");
+	 &push("ebx");
+
+	&xor($c0,$c0);
+	 &mov("eax",&DWP(0,$a,"",0));	# load the first word 
+	&xor($c1,$c1);
+	 &mov("edx",&DWP(0,$b,"",0));	# load the first second 
+
+	for ($i=0; $i<$tot; $i++)
+		{
+		$ai=$as;
+		$bi=$bs;
+		$end=$be+1;
+
+		&comment("################## Calculate word $i"); 
+
+		for ($j=$bs; $j<$end; $j++)
+			{
+			&xor($c2,$c2) if ($j == $bs);
+			if (($j+1) == $end)
+				{
+				$v=1;
+				$v=2 if (($i+1) == $tot);
+				}
+			else
+				{ $v=0; }
+			if (($j+1) != $end)
+				{
+				$na=($ai-1);
+				$nb=($bi+1);
+				}
+			else
+				{
+				$na=$as+($i < ($num-1));
+				$nb=$bs+($i >= ($num-1));
+				}
+#printf STDERR "[$ai,$bi] -> [$na,$nb]\n";
+			&mul_add_c($a,$ai,$b,$bi,$c0,$c1,$c2,$v,$i,$na,$nb);
+			if ($v)
+				{
+				&comment("saved r[$i]");
+				# &mov("eax",&wparam(0));
+				# &mov(&DWP($i*4,"eax","",0),$c0);
+				($c0,$c1,$c2)=($c1,$c2,$c0);
+				}
+			$ai--;
+			$bi++;
+			}
+		$as++ if ($i < ($num-1));
+		$ae++ if ($i >= ($num-1));
+
+		$bs++ if ($i >= ($num-1));
+		$be++ if ($i < ($num-1));
+		}
+	&comment("save r[$i]");
+	# &mov("eax",&wparam(0));
+	&mov(&DWP($i*4,"eax","",0),$c0);
+
+	&pop("ebx");
+	&pop("ebp");
+	&pop("edi");
+	&pop("esi");
+	&ret();
+	&function_end_B($name);
+	}
+
+sub bn_sqr_comba
+	{
+	local($name,$num)=@_;
+	local($r,$a,$c0,$c1,$c2)=@_;
+	local($i,$as,$ae,$bs,$be,$ai,$bi);
+	local($b,$tot,$end,$half);
+
+	&function_begin_B($name,"");
+
+	$c0="ebx";
+	$c1="ecx";
+	$c2="ebp";
+	$a="esi";
+	$r="edi";
+
+	&push("esi");
+	 &push("edi");
+	&push("ebp");
+	 &push("ebx");
+	&mov($r,&wparam(0));
+	 &mov($a,&wparam(1));
+	&xor($c0,$c0);
+	 &xor($c1,$c1);
+	&mov("eax",&DWP(0,$a,"",0)); # load the first word
+
+	$as=0;
+	$ae=0;
+	$bs=0;
+	$be=0;
+	$tot=$num+$num-1;
+
+	for ($i=0; $i<$tot; $i++)
+		{
+		$ai=$as;
+		$bi=$bs;
+		$end=$be+1;
+
+		&comment("############### Calculate word $i");
+		for ($j=$bs; $j<$end; $j++)
+			{
+			&xor($c2,$c2) if ($j == $bs);
+			if (($ai-1) < ($bi+1))
+				{
+				$v=1;
+				$v=2 if ($i+1) == $tot;
+				}
+			else
+				{ $v=0; }
+			if (!$v)
+				{
+				$na=$ai-1;
+				$nb=$bi+1;
+				}
+			else
+				{
+				$na=$as+($i < ($num-1));
+				$nb=$bs+($i >= ($num-1));
+				}
+			if ($ai == $bi)
+				{
+				&sqr_add_c($r,$a,$ai,$bi,
+					$c0,$c1,$c2,$v,$i,$na,$nb);
+				}
+			else
+				{
+				&sqr_add_c2($r,$a,$ai,$bi,
+					$c0,$c1,$c2,$v,$i,$na,$nb);
+				}
+			if ($v)
+				{
+				&comment("saved r[$i]");
+				#&mov(&DWP($i*4,$r,"",0),$c0);
+				($c0,$c1,$c2)=($c1,$c2,$c0);
+				last;
+				}
+			$ai--;
+			$bi++;
+			}
+		$as++ if ($i < ($num-1));
+		$ae++ if ($i >= ($num-1));
+
+		$bs++ if ($i >= ($num-1));
+		$be++ if ($i < ($num-1));
+		}
+	&mov(&DWP($i*4,$r,"",0),$c0);
+	&pop("ebx");
+	&pop("ebp");
+	&pop("edi");
+	&pop("esi");
+	&ret();
+	&function_end_B($name);
+	}
+
+1;
diff --git a/openssl/crypto/bn/asm/x86/div.pl b/openssl/crypto/bn/asm/x86/div.pl
new file mode 100644
index 000000000..0e90152ca
--- /dev/null
+++ b/openssl/crypto/bn/asm/x86/div.pl
@@ -0,0 +1,15 @@
+#!/usr/local/bin/perl
+# x86 assember
+
+sub bn_div_words
+	{
+	local($name)=@_;
+
+	&function_begin($name,"");
+	&mov("edx",&wparam(0));	#
+	&mov("eax",&wparam(1));	#
+	&mov("ebx",&wparam(2));	#
+	&div("ebx");
+	&function_end($name);
+	}
+1;
diff --git a/openssl/crypto/bn/asm/x86/f b/openssl/crypto/bn/asm/x86/f
new file mode 100644
index 000000000..22e411222
--- /dev/null
+++ b/openssl/crypto/bn/asm/x86/f
@@ -0,0 +1,3 @@
+#!/usr/local/bin/perl
+# x86 assember
+
diff --git a/openssl/crypto/bn/asm/x86/mul.pl b/openssl/crypto/bn/asm/x86/mul.pl
new file mode 100644
index 000000000..674cb9b05
--- /dev/null
+++ b/openssl/crypto/bn/asm/x86/mul.pl
@@ -0,0 +1,77 @@
+#!/usr/local/bin/perl
+# x86 assember
+
+sub bn_mul_words
+	{
+	local($name)=@_;
+
+	&function_begin($name,"");
+
+	&comment("");
+	$Low="eax";
+	$High="edx";
+	$a="ebx";
+	$w="ecx";
+	$r="edi";
+	$c="esi";
+	$num="ebp";
+
+	&xor($c,$c);		# clear carry
+	&mov($r,&wparam(0));	#
+	&mov($a,&wparam(1));	#
+	&mov($num,&wparam(2));	#
+	&mov($w,&wparam(3));	#
+
+	&and($num,0xfffffff8);	# num / 8
+	&jz(&label("mw_finish"));
+
+	&set_label("mw_loop",0);
+	for ($i=0; $i<32; $i+=4)
+		{
+		&comment("Round $i");
+
+		 &mov("eax",&DWP($i,$a,"",0)); 	# *a
+		&mul($w);			# *a * w
+		&add("eax",$c);			# L(t)+=c
+		 # XXX
+
+		&adc("edx",0);			# H(t)+=carry
+		 &mov(&DWP($i,$r,"",0),"eax");	# *r= L(t);
+
+		&mov($c,"edx");			# c=  H(t);
+		}
+
+	&comment("");
+	&add($a,32);
+	&add($r,32);
+	&sub($num,8);
+	&jz(&label("mw_finish"));
+	&jmp(&label("mw_loop"));
+
+	&set_label("mw_finish",0);
+	&mov($num,&wparam(2));	# get num
+	&and($num,7);
+	&jnz(&label("mw_finish2"));
+	&jmp(&label("mw_end"));
+
+	&set_label("mw_finish2",1);
+	for ($i=0; $i<7; $i++)
+		{
+		&comment("Tail Round $i");
+		 &mov("eax",&DWP($i*4,$a,"",0));# *a
+		&mul($w);			# *a * w
+		&add("eax",$c);			# L(t)+=c
+		 # XXX
+		&adc("edx",0);			# H(t)+=carry
+		 &mov(&DWP($i*4,$r,"",0),"eax");# *r= L(t);
+		&mov($c,"edx");			# c=  H(t);
+		 &dec($num) if ($i != 7-1);
+		&jz(&label("mw_end")) if ($i != 7-1);
+		}
+	&set_label("mw_end",0);
+	&mov("eax",$c);
+
+	&function_end($name);
+	}
+
+1;
diff --git a/openssl/crypto/bn/asm/x86/mul_add.pl b/openssl/crypto/bn/asm/x86/mul_add.pl
new file mode 100644
index 000000000..61830d3a9
--- /dev/null
+++ b/openssl/crypto/bn/asm/x86/mul_add.pl
@@ -0,0 +1,87 @@
+#!/usr/local/bin/perl
+# x86 assember
+
+sub bn_mul_add_words
+	{
+	local($name)=@_;
+
+	&function_begin($name,"");
+
+	&comment("");
+	$Low="eax";
+	$High="edx";
+	$a="ebx";
+	$w="ebp";
+	$r="edi";
+	$c="esi";
+
+	&xor($c,$c);		# clear carry
+	&mov($r,&wparam(0));	#
+
+	&mov("ecx",&wparam(2));	#
+	&mov($a,&wparam(1));	#
+
+	&and("ecx",0xfffffff8);	# num / 8
+	&mov($w,&wparam(3));	#
+
+	&push("ecx");		# Up the stack for a tmp variable
+
+	&jz(&label("maw_finish"));
+
+	&set_label("maw_loop",0);
+
+	&mov(&swtmp(0),"ecx");	#
+
+	for ($i=0; $i<32; $i+=4)
+		{
+		&comment("Round $i");
+
+		 &mov("eax",&DWP($i,$a,"",0)); 	# *a
+		&mul($w);			# *a * w
+		&add("eax",$c);		# L(t)+= *r
+		 &mov($c,&DWP($i,$r,"",0));	# L(t)+= *r
+		&adc("edx",0);			# H(t)+=carry
+		 &add("eax",$c);		# L(t)+=c
+		&adc("edx",0);			# H(t)+=carry
+		 &mov(&DWP($i,$r,"",0),"eax");	# *r= L(t);
+		&mov($c,"edx");			# c=  H(t);
+		}
+
+	&comment("");
+	&mov("ecx",&swtmp(0));	#
+	&add($a,32);
+	&add($r,32);
+	&sub("ecx",8);
+	&jnz(&label("maw_loop"));
+
+	&set_label("maw_finish",0);
+	&mov("ecx",&wparam(2));	# get num
+	&and("ecx",7);
+	&jnz(&label("maw_finish2"));	# helps branch prediction
+	&jmp(&label("maw_end"));
+
+	&set_label("maw_finish2",1);
+	for ($i=0; $i<7; $i++)
+		{
+		&comment("Tail Round $i");
+		 &mov("eax",&DWP($i*4,$a,"",0));# *a
+		&mul($w);			# *a * w
+		&add("eax",$c);			# L(t)+=c
+		 &mov($c,&DWP($i*4,$r,"",0));	# L(t)+= *r
+		&adc("edx",0);			# H(t)+=carry
+		 &add("eax",$c);
+		&adc("edx",0);			# H(t)+=carry
+		 &dec("ecx") if ($i != 7-1);
+		&mov(&DWP($i*4,$r,"",0),"eax");	# *r= L(t);
+		 &mov($c,"edx");			# c=  H(t);
+		&jz(&label("maw_end")) if ($i != 7-1);
+		}
+	&set_label("maw_end",0);
+	&mov("eax",$c);
+
+	&pop("ecx");	# clear variable from
+
+	&function_end($name);
+	}
+
+1;
diff --git a/openssl/crypto/bn/asm/x86/sqr.pl b/openssl/crypto/bn/asm/x86/sqr.pl
new file mode 100644
index 000000000..1f90993cf
--- /dev/null
+++ b/openssl/crypto/bn/asm/x86/sqr.pl
@@ -0,0 +1,60 @@
+#!/usr/local/bin/perl
+# x86 assember
+
+sub bn_sqr_words
+	{
+	local($name)=@_;
+
+	&function_begin($name,"");
+
+	&comment("");
+	$r="esi";
+	$a="edi";
+	$num="ebx";
+
+	&mov($r,&wparam(0));	#
+	&mov($a,&wparam(1));	#
+	&mov($num,&wparam(2));	#
+
+	&and($num,0xfffffff8);	# num / 8
+	&jz(&label("sw_finish"));
+
+	&set_label("sw_loop",0);
+	for ($i=0; $i<32; $i+=4)
+		{
+		&comment("Round $i");
+		&mov("eax",&DWP($i,$a,"",0)); 	# *a
+		 # XXX
+		&mul("eax");			# *a * *a
+		&mov(&DWP($i*2,$r,"",0),"eax");	#
+		 &mov(&DWP($i*2+4,$r,"",0),"edx");#
+		}
+
+	&comment("");
+	&add($a,32);
+	&add($r,64);
+	&sub($num,8);
+	&jnz(&label("sw_loop"));
+
+	&set_label("sw_finish",0);
+	&mov($num,&wparam(2));	# get num
+	&and($num,7);
+	&jz(&label("sw_end"));
+
+	for ($i=0; $i<7; $i++)
+		{
+		&comment("Tail Round $i");
+		&mov("eax",&DWP($i*4,$a,"",0));	# *a
+		 # XXX
+		&mul("eax");			# *a * *a
+		&mov(&DWP($i*8,$r,"",0),"eax");	#
+		 &dec($num) if ($i != 7-1);
+		&mov(&DWP($i*8+4,$r,"",0),"edx");
+		 &jz(&label("sw_end")) if ($i != 7-1);
+		}
+	&set_label("sw_end",0);
+
+	&function_end($name);
+	}
+
+1;
diff --git a/openssl/crypto/bn/asm/x86/sub.pl b/openssl/crypto/bn/asm/x86/sub.pl
new file mode 100644
index 000000000..837b0e1b0
--- /dev/null
+++ b/openssl/crypto/bn/asm/x86/sub.pl
@@ -0,0 +1,76 @@
+#!/usr/local/bin/perl
+# x86 assember
+
+sub bn_sub_words
+	{
+	local($name)=@_;
+
+	&function_begin($name,"");
+
+	&comment("");
+	$a="esi";
+	$b="edi";
+	$c="eax";
+	$r="ebx";
+	$tmp1="ecx";
+	$tmp2="edx";
+	$num="ebp";
+
+	&mov($r,&wparam(0));	# get r
+	 &mov($a,&wparam(1));	# get a
+	&mov($b,&wparam(2));	# get b
+	 &mov($num,&wparam(3));	# get num
+	&xor($c,$c);		# clear carry
+	 &and($num,0xfffffff8);	# num / 8
+
+	&jz(&label("aw_finish"));
+
+	&set_label("aw_loop",0);
+	for ($i=0; $i<8; $i++)
+		{
+		&comment("Round $i");
+
+		&mov($tmp1,&DWP($i*4,$a,"",0)); 	# *a
+		 &mov($tmp2,&DWP($i*4,$b,"",0)); 	# *b
+		&sub($tmp1,$c);
+		 &mov($c,0);
+		&adc($c,$c);
+		 &sub($tmp1,$tmp2);
+		&adc($c,0);
+		 &mov(&DWP($i*4,$r,"",0),$tmp1); 	# *r
+		}
+
+	&comment("");
+	&add($a,32);
+	 &add($b,32);
+	&add($r,32);
+	 &sub($num,8);
+	&jnz(&label("aw_loop"));
+
+	&set_label("aw_finish",0);
+	&mov($num,&wparam(3));	# get num
+	&and($num,7);
+	 &jz(&label("aw_end"));
+
+	for ($i=0; $i<7; $i++)
+		{
+		&comment("Tail Round $i");
+		&mov($tmp1,&DWP($i*4,$a,"",0));	# *a
+		 &mov($tmp2,&DWP($i*4,$b,"",0));# *b
+		&sub($tmp1,$c);
+		 &mov($c,0);
+		&adc($c,$c);
+		 &sub($tmp1,$tmp2);
+		&adc($c,0);
+		 &dec($num) if ($i != 6);
+		&mov(&DWP($i*4,$r,"",0),$tmp1);	# *a
+		 &jz(&label("aw_end")) if ($i != 6);
+		}
+	&set_label("aw_end",0);
+
+#	&mov("eax",$c);		# $c is "eax"
+
+	&function_end($name);
+	}
+
+1;