"Monolithic" x86 assembler replacement for aes_core.c. Up to +15% better

performance on recent microarchitectures.

2 files changed, 1193 insertions, 747 deletions

diff --git a/crypto/aes/asm/aes-586.pl b/crypto/aes/asm/aes-586.pl
index c61ed6bdb2..688fda21ff 100755
--- a/crypto/aes/asm/aes-586.pl
+++ b/crypto/aes/asm/aes-586.pl
@@ -6,15 +6,21 @@
 # forms are granted according to the OpenSSL license.
 # ====================================================================
 #
+# Version 2.0.
+#
 # You might fail to appreciate this module performance from the first
-# try. If compared to "vanilla" linux-ia32-icc target, i.e. Intel C
-# without -KPIC, performance appears to be virtually identical... But
-# try to configure with shared library support... Aha! Intel compiler
-# "suddenly" lags behind by 30% [on P4]:-) 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 implementations] this module was explicitly designed
-# to be safe to use even in shared library context...
+# 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.
@@ -30,212 +36,326 @@
 # 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...
+#
+# Current ECB performance numbers for 128-bit key in cycles per byte
+# [measure commonly used by AES benchmarkers] are:
+#
+#		small footprint		fully unrolled
+# P4[-3]	23[24]			22[23]
+# AMD K8	19			18
+# PIII		26(*)			23
+# Pentium	63(*)			52
+#
+# (*)	Performance difference between small footprint code and fully
+#	unrolled in more commonly used CBC mode is not as big, 7% for
+#	PIII and 15% for Pentium, which I consider tolerable.
 
 push(@INC,"perlasm","../../perlasm");
 require "x86asm.pl";
 
 &asm_init($ARGV[0],"aes-586.pl",$ARGV[$#ARGV] eq "386");
 
-$small_footprint=1;	# $small_footprint=1 code is 4-6% slower, but
-			# 5 times smaller! I default to compact code.
+$small_footprint=1;	# $small_footprint=1 code is ~5% slower [on
+			# recent µ-archs], but ~5 times smaller!
+			# I favor compact code, because it minimizes
+			# cache contention...
+$vertical_spin=0;	# shift "verticaly" defaults to 0, because of
+			# its proof-of-concept status, see below...
+
 $s0="eax";
 $s1="ebx";
 $s2="ecx";
 $s3="edx";
+$key="esi";
+$acc="edi";
+
+if ($vertical_spin) {
+	# I need high parts of volatile registers to be accessible...
+	$s1="esi";	$key="ebx";
+	$s2="edi";	$acc="ecx";
+}
+# 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,$s[3]);				# copy s3
+	&mov	(&DWP(0,"esp"),$s[2]);			# save s2
+	&mov	($v1,$s[0]);				# copy s0
+	&mov	(&DWP(4,"esp"),$s[1]);			# save s1
+
+	&movz	($s[2],&HB($s[0]));
+	&and	($s[0],0xFF);
+	&mov	($s[0],&DWP(1024*0,$te,$s[0],4));	# s0>>0
+	&shr	($v1,16);
+	&mov	($s[3],&DWP(1024*1,$te,$s[2],4));	# s0>>8
+	&movz	($s[1],&HB($v1));
+	&and	($v1,0xFF);
+	&mov	($s[2],&DWP(1024*2,$te,$v1,4));		# s0>>16
+	 &mov	($v1,$v0);
+	&mov	($s[1],&DWP(1024*3,$te,$s[1],4));	# s0>>24
+
+	&and	($v0,0xFF);
+	&xor	($s[3],&DWP(1024*0,$te,$v0,4));		# s3>>0
+	&movz	($v0,&HB($v1));
+	&shr	($v1,16);
+	&xor	($s[2],&DWP(1024*1,$te,$v0,4));		# s3>>8
+	&movz	($v0,&HB($v1));
+	&and	($v1,0xFF);
+	&xor	($s[1],&DWP(1024*2,$te,$v1,4));		# s3>>16
+	 &mov	($v1,&DWP(0,"esp"));			# restore s2
+	&xor	($s[0],&DWP(1024*3,$te,$v0,4));		# s3>>24
+
+	&mov	($v0,$v1);
+	&and	($v1,0xFF);
+	&xor	($s[2],&DWP(1024*0,$te,$v1,4));		# s2>>0
+	&movz	($v1,&HB($v0));
+	&shr	($v0,16);
+	&xor	($s[1],&DWP(1024*1,$te,$v1,4));		# s2>>8
+	&movz	($v1,&HB($v0));
+	&and	($v0,0xFF);
+	&xor	($s[0],&DWP(1024*2,$te,$v0,4));		# s2>>16
+	 &mov	($v0,&DWP(4,"esp"));			# restore s1
+	&xor	($s[3],&DWP(1024*3,$te,$v1,4));		# s2>>24
+
+	&mov	($v1,$v0);
+	&and	($v0,0xFF);
+	&xor	($s[1],&DWP(1024*0,$te,$v0,4));		# s1>>0
+	&movz	($v0,&HB($v1));
+	&shr	($v1,16);
+	&xor	($s[0],&DWP(1024*1,$te,$v0,4));		# s1>>8
+	&movz	($v0,&HB($v1));
+	&and	($v1,0xFF);
+	&xor	($s[3],&DWP(1024*2,$te,$v1,4));		# s1>>16
+	 &mov	($key,&DWP(12,"esp"));			# reincarnate v1 as key
+	&xor	($s[2],&DWP(1024*3,$te,$v0,4));		# s1>>24
+}
 
 sub encstep()
 { my ($i,$te,@s) = @_;
-  my $tmp,$out;
+  my $tmp = $key;
+  my $out = $i==3?$s[0]:$acc;
 
-	# lines marked with ## denote same $sN...
-	if ($i==3)  {	&mov	("edi",&DWP(12,"esp"));
-			&movz	($out=$s[0],&HB($s[0]));	}	##
-	else        {	&mov	($out="esi",$s[0]);
-			&shr	($out,24);			}
+	# lines marked with #%e?x[i] denote "reordered" instructions...
+	if ($i==3)  {	&mov	($key,&DWP(12,"esp"));		}##%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(1024*0,$te,$out,4));
 
-	if ($i==2)  {	&movz	($tmp="edi",&LB($s[1]));	}	##
-	else        {	$i==3?$tmp=$s[1]:&mov($tmp="edi",$s[1]);
-			&shr	($tmp,16);
-			&and	($tmp,0xFF);			}
+	if ($i==3)  {	$tmp=$s[1];				}##%eax
+			&movz	($tmp,&HB($s[1]));
 			&xor	($out,&DWP(1024*1,$te,$tmp,4));
 
-	if ($i==3)  {	$tmp=$s[2]; &mov ($s[1],&DWP(0,"esp"));	}
-	else        {	$tmp="edi";				}
-			&movz	($tmp,&HB($s[2]));
+	if ($i==3)  {	$tmp=$s[2]; &mov ($s[1],&DWP(0,"esp"));	}##%ebx
+	else        {	&mov	($tmp,$s[2]);
+			&shr	($tmp,16);			}
+	if ($i==2)  {	&and	($s[1],0xFF);			}#%edx[2]
+			&and	($tmp,0xFF);
 			&xor	($out,&DWP(1024*2,$te,$tmp,4));
-	if ($i==1)  {	&shr	($s[2],16);			}	##
 
-	if ($i==3)  {	$tmp=$s[3]; &mov ($s[2],&DWP(4,"esp"));	}
-	else        {	&mov	($tmp="edi",$s[3]);		} 
-			&and	($tmp,0xFF);
+	if ($i==3)  {	$tmp=$s[3]; &mov ($s[2],&DWP(4,"esp"));	}##%ecx
+	elsif($i==2){	&movz	($tmp,&HB($s[3]));		}#%ebx[2]
+	else        {	&mov	($tmp,$s[3]); 
+			&shr	($tmp,24)			}
 			&xor	($out,&DWP(1024*3,$te,$tmp,4));
 	if ($i<2)   {	&mov	(&DWP(4*$i,"esp"),$out);	}
-	if ($i==3)  {	&mov	($s[3],"esi");			}
+	if ($i==3)  {	&mov	($s[3],$acc);			}
+			&comment();
 }
 
 sub enclast()
 { my ($i,$te,@s)=@_;
-  my $tmp,$out;
+  my $tmp = $key;
+  my $out = $i==3?$s[0]:$acc;
 
-	if ($i==3)  {	&mov	("edi",&DWP(12,"esp"));
-			&movz	($out=$s[0],&HB($s[0]));	}	##
-	else        {	&mov	($out="esi",$s[0]);
-			&shr	($out,24);			}
-			&mov	($out,&DWP(0,$te,$out,4));
-			&and	($out,0xff000000);
+	if ($i==3)  {	&mov	($key,&DWP(12,"esp"));		}##%edx
+	else        {	&mov	($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	($out,&DWP(1024*0,$te,$out,4));
+			&and	($out,0x000000ff);
 
-	if ($i==2)  {	&movz	($tmp="edi",&LB($s[1]));	}	##
-	else        {	$i==3?$tmp=$s[1]:&mov($tmp="edi",$s[1]);
-			&shr	($tmp,16);
-			&and	($tmp,0xFF);			}
+	if ($i==3)  {	$tmp=$s[1];				}##%eax
+			&movz	($tmp,&HB($s[1]));
 			&mov	($tmp,&DWP(0,$te,$tmp,4));
-			&and	($tmp,0x00ff0000);
+			&and	($tmp,0x0000ff00);
 			&xor	($out,$tmp);
 
-	if ($i==3)  {	$tmp=$s[2]; &mov ($s[1],&DWP(0,"esp"));	}
-	else        {	$tmp="edi";				}
-			&movz	($tmp,&HB($s[2]));
+	if ($i==3)  {	$tmp=$s[2]; &mov ($s[1],&DWP(0,"esp"));	}##%ebx
+	else        {	mov	($tmp,$s[2]);
+			&shr	($tmp,16);			}
+	if ($i==2)  {	&and	($s[1],0xFF);			}#%edx[2]
+			&and	($tmp,0xFF);
 			&mov	($tmp,&DWP(0,$te,$tmp,4));
-			&and	($tmp,0x0000ff00);
-	if ($i==1)  {	&shr	($s[2],16);			}	##
+			&and	($tmp,0x00ff0000);
 			&xor	($out,$tmp);
 
-	if ($i==3)  {	$tmp=$s[3]; &mov ($s[2],&DWP(4,"esp"));	}
-	else        {	&mov	($tmp="edi",$s[3]);		} 
-			&and	($tmp,0xFF);
+	if ($i==3)  {	$tmp=$s[3]; &mov ($s[2],&DWP(4,"esp"));	}##%ecx
+	elsif($i==2){	&movz	($tmp,&HB($s[3]));		}#%ebx[2]
+	else        {	&mov	($tmp,$s[3]);
+			&shr	($tmp,24);			}
 			&mov	($tmp,&DWP(0,$te,$tmp,4));
-			&and	($tmp,0x000000ff);
+			&and	($tmp,0xff000000);
 			&xor	($out,$tmp);
 	if ($i<2)   {	&mov	(&DWP(4*$i,"esp"),$out);	}
-	if ($i==3)  {	&mov	($s[3],"esi");			}
+	if ($i==3)  {	&mov	($s[3],$acc);			}
 }
 
 # void AES_encrypt (const void *inp,void *out,const AES_KEY *key);
 &public_label("AES_Te");
 &function_begin("AES_encrypt");
-	&mov	("esi",&wparam(0));		# load inp
-	&mov	("edi",&wparam(2));		# load key
+	&mov	($acc,&wparam(0));		# load inp
+	&mov	($key,&wparam(2));		# load key
 
         &call   (&label("pic_point"));          # make it PIC!
-&set_label("pic_point");
+	&set_label("pic_point");
         &blindpop("ebp");
         &lea    ("ebp",&DWP(&label("AES_Te")."-".&label("pic_point"),"ebp"));
 
 	# allocate aligned stack frame
-	&mov	("eax","esp");
+	&mov	($s0,"esp");
 	&sub	("esp",20);
 	&and	("esp",-16);
 
-	&mov	(&DWP(12,"esp"),"edi");		# save key
-	&mov	(&DWP(16,"esp"),"eax");		# save %esp
-
-	&mov	($s0,&DWP(0,"esi"));		# load input data
-	&mov	($s1,&DWP(4,"esi"));
-	&mov	($s2,&DWP(8,"esi"));
-	&mov	($s3,&DWP(12,"esi"));
-	#
-	# It's perfectly possible to implement algorithm as
-	# little-endian and get rid of bswaps... It would give
-	# less than 1% performance improvement, so I judge it
-	# doesn't worth the trouble...
-	#
-	&bswap	($s0);
-	&bswap	($s1);
-	&bswap	($s2);
-	&bswap	($s3);
-	&xor	($s0,&DWP(0,"edi"));
-	&xor	($s1,&DWP(4,"edi"));
-	&xor	($s2,&DWP(8,"edi"));
-	&xor	($s3,&DWP(12,"edi"));
-
-	&mov	("esi",&DWP(240,"edi"));	# load key->rounds
+	&mov	(&DWP(12,"esp"),$key);		# save key
+	&mov	(&DWP(16,"esp"),$s0);		# save %esp
+
+	&mov	($s0,&DWP(0,$acc));		# load input data
+	&mov	($s1,&DWP(4,$acc));
+	&mov	($s2,&DWP(8,$acc));
+	&mov	($s3,&DWP(12,$acc));
+
+	&xor	($s0,&DWP(0,$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 ($small_footprint) {
-	    &lea	("esi",&DWP(-2,"esi","esi"));
-	    &lea	("esi",&DWP(0,"edi","esi",8));
-	    &mov	(&DWP(8,"esp"),"esi");	# end of key schedule
+	    &lea	($acc,&DWP(-2,$acc,$acc));
+	    &lea	($acc,&DWP(0,$key,$acc,8));
+	    &mov	(&DWP(8,"esp"),$acc);	# end of key schedule
 	    &align	(4);
 	    &set_label("loop");
-		&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);
-		&add	("edi",16);			# advance rd_key
-		&xor	($s0,&DWP(0,"edi"));
-		&xor	($s1,&DWP(4,"edi"));
-		&xor	($s2,&DWP(8,"edi"));
-		&xor	($s3,&DWP(12,"edi"));
-	    &cmp	("edi",&DWP(8,"esp"));
-	    &mov	(&DWP(12,"esp"),"edi");
+		if ($vertical_spin) {
+		    &encvert("ebp",$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,$s2);
+		}
+		&add	($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	($key,&DWP(8,"esp"));
+	    &mov	(&DWP(12,"esp"),$key);
 	    &jb		(&label("loop"));
 	}
 	else {
-	    &cmp	("esi",10);
+	    &cmp	($acc,10);
 	    &jle	(&label("10rounds"));
-	    &cmp	("esi",12);
+	    &cmp	($acc,12);
 	    &jle	(&label("12rounds"));
 
 	&set_label("14rounds");
 	    for ($i=1;$i<3;$i++) {
-		&encstep(0,"ebp",$s0,$s1,$s2,$s3);
-		&encstep(1,"ebp",$s1,$s2,$s3,$s0);
-		&encstep(2,"ebp",$s2,$s3,$s0,$s1);
-		&encstep(3,"ebp",$s3,$s0,$s1,$s2);
-		&xor	($s0,&DWP(16*$i+0,"edi"));
-		&xor	($s1,&DWP(16*$i+4,"edi"));
-		&xor	($s2,&DWP(16*$i+8,"edi"));
-		&xor	($s3,&DWP(16*$i+12,"edi"));
+		if ($vertical_spin) {
+		    &encvert("ebp",$s0,$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);
+		}
+		&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	("edi",32);
-	    &mov	(&DWP(12,"esp"),"edi");		# advance rd_key
+	    &add	($key,32);
+	    &mov	(&DWP(12,"esp"),$key);		# advance rd_key
 	&set_label("12rounds");
 	    for ($i=1;$i<3;$i++) {
-		&encstep(0,"ebp",$s0,$s1,$s2,$s3);
-		&encstep(1,"ebp",$s1,$s2,$s3,$s0);
-		&encstep(2,"ebp",$s2,$s3,$s0,$s1);
-		&encstep(3,"ebp",$s3,$s0,$s1,$s2);
-		&xor	($s0,&DWP(16*$i+0,"edi"));
-		&xor	($s1,&DWP(16*$i+4,"edi"));
-		&xor	($s2,&DWP(16*$i+8,"edi"));
-		&xor	($s3,&DWP(16*$i+12,"edi"));
+		if ($vertical_spin) {
+		    &encvert("ebp",$s0,$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);
+		}
+		&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	("edi",32);
-	    &mov	(&DWP(12,"esp"),"edi");		# advance rd_key
+	    &add	($key,32);
+	    &mov	(&DWP(12,"esp"),$key);		# advance rd_key
 	&set_label("10rounds");
 	    for ($i=1;$i<10;$i++) {
-		&encstep(0,"ebp",$s0,$s1,$s2,$s3);
-		&encstep(1,"ebp",$s1,$s2,$s3,$s0);
-		&encstep(2,"ebp",$s2,$s3,$s0,$s1);
-		&encstep(3,"ebp",$s3,$s0,$s1,$s2);
-		&xor	($s0,&DWP(16*$i+0,"edi"));
-		&xor	($s1,&DWP(16*$i+4,"edi"));
-		&xor	($s2,&DWP(16*$i+8,"edi"));
-		&xor	($s3,&DWP(16*$i+12,"edi"));
+		if ($vertical_spin) {
+		    &encvert("ebp",$s0,$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);
+		}
+		&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	("ebp",4*1024);			# skip to Te4
+	if ($vertical_spin) {
+	    # "reincarnate" some registers for "horizontal" spin...
+	    &mov	($s1="ebx",$key="esi");
+	    &mov	($s2="ecx",$acc="edi");
+	}
 	&enclast(0,"ebp",$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);
 
 	&mov	("esp",&DWP(16,"esp"));		# restore %esp
-	&add	("edi",$small_footprint?16:160);
-	&xor	($s0,&DWP(0,"edi"));
-	&xor	($s1,&DWP(4,"edi"));
-	&xor	($s2,&DWP(8,"edi"));
-	&xor	($s3,&DWP(12,"edi"));
-	&bswap	($s0);
-	&bswap	($s1);
-	&bswap	($s2);
-	&bswap	($s3);
-	&mov	("edi",&wparam(1));		# load out
-	&mov	(&DWP(0,"edi"),$s0);		# write output data
-	&mov	(&DWP(4,"edi"),$s1);
-	&mov	(&DWP(8,"edi"),$s2);
-	&mov	(&DWP(12,"edi"),$s3);
+	&add	($key,$small_footprint?16:160);
+	&xor	($s0,&DWP(0,$key));
+	&xor	($s1,&DWP(4,$key));
+	&xor	($s2,&DWP(8,$key));
+	&xor	($s3,&DWP(12,$key));
+
+	&mov	($acc,&wparam(1));		# load out
+	&mov	(&DWP(0,$acc),$s0);		# write output data
+	&mov	(&DWP(4,$acc),$s1);
+	&mov	(&DWP(8,$acc),$s2);
+	&mov	(&DWP(12,$acc),$s3);
 
 	&pop	("edi");
 	&pop	("esi");
@@ -244,265 +364,265 @@ sub enclast()
 	&ret	();
 
 &set_label("AES_Te",64);	# Yes! I keep it in the code segment!
-	&data_word(0xc66363a5, 0xf87c7c84, 0xee777799, 0xf67b7b8d);
-	&data_word(0xfff2f20d, 0xd66b6bbd, 0xde6f6fb1, 0x91c5c554);
-	&data_word(0x60303050, 0x02010103, 0xce6767a9, 0x562b2b7d);
-	&data_word(0xe7fefe19, 0xb5d7d762, 0x4dababe6, 0xec76769a);
-	&data_word(0x8fcaca45, 0x1f82829d, 0x89c9c940, 0xfa7d7d87);
-	&data_word(0xeffafa15, 0xb25959eb, 0x8e4747c9, 0xfbf0f00b);
-	&data_word(0x41adadec, 0xb3d4d467, 0x5fa2a2fd, 0x45afafea);
-	&data_word(0x239c9cbf, 0x53a4a4f7, 0xe4727296, 0x9bc0c05b);
-	&data_word(0x75b7b7c2, 0xe1fdfd1c, 0x3d9393ae, 0x4c26266a);
-	&data_word(0x6c36365a, 0x7e3f3f41, 0xf5f7f702, 0x83cccc4f);
-	&data_word(0x6834345c, 0x51a5a5f4, 0xd1e5e534, 0xf9f1f108);
-	&data_word(0xe2717193, 0xabd8d873, 0x62313153, 0x2a15153f);
-	&data_word(0x0804040c, 0x95c7c752, 0x46232365, 0x9dc3c35e);
-	&data_word(0x30181828, 0x379696a1, 0x0a05050f, 0x2f9a9ab5);
-	&data_word(0x0e070709, 0x24121236, 0x1b80809b, 0xdfe2e23d);
-	&data_word(0xcdebeb26, 0x4e272769, 0x7fb2b2cd, 0xea75759f);
-	&data_word(0x1209091b, 0x1d83839e, 0x582c2c74, 0x341a1a2e);
-	&data_word(0x361b1b2d, 0xdc6e6eb2, 0xb45a5aee, 0x5ba0a0fb);
-	&data_word(0xa45252f6, 0x763b3b4d, 0xb7d6d661, 0x7db3b3ce);
-	&data_word(0x5229297b, 0xdde3e33e, 0x5e2f2f71, 0x13848497);
-	&data_word(0xa65353f5, 0xb9d1d168, 0x00000000, 0xc1eded2c);
-	&data_word(0x40202060, 0xe3fcfc1f, 0x79b1b1c8, 0xb65b5bed);
-	&data_word(0xd46a6abe, 0x8dcbcb46, 0x67bebed9, 0x7239394b);
-	&data_word(0x944a4ade, 0x984c4cd4, 0xb05858e8, 0x85cfcf4a);
-	&data_word(0xbbd0d06b, 0xc5efef2a, 0x4faaaae5, 0xedfbfb16);
-	&data_word(0x864343c5, 0x9a4d4dd7, 0x66333355, 0x11858594);
-	&data_word(0x8a4545cf, 0xe9f9f910, 0x04020206, 0xfe7f7f81);
-	&data_word(0xa05050f0, 0x783c3c44, 0x259f9fba, 0x4ba8a8e3);
-	&data_word(0xa25151f3, 0x5da3a3fe, 0x804040c0, 0x058f8f8a);
-	&data_word(0x3f9292ad, 0x219d9dbc, 0x70383848, 0xf1f5f504);
-	&data_word(0x63bcbcdf, 0x77b6b6c1, 0xafdada75, 0x42212163);
-	&data_word(0x20101030, 0xe5ffff1a, 0xfdf3f30e, 0xbfd2d26d);
-	&data_word(0x81cdcd4c, 0x180c0c14, 0x26131335, 0xc3ecec2f);
-	&data_word(0xbe5f5fe1, 0x359797a2, 0x884444cc, 0x2e171739);
-	&data_word(0x93c4c457, 0x55a7a7f2, 0xfc7e7e82, 0x7a3d3d47);
-	&data_word(0xc86464ac, 0xba5d5de7, 0x3219192b, 0xe6737395);
-	&data_word(0xc06060a0, 0x19818198, 0x9e4f4fd1, 0xa3dcdc7f);
-	&data_word(0x44222266, 0x542a2a7e, 0x3b9090ab, 0x0b888883);
-	&data_word(0x8c4646ca, 0xc7eeee29, 0x6bb8b8d3, 0x2814143c);
-	&data_word(0xa7dede79, 0xbc5e5ee2, 0x160b0b1d, 0xaddbdb76);
-	&data_word(0xdbe0e03b, 0x64323256, 0x743a3a4e, 0x140a0a1e);
-	&data_word(0x924949db, 0x0c06060a, 0x4824246c, 0xb85c5ce4);
-	&data_word(0x9fc2c25d, 0xbdd3d36e, 0x43acacef, 0xc46262a6);
-	&data_word(0x399191a8, 0x319595a4, 0xd3e4e437, 0xf279798b);
-	&data_word(0xd5e7e732, 0x8bc8c843, 0x6e373759, 0xda6d6db7);
-	&data_word(0x018d8d8c, 0xb1d5d564, 0x9c4e4ed2, 0x49a9a9e0);
-	&data_word(0xd86c6cb4, 0xac5656fa, 0xf3f4f407, 0xcfeaea25);
-	&data_word(0xca6565af, 0xf47a7a8e, 0x47aeaee9, 0x10080818);
-	&data_word(0x6fbabad5, 0xf0787888, 0x4a25256f, 0x5c2e2e72);
-	&data_word(0x381c1c24, 0x57a6a6f1, 0x73b4b4c7, 0x97c6c651);
-	&data_word(0xcbe8e823, 0xa1dddd7c, 0xe874749c, 0x3e1f1f21);
-	&data_word(0x964b4bdd, 0x61bdbddc, 0x0d8b8b86, 0x0f8a8a85);
-	&data_word(0xe0707090, 0x7c3e3e42, 0x71b5b5c4, 0xcc6666aa);
-	&data_word(0x904848d8, 0x06030305, 0xf7f6f601, 0x1c0e0e12);
-	&data_word(0xc26161a3, 0x6a35355f, 0xae5757f9, 0x69b9b9d0);
-	&data_word(0x17868691, 0x99c1c158, 0x3a1d1d27, 0x279e9eb9);
-	&data_word(0xd9e1e138, 0xebf8f813, 0x2b9898b3, 0x22111133);
-	&data_word(0xd26969bb, 0xa9d9d970, 0x078e8e89, 0x339494a7);
-	&data_word(0x2d9b9bb6, 0x3c1e1e22, 0x15878792, 0xc9e9e920);
-	&data_word(0x87cece49, 0xaa5555ff, 0x50282878, 0xa5dfdf7a);
-	&data_word(0x038c8c8f, 0x59a1a1f8, 0x09898980, 0x1a0d0d17);
-	&data_word(0x65bfbfda, 0xd7e6e631, 0x844242c6, 0xd06868b8);
-	&data_word(0x824141c3, 0x299999b0, 0x5a2d2d77, 0x1e0f0f11);
-	&data_word(0x7bb0b0cb, 0xa85454fc, 0x6dbbbbd6, 0x2c16163a);
+	&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);
+	&data_word(0xd5baba6f, 0x887878f0, 0x6f25254a, 0x722e2e5c);
+	&data_word(0x241c1c38, 0xf1a6a657, 0xc7b4b473, 0x51c6c697);
+	&data_word(0x23e8e8cb, 0x7cdddda1, 0x9c7474e8, 0x211f1f3e);
+	&data_word(0xdd4b4b96, 0xdcbdbd61, 0x868b8b0d, 0x858a8a0f);
+	&data_word(0x907070e0, 0x423e3e7c, 0xc4b5b571, 0xaa6666cc);
+	&data_word(0xd8484890, 0x05030306, 0x01f6f6f7, 0x120e0e1c);
+	&data_word(0xa36161c2, 0x5f35356a, 0xf95757ae, 0xd0b9b969);
+	&data_word(0x91868617, 0x58c1c199, 0x271d1d3a, 0xb99e9e27);
+	&data_word(0x38e1e1d9, 0x13f8f8eb, 0xb398982b, 0x33111122);
+	&data_word(0xbb6969d2, 0x70d9d9a9, 0x898e8e07, 0xa7949433);
+	&data_word(0xb69b9b2d, 0x221e1e3c, 0x92878715, 0x20e9e9c9);
+	&data_word(0x49cece87, 0xff5555aa, 0x78282850, 0x7adfdfa5);
+	&data_word(0x8f8c8c03, 0xf8a1a159, 0x80898909, 0x170d0d1a);
+	&data_word(0xdabfbf65, 0x31e6e6d7, 0xc6424284, 0xb86868d0);
+	&data_word(0xc3414182, 0xb0999929, 0x772d2d5a, 0x110f0f1e);
+	&data_word(0xcbb0b07b, 0xfc5454a8, 0xd6bbbb6d, 0x3a16162c);
 #Te1:
-	&data_word(0xa5c66363, 0x84f87c7c, 0x99ee7777, 0x8df67b7b);
-	&data_word(0x0dfff2f2, 0xbdd66b6b, 0xb1de6f6f, 0x5491c5c5);
-	&data_word(0x50603030, 0x03020101, 0xa9ce6767, 0x7d562b2b);
-	&data_word(0x19e7fefe, 0x62b5d7d7, 0xe64dabab, 0x9aec7676);
-	&data_word(0x458fcaca, 0x9d1f8282, 0x4089c9c9, 0x87fa7d7d);
-	&data_word(0x15effafa, 0xebb25959, 0xc98e4747, 0x0bfbf0f0);
-	&data_word(0xec41adad, 0x67b3d4d4, 0xfd5fa2a2, 0xea45afaf);
-	&data_word(0xbf239c9c, 0xf753a4a4, 0x96e47272, 0x5b9bc0c0);
-	&data_word(0xc275b7b7, 0x1ce1fdfd, 0xae3d9393, 0x6a4c2626);
-	&data_word(0x5a6c3636, 0x417e3f3f, 0x02f5f7f7, 0x4f83cccc);
-	&data_word(0x5c683434, 0xf451a5a5, 0x34d1e5e5, 0x08f9f1f1);
-	&data_word(0x93e27171, 0x73abd8d8, 0x53623131, 0x3f2a1515);
-	&data_word(0x0c080404, 0x5295c7c7, 0x65462323, 0x5e9dc3c3);
-	&data_word(0x28301818, 0xa1379696, 0x0f0a0505, 0xb52f9a9a);
-	&data_word(0x090e0707, 0x36241212, 0x9b1b8080, 0x3ddfe2e2);
-	&data_word(0x26cdebeb, 0x694e2727, 0xcd7fb2b2, 0x9fea7575);
-	&data_word(0x1b120909, 0x9e1d8383, 0x74582c2c, 0x2e341a1a);
-	&data_word(0x2d361b1b, 0xb2dc6e6e, 0xeeb45a5a, 0xfb5ba0a0);
-	&data_word(0xf6a45252, 0x4d763b3b, 0x61b7d6d6, 0xce7db3b3);
-	&data_word(0x7b522929, 0x3edde3e3, 0x715e2f2f, 0x97138484);
-	&data_word(0xf5a65353, 0x68b9d1d1, 0x00000000, 0x2cc1eded);
-	&data_word(0x60402020, 0x1fe3fcfc, 0xc879b1b1, 0xedb65b5b);
-	&data_word(0xbed46a6a, 0x468dcbcb, 0xd967bebe, 0x4b723939);
-	&data_word(0xde944a4a, 0xd4984c4c, 0xe8b05858, 0x4a85cfcf);
-	&data_word(0x6bbbd0d0, 0x2ac5efef, 0xe54faaaa, 0x16edfbfb);
-	&data_word(0xc5864343, 0xd79a4d4d, 0x55663333, 0x94118585);
-	&data_word(0xcf8a4545, 0x10e9f9f9, 0x06040202, 0x81fe7f7f);
-	&data_word(0xf0a05050, 0x44783c3c, 0xba259f9f, 0xe34ba8a8);
-	&data_word(0xf3a25151, 0xfe5da3a3, 0xc0804040, 0x8a058f8f);
-	&data_word(0xad3f9292, 0xbc219d9d, 0x48703838, 0x04f1f5f5);
-	&data_word(0xdf63bcbc, 0xc177b6b6, 0x75afdada, 0x63422121);
-	&data_word(0x30201010, 0x1ae5ffff, 0x0efdf3f3, 0x6dbfd2d2);
-	&data_word(0x4c81cdcd, 0x14180c0c, 0x35261313, 0x2fc3ecec);
-	&data_word(0xe1be5f5f, 0xa2359797, 0xcc884444, 0x392e1717);
-	&data_word(0x5793c4c4, 0xf255a7a7, 0x82fc7e7e, 0x477a3d3d);
-	&data_word(0xacc86464, 0xe7ba5d5d, 0x2b321919, 0x95e67373);
-	&data_word(0xa0c06060, 0x98198181, 0xd19e4f4f, 0x7fa3dcdc);
-	&data_word(0x66442222, 0x7e542a2a, 0xab3b9090, 0x830b8888);
-	&data_word(0xca8c4646, 0x29c7eeee, 0xd36bb8b8, 0x3c281414);
-	&data_word(0x79a7dede, 0xe2bc5e5e, 0x1d160b0b, 0x76addbdb);
-	&data_word(0x3bdbe0e0, 0x56643232, 0x4e743a3a, 0x1e140a0a);
-	&data_word(0xdb924949, 0x0a0c0606, 0x6c482424, 0xe4b85c5c);
-	&data_word(0x5d9fc2c2, 0x6ebdd3d3, 0xef43acac, 0xa6c46262);
-	&data_word(0xa8399191, 0xa4319595, 0x37d3e4e4, 0x8bf27979);
-	&data_word(0x32d5e7e7, 0x438bc8c8, 0x596e3737, 0xb7da6d6d);
-	&data_word(0x8c018d8d, 0x64b1d5d5, 0xd29c4e4e, 0xe049a9a9);
-	&data_word(0xb4d86c6c, 0xfaac5656, 0x07f3f4f4, 0x25cfeaea);
-	&data_word(0xafca6565, 0x8ef47a7a, 0xe947aeae, 0x18100808);
-	&data_word(0xd56fbaba, 0x88f07878, 0x6f4a2525, 0x725c2e2e);
-	&data_word(0x24381c1c, 0xf157a6a6, 0xc773b4b4, 0x5197c6c6);
-	&data_word(0x23cbe8e8, 0x7ca1dddd, 0x9ce87474, 0x213e1f1f);
-	&data_word(0xdd964b4b, 0xdc61bdbd, 0x860d8b8b, 0x850f8a8a);
-	&data_word(0x90e07070, 0x427c3e3e, 0xc471b5b5, 0xaacc6666);
-	&data_word(0xd8904848, 0x05060303, 0x01f7f6f6, 0x121c0e0e);
-	&data_word(0xa3c26161, 0x5f6a3535, 0xf9ae5757, 0xd069b9b9);
-	&data_word(0x91178686, 0x5899c1c1, 0x273a1d1d, 0xb9279e9e);
-	&data_word(0x38d9e1e1, 0x13ebf8f8, 0xb32b9898, 0x33221111);
-	&data_word(0xbbd26969, 0x70a9d9d9, 0x89078e8e, 0xa7339494);
-	&data_word(0xb62d9b9b, 0x223c1e1e, 0x92158787, 0x20c9e9e9);
-	&data_word(0x4987cece, 0xffaa5555, 0x78502828, 0x7aa5dfdf);
-	&data_word(0x8f038c8c, 0xf859a1a1, 0x80098989, 0x171a0d0d);
-	&data_word(0xda65bfbf, 0x31d7e6e6, 0xc6844242, 0xb8d06868);
-	&data_word(0xc3824141, 0xb0299999, 0x775a2d2d, 0x111e0f0f);
-	&data_word(0xcb7bb0b0, 0xfca85454, 0xd66dbbbb, 0x3a2c1616);
+	&data_word(0x6363c6a5, 0x7c7cf884, 0x7777ee99, 0x7b7bf68d);
+	&data_word(0xf2f2ff0d, 0x6b6bd6bd, 0x6f6fdeb1, 0xc5c59154);
+	&data_word(0x30306050, 0x01010203, 0x6767cea9, 0x2b2b567d);
+	&data_word(0xfefee719, 0xd7d7b562, 0xabab4de6, 0x7676ec9a);
+	&data_word(0xcaca8f45, 0x82821f9d, 0xc9c98940, 0x7d7dfa87);
+	&data_word(0xfafaef15, 0x5959b2eb, 0x47478ec9, 0xf0f0fb0b);
+	&data_word(0xadad41ec, 0xd4d4b367, 0xa2a25ffd, 0xafaf45ea);
+	&data_word(0x9c9c23bf, 0xa4a453f7, 0x7272e496, 0xc0c09b5b);
+	&data_word(0xb7b775c2, 0xfdfde11c, 0x93933dae, 0x26264c6a);
+	&data_word(0x36366c5a, 0x3f3f7e41, 0xf7f7f502, 0xcccc834f);
+	&data_word(0x3434685c, 0xa5a551f4, 0xe5e5d134, 0xf1f1f908);
+	&data_word(0x7171e293, 0xd8d8ab73, 0x31316253, 0x15152a3f);
+	&data_word(0x0404080c, 0xc7c79552, 0x23234665, 0xc3c39d5e);
+	&data_word(0x18183028, 0x969637a1, 0x05050a0f, 0x9a9a2fb5);
+	&data_word(0x07070e09, 0x12122436, 0x80801b9b, 0xe2e2df3d);
+	&data_word(0xebebcd26, 0x27274e69, 0xb2b27fcd, 0x7575ea9f);
+	&data_word(0x0909121b, 0x83831d9e, 0x2c2c5874, 0x1a1a342e);
+	&data_word(0x1b1b362d, 0x6e6edcb2, 0x5a5ab4ee, 0xa0a05bfb);
+	&data_word(0x5252a4f6, 0x3b3b764d, 0xd6d6b761, 0xb3b37dce);
+	&data_word(0x2929527b, 0xe3e3dd3e, 0x2f2f5e71, 0x84841397);
+	&data_word(0x5353a6f5, 0xd1d1b968, 0x00000000, 0xededc12c);
+	&data_word(0x20204060, 0xfcfce31f, 0xb1b179c8, 0x5b5bb6ed);
+	&data_word(0x6a6ad4be, 0xcbcb8d46, 0xbebe67d9, 0x3939724b);
+	&data_word(0x4a4a94de, 0x4c4c98d4, 0x5858b0e8, 0xcfcf854a);
+	&data_word(0xd0d0bb6b, 0xefefc52a, 0xaaaa4fe5, 0xfbfbed16);
+	&data_word(0x434386c5, 0x4d4d9ad7, 0x33336655, 0x85851194);
+	&data_word(0x45458acf, 0xf9f9e910, 0x02020406, 0x7f7ffe81);
+	&data_word(0x5050a0f0, 0x3c3c7844, 0x9f9f25ba, 0xa8a84be3);
+	&data_word(0x5151a2f3, 0xa3a35dfe, 0x404080c0, 0x8f8f058a);
+	&data_word(0x92923fad, 0x9d9d21bc, 0x38387048, 0xf5f5f104);
+	&data_word(0xbcbc63df, 0xb6b677c1, 0xdadaaf75, 0x21214263);
+	&data_word(0x10102030, 0xffffe51a, 0xf3f3fd0e, 0xd2d2bf6d);
+	&data_word(0xcdcd814c, 0x0c0c1814, 0x13132635, 0xececc32f);
+	&data_word(0x5f5fbee1, 0x979735a2, 0x444488cc, 0x17172e39);
+	&data_word(0xc4c49357, 0xa7a755f2, 0x7e7efc82, 0x3d3d7a47);
+	&data_word(0x6464c8ac, 0x5d5dbae7, 0x1919322b, 0x7373e695);
+	&data_word(0x6060c0a0, 0x81811998, 0x4f4f9ed1, 0xdcdca37f);
+	&data_word(0x22224466, 0x2a2a547e, 0x90903bab, 0x88880b83);
+	&data_word(0x46468cca, 0xeeeec729, 0xb8b86bd3, 0x1414283c);
+	&data_word(0xdedea779, 0x5e5ebce2, 0x0b0b161d, 0xdbdbad76);
+	&data_word(0xe0e0db3b, 0x32326456, 0x3a3a744e, 0x0a0a141e);
+	&data_word(0x494992db, 0x06060c0a, 0x2424486c, 0x5c5cb8e4);
+	&data_word(0xc2c29f5d, 0xd3d3bd6e, 0xacac43ef, 0x6262c4a6);
+	&data_word(0x919139a8, 0x959531a4, 0xe4e4d337, 0x7979f28b);
+	&data_word(0xe7e7d532, 0xc8c88b43, 0x37376e59, 0x6d6ddab7);
+	&data_word(0x8d8d018c, 0xd5d5b164, 0x4e4e9cd2, 0xa9a949e0);
+	&data_word(0x6c6cd8b4, 0x5656acfa, 0xf4f4f307, 0xeaeacf25);
+	&data_word(0x6565caaf, 0x7a7af48e, 0xaeae47e9, 0x08081018);
+	&data_word(0xbaba6fd5, 0x7878f088, 0x25254a6f, 0x2e2e5c72);
+	&data_word(0x1c1c3824, 0xa6a657f1, 0xb4b473c7, 0xc6c69751);
+	&data_word(0xe8e8cb23, 0xdddda17c, 0x7474e89c, 0x1f1f3e21);
+	&data_word(0x4b4b96dd, 0xbdbd61dc, 0x8b8b0d86, 0x8a8a0f85);
+	&data_word(0x7070e090, 0x3e3e7c42, 0xb5b571c4, 0x6666ccaa);
+	&data_word(0x484890d8, 0x03030605, 0xf6f6f701, 0x0e0e1c12);
+	&data_word(0x6161c2a3, 0x35356a5f, 0x5757aef9, 0xb9b969d0);
+	&data_word(0x86861791, 0xc1c19958, 0x1d1d3a27, 0x9e9e27b9);
+	&data_word(0xe1e1d938, 0xf8f8eb13, 0x98982bb3, 0x11112233);
+	&data_word(0x6969d2bb, 0xd9d9a970, 0x8e8e0789, 0x949433a7);
+	&data_word(0x9b9b2db6, 0x1e1e3c22, 0x87871592, 0xe9e9c920);
+	&data_word(0xcece8749, 0x5555aaff, 0x28285078, 0xdfdfa57a);
+	&data_word(0x8c8c038f, 0xa1a159f8, 0x89890980, 0x0d0d1a17);
+	&data_word(0xbfbf65da, 0xe6e6d731, 0x424284c6, 0x6868d0b8);
+	&data_word(0x414182c3, 0x999929b0, 0x2d2d5a77, 0x0f0f1e11);
+	&data_word(0xb0b07bcb, 0x5454a8fc, 0xbbbb6dd6, 0x16162c3a);
 #Te2:
-	&data_word(0x63a5c663, 0x7c84f87c, 0x7799ee77, 0x7b8df67b);
-	&data_word(0xf20dfff2, 0x6bbdd66b, 0x6fb1de6f, 0xc55491c5);
-	&data_word(0x30506030, 0x01030201, 0x67a9ce67, 0x2b7d562b);
-	&data_word(0xfe19e7fe, 0xd762b5d7, 0xabe64dab, 0x769aec76);
-	&data_word(0xca458fca, 0x829d1f82, 0xc94089c9, 0x7d87fa7d);
-	&data_word(0xfa15effa, 0x59ebb259, 0x47c98e47, 0xf00bfbf0);
-	&data_word(0xadec41ad, 0xd467b3d4, 0xa2fd5fa2, 0xafea45af);
-	&data_word(0x9cbf239c, 0xa4f753a4, 0x7296e472, 0xc05b9bc0);
-	&data_word(0xb7c275b7, 0xfd1ce1fd, 0x93ae3d93, 0x266a4c26);
-	&data_word(0x365a6c36, 0x3f417e3f, 0xf702f5f7, 0xcc4f83cc);
-	&data_word(0x345c6834, 0xa5f451a5, 0xe534d1e5, 0xf108f9f1);
-	&data_word(0x7193e271, 0xd873abd8, 0x31536231, 0x153f2a15);
-	&data_word(0x040c0804, 0xc75295c7, 0x23654623, 0xc35e9dc3);
-	&data_word(0x18283018, 0x96a13796, 0x050f0a05, 0x9ab52f9a);
-	&data_word(0x07090e07, 0x12362412, 0x809b1b80, 0xe23ddfe2);
-	&data_word(0xeb26cdeb, 0x27694e27, 0xb2cd7fb2, 0x759fea75);
-	&data_word(0x091b1209, 0x839e1d83, 0x2c74582c, 0x1a2e341a);
-	&data_word(0x1b2d361b, 0x6eb2dc6e, 0x5aeeb45a, 0xa0fb5ba0);
-	&data_word(0x52f6a452, 0x3b4d763b, 0xd661b7d6, 0xb3ce7db3);
-	&data_word(0x297b5229, 0xe33edde3, 0x2f715e2f, 0x84971384);
-	&data_word(0x53f5a653, 0xd168b9d1, 0x00000000, 0xed2cc1ed);
-	&data_word(0x20604020, 0xfc1fe3fc, 0xb1c879b1, 0x5bedb65b);
-	&data_word(0x6abed46a, 0xcb468dcb, 0xbed967be, 0x394b7239);
-	&data_word(0x4ade944a, 0x4cd4984c, 0x58e8b058, 0xcf4a85cf);
-	&data_word(0xd06bbbd0, 0xef2ac5ef, 0xaae54faa, 0xfb16edfb);
-	&data_word(0x43c58643, 0x4dd79a4d, 0x33556633, 0x85941185);
-	&data_word(0x45cf8a45, 0xf910e9f9, 0x02060402, 0x7f81fe7f);
-	&data_word(0x50f0a050, 0x3c44783c, 0x9fba259f, 0xa8e34ba8);
-	&data_word(0x51f3a251, 0xa3fe5da3, 0x40c08040, 0x8f8a058f);
-	&data_word(0x92ad3f92, 0x9dbc219d, 0x38487038, 0xf504f1f5);
-	&data_word(0xbcdf63bc, 0xb6c177b6, 0xda75afda, 0x21634221);
-	&data_word(0x10302010, 0xff1ae5ff, 0xf30efdf3, 0xd26dbfd2);
-	&data_word(0xcd4c81cd, 0x0c14180c, 0x13352613, 0xec2fc3ec);
-	&data_word(0x5fe1be5f, 0x97a23597, 0x44cc8844, 0x17392e17);
-	&data_word(0xc45793c4, 0xa7f255a7, 0x7e82fc7e, 0x3d477a3d);
-	&data_word(0x64acc864, 0x5de7ba5d, 0x192b3219, 0x7395e673);
-	&data_word(0x60a0c060, 0x81981981, 0x4fd19e4f, 0xdc7fa3dc);
-	&data_word(0x22664422, 0x2a7e542a, 0x90ab3b90, 0x88830b88);
-	&data_word(0x46ca8c46, 0xee29c7ee, 0xb8d36bb8, 0x143c2814);
-	&data_word(0xde79a7de, 0x5ee2bc5e, 0x0b1d160b, 0xdb76addb);
-	&data_word(0xe03bdbe0, 0x32566432, 0x3a4e743a, 0x0a1e140a);
-	&data_word(0x49db9249, 0x060a0c06, 0x246c4824, 0x5ce4b85c);
-	&data_word(0xc25d9fc2, 0xd36ebdd3, 0xacef43ac, 0x62a6c462);
-	&data_word(0x91a83991, 0x95a43195, 0xe437d3e4, 0x798bf279);
-	&data_word(0xe732d5e7, 0xc8438bc8, 0x37596e37, 0x6db7da6d);
-	&data_word(0x8d8c018d, 0xd564b1d5, 0x4ed29c4e, 0xa9e049a9);
-	&data_word(0x6cb4d86c, 0x56faac56, 0xf407f3f4, 0xea25cfea);
-	&data_word(0x65afca65, 0x7a8ef47a, 0xaee947ae, 0x08181008);
-	&data_word(0xbad56fba, 0x7888f078, 0x256f4a25, 0x2e725c2e);
-	&data_word(0x1c24381c, 0xa6f157a6, 0xb4c773b4, 0xc65197c6);
-	&data_word(0xe823cbe8, 0xdd7ca1dd, 0x749ce874, 0x1f213e1f);
-	&data_word(0x4bdd964b, 0xbddc61bd, 0x