ghash-x86.pl: "528B" variant of gcm_ghash_4bit_mmx gives 20-40%
authorAndy Polyakov <appro@openssl.org>
Fri, 4 Jun 2010 13:21:01 +0000 (13:21 +0000)
committerAndy Polyakov <appro@openssl.org>
Fri, 4 Jun 2010 13:21:01 +0000 (13:21 +0000)
improvement.

crypto/modes/asm/ghash-x86.pl

index d74c337..9fa4da1 100644 (file)
@@ -7,23 +7,25 @@
 # details see http://www.openssl.org/~appro/cryptogams/.
 # ====================================================================
 #
-# March, May 2010
+# March, May, June 2010
 #
 # The module implements "4-bit" GCM GHASH function and underlying
 # single multiplication operation in GF(2^128). "4-bit" means that it
 # uses 256 bytes per-key table [+64/128 bytes fixed table]. It has two
 # code paths: vanilla x86 and vanilla MMX. Former will be executed on
-# 486 and Pentium, latter on all others. Performance results are for
-# streamed GHASH subroutine and are expressed in cycles per processed
-# byte, less is better:
+# 486 and Pentium, latter on all others. MMX GHASH features so called
+# "528B" variant of "4-bit" method utilizing additional 256+16 bytes
+# of per-key storage [+512 bytes shared table]. Performance results
+# are for streamed GHASH subroutine and are expressed in cycles per
+# processed byte, less is better:
 #
 #              gcc 2.95.3(*)   MMX assembler   x86 assembler
 #
 # Pentium      100/112(**)     -               50
-# PIII         63 /77          14.5            24
-# P4           96 /122         24.5            84(***)
-# Opteron      50 /71          14.5            30
-# Core2                54 /68          10.5            18
+# PIII         63 /77          12.2            24
+# P4           96 /122         18.0            84(***)
+# Opteron      50 /71          10.1            30
+# Core2                54 /68          8.6             18
 #
 # (*)  gcc 3.4.x was observed to generate few percent slower code,
 #      which is one of reasons why 2.95.3 results were chosen,
@@ -33,7 +35,7 @@
 #      position-independent;
 # (***)        see comment in non-MMX routine for further details;
 #
-# To summarize, it's >2-4 times faster than gcc-generated code. To
+# To summarize, it's >2-5 times faster than gcc-generated code. To
 # anchor it to something else SHA1 assembler processes one byte in
 # 11-13 cycles on contemporary x86 cores. As for choice of MMX in
 # particular, see comment at the end of the file...
@@ -318,6 +320,10 @@ if (!$x86only) {{{
 
 &static_label("rem_4bit");
 
+if (0) {{      # "May" MMX version is kept for reference...
+
+$S=12;         # shift factor for rem_4bit
+
 &function_begin_B("_mmx_gmult_4bit_inner");
 # MMX version performs 3.5 times better on P4 (see comment in non-MMX
 # routine for further details), 100% better on Opteron, ~70% better
@@ -465,6 +471,329 @@ if (!$x86only) {{{
        &stack_pop(4+1);
 &function_end("gcm_ghash_4bit_mmx");
 \f
+}} else {{     # "June" MMX version...
+               # ... has "April" gcm_gmult_4bit_mmx with folded loop.
+               # This is done to conserve code size...
+$S=16;         # shift factor for rem_4bit
+
+sub mmx_loop() {
+# MMX version performs 2.8 times better on P4 (see comment in non-MMX
+# routine for further details), 40% better on Opteron and Core2, 50%
+# better on PIII... In other words effort is considered to be well
+# spent...
+    my $inp = shift;
+    my $rem_4bit = shift;
+    my $cnt = $Zhh;
+    my $nhi = $Zhl;
+    my $nlo = $Zlh;
+    my $rem = $Zll;
+
+    my ($Zlo,$Zhi) = ("mm0","mm1");
+    my $tmp = "mm2";
+
+       &xor    ($nlo,$nlo);    # avoid partial register stalls on PIII
+       &mov    ($nhi,$Zll);
+       &mov    (&LB($nlo),&LB($nhi));
+       &mov    ($cnt,14);
+       &shl    (&LB($nlo),4);
+       &and    ($nhi,0xf0);
+       &movq   ($Zlo,&QWP(8,$Htbl,$nlo));
+       &movq   ($Zhi,&QWP(0,$Htbl,$nlo));
+       &movd   ($rem,$Zlo);
+       &jmp    (&label("mmx_loop"));
+
+    &set_label("mmx_loop",16);
+       &psrlq  ($Zlo,4);
+       &and    ($rem,0xf);
+       &movq   ($tmp,$Zhi);
+       &psrlq  ($Zhi,4);
+       &pxor   ($Zlo,&QWP(8,$Htbl,$nhi));
+       &mov    (&LB($nlo),&BP(0,$inp,$cnt));
+       &psllq  ($tmp,60);
+       &pxor   ($Zhi,&QWP(0,$rem_4bit,$rem,8));
+       &dec    ($cnt);
+       &movd   ($rem,$Zlo);
+       &pxor   ($Zhi,&QWP(0,$Htbl,$nhi));
+       &mov    ($nhi,$nlo);
+       &pxor   ($Zlo,$tmp);
+       &js     (&label("mmx_break"));
+
+       &shl    (&LB($nlo),4);
+       &and    ($rem,0xf);
+       &psrlq  ($Zlo,4);
+       &and    ($nhi,0xf0);
+       &movq   ($tmp,$Zhi);
+       &psrlq  ($Zhi,4);
+       &pxor   ($Zlo,&QWP(8,$Htbl,$nlo));
+       &psllq  ($tmp,60);
+       &pxor   ($Zhi,&QWP(0,$rem_4bit,$rem,8));
+       &movd   ($rem,$Zlo);
+       &pxor   ($Zhi,&QWP(0,$Htbl,$nlo));
+       &pxor   ($Zlo,$tmp);
+       &jmp    (&label("mmx_loop"));
+
+    &set_label("mmx_break",16);
+       &shl    (&LB($nlo),4);
+       &and    ($rem,0xf);
+       &psrlq  ($Zlo,4);
+       &and    ($nhi,0xf0);
+       &movq   ($tmp,$Zhi);
+       &psrlq  ($Zhi,4);
+       &pxor   ($Zlo,&QWP(8,$Htbl,$nlo));
+       &psllq  ($tmp,60);
+       &pxor   ($Zhi,&QWP(0,$rem_4bit,$rem,8));
+       &movd   ($rem,$Zlo);
+       &pxor   ($Zhi,&QWP(0,$Htbl,$nlo));
+       &pxor   ($Zlo,$tmp);
+
+       &psrlq  ($Zlo,4);
+       &and    ($rem,0xf);
+       &movq   ($tmp,$Zhi);
+       &psrlq  ($Zhi,4);
+       &pxor   ($Zlo,&QWP(8,$Htbl,$nhi));
+       &psllq  ($tmp,60);
+       &pxor   ($Zhi,&QWP(0,$rem_4bit,$rem,8));
+       &movd   ($rem,$Zlo);
+       &pxor   ($Zhi,&QWP(0,$Htbl,$nhi));
+       &pxor   ($Zlo,$tmp);
+
+       &psrlq  ($Zlo,32);      # lower part of Zlo is already there
+       &movd   ($Zhl,$Zhi);
+       &psrlq  ($Zhi,32);
+       &movd   ($Zlh,$Zlo);
+       &movd   ($Zhh,$Zhi);
+
+       &bswap  ($Zll);
+       &bswap  ($Zhl);
+       &bswap  ($Zlh);
+       &bswap  ($Zhh);
+}
+
+&function_begin("gcm_gmult_4bit_mmx");
+       &mov    ($inp,&wparam(0));      # load Xi
+       &mov    ($Htbl,&wparam(1));     # load Htable
+
+       &call   (&label("pic_point"));
+       &set_label("pic_point");
+       &blindpop("eax");
+       &lea    ("eax",&DWP(&label("rem_4bit")."-".&label("pic_point"),"eax"));
+
+       &movz   ($Zll,&BP(15,$inp));
+
+       &mmx_loop($inp,"eax");
+
+       &emms   ();
+       &mov    (&DWP(12,$inp),$Zll);
+       &mov    (&DWP(4,$inp),$Zhl);
+       &mov    (&DWP(8,$inp),$Zlh);
+       &mov    (&DWP(0,$inp),$Zhh);
+&function_end("gcm_gmult_4bit_mmx");
+\f
+######################################################################
+# Below subroutine is "528B" variant of "4-bit" GCM GHASH function
+# (see gcm128.c for details). It provides further 20-40% performance
+# improvement over *previous* version of this module.
+
+&static_label("rem_8bit");
+
+&function_begin("gcm_ghash_4bit_mmx");
+{ my ($Zlo,$Zhi) = ("mm7","mm6");
+  my $rem_8bit = "esi";
+  my $Htbl = "ebx";
+
+    # parameter block
+    &mov       ("eax",&wparam(0));             # Xi
+    &mov       ("ebx",&wparam(1));             # Htable
+    &mov       ("ecx",&wparam(2));             # inp
+    &mov       ("edx",&wparam(3));             # len
+    &mov       ("ebp","esp");                  # original %esp
+    &call      (&label("pic_point"));
+    &set_label ("pic_point");
+    &blindpop  ($rem_8bit);
+    &lea       ($rem_8bit,&DWP(&label("rem_8bit")."-".&label("pic_point"),$rem_8bit));
+
+    &sub       ("esp",512+16+16);              # allocate stack frame...
+    &and       ("esp",-64);                    # ...and align it
+    &sub       ("esp",16);                     # place for (u8)(H[]<<4)
+
+    &add       ("edx","ecx");                  # pointer to the end of input
+    &mov       (&DWP(528+16+0,"esp"),"eax");   # save Xi
+    &mov       (&DWP(528+16+8,"esp"),"edx");   # save inp+len
+    &mov       (&DWP(528+16+12,"esp"),"ebp");  # save original %esp
+
+    { my @lo  = ("mm0","mm1","mm2");
+      my @hi  = ("mm3","mm4","mm5");
+      my @tmp = ("mm6","mm7");
+      my $off1=0,$off2=0,$i;
+
+      &add     ($Htbl,128);                    # optimize for size
+      &lea     ("edi",&DWP(16+128,"esp"));
+      &lea     ("ebp",&DWP(16+256+128,"esp"));
+
+      # decompose Htable (low and high parts are kept separately),
+      # generate Htable>>4, save to stack...
+      for ($i=0;$i<18;$i++) {
+
+       &mov    ("edx",&DWP(16*$i+8-128,$Htbl))         if ($i<16);
+       &movq   ($lo[0],&QWP(16*$i+8-128,$Htbl))        if ($i<16);
+       &psllq  ($tmp[1],60)                            if ($i>1);
+       &movq   ($hi[0],&QWP(16*$i+0-128,$Htbl))        if ($i<16);
+       &por    ($lo[2],$tmp[1])                        if ($i>1);
+       &movq   (&QWP($off1-128,"edi"),$lo[1])          if ($i>0 && $i<17);
+       &psrlq  ($lo[1],4)                              if ($i>0 && $i<17);
+       &movq   (&QWP($off1,"edi"),$hi[1])              if ($i>0 && $i<17);
+       &movq   ($tmp[0],$hi[1])                        if ($i>0 && $i<17);
+       &movq   (&QWP($off2-128,"ebp"),$lo[2])          if ($i>1);
+       &psrlq  ($hi[1],4)                              if ($i>0 && $i<17);
+       &movq   (&QWP($off2,"ebp"),$hi[2])              if ($i>1);
+       &shl    ("edx",4)                               if ($i<16);
+       &mov    (&BP($i,"esp"),&LB("edx"))              if ($i<16);
+
+       unshift (@lo,pop(@lo));                 # "rotate" registers
+       unshift (@hi,pop(@hi));
+       unshift (@tmp,pop(@tmp));
+       $off1 += 8      if ($i>0);
+       $off2 += 8      if ($i>1);
+      }
+    }
+
+    &movq      ($Zhi,&QWP(0,"eax"));
+    &mov       ("ebx",&DWP(8,"eax"));
+    &mov       ("edx",&DWP(12,"eax"));         # load Xi
+
+&set_label("outer",16);
+  { my $nlo = "eax";
+    my $dat = "edx";
+    my @nhi = ("edi","ebp");
+    my @rem = ("ebx","ecx");
+    my @red = ("mm0","mm1","mm2");
+    my $tmp = "mm3";
+
+    &xor       ($dat,&DWP(12,"ecx"));          # merge input
+    &xor       ("ebx",&DWP(8,"ecx"));
+    &pxor      ($Zhi,&QWP(0,"ecx"));
+    &lea       ("ecx",&DWP(16,"ecx"));         # inp+=16
+    #&mov      (&DWP(528+12,"esp"),$dat);      # save inp^Xi
+    &mov       (&DWP(528+8,"esp"),"ebx");
+    &movq      (&QWP(528+0,"esp"),$Zhi);
+    &mov       (&DWP(528+16+4,"esp"),"ecx");   # save inp
+
+    &xor       ($nlo,$nlo);
+    &rol       ($dat,8);
+    &mov       (&LB($nlo),&LB($dat));
+    &mov       ($nhi[1],$nlo);
+    &and       (&LB($nlo),0x0f);
+    &shr       ($nhi[1],4);
+    &pxor      ($red[0],$red[0]);
+    &rol       ($dat,8);                               # next byte
+    &pxor      ($red[1],$red[1]);
+    &pxor      ($red[2],$red[2]);
+
+    # Just like in "May" verson modulo-schedule for critical path in
+    # 'Z.hi ^= rem_8bit[Z.lo&0xff^((u8)H[nhi]<<4)]<<48'. Final xor
+    # is scheduled so late that rem_8bit is shifted *right* by 16,
+    # which is why last argument to pinsrw is 2, which corresponds to
+    # <<32...
+    for ($j=11,$i=0;$i<15;$i++) {
+
+      if ($i>0) {
+       &pxor   ($Zlo,&QWP(16,"esp",$nlo,8));           # Z^=H[nlo]
+       &rol    ($dat,8);                               # next byte
+       &pxor   ($Zhi,&QWP(16+128,"esp",$nlo,8));
+
+       &pxor   ($Zlo,$tmp);
+       &pxor   ($Zhi,&QWP(16+256+128,"esp",$nhi[0],8));
+       &xor    (&LB($rem[1]),&BP(0,"esp",$nhi[0]));    # rem^H[nhi]<<4
+      } else {
+       &movq   ($Zlo,&QWP(16,"esp",$nlo,8));
+       &movq   ($Zhi,&QWP(16+128,"esp",$nlo,8));
+      }
+
+       &mov    (&LB($nlo),&LB($dat));
+       &mov    ($dat,&DWP(528+$j,"esp"))       if (--$j%4==0);
+
+       &movd   ($rem[0],$Zlo);
+       &movz   ($rem[1],&LB($rem[1]))          if ($i>0);
+       &psrlq  ($Zlo,8);
+
+       &movq   ($tmp,$Zhi);
+       &mov    ($nhi[0],$nlo);
+       &psrlq  ($Zhi,8);
+
+       &pxor   ($Zlo,&QWP(16+256+0,"esp",$nhi[1],8));  # Z^=H[nhi]>>4
+       &and    (&LB($nlo),0x0f);
+       &psllq  ($tmp,56);
+
+       &pxor   ($Zhi,$red[1])                          if ($i>1);
+       &shr    ($nhi[0],4);
+       &pinsrw ($red[0],&WP(0,$rem_8bit,$rem[1],2),2)  if ($i>0);
+
+       unshift (@red,pop(@red));                       # "rotate" registers
+       unshift (@rem,pop(@rem));
+       unshift (@nhi,pop(@nhi));
+    }
+
+    &pxor      ($Zlo,&QWP(16,"esp",$nlo,8));           # Z^=H[nlo]
+    &pxor      ($Zhi,&QWP(16+128,"esp",$nlo,8));
+    &xor       (&LB($rem[1]),&BP(0,"esp",$nhi[0]));    #$rem[0]);                      # rem^H[nhi]<<4
+
+    &pxor      ($Zlo,$tmp);
+    &pxor      ($Zhi,&QWP(16+256+128,"esp",$nhi[0],8));
+    &movz      ($rem[1],&LB($rem[1]));
+
+    &pxor      ($red[2],$red[2]);                      # clear 2nd word
+    &psllq     ($red[1],4);
+
+    &movd      ($rem[0],$Zlo);
+    &psrlq     ($Zlo,4);
+
+    &movq      ($tmp,$Zhi);
+    &psrlq     ($Zhi,4);
+    &shl       ($rem[0],4);
+
+    &pxor      ($Zlo,&QWP(16,"esp",$nhi[1],8));        # Z^=H[nhi]
+    &psllq     ($tmp,60);
+    &movz      ($rem[0],&LB($rem[0]));
+
+    &pxor      ($Zlo,$tmp);
+    &pxor      ($Zhi,&QWP(16+128,"esp",$nhi[1],8));
+
+    &pinsrw    ($red[0],&WP(0,$rem_8bit,$rem[1],2),2);
+    &pxor      ($Zhi,$red[1]);
+
+    &movd      ($dat,$Zlo);
+    &pinsrw    ($red[2],&WP(0,$rem_8bit,$rem[0],2),3);
+
+    &psllq     ($red[0],12);
+    &pxor      ($Zhi,$red[0]);
+    &psrlq     ($Zlo,32);
+    &pxor      ($Zhi,$red[2]);
+
+    &mov       ("ecx",&DWP(528+16+4,"esp"));   # restore inp
+    &movd      ("ebx",$Zlo);
+    &movq      ($tmp,$Zhi);                    # 01234567
+    &psllw     ($Zhi,8);                       # 1.3.5.7.
+    &psrlw     ($tmp,8);                       # .0.2.4.6
+    &por       ($Zhi,$tmp);                    # 10325476
+    &bswap     ($dat);
+    &pshufw    ($Zhi,$Zhi,0b00011011);         # 76543210
+    &bswap     ("ebx");
+    
+    &cmp       ("ecx",&DWP(528+16+8,"esp"));   # are we done?
+    &jne       (&label("outer"));
+  }
+
+    &mov       ("eax",&DWP(528+16+0,"esp"));   # restore Xi
+    &mov       (&DWP(12,"eax"),"edx");
+    &mov       (&DWP(8,"eax"),"ebx");
+    &movq      (&QWP(0,"eax"),$Zhi);
+
+    &mov       ("esp",&DWP(528+16+12,"esp"));  # restore original %esp
+    &emms      ();
+}
+&function_end("gcm_ghash_4bit_mmx");
+}}
+\f
 if ($sse2) {{
 ######################################################################
 # PCLMULQDQ version.
@@ -936,10 +1265,43 @@ my ($Xhi,$Xi)=@_;
 }}     # $sse2
 
 &set_label("rem_4bit",64);
-       &data_word(0,0x0000<<12,0,0x1C20<<12,0,0x3840<<12,0,0x2460<<12);
-       &data_word(0,0x7080<<12,0,0x6CA0<<12,0,0x48C0<<12,0,0x54E0<<12);
-       &data_word(0,0xE100<<12,0,0xFD20<<12,0,0xD940<<12,0,0xC560<<12);
-       &data_word(0,0x9180<<12,0,0x8DA0<<12,0,0xA9C0<<12,0,0xB5E0<<12);
+       &data_word(0,0x0000<<$S,0,0x1C20<<$S,0,0x3840<<$S,0,0x2460<<$S);
+       &data_word(0,0x7080<<$S,0,0x6CA0<<$S,0,0x48C0<<$S,0,0x54E0<<$S);
+       &data_word(0,0xE100<<$S,0,0xFD20<<$S,0,0xD940<<$S,0,0xC560<<$S);
+       &data_word(0,0x9180<<$S,0,0x8DA0<<$S,0,0xA9C0<<$S,0,0xB5E0<<$S);
+&set_label("rem_8bit",64);
+       &data_short(0x0000,0x01C2,0x0384,0x0246,0x0708,0x06CA,0x048C,0x054E);
+       &data_short(0x0E10,0x0FD2,0x0D94,0x0C56,0x0918,0x08DA,0x0A9C,0x0B5E);
+       &data_short(0x1C20,0x1DE2,0x1FA4,0x1E66,0x1B28,0x1AEA,0x18AC,0x196E);
+       &data_short(0x1230,0x13F2,0x11B4,0x1076,0x1538,0x14FA,0x16BC,0x177E);
+       &data_short(0x3840,0x3982,0x3BC4,0x3A06,0x3F48,0x3E8A,0x3CCC,0x3D0E);
+       &data_short(0x3650,0x3792,0x35D4,0x3416,0x3158,0x309A,0x32DC,0x331E);
+       &data_short(0x2460,0x25A2,0x27E4,0x2626,0x2368,0x22AA,0x20EC,0x212E);
+       &data_short(0x2A70,0x2BB2,0x29F4,0x2836,0x2D78,0x2CBA,0x2EFC,0x2F3E);
+       &data_short(0x7080,0x7142,0x7304,0x72C6,0x7788,0x764A,0x740C,0x75CE);
+       &data_short(0x7E90,0x7F52,0x7D14,0x7CD6,0x7998,0x785A,0x7A1C,0x7BDE);
+       &data_short(0x6CA0,0x6D62,0x6F24,0x6EE6,0x6BA8,0x6A6A,0x682C,0x69EE);
+       &data_short(0x62B0,0x6372,0x6134,0x60F6,0x65B8,0x647A,0x663C,0x67FE);
+       &data_short(0x48C0,0x4902,0x4B44,0x4A86,0x4FC8,0x4E0A,0x4C4C,0x4D8E);
+       &data_short(0x46D0,0x4712,0x4554,0x4496,0x41D8,0x401A,0x425C,0x439E);
+       &data_short(0x54E0,0x5522,0x5764,0x56A6,0x53E8,0x522A,0x506C,0x51AE);
+       &data_short(0x5AF0,0x5B32,0x5974,0x58B6,0x5DF8,0x5C3A,0x5E7C,0x5FBE);
+       &data_short(0xE100,0xE0C2,0xE284,0xE346,0xE608,0xE7CA,0xE58C,0xE44E);
+       &data_short(0xEF10,0xEED2,0xEC94,0xED56,0xE818,0xE9DA,0xEB9C,0xEA5E);
+       &data_short(0xFD20,0xFCE2,0xFEA4,0xFF66,0xFA28,0xFBEA,0xF9AC,0xF86E);
+       &data_short(0xF330,0xF2F2,0xF0B4,0xF176,0xF438,0xF5FA,0xF7BC,0xF67E);
+       &data_short(0xD940,0xD882,0xDAC4,0xDB06,0xDE48,0xDF8A,0xDDCC,0xDC0E);
+       &data_short(0xD750,0xD692,0xD4D4,0xD516,0xD058,0xD19A,0xD3DC,0xD21E);
+       &data_short(0xC560,0xC4A2,0xC6E4,0xC726,0xC268,0xC3AA,0xC1EC,0xC02E);
+       &data_short(0xCB70,0xCAB2,0xC8F4,0xC936,0xCC78,0xCDBA,0xCFFC,0xCE3E);
+       &data_short(0x9180,0x9042,0x9204,0x93C6,0x9688,0x974A,0x950C,0x94CE);
+       &data_short(0x9F90,0x9E52,0x9C14,0x9DD6,0x9898,0x995A,0x9B1C,0x9ADE);
+       &data_short(0x8DA0,0x8C62,0x8E24,0x8FE6,0x8AA8,0x8B6A,0x892C,0x88EE);
+       &data_short(0x83B0,0x8272,0x8034,0x81F6,0x84B8,0x857A,0x873C,0x86FE);
+       &data_short(0xA9C0,0xA802,0xAA44,0xAB86,0xAEC8,0xAF0A,0xAD4C,0xAC8E);
+       &data_short(0xA7D0,0xA612,0xA454,0xA596,0xA0D8,0xA11A,0xA35C,0xA29E);
+       &data_short(0xB5E0,0xB422,0xB664,0xB7A6,0xB2E8,0xB32A,0xB16C,0xB0AE);
+       &data_short(0xBBF0,0xBA32,0xB874,0xB9B6,0xBCF8,0xBD3A,0xBF7C,0xBEBE);
 }}}    # !$x86only
 
 &asciz("GHASH for x86, CRYPTOGAMS by <appro\@openssl.org>");
@@ -957,13 +1319,12 @@ my ($Xhi,$Xi)=@_;
 # per processed byte out of 64KB block. Recall that this number accounts
 # even for 64KB table setup overhead. As discussed in gcm128.c we choose
 # to be more conservative in respect to lookup table sizes, but how
-# do the results compare? As per table in the beginning, minimalistic
-# MMX version delivers ~11 cycles on same platform. As also discussed in
-# gcm128.c, next in line "8-bit Shoup's" method should deliver twice the
-# performance of "4-bit" one. It should be also be noted that in SSE2
-# case improvement can be "super-linear," i.e. more than twice, mostly
-# because >>8 maps to single instruction on SSE2 register. This is
-# unlike "4-bit" case when >>4 maps to same amount of instructions in
-# both MMX and SSE2 cases. Bottom line is that switch to SSE2 is
-# considered to be justifiable only in case we choose to implement
-# "8-bit" method...
+# do the results compare? Minimalistic "256B" MMX version delivers ~11
+# cycles on same platform. As also discussed in gcm128.c, next in line
+# "8-bit Shoup's" method should deliver twice the performance of "4-bit"
+# one. It should be also be noted that in SSE2 case improvement can be
+# "super-linear," i.e. more than twice, mostly because >>8 maps to
+# single instruction on SSE2 register. This is unlike "4-bit" case when
+# >>4 maps to same amount of instructions in both MMX and SSE2 cases.
+# Bottom line is that switch to SSE2 is considered to be justifiable
+# only in case we choose to implement "8-bit" method...