+# 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 ($__s0,$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,$__s0); # s[0]=t[0]
+ &movz ($v0,&LB($s3)); # 13,12, 7, 6*
+ &shr ($s3,16); # , ,13,12
+ &xor ($s2,&DWP(2,$te,$v0,8)); # 6
+ &mov ($key,$__key); # reincarnate v0 as key
+ &and ($s3,0xff); # , ,13,12*
+ &mov ($s3,&DWP(0,$te,$s3,8)); # 12
+ &xor ($s3,$s2); # s[2]=t[3] collected
+ &mov ($s2,$v1); # s[2]=t[2]
+}
+
+# More experimental code... SSE one... Even though this one eliminates
+# *all* references to stack, it's not faster...
+sub sse_encbody()
+{
+ &movz ($acc,&LB("eax")); # 0
+ &mov ("ecx",&DWP(0,$tbl,$acc,8)); # 0
+ &pshufw ("mm2","mm0",0x0d); # 7, 6, 3, 2
+ &movz ("edx",&HB("eax")); # 1
+ &mov ("edx",&DWP(3,$tbl,"edx",8)); # 1
+ &shr ("eax",16); # 5, 4
+
+ &movz ($acc,&LB("ebx")); # 10
+ &xor ("ecx",&DWP(2,$tbl,$acc,8)); # 10
+ &pshufw ("mm6","mm4",0x08); # 13,12, 9, 8
+ &movz ($acc,&HB("ebx")); # 11
+ &xor ("edx",&DWP(1,$tbl,$acc,8)); # 11
+ &shr ("ebx",16); # 15,14
+
+ &movz ($acc,&HB("eax")); # 5
+ &xor ("ecx",&DWP(3,$tbl,$acc,8)); # 5
+ &movq ("mm3",QWP(16,$key));
+ &movz ($acc,&HB("ebx")); # 15
+ &xor ("ecx",&DWP(1,$tbl,$acc,8)); # 15
+ &movd ("mm0","ecx"); # t[0] collected
+
+ &movz ($acc,&LB("eax")); # 4
+ &mov ("ecx",&DWP(0,$tbl,$acc,8)); # 4
+ &movd ("eax","mm2"); # 7, 6, 3, 2
+ &movz ($acc,&LB("ebx")); # 14
+ &xor ("ecx",&DWP(2,$tbl,$acc,8)); # 14
+ &movd ("ebx","mm6"); # 13,12, 9, 8
+
+ &movz ($acc,&HB("eax")); # 3
+ &xor ("ecx",&DWP(1,$tbl,$acc,8)); # 3
+ &movz ($acc,&HB("ebx")); # 9
+ &xor ("ecx",&DWP(3,$tbl,$acc,8)); # 9
+ &movd ("mm1","ecx"); # t[1] collected
+
+ &movz ($acc,&LB("eax")); # 2
+ &mov ("ecx",&DWP(2,$tbl,$acc,8)); # 2
+ &shr ("eax",16); # 7, 6
+ &punpckldq ("mm0","mm1"); # t[0,1] collected
+ &movz ($acc,&LB("ebx")); # 8
+ &xor ("ecx",&DWP(0,$tbl,$acc,8)); # 8
+ &shr ("ebx",16); # 13,12
+
+ &movz ($acc,&HB("eax")); # 7
+ &xor ("ecx",&DWP(1,$tbl,$acc,8)); # 7
+ &pxor ("mm0","mm3");
+ &movz ("eax",&LB("eax")); # 6
+ &xor ("edx",&DWP(2,$tbl,"eax",8)); # 6
+ &pshufw ("mm1","mm0",0x08); # 5, 4, 1, 0
+ &movz ($acc,&HB("ebx")); # 13
+ &xor ("ecx",&DWP(3,$tbl,$acc,8)); # 13
+ &xor ("ecx",&DWP(24,$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
+}
+
+######################################################################
+# "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;
+
+ # $Fn is used in first compact round and its purpose is to
+ # void restoration of some values from stack, so that after
+ # 4xenccompact with extra argument $key value is left there...
+ if ($i==3) { &$Fn ($key,$__key); }##%edx
+ else { &mov ($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 ($out,&BP(-128,$te,$out,1));
+
+ if ($i==3) { $tmp=$s[1]; }##%eax
+ &movz ($tmp,&HB($s[1]));
+ &movz ($tmp,&BP(-128,$te,$tmp,1));
+ &shl ($tmp,8);
+ &xor ($out,$tmp);
+
+ if ($i==3) { $tmp=$s[2]; &mov ($s[1],$__s0); }##%ebx
+ else { &mov ($tmp,$s[2]);
+ &shr ($tmp,16); }
+ if ($i==2) { &and ($s[1],0xFF); }#%edx[2]
+ &and ($tmp,0xFF);
+ &movz ($tmp,&BP(-128,$te,$tmp,1));
+ &shl ($tmp,16);
+ &xor ($out,$tmp);
+
+ if ($i==3) { $tmp=$s[3]; &mov ($s[2],$__s1); }##%ecx
+ elsif($i==2){ &movz ($tmp,&HB($s[3])); }#%ebx[2]
+ else { &mov ($tmp,$s[3]);
+ &shr ($tmp,24); }
+ &movz ($tmp,&BP(-128,$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); }
+ &comment();
+}
+
+sub enctransform()
+{ my @s = ($s0,$s1,$s2,$s3);
+ my $i = shift;
+ my $tmp = $tbl;
+ my $r2 = $key ;
+
+ &mov ($acc,$s[$i]);
+ &and ($acc,0x80808080);
+ &mov ($tmp,$acc);
+ &shr ($tmp,7);
+ &lea ($r2,&DWP(0,$s[$i],$s[$i]));
+ &sub ($acc,$tmp);
+ &and ($r2,0xfefefefe);
+ &and ($acc,0x1b1b1b1b);
+ &mov ($tmp,$s[$i]);
+ &xor ($acc,$r2); # r2
+
+ &xor ($s[$i],$acc); # r0 ^ r2
+ &rotl ($s[$i],24);
+ &xor ($s[$i],$acc) # ROTATE(r2^r0,24) ^ r2
+ &rotr ($tmp,16);
+ &xor ($s[$i],$tmp);
+ &rotr ($tmp,8);
+ &xor ($s[$i],$tmp);
+}
+
+&function_begin_B("_x86_AES_encrypt_compact");
+ # note that caller is expected to allocate stack frame for me!
+ &mov ($__key,$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 ($__end,$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 ($key,$__key);
+ &mov ($tbl,$__tbl);
+ &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,$__end);
+ &mov ($__key,$key);
+ &jb (&label("loop"));
+
+ &enccompact(0,$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" SSE block function.
+######################################################################
+#
+# Performance is not actually extraordinary in comparison to pure
+# x86 code. In particular encrypt performance is virtually the same.
+# Decrypt performance on the other hand is 15-20% better on newer
+# µ-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 SSE-only CPUs:-)
+
+sub sse_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 ($acc,&LB("eax")); # 0
+ &movz ("ecx",&BP(-128,$tbl,$acc,1)); # 0
+ &pshufw ("mm2","mm0",0x0d); # 7, 6, 3, 2
+ &movz ("edx",&HB("eax")); # 1
+ &movz ("edx",&BP(-128,$tbl,"edx",1)); # 1
+ &shl ("edx",8); # 1
+ &shr ("eax",16); # 5, 4
+
+ &movz ($acc,&LB("ebx")); # 10
+ &movz ($acc,&BP(-128,$tbl,$acc,1)); # 10
+ &shl ($acc,16); # 10
+ &or ("ecx",$acc); # 10
+ &pshufw ("mm6","mm4",0x08); # 13,12, 9, 8
+ &movz ($acc,&HB("ebx")); # 11
+ &movz ($acc,&BP(-128,$tbl,$acc,1)); # 11
+ &shl ($acc,24); # 11
+ &or ("edx",$acc); # 11
+ &shr ("ebx",16); # 15,14
+
+ &movz ($acc,&HB("eax")); # 5
+ &movz ($acc,&BP(-128,$tbl,$acc,1)); # 5
+ &shl ($acc,8); # 5
+ &or ("ecx",$acc); # 5
+ &movz ($acc,&HB("ebx")); # 15
+ &movz ($acc,&BP(-128,$tbl,$acc,1)); # 15
+ &shl ($acc,24); # 15
+ &or ("ecx",$acc); # 15
+ &movd ("mm0","ecx"); # t[0] collected
+
+ &movz ($acc,&LB("eax")); # 4
+ &movz ("ecx",&BP(-128,$tbl,$acc,1)); # 4
+ &movd ("eax","mm2"); # 7, 6, 3, 2
+ &movz ($acc,&LB("ebx")); # 14
+ &movz ($acc,&BP(-128,$tbl,$acc,1)); # 14
+ &shl ($acc,16); # 14
+ &or ("ecx",$acc); # 14
+
+ &movd ("ebx","mm6"); # 13,12, 9, 8
+ &movz ($acc,&HB("eax")); # 3
+ &movz ($acc,&BP(-128,$tbl,$acc,1)); # 3
+ &shl ($acc,24); # 3
+ &or ("ecx",$acc); # 3
+ &movz ($acc,&HB("ebx")); # 9
+ &movz ($acc,&BP(-128,$tbl,$acc,1)); # 9
+ &shl ($acc,8); # 9
+ &or ("ecx",$acc); # 9
+ &movd ("mm1","ecx"); # t[1] collected
+
+ &movz ($acc,&LB("ebx")); # 8
+ &movz ("ecx",&BP(-128,$tbl,$acc,1)); # 8
+ &shr ("ebx",16); # 13,12
+ &movz ($acc,&LB("eax")); # 2
+ &movz ($acc,&BP(-128,$tbl,$acc,1)); # 2
+ &shl ($acc,16); # 2
+ &or ("ecx",$acc); # 2
+ &shr ("eax",16); # 7, 6
+
+ &punpckldq ("mm0","mm1"); # t[0,1] collected
+
+ &movz ($acc,&HB("eax")); # 7
+ &movz ($acc,&BP(-128,$tbl,$acc,1)); # 7
+ &shl ($acc,24); # 7
+ &or ("ecx",$acc); # 7
+ &and ("eax",0xff); # 6
+ &movz ("eax",&BP(-128,$tbl,"eax",1)); # 6
+ &shl ("eax",16); # 6
+ &or ("edx","eax"); # 6
+ &movz ($acc,&HB("ebx")); # 13
+ &movz ($acc,&BP(-128,$tbl,$acc,1)); # 13
+ &shl ($acc,8); # 13
+ &or ("ecx",$acc); # 13
+ &movd ("mm4","ecx"); # t[2] collected
+ &and ("ebx",0xff); # 12
+ &movz ("ebx",&BP(-128,$tbl,"ebx",1)); # 12
+ &or ("edx","ebx"); # 12
+ &movd ("mm5","edx"); # t[3] collected
+
+ &punpckldq ("mm4","mm5"); # t[2,3] collected
+}
+
+ if (!$x86only) {
+&function_begin_B("_sse_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 ($__end,$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,$tbl));
+
+ &set_label("loop",16);
+ &sse_enccompact();
+ &add ($key,16);
+ &cmp ($key,$__end);
+ &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");
+ &pshufw ("mm2","mm0",0xb1); &pshufw ("mm6","mm4",0xb1);# ROTATE(r0,16)
+ &paddb ("mm0","mm0"); &paddb ("mm4","mm4");
+ &pxor ("mm0","mm3"); &pxor ("mm4","mm7"); # = r2
+ &pshufw ("mm3","mm2",0xb1); &pshufw ("mm7","mm6",0xb1);# r0
+ &pxor ("mm1","mm0"); &pxor ("mm5","mm4"); # r0^r2
+ &pxor ("mm0","mm2"); &pxor ("mm4","mm6"); # ^= ROTATE(r0,16)
+
+ &movq ("mm2","mm3"); &movq ("mm6","mm7");
+ &pslld ("mm3",8); &pslld ("mm7",8);
+ &psrld ("mm2",24); &psrld ("mm6",24);
+ &pxor ("mm0","mm3"); &pxor ("mm4","mm7"); # ^= r0<<8
+ &pxor ("mm0","mm2"); &pxor ("mm4","mm6"); # ^= r0>>24
+
+ &movq ("mm3","mm1"); &movq ("mm7","mm5");
+ &movq ("mm2",&QWP(0,$key)); &movq ("mm6",&QWP(8,$key));
+ &psrld ("mm1",8); &psrld ("mm5",8);
+ &mov ($s0,&DWP(0-128,$tbl));
+ &pslld ("mm3",24); &pslld ("mm7",24);
+ &mov ($s1,&DWP(64-128,$tbl));
+ &pxor ("mm0","mm1"); &pxor ("mm4","mm5"); # ^= (r2^r0)<<8
+ &mov ($s2,&DWP(128-128,$tbl));
+ &pxor ("mm0","mm3"); &pxor ("mm4","mm7"); # ^= (r2^r0)>>24
+ &mov ($s3,&DWP(192-128,$tbl));
+
+ &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("_sse_AES_encrypt_compact");
+ }
+
+######################################################################
+# Vanilla block function.
+######################################################################
+