Next generation aes-586.pl featuring AES_[en|de]crypt, accessing exclusively
authorAndy Polyakov <appro@openssl.org>
Mon, 31 Jul 2006 20:03:56 +0000 (20:03 +0000)
committerAndy Polyakov <appro@openssl.org>
Mon, 31 Jul 2006 20:03:56 +0000 (20:03 +0000)
256 byte S-box. AES_cbc_encrypt needs further work as it should also use
slow routines when processing smaller amount of data.

crypto/aes/asm/aes-586.pl

index d7f64554ff762b5694291e9ed268131ceb1cce42..07243ab4482cf08b39525eae19bdfe6cfca6d1e0 100755 (executable)
@@ -6,7 +6,7 @@
 # forms are granted according to the OpenSSL license.
 # ====================================================================
 #
-# Version 3.7.
+# Version 4.0.
 #
 # You might fail to appreciate this module performance from the first
 # try. If compared to "vanilla" linux-ia32-icc target, i.e. considered
 # and decrypt performance becomes asymmetric. Encrypt performance
 # drops by 10-12%, while decrypt - by 20-25%:-( 256 bytes S-box is
 # aggressively pre-fetched.
+#
+# Version 4.0 effectively rolls back to 3.6 and instead implements
+# additional set of functions, _[x86|mmx]_AES_[en|de]crypt_compact,
+# which use exclusively 256 byte S-box. These functions are to be
+# called in modes not concealing plain text, such as ECB, or when
+# we're asked to process smaller amount of data [or unconditionally
+# on hyper-threading CPU]. Currently it's called unconditionally from
+# AES_[en|de]crypt, which affects all modes, but CBC. CBC routine
+# still needs to be modified to switch between slower and faster
+# mode when appropriate... But in either case benchmark landscape
+# changes dramatically and below numbers are CPU cycles per processed
+# byte for 128-bit key.
+#
+#              ECB encrypt     ECB decrypt     CBC large chunk
+# P4           57[60]          84[100]         23
+# AMD K8       48[44]          70[79]          18
+# PIII         41[50]          61[91]          24
+# Pentium      120             160             77
 
 push(@INC,"perlasm","../../perlasm");
 require "x86asm.pl";
@@ -100,6 +118,9 @@ $s2="ecx";
 $s3="edx";
 $key="edi";
 $acc="esi";
+$tbl="ebp";
+
+sub _data_word() { my $i; while(defined($i=shift)) { &data_word($i,$i); } }
 
 $compromise=0;         # $compromise=128 abstains from copying key
                        # schedule to stack when encrypting inputs
@@ -112,9 +133,9 @@ $small_footprint=1; # $small_footprint=1 code is ~5% slower [on
                        # I favor compact code to minimize cache
                        # contention and in hope to "collect" 5% back
                        # in real-life applications...
+
 $vertical_spin=0;      # shift "verticaly" defaults to 0, because of
                        # its proof-of-concept status...
-
 # Note that there is no decvert(), as well as last encryption round is
 # performed with "horizontal" shifts. This is because this "vertical"
 # implementation [one which groups shifts on a given $s[i] to form a
@@ -181,73 +202,156 @@ sub encvert()
        &xor    ($s[2],&DWP(1,$te,$v0,8));              # s1>>24
 }
 
-sub encstep()
-{ my ($i,$te,@s) = @_;
-  my $tmp = $key;
-  my $out = $i==3?$s[0]:$acc;
-
-       # lines marked with #%e?x[i] denote "reordered" instructions...
-       if ($i==3)  {   &mov    ($key,&DWP(20,"esp"));          }##%edx
-       else        {   &mov    ($out,$s[0]);
-                       &and    ($out,0xFF);                    }
-       if ($i==1)  {   &shr    ($s[0],16);                     }#%ebx[1]
-       if ($i==2)  {   &shr    ($s[0],24);                     }#%ecx[2]
-                       &mov    ($out,&DWP(0,$te,$out,8));
-
-       if ($i==3)  {   $tmp=$s[1];                             }##%eax
-                       &movz   ($tmp,&HB($s[1]));
-                       &xor    ($out,&DWP(3,$te,$tmp,8));
-
-       if ($i==3)  {   $tmp=$s[2]; &mov ($s[1],&DWP(4,"esp")); }##%ebx
-       else        {   &mov    ($tmp,$s[2]);
-                       &shr    ($tmp,16);                      }
-       if ($i==2)  {   &and    ($s[1],0xFF);                   }#%edx[2]
-                       &and    ($tmp,0xFF);
-                       &xor    ($out,&DWP(2,$te,$tmp,8));
+# Another experimental routine, which features "horizontal spin," but
+# eliminates one reference to stack. Strangely enough runs slower...
+sub enchoriz()
+{ my $v0 = $key, $v1 = $acc;
+
+       &movz   ($v0,&LB($s0));                 #  3, 2, 1, 0*
+       &rotr   ($s2,8);                        #  8,11,10, 9
+       &mov    ($v1,&DWP(0,$te,$v0,8));        #  0
+       &movz   ($v0,&HB($s1));                 #  7, 6, 5*, 4
+       &rotr   ($s3,16);                       # 13,12,15,14
+       &xor    ($v1,&DWP(3,$te,$v0,8));        #  5
+       &movz   ($v0,&HB($s2));                 #  8,11,10*, 9
+       &rotr   ($s0,16);                       #  1, 0, 3, 2
+       &xor    ($v1,&DWP(2,$te,$v0,8));        # 10
+       &movz   ($v0,&HB($s3));                 # 13,12,15*,14
+       &xor    ($v1,&DWP(1,$te,$v0,8));        # 15, t[0] collected
+       &mov    (&DWP(4,"esp"),$v1);            # t[0] saved
+
+       &movz   ($v0,&LB($s1));                 #  7, 6, 5, 4*
+       &shr    ($s1,16);                       #  -, -, 7, 6
+       &mov    ($v1,&DWP(0,$te,$v0,8));        #  4
+       &movz   ($v0,&LB($s3));                 # 13,12,15,14*
+       &xor    ($v1,&DWP(2,$te,$v0,8));        # 14
+       &movz   ($v0,&HB($s0));                 #  1, 0, 3*, 2
+       &and    ($s3,0xffff0000);               # 13,12, -, -
+       &xor    ($v1,&DWP(1,$te,$v0,8));        #  3
+       &movz   ($v0,&LB($s2));                 #  8,11,10, 9*
+       &or     ($s3,$s1);                      # 13,12, 7, 6
+       &xor    ($v1,&DWP(3,$te,$v0,8));        #  9, t[1] collected
+       &mov    ($s1,$v1);                      #  s[1]=t[1]
+
+       &movz   ($v0,&LB($s0));                 #  1, 0, 3, 2*
+       &shr    ($s2,16);                       #  -, -, 8,11
+       &mov    ($v1,&DWP(2,$te,$v0,8));        #  2
+       &movz   ($v0,&HB($s3));                 # 13,12, 7*, 6
+       &xor    ($v1,&DWP(1,$te,$v0,8));        #  7
+       &movz   ($v0,&HB($s2));                 #  -, -, 8*,11
+       &xor    ($v1,&DWP(0,$te,$v0,8));        #  8
+       &mov    ($v0,$s3);
+       &shr    ($v0,24);                       # 13
+       &xor    ($v1,&DWP(3,$te,$v0,8));        # 13, t[2] collected
+
+       &movz   ($v0,&LB($s2));                 #  -, -, 8,11*
+       &shr    ($s0,24);                       #  1*
+       &mov    ($s2,&DWP(1,$te,$v0,8));        # 11
+       &xor    ($s2,&DWP(3,$te,$s0,8));        #  1
+       &mov    ($s0,&DWP(4,"esp"));            # s[0]=t[0]
+       &movz   ($v0,&LB($s3));                 # 13,12, 7, 6*
+       &shr    ($s3,16);                       #   ,  ,13,12
+       &xor    ($s2,&DWP(2,$te,$v0,8));        #  6
+       &mov    ($key,&DWP(20,"esp"));          # reincarnate v0 as key
+       &and    ($s3,0xff);                     #   ,  ,13,12*
+       &mov    ($s3,&DWP(0,$te,$s3,8));        # 12
+       &xor    ($s3,$s2);                      # s[2]=t[3] collected
+       &mov    ($s2,$v1);                      # s[2]=t[2]
+}
 
-       if ($i==3)  {   $tmp=$s[3]; &mov ($s[2],&DWP(8,"esp")); }##%ecx
-       elsif($i==2){   &movz   ($tmp,&HB($s[3]));              }#%ebx[2]
-       else        {   &mov    ($tmp,$s[3]); 
-                       &shr    ($tmp,24)                       }
-                       &xor    ($out,&DWP(1,$te,$tmp,8));
-       if ($i<2)   {   &mov    (&DWP(4+4*$i,"esp"),$out);      }
-       if ($i==3)  {   &mov    ($s[3],$acc);                   }
-                       &comment();
+# More experimental code... MMX one... Even though this one eliminates
+# *all* references to stack, it's not faster...
+sub mmx_encbody()
+{
+       &movz   ("esi",&LB("eax"));             #  0
+       &mov    ("ecx",&DWP(0,$tbl,"esi",8));   #  0
+       &pshufw ("mm2","mm0",0x0d);             #  7, 6, 3, 2
+       &movz   ("edx",&HB("eax"));             #  1
+       &mov    ("edx",&DWP(3,$tbl,"edx",8));   #  1
+       &shr    ("eax",16);                     #  5, 4
+
+       &movz   ("esi",&LB("ebx"));             # 10
+       &xor    ("ecx",&DWP(2,$tbl,"esi",8));   # 10
+       &pshufw ("mm6","mm4",0x08);             # 13,12, 9, 8
+       &movz   ("esi",&HB("ebx"));             # 11
+       &xor    ("edx",&DWP(1,$tbl,"esi",8));   # 11
+       &shr    ("ebx",16);                     # 15,14
+
+       &movz   ("esi",&HB("eax"));             #  5
+       &xor    ("ecx",&DWP(3,$tbl,"esi",8));   #  5
+       &movq   ("mm3",QWP(16,$key));
+       &movz   ("esi",&HB("ebx"));             # 15
+       &xor    ("ecx",&DWP(1,$tbl,"esi",8));   # 15
+       &movd   ("mm0","ecx");                  # t[0] collected
+
+       &movz   ("esi",&LB("eax"));             #  4
+       &mov    ("ecx",&DWP(0,$tbl,"esi",8));   #  4
+       &movd   ("eax","mm2");                  #  7, 6, 3, 2
+       &movz   ("esi",&LB("ebx"));             # 14
+       &xor    ("ecx",&DWP(2,$tbl,"esi",8));   # 14
+       &movd   ("ebx","mm6");                  # 13,12, 9, 8
+
+       &movz   ("esi",&HB("eax"));             #  3
+       &xor    ("ecx",&DWP(1,$tbl,"esi",8));   #  3
+       &movz   ("esi",&HB("ebx"));             #  9
+       &xor    ("ecx",&DWP(3,$tbl,"esi",8));   #  9
+       &movd   ("mm1","ecx");                  # t[1] collected
+
+       &movz   ("esi",&LB("eax"));             #  2
+       &mov    ("ecx",&DWP(2,$tbl,"esi",8));   #  2
+       &shr    ("eax",16);                     #  7, 6
+       &punpckldq      ("mm0","mm1");          # t[0,1] collected
+       &movz   ("esi",&LB("ebx"));             #  8
+       &xor    ("ecx",&DWP(0,$tbl,"esi",8));   #  8
+       &shr    ("ebx",16);                     # 13,12
+
+       &movz   ("esi",&HB("eax"));             #  7
+       &xor    ("ecx",&DWP(1,$tbl,"esi",8));   #  7
+       &pxor   ("mm0","mm3");
+       &movz   ("eax",&LB("eax"));             #  6
+       &xor    ("edx",&DWP(2,$tbl,"eax",8));   #  6
+       &pshufw ("mm1","mm0",0x08);             #  5, 4, 1, 0
+       &movz   ("esi",&HB("ebx"));             # 13
+       &xor    ("ecx",&DWP(3,$tbl,"esi",8));   # 13
+       &xor    ("ecx",&DWP(24,$key));          # t[2]
+       &movd   ("mm4","ecx");                  # t[2] collected
+       &movz   ("ebx",&LB("ebx"));             # 12
+       &xor    ("edx",&DWP(0,$tbl,"ebx",8));   # 12
+       &shr    ("ecx",16);
+       &movd   ("eax","mm1");                  #  5, 4, 1, 0
+       &mov    ("ebx",&DWP(28,$key));          # t[3]
+       &xor    ("ebx","edx");
+       &movd   ("mm5","ebx");                  # t[3] collected
+       &and    ("ebx",0xffff0000);
+       &or     ("ebx","ecx");
+
+       &punpckldq      ("mm4","mm5");          # t[2,3] collected
 }
 
-sub enclast()
+######################################################################
+# "Compact" block function
+######################################################################
+
+sub enccompact()
 { my $Fn = mov;
   while ($#_>5) { pop(@_); $Fn=sub{}; }
   my ($i,$te,@s)=@_;
   my $tmp = $key;
   my $out = $i==3?$s[0]:$acc;
 
-       if ($i==0)      # prefetch 256-byte Te4
-       {   &lea        ($te,&DWP(2048+128,$te));
-           &mov        ($tmp,&DWP(0-128,$te));
-           &mov        ($acc,&DWP(32-128,$te));
-           &mov        ($tmp,&DWP(64-128,$te));
-           &mov        ($acc,&DWP(96-128,$te));
-           &mov        ($tmp,&DWP(128-128,$te));
-           &mov        ($acc,&DWP(160-128,$te));
-           &mov        ($tmp,&DWP(192-128,$te));
-           &mov        ($acc,&DWP(224-128,$te));
-           &lea        ($te,&DWP(-128,$te));
-       }
-
        # $Fn is used in first compact round and its purpose is to
        # void restoration of some values from stack, so that after
-       # 4xenclast with extra argument $key value is left there...
+       # 4xenccompact with extra argument $key value is left there...
        if ($i==3)  {   &$Fn    ($key,&DWP(20,"esp"));          }##%edx
        else        {   &mov    ($out,$s[0]);                   }
                        &and    ($out,0xFF);
        if ($i==1)  {   &shr    ($s[0],16);                     }#%ebx[1]
        if ($i==2)  {   &shr    ($s[0],24);                     }#%ecx[2]
-                       &movz   ($out,&DWP(0,$te,$out,1));
+                       &movz   ($out,&DWP(-128,$te,$out,1));
 
        if ($i==3)  {   $tmp=$s[1];                             }##%eax
                        &movz   ($tmp,&HB($s[1]));
-                       &movz   ($tmp,&DWP(0,$te,$tmp,1));
+                       &movz   ($tmp,&DWP(-128,$te,$tmp,1));
                        &shl    ($tmp,8);
                        &xor    ($out,$tmp);
 
@@ -256,7 +360,7 @@ sub enclast()
                        &shr    ($tmp,16);                      }
        if ($i==2)  {   &and    ($s[1],0xFF);                   }#%edx[2]
                        &and    ($tmp,0xFF);
-                       &movz   ($tmp,&DWP(0,$te,$tmp,1));
+                       &movz   ($tmp,&DWP(-128,$te,$tmp,1));
                        &shl    ($tmp,16);
                        &xor    ($out,$tmp);
 
@@ -264,20 +368,19 @@ sub enclast()
        elsif($i==2){   &movz   ($tmp,&HB($s[3]));              }#%ebx[2]
        else        {   &mov    ($tmp,$s[3]);
                        &shr    ($tmp,24);                      }
-                       &movz   ($tmp,&DWP(0,$te,$tmp,1));
+                       &movz   ($tmp,&DWP(-128,$te,$tmp,1));
                        &shl    ($tmp,24);
                        &xor    ($out,$tmp);
        if ($i<2)   {   &mov    (&DWP(4+4*$i,"esp"),$out);      }
-       if ($i==3)  {   &mov    ($s[3],$acc);
-                       &lea    ($te,&DWP(-2048,$te));          }
+       if ($i==3)  {   &mov    ($s[3],$acc);                   }
        &comment();
 }
 
 sub enctransform()
 { my @s = ($s0,$s1,$s2,$s3);
   my $i = shift;
-  my $tmp = $key;
-  my $r2 = "ebp";
+  my $tmp = $tbl;
+  my $r2  = $key ;
 
        &mov    ($acc,$s[$i]);
        &and    ($acc,0x80808080);
@@ -289,18 +392,317 @@ sub enctransform()
        &lea    ($r2,&DWP(0,$r2,$r2));
        &and    ($acc,0x1b1b1b1b);
        &mov    ($tmp,$s[$i]);
-       &xor    ($r2,$acc);     # r2
+       &xor    ($acc,$r2);     # r2
 
-       &xor    ($s[$i],$r2);   # r0 ^ r2
+       &xor    ($s[$i],$acc);  # r0 ^ r2
        &rotl   ($s[$i],24);
-       &xor    ($s[$i],$r2)    # ROTATE(r2^r0,24) ^ r2
+       &xor    ($s[$i],$acc)   # ROTATE(r2^r0,24) ^ r2
        &rotr   ($tmp,16);
        &xor    ($s[$i],$tmp);
        &rotr   ($tmp,8);
        &xor    ($s[$i],$tmp);
 }
 
-sub enclast_large()
+&public_label("AES_Te");
+&function_begin_B("_x86_AES_encrypt_compact");
+       # note that caller is expected to allocate stack frame for me!
+       &mov    (&DWP(20,"esp"),$key);          # save key
+
+       &xor    ($s0,&DWP(0,$key));             # xor with key
+       &xor    ($s1,&DWP(4,$key));
+       &xor    ($s2,&DWP(8,$key));
+       &xor    ($s3,&DWP(12,$key));
+
+       &mov    ($acc,&DWP(240,$key));          # load key->rounds
+       &lea    ($acc,&DWP(-2,$acc,$acc));
+       &lea    ($acc,&DWP(0,$key,$acc,8));
+       &mov    (&DWP(24,"esp"),$acc);          # end of key schedule
+
+       # prefetch Te4
+       &mov    ($key,&DWP(0-128,$tbl));
+       &mov    ($acc,&DWP(32-128,$tbl));
+       &mov    ($key,&DWP(64-128,$tbl));
+       &mov    ($acc,&DWP(96-128,$tbl));
+       &mov    ($key,&DWP(128-128,$tbl));
+       &mov    ($acc,&DWP(160-128,$tbl));
+       &mov    ($key,&DWP(192-128,$tbl));
+       &mov    ($acc,&DWP(224-128,$tbl));
+
+       &set_label("loop",16);
+
+               &enccompact(0,$tbl,$s0,$s1,$s2,$s3,1);
+               &enccompact(1,$tbl,$s1,$s2,$s3,$s0,1);
+               &enccompact(2,$tbl,$s2,$s3,$s0,$s1,1);
+               &enccompact(3,$tbl,$s3,$s0,$s1,$s2,1);
+               &enctransform(2);
+               &enctransform(3);
+               &enctransform(0);
+               &enctransform(1);
+               &mov    ($key,&DWP(20,"esp"));
+               &mov    ($tbl,&DWP(28,"esp"));
+               &add    ($key,16);              # advance rd_key
+               &xor    ($s0,&DWP(0,$key));
+               &xor    ($s1,&DWP(4,$key));
+               &xor    ($s2,&DWP(8,$key));
+               &xor    ($s3,&DWP(12,$key));
+
+       &cmp    ($key,&DWP(24,"esp"));
+       &mov    (&DWP(20,"esp"),$key);
+       &jb     (&label("loop"));
+
+       &enccompact(0,$tbl,$s0,$s1,$s2,$s3);
+       &enccompact(1,$tbl,$s1,$s2,$s3,$s0);
+       &enccompact(2,$tbl,$s2,$s3,$s0,$s1);
+       &enccompact(3,$tbl,$s3,$s0,$s1,$s2);
+
+       &xor    ($s0,&DWP(16,$key));
+       &xor    ($s1,&DWP(20,$key));
+       &xor    ($s2,&DWP(24,$key));
+       &xor    ($s3,&DWP(28,$key));
+
+       &ret    ();
+&function_end_B("_x86_AES_encrypt_compact");
+
+######################################################################
+# "Compact" MMX block function.
+######################################################################
+#
+# Performance is not actually extraordinary in comparison to pure
+# x86 code. In particular encrypt performance is virtually the same.
+# same. Decrypt performance on the other hand is 15-20% better on
+# newer Âµ-archs [but we're thankful for *any* improvement here], and
+# ~50% better on PIII:-) And additionally on the pros side this code
+# eliminates redundant references to stack and thus relieves/
+# minimizes the pressure on the memory bus.
+#
+# MMX register layout                           lsb
+# +--+--+--+--+--+--+--+--+--+--+--+--+--+--+--+--+
+# |          mm4          |          mm0          |
+# +--+--+--+--+--+--+--+--+--+--+--+--+--+--+--+--+
+# |     s3    |     s2    |     s1    |     s0    |    
+# +--+--+--+--+--+--+--+--+--+--+--+--+--+--+--+--+
+# |15|14|13|12|11|10| 9| 8| 7| 6| 5| 4| 3| 2| 1| 0|
+# +--+--+--+--+--+--+--+--+--+--+--+--+--+--+--+--+
+#
+# Indexes translate as s[N/4]>>(8*(N%4)), e.g. 5 means s1>>8.
+# In this terms encryption and decryption "compact" permutation
+# matrices can be depicted as following:
+#
+# encryption              lsb  # decryption              lsb
+# +----++----+----+----+----+  # +----++----+----+----+----+
+# | t0 || 15 | 10 |  5 |  0 |  # | t0 ||  7 | 10 | 13 |  0 |
+# +----++----+----+----+----+  # +----++----+----+----+----+
+# | t1 ||  3 | 14 |  9 |  4 |  # | t1 || 11 | 14 |  1 |  4 |
+# +----++----+----+----+----+  # +----++----+----+----+----+
+# | t2 ||  7 |  2 | 13 |  8 |  # | t2 || 15 |  2 |  5 |  8 |
+# +----++----+----+----+----+  # +----++----+----+----+----+
+# | t3 || 11 |  6 |  1 | 12 |  # | t3 ||  3 |  6 |  9 | 12 |
+# +----++----+----+----+----+  # +----++----+----+----+----+
+#
+######################################################################
+# Why not xmm registers? Short answer. It was actually tested and
+# was not any faster, but *contrary*, most notably on Intel CPUs.
+# Longer answer. Main advantage of using mm registers is that movd
+# latency is lower, especially on Intel P4. While arithmetic
+# instructions are twice as many, they can be scheduled every cycle
+# and not every second one when they are operating on xmm register,
+# so that "arithmetic throughput" remains virtually the same. And
+# finally the code can be executed even on elder MMX-only CPUs:-)
+
+sub mmx_enccompact()
+{
+       &pshufw ("mm1","mm0",0x08);             #  5, 4, 1, 0
+       &pshufw ("mm5","mm4",0x0d);             # 15,14,11,10
+       &movd   ("eax","mm1");                  #  5, 4, 1, 0
+       &movd   ("ebx","mm5");                  # 15,14,11,10
+
+       &movz   ("esi",&LB("eax"));             #  0
+       &movz   ("ecx",&DWP(-128,$tbl,"esi",1));#  0
+       &pshufw ("mm2","mm0",0x0d);             #  7, 6, 3, 2
+       &movz   ("edx",&HB("eax"));             #  1
+       &movz   ("edx",&DWP(-128,$tbl,"edx",1));#  1
+       &shl    ("edx",8);                      #  1
+       &shr    ("eax",16);                     #  5, 4
+
+       &movz   ("esi",&LB("ebx"));             # 10
+       &movz   ("esi",&DWP(-128,$tbl,"esi",1));# 10
+       &shl    ("esi",16);                     # 10
+       &or     ("ecx","esi");                  # 10
+       &pshufw ("mm6","mm4",0x08);             # 13,12, 9, 8
+       &movz   ("esi",&HB("ebx"));             # 11
+       &movz   ("esi",&DWP(-128,$tbl,"esi",1));# 11
+       &shl    ("esi",24);                     # 11
+       &or     ("edx","esi");                  # 11
+       &shr    ("ebx",16);                     # 15,14
+
+       &movz   ("esi",&HB("eax"));             #  5
+       &movz   ("esi",&DWP(-128,$tbl,"esi",1));#  5
+       &shl    ("esi",8);                      #  5
+       &or     ("ecx","esi");                  #  5
+       &movz   ("esi",&HB("ebx"));             # 15
+       &movz   ("esi",&DWP(-128,$tbl,"esi",1));# 15
+       &shl    ("esi",24);                     # 15
+       &or     ("ecx","esi");                  # 15
+       &movd   ("mm0","ecx");                  # t[0] collected
+
+       &movz   ("esi",&LB("eax"));             #  4
+       &movz   ("ecx",&DWP(-128,$tbl,"esi",1));#  4
+       &movd   ("eax","mm2");                  #  7, 6, 3, 2
+       &movz   ("esi",&LB("ebx"));             # 14
+       &movz   ("esi",&DWP(-128,$tbl,"esi",1));# 14
+       &shl    ("esi",16);                     # 14
+       &or     ("ecx","esi");                  # 14
+
+       &movd   ("ebx","mm6");                  # 13,12, 9, 8
+       &movz   ("esi",&HB("eax"));             #  3
+       &movz   ("esi",&DWP(-128,$tbl,"esi",1));#  3
+       &shl    ("esi",24);                     #  3
+       &or     ("ecx","esi");                  #  3
+       &movz   ("esi",&HB("ebx"));             #  9
+       &movz   ("esi",&DWP(-128,$tbl,"esi",1));#  9
+       &shl    ("esi",8);                      #  9
+       &or     ("ecx","esi");                  #  9
+       &movd   ("mm1","ecx");                  # t[1] collected
+
+       &movz   ("esi",&LB("ebx"));             #  8
+       &movz   ("ecx",&DWP(-128,$tbl,"esi",1));#  8
+       &shr    ("ebx",16);                     # 13,12
+       &movz   ("esi",&LB("eax"));             #  2
+       &movz   ("esi",&DWP(-128,$tbl,"esi",1));#  2
+       &shl    ("esi",16);                     #  2
+       &or     ("ecx","esi");                  #  2
+       &shr    ("eax",16);                     #  7, 6
+
+       &punpckldq      ("mm0","mm1");          # t[0,1] collected
+
+       &movz   ("esi",&HB("eax"));             #  7
+       &movz   ("esi",&DWP(-128,$tbl,"esi",1));#  7
+       &shl    ("esi",24);                     #  7
+       &or     ("ecx","esi");                  #  7
+       &and    ("eax",0xff);                   #  6
+       &movz   ("eax",&DWP(-128,$tbl,"eax",1));#  6
+       &shl    ("eax",16);                     #  6
+       &or     ("edx","eax");                  #  6
+       &movz   ("esi",&HB("ebx"));             # 13
+       &movz   ("esi",&DWP(-128,$tbl,"esi",1));# 13
+       &shl    ("esi",8);                      # 13
+       &or     ("ecx","esi");                  # 13
+       &movd   ("mm4","ecx");                  # t[2] collected
+       &and    ("ebx",0xff);                   # 12
+       &movz   ("ebx",&DWP(-128,$tbl,"ebx",1));# 12
+       &or     ("edx","ebx");                  # 12
+       &movd   ("mm5","edx");                  # t[3] collected
+
+       &punpckldq      ("mm4","mm5");          # t[2,3] collected
+}
+
+&public_label("AES_Te");
+&function_begin_B("_mmx_AES_encrypt_compact");
+       &pxor   ("mm0",&QWP(0,$key));   #  7, 6, 5, 4, 3, 2, 1, 0
+       &pxor   ("mm4",&QWP(8,$key));   # 15,14,13,12,11,10, 9, 8
+
+       # note that caller is expected to allocate stack frame for me!
+       &mov    ($acc,&DWP(240,$key));          # load key->rounds
+       &lea    ($acc,&DWP(-2,$acc,$acc));
+       &lea    ($acc,&DWP(0,$key,$acc,8));
+       &mov    (&DWP(24,"esp"),$acc);          # end of key schedule
+
+       &mov    ($s0,0x1b1b1b1b);               # magic constant
+       &mov    (&DWP(8,"esp"),$s0);
+       &mov    (&DWP(12,"esp"),$s0);
+
+       # prefetch Te4
+       &mov    ($s0,&DWP(0-128,$tbl));
+       &mov    ($s1,&DWP(32-128,$tbl));
+       &mov    ($s2,&DWP(64-128,$tbl));
+       &mov    ($s3,&DWP(96-128,$tbl));
+       &mov    ($s0,&DWP(128-128,$tbl));
+       &mov    ($s1,&DWP(160-128,$tbl));
+       &mov    ($s2,&DWP(192-128,$tbl));
+       &mov    ($s3,&DWP(224-128,$tbl));
+
+       &set_label("loop",16);
+               &mmx_enccompact();
+               &add    ($key,16);
+               &cmp    ($key,&DWP(24,"esp"));
+               &ja     (&label("out"));
+
+               &movq   ("mm2",&QWP(8,"esp"));
+               &pxor   ("mm3","mm3");          &pxor   ("mm7","mm7");
+               &movq   ("mm1","mm0");          &movq   ("mm5","mm4");  # r0
+               &pcmpgtb("mm3","mm0");          &pcmpgtb("mm7","mm4");
+               &pand   ("mm3","mm2");          &pand   ("mm7","mm2");
+               &movq   ("mm2","mm0");          &movq   ("mm6","mm4");  # r0
+               &paddb  ("mm0","mm0");          &paddb  ("mm4","mm4");
+               &pxor   ("mm0","mm3");          &pxor   ("mm4","mm7");  # = r2
+               &movq   ("mm3","mm2");          &movq   ("mm7","mm6");
+               &pxor   ("mm1","mm0");          &pxor   ("mm5","mm4");  # r2^r0
+
+               &pslld  ("mm3",8);              &pslld  ("mm7",8);
+               &psrld  ("mm2",16);             &psrld  ("mm6",16);
+               &pxor   ("mm0","mm3");          &pxor   ("mm4","mm7");  # ^= r0<<8
+               &pxor   ("mm0","mm2");          &pxor   ("mm4","mm6");  # ^= r0>>16
+               &pslld  ("mm3",8);              &pslld  ("mm7",8);
+               &psrld  ("mm2",8);              &psrld  ("mm6",8);
+               &pxor   ("mm0","mm3");          &pxor   ("mm4","mm7");  # ^= r0<<16
+               &movq   ("mm3","mm1");          &movq   ("mm7","mm5");
+               &pxor   ("mm0","mm2");          &pxor   ("mm4","mm6");  # ^= r0>>24
+
+               &psrld  ("mm1",8);              &psrld  ("mm5",8);
+               &movq   ("mm2",&QWP(0,$key));   &movq   ("mm6",&QWP(8,$key));
+               &pslld  ("mm3",24);             &pslld  ("mm7",24);
+               &pxor   ("mm0","mm1");          &pxor   ("mm4","mm5");  # ^= (r2^r0)<<8
+               &pxor   ("mm0","mm3");          &pxor   ("mm4","mm7");  # ^= (r2^r0)>>24
+
+               &pxor   ("mm0","mm2");          &pxor   ("mm4","mm6");
+       &jmp    (&label("loop"));
+
+       &set_label("out",16);
+       &pxor   ("mm0",&QWP(0,$key));
+       &pxor   ("mm4",&QWP(8,$key));
+
+       &ret    ();
+&function_end_B("_mmx_AES_encrypt_compact");
+
+######################################################################
+# Vanilla block function.
+######################################################################
+
+sub encstep()
+{ my ($i,$te,@s) = @_;
+  my $tmp = $key;
+  my $out = $i==3?$s[0]:$acc;
+
+       # lines marked with #%e?x[i] denote "reordered" instructions...
+       if ($i==3)  {   &mov    ($key,&DWP(20,"esp"));          }##%edx
+       else        {   &mov    ($out,$s[0]);
+                       &and    ($out,0xFF);                    }
+       if ($i==1)  {   &shr    ($s[0],16);                     }#%ebx[1]
+       if ($i==2)  {   &shr    ($s[0],24);                     }#%ecx[2]
+                       &mov    ($out,&DWP(0,$te,$out,8));
+
+       if ($i==3)  {   $tmp=$s[1];                             }##%eax
+                       &movz   ($tmp,&HB($s[1]));
+                       &xor    ($out,&DWP(3,$te,$tmp,8));
+
+       if ($i==3)  {   $tmp=$s[2]; &mov ($s[1],&DWP(4,"esp")); }##%ebx
+       else        {   &mov    ($tmp,$s[2]);
+                       &shr    ($tmp,16);                      }
+       if ($i==2)  {   &and    ($s[1],0xFF);                   }#%edx[2]
+                       &and    ($tmp,0xFF);
+                       &xor    ($out,&DWP(2,$te,$tmp,8));
+
+       if ($i==3)  {   $tmp=$s[3]; &mov ($s[2],&DWP(8,"esp")); }##%ecx
+       elsif($i==2){   &movz   ($tmp,&HB($s[3]));              }#%ebx[2]
+       else        {   &mov    ($tmp,$s[3]); 
+                       &shr    ($tmp,24)                       }
+                       &xor    ($out,&DWP(1,$te,$tmp,8));
+       if ($i<2)   {   &mov    (&DWP(4+4*$i,"esp"),$out);      }
+       if ($i==3)  {   &mov    ($s[3],$acc);                   }
+                       &comment();
+}
+
+sub enclast()
 { my ($i,$te,@s)=@_;
   my $tmp = $key;
   my $out = $i==3?$s[0]:$acc;
@@ -339,8 +741,6 @@ sub enclast_large()
        if ($i==3)  {   &mov    ($s[3],$acc);                   }
 }
 
-sub _data_word() { my $i; while(defined($i=shift)) { &data_word($i,$i); } }
-
 &public_label("AES_Te");
 &function_begin_B("_x86_AES_encrypt");
        if ($vertical_spin) {
@@ -357,40 +757,21 @@ sub _data_word() { my $i; while(defined($i=shift)) { &data_word($i,$i); } }
        &xor    ($s2,&DWP(8,$key));
        &xor    ($s3,&DWP(12,$key));
 
-       # not really last round, just "compact" one...
-       &enclast(0,"ebp",$s0,$s1,$s2,$s3,1);
-       &enclast(1,"ebp",$s1,$s2,$s3,$s0,1);
-       &enclast(2,"ebp",$s2,$s3,$s0,$s1,1);
-       &enclast(3,"ebp",$s3,$s0,$s1,$s2,1);
-       &enctransform(2);
-       &enctransform(3);
-       &enctransform(0);
-       &enctransform(1);
-       &mov    ($key,&DWP(20,"esp"));
-       &mov    ("ebp",&DWP(28,"esp"));
-       &xor    ($s0,&DWP(16,$key));
-       &xor    ($s1,&DWP(20,$key));
-       &xor    ($s2,&DWP(24,$key));
-       &xor    ($s3,&DWP(28,$key));
-
        &mov    ($acc,&DWP(240,$key));          # load key->rounds
 
        if ($small_footprint) {
            &lea        ($acc,&DWP(-2,$acc,$acc));
            &lea        ($acc,&DWP(0,$key,$acc,8));
-           &add        ($key,16);
-           &mov        (&DWP(20,"esp"),$key);
            &mov        (&DWP(24,"esp"),$acc);  # end of key schedule
 
-           &align      (4);
-           &set_label("loop");
+           &set_label("loop",16);
                if ($vertical_spin) {
-                   &encvert("ebp",$s0,$s1,$s2,$s3);
+                   &encvert($tbl,$s0,$s1,$s2,$s3);
                } else {
-                   &encstep(0,"ebp",$s0,$s1,$s2,$s3);
-                   &encstep(1,"ebp",$s1,$s2,$s3,$s0);
-                   &encstep(2,"ebp",$s2,$s3,$s0,$s1);
-                   &encstep(3,"ebp",$s3,$s0,$s1,$s2);
+                   &encstep(0,$tbl,$s0,$s1,$s2,$s3);
+                   &encstep(1,$tbl,$s1,$s2,$s3,$s0);
+                   &encstep(2,$tbl,$s2,$s3,$s0,$s1);
+                   &encstep(3,$tbl,$s3,$s0,$s1,$s2);
                }
                &add    ($key,16);              # advance rd_key
                &xor    ($s0,&DWP(0,$key));
@@ -407,15 +788,15 @@ sub _data_word() { my $i; while(defined($i=shift)) { &data_word($i,$i); } }
            &cmp        ($acc,12);
            &jle        (&label("12rounds"));
 
-       &set_label("14rounds");
-           for ($i=2;$i<4;$i++) {
+       &set_label("14rounds",4);
+           for ($i=1;$i<3;$i++) {
                if ($vertical_spin) {
-                   &encvert("ebp",$s0,$s1,$s2,$s3);
+                   &encvert($tbl,$s0,$s1,$s2,$s3);
                } else {
-                   &encstep(0,"ebp",$s0,$s1,$s2,$s3);
-                   &encstep(1,"ebp",$s1,$s2,$s3,$s0);
-                   &encstep(2,"ebp",$s2,$s3,$s0,$s1);
-                   &encstep(3,"ebp",$s3,$s0,$s1,$s2);
+                   &encstep(0,$tbl,$s0,$s1,$s2,$s3);
+                   &encstep(1,$tbl,$s1,$s2,$s3,$s0);
+                   &encstep(2,$tbl,$s2,$s3,$s0,$s1);
+                   &encstep(3,$tbl,$s3,$s0,$s1,$s2);
                }
                &xor    ($s0,&DWP(16*$i+0,$key));
                &xor    ($s1,&DWP(16*$i+4,$key));
@@ -424,15 +805,15 @@ sub _data_word() { my $i; while(defined($i=shift)) { &data_word($i,$i); } }
            }
            &add        ($key,32);
            &mov        (&DWP(20,"esp"),$key);  # advance rd_key
-       &set_label("12rounds");
-           for ($i=2;$i<4;$i++) {
+       &set_label("12rounds",4);
+           for ($i=1;$i<3;$i++) {
                if ($vertical_spin) {
-                   &encvert("ebp",$s0,$s1,$s2,$s3);
+                   &encvert($tbl,$s0,$s1,$s2,$s3);
                } else {
-                   &encstep(0,"ebp",$s0,$s1,$s2,$s3);
-                   &encstep(1,"ebp",$s1,$s2,$s3,$s0);
-                   &encstep(2,"ebp",$s2,$s3,$s0,$s1);
-                   &encstep(3,"ebp",$s3,$s0,$s1,$s2);
+                   &encstep(0,$tbl,$s0,$s1,$s2,$s3);
+                   &encstep(1,$tbl,$s1,$s2,$s3,$s0);
+                   &encstep(2,$tbl,$s2,$s3,$s0,$s1);
+                   &encstep(3,$tbl,$s3,$s0,$s1,$s2);
                }
                &xor    ($s0,&DWP(16*$i+0,$key));
                &xor    ($s1,&DWP(16*$i+4,$key));
@@ -441,15 +822,15 @@ sub _data_word() { my $i; while(defined($i=shift)) { &data_word($i,$i); } }
            }
            &add        ($key,32);
            &mov        (&DWP(20,"esp"),$key);  # advance rd_key
-       &set_label("10rounds");
-           for ($i=2;$i<10;$i++) {
+       &set_label("10rounds",4);
+           for ($i=1;$i<10;$i++) {
                if ($vertical_spin) {
-                   &encvert("ebp",$s0,$s1,$s2,$s3);
+                   &encvert($tbl,$s0,$s1,$s2,$s3);
                } else {
-                   &encstep(0,"ebp",$s0,$s1,$s2,$s3);
-                   &encstep(1,"ebp",$s1,$s2,$s3,$s0);
-                   &encstep(2,"ebp",$s2,$s3,$s0,$s1);
-                   &encstep(3,"ebp",$s3,$s0,$s1,$s2);
+                   &encstep(0,$tbl,$s0,$s1,$s2,$s3);
+                   &encstep(1,$tbl,$s1,$s2,$s3,$s0);
+                   &encstep(2,$tbl,$s2,$s3,$s0,$s1);
+                   &encstep(3,$tbl,$s3,$s0,$s1,$s2);
                }
                &xor    ($s0,&DWP(16*$i+0,$key));
                &xor    ($s1,&DWP(16*$i+4,$key));
@@ -463,10 +844,10 @@ sub _data_word() { my $i; while(defined($i=shift)) { &data_word($i,$i); } }
            &mov        ($s1="ebx",$key="edi");
            &mov        ($s2="ecx",$acc="esi");
        }
-       &enclast(0,"ebp",$s0,$s1,$s2,$s3);
-       &enclast(1,"ebp",$s1,$s2,$s3,$s0);
-       &enclast(2,"ebp",$s2,$s3,$s0,$s1);
-       &enclast(3,"ebp",$s3,$s0,$s1,$s2);
+       &enclast(0,$tbl,$s0,$s1,$s2,$s3);
+       &enclast(1,$tbl,$s1,$s2,$s3,$s0);
+       &enclast(2,$tbl,$s2,$s3,$s0,$s1);
+       &enclast(3,$tbl,$s3,$s0,$s1,$s2);
 
        &add    ($key,$small_footprint?16:160);
        &xor    ($s0,&DWP(0,$key));
@@ -476,7 +857,7 @@ sub _data_word() { my $i; while(defined($i=shift)) { &data_word($i,$i); } }
 
        &ret    ();
 
-&set_label("AES_Te",64);       # Yes! I keep it in the code segment!
+&set_label("AES_Te",1024);     # Yes! I keep it in the code segment!
        &_data_word(0xa56363c6, 0x847c7cf8, 0x997777ee, 0x8d7b7bf6);
        &_data_word(0x0df2f2ff, 0xbd6b6bd6, 0xb16f6fde, 0x54c5c591);
        &_data_word(0x50303060, 0x03010102, 0xa96767ce, 0x7d2b2b56);
@@ -541,7 +922,107 @@ sub _data_word() { my $i; while(defined($i=shift)) { &data_word($i,$i); } }
        &_data_word(0xdabfbf65, 0x31e6e6d7, 0xc6424284, 0xb86868d0);
        &_data_word(0xc3414182, 0xb0999929, 0x772d2d5a, 0x110f0f1e);
        &_data_word(0xcbb0b07b, 0xfc5454a8, 0xd6bbbb6d, 0x3a16162c);
-#Te4
+
+#Te4   # four copies of Te4 to choose from to avoid L1 aliasing
+       &data_byte(0x63, 0x7c, 0x77, 0x7b, 0xf2, 0x6b, 0x6f, 0xc5);
+       &data_byte(0x30, 0x01, 0x67, 0x2b, 0xfe, 0xd7, 0xab, 0x76);
+       &data_byte(0xca, 0x82, 0xc9, 0x7d, 0xfa, 0x59, 0x47, 0xf0);
+       &data_byte(0xad, 0xd4, 0xa2, 0xaf, 0x9c, 0xa4, 0x72, 0xc0);
+       &data_byte(0xb7, 0xfd, 0x93, 0x26, 0x36, 0x3f, 0xf7, 0xcc);
+       &data_byte(0x34, 0xa5, 0xe5, 0xf1, 0x71, 0xd8, 0x31, 0x15);
+       &data_byte(0x04, 0xc7, 0x23, 0xc3, 0x18, 0x96, 0x05, 0x9a);
+       &data_byte(0x07, 0x12, 0x80, 0xe2, 0xeb, 0x27, 0xb2, 0x75);
+       &data_byte(0x09, 0x83, 0x2c, 0x1a, 0x1b, 0x6e, 0x5a, 0xa0);
+       &data_byte(0x52, 0x3b, 0xd6, 0xb3, 0x29, 0xe3, 0x2f, 0x84);
+       &data_byte(0x53, 0xd1, 0x00, 0xed, 0x20, 0xfc, 0xb1, 0x5b);
+       &data_byte(0x6a, 0xcb, 0xbe, 0x39, 0x4a, 0x4c, 0x58, 0xcf);
+       &data_byte(0xd0, 0xef, 0xaa, 0xfb, 0x43, 0x4d, 0x33, 0x85);
+       &data_byte(0x45, 0xf9, 0x02, 0x7f, 0x50, 0x3c, 0x9f, 0xa8);
+       &data_byte(0x51, 0xa3, 0x40, 0x8f, 0x92, 0x9d, 0x38, 0xf5);
+       &data_byte(0xbc, 0xb6, 0xda, 0x21, 0x10, 0xff, 0xf3, 0xd2);
+       &data_byte(0xcd, 0x0c, 0x13, 0xec, 0x5f, 0x97, 0x44, 0x17);
+       &data_byte(0xc4, 0xa7, 0x7e, 0x3d, 0x64, 0x5d, 0x19, 0x73);
+       &data_byte(0x60, 0x81, 0x4f, 0xdc, 0x22, 0x2a, 0x90, 0x88);
+       &data_byte(0x46, 0xee, 0xb8, 0x14, 0xde, 0x5e, 0x0b, 0xdb);
+       &data_byte(0xe0, 0x32, 0x3a, 0x0a, 0x49, 0x06, 0x24, 0x5c);
+       &data_byte(0xc2, 0xd3, 0xac, 0x62, 0x91, 0x95, 0xe4, 0x79);
+       &data_byte(0xe7, 0xc8, 0x37, 0x6d, 0x8d, 0xd5, 0x4e, 0xa9);
+       &data_byte(0x6c, 0x56, 0xf4, 0xea, 0x65, 0x7a, 0xae, 0x08);
+       &data_byte(0xba, 0x78, 0x25, 0x2e, 0x1c, 0xa6, 0xb4, 0xc6);
+       &data_byte(0xe8, 0xdd, 0x74, 0x1f, 0x4b, 0xbd, 0x8b, 0x8a);
+       &data_byte(0x70, 0x3e, 0xb5, 0x66, 0x48, 0x03, 0xf6, 0x0e);
+       &data_byte(0x61, 0x35, 0x57, 0xb9, 0x86, 0xc1, 0x1d, 0x9e);
+       &data_byte(0xe1, 0xf8, 0x98, 0x11, 0x69, 0xd9, 0x8e, 0x94);
+       &data_byte(0x9b, 0x1e, 0x87, 0xe9, 0xce, 0x55, 0x28, 0xdf);
+       &data_byte(0x8c, 0xa1, 0x89, 0x0d, 0xbf, 0xe6, 0x42, 0x68);
+       &data_byte(0x41, 0x99, 0x2d, 0x0f, 0xb0, 0x54, 0xbb, 0x16);
+
+       &data_byte(0x63, 0x7c, 0x77, 0x7b, 0xf2, 0x6b, 0x6f, 0xc5);
+       &data_byte(0x30, 0x01, 0x67, 0x2b, 0xfe, 0xd7, 0xab, 0x76);
+       &data_byte(0xca, 0x82, 0xc9, 0x7d, 0xfa, 0x59, 0x47, 0xf0);
+       &data_byte(0xad, 0xd4, 0xa2, 0xaf, 0x9c, 0xa4, 0x72, 0xc0);
+       &data_byte(0xb7, 0xfd, 0x93, 0x26, 0x36, 0x3f, 0xf7, 0xcc);
+       &data_byte(0x34, 0xa5, 0xe5, 0xf1, 0x71, 0xd8, 0x31, 0x15);
+       &data_byte(0x04, 0xc7, 0x23, 0xc3, 0x18, 0x96, 0x05, 0x9a);
+       &data_byte(0x07, 0x12, 0x80, 0xe2, 0xeb, 0x27, 0xb2, 0x75);
+       &data_byte(0x09, 0x83, 0x2c, 0x1a, 0x1b, 0x6e, 0x5a, 0xa0);
+       &data_byte(0x52, 0x3b, 0xd6, 0xb3, 0x29, 0xe3, 0x2f, 0x84);
+       &data_byte(0x53, 0xd1, 0x00, 0xed, 0x20, 0xfc, 0xb1, 0x5b);
+       &data_byte(0x6a, 0xcb, 0xbe, 0x39, 0x4a, 0x4c, 0x58, 0xcf);
+       &data_byte(0xd0, 0xef, 0xaa, 0xfb, 0x43, 0x4d, 0x33, 0x85);
+       &data_byte(0x45, 0xf9, 0x02, 0x7f, 0x50, 0x3c, 0x9f, 0xa8);
+       &data_byte(0x51, 0xa3, 0x40, 0x8f, 0x92, 0x9d, 0x38, 0xf5);
+       &data_byte(0xbc, 0xb6, 0xda, 0x21, 0x10, 0xff, 0xf3, 0xd2);
+       &data_byte(0xcd, 0x0c, 0x13, 0xec, 0x5f, 0x97, 0x44, 0x17);
+       &data_byte(0xc4, 0xa7, 0x7e, 0x3d, 0x64, 0x5d, 0x19, 0x73);
+       &data_byte(0x60, 0x81, 0x4f, 0xdc, 0x22, 0x2a, 0x90, 0x88);
+       &data_byte(0x46, 0xee, 0xb8, 0x14, 0xde, 0x5e, 0x0b, 0xdb);
+       &data_byte(0xe0, 0x32, 0x3a, 0x0a, 0x49, 0x06, 0x24, 0x5c);
+       &data_byte(0xc2, 0xd3, 0xac, 0x62, 0x91, 0x95, 0xe4, 0x79);
+       &data_byte(0xe7, 0xc8, 0x37, 0x6d, 0x8d, 0xd5, 0x4e, 0xa9);
+       &data_byte(0x6c, 0x56, 0xf4, 0xea, 0x65, 0x7a, 0xae, 0x08);
+       &data_byte(0xba, 0x78, 0x25, 0x2e, 0x1c, 0xa6, 0xb4, 0xc6);
+       &data_byte(0xe8, 0xdd, 0x74, 0x1f, 0x4b, 0xbd, 0x8b, 0x8a);
+       &data_byte(0x70, 0x3e, 0xb5, 0x66, 0x48, 0x03, 0xf6, 0x0e);
+       &data_byte(0x61, 0x35, 0x57, 0xb9, 0x86, 0xc1, 0x1d, 0x9e);
+       &data_byte(0xe1, 0xf8, 0x98, 0x11, 0x69, 0xd9, 0x8e, 0x94);
+       &data_byte(0x9b, 0x1e, 0x87, 0xe9, 0xce, 0x55, 0x28, 0xdf);
+       &data_byte(0x8c, 0xa1, 0x89, 0x0d, 0xbf, 0xe6, 0x42, 0x68);
+       &data_byte(0x41, 0x99, 0x2d, 0x0f, 0xb0, 0x54, 0xbb, 0x16);
+
+       &data_byte(0x63, 0x7c, 0x77, 0x7b, 0xf2, 0x6b, 0x6f, 0xc5);
+       &data_byte(0x30, 0x01, 0x67, 0x2b, 0xfe, 0xd7, 0xab, 0x76);
+       &data_byte(0xca, 0x82, 0xc9, 0x7d, 0xfa, 0x59, 0x47, 0xf0);
+       &data_byte(0xad, 0xd4, 0xa2, 0xaf, 0x9c, 0xa4, 0x72, 0xc0);
+       &data_byte(0xb7, 0xfd, 0x93, 0x26, 0x36, 0x3f, 0xf7, 0xcc);
+       &data_byte(0x34, 0xa5, 0xe5, 0xf1, 0x71, 0xd8, 0x31, 0x15);
+       &data_byte(0x04, 0xc7, 0x23, 0xc3, 0x18, 0x96, 0x05, 0x9a);
+       &data_byte(0x07, 0x12, 0x80, 0xe2, 0xeb, 0x27, 0xb2, 0x75);
+       &data_byte(0x09, 0x83, 0x2c, 0x1a, 0x1b, 0x6e, 0x5a, 0xa0);
+       &data_byte(0x52, 0x3b, 0xd6, 0xb3, 0x29, 0xe3, 0x2f, 0x84);
+       &data_byte(0x53, 0xd1, 0x00, 0xed, 0x20, 0xfc, 0xb1, 0x5b);
+       &data_byte(0x6a, 0xcb, 0xbe, 0x39, 0x4a, 0x4c, 0x58, 0xcf);
+       &data_byte(0xd0, 0xef, 0xaa, 0xfb, 0x43, 0x4d, 0x33, 0x85);
+       &data_byte(0x45, 0xf9, 0x02, 0x7f, 0x50, 0x3c, 0x9f, 0xa8);
+       &data_byte(0x51, 0xa3, 0x40, 0x8f, 0x92, 0x9d, 0x38, 0xf5);
+       &data_byte(0xbc, 0xb6, 0xda, 0x21, 0x10, 0xff, 0xf3, 0xd2);
+       &data_byte(0xcd, 0x0c, 0x13, 0xec, 0x5f, 0x97, 0x44, 0x17);
+       &data_byte(0xc4, 0xa7, 0x7e, 0x3d, 0x64, 0x5d, 0x19, 0x73);
+       &data_byte(0x60, 0x81, 0x4f, 0xdc, 0x22, 0x2a, 0x90, 0x88);
+       &data_byte(0x46, 0xee, 0xb8, 0x14, 0xde, 0x5e, 0x0b, 0xdb);
+       &data_byte(0xe0, 0x32, 0x3a, 0x0a, 0x49, 0x06, 0x24, 0x5c);
+       &data_byte(0xc2, 0xd3, 0xac, 0x62, 0x91, 0x95, 0xe4, 0x79);
+       &data_byte(0xe7, 0xc8, 0x37, 0x6d, 0x8d, 0xd5, 0x4e, 0xa9);
+       &data_byte(0x6c, 0x56, 0xf4, 0xea, 0x65, 0x7a, 0xae, 0x08);
+       &data_byte(0xba, 0x78, 0x25, 0x2e, 0x1c, 0xa6, 0xb4, 0xc6);
+       &data_byte(0xe8, 0xdd, 0x74, 0x1f, 0x4b, 0xbd, 0x8b, 0x8a);
+       &data_byte(0x70, 0x3e, 0xb5, 0x66, 0x48, 0x03, 0xf6, 0x0e);
+       &data_byte(0x61, 0x35, 0x57, 0xb9, 0x86, 0xc1, 0x1d, 0x9e);
+       &data_byte(0xe1, 0xf8, 0x98, 0x11, 0x69, 0xd9, 0x8e, 0x94);
+       &data_byte(0x9b, 0x1e, 0x87, 0xe9, 0xce, 0x55, 0x28, 0xdf);
+       &data_byte(0x8c, 0xa1, 0x89, 0x0d, 0xbf, 0xe6, 0x42, 0x68);
+       &data_byte(0x41, 0x99, 0x2d, 0x0f, 0xb0, 0x54, 0xbb, 0x16);
+
        &data_byte(0x63, 0x7c, 0x77, 0x7b, 0xf2, 0x6b, 0x6f, 0xc5);
        &data_byte(0x30, 0x01, 0x67, 0x2b, 0xfe, 0xd7, 0xab, 0x76);
        &data_byte(0xca, 0x82, 0xc9, 0x7d, 0xfa, 0x59, 0x47, 0xf0);
@@ -588,100 +1069,82 @@ sub _data_word() { my $i; while(defined($i=shift)) { &data_word($i,$i); } }
 
        &mov    ($s0,"esp");
        &sub    ("esp",36);
-       &and    ("esp",-64);
-       &add    ("esp",4);
-       &mov    (&DWP(28,"esp"),$s0);
+       &and    ("esp",-64);                    # align to cache-line
+
+       # place stack frame just "above" the key schedule
+       &lea    ($s1,&DWP(-64-63,$key));
+       &sub    ($s1,"esp");
+       &neg    ($s1);
+       &and    ($s1,0x3C0);    # modulo 1024, but aligned to cache-line
+       &sub    ("esp",$s1);
+       &add    ("esp",4);      # 4 is reserved for caller's return address
+       &mov    (&DWP(28,"esp"),$s0);           # save stack pointer
 
        &call   (&label("pic_point"));          # make it PIC!
        &set_label("pic_point");
-       &blindpop("ebp");
-       &lea    ("ebp",&DWP(&label("AES_Te")."-".&label("pic_point"),"ebp"));
-
+       &blindpop($tbl);
+       &picmeup($s0,"OPENSSL_ia32cap_P",$tbl,&label("pic_point"));
+       &lea    ($tbl,&DWP(&label("AES_Te")."-".&label("pic_point"),$tbl));
+       # pick Te4 copy which can't "overlap" with stack frame or key schedule
+       &lea    ($s1,&DWP(768,"esp"));
+       &and    ($s1,0x300);
+       &lea    ($tbl,&DWP(2048+128,$tbl,$s1));
+
+       &bt     (&DWP(0,$s0),23);               # check for MMX bit
+       &jc     (&label("mmx"));
+
+       &mov    (&DWP(24,"esp"),$tbl);
        &mov    ($s0,&DWP(0,$acc));             # load input data
        &mov    ($s1,&DWP(4,$acc));
        &mov    ($s2,&DWP(8,$acc));
        &mov    ($s3,&DWP(12,$acc));
-
-       &mov    (&DWP(24,"esp"),"ebp");
-
-       &call   ("_x86_AES_encrypt");
-
-       &mov    ("esp",&DWP(28,"esp"));
-
+       &call   ("_x86_AES_encrypt_compact");
+       &mov    ("esp",&DWP(28,"esp"));         # restore stack pointer
        &mov    ($acc,&wparam(1));              # load out
        &mov    (&DWP(0,$acc),$s0);             # write output data
        &mov    (&DWP(4,$acc),$s1);
        &mov    (&DWP(8,$acc),$s2);
        &mov    (&DWP(12,$acc),$s3);
-&function_end("AES_encrypt");
-
-#------------------------------------------------------------------#
-
-sub decstep()
-{ my ($i,$td,@s) = @_;
-  my $tmp = $key;
-  my $out = $i==3?$s[0]:$acc;
+       &jmp    (&label("ret"));
 
-       # no instructions are reordered, as performance appears
-       # optimal... or rather that all attempts to reorder didn't
-       # result in better performance [which by the way is not a
-       # bit lower than ecryption].
-       if($i==3)   {   &mov    ($key,&DWP(20,"esp"));          }
-       else        {   &mov    ($out,$s[0]);                   }
-                       &and    ($out,0xFF);
-                       &mov    ($out,&DWP(0,$td,$out,8));
+       &set_label("mmx",16);
+       &movq   ("mm0",&QWP(0,$acc));
+       &movq   ("mm4",&QWP(8,$acc));
+       &call   ("_mmx_AES_encrypt_compact");
+       &mov    ("esp",&DWP(28,"esp"));         # restore stack pointer
+       &mov    ($acc,&wparam(1));              # load out
+       &movq   (&QWP(0,$acc),"mm0");           # write output data
+       &movq   (&QWP(8,$acc),"mm4");
+       &emms   ();
 
-       if ($i==3)  {   $tmp=$s[1];                             }
-                       &movz   ($tmp,&HB($s[1]));
-                       &xor    ($out,&DWP(3,$td,$tmp,8));
+&set_label("ret",4);
+&function_end("AES_encrypt");
 
-       if ($i==3)  {   $tmp=$s[2]; &mov ($s[1],$acc);          }
-       else        {   &mov    ($tmp,$s[2]);                   }
-                       &shr    ($tmp,16);
-                       &and    ($tmp,0xFF);
-                       &xor    ($out,&DWP(2,$td,$tmp,8));
+#--------------------------------------------------------------------#
 
-       if ($i==3)  {   $tmp=$s[3]; &mov ($s[2],&DWP(8,"esp")); }
-       else        {   &mov    ($tmp,$s[3]);                   }
-                       &shr    ($tmp,24);
-                       &xor    ($out,&DWP(1,$td,$tmp,8));
-       if ($i<2)   {   &mov    (&DWP(4+4*$i,"esp"),$out);      }
-       if ($i==3)  {   &mov    ($s[3],&DWP(4,"esp"));          }
-                       &comment();
-}
+######################################################################
+# "Compact" block function
+######################################################################
 
-sub declast()
+sub deccompact()
 { my $Fn = mov;
   while ($#_>5) { pop(@_); $Fn=sub{}; }
   my ($i,$td,@s)=@_;
   my $tmp = $key;
   my $out = $i==3?$s[0]:$acc;
 
-       if ($i==0)      # prefetch 256-byte Td4
-       {   &lea        ($td,&DWP(2048+128,$td));
-           &mov        ($tmp,&DWP(0-128,$td));
-           &mov        ($acc,&DWP(32-128,$td));
-           &mov        ($tmp,&DWP(64-128,$td));
-           &mov        ($acc,&DWP(96-128,$td));
-           &mov        ($tmp,&DWP(128-128,$td));
-           &mov        ($acc,&DWP(160-128,$td));
-           &mov        ($tmp,&DWP(192-128,$td));
-           &mov        ($acc,&DWP(224-128,$td));
-           &lea        ($td,&DWP(-128,$td));
-       }
-
        # $Fn is used in first compact round and its purpose is to
        # void restoration of some values from stack, so that after
-       # 4xenclast with extra argument $key, $s0 and $s1 values
+       # 4xdeccompact with extra argument $key, $s0 and $s1 values
        # are left there...
        if($i==3)   {   &$Fn    ($key,&DWP(20,"esp"));          }
        else        {   &mov    ($out,$s[0]);                   }
                        &and    ($out,0xFF);
-                       &movz   ($out,&DWP(0,$td,$out,1));
+                       &movz   ($out,&DWP(-128,$td,$out,1));
 
        if ($i==3)  {   $tmp=$s[1];                             }
                        &movz   ($tmp,&HB($s[1]));
-                       &movz   ($tmp,&DWP(0,$td,$tmp,1));
+                       &movz   ($tmp,&DWP(-128,$td,$tmp,1));
                        &shl    ($tmp,8);
                        &xor    ($out,$tmp);
 
@@ -689,19 +1152,18 @@ sub declast()
        else        {   mov     ($tmp,$s[2]);                   }
                        &shr    ($tmp,16);
                        &and    ($tmp,0xFF);
-                       &movz   ($tmp,&DWP(0,$td,$tmp,1));
+                       &movz   ($tmp,&DWP(-128,$td,$tmp,1));
                        &shl    ($tmp,16);
                        &xor    ($out,$tmp);
 
        if ($i==3)  {   $tmp=$s[3]; &$Fn ($s[2],&DWP(8,"esp")); }
        else        {   &mov    ($tmp,$s[3]);                   }
                        &shr    ($tmp,24);
-                       &movz   ($tmp,&DWP(0,$td,$tmp,1));
+                       &movz   ($tmp,&DWP(-128,$td,$tmp,1));
                        &shl    ($tmp,24);
                        &xor    ($out,$tmp);
        if ($i<2)   {   &mov    (&DWP(4+4*$i,"esp"),$out);      }
-       if ($i==3)  {   &$Fn    ($s[3],&DWP(4,"esp"));
-                       &lea    ($td,&DWP(-2048,$td));          }
+       if ($i==3)  {   &$Fn    ($s[3],&DWP(4,"esp"));          }
 }
 
 # must be called with 2,3,0,1 as argument sequence!!!
@@ -711,7 +1173,7 @@ sub dectransform()
   my $tmp = $key;
   my $tp2 = @s[($i+2)%4]; $tp2 = @s[2] if ($i==1);
   my $tp4 = @s[($i+3)%4]; $tp4 = @s[3] if ($i==1);
-  my $tp8 = "ebp";
+  my $tp8 = $tbl;
 
        &mov    ($acc,$s[$i]);
        &and    ($acc,0x80808080);
@@ -720,7 +1182,7 @@ sub dectransform()
        &shr    ($tmp,7);
        &and    ($tp2,0x7f7f7f7f);
        &sub    ($acc,$tmp);
-       &lea    ($tp2,&DWP(0,$tp2,$tp2));
+       &add    ($tp2,$tp2);
        &and    ($acc,0x1b1b1b1b);
        &xor    ($acc,$tp2);
        &mov    ($tp2,$acc);
@@ -728,10 +1190,11 @@ sub dectransform()
        &and    ($acc,0x80808080);
        &mov    ($tmp,$acc);
        &mov    ($tp4,$tp2);
+        &xor   ($tp2,$s[$i]);  # tp2^tp1
        &shr    ($tmp,7);
        &and    ($tp4,0x7f7f7f7f);
        &sub    ($acc,$tmp);
-       &lea    ($tp4,&DWP(0,$tp4,$tp4));
+       &add    ($tp4,$tp4);
        &and    ($acc,0x1b1b1b1b);
        &xor    ($acc,$tp4);
        &mov    ($tp4,$acc);
@@ -739,33 +1202,36 @@ sub dectransform()
        &and    ($acc,0x80808080);
        &mov    ($tmp,$acc);
        &mov    ($tp8,$tp4);
+        &xor   ($tp4,$s[$i]);  # tp4^tp1
        &shr    ($tmp,7);
        &and    ($tp8,0x7f7f7f7f);
        &sub    ($acc,$tmp);
-       &lea    ($tp8,&DWP(0,$tp8,$tp8));
+       &add    ($tp8,$tp8);
        &and    ($acc,0x1b1b1b1b);
+        &rotl  ($s[$i],8);     # = ROTATE(tp1,8)
        &xor    ($tp8,$acc);
 
-       &exch   ($s[$i],$tp8);
-       &xor    ($tp8,$s[$i]);  # tp8 ^ tp1
-       &xor    ($s[$i],$tp4);
-       &xor    ($s[$i],$tp2);  # tp8 ^ tp4 ^ tp2
-       &xor    ($tp4,$tp8);
+       &xor    ($s[$i],$tp2);
        &xor    ($tp2,$tp8);
-       &rotl   ($tp4,16);
        &xor    ($s[$i],$tp4);
        &rotl   ($tp2,24);
-       &xor    ($s[$i],$tp2);
+       &xor    ($tp4,$tp8);
+       &xor    ($s[$i],$tp8);  # ^= tp8^(tp4^tp1)^(tp2^tp1)
+       &rotl   ($tp4,16);
+       &xor    ($s[$i],$tp2);  # ^= ROTATE(tp8^tp2^tp1,24)
        &rotl   ($tp8,8);
-       &xor    ($s[$i],$tp8);
+       &xor    ($s[$i],$tp4);  # ^= ROTATE(tp8^tp4^tp1,16)
+       &xor    ($s[$i],$tp8);  # ^= ROTATE(tp8,8)
 
        &mov    ($s[0],&DWP(4,"esp"))           if($i==2); #prefetch $s0
        &mov    ($s[1],&DWP(8,"esp"))           if($i==3); #prefetch $s1
+       &mov    ($s[2],&DWP(12,"esp"))          if($i==1);
+       &mov    ($s[3],&DWP(16,"esp"))          if($i==1);
        &mov    (&DWP(4+4*$i,"esp"),$s[$i])     if($i>=2);
 }
 
 &public_label("AES_Td");
-&function_begin_B("_x86_AES_decrypt");
+&function_begin_B("_x86_AES_decrypt_compact");
        # note that caller is expected to allocate stack frame for me!
        &mov    (&DWP(20,"esp"),$key);          # save key
 
@@ -774,38 +1240,350 @@ sub dectransform()
        &xor    ($s2,&DWP(8,$key));
        &xor    ($s3,&DWP(12,$key));
 
-       # not really last round, just "compact" one...
-       &declast(0,"ebp",$s0,$s3,$s2,$s1,1);
-       &declast(1,"ebp",$s1,$s0,$s3,$s2,1);
-       &declast(2,"ebp",$s2,$s1,$s0,$s3,1);
-       &declast(3,"ebp",$s3,$s2,$s1,$s0,1);
-       &dectransform(2);
-       &dectransform(3);
-       &dectransform(0);
-       &dectransform(1);
-       &mov    ($key,&DWP(20,"esp"));
-       &mov    ($s2,&DWP(12,"esp"));
-       &mov    ($s3,&DWP(16,"esp"));
-       &mov    ("ebp",&DWP(28,"esp"));
+       &mov    ($acc,&DWP(240,$key));          # load key->rounds
+
+       &lea    ($acc,&DWP(-2,$acc,$acc));
+       &lea    ($acc,&DWP(0,$key,$acc,8));
+       &mov    (&DWP(24,"esp"),$acc);          # end of key schedule
+
+       # prefetch Td4
+       &mov    ($key,&DWP(0-128,$tbl));
+       &mov    ($acc,&DWP(32-128,$tbl));
+       &mov    ($key,&DWP(64-128,$tbl));
+       &mov    ($acc,&DWP(96-128,$tbl));
+       &mov    ($key,&DWP(128-128,$tbl));
+       &mov    ($acc,&DWP(160-128,$tbl));
+       &mov    ($key,&DWP(192-128,$tbl));
+       &mov    ($acc,&DWP(224-128,$tbl));
+
+       &set_label("loop",16);
+
+               &deccompact(0,$tbl,$s0,$s3,$s2,$s1,1);
+               &deccompact(1,$tbl,$s1,$s0,$s3,$s2,1);
+               &deccompact(2,$tbl,$s2,$s1,$s0,$s3,1);
+               &deccompact(3,$tbl,$s3,$s2,$s1,$s0,1);
+               &dectransform(2);
+               &dectransform(3);
+               &dectransform(0);
+               &dectransform(1);
+               &mov    ($key,&DWP(20,"esp"));
+               &mov    ($tbl,&DWP(28,"esp"));
+               &add    ($key,16);              # advance rd_key
+               &xor    ($s0,&DWP(0,$key));
+               &xor    ($s1,&DWP(4,$key));
+               &xor    ($s2,&DWP(8,$key));
+               &xor    ($s3,&DWP(12,$key));
+
+       &cmp    ($key,&DWP(24,"esp"));
+       &mov    (&DWP(20,"esp"),$key);
+       &jb     (&label("loop"));
+
+       &deccompact(0,$tbl,$s0,$s3,$s2,$s1);
+       &deccompact(1,$tbl,$s1,$s0,$s3,$s2);
+       &deccompact(2,$tbl,$s2,$s1,$s0,$s3);
+       &deccompact(3,$tbl,$s3,$s2,$s1,$s0);
+
        &xor    ($s0,&DWP(16,$key));
        &xor    ($s1,&DWP(20,$key));
        &xor    ($s2,&DWP(24,$key));
        &xor    ($s3,&DWP(28,$key));
 
+       &ret    ();
+&function_end_B("_x86_AES_decrypt_compact");
+
+######################################################################
+# "Compact" MMX block function.
+######################################################################
+
+sub mmx_deccompact()
+{
+       &pshufw ("mm1","mm0",0x0c);             #  7, 6, 1, 0
+       &movd   ("eax","mm1");                  #  7, 6, 1, 0
+
+       &pshufw ("mm5","mm4",0x09);             # 13,12,11,10
+       &movz   ("esi",&LB("eax"));             #  0
+       &movz   ("ecx",&DWP(-128,$tbl,"esi",1));#  0
+       &movd   ("ebx","mm5");                  # 13,12,11,10
+       &movz   ("edx",&HB("eax"));             #  1
+       &movz   ("edx",&DWP(-128,$tbl,"edx",1));#  1
+       &shl    ("edx",8);                      #  1
+
+       &pshufw ("mm2","mm0",0x06);             #  3, 2, 5, 4
+       &movz   ("esi",&LB("ebx"));             # 10
+       &movz   ("esi",&DWP(-128,$tbl,"esi",1));# 10
+       &shl    ("esi",16);                     # 10
+       &or     ("ecx","esi");                  # 10
+       &shr    ("eax",16);                     #  7, 6
+       &movz   ("esi",&HB("ebx"));             # 11
+       &movz   ("esi",&DWP(-128,$tbl,"esi",1));# 11
+       &shl    ("esi",24);                     # 11
+       &or     ("edx","esi");                  # 11
+       &shr    ("ebx",16);                     # 13,12
+
+       &pshufw ("mm6","mm4",0x03);             # 9, 8,15,14
+       &movz   ("esi",&HB("eax"));             #  7
+       &movz   ("esi",&DWP(-128,$tbl,"esi",1));#  7
+       &shl    ("esi",24);                     #  7
+       &or     ("ecx","esi");                  #  7
+       &movz   ("esi",&HB("ebx"));             # 13
+       &movz   ("esi",&DWP(-128,$tbl,"esi",1));# 13
+       &shl    ("esi",8);                      # 13
+       &or     ("ecx","esi");                  # 13
+       &movd   ("mm0","ecx");                  # t[0] collected
+
+       &movz   ("esi",&LB("eax"));             #  6
+       &movd   ("eax","mm2");                  #  3, 2, 5, 4
+       &movz   ("ecx",&DWP(-128,$tbl,"esi",1));#  6
+       &shl    ("ecx",16);                     #  6
+       &movz   ("esi",&LB("ebx"));             # 12
+       &movd   ("ebx","mm6");                  #  9, 8,15,14
+       &movz   ("esi",&DWP(-128,$tbl,"esi",1));# 12
+       &or     ("ecx","esi");                  # 12
+
+       &movz   ("esi",&LB("eax"));             #  4
+       &movz   ("esi",&DWP(-128,$tbl,"esi",1));#  4
+       &or     ("edx","esi");                  #  4
+       &movz   ("esi",&LB("ebx"));             # 14
+       &movz   ("esi",&DWP(-128,$tbl,"esi",1));# 14
+       &shl    ("esi",16);                     # 14
+       &or     ("edx","esi");                  # 14
+       &movd   ("mm1","edx");                  # t[1] collected
+
+       &movz   ("esi",&HB("eax"));             #  5
+       &movz   ("edx",&DWP(-128,$tbl,"esi",1));#  5
+       &shl    ("edx",8);                      #  5
+       &movz   ("esi",&HB("ebx"));             # 15
+       &shr    ("eax",16);                     #  3, 2
+       &movz   ("esi",&DWP(-128,$tbl,"esi",1));# 15
+       &shl    ("esi",24);                     # 15
+       &or     ("edx","esi");                  # 15
+       &shr    ("ebx",16);                     #  9, 8
+
+       &punpckldq      ("mm0","mm1");          # t[0,1] collected
+
+       &movz   ("esi",&HB("ebx"));             #  9
+       &movz   ("esi",&DWP(-128,$tbl,"esi",1));#  9
+       &shl    ("esi",8);                      #  9
+       &or     ("ecx","esi");                  #  9
+       &and    ("ebx",0xff);                   #  8
+       &movz   ("ebx",&DWP(-128,$tbl,"ebx",1));#  8
+       &or     ("edx","ebx");                  #  8
+       &movz   ("esi",&LB("eax"));             #  2
+       &movz   ("esi",&DWP(-128,$tbl,"esi",1));#  2
+       &shl    ("esi",16);                     #  2
+       &or     ("edx","esi");                  #  2
+       &movd   ("mm4","edx");                  # t[2] collected
+       &movz   ("eax",&HB("eax"));             #  3
+       &movz   ("eax",&DWP(-128,$tbl,"eax",1));#  3
+       &shl    ("eax",24);                     #  3
+       &or     ("ecx","eax");                  #  3
+       &movd   ("mm5","ecx");                  # t[3] collected
+
+       &punpckldq      ("mm4","mm5");          # t[2,3] collected
+}
+
+&public_label("AES_Td");
+&function_begin_B("_mmx_AES_decrypt_compact");
+       &pxor   ("mm0",&QWP(0,$key));   #  7, 6, 5, 4, 3, 2, 1, 0
+       &pxor   ("mm4",&QWP(8,$key));   # 15,14,13,12,11,10, 9, 8
+
+       # note that caller is expected to allocate stack frame for me!
+       &mov    ($acc,&DWP(240,$key));          # load key->rounds
+       &lea    ($acc,&DWP(-2,$acc,$acc));
+       &lea    ($acc,&DWP(0,$key,$acc,8));
+       &mov    (&DWP(24,"esp"),$acc);          # end of key schedule
+
+       &mov    ($s0,0x1b1b1b1b);               # magic constant
+       &mov    (&DWP(8,"esp"),$s0);
+       &mov    (&DWP(12,"esp"),$s0);
+
+       # prefetch Td4
+       &mov    ($s0,&DWP(0-128,$tbl));
+       &mov    ($s1,&DWP(32-128,$tbl));
+       &mov    ($s2,&DWP(64-128,$tbl));
+       &mov    ($s3,&DWP(96-128,$tbl));
+       &mov    ($s0,&DWP(128-128,$tbl));
+       &mov    ($s1,&DWP(160-128,$tbl));
+       &mov    ($s2,&DWP(192-128,$tbl));
+       &mov    ($s3,&DWP(224-128,$tbl));
+
+       &align  (4);
+       &set_label("loop",16);
+               &mmx_deccompact();
+               &add    ($key,16);
+               &cmp    ($key,&DWP(24,"esp"));
+               &ja     (&label("out"));
+
+               # ROTATE(x^y,N) == ROTATE(x,N)^ROTATE(y,N)
+               &movq   ("mm3","mm0");          &movq   ("mm7","mm4");
+               &movq   ("mm2","mm0",1);        &movq   ("mm6","mm4",1);
+               &movq   ("mm1","mm0");          &movq   ("mm5","mm4");
+               &pshufw ("mm0","mm0",0xb1);     &pshufw ("mm4","mm4",0xb1);# = ROTATE(tp0,16)
+               &pslld  ("mm2",8);              &pslld  ("mm6",8);
+               &psrld  ("mm3",8);              &psrld  ("mm7",8);
+               &pxor   ("mm0","mm2");          &pxor   ("mm4","mm6");  # ^= tp0<<8
+               &pxor   ("mm0","mm3");          &pxor   ("mm4","mm7");  # ^= tp0>>8
+               &pslld  ("mm2",16);             &pslld  ("mm6",16);
+               &psrld  ("mm3",16);             &psrld  ("mm7",16);
+               &pxor   ("mm0","mm2");          &pxor   ("mm4","mm6");  # ^= tp0<<24
+               &pxor   ("mm0","mm3");          &pxor   ("mm4","mm7");  # ^= tp0>>24
+
+               &movq   ("mm3",&QWP(8,"esp"));
+               &pxor   ("mm2","mm2");          &pxor   ("mm6","mm6");
+               &pcmpgtb("mm2","mm1");          &pcmpgtb("mm6","mm5");
+               &pand   ("mm2","mm3");          &pand   ("mm6","mm3");
+               &paddb  ("mm1","mm1");          &paddb  ("mm5","mm5");
+               &pxor   ("mm1","mm2");          &pxor   ("mm5","mm6");  # tp2
+               &movq   ("mm3","mm1");          &movq   ("mm7","mm5");
+               &movq   ("mm2","mm1");          &movq   ("mm6","mm5");
+               &pxor   ("mm0","mm1");          &pxor   ("mm4","mm5");  # ^= tp2
+               &pslld  ("mm3",24);             &pslld  ("mm7",24);
+               &psrld  ("mm2",8);              &psrld  ("mm6",8);
+               &pxor   ("mm0","mm3");          &pxor   ("mm4","mm7");  # ^= tp2<<24
+               &pxor   ("mm0","mm2");          &pxor   ("mm4","mm6");  # ^= tp2>>8
+
+               &movq   ("mm2",&QWP(8,"esp"));
+               &pxor   ("mm3","mm3");          &pxor   ("mm7","mm7");
+               &pcmpgtb("mm3","mm1");          &pcmpgtb("mm7","mm5");
+               &pand   ("mm3","mm2");          &pand   ("mm7","mm2");
+               &paddb  ("mm1","mm1");          &paddb  ("mm5","mm5");
+               &pxor   ("mm1","mm3");          &pxor   ("mm5","mm7");  # tp4
+               &pshufw ("mm3","mm1",0xb1);     &pshufw ("mm7","mm5",0xb1);
+               &pxor   ("mm0","mm1");          &pxor   ("mm4","mm5");  # ^= tp4
+               &pxor   ("mm0","mm3");          &pxor   ("mm4","mm7");  # ^= ROTATE(tp4,16)     
+
+               &pxor   ("mm3","mm3");          &pxor   ("mm7","mm7");
+               &pcmpgtb("mm3","mm1");          &pcmpgtb("mm7","mm5");
+               &pand   ("mm3","mm2");          &pand   ("mm7","mm2");
+               &paddb  ("mm1","mm1");          &paddb  ("mm5","mm5");
+               &pxor   ("mm1","mm3");          &pxor   ("mm5","mm7");  # tp8
+               &pxor   ("mm0","mm1");          &pxor   ("mm4","mm5");  # ^= tp8
+               &movq   ("mm3","mm1");          &movq   ("mm7","mm5");
+               &pshufw ("mm2","mm1",0xb1);     &pshufw ("mm6","mm5",0xb1);
+               &pxor   ("mm0","mm2");          &pxor   ("mm4","mm6");  # ^= ROTATE(tp8,16)
+               &pslld  ("mm1",8);              &pslld  ("mm5",8);
+               &psrld  ("mm3",8);              &psrld  ("mm7",8);
+               &movq   ("mm2",&QWP(0,$key));   &movq   ("mm6",&DWP(8,$key));
+               &pxor   ("mm0","mm1");          &pxor   ("mm4","mm5");  # ^= tp8<<8
+               &pxor   ("mm0","mm3");          &pxor   ("mm4","mm7");  # ^= tp8>>8
+               &pslld  ("mm1",16);             &pslld  ("mm5",16);
+               &psrld  ("mm3",16);             &psrld  ("mm7",16);
+               &pxor   ("mm0","mm1");          &pxor   ("mm4","mm5");  # ^= tp8<<24
+               &pxor   ("mm0","mm3");          &pxor   ("mm4","mm7");  # ^= tp8>>24
+
+               &pxor   ("mm0","mm2");          &pxor   ("mm4","mm6");
+       &jmp    (&label("loop"));
+
+       &set_label("out",16);
+       &pxor   ("mm0",&QWP(0,$key));
+       &pxor   ("mm4",&QWP(8,$key));
+
+       &ret    ();
+&function_end_B("_mmx_AES_decrypt_compact");
+
+######################################################################
+# Vanilla block function.
+######################################################################
+
+sub decstep()
+{ my ($i,$td,@s) = @_;
+  my $tmp = $key;
+  my $out = $i==3?$s[0]:$acc;
+
+       # no instructions are reordered, as performance appears
+       # optimal... or rather that all attempts to reorder didn't
+       # result in better performance [which by the way is not a
+       # bit lower than ecryption].
+       if($i==3)   {   &mov    ($key,&DWP(20,"esp"));          }
+       else        {   &mov    ($out,$s[0]);                   }
+                       &and    ($out,0xFF);
+                       &mov    ($out,&DWP(0,$td,$out,8));
+
+       if ($i==3)  {   $tmp=$s[1];                             }
+                       &movz   ($tmp,&HB($s[1]));
+                       &xor    ($out,&DWP(3,$td,$tmp,8));
+
+       if ($i==3)  {   $tmp=$s[2]; &mov ($s[1],$acc);          }
+       else        {   &mov    ($tmp,$s[2]);                   }
+                       &shr    ($tmp,16);
+                       &and    ($tmp,0xFF);
+                       &xor    ($out,&DWP(2,$td,$tmp,8));
+
+       if ($i==3)  {   $tmp=$s[3]; &mov ($s[2],&DWP(8,"esp")); }
+       else        {   &mov    ($tmp,$s[3]);                   }
+                       &shr    ($tmp,24);
+                       &xor    ($out,&DWP(1,$td,$tmp,8));
+       if ($i<2)   {   &mov    (&DWP(4+4*$i,"esp"),$out);      }
+       if ($i==3)  {   &mov    ($s[3],&DWP(4,"esp"));          }
+                       &comment();
+}
+
+sub declast()
+{ my ($i,$td,@s)=@_;
+  my $tmp = $key;
+  my $out = $i==3?$s[0]:$acc;
+
+       if($i==0)   {   &lea    ($td,&DWP(2048+128,$td));
+                       &mov    ($tmp,&DWP(0-128,$td));
+                       &mov    ($acc,&DWP(32-128,$td));
+                       &mov    ($tmp,&DWP(64-128,$td));
+                       &mov    ($acc,&DWP(96-128,$td));
+                       &mov    ($tmp,&DWP(128-128,$td));
+                       &mov    ($acc,&DWP(160-128,$td));
+                       &mov    ($tmp,&DWP(192-128,$td));
+                       &mov    ($acc,&DWP(224-128,$td));
+                       &lea    ($td,&DWP(-128,$td));           }
+       if($i==3)   {   &mov    ($key,&DWP(20,"esp"));          }
+       else        {   &mov    ($out,$s[0]);                   }
+                       &and    ($out,0xFF);
+                       &movz   ($out,&DWP(0,$td,$out,1));
+
+       if ($i==3)  {   $tmp=$s[1];                             }
+                       &movz   ($tmp,&HB($s[1]));
+                       &movz   ($tmp,&DWP(0,$td,$tmp,1));
+                       &shl    ($tmp,8);
+                       &xor    ($out,$tmp);
+
+       if ($i==3)  {   $tmp=$s[2]; &mov ($s[1],$acc);          }
+       else        {   mov     ($tmp,$s[2]);                   }
+                       &shr    ($tmp,16);
+                       &and    ($tmp,0xFF);
+                       &movz   ($tmp,&DWP(0,$td,$tmp,1));
+                       &shl    ($tmp,16);
+                       &xor    ($out,$tmp);
+
+       if ($i==3)  {   $tmp=$s[3]; &mov ($s[2],&DWP(8,"esp")); }
+       else        {   &mov    ($tmp,$s[3]);                   }
+                       &shr    ($tmp,24);
+                       &movz   ($tmp,&DWP(0,$td,$tmp,1));
+                       &shl    ($tmp,24);
+                       &xor    ($out,$tmp);
+       if ($i<2)   {   &mov    (&DWP(4+4*$i,"esp"),$out);      }
+       if ($i==3)  {   &mov    ($s[3],&DWP(4,"esp"));
+                       &lea    ($td,&DWP(-2048,$td));          }
+}
+
+&public_label("AES_Td");
+&function_begin_B("_x86_AES_decrypt");
+       # note that caller is expected to allocate stack frame for me!
+       &mov    (&DWP(20,"esp"),$key);          # save key
+
+       &xor    ($s0,&DWP(0,$key));             # xor with key
+       &xor    ($s1,&DWP(4,$key));
+       &xor    ($s2,&DWP(8,$key));
+       &xor    ($s3,&DWP(12,$key));
+
        &mov    ($acc,&DWP(240,$key));          # load key->rounds
 
        if ($small_footprint) {
            &lea        ($acc,&DWP(-2,$acc,$acc));
            &lea        ($acc,&DWP(0,$key,$acc,8));
-           &add        ($key,16);
-           &mov        (&DWP(20,"esp"),$key);
            &mov        (&DWP(24,"esp"),$acc);  # end of key schedule
-           &align      (4);
-           &set_label("loop");
-               &decstep(0,"ebp",$s0,$s3,$s2,$s1);
-               &decstep(1,"ebp",$s1,$s0,$s3,$s2);
-               &decstep(2,"ebp",$s2,$s1,$s0,$s3);
-               &decstep(3,"ebp",$s3,$s2,$s1,$s0);
+           &set_label("loop",16);
+               &decstep(0,$tbl,$s0,$s3,$s2,$s1);
+               &decstep(1,$tbl,$s1,$s0,$s3,$s2);
+               &decstep(2,$tbl,$s2,$s1,$s0,$s3);
+               &decstep(3,$tbl,$s3,$s2,$s1,$s0);
                &add    ($key,16);              # advance rd_key
                &xor    ($s0,&DWP(0,$key));
                &xor    ($s1,&DWP(4,$key));
@@ -821,12 +1599,12 @@ sub dectransform()
            &cmp        ($acc,12);
            &jle        (&label("12rounds"));
 
-       &set_label("14rounds");
-           for ($i=2;$i<4;$i++) {
-               &decstep(0,"ebp",$s0,$s3,$s2,$s1);
-               &decstep(1,"ebp",$s1,$s0,$s3,$s2);
-               &decstep(2,"ebp",$s2,$s1,$s0,$s3);
-               &decstep(3,"ebp",$s3,$s2,$s1,$s0);
+       &set_label("14rounds",4);
+           for ($i=1;$i<3;$i++) {
+               &decstep(0,$tbl,$s0,$s3,$s2,$s1);
+               &decstep(1,$tbl,$s1,$s0,$s3,$s2);
+               &decstep(2,$tbl,$s2,$s1,$s0,$s3);
+               &decstep(3,$tbl,$s3,$s2,$s1,$s0);
                &xor    ($s0,&DWP(16*$i+0,$key));
                &xor    ($s1,&DWP(16*$i+4,$key));
                &xor    ($s2,&DWP(16*$i+8,$key));
@@ -834,12 +1612,12 @@ sub dectransform()
            }
            &add        ($key,32);
            &mov        (&DWP(20,"esp"),$key);  # advance rd_key
-       &set_label("12rounds");
-           for ($i=2;$i<4;$i++) {
-               &decstep(0,"ebp",$s0,$s3,$s2,$s1);
-               &decstep(1,"ebp",$s1,$s0,$s3,$s2);
-               &decstep(2,"ebp",$s2,$s1,$s0,$s3);
-               &decstep(3,"ebp",$s3,$s2,$s1,$s0);
+       &set_label("12rounds",4);
+           for ($i=1;$i<3;$i++) {
+               &decstep(0,$tbl,$s0,$s3,$s2,$s1);
+               &decstep(1,$tbl,$s1,$s0,$s3,$s2);
+               &decstep(2,$tbl,$s2,$s1,$s0,$s3);
+               &decstep(3,$tbl,$s3,$s2,$s1,$s0);
                &xor    ($s0,&DWP(16*$i+0,$key));
                &xor    ($s1,&DWP(16*$i+4,$key));
                &xor    ($s2,&DWP(16*$i+8,$key));
@@ -847,12 +1625,12 @@ sub dectransform()
            }
            &add        ($key,32);
            &mov        (&DWP(20,"esp"),$key);  # advance rd_key
-       &set_label("10rounds");
-           for ($i=2;$i<10;$i++) {
-               &decstep(0,"ebp",$s0,$s3,$s2,$s1);
-               &decstep(1,"ebp",$s1,$s0,$s3,$s2);
-               &decstep(2,"ebp",$s2,$s1,$s0,$s3);
-               &decstep(3,"ebp",$s3,$s2,$s1,$s0);
+       &set_label("10rounds",4);
+           for ($i=1;$i<10;$i++) {
+               &decstep(0,$tbl,$s0,$s3,$s2,$s1);
+               &decstep(1,$tbl,$s1,$s0,$s3,$s2);
+               &decstep(2,$tbl,$s2,$s1,$s0,$s3);
+               &decstep(3,$tbl,$s3,$s2,$s1,$s0);
                &xor    ($s0,&DWP(16*$i+0,$key));
                &xor    ($s1,&DWP(16*$i+4,$key));
                &xor    ($s2,&DWP(16*$i+8,$key));
@@ -860,10 +1638,10 @@ sub dectransform()
            }
        }
 
-       &declast(0,"ebp",$s0,$s3,$s2,$s1);
-       &declast(1,"ebp",$s1,$s0,$s3,$s2);
-       &declast(2,"ebp",$s2,$s1,$s0,$s3);
-       &declast(3,"ebp",$s3,$s2,$s1,$s0);
+       &declast(0,$tbl,$s0,$s3,$s2,$s1);
+       &declast(1,$tbl,$s1,$s0,$s3,$s2);
+       &declast(2,$tbl,$s2,$s1,$s0,$s3);
+       &declast(3,$tbl,$s3,$s2,$s1,$s0);
 
        &add    ($key,$small_footprint?16:160);
        &xor    ($s0,&DWP(0,$key));
@@ -873,7 +1651,7 @@ sub dectransform()
 
        &ret    ();
 
-&set_label("AES_Td",64);       # Yes! I keep it in the code segment!
+&set_label("AES_Td",1024);     # Yes! I keep it in the code segment!
        &_data_word(0x50a7f451, 0x5365417e, 0xc3a4171a, 0x965e273a);
        &_data_word(0xcb6bab3b, 0xf1459d1f, 0xab58faac, 0x9303e34b);
        &_data_word(0x55fa3020, 0xf66d76ad, 0x9176cc88, 0x254c02f5);
@@ -938,7 +1716,107 @@ sub dectransform()
        &_data_word(0x72c31d16, 0x0c25e2bc, 0x8b493c28, 0x41950dff);
        &_data_word(0x7101a839, 0xdeb30c08, 0x9ce4b4d8, 0x90c15664);
        &_data_word(0x6184cb7b, 0x70b632d5, 0x745c6c48, 0x4257b8d0);
-#Td4:
+
+#Td4:  # four copies of Td4 to choose from to avoid L1 aliasing
+       &data_byte(0x52, 0x09, 0x6a, 0xd5, 0x30, 0x36, 0xa5, 0x38);
+       &data_byte(0xbf, 0x40, 0xa3, 0x9e, 0x81, 0xf3, 0xd7, 0xfb);
+       &data_byte(0x7c, 0xe3, 0x39, 0x82, 0x9b, 0x2f, 0xff, 0x87);
+       &data_byte(0x34, 0x8e, 0x43, 0x44, 0xc4, 0xde, 0xe9, 0xcb);
+       &data_byte(0x54, 0x7b, 0x94, 0x32, 0xa6, 0xc2, 0x23, 0x3d);
+       &data_byte(0xee, 0x4c, 0x95, 0x0b, 0x42, 0xfa, 0xc3, 0x4e);
+       &data_byte(0x08, 0x2e, 0xa1, 0x66, 0x28, 0xd9, 0x24, 0xb2);
+       &data_byte(0x76, 0x5b, 0xa2, 0x49, 0x6d, 0x8b, 0xd1, 0x25);
+       &data_byte(0x72, 0xf8, 0xf6, 0x64, 0x86, 0x68, 0x98, 0x16);
+       &data_byte(0xd4, 0xa4, 0x5c, 0xcc, 0x5d, 0x65, 0xb6, 0x92);
+       &data_byte(0x6c, 0x70, 0x48, 0x50, 0xfd, 0xed, 0xb9, 0xda);
+       &data_byte(0x5e, 0x15, 0x46, 0x57, 0xa7, 0x8d, 0x9d, 0x84);
+       &data_byte(0x90, 0xd8, 0xab, 0x00, 0x8c, 0xbc, 0xd3, 0x0a);
+       &data_byte(0xf7, 0xe4, 0x58, 0x05, 0xb8, 0xb3, 0x45, 0x06);
+       &data_byte(0xd0, 0x2c, 0x1e, 0x8f, 0xca, 0x3f, 0x0f, 0x02);
+       &data_byte(0xc1, 0xaf, 0xbd, 0x03, 0x01, 0x13, 0x8a, 0x6b);
+       &data_byte(0x3a, 0x91, 0x11, 0x41, 0x4f, 0x67, 0xdc, 0xea);
+       &data_byte(0x97, 0xf2, 0xcf, 0xce, 0xf0, 0xb4, 0xe6, 0x73);
+       &data_byte(0x96, 0xac, 0x74, 0x22, 0xe7, 0xad, 0x35, 0x85);
+       &data_byte(0xe2, 0xf9, 0x37, 0xe8, 0x1c, 0x75, 0xdf, 0x6e);
+       &data_byte(0x47, 0xf1, 0x1a, 0x71, 0x1d, 0x29, 0xc5, 0x89);
+       &data_byte(0x6f, 0xb7, 0x62, 0x0e, 0xaa, 0x18, 0xbe, 0x1b);
+       &data_byte(0xfc, 0x56, 0x3e, 0x4b, 0xc6, 0xd2, 0x79, 0x20);
+       &data_byte(0x9a, 0xdb, 0xc0, 0xfe, 0x78, 0xcd, 0x5a, 0xf4);
+       &data_byte(0x1f, 0xdd, 0xa8, 0x33, 0x88, 0x07, 0xc7, 0x31);
+       &data_byte(0xb1, 0x12, 0x10, 0x59, 0x27, 0x80, 0xec, 0x5f);
+       &data_byte(0x60, 0x51, 0x7f, 0xa9, 0x19, 0xb5, 0x4a, 0x0d);
+       &data_byte(0x2d, 0xe5, 0x7a, 0x9f, 0x93, 0xc9, 0x9c, 0xef);
+       &data_byte(0xa0, 0xe0, 0x3b, 0x4d, 0xae, 0x2a, 0xf5, 0xb0);
+       &data_byte(0xc8, 0xeb, 0xbb, 0x3c, 0x83, 0x53, 0x99, 0x61);
+       &data_byte(0x17, 0x2b, 0x04, 0x7e, 0xba, 0x77, 0xd6, 0x26);
+       &data_byte(0xe1, 0x69, 0x14, 0x63, 0x55, 0x21, 0x0c, 0x7d);
+
+       &data_byte(0x52, 0x09, 0x6a, 0xd5, 0x30, 0x36, 0xa5, 0x38);
+       &data_byte(0xbf, 0x40, 0xa3, 0x9e, 0x81, 0xf3, 0xd7, 0xfb);
+       &data_byte(0x7c, 0xe3, 0x39, 0x82, 0x9b, 0x2f, 0xff, 0x87);
+       &data_byte(0x34, 0x8e, 0x43, 0x44, 0xc4, 0xde, 0xe9, 0xcb);
+       &data_byte(0x54, 0x7b, 0x94, 0x32, 0xa6, 0xc2, 0x23, 0x3d);
+       &data_byte(0xee, 0x4c, 0x95, 0x0b, 0x42, 0xfa, 0xc3, 0x4e);
+       &data_byte(0x08, 0x2e, 0xa1, 0x66, 0x28, 0xd9, 0x24, 0xb2);
+       &data_byte(0x76, 0x5b, 0xa2, 0x49, 0x6d, 0x8b, 0xd1, 0x25);
+       &data_byte(0x72, 0xf8, 0xf6, 0x64, 0x86, 0x68, 0x98, 0x16);
+       &data_byte(0xd4, 0xa4, 0x5c, 0xcc, 0x5d, 0x65, 0xb6, 0x92);
+       &data_byte(0x6c, 0x70, 0x48, 0x50, 0xfd, 0xed, 0xb9, 0xda);
+       &data_byte(0x5e, 0x15, 0x46, 0x57, 0xa7, 0x8d, 0x9d, 0x84);
+       &data_byte(0x90, 0xd8, 0xab, 0x00, 0x8c, 0xbc, 0xd3, 0x0a);
+       &data_byte(0xf7, 0xe4, 0x58, 0x05, 0xb8, 0xb3, 0x45, 0x06);
+       &data_byte(0xd0, 0x2c, 0x1e, 0x8f, 0xca, 0x3f, 0x0f, 0x02);
+       &data_byte(0xc1, 0xaf, 0xbd, 0x03, 0x01, 0x13, 0x8a, 0x6b);
+       &data_byte(0x3a, 0x91, 0x11, 0x41, 0x4f, 0x67, 0xdc, 0xea);
+       &data_byte(0x97, 0xf2, 0xcf, 0xce, 0xf0, 0xb4, 0xe6, 0x73);
+       &data_byte(0x96, 0xac, 0x74, 0x22, 0xe7, 0xad, 0x35, 0x85);
+       &data_byte(0xe2, 0xf9, 0x37, 0xe8, 0x1c, 0x75, 0xdf, 0x6e);
+       &data_byte(0x47, 0xf1, 0x1a, 0x71, 0x1d, 0x29, 0xc5, 0x89);
+       &data_byte(0x6f, 0xb7, 0x62, 0x0e, 0xaa, 0x18, 0xbe, 0x1b);
+       &data_byte(0xfc, 0x56, 0x3e, 0x4b, 0xc6, 0xd2, 0x79, 0x20);
+       &data_byte(0x9a, 0xdb, 0xc0, 0xfe, 0x78, 0xcd, 0x5a, 0xf4);
+       &data_byte(0x1f, 0xdd, 0xa8, 0x33, 0x88, 0x07, 0xc7, 0x31);
+       &data_byte(0xb1, 0x12, 0x10, 0x59, 0x27, 0x80, 0xec, 0x5f);
+       &data_byte(0x60, 0x51, 0x7f, 0xa9, 0x19, 0xb5, 0x4a, 0x0d);
+       &data_byte(0x2d, 0xe5, 0x7a, 0x9f, 0x93, 0xc9, 0x9c, 0xef);
+       &data_byte(0xa0, 0xe0, 0x3b, 0x4d, 0xae, 0x2a, 0xf5, 0xb0);
+       &data_byte(0xc8, 0xeb, 0xbb, 0x3c, 0x83, 0x53, 0x99, 0x61);
+       &data_byte(0x17, 0x2b, 0x04, 0x7e, 0xba, 0x77, 0xd6, 0x26);
+       &data_byte(0xe1, 0x69, 0x14, 0x63, 0x55, 0x21, 0x0c, 0x7d);
+
+       &data_byte(0x52, 0x09, 0x6a, 0xd5, 0x30, 0x36, 0xa5, 0x38);
+       &data_byte(0xbf, 0x40, 0xa3, 0x9e, 0x81, 0xf3, 0xd7, 0xfb);
+       &data_byte(0x7c, 0xe3, 0x39, 0x82, 0x9b, 0x2f, 0xff, 0x87);
+       &data_byte(0x34, 0x8e, 0x43, 0x44, 0xc4, 0xde, 0xe9, 0xcb);
+       &data_byte(0x54, 0x7b, 0x94, 0x32, 0xa6, 0xc2, 0x23, 0x3d);
+       &data_byte(0xee, 0x4c, 0x95, 0x0b, 0x42, 0xfa, 0xc3, 0x4e);
+       &data_byte(0x08, 0x2e, 0xa1, 0x66, 0x28, 0xd9, 0x24, 0xb2);
+       &data_byte(0x76, 0x5b, 0xa2, 0x49, 0x6d, 0x8b, 0xd1, 0x25);
+       &data_byte(0x72, 0xf8, 0xf6, 0x64, 0x86, 0x68, 0x98, 0x16);
+       &data_byte(0xd4, 0xa4, 0x5c, 0xcc, 0x5d, 0x65, 0xb6, 0x92);
+       &data_byte(0x6c, 0x70, 0x48, 0x50, 0xfd, 0xed, 0xb9, 0xda);
+       &data_byte(0x5e, 0x15, 0x46, 0x57, 0xa7, 0x8d, 0x9d, 0x84);
+       &data_byte(0x90, 0xd8, 0xab, 0x00, 0x8c, 0xbc, 0xd3, 0x0a);
+       &data_byte(0xf7, 0xe4, 0x58, 0x05, 0xb8, 0xb3, 0x45, 0x06);
+       &data_byte(0xd0, 0x2c, 0x1e, 0x8f, 0xca, 0x3f, 0x0f, 0x02);
+       &data_byte(0xc1, 0xaf, 0xbd, 0x03, 0x01, 0x13, 0x8a, 0x6b);
+       &data_byte(0x3a, 0x91, 0x11, 0x41, 0x4f, 0x67, 0xdc, 0xea);
+       &data_byte(0x97, 0xf2, 0xcf, 0xce, 0xf0, 0xb4, 0xe6, 0x73);
+       &data_byte(0x96, 0xac, 0x74, 0x22, 0xe7, 0xad, 0x35, 0x85);
+       &data_byte(0xe2, 0xf9, 0x37, 0xe8, 0x1c, 0x75, 0xdf, 0x6e);
+       &data_byte(0x47, 0xf1, 0x1a, 0x71, 0x1d, 0x29, 0xc5, 0x89);
+       &data_byte(0x6f, 0xb7, 0x62, 0x0e, 0xaa, 0x18, 0xbe, 0x1b);
+       &data_byte(0xfc, 0x56, 0x3e, 0x4b, 0xc6, 0xd2, 0x79, 0x20);
+       &data_byte(0x9a, 0xdb, 0xc0, 0xfe, 0x78, 0xcd, 0x5a, 0xf4);
+       &data_byte(0x1f, 0xdd, 0xa8, 0x33, 0x88, 0x07, 0xc7, 0x31);
+       &data_byte(0xb1, 0x12, 0x10, 0x59, 0x27, 0x80, 0xec, 0x5f);
+       &data_byte(0x60, 0x51, 0x7f, 0xa9, 0x19, 0xb5, 0x4a, 0x0d);
+       &data_byte(0x2d, 0xe5, 0x7a, 0x9f, 0x93, 0xc9, 0x9c, 0xef);
+       &data_byte(0xa0, 0xe0, 0x3b, 0x4d, 0xae, 0x2a, 0xf5, 0xb0);
+       &data_byte(0xc8, 0xeb, 0xbb, 0x3c, 0x83, 0x53, 0x99, 0x61);
+       &data_byte(0x17, 0x2b, 0x04, 0x7e, 0xba, 0x77, 0xd6, 0x26);
+       &data_byte(0xe1, 0x69, 0x14, 0x63, 0x55, 0x21, 0x0c, 0x7d);
+
        &data_byte(0x52, 0x09, 0x6a, 0xd5, 0x30, 0x36, 0xa5, 0x38);
        &data_byte(0xbf, 0x40, 0xa3, 0x9e, 0x81, 0xf3, 0xd7, 0xfb);
        &data_byte(0x7c, 0xe3, 0x39, 0x82, 0x9b, 0x2f, 0xff, 0x87);
@@ -981,31 +1859,55 @@ sub dectransform()
 
        &mov    ($s0,"esp");
        &sub    ("esp",36);
-       &and    ("esp",-64);
-       &add    ("esp",4);
-       &mov    (&DWP(28,"esp"),$s0);
+       &and    ("esp",-64);                    # align to cache-line
+
+       # place stack frame just "above" the key schedule
+       &lea    ($s1,&DWP(-64-63,$key));
+       &sub    ($s1,"esp");
+       &neg    ($s1);
+       &and    ($s1,0x3C0);    # modulo 1024, but aligned to cache-line
+       &sub    ("esp",$s1);
+       &add    ("esp",4);      # 4 is reserved for caller's return address
+       &mov    (&DWP(28,"esp"),$s0);           # save stack pointer
 
        &call   (&label("pic_point"));          # make it PIC!
        &set_label("pic_point");
-       &blindpop("ebp");
-       &lea    ("ebp",&DWP(&label("AES_Td")."-".&label("pic_point"),"ebp"));
-
+       &blindpop($tbl);
+       &picmeup($s0,"OPENSSL_ia32cap_P",$tbl,&label("pic_point"));
+       &lea    ($tbl,&DWP(&label("AES_Td")."-".&label("pic_point"),$tbl));
+       # pick Td4 copy which can't "overlap" with stack frame or key schedule
+       &lea    ($s1,&DWP(768,"esp"));
+       &and    ($s1,0x300);
+       &lea    ($tbl,&DWP(2048+128,$tbl,$s1));
+
+       &bt     (&DWP(0,$s0),23);               # check for MMX bit
+       &jc     (&label("mmx"));
+
+       &mov    (&DWP(24,"esp"),$tbl);
        &mov    ($s0,&DWP(0,$acc));             # load input data
        &mov    ($s1,&DWP(4,$acc));
        &mov    ($s2,&DWP(8,$acc));
        &mov    ($s3,&DWP(12,$acc));
-
-       &mov    (&DWP(24,"esp"),"ebp");
-
-       &call   ("_x86_AES_decrypt");
-
-       &mov    ("esp",&DWP(28,"esp"));
-
+       &call   ("_x86_AES_decrypt_compact");
+       &mov    ("esp",&DWP(28,"esp"));         # restore stack pointer
        &mov    ($acc,&wparam(1));              # load out
        &mov    (&DWP(0,$acc),$s0);             # write output data
        &mov    (&DWP(4,$acc),$s1);
        &mov    (&DWP(8,$acc),$s2);
        &mov    (&DWP(12,$acc),$s3);
+       &jmp    (&label("ret"));
+
+       &set_label("mmx",16);
+       &movq   ("mm0",&QWP(0,$acc));
+       &movq   ("mm4",&QWP(8,$acc));
+       &call   ("_mmx_AES_decrypt_compact");
+       &mov    ("esp",&DWP(28,"esp"));         # restore stack pointer
+       &mov    ($acc,&wparam(1));              # load out
+       &movq   (&QWP(0,$acc),"mm0");           # write output data
+       &movq   (&QWP(8,$acc),"mm4");
+       &emms   ();
+
+&set_label("ret",4);
 &function_end("AES_decrypt");
 
 # void AES_cbc_encrypt (const void char *inp, unsigned char *out,
@@ -1041,7 +1943,7 @@ my $mark=&DWP(72+240,"esp");      #copy of aes_key->rounds
 
        &call   (&label("pic_point"));          # make it PIC!
        &set_label("pic_point");
-       &blindpop("ebp");
+       &blindpop($tbl);
 
        &pushf  ();
        &cld    ();
@@ -1049,18 +1951,18 @@ my $mark=&DWP(72+240,"esp");    #copy of aes_key->rounds
        &cmp    (&wparam(5),0);
        &je     (&label("DECRYPT"));
 
-       &lea    ("ebp",&DWP(&label("AES_Te")."-".&label("pic_point"),"ebp"));
+       &lea    ($tbl,&DWP(&label("AES_Te")."-".&label("pic_point"),$tbl));
 
        # allocate aligned stack frame...
        &lea    ($key,&DWP(-76-244,"esp"));
        &and    ($key,-64);
 
        # ... and make sure it doesn't alias with AES_Te modulo 4096
-       &mov    ($s0,"ebp");
-       &lea    ($s1,&DWP(2048+256,"ebp"));
+       &mov    ($s0,$tbl);
+       &lea    ($s1,&DWP(2048,$tbl));
        &mov    ($s3,$key);
        &and    ($s0,0xfff);            # s = %ebp&0xfff
-       &and    ($s1,0xfff);            # e = (%ebp+2048+256)&0xfff
+       &and    ($s1,0xfff);            # e = (%ebp+2048)&0xfff
        &and    ($s3,0xfff);            # p = %esp&0xfff
 
        &cmp    ($s3,$s1);              # if (p>=e) %esp =- (p-e);
@@ -1099,11 +2001,11 @@ my $mark=&DWP(72+240,"esp");    #copy of aes_key->rounds
        # do we copy key schedule to stack?
        &mov    ($s1 eq "ebx" ? $s1 : "",$s3);
        &mov    ($s2 eq "ecx" ? $s2 : "",244/4);
-       &sub    ($s1,"ebp");
+       &sub    ($s1,$tbl);
        &mov    ("esi",$s3);
        &and    ($s1,0xfff);
        &lea    ("edi",$aes_key);
-       &cmp    ($s1,2048+256);
+       &cmp    ($s1,2048);
        &jb     (&label("do_ecopy"));
        &cmp    ($s1,4096-244);
        &jb     (&label("skip_ecopy"));
@@ -1114,18 +2016,18 @@ my $mark=&DWP(72+240,"esp");    #copy of aes_key->rounds
        &set_label("skip_ecopy");
 
        &mov    ($acc,$s0);
-       &mov    ($key,18);
+       &mov    ($key,16);
        &align  (4);
        &set_label("prefetch_te");
-               &mov    ($s0,&DWP(0,"ebp"));
-               &mov    ($s1,&DWP(32,"ebp"));
-               &mov    ($s2,&DWP(64,"ebp"));
-               &mov    ($s3,&DWP(96,"ebp"));
-               &lea    ("ebp",&DWP(128,"ebp"));
+               &mov    ($s0,&DWP(0,$tbl));
+               &mov    ($s1,&DWP(32,$tbl));
+               &mov    ($s2,&DWP(64,$tbl));
+               &mov    ($s3,&DWP(96,$tbl));
+               &lea    ($tbl,&DWP(128,$tbl));
                &dec    ($key);
        &jnz    (&label("prefetch_te"));
-       &sub    ("ebp",2048+256);
-       &mov    (&DWP(24,"esp"),"ebp");
+       &sub    ($tbl,2048);
+       &mov    (&DWP(24,"esp"),$tbl);
 
        &mov    ($s2,$_len);
        &mov    ($key,$_ivp);
@@ -1222,15 +2124,15 @@ my $mark=&DWP(72+240,"esp");    #copy of aes_key->rounds
 #----------------------------- DECRYPT -----------------------------#
 &align (4);
 &set_label("DECRYPT");
-       &lea    ("ebp",&DWP(&label("AES_Td")."-".&label("pic_point"),"ebp"));
+       &lea    ($tbl,&DWP(&label("AES_Td")."-".&label("pic_point"),$tbl));
 
        # allocate aligned stack frame...
        &lea    ($key,&DWP(-64-244,"esp"));
        &and    ($key,-64);
 
        # ... and make sure it doesn't alias with AES_Td modulo 4096
-       &mov    ($s0,"ebp");
-       &lea    ($s1,&DWP(2048+256,"ebp"));
+       &mov    ($s0,$tbl);
+       &lea    ($s1,&DWP(2048+256,$tbl));
        &mov    ($s3,$key);
        &and    ($s0,0xfff);            # s = %ebp&0xfff
        &and    ($s1,0xfff);            # e = (%ebp+2048+256)&0xfff
@@ -1272,7 +2174,7 @@ my $mark=&DWP(72+240,"esp");      #copy of aes_key->rounds
        # do we copy key schedule to stack?
        &mov    ($s1 eq "ebx" ? $s1 : "",$s3);
        &mov    ($s2 eq "ecx" ? $s2 : "",244/4);
-       &sub    ($s1,"ebp");
+       &sub    ($s1,$tbl);
        &mov    ("esi",$s3);
        &and    ($s1,0xfff);
        &lea    ("edi",$aes_key);
@@ -1290,15 +2192,15 @@ my $mark=&DWP(72+240,"esp");    #copy of aes_key->rounds
        &mov    ($key,18);
        &align  (4);
        &set_label("prefetch_td");
-               &mov    ($s0,&DWP(0,"ebp"));
-               &mov    ($s1,&DWP(32,"ebp"));
-               &mov    ($s2,&DWP(64,"ebp"));
-               &mov    ($s3,&DWP(96,"ebp"));
-               &lea    ("ebp",&DWP(128,"ebp"));
+               &mov    ($s0,&DWP(0,$tbl));
+               &mov    ($s1,&DWP(32,$tbl));
+               &mov    ($s2,&DWP(64,$tbl));
+               &mov    ($s3,&DWP(96,$tbl));
+               &lea    ($tbl,&DWP(128,$tbl));
                &dec    ($key);
        &jnz    (&label("prefetch_td"));
-       &sub    ("ebp",2048+256);
-       &mov    (&DWP(24,"esp"),"ebp");
+       &sub    ($tbl,2048+256);
+       &mov    (&DWP(24,"esp"),$tbl);
 
        &cmp    ($acc,$_out);
        &je     (&label("dec_in_place"));       # in-place processing...
@@ -1455,27 +2357,27 @@ my $mark=&DWP(72+240,"esp");    #copy of aes_key->rounds
 sub enckey()
 {
        &movz   ("esi",&LB("edx"));             # rk[i]>>0
-       &mov    ("ebx",&DWP(2,"ebp","esi",8));
+       &mov    ("ebx",&DWP(2,$tbl,"esi",8));
        &movz   ("esi",&HB("edx"));             # rk[i]>>8
        &and    ("ebx",0xFF000000);
        &xor    ("eax","ebx");
 
-       &mov    ("ebx",&DWP(2,"ebp","esi",8));
+       &mov    ("ebx",&DWP(2,$tbl,"esi",8));
        &shr    ("edx",16);
        &and    ("ebx",0x000000FF);
        &movz   ("esi",&LB("edx"));             # rk[i]>>16
        &xor    ("eax","ebx");
 
-       &mov    ("ebx",&DWP(0,"ebp","esi",8));
+       &mov    ("ebx",&DWP(0,$tbl,"esi",8));
        &movz   ("esi",&HB("edx"));             # rk[i]>>24
        &and    ("ebx",0x0000FF00);
        &xor    ("eax","ebx");
 
-       &mov    ("ebx",&DWP(0,"ebp","esi",8));
+       &mov    ("ebx",&DWP(0,$tbl,"esi",8));
        &and    ("ebx",0x00FF0000);
        &xor    ("eax","ebx");
 
-       &xor    ("eax",&DWP(2048+256,"ebp","ecx",4));   # rcon
+       &xor    ("eax",&DWP(2048+1024,$tbl,"ecx",4));   # rcon
 }
 
 # int AES_set_encrypt_key(const unsigned char *userKey, const int bits,
@@ -1492,8 +2394,8 @@ sub enckey()
 
        &call   (&label("pic_point"));
        &set_label("pic_point");
-       &blindpop("ebp");
-       &lea    ("ebp",&DWP(&label("AES_Te")."-".&label("pic_point"),"ebp"));
+       &blindpop($tbl);
+       &lea    ($tbl,&DWP(&label("AES_Te")."-".&label("pic_point"),$tbl));
 
        &mov    ("ecx",&wparam(1));             # number of bits in key
        &cmp    ("ecx",128);
@@ -1634,23 +2536,23 @@ sub enckey()
                &mov    ("edx","eax");
                &mov    ("eax",&DWP(16,"edi"));         # rk[4]
                &movz   ("esi",&LB("edx"));             # rk[11]>>0
-               &mov    ("ebx",&DWP(2,"ebp","esi",8));
+               &mov    ("ebx",&DWP(2,$tbl,"esi",8));
                &movz   ("esi",&HB("edx"));             # rk[11]>>8
                &and    ("ebx",0x000000FF);
                &xor    ("eax","ebx");
 
-               &mov    ("ebx",&DWP(0,"ebp","esi",8));
+               &mov    ("ebx",&DWP(0,$tbl,"esi",8));
                &shr    ("edx",16);
                &and    ("ebx",0x0000FF00);
                &movz   ("esi",&LB("edx"));             # rk[11]>>16
                &xor    ("eax","ebx");
 
-               &mov    ("ebx",&DWP(0,"ebp","esi",8));
+               &mov    ("ebx",&DWP(0,$tbl,"esi",8));
                &movz   ("esi",&HB("edx"));             # rk[11]>>24
                &and    ("ebx",0x00FF0000);
                &xor    ("eax","ebx");
 
-               &mov    ("ebx",&DWP(2,"ebp","esi",8));
+               &mov    ("ebx",&DWP(2,$tbl,"esi",8));
                &and    ("ebx",0xFF000000);
                &xor    ("eax","ebx");
 
@@ -1750,9 +2652,9 @@ sub deckey()
 
        &call   (&label("pic_point"));
        &set_label("pic_point");
-       blindpop("ebp");
-       &lea    ("edi",&DWP(&label("AES_Td")."-".&label("pic_point"),"ebp"));
-       &lea    ("ebp",&DWP(&label("AES_Te")."-".&label("pic_point"),"ebp"));
+       blindpop($tbl);
+       &lea    ("edi",&DWP(&label("AES_Td")."-".&label("pic_point"),$tbl));
+       &lea    ($tbl,&DWP(&label("AES_Te")."-".&label("pic_point"),$tbl));
 
        &mov    ("esi",&wparam(2));
        &mov    ("ecx",&DWP(240,"esi"));        # pull number of rounds
@@ -1760,10 +2662,10 @@ sub deckey()
        &align  (4);
        &set_label("permute");                  # permute the key schedule
                &add    ("esi",16);
-               &deckey (0,"esi","ebp","edi");
-               &deckey (4,"esi","ebp","edi");
-               &deckey (8,"esi","ebp","edi");
-               &deckey (12,"esi","ebp","edi");
+               &deckey (0,"esi",$tbl,"edi");
+               &deckey (4,"esi",$tbl,"edi");
+               &deckey (8,"esi",$tbl,"edi");
+               &deckey (12,"esi",$tbl,"edi");
                &dec    ("ecx");
        &jnz    (&label("permute"));