aes-586.pl: Atom-specific optimization, +44/29%, minor improvement on others.

vpaes-x86.pl: minor performance squeeze.

2 files changed, 180 insertions, 174 deletions

diff --git a/crypto/aes/asm/aes-586.pl b/crypto/aes/asm/aes-586.pl
index 406cd222f5..1c1e23e575 100755
--- a/crypto/aes/asm/aes-586.pl
+++ b/crypto/aes/asm/aes-586.pl
@@ -103,11 +103,12 @@
 # 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
+# P4		52[54]		83[95]		23
+# AMD K8	46[41]		66[70]		18
+# PIII		41[50]		60[77]		24
+# Core 2	31[36]		45[64]		18.5
+# Atom		76[100]		96[138]		60
+# Pentium	115		150		77
 #
 # Version 4.1 switches to compact S-box even in key schedule setup.
 #
@@ -476,24 +477,25 @@ sub enctransform()
   my $tmp = $tbl;
   my $r2  = $key ;
 
-	&mov	($acc,$s[$i]);
-	&and	($acc,0x80808080);
-	&mov	($tmp,$acc);
-	&shr	($tmp,7);
+	&and	($tmp,$s[$i]);
 	&lea	($r2,&DWP(0,$s[$i],$s[$i]));
-	&sub	($acc,$tmp);
+	&mov	($acc,$tmp);
+	&shr	($tmp,7);
 	&and	($r2,0xfefefefe);
-	&and	($acc,0x1b1b1b1b);
+	&sub	($acc,$tmp);
 	&mov	($tmp,$s[$i]);
+	&and	($acc,0x1b1b1b1b);
+	&rotr	($tmp,16);
 	&xor	($acc,$r2);	# r2
+	&mov	($r2,$s[$i]);
 
 	&xor	($s[$i],$acc);	# r0 ^ r2
+	&rotr	($r2,16+8);
+	&xor	($acc,$tmp);
 	&rotl	($s[$i],24);
+	&xor	($acc,$r2);
+	&mov	($tmp,0x80808080)	if ($i!=1);
 	&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");
@@ -526,6 +528,7 @@ sub enctransform()
 		&enccompact(1,$tbl,$s1,$s2,$s3,$s0,1);
 		&enccompact(2,$tbl,$s2,$s3,$s0,$s1,1);
 		&enccompact(3,$tbl,$s3,$s0,$s1,$s2,1);
+		&mov	($tbl,0x80808080);
 		&enctransform(2);
 		&enctransform(3);
 		&enctransform(0);
@@ -607,82 +610,84 @@ sub sse_enccompact()
 	&pshufw	("mm5","mm4",0x0d);		# 15,14,11,10
 	&movd	("eax","mm1");			#  5, 4, 1, 0
 	&movd	("ebx","mm5");			# 15,14,11,10
+	&mov	($__key,$key);
 
 	&movz	($acc,&LB("eax"));		#  0
-	&movz	("ecx",&BP(-128,$tbl,$acc,1));	#  0
-	&pshufw	("mm2","mm0",0x0d);		#  7, 6, 3, 2
 	&movz	("edx",&HB("eax"));		#  1
+	&pshufw	("mm2","mm0",0x0d);		#  7, 6, 3, 2
+	&movz	("ecx",&BP(-128,$tbl,$acc,1));	#  0
+	&movz	($key,&LB("ebx"));		# 10
 	&movz	("edx",&BP(-128,$tbl,"edx",1));	#  1
-	&shl	("edx",8);			#  1
 	&shr	("eax",16);			#  5, 4
+	&shl	("edx",8);			#  1
 
-	&movz	($acc,&LB("ebx"));		# 10
-	&movz	($acc,&BP(-128,$tbl,$acc,1));	# 10
+	&movz	($acc,&BP(-128,$tbl,$key,1));	# 10
+	&movz	($key,&HB("ebx"));		# 11
 	&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
+	&or	("ecx",$acc);			# 10
+	&movz	($acc,&BP(-128,$tbl,$key,1));	# 11
+	&movz	($key,&HB("eax"));		#  5
 	&shl	($acc,24);			# 11
-	&or	("edx",$acc);			# 11
 	&shr	("ebx",16);			# 15,14
+	&or	("edx",$acc);			# 11
 
-	&movz	($acc,&HB("eax"));		#  5
-	&movz	($acc,&BP(-128,$tbl,$acc,1));	#  5
+	&movz	($acc,&BP(-128,$tbl,$key,1));	#  5
+	&movz	($key,&HB("ebx"));		# 15
 	&shl	($acc,8);			#  5
 	&or	("ecx",$acc);			#  5
-	&movz	($acc,&HB("ebx"));		# 15
-	&movz	($acc,&BP(-128,$tbl,$acc,1));	# 15
+	&movz	($acc,&BP(-128,$tbl,$key,1));	# 15
+	&movz	($key,&LB("eax"));		#  4
 	&shl	($acc,24);			# 15
 	&or	("ecx",$acc);			# 15
-	&movd	("mm0","ecx");			# t[0] collected
 
-	&movz	($acc,&LB("eax"));		#  4
-	&movz	("ecx",&BP(-128,$tbl,$acc,1));	#  4
+	&movz	($acc,&BP(-128,$tbl,$key,1));	#  4
+	&movz	($key,&LB("ebx"));		# 14
 	&movd	("eax","mm2");			#  7, 6, 3, 2
-	&movz	($acc,&LB("ebx"));		# 14
-	&movz	($acc,&BP(-128,$tbl,$acc,1));	# 14
-	&shl	($acc,16);			# 14
+	&movd	("mm0","ecx");			# t[0] collected
+	&movz	("ecx",&BP(-128,$tbl,$key,1));	# 14
+	&movz	($key,&HB("eax"));		#  3
+	&shl	("ecx",16);			# 14
+	&movd	("ebx","mm6");			# 13,12, 9, 8
 	&or	("ecx",$acc);			# 14
 
-	&movd	("ebx","mm6");			# 13,12, 9, 8
-	&movz	($acc,&HB("eax"));		#  3
-	&movz	($acc,&BP(-128,$tbl,$acc,1));	#  3
+	&movz	($acc,&BP(-128,$tbl,$key,1));	#  3
+	&movz	($key,&HB("ebx"));		#  9
 	&shl	($acc,24);			#  3
 	&or	("ecx",$acc);			#  3
-	&movz	($acc,&HB("ebx"));		#  9
-	&movz	($acc,&BP(-128,$tbl,$acc,1));	#  9
+	&movz	($acc,&BP(-128,$tbl,$key,1));	#  9
+	&movz	($key,&LB("ebx"));		#  8
 	&shl	($acc,8);			#  9
+	&shr	("ebx",16);			# 13,12
 	&or	("ecx",$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",$acc);			#  2
+	&movz	($acc,&BP(-128,$tbl,$key,1));	#  8
+	&movz	($key,&LB("eax"));		#  2
 	&shr	("eax",16);			#  7, 6
+	&movd	("mm1","ecx");			# t[1] collected
+	&movz	("ecx",&BP(-128,$tbl,$key,1));	#  2
+	&movz	($key,&HB("eax"));		#  7
+	&shl	("ecx",16);			#  2
+	&and	("eax",0xff);			#  6
+	&or	("ecx",$acc);			#  2
 
 	&punpckldq	("mm0","mm1");		# t[0,1] collected
 
-	&movz	($acc,&HB("eax"));		#  7
-	&movz	($acc,&BP(-128,$tbl,$acc,1));	#  7
+	&movz	($acc,&BP(-128,$tbl,$key,1));	#  7
+	&movz	($key,&HB("ebx"));		# 13
 	&shl	($acc,24);			#  7
-	&or	("ecx",$acc);			#  7
-	&and	("eax",0xff);			#  6
+	&and	("ebx",0xff);			# 12
 	&movz	("eax",&BP(-128,$tbl,"eax",1));	#  6
+	&or	("ecx",$acc);			#  7
 	&shl	("eax",16);			#  6
+	&movz	($acc,&BP(-128,$tbl,$key,1));	# 13
 	&or	("edx","eax");			#  6
-	&movz	($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	("ecx",$acc);			# 13
 	&or	("edx","ebx");			# 12
+	&mov	($key,$__key);
+	&movd	("mm4","ecx");			# t[2] collected
 	&movd	("mm5","edx");			# t[3] collected
 
 	&punpckldq	("mm4","mm5");		# t[2,3] collected
@@ -1270,30 +1275,30 @@ sub dectransform()
   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);
+	&mov	($tmp,0x80808080);
+	&and	($tmp,$s[$i]);
+	&mov	($acc,$tmp);
 	&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);
+	&xor	($tp2,$acc);
+	&mov	($tmp,0x80808080);
 
-	&and	($acc,0x80808080);
-	&mov	($tmp,$acc);
+	&and	($tmp,$tp2);
+	&mov	($acc,$tmp);
 	&shr	($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);
+	&xor	($tp4,$acc);
+	&mov	($tmp,0x80808080);
 
-	&and	($acc,0x80808080);
-	&mov	($tmp,$acc);
+	&and	($tmp,$tp4);
+	&mov	($acc,$tmp);
 	&shr	($tmp,7);
 	&lea	($tp8,&DWP(0,$tp4,$tp4));
 	&sub	($acc,$tmp);
@@ -1305,13 +1310,13 @@ sub dectransform()
 
 	&xor	($s[$i],$tp2);
 	&xor	($tp2,$tp8);
-	&rotl	($tp2,24);
 	&xor	($s[$i],$tp4);
 	&xor	($tp4,$tp8);
-	&rotl	($tp4,16);
+	&rotl	($tp2,24);
 	&xor	($s[$i],$tp8);	# ^= tp8^(tp4^tp1)^(tp2^tp1)
-	&rotl	($tp8,8);
+	&rotl	($tp4,16);
 	&xor	($s[$i],$tp2);	# ^= ROTATE(tp8^tp2^tp1,24)
+	&rotl	($tp8,8);
 	&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 $s1
@@ -1389,85 +1394,87 @@ sub dectransform()
 sub sse_deccompact()
 {
 	&pshufw	("mm1","mm0",0x0c);		#  7, 6, 1, 0
+	&pshufw	("mm5","mm4",0x09);		# 13,12,11,10
 	&movd	("eax","mm1");			#  7, 6, 1, 0
+	&movd	("ebx","mm5");			# 13,12,11,10
+	&mov	($__key,$key);
 
-	&pshufw	("mm5","mm4",0x09);		# 13,12,11,10
 	&movz	($acc,&LB("eax"));		#  0
-	&movz	("ecx",&BP(-128,$tbl,$acc,1));	#  0
-	&movd	("ebx","mm5");			# 13,12,11,10
 	&movz	("edx",&HB("eax"));		#  1
+	&pshufw	("mm2","mm0",0x06);		#  3, 2, 5, 4
+	&movz	("ecx",&BP(-128,$tbl,$acc,1));	#  0
+	&movz	($key,&LB("ebx"));		# 10
 	&movz	("edx",&BP(-128,$tbl,"edx",1));	#  1
+	&shr	("eax",16);			#  7, 6
 	&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
+	&movz	($acc,&BP(-128,$tbl,$key,1));	# 10
+	&movz	($key,&HB("ebx"));		# 11
 	&shl	($acc,16);			# 10
+	&pshufw	("mm6","mm4",0x03);		# 9, 8,15,14
 	&or	("ecx",$acc);			# 10
-	&shr	("eax",16);			#  7, 6
-	&movz	($acc,&HB("ebx"));		# 11
-	&movz	($acc,&BP(-128,$tbl,$acc,1));	# 11
+	&movz	($acc,&BP(-128,$tbl,$key,1));	# 11
+	&movz	($key,&HB("eax"));		#  7
 	&shl	($acc,24);			# 11
-	&or	("edx",$acc);			# 11
 	&shr	("ebx",16);			# 13,12
+	&or	("edx",$acc);			# 11
 
-	&pshufw	("mm6","mm4",0x03);		# 9, 8,15,14
-	&movz	($acc,&HB("eax"));		#  7
-	&movz	($acc,&BP(-128,$tbl,$acc,1));	#  7
+	&movz	($acc,&BP(-128,$tbl,$key,1));	#  7
+	&movz	($key,&HB("ebx"));		# 13
 	&shl	($acc,24);			#  7
 	&or	("ecx",$acc);			#  7
-	&movz	($acc,&HB("ebx"));		# 13
-	&movz	($acc,&BP(-128,$tbl,$acc,1));	# 13
+	&movz	($acc,&BP(-128,$tbl,$key,1));	# 13
+	&movz	($key,&LB("eax"));		#  6
 	&shl	($acc,8);			# 13
+	&movd	("eax","mm2");			#  3, 2, 5, 4
 	&or	("ecx",$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
+	&movz	($acc,&BP(-128,$tbl,$key,1));	#  6
+	&movz	($key,&LB("ebx"));		# 12
+	&shl	($acc,16);			#  6
 	&movd	("ebx","mm6");			#  9, 8,15,14
-	&movz	($acc,&BP(-128,$tbl,$acc,1));	# 12
+	&movd	("mm0","ecx");			# t[0] collected
+	&movz	("ecx",&BP(-128,$tbl,$key,1));	# 12
+	&movz	($key,&LB("eax"));		#  4
 	&or	("ecx",$acc);			# 12
 
-	&movz	($acc,&LB("eax"));		#  4
-	&movz	($acc,&BP(-128,$tbl,$acc,1));	#  4
+	&movz	($acc,&BP(-128,$tbl,$key,1));	#  4
+	&movz	($key,&LB("ebx"));		# 14
 	&or	("edx",$acc);			#  4
-	&movz	($acc,&LB("ebx"));		# 14
-	&movz	($acc,&BP(-128,$tbl,$acc,1));	# 14
+	&movz	($acc,&BP(-128,$tbl,$key,1));	# 14
+	&movz	($key,&HB("eax"));		#  5
 	&shl	($acc,16);			# 14
+	&shr	("eax",16);			#  3, 2
 	&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",$acc);			# 15
+	&movz	($acc,&BP(-128,$tbl,$key,1));	#  5
+	&movz	($key,&HB("ebx"));		# 15
 	&shr	("ebx",16);			#  9, 8
+	&shl	($acc,8);			#  5
+	&movd	("mm1","edx");			# t[1] collected
+	&movz	("edx",&BP(-128,$tbl,$key,1));	# 15
+	&movz	($key,&HB("ebx"));		#  9
+	&shl	("edx",24);			# 15
+	&and	("ebx",0xff);			#  8
+	&or	("edx",$acc);			# 15
 
 	&punpckldq	("mm0","mm1");		# t[0,1] collected
 
-	&movz	($acc,&HB("ebx"));		#  9
-	&movz	($acc,&BP(-128,$tbl,$acc,1));	#  9
+	&movz	($acc,&BP(-128,$tbl,$key,1));	#  9
+	&movz	($key,&LB("eax"));		#  2
 	&shl	($acc,8);			#  9
-	&or	("ecx",$acc);			#  9
-	&and	("ebx",0xff);			#  8
+	&movz	("eax",&HB("eax"));		#  3
 	&movz	("ebx",&BP(-128,$tbl,"ebx",1));	#  8
+	&or	("ecx",$acc);			#  9
+	&movz	($acc,&BP(-128,$tbl,$key,1));	#  2
 	&or	("edx","ebx");			#  8
-	&movz	($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
+	&or	("edx",$acc);			#  2
 	&shl	("eax",24);			#  3
 	&or	("ecx","eax");			#  3
+	&mov	($key,$__key);
+	&movd	("mm4","edx");			# t[2] collected
 	&movd	("mm5","ecx");			# t[3] collected
 
 	&punpckldq	("mm4","mm5");		# t[2,3] collected
@@ -2865,32 +2872,32 @@ sub deckey()
 { my ($i,$key,$tp1,$tp2,$tp4,$tp8) = @_;
   my $tmp = $tbl;
 
-	&mov	($acc,$tp1);
-	&and	($acc,0x80808080);
-	&mov	($tmp,$acc);
-	&shr	($tmp,7);
+	&mov	($tmp,0x80808080);
+	&and	($tmp,$tp1);
 	&lea	($tp2,&DWP(0,$tp1,$tp1));
+	&mov	($acc,$tmp);
+	&shr	($tmp,7);
 	&sub	($acc,$tmp);
 	&and	($tp2,0xfefefefe);
 	&and	($acc,0x1b1b1b1b);
-	&xor	($acc,$tp2);
-	&mov	($tp2,$acc);
+	&xor	($tp2,$acc);
+	&mov	($tmp,0x80808080);
 
-	&and	($acc,0x80808080);
-	&mov	($tmp,$acc);
-	&shr	($tmp,7);
+	&and	($tmp,$tp2);
 	&lea	($tp4,&DWP(0,$tp2,$tp2));
+	&mov	($acc,$tmp);
+	&shr	($tmp,7);
 	&sub	($acc,$tmp);
 	&and	($tp4,0xfefefefe);
 	&and	($acc,0x1b1b1b1b);
 	 &xor	($tp2,$tp1);	# tp2^tp1
-	&xor	($acc,$tp4);
-	&mov	($tp4,$acc);
+	&xor	($tp4,$acc);
+	&mov	($tmp,0x80808080);
 
-	&and	($acc,0x80808080);
-	&mov	($tmp,$acc);
-	&shr	($tmp,7);
+	&and	($tmp,$tp4);
 	&lea	($tp8,&DWP(0,$tp4,$tp4));
+	&mov	($acc,$tmp);
+	&shr	($tmp,7);
 	 &xor	($tp4,$tp1);	# tp4^tp1
 	&sub	($acc,$tmp);
 	&and	($tp8,0xfefefefe);
diff --git a/crypto/aes/asm/vpaes-x86.pl b/crypto/aes/asm/vpaes-x86.pl
index 1533e2c304..433912ff57 100644
--- a/crypto/aes/asm/vpaes-x86.pl
+++ b/crypto/aes/asm/vpaes-x86.pl
@@ -27,9 +27,9 @@
 #
 #		aes-586.pl		vpaes-x86.pl
 #
-# Core 2(**)	29.1/42.3/18.3		22.0/25.6(***)
-# Nehalem	27.9/40.4/18.1		10.3/12.0
-# Atom		102./119./60.1		64.5/85.3(***)
+# Core 2(**)	28.1/41.4/18.3		21.9/25.2(***)
+# Nehalem	27.9/40.4/18.1		10.2/11.9
+# Atom		70.7/92.1/60.1		61.1/81.0(***)
 #
 # (*)	"Hyper-threading" in the context refers rather to cache shared
 #	among multiple cores, than to specifically Intel HTT. As vast
@@ -40,8 +40,8 @@
 # (**)	"Core 2" refers to initial 65nm design, a.k.a. Conroe.
 #
 # (***)	Less impressive improvement on Core 2 and Atom is due to slow
-#	pshufb,	yet it's respectable +32%/65%  improvement on Core 2
-#	and +58%/40% on Atom (as implied, over "hyper-threading-safe"
+#	pshufb,	yet it's respectable +28%/64%  improvement on Core 2
+#	and +15% on Atom (as implied, over "hyper-threading-safe"
 #	code path).
 #
 #						<appro@openssl.org>
@@ -183,35 +183,35 @@ $k_dsbo=0x2c0;		# decryption sbox final output
 	&movdqa	("xmm1","xmm6")
 	&movdqa	("xmm2",&QWP($k_ipt,$const));
 	&pandn	("xmm1","xmm0");
-	&movdqu	("xmm5",&QWP(0,$key));
-	&psrld	("xmm1",4);
 	&pand	("xmm0","xmm6");
+	&movdqu	("xmm5",&QWP(0,$key));
 	&pshufb	("xmm2","xmm0");
 	&movdqa	("xmm0",&QWP($k_ipt+16,$const));
-	&pshufb	("xmm0","xmm1");
 	&pxor	("xmm2","xmm5");
-	&pxor	("xmm0","xmm2");
+	&psrld	("xmm1",4);
 	&add	($key,16);
+	&pshufb	("xmm0","xmm1");
 	&lea	($base,&DWP($k_mc_backward,$const));
+	&pxor	("xmm0","xmm2");
 	&jmp	(&label("enc_entry"));
 
 
 &set_label("enc_loop",16);
 	# middle of middle round
 	&movdqa	("xmm4",&QWP($k_sb1,$const));	# 4 : sb1u
-	&pshufb	("xmm4","xmm2");		# 4 = sb1u
-	&pxor	("xmm4","xmm5");		# 4 = sb1u + k
 	&movdqa	("xmm0",&QWP($k_sb1+16,$const));# 0 : sb1t
+	&pshufb	("xmm4","xmm2");		# 4 = sb1u
 	&pshufb	("xmm0","xmm3");		# 0 = sb1t
-	&pxor	("xmm0","xmm4");		# 0 = A
+	&pxor	("xmm4","xmm5");		# 4 = sb1u + k
 	&movdqa	("xmm5",&QWP($k_sb2,$const));	# 4 : sb2u
-	&pshufb	("xmm5","xmm2");		# 4 = sb2u
+	&pxor	("xmm0","xmm4");		# 0 = A
 	&movdqa	("xmm1",&QWP(-0x40,$base,$magic));# .Lk_mc_forward[]
+	&pshufb	("xmm5","xmm2");		# 4 = sb2u
 	&movdqa	("xmm2",&QWP($k_sb2+16,$const));# 2 : sb2t
-	&pshufb	("xmm2","xmm3");		# 2 = sb2t
-	&pxor	("xmm2","xmm5");		# 2 = 2A
 	&movdqa	("xmm4",&QWP(0,$base,$magic));	# .Lk_mc_backward[]
+	&pshufb	("xmm2","xmm3");		# 2 = sb2t
 	&movdqa	("xmm3","xmm0");		# 3 = A
+	&pxor	("xmm2","xmm5");		# 2 = 2A
 	&pshufb	("xmm0","xmm1");		# 0 = B
 	&add	($key,16);			# next key
 	&pxor	("xmm0","xmm2");		# 0 = 2A+B
@@ -220,30 +220,30 @@ $k_dsbo=0x2c0;		# decryption sbox final output
 	&pxor	("xmm3","xmm0");		# 3 = 2A+B+D
 	&pshufb	("xmm0","xmm1");		# 0 = 2B+C
 	&and	($magic,0x30);			# ... mod 4
-	&pxor	("xmm0","xmm3");		# 0 = 2A+3B+C+D
 	&sub	($round,1);			# nr--
+	&pxor	("xmm0","xmm3");		# 0 = 2A+3B+C+D
 
 &set_label("enc_entry");
 	# top of round
 	&movdqa	("xmm1","xmm6");		# 1 : i
+	&movdqa	("xmm5",&QWP($k_inv+16,$const));# 2 : a/k
 	&pandn	("xmm1","xmm0");		# 1 = i<<4
 	&psrld	("xmm1",4);			# 1 = i
 	&pand	("xmm0","xmm6");		# 0 = k
-	&movdqa	("xmm5",&QWP($k_inv+16,$const));# 2 : a/k
 	&pshufb	("xmm5","xmm0");		# 2 = a/k
-	&pxor	("xmm0","xmm1");		# 0 = j
 	&movdqa	("xmm3","xmm7");		# 3 : 1/i
+	&pxor	("xmm0","xmm1");		# 0 = j
 	&pshufb	("xmm3","xmm1");		# 3 = 1/i
-	&pxor	("xmm3","xmm5");		# 3 = iak = 1/i + a/k
 	&movdqa	("xmm4","xmm7");		# 4 : 1/j
+	&pxor	("xmm3","xmm5");		# 3 = iak = 1/i + a/k
 	&pshufb	("xmm4","xmm0");		# 4 = 1/j
-	&pxor	("xmm4","xmm5");		# 4 = jak = 1/j + a/k
 	&movdqa	("xmm2","xmm7");		# 2 : 1/iak
+	&pxor	("xmm4","xmm5");		# 4 = jak = 1/j + a/k
 	&pshufb	("xmm2","xmm3");		# 2 = 1/iak
-	&pxor	("xmm2","xmm0");		# 2 = io
 	&movdqa	("xmm3","xmm7");		# 3 : 1/jak
-	&movdqu	("xmm5",&QWP(0,$key));
+	&pxor	("xmm2","xmm0");		# 2 = io
 	&pshufb	("xmm3","xmm4");		# 3 = 1/jak
+	&movdqu	("xmm5",&QWP(0,$key));
 	&pxor	("xmm3","xmm1");		# 3 = jo
 	&jnz	(&label("enc_loop"));
 
@@ -265,8 +265,8 @@ $k_dsbo=0x2c0;		# decryption sbox final output
 ##  Same API as encryption core.
 ##
 &function_begin_B("_vpaes_decrypt_core");
-	&mov	($round,&DWP(240,$key));
 	&lea	($base,&DWP($k_dsbd,$const));
+	&mov	($round,&DWP(240,$key));
 	&movdqa	("xmm1","xmm6");
 	&movdqa	("xmm2",&QWP($k_dipt-$k_dsbd,$base));
 	&pandn	("xmm1","xmm0");
@@ -292,62 +292,61 @@ $k_dsbo=0x2c0;		# decryption sbox final output
 ##  Inverse mix columns
 ##
 	&movdqa	("xmm4",&QWP(-0x20,$base));	# 4 : sb9u
+	&movdqa	("xmm1",&QWP(-0x10,$base));	# 0 : sb9t
 	&pshufb	("xmm4","xmm2");		# 4 = sb9u
+	&pshufb	("xmm1","xmm3");		# 0 = sb9t
 	&pxor	("xmm4","xmm0");
-	&movdqa	("xmm0",&QWP(-0x10,$base));	# 0 : sb9t
-	&pshufb	("xmm0","xmm3");		# 0 = sb9t
-	&pxor	("xmm0","xmm4");		# 0 = ch
 	&add	($key,16);			# next round key
+	&pxor	("xmm1","xmm4");		# 0 = ch
 
-	&pshufb	("xmm0","xmm5");		# MC ch
 	&movdqa	("xmm4",&QWP(0,$base));		# 4 : sbdu
+	&pshufb	("xmm1","xmm5");		# MC ch
 	&pshufb	("xmm4","xmm2");		# 4 = sbdu
-	&pxor	("xmm4","xmm0");		# 4 = ch
 	&movdqa	("xmm0",&QWP(0x10,$base));	# 0 : sbdt
+	&pxor	("xmm4","xmm1");		# 4 = ch
 	&pshufb	("xmm0","xmm3");		# 0 = sbdt
-	&pxor	("xmm0","xmm4");		# 0 = ch
 	&sub	($round,1);			# nr--
+	&pxor	("xmm0","xmm4");		# 0 = ch
 
-	&pshufb	("xmm0","xmm5");		# MC ch
 	&movdqa	("xmm4",&QWP(0x20,$base));	# 4 : sbbu
+	&pshufb	("xmm0","xmm5");		# MC ch
+	&movdqa	("xmm1",&QWP(0x30,$base));	# 0 : sbbt
 	&pshufb	("xmm4","xmm2");		# 4 = sbbu
+	&pshufb	("xmm1","xmm3");		# 0 = sbbt
 	&pxor	("xmm4","xmm0");		# 4 = ch
-	&movdqa	("xmm0",&QWP(0x30,$base));	# 0 : sbbt
-	&pshufb	("xmm0","xmm3");		# 0 = sbbt
-	&pxor	("xmm0","xmm4");		# 0 = ch
+	&pxor	("xmm1","xmm4");		# 0 = ch
 
-	&pshufb	("xmm0","xmm5");		# MC ch
 	&movdqa	("xmm4",&QWP(0x40,$base));	# 4 : sbeu
-	&pshufb	("xmm4","xmm2");		# 4 = sbeu
-	&pxor	("xmm4","xmm0");		# 4 = ch
+	&pshufb	("xmm1","xmm5");		# MC ch
 	&movdqa	("xmm0",&QWP(0x50,$base));	# 0 : sbet
+	&pshufb	("xmm4","xmm2");		# 4 = sbeu
 	&pshufb	("xmm0","xmm3");		# 0 = sbet
-	&pxor	("xmm0","xmm4");		# 0 = ch
-
 	&palignr("xmm5","xmm5",12);
+	&pxor	("xmm4","xmm1");		# 4 = ch
+	&pxor	("xmm0","xmm4");		# 0 = ch
 
 &set_label("dec_entry");
 	# top of round
 	&movdqa	("xmm1","xmm6");		# 1 : i
 	&pandn	("xmm1","xmm0");		# 1 = i<<4
+	&movdqa	("xmm2",&QWP($k_inv+16,$const));# 2 : a/k
 	&psrld	("xmm1",4);			# 1 = i
 	&pand	("xmm0","xmm6");		# 0 = k
-	&movdqa	("xmm2",&QWP($k_inv+16,$const));# 2 : a/k
 	&pshufb	("xmm2","xmm0");		# 2 = a/k
-	&pxor	("xmm0","xmm1");		# 0 = j
 	&movdqa	("xmm3","xmm7");		# 3 : 1/i
+	&pxor	("xmm0","xmm1");		# 0 = j
 	&pshufb	("xmm3","xmm1");		# 3 = 1/i
-	&pxor	("xmm3","xmm2");		# 3 = iak = 1/i + a/k
 	&movdqa	("xmm4","xmm7");		# 4 : 1/j
+	&pxor	("xmm3","xmm2");		# 3 = iak = 1/i + a/k
 	&pshufb	("xmm4","xmm0");		# 4 = 1/j
 	&pxor	("xmm4","xmm2");		# 4 = jak = 1/j + a/k
 	&movdqa	("xmm2","xmm7");		# 2 : 1/iak
 	&pshufb	("xmm2","xmm3");		# 2 = 1/iak
-	&pxor	("xmm2","xmm0");		# 2 = io
 	&movdqa	("xmm3","xmm7");		# 3 : 1/jak
+	&pxor	("xmm2","xmm0");		# 2 = io
 	&pshufb	("xmm3","xmm4");		# 3 = 1/jak
-	&pxor	("xmm3","xmm1");		# 3 = jo
 	&movdqu	("xmm0",&QWP(0,$key));
+	&pxor	("xmm3","xmm1");		# 3 = jo
 	&jnz	(&label("dec_loop"));
 
 	# middle of last round
@@ -542,12 +541,12 @@ $k_dsbo=0x2c0;		# decryption sbox final output
 ##    %xmm0: b+c+d  b+c  b  a
 ##
 &function_begin_B("_vpaes_schedule_192_smear");
-	&pshufd	("xmm0","xmm6",0x80);		# d c 0 0 -> c 0 0 0
-	&pxor	("xmm6","xmm0");		# -> c+d c 0 0
+	&pshufd	("xmm1","xmm6",0x80);		# d c 0 0 -> c 0 0 0
 	&pshufd	("xmm0","xmm7",0xFE);		# b a _ _ -> b b b a
+	&pxor	("xmm6","xmm1");		# -> c+d c 0 0
+	&pxor	("xmm1","xmm1");
 	&pxor	("xmm6","xmm0");		# -> b+c+d b+c b a
 	&movdqa	("xmm0","xmm6");
-	&pxor	("xmm1","xmm1");
 	&movhlps("xmm6","xmm1");		# clobber low side with zeros
 	&ret	();
 &function_end_B("_vpaes_schedule_192_smear");