rc4-586.pl: optimize even further...
authorAndy Polyakov <appro@openssl.org>
Fri, 27 May 2011 09:46:19 +0000 (09:46 +0000)
committerAndy Polyakov <appro@openssl.org>
Fri, 27 May 2011 09:46:19 +0000 (09:46 +0000)
crypto/rc4/asm/rc4-586.pl

index 0c4cac4e8986e24970ca04fe381526a1b2df1444..4b8bc78b33b294960cc1bdf56354a0ad55ce0a54 100644 (file)
 # Pentium III  7.8(*)
 # Intel P4     7.5
 #
-# Opteron      6.4/+14%                # new MMX numbers
-# Core2                5.8/+50%(**)
-# Westmere     5.5/+80%(**)
-# Sandy Bridge 5.4/0%
+# Opteron      6.1/+20%                # new MMX numbers
+# Core2                5.3/+67%(**)
+# Westmere     5.1/+94%(**)
+# Sandy Bridge 5.0/+8%
 #
 # (*)  PIII can actually deliver 6.6 cycles per byte with MMX code,
 #      but this specific code performs poorly on Core2. And vice
@@ -126,8 +126,7 @@ if ($alt=0) {
   $RC4_loop_mmx = sub {
     my $i=shift;
 
-       &add    ($yy,$tx);
-       &movz   ($yy,&LB($yy));                         # (*)
+       &add    (&LB($yy),&LB($tx));
        &psllq  ("mm1",8*(($i-1)&7))                    if (abs($i)!=1);
        &mov    ($ty,&DWP(0,$dat,$yy,4));
        &mov    (&DWP(0,$dat,$yy,4),$tx);
@@ -204,6 +203,9 @@ if ($alt=0) {
                &$RC4_loop_mmx(0);
        &set_label("loop_mmx_enter");
                for     ($i=1;$i<8;$i++) { &$RC4_loop_mmx($i); }
+               &mov    ($ty,$yy);
+               &xor    ($yy,$yy);              # this is second key to Core2
+               &mov    (&LB($yy),&LB($ty));    # and Westmere performance...
                &cmp    ($inp,&DWP(-4,$dat));
                &lea    ($inp,&DWP(8,$inp));
        &jb     (&label("loop_mmx"));