1 files changed, 151 insertions, 11 deletions

diff --git a/src/lib/libcrypto/rc4/asm/rc4-586.pl b/src/lib/libcrypto/rc4/asm/rc4-586.pl
index 38a44a70ef..5c9ac6ad28 100644
--- a/src/lib/libcrypto/rc4/asm/rc4-586.pl
+++ b/src/lib/libcrypto/rc4/asm/rc4-586.pl
@@ -28,6 +28,34 @@
 #
 #                                       <appro@fy.chalmers.se>
 
+# May 2011
+#
+# Optimize for Core2 and Westmere [and incidentally Opteron]. Current
+# performance in cycles per processed byte (less is better) and
+# improvement relative to previous version of this module is:
+#
+# Pentium       10.2                    # original numbers
+# Pentium III   7.8(*)
+# Intel P4      7.5
+#
+# Opteron       6.1/+20%                # new MMX numbers
+# Core2         5.3/+67%(**)
+# Westmere      5.1/+94%(**)
+# Sandy Bridge  5.0/+8%
+# Atom          12.6/+6%
+#
+# (*)   PIII can actually deliver 6.6 cycles per byte with MMX code,
+#       but this specific code performs poorly on Core2. And vice
+#       versa, below MMX/SSE code delivering 5.8/7.1 on Core2 performs
+#       poorly on PIII, at 8.0/14.5:-( As PIII is not a "hot" CPU
+#       [anymore], I chose to discard PIII-specific code path and opt
+#       for original IALU-only code, which is why MMX/SSE code path
+#       is guarded by SSE2 bit (see below), not MMX/SSE.
+# (**)  Performance vs. block size on Core2 and Westmere had a maximum
+#       at ... 64 bytes block size. And it was quite a maximum, 40-60%
+#       in comparison to largest 8KB block size. Above improvement
+#       coefficients are for the largest block size.
+
 $0 =~ m/(.*[\/\\])[^\/\\]+$/; $dir=$1;
 push(@INC,"${dir}","${dir}../../perlasm");
 require "x86asm.pl";
@@ -62,6 +90,68 @@ sub RC4_loop {
         &$func  ($out,&DWP(0,$dat,$ty,4));
 }
 
+if ($alt=0) {
+  # >20% faster on Atom and Sandy Bridge[!], 8% faster on Opteron,
+  # but ~40% slower on Core2 and Westmere... Attempt to add movz
+  # brings down Opteron by 25%, Atom and Sandy Bridge by 15%, yet
+  # on Core2 with movz it's almost 20% slower than below alternative
+  # code... Yes, it's a total mess...
+  my @XX=($xx,$out);
+  $RC4_loop_mmx = sub {         # SSE actually...
+    my $i=shift;
+    my $j=$i<=0?0:$i>>1;
+    my $mm=$i<=0?"mm0":"mm".($i&1);
+
+        &add    (&LB($yy),&LB($tx));
+        &lea    (@XX[1],&DWP(1,@XX[0]));
+        &pxor   ("mm2","mm0")                           if ($i==0);
+        &psllq  ("mm1",8)                               if ($i==0);
+        &and    (@XX[1],0xff);
+        &pxor   ("mm0","mm0")                           if ($i<=0);
+        &mov    ($ty,&DWP(0,$dat,$yy,4));
+        &mov    (&DWP(0,$dat,$yy,4),$tx);
+        &pxor   ("mm1","mm2")                           if ($i==0);
+        &mov    (&DWP(0,$dat,$XX[0],4),$ty);
+        &add    (&LB($ty),&LB($tx));
+        &movd   (@XX[0],"mm7")                          if ($i==0);
+        &mov    ($tx,&DWP(0,$dat,@XX[1],4));
+        &pxor   ("mm1","mm1")                           if ($i==1);
+        &movq   ("mm2",&QWP(0,$inp))                    if ($i==1);
+        &movq   (&QWP(-8,(@XX[0],$inp)),"mm1")          if ($i==0);
+        &pinsrw ($mm,&DWP(0,$dat,$ty,4),$j);
+
+        push    (@XX,shift(@XX))                        if ($i>=0);
+  }
+} else {
+  # Using pinsrw here improves performane on Intel CPUs by 2-3%, but
+  # brings down AMD by 7%...
+  $RC4_loop_mmx = sub {
+    my $i=shift;
+
+        &add    (&LB($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),$tx);
+        &mov    (&DWP(0,$dat,$xx,4),$ty);
+        &inc    ($xx);
+        &add    ($ty,$tx);
+        &movz   ($xx,&LB($xx));                         # (*)
+        &movz   ($ty,&LB($ty));                         # (*)
+        &pxor   ("mm2",$i==1?"mm0":"mm1")               if ($i>=0);
+        &movq   ("mm0",&QWP(0,$inp))                    if ($i<=0);
+        &movq   (&QWP(-8,($out,$inp)),"mm2")            if ($i==0);
+        &mov    ($tx,&DWP(0,$dat,$xx,4));
+        &movd   ($i>0?"mm1":"mm2",&DWP(0,$dat,$ty,4));
+
+        # (*)   This is the key to Core2 and Westmere performance.
+        #       Whithout movz out-of-order execution logic confuses
+        #       itself and fails to reorder loads and stores. Problem
+        #       appears to be fixed in Sandy Bridge...
+  }
+}
+
+&external_label("OPENSSL_ia32cap_P");
+
 # void RC4(RC4_KEY *key,size_t len,const unsigned char *inp,unsigned char *out);
 &function_begin("RC4");
         &mov    ($dat,&wparam(0));      # load key schedule pointer
@@ -94,11 +184,56 @@ sub RC4_loop {
         &and    ($ty,-4);               # how many 4-byte chunks?
         &jz     (&label("loop1"));
 
+        &test   ($ty,-8);
+        &mov    (&wparam(3),$out);      # $out as accumulator in these loops
+        &jz     (&label("go4loop4"));
+
+        &picmeup($out,"OPENSSL_ia32cap_P");
+        &bt     (&DWP(0,$out),26);      # check SSE2 bit [could have been MMX]
+        &jnc    (&label("go4loop4"));
+
+        &mov    ($out,&wparam(3))       if (!$alt);
+        &movd   ("mm7",&wparam(3))      if ($alt);
+        &and    ($ty,-8);
+        &lea    ($ty,&DWP(-8,$inp,$ty));
+        &mov    (&DWP(-4,$dat),$ty);    # save input+(len/8)*8-8
+
+        &$RC4_loop_mmx(-1);
+        &jmp(&label("loop_mmx_enter"));
+
+        &set_label("loop_mmx",16);
+                &$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,$inp));
+        &jb     (&label("loop_mmx"));
+
+    if ($alt) {
+        &movd   ($out,"mm7");
+        &pxor   ("mm2","mm0");
+        &psllq  ("mm1",8);
+        &pxor   ("mm1","mm2");
+        &movq   (&QWP(-8,$out,$inp),"mm1");
+    } else {
+        &psllq  ("mm1",56);
+        &pxor   ("mm2","mm1");
+        &movq   (&QWP(-8,$out,$inp),"mm2");
+    }
+        &emms   ();
+
+        &cmp    ($inp,&wparam(1));      # compare to input+len
+        &je     (&label("done"));
+        &jmp    (&label("loop1"));
+
+&set_label("go4loop4",16);
         &lea    ($ty,&DWP(-4,$inp,$ty));
         &mov    (&wparam(2),$ty);       # save input+(len/4)*4-4
-        &mov    (&wparam(3),$out);      # $out as accumulator in this loop
 
-        &set_label("loop4",16);
+        &set_label("loop4");
                 for ($i=0;$i<4;$i++) { RC4_loop($i); }
                 &ror    ($out,8);
                 &xor    ($out,&DWP(0,$inp));
@@ -151,7 +286,7 @@ sub RC4_loop {
 
 &set_label("done");
         &dec    (&LB($xx));
-        &mov    (&BP(-4,$dat),&LB($yy));        # save key->y
+        &mov    (&DWP(-4,$dat),$yy);            # save key->y
         &mov    (&BP(-8,$dat),&LB($xx));        # save key->x
 &set_label("abort");
 &function_end("RC4");
@@ -164,10 +299,8 @@ $idi="ebp";
 $ido="ecx";
 $idx="edx";
 
-&external_label("OPENSSL_ia32cap_P");
-
 # void RC4_set_key(RC4_KEY *key,int len,const unsigned char *data);
-&function_begin("RC4_set_key");
+&function_begin("private_RC4_set_key");
         &mov    ($out,&wparam(0));              # load key
         &mov    ($idi,&wparam(1));              # load len
         &mov    ($inp,&wparam(2));              # load data
@@ -245,7 +378,7 @@ $idx="edx";
         &xor    ("eax","eax");
         &mov    (&DWP(-8,$out),"eax");          # key->x=0;
         &mov    (&DWP(-4,$out),"eax");          # key->y=0;
-&function_end("RC4_set_key");
+&function_end("private_RC4_set_key");
 
 # const char *RC4_options(void);
 &function_begin_B("RC4_options");
@@ -254,14 +387,21 @@ $idx="edx";
         &blindpop("eax");
         &lea    ("eax",&DWP(&label("opts")."-".&label("pic_point"),"eax"));
         &picmeup("edx","OPENSSL_ia32cap_P");
-        &bt     (&DWP(0,"edx"),20);
-        &jnc    (&label("skip"));
-          &add  ("eax",12);
-        &set_label("skip");
+        &mov    ("edx",&DWP(0,"edx"));
+        &bt     ("edx",20);
+        &jc     (&label("1xchar"));
+        &bt     ("edx",26);
+        &jnc    (&label("ret"));
+        &add    ("eax",25);
+        &ret    ();
+&set_label("1xchar");
+        &add    ("eax",12);
+&set_label("ret");
         &ret    ();
 &set_label("opts",64);
 &asciz  ("rc4(4x,int)");
 &asciz  ("rc4(1x,char)");
+&asciz  ("rc4(8x,mmx)");
 &asciz  ("RC4 for x86, CRYPTOGAMS by <appro\@openssl.org>");
 &align  (64);
 &function_end_B("RC4_options");