From 1121ba1b74f4ac132aa36e6363ac3c3c2e23bfb6 Mon Sep 17 00:00:00 2001
From: Andy Polyakov <appro@openssl.org>
Date: Sun, 15 Jul 2012 13:29:23 +0000
Subject: [PATCH] wp-mmx.pl: ~10% performance improvement. (cherry picked from
 commit ae007d4d09f2ed9940c6e581bab9290d78615506)

---
 crypto/whrlpool/asm/wp-mmx.pl | 42 ++++++++++++++++++-----------------
 1 file changed, 22 insertions(+), 20 deletions(-)

diff --git a/crypto/whrlpool/asm/wp-mmx.pl b/crypto/whrlpool/asm/wp-mmx.pl
index cb2381c22b..c584e5b92b 100644
--- a/crypto/whrlpool/asm/wp-mmx.pl
+++ b/crypto/whrlpool/asm/wp-mmx.pl
@@ -118,34 +118,36 @@ $tbl="ebp";
 	&movq	(@mm[0],&QWP(2048*$SCALE,$tbl,"esi",8));	# rc[r]
 	&mov	("eax",&DWP(0,"esp"));
 	&mov	("ebx",&DWP(4,"esp"));
+	&movz	("ecx",&LB("eax"));
+	&movz	("edx",&HB("eax"));
 for($i=0;$i<8;$i++) {
     my $func = ($i==0)? \&movq : \&pxor;
-	&movb	(&LB("ecx"),&LB("eax"));
-	&movb	(&LB("edx"),&HB("eax"));
+	&shr	("eax",16);
 	&scale	("esi","ecx");
+	&movz	("ecx",&LB("eax"));
 	&scale	("edi","edx");
-	&shr	("eax",16);
+	&movz	("edx",&HB("eax"));
 	&pxor	(@mm[0],&QWP(&row(0),$tbl,"esi",8));
 	&$func	(@mm[1],&QWP(&row(1),$tbl,"edi",8));
-	&movb	(&LB("ecx"),&LB("eax"));
-	&movb	(&LB("edx"),&HB("eax"));
 	&mov	("eax",&DWP(($i+1)*8,"esp"));
 	&scale	("esi","ecx");
+	&movz	("ecx",&LB("ebx"));
 	&scale	("edi","edx");
+	&movz	("edx",&HB("ebx"));
 	&$func	(@mm[2],&QWP(&row(2),$tbl,"esi",8));
 	&$func	(@mm[3],&QWP(&row(3),$tbl,"edi",8));
-	&movb	(&LB("ecx"),&LB("ebx"));
-	&movb	(&LB("edx"),&HB("ebx"));
+	&shr	("ebx",16);
 	&scale	("esi","ecx");
+	&movz	("ecx",&LB("ebx"));
 	&scale	("edi","edx");
-	&shr	("ebx",16);
+	&movz	("edx",&HB("ebx"));
 	&$func	(@mm[4],&QWP(&row(4),$tbl,"esi",8));
 	&$func	(@mm[5],&QWP(&row(5),$tbl,"edi",8));
-	&movb	(&LB("ecx"),&LB("ebx"));
-	&movb	(&LB("edx"),&HB("ebx"));
 	&mov	("ebx",&DWP(($i+1)*8+4,"esp"));
 	&scale	("esi","ecx");
+	&movz	("ecx",&LB("eax"));
 	&scale	("edi","edx");
+	&movz	("edx",&HB("eax"));
 	&$func	(@mm[6],&QWP(&row(6),$tbl,"esi",8));
 	&$func	(@mm[7],&QWP(&row(7),$tbl,"edi",8));
     push(@mm,shift(@mm));
@@ -154,32 +156,32 @@ for($i=0;$i<8;$i++) {
 	for($i=0;$i<8;$i++) { &movq(&QWP($i*8,"esp"),@mm[$i]); }    # K=L
 
 for($i=0;$i<8;$i++) {
-	&movb	(&LB("ecx"),&LB("eax"));
-	&movb	(&LB("edx"),&HB("eax"));
+	&shr	("eax",16);
 	&scale	("esi","ecx");
+	&movz	("ecx",&LB("eax"));
 	&scale	("edi","edx");
-	&shr	("eax",16);
+	&movz	("edx",&HB("eax"));
 	&pxor	(@mm[0],&QWP(&row(0),$tbl,"esi",8));
 	&pxor	(@mm[1],&QWP(&row(1),$tbl,"edi",8));
-	&movb	(&LB("ecx"),&LB("eax"));
-	&movb	(&LB("edx"),&HB("eax"));
 	&mov	("eax",&DWP(64+($i+1)*8,"esp"))		if ($i<7);
 	&scale	("esi","ecx");
+	&movz	("ecx",&LB("ebx"));
 	&scale	("edi","edx");
+	&movz	("edx",&HB("ebx"));
 	&pxor	(@mm[2],&QWP(&row(2),$tbl,"esi",8));
 	&pxor	(@mm[3],&QWP(&row(3),$tbl,"edi",8));
-	&movb	(&LB("ecx"),&LB("ebx"));
-	&movb	(&LB("edx"),&HB("ebx"));
+	&shr	("ebx",16);
 	&scale	("esi","ecx");
+	&movz	("ecx",&LB("ebx"));
 	&scale	("edi","edx");
-	&shr	("ebx",16);
+	&movz	("edx",&HB("ebx"));
 	&pxor	(@mm[4],&QWP(&row(4),$tbl,"esi",8));
 	&pxor	(@mm[5],&QWP(&row(5),$tbl,"edi",8));
-	&movb	(&LB("ecx"),&LB("ebx"));
-	&movb	(&LB("edx"),&HB("ebx"));
 	&mov	("ebx",&DWP(64+($i+1)*8+4,"esp"))	if ($i<7);
 	&scale	("esi","ecx");
+	&movz	("ecx",&LB("eax"));
 	&scale	("edi","edx");
+	&movz	("edx",&HB("eax"));
 	&pxor	(@mm[6],&QWP(&row(6),$tbl,"esi",8));
 	&pxor	(@mm[7],&QWP(&row(7),$tbl,"edi",8));
     push(@mm,shift(@mm));
-- 
2.25.1