# details see http://www.openssl.org/~appro/cryptogams/.
# ====================================================================
#
-# March 2010
+# March, 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 [+128 bytes shared table]. Performance
-# results are for streamed GHASH subroutine and are expressed in
-# cycles per processed byte, less is better:
+# single multiplication operation in GF(2^128). "4-bit" means that
+# it uses 256 bytes per-key table [+128 bytes shared table]. GHASH
+# function features so called "528B" variant utilizing additional
+# 256+16 bytes of per-key storage [+512 bytes shared table].
+# Performance results are for this streamed GHASH subroutine and are
+# expressed in cycles per processed byte, less is better:
#
-# gcc 3.4.x assembler
+# gcc 3.4.x(*) assembler
#
-# Opteron 18.5 10.2 +80%
-# Core2 26.0 16.4 +58%
+# P4 28.6 14.0 +100%
+# Opteron 19.3 7.7 +150%
+# Core2 17.8 8.1(**) +120%
+# Atom 31.6 16.8 +88%
+# VIA Nano 21.8 10.1 +115%
+#
+# (*) comparison is not completely fair, because C results are
+# for vanilla "256B" implementation, while assembler results
+# are for "528B";-)
+# (**) it's mystery [to me] why Core2 result is not same as for
+# Opteron;
+
+# May 2010
+#
+# Add PCLMULQDQ version performing at 2.02 cycles per processed byte.
+# See ghash-x86.pl for background information and details about coding
+# techniques.
+#
+# Special thanks to David Woodhouse <dwmw2@infradead.org> for
+# providing access to a Westmere-based system on behalf of Intel
+# Open Source Technology Centre.
+
+# December 2012
+#
+# Overhaul: aggregate Karatsuba post-processing, improve ILP in
+# reduction_alg9, increase reduction aggregate factor to 4x. As for
+# the latter. ghash-x86.pl discusses that it makes lesser sense to
+# increase aggregate factor. Then why increase here? Critical path
+# consists of 3 independent pclmulqdq instructions, Karatsuba post-
+# processing and reduction. "On top" of this we lay down aggregated
+# multiplication operations, triplets of independent pclmulqdq's. As
+# issue rate for pclmulqdq is limited, it makes lesser sense to
+# aggregate more multiplications than it takes to perform remaining
+# non-multiplication operations. 2x is near-optimal coefficient for
+# contemporary Intel CPUs (therefore modest improvement coefficient),
+# but not for Bulldozer. Latter is because logical SIMD operations
+# are twice as slow in comparison to Intel, so that critical path is
+# longer. A CPU with higher pclmulqdq issue rate would also benefit
+# from higher aggregate factor...
+#
+# Westmere 1.76(+14%)
+# Sandy Bridge 1.79(+9%)
+# Ivy Bridge 1.79(+8%)
+# Bulldozer 1.52(+25%)
$flavour = shift;
$output = shift;
( $xlate="${dir}../../perlasm/x86_64-xlate.pl" and -f $xlate) or
die "can't locate x86_64-xlate.pl";
-open STDOUT,"| $^X $xlate $flavour $output";
+open OUT,"| \"$^X\" $xlate $flavour $output";
+*STDOUT=*OUT;
+
+$do4xaggr=1;
# common register layout
$nlo="%rax";
$tmp="%r10";
$rem_4bit = "%r11";
-# per-function register layout
$Xi="%rdi";
$Htbl="%rsi";
+# per-function register layout
$cnt="%rcx";
$rem="%rdx";
-sub lo() { my $r=shift; $r =~ s/%[er]([a-d])x/%\1l/;
- $r =~ s/%[er]([sd]i)/%\1l/;
+sub LB() { my $r=shift; $r =~ s/%[er]([a-d])x/%\1l/ or
+ $r =~ s/%[er]([sd]i)/%\1l/ or
+ $r =~ s/%[er](bp)/%\1l/ or
$r =~ s/%(r[0-9]+)[d]?/%\1b/; $r; }
+sub AUTOLOAD() # thunk [simplified] 32-bit style perlasm
+{ my $opcode = $AUTOLOAD; $opcode =~ s/.*:://;
+ my $arg = pop;
+ $arg = "\$$arg" if ($arg*1 eq $arg);
+ $code .= "\t$opcode\t".join(',',$arg,reverse @_)."\n";
+}
+\f
{ my $N;
sub loop() {
my $inp = shift;
$code.=<<___;
xor $nlo,$nlo
xor $nhi,$nhi
- mov `&lo("$Zlo")`,`&lo("$nlo")`
- mov `&lo("$Zlo")`,`&lo("$nhi")`
- shl \$4,`&lo("$nlo")`
+ mov `&LB("$Zlo")`,`&LB("$nlo")`
+ mov `&LB("$Zlo")`,`&LB("$nhi")`
+ shl \$4,`&LB("$nlo")`
mov \$14,$cnt
mov 8($Htbl,$nlo),$Zlo
mov ($Htbl,$nlo),$Zhi
- and \$0xf0,`&lo("$nhi")`
+ and \$0xf0,`&LB("$nhi")`
mov $Zlo,$rem
jmp .Loop$N
shr \$4,$Zlo
and \$0xf,$rem
mov $Zhi,$tmp
- mov ($inp,$cnt),`&lo("$nlo")`
+ mov ($inp,$cnt),`&LB("$nlo")`
shr \$4,$Zhi
xor 8($Htbl,$nhi),$Zlo
shl \$60,$tmp
xor ($Htbl,$nhi),$Zhi
- mov `&lo("$nlo")`,`&lo("$nhi")`
+ mov `&LB("$nlo")`,`&LB("$nhi")`
xor ($rem_4bit,$rem,8),$Zhi
mov $Zlo,$rem
- shl \$4,`&lo("$nlo")`
+ shl \$4,`&LB("$nlo")`
xor $tmp,$Zlo
dec $cnt
js .Lbreak$N
xor 8($Htbl,$nlo),$Zlo
shl \$60,$tmp
xor ($Htbl,$nlo),$Zhi
- and \$0xf0,`&lo("$nhi")`
+ and \$0xf0,`&LB("$nhi")`
xor ($rem_4bit,$rem,8),$Zhi
mov $Zlo,$rem
xor $tmp,$Zlo
xor 8($Htbl,$nlo),$Zlo
shl \$60,$tmp
xor ($Htbl,$nlo),$Zhi
- and \$0xf0,`&lo("$nhi")`
+ and \$0xf0,`&LB("$nhi")`
xor ($rem_4bit,$rem,8),$Zhi
mov $Zlo,$rem
xor $tmp,$Zlo
ret
.size gcm_gmult_4bit,.-gcm_gmult_4bit
___
-
-
+\f
# per-function register layout
-$inp="%rdi";
-$len="%rsi";
-$Xi="%rdx";
-$Htbl="%rcx";
-
-$cnt="%rbp";
-$rem="%r12";
+$inp="%rdx";
+$len="%rcx";
+$rem_8bit=$rem_4bit;
$code.=<<___;
.globl gcm_ghash_4bit
push %rbx
push %rbp
push %r12
+ push %r13
+ push %r14
+ push %r15
+ sub \$280,%rsp
.Lghash_prologue:
-
- mov 8($Xi),$Zlo
- mov ($Xi),$Zhi
- add $inp,$len
- lea .Lrem_4bit(%rip),$rem_4bit
-.align 4
-.Louter_loop:
- xor 8($inp),$Zlo
- xor ($inp),$Zhi
- lea 16($inp),$inp
- mov $Zlo,8($Xi)
- mov $Zhi,($Xi)
- shr \$56,$Zlo
+ mov $inp,%r14 # reassign couple of args
+ mov $len,%r15
___
- &loop ($Xi);
-$code.=<<___;
- cmp $len,$inp
- jb .Louter_loop
+{ my $inp="%r14";
+ my $dat="%edx";
+ my $len="%r15";
+ my @nhi=("%ebx","%ecx");
+ my @rem=("%r12","%r13");
+ my $Hshr4="%rbp";
+
+ &sub ($Htbl,-128); # size optimization
+ &lea ($Hshr4,"16+128(%rsp)");
+ { my @lo =($nlo,$nhi);
+ my @hi =($Zlo,$Zhi);
+
+ &xor ($dat,$dat);
+ for ($i=0,$j=-2;$i<18;$i++,$j++) {
+ &mov ("$j(%rsp)",&LB($dat)) if ($i>1);
+ &or ($lo[0],$tmp) if ($i>1);
+ &mov (&LB($dat),&LB($lo[1])) if ($i>0 && $i<17);
+ &shr ($lo[1],4) if ($i>0 && $i<17);
+ &mov ($tmp,$hi[1]) if ($i>0 && $i<17);
+ &shr ($hi[1],4) if ($i>0 && $i<17);
+ &mov ("8*$j($Hshr4)",$hi[0]) if ($i>1);
+ &mov ($hi[0],"16*$i+0-128($Htbl)") if ($i<16);
+ &shl (&LB($dat),4) if ($i>0 && $i<17);
+ &mov ("8*$j-128($Hshr4)",$lo[0]) if ($i>1);
+ &mov ($lo[0],"16*$i+8-128($Htbl)") if ($i<16);
+ &shl ($tmp,60) if ($i>0 && $i<17);
+
+ push (@lo,shift(@lo));
+ push (@hi,shift(@hi));
+ }
+ }
+ &add ($Htbl,-128);
+ &mov ($Zlo,"8($Xi)");
+ &mov ($Zhi,"0($Xi)");
+ &add ($len,$inp); # pointer to the end of data
+ &lea ($rem_8bit,".Lrem_8bit(%rip)");
+ &jmp (".Louter_loop");
+
+$code.=".align 16\n.Louter_loop:\n";
+ &xor ($Zhi,"($inp)");
+ &mov ("%rdx","8($inp)");
+ &lea ($inp,"16($inp)");
+ &xor ("%rdx",$Zlo);
+ &mov ("($Xi)",$Zhi);
+ &mov ("8($Xi)","%rdx");
+ &shr ("%rdx",32);
+
+ &xor ($nlo,$nlo);
+ &rol ($dat,8);
+ &mov (&LB($nlo),&LB($dat));
+ &movz ($nhi[0],&LB($dat));
+ &shl (&LB($nlo),4);
+ &shr ($nhi[0],4);
+
+ for ($j=11,$i=0;$i<15;$i++) {
+ &rol ($dat,8);
+ &xor ($Zlo,"8($Htbl,$nlo)") if ($i>0);
+ &xor ($Zhi,"($Htbl,$nlo)") if ($i>0);
+ &mov ($Zlo,"8($Htbl,$nlo)") if ($i==0);
+ &mov ($Zhi,"($Htbl,$nlo)") if ($i==0);
+
+ &mov (&LB($nlo),&LB($dat));
+ &xor ($Zlo,$tmp) if ($i>0);
+ &movzw ($rem[1],"($rem_8bit,$rem[1],2)") if ($i>0);
+
+ &movz ($nhi[1],&LB($dat));
+ &shl (&LB($nlo),4);
+ &movzb ($rem[0],"(%rsp,$nhi[0])");
+
+ &shr ($nhi[1],4) if ($i<14);
+ &and ($nhi[1],0xf0) if ($i==14);
+ &shl ($rem[1],48) if ($i>0);
+ &xor ($rem[0],$Zlo);
+
+ &mov ($tmp,$Zhi);
+ &xor ($Zhi,$rem[1]) if ($i>0);
+ &shr ($Zlo,8);
+
+ &movz ($rem[0],&LB($rem[0]));
+ &mov ($dat,"$j($Xi)") if (--$j%4==0);
+ &shr ($Zhi,8);
+
+ &xor ($Zlo,"-128($Hshr4,$nhi[0],8)");
+ &shl ($tmp,56);
+ &xor ($Zhi,"($Hshr4,$nhi[0],8)");
+
+ unshift (@nhi,pop(@nhi)); # "rotate" registers
+ unshift (@rem,pop(@rem));
+ }
+ &movzw ($rem[1],"($rem_8bit,$rem[1],2)");
+ &xor ($Zlo,"8($Htbl,$nlo)");
+ &xor ($Zhi,"($Htbl,$nlo)");
+
+ &shl ($rem[1],48);
+ &xor ($Zlo,$tmp);
+
+ &xor ($Zhi,$rem[1]);
+ &movz ($rem[0],&LB($Zlo));
+ &shr ($Zlo,4);
+ &mov ($tmp,$Zhi);
+ &shl (&LB($rem[0]),4);
+ &shr ($Zhi,4);
+
+ &xor ($Zlo,"8($Htbl,$nhi[0])");
+ &movzw ($rem[0],"($rem_8bit,$rem[0],2)");
+ &shl ($tmp,60);
+
+ &xor ($Zhi,"($Htbl,$nhi[0])");
+ &xor ($Zlo,$tmp);
+ &shl ($rem[0],48);
+
+ &bswap ($Zlo);
+ &xor ($Zhi,$rem[0]);
+
+ &bswap ($Zhi);
+ &cmp ($inp,$len);
+ &jb (".Louter_loop");
+}
+$code.=<<___;
mov $Zlo,8($Xi)
mov $Zhi,($Xi)
- mov 0(%rsp),%r12
- mov 8(%rsp),%rbp
- mov 16(%rsp),%rbx
- lea 24(%rsp),%rsp
+ lea 280(%rsp),%rsi
+ mov 0(%rsi),%r15
+ mov 8(%rsi),%r14
+ mov 16(%rsi),%r13
+ mov 24(%rsi),%r12
+ mov 32(%rsi),%rbp
+ mov 40(%rsi),%rbx
+ lea 48(%rsi),%rsp
.Lghash_epilogue:
ret
.size gcm_ghash_4bit,.-gcm_ghash_4bit
+___
+\f
+######################################################################
+# PCLMULQDQ version.
+
+@_4args=$win64? ("%rcx","%rdx","%r8", "%r9") : # Win64 order
+ ("%rdi","%rsi","%rdx","%rcx"); # Unix order
+
+($Xi,$Xhi)=("%xmm0","%xmm1"); $Hkey="%xmm2";
+($T1,$T2,$T3)=("%xmm3","%xmm4","%xmm5");
+
+sub clmul64x64_T2 { # minimal register pressure
+my ($Xhi,$Xi,$Hkey,$HK)=@_;
+
+if (!defined($HK)) { $HK = $T2;
+$code.=<<___;
+ movdqa $Xi,$Xhi #
+ pshufd \$0b01001110,$Xi,$T1
+ pshufd \$0b01001110,$Hkey,$T2
+ pxor $Xi,$T1 #
+ pxor $Hkey,$T2
+___
+} else {
+$code.=<<___;
+ movdqa $Xi,$Xhi #
+ pshufd \$0b01001110,$Xi,$T1
+ pxor $Xi,$T1 #
+___
+}
+$code.=<<___;
+ pclmulqdq \$0x00,$Hkey,$Xi #######
+ pclmulqdq \$0x11,$Hkey,$Xhi #######
+ pclmulqdq \$0x00,$HK,$T1 #######
+ pxor $Xi,$T1 #
+ pxor $Xhi,$T1 #
+
+ movdqa $T1,$T2 #
+ psrldq \$8,$T1
+ pslldq \$8,$T2 #
+ pxor $T1,$Xhi
+ pxor $T2,$Xi #
+___
+}
+
+sub reduction_alg9 { # 17/11 times faster than Intel version
+my ($Xhi,$Xi) = @_;
+
+$code.=<<___;
+ # 1st phase
+ movdqa $Xi,$T2 #
+ movdqa $Xi,$T1
+ psllq \$5,$Xi
+ pxor $Xi,$T1 #
+ psllq \$1,$Xi
+ pxor $T1,$Xi #
+ psllq \$57,$Xi #
+ movdqa $Xi,$T1 #
+ pslldq \$8,$Xi
+ psrldq \$8,$T1 #
+ pxor $T2,$Xi
+ pxor $T1,$Xhi #
+
+ # 2nd phase
+ movdqa $Xi,$T2
+ psrlq \$1,$Xi
+ pxor $T2,$Xhi #
+ pxor $Xi,$T2
+ psrlq \$5,$Xi
+ pxor $T2,$Xi #
+ psrlq \$1,$Xi #
+ pxor $Xhi,$Xi #
+___
+}
+\f
+{ my ($Htbl,$Xip)=@_4args;
+
+$code.=<<___;
+.globl gcm_init_clmul
+.type gcm_init_clmul,\@abi-omnipotent
+.align 16
+gcm_init_clmul:
+ movdqu ($Xip),$Hkey
+ pshufd \$0b01001110,$Hkey,$Hkey # dword swap
+
+ # <<1 twist
+ pshufd \$0b11111111,$Hkey,$T2 # broadcast uppermost dword
+ movdqa $Hkey,$T1
+ psllq \$1,$Hkey
+ pxor $T3,$T3 #
+ psrlq \$63,$T1
+ pcmpgtd $T2,$T3 # broadcast carry bit
+ pslldq \$8,$T1
+ por $T1,$Hkey # H<<=1
+
+ # magic reduction
+ pand .L0x1c2_polynomial(%rip),$T3
+ pxor $T3,$Hkey # if(carry) H^=0x1c2_polynomial
+
+ # calculate H^2
+ movdqa $Hkey,$Xi
+___
+ &clmul64x64_T2 ($Xhi,$Xi,$Hkey);
+ &reduction_alg9 ($Xhi,$Xi);
+$code.=<<___;
+ pshufd \$0b01001110,$Hkey,$T1
+ pshufd \$0b01001110,$Xi,$T2
+ pxor $Hkey,$T1 # Karatsuba pre-processing
+ movdqu $Hkey,0x00($Htbl) # save H
+ pxor $Xi,$T2 # Karatsuba pre-processing
+ movdqu $Xi,0x10($Htbl) # save H^2
+ palignr \$8,$T1,$T2 # low part is H.lo^H.hi...
+ movdqu $T2,0x20($Htbl) # save Karatsuba "salt"
+___
+if ($do4xaggr) {
+ &clmul64x64_T2 ($Xhi,$Xi,$Hkey); # H^3
+ &reduction_alg9 ($Xhi,$Xi);
+$code.=<<___;
+ movdqa $Xi,$T3
+___
+ &clmul64x64_T2 ($Xhi,$Xi,$Hkey); # H^4
+ &reduction_alg9 ($Xhi,$Xi);
+$code.=<<___;
+ pshufd \$0b01001110,$T3,$T1
+ pshufd \$0b01001110,$Xi,$T2
+ pxor $T3,$T1 # Karatsuba pre-processing
+ movdqu $T3,0x30($Htbl) # save H^3
+ pxor $Xi,$T2 # Karatsuba pre-processing
+ movdqu $Xi,0x40($Htbl) # save H^4
+ palignr \$8,$T1,$T2 # low part is H.lo^H.hi...
+ movdqu $T2,0x50($Htbl) # save Karatsuba "salt"
+___
+}
+$code.=<<___;
+ ret
+.size gcm_init_clmul,.-gcm_init_clmul
+___
+}
+
+{ my ($Xip,$Htbl)=@_4args;
+
+$code.=<<___;
+.globl gcm_gmult_clmul
+.type gcm_gmult_clmul,\@abi-omnipotent
+.align 16
+gcm_gmult_clmul:
+ movdqu ($Xip),$Xi
+ movdqa .Lbswap_mask(%rip),$T3
+ movdqu ($Htbl),$Hkey
+ movdqu 0x20($Htbl),$T2
+ pshufb $T3,$Xi
+___
+ &clmul64x64_T2 ($Xhi,$Xi,$Hkey,$T2);
+$code.=<<___ if (0 || (&reduction_alg9($Xhi,$Xi)&&0));
+ # experimental alternative. special thing about is that there
+ # no dependency between the two multiplications...
+ mov \$`0xE1<<1`,%eax
+ mov \$0xA040608020C0E000,%r10 # ((7..0)·0xE0)&0xff
+ mov \$0x07,%r11d
+ movq %rax,$T1
+ movq %r10,$T2
+ movq %r11,$T3 # borrow $T3
+ pand $Xi,$T3
+ pshufb $T3,$T2 # ($Xi&7)·0xE0
+ movq %rax,$T3
+ pclmulqdq \$0x00,$Xi,$T1 # ·(0xE1<<1)
+ pxor $Xi,$T2
+ pslldq \$15,$T2
+ paddd $T2,$T2 # <<(64+56+1)
+ pxor $T2,$Xi
+ pclmulqdq \$0x01,$T3,$Xi
+ movdqa .Lbswap_mask(%rip),$T3 # reload $T3
+ psrldq \$1,$T1
+ pxor $T1,$Xhi
+ pslldq \$7,$Xi
+ pxor $Xhi,$Xi
+___
+$code.=<<___;
+ pshufb $T3,$Xi
+ movdqu $Xi,($Xip)
+ ret
+.size gcm_gmult_clmul,.-gcm_gmult_clmul
+___
+}
+\f
+{ my ($Xip,$Htbl,$inp,$len)=@_4args;
+ my ($Xln,$Xmn,$Xhn,$Hkey2,$HK) = map("%xmm$_",(6..10));
+
+$code.=<<___;
+.globl gcm_ghash_clmul
+.type gcm_ghash_clmul,\@abi-omnipotent
+.align 32
+gcm_ghash_clmul:
+___
+$code.=<<___ if ($win64);
+ lea -0x88(%rsp),%rax
+.LSEH_begin_gcm_ghash_clmul:
+ # I can't trust assembler to use specific encoding:-(
+ .byte 0x48,0x8d,0x60,0xe0 #lea -0x20(%rax),%rsp
+ .byte 0x0f,0x29,0x70,0xe0 #movaps %xmm6,-0x20(%rax)
+ .byte 0x0f,0x29,0x78,0xf0 #movaps %xmm7,-0x10(%rax)
+ .byte 0x44,0x0f,0x29,0x00 #movaps %xmm8,0(%rax)
+ .byte 0x44,0x0f,0x29,0x48,0x10 #movaps %xmm9,0x10(%rax)
+ .byte 0x44,0x0f,0x29,0x50,0x20 #movaps %xmm10,0x20(%rax)
+ .byte 0x44,0x0f,0x29,0x58,0x30 #movaps %xmm11,0x30(%rax)
+ .byte 0x44,0x0f,0x29,0x60,0x40 #movaps %xmm12,0x40(%rax)
+ .byte 0x44,0x0f,0x29,0x68,0x50 #movaps %xmm13,0x50(%rax)
+ .byte 0x44,0x0f,0x29,0x70,0x60 #movaps %xmm14,0x60(%rax)
+ .byte 0x44,0x0f,0x29,0x78,0x70 #movaps %xmm15,0x70(%rax)
+___
+$code.=<<___;
+ movdqa .Lbswap_mask(%rip),$T3
+ mov \$0xA040608020C0E000,%rax # ((7..0)·0xE0)&0xff
+
+ movdqu ($Xip),$Xi
+ movdqu ($Htbl),$Hkey
+ movdqu 0x20($Htbl),$HK
+ pshufb $T3,$Xi
+
+ sub \$0x10,$len
+ jz .Lodd_tail
+
+ movdqu 0x10($Htbl),$Hkey2
+___
+if ($do4xaggr) {
+my ($Xl,$Xm,$Xh,$Hkey3,$Hkey4)=map("%xmm$_",(11..15));
+
+$code.=<<___;
+ cmp \$0x30,$len
+ jb .Lskip4x
+
+ sub \$0x30,$len
+ movdqu 0x30($Htbl),$Hkey3
+ movdqu 0x40($Htbl),$Hkey4
+
+ #######
+ # Xi+4 =[(H*Ii+3) + (H^2*Ii+2) + (H^3*Ii+1) + H^4*(Ii+Xi)] mod P
+ #
+ movdqu 0x30($inp),$Xln
+ movdqu 0x20($inp),$Xl
+ pshufb $T3,$Xln
+ pshufb $T3,$Xl
+ movdqa $Xln,$Xhn
+ pshufd \$0b01001110,$Xln,$Xmn
+ pxor $Xln,$Xmn
+ pclmulqdq \$0x00,$Hkey,$Xln
+ pclmulqdq \$0x11,$Hkey,$Xhn
+ pclmulqdq \$0x00,$HK,$Xmn
+
+ movdqa $Xl,$Xh
+ pshufd \$0b01001110,$Xl,$Xm
+ pxor $Xl,$Xm
+ pclmulqdq \$0x00,$Hkey2,$Xl
+ pclmulqdq \$0x11,$Hkey2,$Xh
+ xorps $Xl,$Xln
+ pclmulqdq \$0x10,$HK,$Xm
+ xorps $Xh,$Xhn
+ movups 0x50($Htbl),$HK
+ xorps $Xm,$Xmn
+
+ movdqu 0x10($inp),$Xl
+ movdqu 0($inp),$T1
+ pshufb $T3,$Xl
+ pshufb $T3,$T1
+ movdqa $Xl,$Xh
+ pshufd \$0b01001110,$Xl,$Xm
+ pxor $T1,$Xi
+ pxor $Xl,$Xm
+ pclmulqdq \$0x00,$Hkey3,$Xl
+ movdqa $Xi,$Xhi
+ pshufd \$0b01001110,$Xi,$T1
+ pxor $Xi,$T1
+ pclmulqdq \$0x11,$Hkey3,$Xh
+ xorps $Xl,$Xln
+ pclmulqdq \$0x00,$HK,$Xm
+ xorps $Xh,$Xhn
+
+ lea 0x40($inp),$inp
+ sub \$0x40,$len
+ jc .Ltail4x
+
+ jmp .Lmod4_loop
+.align 32
+.Lmod4_loop:
+ pclmulqdq \$0x00,$Hkey4,$Xi
+ xorps $Xm,$Xmn
+ movdqu 0x30($inp),$Xl
+ pshufb $T3,$Xl
+ pclmulqdq \$0x11,$Hkey4,$Xhi
+ xorps $Xln,$Xi
+ movdqu 0x20($inp),$Xln
+ movdqa $Xl,$Xh
+ pshufd \$0b01001110,$Xl,$Xm
+ pclmulqdq \$0x10,$HK,$T1
+ xorps $Xhn,$Xhi
+ pxor $Xl,$Xm
+ pshufb $T3,$Xln
+ movups 0x20($Htbl),$HK
+ pclmulqdq \$0x00,$Hkey,$Xl
+ xorps $Xmn,$T1
+ movdqa $Xln,$Xhn
+ pshufd \$0b01001110,$Xln,$Xmn
+
+ pxor $Xi,$T1 # aggregated Karatsuba post-processing
+ pxor $Xln,$Xmn
+ pxor $Xhi,$T1 #
+ movdqa $T1,$T2 #
+ pslldq \$8,$T1
+ pclmulqdq \$0x11,$Hkey,$Xh
+ psrldq \$8,$T2 #
+ pxor $T1,$Xi
+ movdqa .L7_mask(%rip),$T1
+ pxor $T2,$Xhi #
+ movq %rax,$T2
+
+ pand $Xi,$T1 # 1st phase
+ pshufb $T1,$T2 #
+ pclmulqdq \$0x00,$HK,$Xm
+ pxor $Xi,$T2 #
+ psllq \$57,$T2 #
+ movdqa $T2,$T1 #
+ pslldq \$8,$T2
+ pclmulqdq \$0x00,$Hkey2,$Xln
+ psrldq \$8,$T1 #
+ pxor $T2,$Xi
+ pxor $T1,$Xhi #
+ movdqu 0($inp),$T1
+
+ movdqa $Xi,$T2 # 2nd phase
+ psrlq \$1,$Xi
+ pclmulqdq \$0x11,$Hkey2,$Xhn
+ xorps $Xl,$Xln
+ movdqu 0x10($inp),$Xl
+ pshufb $T3,$Xl
+ pclmulqdq \$0x10,$HK,$Xmn
+ xorps $Xh,$Xhn
+ movups 0x50($Htbl),$HK
+ pshufb $T3,$T1
+ pxor $T2,$Xhi #
+ pxor $Xi,$T2
+ psrlq \$5,$Xi
+
+ movdqa $Xl,$Xh
+ pxor $Xm,$Xmn
+ pshufd \$0b01001110,$Xl,$Xm
+ pxor $Xl,$Xm
+ pclmulqdq \$0x00,$Hkey3,$Xl
+ pxor $T2,$Xi #
+ pxor $T1,$Xhi
+ psrlq \$1,$Xi #
+ pclmulqdq \$0x11,$Hkey3,$Xh
+ xorps $Xl,$Xln
+ pxor $Xhi,$Xi #
+
+ pclmulqdq \$0x00,$HK,$Xm
+ xorps $Xh,$Xhn
+ movdqa $Xi,$Xhi
+ pshufd \$0b01001110,$Xi,$T1
+ pxor $Xi,$T1
+
+ lea 0x40($inp),$inp
+ sub \$0x40,$len
+ jnc .Lmod4_loop
+
+.Ltail4x:
+ pclmulqdq \$0x00,$Hkey4,$Xi
+ xorps $Xm,$Xmn
+ pclmulqdq \$0x11,$Hkey4,$Xhi
+ xorps $Xln,$Xi
+ pclmulqdq \$0x10,$HK,$T1
+ xorps $Xhn,$Xhi
+ pxor $Xi,$Xhi # aggregated Karatsuba post-processing
+ pxor $Xmn,$T1
+
+ pxor $Xhi,$T1 #
+ pxor $Xi,$Xhi
+
+ movdqa $T1,$T2 #
+ psrldq \$8,$T1
+ pslldq \$8,$T2 #
+ pxor $T1,$Xhi
+ pxor $T2,$Xi #
+___
+ &reduction_alg9($Xhi,$Xi);
+$code.=<<___;
+ add \$0x40,$len
+ jz .Ldone
+ movdqu 0x20($Htbl),$HK
+ sub \$0x10,$len
+ jz .Lodd_tail
+.Lskip4x:
+___
+}
+$code.=<<___;
+ #######
+ # Xi+2 =[H*(Ii+1 + Xi+1)] mod P =
+ # [(H*Ii+1) + (H*Xi+1)] mod P =
+ # [(H*Ii+1) + H^2*(Ii+Xi)] mod P
+ #
+ movdqu ($inp),$T1 # Ii
+ movdqu 16($inp),$Xln # Ii+1
+ pshufb $T3,$T1
+ pshufb $T3,$Xln
+ pxor $T1,$Xi # Ii+Xi
+
+ movdqa $Xln,$Xhn
+ pshufd \$0b01001110,$Xln,$Xmn
+ pxor $Xln,$Xmn
+ pclmulqdq \$0x00,$Hkey,$Xln
+ pclmulqdq \$0x11,$Hkey,$Xhn
+ pclmulqdq \$0x00,$HK,$Xmn
+
+ movdqa $Xi,$Xhi
+ pshufd \$0b01001110,$Xi,$T1 #
+ pxor $Xi,$T1 #
+
+ lea 32($inp),$inp # i+=2
+ sub \$0x20,$len
+ jbe .Leven_tail
+ jmp .Lmod_loop
+
+.align 32
+.Lmod_loop:
+ pclmulqdq \$0x00,$Hkey2,$Xi
+ pclmulqdq \$0x11,$Hkey2,$Xhi
+ movdqu ($inp),$T2 # Ii
+ pclmulqdq \$0x10,$HK,$T1
+ pshufb $T3,$T2
+
+ pxor $Xln,$Xi # (H*Ii+1) + H^2*(Ii+Xi)
+ movdqu 16($inp),$Xln # Ii+1
+ pxor $Xhn,$Xhi
+
+ pxor $Xi,$Xmn # aggregated Karatsuba post-processing
+ pxor $Xhi,$Xmn
+ pxor $T2,$Xhi # "Ii+Xi", consume early
+ pxor $Xmn,$T1
+ pshufb $T3,$Xln
+ movdqa $T1,$T2 #
+ psrldq \$8,$T1
+ pslldq \$8,$T2 #
+ pxor $T1,$Xhi
+ pxor $T2,$Xi #
+
+ movdqa $Xln,$Xhn #
+ pshufd \$0b01001110,$Xln,$Xmn
+ pxor $Xln,$Xmn #
+
+ movdqa $Xi,$T2 # 1st phase
+ movdqa $Xi,$T1
+ psllq \$5,$Xi
+ pclmulqdq \$0x00,$Hkey,$Xln #######
+ pxor $Xi,$T1 #
+ psllq \$1,$Xi
+ pxor $T1,$Xi #
+ psllq \$57,$Xi #
+ movdqa $Xi,$T1 #
+ pslldq \$8,$Xi
+ psrldq \$8,$T1 #
+ pxor $T2,$Xi
+ pxor $T1,$Xhi #
+
+ pclmulqdq \$0x11,$Hkey,$Xhn #######
+ movdqa $Xi,$T2 # 2nd phase
+ psrlq \$1,$Xi
+ pxor $T2,$Xhi #
+ pxor $Xi,$T2
+ psrlq \$5,$Xi
+ pxor $T2,$Xi #
+ psrlq \$1,$Xi #
+ pclmulqdq \$0x00,$HK,$Xmn #######
+ pxor $Xhi,$Xi #
+
+ movdqa $Xi,$Xhi
+ pshufd \$0b01001110,$Xi,$T1 #
+ pxor $Xi,$T1 #
+
+ lea 32($inp),$inp
+ sub \$0x20,$len
+ ja .Lmod_loop
+
+.Leven_tail:
+ pclmulqdq \$0x00,$Hkey2,$Xi
+ pclmulqdq \$0x11,$Hkey2,$Xhi
+ pclmulqdq \$0x10,$HK,$T1
+
+ pxor $Xln,$Xi # (H*Ii+1) + H^2*(Ii+Xi)
+ pxor $Xhn,$Xhi
+ pxor $Xi,$Xmn
+ pxor $Xhi,$Xmn
+ pxor $Xmn,$T1
+ movdqa $T1,$T2 #
+ psrldq \$8,$T1
+ pslldq \$8,$T2 #
+ pxor $T1,$Xhi
+ pxor $T2,$Xi #
+___
+ &reduction_alg9 ($Xhi,$Xi);
+$code.=<<___;
+ test $len,$len
+ jnz .Ldone
+
+.Lodd_tail:
+ movdqu ($inp),$T1 # Ii
+ pshufb $T3,$T1
+ pxor $T1,$Xi # Ii+Xi
+___
+ &clmul64x64_T2 ($Xhi,$Xi,$Hkey,$HK); # H*(Ii+Xi)
+ &reduction_alg9 ($Xhi,$Xi);
+$code.=<<___;
+.Ldone:
+ pshufb $T3,$Xi
+ movdqu $Xi,($Xip)
+___
+$code.=<<___ if ($win64);
+ movaps (%rsp),%xmm6
+ movaps 0x10(%rsp),%xmm7
+ movaps 0x20(%rsp),%xmm8
+ movaps 0x30(%rsp),%xmm9
+ movaps 0x40(%rsp),%xmm10
+ movaps 0x50(%rsp),%xmm11
+ movaps 0x60(%rsp),%xmm12
+ movaps 0x70(%rsp),%xmm13
+ movaps 0x80(%rsp),%xmm14
+ movaps 0x90(%rsp),%xmm15
+ lea 0xa8(%rsp),%rsp
+___
+$code.=<<___;
+ ret
+.LSEH_end_gcm_ghash_clmul:
+.size gcm_ghash_clmul,.-gcm_ghash_clmul
+___
+}
+
+$code.=<<___;
.align 64
-.type rem_4bit,\@object
+.Lbswap_mask:
+ .byte 15,14,13,12,11,10,9,8,7,6,5,4,3,2,1,0
+.L0x1c2_polynomial:
+ .byte 1,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0xc2
+.L7_mask:
+ .long 7,0,7,0
+.L7_mask_poly:
+ .long 7,0,`0xE1<<1`,0
+.align 64
+.type .Lrem_4bit,\@object
.Lrem_4bit:
.long 0,`0x0000<<16`,0,`0x1C20<<16`,0,`0x3840<<16`,0,`0x2460<<16`
.long 0,`0x7080<<16`,0,`0x6CA0<<16`,0,`0x48C0<<16`,0,`0x54E0<<16`
.long 0,`0xE100<<16`,0,`0xFD20<<16`,0,`0xD940<<16`,0,`0xC560<<16`
.long 0,`0x9180<<16`,0,`0x8DA0<<16`,0,`0xA9C0<<16`,0,`0xB5E0<<16`
+.type .Lrem_8bit,\@object
+.Lrem_8bit:
+ .value 0x0000,0x01C2,0x0384,0x0246,0x0708,0x06CA,0x048C,0x054E
+ .value 0x0E10,0x0FD2,0x0D94,0x0C56,0x0918,0x08DA,0x0A9C,0x0B5E
+ .value 0x1C20,0x1DE2,0x1FA4,0x1E66,0x1B28,0x1AEA,0x18AC,0x196E
+ .value 0x1230,0x13F2,0x11B4,0x1076,0x1538,0x14FA,0x16BC,0x177E
+ .value 0x3840,0x3982,0x3BC4,0x3A06,0x3F48,0x3E8A,0x3CCC,0x3D0E
+ .value 0x3650,0x3792,0x35D4,0x3416,0x3158,0x309A,0x32DC,0x331E
+ .value 0x2460,0x25A2,0x27E4,0x2626,0x2368,0x22AA,0x20EC,0x212E
+ .value 0x2A70,0x2BB2,0x29F4,0x2836,0x2D78,0x2CBA,0x2EFC,0x2F3E
+ .value 0x7080,0x7142,0x7304,0x72C6,0x7788,0x764A,0x740C,0x75CE
+ .value 0x7E90,0x7F52,0x7D14,0x7CD6,0x7998,0x785A,0x7A1C,0x7BDE
+ .value 0x6CA0,0x6D62,0x6F24,0x6EE6,0x6BA8,0x6A6A,0x682C,0x69EE
+ .value 0x62B0,0x6372,0x6134,0x60F6,0x65B8,0x647A,0x663C,0x67FE
+ .value 0x48C0,0x4902,0x4B44,0x4A86,0x4FC8,0x4E0A,0x4C4C,0x4D8E
+ .value 0x46D0,0x4712,0x4554,0x4496,0x41D8,0x401A,0x425C,0x439E
+ .value 0x54E0,0x5522,0x5764,0x56A6,0x53E8,0x522A,0x506C,0x51AE
+ .value 0x5AF0,0x5B32,0x5974,0x58B6,0x5DF8,0x5C3A,0x5E7C,0x5FBE
+ .value 0xE100,0xE0C2,0xE284,0xE346,0xE608,0xE7CA,0xE58C,0xE44E
+ .value 0xEF10,0xEED2,0xEC94,0xED56,0xE818,0xE9DA,0xEB9C,0xEA5E
+ .value 0xFD20,0xFCE2,0xFEA4,0xFF66,0xFA28,0xFBEA,0xF9AC,0xF86E
+ .value 0xF330,0xF2F2,0xF0B4,0xF176,0xF438,0xF5FA,0xF7BC,0xF67E
+ .value 0xD940,0xD882,0xDAC4,0xDB06,0xDE48,0xDF8A,0xDDCC,0xDC0E
+ .value 0xD750,0xD692,0xD4D4,0xD516,0xD058,0xD19A,0xD3DC,0xD21E
+ .value 0xC560,0xC4A2,0xC6E4,0xC726,0xC268,0xC3AA,0xC1EC,0xC02E
+ .value 0xCB70,0xCAB2,0xC8F4,0xC936,0xCC78,0xCDBA,0xCFFC,0xCE3E
+ .value 0x9180,0x9042,0x9204,0x93C6,0x9688,0x974A,0x950C,0x94CE
+ .value 0x9F90,0x9E52,0x9C14,0x9DD6,0x9898,0x995A,0x9B1C,0x9ADE
+ .value 0x8DA0,0x8C62,0x8E24,0x8FE6,0x8AA8,0x8B6A,0x892C,0x88EE
+ .value 0x83B0,0x8272,0x8034,0x81F6,0x84B8,0x857A,0x873C,0x86FE
+ .value 0xA9C0,0xA802,0xAA44,0xAB86,0xAEC8,0xAF0A,0xAD4C,0xAC8E
+ .value 0xA7D0,0xA612,0xA454,0xA596,0xA0D8,0xA11A,0xA35C,0xA29E
+ .value 0xB5E0,0xB422,0xB664,0xB7A6,0xB2E8,0xB32A,0xB16C,0xB0AE
+ .value 0xBBF0,0xBA32,0xB874,0xB9B6,0xBCF8,0xBD3A,0xBF7C,0xBEBE
+
.asciz "GHASH for x86_64, CRYPTOGAMS by <appro\@openssl.org>"
.align 64
___
-
+\f
# EXCEPTION_DISPOSITION handler (EXCEPTION_RECORD *rec,ULONG64 frame,
# CONTEXT *context,DISPATCHER_CONTEXT *disp)
if ($win64) {
.rva .LSEH_end_gcm_ghash_4bit
.rva .LSEH_info_gcm_ghash_4bit
+ .rva .LSEH_begin_gcm_ghash_clmul
+ .rva .LSEH_end_gcm_ghash_clmul
+ .rva .LSEH_info_gcm_ghash_clmul
+
.section .xdata
.align 8
.LSEH_info_gcm_gmult_4bit:
.byte 9,0,0,0
.rva se_handler
.rva .Lghash_prologue,.Lghash_epilogue # HandlerData
+.LSEH_info_gcm_ghash_clmul:
+ .byte 0x01,0x33,0x16,0x00
+ .byte 0x33,0xf8,0x09,0x00 #movaps 0x90(rsp),xmm15
+ .byte 0x2e,0xe8,0x08,0x00 #movaps 0x80(rsp),xmm14
+ .byte 0x29,0xd8,0x07,0x00 #movaps 0x70(rsp),xmm13
+ .byte 0x24,0xc8,0x06,0x00 #movaps 0x60(rsp),xmm12
+ .byte 0x1f,0xb8,0x05,0x00 #movaps 0x50(rsp),xmm11
+ .byte 0x1a,0xa8,0x04,0x00 #movaps 0x40(rsp),xmm10
+ .byte 0x15,0x98,0x03,0x00 #movaps 0x30(rsp),xmm9
+ .byte 0x10,0x88,0x02,0x00 #movaps 0x20(rsp),xmm8
+ .byte 0x0c,0x78,0x01,0x00 #movaps 0x10(rsp),xmm7
+ .byte 0x08,0x68,0x00,0x00 #movaps 0x00(rsp),xmm6
+ .byte 0x04,0x01,0x15,0x00 #sub 0xa8,rsp
___
}
+\f
$code =~ s/\`([^\`]*)\`/eval($1)/gem;
print $code;