From b943b7d2c7be535a996b7103b6fe7011735e7f0b Mon Sep 17 00:00:00 2001
From: Andy Polyakov <appro@openssl.org>
Date: Sun, 4 May 2014 16:56:00 +0200
Subject: [PATCH 1/1] md5/asm/md5-[586|x86_64].pl: +15% on Atom.

[MD5 is hardly relevant, just cleaning up repository]
---
 crypto/md5/asm/md5-586.pl    | 17 ++++++++---------
 crypto/md5/asm/md5-x86_64.pl | 13 ++++++++++++-
 2 files changed, 20 insertions(+), 10 deletions(-)

diff --git a/crypto/md5/asm/md5-586.pl b/crypto/md5/asm/md5-586.pl
index 6cb66bb499..4971357173 100644
--- a/crypto/md5/asm/md5-586.pl
+++ b/crypto/md5/asm/md5-586.pl
@@ -56,14 +56,14 @@ sub R0
 	&lea($a,&DWP($t,$a,$tmp2,1));
 
 	&xor($tmp1,$d); # F function - part 4
+	&mov($tmp2,&DWP($xo[$ki+1]*4,$K,"",0)) if ($pos != 2);
 
 	&add($a,$tmp1);
-	&mov($tmp1,&Np($c)) if $pos < 1;	# next tmp1 for R0
-	&mov($tmp1,&Np($c)) if $pos == 1;	# next tmp1 for R1
 
 	&rotl($a,$s);
 
-	&mov($tmp2,&DWP($xo[$ki+1]*4,$K,"",0)) if ($pos != 2);
+	&mov($tmp1,&Np($c)) if $pos < 1;	# next tmp1 for R0
+	&mov($tmp1,&Np($c)) if $pos == 1;	# next tmp1 for R1
 
 	&add($a,$b);
 	}
@@ -74,13 +74,12 @@ sub R1
 
 	&comment("R1 $ki");
 
-	&lea($a,&DWP($t,$a,$tmp2,1));
-
 	&xor($tmp1,$b); # G function - part 2
 	&and($tmp1,$d); # G function - part 3
+	&lea($a,&DWP($t,$a,$tmp2,1));
 
-	&mov($tmp2,&DWP($xo[$ki+1]*4,$K,"",0)) if ($pos != 2);
 	&xor($tmp1,$c);			# G function - part 4
+	&mov($tmp2,&DWP($xo[$ki+1]*4,$K,"",0)) if ($pos != 2);
 
 	&add($a,$tmp1);
 	&mov($tmp1,&Np($c)) if $pos < 1;	# G function - part 1
@@ -108,10 +107,10 @@ if (($n & 1) == 0)
 	&lea($a,&DWP($t,$a,$tmp2,1));
 
 	&add($a,$tmp1);
+	&mov($tmp2,&DWP($xo[$ki+1]*4,$K,"",0));
 
 	&rotl($a,$s);
 
-	&mov($tmp2,&DWP($xo[$ki+1]*4,$K,"",0));
 	&mov($tmp1,&Np($c));
 	}
 else
@@ -120,11 +119,11 @@ else
 	# make sure to do 'D' first, not 'B', else we clash with
 	# the last add from the previous round.
 
-	&lea($a,&DWP($t,$a,$tmp2,1));
-
 	&add($b,$c);			# MOVED FORWARD
 	&xor($tmp1,$d); # H function - part 2
 
+	&lea($a,&DWP($t,$a,$tmp2,1));
+
 	&xor($tmp1,$b); # H function - part 3
 	&mov($tmp2,&DWP($xo[$ki+1]*4,$K,"",0)) if ($pos != 2);
 
diff --git a/crypto/md5/asm/md5-x86_64.pl b/crypto/md5/asm/md5-x86_64.pl
index a79d0e46ea..f7dae27576 100755
--- a/crypto/md5/asm/md5-x86_64.pl
+++ b/crypto/md5/asm/md5-x86_64.pl
@@ -47,8 +47,8 @@ sub round2_step
     $code .= " mov	%edx,		%r12d		/* (NEXT STEP) z' = %edx */\n" if ($pos == -1);
     $code .= <<EOF;
 	not	%r11d				/* not z */
-	lea	$T_i($dst,%r10d),$dst		/* Const + dst + ... */
 	and	$x,		%r12d		/* x & z */
+	lea	$T_i($dst,%r10d),$dst		/* Const + dst + ... */
 	and	$y,		%r11d		/* y & (not z) */
 	mov	$k_next*4(%rsi),%r10d		/* (NEXT STEP) X[$k_next] */
 	or	%r11d,		%r12d		/* (y & (not z)) | (x & z) */
@@ -65,6 +65,7 @@ EOF
 #   %r10d = X[k_next]
 #   %r11d = y' (copy of y for the next step)
 # Each round3_step() takes about 4.2 clocks (8 instructions, 1.9 IPC)
+{ my $round3_alter=0;
 sub round3_step
 {
     my ($pos, $dst, $x, $y, $z, $k_next, $T_i, $s) = @_;
@@ -75,10 +76,20 @@ sub round3_step
 	mov	$k_next*4(%rsi),%r10d		/* (NEXT STEP) X[$k_next] */
 	xor	$x,		%r11d		/* x ^ ... */
 	add	%r11d,		$dst		/* dst += ... */
+EOF
+    $code .= <<EOF if ($round3_alter);
 	rol	\$$s,		$dst		/* dst <<< s */
 	mov	$x,		%r11d		/* (NEXT STEP) y' = $x */
+EOF
+    $code .= <<EOF if (!$round3_alter);
+	mov	$x,		%r11d		/* (NEXT STEP) y' = $x */
+	rol	\$$s,		$dst		/* dst <<< s */
+EOF
+    $code .= <<EOF;
 	add	$x,		$dst		/* dst += x */
 EOF
+    $round3_alter^=1;
+}
 }
 
 # round4_step() does:
-- 
2.34.1