aes/asm/bsaes-x86_64.pl

34 # - this module doesn't implement key setup subroutine, instead it
35 #   relies on conversion of "conventional" key schedule as returned
38 #   to skip one shiftrows(), reduce bit-sliced key schedule and
43 # of 4096-byte buffer with 128-bit key is:
62 # As for key schedule conversion subroutine. Interface to OpenSSL
85 # one byte out of 4096-byte buffer with 128-bit key is:
114 my ($inp,$out,$len,$key,$ivp)=("%rdi","%rsi","%rdx","%rcx");
119 my ($key,$rounds,$const)=("%rax","%r10d","%r11");
438 	pxor	0x00($key),@x[0]
439 	pxor	0x10($key),@x[1]
440 	pxor	0x20($key),@x[2]
441 	pxor	0x30($key),@x[3]
444 	pxor	0x40($key),@x[4]
445 	pxor	0x50($key),@x[5]
448 	pxor	0x60($key),@x[6]
449 	pxor	0x70($key),@x[7]
454 	lea	0x80($key),$key
748 	pxor	0x00($key),@b[0]
749 	pxor	0x10($key),@b[1]
750 	pxor	0x20($key),@b[4]
751 	pxor	0x30($key),@b[6]
752 	pxor	0x40($key),@b[3]
753 	pxor	0x50($key),@b[7]
754 	pxor	0x60($key),@b[2]
755 	pxor	0x70($key),@b[5]
821 	movdqa	($key), @XMM[9]		# round 0 key
822 	lea	0x10($key), $key
824 	pxor	@XMM[9], @XMM[0]	# xor with round0 key
868 	movdqa	($key), @XMM[8]		# last round key
885 	movdqa	($key), @XMM[9]		# round 0 key
886 	lea	0x10($key), $key
888 	pxor	@XMM[9], @XMM[0]	# xor with round0 key
930 	movdqa	($key), @XMM[8]		# last round key
975 	movdqu	($inp), %xmm7		# load round 0 key
984 	movdqu	($inp), %xmm6		# load round 1 key
985 	movdqa	%xmm7, ($out)		# save round 0 key
1021 	 movdqa	%xmm8, 0x00($out)	# write bit-sliced round key
1037 	 movdqu	($inp), %xmm6		# load next round key
1050 	#movdqa	%xmm6, ($out)		# don't save last round key
1064 	mov	$inp,%rcx		# pass key
1065 	mov	$out,%rax		# pass key schedule
1067 	pxor	%xmm6,%xmm7		# fix up last round key
1068 	movdqa	%xmm7,(%rax)		# save last round key
1085 	mov	$key, %rax		# pass the $key
1110 	mov	$inp,%rcx		# pass key
1111 	mov	$out,%rax		# pass key schedule
1113 	pxor	($out),%xmm7		# fix up round 0 key
1114 	movdqa	%xmm6,(%rax)		# save last round key
1132 	mov	$key, %rax		# pass the $key
1160 my ($inp,$out,$len,$key)=("%r12","%r13","%r14","%r15");
1207 	mov	$arg4,$key
1212 	shl	\$7,%rax		# 128 bytes per inner round key
1213 	sub	\$`128-32`,%rax		# size of bit-sliced key schedule
1215 	mov	%rsp,%rax		# pass key schedule
1216 	mov	$key,%rcx		# pass key
1219 	pxor	%xmm6,%xmm7		# fix up last round key
1220 	movdqa	%xmm7,(%rax)		# save last round key
1230 	mov	%rsp, %rax		# pass key schedule
1254 	mov	%rsp, %rax		# pass key schedule
1329 	lea	($key), $arg3
1339 .Lecb_enc_bzero:			# wipe key schedule [if any]
1426 	mov	$arg4,$key
1431 	shl	\$7,%rax		# 128 bytes per inner round key
1432 	sub	\$`128-32`,%rax		# size of bit-sliced key schedule
1434 	mov	%rsp,%rax		# pass key schedule
1435 	mov	$key,%rcx		# pass key
1438 	pxor	(%rsp),%xmm7		# fix up 0 round key
1439 	movdqa	%xmm6,(%rax)		# save last round key
1450 	mov	%rsp, %rax		# pass key schedule
1474 	mov	%rsp, %rax		# pass key schedule
1549 	lea	($key), $arg3
1559 .Lecb_dec_bzero:			# wipe key schedule [if any]
1660 	mov	$arg4, $key
1665 	shl	\$7, %rax		# 128 bytes per inner round key
1666 	sub	\$`128-32`, %rax	# size of bit-sliced key schedule
1669 	mov	%rsp, %rax		# pass key schedule
1670 	mov	$key, %rcx		# pass key
1673 	pxor	(%rsp),%xmm7		# fix up 0 round key
1674 	movdqa	%xmm6,(%rax)		# save last round key
1686 	mov	%rsp, %rax		# pass key schedule
1727 	mov	%rsp, %rax		# pass key schedule
1857 	lea	($key), $arg3
1867 .Lcbc_dec_bzero:			# wipe key schedule [if any]
1956 	mov	$arg4, $key
1962 	shl	\$7, %rax		# 128 bytes per inner round key
1963 	sub	\$`128-32`, %rax	# size of bit-sliced key schedule
1966 	mov	%rsp, %rax		# pass key schedule
1967 	mov	$key, %rcx		# pass key
1970 	pxor	%xmm6,%xmm7		# fix up last round key
1971 	movdqa	%xmm7,(%rax)		# save last round key
1973 	movdqa	(%rsp), @XMM[9]		# load round0 key
1979 	movdqa	@XMM[9], (%rsp)		# save adjusted round0 key
2001 	movdqa	(%rsp), @XMM[9]		# round 0 key
2002 	lea	0x10(%rsp), %rax	# pass key schedule
2004 	pxor	@XMM[9], @XMM[0]	# xor with round0 key
2099 	lea	($key), $arg3
2117 .Lctr_enc_bzero:			# wipe key schedule [if any]
2214 	mov	$arg4, $key
2221 	mov	240($key), %eax		# rounds
2225 	shl	\$7, %rax		# 128 bytes per inner round key
2226 	sub	\$`128-32`, %rax	# size of bit-sliced key schedule
2229 	mov	%rsp, %rax		# pass key schedule
2230 	mov	$key, %rcx		# pass key
2233 	pxor	%xmm6, %xmm7		# fix up last round key
2234 	movdqa	%xmm7, (%rax)		# save last round key
2276 	lea	0x80(%rsp), %rax	# pass key schedule
2344 	lea	0x80(%rsp), %rax	# pass key schedule
2372 	lea	0x80(%rsp), %rax	# pass key schedule
2398 	lea	0x80(%rsp), %rax	# pass key schedule
2422 	lea	0x80(%rsp), %rax	# pass key schedule
2444 	lea	0x80(%rsp), %rax	# pass key schedule
2464 	lea	0x80(%rsp), %rax	# pass key schedule
2484 	lea	($key), $arg3
2488 	#lea	0x80(%rsp), %rax	# pass key schedule
2517 	lea	($key), $arg3
2525 .Lxts_enc_bzero:			# wipe key schedule [if any]
2612 	mov	$arg4, $key
2619 	mov	240($key), %eax		# rounds
2623 	shl	\$7, %rax		# 128 bytes per inner round key
2624 	sub	\$`128-32`, %rax	# size of bit-sliced key schedule
2627 	mov	%rsp, %rax		# pass key schedule
2628 	mov	$key, %rcx		# pass key
2631 	pxor	(%rsp), %xmm7		# fix up round 0 key
2632 	movdqa	%xmm6, (%rax)		# save last round key
2681 	lea	0x80(%rsp), %rax	# pass key schedule
2749 	lea	0x80(%rsp), %rax	# pass key schedule
2777 	lea	0x80(%rsp), %rax	# pass key schedule
2803 	lea	0x80(%rsp), %rax	# pass key schedule
2827 	lea	0x80(%rsp), %rax	# pass key schedule
2849 	lea	0x80(%rsp), %rax	# pass key schedule
2869 	lea	0x80(%rsp), %rax	# pass key schedule
2889 	lea	($key), $arg3
2893 	#lea	0x80(%rsp), %rax	# pass key schedule
2920 	lea	($key), $arg3
2941 	lea	($key), $arg3
2949 .Lxts_dec_bzero:			# wipe key schedule [if any]