aes/asm/aesni-mb-x86_64.pl

90 ($rounds,$one,$sink,$offset)=("%eax","%ecx","%rbp","%rbx");
187 	xor	$offset,$offset
192 	add	\$16,$offset
195 	sub	$offset,$sink
198 	prefetcht0	31(@inptr[0],$offset)	# prefetch input
199 	prefetcht0	31(@inptr[1],$offset)
201 	prefetcht0	31(@inptr[2],$offset)
202 	prefetcht0	31(@inptr[2],$offset)
223 	prefetcht0	15(@outptr[0],$offset)	# prefetch output
224 	prefetcht0	15(@outptr[1],$offset)
226 	prefetcht0	15(@outptr[2],$offset)
227 	prefetcht0	15(@outptr[3],$offset)
286 	 movdqu		(@inptr[0],$offset),@inp[0]
290 	 movdqu		(@inptr[1],$offset),@inp[1]
293 	 movdqu		(@inptr[2],$offset),@inp[2]
296 	 movdqu		(@inptr[3],$offset),@inp[3]
302 	movups		@out[0],-16(@outptr[0],$offset)
304 	movups		@out[1],-16(@outptr[1],$offset)
306 	movups		@out[2],-16(@outptr[2],$offset)
308 	movups		@out[3],-16(@outptr[3],$offset)
442 	xor	$offset,$offset
447 	add	\$16,$offset
450 	sub	$offset,$sink
453 	prefetcht0	31(@inptr[0],$offset)	# prefetch input
454 	prefetcht0	31(@inptr[1],$offset)
456 	prefetcht0	31(@inptr[2],$offset)
457 	prefetcht0	31(@inptr[3],$offset)
478 	prefetcht0	15(@outptr[0],$offset)	# prefetch output
479 	prefetcht0	15(@outptr[1],$offset)
481 	prefetcht0	15(@outptr[2],$offset)
482 	prefetcht0	15(@outptr[3],$offset)
550 	 movdqu		-16(@inptr[0],$offset),@inp[0]	# load next IV
551 	 movdqu		-16(@inptr[1],$offset),@inp[1]
554 	 movdqu		-16(@inptr[2],$offset),@inp[2]
555 	 movdqu		-16(@inptr[3],$offset),@inp[3]
557 	movups		@out[0],-16(@outptr[0],$offset)
558 	 movdqu		(@inptr[0],$offset),@out[0]
559 	movups		@out[1],-16(@outptr[1],$offset)
560 	 movdqu		(@inptr[1],$offset),@out[1]
562 	movups		@out[2],-16(@outptr[2],$offset)
563 	 movdqu		(@inptr[2],$offset),@out[2]
565 	movups		@out[3],-16(@outptr[3],$offset)
566 	 movdqu		(@inptr[3],$offset),@out[3]
666   my $temp = $i ? $offload : $offset;
719 	 mov		64+8*$i(%rsp),$offset
731 	 lea		(@ptr[$i],$offset),$offset
734 	 cmovg		%rsp,$offset			# sink output
736 	 sub		@ptr[$i],$offset
739 	 mov		$offset,64+8*$i(%rsp)
742 	 lea		16(@ptr[$i],$offset),@ptr[$i]	# switch to output
807 	 mov		64(%rsp),$offset		# pre-load 1st offset
829 	 sub		$offset,@ptr[0]			# switch to input
941   my $temp = $i ? $offload : $offset;
1004 	 mov		64+8*$i(%rsp),$offset
1016 	 lea		(@ptr[$i],$offset),$offset
1019 	 cmovg		%rsp,$offset			# sink output
1021 	 sub		@ptr[$i],$offset
1024 	 mov		$offset,64+8*$i(%rsp)
1027 	 lea		16(@ptr[$i],$offset),@ptr[$i]	# switch to output
1092 	 mov		64(%rsp),$offset		# pre-load 1st offset
1121 	 sub		$offset,@ptr[0]			# switch to input