aes-586.pl

#!/usr/bin/env perl
#
# ====================================================================
# 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.
# ====================================================================
#
# 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
# to be *the* best Intel C compiler without -KPIC, performance appears
# to be virtually identical... But try to re-configure with shared
# library support... Aha! Intel compiler "suddenly" lags behind by 30%
# [on P4, more on others]:-) And if compared to position-independent
# code generated by GNU C, this code performs *more* than *twice* as
# fast! Yes, all this buzz about PIC means that unlike other hand-
# coded implementations, this one was explicitly designed to be safe
# to use even in shared library context... This also means that this
# code isn't necessarily absolutely fastest "ever," because in order
# to achieve position independence an extra register has to be
# off-loaded to stack, which affects the benchmark result.
#
# Special note about instruction choice. Do you recall RC4_INT code
# performing poorly on P4? It might be the time to figure out why.
# RC4_INT code implies effective address calculations in base+offset*4
# form. Trouble is that it seems that offset scaling turned to be
# critical path... At least eliminating scaling resulted in 2.8x RC4
# performance improvement [as you might recall]. As AES code is hungry
# for scaling too, I [try to] avoid the latter by favoring off-by-2
# shifts and masking the result with 0xFF<<2 instead of "boring" 0xFF.
#
# As was shown by Dean Gaudet <dean@arctic.org>, the above note turned
# void. Performance improvement with off-by-2 shifts was observed on
# intermediate implementation, which was spilling yet another register
# to stack... Final offset*4 code below runs just a tad faster on P4,
# but exhibits up to 10% improvement on other cores.
#
# Second version is "monolithic" replacement for aes_core.c, which in
# addition to AES_[de|en]crypt implements AES_set_[de|en]cryption_key.
# This made it possible to implement little-endian variant of the
# algorithm without modifying the base C code. Motivating factor for
# the undertaken effort was that it appeared that in tight IA-32
# register window little-endian flavor could achieve slightly higher
# Instruction Level Parallelism, and it indeed resulted in up to 15%
# better performance on most recent µ-archs...
#
# Third version adds AES_cbc_encrypt implementation, which resulted in
# up to 40% performance imrovement of CBC benchmark results. 40% was
# observed on P4 core, where "overall" imrovement coefficient, i.e. if
# compared to PIC generated by GCC and in CBC mode, was observed to be
# as large as 4x:-) CBC performance is virtually identical to ECB now
# and on some platforms even better, e.g. 17.6 "small" cycles/byte on
# Opteron, because certain function prologues and epilogues are
# effectively taken out of the loop...
#
# Version 3.2 implements compressed tables and prefetch of these tables
# in CBC[!] mode. Former means that 3/4 of table references are now
# misaligned, which unfortunately has negative impact on elder IA-32
# implementations, Pentium suffered 30% penalty, PIII - 10%.
#
# Version 3.3 avoids L1 cache aliasing between stack frame and
# S-boxes, and 3.4 - L1 cache aliasing even between key schedule. The
# latter is achieved by copying the key schedule to controlled place in
# stack. This unfortunately has rather strong impact on small block CBC
# performance, ~2x deterioration on 16-byte block if compared to 3.3.
#
# Version 3.5 checks if there is L1 cache aliasing between user-supplied
# key schedule and S-boxes and abstains from copying the former if
# there is no. This allows end-user to consciously retain small block
# performance by aligning key schedule in specific manner.
#
# Version 3.6 compresses Td4 to 256 bytes and prefetches it in ECB.
#
# Current ECB performance numbers for 128-bit key in CPU cycles per
# processed byte [measure commonly used by AES benchmarkers] are:
#
#		small footprint		fully unrolled
# P4		24			22
# AMD K8	20			19
# PIII		25			23
# Pentium	81			78
#
# 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
# 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 process 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 masks 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!

push(@INC,"perlasm","../../perlasm");
require "x86asm.pl";

&asm_init($ARGV[0],"aes-586.pl",$x86only = $ARGV[$#ARGV] eq "386");

$s0="eax";
$s1="ebx";
$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");

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
# "column," unlike "horizontal" one, which groups shifts on different
# $s[i] to form a "row"] is work in progress. It was observed to run
# few percents faster on Intel cores, but not AMD. On AMD K8 core it's
# whole 12% slower:-( So we face a trade-off... Shall it be resolved
# some day? Till then the code is considered experimental and by
# default remains dormant...

sub encvert()
{ my ($te,@s) = @_;
  my $v0 = $acc, $v1 = $key;

	&mov	($v0,$s[3]);				# copy s3
	&mov	(&DWP(4,"esp"),$s[2]);			# save s2
	&mov	($v1,$s[0]);				# copy s0
	&mov	(&DWP(8,"esp"),$s[1]);			# save s1

	&movz	($s[2],&HB($s[0]));
	&and	($s[0],0xFF);
	&mov	($s[0],&DWP(0,$te,$s[0],8));		# s0>>0
	&shr	($v1,16);
	&mov	($s[3],&DWP(3,$te,$s[2],8));		# s0>>8
	&movz	($s[1],&HB($v1));
	&and	($v1,0xFF);
	&mov	($s[2],&DWP(2,$te,$v1,8));		# s0>>16
	 &mov	($v1,$v0);
	&mov	($s[1],&DWP(1,$te,$s[1],8));		# s0>>24

	&and	($v0,0xFF);
	&xor	($s[3],&DWP(0,$te,$v0,8));		# s3>>0
	&movz	($v0,&HB($v1));
	&shr	($v1,16);
	&xor	($s[2],&DWP(3,$te,$v0,8));		# s3>>8
	&movz	($v0,&HB($v1));
	&and	($v1,0xFF);
	&xor	($s[1],&DWP(2,$te,$v1,8));		# s3>>16
	 &mov	($v1,&DWP(4,"esp"));			# restore s2
	&xor	($s[0],&DWP(1,$te,$v0,8));		# s3>>24

	&mov	($v0,$v1);
	&and	($v1,0xFF);
	&xor	($s[2],&DWP(0,$te,$v1,8));		# s2>>0
	&movz	($v1,&HB($v0));
	&shr	($v0,16);
	&xor	($s[1],&DWP(3,$te,$v1,8));		# s2>>8
	&movz	($v1,&HB($v0));
	&and	($v0,0xFF);
	&xor	($s[0],&DWP(2,$te,$v0,8));		# s2>>16
	 &mov	($v0,&DWP(8,"esp"));			# restore s1
	&xor	($s[3],&DWP(1,$te,$v1,8));		# s2>>24

	&mov	($v1,$v0);
	&and	($v0,0xFF);
	&xor	($s[1],&DWP(0,$te,$v0,8));		# s1>>0
	&movz	($v0,&HB($v1));
	&shr	($v1,16);
	&xor	($s[0],&DWP(3,$te,$v0,8));		# s1>>8
	&movz	($v0,&HB($v1));
	&and	($v1,0xFF);
	&xor	($s[3],&DWP(2,$te,$v1,8));		# s1>>16
	 &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);
	&mov	($r2,$s[$i]);
	&shr	($tmp,7);
	&and	($r2,0x7f7f7f7f);
	&sub	($acc,$tmp);
	&lea	($r2,&DWP(0,$r2,$r2));
	&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);
}

&public_label("AES_Te");
&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 (!$x86only) {
&public_label("AES_Te");
&function_begin_B("_sse_AES_encrypt_compact");
	&pxor	("mm0",&QWP(0,$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,$__key);			}##%edx
	else        {	&mov	($out,$s[0]);
			&and	($out,0xFF);			}
	if ($i==1)  {	&shr	($s[0],16);			}#%ebx[1]
	if ($i==2)  {	&shr	($s[0],24);			}#%ecx[2]
			&mov	($out,&DWP(0,$te,$out,8));

	if ($i==3)  {	$tmp=$s[1];				}##%eax
			&movz	($tmp,&HB($s[1]));
			&xor	($out,&DWP(3,$te,$tmp,8));

	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 ($s[2],$__s1);		}##%ecx
	elsif($i==2){	&movz	($tmp,&HB($s[3]));		}#%ebx[2]
	else        {	&mov	($tmp,$s[3]); 
			&shr	($tmp,24)			}
			&xor	($out,&DWP(1,$te,$tmp,8));
	if ($i<2)   {	&mov	(&DWP(4+4*$i,"esp"),$out);	}
	if ($i==3)  {	&mov	($s[3],$acc);			}
			&comment();
}

sub enclast()
{ my ($i,$te,@s)=@_;
  my $tmp = $key;
  my $out = $i==3?$s[0]:$acc;

	if ($i==3)  {	&mov	($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]
			&mov	($out,&DWP(2,$te,$out,8));
			&and	($out,0x000000ff);

	if ($i==3)  {	$tmp=$s[1];				}##%eax
			&movz	($tmp,&HB($s[1]));
			&mov	($tmp,&DWP(0,$te,$tmp,8));
			&and	($tmp,0x0000ff00);
			&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);
			&mov	($tmp,&DWP(0,$te,$tmp,8));
			&and	($tmp,0x00ff0000);
			&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);			}
			&mov	($tmp,&DWP(2,$te,$tmp,8));
			&and	($tmp,0xff000000);
			&xor	($out,$tmp);
	if ($i<2)   {	&mov	(&DWP(4+4*$i,"esp"),$out);	}
	if ($i==3)  {	&mov	($s[3],$acc);			}
}

&public_label("AES_Te");
&function_begin_B("_x86_AES_encrypt");
	if ($vertical_spin) {
		# I need high parts of volatile registers to be accessible...
		&exch	($s1="edi",$key="ebx");
		&mov	($s2="esi",$acc="ecx");
	}

	# 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

	if ($small_footprint) {
	    &lea	($acc,&DWP(-2,$acc,$acc));
	    &lea	($acc,&DWP(0,$key,$acc,8));
	    &mov	($__end,$acc);		# end of key schedule

	    &set_label("loop",16);
		if ($vertical_spin) {
		    &encvert($tbl,$s0,$s1,$s2,$s3);
		} else {
		    &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,$__end);
	    &mov	($__key,$key);
	    &jb		(&label("loop"));
	}
	else {
	    &cmp	($acc,10);
	    &jle	(&label("10rounds"));
	    &cmp	($acc,12);
	    &jle	(&label("12rounds"));

	&set_label("14rounds",4);
	    for ($i=1;$i<3;$i++) {
		if ($vertical_spin) {
		    &encvert($tbl,$s0,$s1,$s2,$s3);
		} else {
		    &encstep(0,$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,$key));
		&xor	($s2,&DWP(16*$i+8,$key));
		&xor	($s3,&DWP(16*$i+12,$key));
	    }
	    &add	($key,32);
	    &mov	($__key,$key);		# advance rd_key
	&set_label("12rounds",4);
	    for ($i=1;$i<3;$i++) {
		if ($vertical_spin) {
		    &encvert($tbl,$s0,$s1,$s2,$s3);
		} else {
		    &encstep(0,$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,$key));
		&xor	($s2,&DWP(16*$i+8,$key));
		&xor	($s3,&DWP(16*$i+12,$key));
	    }
	    &add	($key,32);
	    &mov	($__key,$key);		# advance rd_key
	&set_label("10rounds",4);
	    for ($i=1;$i<10;$i++) {
		if ($vertical_spin) {
		    &encvert($tbl,$s0,$s1,$s2,$s3);
		} else {
		    &encstep(0,$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,$key));
		&xor	($s2,&DWP(16*$i+8,$key));
		&xor	($s3,&DWP(16*$i+12,$key));
	    }
	}

	if ($vertical_spin) {
	    # "reincarnate" some registers for "horizontal" spin...
	    &mov	($s1="ebx",$key="edi");
	    &mov	($s2="ecx",$acc="esi");
	}
	&enclast(0,$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));
	&xor	($s1,&DWP(4,$key));
	&xor	($s2,&DWP(8,$key));
	&xor	($s3,&DWP(12,$key));

	&ret	();

&set_label("AES_Te",64);	# Yes! I keep it in the code segment!
	&_data_word(0xa56363c6, 0x847c7cf8, 0x997777ee, 0x8d7b7bf6);
	&_data_word(0x0df2f2ff, 0xbd6b6bd6, 0xb16f6fde, 0x54c5c591);
	&_data_word(0x50303060, 0x03010102, 0xa96767ce, 0x7d2b2b56);
	&_data_word(0x19fefee7, 0x62d7d7b5, 0xe6abab4d, 0x9a7676ec);
	&_data_word(0x45caca8f, 0x9d82821f, 0x40c9c989, 0x877d7dfa);
	&_data_word(0x15fafaef, 0xeb5959b2, 0xc947478e, 0x0bf0f0fb);
	&_data_word(0xecadad41, 0x67d4d4b3, 0xfda2a25f, 0xeaafaf45);
	&_data_word(0xbf9c9c23, 0xf7a4a453, 0x967272e4, 0x5bc0c09b);
	&_data_word(0xc2b7b775, 0x1cfdfde1, 0xae93933d, 0x6a26264c);
	&_data_word(0x5a36366c, 0x413f3f7e, 0x02f7f7f5, 0x4fcccc83);
	&_data_word(0x5c343468, 0xf4a5a551, 0x34e5e5d1, 0x08f1f1f9);
	&_data_word(0x937171e2, 0x73d8d8ab, 0x53313162, 0x3f15152a);
	&_data_word(0x0c040408, 0x52c7c795, 0x65232346, 0x5ec3c39d);
	&_data_word(0x28181830, 0xa1969637, 0x0f05050a, 0xb59a9a2f);
	&_data_word(0x0907070e, 0x36121224, 0x9b80801b, 0x3de2e2df);
	&_data_word(0x26ebebcd, 0x6927274e, 0xcdb2b27f, 0x9f7575ea);
	&_data_word(0x1b090912, 0x9e83831d, 0x742c2c58, 0x2e1a1a34);
	&_data_word(0x2d1b1b36, 0xb26e6edc, 0xee5a5ab4, 0xfba0a05b);
	&_data_word(0xf65252a4, 0x4d3b3b76, 0x61d6d6b7, 0xceb3b37d);
	&_data_word(0x7b292952, 0x3ee3e3dd, 0x712f2f5e, 0x97848413);
	&_data_word(0xf55353a6, 0x68d1d1b9, 0x00000000, 0x2cededc1);
	&_data_word(0x60202040, 0x1ffcfce3, 0xc8b1b179, 0xed5b5bb6);
	&_data_word(0xbe6a6ad4, 0x46cbcb8d, 0xd9bebe67, 0x4b393972);
	&_data_word(0xde4a4a94, 0xd44c4c98, 0xe85858b0, 0x4acfcf85);
	&_data_word(0x6bd0d0bb, 0x2aefefc5, 0xe5aaaa4f, 0x16fbfbed);
	&_data_word(0xc5434386, 0xd74d4d9a, 0x55333366, 0x94858511);
	&_data_word(0xcf45458a, 0x10f9f9e9, 0x06020204, 0x817f7ffe);
	&_data_word(0xf05050a0, 0x443c3c78, 0xba9f9f25, 0xe3a8a84b);
	&_data_word(0xf35151a2, 0xfea3a35d, 0xc0404080, 0x8a8f8f05);
	&_data_word(0xad92923f, 0xbc9d9d21, 0x48383870, 0x04f5f5f1);
	&_data_word(0xdfbcbc63, 0xc1b6b677, 0x75dadaaf, 0x63212142);
	&_data_word(0x30101020, 0x1affffe5, 0x0ef3f3fd, 0x6dd2d2bf);
	&_data_word(0x4ccdcd81, 0x140c0c18, 0x35131326, 0x2fececc3);
	&_data_word(0xe15f5fbe, 0xa2979735, 0xcc444488, 0x3917172e);
	&_data_word(0x57c4c493, 0xf2a7a755, 0x827e7efc, 0x473d3d7a);
	&_data_word(0xac6464c8, 0xe75d5dba, 0x2b191932, 0x957373e6);
	&_data_word(0xa06060c0, 0x98818119, 0xd14f4f9e, 0x7fdcdca3);
	&_data_word(0x66222244, 0x7e2a2a54, 0xab90903b, 0x8388880b);
	&_data_word(0xca46468c, 0x29eeeec7, 0xd3b8b86b, 0x3c141428);
	&_data_word(0x79dedea7, 0xe25e5ebc, 0x1d0b0b16, 0x76dbdbad);
	&_data_word(0x3be0e0db, 0x56323264, 0x4e3a3a74, 0x1e0a0a14);
	&_data_word(0xdb494992, 0x0a06060c, 0x6c242448, 0xe45c5cb8);
	&_data_word(0x5dc2c29f, 0x6ed3d3bd, 0xefacac43, 0xa66262c4);
	&_data_word(0xa8919139, 0xa4959531, 0x37e4e4d3, 0x8b7979f2);
	&_data_word(0x32e7e7d5, 0x43c8c88b, 0x5937376e, 0xb76d6dda);
	&_data_word(0x8c8d8d01, 0x64d5d5b1, 0xd24e4e9c, 0xe0a9a949);
	&_data_word(0xb46c6cd8, 0xfa5656ac, 0x07f4f4f3, 0x25eaeacf);
	&_data_word(0xaf6565ca, 0x8e7a7af4, 0xe9aeae47, 0x18080810);