SHA256/512 for x86.

#!/usr/bin/env perl

#

# ====================================================================

# Written by Andy Polyakov <appro@fy.chalmers.se> for the OpenSSL

# project. The module is, however, dual licensed under OpenSSL and

# CRYPTOGAMS licenses depending on where you obtain it. For further

# details see http://www.openssl.org/~appro/cryptogams/.

# ====================================================================

#

# SHA256 block transform for x86. September 2007.

#

# Performance in clock cycles per processed byte (less is better):

#

#		Pentium	PIII	P4	AMD K8	Core2

# gcc		46	36	41	27	26

# icc		57	33	38	25	23	

# x86 asm	40	30	35	20	20

# x86_64 asm(*)	-	-	21	15.8	16.5

#

# (*) x86_64 assembler performance is presented for reference

#     purposes.

#

# Performance improvement over compiler generated code varies from

# 10% to 40% [see above]. Not very impressive on some -archs, but

# it's 5 times smaller and optimizies amount of writes.

$0 =~ m/(.*[\/\\])[^\/\\]+$/; $dir=$1;

push(@INC,"${dir}","${dir}../../perlasm");

require "x86asm.pl";

&asm_init($ARGV[0],"sha512-586.pl",$ARGV[$#ARGV] eq "386");

$A="eax";

$E="edx";

$T="ebx";

$Aoff=&DWP(0,"esp");

$Boff=&DWP(4,"esp");

$Coff=&DWP(8,"esp");

$Doff=&DWP(12,"esp");

$Eoff=&DWP(16,"esp");

$Foff=&DWP(20,"esp");

$Goff=&DWP(24,"esp");

$Hoff=&DWP(28,"esp");

$Xoff=&DWP(32,"esp");

$K256="ebp";

sub BODY_00_15() {

	&mov	("ecx",$E);

	&ror	("ecx",6);

	&mov	("edi",$E);

	&ror	("edi",11);

	 &mov	("esi",$Foff);

	&xor	("ecx","edi");

	&ror	("edi",25-11);

	&xor	("ecx","edi");	# Sigma1(e)

	 &mov	("edi",$Goff);

	&add	($T,"ecx");	# T += Sigma1(e)

	 &mov	($Eoff,$E);	# modulo-scheduled

	&xor	("esi","edi");

	 &mov	("ecx",$A);

	&and	("esi",$E);

	 &mov	($E,$Doff);	# e becomes d, which is e in next iteration

	&xor	("esi","edi");	# Ch(e,f,g)

	 &mov	("edi",$A);

	&add	($T,"esi");	# T += Ch(e,f,g)

	&ror	("ecx",2);

	 &add	($T,$Hoff);	# T += h

	&ror	("edi",13);

	 &mov	("esi",$Boff);

	&xor	("ecx","edi");

	&ror	("edi",22-13);

	 &add	($E,$T);	# d += T

	&xor	("ecx","edi");	# Sigma0(a)

	 &mov	("edi",$Coff);

	&add	($T,"ecx");	# T += Sigma0(a)

	 &mov	($Aoff,$A);	# modulo-scheduled

	&mov	("ecx",$A);

	 &sub	("esp",4);

	&or	($A,"esi");	# a becomes h, which is a in next iteration

	&and	("ecx","esi");

	&and	($A,"edi");

	 &mov	("esi",&DWP(0,$K256));

	&or	($A,"ecx");	# h=Maj(a,b,c)

	&add	($K256,4);

	&add	($A,$T);	# h += T

	&add	($E,"esi");	# d += K256[i]

	&add	($A,"esi");	# h += K256[i]

}

&function_begin("sha256_block_data_order",16);

	&mov	("esi",wparam(0));	# ctx

	&mov	("edi",wparam(1));	# inp

	&mov	("eax",wparam(2));	# num

	&mov	("ebx","esp");		# saved sp

	&call	(&label("pic_point"));	# make it PIC!

&set_label("pic_point");

	&blindpop($K256);

	&lea	($K256,&DWP(&label("K256")."-".&label("pic_point"),$K256));

	&sub	("esp",16);

	&and	("esp",-64);

	&shl	("eax",6);

	&add	("eax","edi");

	&mov	(&DWP(0,"esp"),"esi");	# ctx

	&mov	(&DWP(4,"esp"),"edi");	# inp

	&mov	(&DWP(8,"esp"),"eax");	# inp+num*128

	&mov	(&DWP(12,"esp"),"ebx");	# saved sp

&set_label("loop",16);

    # copy input block to stack reversing byte and dword order

    for($i=0;$i<4;$i++) {

	&mov	("eax",&DWP($i*16+0,"edi"));

	&mov	("ebx",&DWP($i*16+4,"edi"));

	&mov	("ecx",&DWP($i*16+8,"edi"));

	&mov	("edx",&DWP($i*16+12,"edi"));

	&bswap	("eax");

	&bswap	("ebx");

	&bswap	("ecx");

	&bswap	("edx");

	&push	("eax");

	&push	("ebx");

	&push	("ecx");

	&push	("edx");

    }

	&add	("edi",64);

	&sub	("esp",4*8);		# place for A,B,C,D,E,F,G,H

	&mov	(&DWP(4*(8+16)+4,"esp"),"edi");

	# copy ctx->h[0-7] to A,B,C,D,E,F,G,H on stack

	&mov	($A,&DWP(0,"esi"));

	&mov	("ebx",&DWP(4,"esi"));

	&mov	("ecx",&DWP(8,"esi"));

	&mov	("edi",&DWP(12,"esi"));

	# &mov	($Aoff,$A);

	&mov	($Boff,"ebx");

	&mov	($Coff,"ecx");

	&mov	($Doff,"edi");

	&mov	($E,&DWP(16,"esi"));	

	&mov	("ebx",&DWP(20,"esi"));

	&mov	("ecx",&DWP(24,"esi"));

	&mov	("edi",&DWP(28,"esi"));

	# &mov	($Eoff,$E);

	&mov	($Foff,"ebx");

	&mov	($Goff,"ecx");

	&mov	($Hoff,"edi");

&set_label("00_15",16);

	&mov	($T,&DWP(4*(8+15),"esp"));

	&BODY_00_15();

	&cmp	("esi",0xc19bf174);

	&jne	(&label("00_15"));

&set_label("16_63",16);

	&mov	($T,&DWP(4*(8+15+16-1),"esp"));

	 &mov	("ecx",&DWP(4*(8+15+16-14),"esp"));

	&mov	("esi",$T);

	&shr	($T,3);

	&ror	("esi",7);

	&xor	($T,"esi");

	&ror	("esi",18-7);

	 &mov	("edi","ecx");

	&xor	($T,"esi");			# T = sigma0(X[-15])

	&shr	("ecx",10);

	 &mov	("esi",&DWP(4*(8+15+16),"esp"));

	&ror	("edi",17);

	&xor	("ecx","edi");

	&ror	("edi",19-17);

	 &add	($T,"esi");			# T += X[-16]

	&xor	("ecx","edi")			# sigma1(X[-2])

	&add	($T,"ecx");			# T += sigma1(X[-2])

	&add	($T,&DWP(4*(8+15+16-9),"esp"));	# T += X[-7]

	&mov	(&DWP(4*(8+15),"esp"),$T);	# save X[0]

	&BODY_00_15();

	&cmp	("esi",0xc67178f2);

	&jne	(&label("16_63"));

	&mov	("esi",&DWP(4*(8+16+64)+0,"esp"));#ctx

	# &mov	($A,$Aoff);

	&mov	("ebx",$Boff);

	&mov	("ecx",$Coff);

	&mov	("edi",$Doff);

	&add	($A,&DWP(0,"esi"));

	&add	("ebx",&DWP(4,"esi"));

	&add	("ecx",&DWP(8,"esi"));

	&add	("edi",&DWP(12,"esi"));

	&mov	(&DWP(0,"esi"),$A);

	&mov	(&DWP(4,"esi"),"ebx");

	&mov	(&DWP(8,"esi"),"ecx");

	&mov	(&DWP(12,"esi"),"edi");

	# &mov	($E,$Eoff);

	&mov	("eax",$Foff);

	&mov	("ebx",$Goff);

	&mov	("ecx",$Hoff);

	&mov	("edi",&DWP(4*(8+16+64)+4,"esp"));#inp

	&add	($E,&DWP(16,"esi"));

	&add	("eax",&DWP(20,"esi"));

	&add	("ebx",&DWP(24,"esi"));

	&add	("ecx",&DWP(28,"esi"));

	&mov	(&DWP(16,"esi"),$E);

	&mov	(&DWP(20,"esi"),"eax");

	&mov	(&DWP(24,"esi"),"ebx");

	&mov	(&DWP(28,"esi"),"ecx");

	&add	("esp",4*(8+16+64));		# destroy frame

	&sub	($K256,4*64);			# rewind K

	&cmp	("edi",&DWP(8,"esp"));		# are we done yet?

	&jb	(&label("loop"));

	&mov	("esp",&DWP(12,"esp"));		# restore sp

&function_end_A();

&set_label("K256",64);	# Yes! I keep it in the code segment!

	&data_word(0x428a2f98,0x71374491,0xb5c0fbcf,0xe9b5dba5);

	&data_word(0x3956c25b,0x59f111f1,0x923f82a4,0xab1c5ed5);

	&data_word(0xd807aa98,0x12835b01,0x243185be,0x550c7dc3);

	&data_word(0x72be5d74,0x80deb1fe,0x9bdc06a7,0xc19bf174);

	&data_word(0xe49b69c1,0xefbe4786,0x0fc19dc6,0x240ca1cc);

	&data_word(0x2de92c6f,0x4a7484aa,0x5cb0a9dc,0x76f988da);

	&data_word(0x983e5152,0xa831c66d,0xb00327c8,0xbf597fc7);

	&data_word(0xc6e00bf3,0xd5a79147,0x06ca6351,0x14292967);

	&data_word(0x27b70a85,0x2e1b2138,0x4d2c6dfc,0x53380d13);

	&data_word(0x650a7354,0x766a0abb,0x81c2c92e,0x92722c85);

	&data_word(0xa2bfe8a1,0xa81a664b,0xc24b8b70,0xc76c51a3);

	&data_word(0xd192e819,0xd6990624,0xf40e3585,0x106aa070);

	&data_word(0x19a4c116,0x1e376c08,0x2748774c,0x34b0bcb5);

	&data_word(0x391c0cb3,0x4ed8aa4a,0x5b9cca4f,0x682e6ff3);

	&data_word(0x748f82ee,0x78a5636f,0x84c87814,0x8cc70208);

	&data_word(0x90befffa,0xa4506ceb,0xbef9a3f7,0xc67178f2);

&function_end_B("sha256_block_data_order");

&asciz("SHA256 block transform for x86, CRYPTOGAMS by <appro\@openssl.org>");

&asm_finish();

#!/usr/bin/env perl

#

# ====================================================================

# Written by Andy Polyakov <appro@fy.chalmers.se> for the OpenSSL

# project. The module is, however, dual licensed under OpenSSL and

# CRYPTOGAMS licenses depending on where you obtain it. For further

# details see http://www.openssl.org/~appro/cryptogams/.

# ====================================================================

#

# SHA512 block transform for x86. September 2007.

#

# Performance in clock cycles per processed byte (less is better):

#

#		Pentium	PIII	P4	AMD K8	Core2

# gcc		100	75	116	54	66

# icc		97	77	95	55	57

# x86 asm	61	56	82	36	40

# SSE2 asm	-	-	38	24	20

# x86_64 asm(*)	-	-	30	10.0	10.5

#

# (*) x86_64 assembler performance is presented for reference

#     purposes.

#

# IALU code-path is optimized for elder Pentiums. On vanilla Pentium

# performance improvement over compiler generated code reaches ~60%,

# while on PIII - ~35%. On newer -archs improvement varies from 15%

# to 50%, but it's less important as they are expected to execute SSE2

# code-path, which is commonly ~2-3x faster [than compiler generated

# code]. SSE2 code-path is as fast as original sha512-sse2.pl, even

# though it does not use 128-bit operations. The latter means that

# SSE2-aware kernel is no longer required to execute the code. Another

# difference is that new code optimizes amount of writes, but at the

# cost of increased data cache "footprint" by 1/2KB.

$0 =~ m/(.*[\/\\])[^\/\\]+$/; $dir=$1;

push(@INC,"${dir}","${dir}../../perlasm");

require "x86asm.pl";

&asm_init($ARGV[0],"sha512-586.pl",$ARGV[$#ARGV] eq "386");

$Tlo=&DWP(0,"esp");	$Thi=&DWP(4,"esp");

$Alo=&DWP(8,"esp");	$Ahi=&DWP(8+4,"esp");

$Blo=&DWP(16,"esp");	$Bhi=&DWP(16+4,"esp");

$Clo=&DWP(24,"esp");	$Chi=&DWP(24+4,"esp");

$Dlo=&DWP(32,"esp");	$Dhi=&DWP(32+4,"esp");

$Elo=&DWP(40,"esp");	$Ehi=&DWP(40+4,"esp");

$Flo=&DWP(48,"esp");	$Fhi=&DWP(48+4,"esp");

$Glo=&DWP(56,"esp");	$Ghi=&DWP(56+4,"esp");

$Hlo=&DWP(64,"esp");	$Hhi=&DWP(64+4,"esp");

$K512="ebp";

$Asse2=&QWP(0,"esp");

$Bsse2=&QWP(8,"esp");

$Csse2=&QWP(16,"esp");

$Dsse2=&QWP(24,"esp");

$Esse2=&QWP(32,"esp");

$Fsse2=&QWP(40,"esp");

$Gsse2=&QWP(48,"esp");

$Hsse2=&QWP(56,"esp");

$A="mm0";	# B-D and

$E="mm4";	# F-H are commonly loaded to respectively mm1-mm3 and

		# mm5-mm7, but it's done on on-demand basis...

sub BODY_00_15_sse2 {

	&movq	("mm5",$Fsse2);			# load f

	&movq	("mm6",$Gsse2);			# load g

	&movq	("mm7",$Hsse2);			# load h

	&movq	("mm1",$E);			# %mm1 is sliding right

	&movq	("mm2",$E);			# %mm2 is sliding left

	&psrlq	("mm1",14);

	&movq	($Esse2,$E);			# module-scheduled save e

	&psllq	("mm2",23);

	&movq	("mm3","mm1");			# %mm3 is T1

	&psrlq	("mm1",4);

	&pxor	("mm3","mm2");

	&psllq	("mm2",23);

	&pxor	("mm3","mm1");

	&psrlq	("mm1",23);

	&pxor	("mm3","mm2");

	&psllq	("mm2",4);

	&pxor	("mm3","mm1");

	&paddq	("mm7",QWP(0,$K512));		# h+=K512[i]

	&pxor	("mm3","mm2");			# T1=Sigma1_512(e)

	&pxor	("mm5","mm6");			# f^=g

	&movq	("mm1",$Bsse2);			# load b

	&pand	("mm5",$E);			# f&=e

	&movq	("mm2",$Csse2);			# load c

	&pxor	("mm5","mm6");			# f^=g

	&movq	($E,$Dsse2);			# e = load d

	&paddq	("mm3","mm5");			# T1+=Ch(e,f,g)

	&paddq	("mm3","mm7");			# T1+=h

	&movq	("mm5",$A);			# %mm5 is sliding right

	&movq	("mm6",$A);			# %mm6 is sliding left

	&paddq	("mm3",&QWP(8*9,"esp"));	# T1+=X[0]

	&psrlq	("mm5",28);

	&paddq	($E,"mm3");			# e += T1

	&psllq	("mm6",25);

	&movq	("mm7","mm5");			# %mm7 is T2

	&psrlq	("mm5",6);

	&pxor	("mm7","mm6");

	&psllq	("mm6",5);

	&pxor	("mm7","mm5");

	&psrlq	("mm5",5);

	&pxor	("mm7","mm6");

	&psllq	("mm6",6);

	&pxor	("mm7","mm5");

	&movq	(&QWP(0,"esp"),$A);		# module-scheduled save a

	&pxor	("mm7","mm6");			# T2=Sigma0_512(a)

	&movq	("mm5",$A);			# %mm5=a

	&por	($A,"mm2");			# a=a|c

	&pand	("mm5","mm2");			# %mm5=a&c

	&pand	($A,"mm1");			# a=(a|c)&b

	&por	("mm5",$A);			# %mm5=(a&c)|((a|c)&b)

	&sub	("esp",8);

	&paddq	("mm7","mm5");			# T2+=Maj(a,b,c)

	&movq	($A,"mm3");			# a=T1

	&mov	(&LB("edx"),&BP(0,$K512));

	&paddq	($A,"mm7");			# a+=T2

	&add	($K512,8);

}

sub BODY_00_15_x86 {

	#define Sigma1(x)	(ROTR((x),14) ^ ROTR((x),18)  ^ ROTR((x),41))

	#	LO		lo>>14^hi<<18 ^ lo>>18^hi<<14 ^ hi>>9^lo<<23

	#	HI		hi>>14^lo<<18 ^ hi>>18^lo<<14 ^ lo>>9^hi<<23

	&mov	("ecx",$Elo);

	&mov	("edx",$Ehi);

	&mov	("esi","ecx");

	&shr	("ecx",9)	# lo>>9

	&mov	("edi","edx");

	&shr	("edx",9)	# hi>>9

	&mov	("ebx","ecx");

	&shl	("esi",14);	# lo<<14

	&mov	("eax","edx");

	&shl	("edi",14);	# hi<<14

	&xor	("ebx","esi");

	&shr	("ecx",14-9);	# lo>>14

	&xor	("eax","edi");

	&shr	("edx",14-9);	# hi>>14

	&xor	("eax","ecx");

	&shl	("esi",18-14);	# lo<<18

	&xor	("ebx","edx");

	&shl	("edi",18-14);	# hi<<18

	&xor	("ebx","esi");

	&shr	("ecx",18-14);	# lo>>18

	&xor	("eax","edi");

	&shr	("edx",18-14);	# hi>>18

	&xor	("eax","ecx");

	&shl	("esi",23-18);	# lo<<23

	&xor	("ebx","edx");

	&shl	("edi",23-18);	# hi<<23

	&xor	("eax","esi");

	&xor	("ebx","edi");			# T1 = Sigma1(e)

	&mov	("ecx",$Flo);

	&mov	("edx",$Fhi);

	&mov	("esi",$Glo);

	&mov	("edi",$Ghi);

	 &add	("eax",$Hlo);

	 &adc	("ebx",$Hhi);			# T1 += h

	&xor	("ecx","esi");

	&xor	("edx","edi");

	&and	("ecx",$Elo);

	&and	("edx",$Ehi);

	 &add	("eax",&DWP(8*(9+15)+0,"esp"));

	 &adc	("ebx",&DWP(8*(9+15)+4,"esp"));	# T1 += X[0]

	&xor	("ecx","esi");

	&xor	("edx","edi");			# Ch(e,f,g) = (f^g)&e)^g

	&mov	("esi",&DWP(0,$K512));

	&mov	("edi",&DWP(4,$K512));		# K[i]

	&add	("eax","ecx");

	&adc	("ebx","edx");			# T1 += Ch(e,f,g)

	&mov	("ecx",$Dlo);

	&mov	("edx",$Dhi);

	&add	("eax","esi");

	&adc	("ebx","edi");			# T1 += K[i]

	&mov	($Tlo,"eax");

	&mov	($Thi,"ebx");			# put T1 away

	&add	("eax","ecx");

	&adc	("ebx","edx");			# d += T1

	#define Sigma0(x)	(ROTR((x),28) ^ ROTR((x),34) ^ ROTR((x),39))

	#	LO		lo>>28^hi<<4  ^ hi>>2^lo<<30 ^ hi>>7^lo<<25

	#	HI		hi>>28^lo<<4  ^ lo>>2^hi<<30 ^ lo>>7^hi<<25

	&mov	("ecx",$Alo);

	&mov	("edx",$Ahi);

	&mov	($Dlo,"eax");

	&mov	($Dhi,"ebx");

	&mov	("esi","ecx");

	&shr	("ecx",2)	# lo>>2

	&mov	("edi","edx");

	&shr	("edx",2)	# hi>>2

	&mov	("ebx","ecx");

	&shl	("esi",4);	# lo<<4

	&mov	("eax","edx");

	&shl	("edi",4);	# hi<<4

	&xor	("ebx","esi");

	&shr	("ecx",7-2);	# lo>>7

	&xor	("eax","edi");

	&shr	("edx",7-2);	# hi>>7

	&xor	("ebx","ecx");

	&shl	("esi",25-4);	# lo<<25

	&xor	("eax","edx");

	&shl	("edi",25-4);	# hi<<25

	&xor	("eax","esi");

	&shr	("ecx",28-7);	# lo>>28

	&xor	("ebx","edi");

	&shr	("edx",28-7);	# hi>>28

	&xor	("eax","ecx");

	&shl	("esi",30-25);	# lo<<30

	&xor	("ebx","edx");

	&shl	("edi",30-25);	# hi<<30

	&xor	("eax","esi");

	&xor	("ebx","edi");			# Sigma0(a)

	&mov	("ecx",$Alo);

	&mov	("edx",$Ahi);

	&mov	("esi",$Blo);

	&mov	("edi",$Bhi);

	&add	("eax",$Tlo);

	&adc	("ebx",$Thi);			# T1 = Sigma0(a)+T1

	&or	("ecx","esi");

	&or	("edx","edi");

	&and	("ecx",$Clo);

	&and	("edx",$Chi);

	&and	("esi",$Alo);

	&and	("edi",$Ahi);

	&or	("ecx","esi");

	&or	("edx","edi");			# Maj(a,b,c) = ((a|b)&c)|(a&b)

	&add	("eax","ecx");

	&adc	("ebx","edx");			# T1 += Maj(a,b,c)

	&mov	($Tlo,"eax");

	&mov	($Thi,"ebx");

	&mov	(&LB("edx"),&BP(0,$K512));	# pre-fetch LSB of *K

	&sub	("esp",8);

	&lea	($K512,&DWP(8,$K512));		# K++

}

&function_begin("sha512_block_data_order",16);

	&mov	("esi",wparam(0));	# ctx

	&mov	("edi",wparam(1));	# inp

	&mov	("eax",wparam(2));	# num

	&mov	("ebx","esp");		# saved sp

	&call	(&label("pic_point"));	# make it PIC!

&set_label("pic_point");

	&blindpop($K512);

	&lea	($K512,&DWP(&label("K512")."-".&label("pic_point"),$K512));

	&sub	("esp",16);

	&and	("esp",-64);

	&shl	("eax",7);

	&add	("eax","edi");

	&mov	(&DWP(0,"esp"),"esi");	# ctx

	&mov	(&DWP(4,"esp"),"edi");	# inp

	&mov	(&DWP(8,"esp"),"eax");	# inp+num*128

	&mov	(&DWP(12,"esp"),"ebx");	# saved sp

	&picmeup("edx","OPENSSL_ia32cap_P",$K512,&label("pic_point"));

	&bt	(&DWP(0,"edx"),26);

	&jnc	(&label("loop_x86"));

	# load ctx->h[0-7]

	&movq	($A,&QWP(0,"esi"));

	&movq	("mm1",&QWP(8,"esi"));

	&movq	("mm2",&QWP(16,"esi"));

	&movq	("mm3",&QWP(24,"esi"));

	&movq	($E,&QWP(32,"esi"));

	&movq	("mm5",&QWP(40,"esi"));

	&movq	("mm6",&QWP(48,"esi"));

	&movq	("mm7",&QWP(56,"esi"));

	&sub	("esp",8*10);

&set_label("loop_sse2",16);

	# &movq	($Asse2,$A);

	&movq	($Bsse2,"mm1");

	&movq	($Csse2,"mm2");

	&movq	($Dsse2,"mm3");

	# &movq	($Esse2,$E);

	&movq	($Fsse2,"mm5");

	&movq	($Gsse2,"mm6");

	&movq	($Hsse2,"mm7");

	&mov	("ecx",&DWP(0,"edi"));

	&mov	("edx",&DWP(4,"edi"));

	&add	("edi",8);

	&bswap	("ecx");

	&bswap	("edx");

	&mov	(&DWP(8*9+4,"esp"),"ecx");

	&mov	(&DWP(8*9+0,"esp"),"edx");

&set_label("00_14_sse2",16);

	&mov	("eax",&DWP(0,"edi"));

	&mov	("ebx",&DWP(4,"edi"));

	&add	("edi",8);

	&bswap	("eax");

	&bswap	("ebx");

	&mov	(&DWP(8*8+4,"esp"),"eax");

	&mov	(&DWP(8*8+0,"esp"),"ebx");

	&BODY_00_15_sse2();

	&cmp	(&LB("edx"),0x35);

	&jne	(&label("00_14_sse2"));

	&BODY_00_15_sse2();

&set_label("16_79_sse2",16);

	&movq	("mm3",&QWP(8*(9+16-1),"esp"));

	&movq	("mm6",&QWP(8*(9+16-14),"esp"));

	&movq	("mm1","mm3");

	&psrlq	("mm3",1);

	&movq	("mm7","mm6");

	&psrlq	("mm6",6);

	&movq	("mm2","mm3");

	&psrlq	("mm3",7-1);

	&movq	("mm5","mm6");

	&psrlq	("mm6",19-6);

	&pxor	("mm2","mm3");

	&psrlq	("mm3",8-7);

	&pxor	("mm5","mm6");

	&psrlq	("mm6",61-19);

	&pxor	("mm2","mm3");

	&movq	("mm3",&QWP(8*(9+16),"esp"));

	&psllq	("mm1",56);

	&pxor	("mm5","mm6");

	&psllq	("mm7",3);

	&pxor	("mm2","mm1");

	&paddq	("mm3",&QWP(8*(9+16-9),"esp"));

	&psllq	("mm1",63-56);