modes/asm/ghash-x86_64.pl

64 $Xi="%rdi";
175 	movzb	15($Xi),$Zlo
178 	&loop	($Xi);
180 	mov	$Zlo,8($Xi)
181 	mov	$Zhi,($Xi)
245 	&mov	($Zlo,"8($Xi)");
246 	&mov	($Zhi,"0($Xi)");
256 	&mov	("($Xi)",$Zhi);
257 	&mov	("8($Xi)","%rdx");
292 	    &mov	($dat,"$j($Xi)")			if (--$j%4==0 && $j>=0);
333 	mov	$Zlo,8($Xi)
334 	mov	$Zhi,($Xi)
356 ($Xi,$Xhi)=("%xmm0","%xmm1");	$Hkey="%xmm2";
360 my ($Xhi,$Xi,$Hkey,$modulo)=@_;
363 	movdqa		$Xi,$Xhi		#
364 	pshufd		\$0b01001110,$Xi,$T1
366 	pxor		$Xi,$T1			#
370 	pclmulqdq	\$0x00,$Hkey,$Xi	#######
373 	pxor		$Xi,$T1			#
380 	pxor		$T2,$Xi			#
385 my ($Xhi,$Xi) = @_;
389 	movdqa		$Xi,$T1			#
390 	psllq		\$1,$Xi
391 	pxor		$T1,$Xi			#
392 	psllq		\$5,$Xi			#
393 	pxor		$T1,$Xi			#
394 	psllq		\$57,$Xi		#
395 	movdqa		$Xi,$T2			#
396 	pslldq		\$8,$Xi
398 	pxor		$T1,$Xi
402 	movdqa		$Xi,$T2
403 	psrlq		\$5,$Xi
404 	pxor		$T2,$Xi			#
405 	psrlq		\$1,$Xi			#
406 	pxor		$T2,$Xi			#
408 	psrlq		\$1,$Xi			#
409 	pxor		$T2,$Xi			#
440 	movdqa		$Hkey,$Xi
442 	&clmul64x64_T2	($Xhi,$Xi,$Hkey);
443 	&reduction_alg9	($Xhi,$Xi);
446 	movdqu		$Xi,16($Htbl)		# save H^2
460 	movdqu		($Xip),$Xi
463 	pshufb		$T3,$Xi
465 	&clmul64x64_T2	($Xhi,$Xi,$Hkey);
466 	&reduction_alg9	($Xhi,$Xi);
468 	pshufb		$T3,$Xi
469 	movdqu		$Xi,($Xip)
503 	movdqu		($Xip),$Xi
505 	pshufb		$T3,$Xi
512 	# Xi+2 =[H*(Ii+1 + Xi+1)] mod P =
513 	#	[(H*Ii+1) + (H*Xi+1)] mod P =
514 	#	[(H*Ii+1) + H^2*(Ii+Xi)] mod P
520 	pxor		$T1,$Xi			# Ii+Xi
524 	movdqa		$Xi,$Xhi		#
525 	pshufd		\$0b01001110,$Xi,$T1
527 	pxor		$Xi,$T1			#
536 	&clmul64x64_T2	($Xhi,$Xi,$Hkey2,1);	# H^2*(Ii+Xi)
539 	pxor		$Xn,$Xi			# (H*Ii+1) + H^2*(Ii+Xi)
551 	 pxor		$T1,$Xhi		# "Ii+Xi", consume early
553 	  movdqa	$Xi,$T1			# 1st phase
554 	  psllq		\$1,$Xi
555 	  pxor		$T1,$Xi			#
556 	  psllq		\$5,$Xi			#
557 	  pxor		$T1,$Xi			#
559 	  psllq		\$57,$Xi		#
560 	  movdqa	$Xi,$T2			#
561 	  pslldq	\$8,$Xi
563 	  pxor		$T1,$Xi
567 	  movdqa	$Xi,$T2			# 2nd phase
568 	  psrlq		\$5,$Xi
569 	  pxor		$T2,$Xi			#
570 	  psrlq		\$1,$Xi			#
571 	  pxor		$T2,$Xi			#
573 	  psrlq		\$1,$Xi			#
574 	  pxor		$T2,$Xi			#
577 	 movdqa		$Xi,$Xhi		#
578 	 pshufd		\$0b01001110,$Xi,$T1
580 	 pxor		$Xi,$T1			#
597 	&clmul64x64_T2	($Xhi,$Xi,$Hkey2,1);	# H^2*(Ii+Xi)
599 	pxor		$Xn,$Xi			# (H*Ii+1) + H^2*(Ii+Xi)
602 	&reduction_alg9	($Xhi,$Xi);
610 	pxor		$T1,$Xi			# Ii+Xi
612 	&clmul64x64_T2	($Xhi,$Xi,$Hkey);	# H*(Ii+Xi)
613 	&reduction_alg9	($Xhi,$Xi);
616 	pshufb		$T3,$Xi
617 	movdqu		$Xi,($Xip)