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

4 files changed, 3552 insertions, 868 deletions

diff --git a/openssl/crypto/aes/asm/aes-586.pl b/openssl/crypto/aes/asm/aes-586.pl
index 3bc46a968..aab40e6f1 100644
--- a/openssl/crypto/aes/asm/aes-586.pl
+++ b/openssl/crypto/aes/asm/aes-586.pl
@@ -2,11 +2,12 @@
 #
 # ====================================================================
 # 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/.
 # ====================================================================
 #
-# Version 3.6.
+# Version 4.3.
 #
 # You might fail to appreciate this module performance from the first
 # try. If compared to "vanilla" linux-ia32-icc target, i.e. considered
@@ -81,11 +82,117 @@
 # AMD K8	20			19
 # PIII		25			23
 # Pentium	81			78
-
-push(@INC,"perlasm","../../perlasm");
+#
+# Version 3.7 reimplements outer rounds as "compact." Meaning that
+# first and last rounds reference compact 256 bytes S-box. This means
+# that first round consumes a lot more CPU cycles and that encrypt
+# and decrypt performance becomes asymmetric. Encrypt performance
+# drops by 10-12%, while decrypt - by 20-25%:-( 256 bytes S-box is
+# aggressively pre-fetched.
+#
+# Version 4.0 effectively rolls back to 3.6 and instead implements
+# additional set of functions, _[x86|sse]_AES_[en|de]crypt_compact,
+# which use exclusively 256 byte S-box. These functions are to be
+# called in modes not concealing plain text, such as ECB, or when
+# we're asked to process smaller amount of data [or unconditionally
+# on hyper-threading CPU]. Currently it's called unconditionally from
+# AES_[en|de]crypt, which affects all modes, but CBC. CBC routine
+# still needs to be modified to switch between slower and faster
+# mode when appropriate... But in either case benchmark landscape
+# changes dramatically and below numbers are CPU cycles per processed
+# byte for 128-bit key.
+#
+#		ECB encrypt	ECB decrypt	CBC large chunk
+# P4		56[60]		84[100]		23
+# AMD K8	48[44]		70[79]		18
+# PIII		41[50]		61[91]		24
+# Core 2	32[38]		45[70]		18.5
+# Pentium	120		160		77
+#
+# Version 4.1 switches to compact S-box even in key schedule setup.
+#
+# Version 4.2 prefetches compact S-box in every SSE round or in other
+# words every cache-line is *guaranteed* to be accessed within ~50
+# cycles window. Why just SSE? Because it's needed on hyper-threading
+# CPU! Which is also why it's prefetched with 64 byte stride. Best
+# part is that it has no negative effect on performance:-)  
+#
+# Version 4.3 implements switch between compact and non-compact block
+# functions in AES_cbc_encrypt depending on how much data was asked
+# to be processed in one stroke.
+#
+######################################################################
+# Timing attacks are classified in two classes: synchronous when
+# attacker consciously initiates cryptographic operation and collects
+# timing data of various character afterwards, and asynchronous when
+# malicious code is executed on same CPU simultaneously with AES,
+# instruments itself and performs statistical analysis of this data.
+#
+# As far as synchronous attacks go the root to the AES timing
+# vulnerability is twofold. Firstly, of 256 S-box elements at most 160
+# are referred to in single 128-bit block operation. Well, in C
+# implementation with 4 distinct tables it's actually as little as 40
+# references per 256 elements table, but anyway... Secondly, even
+# though S-box elements are clustered into smaller amount of cache-
+# lines, smaller than 160 and even 40, it turned out that for certain
+# plain-text pattern[s] or simply put chosen plain-text and given key
+# few cache-lines remain unaccessed during block operation. Now, if
+# attacker can figure out this access pattern, he can deduct the key
+# [or at least part of it]. The natural way to mitigate this kind of
+# attacks is to minimize the amount of cache-lines in S-box and/or
+# prefetch them to ensure that every one is accessed for more uniform
+# timing. But note that *if* plain-text was concealed in such way that
+# input to block function is distributed *uniformly*, then attack
+# wouldn't apply. Now note that some encryption modes, most notably
+# CBC, do mask the plain-text in this exact way [secure cipher output
+# is distributed uniformly]. Yes, one still might find input that
+# would reveal the information about given key, but if amount of
+# candidate inputs to be tried is larger than amount of possible key
+# combinations then attack becomes infeasible. This is why revised
+# AES_cbc_encrypt "dares" to switch to larger S-box when larger chunk
+# of data is to be processed in one stroke. The current size limit of
+# 512 bytes is chosen to provide same [diminishigly low] probability
+# for cache-line to remain untouched in large chunk operation with
+# large S-box as for single block operation with compact S-box and
+# surely needs more careful consideration...
+#
+# As for asynchronous attacks. There are two flavours: attacker code
+# being interleaved with AES on hyper-threading CPU at *instruction*
+# level, and two processes time sharing single core. As for latter.
+# Two vectors. 1. Given that attacker process has higher priority,
+# yield execution to process performing AES just before timer fires
+# off the scheduler, immediately regain control of CPU and analyze the
+# cache state. For this attack to be efficient attacker would have to
+# effectively slow down the operation by several *orders* of magnitute,
+# by ratio of time slice to duration of handful of AES rounds, which
+# unlikely to remain unnoticed. Not to mention that this also means
+# that he would spend correspondigly more time to collect enough
+# statistical data to mount the attack. It's probably appropriate to
+# say that if adeversary reckons that this attack is beneficial and
+# risks to be noticed, you probably have larger problems having him
+# mere opportunity. In other words suggested code design expects you
+# to preclude/mitigate this attack by overall system security design.
+# 2. Attacker manages to make his code interrupt driven. In order for
+# this kind of attack to be feasible, interrupt rate has to be high
+# enough, again comparable to duration of handful of AES rounds. But
+# is there interrupt source of such rate? Hardly, not even 1Gbps NIC
+# generates interrupts at such raging rate...
+#
+# And now back to the former, hyper-threading CPU or more specifically
+# Intel P4. Recall that asynchronous attack implies that malicious
+# code instruments itself. And naturally instrumentation granularity
+# has be noticeably lower than duration of codepath accessing S-box.
+# Given that all cache-lines are accessed during that time that is.
+# Current implementation accesses *all* cache-lines within ~50 cycles
+# window, which is actually *less* than RDTSC latency on Intel P4!
+
+$0 =~ m/(.*[\/\\])[^\/\\]+$/; $dir=$1;
+push(@INC,"${dir}","${dir}../../perlasm");
 require "x86asm.pl";
 
-&asm_init($ARGV[0],"aes-586.pl",$ARGV[$#ARGV] eq "386");
+&asm_init($ARGV[0],"aes-586.pl",$x86only = $ARGV[$#ARGV] eq "386");
+&static_label("AES_Te");
+&static_label("AES_Td");
 
 $s0="eax";
 $s1="ebx";
@@ -93,21 +200,36 @@ $s2="ecx";
 $s3="edx";
 $key="edi";
 $acc="esi";
+$tbl="ebp";
+
+# stack frame layout in _[x86|sse]_AES_* routines, frame is allocated
+# by caller
+$__ra=&DWP(0,"esp");	# return address
+$__s0=&DWP(4,"esp");	# s0 backing store
+$__s1=&DWP(8,"esp");	# s1 backing store
+$__s2=&DWP(12,"esp");	# s2 backing store
+$__s3=&DWP(16,"esp");	# s3 backing store
+$__key=&DWP(20,"esp");	# pointer to key schedule
+$__end=&DWP(24,"esp");	# pointer to end of key schedule
+$__tbl=&DWP(28,"esp");	# %ebp backing store
+
+# stack frame layout in AES_[en|crypt] routines, which differs from
+# above by 4 and overlaps by %ebp backing store
+$_tbl=&DWP(24,"esp");
+$_esp=&DWP(28,"esp");
 
-$compromise=0;		# $compromise=128 abstains from copying key
-			# schedule to stack when encrypting inputs
-			# shorter than 128 bytes at the cost of
-			# risksing aliasing with S-boxes. In return
-			# you get way better, up to +70%, small block
-			# performance.
+sub _data_word() { my $i; while(defined($i=shift)) { &data_word($i,$i); } }
+
+$speed_limit=512;	# chunks smaller than $speed_limit are
+			# processed with compact routine in CBC mode
 $small_footprint=1;	# $small_footprint=1 code is ~5% slower [on
 			# recent µ-archs], but ~5 times smaller!
 			# I favor compact code to minimize cache
 			# contention and in hope to "collect" 5% back
 			# in real-life applications...
+
 $vertical_spin=0;	# shift "verticaly" defaults to 0, because of
 			# its proof-of-concept status...
-
 # Note that there is no decvert(), as well as last encryption round is
 # performed with "horizontal" shifts. This is because this "vertical"
 # implementation [one which groups shifts on a given $s[i] to form a
@@ -170,17 +292,484 @@ sub encvert()
 	&movz	($v0,&HB($v1));
 	&and	($v1,0xFF);
 	&xor	($s[3],&DWP(2,$te,$v1,8));		# s1>>16
-	 &mov	($key,&DWP(12,"esp"));			# reincarnate v1 as key
+	 &mov	($key,$__key);				# reincarnate v1 as key
 	&xor	($s[2],&DWP(1,$te,$v0,8));		# s1>>24
 }
 
+# Another experimental routine, which features "horizontal spin," but
+# eliminates one reference to stack. Strangely enough runs slower...
+sub enchoriz()
+{ my $v0 = $key, $v1 = $acc;
+
+	&movz	($v0,&LB($s0));			#  3, 2, 1, 0*
+	&rotr	($s2,8);			#  8,11,10, 9
+	&mov	($v1,&DWP(0,$te,$v0,8));	#  0
+	&movz	($v0,&HB($s1));			#  7, 6, 5*, 4
+	&rotr	($s3,16);			# 13,12,15,14
+	&xor	($v1,&DWP(3,$te,$v0,8));	#  5
+	&movz	($v0,&HB($s2));			#  8,11,10*, 9
+	&rotr	($s0,16);			#  1, 0, 3, 2
+	&xor	($v1,&DWP(2,$te,$v0,8));	# 10
+	&movz	($v0,&HB($s3));			# 13,12,15*,14
+	&xor	($v1,&DWP(1,$te,$v0,8));	# 15, t[0] collected
+	&mov	($__s0,$v1);			# t[0] saved
+
+	&movz	($v0,&LB($s1));			#  7, 6, 5, 4*
+	&shr	($s1,16);			#  -, -, 7, 6
+	&mov	($v1,&DWP(0,$te,$v0,8));	#  4
+	&movz	($v0,&LB($s3));			# 13,12,15,14*
+	&xor	($v1,&DWP(2,$te,$v0,8));	# 14
+	&movz	($v0,&HB($s0));			#  1, 0, 3*, 2
+	&and	($s3,0xffff0000);		# 13,12, -, -
+	&xor	($v1,&DWP(1,$te,$v0,8));	#  3
+	&movz	($v0,&LB($s2));			#  8,11,10, 9*
+	&or	($s3,$s1);			# 13,12, 7, 6
+	&xor	($v1,&DWP(3,$te,$v0,8));	#  9, t[1] collected
+	&mov	($s1,$v1);			#  s[1]=t[1]
+
+	&movz	($v0,&LB($s0));			#  1, 0, 3, 2*
+	&shr	($s2,16);			#  -, -, 8,11
+	&mov	($v1,&DWP(2,$te,$v0,8));	#  2
+	&movz	($v0,&HB($s3));			# 13,12, 7*, 6
+	&xor	($v1,&DWP(1,$te,$v0,8));	#  7
+	&movz	($v0,&HB($s2));			#  -, -, 8*,11
+	&xor	($v1,&DWP(0,$te,$v0,8));	#  8
+	&mov	($v0,$s3);
+	&shr	($v0,24);			# 13
+	&xor	($v1,&DWP(3,$te,$v0,8));	# 13, t[2] collected
+
+	&movz	($v0,&LB($s2));			#  -, -, 8,11*
+	&shr	($s0,24);			#  1*
+	&mov	($s2,&DWP(1,$te,$v0,8));	# 11
+	&xor	($s2,&DWP(3,$te,$s0,8));	#  1
+	&mov	($s0,$__s0);			# s[0]=t[0]
+	&movz	($v0,&LB($s3));			# 13,12, 7, 6*
+	&shr	($s3,16);			#   ,  ,13,12
+	&xor	($s2,&DWP(2,$te,$v0,8));	#  6
+	&mov	($key,$__key);			# reincarnate v0 as key
+	&and	($s3,0xff);			#   ,  ,13,12*
+	&mov	($s3,&DWP(0,$te,$s3,8));	# 12
+	&xor	($s3,$s2);			# s[2]=t[3] collected
+	&mov	($s2,$v1);			# s[2]=t[2]
+}
+
+# More experimental code... SSE one... Even though this one eliminates
+# *all* references to stack, it's not faster...
+sub sse_encbody()
+{
+	&movz	($acc,&LB("eax"));		#  0
+	&mov	("ecx",&DWP(0,$tbl,$acc,8));	#  0
+	&pshufw	("mm2","mm0",0x0d);		#  7, 6, 3, 2
+	&movz	("edx",&HB("eax"));		#  1
+	&mov	("edx",&DWP(3,$tbl,"edx",8));	#  1
+	&shr	("eax",16);			#  5, 4
+
+	&movz	($acc,&LB("ebx"));		# 10
+	&xor	("ecx",&DWP(2,$tbl,$acc,8));	# 10
+	&pshufw	("mm6","mm4",0x08);		# 13,12, 9, 8
+	&movz	($acc,&HB("ebx"));		# 11
+	&xor	("edx",&DWP(1,$tbl,$acc,8));	# 11
+	&shr	("ebx",16);			# 15,14
+
+	&movz	($acc,&HB("eax"));		#  5
+	&xor	("ecx",&DWP(3,$tbl,$acc,8));	#  5
+	&movq	("mm3",QWP(16,$key));
+	&movz	($acc,&HB("ebx"));		# 15
+	&xor	("ecx",&DWP(1,$tbl,$acc,8));	# 15
+	&movd	("mm0","ecx");			# t[0] collected
+
+	&movz	($acc,&LB("eax"));		#  4
+	&mov	("ecx",&DWP(0,$tbl,$acc,8));	#  4
+	&movd	("eax","mm2");			#  7, 6, 3, 2
+	&movz	($acc,&LB("ebx"));		# 14
+	&xor	("ecx",&DWP(2,$tbl,$acc,8));	# 14
+	&movd	("ebx","mm6");			# 13,12, 9, 8
+
+	&movz	($acc,&HB("eax"));		#  3
+	&xor	("ecx",&DWP(1,$tbl,$acc,8));	#  3
+	&movz	($acc,&HB("ebx"));		#  9
+	&xor	("ecx",&DWP(3,$tbl,$acc,8));	#  9
+	&movd	("mm1","ecx");			# t[1] collected
+
+	&movz	($acc,&LB("eax"));		#  2
+	&mov	("ecx",&DWP(2,$tbl,$acc,8));	#  2
+	&shr	("eax",16);			#  7, 6
+	&punpckldq	("mm0","mm1");		# t[0,1] collected
+	&movz	($acc,&LB("ebx"));		#  8
+	&xor	("ecx",&DWP(0,$tbl,$acc,8));	#  8
+	&shr	("ebx",16);			# 13,12
+
+	&movz	($acc,&HB("eax"));		#  7
+	&xor	("ecx",&DWP(1,$tbl,$acc,8));	#  7
+	&pxor	("mm0","mm3");
+	&movz	("eax",&LB("eax"));		#  6
+	&xor	("edx",&DWP(2,$tbl,"eax",8));	#  6
+	&pshufw	("mm1","mm0",0x08);		#  5, 4, 1, 0
+	&movz	($acc,&HB("ebx"));		# 13
+	&xor	("ecx",&DWP(3,$tbl,$acc,8));	# 13
+	&xor	("ecx",&DWP(24,$key));		# t[2]
+	&movd	("mm4","ecx");			# t[2] collected
+	&movz	("ebx",&LB("ebx"));		# 12
+	&xor	("edx",&DWP(0,$tbl,"ebx",8));	# 12
+	&shr	("ecx",16);
+	&movd	("eax","mm1");			#  5, 4, 1, 0
+	&mov	("ebx",&DWP(28,$key));		# t[3]
+	&xor	("ebx","edx");
+	&movd	("mm5","ebx");			# t[3] collected
+	&and	("ebx",0xffff0000);
+	&or	("ebx","ecx");
+
+	&punpckldq	("mm4","mm5");		# t[2,3] collected
+}
+
+######################################################################
+# "Compact" block function
+######################################################################
+
+sub enccompact()
+{ my $Fn = mov;
+  while ($#_>5) { pop(@_); $Fn=sub{}; }
+  my ($i,$te,@s)=@_;
+  my $tmp = $key;
+  my $out = $i==3?$s[0]:$acc;
+
+	# $Fn is used in first compact round and its purpose is to
+	# void restoration of some values from stack, so that after
+	# 4xenccompact with extra argument $key value is left there...
+	if ($i==3)  {	&$Fn	($key,$__key);			}##%edx
+	else        {	&mov	($out,$s[0]);			}
+			&and	($out,0xFF);
+	if ($i==1)  {	&shr	($s[0],16);			}#%ebx[1]
+	if ($i==2)  {	&shr	($s[0],24);			}#%ecx[2]
+			&movz	($out,&BP(-128,$te,$out,1));
+
+	if ($i==3)  {	$tmp=$s[1];				}##%eax
+			&movz	($tmp,&HB($s[1]));
+			&movz	($tmp,&BP(-128,$te,$tmp,1));
+			&shl	($tmp,8);
+			&xor	($out,$tmp);
+
+	if ($i==3)  {	$tmp=$s[2]; &mov ($s[1],$__s0);		}##%ebx
+	else        {	&mov	($tmp,$s[2]);
+			&shr	($tmp,16);			}
+	if ($i==2)  {	&and	($s[1],0xFF);			}#%edx[2]
+			&and	($tmp,0xFF);
+			&movz	($tmp,&BP(-128,$te,$tmp,1));
+			&shl	($tmp,16);
+			&xor	($out,$tmp);
+
+	if ($i==3)  {	$tmp=$s[3]; &mov ($s[2],$__s1);		}##%ecx
+	elsif($i==2){	&movz	($tmp,&HB($s[3]));		}#%ebx[2]
+	else        {	&mov	($tmp,$s[3]);
+			&shr	($tmp,24);			}
+			&movz	($tmp,&BP(-128,$te,$tmp,1));
+			&shl	($tmp,24);
+			&xor	($out,$tmp);
+	if ($i<2)   {	&mov	(&DWP(4+4*$i,"esp"),$out);	}
+	if ($i==3)  {	&mov	($s[3],$acc);			}
+	&comment();
+}
+
+sub enctransform()
+{ my @s = ($s0,$s1,$s2,$s3);
+  my $i = shift;
+  my $tmp = $tbl;
+  my $r2  = $key ;
+
+	&mov	($acc,$s[$i]);
+	&and	($acc,0x80808080);
+	&mov	($tmp,$acc);
+	&shr	($tmp,7);
+	&lea	($r2,&DWP(0,$s[$i],$s[$i]));
+	&sub	($acc,$tmp);
+	&and	($r2,0xfefefefe);
+	&and	($acc,0x1b1b1b1b);
+	&mov	($tmp,$s[$i]);
+	&xor	($acc,$r2);	# r2
+
+	&xor	($s[$i],$acc);	# r0 ^ r2
+	&rotl	($s[$i],24);
+	&xor	($s[$i],$acc)	# ROTATE(r2^r0,24) ^ r2
+	&rotr	($tmp,16);
+	&xor	($s[$i],$tmp);
+	&rotr	($tmp,8);
+	&xor	($s[$i],$tmp);
+}
+
+&function_begin_B("_x86_AES_encrypt_compact");
+	# note that caller is expected to allocate stack frame for me!
+	&mov	($__key,$key);			# save key
+
+	&xor	($s0,&DWP(0,$key));		# xor with key
+	&xor	($s1,&DWP(4,$key));
+	&xor	($s2,&DWP(8,$key));
+	&xor	($s3,&DWP(12,$key));
+
+	&mov	($acc,&DWP(240,$key));		# load key->rounds
+	&lea	($acc,&DWP(-2,$acc,$acc));
+	&lea	($acc,&DWP(0,$key,$acc,8));
+	&mov	($__end,$acc);			# end of key schedule
+
+	# prefetch Te4
+	&mov	($key,&DWP(0-128,$tbl));
+	&mov	($acc,&DWP(32-128,$tbl));
+	&mov	($key,&DWP(64-128,$tbl));
+	&mov	($acc,&DWP(96-128,$tbl));
+	&mov	($key,&DWP(128-128,$tbl));
+	&mov	($acc,&DWP(160-128,$tbl));
+	&mov	($key,&DWP(192-128,$tbl));
+	&mov	($acc,&DWP(224-128,$tbl));
+
+	&set_label("loop",16);
+
+		&enccompact(0,$tbl,$s0,$s1,$s2,$s3,1);
+		&enccompact(1,$tbl,$s1,$s2,$s3,$s0,1);
+		&enccompact(2,$tbl,$s2,$s3,$s0,$s1,1);
+		&enccompact(3,$tbl,$s3,$s0,$s1,$s2,1);
+		&enctransform(2);
+		&enctransform(3);
+		&enctransform(0);
+		&enctransform(1);
+		&mov 	($key,$__key);
+		&mov	($tbl,$__tbl);
+		&add	($key,16);		# advance rd_key
+		&xor	($s0,&DWP(0,$key));
+		&xor	($s1,&DWP(4,$key));
+		&xor	($s2,&DWP(8,$key));
+		&xor	($s3,&DWP(12,$key));
+
+	&cmp	($key,$__end);
+	&mov	($__key,$key);
+	&jb	(&label("loop"));
+
+	&enccompact(0,$tbl,$s0,$s1,$s2,$s3);
+	&enccompact(1,$tbl,$s1,$s2,$s3,$s0);
+	&enccompact(2,$tbl,$s2,$s3,$s0,$s1);
+	&enccompact(3,$tbl,$s3,$s0,$s1,$s2);
+
+	&xor	($s0,&DWP(16,$key));
+	&xor	($s1,&DWP(20,$key));
+	&xor	($s2,&DWP(24,$key));
+	&xor	($s3,&DWP(28,$key));
+
+	&ret	();
+&function_end_B("_x86_AES_encrypt_compact");
+
+######################################################################
+# "Compact" SSE block function.
+######################################################################
+#
+# Performance is not actually extraordinary in comparison to pure
+# x86 code. In particular encrypt performance is virtually the same.
+# Decrypt performance on the other hand is 15-20% better on newer
+# µ-archs [but we're thankful for *any* improvement here], and ~50%
+# better on PIII:-) And additionally on the pros side this code
+# eliminates redundant references to stack and thus relieves/
+# minimizes the pressure on the memory bus.
+#
+# MMX register layout                           lsb
+# +--+--+--+--+--+--+--+--+--+--+--+--+--+--+--+--+
+# |          mm4          |          mm0          |
+# +--+--+--+--+--+--+--+--+--+--+--+--+--+--+--+--+
+# |     s3    |     s2    |     s1    |     s0    |    
+# +--+--+--+--+--+--+--+--+--+--+--+--+--+--+--+--+
+# |15|14|13|12|11|10| 9| 8| 7| 6| 5| 4| 3| 2| 1| 0|
+# +--+--+--+--+--+--+--+--+--+--+--+--+--+--+--+--+
+#
+# Indexes translate as s[N/4]>>(8*(N%4)), e.g. 5 means s1>>8.
+# In this terms encryption and decryption "compact" permutation
+# matrices can be depicted as following:
+#
+# encryption              lsb	# decryption              lsb
+# +----++----+----+----+----+	# +----++----+----+----+----+
+# | t0 || 15 | 10 |  5 |  0 |	# | t0 ||  7 | 10 | 13 |  0 |
+# +----++----+----+----+----+	# +----++----+----+----+----+
+# | t1 ||  3 | 14 |  9 |  4 |	# | t1 || 11 | 14 |  1 |  4 |
+# +----++----+----+----+----+	# +----++----+----+----+----+
+# | t2 ||  7 |  2 | 13 |  8 |	# | t2 || 15 |  2 |  5 |  8 |
+# +----++----+----+----+----+	# +----++----+----+----+----+
+# | t3 || 11 |  6 |  1 | 12 |	# | t3 ||  3 |  6 |  9 | 12 |
+# +----++----+----+----+----+	# +----++----+----+----+----+
+#
+######################################################################
+# Why not xmm registers? Short answer. It was actually tested and
+# was not any faster, but *contrary*, most notably on Intel CPUs.
+# Longer answer. Main advantage of using mm registers is that movd
+# latency is lower, especially on Intel P4. While arithmetic
+# instructions are twice as many, they can be scheduled every cycle
+# and not every second one when they are operating on xmm register,
+# so that "arithmetic throughput" remains virtually the same. And
+# finally the code can be executed even on elder SSE-only CPUs:-)
+
+sub sse_enccompact()
+{
+	&pshufw	("mm1","mm0",0x08);		#  5, 4, 1, 0
+	&pshufw	("mm5","mm4",0x0d);		# 15,14,11,10
+	&movd	("eax","mm1");			#  5, 4, 1, 0
+	&movd	("ebx","mm5");			# 15,14,11,10
+
+	&movz	($acc,&LB("eax"));		#  0
+	&movz	("ecx",&BP(-128,$tbl,$acc,1));	#  0
+	&pshufw	("mm2","mm0",0x0d);		#  7, 6, 3, 2
+	&movz	("edx",&HB("eax"));		#  1
+	&movz	("edx",&BP(-128,$tbl,"edx",1));	#  1
+	&shl	("edx",8);			#  1
+	&shr	("eax",16);			#  5, 4
+
+	&movz	($acc,&LB("ebx"));		# 10
+	&movz	($acc,&BP(-128,$tbl,$acc,1));	# 10
+	&shl	($acc,16);			# 10
+	&or	("ecx",$acc);			# 10
+	&pshufw	("mm6","mm4",0x08);		# 13,12, 9, 8
+	&movz	($acc,&HB("ebx"));		# 11
+	&movz	($acc,&BP(-128,$tbl,$acc,1));	# 11
+	&shl	($acc,24);			# 11
+	&or	("edx",$acc);			# 11
+	&shr	("ebx",16);			# 15,14
+
+	&movz	($acc,&HB("eax"));		#  5
+	&movz	($acc,&BP(-128,$tbl,$acc,1));	#  5
+	&shl	($acc,8);			#  5
+	&or	("ecx",$acc);			#  5
+	&movz	($acc,&HB("ebx"));		# 15
+	&movz	($acc,&BP(-128,$tbl,$acc,1));	# 15
+	&shl	($acc,24);			# 15
+	&or	("ecx",$acc);			# 15
+	&movd	("mm0","ecx");			# t[0] collected
+
+	&movz	($acc,&LB("eax"));		#  4
+	&movz	("ecx",&BP(-128,$tbl,$acc,1));	#  4
+	&movd	("eax","mm2");			#  7, 6, 3, 2
+	&movz	($acc,&LB("ebx"));		# 14
+	&movz	($acc,&BP(-128,$tbl,$acc,1));	# 14
+	&shl	($acc,16);			# 14
+	&or	("ecx",$acc);			# 14
+
+	&movd	("ebx","mm6");			# 13,12, 9, 8
+	&movz	($acc,&HB("eax"));		#  3
+	&movz	($acc,&BP(-128,$tbl,$acc,1));	#  3
+	&shl	($acc,24);			#  3
+	&or	("ecx",$acc);			#  3
+	&movz	($acc,&HB("ebx"));		#  9
+	&movz	($acc,&BP(-128,$tbl,$acc,1));	#  9
+	&shl	($acc,8);			#  9
+	&or	("ecx",$acc);			#  9
+	&movd	("mm1","ecx");			# t[1] collected
+
+	&movz	($acc,&LB("ebx"));		#  8
+	&movz	("ecx",&BP(-128,$tbl,$acc,1));	#  8
+	&shr	("ebx",16);			# 13,12
+	&movz	($acc,&LB("eax"));		#  2
+	&movz	($acc,&BP(-128,$tbl,$acc,1));	#  2
+	&shl	($acc,16);			#  2
+	&or	("ecx",$acc);			#  2
+	&shr	("eax",16);			#  7, 6
+
+	&punpckldq	("mm0","mm1");		# t[0,1] collected
+
+	&movz	($acc,&HB("eax"));		#  7
+	&movz	($acc,&BP(-128,$tbl,$acc,1));	#  7
+	&shl	($acc,24);			#  7
+	&or	("ecx",$acc);			#  7
+	&and	("eax",0xff);			#  6
+	&movz	("eax",&BP(-128,$tbl,"eax",1));	#  6
+	&shl	("eax",16);			#  6
+	&or	("edx","eax");			#  6
+	&movz	($acc,&HB("ebx"));		# 13
+	&movz	($acc,&BP(-128,$tbl,$acc,1));	# 13
+	&shl	($acc,8);			# 13
+	&or	("ecx",$acc);			# 13
+	&movd	("mm4","ecx");			# t[2] collected
+	&and	("ebx",0xff);			# 12
+	&movz	("ebx",&BP(-128,$tbl,"ebx",1));	# 12
+	&or	("edx","ebx");			# 12
+	&movd	("mm5","edx");			# t[3] collected
+
+	&punpckldq	("mm4","mm5");		# t[2,3] collected
+}
+
+					if (!$x86only) {
+&function_begin_B("_sse_AES_encrypt_compact");
+	&pxor	("mm0",&QWP(0,$key));	#  7, 6, 5, 4, 3, 2, 1, 0
+	&pxor	("mm4",&QWP(8,$key));	# 15,14,13,12,11,10, 9, 8
+
+	# note that caller is expected to allocate stack frame for me!
+	&mov	($acc,&DWP(240,$key));		# load key->rounds
+	&lea	($acc,&DWP(-2,$acc,$acc));
+	&lea	($acc,&DWP(0,$key,$acc,8));
+	&mov	($__end,$acc);			# end of key schedule
+
+	&mov	($s0,0x1b1b1b1b);		# magic constant
+	&mov	(&DWP(8,"esp"),$s0);
+	&mov	(&DWP(12,"esp"),$s0);
+
+	# prefetch Te4
+	&mov	($s0,&DWP(0-128,$tbl));
+	&mov	($s1,&DWP(32-128,$tbl));
+	&mov	($s2,&DWP(64-128,$tbl));
+	&mov	($s3,&DWP(96-128,$tbl));
+	&mov	($s0,&DWP(128-128,$tbl));
+	&mov	($s1,&DWP(160-128,$tbl));
+	&mov	($s2,&DWP(192-128,$tbl));
+	&mov	($s3,&DWP(224-128,$tbl));
+
+	&set_label("loop",16);
+		&sse_enccompact();
+		&add	($key,16);
+		&cmp	($key,$__end);
+		&ja	(&label("out"));
+
+		&movq	("mm2",&QWP(8,"esp"));
+		&pxor	("mm3","mm3");		&pxor	("mm7","mm7");
+		&movq	("mm1","mm0");		&movq	("mm5","mm4");	# r0
+		&pcmpgtb("mm3","mm0");		&pcmpgtb("mm7","mm4");
+		&pand	("mm3","mm2");		&pand	("mm7","mm2");
+		&pshufw	("mm2","mm0",0xb1);	&pshufw	("mm6","mm4",0xb1);# ROTATE(r0,16)
+		&paddb	("mm0","mm0");		&paddb	("mm4","mm4");
+		&pxor	("mm0","mm3");		&pxor	("mm4","mm7");	# = r2
+		&pshufw	("mm3","mm2",0xb1);	&pshufw	("mm7","mm6",0xb1);# r0
+		&pxor	("mm1","mm0");		&pxor	("mm5","mm4");	# r0^r2
+		&pxor	("mm0","mm2");		&pxor	("mm4","mm6");	# ^= ROTATE(r0,16)
+
+		&movq	("mm2","mm3");		&movq	("mm6","mm7");
+		&pslld	("mm3",8);		&pslld	("mm7",8);
+		&psrld	("mm2",24);		&psrld	("mm6",24);
+		&pxor	("mm0","mm3");		&pxor	("mm4","mm7");	# ^= r0<<8
+		&pxor	("mm0","mm2");		&pxor	("mm4","mm6");	# ^= r0>>24
+
+		&movq	("mm3","mm1");		&movq	("mm7","mm5");
+		&movq	("mm2",&QWP(0,$key));	&movq	("mm6",&QWP(8,$key));
+		&psrld	("mm1",8);		&psrld	("mm5",8);
+		&mov	($s0,&DWP(0-128,$tbl));
+		&pslld	("mm3",24);		&pslld	("mm7",24);
+		&mov	($s1,&DWP(64-128,$tbl));
+		&pxor	("mm0","mm1");		&pxor	("mm4","mm5");	# ^= (r2^r0)<<8
+		&mov	($s2,&DWP(128-128,$tbl));
+		&pxor	("mm0","mm3");		&pxor	("mm4","mm7");	# ^= (r2^r0)>>24
+		&mov	($s3,&DWP(192-128,$tbl));
+
+		&pxor	("mm0","mm2");		&pxor	("mm4","mm6");
+	&jmp	(&label("loop"));
+
+	&set_label("out",16);
+	&pxor	("mm0",&QWP(0,$key));
+	&pxor	("mm4",&QWP(8,$key));
+
+	&ret	();
+&function_end_B("_sse_AES_encrypt_compact");
+					}
+
+######################################################################
+# Vanilla block function.
+######################################################################
+
 sub encstep()
 { my ($i,$te,@s) = @_;
   my $tmp = $key;
   my $out = $i==3?$s[0]:$acc;
 
 	# lines marked with #%e?x[i] denote "reordered" instructions...
-	if ($i==3)  {	&mov	($key,&DWP(12,"esp"));		}##%edx
+	if ($i==3)  {	&mov	($key,$__key);			}##%edx
 	else        {	&mov	($out,$s[0]);
 			&and	($out,0xFF);			}
 	if ($i==1)  {	&shr	($s[0],16);			}#%ebx[1]
@@ -191,14 +780,14 @@ sub encstep()
 			&movz	($tmp,&HB($s[1]));
 			&xor	($out,&DWP(3,$te,$tmp,8));
 
-	if ($i==3)  {	$tmp=$s[2]; &mov ($s[1],&DWP(4,"esp"));	}##%ebx
+	if ($i==3)  {	$tmp=$s[2]; &mov ($s[1],$__s0);		}##%ebx
 	else        {	&mov	($tmp,$s[2]);
 			&shr	($tmp,16);			}
 	if ($i==2)  {	&and	($s[1],0xFF);			}#%edx[2]
 			&and	($tmp,0xFF);
 			&xor	($out,&DWP(2,$te,$tmp,8));
 
-	if ($i==3)  {	$tmp=$s[3]; &mov ($s[2],&DWP(8,"esp"));	}##%ecx
+	if ($i==3)  {	$tmp=$s[3]; &mov ($s[2],$__s1);		}##%ecx
 	elsif($i==2){	&movz	($tmp,&HB($s[3]));		}#%ebx[2]
 	else        {	&mov	($tmp,$s[3]); 
 			&shr	($tmp,24)			}
@@ -213,7 +802,7 @@ sub enclast()
   my $tmp = $key;
   my $out = $i==3?$s[0]:$acc;
 
-	if ($i==3)  {	&mov	($key,&DWP(12,"esp"));		}##%edx
+	if ($i==3)  {	&mov	($key,$__key);			}##%edx
 	else        {	&mov	($out,$s[0]);			}
 			&and	($out,0xFF);
 	if ($i==1)  {	&shr	($s[0],16);			}#%ebx[1]
@@ -227,8 +816,8 @@ sub enclast()
 			&and	($tmp,0x0000ff00);
 			&xor	($out,$tmp);
 
-	if ($i==3)  {	$tmp=$s[2]; &mov ($s[1],&DWP(4,"esp"));	}##%ebx
-	else        {	mov	($tmp,$s[2]);
+	if ($i==3)  {	$tmp=$s[2]; &mov ($s[1],$__s0);		}##%ebx
+	else        {	&mov	($tmp,$s[2]);
 			&shr	($tmp,16);			}
 	if ($i==2)  {	&and	($s[1],0xFF);			}#%edx[2]
 			&and	($tmp,0xFF);
@@ -236,7 +825,7 @@ sub enclast()
 			&and	($tmp,0x00ff0000);
 			&xor	($out,$tmp);
 
-	if ($i==3)  {	$tmp=$s[3]; &mov ($s[2],&DWP(8,"esp"));	}##%ecx
+	if ($i==3)  {	$tmp=$s[3]; &mov ($s[2],$__s1);		}##%ecx
 	elsif($i==2){	&movz	($tmp,&HB($s[3]));		}#%ebx[2]
 	else        {	&mov	($tmp,$s[3]);
 			&shr	($tmp,24);			}
@@ -247,9 +836,6 @@ sub enclast()
 	if ($i==3)  {	&mov	($s[3],$acc);			}
 }
 
-sub _data_word() { my $i; while(defined($i=shift)) { &data_word($i,$i); } }
-
-&public_label("AES_Te");
 &function_begin_B("_x86_AES_encrypt");
 	if ($vertical_spin) {
 		# I need high parts of volatile registers to be accessible...
@@ -258,7 +844,7 @@ sub _data_word() { my $i; while(defined($i=shift)) { &data_word($i,$i); } }
 	}
 
 	# note that caller is expected to allocate stack frame for me!
-	&mov	(&DWP(12,"esp"),$key);		# save key
+	&mov	($__key,$key);			# save key
 
 	&xor	($s0,&DWP(0,$key));		# xor with key
 	&xor	($s1,&DWP(4,$key));
@@ -270,24 +856,24 @@ sub _data_word() { my $i; while(defined($i=shift)) { &data_word($i,$i); } }
 	if ($small_footprint) {
 	    &lea	($acc,&DWP(-2,$acc,$acc));
 	    &lea	($acc,&DWP(0,$key,$acc,8));
-	    &mov	(&DWP(16,"esp"),$acc);	# end of key schedule
-	    &align	(4);
-	    &set_label("loop");
+	    &mov	($__end,$acc);		# end of key schedule
+
+	    &set_label("loop",16);
 		if ($vertical_spin) {
-		    &encvert("ebp",$s0,$s1,$s2,$s3);
+		    &encvert($tbl,$s0,$s1,$s2,$s3);
 		} else {
-		    &encstep(0,"ebp",$s0,$s1,$s2,$s3);
-		    &encstep(1,"ebp",$s1,$s2,$s3,$s0);
-		    &encstep(2,"ebp",$s2,$s3,$s0,$s1);
-		    &encstep(3,"ebp",$s3,$s0,$s1,$s2);
+		    &encstep(0,$tbl,$s0,$s1,$s2,$s3);
+		    &encstep(1,$tbl,$s1,$s2,$s3,$s0);
+		    &encstep(2,$tbl,$s2,$s3,$s0,$s1);
+		    &encstep(3,$tbl,$s3,$s0,$s1,$s2);
 		}
 		&add	($key,16);		# advance rd_key
 		&xor	($s0,&DWP(0,$key));
 		&xor	($s1,&DWP(4,$key));
 		&xor	($s2,&DWP(8,$key));
 		&xor	($s3,&DWP(12,$key));
-	    &cmp	($key,&DWP(16,"esp"));
-	    &mov	(&DWP(12,"esp"),$key);
+	    &cmp	($key,$__end);
+	    &mov	($__key,$key);
 	    &jb		(&label("loop"));
 	}
 	else {
@@ -296,15 +882,15 @@ sub _data_word() { my $i; while(defined($i=shift)) { &data_word($i,$i); } }
 	    &cmp	($acc,12);
 	    &jle	(&label("12rounds"));
 
-	&set_label("14rounds");
+	&set_label("14rounds",4);
 	    for ($i=1;$i<3;$i++) {
 		if ($vertical_spin) {
-		    &encvert("ebp",$s0,$s1,$s2,$s3);
+		    &encvert($tbl,$s0,$s1,$s2,$s3);
 		} else {
-		    &encstep(0,"ebp",$s0,$s1,$s2,$s3);
-		    &encstep(1,"ebp",$s1,$s2,$s3,$s0);
-		    &encstep(2,"ebp",$s2,$s3,$s0,$s1);
-		    &encstep(3,"ebp",$s3,$s0,$s1,$s2);
+		    &encstep(0,$tbl,$s0,$s1,$s2,$s3);
+		    &encstep(1,$tbl,$s1,$s2,$s3,$s0);
+		    &encstep(2,$tbl,$s2,$s3,$s0,$s1);
+		    &encstep(3,$tbl,$s3,$s0,$s1,$s2);
 		}
 		&xor	($s0,&DWP(16*$i+0,$key));
 		&xor	($s1,&DWP(16*$i+4,$key));
@@ -312,16 +898,16 @@ sub _data_word() { my $i; while(defined($i=shift)) { &data_word($i,$i); } }
 		&xor	($s3,&DWP(16*$i+12,$key));
 	    }
 	    &add	($key,32);
-	    &mov	(&DWP(12,"esp"),$key);	# advance rd_key
-	&set_label("12rounds");
+	    &mov	($__key,$key);		# advance rd_key
+	&set_label("12rounds",4);
 	    for ($i=1;$i<3;$i++) {
 		if ($vertical_spin) {
-		    &encvert("ebp",$s0,$s1,$s2,$s3);
+		    &encvert($tbl,$s0,$s1,$s2,$s3);
 		} else {
-		    &encstep(0,"ebp",$s0,$s1,$s2,$s3);
-		    &encstep(1,"ebp",$s1,$s2,$s3,$s0);
-		    &encstep(2,"ebp",$s2,$s3,$s0,$s1);
-		    &encstep(3,"ebp",$s3,$s0,$s1,$s2);
+		    &encstep(0,$tbl,$s0,$s1,$s2,$s3);
+		    &encstep(1,$tbl,$s1,$s2,$s3,$s0);
+		    &encstep(2,$tbl,$s2,$s3,$s0,$s1);
+		    &encstep(3,$tbl,$s3,$s0,$s1,$s2);
 		}
 		&xor	($s0,&DWP(16*$i+0,$key));
 		&xor	($s1,&DWP(16*$i+4,$key));
@@ -329,16 +915,16 @@ sub _data_word() { my $i; while(defined($i=shift)) { &data_word($i,$i); } }
 		&xor	($s3,&DWP(16*$i+12,$key));
 	    }
 	    &add	($key,32);
-	    &mov	(&DWP(12,"esp"),$key);	# advance rd_key
-	&set_label("10rounds");
+	    &mov	($__key,$key);		# advance rd_key
+	&set_label("10rounds",4);
 	    for ($i=1;$i<10;$i++) {
 		if ($vertical_spin) {
-		    &encvert("ebp",$s0,$s1,$s2,$s3);
+		    &encvert($tbl,$s0,$s1,$s2,$s3);
 		} else {
-		    &encstep(0,"ebp",$s0,$s1,$s2,$s3);
-		    &encstep(1,"ebp",$s1,$s2,$s3,$s0);
-		    &encstep(2,"ebp",$s2,$s3,$s0,$s1);
-		    &encstep(3,"ebp",$s3,$s0,$s1,$s2);
+		    &encstep(0,$tbl,$s0,$s1,$s2,$s3);
+		    &encstep(1,$tbl,$s1,$s2,$s3,$s0);
+		    &encstep(2,$tbl,$s2,$s3,$s0,$s1);
+		    &encstep(3,$tbl,$s3,$s0,$s1,$s2);
 		}
 		&xor	($s0,&DWP(16*$i+0,$key));
 		&xor	($s1,&DWP(16*$i+4,$key));
@@ -352,10 +938,10 @@ sub _data_word() { my $i; while(defined($i=shift)) { &data_word($i,$i); } }
 	    &mov	($s1="ebx",$key="edi");
 	    &mov	($s2="ecx",$acc="esi");
 	}
-	&enclast(0,"ebp",$s0,$s1,$s2,$s3);
-	&enclast(1,"ebp",$s1,$s2,$s3,$s0);
-	&enclast(2,"ebp",$s2,$s3,$s0,$s1);
-	&enclast(3,"ebp",$s3,$s0,$s1,$s2);
+	&enclast(0,$tbl,$s0,$s1,$s2,$s3);
+	&enclast(1,$tbl,$s1,$s2,$s3,$s0);
+	&enclast(2,$tbl,$s2,$s3,$s0,$s1);
+	&enclast(3,$tbl,$s3,$s0,$s1,$s2);
 
 	&add	($key,$small_footprint?16:160);
 	&xor	($s0,&DWP(0,$key));
@@ -430,38 +1016,198 @@ sub _data_word() { my $i; while(defined($i=shift)) { &data_word($i,$i); } }
 	&_data_word(0xdabfbf65, 0x31e6e6d7, 0xc6424284, 0xb86868d0);
 	&_data_word(0xc3414182, 0xb0999929, 0x772d2d5a, 0x110f0f1e);
 	&_data_word(0xcbb0b07b, 0xfc5454a8, 0xd6bbbb6d, 0x3a16162c);
+
+#Te4	# four copies of Te4 to choose from to avoid L1 aliasing
+	&data_byte(0x63, 0x7c, 0x77, 0x7b, 0xf2, 0x6b, 0x6f, 0xc5);
+	&data_byte(0x30, 0x01, 0x67, 0x2b, 0xfe, 0xd7, 0xab, 0x76);
+	&data_byte(0xca, 0x82, 0xc9, 0x7d, 0xfa, 0x59, 0x47, 0xf0);
+	&data_byte(0xad, 0xd4, 0xa2, 0xaf, 0x9c, 0xa4, 0x72, 0xc0);
+	&data_byte(0xb7, 0xfd, 0x93, 0x26, 0x36, 0x3f, 0xf7, 0xcc);
+	&data_byte(0x34, 0xa5, 0xe5, 0xf1, 0x71, 0xd8, 0x31, 0x15);
+	&data_byte(0x04, 0xc7, 0x23, 0xc3, 0x18, 0x96, 0x05, 0x9a);
+	&data_byte(0x07, 0x12, 0x80, 0xe2, 0xeb, 0x27, 0xb2, 0x75);
+	&data_byte(0x09, 0x83, 0x2c, 0x1a, 0x1b, 0x6e, 0x5a, 0xa0);
+	&data_byte(0x52, 0x3b, 0xd6, 0xb3, 0x29, 0xe3, 0x2f, 0x84);
+	&data_byte(0x53, 0xd1, 0x00, 0xed, 0x20, 0xfc, 0xb1, 0x5b);
+	&data_byte(0x6a, 0xcb, 0xbe, 0x39, 0x4a, 0x4c, 0x58, 0xcf);
+	&data_byte(0xd0, 0xef, 0xaa, 0xfb, 0x43, 0x4d, 0x33, 0x85);
+	&data_byte(0x45, 0xf9, 0x02, 0x7f, 0x50, 0x3c, 0x9f, 0xa8);
+	&data_byte(0x51, 0xa3, 0x40, 0x8f, 0x92, 0x9d, 0x38, 0xf5);
+	&data_byte(0xbc, 0xb6, 0xda, 0x21, 0x10, 0xff, 0xf3, 0xd2);
+	&data_byte(0xcd, 0x0c, 0x13, 0xec, 0x5f, 0x97, 0x44, 0x17);
+	&data_byte(0xc4, 0xa7, 0x7e, 0x3d, 0x64, 0x5d, 0x19, 0x73);
+	&data_byte(0x60, 0x81, 0x4f, 0xdc, 0x22, 0x2a, 0x90, 0x88);
+	&data_byte(0x46, 0xee, 0xb8, 0x14, 0xde, 0x5e, 0x0b, 0xdb);
+	&data_byte(0xe0, 0x32, 0x3a, 0x0a, 0x49, 0x06, 0x24, 0x5c);
+	&data_byte(0xc2, 0xd3, 0xac, 0x62, 0x91, 0x95, 0xe4, 0x79);
+	&data_byte(0xe7, 0xc8, 0x37, 0x6d, 0x8d, 0xd5, 0x4e, 0xa9);
+	&data_byte(0x6c, 0x56, 0xf4, 0xea, 0x65, 0x7a, 0xae, 0x08);
+	&data_byte(0xba, 0x78, 0x25, 0x2e, 0x1c, 0xa6, 0xb4, 0xc6);
+	&data_byte(0xe8, 0xdd, 0x74, 0x1f, 0x4b, 0xbd, 0x8b, 0x8a);
+	&data_byte(0x70, 0x3e, 0xb5, 0x66, 0x48, 0x03, 0xf6, 0x0e);
+	&data_byte(0x61, 0x35, 0x57, 0xb9, 0x86, 0xc1, 0x1d, 0x9e);
+	&data_byte(0xe1, 0xf8, 0x98, 0x11, 0x69, 0xd9, 0x8e, 0x94);
+	&data_byte(0x9b, 0x1e, 0x87, 0xe9, 0xce, 0x55, 0x28, 0xdf);
+	&data_byte(0x8c, 0xa1, 0x89, 0x0d, 0xbf, 0xe6, 0x42, 0x68);
+	&data_byte(0x41, 0x99, 0x2d, 0x0f, 0xb0, 0x54, 0xbb, 0x16);
+
+	&data_byte(0x63, 0x7c, 0x77, 0x7b, 0xf2, 0x6b, 0x6f, 0xc5);
+	&data_byte(0x30, 0x01, 0x67, 0x2b, 0xfe, 0xd7, 0xab, 0x76);
+	&data_byte(0xca, 0x82, 0xc9, 0x7d, 0xfa, 0x59, 0x47, 0xf0);
+	&data_byte(0xad, 0xd4, 0xa2, 0xaf, 0x9c, 0xa4, 0x72, 0xc0);
+	&data_byte(0xb7, 0xfd, 0x93, 0x26, 0x36, 0x3f, 0xf7, 0xcc);
+	&data_byte(0x34, 0xa5, 0xe5, 0xf1, 0x71, 0xd8, 0x31, 0x15);
+	&data_byte(0x04, 0xc7, 0x23, 0xc3, 0x18, 0x96, 0x05, 0x9a);
+	&data_byte(0x07, 0x12, 0x80, 0xe2, 0xeb, 0x27, 0xb2, 0x75);
+	&data_byte(0x09, 0x83, 0x2c, 0x1a, 0x1b, 0x6e, 0x5a, 0xa0);
+	&data_byte(0x52, 0x3b, 0xd6, 0xb3, 0x29, 0xe3, 0x2f, 0x84);
+	&data_byte(0x53, 0xd1, 0x00, 0xed, 0x20, 0xfc, 0xb1, 0x5b);
+	&data_byte(0x6a, 0xcb, 0xbe, 0x39, 0x4a, 0x4c, 0x58, 0xcf);
+	&data_byte(0xd0, 0xef, 0xaa, 0xfb, 0x43, 0x4d, 0x33, 0x85);
+	&data_byte(0x45, 0xf9, 0x02, 0x7f, 0x50, 0x3c, 0x9f, 0xa8);
+	&data_byte(0x51, 0xa3, 0x40, 0x8f, 0x92, 0x9d, 0x38, 0xf5);
+	&data_byte(0xbc, 0xb6, 0xda, 0x21, 0x10, 0xff, 0xf3, 0xd2);
+	&data_byte(0xcd, 0x0c, 0x13, 0xec, 0x5f, 0x97, 0x44, 0x17);
+	&data_byte(0xc4, 0xa7, 0x7e, 0x3d, 0x64, 0x5d, 0x19, 0x73);
+	&data_byte(0x60, 0x81, 0x4f, 0xdc, 0x22, 0x2a, 0x90, 0x88);
+	&data_byte(0x46, 0xee, 0xb8, 0x14, 0xde, 0x5e, 0x0b, 0xdb);
+	&data_byte(0xe0, 0x32, 0x3a, 0x0a, 0x49, 0x06, 0x24, 0x5c);
+	&data_byte(0xc2, 0xd3, 0xac, 0x62, 0x91, 0x95, 0xe4, 0x79);
+	&data_byte(0xe7, 0xc8, 0x37, 0x6d, 0x8d, 0xd5, 0x4e, 0xa9);
+	&data_byte(0x6c, 0x56, 0xf4, 0xea, 0x65, 0x7a, 0xae, 0x08);
+	&data_byte(0xba, 0x78, 0x25, 0x2e, 0x1c, 0xa6, 0xb4, 0xc6);
+	&data_byte(0xe8, 0xdd, 0x74, 0x1f, 0x4b, 0xbd, 0x8b, 0x8a);
+	&data_byte(0x70, 0x3e, 0xb5, 0x66, 0x48, 0x03, 0xf6, 0x0e);
+	&data_byte(0x61, 0x35, 0x57, 0xb9, 0x86, 0xc1, 0x1d, 0x9e);
+	&data_byte(0xe1, 0xf8, 0x98, 0x11, 0x69, 0xd9, 0x8e, 0x94);
+	&data_byte(0x9b, 0x1e, 0x87, 0xe9, 0xce, 0x55, 0x28, 0xdf);
+	&data_byte(0x8c, 0xa1, 0x89, 0x0d, 0xbf, 0xe6, 0x42, 0x68);
+	&data_byte(0x41, 0x99, 0x2d, 0x0f, 0xb0, 0x54, 0xbb, 0x16);
+
+	&data_byte(0x63, 0x7c, 0x77, 0x7b, 0xf2, 0x6b, 0x6f, 0xc5);
+	&data_byte(0x30, 0x01, 0x67, 0x2b, 0xfe, 0xd7, 0xab, 0x76);
+	&data_byte(0xca, 0x82, 0xc9, 0x7d, 0xfa, 0x59, 0x47, 0xf0);
+	&data_byte(0xad, 0xd4, 0xa2, 0xaf, 0x9c, 0xa4, 0x72, 0xc0);
+	&data_byte(0xb7, 0xfd, 0x93, 0x26, 0x36, 0x3f, 0xf7, 0xcc);
+	&data_byte(0x34, 0xa5, 0xe5, 0xf1, 0x71, 0xd8, 0x31, 0x15);
+	&data_byte(0x04, 0xc7, 0x23, 0xc3, 0x18, 0x96, 0x05, 0x9a);
+	&data_byte(0x07, 0x12, 0x80, 0xe2, 0xeb, 0x27, 0xb2, 0x75);
+	&data_byte(0x09, 0x83, 0x2c, 0x1a, 0x1b, 0x6e, 0x5a, 0xa0);
+	&data_byte(0x52, 0x3b, 0xd6, 0xb3, 0x29, 0xe3, 0x2f, 0x84);
+	&data_byte(0x53, 0xd1, 0x00, 0xed, 0x20, 0xfc, 0xb1, 0x5b);
+	&data_byte(0x6a, 0xcb, 0xbe, 0x39, 0x4a, 0x4c, 0x58, 0xcf);
+	&data_byte(0xd0, 0xef, 0xaa, 0xfb, 0x43, 0x4d, 0x33, 0x85);
+	&data_byte(0x45, 0xf9, 0x02, 0x7f, 0x50, 0x3c, 0x9f, 0xa8);
+	&data_byte(0x51, 0xa3, 0x40, 0x8f, 0x92, 0x9d, 0x38, 0xf5);
+	&data_byte(0xbc, 0xb6, 0xda, 0x21, 0x10, 0xff, 0xf3, 0xd2);
+	&data_byte(0xcd, 0x0c, 0x13, 0xec, 0x5f, 0x97, 0x44, 0x17);
+	&data_byte(0xc4, 0xa7, 0x7e, 0x3d, 0x64, 0x5d, 0x19, 0x73);
+	&data_byte(0x60, 0x81, 0x4f, 0xdc, 0x22, 0x2a, 0x90, 0x88);
+	&data_byte(0x46, 0xee, 0xb8, 0x14, 0xde, 0x5e, 0x0b, 0xdb);
+	&data_byte(0xe0, 0x32, 0x3a, 0x0a, 0x49, 0x06, 0x24, 0x5c);
+	&data_byte(0xc2, 0xd3, 0xac, 0x62, 0x91, 0x95, 0xe4, 0x79);
+	&data_byte(0xe7, 0xc8, 0x37, 0x6d, 0x8d, 0xd5, 0x4e, 0xa9);
+	&data_byte(0x6c, 0x56, 0xf4, 0xea, 0x65, 0x7a, 0xae, 0x08);
+	&data_byte(0xba, 0x78, 0x25, 0x2e, 0x1c, 0xa6, 0xb4, 0xc6);
+	&data_byte(0xe8, 0xdd, 0x74, 0x1f, 0x4b, 0xbd, 0x8b, 0x8a);
+	&data_byte(0x70, 0x3e, 0xb5, 0x66, 0x48, 0x03, 0xf6, 0x0e);
+	&data_byte(0x61, 0x35, 0x57, 0xb9, 0x86, 0xc1, 0x1d, 0x9e);
+	&data_byte(0xe1, 0xf8, 0x98, 0x11, 0x69, 0xd9, 0x8e, 0x94);
+	&data_byte(0x9b, 0x1e, 0x87, 0xe9, 0xce, 0x55, 0x28, 0xdf);
+	&data_byte(0x8c, 0xa1, 0x89, 0x0d, 0xbf, 0xe6, 0x42, 0x68);
+	&data_byte(0x41, 0x99, 0x2d, 0x0f, 0xb0, 0x54, 0xbb, 0x16);
+
+	&data_byte(0x63, 0x7c, 0x77, 0x7b, 0xf2, 0x6b, 0x6f, 0xc5);
+	&data_byte(0x30, 0x01, 0x67, 0x2b, 0xfe, 0xd7, 0xab, 0x76);
+	&data_byte(0xca, 0x82, 0xc9, 0x7d, 0xfa, 0x59, 0x47, 0xf0);
+	&data_byte(0xad, 0xd4, 0xa2, 0xaf, 0x9c, 0xa4, 0x72, 0xc0);
+	&data_byte(0xb7, 0xfd, 0x93, 0x26, 0x36, 0x3f, 0xf7, 0xcc);
+	&data_byte(0x34, 0xa5, 0xe5, 0xf1, 0x71, 0xd8, 0x31, 0x15);
+	&data_byte(0x04, 0xc7, 0x23, 0xc3, 0x18, 0x96, 0x05, 0x9a);
+	&data_byte(0x07, 0x12, 0x80, 0xe2, 0xeb, 0x27, 0xb2, 0x75);
+	&data_byte(0x09, 0x83, 0x2c, 0x1a, 0x1b, 0x6e, 0x5a, 0xa0);
+	&data_byte(0x52, 0x3b, 0xd6, 0xb3, 0x29, 0xe3, 0x2f, 0x84);
+	&data_byte(0x53, 0xd1, 0x00, 0xed, 0x20, 0xfc, 0xb1, 0x5b);
+	&data_byte(0x6a, 0xcb, 0xbe, 0x39, 0x4a, 0x4c, 0x58, 0xcf);
+	&data_byte(0xd0, 0xef, 0xaa, 0xfb, 0x43, 0x4d, 0x33, 0x85);
+	&data_byte(0x45, 0xf9, 0x02, 0x7f, 0x50, 0x3c, 0x9f, 0xa8);
+	&data_byte(0x51, 0xa3, 0x40, 0x8f, 0x92, 0x9d, 0x38, 0xf5);
+	&data_byte(0xbc, 0xb6, 0xda, 0x21, 0x10, 0xff, 0xf3, 0xd2);
+	&data_byte(0xcd, 0x0c, 0x13, 0xec, 0x5f, 0x97, 0x44, 0x17);
+	&data_byte(0xc4, 0xa7, 0x7e, 0x3d, 0x64, 0x5d, 0x19, 0x73);
+	&data_byte(0x60, 0x81, 0x4f, 0xdc, 0x22, 0x2a, 0x90, 0x88);
+	&data_byte(0x46, 0xee, 0xb8, 0x14, 0xde, 0x5e, 0x0b, 0xdb);
+	&data_byte(0xe0, 0x32, 0x3a, 0x0a, 0x49, 0x06, 0x24, 0x5c);
+	&data_byte(0xc2, 0xd3, 0xac, 0x62, 0x91, 0x95, 0xe4, 0x79);
+	&data_byte(0xe7, 0xc8, 0x37, 0x6d, 0x8d, 0xd5, 0x4e, 0xa9);
+	&data_byte(0x6c, 0x56, 0xf4, 0xea, 0x65, 0x7a, 0xae, 0x08);
+	&data_byte(0xba, 0x78, 0x25, 0x2e, 0x1c, 0xa6, 0xb4, 0xc6);
+	&data_byte(0xe8, 0xdd, 0x74, 0x1f, 0x4b, 0xbd, 0x8b, 0x8a);
+	&data_byte(0x70, 0x3e, 0xb5, 0x66, 0x48, 0x03, 0xf6, 0x0e);
+	&data_byte(0x61, 0x35, 0x57, 0xb9, 0x86, 0xc1, 0x1d, 0x9e);
+	&data_byte(0xe1, 0xf8, 0x98, 0x11, 0x69, 0xd9, 0x8e, 0x94);
+	&data_byte(0x9b, 0x1e, 0x87, 0xe9, 0xce, 0x55, 0x28, 0xdf);
+	&data_byte(0x8c, 0xa1, 0x89, 0x0d, 0xbf, 0xe6, 0x42, 0x68);
+	&data_byte(0x41, 0x99, 0x2d, 0x0f, 0xb0, 0x54, 0xbb, 0x16);
 #rcon:
 	&data_word(0x00000001, 0x00000002, 0x00000004, 0x00000008);
 	&data_word(0x00000010, 0x00000020, 0x00000040, 0x00000080);
-	&data_word(0x0000001b, 0x00000036, 0, 0, 0, 0, 0, 0);
+	&data_word(0x0000001b, 0x00000036, 0x00000000, 0x00000000);
+	&data_word(0x00000000, 0x00000000, 0x00000000, 0x00000000);
 &function_end_B("_x86_AES_encrypt");
 
 # void AES_encrypt (const void *inp,void *out,const AES_KEY *key);
-&public_label("AES_Te");
 &function_begin("AES_encrypt");
 	&mov	($acc,&wparam(0));		# load inp
 	&mov	($key,&wparam(2));		# load key
 
 	&mov	($s0,"esp");
-	&sub	("esp",24);
-	&and	("esp",-64);
-	&add	("esp",4);
-	&mov	(&DWP(16,"esp"),$s0);
+	&sub	("esp",36);
+	&and	("esp",-64);			# align to cache-line
+
+	# place stack frame just "above" the key schedule
+	&lea	($s1,&DWP(-64-63,$key));
+	&sub	($s1,"esp");
+	&neg	($s1);
+	&and	($s1,0x3C0);	# modulo 1024, but aligned to cache-line
+	&sub	("esp",$s1);
+	&add	("esp",4);	# 4 is reserved for caller's return address
+	&mov	($_esp,$s0);			# save stack pointer
 
 	&call   (&label("pic_point"));          # make it PIC!
 	&set_label("pic_point");
-	&blindpop("ebp");
-	&lea    ("ebp",&DWP(&label("AES_Te")."-".&label("pic_point"),"ebp"));
-
+	&blindpop($tbl);
+	&picmeup($s0,"OPENSSL_ia32cap_P",$tbl,&label("pic_point")) if (!$x86only);
+	&lea    ($tbl,&DWP(&label("AES_Te")."-".&label("pic_point"),$tbl));
+
+	# pick Te4 copy which can't "overlap" with stack frame or key schedule
+	&lea	($s1,&DWP(768-4,"esp"));
+	&sub	($s1,$tbl);
+	&and	($s1,0x300);
+	&lea	($tbl,&DWP(2048+128,$tbl,$s1));
+
+					if (!$x86only) {
+	&bt	(&DWP(0,$s0),25);	# check for SSE bit
+	&jnc	(&label("x86"));
+
+	&movq	("mm0",&QWP(0,$acc));
+	&movq	("mm4",&QWP(8,$acc));
+	&call	("_sse_AES_encrypt_compact");
+	&mov	("esp",$_esp);			# restore stack pointer
+	&mov	($acc,&wparam(1));		# load out
+	&movq	(&QWP(0,$acc),"mm0");		# write output data
+	&movq	(&QWP(8,$acc),"mm4");
+	&emms	();
+	&function_end_A();
+					}
+	&set_label("x86",16);
+	&mov	($_tbl,$tbl);
 	&mov	($s0,&DWP(0,$acc));		# load input data
 	&mov	($s1,&DWP(4,$acc));
 	&mov	($s2,&DWP(8,$acc));
 	&mov	($s3,&DWP(12,$acc));
-
-	&call	("_x86_AES_encrypt");
-
-	&mov	("esp",&DWP(16,"esp"));
-
+	&call	("_x86_AES_encrypt_compact");
+	&mov	("esp",$_esp);			# restore stack pointer
 	&mov	($acc,&wparam(1));		# load out
 	&mov	(&DWP(0,$acc),$s0);		# write output data
 	&mov	(&DWP(4,$acc),$s1);
@@ -469,7 +1215,370 @@ sub _data_word() { my $i; while(defined($i=shift)) { &data_word($i,$i); } }
 	&mov	(&DWP(12,$acc),$s3);
 &function_end("AES_encrypt");
 
-#------------------------------------------------------------------#
+#--------------------------------------------------------------------#
+
+######################################################################
+# "Compact" block function
+######################################################################
+
+sub deccompact()
+{ my $Fn = mov;
+  while ($#_>5) { pop(@_); $Fn=sub{}; }
+  my ($i,$td,@s)=@_;
+  my $tmp = $key;
+  my $out = $i==3?$s[0]:$acc;
+
+	# $Fn is used in first compact round and its purpose is to
+	# void restoration of some values from stack, so that after
+	# 4xdeccompact with extra argument $key, $s0 and $s1 values
+	# are left there...
+	if($i==3)   {	&$Fn	($key,$__key);			}
+	else        {	&mov	($out,$s[0]);			}
+			&and	($out,0xFF);
+			&movz	($out,&BP(-128,$td,$out,1));
+
+	if ($i==3)  {	$tmp=$s[1];				}
+			&movz	($tmp,&HB($s[1]));
+			&movz	($tmp,&BP(-128,$td,$tmp,1));
+			&shl	($tmp,8);
+			&xor	($out,$tmp);
+
+	if ($i==3)  {	$tmp=$s[2]; &mov ($s[1],$acc);		}
+	else        {	mov	($tmp,$s[2]);			}
+			&shr	($tmp,16);
+			&and	($tmp,0xFF);
+			&movz	($tmp,&BP(-128,$td,$tmp,1));
+			&shl	($tmp,16);
+			&xor	($out,$tmp);
+
+	if ($i==3)  {	$tmp=$s[3]; &$Fn ($s[2],$__s1);		}
+	else        {	&mov	($tmp,$s[3]);			}
+			&shr	($tmp,24);
+			&movz	($tmp,&BP(-128,$td,$tmp,1));
+			&shl	($tmp,24);
+			&xor	($out,$tmp);
+	if ($i<2)   {	&mov	(&DWP(4+4*$i,"esp"),$out);	}
+	if ($i==3)  {	&$Fn	($s[3],$__s0);			}
+}
+
+# must be called with 2,3,0,1 as argument sequence!!!
+sub dectransform()
+{ my @s = ($s0,$s1,$s2,$s3);
+  my $i = shift;
+  my $tmp = $key;
+  my $tp2 = @s[($i+2)%4]; $tp2 = @s[2] if ($i==1);
+  my $tp4 = @s[($i+3)%4]; $tp4 = @s[3] if ($i==1);
+  my $tp8 = $tbl;
+
+	&mov	($acc,$s[$i]);
+	&and	($acc,0x80808080);
+	&mov	($tmp,$acc);
+	&shr	($tmp,7);
+	&lea	($tp2,&DWP(0,$s[$i],$s[$i]));
+	&sub	($acc,$tmp);
+	&and	($tp2,0xfefefefe);
+	&and	($acc,0x1b1b1b1b);
+	&xor	($acc,$tp2);
+	&mov	($tp2,$acc);
+
+	&and	($acc,0x80808080);
+	&mov	($tmp,$acc);
+	&shr	($tmp,7);
+	&lea	($tp4,&DWP(0,$tp2,$tp2));
+	&sub	($acc,$tmp);
+	&and	($tp4,0xfefefefe);
+	&and	($acc,0x1b1b1b1b);
+	 &xor	($tp2,$s[$i]);	# tp2^tp1
+	&xor	($acc,$tp4);
+	&mov	($tp4,$acc);
+
+	&and	($acc,0x80808080);
+	&mov	($tmp,$acc);
+	&shr	($tmp,7);
+	&lea	($tp8,&DWP(0,$tp4,$tp4));
+	&sub	($acc,$tmp);
+	&and	($tp8,0xfefefefe);
+	&and	($acc,0x1b1b1b1b);
+	 &xor	($tp4,$s[$i]);	# tp4^tp1
+	 &rotl	($s[$i],8);	# = ROTATE(tp1,8)
+	&xor	($tp8,$acc);
+
+	&xor	($s[$i],$tp2);
+	&xor	($tp2,$tp8);
+	&rotl	($tp2,24);
+	&xor	($s[$i],$tp4);
+	&xor	($tp4,$tp8);
+	&rotl	($tp4,16);
+	&xor	($s[$i],$tp8);	# ^= tp8^(tp4^tp1)^(tp2^tp1)
+	&rotl	($tp8,8);
+	&xor	($s[$i],$tp2);	# ^= ROTATE(tp8^tp2^tp1,24)
+	&xor	($s[$i],$tp4);	# ^= ROTATE(tp8^tp4^tp1,16)
+	 &mov	($s[0],$__s0)			if($i==2); #prefetch $s0
+	 &mov	($s[1],$__s1)			if($i==3); #prefetch $s1
+	 &mov	($s[2],$__s2)			if($i==1);
+	&xor	($s[$i],$tp8);	# ^= ROTATE(tp8,8)
+
+	&mov	($s[3],$__s3)			if($i==1);
+	&mov	(&DWP(4+4*$i,"esp"),$s[$i])	if($i>=2);
+}
+
+&function_begin_B("_x86_AES_decrypt_compact");
+	# note that caller is expected to allocate stack frame for me!
+	&mov	($__key,$key);			# save key
+
+	&xor	($s0,&DWP(0,$key));		# xor with key
+	&xor	($s1,&DWP(4,$key));
+	&xor	($s2,&DWP(8,$key));
+	&xor	($s3,&DWP(12,$key));
+
+	&mov	($acc,&DWP(240,$key));		# load key->rounds
+
+	&lea	($acc,&DWP(-2,$acc,$acc));
+	&lea	($acc,&DWP(0,$key,$acc,8));
+	&mov	($__end,$acc);			# end of key schedule
+
+	# prefetch Td4
+	&mov	($key,&DWP(0-128,$tbl));
+	&mov	($acc,&DWP(32-128,$tbl));
+	&mov	($key,&DWP(64-128,$tbl));
+	&mov	($acc,&DWP(96-128,$tbl));
+	&mov	($key,&DWP(128-128,$tbl));
+	&mov	($acc,&DWP(160-128,$tbl));
+	&mov	($key,&DWP(192-128,$tbl));
+	&mov	($acc,&DWP(224-128,$tbl));
+
+	&set_label("loop",16);
+
+		&deccompact(0,$tbl,$s0,$s3,$s2,$s1,1);
+		&deccompact(1,$tbl,$s1,$s0,$s3,$s2,1);
+		&deccompact(2,$tbl,$s2,$s1,$s0,$s3,1);
+		&deccompact(3,$tbl,$s3,$s2,$s1,$s0,1);
+		&dectransform(2);
+		&dectransform(3);
+		&dectransform(0);
+		&dectransform(1);
+		&mov 	($key,$__key);
+		&mov	($tbl,$__tbl);
+		&add	($key,16);		# advance rd_key
+		&xor	($s0,&DWP(0,$key));
+		&xor	($s1,&DWP(4,$key));
+		&xor	($s2,&DWP(8,$key));
+		&xor	($s3,&DWP(12,$key));
+
+	&cmp	($key,$__end);
+	&mov	($__key,$key);
+	&jb	(&label("loop"));
+
+	&deccompact(0,$tbl,$s0,$s3,$s2,$s1);
+	&deccompact(1,$tbl,$s1,$s0,$s3,$s2);
+	&deccompact(2,$tbl,$s2,$s1,$s0,$s3);
+	&deccompact(3,$tbl,$s3,$s2,$s1,$s0);
+
+	&xor	($s0,&DWP(16,$key));
+	&xor	($s1,&DWP(20,$key));
+	&xor	($s2,&DWP(24,$key));
+	&xor	($s3,&DWP(28,$key));
+
+	&ret	();
+&function_end_B("_x86_AES_decrypt_compact");
+
+######################################################################
+# "Compact" SSE block function.
+######################################################################
+
+sub sse_deccompact()
+{
+	&pshufw	("mm1","mm0",0x0c);		#  7, 6, 1, 0
+	&movd	("eax","mm1");			#  7, 6, 1, 0
+
+	&pshufw	("mm5","mm4",0x09);		# 13,12,11,10
+	&movz	($acc,&LB("eax"));		#  0
+	&movz	("ecx",&BP(-128,$tbl,$acc,1));	#  0
+	&movd	("ebx","mm5");			# 13,12,11,10
+	&movz	("edx",&HB("eax"));		#  1
+	&movz	("edx",&BP(-128,$tbl,"edx",1));	#  1
+	&shl	("edx",8);			#  1
+
+	&pshufw	("mm2","mm0",0x06);		#  3, 2, 5, 4
+	&movz	($acc,&LB("ebx"));		# 10
+	&movz	($acc,&BP(-128,$tbl,$acc,1));	# 10
+	&shl	($acc,16);			# 10
+	&or	("ecx",$acc);			# 10
+	&shr	("eax",16);			#  7, 6
+	&movz	($acc,&HB("ebx"));		# 11
+	&movz	($acc,&BP(-128,$tbl,$acc,1));	# 11
+	&shl	($acc,24);			# 11
+	&or	("edx",$acc);			# 11
+	&shr	("ebx",16);			# 13,12
+
+	&pshufw	("mm6","mm4",0x03);		# 9, 8,15,14
+	&movz	($acc,&HB("eax"));		#  7
+	&movz	($acc,&BP(-128,$tbl,$acc,1));	#  7
+	&shl	($acc,24);			#  7
+	&or	("ecx",$acc);			#  7
+	&movz	($acc,&HB("ebx"));		# 13
+	&movz	($acc,&BP(-128,$tbl,$acc,1));	# 13
+	&shl	($acc,8);			# 13
+	&or	("ecx",$acc);			# 13
+	&movd	("mm0","ecx");			# t[0] collected
+
+	&movz	($acc,&LB("eax"));		#  6
+	&movd	("eax","mm2");			#  3, 2, 5, 4
+	&movz	("ecx",&BP(-128,$tbl,$acc,1));	#  6
+	&shl	("ecx",16);			#  6
+	&movz	($acc,&LB("ebx"));		# 12
+	&movd	("ebx","mm6");			#  9, 8,15,14
+	&movz	($acc,&BP(-128,$tbl,$acc,1));	# 12
+	&or	("ecx",$acc);			# 12
+
+	&movz	($acc,&LB("eax"));		#  4
+	&movz	($acc,&BP(-128,$tbl,$acc,1));	#  4
+	&or	("edx",$acc);			#  4
+	&movz	($acc,&LB("ebx"));		# 14
+	&movz	($acc,&BP(-128,$tbl,$acc,1));	# 14
+	&shl	($acc,16);			# 14
+	&or	("edx",$acc);			# 14
+	&movd	("mm1","edx");			# t[1] collected
+
+	&movz	($acc,&HB("eax"));		#  5
+	&movz	("edx",&BP(-128,$tbl,$acc,1));	#  5
+	&shl	("edx",8);			#  5
+	&movz	($acc,&HB("ebx"));		# 15
+	&shr	("eax",16);			#  3, 2
+	&movz	($acc,&BP(-128,$tbl,$acc,1));	# 15
+	&shl	($acc,24);			# 15
+	&or	("edx",$acc);			# 15
+	&shr	("ebx",16);			#  9, 8
+
+	&punpckldq	("mm0","mm1");		# t[0,1] collected
+
+	&movz	($acc,&HB("ebx"));		#  9
+	&movz	($acc,&BP(-128,$tbl,$acc,1));	#  9
+	&shl	($acc,8);			#  9
+	&or	("ecx",$acc);			#  9
+	&and	("ebx",0xff);			#  8
+	&movz	("ebx",&BP(-128,$tbl,"ebx",1));	#  8
+	&or	("edx","ebx");			#  8
+	&movz	($acc,&LB("eax"));		#  2
+	&movz	($acc,&BP(-128,$tbl,$acc,1));	#  2
+	&shl	($acc,16);			#  2
+	&or	("edx",$acc);			#  2
+	&movd	("mm4","edx");			# t[2] collected
+	&movz	("eax",&HB("eax"));		#  3
+	&movz	("eax",&BP(-128,$tbl,"eax",1));	#  3
+	&shl	("eax",24);			#  3
+	&or	("ecx","eax");			#  3
+	&movd	("mm5","ecx");			# t[3] collected
+
+	&punpckldq	("mm4","mm5");		# t[2,3] collected
+}
+
+					if (!$x86only) {
+&function_begin_B("_sse_AES_decrypt_compact");
+	&pxor	("mm0",&QWP(0,$key));	#  7, 6, 5, 4, 3, 2, 1, 0
+	&pxor	("mm4",&QWP(8,$key));	# 15,14,13,12,11,10, 9, 8
+
+	# note that caller is expected to allocate stack frame for me!
+	&mov	($acc,&DWP(240,$key));		# load key->rounds
+	&lea	($acc,&DWP(-2,$acc,$acc));
+	&lea	($acc,&DWP(0,$key,$acc,8));
+	&mov	($__end,$acc);			# end of key schedule
+
+	&mov	($s0,0x1b1b1b1b);		# magic constant
+	&mov	(&DWP(8,"esp"),$s0);
+	&mov	(&DWP(12,"esp"),$s0);
+
+	# prefetch Td4
+	&mov	($s0,&DWP(0-128,$tbl));
+	&mov	($s1,&DWP(32-128,$tbl));
+	&mov	($s2,&DWP(64-128,$tbl));
+	&mov	($s3,&DWP(96-128,$tbl));
+	&mov	($s0,&DWP(128-128,$tbl));
+	&mov	($s1,&DWP(160-128,$tbl));
+	&mov	($s2,&DWP(192-128,$tbl));
+	&mov	($s3,&DWP(224-128,$tbl));
+
+	&set_label("loop",16);
+		&sse_deccompact();
+		&add	($key,16);
+		&cmp	($key,$__end);
+		&ja	(&label("out"));
+
+		# ROTATE(x^y,N) == ROTATE(x,N)^ROTATE(y,N)
+		&movq	("mm3","mm0");		&movq	("mm7","mm4");
+		&movq	("mm2","mm0",1);	&movq	("mm6","mm4",1);
+		&movq	("mm1","mm0");		&movq	("mm5","mm4");
+		&pshufw	("mm0","mm0",0xb1);	&pshufw	("mm4","mm4",0xb1);# = ROTATE(tp0,16)
+		&pslld	("mm2",8);		&pslld	("mm6",8);
+		&psrld	("mm3",8);		&psrld	("mm7",8);
+		&pxor	("mm0","mm2");		&pxor	("mm4","mm6");	# ^= tp0<<8
+		&pxor	("mm0","mm3");		&pxor	("mm4","mm7");	# ^= tp0>>8
+		&pslld	("mm2",16);		&pslld	("mm6",16);
+		&psrld	("mm3",16);		&psrld	("mm7",16);
+		&pxor	("mm0","mm2");		&pxor	("mm4","mm6");	# ^= tp0<<24
+		&pxor	("mm0","mm3");		&pxor	("mm4","mm7");	# ^= tp0>>24
+
+		&movq	("mm3",&QWP(8,"esp"));
+		&pxor	("mm2","mm2");		&pxor	("mm6","mm6");
+		&pcmpgtb("mm2","mm1");		&pcmpgtb("mm6","mm5");
+		&pand	("mm2","mm3");		&pand	("mm6","mm3");
+		&paddb	("mm1","mm1");		&paddb	("mm5","mm5");
+		&pxor	("mm1","mm2");		&pxor	("mm5","mm6");	# tp2
+		&movq	("mm3","mm1");		&movq	("mm7","mm5");
+		&movq	("mm2","mm1");		&movq	("mm6","mm5");
+		&pxor	("mm0","mm1");		&pxor	("mm4","mm5");	# ^= tp2
+		&pslld	("mm3",24);		&pslld	("mm7",24);
+		&psrld	("mm2",8);		&psrld	("mm6",8);
+		&pxor	("mm0","mm3");		&pxor	("mm4","mm7");	# ^= tp2<<24
+		&pxor	("mm0","mm2");		&pxor	("mm4","mm6");	# ^= tp2>>8
+
+		&movq	("mm2",&QWP(8,"esp"));
+		&pxor	("mm3","mm3");		&pxor	("mm7","mm7");
+		&pcmpgtb("mm3","mm1");		&pcmpgtb("mm7","mm5");
+		&pand	("mm3","mm2");		&pand	("mm7","mm2");
+		&paddb	("mm1","mm1");		&paddb	("mm5","mm5");
+		&pxor	("mm1","mm3");		&pxor	("mm5","mm7");	# tp4
+		&pshufw	("mm3","mm1",0xb1);	&pshufw	("mm7","mm5",0xb1);
+		&pxor	("mm0","mm1");		&pxor	("mm4","mm5");	# ^= tp4
+		&pxor	("mm0","mm3");		&pxor	("mm4","mm7");	# ^= ROTATE(tp4,16)	
+
+		&pxor	("mm3","mm3");		&pxor	("mm7","mm7");
+		&pcmpgtb("mm3","mm1");		&pcmpgtb("mm7","mm5");
+		&pand	("mm3","mm2");		&pand	("mm7","mm2");
+		&paddb	("mm1","mm1");		&paddb	("mm5","mm5");
+		&pxor	("mm1","mm3");		&pxor	("mm5","mm7");	# tp8
+		&pxor	("mm0","mm1");		&pxor	("mm4","mm5");	# ^= tp8
+		&movq	("mm3","mm1");		&movq	("mm7","mm5");
+		&pshufw	("mm2","mm1",0xb1);	&pshufw	("mm6","mm5",0xb1);
+		&pxor	("mm0","mm2");		&pxor	("mm4","mm6");	# ^= ROTATE(tp8,16)
+		&pslld	("mm1",8);		&pslld	("mm5",8);
+		&psrld	("mm3",8);		&psrld	("mm7",8);
+		&movq	("mm2",&QWP(0,$key));	&movq	("mm6",&QWP(8,$key));
+		&pxor	("mm0","mm1");		&pxor	("mm4","mm5");	# ^= tp8<<8
+		&pxor	("mm0","mm3");		&pxor	("mm4","mm7");	# ^= tp8>>8
+		&mov	($s0,&DWP(0-128,$tbl));
+		&pslld	("mm1",16);		&pslld	("mm5",16);
+		&mov	($s1,&DWP(64-128,$tbl));
+		&psrld	("mm3",16);		&psrld	("mm7",16);
+		&mov	($s2,&DWP(128-128,$tbl));
+		&pxor	("mm0","mm1");		&pxor	("mm4","mm5");	# ^= tp8<<24
+		&mov	($s3,&DWP(192-128,$tbl));
+		&pxor	("mm0","mm3");		&pxor	("mm4","mm7");	# ^= tp8>>24
+
+		&pxor	("mm0","mm2");		&pxor	("mm4","mm6");
+	&jmp	(&label("loop"));
+
+	&set_label("out",16);
+	&pxor	("mm0",&QWP(0,$key));
+	&pxor	("mm4",&QWP(8,$key));
+
+	&ret	();
+&function_end_B("_sse_AES_decrypt_compact");
+					}
+
+######################################################################
+# Vanilla block function.
+######################################################################
 
 sub decstep()
 { my ($i,$td,@s) = @_;
@@ -480,7 +1589,7 @@ sub decstep()
 	# optimal... or rather that all attempts to reorder didn't
 	# result in better performance [which by the way is not a
 	# bit lower than ecryption].
-	if($i==3)   {	&mov	($key,&DWP(12,"esp"));		}
+	if($i==3)   {	&mov	($key,$__key);			}
 	else        {	&mov	($out,$s[0]);			}
 			&and	($out,0xFF);
 			&mov	($out,&DWP(0,$td,$out,8));
@@ -495,12 +1604,12 @@ sub decstep()
 			&and	($tmp,0xFF);
 			&xor	($out,&DWP(2,$td,$tmp,8));
 
-	if ($i==3)  {	$tmp=$s[3]; &mov ($s[2],&DWP(8,"esp"));	}
+	if ($i==3)  {	$tmp=$s[3]; &mov ($s[2],$__s1);		}
 	else        {	&mov	($tmp,$s[3]);			}
 			&shr	($tmp,24);
 			&xor	($out,&DWP(1,$td,$tmp,8));
 	if ($i<2)   {	&mov	(&DWP(4+4*$i,"esp"),$out);	}
-	if ($i==3)  {	&mov	($s[3],&DWP(4,"esp"));		}
+	if ($i==3)  {	&mov	($s[3],$__s0);			}
 			&comment();
 }
 
@@ -509,14 +1618,24 @@ sub declast()
   my $tmp = $key;
   my $out = $i==3?$s[0]:$acc;
 
-	if($i==3)   {	&mov	($key,&DWP(12,"esp"));		}
+	if($i==0)   {	&lea	($td,&DWP(2048+128,$td));
+			&mov	($tmp,&DWP(0-128,$td));
+			&mov	($acc,&DWP(32-128,$td));
+			&mov	($tmp,&DWP(64-128,$td));
+			&mov	($acc,&DWP(96-128,$td));
+			&mov	($tmp,&DWP(128-128,$td));
+			&mov	($acc,&DWP(160-128,$td));
+			&mov	($tmp,&DWP(192-128,$td));
+			&mov	($acc,&DWP(224-128,$td));
+			&lea	($td,&DWP(-128,$td));		}
+	if($i==3)   {	&mov	($key,$__key);			}
 	else        {	&mov	($out,$s[0]);			}
 			&and	($out,0xFF);
-			&movz	($out,&BP(2048,$td,$out,1));
+			&movz	($out,&BP(0,$td,$out,1));
 
 	if ($i==3)  {	$tmp=$s[1];				}
 			&movz	($tmp,&HB($s[1]));
-			&movz	($tmp,&BP(2048,$td,$tmp,1));
+			&movz	($tmp,&BP(0,$td,$tmp,1));
 			&shl	($tmp,8);
 			&xor	($out,$tmp);
 
@@ -524,24 +1643,24 @@ sub declast()
 	else        {	mov	($tmp,$s[2]);			}
 			&shr	($tmp,16);
 			&and	($tmp,0xFF);
-			&movz	($tmp,&BP(2048,$td,$tmp,1));
+			&movz	($tmp,&BP(0,$td,$tmp,1));
 			&shl	($tmp,16);
 			&xor	($out,$tmp);
 
-	if ($i==3)  {	$tmp=$s[3]; &mov ($s[2],&DWP(8,"esp"));	}
+	if ($i==3)  {	$tmp=$s[3]; &mov ($s[2],$__s1);		}
 	else        {	&mov	($tmp,$s[3]);			}
 			&shr	($tmp,24);
-			&movz	($tmp,&BP(2048,$td,$tmp,1));
+			&movz	($tmp,&BP(0,$td,$tmp,1));
 			&shl	($tmp,24);
 			&xor	($out,$tmp);
 	if ($i<2)   {	&mov	(&DWP(4+4*$i,"esp"),$out);	}
-	if ($i==3)  {	&mov	($s[3],&DWP(4,"esp"));		}
+	if ($i==3)  {	&mov	($s[3],$__s0);
+			&lea	($td,&DWP(-2048,$td));		}
 }
 
-&public_label("AES_Td");
 &function_begin_B("_x86_AES_decrypt");
 	# note that caller is expected to allocate stack frame for me!
-	&mov	(&DWP(12,"esp"),$key);		# save key
+	&mov	($__key,$key);			# save key
 
 	&xor	($s0,&DWP(0,$key));		# xor with key
 	&xor	($s1,&DWP(4,$key));
@@ -553,20 +1672,19 @@ sub declast()
 	if ($small_footprint) {
 	    &lea	($acc,&DWP(-2,$acc,$acc));
 	    &lea	($acc,&DWP(0,$key,$acc,8));
-	    &mov	(&DWP(16,"esp"),$acc);	# end of key schedule
-	    &align	(4);
-	    &set_label("loop");
-		&decstep(0,"ebp",$s0,$s3,$s2,$s1);
-		&decstep(1,"ebp",$s1,$s0,$s3,$s2);
-		&decstep(2,"ebp",$s2,$s1,$s0,$s3);
-		&decstep(3,"ebp",$s3,$s2,$s1,$s0);
+	    &mov	($__end,$acc);		# end of key schedule
+	    &set_label("loop",16);
+		&decstep(0,$tbl,$s0,$s3,$s2,$s1);
+		&decstep(1,$tbl,$s1,$s0,$s3,$s2);
+		&decstep(2,$tbl,$s2,$s1,$s0,$s3);
+		&decstep(3,$tbl,$s3,$s2,$s1,$s0);
 		&add	($key,16);		# advance rd_key
 		&xor	($s0,&DWP(0,$key));
 		&xor	($s1,&DWP(4,$key));
 		&xor	($s2,&DWP(8,$key));
 		&xor	($s3,&DWP(12,$key));
-	    &cmp	($key,&DWP(16,"esp"));
-	    &mov	(&DWP(12,"esp"),$key);
+	    &cmp	($key,$__end);
+	    &mov	($__key,$key);
 	    &jb		(&label("loop"));
 	}
 	else {
@@ -575,38 +1693,38 @@ sub declast()
 	    &cmp	($acc,12);
 	    &jle	(&label("12rounds"));
 
-	&set_label("14rounds");
+	&set_label("14rounds",4);
 	    for ($i=1;$i<3;$i++) {
-		&decstep(0,"ebp",$s0,$s3,$s2,$s1);
-		&decstep(1,"ebp",$s1,$s0,$s3,$s2);
-		&decstep(2,"ebp",$s2,$s1,$s0,$s3);
-		&decstep(3,"ebp",$s3,$s2,$s1,$s0);
+		&decstep(0,$tbl,$s0,$s3,$s2,$s1);
+		&decstep(1,$tbl,$s1,$s0,$s3,$s2);
+		&decstep(2,$tbl,$s2,$s1,$s0,$s3);
+		&decstep(3,$tbl,$s3,$s2,$s1,$s0);
 		&xor	($s0,&DWP(16*$i+0,$key));
 		&xor	($s1,&DWP(16*$i+4,$key));
 		&xor	($s2,&DWP(16*$i+8,$key));
 		&xor	($s3,&DWP(16*$i+12,$key));
 	    }
 	    &add	($key,32);
-	    &mov	(&DWP(12,"esp"),$key);	# advance rd_key
-	&set_label("12rounds");
+	    &mov	($__key,$key);		# advance rd_key
+	&set_label("12rounds",4);
 	    for ($i=1;$i<3;$i++) {
-		&decstep(0,"ebp",$s0,$s3,$s2,$s1);
-		&decstep(1,"ebp",$s1,$s0,$s3,$s2);
-		&decstep(2,"ebp",$s2,$s1,$s0,$s3);
-		&decstep(3,"ebp",$s3,$s2,$s1,$s0);
+		&decstep(0,$tbl,$s0,$s3,$s2,$s1);
+		&decstep(1,$tbl,$s1,$s0,$s3,$s2);
+		&decstep(2,$tbl,$s2,$s1,$s0,$s3);
+		&decstep(3,$tbl,$s3,$s2,$s1,$s0);
 		&xor	($s0,&DWP(16*$i+0,$key));
 		&xor	($s1,&DWP(16*$i+4,$key));
 		&xor	($s2,&DWP(16*$i+8,$key));
 		&xor	($s3,&DWP(16*$i+12,$key));
 	    }
 	    &add	($key,32);
-	    &mov	(&DWP(12,"esp"),$key);	# advance rd_key
-	&set_label("10rounds");
+	    &mov	($__key,$key);		# advance rd_key
+	&set_label("10rounds",4);
 	    for ($i=1;$i<10;$i++) {
-		&decstep(0,"ebp",$s0,$s3,$s2,$s1);
-		&decstep(1,"ebp",$s1,$s0,$s3,$s2);
-		&decstep(2,"ebp",$s2,$s1,$s0,$s3);
-		&decstep(3,"ebp",$s3,$s2,$s1,$s0);
+		&decstep(0,$tbl,$s0,$s3,$s2,$s1);
+		&decstep(1,$tbl,$s1,$s0,$s3,$s2);
+		&decstep(2,$tbl,$s2,$s1,$s0,$s3);
+		&decstep(3,$tbl,$s3,$s2,$s1,$s0);
 		&xor	($s0,&DWP(16*$i+0,$key));
 		&xor	($s1,&DWP(16*$i+4,$key));
 		&xor	($s2,&DWP(16*$i+8,$key));
@@ -614,10 +1732,10 @@ sub declast()
 	    }
 	}
 
-	&declast(0,"ebp",$s0,$s3,$s2,$s1);
-	&declast(1,"ebp",$s1,$s0,$s3,$s2);
-	&declast(2,"ebp",$s2,$s1,$s0,$s3);
-	&declast(3,"ebp",$s3,$s2,$s1,$s0);
+	&declast(0,$tbl,$s0,$s3,$s2,$s1);
+	&declast(1,$tbl,$s1,$s0,$s3,$s2);
+	&declast(2,$tbl,$s2,$s1,$s0,$s3);
+	&declast(3,$tbl,$s3,$s2,$s1,$s0);
 
 	&add	($key,$small_footprint?16:160);
 	&xor	($s0,&DWP(0,$key));
@@ -692,7 +1810,107 @@ sub declast()
 	&_data_word(0x72c31d16, 0x0c25e2bc, 0x8b493c28, 0x41950dff);
 	&_data_word(0x7101a839, 0xdeb30c08, 0x9ce4b4d8, 0x90c15664);
 	&_data_word(0x6184cb7b, 0x70b632d5, 0x745c6c48, 0x4257b8d0);
-#Td4:
+
+#Td4:	# four copies of Td4 to choose from to avoid L1 aliasing
+	&data_byte(0x52, 0x09, 0x6a, 0xd5, 0x30, 0x36, 0xa5, 0x38);
+	&data_byte(0xbf, 0x40, 0xa3, 0x9e, 0x81, 0xf3, 0xd7, 0xfb);
+	&data_byte(0x7c, 0xe3, 0x39, 0x82, 0x9b, 0x2f, 0xff, 0x87);
+	&data_byte(0x34, 0x8e, 0x43, 0x44, 0xc4, 0xde, 0xe9, 0xcb);
+	&data_byte(0x54, 0x7b, 0x94, 0x32, 0xa6, 0xc2, 0x23, 0x3d);
+	&data_byte(0xee, 0x4c, 0x95, 0x0b, 0x42, 0xfa, 0xc3, 0x4e);
+	&data_byte(0x08, 0x2e, 0xa1, 0x66, 0x28, 0xd9, 0x24, 0xb2);
+	&data_byte(0x76, 0x5b, 0xa2, 0x49, 0x6d, 0x8b, 0xd1, 0x25);
+	&data_byte(0x72, 0xf8, 0xf6, 0x64, 0x86, 0x68, 0x98, 0x16);
+	&data_byte(0xd4, 0xa4, 0x5c, 0xcc, 0x5d, 0x65, 0xb6, 0x92);
+	&data_byte(0x6c, 0x70, 0x48, 0x50, 0xfd, 0xed, 0xb9, 0xda);
+	&data_byte(0x5e, 0x15, 0x46, 0x57, 0xa7, 0x8d, 0x9d, 0x84);
+	&data_byte(0x90, 0xd8, 0xab, 0x00, 0x8c, 0xbc, 0xd3, 0x0a);
+	&data_byte(0xf7, 0xe4, 0x58, 0x05, 0xb8, 0xb3, 0x45, 0x06);
+	&data_byte(0xd0, 0x2c, 0x1e, 0x8f, 0xca, 0x3f, 0x0f, 0x02);
+	&data_byte(0xc1, 0xaf, 0xbd, 0x03, 0x01, 0x13, 0x8a, 0x6b);
+	&data_byte(0x3a, 0x91, 0x11, 0x41, 0x4f, 0x67, 0xdc, 0xea);
+	&data_byte(0x97, 0xf2, 0xcf, 0xce, 0xf0, 0xb4, 0xe6, 0x73);
+	&data_byte(0x96, 0xac, 0x74, 0x22, 0xe7, 0xad, 0x35, 0x85);
+	&data_byte(0xe2, 0xf9, 0x37, 0xe8, 0x1c, 0x75, 0xdf, 0x6e);
+	&data_byte(0x47, 0xf1, 0x1a, 0x71, 0x1d, 0x29, 0xc5, 0x89);
+	&data_byte(0x6f, 0xb7, 0x62, 0x0e, 0xaa, 0x18, 0xbe, 0x1b);
+	&data_byte(0xfc, 0x56, 0x3e, 0x4b, 0xc6, 0xd2, 0x79, 0x20);
+	&data_byte(0x9a, 0xdb, 0xc0, 0xfe, 0x78, 0xcd, 0x5a, 0xf4);
+	&data_byte(0x1f, 0xdd, 0xa8, 0x33, 0x88, 0x07, 0xc7, 0x31);
+	&data_byte(0xb1, 0x12, 0x10, 0x59, 0x27, 0x80, 0xec, 0x5f);
+	&data_byte(0x60, 0x51, 0x7f, 0xa9, 0x19, 0xb5, 0x4a, 0x0d);
+	&data_byte(0x2d, 0xe5, 0x7a, 0x9f, 0x93, 0xc9, 0x9c, 0xef);
+	&data_byte(0xa0, 0xe0, 0x3b, 0x4d, 0xae, 0x2a, 0xf5, 0xb0);
+	&data_byte(0xc8, 0xeb, 0xbb, 0x3c, 0x83, 0x53, 0x99, 0x61);
+	&data_byte(0x17, 0x2b, 0x04, 0x7e, 0xba, 0x77, 0xd6, 0x26);
+	&data_byte(0xe1, 0x69, 0x14, 0x63, 0x55, 0x21, 0x0c, 0x7d);
+
+	&data_byte(0x52, 0x09, 0x6a, 0xd5, 0x30, 0x36, 0xa5, 0x38);
+	&data_byte(0xbf, 0x40, 0xa3, 0x9e, 0x81, 0xf3, 0xd7, 0xfb);
+	&data_byte(0x7c, 0xe3, 0x39, 0x82, 0x9b, 0x2f, 0xff, 0x87);
+	&data_byte(0x34, 0x8e, 0x43, 0x44, 0xc4, 0xde, 0xe9, 0xcb);
+	&data_byte(0x54, 0x7b, 0x94, 0x32, 0xa6, 0xc2, 0x23, 0x3d);
+	&data_byte(0xee, 0x4c, 0x95, 0x0b, 0x42, 0xfa, 0xc3, 0x4e);
+	&data_byte(0x08, 0x2e, 0xa1, 0x66, 0x28, 0xd9, 0x24, 0xb2);
+	&data_byte(0x76, 0x5b, 0xa2, 0x49, 0x6d, 0x8b, 0xd1, 0x25);
+	&data_byte(0x72, 0xf8, 0xf6, 0x64, 0x86, 0x68, 0x98, 0x16);
+	&data_byte(0xd4, 0xa4, 0x5c, 0xcc, 0x5d, 0x65, 0xb6, 0x92);
+	&data_byte(0x6c, 0x70, 0x48, 0x50, 0xfd, 0xed, 0xb9, 0xda);
+	&data_byte(0x5e, 0x15, 0x46, 0x57, 0xa7, 0x8d, 0x9d, 0x84);
+	&data_byte(0x90, 0xd8, 0xab, 0x00, 0x8c, 0xbc, 0xd3, 0x0a);
+	&data_byte(0xf7, 0xe4, 0x58, 0x05, 0xb8, 0xb3, 0x45, 0x06);
+	&data_byte(0xd0, 0x2c, 0x1e, 0x8f, 0xca, 0x3f, 0x0f, 0x02);
+	&data_byte(0xc1, 0xaf, 0xbd, 0x03, 0x01, 0x13, 0x8a, 0x6b);
+	&data_byte(0x3a, 0x91, 0x11, 0x41, 0x4f, 0x67, 0xdc, 0xea);
+	&data_byte(0x97, 0xf2, 0xcf, 0xce, 0xf0, 0xb4, 0xe6, 0x73);
+	&data_byte(0x96, 0xac, 0x74, 0x22, 0xe7, 0xad, 0x35, 0x85);
+	&data_byte(0xe2, 0xf9, 0x37, 0xe8, 0x1c, 0x75, 0xdf, 0x6e);
+	&data_byte(0x47, 0xf1, 0x1a, 0x71, 0x1d, 0x29, 0xc5, 0x89);
+	&data_byte(0x6f, 0xb7, 0x62, 0x0e, 0xaa, 0x18, 0xbe, 0x1b);
+	&data_byte(0xfc, 0x56, 0x3e, 0x4b, 0xc6, 0xd2, 0x79, 0x20);
+	&data_byte(0x9a, 0xdb, 0xc0, 0xfe, 0x78, 0xcd, 0x5a, 0xf4);
+	&data_byte(0x1f, 0xdd, 0xa8, 0x33, 0x88, 0x07, 0xc7, 0x31);
+	&data_byte(0xb1, 0x12, 0x10, 0x59, 0x27, 0x80, 0xec, 0x5f);
+	&data_byte(0x60, 0x51, 0x7f, 0xa9, 0x19, 0xb5, 0x4a, 0x0d);
+	&data_byte(0x2d, 0xe5, 0x7a, 0x9f, 0x93, 0xc9, 0x9c, 0xef);
+	&data_byte(0xa0, 0xe0, 0x3b, 0x4d, 0xae, 0x2a, 0xf5, 0xb0);
+	&data_byte(0xc8, 0xeb, 0xbb, 0x3c, 0x83, 0x53, 0x99, 0x61);
+	&data_byte(0x17, 0x2b, 0x04, 0x7e, 0xba, 0x77, 0xd6, 0x26);
+	&data_byte(0xe1, 0x69, 0x14, 0x63, 0x55, 0x21, 0x0c, 0x7d);
+
+	&data_byte(0x52, 0x09, 0x6a, 0xd5, 0x30, 0x36, 0xa5, 0x38);
+	&data_byte(0xbf, 0x40, 0xa3, 0x9e, 0x81, 0xf3, 0xd7, 0xfb);
+	&data_byte(0x7c, 0xe3, 0x39, 0x82, 0x9b, 0x2f, 0xff, 0x87);
+	&data_byte(0x34, 0x8e, 0x43, 0x44, 0xc4, 0xde, 0xe9, 0xcb);
+	&data_byte(0x54, 0x7b, 0x94, 0x32, 0xa6, 0xc2, 0x23, 0x3d);
+	&data_byte(0xee, 0x4c, 0x95, 0x0b, 0x42, 0xfa, 0xc3, 0x4e);
+	&data_byte(0x08, 0x2e, 0xa1, 0x66, 0x28, 0xd9, 0x24, 0xb2);
+	&data_byte(0x76, 0x5b, 0xa2, 0x49, 0x6d, 0x8b, 0xd1, 0x25);
+	&data_byte(0x72, 0xf8, 0xf6, 0x64, 0x86, 0x68, 0x98, 0x16);
+	&data_byte(0xd4, 0xa4, 0x5c, 0xcc, 0x5d, 0x65, 0xb6, 0x92);
+	&data_byte(0x6c, 0x70, 0x48, 0x50, 0xfd, 0xed, 0xb9, 0xda);
+	&data_byte(0x5e, 0x15, 0x46, 0x57, 0xa7, 0x8d, 0x9d, 0x84);
+	&data_byte(0x90, 0xd8, 0xab, 0x00, 0x8c, 0xbc, 0xd3, 0x0a);
+	&data_byte(0xf7, 0xe4, 0x58, 0x05, 0xb8, 0xb3, 0x45, 0x06);
+	&data_byte(0xd0, 0x2c, 0x1e, 0x8f, 0xca, 0x3f, 0x0f, 0x02);
+	&data_byte(0xc1, 0xaf, 0xbd, 0x03, 0x01, 0x13, 0x8a, 0x6b);
+	&data_byte(0x3a, 0x91, 0x11, 0x41, 0x4f, 0x67, 0xdc, 0xea);
+	&data_byte(0x97, 0xf2, 0xcf, 0xce, 0xf0, 0xb4, 0xe6, 0x73);
+	&data_byte(0x96, 0xac, 0x74, 0x22, 0xe7, 0xad, 0x35, 0x85);
+	&data_byte(0xe2, 0xf9, 0x37, 0xe8, 0x1c, 0x75, 0xdf, 0x6e);
+	&data_byte(0x47, 0xf1, 0x1a, 0x71, 0x1d, 0x29, 0xc5, 0x89);
+	&data_byte(0x6f, 0xb7, 0x62, 0x0e, 0xaa, 0x18, 0xbe, 0x1b);
+	&data_byte(0xfc, 0x56, 0x3e, 0x4b, 0xc6, 0xd2, 0x79, 0x20);
+	&data_byte(0x9a, 0xdb, 0xc0, 0xfe, 0x78, 0xcd, 0x5a, 0xf4);
+	&data_byte(0x1f, 0xdd, 0xa8, 0x33, 0x88, 0x07, 0xc7, 0x31);
+	&data_byte(0xb1, 0x12, 0x10, 0x59, 0x27, 0x80, 0xec, 0x5f);
+	&data_byte(0x60, 0x51, 0x7f, 0xa9, 0x19, 0xb5, 0x4a, 0x0d);
+	&data_byte(0x2d, 0xe5, 0x7a, 0x9f, 0x93, 0xc9, 0x9c, 0xef);
+	&data_byte(0xa0, 0xe0, 0x3b, 0x4d, 0xae, 0x2a, 0xf5, 0xb0);
+	&data_byte(0xc8, 0xeb, 0xbb, 0x3c, 0x83, 0x53, 0x99, 0x61);
+	&data_byte(0x17, 0x2b, 0x04, 0x7e, 0xba, 0x77, 0xd6, 0x26);
+	&data_byte(0xe1, 0x69, 0x14, 0x63, 0x55, 0x21, 0x0c, 0x7d);
+
 	&data_byte(0x52, 0x09, 0x6a, 0xd5, 0x30, 0x36, 0xa5, 0x38);
 	&data_byte(0xbf, 0x40, 0xa3, 0x9e, 0x81, 0xf3, 0xd7, 0xfb);
 	&data_byte(0x7c, 0xe3, 0x39, 0x82, 0x9b, 0x2f, 0xff, 0x87);
@@ -728,43 +1946,57 @@ sub declast()
 &function_end_B("_x86_AES_decrypt");
 
 # void AES_decrypt (const void *inp,void *out,const AES_KEY *key);
-&public_label("AES_Td");
 &function_begin("AES_decrypt");
 	&mov	($acc,&wparam(0));		# load inp
 	&mov	($key,&wparam(2));		# load key
 
 	&mov	($s0,"esp");
-	&sub	("esp",24);
-	&and	("esp",-64);
-	&add	("esp",4);
-	&mov	(&DWP(16,"esp"),$s0);
+	&sub	("esp",36);
+	&and	("esp",-64);			# align to cache-line
+
+	# place stack frame just "above" the key schedule
+	&lea	($s1,&DWP(-64-63,$key));
+	&sub	($s1,"esp");
+	&neg	($s1);
+	&and	($s1,0x3C0);	# modulo 1024, but aligned to cache-line
+	&sub	("esp",$s1);
+	&add	("esp",4);	# 4 is reserved for caller's return address
+	&mov	($_esp,$s0);	# save stack pointer
 
 	&call   (&label("pic_point"));          # make it PIC!
 	&set_label("pic_point");
-	&blindpop("ebp");
-	&lea    ("ebp",&DWP(&label("AES_Td")."-".&label("pic_point"),"ebp"));
-
-	# prefetch Td4
-	&lea	("ebp",&DWP(2048+128,"ebp"));
-	&mov	($s0,&DWP(0-128,"ebp"));
-	&mov	($s1,&DWP(32-128,"ebp"));
-	&mov	($s2,&DWP(64-128,"ebp"));
-	&mov	($s3,&DWP(96-128,"ebp"));
-	&mov	($s0,&DWP(128-128,"ebp"));
-	&mov	($s1,&DWP(160-128,"ebp"));
-	&mov	($s2,&DWP(192-128,"ebp"));
-	&mov	($s3,&DWP(224-128,"ebp"));
-	&lea	("ebp",&DWP(-2048-128,"ebp"));
-
+	&blindpop($tbl);
+	&picmeup($s0,"OPENSSL_ia32cap_P",$tbl,&label("pic_point")) if(!$x86only);
+	&lea    ($tbl,&DWP(&label("AES_Td")."-".&label("pic_point"),$tbl));
+
+	# pick Td4 copy which can't "overlap" with stack frame or key schedule
+	&lea	($s1,&DWP(768-4,"esp"));
+	&sub	($s1,$tbl);
+	&and	($s1,0x300);
+	&lea	($tbl,&DWP(2048+128,$tbl,$s1));
+
+					if (!$x86only) {
+	&bt	(&DWP(0,$s0),25);	# check for SSE bit
+	&jnc	(&label("x86"));
+
+	&movq	("mm0",&QWP(0,$acc));
+	&movq	("mm4",&QWP(8,$acc));
+	&call	("_sse_AES_decrypt_compact");
+	&mov	("esp",$_esp);			# restore stack pointer
+	&mov	($acc,&wparam(1));		# load out
+	&movq	(&QWP(0,$acc),"mm0");		# write output data
+	&movq	(&QWP(8,$acc),"mm4");
+	&emms	();
+	&function_end_A();
+					}
+	&set_label("x86",16);
+	&mov	($_tbl,$tbl);
 	&mov	($s0,&DWP(0,$acc));		# load input data
 	&mov	($s1,&DWP(4,$acc));
 	&mov	($s2,&DWP(8,$acc));
 	&mov	($s3,&DWP(12,$acc));
-
-	&call	("_x86_AES_decrypt");
-
-	&mov	("esp",&DWP(16,"esp"));
-
+	&call	("_x86_AES_decrypt_compact");
+	&mov	("esp",$_esp);			# restore stack pointer
 	&mov	($acc,&wparam(1));		# load out
 	&mov	(&DWP(0,$acc),$s0);		# write output data
 	&mov	(&DWP(4,$acc),$s1);
@@ -777,126 +2009,136 @@ sub declast()
 #			unsigned char *ivp,const int enc);
 {
 # stack frame layout
-# -4(%esp)	0(%esp)		return address
-# 0(%esp)	4(%esp)		tmp1
-# 4(%esp)	8(%esp)		tmp2
-# 8(%esp)	12(%esp)	key
-# 12(%esp)	16(%esp)	end of key schedule
-my $_esp=&DWP(16,"esp");	#saved %esp
-my $_inp=&DWP(20,"esp");	#copy of wparam(0)
-my $_out=&DWP(24,"esp");	#copy of wparam(1)
-my $_len=&DWP(28,"esp");	#copy of wparam(2)
-my $_key=&DWP(32,"esp");	#copy of wparam(3)
-my $_ivp=&DWP(36,"esp");	#copy of wparam(4)
-my $_tmp=&DWP(40,"esp");	#volatile variable
-my $ivec=&DWP(44,"esp");	#ivec[16]
-my $aes_key=&DWP(60,"esp");	#copy of aes_key
-my $mark=&DWP(60+240,"esp");	#copy of aes_key->rounds
-
-&public_label("AES_Te");
-&public_label("AES_Td");
+#             -4(%esp)		# return address	 0(%esp)
+#              0(%esp)		# s0 backing store	 4(%esp)	
+#              4(%esp)		# s1 backing store	 8(%esp)
+#              8(%esp)		# s2 backing store	12(%esp)
+#             12(%esp)		# s3 backing store	16(%esp)
+#             16(%esp)		# key backup		20(%esp)
+#             20(%esp)		# end of key schedule	24(%esp)
+#             24(%esp)		# %ebp backup		28(%esp)
+#             28(%esp)		# %esp backup
+my $_inp=&DWP(32,"esp");	# copy of wparam(0)
+my $_out=&DWP(36,"esp");	# copy of wparam(1)
+my $_len=&DWP(40,"esp");	# copy of wparam(2)
+my $_key=&DWP(44,"esp");	# copy of wparam(3)
+my $_ivp=&DWP(48,"esp");	# copy of wparam(4)
+my $_tmp=&DWP(52,"esp");	# volatile variable
+#
+my $ivec=&DWP(60,"esp");	# ivec[16]
+my $aes_key=&DWP(76,"esp");	# copy of aes_key
+my $mark=&DWP(76+240,"esp");	# copy of aes_key->rounds
+
 &function_begin("AES_cbc_encrypt");
 	&mov	($s2 eq "ecx"? $s2 : "",&wparam(2));	# load len
 	&cmp	($s2,0);
-	&je	(&label("enc_out"));
+	&je	(&label("drop_out"));
 
 	&call   (&label("pic_point"));		# make it PIC!
 	&set_label("pic_point");
-	&blindpop("ebp");
-
-	&pushf	();
-	&cld	();
+	&blindpop($tbl);
+	&picmeup($s0,"OPENSSL_ia32cap_P",$tbl,&label("pic_point")) if(!$x86only);
 
 	&cmp	(&wparam(5),0);
-	&je	(&label("DECRYPT"));
-
-	&lea    ("ebp",&DWP(&label("AES_Te")."-".&label("pic_point"),"ebp"));
+	&lea    ($tbl,&DWP(&label("AES_Te")."-".&label("pic_point"),$tbl));
+	&jne	(&label("picked_te"));
+	&lea	($tbl,&DWP(&label("AES_Td")."-".&label("AES_Te"),$tbl));
+	&set_label("picked_te");
 
-	# allocate aligned stack frame...
-	&lea	($key,&DWP(-64-244,"esp"));
-	&and	($key,-64);
+	# one can argue if this is required
+	&pushf	();
+	&cld	();
 
-	# ... and make sure it doesn't alias with AES_Te modulo 4096
-	&mov	($s0,"ebp");
-	&lea	($s1,&DWP(2048,"ebp"));
-	&mov	($s3,$key);
+	&cmp	($s2,$speed_limit);
+	&jb	(&label("slow_way"));
+	&test	($s2,15);
+	&jnz	(&label("slow_way"));
+					if (!$x86only) {
+	&bt	(&DWP(0,$s0),28);	# check for hyper-threading bit
+	&jc	(&label("slow_way"));
+					}
+	# pre-allocate aligned stack frame...
+	&lea	($acc,&DWP(-80-244,"esp"));
+	&and	($acc,-64);
+
+	# ... and make sure it doesn't alias with $tbl modulo 4096
+	&mov	($s0,$tbl);
+	&lea	($s1,&DWP(2048+256,$tbl));
+	&mov	($s3,$acc);
 	&and	($s0,0xfff);		# s = %ebp&0xfff
-	&and	($s1,0xfff);		# e = (%ebp+2048)&0xfff
+	&and	($s1,0xfff);		# e = (%ebp+2048+256)&0xfff
 	&and	($s3,0xfff);		# p = %esp&0xfff
 
 	&cmp	($s3,$s1);		# if (p>=e) %esp =- (p-e);
-	&jb	(&label("te_break_out"));
+	&jb	(&label("tbl_break_out"));
 	&sub	($s3,$s1);
-	&sub	($key,$s3);
-	&jmp	(&label("te_ok"));
-	&set_label("te_break_out");	# else %esp -= (p-s)&0xfff + framesz;
+	&sub	($acc,$s3);
+	&jmp	(&label("tbl_ok"));
+	&set_label("tbl_break_out",4);	# else %esp -= (p-s)&0xfff + framesz;
 	&sub	($s3,$s0);
 	&and	($s3,0xfff);
-	&add	($s3,64+256);
-	&sub	($key,$s3);
-	&align	(4);
-	&set_label("te_ok");
-
-	&mov	($s0,&wparam(0));	# load inp
-	&mov	($s1,&wparam(1));	# load out
-	&mov	($s3,&wparam(3));	# load key
-	&mov	($acc,&wparam(4));	# load ivp
+	&add	($s3,384);
+	&sub	($acc,$s3);
+	&set_label("tbl_ok",4);
 
-	&exch	("esp",$key);
+	&lea	($s3,&wparam(0));	# obtain pointer to parameter block
+	&exch	("esp",$acc);		# allocate stack frame
 	&add	("esp",4);		# reserve for return address!
-	&mov	($_esp,$key);		# save %esp
+	&mov	($_tbl,$tbl);		# save %ebp
+	&mov	($_esp,$acc);		# save %esp
+
+	&mov	($s0,&DWP(0,$s3));	# load inp
+	&mov	($s1,&DWP(4,$s3));	# load out
+	#&mov	($s2,&DWP(8,$s3));	# load len
+	&mov	($key,&DWP(12,$s3));	# load key
+	&mov	($acc,&DWP(16,$s3));	# load ivp
+	&mov	($s3,&DWP(20,$s3));	# load enc flag
 
 	&mov	($_inp,$s0);		# save copy of inp
 	&mov	($_out,$s1);		# save copy of out
 	&mov	($_len,$s2);		# save copy of len
-	&mov	($_key,$s3);		# save copy of key
+	&mov	($_key,$key);		# save copy of key
 	&mov	($_ivp,$acc);		# save copy of ivp
 
 	&mov	($mark,0);		# copy of aes_key->rounds = 0;
-	if ($compromise) {
-		&cmp	($s2,$compromise);
-		&jb	(&label("skip_ecopy"));
-	}
 	# do we copy key schedule to stack?
-	&mov	($s1 eq "ebx" ? $s1 : "",$s3);
+	&mov	($s1 eq "ebx" ? $s1 : "",$key);
 	&mov	($s2 eq "ecx" ? $s2 : "",244/4);
-	&sub	($s1,"ebp");
-	&mov	("esi",$s3);
+	&sub	($s1,$tbl);
+	&mov	("esi",$key);
 	&and	($s1,0xfff);
 	&lea	("edi",$aes_key);
-	&cmp	($s1,2048);
-	&jb	(&label("do_ecopy"));
+	&cmp	($s1,2048+256);
+	&jb	(&label("do_copy"));
 	&cmp	($s1,4096-244);
-	&jb	(&label("skip_ecopy"));
-	&align	(4);
-	&set_label("do_ecopy");
+	&jb	(&label("skip_copy"));
+	&set_label("do_copy",4);
 		&mov	($_key,"edi");
 		&data_word(0xA5F3F689);	# rep movsd
-	&set_label("skip_ecopy");
+	&set_label("skip_copy");
 
-	&mov	($acc,$s0);
 	&mov	($key,16);
-	&align	(4);
-	&set_label("prefetch_te");
-		&mov	($s0,&DWP(0,"ebp"));
-		&mov	($s1,&DWP(32,"ebp"));
-		&mov	($s2,&DWP(64,"ebp"));
-		&mov	($s3,&DWP(96,"ebp"));
-		&lea	("ebp",&DWP(128,"ebp"));
-		&dec	($key);
-	&jnz	(&label("prefetch_te"));
-	&sub	("ebp",2048);
-
-	&mov	($s2,$_len);
+	&set_label("prefetch_tbl",4);
+		&mov	($s0,&DWP(0,$tbl));
+		&mov	($s1,&DWP(32,$tbl));
+		&mov	($s2,&DWP(64,$tbl));
+		&mov	($acc,&DWP(96,$tbl));
+		&lea	($tbl,&DWP(128,$tbl));
+		&sub	($key,1);
+	&jnz	(&label("prefetch_tbl"));
+	&sub	($tbl,2048);
+
+	&mov	($acc,$_inp);
 	&mov	($key,$_ivp);
-	&test	($s2,0xFFFFFFF0);
-	&jz	(&label("enc_tail"));		# short input...
 
+	&cmp	($s3,0);
+	&je	(&label("fast_decrypt"));
+
+#----------------------------- ENCRYPT -----------------------------#
 	&mov	($s0,&DWP(0,$key));		# load iv
 	&mov	($s1,&DWP(4,$key));
 
-	&align	(4);
-	&set_label("enc_loop");
+	&set_label("fast_enc_loop",16);
 		&mov	($s2,&DWP(8,$key));
 		&mov	($s3,&DWP(12,$key));
 
@@ -916,22 +2158,16 @@ my $mark=&DWP(60+240,"esp");	#copy of aes_key->rounds
 		&mov	(&DWP(8,$key),$s2);
 		&mov	(&DWP(12,$key),$s3);
 
+		&lea	($acc,&DWP(16,$acc));	# advance inp
 		&mov	($s2,$_len);		# load len
-
-		&lea	($acc,&DWP(16,$acc));
 		&mov	($_inp,$acc);		# save inp
-
-		&lea	($s3,&DWP(16,$key));
+		&lea	($s3,&DWP(16,$key));	# advance out
 		&mov	($_out,$s3);		# save out
-
-		&sub	($s2,16);
-		&test	($s2,0xFFFFFFF0);
+		&sub	($s2,16);		# decrease len
 		&mov	($_len,$s2);		# save len
-	&jnz	(&label("enc_loop"));
-	&test	($s2,15);
-	&jnz	(&label("enc_tail"));
+	&jnz	(&label("fast_enc_loop"));
 	&mov	($acc,$_ivp);		# load ivp
-	&mov	($s2,&DWP(8,$key));	# restore last dwords
+	&mov	($s2,&DWP(8,$key));	# restore last 2 dwords
 	&mov	($s3,&DWP(12,$key));
 	&mov	(&DWP(0,$acc),$s0);	# save ivec
 	&mov	(&DWP(4,$acc),$s1);
@@ -949,125 +2185,20 @@ my $mark=&DWP(60+240,"esp");	#copy of aes_key->rounds
 	&set_label("skip_ezero")
 	&mov	("esp",$_esp);
 	&popf	();
-    &set_label("enc_out");
+    &set_label("drop_out");
 	&function_end_A();
 	&pushf	();			# kludge, never executed
 
-    &align	(4);
-    &set_label("enc_tail");
-	&mov	($s0,$key eq "edi" ? $key : "");
-	&mov	($key,$_out);			# load out
-	&push	($s0);				# push ivp
-	&mov	($s1,16);
-	&sub	($s1,$s2);
-	&cmp	($key,$acc);			# compare with inp
-	&je	(&label("enc_in_place"));
-	&align	(4);
-	&data_word(0xA4F3F689);	# rep movsb	# copy input
-	&jmp	(&label("enc_skip_in_place"));
-    &set_label("enc_in_place");
-	&lea	($key,&DWP(0,$key,$s2));
-    &set_label("enc_skip_in_place");
-	&mov	($s2,$s1);
-	&xor	($s0,$s0);
-	&align	(4);
-	&data_word(0xAAF3F689);	# rep stosb	# zero tail
-	&pop	($key);				# pop ivp
-
-	&mov	($acc,$_out);			# output as input
-	&mov	($s0,&DWP(0,$key));
-	&mov	($s1,&DWP(4,$key));
-	&mov	($_len,16);			# len=16
-	&jmp	(&label("enc_loop"));		# one more spin...
-
 #----------------------------- DECRYPT -----------------------------#
-&align	(4);
-&set_label("DECRYPT");
-	&lea    ("ebp",&DWP(&label("AES_Td")."-".&label("pic_point"),"ebp"));
-
-	# allocate aligned stack frame...
-	&lea	($key,&DWP(-64-244,"esp"));
-	&and	($key,-64);
-
-	# ... and make sure it doesn't alias with AES_Td modulo 4096
-	&mov	($s0,"ebp");
-	&lea	($s1,&DWP(2048+256,"ebp"));
-	&mov	($s3,$key);
-	&and	($s0,0xfff);		# s = %ebp&0xfff
-	&and	($s1,0xfff);		# e = (%ebp+2048+256)&0xfff
-	&and	($s3,0xfff);		# p = %esp&0xfff
-
-	&cmp	($s3,$s1);		# if (p>=e) %esp =- (p-e);
-	&jb	(&label("td_break_out"));
-	&sub	($s3,$s1);
-	&sub	($key,$s3);
-	&jmp	(&label("td_ok"));
-	&set_label("td_break_out");	# else %esp -= (p-s)&0xfff + framesz;
-	&sub	($s3,$s0);
-	&and	($s3,0xfff);
-	&add	($s3,64+256);
-	&sub	($key,$s3);
-	&align	(4);
-	&set_label("td_ok");
-
-	&mov	($s0,&wparam(0));	# load inp
-	&mov	($s1,&wparam(1));	# load out
-	&mov	($s3,&wparam(3));	# load key
-	&mov	($acc,&wparam(4));	# load ivp
-
-	&exch	("esp",$key);
-	&add	("esp",4);		# reserve for return address!
-	&mov	($_esp,$key);		# save %esp
-
-	&mov	($_inp,$s0);		# save copy of inp
-	&mov	($_out,$s1);		# save copy of out
-	&mov	($_len,$s2);		# save copy of len
-	&mov	($_key,$s3);		# save copy of key
-	&mov	($_ivp,$acc);		# save copy of ivp
-
-	&mov	($mark,0);		# copy of aes_key->rounds = 0;
-	if ($compromise) {
-		&cmp	($s2,$compromise);
-		&jb	(&label("skip_dcopy"));
-	}
-	# do we copy key schedule to stack?
-	&mov	($s1 eq "ebx" ? $s1 : "",$s3);
-	&mov	($s2 eq "ecx" ? $s2 : "",244/4);
-	&sub	($s1,"ebp");
-	&mov	("esi",$s3);
-	&and	($s1,0xfff);
-	&lea	("edi",$aes_key);
-	&cmp	($s1,2048+256);
-	&jb	(&label("do_dcopy"));
-	&cmp	($s1,4096-244);
-	&jb	(&label("skip_dcopy"));
-	&align	(4);
-	&set_label("do_dcopy");
-		&mov	($_key,"edi");
-		&data_word(0xA5F3F689);	# rep movsd
-	&set_label("skip_dcopy");
-
-	&mov	($acc,$s0);
-	&mov	($key,18);
-	&align	(4);
-	&set_label("prefetch_td");
-		&mov	($s0,&DWP(0,"ebp"));
-		&mov	($s1,&DWP(32,"ebp"));
-		&mov	($s2,&DWP(64,"ebp"));
-		&mov	($s3,&DWP(96,"ebp"));
-		&lea	("ebp",&DWP(128,"ebp"));
-		&dec	($key);
-	&jnz	(&label("prefetch_td"));
-	&sub	("ebp",2048+256);
+&set_label("fast_decrypt",16);
 
 	&cmp	($acc,$_out);
-	&je	(&label("dec_in_place"));	# in-place processing...
+	&je	(&label("fast_dec_in_place"));	# in-place processing...
 
-	&mov	($key,$_ivp);		# load ivp
 	&mov	($_tmp,$key);
 
 	&align	(4);
-	&set_label("dec_loop");
+	&set_label("fast_dec_loop",16);
 		&mov	($s0,&DWP(0,$acc));	# read input
 		&mov	($s1,&DWP(4,$acc));
 		&mov	($s2,&DWP(8,$acc));
@@ -1083,27 +2214,24 @@ my $mark=&DWP(60+240,"esp");	#copy of aes_key->rounds
 		&xor	($s2,&DWP(8,$key));
 		&xor	($s3,&DWP(12,$key));
 
-		&sub	($acc,16);
-		&jc	(&label("dec_partial"));
-		&mov	($_len,$acc);		# save len
-		&mov	($acc,$_inp);		# load inp
 		&mov	($key,$_out);		# load out
+		&mov	($acc,$_inp);		# load inp
 
 		&mov	(&DWP(0,$key),$s0);	# write output
 		&mov	(&DWP(4,$key),$s1);
 		&mov	(&DWP(8,$key),$s2);
 		&mov	(&DWP(12,$key),$s3);
 
+		&mov	($s2,$_len);		# load len
 		&mov	($_tmp,$acc);		# save ivp
-		&lea	($acc,&DWP(16,$acc));
+		&lea	($acc,&DWP(16,$acc));	# advance inp
 		&mov	($_inp,$acc);		# save inp
-
-		&lea	($key,&DWP(16,$key));
+		&lea	($key,&DWP(16,$key));	# advance out
 		&mov	($_out,$key);		# save out
-
-	&jnz	(&label("dec_loop"));
+		&sub	($s2,16);		# decrease len
+		&mov	($_len,$s2);		# save len
+	&jnz	(&label("fast_dec_loop"));
 	&mov	($key,$_tmp);		# load temp ivp
-    &set_label("dec_end");
 	&mov	($acc,$_ivp);		# load user ivp
 	&mov	($s0,&DWP(0,$key));	# load iv
 	&mov	($s1,&DWP(4,$key));
@@ -1113,31 +2241,16 @@ my $mark=&DWP(60+240,"esp");	#copy of aes_key->rounds
 	&mov	(&DWP(4,$acc),$s1);
 	&mov	(&DWP(8,$acc),$s2);
 	&mov	(&DWP(12,$acc),$s3);
-	&jmp	(&label("dec_out"));
+	&jmp	(&label("fast_dec_out"));
 
-    &align	(4);
-    &set_label("dec_partial");
-	&lea	($key,$ivec);
-	&mov	(&DWP(0,$key),$s0);	# dump output to stack
-	&mov	(&DWP(4,$key),$s1);
-	&mov	(&DWP(8,$key),$s2);
-	&mov	(&DWP(12,$key),$s3);
-	&lea	($s2 eq "ecx" ? $s2 : "",&DWP(16,$acc));
-	&mov	($acc eq "esi" ? $acc : "",$key);
-	&mov	($key eq "edi" ? $key : "",$_out);	# load out
-	&data_word(0xA4F3F689);	# rep movsb		# copy output
-	&mov	($key,$_inp);				# use inp as temp ivp
-	&jmp	(&label("dec_end"));
-
-    &align	(4);
-    &set_label("dec_in_place");
-	&set_label("dec_in_place_loop");
-		&lea	($key,$ivec);
+    &set_label("fast_dec_in_place",16);
+	&set_label("fast_dec_in_place_loop");
 		&mov	($s0,&DWP(0,$acc));	# read input
 		&mov	($s1,&DWP(4,$acc));
 		&mov	($s2,&DWP(8,$acc));
 		&mov	($s3,&DWP(12,$acc));
 
+		&lea	($key,$ivec);
 		&mov	(&DWP(0,$key),$s0);	# copy to temp
 		&mov	(&DWP(4,$key),$s1);
 		&mov	(&DWP(8,$key),$s2);
@@ -1158,7 +2271,7 @@ my $mark=&DWP(60+240,"esp");	#copy of aes_key->rounds
 		&mov	(&DWP(8,$acc),$s2);
 		&mov	(&DWP(12,$acc),$s3);
 
-		&lea	($acc,&DWP(16,$acc));
+		&lea	($acc,&DWP(16,$acc));	# advance out
 		&mov	($_out,$acc);		# save out
 
 		&lea	($acc,$ivec);
@@ -1173,40 +2286,340 @@ my $mark=&DWP(60+240,"esp");	#copy of aes_key->rounds
 		&mov	(&DWP(12,$key),$s3);
 
 		&mov	($acc,$_inp);		# load inp
+		&mov	($s2,$_len);		# load len
+		&lea	($acc,&DWP(16,$acc));	# advance inp
+		&mov	($_inp,$acc);		# save inp
+		&sub	($s2,16);		# decrease len
+		&mov	($_len,$s2);		# save len
+	&jnz	(&label("fast_dec_in_place_loop"));
+
+    &set_label("fast_dec_out",4);
+	&cmp	($mark,0);		# was the key schedule copied?
+	&mov	("edi",$_key);
+	&je	(&label("skip_dzero"));
+	# zero copy of key schedule
+	&mov	("ecx",240/4);
+	&xor	("eax","eax");
+	&align	(4);
+	&data_word(0xABF3F689);	# rep stosd
+	&set_label("skip_dzero")
+	&mov	("esp",$_esp);
+	&popf	();
+	&function_end_A();
+	&pushf	();			# kludge, never executed
+
+#--------------------------- SLOW ROUTINE ---------------------------#
+&set_label("slow_way",16);
+
+	&mov	($s0,&DWP(0,$s0)) if (!$x86only);# load OPENSSL_ia32cap
+	&mov	($key,&wparam(3));	# load key
+
+	# pre-allocate aligned stack frame...
+	&lea	($acc,&DWP(-80,"esp"));
+	&and	($acc,-64);
+
+	# ... and make sure it doesn't alias with $key modulo 1024
+	&lea	($s1,&DWP(-80-63,$key));
+	&sub	($s1,$acc);
+	&neg	($s1);
+	&and	($s1,0x3C0);	# modulo 1024, but aligned to cache-line
+	&sub	($acc,$s1);
+
+	# pick S-box copy which can't overlap with stack frame or $key
+	&lea	($s1,&DWP(768,$acc));
+	&sub	($s1,$tbl);
+	&and	($s1,0x300);
+	&lea	($tbl,&DWP(2048+128,$tbl,$s1));
+
+	&lea	($s3,&wparam(0));	# pointer to parameter block
+
+	&exch	("esp",$acc);
+	&add	("esp",4);		# reserve for return address!
+	&mov	($_tbl,$tbl);		# save %ebp
+	&mov	($_esp,$acc);		# save %esp
+	&mov	($_tmp,$s0);		# save OPENSSL_ia32cap
+
+	&mov	($s0,&DWP(0,$s3));	# load inp
+	&mov	($s1,&DWP(4,$s3));	# load out
+	#&mov	($s2,&DWP(8,$s3));	# load len
+	#&mov	($key,&DWP(12,$s3));	# load key
+	&mov	($acc,&DWP(16,$s3));	# load ivp
+	&mov	($s3,&DWP(20,$s3));	# load enc flag
+
+	&mov	($_inp,$s0);		# save copy of inp
+	&mov	($_out,$s1);		# save copy of out
+	&mov	($_len,$s2);		# save copy of len
+	&mov	($_key,$key);		# save copy of key
+	&mov	($_ivp,$acc);		# save copy of ivp
+
+	&mov	($key,$acc);
+	&mov	($acc,$s0);
+
+	&cmp	($s3,0);
+	&je	(&label("slow_decrypt"));
+
+#--------------------------- SLOW ENCRYPT ---------------------------#
+	&cmp	($s2,16);
+	&mov	($s3,$s1);
+	&jb	(&label("slow_enc_tail"));
+
+					if (!$x86only) {
+	&bt	($_tmp,25);		# check for SSE bit
+	&jnc	(&label("slow_enc_x86"));
 
-		&lea	($acc,&DWP(16,$acc));
+	&movq	("mm0",&QWP(0,$key));	# load iv
+	&movq	("mm4",&QWP(8,$key));
+
+	&set_label("slow_enc_loop_sse",16);
+		&pxor	("mm0",&QWP(0,$acc));	# xor input data
+		&pxor	("mm4",&QWP(8,$acc));
+
+		&mov	($key,$_key);
+		&call	("_sse_AES_encrypt_compact");
+
+		&mov	($acc,$_inp);		# load inp
+		&mov	($key,$_out);		# load out
+		&mov	($s2,$_len);		# load len
+
+		&movq	(&QWP(0,$key),"mm0");	# save output data
+		&movq	(&QWP(8,$key),"mm4");
+
+		&lea	($acc,&DWP(16,$acc));	# advance inp
 		&mov	($_inp,$acc);		# save inp
+		&lea	($s3,&DWP(16,$key));	# advance out
+		&mov	($_out,$s3);		# save out
+		&sub	($s2,16);		# decrease len
+		&cmp	($s2,16);
+		&mov	($_len,$s2);		# save len
+	&jae	(&label("slow_enc_loop_sse"));
+	&test	($s2,15);
+	&jnz	(&label("slow_enc_tail"));
+	&mov	($acc,$_ivp);		# load ivp
+	&movq	(&QWP(0,$acc),"mm0");	# save ivec
+	&movq	(&QWP(8,$acc),"mm4");
+	&emms	();
+	&mov	("esp",$_esp);
+	&popf	();
+	&function_end_A();
+	&pushf	();			# kludge, never executed
+					}
+    &set_label("slow_enc_x86",16);
+	&mov	($s0,&DWP(0,$key));	# load iv
+	&mov	($s1,&DWP(4,$key));
+
+	&set_label("slow_enc_loop_x86",4);
+		&mov	($s2,&DWP(8,$key));
+		&mov	($s3,&DWP(12,$key));
+
+		&xor	($s0,&DWP(0,$acc));	# xor input data
+		&xor	($s1,&DWP(4,$acc));
+		&xor	($s2,&DWP(8,$acc));
+		&xor	($s3,&DWP(12,$acc));
+
+		&mov	($key,$_key);		# load key
+		&call	("_x86_AES_encrypt_compact");
+
+		&mov	($acc,$_inp);		# load inp
+		&mov	($key,$_out);		# load out
+
+		&mov	(&DWP(0,$key),$s0);	# save output data
+		&mov	(&DWP(4,$key),$s1);
+		&mov	(&DWP(8,$key),$s2);
+		&mov	(&DWP(12,$key),$s3);
 
 		&mov	($s2,$_len);		# load len
-		&sub	($s2,16);
-		&jc	(&label("dec_in_place_partial"));
+		&lea	($acc,&DWP(16,$acc));	# advance inp
+		&mov	($_inp,$acc);		# save inp
+		&lea	($s3,&DWP(16,$key));	# advance out
+		&mov	($_out,$s3);		# save out
+		&sub	($s2,16);		# decrease len
+		&cmp	($s2,16);
 		&mov	($_len,$s2);		# save len
-	&jnz	(&label("dec_in_place_loop"));
-	&jmp	(&label("dec_out"));
-
-    &align	(4);
-    &set_label("dec_in_place_partial");
-	# one can argue if this is actually required...
-	&mov	($key eq "edi" ? $key : "",$_out);
-	&lea	($acc eq "esi" ? $acc : "",$ivec);
+	&jae	(&label("slow_enc_loop_x86"));
+	&test	($s2,15);
+	&jnz	(&label("slow_enc_tail"));
+	&mov	($acc,$_ivp);		# load ivp
+	&mov	($s2,&DWP(8,$key));	# restore last dwords
+	&mov	($s3,&DWP(12,$key));
+	&mov	(&DWP(0,$acc),$s0);	# save ivec
+	&mov	(&DWP(4,$acc),$s1);
+	&mov	(&DWP(8,$acc),$s2);
+	&mov	(&DWP(12,$acc),$s3);
+
+	&mov	("esp",$_esp);
+	&popf	();
+	&function_end_A();
+	&pushf	();			# kludge, never executed
+
+    &set_label("slow_enc_tail",16);
+	&emms	()	if (!$x86only);
+	&mov	($key eq "edi"? $key:"",$s3);	# load out to edi
+	&mov	($s1,16);
+	&sub	($s1,$s2);
+	&cmp	($key,$acc eq "esi"? $acc:"");	# compare with inp
+	&je	(&label("enc_in_place"));
+	&align	(4);
+	&data_word(0xA4F3F689);	# rep movsb	# copy input
+	&jmp	(&label("enc_skip_in_place"));
+    &set_label("enc_in_place");
 	&lea	($key,&DWP(0,$key,$s2));
-	&lea	($acc,&DWP(16,$acc,$s2));
-	&neg	($s2 eq "ecx" ? $s2 : "");
-	&data_word(0xA4F3F689);	# rep movsb	# restore tail
-
-    &align	(4);
-    &set_label("dec_out");
-    &cmp	($mark,0);		# was the key schedule copied?
-    &mov	("edi",$_key);
-    &je		(&label("skip_dzero"));
-    # zero copy of key schedule
-    &mov	("ecx",240/4);
-    &xor	("eax","eax");
-    &align	(4);
-    &data_word(0xABF3F689);	# rep stosd
-    &set_label("skip_dzero")
-    &mov	("esp",$_esp);
-    &popf	();
+    &set_label("enc_skip_in_place");
+	&mov	($s2,$s1);
+	&xor	($s0,$s0);
+	&align	(4);
+	&data_word(0xAAF3F689);	# rep stosb	# zero tail
+
+	&mov	($key,$_ivp);			# restore ivp
+	&mov	($acc,$s3);			# output as input
+	&mov	($s0,&DWP(0,$key));
+	&mov	($s1,&DWP(4,$key));
+	&mov	($_len,16);			# len=16
+	&jmp	(&label("slow_enc_loop_x86"));	# one more spin...
+
+#--------------------------- SLOW DECRYPT ---------------------------#
+&set_label("slow_decrypt",16);
+					if (!$x86only) {
+	&bt	($_tmp,25);		# check for SSE bit
+	&jnc	(&label("slow_dec_loop_x86"));
+
+	&set_label("slow_dec_loop_sse",4);
+		&movq	("mm0",&QWP(0,$acc));	# read input
+		&movq	("mm4",&QWP(8,$acc));
+
+		&mov	($key,$_key);
+		&call	("_sse_AES_decrypt_compact");
+
+		&mov	($acc,$_inp);		# load inp
+		&lea	($s0,$ivec);
+		&mov	($s1,$_out);		# load out
+		&mov	($s2,$_len);		# load len
+		&mov	($key,$_ivp);		# load ivp
+
+		&movq	("mm1",&QWP(0,$acc));	# re-read input
+		&movq	("mm5",&QWP(8,$acc));
+
+		&pxor	("mm0",&QWP(0,$key));	# xor iv
+		&pxor	("mm4",&QWP(8,$key));
+
+		&movq	(&QWP(0,$key),"mm1");	# copy input to iv
+		&movq	(&QWP(8,$key),"mm5");
+
+		&sub	($s2,16);		# decrease len
+		&jc	(&label("slow_dec_partial_sse"));
+
+		&movq	(&QWP(0,$s1),"mm0");	# write output
+		&movq	(&QWP(8,$s1),"mm4");
+
+		&lea	($s1,&DWP(16,$s1));	# advance out
+		&mov	($_out,$s1);		# save out
+		&lea	($acc,&DWP(16,$acc));	# advance inp
+		&mov	($_inp,$acc);		# save inp
+		&mov	($_len,$s2);		# save len
+	&jnz	(&label("slow_dec_loop_sse"));
+	&emms	();
+	&mov	("esp",$_esp);
+	&popf	();
+	&function_end_A();
+	&pushf	();			# kludge, never executed
+
+    &set_label("slow_dec_partial_sse",16);
+	&movq	(&QWP(0,$s0),"mm0");	# save output to temp
+	&movq	(&QWP(8,$s0),"mm4");
+	&emms	();
+
+	&add	($s2 eq "ecx" ? "ecx":"",16);
+	&mov	("edi",$s1);		# out
+	&mov	("esi",$s0);		# temp
+	&align	(4);
+	&data_word(0xA4F3F689);		# rep movsb # copy partial output
+
+	&mov	("esp",$_esp);
+	&popf	();
+	&function_end_A();
+	&pushf	();			# kludge, never executed
+					}
+	&set_label("slow_dec_loop_x86",16);
+		&mov	($s0,&DWP(0,$acc));	# read input
+		&mov	($s1,&DWP(4,$acc));
+		&mov	($s2,&DWP(8,$acc));
+		&mov	($s3,&DWP(12,$acc));
+
+		&lea	($key,$ivec);
+		&mov	(&DWP(0,$key),$s0);	# copy to temp
+		&mov	(&DWP(4,$key),$s1);
+		&mov	(&DWP(8,$key),$s2);
+		&mov	(&DWP(12,$key),$s3);
+
+		&mov	($key,$_key);		# load key
+		&call	("_x86_AES_decrypt_compact");
+
+		&mov	($key,$_ivp);		# load ivp
+		&mov	($acc,$_len);		# load len
+		&xor	($s0,&DWP(0,$key));	# xor iv
+		&xor	($s1,&DWP(4,$key));
+		&xor	($s2,&DWP(8,$key));
+		&xor	($s3,&DWP(12,$key));
+
+		&sub	($acc,16);
+		&jc	(&label("slow_dec_partial_x86"));
+
+		&mov	($_len,$acc);		# save len
+		&mov	($acc,$_out);		# load out
+
+		&mov	(&DWP(0,$acc),$s0);	# write output
+		&mov	(&DWP(4,$acc),$s1);
+		&mov	(&DWP(8,$acc),$s2);
+		&mov	(&DWP(12,$acc),$s3);
+
+		&lea	($acc,&DWP(16,$acc));	# advance out
+		&mov	($_out,$acc);		# save out
+
+		&lea	($acc,$ivec);
+		&mov	($s0,&DWP(0,$acc));	# read temp
+		&mov	($s1,&DWP(4,$acc));
+		&mov	($s2,&DWP(8,$acc));
+		&mov	($s3,&DWP(12,$acc));
+
+		&mov	(&DWP(0,$key),$s0);	# copy it to iv
+		&mov	(&DWP(4,$key),$s1);
+		&mov	(&DWP(8,$key),$s2);
+		&mov	(&DWP(12,$key),$s3);
+
+		&mov	($acc,$_inp);		# load inp
+		&lea	($acc,&DWP(16,$acc));	# advance inp
+		&mov	($_inp,$acc);		# save inp
+	&jnz	(&label("slow_dec_loop_x86"));
+	&mov	("esp",$_esp);
+	&popf	();
+	&function_end_A();
+	&pushf	();			# kludge, never executed
+
+    &set_label("slow_dec_partial_x86",16);
+	&lea	($acc,$ivec);
+	&mov	(&DWP(0,$acc),$s0);	# save output to temp
+	&mov	(&DWP(4,$acc),$s1);
+	&mov	(&DWP(8,$acc),$s2);
+	&mov	(&DWP(12,$acc),$s3);
+
+	&mov	($acc,$_inp);
+	&mov	($s0,&DWP(0,$acc));	# re-read input
+	&mov	($s1,&DWP(4,$acc));
+	&mov	($s2,&DWP(8,$acc));
+	&mov	($s3,&DWP(12,$acc));
+
+	&mov	(&DWP(0,$key),$s0);	# copy it to iv
+	&mov	(&DWP(4,$key),$s1);
+	&mov	(&DWP(8,$key),$s2);
+	&mov	(&DWP(12,$key),$s3);
+
+	&mov	("ecx",$_len);
+	&mov	("edi",$_out);
+	&lea	("esi",$ivec);
+	&align	(4);
+	&data_word(0xA4F3F689);		# rep movsb # copy partial output
+
+	&mov	("esp",$_esp);
+	&popf	();
 &function_end("AES_cbc_encrypt");
 }
 
@@ -1215,35 +2628,31 @@ my $mark=&DWP(60+240,"esp");	#copy of aes_key->rounds
 sub enckey()
 {
 	&movz	("esi",&LB("edx"));		# rk[i]>>0
-	&mov	("ebx",&DWP(2,"ebp","esi",8));
+	&movz	("ebx",&BP(-128,$tbl,"esi",1));
 	&movz	("esi",&HB("edx"));		# rk[i]>>8
-	&and	("ebx",0xFF000000);
+	&shl	("ebx",24);
 	&xor	("eax","ebx");
 
-	&mov	("ebx",&DWP(2,"ebp","esi",8));
+	&movz	("ebx",&BP(-128,$tbl,"esi",1));
 	&shr	("edx",16);
-	&and	("ebx",0x000000FF);
 	&movz	("esi",&LB("edx"));		# rk[i]>>16
 	&xor	("eax","ebx");
 
-	&mov	("ebx",&DWP(0,"ebp","esi",8));
+	&movz	("ebx",&BP(-128,$tbl,"esi",1));
 	&movz	("esi",&HB("edx"));		# rk[i]>>24
-	&and	("ebx",0x0000FF00);
+	&shl	("ebx",8);
 	&xor	("eax","ebx");
 
-	&mov	("ebx",&DWP(0,"ebp","esi",8));
-	&and	("ebx",0x00FF0000);
+	&movz	("ebx",&BP(-128,$tbl,"esi",1));
+	&shl	("ebx",16);
 	&xor	("eax","ebx");
 
-	&xor	("eax",&DWP(2048,"ebp","ecx",4));	# rcon
+	&xor	("eax",&DWP(1024-128,$tbl,"ecx",4));	# rcon
 }
 
-# int AES_set_encrypt_key(const unsigned char *userKey, const int bits,
-#                        AES_KEY *key)
-&public_label("AES_Te");
-&function_begin("AES_set_encrypt_key");
-	&mov	("esi",&wparam(0));		# user supplied key
-	&mov	("edi",&wparam(2));		# private key schedule
+&function_begin("_x86_AES_set_encrypt_key");
+	&mov	("esi",&wparam(1));		# user supplied key
+	&mov	("edi",&wparam(3));		# private key schedule
 
 	&test	("esi",-1);
 	&jz	(&label("badpointer"));
@@ -1252,10 +2661,21 @@ sub enckey()
 
 	&call	(&label("pic_point"));
 	&set_label("pic_point");
-	&blindpop("ebp");
-	&lea	("ebp",&DWP(&label("AES_Te")."-".&label("pic_point"),"ebp"));
-
-	&mov	("ecx",&wparam(1));		# number of bits in key
+	&blindpop($tbl);
+	&lea	($tbl,&DWP(&label("AES_Te")."-".&label("pic_point"),$tbl));
+	&lea	($tbl,&DWP(2048+128,$tbl));
+
+	# prefetch Te4
+	&mov	("eax",&DWP(0-128,$tbl));
+	&mov	("ebx",&DWP(32-128,$tbl));
+	&mov	("ecx",&DWP(64-128,$tbl));
+	&mov	("edx",&DWP(96-128,$tbl));
+	&mov	("eax",&DWP(128-128,$tbl));
+	&mov	("ebx",&DWP(160-128,$tbl));
+	&mov	("ecx",&DWP(192-128,$tbl));
+	&mov	("edx",&DWP(224-128,$tbl));
+
+	&mov	("ecx",&wparam(2));		# number of bits in key
 	&cmp	("ecx",128);
 	&je	(&label("10rounds"));
 	&cmp	("ecx",192);
@@ -1394,24 +2814,23 @@ sub enckey()
 		&mov	("edx","eax");
 		&mov	("eax",&DWP(16,"edi"));		# rk[4]
 		&movz	("esi",&LB("edx"));		# rk[11]>>0
-		&mov	("ebx",&DWP(2,"ebp","esi",8));
+		&movz	("ebx",&BP(-128,$tbl,"esi",1));
 		&movz	("esi",&HB("edx"));		# rk[11]>>8
-		&and	("ebx",0x000000FF);
 		&xor	("eax","ebx");
 
-		&mov	("ebx",&DWP(0,"ebp","esi",8));
+		&movz	("ebx",&BP(-128,$tbl,"esi",1));
 		&shr	("edx",16);
-		&and	("ebx",0x0000FF00);
+		&shl	("ebx",8);
 		&movz	("esi",&LB("edx"));		# rk[11]>>16
 		&xor	("eax","ebx");
 
-		&mov	("ebx",&DWP(0,"ebp","esi",8));
+		&movz	("ebx",&BP(-128,$tbl,"esi",1));
 		&movz	("esi",&HB("edx"));		# rk[11]>>24
-		&and	("ebx",0x00FF0000);
+		&shl	("ebx",16);
 		&xor	("eax","ebx");
 
-		&mov	("ebx",&DWP(2,"ebp","esi",8));
-		&and	("ebx",0xFF000000);
+		&movz	("ebx",&BP(-128,$tbl,"esi",1));
+		&shl	("ebx",24);
 		&xor	("eax","ebx");
 
 		&mov	(&DWP(48,"edi"),"eax");		# rk[12]
@@ -1433,43 +2852,74 @@ sub enckey()
     &set_label("badpointer");
 	&mov	("eax",-1);
     &set_label("exit");
-&function_end("AES_set_encrypt_key");
+&function_end("_x86_AES_set_encrypt_key");
 
-sub deckey()
-{ my ($i,$ptr,$te,$td) = @_;
+# int AES_set_encrypt_key(const unsigned char *userKey, const int bits,
+#                        AES_KEY *key)
+&function_begin_B("AES_set_encrypt_key");
+	&call	("_x86_AES_set_encrypt_key");
+	&ret	();
+&function_end_B("AES_set_encrypt_key");
 
-	&mov	("eax",&DWP($i,$ptr));
-	&mov	("edx","eax");
-	&movz	("ebx",&HB("eax"));
-	&shr	("edx",16);
-	&and	("eax",0xFF);
-	&movz	("eax",&BP(2,$te,"eax",8));
-	&movz	("ebx",&BP(2,$te,"ebx",8));
-	&mov	("eax",&DWP(0,$td,"eax",8));
-	&xor	("eax",&DWP(3,$td,"ebx",8));
-	&movz	("ebx",&HB("edx"));
-	&and	("edx",0xFF);
-	&movz	("edx",&BP(2,$te,"edx",8));
-	&movz	("ebx",&BP(2,$te,"ebx",8));
-	&xor	("eax",&DWP(2,$td,"edx",8));
-	&xor	("eax",&DWP(1,$td,"ebx",8));
-	&mov	(&DWP($i,$ptr),"eax");
+sub deckey()
+{ my ($i,$key,$tp1,$tp2,$tp4,$tp8) = @_;
+  my $tmp = $tbl;
+
+	&mov	($acc,$tp1);
+	&and	($acc,0x80808080);
+	&mov	($tmp,$acc);
+	&shr	($tmp,7);
+	&lea	($tp2,&DWP(0,$tp1,$tp1));
+	&sub	($acc,$tmp);
+	&and	($tp2,0xfefefefe);
+	&and	($acc,0x1b1b1b1b);
+	&xor	($acc,$tp2);
+	&mov	($tp2,$acc);
+
+	&and	($acc,0x80808080);
+	&mov	($tmp,$acc);
+	&shr	($tmp,7);
+	&lea	($tp4,&DWP(0,$tp2,$tp2));
+	&sub	($acc,$tmp);
+	&and	($tp4,0xfefefefe);
+	&and	($acc,0x1b1b1b1b);
+	 &xor	($tp2,$tp1);	# tp2^tp1
+	&xor	($acc,$tp4);
+	&mov	($tp4,$acc);
+
+	&and	($acc,0x80808080);
+	&mov	($tmp,$acc);
+	&shr	($tmp,7);
+	&lea	($tp8,&DWP(0,$tp4,$tp4));
+	 &xor	($tp4,$tp1);	# tp4^tp1
+	&sub	($acc,$tmp);
+	&and	($tp8,0xfefefefe);
+	&and	($acc,0x1b1b1b1b);
+	 &rotl	($tp1,8);	# = ROTATE(tp1,8)
+	&xor	($tp8,$acc);
+
+	&mov	($tmp,&DWP(4*($i+1),$key));	# modulo-scheduled load
+
+	&xor	($tp1,$tp2);
+	&xor	($tp2,$tp8);
+	&xor	($tp1,$tp4);
+	&rotl	($tp2,24);
+	&xor	($tp4,$tp8);
+	&xor	($tp1,$tp8);	# ^= tp8^(tp4^tp1)^(tp2^tp1)
+	&rotl	($tp4,16);
+	&xor	($tp1,$tp2);	# ^= ROTATE(tp8^tp2^tp1,24)
+	&rotl	($tp8,8);
+	&xor	($tp1,$tp4);	# ^= ROTATE(tp8^tp4^tp1,16)
+	&mov	($tp2,$tmp);
+	&xor	($tp1,$tp8);	# ^= ROTATE(tp8,8)
+
+	&mov	(&DWP(4*$i,$key),$tp1);
 }
 
 # int AES_set_decrypt_key(const unsigned char *userKey, const int bits,
 #                        AES_KEY *key)
-&public_label("AES_Td");
-&public_label("AES_Te");
 &function_begin_B("AES_set_decrypt_key");
-	&mov	("eax",&wparam(0));
-	&mov	("ecx",&wparam(1));
-	&mov	("edx",&wparam(2));
-	&sub	("esp",12);
-	&mov	(&DWP(0,"esp"),"eax");
-	&mov	(&DWP(4,"esp"),"ecx");
-	&mov	(&DWP(8,"esp"),"edx");
-	&call	("AES_set_encrypt_key");
-	&add	("esp",12);
+	&call	("_x86_AES_set_encrypt_key");
 	&cmp	("eax",0);
 	&je	(&label("proceed"));
 	&ret	();
@@ -1485,8 +2935,7 @@ sub deckey()
 	&lea	("ecx",&DWP(0,"","ecx",4));
 	&lea	("edi",&DWP(0,"esi","ecx",4));	# pointer to last chunk
 
-	&align	(4);
-	&set_label("invert");			# invert order of chunks
+	&set_label("invert",4);			# invert order of chunks
 		&mov	("eax",&DWP(0,"esi"));
 		&mov	("ebx",&DWP(4,"esi"));
 		&mov	("ecx",&DWP(0,"edi"));
@@ -1508,26 +2957,24 @@ sub deckey()
 		&cmp	("esi","edi");
 	&jne	(&label("invert"));
 
-	&call	(&label("pic_point"));
-	&set_label("pic_point");
-	blindpop("ebp");
-	&lea	("edi",&DWP(&label("AES_Td")."-".&label("pic_point"),"ebp"));
-	&lea	("ebp",&DWP(&label("AES_Te")."-".&label("pic_point"),"ebp"));
-
-	&mov	("esi",&wparam(2));
-	&mov	("ecx",&DWP(240,"esi"));	# pull number of rounds
-	&dec	("ecx");
-	&align	(4);
-	&set_label("permute");			# permute the key schedule
-		&add	("esi",16);
-		&deckey	(0,"esi","ebp","edi");
-		&deckey	(4,"esi","ebp","edi");
-		&deckey	(8,"esi","ebp","edi");
-		&deckey	(12,"esi","ebp","edi");
-		&dec	("ecx");
-	&jnz	(&label("permute"));
+	&mov	($key,&wparam(2));
+	&mov	($acc,&DWP(240,$key));		# pull number of rounds
+	&lea	($acc,&DWP(-2,$acc,$acc));
+	&lea	($acc,&DWP(0,$key,$acc,8));
+	&mov	(&wparam(2),$acc);
+
+	&mov	($s0,&DWP(16,$key));		# modulo-scheduled load
+	&set_label("permute",4);		# permute the key schedule
+		&add	($key,16);
+		&deckey	(0,$key,$s0,$s1,$s2,$s3);
+		&deckey	(1,$key,$s1,$s2,$s3,$s0);
+		&deckey	(2,$key,$s2,$s3,$s0,$s1);
+		&deckey	(3,$key,$s3,$s0,$s1,$s2);
+		&cmp	($key,&wparam(2));
+	&jb	(&label("permute"));
 
 	&xor	("eax","eax");			# return success
 &function_end("AES_set_decrypt_key");
+&asciz("AES for x86, CRYPTOGAMS by <appro\@openssl.org>");
 
 &asm_finish();
diff --git a/openssl/crypto/aes/asm/aes-armv4.pl b/openssl/crypto/aes/asm/aes-armv4.pl
index 15742c1ec..690244111 100644
--- a/openssl/crypto/aes/asm/aes-armv4.pl
+++ b/openssl/crypto/aes/asm/aes-armv4.pl
@@ -1024,6 +1024,7 @@ _armv4_AES_decrypt:
 	mov	pc,lr			@ return
 .size	_armv4_AES_decrypt,.-_armv4_AES_decrypt
 .asciz	"AES for ARMv4, CRYPTOGAMS by <appro\@openssl.org>"
+.align	2
 ___
 
 $code =~ s/\bbx\s+lr\b/.word\t0xe12fff1e/gm;	# make it possible to compile with -march=armv4
diff --git a/openssl/crypto/aes/asm/aes-s390x.pl b/openssl/crypto/aes/asm/aes-s390x.pl
index 4b27afd92..7e0188929 100644
--- a/openssl/crypto/aes/asm/aes-s390x.pl
+++ b/openssl/crypto/aes/asm/aes-s390x.pl
@@ -765,6 +765,11 @@ $code.=<<___ if (!$softonly);
 	srl	%r5,6
 	ar	%r5,%r0
 
+	larl	%r1,OPENSSL_s390xcap_P
+	lg	%r0,0(%r1)
+	tmhl	%r0,0x4000	# check for message-security assist
+	jz	.Lekey_internal
+
 	lghi	%r0,0		# query capability vector
 	la	%r1,16($sp)
 	.long	0xb92f0042	# kmc %r4,%r2
@@ -1323,6 +1328,7 @@ $code.=<<___;
 4:	ex	$len,0($s1)
 	j	.Lcbc_dec_exit
 .size	AES_cbc_encrypt,.-AES_cbc_encrypt
+.comm  OPENSSL_s390xcap_P,8,8
 ___
 }
 $code.=<<___;
diff --git a/openssl/crypto/aes/asm/aes-x86_64.pl b/openssl/crypto/aes/asm/aes-x86_64.pl
index f616f1751..a545e892a 100644
--- a/openssl/crypto/aes/asm/aes-x86_64.pl
+++ b/openssl/crypto/aes/asm/aes-x86_64.pl
@@ -2,11 +2,12 @@
 #
 # ====================================================================
 # Written by Andy Polyakov <appro@fy.chalmers.se> for the OpenSSL
-# project. Rights for redistribution and usage in source and binary
-# forms are granted according to the OpenSSL license.
+# project. The module is, however, dual licensed under OpenSSL and
+# CRYPTOGAMS licenses depending on where you obtain it. For further
+# details see http://www.openssl.org/~appro/cryptogams/.
 # ====================================================================
 #
-# Version 1.2.
+# Version 2.1.
 #
 # aes-*-cbc benchmarks are improved by >70% [compared to gcc 3.3.2 on
 # Opteron 240 CPU] plus all the bells-n-whistles from 32-bit version
@@ -17,17 +18,29 @@
 #
 # Performance in number of cycles per processed byte for 128-bit key:
 #
-#		ECB		CBC encrypt
-# AMD64		13.7		13.0(*)
-# EM64T		20.2		18.6(*)
+#		ECB encrypt	ECB decrypt	CBC large chunk
+# AMD64		33		41		13.0
+# EM64T		38		59		18.6(*)
+# Core 2	30		43		14.5(*)
 #
-# (*)	CBC benchmarks are better than ECB thanks to custom ABI used
-#	by the private block encryption function.
+# (*) with hyper-threading off
+
+$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 $flavour $output";
 
 $verticalspin=1;	# unlike 32-bit version $verticalspin performs
 			# ~15% better on both AMD and Intel cores
-$output=shift;
-open STDOUT,"| $^X ../perlasm/x86_64-xlate.pl $output";
+$speed_limit=512;	# see aes-586.pl for details
 
 $code=".text\n";
 
@@ -35,9 +48,9 @@ $s0="%eax";
 $s1="%ebx";
 $s2="%ecx";
 $s3="%edx";
-$acc0="%esi";
-$acc1="%edi";
-$acc2="%ebp";
+$acc0="%esi";	$mask80="%rsi";
+$acc1="%edi";	$maskfe="%rdi";
+$acc2="%ebp";	$mask1b="%rbp";
 $inp="%r8";
 $out="%r9";
 $t0="%r10d";
@@ -51,6 +64,8 @@ sub hi() { my $r=shift;	$r =~ s/%[er]([a-d])x/%\1h/;	$r; }
 sub lo() { my $r=shift;	$r =~ s/%[er]([a-d])x/%\1l/;
 			$r =~ s/%[er]([sd]i)/%\1l/;
 			$r =~ s/%(r[0-9]+)[d]?/%\1b/;	$r; }
+sub LO() { my $r=shift; $r =~ s/%r([a-z]+)/%e\1/;
+			$r =~ s/%r([0-9]+)/%r\1d/;	$r; }
 sub _data_word()
 { my $i;
     while(defined($i=shift)) { $code.=sprintf".long\t0x%08x,0x%08x\n",$i,$i; }
@@ -138,22 +153,17 @@ $code.=<<___;
 	movzb	`&lo("$s0")`,$acc0
 	movzb	`&lo("$s1")`,$acc1
 	movzb	`&lo("$s2")`,$acc2
-	mov	2($sbox,$acc0,8),$t0
-	mov	2($sbox,$acc1,8),$t1
-	mov	2($sbox,$acc2,8),$t2
-
-	and	\$0x000000ff,$t0
-	and	\$0x000000ff,$t1
-	and	\$0x000000ff,$t2
+	movzb	2($sbox,$acc0,8),$t0
+	movzb	2($sbox,$acc1,8),$t1
+	movzb	2($sbox,$acc2,8),$t2
 
 	movzb	`&lo("$s3")`,$acc0
 	movzb	`&hi("$s1")`,$acc1
 	movzb	`&hi("$s2")`,$acc2
-	mov	2($sbox,$acc0,8),$t3
+	movzb	2($sbox,$acc0,8),$t3
 	mov	0($sbox,$acc1,8),$acc1	#$t0
 	mov	0($sbox,$acc2,8),$acc2	#$t1
 
-	and	\$0x000000ff,$t3
 	and	\$0x0000ff00,$acc1
 	and	\$0x0000ff00,$acc2
 
@@ -345,6 +355,234 @@ $code.=<<___;
 .size	_x86_64_AES_encrypt,.-_x86_64_AES_encrypt
 ___
 
+# it's possible to implement this by shifting tN by 8, filling least
+# significant byte with byte load and finally bswap-ing at the end,
+# but such partial register load kills Core 2...
+sub enccompactvert()
+{ my ($t3,$t4,$t5)=("%r8d","%r9d","%r13d");
+
+$code.=<<___;
+	movzb	`&lo("$s0")`,$t0
+	movzb	`&lo("$s1")`,$t1
+	movzb	`&lo("$s2")`,$t2
+	movzb	($sbox,$t0,1),$t0
+	movzb	($sbox,$t1,1),$t1
+	movzb	($sbox,$t2,1),$t2
+
+	movzb	`&lo("$s3")`,$t3
+	movzb	`&hi("$s1")`,$acc0
+	movzb	`&hi("$s2")`,$acc1
+	movzb	($sbox,$t3,1),$t3
+	movzb	($sbox,$acc0,1),$t4	#$t0
+	movzb	($sbox,$acc1,1),$t5	#$t1
+
+	movzb	`&hi("$s3")`,$acc2
+	movzb	`&hi("$s0")`,$acc0
+	shr	\$16,$s2
+	movzb	($sbox,$acc2,1),$acc2	#$t2
+	movzb	($sbox,$acc0,1),$acc0	#$t3
+	shr	\$16,$s3
+
+	movzb	`&lo("$s2")`,$acc1
+	shl	\$8,$t4
+	shl	\$8,$t5
+	movzb	($sbox,$acc1,1),$acc1	#$t0
+	xor	$t4,$t0
+	xor	$t5,$t1
+
+	movzb	`&lo("$s3")`,$t4
+	shr	\$16,$s0
+	shr	\$16,$s1
+	movzb	`&lo("$s0")`,$t5
+	shl	\$8,$acc2
+	shl	\$8,$acc0
+	movzb	($sbox,$t4,1),$t4	#$t1
+	movzb	($sbox,$t5,1),$t5	#$t2
+	xor	$acc2,$t2
+	xor	$acc0,$t3
+
+	movzb	`&lo("$s1")`,$acc2
+	movzb	`&hi("$s3")`,$acc0
+	shl	\$16,$acc1
+	movzb	($sbox,$acc2,1),$acc2	#$t3
+	movzb	($sbox,$acc0,1),$acc0	#$t0
+	xor	$acc1,$t0
+
+	movzb	`&hi("$s0")`,$acc1
+	shr	\$8,$s2
+	shr	\$8,$s1
+	movzb	($sbox,$acc1,1),$acc1	#$t1
+	movzb	($sbox,$s2,1),$s3	#$t3
+	movzb	($sbox,$s1,1),$s2	#$t2
+	shl	\$16,$t4
+	shl	\$16,$t5
+	shl	\$16,$acc2
+	xor	$t4,$t1
+	xor	$t5,$t2
+	xor	$acc2,$t3
+
+	shl	\$24,$acc0
+	shl	\$24,$acc1
+	shl	\$24,$s3
+	xor	$acc0,$t0
+	shl	\$24,$s2
+	xor	$acc1,$t1
+	mov	$t0,$s0
+	mov	$t1,$s1
+	xor	$t2,$s2
+	xor	$t3,$s3
+___
+}
+
+sub enctransform_ref()
+{ my $sn = shift;
+  my ($acc,$r2,$tmp)=("%r8d","%r9d","%r13d");
+
+$code.=<<___;
+	mov	$sn,$acc
+	and	\$0x80808080,$acc
+	mov	$acc,$tmp
+	shr	\$7,$tmp
+	lea	($sn,$sn),$r2
+	sub	$tmp,$acc
+	and	\$0xfefefefe,$r2
+	and	\$0x1b1b1b1b,$acc
+	mov	$sn,$tmp
+	xor	$acc,$r2
+
+	xor	$r2,$sn
+	rol	\$24,$sn
+	xor	$r2,$sn
+	ror	\$16,$tmp
+	xor	$tmp,$sn
+	ror	\$8,$tmp
+	xor	$tmp,$sn
+___
+}
+
+# unlike decrypt case it does not pay off to parallelize enctransform
+sub enctransform()
+{ my ($t3,$r20,$r21)=($acc2,"%r8d","%r9d");
+
+$code.=<<___;
+	mov	$s0,$acc0
+	mov	$s1,$acc1
+	and	\$0x80808080,$acc0
+	and	\$0x80808080,$acc1
+	mov	$acc0,$t0
+	mov	$acc1,$t1
+	shr	\$7,$t0
+	lea	($s0,$s0),$r20
+	shr	\$7,$t1
+	lea	($s1,$s1),$r21
+	sub	$t0,$acc0
+	sub	$t1,$acc1
+	and	\$0xfefefefe,$r20
+	and	\$0xfefefefe,$r21
+	and	\$0x1b1b1b1b,$acc0
+	and	\$0x1b1b1b1b,$acc1
+	mov	$s0,$t0
+	mov	$s1,$t1
+	xor	$acc0,$r20
+	xor	$acc1,$r21
+
+	xor	$r20,$s0
+	xor	$r21,$s1
+	 mov	$s2,$acc0
+	 mov	$s3,$acc1
+	rol	\$24,$s0
+	rol	\$24,$s1
+	 and	\$0x80808080,$acc0
+	 and	\$0x80808080,$acc1
+	xor	$r20,$s0
+	xor	$r21,$s1
+	 mov	$acc0,$t2
+	 mov	$acc1,$t3
+	ror	\$16,$t0
+	ror	\$16,$t1
+	 shr	\$7,$t2
+	 lea	($s2,$s2),$r20
+	xor	$t0,$s0
+	xor	$t1,$s1
+	 shr	\$7,$t3
+	 lea	($s3,$s3),$r21
+	ror	\$8,$t0
+	ror	\$8,$t1
+	 sub	$t2,$acc0
+	 sub	$t3,$acc1
+	xor	$t0,$s0
+	xor	$t1,$s1
+
+	and	\$0xfefefefe,$r20
+	and	\$0xfefefefe,$r21
+	and	\$0x1b1b1b1b,$acc0
+	and	\$0x1b1b1b1b,$acc1
+	mov	$s2,$t2
+	mov	$s3,$t3
+	xor	$acc0,$r20
+	xor	$acc1,$r21
+
+	xor	$r20,$s2
+	xor	$r21,$s3
+	rol	\$24,$s2
+	rol	\$24,$s3
+	xor	$r20,$s2
+	xor	$r21,$s3
+	mov	0($sbox),$acc0			# prefetch Te4
+	ror	\$16,$t2
+	ror	\$16,$t3
+	mov	64($sbox),$acc1
+	xor	$t2,$s2
+	xor	$t3,$s3
+	mov	128($sbox),$r20
+	ror	\$8,$t2
+	ror	\$8,$t3
+	mov	192($sbox),$r21
+	xor	$t2,$s2
+	xor	$t3,$s3
+___
+}
+
+$code.=<<___;
+.type	_x86_64_AES_encrypt_compact,\@abi-omnipotent
+.align	16
+_x86_64_AES_encrypt_compact:
+	lea	128($sbox),$inp			# size optimization
+	mov	0-128($inp),$acc1		# prefetch Te4
+	mov	32-128($inp),$acc2
+	mov	64-128($inp),$t0
+	mov	96-128($inp),$t1
+	mov	128-128($inp),$acc1
+	mov	160-128($inp),$acc2
+	mov	192-128($inp),$t0
+	mov	224-128($inp),$t1
+	jmp	.Lenc_loop_compact
+.align	16
+.Lenc_loop_compact:
+		xor	0($key),$s0		# xor with key
+		xor	4($key),$s1
+		xor	8($key),$s2
+		xor	12($key),$s3
+		lea	16($key),$key
+___
+		&enccompactvert();
+$code.=<<___;
+		cmp	16(%rsp),$key
+		je	.Lenc_compact_done
+___
+		&enctransform();
+$code.=<<___;
+	jmp	.Lenc_loop_compact
+.align	16
+.Lenc_compact_done:
+	xor	0($key),$s0
+	xor	4($key),$s1
+	xor	8($key),$s2
+	xor	12($key),$s3
+	.byte	0xf3,0xc3			# rep ret
+.size	_x86_64_AES_encrypt_compact,.-_x86_64_AES_encrypt_compact
+___
+
 # void AES_encrypt (const void *inp,void *out,const AES_KEY *key);
 $code.=<<___;
 .globl	AES_encrypt
@@ -358,31 +596,57 @@ AES_encrypt:
 	push	%r14
 	push	%r15
 
-	mov	%rdx,$key
-	mov	%rdi,$inp
-	mov	%rsi,$out
-
-	.picmeup	$sbox
-	lea	AES_Te-.($sbox),$sbox
-
-	mov	0($inp),$s0
-	mov	4($inp),$s1
-	mov	8($inp),$s2
-	mov	12($inp),$s3
+	# allocate frame "above" key schedule
+	mov	%rsp,%r10
+	lea	-63(%rdx),%rcx	# %rdx is key argument
+	and	\$-64,%rsp
+	sub	%rsp,%rcx
+	neg	%rcx
+	and	\$0x3c0,%rcx
+	sub	%rcx,%rsp
+	sub	\$32,%rsp
 
-	call	_x86_64_AES_encrypt
+	mov	%rsi,16(%rsp)	# save out
+	mov	%r10,24(%rsp)	# save real stack pointer
+.Lenc_prologue:
 
-	mov	$s0,0($out)
+	mov	%rdx,$key
+	mov	240($key),$rnds	# load rounds
+
+	mov	0(%rdi),$s0	# load input vector
+	mov	4(%rdi),$s1
+	mov	8(%rdi),$s2
+	mov	12(%rdi),$s3
+
+	shl	\$4,$rnds
+	lea	($key,$rnds),%rbp
+	mov	$key,(%rsp)	# key schedule
+	mov	%rbp,8(%rsp)	# end of key schedule
+
+	# pick Te4 copy which can't "overlap" with stack frame or key schedule
+	lea	.LAES_Te+2048(%rip),$sbox
+	lea	768(%rsp),%rbp
+	sub	$sbox,%rbp
+	and	\$0x300,%rbp
+	lea	($sbox,%rbp),$sbox
+
+	call	_x86_64_AES_encrypt_compact
+
+	mov	16(%rsp),$out	# restore out
+	mov	24(%rsp),%rsi	# restore saved stack pointer
+	mov	$s0,0($out)	# write output vector
 	mov	$s1,4($out)
 	mov	$s2,8($out)
 	mov	$s3,12($out)
 
-	pop	%r15
-	pop	%r14
-	pop	%r13
-	pop	%r12
-	pop	%rbp
-	pop	%rbx
+	mov	(%rsi),%r15
+	mov	8(%rsi),%r14
+	mov	16(%rsi),%r13
+	mov	24(%rsi),%r12
+	mov	32(%rsi),%rbp
+	mov	40(%rsi),%rbx
+	lea	48(%rsi),%rsp
+.Lenc_epilogue:
 	ret
 .size	AES_encrypt,.-AES_encrypt
 ___
@@ -453,19 +717,20 @@ sub declastvert()
 { my $t3="%r8d";	# zaps $inp!
 
 $code.=<<___;
+	lea	2048($sbox),$sbox	# size optimization
 	movzb	`&lo("$s0")`,$acc0
 	movzb	`&lo("$s1")`,$acc1
 	movzb	`&lo("$s2")`,$acc2
-	movzb	2048($sbox,$acc0,1),$t0
-	movzb	2048($sbox,$acc1,1),$t1
-	movzb	2048($sbox,$acc2,1),$t2
+	movzb	($sbox,$acc0,1),$t0
+	movzb	($sbox,$acc1,1),$t1
+	movzb	($sbox,$acc2,1),$t2
 
 	movzb	`&lo("$s3")`,$acc0
 	movzb	`&hi("$s3")`,$acc1
 	movzb	`&hi("$s0")`,$acc2
-	movzb	2048($sbox,$acc0,1),$t3
-	movzb	2048($sbox,$acc1,1),$acc1	#$t0
-	movzb	2048($sbox,$acc2,1),$acc2	#$t1
+	movzb	($sbox,$acc0,1),$t3
+	movzb	($sbox,$acc1,1),$acc1	#$t0
+	movzb	($sbox,$acc2,1),$acc2	#$t1
 
 	shl	\$8,$acc1
 	shl	\$8,$acc2
@@ -477,8 +742,8 @@ $code.=<<___;
 	movzb	`&hi("$s1")`,$acc0
 	movzb	`&hi("$s2")`,$acc1
 	shr	\$16,$s0
-	movzb	2048($sbox,$acc0,1),$acc0	#$t2
-	movzb	2048($sbox,$acc1,1),$acc1	#$t3
+	movzb	($sbox,$acc0,1),$acc0	#$t2
+	movzb	($sbox,$acc1,1),$acc1	#$t3
 
 	shl	\$8,$acc0
 	shl	\$8,$acc1
@@ -490,9 +755,9 @@ $code.=<<___;
 	movzb	`&lo("$s2")`,$acc0
 	movzb	`&lo("$s3")`,$acc1
 	movzb	`&lo("$s0")`,$acc2
-	movzb	2048($sbox,$acc0,1),$acc0	#$t0
-	movzb	2048($sbox,$acc1,1),$acc1	#$t1
-	movzb	2048($sbox,$acc2,1),$acc2	#$t2
+	movzb	($sbox,$acc0,1),$acc0	#$t0
+	movzb	($sbox,$acc1,1),$acc1	#$t1
+	movzb	($sbox,$acc2,1),$acc2	#$t2
 
 	shl	\$16,$acc0
 	shl	\$16,$acc1
@@ -505,9 +770,9 @@ $code.=<<___;
 	movzb	`&lo("$s1")`,$acc0
 	movzb	`&hi("$s1")`,$acc1
 	movzb	`&hi("$s2")`,$acc2
-	movzb	2048($sbox,$acc0,1),$acc0	#$t3
-	movzb	2048($sbox,$acc1,1),$acc1	#$t0
-	movzb	2048($sbox,$acc2,1),$acc2	#$t1
+	movzb	($sbox,$acc0,1),$acc0	#$t3
+	movzb	($sbox,$acc1,1),$acc1	#$t0
+	movzb	($sbox,$acc2,1),$acc2	#$t1
 
 	shl	\$16,$acc0
 	shl	\$24,$acc1
@@ -520,8 +785,8 @@ $code.=<<___;
 	movzb	`&hi("$s3")`,$acc0
 	movzb	`&hi("$s0")`,$acc1
 	mov	16+12($key),$s3
-	movzb	2048($sbox,$acc0,1),$acc0	#$t2
-	movzb	2048($sbox,$acc1,1),$acc1	#$t3
+	movzb	($sbox,$acc0,1),$acc0	#$t2
+	movzb	($sbox,$acc1,1),$acc1	#$t3
 	mov	16+0($key),$s0
 
 	shl	\$24,$acc0
@@ -532,6 +797,7 @@ $code.=<<___;
 
 	mov	16+4($key),$s1
 	mov	16+8($key),$s2
+	lea	-2048($sbox),$sbox
 	xor	$t0,$s0
 	xor	$t1,$s1
 	xor	$t2,$s2
@@ -659,6 +925,260 @@ $code.=<<___;
 .size	_x86_64_AES_decrypt,.-_x86_64_AES_decrypt
 ___
 
+sub deccompactvert()
+{ my ($t3,$t4,$t5)=("%r8d","%r9d","%r13d");
+
+$code.=<<___;
+	movzb	`&lo("$s0")`,$t0
+	movzb	`&lo("$s1")`,$t1
+	movzb	`&lo("$s2")`,$t2
+	movzb	($sbox,$t0,1),$t0
+	movzb	($sbox,$t1,1),$t1
+	movzb	($sbox,$t2,1),$t2
+
+	movzb	`&lo("$s3")`,$t3
+	movzb	`&hi("$s3")`,$acc0
+	movzb	`&hi("$s0")`,$acc1
+	movzb	($sbox,$t3,1),$t3
+	movzb	($sbox,$acc0,1),$t4	#$t0
+	movzb	($sbox,$acc1,1),$t5	#$t1
+
+	movzb	`&hi("$s1")`,$acc2
+	movzb	`&hi("$s2")`,$acc0
+	shr	\$16,$s2
+	movzb	($sbox,$acc2,1),$acc2	#$t2
+	movzb	($sbox,$acc0,1),$acc0	#$t3
+	shr	\$16,$s3
+
+	movzb	`&lo("$s2")`,$acc1
+	shl	\$8,$t4
+	shl	\$8,$t5
+	movzb	($sbox,$acc1,1),$acc1	#$t0
+	xor	$t4,$t0
+	xor	$t5,$t1
+
+	movzb	`&lo("$s3")`,$t4
+	shr	\$16,$s0
+	shr	\$16,$s1
+	movzb	`&lo("$s0")`,$t5
+	shl	\$8,$acc2
+	shl	\$8,$acc0
+	movzb	($sbox,$t4,1),$t4	#$t1
+	movzb	($sbox,$t5,1),$t5	#$t2
+	xor	$acc2,$t2
+	xor	$acc0,$t3
+
+	movzb	`&lo("$s1")`,$acc2
+	movzb	`&hi("$s1")`,$acc0
+	shl	\$16,$acc1
+	movzb	($sbox,$acc2,1),$acc2	#$t3
+	movzb	($sbox,$acc0,1),$acc0	#$t0
+	xor	$acc1,$t0
+
+	movzb	`&hi("$s2")`,$acc1
+	shl	\$16,$t4
+	shl	\$16,$t5
+	movzb	($sbox,$acc1,1),$s1	#$t1
+	xor	$t4,$t1
+	xor	$t5,$t2
+
+	movzb	`&hi("$s3")`,$acc1
+	shr	\$8,$s0
+	shl	\$16,$acc2
+	movzb	($sbox,$acc1,1),$s2	#$t2
+	movzb	($sbox,$s0,1),$s3	#$t3
+	xor	$acc2,$t3
+
+	shl	\$24,$acc0
+	shl	\$24,$s1
+	shl	\$24,$s2
+	xor	$acc0,$t0
+	shl	\$24,$s3
+	xor	$t1,$s1
+	mov	$t0,$s0
+	xor	$t2,$s2
+	xor	$t3,$s3
+___
+}
+
+# parallelized version! input is pair of 64-bit values: %rax=s1.s0
+# and %rcx=s3.s2, output is four 32-bit values in %eax=s0, %ebx=s1,
+# %ecx=s2 and %edx=s3.
+sub dectransform()
+{ my ($tp10,$tp20,$tp40,$tp80,$acc0)=("%rax","%r8", "%r9", "%r10","%rbx");
+  my ($tp18,$tp28,$tp48,$tp88,$acc8)=("%rcx","%r11","%r12","%r13","%rdx");
+  my $prefetch = shift;
+
+$code.=<<___;
+	mov	$tp10,$acc0
+	mov	$tp18,$acc8
+	and	$mask80,$acc0
+	and	$mask80,$acc8
+	mov	$acc0,$tp40
+	mov	$acc8,$tp48
+	shr	\$7,$tp40
+	lea	($tp10,$tp10),$tp20
+	shr	\$7,$tp48
+	lea	($tp18,$tp18),$tp28
+	sub	$tp40,$acc0
+	sub	$tp48,$acc8
+	and	$maskfe,$tp20
+	and	$maskfe,$tp28
+	and	$mask1b,$acc0
+	and	$mask1b,$acc8
+	xor	$tp20,$acc0
+	xor	$tp28,$acc8
+	mov	$acc0,$tp20
+	mov	$acc8,$tp28
+
+	and	$mask80,$acc0
+	and	$mask80,$acc8
+	mov	$acc0,$tp80
+	mov	$acc8,$tp88
+	shr	\$7,$tp80
+	lea	($tp20,$tp20),$tp40
+	shr	\$7,$tp88
+	lea	($tp28,$tp28),$tp48
+	sub	$tp80,$acc0
+	sub	$tp88,$acc8
+	and	$maskfe,$tp40
+	and	$maskfe,$tp48
+	and	$mask1b,$acc0
+	and	$mask1b,$acc8
+	xor	$tp40,$acc0
+	xor	$tp48,$acc8
+	mov	$acc0,$tp40
+	mov	$acc8,$tp48
+
+	and	$mask80,$acc0
+	and	$mask80,$acc8
+	mov	$acc0,$tp80
+	mov	$acc8,$tp88
+	shr	\$7,$tp80
+	 xor	$tp10,$tp20		# tp2^=tp1
+	shr	\$7,$tp88
+	 xor	$tp18,$tp28		# tp2^=tp1
+	sub	$tp80,$acc0
+	sub	$tp88,$acc8
+	lea	($tp40,$tp40),$tp80
+	lea	($tp48,$tp48),$tp88
+	 xor	$tp10,$tp40		# tp4^=tp1
+	 xor	$tp18,$tp48		# tp4^=tp1
+	and	$maskfe,$tp80
+	and	$maskfe,$tp88
+	and	$mask1b,$acc0
+	and	$mask1b,$acc8
+	xor	$acc0,$tp80
+	xor	$acc8,$tp88
+
+	xor	$tp80,$tp10		# tp1^=tp8
+	xor	$tp88,$tp18		# tp1^=tp8
+	xor	$tp80,$tp20		# tp2^tp1^=tp8
+	xor	$tp88,$tp28		# tp2^tp1^=tp8
+	mov	$tp10,$acc0
+	mov	$tp18,$acc8
+	xor	$tp80,$tp40		# tp4^tp1^=tp8
+	xor	$tp88,$tp48		# tp4^tp1^=tp8
+	shr	\$32,$acc0
+	shr	\$32,$acc8
+	xor	$tp20,$tp80		# tp8^=tp8^tp2^tp1=tp2^tp1
+	xor	$tp28,$tp88		# tp8^=tp8^tp2^tp1=tp2^tp1
+	rol	\$8,`&LO("$tp10")`	# ROTATE(tp1^tp8,8)
+	rol	\$8,`&LO("$tp18")`	# ROTATE(tp1^tp8,8)
+	xor	$tp40,$tp80		# tp2^tp1^=tp8^tp4^tp1=tp8^tp4^tp2
+	xor	$tp48,$tp88		# tp2^tp1^=tp8^tp4^tp1=tp8^tp4^tp2
+
+	rol	\$8,`&LO("$acc0")`	# ROTATE(tp1^tp8,8)
+	rol	\$8,`&LO("$acc8")`	# ROTATE(tp1^tp8,8)
+	xor	`&LO("$tp80")`,`&LO("$tp10")`
+	xor	`&LO("$tp88")`,`&LO("$tp18")`
+	shr	\$32,$tp80
+	shr	\$32,$tp88
+	xor	`&LO("$tp80")`,`&LO("$acc0")`
+	xor	`&LO("$tp88")`,`&LO("$acc8")`
+
+	mov	$tp20,$tp80
+	mov	$tp28,$tp88
+	shr	\$32,$tp80
+	shr	\$32,$tp88
+	rol	\$24,`&LO("$tp20")`	# ROTATE(tp2^tp1^tp8,24)
+	rol	\$24,`&LO("$tp28")`	# ROTATE(tp2^tp1^tp8,24)
+	rol	\$24,`&LO("$tp80")`	# ROTATE(tp2^tp1^tp8,24)
+	rol	\$24,`&LO("$tp88")`	# ROTATE(tp2^tp1^tp8,24)
+	xor	`&LO("$tp20")`,`&LO("$tp10")`
+	xor	`&LO("$tp28")`,`&LO("$tp18")`
+	mov	$tp40,$tp20
+	mov	$tp48,$tp28
+	xor	`&LO("$tp80")`,`&LO("$acc0")`
+	xor	`&LO("$tp88")`,`&LO("$acc8")`
+
+	`"mov	0($sbox),$mask80"	if ($prefetch)`
+	shr	\$32,$tp20
+	shr	\$32,$tp28
+	`"mov	64($sbox),$maskfe"	if ($prefetch)`
+	rol	\$16,`&LO("$tp40")`	# ROTATE(tp4^tp1^tp8,16)
+	rol	\$16,`&LO("$tp48")`	# ROTATE(tp4^tp1^tp8,16)
+	`"mov	128($sbox),$mask1b"	if ($prefetch)`
+	rol	\$16,`&LO("$tp20")`	# ROTATE(tp4^tp1^tp8,16)
+	rol	\$16,`&LO("$tp28")`	# ROTATE(tp4^tp1^tp8,16)
+	`"mov	192($sbox),$tp80"	if ($prefetch)`
+	xor	`&LO("$tp40")`,`&LO("$tp10")`
+	xor	`&LO("$tp48")`,`&LO("$tp18")`
+	`"mov	256($sbox),$tp88"	if ($prefetch)`
+	xor	`&LO("$tp20")`,`&LO("$acc0")`
+	xor	`&LO("$tp28")`,`&LO("$acc8")`
+___
+}
+
+$code.=<<___;
+.type	_x86_64_AES_decrypt_compact,\@abi-omnipotent
+.align	16
+_x86_64_AES_decrypt_compact:
+	lea	128($sbox),$inp			# size optimization
+	mov	0-128($inp),$acc1		# prefetch Td4
+	mov	32-128($inp),$acc2
+	mov	64-128($inp),$t0
+	mov	96-128($inp),$t1
+	mov	128-128($inp),$acc1
+	mov	160-128($inp),$acc2
+	mov	192-128($inp),$t0
+	mov	224-128($inp),$t1
+	jmp	.Ldec_loop_compact
+
+.align	16
+.Ldec_loop_compact:
+		xor	0($key),$s0		# xor with key
+		xor	4($key),$s1
+		xor	8($key),$s2
+		xor	12($key),$s3
+		lea	16($key),$key
+___
+		&deccompactvert();
+$code.=<<___;
+		cmp	16(%rsp),$key
+		je	.Ldec_compact_done
+
+		mov	256+0($sbox),$mask80
+		shl	\$32,%rbx
+		shl	\$32,%rdx
+		mov	256+8($sbox),$maskfe
+		or	%rbx,%rax
+		or	%rdx,%rcx
+		mov	256+16($sbox),$mask1b
+___
+		&dectransform(1);
+$code.=<<___;
+	jmp	.Ldec_loop_compact
+.align	16
+.Ldec_compact_done:
+	xor	0($key),$s0
+	xor	4($key),$s1
+	xor	8($key),$s2
+	xor	12($key),$s3
+	.byte	0xf3,0xc3			# rep ret
+.size	_x86_64_AES_decrypt_compact,.-_x86_64_AES_decrypt_compact
+___
+
 # void AES_decrypt (const void *inp,void *out,const AES_KEY *key);
 $code.=<<___;
 .globl	AES_decrypt
@@ -672,43 +1192,59 @@ AES_decrypt:
 	push	%r14
 	push	%r15
 
-	mov	%rdx,$key
-	mov	%rdi,$inp
-	mov	%rsi,$out
+	# allocate frame "above" key schedule
+	mov	%rsp,%r10
+	lea	-63(%rdx),%rcx	# %rdx is key argument
+	and	\$-64,%rsp
+	sub	%rsp,%rcx
+	neg	%rcx
+	and	\$0x3c0,%rcx
+	sub	%rcx,%rsp
+	sub	\$32,%rsp
+
+	mov	%rsi,16(%rsp)	# save out
+	mov	%r10,24(%rsp)	# save real stack pointer
+.Ldec_prologue:
 
-	.picmeup	$sbox
-	lea	AES_Td-.($sbox),$sbox
-
-	# prefetch Td4
-	lea	2048+128($sbox),$sbox;
-	mov	0-128($sbox),$s0
-	mov	32-128($sbox),$s1
-	mov	64-128($sbox),$s2
-	mov	96-128($sbox),$s3
-	mov	128-128($sbox),$s0
-	mov	160-128($sbox),$s1
-	mov	192-128($sbox),$s2
-	mov	224-128($sbox),$s3
-	lea	-2048-128($sbox),$sbox;
-
-	mov	0($inp),$s0
-	mov	4($inp),$s1
-	mov	8($inp),$s2
-	mov	12($inp),$s3
-
-	call	_x86_64_AES_decrypt
-
-	mov	$s0,0($out)
+	mov	%rdx,$key
+	mov	240($key),$rnds	# load rounds
+
+	mov	0(%rdi),$s0	# load input vector
+	mov	4(%rdi),$s1
+	mov	8(%rdi),$s2
+	mov	12(%rdi),$s3
+
+	shl	\$4,$rnds
+	lea	($key,$rnds),%rbp
+	mov	$key,(%rsp)	# key schedule
+	mov	%rbp,8(%rsp)	# end of key schedule
+
+	# pick Td4 copy which can't "overlap" with stack frame or key schedule
+	lea	.LAES_Td+2048(%rip),$sbox
+	lea	768(%rsp),%rbp
+	sub	$sbox,%rbp
+	and	\$0x300,%rbp
+	lea	($sbox,%rbp),$sbox
+	shr	\$3,%rbp	# recall "magic" constants!
+	add	%rbp,$sbox
+
+	call	_x86_64_AES_decrypt_compact
+
+	mov	16(%rsp),$out	# restore out
+	mov	24(%rsp),%rsi	# restore saved stack pointer
+	mov	$s0,0($out)	# write output vector
 	mov	$s1,4($out)
 	mov	$s2,8($out)
 	mov	$s3,12($out)
 
-	pop	%r15
-	pop	%r14
-	pop	%r13
-	pop	%r12
-	pop	%rbp
-	pop	%rbx
+	mov	(%rsi),%r15
+	mov	8(%rsi),%r14
+	mov	16(%rsi),%r13
+	mov	24(%rsi),%r12
+	mov	32(%rsi),%rbp
+	mov	40(%rsi),%rbx
+	lea	48(%rsi),%rsp
+.Ldec_epilogue:
 	ret
 .size	AES_decrypt,.-AES_decrypt
 ___
@@ -718,27 +1254,26 @@ sub enckey()
 {
 $code.=<<___;
 	movz	%dl,%esi		# rk[i]>>0
-	mov	2(%rbp,%rsi,8),%ebx
+	movzb	-128(%rbp,%rsi),%ebx
 	movz	%dh,%esi		# rk[i]>>8
-	and	\$0xFF000000,%ebx
+	shl	\$24,%ebx
 	xor	%ebx,%eax
 
-	mov	2(%rbp,%rsi,8),%ebx
+	movzb	-128(%rbp,%rsi),%ebx
 	shr	\$16,%edx
-	and	\$0x000000FF,%ebx
 	movz	%dl,%esi		# rk[i]>>16
 	xor	%ebx,%eax
 
-	mov	0(%rbp,%rsi,8),%ebx
+	movzb	-128(%rbp,%rsi),%ebx
 	movz	%dh,%esi		# rk[i]>>24
-	and	\$0x0000FF00,%ebx
+	shl	\$8,%ebx
 	xor	%ebx,%eax
 
-	mov	0(%rbp,%rsi,8),%ebx
-	and	\$0x00FF0000,%ebx
+	movzb	-128(%rbp,%rsi),%ebx
+	shl	\$16,%ebx
 	xor	%ebx,%eax
 
-	xor	2048(%rbp,%rcx,4),%eax		# rcon
+	xor	1024-128(%rbp,%rcx,4),%eax		# rcon
 ___
 }
 
@@ -751,7 +1286,29 @@ $code.=<<___;
 AES_set_encrypt_key:
 	push	%rbx
 	push	%rbp
+	push	%r12			# redundant, but allows to share 
+	push	%r13			# exception handler...
+	push	%r14
+	push	%r15
+	sub	\$8,%rsp
+.Lenc_key_prologue:
+
+	call	_x86_64_AES_set_encrypt_key
+
+	mov	8(%rsp),%r15
+	mov	16(%rsp),%r14
+	mov	24(%rsp),%r13
+	mov	32(%rsp),%r12
+	mov	40(%rsp),%rbp
+	mov	48(%rsp),%rbx
+	add	\$56,%rsp
+.Lenc_key_epilogue:
+	ret
+.size	AES_set_encrypt_key,.-AES_set_encrypt_key
 
+.type	_x86_64_AES_set_encrypt_key,\@abi-omnipotent
+.align	16
+_x86_64_AES_set_encrypt_key:
 	mov	%esi,%ecx			# %ecx=bits
 	mov	%rdi,%rsi			# %rsi=userKey
 	mov	%rdx,%rdi			# %rdi=key
@@ -761,8 +1318,18 @@ AES_set_encrypt_key:
 	test	\$-1,%rdi
 	jz	.Lbadpointer
 
-	.picmeup %rbp
-	lea	AES_Te-.(%rbp),%rbp
+	lea	.LAES_Te(%rip),%rbp
+	lea	2048+128(%rbp),%rbp
+
+	# prefetch Te4
+	mov	0-128(%rbp),%eax
+	mov	32-128(%rbp),%ebx
+	mov	64-128(%rbp),%r8d
+	mov	96-128(%rbp),%edx
+	mov	128-128(%rbp),%eax
+	mov	160-128(%rbp),%ebx
+	mov	192-128(%rbp),%r8d
+	mov	224-128(%rbp),%edx
 
 	cmp	\$128,%ecx
 	je	.L10rounds
@@ -774,15 +1341,12 @@ AES_set_encrypt_key:
 	jmp	.Lexit
 
 .L10rounds:
-	mov	0(%rsi),%eax			# copy first 4 dwords
-	mov	4(%rsi),%ebx
-	mov	8(%rsi),%ecx
-	mov	12(%rsi),%edx
-	mov	%eax,0(%rdi)
-	mov	%ebx,4(%rdi)
-	mov	%ecx,8(%rdi)
-	mov	%edx,12(%rdi)
+	mov	0(%rsi),%rax			# copy first 4 dwords
+	mov	8(%rsi),%rdx
+	mov	%rax,0(%rdi)
+	mov	%rdx,8(%rdi)
 
+	shr	\$32,%rdx
 	xor	%ecx,%ecx
 	jmp	.L10shortcut
 .align	4
@@ -810,19 +1374,14 @@ $code.=<<___;
 	jmp	.Lexit
 
 .L12rounds:
-	mov	0(%rsi),%eax			# copy first 6 dwords
-	mov	4(%rsi),%ebx
-	mov	8(%rsi),%ecx
-	mov	12(%rsi),%edx
-	mov	%eax,0(%rdi)
-	mov	%ebx,4(%rdi)
-	mov	%ecx,8(%rdi)
-	mov	%edx,12(%rdi)
-	mov	16(%rsi),%ecx
-	mov	20(%rsi),%edx
-	mov	%ecx,16(%rdi)
-	mov	%edx,20(%rdi)
-
+	mov	0(%rsi),%rax			# copy first 6 dwords
+	mov	8(%rsi),%rbx
+	mov	16(%rsi),%rdx
+	mov	%rax,0(%rdi)
+	mov	%rbx,8(%rdi)
+	mov	%rdx,16(%rdi)
+
+	shr	\$32,%rdx
 	xor	%ecx,%ecx
 	jmp	.L12shortcut
 .align	4
@@ -858,30 +1417,23 @@ $code.=<<___;
 	jmp	.Lexit
 
 .L14rounds:		
-	mov	0(%rsi),%eax			# copy first 8 dwords
-	mov	4(%rsi),%ebx
-	mov	8(%rsi),%ecx
-	mov	12(%rsi),%edx
-	mov	%eax,0(%rdi)
-	mov	%ebx,4(%rdi)
-	mov	%ecx,8(%rdi)
-	mov	%edx,12(%rdi)
-	mov	16(%rsi),%eax
-	mov	20(%rsi),%ebx
-	mov	24(%rsi),%ecx
-	mov	28(%rsi),%edx
-	mov	%eax,16(%rdi)
-	mov	%ebx,20(%rdi)
-	mov	%ecx,24(%rdi)
-	mov	%edx,28(%rdi)
-
+	mov	0(%rsi),%rax			# copy first 8 dwords
+	mov	8(%rsi),%rbx
+	mov	16(%rsi),%rcx
+	mov	24(%rsi),%rdx
+	mov	%rax,0(%rdi)
+	mov	%rbx,8(%rdi)
+	mov	%rcx,16(%rdi)
+	mov	%rdx,24(%rdi)
+
+	shr	\$32,%rdx
 	xor	%ecx,%ecx
 	jmp	.L14shortcut
 .align	4
 .L14loop:
+		mov	0(%rdi),%eax			# rk[0]
 		mov	28(%rdi),%edx			# rk[4]
 .L14shortcut:
-		mov	0(%rdi),%eax			# rk[0]
 ___
 		&enckey	();
 $code.=<<___;
@@ -900,24 +1452,23 @@ $code.=<<___;
 		mov	%eax,%edx
 		mov	16(%rdi),%eax			# rk[4]
 		movz	%dl,%esi			# rk[11]>>0
-		mov	2(%rbp,%rsi,8),%ebx
+		movzb	-128(%rbp,%rsi),%ebx
 		movz	%dh,%esi			# rk[11]>>8
-		and	\$0x000000FF,%ebx
 		xor	%ebx,%eax
 
-		mov	0(%rbp,%rsi,8),%ebx
+		movzb	-128(%rbp,%rsi),%ebx
 		shr	\$16,%edx
-		and	\$0x0000FF00,%ebx
+		shl	\$8,%ebx
 		movz	%dl,%esi			# rk[11]>>16
 		xor	%ebx,%eax
 
-		mov	0(%rbp,%rsi,8),%ebx
+		movzb	-128(%rbp,%rsi),%ebx
 		movz	%dh,%esi			# rk[11]>>24
-		and	\$0x00FF0000,%ebx
+		shl	\$16,%ebx
 		xor	%ebx,%eax
 
-		mov	2(%rbp,%rsi,8),%ebx
-		and	\$0xFF000000,%ebx
+		movzb	-128(%rbp,%rsi),%ebx
+		shl	\$24,%ebx
 		xor	%ebx,%eax
 
 		mov	%eax,48(%rdi)			# rk[12]
@@ -938,31 +1489,61 @@ $code.=<<___;
 .Lbadpointer:
 	mov	\$-1,%rax
 .Lexit:
-	pop	%rbp
-	pop	%rbx
-	ret
-.size	AES_set_encrypt_key,.-AES_set_encrypt_key
+	.byte	0xf3,0xc3			# rep ret
+.size	_x86_64_AES_set_encrypt_key,.-_x86_64_AES_set_encrypt_key
 ___
 
-sub deckey()
+sub deckey_ref()
 { my ($i,$ptr,$te,$td) = @_;
+  my ($tp1,$tp2,$tp4,$tp8,$acc)=("%eax","%ebx","%edi","%edx","%r8d");
 $code.=<<___;
-	mov	$i($ptr),%eax
-	mov	%eax,%edx
-	movz	%ah,%ebx
-	shr	\$16,%edx
-	and	\$0xFF,%eax
-	movzb	2($te,%rax,8),%rax
-	movzb	2($te,%rbx,8),%rbx
-	mov	0($td,%rax,8),%eax
-	xor	3($td,%rbx,8),%eax
-	movzb	%dh,%ebx
-	and	\$0xFF,%edx
-	movzb	2($te,%rdx,8),%rdx
-	movzb	2($te,%rbx,8),%rbx
-	xor	2($td,%rdx,8),%eax
-	xor	1($td,%rbx,8),%eax
-	mov	%eax,$i($ptr)
+	mov	$i($ptr),$tp1
+	mov	$tp1,$acc
+	and	\$0x80808080,$acc
+	mov	$acc,$tp4
+	shr	\$7,$tp4
+	lea	0($tp1,$tp1),$tp2
+	sub	$tp4,$acc
+	and	\$0xfefefefe,$tp2
+	and	\$0x1b1b1b1b,$acc
+	xor	$tp2,$acc
+	mov	$acc,$tp2
+
+	and	\$0x80808080,$acc
+	mov	$acc,$tp8
+	shr	\$7,$tp8
+	lea	0($tp2,$tp2),$tp4
+	sub	$tp8,$acc
+	and	\$0xfefefefe,$tp4
+	and	\$0x1b1b1b1b,$acc
+	 xor	$tp1,$tp2		# tp2^tp1
+	xor	$tp4,$acc
+	mov	$acc,$tp4
+
+	and	\$0x80808080,$acc
+	mov	$acc,$tp8
+	shr	\$7,$tp8
+	sub	$tp8,$acc
+	lea	0($tp4,$tp4),$tp8
+	 xor	$tp1,$tp4		# tp4^tp1
+	and	\$0xfefefefe,$tp8
+	and	\$0x1b1b1b1b,$acc
+	xor	$acc,$tp8
+
+	xor	$tp8,$tp1		# tp1^tp8
+	rol	\$8,$tp1		# ROTATE(tp1^tp8,8)
+	xor	$tp8,$tp2		# tp2^tp1^tp8
+	xor	$tp8,$tp4		# tp4^tp1^tp8
+	xor	$tp2,$tp8
+	xor	$tp4,$tp8		# tp8^(tp8^tp4^tp1)^(tp8^tp2^tp1)=tp8^tp4^tp2
+
+	xor	$tp8,$tp1
+	rol	\$24,$tp2		# ROTATE(tp2^tp1^tp8,24)
+	xor	$tp2,$tp1
+	rol	\$16,$tp4		# ROTATE(tp4^tp1^tp8,16)
+	xor	$tp4,$tp1
+
+	mov	$tp1,$i($ptr)
 ___
 }
 
@@ -973,19 +1554,23 @@ $code.=<<___;
 .type	AES_set_decrypt_key,\@function,3
 .align	16
 AES_set_decrypt_key:
-	push	%rdx
-	call	AES_set_encrypt_key
-	cmp	\$0,%eax
-	je	.Lproceed
-	lea	24(%rsp),%rsp
-	ret
-.Lproceed:
+	push	%rbx
+	push	%rbp
+	push	%r12
+	push	%r13
+	push	%r14
+	push	%r15
+	push	%rdx			# save key schedule
+.Ldec_key_prologue:
+
+	call	_x86_64_AES_set_encrypt_key
 	mov	(%rsp),%r8		# restore key schedule
-	mov	%rbx,(%rsp)
+	cmp	\$0,%eax
+	jne	.Labort
 
-	mov	240(%r8),%ecx		# pull number of rounds
+	mov	240(%r8),%r14d		# pull number of rounds
 	xor	%rdi,%rdi
-	lea	(%rdi,%rcx,4),%rcx
+	lea	(%rdi,%r14d,4),%rcx
 	mov	%r8,%rsi
 	lea	(%r8,%rcx,4),%rdi	# pointer to last chunk
 .align	4
@@ -1003,27 +1588,39 @@ AES_set_decrypt_key:
 		cmp	%rsi,%rdi
 	jne	.Linvert
 
-	.picmeup %r9
-	lea	AES_Td-.(%r9),%rdi
-	lea	AES_Te-AES_Td(%rdi),%r9
+	lea	.LAES_Te+2048+1024(%rip),%rax	# rcon
 
-	mov	%r8,%rsi
-	mov	240(%r8),%ecx		# pull number of rounds
-	sub	\$1,%ecx
+	mov	40(%rax),$mask80
+	mov	48(%rax),$maskfe
+	mov	56(%rax),$mask1b
+
+	mov	%r8,$key
+	sub	\$1,%r14d
 .align	4
 .Lpermute:
-		lea	16(%rsi),%rsi
+		lea	16($key),$key
+		mov	0($key),%rax
+		mov	8($key),%rcx
 ___
-		&deckey	(0,"%rsi","%r9","%rdi");
-		&deckey	(4,"%rsi","%r9","%rdi");
-		&deckey	(8,"%rsi","%r9","%rdi");
-		&deckey	(12,"%rsi","%r9","%rdi");
+		&dectransform ();
 $code.=<<___;
-		sub	\$1,%ecx
+		mov	%eax,0($key)
+		mov	%ebx,4($key)
+		mov	%ecx,8($key)
+		mov	%edx,12($key)
+		sub	\$1,%r14d
 	jnz	.Lpermute
 
 	xor	%rax,%rax
-	pop	%rbx
+.Labort:
+	mov	8(%rsp),%r15
+	mov	16(%rsp),%r14
+	mov	24(%rsp),%r13
+	mov	32(%rsp),%r12
+	mov	40(%rsp),%rbp
+	mov	48(%rsp),%rbx
+	add	\$56,%rsp
+.Ldec_key_epilogue:
 	ret
 .size	AES_set_decrypt_key,.-AES_set_decrypt_key
 ___
@@ -1034,47 +1631,59 @@ ___
 {
 # stack frame layout
 # -8(%rsp)		return address
-my $_rsp="0(%rsp)";		# saved %rsp
-my $_len="8(%rsp)";		# copy of 3rd parameter, length
-my $_key="16(%rsp)";		# copy of 4th parameter, key
-my $_ivp="24(%rsp)";		# copy of 5th parameter, ivp
-my $keyp="32(%rsp)";		# one to pass as $key
-my $ivec="40(%rsp)";		# ivec[16]
-my $aes_key="56(%rsp)";		# copy of aes_key
-my $mark="56+240(%rsp)";	# copy of aes_key->rounds
+my $keyp="0(%rsp)";		# one to pass as $key
+my $keyend="8(%rsp)";		# &(keyp->rd_key[4*keyp->rounds])
+my $_rsp="16(%rsp)";		# saved %rsp
+my $_inp="24(%rsp)";		# copy of 1st parameter, inp
+my $_out="32(%rsp)";		# copy of 2nd parameter, out
+my $_len="40(%rsp)";		# copy of 3rd parameter, length
+my $_key="48(%rsp)";		# copy of 4th parameter, key
+my $_ivp="56(%rsp)";		# copy of 5th parameter, ivp
+my $ivec="64(%rsp)";		# ivec[16]
+my $aes_key="80(%rsp)";		# copy of aes_key
+my $mark="80+240(%rsp)";	# copy of aes_key->rounds
 
 $code.=<<___;
 .globl	AES_cbc_encrypt
 .type	AES_cbc_encrypt,\@function,6
 .align	16
+.extern	OPENSSL_ia32cap_P
 AES_cbc_encrypt:
 	cmp	\$0,%rdx	# check length
-	je	.Lcbc_just_ret
+	je	.Lcbc_epilogue
+	pushfq
 	push	%rbx
 	push	%rbp
 	push	%r12
 	push	%r13
 	push	%r14
 	push	%r15
-	pushfq
+.Lcbc_prologue:
+
 	cld
 	mov	%r9d,%r9d	# clear upper half of enc
 
-	.picmeup $sbox
-.Lcbc_pic_point:
-
+	lea	.LAES_Te(%rip),$sbox
 	cmp	\$0,%r9
-	je	.LDECRYPT
-
-	lea	AES_Te-.Lcbc_pic_point($sbox),$sbox
+	jne	.Lcbc_picked_te
+	lea	.LAES_Td(%rip),$sbox
+.Lcbc_picked_te:
+
+	mov	OPENSSL_ia32cap_P(%rip),%r10d
+	cmp	\$$speed_limit,%rdx
+	jb	.Lcbc_slow_prologue
+	test	\$15,%rdx
+	jnz	.Lcbc_slow_prologue
+	bt	\$28,%r10d
+	jc	.Lcbc_slow_prologue
 
 	# allocate aligned stack frame...
-	lea	-64-248(%rsp),$key
+	lea	-88-248(%rsp),$key
 	and	\$-64,$key
 
-	# ... and make it doesn't alias with AES_Te modulo 4096
+	# ... and make sure it doesn't alias with AES_T[ed] modulo 4096
 	mov	$sbox,%r10
-	lea	2048($sbox),%r11
+	lea	2304($sbox),%r11
 	mov	$key,%r12
 	and	\$0xFFF,%r10	# s = $sbox&0xfff
 	and	\$0xFFF,%r11	# e = ($sbox+2048)&0xfff
@@ -1094,22 +1703,27 @@ AES_cbc_encrypt:
 .Lcbc_te_ok:
 
 	xchg	%rsp,$key
-	add	\$8,%rsp	# reserve for return address!
+	#add	\$8,%rsp	# reserve for return address!
 	mov	$key,$_rsp	# save %rsp
+.Lcbc_fast_body:
+	mov	%rdi,$_inp	# save copy of inp
+	mov	%rsi,$_out	# save copy of out
 	mov	%rdx,$_len	# save copy of len
 	mov	%rcx,$_key	# save copy of key
 	mov	%r8,$_ivp	# save copy of ivp
 	movl	\$0,$mark	# copy of aes_key->rounds = 0;
 	mov	%r8,%rbp	# rearrange input arguments
+	mov	%r9,%rbx
 	mov	%rsi,$out
 	mov	%rdi,$inp
 	mov	%rcx,$key
 
+	mov	240($key),%eax		# key->rounds
 	# do we copy key schedule to stack?
 	mov	$key,%r10
 	sub	$sbox,%r10
 	and	\$0xfff,%r10
-	cmp	\$2048,%r10
+	cmp	\$2304,%r10
 	jb	.Lcbc_do_ecopy
 	cmp	\$4096-248,%r10
 	jb	.Lcbc_skip_ecopy
@@ -1120,12 +1734,11 @@ AES_cbc_encrypt:
 		lea	$aes_key,$key
 		mov	\$240/8,%ecx
 		.long	0x90A548F3	# rep movsq
-		mov	(%rsi),%eax	# copy aes_key->rounds
-		mov	%eax,(%rdi)
+		mov	%eax,(%rdi)	# copy aes_key->rounds
 .Lcbc_skip_ecopy:
 	mov	$key,$keyp	# save key pointer
 
-	mov	\$16,%ecx
+	mov	\$18,%ecx
 .align	4
 .Lcbc_prefetch_te:
 		mov	0($sbox),%r10
@@ -1135,184 +1748,77 @@ AES_cbc_encrypt:
 		lea	128($sbox),$sbox
 		sub	\$1,%ecx
 	jnz	.Lcbc_prefetch_te
-	sub	\$2048,$sbox
+	lea	-2304($sbox),$sbox
 
-	test	\$-16,%rdx		# check upon length
-	mov	%rdx,%r10
+	cmp	\$0,%rbx
+	je	.LFAST_DECRYPT
+
+#----------------------------- ENCRYPT -----------------------------#
 	mov	0(%rbp),$s0		# load iv
 	mov	4(%rbp),$s1
 	mov	8(%rbp),$s2
 	mov	12(%rbp),$s3
-	jz	.Lcbc_enc_tail		# short input...
 
 .align	4
-.Lcbc_enc_loop:
+.Lcbc_fast_enc_loop:
 		xor	0($inp),$s0
 		xor	4($inp),$s1
 		xor	8($inp),$s2
 		xor	12($inp),$s3
-		mov	$inp,$ivec	# if ($verticalspin) save inp
-
 		mov	$keyp,$key	# restore key
+		mov	$inp,$_inp	# if ($verticalspin) save inp
+
 		call	_x86_64_AES_encrypt
 
-		mov	$ivec,$inp	# if ($verticalspin) restore inp
+		mov	$_inp,$inp	# if ($verticalspin) restore inp
+		mov	$_len,%r10
 		mov	$s0,0($out)
 		mov	$s1,4($out)
 		mov	$s2,8($out)
 		mov	$s3,12($out)
 
-		mov	$_len,%r10
 		lea	16($inp),$inp
 		lea	16($out),$out
 		sub	\$16,%r10
 		test	\$-16,%r10
 		mov	%r10,$_len
-	jnz	.Lcbc_enc_loop
-	test	\$15,%r10
-	jnz	.Lcbc_enc_tail
+	jnz	.Lcbc_fast_enc_loop
 	mov	$_ivp,%rbp	# restore ivp
 	mov	$s0,0(%rbp)	# save ivec
 	mov	$s1,4(%rbp)
 	mov	$s2,8(%rbp)
 	mov	$s3,12(%rbp)
 
-.align	4
-.Lcbc_cleanup:
-	cmpl	\$0,$mark	# was the key schedule copied?
-	lea	$aes_key,%rdi
-	mov	$_rsp,%rsp
-	je	.Lcbc_exit
-		mov	\$240/8,%ecx
-		xor	%rax,%rax
-		.long	0x90AB48F3	# rep stosq
-.Lcbc_exit:
-	popfq
-	pop	%r15
-	pop	%r14
-	pop	%r13
-	pop	%r12
-	pop	%rbp
-	pop	%rbx
-.Lcbc_just_ret:
-	ret
-.align	4
-.Lcbc_enc_tail:
-	mov	%rax,%r11
-	mov	%rcx,%r12
-	mov	%r10,%rcx
-	mov	$inp,%rsi
-	mov	$out,%rdi
-	.long	0xF689A4F3		# rep movsb
-	mov	\$16,%rcx		# zero tail
-	sub	%r10,%rcx
-	xor	%rax,%rax
-	.long	0xF689AAF3		# rep stosb
-	mov	$out,$inp		# this is not a mistake!
-	movq	\$16,$_len		# len=16
-	mov	%r11,%rax
-	mov	%r12,%rcx
-	jmp	.Lcbc_enc_loop		# one more spin...
+	jmp	.Lcbc_fast_cleanup
+
 #----------------------------- DECRYPT -----------------------------#
 .align	16
-.LDECRYPT:
-	lea	AES_Td-.Lcbc_pic_point($sbox),$sbox
-
-	# allocate aligned stack frame...
-	lea	-64-248(%rsp),$key
-	and	\$-64,$key
-
-	# ... and make it doesn't alias with AES_Td modulo 4096
-	mov	$sbox,%r10
-	lea	2304($sbox),%r11
-	mov	$key,%r12
-	and	\$0xFFF,%r10	# s = $sbox&0xfff
-	and	\$0xFFF,%r11	# e = ($sbox+2048+256)&0xfff
-	and	\$0xFFF,%r12	# p = %rsp&0xfff
-
-	cmp	%r11,%r12	# if (p=>e) %rsp =- (p-e);
-	jb	.Lcbc_td_break_out
-	sub	%r11,%r12
-	sub	%r12,$key
-	jmp	.Lcbc_td_ok
-.Lcbc_td_break_out:		# else %rsp -= (p-s)&0xfff + framesz
-	sub	%r10,%r12
-	and	\$0xFFF,%r12
-	add	\$320,%r12
-	sub	%r12,$key
-.align	4
-.Lcbc_td_ok:
-
-	xchg	%rsp,$key
-	add	\$8,%rsp	# reserve for return address!
-	mov	$key,$_rsp	# save %rsp
-	mov	%rdx,$_len	# save copy of len
-	mov	%rcx,$_key	# save copy of key
-	mov	%r8,$_ivp	# save copy of ivp
-	movl	\$0,$mark	# copy of aes_key->rounds = 0;
-	mov	%r8,%rbp	# rearrange input arguments
-	mov	%rsi,$out
-	mov	%rdi,$inp
-	mov	%rcx,$key
-
-	# do we copy key schedule to stack?
-	mov	$key,%r10
-	sub	$sbox,%r10
-	and	\$0xfff,%r10
-	cmp	\$2304,%r10
-	jb	.Lcbc_do_dcopy
-	cmp	\$4096-248,%r10
-	jb	.Lcbc_skip_dcopy
-.align	4
-.Lcbc_do_dcopy:
-		mov	$key,%rsi
-		lea	$aes_key,%rdi
-		lea	$aes_key,$key
-		mov	\$240/8,%ecx
-		.long	0x90A548F3	# rep movsq
-		mov	(%rsi),%eax	# copy aes_key->rounds
-		mov	%eax,(%rdi)
-.Lcbc_skip_dcopy:
-	mov	$key,$keyp	# save key pointer
-
-	mov	\$18,%ecx
-.align	4
-.Lcbc_prefetch_td:
-		mov	0($sbox),%r10
-		mov	32($sbox),%r11
-		mov	64($sbox),%r12
-		mov	96($sbox),%r13
-		lea	128($sbox),$sbox
-		sub	\$1,%ecx
-	jnz	.Lcbc_prefetch_td
-	sub	\$2304,$sbox
-
+.LFAST_DECRYPT:
 	cmp	$inp,$out
-	je	.Lcbc_dec_in_place
+	je	.Lcbc_fast_dec_in_place
 
 	mov	%rbp,$ivec
 .align	4
-.Lcbc_dec_loop:
-		mov	0($inp),$s0		# read input
+.Lcbc_fast_dec_loop:
+		mov	0($inp),$s0	# read input
 		mov	4($inp),$s1
 		mov	8($inp),$s2
 		mov	12($inp),$s3
-		mov	$inp,8+$ivec	# if ($verticalspin) save inp
-
 		mov	$keyp,$key	# restore key
+		mov	$inp,$_inp	# if ($verticalspin) save inp
+
 		call	_x86_64_AES_decrypt
 
 		mov	$ivec,%rbp	# load ivp
-		mov	8+$ivec,$inp	# if ($verticalspin) restore inp
+		mov	$_inp,$inp	# if ($verticalspin) restore inp
+		mov	$_len,%r10	# load len
 		xor	0(%rbp),$s0	# xor iv
 		xor	4(%rbp),$s1
 		xor	8(%rbp),$s2
 		xor	12(%rbp),$s3
 		mov	$inp,%rbp	# current input, next iv
 
-		mov	$_len,%r10	# load len
 		sub	\$16,%r10
-		jc	.Lcbc_dec_partial
 		mov	%r10,$_len	# update len
 		mov	%rbp,$ivec	# update ivp
 
@@ -1323,81 +1829,281 @@ AES_cbc_encrypt:
 
 		lea	16($inp),$inp
 		lea	16($out),$out
-	jnz	.Lcbc_dec_loop
-.Lcbc_dec_end:
+	jnz	.Lcbc_fast_dec_loop
 	mov	$_ivp,%r12		# load user ivp
 	mov	0(%rbp),%r10		# load iv
 	mov	8(%rbp),%r11
 	mov	%r10,0(%r12)		# copy back to user
 	mov	%r11,8(%r12)
-	jmp	.Lcbc_cleanup
-
-.align	4
-.Lcbc_dec_partial:
-	mov	$s0,0+$ivec		# dump output to stack
-	mov	$s1,4+$ivec
-	mov	$s2,8+$ivec
-	mov	$s3,12+$ivec
-	mov	$out,%rdi
-	lea	$ivec,%rsi
-	mov	\$16,%rcx
-	add	%r10,%rcx		# number of bytes to copy
-	.long	0xF689A4F3		# rep movsb
-	jmp	.Lcbc_dec_end
+	jmp	.Lcbc_fast_cleanup
 
 .align	16
-.Lcbc_dec_in_place:
+.Lcbc_fast_dec_in_place:
+	mov	0(%rbp),%r10		# copy iv to stack
+	mov	8(%rbp),%r11
+	mov	%r10,0+$ivec
+	mov	%r11,8+$ivec
+.align	4
+.Lcbc_fast_dec_in_place_loop:
 		mov	0($inp),$s0	# load input
 		mov	4($inp),$s1
 		mov	8($inp),$s2
 		mov	12($inp),$s3
+		mov	$keyp,$key	# restore key
+		mov	$inp,$_inp	# if ($verticalspin) save inp
 
-		mov	$inp,$ivec	# if ($verticalspin) save inp
-		mov	$keyp,$key
 		call	_x86_64_AES_decrypt
 
-		mov	$ivec,$inp	# if ($verticalspin) restore inp
-		mov	$_ivp,%rbp
-		xor	0(%rbp),$s0
-		xor	4(%rbp),$s1
-		xor	8(%rbp),$s2
-		xor	12(%rbp),$s3
+		mov	$_inp,$inp	# if ($verticalspin) restore inp
+		mov	$_len,%r10
+		xor	0+$ivec,$s0
+		xor	4+$ivec,$s1
+		xor	8+$ivec,$s2
+		xor	12+$ivec,$s3
+
+		mov	0($inp),%r11	# load input
+		mov	8($inp),%r12
+		sub	\$16,%r10
+		jz	.Lcbc_fast_dec_in_place_done
 
-		mov	0($inp),%r10	# copy input to iv
-		mov	8($inp),%r11
-		mov	%r10,0(%rbp)
-		mov	%r11,8(%rbp)
+		mov	%r11,0+$ivec	# copy input to iv
+		mov	%r12,8+$ivec
 
 		mov	$s0,0($out)	# save output [zaps input]
 		mov	$s1,4($out)
 		mov	$s2,8($out)
 		mov	$s3,12($out)
 
-		mov	$_len,%rcx
 		lea	16($inp),$inp
 		lea	16($out),$out
-		sub	\$16,%rcx
-		jc	.Lcbc_dec_in_place_partial
-		mov	%rcx,$_len
-	jnz	.Lcbc_dec_in_place
-	jmp	.Lcbc_cleanup
+		mov	%r10,$_len
+	jmp	.Lcbc_fast_dec_in_place_loop
+.Lcbc_fast_dec_in_place_done:
+	mov	$_ivp,%rdi
+	mov	%r11,0(%rdi)	# copy iv back to user
+	mov	%r12,8(%rdi)
+
+	mov	$s0,0($out)	# save output [zaps input]
+	mov	$s1,4($out)
+	mov	$s2,8($out)
+	mov	$s3,12($out)
 
 .align	4
-.Lcbc_dec_in_place_partial:
-	# one can argue if this is actually required
-	lea	($out,%rcx),%rdi
-	lea	(%rbp,%rcx),%rsi
-	neg	%rcx
-	.long	0xF689A4F3	# rep movsb	# restore tail
-	jmp	.Lcbc_cleanup
+.Lcbc_fast_cleanup:
+	cmpl	\$0,$mark	# was the key schedule copied?
+	lea	$aes_key,%rdi
+	je	.Lcbc_exit
+		mov	\$240/8,%ecx
+		xor	%rax,%rax
+		.long	0x90AB48F3	# rep stosq
+
+	jmp	.Lcbc_exit
+
+#--------------------------- SLOW ROUTINE ---------------------------#
+.align	16
+.Lcbc_slow_prologue:
+	# allocate aligned stack frame...
+	lea	-88(%rsp),%rbp
+	and	\$-64,%rbp
+	# ... just "above" key schedule
+	lea	-88-63(%rcx),%r10
+	sub	%rbp,%r10
+	neg	%r10
+	and	\$0x3c0,%r10
+	sub	%r10,%rbp
+
+	xchg	%rsp,%rbp
+	#add	\$8,%rsp	# reserve for return address!
+	mov	%rbp,$_rsp	# save %rsp
+.Lcbc_slow_body:
+	#mov	%rdi,$_inp	# save copy of inp
+	#mov	%rsi,$_out	# save copy of out
+	#mov	%rdx,$_len	# save copy of len
+	#mov	%rcx,$_key	# save copy of key
+	mov	%r8,$_ivp	# save copy of ivp
+	mov	%r8,%rbp	# rearrange input arguments
+	mov	%r9,%rbx
+	mov	%rsi,$out
+	mov	%rdi,$inp
+	mov	%rcx,$key
+	mov	%rdx,%r10
+
+	mov	240($key),%eax
+	mov	$key,$keyp	# save key pointer
+	shl	\$4,%eax
+	lea	($key,%rax),%rax
+	mov	%rax,$keyend
+
+	# pick Te4 copy which can't "overlap" with stack frame or key scdedule
+	lea	2048($sbox),$sbox
+	lea	768-8(%rsp),%rax
+	sub	$sbox,%rax
+	and	\$0x300,%rax
+	lea	($sbox,%rax),$sbox
+
+	cmp	\$0,%rbx
+	je	.LSLOW_DECRYPT
+
+#--------------------------- SLOW ENCRYPT ---------------------------#
+	test	\$-16,%r10		# check upon length
+	mov	0(%rbp),$s0		# load iv
+	mov	4(%rbp),$s1
+	mov	8(%rbp),$s2
+	mov	12(%rbp),$s3
+	jz	.Lcbc_slow_enc_tail	# short input...
+
+.align	4
+.Lcbc_slow_enc_loop:
+		xor	0($inp),$s0
+		xor	4($inp),$s1
+		xor	8($inp),$s2
+		xor	12($inp),$s3
+		mov	$keyp,$key	# restore key
+		mov	$inp,$_inp	# save inp
+		mov	$out,$_out	# save out
+		mov	%r10,$_len	# save len
+
+		call	_x86_64_AES_encrypt_compact
+
+		mov	$_inp,$inp	# restore inp
+		mov	$_out,$out	# restore out
+		mov	$_len,%r10	# restore len
+		mov	$s0,0($out)
+		mov	$s1,4($out)
+		mov	$s2,8($out)
+		mov	$s3,12($out)
+
+		lea	16($inp),$inp
+		lea	16($out),$out
+		sub	\$16,%r10
+		test	\$-16,%r10
+	jnz	.Lcbc_slow_enc_loop
+	test	\$15,%r10
+	jnz	.Lcbc_slow_enc_tail
+	mov	$_ivp,%rbp	# restore ivp
+	mov	$s0,0(%rbp)	# save ivec
+	mov	$s1,4(%rbp)
+	mov	$s2,8(%rbp)
+	mov	$s3,12(%rbp)
+
+	jmp	.Lcbc_exit
+
+.align	4
+.Lcbc_slow_enc_tail:
+	mov	%rax,%r11
+	mov	%rcx,%r12
+	mov	%r10,%rcx
+	mov	$inp,%rsi
+	mov	$out,%rdi
+	.long	0x9066A4F3		# rep movsb
+	mov	\$16,%rcx		# zero tail
+	sub	%r10,%rcx
+	xor	%rax,%rax
+	.long	0x9066AAF3		# rep stosb
+	mov	$out,$inp		# this is not a mistake!
+	mov	\$16,%r10		# len=16
+	mov	%r11,%rax
+	mov	%r12,%rcx
+	jmp	.Lcbc_slow_enc_loop	# one more spin...
+#--------------------------- SLOW DECRYPT ---------------------------#
+.align	16
+.LSLOW_DECRYPT:
+	shr	\$3,%rax
+	add	%rax,$sbox		# recall "magic" constants!
+
+	mov	0(%rbp),%r11		# copy iv to stack
+	mov	8(%rbp),%r12
+	mov	%r11,0+$ivec
+	mov	%r12,8+$ivec
+
+.align	4
+.Lcbc_slow_dec_loop:
+		mov	0($inp),$s0	# load input
+		mov	4($inp),$s1
+		mov	8($inp),$s2
+		mov	12($inp),$s3
+		mov	$keyp,$key	# restore key
+		mov	$inp,$_inp	# save inp
+		mov	$out,$_out	# save out
+		mov	%r10,$_len	# save len
+
+		call	_x86_64_AES_decrypt_compact
+
+		mov	$_inp,$inp	# restore inp
+		mov	$_out,$out	# restore out
+		mov	$_len,%r10
+		xor	0+$ivec,$s0
+		xor	4+$ivec,$s1
+		xor	8+$ivec,$s2
+		xor	12+$ivec,$s3
+
+		mov	0($inp),%r11	# load input
+		mov	8($inp),%r12
+		sub	\$16,%r10
+		jc	.Lcbc_slow_dec_partial
+		jz	.Lcbc_slow_dec_done
+
+		mov	%r11,0+$ivec	# copy input to iv
+		mov	%r12,8+$ivec
+
+		mov	$s0,0($out)	# save output [can zap input]
+		mov	$s1,4($out)
+		mov	$s2,8($out)
+		mov	$s3,12($out)
+
+		lea	16($inp),$inp
+		lea	16($out),$out
+	jmp	.Lcbc_slow_dec_loop
+.Lcbc_slow_dec_done:
+	mov	$_ivp,%rdi
+	mov	%r11,0(%rdi)		# copy iv back to user
+	mov	%r12,8(%rdi)
+
+	mov	$s0,0($out)		# save output [can zap input]
+	mov	$s1,4($out)
+	mov	$s2,8($out)
+	mov	$s3,12($out)
+
+	jmp	.Lcbc_exit
+
+.align	4
+.Lcbc_slow_dec_partial:
+	mov	$_ivp,%rdi
+	mov	%r11,0(%rdi)		# copy iv back to user
+	mov	%r12,8(%rdi)
+
+	mov	$s0,0+$ivec		# save output to stack
+	mov	$s1,4+$ivec
+	mov	$s2,8+$ivec
+	mov	$s3,12+$ivec
+
+	mov	$out,%rdi
+	lea	$ivec,%rsi
+	lea	16(%r10),%rcx
+	.long	0x9066A4F3	# rep movsb
+	jmp	.Lcbc_exit
+
+.align	16
+.Lcbc_exit:
+	mov	$_rsp,%rsi
+	mov	(%rsi),%r15
+	mov	8(%rsi),%r14
+	mov	16(%rsi),%r13
+	mov	24(%rsi),%r12
+	mov	32(%rsi),%rbp
+	mov	40(%rsi),%rbx
+	lea	48(%rsi),%rsp
+.Lcbc_popfq:
+	popfq
+.Lcbc_epilogue:
+	ret
 .size	AES_cbc_encrypt,.-AES_cbc_encrypt
 ___
 }
 
 $code.=<<___;
-.globl	AES_Te
 .align	64
-AES_Te:
+.LAES_Te:
 ___
 	&_data_word(0xa56363c6, 0x847c7cf8, 0x997777ee, 0x8d7b7bf6);
 	&_data_word(0x0df2f2ff, 0xbd6b6bd6, 0xb16f6fde, 0x54c5c591);
@@ -1463,16 +2169,149 @@ ___
 	&_data_word(0xdabfbf65, 0x31e6e6d7, 0xc6424284, 0xb86868d0);
 	&_data_word(0xc3414182, 0xb0999929, 0x772d2d5a, 0x110f0f1e);
 	&_data_word(0xcbb0b07b, 0xfc5454a8, 0xd6bbbb6d, 0x3a16162c);
+
+#Te4	# four copies of Te4 to choose from to avoid L1 aliasing
+	&data_byte(0x63, 0x7c, 0x77, 0x7b, 0xf2, 0x6b, 0x6f, 0xc5);
+	&data_byte(0x30, 0x01, 0x67, 0x2b, 0xfe, 0xd7, 0xab, 0x76);
+	&data_byte(0xca, 0x82, 0xc9, 0x7d, 0xfa, 0x59, 0x47, 0xf0);
+	&data_byte(0xad, 0xd4, 0xa2, 0xaf, 0x9c, 0xa4, 0x72, 0xc0);
+	&data_byte(0xb7, 0xfd, 0x93, 0x26, 0x36, 0x3f, 0xf7, 0xcc);
+	&data_byte(0x34, 0xa5, 0xe5, 0xf1, 0x71, 0xd8, 0x31, 0x15);
+	&data_byte(0x04, 0xc7, 0x23, 0xc3, 0x18, 0x96, 0x05, 0x9a);
+	&data_byte(0x07, 0x12, 0x80, 0xe2, 0xeb, 0x27, 0xb2, 0x75);
+	&data_byte(0x09, 0x83, 0x2c, 0x1a, 0x1b, 0x6e, 0x5a, 0xa0);
+	&data_byte(0x52, 0x3b, 0xd6, 0xb3, 0x29, 0xe3, 0x2f, 0x84);
+	&data_byte(0x53, 0xd1, 0x00, 0xed, 0x20, 0xfc, 0xb1, 0x5b);
+	&data_byte(0x6a, 0xcb, 0xbe, 0x39, 0x4a, 0x4c, 0x58, 0xcf);
+	&data_byte(0xd0, 0xef, 0xaa, 0xfb, 0x43, 0x4d, 0x33, 0x85);
+	&data_byte(0x45, 0xf9, 0x02, 0x7f, 0x50, 0x3c, 0x9f, 0xa8);
+	&data_byte(0x51, 0xa3, 0x40, 0x8f, 0x92, 0x9d, 0x38, 0xf5);
+	&data_byte(0xbc, 0xb6, 0xda, 0x21, 0x10, 0xff, 0xf3, 0xd2);
+	&data_byte(0xcd, 0x0c, 0x13, 0xec, 0x5f, 0x97, 0x44, 0x17);
+	&data_byte(0xc4, 0xa7, 0x7e, 0x3d, 0x64, 0x5d, 0x19, 0x73);
+	&data_byte(0x60, 0x81, 0x4f, 0xdc, 0x22, 0x2a, 0x90, 0x88);
+	&data_byte(0x46, 0xee, 0xb8, 0x14, 0xde, 0x5e, 0x0b, 0xdb);
+	&data_byte(0xe0, 0x32, 0x3a, 0x0a, 0x49, 0x06, 0x24, 0x5c);
+	&data_byte(0xc2, 0xd3, 0xac, 0x62, 0x91, 0x95, 0xe4, 0x79);
+	&data_byte(0xe7, 0xc8, 0x37, 0x6d, 0x8d, 0xd5, 0x4e, 0xa9);
+	&data_byte(0x6c, 0x56, 0xf4, 0xea, 0x65, 0x7a, 0xae, 0x08);
+	&data_byte(0xba, 0x78, 0x25, 0x2e, 0x1c, 0xa6, 0xb4, 0xc6);
+	&data_byte(0xe8, 0xdd, 0x74, 0x1f, 0x4b, 0xbd, 0x8b, 0x8a);
+	&data_byte(0x70, 0x3e, 0xb5, 0x66, 0x48, 0x03, 0xf6, 0x0e);
+	&data_byte(0x61, 0x35, 0x57, 0xb9, 0x86, 0xc1, 0x1d, 0x9e);
+	&data_byte(0xe1, 0xf8, 0x98, 0x11, 0x69, 0xd9, 0x8e, 0x94);
+	&data_byte(0x9b, 0x1e, 0x87, 0xe9, 0xce, 0x55, 0x28, 0xdf);
+	&data_byte(0x8c, 0xa1, 0x89, 0x0d, 0xbf, 0xe6, 0x42, 0x68);
+	&data_byte(0x41, 0x99, 0x2d, 0x0f, 0xb0, 0x54, 0xbb, 0x16);
+
+	&data_byte(0x63, 0x7c, 0x77, 0x7b, 0xf2, 0x6b, 0x6f, 0xc5);
+	&data_byte(0x30, 0x01, 0x67, 0x2b, 0xfe, 0xd7, 0xab, 0x76);
+	&data_byte(0xca, 0x82, 0xc9, 0x7d, 0xfa, 0x59, 0x47, 0xf0);
+	&data_byte(0xad, 0xd4, 0xa2, 0xaf, 0x9c, 0xa4, 0x72, 0xc0);
+	&data_byte(0xb7, 0xfd, 0x93, 0x26, 0x36, 0x3f, 0xf7, 0xcc);
+	&data_byte(0x34, 0xa5, 0xe5, 0xf1, 0x71, 0xd8, 0x31, 0x15);
+	&data_byte(0x04, 0xc7, 0x23, 0xc3, 0x18, 0x96, 0x05, 0x9a);
+	&data_byte(0x07, 0x12, 0x80, 0xe2, 0xeb, 0x27, 0xb2, 0x75);
+	&data_byte(0x09, 0x83, 0x2c, 0x1a, 0x1b, 0x6e, 0x5a, 0xa0);
+	&data_byte(0x52, 0x3b, 0xd6, 0xb3, 0x29, 0xe3, 0x2f, 0x84);
+	&data_byte(0x53, 0xd1, 0x00, 0xed, 0x20, 0xfc, 0xb1, 0x5b);
+	&data_byte(0x6a, 0xcb, 0xbe, 0x39, 0x4a, 0x4c, 0x58, 0xcf);
+	&data_byte(0xd0, 0xef, 0xaa, 0xfb, 0x43, 0x4d, 0x33, 0x85);
+	&data_byte(0x45, 0xf9, 0x02, 0x7f, 0x50, 0x3c, 0x9f, 0xa8);
+	&data_byte(0x51, 0xa3, 0x40, 0x8f, 0x92, 0x9d, 0x38, 0xf5);
+	&data_byte(0xbc, 0xb6, 0xda, 0x21, 0x10, 0xff, 0xf3, 0xd2);
+	&data_byte(0xcd, 0x0c, 0x13, 0xec, 0x5f, 0x97, 0x44, 0x17);
+	&data_byte(0xc4, 0xa7, 0x7e, 0x3d, 0x64, 0x5d, 0x19, 0x73);
+	&data_byte(0x60, 0x81, 0x4f, 0xdc, 0x22, 0x2a, 0x90, 0x88);
+	&data_byte(0x46, 0xee, 0xb8, 0x14, 0xde, 0x5e, 0x0b, 0xdb);
+	&data_byte(0xe0, 0x32, 0x3a, 0x0a, 0x49, 0x06, 0x24, 0x5c);
+	&data_byte(0xc2, 0xd3, 0xac, 0x62, 0x91, 0x95, 0xe4, 0x79);
+	&data_byte(0xe7, 0xc8, 0x37, 0x6d, 0x8d, 0xd5, 0x4e, 0xa9);
+	&data_byte(0x6c, 0x56, 0xf4, 0xea, 0x65, 0x7a, 0xae, 0x08);
+	&data_byte(0xba, 0x78, 0x25, 0x2e, 0x1c, 0xa6, 0xb4, 0xc6);
+	&data_byte(0xe8, 0xdd, 0x74, 0x1f, 0x4b, 0xbd, 0x8b, 0x8a);
+	&data_byte(0x70, 0x3e, 0xb5, 0x66, 0x48, 0x03, 0xf6, 0x0e);
+	&data_byte(0x61, 0x35, 0x57, 0xb9, 0x86, 0xc1, 0x1d, 0x9e);
+	&data_byte(0xe1, 0xf8, 0x98, 0x11, 0x69, 0xd9, 0x8e, 0x94);
+	&data_byte(0x9b, 0x1e, 0x87, 0xe9, 0xce, 0x55, 0x28, 0xdf);
+	&data_byte(0x8c, 0xa1, 0x89, 0x0d, 0xbf, 0xe6, 0x42, 0x68);
+	&data_byte(0x41, 0x99, 0x2d, 0x0f, 0xb0, 0x54, 0xbb, 0x16);
+
+	&data_byte(0x63, 0x7c, 0x77, 0x7b, 0xf2, 0x6b, 0x6f, 0xc5);
+	&data_byte(0x30, 0x01, 0x67, 0x2b, 0xfe, 0xd7, 0xab, 0x76);
+	&data_byte(0xca, 0x82, 0xc9, 0x7d, 0xfa, 0x59, 0x47, 0xf0);
+	&data_byte(0xad, 0xd4, 0xa2, 0xaf, 0x9c, 0xa4, 0x72, 0xc0);
+	&data_byte(0xb7, 0xfd, 0x93, 0x26, 0x36, 0x3f, 0xf7, 0xcc);
+	&data_byte(0x34, 0xa5, 0xe5, 0xf1, 0x71, 0xd8, 0x31, 0x15);
+	&data_byte(0x04, 0xc7, 0x23, 0xc3, 0x18, 0x96, 0x05, 0x9a);
+	&data_byte(0x07, 0x12, 0x80, 0xe2, 0xeb, 0x27, 0xb2, 0x75);
+	&data_byte(0x09, 0x83, 0x2c, 0x1a, 0x1b, 0x6e, 0x5a, 0xa0);
+	&data_byte(0x52, 0x3b, 0xd6, 0xb3, 0x29, 0xe3, 0x2f, 0x84);
+	&data_byte(0x53, 0xd1, 0x00, 0xed, 0x20, 0xfc, 0xb1, 0x5b);
+	&data_byte(0x6a, 0xcb, 0xbe, 0x39, 0x4a, 0x4c, 0x58, 0xcf);
+	&data_byte(0xd0, 0xef, 0xaa, 0xfb, 0x43, 0x4d, 0x33, 0x85);
+	&data_byte(0x45, 0xf9, 0x02, 0x7f, 0x50, 0x3c, 0x9f, 0xa8);
+	&data_byte(0x51, 0xa3, 0x40, 0x8f, 0x92, 0x9d, 0x38, 0xf5);
+	&data_byte(0xbc, 0xb6, 0xda, 0x21, 0x10, 0xff, 0xf3, 0xd2);
+	&data_byte(0xcd, 0x0c, 0x13, 0xec, 0x5f, 0x97, 0x44, 0x17);
+	&data_byte(0xc4, 0xa7, 0x7e, 0x3d, 0x64, 0x5d, 0x19, 0x73);
+	&data_byte(0x60, 0x81, 0x4f, 0xdc, 0x22, 0x2a, 0x90, 0x88);
+	&data_byte(0x46, 0xee, 0xb8, 0x14, 0xde, 0x5e, 0x0b, 0xdb);
+	&data_byte(0xe0, 0x32, 0x3a, 0x0a, 0x49, 0x06, 0x24, 0x5c);
+	&data_byte(0xc2, 0xd3, 0xac, 0x62, 0x91, 0x95, 0xe4, 0x79);
+	&data_byte(0xe7, 0xc8, 0x37, 0x6d, 0x8d, 0xd5, 0x4e, 0xa9);
+	&data_byte(0x6c, 0x56, 0xf4, 0xea, 0x65, 0x7a, 0xae, 0x08);
+	&data_byte(0xba, 0x78, 0x25, 0x2e, 0x1c, 0xa6, 0xb4, 0xc6);
+	&data_byte(0xe8, 0xdd, 0x74, 0x1f, 0x4b, 0xbd, 0x8b, 0x8a);
+	&data_byte(0x70, 0x3e, 0xb5, 0x66, 0x48, 0x03, 0xf6, 0x0e);
+	&data_byte(0x61, 0x35, 0x57, 0xb9, 0x86, 0xc1, 0x1d, 0x9e);
+	&data_byte(0xe1, 0xf8, 0x98, 0x11, 0x69, 0xd9, 0x8e, 0x94);
+	&data_byte(0x9b, 0x1e, 0x87, 0xe9, 0xce, 0x55, 0x28, 0xdf);
+	&data_byte(0x8c, 0xa1, 0x89, 0x0d, 0xbf, 0xe6, 0x42, 0x68);
+	&data_byte(0x41, 0x99, 0x2d, 0x0f, 0xb0, 0x54, 0xbb, 0x16);
+
+	&data_byte(0x63, 0x7c, 0x77, 0x7b, 0xf2, 0x6b, 0x6f, 0xc5);
+	&data_byte(0x30, 0x01, 0x67, 0x2b, 0xfe, 0xd7, 0xab, 0x76);
+	&data_byte(0xca, 0x82, 0xc9, 0x7d, 0xfa, 0x59, 0x47, 0xf0);
+	&data_byte(0xad, 0xd4, 0xa2, 0xaf, 0x9c, 0xa4, 0x72, 0xc0);
+	&data_byte(0xb7, 0xfd, 0x93, 0x26, 0x36, 0x3f, 0xf7, 0xcc);
+	&data_byte(0x34, 0xa5, 0xe5, 0xf1, 0x71, 0xd8, 0x31, 0x15);
+	&data_byte(0x04, 0xc7, 0x23, 0xc3, 0x18, 0x96, 0x05, 0x9a);
+	&data_byte(0x07, 0x12, 0x80, 0xe2, 0xeb, 0x27, 0xb2, 0x75);
+	&data_byte(0x09, 0x83, 0x2c, 0x1a, 0x1b, 0x6e, 0x5a, 0xa0);
+	&data_byte(0x52, 0x3b, 0xd6, 0xb3, 0x29, 0xe3, 0x2f, 0x84);
+	&data_byte(0x53, 0xd1, 0x00, 0xed, 0x20, 0xfc, 0xb1, 0x5b);
+	&data_byte(0x6a, 0xcb, 0xbe, 0x39, 0x4a, 0x4c, 0x58, 0xcf);
+	&data_byte(0xd0, 0xef, 0xaa, 0xfb, 0x43, 0x4d, 0x33, 0x85);
+	&data_byte(0x45, 0xf9, 0x02, 0x7f, 0x50, 0x3c, 0x9f, 0xa8);
+	&data_byte(0x51, 0xa3, 0x40, 0x8f, 0x92, 0x9d, 0x38, 0xf5);
+	&data_byte(0xbc, 0xb6, 0xda, 0x21, 0x10, 0xff, 0xf3, 0xd2);
+	&data_byte(0xcd, 0x0c, 0x13, 0xec, 0x5f, 0x97, 0x44, 0x17);
+	&data_byte(0xc4, 0xa7, 0x7e, 0x3d, 0x64, 0x5d, 0x19, 0x73);
+	&data_byte(0x60, 0x81, 0x4f, 0xdc, 0x22, 0x2a, 0x90, 0x88);
+	&data_byte(0x46, 0xee, 0xb8, 0x14, 0xde, 0x5e, 0x0b, 0xdb);
+	&data_byte(0xe0, 0x32, 0x3a, 0x0a, 0x49, 0x06, 0x24, 0x5c);
+	&data_byte(0xc2, 0xd3, 0xac, 0x62, 0x91, 0x95, 0xe4, 0x79);
+	&data_byte(0xe7, 0xc8, 0x37, 0x6d, 0x8d, 0xd5, 0x4e, 0xa9);
+	&data_byte(0x6c, 0x56, 0xf4, 0xea, 0x65, 0x7a, 0xae, 0x08);
+	&data_byte(0xba, 0x78, 0x25, 0x2e, 0x1c, 0xa6, 0xb4, 0xc6);
+	&data_byte(0xe8, 0xdd, 0x74, 0x1f, 0x4b, 0xbd, 0x8b, 0x8a);
+	&data_byte(0x70, 0x3e, 0xb5, 0x66, 0x48, 0x03, 0xf6, 0x0e);
+	&data_byte(0x61, 0x35, 0x57, 0xb9, 0x86, 0xc1, 0x1d, 0x9e);
+	&data_byte(0xe1, 0xf8, 0x98, 0x11, 0x69, 0xd9, 0x8e, 0x94);
+	&data_byte(0x9b, 0x1e, 0x87, 0xe9, 0xce, 0x55, 0x28, 0xdf);
+	&data_byte(0x8c, 0xa1, 0x89, 0x0d, 0xbf, 0xe6, 0x42, 0x68);
+	&data_byte(0x41, 0x99, 0x2d, 0x0f, 0xb0, 0x54, 0xbb, 0x16);
 #rcon:
 $code.=<<___;
 	.long	0x00000001, 0x00000002, 0x00000004, 0x00000008
 	.long	0x00000010, 0x00000020, 0x00000040, 0x00000080
-	.long	0x0000001b, 0x00000036, 0, 0, 0, 0, 0, 0
+	.long	0x0000001b, 0x00000036, 0x80808080, 0x80808080
+	.long	0xfefefefe, 0xfefefefe, 0x1b1b1b1b, 0x1b1b1b1b
 ___
 $code.=<<___;
-.globl	AES_Td
 .align	64
-AES_Td:
+.LAES_Td:
 ___
 	&_data_word(0x50a7f451, 0x5365417e, 0xc3a4171a, 0x965e273a);
 	&_data_word(0xcb6bab3b, 0xf1459d1f, 0xab58faac, 0x9303e34b);
@@ -1538,7 +2377,116 @@ ___
 	&_data_word(0x72c31d16, 0x0c25e2bc, 0x8b493c28, 0x41950dff);
 	&_data_word(0x7101a839, 0xdeb30c08, 0x9ce4b4d8, 0x90c15664);
 	&_data_word(0x6184cb7b, 0x70b632d5, 0x745c6c48, 0x4257b8d0);
-#Td4:
+
+#Td4:	# four copies of Td4 to choose from to avoid L1 aliasing
+	&data_byte(0x52, 0x09, 0x6a, 0xd5, 0x30, 0x36, 0xa5, 0x38);
+	&data_byte(0xbf, 0x40, 0xa3, 0x9e, 0x81, 0xf3, 0xd7, 0xfb);
+	&data_byte(0x7c, 0xe3, 0x39, 0x82, 0x9b, 0x2f, 0xff, 0x87);
+	&data_byte(0x34, 0x8e, 0x43, 0x44, 0xc4, 0xde, 0xe9, 0xcb);
+	&data_byte(0x54, 0x7b, 0x94, 0x32, 0xa6, 0xc2, 0x23, 0x3d);
+	&data_byte(0xee, 0x4c, 0x95, 0x0b, 0x42, 0xfa, 0xc3, 0x4e);
+	&data_byte(0x08, 0x2e, 0xa1, 0x66, 0x28, 0xd9, 0x24, 0xb2);
+	&data_byte(0x76, 0x5b, 0xa2, 0x49, 0x6d, 0x8b, 0xd1, 0x25);
+	&data_byte(0x72, 0xf8, 0xf6, 0x64, 0x86, 0x68, 0x98, 0x16);
+	&data_byte(0xd4, 0xa4, 0x5c, 0xcc, 0x5d, 0x65, 0xb6, 0x92);
+	&data_byte(0x6c, 0x70, 0x48, 0x50, 0xfd, 0xed, 0xb9, 0xda);
+	&data_byte(0x5e, 0x15, 0x46, 0x57, 0xa7, 0x8d, 0x9d, 0x84);
+	&data_byte(0x90, 0xd8, 0xab, 0x00, 0x8c, 0xbc, 0xd3, 0x0a);
+	&data_byte(0xf7, 0xe4, 0x58, 0x05, 0xb8, 0xb3, 0x45, 0x06);
+	&data_byte(0xd0, 0x2c, 0x1e, 0x8f, 0xca, 0x3f, 0x0f, 0x02);
+	&data_byte(0xc1, 0xaf, 0xbd, 0x03, 0x01, 0x13, 0x8a, 0x6b);
+	&data_byte(0x3a, 0x91, 0x11, 0x41, 0x4f, 0x67, 0xdc, 0xea);
+	&data_byte(0x97, 0xf2, 0xcf, 0xce, 0xf0, 0xb4, 0xe6, 0x73);
+	&data_byte(0x96, 0xac, 0x74, 0x22, 0xe7, 0xad, 0x35, 0x85);
+	&data_byte(0xe2, 0xf9, 0x37, 0xe8, 0x1c, 0x75, 0xdf, 0x6e);
+	&data_byte(0x47, 0xf1, 0x1a, 0x71, 0x1d, 0x29, 0xc5, 0x89);
+	&data_byte(0x6f, 0xb7, 0x62, 0x0e, 0xaa, 0x18, 0xbe, 0x1b);
+	&data_byte(0xfc, 0x56, 0x3e, 0x4b, 0xc6, 0xd2, 0x79, 0x20);
+	&data_byte(0x9a, 0xdb, 0xc0, 0xfe, 0x78, 0xcd, 0x5a, 0xf4);
+	&data_byte(0x1f, 0xdd, 0xa8, 0x33, 0x88, 0x07, 0xc7, 0x31);
+	&data_byte(0xb1, 0x12, 0x10, 0x59, 0x27, 0x80, 0xec, 0x5f);
+	&data_byte(0x60, 0x51, 0x7f, 0xa9, 0x19, 0xb5, 0x4a, 0x0d);
+	&data_byte(0x2d, 0xe5, 0x7a, 0x9f, 0x93, 0xc9, 0x9c, 0xef);
+	&data_byte(0xa0, 0xe0, 0x3b, 0x4d, 0xae, 0x2a, 0xf5, 0xb0);
+	&data_byte(0xc8, 0xeb, 0xbb, 0x3c, 0x83, 0x53, 0x99, 0x61);
+	&data_byte(0x17, 0x2b, 0x04, 0x7e, 0xba, 0x77, 0xd6, 0x26);
+	&data_byte(0xe1, 0x69, 0x14, 0x63, 0x55, 0x21, 0x0c, 0x7d);
+$code.=<<___;
+	.long	0x80808080, 0x80808080, 0xfefefefe, 0xfefefefe
+	.long	0x1b1b1b1b, 0x1b1b1b1b, 0, 0
+___
+	&data_byte(0x52, 0x09, 0x6a, 0xd5, 0x30, 0x36, 0xa5, 0x38);
+	&data_byte(0xbf, 0x40, 0xa3, 0x9e, 0x81, 0xf3, 0xd7, 0xfb);
+	&data_byte(0x7c, 0xe3, 0x39, 0x82, 0x9b, 0x2f, 0xff, 0x87);
+	&data_byte(0x34, 0x8e, 0x43, 0x44, 0xc4, 0xde, 0xe9, 0xcb);
+	&data_byte(0x54, 0x7b, 0x94, 0x32, 0xa6, 0xc2, 0x23, 0x3d);
+	&data_byte(0xee, 0x4c, 0x95, 0x0b, 0x42, 0xfa, 0xc3, 0x4e);
+	&data_byte(0x08, 0x2e, 0xa1, 0x66, 0x28, 0xd9, 0x24, 0xb2);
+	&data_byte(0x76, 0x5b, 0xa2, 0x49, 0x6d, 0x8b, 0xd1, 0x25);
+	&data_byte(0x72, 0xf8, 0xf6, 0x64, 0x86, 0x68, 0x98, 0x16);
+	&data_byte(0xd4, 0xa4, 0x5c, 0xcc, 0x5d, 0x65, 0xb6, 0x92);
+	&data_byte(0x6c, 0x70, 0x48, 0x50, 0xfd, 0xed, 0xb9, 0xda);
+	&data_byte(0x5e, 0x15, 0x46, 0x57, 0xa7, 0x8d, 0x9d, 0x84);
+	&data_byte(0x90, 0xd8, 0xab, 0x00, 0x8c, 0xbc, 0xd3, 0x0a);
+	&data_byte(0xf7, 0xe4, 0x58, 0x05, 0xb8, 0xb3, 0x45, 0x06);
+	&data_byte(0xd0, 0x2c, 0x1e, 0x8f, 0xca, 0x3f, 0x0f, 0x02);
+	&data_byte(0xc1, 0xaf, 0xbd, 0x03, 0x01, 0x13, 0x8a, 0x6b);
+	&data_byte(0x3a, 0x91, 0x11, 0x41, 0x4f, 0x67, 0xdc, 0xea);
+	&data_byte(0x97, 0xf2, 0xcf, 0xce, 0xf0, 0xb4, 0xe6, 0x73);
+	&data_byte(0x96, 0xac, 0x74, 0x22, 0xe7, 0xad, 0x35, 0x85);
+	&data_byte(0xe2, 0xf9, 0x37, 0xe8, 0x1c, 0x75, 0xdf, 0x6e);
+	&data_byte(0x47, 0xf1, 0x1a, 0x71, 0x1d, 0x29, 0xc5, 0x89);
+	&data_byte(0x6f, 0xb7, 0x62, 0x0e, 0xaa, 0x18, 0xbe, 0x1b);
+	&data_byte(0xfc, 0x56, 0x3e, 0x4b, 0xc6, 0xd2, 0x79, 0x20);
+	&data_byte(0x9a, 0xdb, 0xc0, 0xfe, 0x78, 0xcd, 0x5a, 0xf4);
+	&data_byte(0x1f, 0xdd, 0xa8, 0x33, 0x88, 0x07, 0xc7, 0x31);
+	&data_byte(0xb1, 0x12, 0x10, 0x59, 0x27, 0x80, 0xec, 0x5f);
+	&data_byte(0x60, 0x51, 0x7f, 0xa9, 0x19, 0xb5, 0x4a, 0x0d);
+	&data_byte(0x2d, 0xe5, 0x7a, 0x9f, 0x93, 0xc9, 0x9c, 0xef);
+	&data_byte(0xa0, 0xe0, 0x3b, 0x4d, 0xae, 0x2a, 0xf5, 0xb0);
+	&data_byte(0xc8, 0xeb, 0xbb, 0x3c, 0x83, 0x53, 0x99, 0x61);
+	&data_byte(0x17, 0x2b, 0x04, 0x7e, 0xba, 0x77, 0xd6, 0x26);
+	&data_byte(0xe1, 0x69, 0x14, 0x63, 0x55, 0x21, 0x0c, 0x7d);
+$code.=<<___;
+	.long	0x80808080, 0x80808080, 0xfefefefe, 0xfefefefe
+	.long	0x1b1b1b1b, 0x1b1b1b1b, 0, 0
+___
+	&data_byte(0x52, 0x09, 0x6a, 0xd5, 0x30, 0x36, 0xa5, 0x38);
+	&data_byte(0xbf, 0x40, 0xa3, 0x9e, 0x81, 0xf3, 0xd7, 0xfb);
+	&data_byte(0x7c, 0xe3, 0x39, 0x82, 0x9b, 0x2f, 0xff, 0x87);
+	&data_byte(0x34, 0x8e, 0x43, 0x44, 0xc4, 0xde, 0xe9, 0xcb);
+	&data_byte(0x54, 0x7b, 0x94, 0x32, 0xa6, 0xc2, 0x23, 0x3d);
+	&data_byte(0xee, 0x4c, 0x95, 0x0b, 0x42, 0xfa, 0xc3, 0x4e);
+	&data_byte(0x08, 0x2e, 0xa1, 0x66, 0x28, 0xd9, 0x24, 0xb2);
+	&data_byte(0x76, 0x5b, 0xa2, 0x49, 0x6d, 0x8b, 0xd1, 0x25);
+	&data_byte(0x72, 0xf8, 0xf6, 0x64, 0x86, 0x68, 0x98, 0x16);
+	&data_byte(0xd4, 0xa4, 0x5c, 0xcc, 0x5d, 0x65, 0xb6, 0x92);
+	&data_byte(0x6c, 0x70, 0x48, 0x50, 0xfd, 0xed, 0xb9, 0xda);
+	&data_byte(0x5e, 0x15, 0x46, 0x57, 0xa7, 0x8d, 0x9d, 0x84);
+	&data_byte(0x90, 0xd8, 0xab, 0x00, 0x8c, 0xbc, 0xd3, 0x0a);
+	&data_byte(0xf7, 0xe4, 0x58, 0x05, 0xb8, 0xb3, 0x45, 0x06);
+	&data_byte(0xd0, 0x2c, 0x1e, 0x8f, 0xca, 0x3f, 0x0f, 0x02);
+	&data_byte(0xc1, 0xaf, 0xbd, 0x03, 0x01, 0x13, 0x8a, 0x6b);
+	&data_byte(0x3a, 0x91, 0x11, 0x41, 0x4f, 0x67, 0xdc, 0xea);
+	&data_byte(0x97, 0xf2, 0xcf, 0xce, 0xf0, 0xb4, 0xe6, 0x73);
+	&data_byte(0x96, 0xac, 0x74, 0x22, 0xe7, 0xad, 0x35, 0x85);
+	&data_byte(0xe2, 0xf9, 0x37, 0xe8, 0x1c, 0x75, 0xdf, 0x6e);
+	&data_byte(0x47, 0xf1, 0x1a, 0x71, 0x1d, 0x29, 0xc5, 0x89);
+	&data_byte(0x6f, 0xb7, 0x62, 0x0e, 0xaa, 0x18, 0xbe, 0x1b);
+	&data_byte(0xfc, 0x56, 0x3e, 0x4b, 0xc6, 0xd2, 0x79, 0x20);
+	&data_byte(0x9a, 0xdb, 0xc0, 0xfe, 0x78, 0xcd, 0x5a, 0xf4);
+	&data_byte(0x1f, 0xdd, 0xa8, 0x33, 0x88, 0x07, 0xc7, 0x31);
+	&data_byte(0xb1, 0x12, 0x10, 0x59, 0x27, 0x80, 0xec, 0x5f);
+	&data_byte(0x60, 0x51, 0x7f, 0xa9, 0x19, 0xb5, 0x4a, 0x0d);
+	&data_byte(0x2d, 0xe5, 0x7a, 0x9f, 0x93, 0xc9, 0x9c, 0xef);
+	&data_byte(0xa0, 0xe0, 0x3b, 0x4d, 0xae, 0x2a, 0xf5, 0xb0);
+	&data_byte(0xc8, 0xeb, 0xbb, 0x3c, 0x83, 0x53, 0x99, 0x61);
+	&data_byte(0x17, 0x2b, 0x04, 0x7e, 0xba, 0x77, 0xd6, 0x26);
+	&data_byte(0xe1, 0x69, 0x14, 0x63, 0x55, 0x21, 0x0c, 0x7d);
+$code.=<<___;
+	.long	0x80808080, 0x80808080, 0xfefefefe, 0xfefefefe
+	.long	0x1b1b1b1b, 0x1b1b1b1b, 0, 0
+___
 	&data_byte(0x52, 0x09, 0x6a, 0xd5, 0x30, 0x36, 0xa5, 0x38);
 	&data_byte(0xbf, 0x40, 0xa3, 0x9e, 0x81, 0xf3, 0xd7, 0xfb);
 	&data_byte(0x7c, 0xe3, 0x39, 0x82, 0x9b, 0x2f, 0xff, 0x87);
@@ -1571,6 +2519,288 @@ ___
 	&data_byte(0xc8, 0xeb, 0xbb, 0x3c, 0x83, 0x53, 0x99, 0x61);
 	&data_byte(0x17, 0x2b, 0x04, 0x7e, 0xba, 0x77, 0xd6, 0x26);
 	&data_byte(0xe1, 0x69, 0x14, 0x63, 0x55, 0x21, 0x0c, 0x7d);
+$code.=<<___;
+	.long	0x80808080, 0x80808080, 0xfefefefe, 0xfefefefe
+	.long	0x1b1b1b1b, 0x1b1b1b1b, 0, 0
+.asciz  "AES for x86_64, CRYPTOGAMS by <appro\@openssl.org>"
+.align	64
+___
+
+# EXCEPTION_DISPOSITION handler (EXCEPTION_RECORD *rec,ULONG64 frame,
+#		CONTEXT *context,DISPATCHER_CONTEXT *disp)
+if ($win64) {
+$rec="%rcx";
+$frame="%rdx";
+$context="%r8";
+$disp="%r9";
+
+$code.=<<___;
+.extern	__imp_RtlVirtualUnwind
+.type	block_se_handler,\@abi-omnipotent
+.align	16
+block_se_handler:
+	push	%rsi
+	push	%rdi
+	push	%rbx
+	push	%rbp
+	push	%r12
+	push	%r13
+	push	%r14
+	push	%r15
+	pushfq
+	sub	\$64,%rsp
+
+	mov	120($context),%rax	# pull context->Rax
+	mov	248($context),%rbx	# pull context->Rip
+
+	mov	8($disp),%rsi		# disp->ImageBase
+	mov	56($disp),%r11		# disp->HandlerData
+
+	mov	0(%r11),%r10d		# HandlerData[0]
+	lea	(%rsi,%r10),%r10	# prologue label
+	cmp	%r10,%rbx		# context->Rip<prologue label
+	jb	.Lin_block_prologue
+
+	mov	152($context),%rax	# pull context->Rsp
+
+	mov	4(%r11),%r10d		# HandlerData[1]
+	lea	(%rsi,%r10),%r10	# epilogue label
+	cmp	%r10,%rbx		# context->Rip>=epilogue label
+	jae	.Lin_block_prologue
+
+	mov	24(%rax),%rax		# pull saved real stack pointer
+	lea	48(%rax),%rax		# adjust...
+
+	mov	-8(%rax),%rbx
+	mov	-16(%rax),%rbp
+	mov	-24(%rax),%r12
+	mov	-32(%rax),%r13
+	mov	-40(%rax),%r14
+	mov	-48(%rax),%r15
+	mov	%rbx,144($context)	# restore context->Rbx
+	mov	%rbp,160($context)	# restore context->Rbp
+	mov	%r12,216($context)	# restore context->R12
+	mov	%r13,224($context)	# restore context->R13
+	mov	%r14,232($context)	# restore context->R14
+	mov	%r15,240($context)	# restore context->R15
+
+.Lin_block_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($context)	# restore context->Rdi
+
+	jmp	.Lcommon_seh_exit
+.size	block_se_handler,.-block_se_handler
+
+.type	key_se_handler,\@abi-omnipotent
+.align	16
+key_se_handler:
+	push	%rsi
+	push	%rdi
+	push	%rbx
+	push	%rbp
+	push	%r12
+	push	%r13
+	push	%r14
+	push	%r15
+	pushfq
+	sub	\$64,%rsp
+
+	mov	120($context),%rax	# pull context->Rax
+	mov	248($context),%rbx	# pull context->Rip
+
+	mov	8($disp),%rsi		# disp->ImageBase
+	mov	56($disp),%r11		# disp->HandlerData
+
+	mov	0(%r11),%r10d		# HandlerData[0]
+	lea	(%rsi,%r10),%r10	# prologue label
+	cmp	%r10,%rbx		# context->Rip<prologue label
+	jb	.Lin_key_prologue
+
+	mov	152($context),%rax	# pull context->Rsp
+
+	mov	4(%r11),%r10d		# HandlerData[1]
+	lea	(%rsi,%r10),%r10	# epilogue label
+	cmp	%r10,%rbx		# context->Rip>=epilogue label
+	jae	.Lin_key_prologue
+
+	lea	56(%rax),%rax
+
+	mov	-8(%rax),%rbx
+	mov	-16(%rax),%rbp
+	mov	-24(%rax),%r12
+	mov	-32(%rax),%r13
+	mov	-40(%rax),%r14
+	mov	-48(%rax),%r15
+	mov	%rbx,144($context)	# restore context->Rbx
+	mov	%rbp,160($context)	# restore context->Rbp
+	mov	%r12,216($context)	# restore context->R12
+	mov	%r13,224($context)	# restore context->R13
+	mov	%r14,232($context)	# restore context->R14
+	mov	%r15,240($context)	# restore context->R15
+
+.Lin_key_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($context)	# restore context->Rdi
+
+	jmp	.Lcommon_seh_exit
+.size	key_se_handler,.-key_se_handler
+
+.type	cbc_se_handler,\@abi-omnipotent
+.align	16
+cbc_se_handler:
+	push	%rsi
+	push	%rdi
+	push	%rbx
+	push	%rbp
+	push	%r12
+	push	%r13
+	push	%r14
+	push	%r15
+	pushfq
+	sub	\$64,%rsp
+
+	mov	120($context),%rax	# pull context->Rax
+	mov	248($context),%rbx	# pull context->Rip
+
+	lea	.Lcbc_prologue(%rip),%r10
+	cmp	%r10,%rbx		# context->Rip<.Lcbc_prologue
+	jb	.Lin_cbc_prologue
+
+	lea	.Lcbc_fast_body(%rip),%r10
+	cmp	%r10,%rbx		# context->Rip<.Lcbc_fast_body
+	jb	.Lin_cbc_frame_setup
+
+	lea	.Lcbc_slow_prologue(%rip),%r10
+	cmp	%r10,%rbx		# context->Rip<.Lcbc_slow_prologue
+	jb	.Lin_cbc_body
+
+	lea	.Lcbc_slow_body(%rip),%r10
+	cmp	%r10,%rbx		# context->Rip<.Lcbc_slow_body
+	jb	.Lin_cbc_frame_setup
+
+.Lin_cbc_body:
+	mov	152($context),%rax	# pull context->Rsp
+
+	lea	.Lcbc_epilogue(%rip),%r10
+	cmp	%r10,%rbx		# context->Rip>=.Lcbc_epilogue
+	jae	.Lin_cbc_prologue
+
+	lea	8(%rax),%rax
+
+	lea	.Lcbc_popfq(%rip),%r10
+	cmp	%r10,%rbx		# context->Rip>=.Lcbc_popfq
+	jae	.Lin_cbc_prologue
+
+	mov	`16-8`(%rax),%rax	# biased $_rsp
+	lea	56(%rax),%rax
+
+.Lin_cbc_frame_setup:
+	mov	-16(%rax),%rbx
+	mov	-24(%rax),%rbp
+	mov	-32(%rax),%r12
+	mov	-40(%rax),%r13
+	mov	-48(%rax),%r14
+	mov	-56(%rax),%r15
+	mov	%rbx,144($context)	# restore context->Rbx
+	mov	%rbp,160($context)	# restore context->Rbp
+	mov	%r12,216($context)	# restore context->R12
+	mov	%r13,224($context)	# restore context->R13
+	mov	%r14,232($context)	# restore context->R14
+	mov	%r15,240($context)	# restore context->R15
+
+.Lin_cbc_prologue:
+	mov	8(%rax),%rdi
+	mov	16(%rax),%rsi
+	mov	%rax,152($context)	# restore context->Rsp
+	mov	%rsi,168($context)	# restore context->Rsi
+	mov	%rdi,176($context)	# restore context->Rdi
+
+.Lcommon_seh_exit:
+
+	mov	40($disp),%rdi		# disp->ContextRecord
+	mov	$context,%rsi		# context
+	mov	\$`1232/8`,%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	cbc_se_handler,.-cbc_se_handler
+
+.section	.pdata
+.align	4
+	.rva	.LSEH_begin_AES_encrypt
+	.rva	.LSEH_end_AES_encrypt
+	.rva	.LSEH_info_AES_encrypt
+
+	.rva	.LSEH_begin_AES_decrypt
+	.rva	.LSEH_end_AES_decrypt
+	.rva	.LSEH_info_AES_decrypt
+
+	.rva	.LSEH_begin_AES_set_encrypt_key
+	.rva	.LSEH_end_AES_set_encrypt_key
+	.rva	.LSEH_info_AES_set_encrypt_key
+
+	.rva	.LSEH_begin_AES_set_decrypt_key
+	.rva	.LSEH_end_AES_set_decrypt_key
+	.rva	.LSEH_info_AES_set_decrypt_key
+
+	.rva	.LSEH_begin_AES_cbc_encrypt
+	.rva	.LSEH_end_AES_cbc_encrypt
+	.rva	.LSEH_info_AES_cbc_encrypt
+
+.section	.xdata
+.align	8
+.LSEH_info_AES_encrypt:
+	.byte	9,0,0,0
+	.rva	block_se_handler
+	.rva	.Lenc_prologue,.Lenc_epilogue	# HandlerData[]
+.LSEH_info_AES_decrypt:
+	.byte	9,0,0,0
+	.rva	block_se_handler
+	.rva	.Ldec_prologue,.Ldec_epilogue	# HandlerData[]
+.LSEH_info_AES_set_encrypt_key:
+	.byte	9,0,0,0
+	.rva	key_se_handler
+	.rva	.Lenc_key_prologue,.Lenc_key_epilogue	# HandlerData[]
+.LSEH_info_AES_set_decrypt_key:
+	.byte	9,0,0,0
+	.rva	key_se_handler
+	.rva	.Ldec_key_prologue,.Ldec_key_epilogue	# HandlerData[]
+.LSEH_info_AES_cbc_encrypt:
+	.byte	9,0,0,0
+	.rva	cbc_se_handler
+___
+}
 
 $code =~ s/\`([^\`]*)\`/eval($1)/gem;