sha1-586.pl 36.2 KB
Newer Older
	 eval(shift(@insns));

	&vpslld	(@X[4],@X[4],2);
	  &vmovdqa	(@X[2],&QWP(64+16*(($Xi-6)%3),"esp")) if ($Xi>5);	# restore X[] from backtrace buffer
	 eval(shift(@insns));
	 eval(shift(@insns));
	&vpxor	(@X[0],@X[0],@X[3]);
	 eval(shift(@insns));
	 eval(shift(@insns));
	 eval(shift(@insns));
	 eval(shift(@insns));

	&vpxor	(@X[0],@X[0],@X[4]);		# "X[0]"^=("X[0]"<<96)<<<2
	 eval(shift(@insns));
	 eval(shift(@insns));
	  &vmovdqa	(@X[4],&QWP(112-16+16*(($Xi)/5),"esp"));	# K_XX_XX
	 eval(shift(@insns));
	 eval(shift(@insns));

	 foreach (@insns) { eval; }	# remaining instructions [if any]

  $Xi++;	push(@X,shift(@X));	# "rotate" X[]
}

sub Xupdate_avx_32_79()
{ use integer;
  my $body = shift;
  my @insns = (&$body,&$body,&$body,&$body);	# 32 to 48 instructions
  my ($a,$b,$c,$d,$e);

	&vpalignr(@X[2],@X[-1&7],@X[-2&7],8);	# compose "X[-6]"
	&vpxor	(@X[0],@X[0],@X[-4&7]);	# "X[0]"="X[-32]"^"X[-16]"
	 eval(shift(@insns));		# body_20_39
	 eval(shift(@insns));
	 eval(shift(@insns));
	 eval(shift(@insns));		# rol

	&vpxor	(@X[0],@X[0],@X[-7&7]);	# "X[0]"^="X[-28]"
	  &vmovdqa	(&QWP(64+16*(($Xi-4)%3),"esp"),@X[-4&7]);	# save X[] to backtrace buffer
	 eval(shift(@insns));
	 eval(shift(@insns));
	 if ($Xi%5) {
	  &vmovdqa	(@X[4],@X[3]);	# "perpetuate" K_XX_XX...
	 } else {			# ... or load next one
	  &vmovdqa	(@X[4],&QWP(112-16+16*($Xi/5),"esp"));
	 }
	  &vpaddd	(@X[3],@X[3],@X[-1&7]);
	 eval(shift(@insns));		# ror
	 eval(shift(@insns));

	&vpxor	(@X[0],@X[0],@X[2]);		# "X[0]"^="X[-6]"
	 eval(shift(@insns));		# body_20_39
	 eval(shift(@insns));
	 eval(shift(@insns));
	 eval(shift(@insns));		# rol

	&vpsrld	(@X[2],@X[0],30);
	  &vmovdqa	(&QWP(0+16*(($Xi-1)&3),"esp"),@X[3]);	# X[]+K xfer to IALU
	 eval(shift(@insns));
	 eval(shift(@insns));
	 eval(shift(@insns));		# ror
	 eval(shift(@insns));

	&vpslld	(@X[0],@X[0],2);
	 eval(shift(@insns));		# body_20_39
	 eval(shift(@insns));
	 eval(shift(@insns));
	 eval(shift(@insns));		# rol
	 eval(shift(@insns));
	 eval(shift(@insns));
	 eval(shift(@insns));		# ror
	 eval(shift(@insns));

	&vpor	(@X[0],@X[0],@X[2]);	# "X[0]"<<<=2
	 eval(shift(@insns));		# body_20_39
	 eval(shift(@insns));
	  &vmovdqa	(@X[2],&QWP(64+16*(($Xi-6)%3),"esp")) if($Xi<19);	# restore X[] from backtrace buffer
	 eval(shift(@insns));
	 eval(shift(@insns));		# rol
	 eval(shift(@insns));
	 eval(shift(@insns));
	 eval(shift(@insns));		# ror
	 eval(shift(@insns));

	 foreach (@insns) { eval; }	# remaining instructions

  $Xi++;	push(@X,shift(@X));	# "rotate" X[]
}

sub Xuplast_avx_80()
{ use integer;
  my $body = shift;
  my @insns = (&$body,&$body,&$body,&$body);	# 32 instructions
  my ($a,$b,$c,$d,$e);

	 eval(shift(@insns));
	  &vpaddd	(@X[3],@X[3],@X[-1&7]);
	 eval(shift(@insns));
	 eval(shift(@insns));
	 eval(shift(@insns));
	 eval(shift(@insns));

	  &vmovdqa	(&QWP(0+16*(($Xi-1)&3),"esp"),@X[3]);	# X[]+K xfer IALU

	 foreach (@insns) { eval; }		# remaining instructions

	&mov	($inp=@T[1],&DWP(192+4,"esp"));
	&cmp	($inp,&DWP(192+8,"esp"));
	&je	(&label("done"));

	&vmovdqa(@X[3],&QWP(112+48,"esp"));	# K_00_19
	&vmovdqa(@X[2],&QWP(112+64,"esp"));	# pbswap mask
	&vmovdqu(@X[-4&7],&QWP(0,$inp));	# load input
	&vmovdqu(@X[-3&7],&QWP(16,$inp));
	&vmovdqu(@X[-2&7],&QWP(32,$inp));
	&vmovdqu(@X[-1&7],&QWP(48,$inp));
	&add	($inp,64);
	&vpshufb(@X[-4&7],@X[-4&7],@X[2]);		# byte swap
	&mov	(&DWP(192+4,"esp"),$inp);
	&vmovdqa(&QWP(112-16,"esp"),@X[3]);	# borrow last backtrace slot

  $Xi=0;
}

sub Xloop_avx()
{ use integer;
  my $body = shift;
  my @insns = (&$body,&$body,&$body,&$body);	# 32 instructions
  my ($a,$b,$c,$d,$e);

	 eval(shift(@insns));
	 eval(shift(@insns));
	&vpshufb	(@X[($Xi-3)&7],@X[($Xi-3)&7],@X[2]);
	 eval(shift(@insns));
	 eval(shift(@insns));
	&vpaddd	(@X[$Xi&7],@X[($Xi-4)&7],@X[3]);
	 eval(shift(@insns));
	 eval(shift(@insns));
	 eval(shift(@insns));
	 eval(shift(@insns));
	&vmovdqa	(&QWP(0+16*$Xi,"esp"),@X[$Xi&7]);	# X[]+K xfer to IALU
	 eval(shift(@insns));
	 eval(shift(@insns));

	foreach (@insns) { eval; }
  $Xi++;
}

sub Xtail_avx()
{ use integer;
  my $body = shift;
  my @insns = (&$body,&$body,&$body,&$body);	# 32 instructions
  my ($a,$b,$c,$d,$e);

	foreach (@insns) { eval; }
}

&set_label("loop",16);
	&Xupdate_avx_16_31(\&body_00_19);
	&Xupdate_avx_16_31(\&body_00_19);
	&Xupdate_avx_16_31(\&body_00_19);
	&Xupdate_avx_16_31(\&body_00_19);
	&Xupdate_avx_32_79(\&body_00_19);
	&Xupdate_avx_32_79(\&body_20_39);
	&Xupdate_avx_32_79(\&body_20_39);
	&Xupdate_avx_32_79(\&body_20_39);
	&Xupdate_avx_32_79(\&body_20_39);
	&Xupdate_avx_32_79(\&body_20_39);
	&Xupdate_avx_32_79(\&body_40_59);
	&Xupdate_avx_32_79(\&body_40_59);
	&Xupdate_avx_32_79(\&body_40_59);
	&Xupdate_avx_32_79(\&body_40_59);
	&Xupdate_avx_32_79(\&body_40_59);
	&Xupdate_avx_32_79(\&body_20_39);
	&Xuplast_avx_80(\&body_20_39);	# can jump to "done"

				$saved_j=$j; @saved_V=@V;

	&Xloop_avx(\&body_20_39);
	&Xloop_avx(\&body_20_39);
	&Xloop_avx(\&body_20_39);

	&mov	(@T[1],&DWP(192,"esp"));	# update context
	&add	($A,&DWP(0,@T[1]));
	&add	(@T[0],&DWP(4,@T[1]));		# $b
	&add	($C,&DWP(8,@T[1]));
	&mov	(&DWP(0,@T[1]),$A);
	&add	($D,&DWP(12,@T[1]));
	&mov	(&DWP(4,@T[1]),@T[0]);
	&add	($E,&DWP(16,@T[1]));
	&mov	(&DWP(8,@T[1]),$C);
	&mov	($B,@T[0]);
	&mov	(&DWP(12,@T[1]),$D);
	&mov	(&DWP(16,@T[1]),$E);

	&jmp	(&label("loop"));

&set_label("done",16);		$j=$saved_j; @V=@saved_V;

	&Xtail_avx(\&body_20_39);
	&Xtail_avx(\&body_20_39);
	&Xtail_avx(\&body_20_39);

	&vzeroall();

	&mov	(@T[1],&DWP(192,"esp"));	# update context
	&add	($A,&DWP(0,@T[1]));
	&mov	("esp",&DWP(192+12,"esp"));	# restore %esp
	&add	(@T[0],&DWP(4,@T[1]));		# $b
	&add	($C,&DWP(8,@T[1]));
	&mov	(&DWP(0,@T[1]),$A);
	&add	($D,&DWP(12,@T[1]));
	&mov	(&DWP(4,@T[1]),@T[0]);
	&add	($E,&DWP(16,@T[1]));
	&mov	(&DWP(8,@T[1]),$C);
	&mov	(&DWP(12,@T[1]),$D);
	&mov	(&DWP(16,@T[1]),$E);
&function_end("_sha1_block_data_order_avx");
}
&set_label("K_XX_XX",64);
&data_word(0x5a827999,0x5a827999,0x5a827999,0x5a827999);	# K_00_19
&data_word(0x6ed9eba1,0x6ed9eba1,0x6ed9eba1,0x6ed9eba1);	# K_20_39
&data_word(0x8f1bbcdc,0x8f1bbcdc,0x8f1bbcdc,0x8f1bbcdc);	# K_40_59
&data_word(0xca62c1d6,0xca62c1d6,0xca62c1d6,0xca62c1d6);	# K_60_79
&data_word(0x00010203,0x04050607,0x08090a0b,0x0c0d0e0f);	# pbswap mask
}
&asciz("SHA1 block transform for x86, CRYPTOGAMS by <appro\@openssl.org>");

&asm_finish();