svn merge -r514:HEAD ^/branches/released .

5 files changed, 1361 insertions, 382 deletions

diff --git a/openssl/crypto/rc4/asm/rc4-586.pl b/openssl/crypto/rc4/asm/rc4-586.pl
index ef7eee766..38a44a70e 100644
--- a/openssl/crypto/rc4/asm/rc4-586.pl
+++ b/openssl/crypto/rc4/asm/rc4-586.pl
@@ -1,14 +1,21 @@
-#!/usr/local/bin/perl
+#!/usr/bin/env perl
+
+# ====================================================================
+# [Re]written by Andy Polyakov <appro@fy.chalmers.se> for the OpenSSL
+# project. The module is, however, dual licensed under OpenSSL and
+# CRYPTOGAMS licenses depending on where you obtain it. For further
+# details see http://www.openssl.org/~appro/cryptogams/.
+# ====================================================================
 
 # At some point it became apparent that the original SSLeay RC4
-# assembler implementation performs suboptimaly on latest IA-32
+# assembler implementation performs suboptimally on latest IA-32
 # microarchitectures. After re-tuning performance has changed as
 # following:
 #
-# Pentium	+0%
-# Pentium III	+17%
-# AMD		+52%(*)
-# P4		+180%(**)
+# Pentium	-10%
+# Pentium III	+12%
+# AMD		+50%(*)
+# P4		+250%(**)
 #
 # (*)	This number is actually a trade-off:-) It's possible to
 #	achieve	+72%, but at the cost of -48% off PIII performance.
@@ -17,214 +24,247 @@
 #	For reference! This code delivers ~80% of rc4-amd64.pl
 #	performance on the same Opteron machine.
 # (**)	This number requires compressed key schedule set up by
-#	RC4_set_key and therefore doesn't apply to 0.9.7 [option for
-#	compressed key schedule is implemented in 0.9.8 and later,
-#	see commentary section in rc4_skey.c for further details].
+#	RC4_set_key [see commentary below for further details].
 #
 #					<appro@fy.chalmers.se>
 
-push(@INC,"perlasm","../../perlasm");
+$0 =~ m/(.*[\/\\])[^\/\\]+$/; $dir=$1;
+push(@INC,"${dir}","${dir}../../perlasm");
 require "x86asm.pl";
 
 &asm_init($ARGV[0],"rc4-586.pl");
 
-$x="eax";
-$y="ebx";
+$xx="eax";
+$yy="ebx";
 $tx="ecx";
 $ty="edx";
-$in="esi";
-$out="edi";
-$d="ebp";
-
-&RC4("RC4");
-
-&asm_finish();
-
-sub RC4_loop
-	{
-	local($n,$p,$char)=@_;
-
-	&comment("Round $n");
-
-	if ($char)
-		{
-		if ($p >= 0)
-			{
-			 &mov($ty,	&swtmp(2));
-			&cmp($ty,	$in);
-			 &jbe(&label("finished"));
-			&inc($in);
-			}
-		else
-			{
-			&add($ty,	8);
-			 &inc($in);
-			&cmp($ty,	$in);
-			 &jb(&label("finished"));
-			&mov(&swtmp(2),	$ty);
-			}
-		}
-	# Moved out
-	# &mov(	$tx,		&DWP(0,$d,$x,4)) if $p < 0;
-
-	&add(	&LB($y),	&LB($tx));
-	&mov(	$ty,		&DWP(0,$d,$y,4));
-	 # XXX
-	&mov(	&DWP(0,$d,$x,4),$ty);
-	 &add(	$ty,		$tx);
-	&mov(	&DWP(0,$d,$y,4),$tx);
-	 &and(	$ty,		0xff);
-	 &inc(	&LB($x));			# NEXT ROUND
-	&mov(	$tx,		&DWP(0,$d,$x,4)) if $p < 1; # NEXT ROUND
-	 &mov(	$ty,		&DWP(0,$d,$ty,4));
-
-	if (!$char)
-		{
-		#moved up into last round
-		if ($p >= 1)
-			{
-			&add(	$out,	8)
-			}
-		&movb(	&BP($n,"esp","",0),	&LB($ty));
-		}
-	else
-		{
-		# Note in+=8 has occured
-		&movb(	&HB($ty),	&BP(-1,$in,"",0));
-		 # XXX
-		&xorb(&LB($ty),		&HB($ty));
-		 # XXX
-		&movb(&BP($n,$out,"",0),&LB($ty));
-		}
+$inp="esi";
+$out="ebp";
+$dat="edi";
+
+sub RC4_loop {
+  my $i=shift;
+  my $func = ($i==0)?*mov:*or;
+
+	&add	(&LB($yy),&LB($tx));
+	&mov	($ty,&DWP(0,$dat,$yy,4));
+	&mov	(&DWP(0,$dat,$yy,4),$tx);
+	&mov	(&DWP(0,$dat,$xx,4),$ty);
+	&add	($ty,$tx);
+	&inc	(&LB($xx));
+	&and	($ty,0xff);
+	&ror	($out,8)	if ($i!=0);
+	if ($i<3) {
+	  &mov	($tx,&DWP(0,$dat,$xx,4));
+	} else {
+	  &mov	($tx,&wparam(3));	# reload [re-biased] out
 	}
-
-
-sub RC4
-	{
-	local($name)=@_;
-
-	&function_begin_B($name,"");
-
-	&mov($ty,&wparam(1));		# len
-	&cmp($ty,0);
-	&jne(&label("proceed"));
-	&ret();
-	&set_label("proceed");
-
-	&comment("");
-
-	&push("ebp");
-	 &push("ebx");
-	&push("esi");
-	 &xor(	$x,	$x);		# avoid partial register stalls
-	&push("edi");
-	 &xor(	$y,	$y);		# avoid partial register stalls
-	&mov(	$d,	&wparam(0));	# key
-	 &mov(	$in,	&wparam(2));
-
-	&movb(	&LB($x),	&BP(0,$d,"",1));
-	 &movb(	&LB($y),	&BP(4,$d,"",1));
-
-	&mov(	$out,	&wparam(3));
-	 &inc(	&LB($x));
-
-	&stack_push(3);	# 3 temp variables
-	 &add(	$d,	8);
-
-	# detect compressed schedule, see commentary section in rc4_skey.c...
-	# in 0.9.7 context ~50 bytes below RC4_CHAR label remain redundant,
-	# as compressed key schedule is set up in 0.9.8 and later.
-	&cmp(&DWP(256,$d),-1);
-	&je(&label("RC4_CHAR"));
-
-	 &lea(	$ty,	&DWP(-8,$ty,$in));
-
-	# check for 0 length input
-
-	 &mov(	&swtmp(2),	$ty);	# this is now address to exit at
-	&mov(	$tx,	&DWP(0,$d,$x,4));
-
-	 &cmp(	$ty,	$in);
-	&jb(	&label("end")); # less than 8 bytes
-
-	&set_label("start");
-
-	# filling DELAY SLOT
-	&add(	$in,	8);
-
-	&RC4_loop(0,-1,0);
-	&RC4_loop(1,0,0);
-	&RC4_loop(2,0,0);
-	&RC4_loop(3,0,0);
-	&RC4_loop(4,0,0);
-	&RC4_loop(5,0,0);
-	&RC4_loop(6,0,0);
-	&RC4_loop(7,1,0);
-	
-	&comment("apply the cipher text");
-	# xor the cipher data with input
-
-	#&add(	$out,	8); #moved up into last round
-
-	&mov(	$tx,	&swtmp(0));
-	 &mov(	$ty,	&DWP(-8,$in,"",0));
-	&xor(	$tx,	$ty);
-	 &mov(	$ty,	&DWP(-4,$in,"",0)); 
-	&mov(	&DWP(-8,$out,"",0),	$tx);
-	 &mov(	$tx,	&swtmp(1));
-	&xor(	$tx,	$ty);
-	 &mov(	$ty,	&swtmp(2));	# load end ptr;
-	&mov(	&DWP(-4,$out,"",0),	$tx);
-	 &mov(	$tx,		&DWP(0,$d,$x,4));
-	&cmp($in,	$ty);
-	 &jbe(&label("start"));
-
-	&set_label("end");
-
-	# There is quite a bit of extra crap in RC4_loop() for this
-	# first round
-	&RC4_loop(0,-1,1);
-	&RC4_loop(1,0,1);
-	&RC4_loop(2,0,1);
-	&RC4_loop(3,0,1);
-	&RC4_loop(4,0,1);
-	&RC4_loop(5,0,1);
-	&RC4_loop(6,1,1);
-
-	&jmp(&label("finished"));
-
-	&align(16);
-	# this is essentially Intel P4 specific codepath, see rc4_skey.c,
-	# and is engaged in 0.9.8 and later context...
-	&set_label("RC4_CHAR");
-
-	&lea	($ty,&DWP(0,$in,$ty));
-	&mov	(&swtmp(2),$ty);
-	&movz	($tx,&BP(0,$d,$x));
-
+	&$func	($out,&DWP(0,$dat,$ty,4));
+}
+
+# void RC4(RC4_KEY *key,size_t len,const unsigned char *inp,unsigned char *out);
+&function_begin("RC4");
+	&mov	($dat,&wparam(0));	# load key schedule pointer
+	&mov	($ty, &wparam(1));	# load len
+	&mov	($inp,&wparam(2));	# load inp
+	&mov	($out,&wparam(3));	# load out
+
+	&xor	($xx,$xx);		# avoid partial register stalls
+	&xor	($yy,$yy);
+
+	&cmp	($ty,0);		# safety net
+	&je	(&label("abort"));
+
+	&mov	(&LB($xx),&BP(0,$dat));	# load key->x
+	&mov	(&LB($yy),&BP(4,$dat));	# load key->y
+	&add	($dat,8);
+
+	&lea	($tx,&DWP(0,$inp,$ty));
+	&sub	($out,$inp);		# re-bias out
+	&mov	(&wparam(1),$tx);	# save input+len
+
+	&inc	(&LB($xx));
+
+	# detect compressed key schedule...
+	&cmp	(&DWP(256,$dat),-1);
+	&je	(&label("RC4_CHAR"));
+
+	&mov	($tx,&DWP(0,$dat,$xx,4));
+
+	&and	($ty,-4);		# how many 4-byte chunks?
+	&jz	(&label("loop1"));
+
+	&lea	($ty,&DWP(-4,$inp,$ty));
+	&mov	(&wparam(2),$ty);	# save input+(len/4)*4-4
+	&mov	(&wparam(3),$out);	# $out as accumulator in this loop
+
+	&set_label("loop4",16);
+		for ($i=0;$i<4;$i++) { RC4_loop($i); }
+		&ror	($out,8);
+		&xor	($out,&DWP(0,$inp));
+		&cmp	($inp,&wparam(2));	# compare to input+(len/4)*4-4
+		&mov	(&DWP(0,$tx,$inp),$out);# $tx holds re-biased out here
+		&lea	($inp,&DWP(4,$inp));
+		&mov	($tx,&DWP(0,$dat,$xx,4));
+	&jb	(&label("loop4"));
+
+	&cmp	($inp,&wparam(1));	# compare to input+len
+	&je	(&label("done"));
+	&mov	($out,&wparam(3));	# restore $out
+
+	&set_label("loop1",16);
+		&add	(&LB($yy),&LB($tx));
+		&mov	($ty,&DWP(0,$dat,$yy,4));
+		&mov	(&DWP(0,$dat,$yy,4),$tx);
+		&mov	(&DWP(0,$dat,$xx,4),$ty);
+		&add	($ty,$tx);
+		&inc	(&LB($xx));
+		&and	($ty,0xff);
+		&mov	($ty,&DWP(0,$dat,$ty,4));
+		&xor	(&LB($ty),&BP(0,$inp));
+		&lea	($inp,&DWP(1,$inp));
+		&mov	($tx,&DWP(0,$dat,$xx,4));
+		&cmp	($inp,&wparam(1));	# compare to input+len
+		&mov	(&BP(-1,$out,$inp),&LB($ty));
+	&jb	(&label("loop1"));
+
+	&jmp	(&label("done"));
+
+# this is essentially Intel P4 specific codepath...
+&set_label("RC4_CHAR",16);
+	&movz	($tx,&BP(0,$dat,$xx));
 	# strangely enough unrolled loop performs over 20% slower...
-	&set_label("RC4_CHAR_loop");
-		&add	(&LB($y),&LB($tx));
-		&movz	($ty,&BP(0,$d,$y));
-		&movb	(&BP(0,$d,$y),&LB($tx));
-		&movb	(&BP(0,$d,$x),&LB($ty));
+	&set_label("cloop1");
+		&add	(&LB($yy),&LB($tx));
+		&movz	($ty,&BP(0,$dat,$yy));
+		&mov	(&BP(0,$dat,$yy),&LB($tx));
+		&mov	(&BP(0,$dat,$xx),&LB($ty));
 		&add	(&LB($ty),&LB($tx));
-		&movz	($ty,&BP(0,$d,$ty));
-		&add	(&LB($x),1);
-		&xorb	(&LB($ty),&BP(0,$in));
-		&lea	($in,&DWP(1,$in));
-		&movz	($tx,&BP(0,$d,$x));
-		&cmp	($in,&swtmp(2));
-		&movb	(&BP(0,$out),&LB($ty));
-		&lea	($out,&DWP(1,$out));
-	&jb	(&label("RC4_CHAR_loop"));
-
-	&set_label("finished");
-	&dec(	$x);
-	 &stack_pop(3);
-	&movb(	&BP(-4,$d,"",0),&LB($y));
-	 &movb(	&BP(-8,$d,"",0),&LB($x));
-
-	&function_end($name);
-	}
+		&movz	($ty,&BP(0,$dat,$ty));
+		&add	(&LB($xx),1);
+		&xor	(&LB($ty),&BP(0,$inp));
+		&lea	($inp,&DWP(1,$inp));
+		&movz	($tx,&BP(0,$dat,$xx));
+		&cmp	($inp,&wparam(1));
+		&mov	(&BP(-1,$out,$inp),&LB($ty));
+	&jb	(&label("cloop1"));
+
+&set_label("done");
+	&dec	(&LB($xx));
+	&mov	(&BP(-4,$dat),&LB($yy));	# save key->y
+	&mov	(&BP(-8,$dat),&LB($xx));	# save key->x
+&set_label("abort");
+&function_end("RC4");
+
+########################################################################
+
+$inp="esi";
+$out="edi";
+$idi="ebp";
+$ido="ecx";
+$idx="edx";
+
+&external_label("OPENSSL_ia32cap_P");
+
+# void RC4_set_key(RC4_KEY *key,int len,const unsigned char *data);
+&function_begin("RC4_set_key");
+	&mov	($out,&wparam(0));		# load key
+	&mov	($idi,&wparam(1));		# load len
+	&mov	($inp,&wparam(2));		# load data
+	&picmeup($idx,"OPENSSL_ia32cap_P");
+
+	&lea	($out,&DWP(2*4,$out));		# &key->data
+	&lea	($inp,&DWP(0,$inp,$idi));	# $inp to point at the end
+	&neg	($idi);
+	&xor	("eax","eax");
+	&mov	(&DWP(-4,$out),$idi);		# borrow key->y
+
+	&bt	(&DWP(0,$idx),20);		# check for bit#20
+	&jc	(&label("c1stloop"));
+
+&set_label("w1stloop",16);
+	&mov	(&DWP(0,$out,"eax",4),"eax");	# key->data[i]=i;
+	&add	(&LB("eax"),1);			# i++;
+	&jnc	(&label("w1stloop"));
+
+	&xor	($ido,$ido);
+	&xor	($idx,$idx);
+
+&set_label("w2ndloop",16);
+	&mov	("eax",&DWP(0,$out,$ido,4));
+	&add	(&LB($idx),&BP(0,$inp,$idi));
+	&add	(&LB($idx),&LB("eax"));
+	&add	($idi,1);
+	&mov	("ebx",&DWP(0,$out,$idx,4));
+	&jnz	(&label("wnowrap"));
+	  &mov	($idi,&DWP(-4,$out));
+	&set_label("wnowrap");
+	&mov	(&DWP(0,$out,$idx,4),"eax");
+	&mov	(&DWP(0,$out,$ido,4),"ebx");
+	&add	(&LB($ido),1);
+	&jnc	(&label("w2ndloop"));
+&jmp	(&label("exit"));
+
+# Unlike all other x86 [and x86_64] implementations, Intel P4 core
+# [including EM64T] was found to perform poorly with above "32-bit" key
+# schedule, a.k.a. RC4_INT. Performance improvement for IA-32 hand-coded
+# assembler turned out to be 3.5x if re-coded for compressed 8-bit one,
+# a.k.a. RC4_CHAR! It's however inappropriate to just switch to 8-bit
+# schedule for x86[_64], because non-P4 implementations suffer from
+# significant performance losses then, e.g. PIII exhibits >2x
+# deterioration, and so does Opteron. In order to assure optimal
+# all-round performance, we detect P4 at run-time and set up compressed
+# key schedule, which is recognized by RC4 procedure.
+
+&set_label("c1stloop",16);
+	&mov	(&BP(0,$out,"eax"),&LB("eax"));	# key->data[i]=i;
+	&add	(&LB("eax"),1);			# i++;
+	&jnc	(&label("c1stloop"));
+
+	&xor	($ido,$ido);
+	&xor	($idx,$idx);
+	&xor	("ebx","ebx");
+
+&set_label("c2ndloop",16);
+	&mov	(&LB("eax"),&BP(0,$out,$ido));
+	&add	(&LB($idx),&BP(0,$inp,$idi));
+	&add	(&LB($idx),&LB("eax"));
+	&add	($idi,1);
+	&mov	(&LB("ebx"),&BP(0,$out,$idx));
+	&jnz	(&label("cnowrap"));
+	  &mov	($idi,&DWP(-4,$out));
+	&set_label("cnowrap");
+	&mov	(&BP(0,$out,$idx),&LB("eax"));
+	&mov	(&BP(0,$out,$ido),&LB("ebx"));
+	&add	(&LB($ido),1);
+	&jnc	(&label("c2ndloop"));
+
+	&mov	(&DWP(256,$out),-1);		# mark schedule as compressed
+
+&set_label("exit");
+	&xor	("eax","eax");
+	&mov	(&DWP(-8,$out),"eax");		# key->x=0;
+	&mov	(&DWP(-4,$out),"eax");		# key->y=0;
+&function_end("RC4_set_key");
+
+# const char *RC4_options(void);
+&function_begin_B("RC4_options");
+	&call	(&label("pic_point"));
+&set_label("pic_point");
+	&blindpop("eax");
+	&lea	("eax",&DWP(&label("opts")."-".&label("pic_point"),"eax"));
+	&picmeup("edx","OPENSSL_ia32cap_P");
+	&bt	(&DWP(0,"edx"),20);
+	&jnc	(&label("skip"));
+	  &add	("eax",12);
+	&set_label("skip");
+	&ret	();
+&set_label("opts",64);
+&asciz	("rc4(4x,int)");
+&asciz	("rc4(1x,char)");
+&asciz	("RC4 for x86, CRYPTOGAMS by <appro\@openssl.org>");
+&align	(64);
+&function_end_B("RC4_options");
+
+&asm_finish();
 
diff --git a/openssl/crypto/rc4/asm/rc4-ia64.S b/openssl/crypto/rc4/asm/rc4-ia64.S
deleted file mode 100644
index 8210c47d0..000000000
--- a/openssl/crypto/rc4/asm/rc4-ia64.S
+++ /dev/null
@@ -1,159 +0,0 @@
-// ====================================================================
-// Written by Andy Polyakov <appro@fy.chalmers.se> for the OpenSSL
-// project.
-//
-// Rights for redistribution and usage in source and binary forms are
-// granted according to the OpenSSL license. Warranty of any kind is
-// disclaimed.
-// ====================================================================
-
-.ident  "rc4-ia64.S, Version 2.0"
-.ident  "IA-64 ISA artwork by Andy Polyakov <appro@fy.chalmers.se>"
-
-// What's wrong with compiler generated code? Because of the nature of
-// C language, compiler doesn't [dare to] reorder load and stores. But
-// being memory-bound, RC4 should benefit from reorder [on in-order-
-// execution core such as IA-64]. But what can we reorder? At the very
-// least we can safely reorder references to key schedule in respect
-// to input and output streams. Secondly, from the first [close] glance
-// it appeared that it's possible to pull up some references to
-// elements of the key schedule itself. Original rationale ["prior
-// loads are not safe only for "degenerated" key schedule, when some
-// elements equal to the same value"] was kind of sloppy. I should have
-// formulated as it really was: if we assume that pulling up reference
-// to key[x+1] is not safe, then it would mean that key schedule would
-// "degenerate," which is never the case. The problem is that this
-// holds true in respect to references to key[x], but not to key[y].
-// Legitimate "collisions" do occur within every 256^2 bytes window.
-// Fortunately there're enough free instruction slots to keep prior
-// reference to key[x+1], detect "collision" and compensate for it.
-// All this without sacrificing a single clock cycle:-) Throughput is
-// ~210MBps on 900MHz CPU, which is is >3x faster than gcc generated
-// code and +30% - if compared to HP-UX C. Unrolling loop below should
-// give >30% on top of that...
-
-.text
-.explicit
-
-#if defined(_HPUX_SOURCE) && !defined(_LP64)
-# define ADDP	addp4
-#else
-# define ADDP	add
-#endif
-
-#ifndef SZ
-#define SZ	4	// this is set to sizeof(RC4_INT)
-#endif
-// SZ==4 seems to be optimal. At least SZ==8 is not any faster, not for
-// assembler implementation, while SZ==1 code is ~30% slower.
-#if SZ==1	// RC4_INT is unsigned char
-# define	LDKEY	ld1
-# define	STKEY	st1
-# define	OFF	0
-#elif SZ==4	// RC4_INT is unsigned int
-# define	LDKEY	ld4
-# define	STKEY	st4
-# define	OFF	2
-#elif SZ==8	// RC4_INT is unsigned long
-# define	LDKEY	ld8
-# define	STKEY	st8
-# define	OFF	3
-#endif
-
-out=r8;		// [expanded] output pointer
-inp=r9;		// [expanded] output pointer
-prsave=r10;
-key=r28;	// [expanded] pointer to RC4_KEY
-ksch=r29;	// (key->data+255)[&~(sizeof(key->data)-1)]
-xx=r30;
-yy=r31;
-
-// void RC4(RC4_KEY *key,size_t len,const void *inp,void *out);
-.global	RC4#
-.proc	RC4#
-.align	32
-.skip	16
-RC4:
-	.prologue
-	.save   ar.pfs,r2
-{ .mii;	alloc	r2=ar.pfs,4,12,0,16
-	.save	pr,prsave
-	mov	prsave=pr
-	ADDP	key=0,in0		};;
-{ .mib;	cmp.eq	p6,p0=0,in1			// len==0?
-	.save	ar.lc,r3
-	mov	r3=ar.lc
-(p6)	br.ret.spnt.many	b0	};;	// emergency exit
-
-	.body
-	.rotr	dat[4],key_x[4],tx[2],rnd[2],key_y[2],ty[1];
-
-{ .mib;	LDKEY	xx=[key],SZ			// load key->x
-	add	in1=-1,in1			// adjust len for loop counter
-	nop.b	0			}
-{ .mib;	ADDP	inp=0,in2
-	ADDP	out=0,in3
-	brp.loop.imp	.Ltop,.Lexit-16	};;
-{ .mmi;	LDKEY	yy=[key]			// load key->y
-	add	ksch=SZ,key
-	mov	ar.lc=in1		}
-{ .mmi;	mov	key_y[1]=r0			// guarantee inequality
-						// in first iteration
-	add	xx=1,xx
-	mov	pr.rot=1<<16		};;
-{ .mii;	nop.m	0
-	dep	key_x[1]=xx,r0,OFF,8
-	mov	ar.ec=3			};;	// note that epilogue counter
-						// is off by 1. I compensate
-						// for this at exit...
-.Ltop:
-// The loop is scheduled for 4*(n+2) spin-rate on Itanium 2, which
-// theoretically gives asymptotic performance of clock frequency
-// divided by 4 bytes per seconds, or 400MBps on 1.6GHz CPU. This is
-// for sizeof(RC4_INT)==4. For smaller RC4_INT STKEY inadvertently
-// splits the last bundle and you end up with 5*n spin-rate:-(
-// Originally the loop was scheduled for 3*n and relied on key
-// schedule to be aligned at 256*sizeof(RC4_INT) boundary. But
-// *(out++)=dat, which maps to st1, had same effect [inadvertent
-// bundle split] and holded the loop back. Rescheduling for 4*n
-// made it possible to eliminate dependence on specific alignment
-// and allow OpenSSH keep "abusing" our API. Reaching for 3*n would
-// require unrolling, sticking to variable shift instruction for
-// collecting output [to avoid starvation for integer shifter] and
-// copying of key schedule to controlled place in stack [so that
-// deposit instruction can serve as substitute for whole
-// key->data+((x&255)<<log2(sizeof(key->data[0])))]...
-{ .mmi;	(p19)	st1	[out]=dat[3],1			// *(out++)=dat
-	(p16)	add	xx=1,xx				// x++
-	(p18)	dep	rnd[1]=rnd[1],r0,OFF,8	}	// ((tx+ty)&255)<<OFF
-{ .mmi;	(p16)	add	key_x[1]=ksch,key_x[1]		// &key[xx&255]
-	(p17)	add	key_y[1]=ksch,key_y[1]	};;	// &key[yy&255]	
-{ .mmi;	(p16)	LDKEY	tx[0]=[key_x[1]]		// tx=key[xx]
-	(p17)	LDKEY	ty[0]=[key_y[1]]		// ty=key[yy]	
-	(p16)	dep	key_x[0]=xx,r0,OFF,8	}	// (xx&255)<<OFF
-{ .mmi;	(p18)	add	rnd[1]=ksch,rnd[1]		// &key[(tx+ty)&255]
-	(p16)	cmp.ne.unc p20,p21=key_x[1],key_y[1] };;
-{ .mmi;	(p18)	LDKEY	rnd[1]=[rnd[1]]			// rnd=key[(tx+ty)&255]
-	(p16)	ld1	dat[0]=[inp],1		}	// dat=*(inp++)
-.pred.rel	"mutex",p20,p21
-{ .mmi;	(p21)	add	yy=yy,tx[1]			// (p16)
-	(p20)	add	yy=yy,tx[0]			// (p16) y+=tx
-	(p21)	mov	tx[0]=tx[1]		};;	// (p16)
-{ .mmi;	(p17)	STKEY	[key_y[1]]=tx[1]		// key[yy]=tx
-	(p17)	STKEY	[key_x[2]]=ty[0]		// key[xx]=ty
-	(p16)	dep	key_y[0]=yy,r0,OFF,8	}	// &key[yy&255]
-{ .mmb;	(p17)	add	rnd[0]=tx[1],ty[0]		// tx+=ty
-	(p18)	xor	dat[2]=dat[2],rnd[1]		// dat^=rnd
-	br.ctop.sptk	.Ltop			};;
-.Lexit:
-{ .mib;	STKEY	[key]=yy,-SZ			// save key->y
-	mov	pr=prsave,0x1ffff
-	nop.b	0			}
-{ .mib;	st1	[out]=dat[3],1			// compensate for truncated
-						// epilogue counter
-	add	xx=-1,xx
-	nop.b	0			};;
-{ .mib;	STKEY	[key]=xx			// save key->x
-	mov	ar.lc=r3
-	br.ret.sptk.many	b0	};;
-.endp	RC4#
diff --git a/openssl/crypto/rc4/asm/rc4-ia64.pl b/openssl/crypto/rc4/asm/rc4-ia64.pl
new file mode 100644
index 000000000..49cd5b5e6
--- /dev/null
+++ b/openssl/crypto/rc4/asm/rc4-ia64.pl
@@ -0,0 +1,755 @@
+#!/usr/bin/env perl
+#
+# ====================================================================
+# Written by David Mosberger <David.Mosberger@acm.org> based on the
+# Itanium optimized Crypto code which was released by HP Labs at
+# http://www.hpl.hp.com/research/linux/crypto/.
+#
+# Copyright (c) 2005 Hewlett-Packard Development Company, L.P.
+#
+# Permission is hereby granted, free of charge, to any person obtaining
+# a copy of this software and associated documentation files (the
+# "Software"), to deal in the Software without restriction, including
+# without limitation the rights to use, copy, modify, merge, publish,
+# distribute, sublicense, and/or sell copies of the Software, and to
+# permit persons to whom the Software is furnished to do so, subject to
+# the following conditions:
+#
+# The above copyright notice and this permission notice shall be
+# included in all copies or substantial portions of the Software.
+
+# THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
+# EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
+# MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
+# NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS BE
+# LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION
+# OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION
+# WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.  */
+
+
+
+# This is a little helper program which generates a software-pipelined
+# for RC4 encryption.  The basic algorithm looks like this:
+#
+#   for (counter = 0; counter < len; ++counter)
+#     {
+#       in = inp[counter];
+#       SI = S[I];
+#       J = (SI + J) & 0xff;
+#       SJ = S[J];
+#       T = (SI + SJ) & 0xff;
+#       S[I] = SJ, S[J] = SI;
+#       ST = S[T];
+#       outp[counter] = in ^ ST;
+#       I = (I + 1) & 0xff;
+#     }
+#
+# Pipelining this loop isn't easy, because the stores to the S[] array
+# need to be observed in the right order.  The loop generated by the
+# code below has the following pipeline diagram:
+#
+#      cycle
+#     | 0 | 1 | 2 | 3 | 4 | 5 | 6 | 7 | 8 | 9 |10 |11 |12 |13 |14 |15 |16 |17 |
+# iter
+#   1: xxx LDI xxx xxx xxx LDJ xxx SWP xxx LDT xxx xxx
+#   2:             xxx LDI xxx xxx xxx LDJ xxx SWP xxx LDT xxx xxx
+#   3:                         xxx LDI xxx xxx xxx LDJ xxx SWP xxx LDT xxx xxx
+#
+#   where:
+# 	LDI = load of S[I]
+# 	LDJ = load of S[J]
+# 	SWP = swap of S[I] and S[J]
+# 	LDT = load of S[T]
+#
+# Note that in the above diagram, the major trouble-spot is that LDI
+# of the 2nd iteration is performed BEFORE the SWP of the first
+# iteration.  Fortunately, this is easy to detect (I of the 1st
+# iteration will be equal to J of the 2nd iteration) and when this
+# happens, we simply forward the proper value from the 1st iteration
+# to the 2nd one.  The proper value in this case is simply the value
+# of S[I] from the first iteration (thanks to the fact that SWP
+# simply swaps the contents of S[I] and S[J]).
+#
+# Another potential trouble-spot is in cycle 7, where SWP of the 1st
+# iteration issues at the same time as the LDI of the 3rd iteration.
+# However, thanks to IA-64 execution semantics, this can be taken
+# care of simply by placing LDI later in the instruction-group than
+# SWP.  IA-64 CPUs will automatically forward the value if they
+# detect that the SWP and LDI are accessing the same memory-location.
+
+# The core-loop that can be pipelined then looks like this (annotated
+# with McKinley/Madison issue port & latency numbers, assuming L1
+# cache hits for the most part):
+
+# operation:	    instruction:		    issue-ports:  latency
+# ------------------  -----------------------------   ------------- -------
+
+# Data = *inp++       ld1 data = [inp], 1             M0-M1         1 cyc     c0
+#                     shladd Iptr = I, KeyTable, 3    M0-M3, I0, I1 1 cyc
+# I = (I + 1) & 0xff  padd1 nextI = I, one            M0-M3, I0, I1 3 cyc
+#                     ;;
+# SI = S[I]           ld8 SI = [Iptr]                 M0-M1         1 cyc     c1 * after SWAP!
+#                     ;;
+#                     cmp.eq.unc pBypass = I, J                                  * after J is valid!
+# J = SI + J          add J = J, SI                   M0-M3, I0, I1 1 cyc     c2
+#                     (pBypass) br.cond.spnt Bypass
+#                     ;;
+# ---------------------------------------------------------------------------------------
+# J = J & 0xff        zxt1 J = J                      I0, I1, 1 cyc           c3
+#                     ;;
+#                     shladd Jptr = J, KeyTable, 3    M0-M3, I0, I1 1 cyc     c4
+#                     ;;
+# SJ = S[J]           ld8 SJ = [Jptr]                 M0-M1         1 cyc     c5
+#                     ;;
+# ---------------------------------------------------------------------------------------
+# T = (SI + SJ)       add T = SI, SJ                  M0-M3, I0, I1 1 cyc     c6
+#                     ;;
+# T = T & 0xff        zxt1 T = T                      I0, I1        1 cyc
+# S[I] = SJ           st8 [Iptr] = SJ                 M2-M3                   c7
+# S[J] = SI           st8 [Jptr] = SI                 M2-M3
+#                     ;;
+#                     shladd Tptr = T, KeyTable, 3    M0-M3, I0, I1 1 cyc     c8
+#                     ;;
+# ---------------------------------------------------------------------------------------
+# T = S[T]            ld8 T = [Tptr]                  M0-M1         1 cyc     c9
+#                     ;;
+# data ^= T           xor data = data, T              M0-M3, I0, I1 1 cyc     c10
+#                     ;;
+# *out++ = Data ^ T   dep word = word, data, 8, POS   I0, I1        1 cyc     c11
+#                     ;;
+# ---------------------------------------------------------------------------------------
+
+# There are several points worth making here:
+
+#   - Note that due to the bypass/forwarding-path, the first two
+#     phases of the loop are strangly mingled together.  In
+#     particular, note that the first stage of the pipeline is
+#     using the value of "J", as calculated by the second stage.
+#   - Each bundle-pair will have exactly 6 instructions.
+#   - Pipelined, the loop can execute in 3 cycles/iteration and
+#     4 stages.  However, McKinley/Madison can issue "st1" to
+#     the same bank at a rate of at most one per 4 cycles.  Thus,
+#     instead of storing each byte, we accumulate them in a word
+#     and then write them back at once with a single "st8" (this
+#     implies that the setup code needs to ensure that the output
+#     buffer is properly aligned, if need be, by encoding the
+#     first few bytes separately).
+#   - There is no space for a "br.ctop" instruction.  For this
+#     reason we can't use module-loop support in IA-64 and have
+#     to do a traditional, purely software-pipelined loop.
+#   - We can't replace any of the remaining "add/zxt1" pairs with
+#     "padd1" because the latency for that instruction is too high
+#     and would push the loop to the point where more bypasses
+#     would be needed, which we don't have space for.
+#   - The above loop runs at around 3.26 cycles/byte, or roughly
+#     440 MByte/sec on a 1.5GHz Madison.  This is well below the
+#     system bus bandwidth and hence with judicious use of
+#     "lfetch" this loop can run at (almost) peak speed even when
+#     the input and output data reside in memory.  The
+#     max. latency that can be tolerated is (PREFETCH_DISTANCE *
+#     L2_LINE_SIZE * 3 cyc), or about 384 cycles assuming (at
+#     least) 1-ahead prefetching of 128 byte cache-lines.  Note
+#     that we do NOT prefetch into L1, since that would only
+#     interfere with the S[] table values stored there.  This is
+#     acceptable because there is a 10 cycle latency between
+#     load and first use of the input data.
+#   - We use a branch to out-of-line bypass-code of cycle-pressure:
+#     we calculate the next J, check for the need to activate the
+#     bypass path, and activate the bypass path ALL IN THE SAME
+#     CYCLE.  If we didn't have these constraints, we could do
+#     the bypass with a simple conditional move instruction.
+#     Fortunately, the bypass paths get activated relatively
+#     infrequently, so the extra branches don't cost all that much
+#     (about 0.04 cycles/byte, measured on a 16396 byte file with
+#     random input data).
+#
+
+$phases = 4;		# number of stages/phases in the pipelined-loop
+$unroll_count = 6;	# number of times we unrolled it
+$pComI = (1 << 0);
+$pComJ = (1 << 1);
+$pComT = (1 << 2);
+$pOut  = (1 << 3);
+
+$NData = 4;
+$NIP = 3;
+$NJP = 2;
+$NI = 2;
+$NSI = 3;
+$NSJ = 2;
+$NT = 2;
+$NOutWord = 2;
+
+#
+# $threshold is the minimum length before we attempt to use the
+# big software-pipelined loop.  It MUST be greater-or-equal
+# to:
+#  		PHASES * (UNROLL_COUNT + 1) + 7
+#
+# The "+ 7" comes from the fact we may have to encode up to
+#   7 bytes separately before the output pointer is aligned.
+#
+$threshold = (3 * ($phases * ($unroll_count + 1)) + 7);
+
+sub I {
+    local *code = shift;
+    local $format = shift;
+    $code .= sprintf ("\t\t".$format."\n", @_);
+}
+
+sub P {
+    local *code = shift;
+    local $format = shift;
+    $code .= sprintf ($format."\n", @_);
+}
+
+sub STOP {
+    local *code = shift;
+    $code .=<<___;
+		;;
+___
+}
+
+sub emit_body {
+    local *c = shift;
+    local *bypass = shift;
+    local ($iteration, $p) = @_;
+
+    local $i0 = $iteration;
+    local $i1 = $iteration - 1;
+    local $i2 = $iteration - 2;
+    local $i3 = $iteration - 3;
+    local $iw0 = ($iteration - 3) / 8;
+    local $iw1 = ($iteration > 3) ? ($iteration - 4) / 8 : 1;
+    local $byte_num = ($iteration - 3) % 8;
+    local $label = $iteration + 1;
+    local $pAny = ($p & 0xf) == 0xf;
+    local $pByp = (($p & $pComI) && ($iteration > 0));
+
+    $c.=<<___;
+//////////////////////////////////////////////////
+___
+
+    if (($p & 0xf) == 0) {
+	$c.="#ifdef HOST_IS_BIG_ENDIAN\n";
+	&I(\$c,"shr.u	OutWord[%u] = OutWord[%u], 32;;",
+				$iw1 % $NOutWord, $iw1 % $NOutWord);
+	$c.="#endif\n";
+	&I(\$c, "st4 [OutPtr] = OutWord[%u], 4", $iw1 % $NOutWord);
+	return;
+    }
+
+    # Cycle 0
+    &I(\$c, "{ .mmi")					      if ($pAny);
+    &I(\$c, "ld1    Data[%u] = [InPtr], 1", $i0 % $NData)     if ($p & $pComI);
+    &I(\$c, "padd1  I[%u] = One, I[%u]", $i0 % $NI, $i1 % $NI)if ($p & $pComI);
+    &I(\$c, "zxt1   J = J")				      if ($p & $pComJ);
+    &I(\$c, "}")					      if ($pAny);
+    &I(\$c, "{ .mmi")					      if ($pAny);
+    &I(\$c, "LKEY   T[%u] = [T[%u]]", $i1 % $NT, $i1 % $NT)   if ($p & $pOut);
+    &I(\$c, "add    T[%u] = SI[%u], SJ[%u]",
+       $i0 % $NT, $i2 % $NSI, $i1 % $NSJ)		      if ($p & $pComT);
+    &I(\$c, "KEYADDR(IPr[%u], I[%u])", $i0 % $NIP, $i1 % $NI) if ($p & $pComI);
+    &I(\$c, "}")					      if ($pAny);
+    &STOP(\$c);
+
+    # Cycle 1
+    &I(\$c, "{ .mmi")					      if ($pAny);
+    &I(\$c, "SKEY   [IPr[%u]] = SJ[%u]", $i2 % $NIP, $i1%$NSJ)if ($p & $pComT);
+    &I(\$c, "SKEY   [JP[%u]] = SI[%u]", $i1 % $NJP, $i2%$NSI) if ($p & $pComT);
+    &I(\$c, "zxt1   T[%u] = T[%u]", $i0 % $NT, $i0 % $NT)     if ($p & $pComT);
+    &I(\$c, "}")					      if ($pAny);
+    &I(\$c, "{ .mmi")					      if ($pAny);
+    &I(\$c, "LKEY   SI[%u] = [IPr[%u]]", $i0 % $NSI, $i0%$NIP)if ($p & $pComI);
+    &I(\$c, "KEYADDR(JP[%u], J)", $i0 % $NJP)		      if ($p & $pComJ);
+    &I(\$c, "xor    Data[%u] = Data[%u], T[%u]",
+       $i3 % $NData, $i3 % $NData, $i1 % $NT)		      if ($p & $pOut);
+    &I(\$c, "}")					      if ($pAny);
+    &STOP(\$c);
+
+    # Cycle 2
+    &I(\$c, "{ .mmi")					      if ($pAny);
+    &I(\$c, "LKEY   SJ[%u] = [JP[%u]]", $i0 % $NSJ, $i0%$NJP) if ($p & $pComJ);
+    &I(\$c, "cmp.eq pBypass, p0 = I[%u], J", $i1 % $NI)	      if ($pByp);
+    &I(\$c, "dep OutWord[%u] = Data[%u], OutWord[%u], BYTE_POS(%u), 8",
+       $iw0%$NOutWord, $i3%$NData, $iw1%$NOutWord, $byte_num) if ($p & $pOut);
+    &I(\$c, "}")					      if ($pAny);
+    &I(\$c, "{ .mmb")					      if ($pAny);
+    &I(\$c, "add    J = J, SI[%u]", $i0 % $NSI)		      if ($p & $pComI);
+    &I(\$c, "KEYADDR(T[%u], T[%u])", $i0 % $NT, $i0 % $NT)    if ($p & $pComT);
+    &P(\$c, "(pBypass)\tbr.cond.spnt.many .rc4Bypass%u",$label)if ($pByp);
+    &I(\$c, "}") if ($pAny);
+    &STOP(\$c);
+
+    &P(\$c, ".rc4Resume%u:", $label)			      if ($pByp);
+    if ($byte_num == 0 && $iteration >= $phases) {
+	&I(\$c, "st8 [OutPtr] = OutWord[%u], 8",
+	   $iw1 % $NOutWord)				      if ($p & $pOut);
+	if ($iteration == (1 + $unroll_count) * $phases - 1) {
+	    if ($unroll_count == 6) {
+		&I(\$c, "mov OutWord[%u] = OutWord[%u]",
+		   $iw1 % $NOutWord, $iw0 % $NOutWord);
+	    }
+	    &I(\$c, "lfetch.nt1 [InPrefetch], %u",
+	       $unroll_count * $phases);
+	    &I(\$c, "lfetch.excl.nt1 [OutPrefetch], %u",
+	       $unroll_count * $phases);
+	    &I(\$c, "br.cloop.sptk.few .rc4Loop");
+	}
+    }
+
+    if ($pByp) {
+	&P(\$bypass, ".rc4Bypass%u:", $label);
+	&I(\$bypass, "sub J = J, SI[%u]", $i0 % $NSI);
+	&I(\$bypass, "nop 0");
+	&I(\$bypass, "nop 0");
+	&I(\$bypass, ";;");
+	&I(\$bypass, "add J = J, SI[%u]", $i1 % $NSI);
+	&I(\$bypass, "mov SI[%u] = SI[%u]", $i0 % $NSI, $i1 % $NSI);
+	&I(\$bypass, "br.sptk.many .rc4Resume%u\n", $label);
+	&I(\$bypass, ";;");
+    }
+}
+
+$code=<<___;
+.ident \"rc4-ia64.s, version 3.0\"
+.ident \"Copyright (c) 2005 Hewlett-Packard Development Company, L.P.\"
+
+#define LCSave		r8
+#define PRSave		r9
+
+/* Inputs become invalid once rotation begins!  */
+
+#define StateTable	in0
+#define DataLen		in1
+#define InputBuffer	in2
+#define OutputBuffer	in3
+
+#define KTable		r14
+#define J		r15
+#define InPtr		r16
+#define OutPtr		r17
+#define InPrefetch	r18
+#define OutPrefetch	r19
+#define One		r20
+#define LoopCount	r21
+#define Remainder	r22
+#define IFinal		r23
+#define EndPtr		r24
+
+#define tmp0		r25
+#define tmp1		r26
+
+#define pBypass		p6
+#define pDone		p7
+#define pSmall		p8
+#define pAligned	p9
+#define pUnaligned	p10
+
+#define pComputeI	pPhase[0]
+#define pComputeJ	pPhase[1]
+#define pComputeT	pPhase[2]
+#define pOutput		pPhase[3]
+
+#define RetVal		r8
+#define L_OK		p7
+#define L_NOK		p8
+
+#define	_NINPUTS	4
+#define	_NOUTPUT	0
+
+#define	_NROTATE	24
+#define	_NLOCALS	(_NROTATE - _NINPUTS - _NOUTPUT)
+
+#ifndef SZ
+# define SZ	4	// this must be set to sizeof(RC4_INT)
+#endif
+
+#if SZ == 1
+# define LKEY			ld1
+# define SKEY			st1
+# define KEYADDR(dst, i)	add dst = i, KTable
+#elif SZ == 2
+# define LKEY			ld2
+# define SKEY			st2
+# define KEYADDR(dst, i)	shladd dst = i, 1, KTable
+#elif SZ == 4
+# define LKEY			ld4
+# define SKEY			st4
+# define KEYADDR(dst, i)	shladd dst = i, 2, KTable
+#else
+# define LKEY			ld8
+# define SKEY			st8
+# define KEYADDR(dst, i)	shladd dst = i, 3, KTable
+#endif
+
+#if defined(_HPUX_SOURCE) && !defined(_LP64)
+# define ADDP	addp4
+#else
+# define ADDP	add
+#endif
+
+/* Define a macro for the bit number of the n-th byte: */
+
+#if defined(_HPUX_SOURCE) || defined(B_ENDIAN)
+# define HOST_IS_BIG_ENDIAN
+# define BYTE_POS(n)	(56 - (8 * (n)))
+#else
+# define BYTE_POS(n)	(8 * (n))
+#endif
+
+/*
+   We must perform the first phase of the pipeline explicitly since
+   we will always load from the stable the first time. The br.cexit
+   will never be taken since regardless of the number of bytes because
+   the epilogue count is 4.
+*/
+/* MODSCHED_RC4 macro was split to _PROLOGUE and _LOOP, because HP-UX
+   assembler failed on original macro with syntax error. <appro> */
+#define MODSCHED_RC4_PROLOGUE						   \\
+	{								   \\
+				ld1		Data[0] = [InPtr], 1;	   \\
+				add		IFinal = 1, I[1];	   \\
+				KEYADDR(IPr[0], I[1]);			   \\
+	} ;;								   \\
+	{								   \\
+				LKEY		SI[0] = [IPr[0]];	   \\
+				mov		pr.rot = 0x10000;	   \\
+				mov		ar.ec = 4;		   \\
+	} ;;								   \\
+	{								   \\
+				add		J = J, SI[0];		   \\
+				zxt1		I[0] = IFinal;		   \\
+				br.cexit.spnt.few .+16; /* never taken */  \\
+	} ;;
+#define MODSCHED_RC4_LOOP(label)					   \\
+label:									   \\
+	{	.mmi;							   \\
+		(pComputeI)	ld1		Data[0] = [InPtr], 1;	   \\
+		(pComputeI)	add		IFinal = 1, I[1];	   \\
+		(pComputeJ)	zxt1		J = J;			   \\
+	}{	.mmi;							   \\
+		(pOutput)	LKEY		T[1] = [T[1]];		   \\
+		(pComputeT)	add		T[0] = SI[2], SJ[1];	   \\
+		(pComputeI)	KEYADDR(IPr[0], I[1]);			   \\
+	} ;;								   \\
+	{	.mmi;							   \\
+		(pComputeT)	SKEY		[IPr[2]] = SJ[1];	   \\
+		(pComputeT)	SKEY		[JP[1]] = SI[2];	   \\
+		(pComputeT)	zxt1		T[0] = T[0];		   \\
+	}{	.mmi;							   \\
+		(pComputeI)	LKEY		SI[0] = [IPr[0]];	   \\
+		(pComputeJ)	KEYADDR(JP[0], J);			   \\
+		(pComputeI)	cmp.eq.unc	pBypass, p0 = I[1], J;	   \\
+	} ;;								   \\
+	{	.mmi;							   \\
+		(pComputeJ)	LKEY		SJ[0] = [JP[0]];	   \\
+		(pOutput)	xor		Data[3] = Data[3], T[1];   \\
+				nop		0x0;			   \\
+	}{	.mmi;							   \\
+		(pComputeT)	KEYADDR(T[0], T[0]);			   \\
+		(pBypass)	mov		SI[0] = SI[1];		   \\
+		(pComputeI)	zxt1		I[0] = IFinal;		   \\
+	} ;;								   \\
+	{	.mmb;							   \\
+		(pOutput)	st1		[OutPtr] = Data[3], 1;	   \\
+		(pComputeI)	add		J = J, SI[0];		   \\
+				br.ctop.sptk.few label;			   \\
+	} ;;
+
+	.text
+
+	.align	32
+
+	.type	RC4, \@function
+	.global	RC4
+
+	.proc	RC4
+	.prologue
+
+RC4:
+	{
+	  	.mmi
+		alloc	r2 = ar.pfs, _NINPUTS, _NLOCALS, _NOUTPUT, _NROTATE
+
+		.rotr Data[4], I[2], IPr[3], SI[3], JP[2], SJ[2], T[2], \\
+		      OutWord[2]
+		.rotp pPhase[4]
+
+		ADDP		InPrefetch = 0, InputBuffer
+		ADDP		KTable = 0, StateTable
+	}
+	{
+		.mmi
+		ADDP		InPtr = 0, InputBuffer
+		ADDP		OutPtr = 0, OutputBuffer
+		mov		RetVal = r0
+	}
+	;;
+	{
+		.mmi
+		lfetch.nt1	[InPrefetch], 0x80
+		ADDP		OutPrefetch = 0, OutputBuffer
+	}
+	{               // Return 0 if the input length is nonsensical
+        	.mib
+		ADDP		StateTable = 0, StateTable
+        	cmp.ge.unc  	L_NOK, L_OK = r0, DataLen
+	(L_NOK) br.ret.sptk.few rp
+	}
+	;;
+	{
+        	.mib
+        	cmp.eq.or  	L_NOK, L_OK = r0, InPtr
+        	cmp.eq.or  	L_NOK, L_OK = r0, OutPtr
+		nop		0x0
+	}
+	{
+		.mib
+        	cmp.eq.or  	L_NOK, L_OK = r0, StateTable
+		nop		0x0
+	(L_NOK) br.ret.sptk.few rp
+	}
+	;;
+		LKEY		I[1] = [KTable], SZ
+/* Prefetch the state-table. It contains 256 elements of size SZ */
+
+#if SZ == 1
+		ADDP		tmp0 = 1*128, StateTable
+#elif SZ == 2
+		ADDP		tmp0 = 3*128, StateTable
+		ADDP		tmp1 = 2*128, StateTable
+#elif SZ == 4
+		ADDP		tmp0 = 7*128, StateTable
+		ADDP		tmp1 = 6*128, StateTable
+#elif SZ == 8
+		ADDP		tmp0 = 15*128, StateTable
+		ADDP		tmp1 = 14*128, StateTable
+#endif
+		;;
+#if SZ >= 8
+		lfetch.fault.nt1		[tmp0], -256	// 15
+		lfetch.fault.nt1		[tmp1], -256;;
+		lfetch.fault.nt1		[tmp0], -256	// 13
+		lfetch.fault.nt1		[tmp1], -256;;
+		lfetch.fault.nt1		[tmp0], -256	// 11
+		lfetch.fault.nt1		[tmp1], -256;;
+		lfetch.fault.nt1		[tmp0], -256	//  9
+		lfetch.fault.nt1		[tmp1], -256;;
+#endif
+#if SZ >= 4
+		lfetch.fault.nt1		[tmp0], -256	//  7
+		lfetch.fault.nt1		[tmp1], -256;;
+		lfetch.fault.nt1		[tmp0], -256	//  5
+		lfetch.fault.nt1		[tmp1], -256;;
+#endif
+#if SZ >= 2
+		lfetch.fault.nt1		[tmp0], -256	//  3
+		lfetch.fault.nt1		[tmp1], -256;;
+#endif
+	{
+		.mii
+		lfetch.fault.nt1		[tmp0]		//  1
+		add		I[1]=1,I[1];;
+		zxt1		I[1]=I[1]
+	}
+	{
+		.mmi
+		lfetch.nt1	[InPrefetch], 0x80
+		lfetch.excl.nt1	[OutPrefetch], 0x80
+		.save		pr, PRSave
+		mov		PRSave = pr
+	} ;;
+	{
+		.mmi
+		lfetch.excl.nt1	[OutPrefetch], 0x80
+		LKEY		J = [KTable], SZ
+		ADDP		EndPtr = DataLen, InPtr
+	}  ;;
+	{
+		.mmi
+		ADDP		EndPtr = -1, EndPtr	// Make it point to
+							// last data byte.
+		mov		One = 1
+		.save		ar.lc, LCSave
+		mov		LCSave = ar.lc
+		.body
+	} ;;
+	{
+		.mmb
+		sub		Remainder = 0, OutPtr
+		cmp.gtu		pSmall, p0 = $threshold, DataLen
+(pSmall)	br.cond.dpnt	.rc4Remainder		// Data too small for
+							// big loop.
+	} ;;
+	{
+		.mmi
+		and		Remainder = 0x7, Remainder
+		;;
+		cmp.eq		pAligned, pUnaligned = Remainder, r0
+		nop		0x0
+	} ;;
+	{
+		.mmb
+.pred.rel	"mutex",pUnaligned,pAligned
+(pUnaligned)	add		Remainder = -1, Remainder
+(pAligned)	sub		Remainder = EndPtr, InPtr
+(pAligned)	br.cond.dptk.many .rc4Aligned
+	} ;;
+	{
+		.mmi
+		nop		0x0
+		nop		0x0
+		mov.i		ar.lc = Remainder
+	}
+
+/* Do the initial few bytes via the compact, modulo-scheduled loop
+   until the output pointer is 8-byte-aligned.  */
+
+		MODSCHED_RC4_PROLOGUE
+		MODSCHED_RC4_LOOP(.RC4AlignLoop)
+
+	{
+		.mib
+		sub		Remainder = EndPtr, InPtr
+		zxt1		IFinal = IFinal
+		clrrrb				// Clear CFM.rrb.pr so
+		;;				// next "mov pr.rot = N"
+						// does the right thing.
+	}
+	{
+		.mmi
+		mov		I[1] = IFinal
+		nop		0x0
+		nop		0x0
+	} ;;
+
+
+.rc4Aligned:
+
+/*
+   Unrolled loop count = (Remainder - ($unroll_count+1)*$phases)/($unroll_count*$phases)
+ */
+
+	{
+		.mlx
+		add	LoopCount = 1 - ($unroll_count + 1)*$phases, Remainder
+		movl		Remainder = 0xaaaaaaaaaaaaaaab
+	} ;;
+	{
+		.mmi
+		setf.sig	f6 = LoopCount		// M2, M3	6 cyc
+		setf.sig	f7 = Remainder		// M2, M3	6 cyc
+		nop		0x0
+	} ;;
+	{
+		.mfb
+		nop		0x0
+		xmpy.hu		f6 = f6, f7
+		nop		0x0
+	} ;;
+	{
+		.mmi
+		getf.sig	LoopCount = f6;;	// M2		5 cyc
+		nop		0x0
+		shr.u		LoopCount = LoopCount, 4
+	} ;;
+	{
+		.mmi
+		nop		0x0
+		nop		0x0
+		mov.i		ar.lc = LoopCount
+	} ;;
+
+/* Now comes the unrolled loop: */
+
+.rc4Prologue:
+___
+
+$iteration = 0;
+
+# Generate the prologue:
+$predicates = 1;
+for ($i = 0; $i < $phases; ++$i) {
+    &emit_body (\$code, \$bypass, $iteration++, $predicates);
+    $predicates = ($predicates << 1) | 1;
+}
+
+$code.=<<___;
+.rc4Loop:
+___
+
+# Generate the body:
+for ($i = 0; $i < $unroll_count*$phases; ++$i) {
+    &emit_body (\$code, \$bypass, $iteration++, $predicates);
+}
+
+$code.=<<___;
+.rc4Epilogue:
+___
+
+# Generate the epilogue:
+for ($i = 0; $i < $phases; ++$i) {
+    $predicates <<= 1;
+    &emit_body (\$code, \$bypass, $iteration++, $predicates);
+}
+
+$code.=<<___;
+	{
+		.mmi
+		lfetch.nt1	[EndPtr]	// fetch line with last byte
+		mov		IFinal = I[1]
+		nop		0x0
+	}
+
+.rc4Remainder:
+	{
+		.mmi
+		sub		Remainder = EndPtr, InPtr	// Calculate
+								// # of bytes
+								// left - 1
+		nop		0x0
+		nop		0x0
+	} ;;
+	{
+		.mib
+		cmp.eq		pDone, p0 = -1, Remainder // done already?
+		mov.i		ar.lc = Remainder
+(pDone)		br.cond.dptk.few .rc4Complete
+	}
+
+/* Do the remaining bytes via the compact, modulo-scheduled loop */
+
+		MODSCHED_RC4_PROLOGUE
+		MODSCHED_RC4_LOOP(.RC4RestLoop)
+
+.rc4Complete:
+	{
+		.mmi
+		add		KTable = -SZ, KTable
+		add		IFinal = -1, IFinal
+		mov		ar.lc = LCSave
+	} ;;
+	{
+		.mii
+		SKEY		[KTable] = J,-SZ
+		zxt1		IFinal = IFinal
+		mov		pr = PRSave, 0x1FFFF
+	} ;;
+	{
+		.mib
+		SKEY		[KTable] = IFinal
+		add		RetVal = 1, r0
+		br.ret.sptk.few	rp
+	} ;;
+___
+
+# Last but not least, emit the code for the bypass-code of the unrolled loop:
+
+$code.=$bypass;
+
+$code.=<<___;
+	.endp RC4
+___
+
+print $code;
diff --git a/openssl/crypto/rc4/asm/rc4-s390x.pl b/openssl/crypto/rc4/asm/rc4-s390x.pl
new file mode 100644
index 000000000..96681fa05
--- /dev/null
+++ b/openssl/crypto/rc4/asm/rc4-s390x.pl
@@ -0,0 +1,205 @@
+#!/usr/bin/env perl
+#
+# ====================================================================
+# Written by Andy Polyakov <appro@fy.chalmers.se> for the OpenSSL
+# project. The module is, however, dual licensed under OpenSSL and
+# CRYPTOGAMS licenses depending on where you obtain it. For further
+# details see http://www.openssl.org/~appro/cryptogams/.
+# ====================================================================
+#
+# February 2009
+#
+# Performance is 2x of gcc 3.4.6 on z10. Coding "secret" is to
+# "cluster" Address Generation Interlocks, so that one pipeline stall
+# resolves several dependencies.
+
+$rp="%r14";
+$sp="%r15";
+$code=<<___;
+.text
+
+___
+
+# void RC4(RC4_KEY *key,size_t len,const void *inp,void *out)
+{
+$acc="%r0";
+$cnt="%r1";
+$key="%r2";
+$len="%r3";
+$inp="%r4";
+$out="%r5";
+
+@XX=("%r6","%r7");
+@TX=("%r8","%r9");
+$YY="%r10";
+$TY="%r11";
+
+$code.=<<___;
+.globl	RC4
+.type	RC4,\@function
+.align	64
+RC4:
+	stmg	%r6,%r11,48($sp)
+	llgc	$XX[0],0($key)
+	llgc	$YY,1($key)
+	la	$XX[0],1($XX[0])
+	nill	$XX[0],0xff
+	srlg	$cnt,$len,3
+	ltgr	$cnt,$cnt
+	llgc	$TX[0],2($XX[0],$key)
+	jz	.Lshort
+	j	.Loop8
+
+.align	64
+.Loop8:
+___
+for ($i=0;$i<8;$i++) {
+$code.=<<___;
+	la	$YY,0($YY,$TX[0])	# $i
+	nill	$YY,255
+	la	$XX[1],1($XX[0])
+	nill	$XX[1],255
+___
+$code.=<<___ if ($i==1);
+	llgc	$acc,2($TY,$key)
+___
+$code.=<<___ if ($i>1);
+	sllg	$acc,$acc,8
+	ic	$acc,2($TY,$key)
+___
+$code.=<<___;
+	llgc	$TY,2($YY,$key)
+	stc	$TX[0],2($YY,$key)
+	llgc	$TX[1],2($XX[1],$key)
+	stc	$TY,2($XX[0],$key)
+	cr	$XX[1],$YY
+	jne	.Lcmov$i
+	la	$TX[1],0($TX[0])
+.Lcmov$i:
+	la	$TY,0($TY,$TX[0])
+	nill	$TY,255
+___
+push(@TX,shift(@TX)); push(@XX,shift(@XX));     # "rotate" registers
+}
+
+$code.=<<___;
+	lg	$TX[1],0($inp)
+	sllg	$acc,$acc,8
+	la	$inp,8($inp)
+	ic	$acc,2($TY,$key)
+	xgr	$acc,$TX[1]
+	stg	$acc,0($out)
+	la	$out,8($out)
+	brct	$cnt,.Loop8
+
+.Lshort:
+	lghi	$acc,7
+	ngr	$len,$acc
+	jz	.Lexit
+	j	.Loop1
+
+.align	16
+.Loop1:
+	la	$YY,0($YY,$TX[0])
+	nill	$YY,255
+	llgc	$TY,2($YY,$key)
+	stc	$TX[0],2($YY,$key)
+	stc	$TY,2($XX[0],$key)
+	ar	$TY,$TX[0]
+	ahi	$XX[0],1
+	nill	$TY,255
+	nill	$XX[0],255
+	llgc	$acc,0($inp)
+	la	$inp,1($inp)
+	llgc	$TY,2($TY,$key)
+	llgc	$TX[0],2($XX[0],$key)
+	xr	$acc,$TY
+	stc	$acc,0($out)
+	la	$out,1($out)
+	brct	$len,.Loop1
+
+.Lexit:
+	ahi	$XX[0],-1
+	stc	$XX[0],0($key)
+	stc	$YY,1($key)
+	lmg	%r6,%r11,48($sp)
+	br	$rp
+.size	RC4,.-RC4
+.string	"RC4 for s390x, CRYPTOGAMS by <appro\@openssl.org>"
+
+___
+}
+
+# void RC4_set_key(RC4_KEY *key,unsigned int len,const void *inp)
+{
+$cnt="%r0";
+$idx="%r1";
+$key="%r2";
+$len="%r3";
+$inp="%r4";
+$acc="%r5";
+$dat="%r6";
+$ikey="%r7";
+$iinp="%r8";
+
+$code.=<<___;
+.globl	RC4_set_key
+.type	RC4_set_key,\@function
+.align	64
+RC4_set_key:
+	stmg	%r6,%r8,48($sp)
+	lhi	$cnt,256
+	la	$idx,0(%r0)
+	sth	$idx,0($key)
+.align	4
+.L1stloop:
+	stc	$idx,2($idx,$key)
+	la	$idx,1($idx)
+	brct	$cnt,.L1stloop
+
+	lghi	$ikey,-256
+	lr	$cnt,$len
+	la	$iinp,0(%r0)
+	la	$idx,0(%r0)
+.align	16
+.L2ndloop:
+	llgc	$acc,2+256($ikey,$key)
+	llgc	$dat,0($iinp,$inp)
+	la	$idx,0($idx,$acc)
+	la	$ikey,1($ikey)
+	la	$idx,0($idx,$dat)
+	nill	$idx,255
+	la	$iinp,1($iinp)
+	tml	$ikey,255
+	llgc	$dat,2($idx,$key)
+	stc	$dat,2+256-1($ikey,$key)
+	stc	$acc,2($idx,$key)
+	jz	.Ldone
+	brct	$cnt,.L2ndloop
+	lr	$cnt,$len
+	la	$iinp,0(%r0)
+	j	.L2ndloop
+.Ldone:
+	lmg	%r6,%r8,48($sp)
+	br	$rp
+.size	RC4_set_key,.-RC4_set_key
+
+___
+}
+
+# const char *RC4_options()
+$code.=<<___;
+.globl	RC4_options
+.type	RC4_options,\@function
+.align	16
+RC4_options:
+	larl	%r2,.Loptions
+	br	%r14
+.size	RC4_options,.-RC4_options
+.section	.rodata
+.Loptions:
+.align	8
+.string	"rc4(8x,char)"
+___
+
+print $code;
diff --git a/openssl/crypto/rc4/asm/rc4-x86_64.pl b/openssl/crypto/rc4/asm/rc4-x86_64.pl
index 00c6fa28a..677be5fe2 100644
--- a/openssl/crypto/rc4/asm/rc4-x86_64.pl
+++ b/openssl/crypto/rc4/asm/rc4-x86_64.pl
@@ -58,14 +58,18 @@
 # fit for Core2 and therefore the code was modified to skip cloop8 on
 # this CPU.
 
-$output=shift;
+$flavour = shift;
+$output  = shift;
+if ($flavour =~ /\./) { $output = $flavour; undef $flavour; }
+
+$win64=0; $win64=1 if ($flavour =~ /[nm]asm|mingw64/ || $output =~ /\.asm$/);
 
 $0 =~ m/(.*[\/\\])[^\/\\]+$/; $dir=$1;
 ( $xlate="${dir}x86_64-xlate.pl" and -f $xlate ) or
 ( $xlate="${dir}../../perlasm/x86_64-xlate.pl" and -f $xlate) or
 die "can't locate x86_64-xlate.pl";
 
-open STDOUT,"| $^X $xlate $output";
+open STDOUT,"| $^X $xlate $flavour $output";
 
 $dat="%rdi";	    # arg1
 $len="%rsi";	    # arg2
@@ -87,8 +91,10 @@ RC4:	or	$len,$len
 	jne	.Lentry
 	ret
 .Lentry:
+	push	%rbx
 	push	%r12
 	push	%r13
+.Lprologue:
 
 	add	\$8,$dat
 	movl	-8($dat),$XX[0]#d
@@ -133,16 +139,8 @@ $code.=<<___;
 	jnz	.Lloop8
 	cmp	\$0,$len
 	jne	.Lloop1
-___
-$code.=<<___;
-.Lexit:
-	sub	\$1,$XX[0]#b
-	movl	$XX[0]#d,-8($dat)
-	movl	$YY#d,-4($dat)
+	jmp	.Lexit
 
-	pop	%r13
-	pop	%r12
-	ret
 .align	16
 .Lloop1:
 	add	$TX[0]#b,$YY#b
@@ -167,9 +165,8 @@ $code.=<<___;
 	movzb	($dat,$XX[0]),$TX[0]#d
 	test	\$-8,$len
 	jz	.Lcloop1
-	cmp	\$0,260($dat)
+	cmpl	\$0,260($dat)
 	jnz	.Lcloop1
-	push	%rbx
 	jmp	.Lcloop8
 .align	16
 .Lcloop8:
@@ -224,7 +221,6 @@ $code.=<<___;
 
 	test	\$-8,$len
 	jnz	.Lcloop8
-	pop	%rbx
 	cmp	\$0,$len
 	jne	.Lcloop1
 	jmp	.Lexit
@@ -249,6 +245,19 @@ $code.=<<___;
 	sub	\$1,$len
 	jnz	.Lcloop1
 	jmp	.Lexit
+
+.align	16
+.Lexit:
+	sub	\$1,$XX[0]#b
+	movl	$XX[0]#d,-8($dat)
+	movl	$YY#d,-4($dat)
+
+	mov	(%rsp),%r13
+	mov	8(%rsp),%r12
+	mov	16(%rsp),%rbx
+	add	\$24,%rsp
+.Lepilogue:
+	ret
 .size	RC4,.-RC4
 ___
 
@@ -333,11 +342,10 @@ RC4_set_key:
 .size	RC4_set_key,.-RC4_set_key
 
 .globl	RC4_options
-.type	RC4_options,\@function,0
+.type	RC4_options,\@abi-omnipotent
 .align	16
 RC4_options:
-	.picmeup %rax
-	lea	.Lopts-.(%rax),%rax
+	lea	.Lopts(%rip),%rax
 	mov	OPENSSL_ia32cap_P(%rip),%edx
 	bt	\$20,%edx
 	jnc	.Ldone
@@ -357,9 +365,139 @@ RC4_options:
 .size	RC4_options,.-RC4_options
 ___
 
-$code =~ s/#([bwd])/$1/gm;
+# EXCEPTION_DISPOSITION handler (EXCEPTION_RECORD *rec,ULONG64 frame,
+#		CONTEXT *context,DISPATCHER_CONTEXT *disp)
+if ($win64) {
+$rec="%rcx";
+$frame="%rdx";
+$context="%r8";
+$disp="%r9";
+
+$code.=<<___;
+.extern	__imp_RtlVirtualUnwind
+.type	stream_se_handler,\@abi-omnipotent
+.align	16
+stream_se_handler:
+	push	%rsi
+	push	%rdi
+	push	%rbx
+	push	%rbp
+	push	%r12
+	push	%r13
+	push	%r14
+	push	%r15
+	pushfq
+	sub	\$64,%rsp
 
-$code =~ s/RC4_set_key/private_RC4_set_key/g if ($ENV{FIPSCANLIB} ne "");
+	mov	120($context),%rax	# pull context->Rax
+	mov	248($context),%rbx	# pull context->Rip
+
+	lea	.Lprologue(%rip),%r10
+	cmp	%r10,%rbx		# context->Rip<prologue label
+	jb	.Lin_prologue
+
+	mov	152($context),%rax	# pull context->Rsp
+
+	lea	.Lepilogue(%rip),%r10
+	cmp	%r10,%rbx		# context->Rip>=epilogue label
+	jae	.Lin_prologue
+
+	lea	24(%rax),%rax
+
+	mov	-8(%rax),%rbx
+	mov	-16(%rax),%r12
+	mov	-24(%rax),%r13
+	mov	%rbx,144($context)	# restore context->Rbx
+	mov	%r12,216($context)	# restore context->R12
+	mov	%r13,224($context)	# restore context->R13
+
+.Lin_prologue:
+	mov	8(%rax),%rdi
+	mov	16(%rax),%rsi
+	mov	%rax,152($context)	# restore context->Rsp
+	mov	%rsi,168($context)	# restore context->Rsi
+	mov	%rdi,176($context)	# restore context->Rdi
+
+	jmp	.Lcommon_seh_exit
+.size	stream_se_handler,.-stream_se_handler
+
+.type	key_se_handler,\@abi-omnipotent
+.align	16
+key_se_handler:
+	push	%rsi
+	push	%rdi
+	push	%rbx
+	push	%rbp
+	push	%r12
+	push	%r13
+	push	%r14
+	push	%r15
+	pushfq
+	sub	\$64,%rsp
+
+	mov	152($context),%rax	# pull context->Rsp
+	mov	8(%rax),%rdi
+	mov	16(%rax),%rsi
+	mov	%rsi,168($context)	# restore context->Rsi
+	mov	%rdi,176($context)	# restore context->Rdi
+
+.Lcommon_seh_exit:
+
+	mov	40($disp),%rdi		# disp->ContextRecord
+	mov	$context,%rsi		# context
+	mov	\$154,%ecx		# sizeof(CONTEXT)
+	.long	0xa548f3fc		# cld; rep movsq
+
+	mov	$disp,%rsi
+	xor	%rcx,%rcx		# arg1, UNW_FLAG_NHANDLER
+	mov	8(%rsi),%rdx		# arg2, disp->ImageBase
+	mov	0(%rsi),%r8		# arg3, disp->ControlPc
+	mov	16(%rsi),%r9		# arg4, disp->FunctionEntry
+	mov	40(%rsi),%r10		# disp->ContextRecord
+	lea	56(%rsi),%r11		# &disp->HandlerData
+	lea	24(%rsi),%r12		# &disp->EstablisherFrame
+	mov	%r10,32(%rsp)		# arg5
+	mov	%r11,40(%rsp)		# arg6
+	mov	%r12,48(%rsp)		# arg7
+	mov	%rcx,56(%rsp)		# arg8, (NULL)
+	call	*__imp_RtlVirtualUnwind(%rip)
+
+	mov	\$1,%eax		# ExceptionContinueSearch
+	add	\$64,%rsp
+	popfq
+	pop	%r15
+	pop	%r14
+	pop	%r13
+	pop	%r12
+	pop	%rbp
+	pop	%rbx
+	pop	%rdi
+	pop	%rsi
+	ret
+.size	key_se_handler,.-key_se_handler
+
+.section	.pdata
+.align	4
+	.rva	.LSEH_begin_RC4
+	.rva	.LSEH_end_RC4
+	.rva	.LSEH_info_RC4
+
+	.rva	.LSEH_begin_RC4_set_key
+	.rva	.LSEH_end_RC4_set_key
+	.rva	.LSEH_info_RC4_set_key
+
+.section	.xdata
+.align	8
+.LSEH_info_RC4:
+	.byte	9,0,0,0
+	.rva	stream_se_handler
+.LSEH_info_RC4_set_key:
+	.byte	9,0,0,0
+	.rva	key_se_handler
+___
+}
+
+$code =~ s/#([bwd])/$1/gm;
 
 print $code;