wp-mmx.pl: ~10% performance improvement.

diff --git a/crypto/whrlpool/asm/wp-mmx.pl b/crypto/whrlpool/asm/wp-mmx.pl
index cb2381c22ba12aed4a9411bb181b7e16e30ec45d..c584e5b92b251a4431027b23ad4f386d32566d00 100644 (file)
--- 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));