+}} else {{ # "June" MMX version...
+ # ... has slower "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),$Zhh);
+&function_end("gcm_gmult_4bit_mmx");
+\f
+######################################################################
+# Below subroutine is "528B" variant of "4-bit" GCM GHASH function
+# (see gcm128.c for details). It provides further 20-40% performance
+# improvement over above mentioned "May" version.
+
+&static_label("rem_8bit");
+
+&function_begin("gcm_ghash_4bit_mmx");
+{ my ($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, (u8)(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 data
+ &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],$red[2]);
+
+ # Just like in "May" verson modulo-schedule for critical path in
+ # 'Z.hi ^= rem_8bit[Z.lo&0xff^((u8)H[nhi]<<4)]<<48'. Final 'pxor'
+ # is scheduled so late that rem_8bit[] has to be shifted *right*
+ # by 16, which is why last argument to pinsrw is 2, which
+ # corresponds to <<32=<<48>>16...
+ for ($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); # Z>>=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",$nhi[0])); # rem^(H[nhi]<<4)
+
+ &pxor ($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 ($Zlo,4); # Z>>=4
+
+ &movq ($tmp,$Zhi);
+ &psrlq ($Zhi,4);
+ &shl ($rem[0],4); # rem<<4
+
+ &pxor ($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); # last is <<48
+
+ &psllq ($red[0],12); # correct by <<16>>4
+ &pxor ($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
projects / openssl.git / blobdiff