From 4b60f4b175158b426c1f1a2f30fac616ded28d21 Mon Sep 17 00:00:00 2001
From: Andy Polyakov <appro@openssl.org>
Date: Sun, 11 Nov 2007 16:25:46 +0000
Subject: [PATCH] rc4-x86_64.pl update [from HEAD].

---
 Configure                    |  10 +--
 TABLE                        |  68 +++++++++---------
 crypto/rc4/Makefile          |   4 +-
 crypto/rc4/asm/rc4-x86_64.pl | 130 ++++++++++++++++++++++++++++++++++-
 4 files changed, 168 insertions(+), 44 deletions(-)

diff --git a/Configure b/Configure
index d260005487..089e4eb76f 100755
--- a/Configure
+++ b/Configure
@@ -114,12 +114,12 @@ my $tlib="-lnsl -lsocket";
 my $bits1="THIRTY_TWO_BIT ";
 my $bits2="SIXTY_FOUR_BIT ";
 
-my $x86_elf_asm="x86cpuid-elf.o:bn86-elf.o co86-elf.o:dx86-elf.o yx86-elf.o:ax86-elf.o:bx86-elf.o:mx86-elf.o:sx86-elf.o s512sse2-elf.o:cx86-elf.o:rx86-elf.o:rm86-elf.o:r586-elf.o";
-my $x86_coff_asm="x86cpuid-cof.o:bn86-cof.o co86-cof.o:dx86-cof.o yx86-cof.o:ax86-cof.o:bx86-cof.o:mx86-cof.o:sx86-cof.o s512sse2-cof.o:cx86-cof.o:rx86-cof.o:rm86-cof.o:r586-cof.o";
-my $x86_out_asm="x86cpuid-out.o:bn86-out.o co86-out.o:dx86-out.o yx86-out.o:ax86-out.o:bx86-out.o:mx86-out.o:sx86-out.o s512sse2-out.o:cx86-out.o:rx86-out.o:rm86-out.o:r586-out.o";
+my $x86_elf_asm="x86cpuid-elf.o:bn86-elf.o co86-elf.o:dx86-elf.o yx86-elf.o:ax86-elf.o:bx86-elf.o:mx86-elf.o:sx86-elf.o s512sse2-elf.o:cx86-elf.o:rx86-elf.o:rm86-elf.o:r586-elf.o rc4_skey.o";
+my $x86_coff_asm="x86cpuid-cof.o:bn86-cof.o co86-cof.o:dx86-cof.o yx86-cof.o:ax86-cof.o:bx86-cof.o:mx86-cof.o:sx86-cof.o s512sse2-cof.o:cx86-cof.o:rx86-cof.o:rm86-cof.o:r586-cof.o rc4_skey.o";
+my $x86_out_asm="x86cpuid-out.o:bn86-out.o co86-out.o:dx86-out.o yx86-out.o:ax86-out.o:bx86-out.o:mx86-out.o:sx86-out.o s512sse2-out.o:cx86-out.o:rx86-out.o:rm86-out.o:r586-out.o rc4_skey.o";
 
 my $x86_64_asm="x86_64cpuid.o:x86_64-gcc.o::aes-x86_64.o::md5-x86_64.o:sha1-x86_64.o sha256-x86_64.o sha512-x86_64.o::rc4-x86_64.o::";
-my $ia64_asm=":bn-ia64.o::aes_core.o aes_cbc.o aes-ia64.o:::sha1-ia64.o sha256-ia64.o sha512-ia64.o::rc4-ia64.o::";
+my $ia64_asm=":bn-ia64.o::aes_core.o aes_cbc.o aes-ia64.o:::sha1-ia64.o sha256-ia64.o sha512-ia64.o::rc4-ia64.o rc4_skey.o::";
 
 my $no_asm="::::::::::";
 
@@ -589,7 +589,7 @@ my $des_enc="des_enc.o fcrypt_b.o";
 my $aes_enc="aes_core.o aes_cbc.o";
 my $bf_enc	="bf_enc.o";
 my $cast_enc="c_enc.o";
-my $rc4_enc="rc4_enc.o";
+my $rc4_enc="rc4_enc.o rc4_skey.o";
 my $rc5_enc="rc5_enc.o";
 my $md5_obj="";
 my $sha1_obj="";
diff --git a/TABLE b/TABLE
index 7fd0d4dcf1..a4e0277615 100644
--- a/TABLE
+++ b/TABLE
@@ -124,7 +124,7 @@ $bf_obj       =
 $md5_obj      = 
 $sha1_obj     = sha1-ia64.o sha256-ia64.o sha512-ia64.o
 $cast_obj     = 
-$rc4_obj      = rc4-ia64.o
+$rc4_obj      = rc4-ia64.o rc4_skey.o
 $rmd160_obj   = 
 $rc5_obj      = 
 $dso_scheme   = dlfcn
@@ -207,7 +207,7 @@ $sha1_obj     = sx86-out.o s512sse2-out.o
 $cast_obj     = cx86-out.o
 $rc4_obj      = rx86-out.o
 $rmd160_obj   = rm86-out.o
-$rc5_obj      = r586-out.o
+$rc5_obj      = r586-out.o rc4_skey.o
 $dso_scheme   = dlfcn
 $shared_target= bsd-shared
 $shared_cflag = -fPIC
@@ -234,7 +234,7 @@ $sha1_obj     = sx86-elf.o s512sse2-elf.o
 $cast_obj     = cx86-elf.o
 $rc4_obj      = rx86-elf.o
 $rmd160_obj   = rm86-elf.o
-$rc5_obj      = r586-elf.o
+$rc5_obj      = r586-elf.o rc4_skey.o
 $dso_scheme   = dlfcn
 $shared_target= bsd-shared
 $shared_cflag = -fPIC
@@ -288,7 +288,7 @@ $sha1_obj     = sx86-cof.o s512sse2-cof.o
 $cast_obj     = cx86-cof.o
 $rc4_obj      = rx86-cof.o
 $rmd160_obj   = rm86-cof.o
-$rc5_obj      = r586-cof.o
+$rc5_obj      = r586-cof.o rc4_skey.o
 $dso_scheme   = dlfcn
 $shared_target= cygwin-shared
 $shared_cflag = -D_WINDLL
@@ -342,7 +342,7 @@ $sha1_obj     = sx86-out.o s512sse2-out.o
 $cast_obj     = cx86-out.o
 $rc4_obj      = rx86-out.o
 $rmd160_obj   = rm86-out.o
-$rc5_obj      = r586-out.o
+$rc5_obj      = r586-out.o rc4_skey.o
 $dso_scheme   = 
 $shared_target= 
 $shared_cflag = 
@@ -855,7 +855,7 @@ $sha1_obj     = sx86-elf.o s512sse2-elf.o
 $cast_obj     = cx86-elf.o
 $rc4_obj      = rx86-elf.o
 $rmd160_obj   = rm86-elf.o
-$rc5_obj      = r586-elf.o
+$rc5_obj      = r586-elf.o rc4_skey.o
 $dso_scheme   = dlfcn
 $shared_target= bsd-gcc-shared
 $shared_cflag = -fPIC
@@ -1098,7 +1098,7 @@ $sha1_obj     = sx86-elf.o s512sse2-elf.o
 $cast_obj     = cx86-elf.o
 $rc4_obj      = rx86-elf.o
 $rmd160_obj   = rm86-elf.o
-$rc5_obj      = r586-elf.o
+$rc5_obj      = r586-elf.o rc4_skey.o
 $dso_scheme   = dlfcn
 $shared_target= bsd-shared
 $shared_cflag = -fPIC
@@ -1287,7 +1287,7 @@ $sha1_obj     = sx86-elf.o s512sse2-elf.o
 $cast_obj     = cx86-elf.o
 $rc4_obj      = rx86-elf.o
 $rmd160_obj   = rm86-elf.o
-$rc5_obj      = r586-elf.o
+$rc5_obj      = r586-elf.o rc4_skey.o
 $dso_scheme   = 
 $shared_target= 
 $shared_cflag = 
@@ -1368,7 +1368,7 @@ $sha1_obj     = sx86-elf.o s512sse2-elf.o
 $cast_obj     = cx86-elf.o
 $rc4_obj      = rx86-elf.o
 $rmd160_obj   = rm86-elf.o
-$rc5_obj      = r586-elf.o
+$rc5_obj      = r586-elf.o rc4_skey.o
 $dso_scheme   = dlfcn
 $shared_target= linux-shared
 $shared_cflag = -fPIC
@@ -1395,7 +1395,7 @@ $sha1_obj     = sx86-elf.o s512sse2-elf.o
 $cast_obj     = cx86-elf.o
 $rc4_obj      = rx86-elf.o
 $rmd160_obj   = rm86-elf.o
-$rc5_obj      = r586-elf.o
+$rc5_obj      = r586-elf.o rc4_skey.o
 $dso_scheme   = dlfcn
 $shared_target= linux-shared
 $shared_cflag = -fPIC
@@ -1476,7 +1476,7 @@ $sha1_obj     = sx86-elf.o s512sse2-elf.o
 $cast_obj     = cx86-elf.o
 $rc4_obj      = rx86-elf.o
 $rmd160_obj   = rm86-elf.o
-$rc5_obj      = r586-elf.o
+$rc5_obj      = r586-elf.o rc4_skey.o
 $dso_scheme   = dlfcn
 $shared_target= linux-shared
 $shared_cflag = -fPIC
@@ -1503,7 +1503,7 @@ $sha1_obj     = sx86-elf.o s512sse2-elf.o
 $cast_obj     = cx86-elf.o
 $rc4_obj      = rx86-elf.o
 $rmd160_obj   = rm86-elf.o
-$rc5_obj      = r586-elf.o
+$rc5_obj      = r586-elf.o rc4_skey.o
 $dso_scheme   = dlfcn
 $shared_target= linux-shared
 $shared_cflag = -fPIC
@@ -1530,7 +1530,7 @@ $sha1_obj     = sx86-elf.o s512sse2-elf.o
 $cast_obj     = cx86-elf.o
 $rc4_obj      = rx86-elf.o
 $rmd160_obj   = rm86-elf.o
-$rc5_obj      = r586-elf.o
+$rc5_obj      = r586-elf.o rc4_skey.o
 $dso_scheme   = dlfcn
 $shared_target= 
 $shared_cflag = 
@@ -1557,7 +1557,7 @@ $sha1_obj     = sx86-elf.o s512sse2-elf.o
 $cast_obj     = cx86-elf.o
 $rc4_obj      = rx86-elf.o
 $rmd160_obj   = rm86-elf.o
-$rc5_obj      = r586-elf.o
+$rc5_obj      = r586-elf.o rc4_skey.o
 $dso_scheme   = dlfcn
 $shared_target= 
 $shared_cflag = 
@@ -1584,7 +1584,7 @@ $sha1_obj     = sx86-elf.o s512sse2-elf.o
 $cast_obj     = cx86-elf.o
 $rc4_obj      = rx86-elf.o
 $rmd160_obj   = rm86-elf.o
-$rc5_obj      = r586-elf.o
+$rc5_obj      = r586-elf.o rc4_skey.o
 $dso_scheme   = 
 $shared_target= 
 $shared_cflag = 
@@ -1719,7 +1719,7 @@ $sha1_obj     = sx86-elf.o s512sse2-elf.o
 $cast_obj     = cx86-elf.o
 $rc4_obj      = rx86-elf.o
 $rmd160_obj   = rm86-elf.o
-$rc5_obj      = r586-elf.o
+$rc5_obj      = r586-elf.o rc4_skey.o
 $dso_scheme   = dlfcn
 $shared_target= linux-shared
 $shared_cflag = 
@@ -1881,7 +1881,7 @@ $sha1_obj     = sx86-elf.o s512sse2-elf.o
 $cast_obj     = cx86-elf.o
 $rc4_obj      = rx86-elf.o
 $rmd160_obj   = rm86-elf.o
-$rc5_obj      = r586-elf.o
+$rc5_obj      = r586-elf.o rc4_skey.o
 $dso_scheme   = 
 $shared_target= 
 $shared_cflag = 
@@ -2014,7 +2014,7 @@ $bf_obj       =
 $md5_obj      = 
 $sha1_obj     = sha1-ia64.o sha256-ia64.o sha512-ia64.o
 $cast_obj     = 
-$rc4_obj      = rc4-ia64.o
+$rc4_obj      = rc4-ia64.o rc4_skey.o
 $rmd160_obj   = 
 $rc5_obj      = 
 $dso_scheme   = dlfcn
@@ -2041,7 +2041,7 @@ $bf_obj       =
 $md5_obj      = 
 $sha1_obj     = sha1-ia64.o sha256-ia64.o sha512-ia64.o
 $cast_obj     = 
-$rc4_obj      = rc4-ia64.o
+$rc4_obj      = rc4-ia64.o rc4_skey.o
 $rmd160_obj   = 
 $rc5_obj      = 
 $dso_scheme   = dlfcn
@@ -2230,7 +2230,7 @@ $bf_obj       =
 $md5_obj      = 
 $sha1_obj     = sha1-ia64.o sha256-ia64.o sha512-ia64.o
 $cast_obj     = 
-$rc4_obj      = rc4-ia64.o
+$rc4_obj      = rc4-ia64.o rc4_skey.o
 $rmd160_obj   = 
 $rc5_obj      = 
 $dso_scheme   = dlfcn
@@ -2257,7 +2257,7 @@ $bf_obj       =
 $md5_obj      = 
 $sha1_obj     = sha1-ia64.o sha256-ia64.o sha512-ia64.o
 $cast_obj     = 
-$rc4_obj      = rc4-ia64.o
+$rc4_obj      = rc4-ia64.o rc4_skey.o
 $rmd160_obj   = 
 $rc5_obj      = 
 $dso_scheme   = dlfcn
@@ -2340,7 +2340,7 @@ $sha1_obj     = sx86-elf.o s512sse2-elf.o
 $cast_obj     = cx86-elf.o
 $rc4_obj      = rx86-elf.o
 $rmd160_obj   = rm86-elf.o
-$rc5_obj      = r586-elf.o
+$rc5_obj      = r586-elf.o rc4_skey.o
 $dso_scheme   = dlfcn
 $shared_target= linux-shared
 $shared_cflag = -fPIC
@@ -2637,7 +2637,7 @@ $sha1_obj     = sx86-out.o s512sse2-out.o
 $cast_obj     = cx86-out.o
 $rc4_obj      = rx86-out.o
 $rmd160_obj   = rm86-out.o
-$rc5_obj      = r586-out.o
+$rc5_obj      = r586-out.o rc4_skey.o
 $dso_scheme   = 
 $shared_target= 
 $shared_cflag = 
@@ -2664,7 +2664,7 @@ $sha1_obj     = sx86-elf.o s512sse2-elf.o
 $cast_obj     = cx86-elf.o
 $rc4_obj      = rx86-elf.o
 $rmd160_obj   = rm86-elf.o
-$rc5_obj      = r586-elf.o
+$rc5_obj      = r586-elf.o rc4_skey.o
 $dso_scheme   = dlfcn
 $shared_target= linux-shared
 $shared_cflag = -fPIC
@@ -2745,7 +2745,7 @@ $sha1_obj     = sx86-elf.o s512sse2-elf.o
 $cast_obj     = cx86-elf.o
 $rc4_obj      = rx86-elf.o
 $rmd160_obj   = rm86-elf.o
-$rc5_obj      = r586-elf.o
+$rc5_obj      = r586-elf.o rc4_skey.o
 $dso_scheme   = dlfcn
 $shared_target= linux-shared
 $shared_cflag = -KPIC
@@ -2770,7 +2770,7 @@ $bf_obj       =
 $md5_obj      = 
 $sha1_obj     = sha1-ia64.o sha256-ia64.o sha512-ia64.o
 $cast_obj     = 
-$rc4_obj      = rc4-ia64.o
+$rc4_obj      = rc4-ia64.o rc4_skey.o
 $rmd160_obj   = 
 $rc5_obj      = 
 $dso_scheme   = dlfcn
@@ -2797,7 +2797,7 @@ $bf_obj       =
 $md5_obj      = 
 $sha1_obj     = sha1-ia64.o sha256-ia64.o sha512-ia64.o
 $cast_obj     = 
-$rc4_obj      = rc4-ia64.o
+$rc4_obj      = rc4-ia64.o rc4_skey.o
 $rmd160_obj   = 
 $rc5_obj      = 
 $dso_scheme   = dlfcn
@@ -2824,7 +2824,7 @@ $bf_obj       =
 $md5_obj      = 
 $sha1_obj     = sha1-ia64.o sha256-ia64.o sha512-ia64.o
 $cast_obj     = 
-$rc4_obj      = rc4-ia64.o
+$rc4_obj      = rc4-ia64.o rc4_skey.o
 $rmd160_obj   = 
 $rc5_obj      = 
 $dso_scheme   = dlfcn
@@ -3015,7 +3015,7 @@ $sha1_obj     = sx86-cof.o s512sse2-cof.o
 $cast_obj     = cx86-cof.o
 $rc4_obj      = rx86-cof.o
 $rmd160_obj   = rm86-cof.o
-$rc5_obj      = r586-cof.o
+$rc5_obj      = r586-cof.o rc4_skey.o
 $dso_scheme   = win32
 $shared_target= cygwin-shared
 $shared_cflag = -D_WINDLL -DOPENSSL_USE_APPLINK
@@ -3420,7 +3420,7 @@ $sha1_obj     = sx86-elf.o s512sse2-elf.o
 $cast_obj     = cx86-elf.o
 $rc4_obj      = rx86-elf.o
 $rmd160_obj   = rm86-elf.o
-$rc5_obj      = r586-elf.o
+$rc5_obj      = r586-elf.o rc4_skey.o
 $dso_scheme   = dlfcn
 $shared_target= svr3-shared
 $shared_cflag = -Kpic
@@ -3447,7 +3447,7 @@ $sha1_obj     = sx86-elf.o s512sse2-elf.o
 $cast_obj     = cx86-elf.o
 $rc4_obj      = rx86-elf.o
 $rmd160_obj   = rm86-elf.o
-$rc5_obj      = r586-elf.o
+$rc5_obj      = r586-elf.o rc4_skey.o
 $dso_scheme   = dlfcn
 $shared_target= svr3-shared
 $shared_cflag = -fPIC
@@ -3663,7 +3663,7 @@ $sha1_obj     = sx86-elf.o s512sse2-elf.o
 $cast_obj     = cx86-elf.o
 $rc4_obj      = rx86-elf.o
 $rmd160_obj   = rm86-elf.o
-$rc5_obj      = r586-elf.o
+$rc5_obj      = r586-elf.o rc4_skey.o
 $dso_scheme   = dlfcn
 $shared_target= solaris-shared
 $shared_cflag = -fPIC
@@ -3987,7 +3987,7 @@ $sha1_obj     = sx86-elf.o s512sse2-elf.o
 $cast_obj     = cx86-elf.o
 $rc4_obj      = rx86-elf.o
 $rmd160_obj   = rm86-elf.o
-$rc5_obj      = r586-elf.o
+$rc5_obj      = r586-elf.o rc4_skey.o
 $dso_scheme   = dlfcn
 $shared_target= svr5-shared
 $shared_cflag = -Kpic
@@ -4014,7 +4014,7 @@ $sha1_obj     = sx86-elf.o s512sse2-elf.o
 $cast_obj     = cx86-elf.o
 $rc4_obj      = rx86-elf.o
 $rmd160_obj   = rm86-elf.o
-$rc5_obj      = r586-elf.o
+$rc5_obj      = r586-elf.o rc4_skey.o
 $dso_scheme   = dlfcn
 $shared_target= gnu-shared
 $shared_cflag = -fPIC
diff --git a/crypto/rc4/Makefile b/crypto/rc4/Makefile
index 7857c95fbf..187ed5c668 100644
--- a/crypto/rc4/Makefile
+++ b/crypto/rc4/Makefile
@@ -10,7 +10,7 @@ INCLUDES=
 CFLAG=-g
 AR=		ar r
 
-RC4_ENC=rc4_enc.o
+RC4_ENC=rc4_enc.o rc4_skey.o
 
 CFLAGS= $(INCLUDES) $(CFLAG)
 ASFLAGS= $(INCLUDES) $(ASFLAG)
@@ -22,7 +22,7 @@ APPS=
 
 LIB=$(TOP)/libcrypto.a
 LIBSRC=rc4_skey.c rc4_enc.c
-LIBOBJ=rc4_skey.o $(RC4_ENC)
+LIBOBJ=$(RC4_ENC)
 
 SRC= $(LIBSRC)
 
diff --git a/crypto/rc4/asm/rc4-x86_64.pl b/crypto/rc4/asm/rc4-x86_64.pl
index 4b990cba07..2d47320485 100755
--- a/crypto/rc4/asm/rc4-x86_64.pl
+++ b/crypto/rc4/asm/rc4-x86_64.pl
@@ -2,8 +2,9 @@
 #
 # ====================================================================
 # 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.
+# 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/.
 # ====================================================================
 #
 # 2.22x RC4 tune-up:-) It should be noted though that my hand [as in
@@ -49,8 +50,22 @@
 # is not implemented, then this final RC4_CHAR code-path should be
 # preferred, as it provides better *all-round* performance].
 
+# Intel Core2 was observed to perform poorly on both code paths:-( It
+# apparently suffers from some kind of partial register stall, which
+# occurs in 64-bit mode only [as virtually identical 32-bit loop was
+# observed to outperform 64-bit one by almost 50%]. Adding two movzb to
+# cloop1 boosts its performance by 80%! This loop appears to be optimal
+# fit for Core2 and therefore the code was modified to skip cloop8 on
+# this CPU.
+
 $output=shift;
-open STDOUT,"| $^X ../perlasm/x86_64-xlate.pl $output";
+
+$0 =~ m/(.*[\/\\])[^\/\\]+$/; $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";
 
 $dat="%rdi";	    # arg1
 $len="%rsi";	    # arg2
@@ -152,6 +167,8 @@ $code.=<<___;
 	movzb	($dat,$XX[0]),$TX[0]#d
 	test	\$-8,$len
 	jz	.Lcloop1
+	cmp	\$0,260($dat)
+	jnz	.Lcloop1
 	push	%rbx
 	jmp	.Lcloop8
 .align	16
@@ -221,6 +238,8 @@ $code.=<<___;
 	movb	$TY#b,($dat,$XX[0])
 	add	$TX[0]#b,$TY#b
 	add	\$1,$XX[0]#b
+	movzb	$TY#b,$TY#d
+	movzb	$XX[0]#b,$XX[0]#d
 	movzb	($dat,$TY),$TY#d
 	movzb	($dat,$XX[0]),$TX[0]#d
 	xorb	($inp),$TY#b
@@ -233,6 +252,111 @@ $code.=<<___;
 .size	RC4,.-RC4
 ___
 
+$idx="%r8";
+$ido="%r9";
+
+$code.=<<___;
+.extern	OPENSSL_ia32cap_P
+.globl	RC4_set_key
+.type	RC4_set_key,\@function,3
+.align	16
+RC4_set_key:
+	lea	8($dat),$dat
+	lea	($inp,$len),$inp
+	neg	$len
+	mov	$len,%rcx
+	xor	%eax,%eax
+	xor	$ido,$ido
+	xor	%r10,%r10
+	xor	%r11,%r11
+
+	mov	OPENSSL_ia32cap_P(%rip),$idx#d
+	bt	\$20,$idx#d
+	jnc	.Lw1stloop
+	bt	\$30,$idx#d
+	setc	$ido#b
+	mov	$ido#d,260($dat)
+	jmp	.Lc1stloop
+
+.align	16
+.Lw1stloop:
+	mov	%eax,($dat,%rax,4)
+	add	\$1,%al
+	jnc	.Lw1stloop
+
+	xor	$ido,$ido
+	xor	$idx,$idx
+.align	16
+.Lw2ndloop:
+	mov	($dat,$ido,4),%r10d
+	add	($inp,$len,1),$idx#b
+	add	%r10b,$idx#b
+	add	\$1,$len
+	mov	($dat,$idx,4),%r11d
+	cmovz	%rcx,$len
+	mov	%r10d,($dat,$idx,4)
+	mov	%r11d,($dat,$ido,4)
+	add	\$1,$ido#b
+	jnc	.Lw2ndloop
+	jmp	.Lexit_key
+
+.align	16
+.Lc1stloop:
+	mov	%al,($dat,%rax)
+	add	\$1,%al
+	jnc	.Lc1stloop
+
+	xor	$ido,$ido
+	xor	$idx,$idx
+.align	16
+.Lc2ndloop:
+	mov	($dat,$ido),%r10b
+	add	($inp,$len),$idx#b
+	add	%r10b,$idx#b
+	add	\$1,$len
+	mov	($dat,$idx),%r11b
+	jnz	.Lcnowrap
+	mov	%rcx,$len
+.Lcnowrap:
+	mov	%r10b,($dat,$idx)
+	mov	%r11b,($dat,$ido)
+	add	\$1,$ido#b
+	jnc	.Lc2ndloop
+	movl	\$-1,256($dat)
+
+.align	16
+.Lexit_key:
+	xor	%eax,%eax
+	mov	%eax,-8($dat)
+	mov	%eax,-4($dat)
+	ret
+.size	RC4_set_key,.-RC4_set_key
+
+.globl	RC4_options
+.type	RC4_options,\@function,0
+.align	16
+RC4_options:
+	.picmeup %rax
+	lea	.Lopts-.(%rax),%rax
+	mov	OPENSSL_ia32cap_P(%rip),%edx
+	bt	\$20,%edx
+	jnc	.Ldone
+	add	\$12,%rax
+	bt	\$30,%edx
+	jnc	.Ldone
+	add	\$13,%rax
+.Ldone:
+	ret
+.align	64
+.Lopts:
+.asciz	"rc4(8x,int)"
+.asciz	"rc4(8x,char)"
+.asciz	"rc4(1x,char)"
+.asciz	"RC4 for x86_64, CRYPTOGAMS by <appro\@openssl.org>"
+.align	64
+.size	RC4_options,.-RC4_options
+___
+
 $code =~ s/#([bwd])/$1/gm;
 
 print $code;
-- 
2.34.1