[deliverable/linux.git] / arch / alpha / lib / csum_ipv6_magic.S

/*
 * arch/alpha/lib/csum_ipv6_magic.S
 * Contributed by Richard Henderson <rth@tamu.edu>
 *
 * unsigned short csum_ipv6_magic(struct in6_addr *saddr,
 *                                struct in6_addr *daddr,
 *                                __u32 len,
 *                                unsigned short proto,
 *                                unsigned int csum);
 *
 * Misalignment handling (which costs 16 instructions / 8 cycles)
 * added by Ivan Kokshaysky <ink@jurassic.park.msu.ru>
 */

	.globl csum_ipv6_magic
	.align 4
	.ent csum_ipv6_magic
	.frame $30,0,$26,0
csum_ipv6_magic:
	.prologue 0

	ldq_u	$0,0($16)	# e0    : load src & dst addr words
	zapnot	$20,15,$20	# .. e1 : zero extend incoming csum
	extqh	$18,1,$4	# e0    : byte swap len & proto while we wait
	ldq_u	$21,7($16)	# .. e1 : handle misalignment

	extbl	$18,1,$5	# e0	:
	ldq_u	$1,8($16)	# .. e1 :
	extbl	$18,2,$6	# e0 	:
	ldq_u	$22,15($16)	# .. e1 :

	extbl	$18,3,$18	# e0	:
	ldq_u	$2,0($17)	# .. e1 :
	sra	$4,32,$4	# e0	:
	ldq_u	$23,7($17)	# .. e1 :

	extql	$0,$16,$0	# e0	:
	ldq_u	$3,8($17)	# .. e1 :
	extqh	$21,$16,$21	# e0	:
	ldq_u	$24,15($17)	# .. e1 :

	sll	$5,16,$5	# e0	:
	or	$0,$21,$0	# .. e1 : 1st src word complete
	extql	$1,$16,$1	# e0	:
	addq	$20,$0,$20	# .. e1 : begin summing the words

	extqh	$22,$16,$22	# e0	:
	cmpult	$20,$0,$0	# .. e1 :
	sll	$6,8,$6		# e0	:
	or	$1,$22,$1	# .. e1 : 2nd src word complete

	extql	$2,$17,$2	# e0	:
	or	$4,$18,$18	# .. e1 :
	extqh	$23,$17,$23	# e0	:
	or	$5,$6,$5	# .. e1 :

	extql	$3,$17,$3	# e0	:
	or	$2,$23,$2	# .. e1 : 1st dst word complete
	extqh	$24,$17,$24	# e0	:
	or	$18,$5,$18	# .. e1 : len complete

	extwh	$19,7,$7	# e0    :
	or	$3,$24,$3	# .. e1 : 2nd dst word complete
	extbl	$19,1,$19	# e0    :
	addq	$20,$1,$20	# .. e1 :

	or	$19,$7,$19	# e0    :
	cmpult	$20,$1,$1	# .. e1 :
	sll	$19,48,$19	# e0    :
	nop			# .. e0 :

	sra	$19,32,$19	# e0    : proto complete
	addq	$20,$2,$20	# .. e1 :
	cmpult	$20,$2,$2	# e0    :
	addq	$20,$3,$20	# .. e1 :

	cmpult	$20,$3,$3	# e0    :
	addq	$20,$18,$20	# .. e1 :
	cmpult	$20,$18,$18	# e0    :
	addq	$20,$19,$20	# .. e1 :

	cmpult	$20,$19,$19	# e0    :
	addq	$0,$1,$0	# .. e1 : merge the carries back into the csum
	addq	$2,$3,$2	# e0    :
	addq	$18,$19,$18	# .. e1 :

	addq	$0,$2,$0	# e0    :
	addq	$20,$18,$20	# .. e1 :
	addq	$0,$20,$0	# e0    :
	unop			#       :

	extwl	$0,2,$2		# e0    : begin folding the 64-bit value
	zapnot	$0,3,$3		# .. e1 :
	extwl	$0,4,$1		# e0    :
	addq	$2,$3,$3	# .. e1 :

	extwl	$0,6,$0		# e0    :
	addq	$3,$1,$3	# .. e1 :
	addq	$0,$3,$0	# e0    :
	unop			#       :

	extwl	$0,2,$1		# e0    : fold 18-bit value
	zapnot	$0,3,$0		# .. e1 :
	addq	$0,$1,$0	# e0    :
	unop			#       :

	extwl	$0,2,$1		# e0    : fold 17-bit value
	zapnot	$0,3,$0		# .. e1 :
	addq	$0,$1,$0	# e0    :
	not	$0,$0		# .. e1 : and complement.

	zapnot	$0,3,$0		# e0    :
	ret			# .. e1 :

	.end csum_ipv6_magic
Commit	Line	Data
1da177e4 LT	1	/*
	2	* arch/alpha/lib/csum_ipv6_magic.S
	3	* Contributed by Richard Henderson <rth@tamu.edu>
	4	*
	5	* unsigned short csum_ipv6_magic(struct in6_addr *saddr,
	6	* struct in6_addr *daddr,
	7	* __u32 len,
	8	* unsigned short proto,
	9	* unsigned int csum);
58ed2f9c IK	10	*
	11	* Misalignment handling (which costs 16 instructions / 8 cycles)
	12	* added by Ivan Kokshaysky <ink@jurassic.park.msu.ru>
1da177e4 LT	13	*/
	14
	15	.globl csum_ipv6_magic
	16	.align 4
	17	.ent csum_ipv6_magic
	18	.frame $30,0,$26,0
	19	csum_ipv6_magic:
	20	.prologue 0
	21
58ed2f9c	22	ldq_u $0,0($16) # e0 : load src & dst addr words
1da177e4 LT	23	zapnot $20,15,$20 # .. e1 : zero extend incoming csum
1da177e4 LT	24	extqh $18,1,$4 # e0 : byte swap len & proto while we wait
58ed2f9c	25	ldq_u $21,7($16) # .. e1 : handle misalignment
1da177e4 LT	26
1da177e4 LT	27	extbl $18,1,$5 # e0 :
58ed2f9c	28	ldq_u $1,8($16) # .. e1 :
1da177e4	29	extbl $18,2,$6 # e0 :
58ed2f9c	30	ldq_u $22,15($16) # .. e1 :
1da177e4 LT	31
1da177e4 LT	32	extbl $18,3,$18 # e0 :
58ed2f9c	33	ldq_u $2,0($17) # .. e1 :
1da177e4	34	sra $4,32,$4 # e0 :
58ed2f9c IK	35	ldq_u $23,7($17) # .. e1 :
	36
	37	extql $0,$16,$0 # e0 :
	38	ldq_u $3,8($17) # .. e1 :
	39	extqh $21,$16,$21 # e0 :
	40	ldq_u $24,15($17) # .. e1 :
	41
1da177e4	42	sll $5,16,$5 # e0 :
58ed2f9c IK	43	or $0,$21,$0 # .. e1 : 1st src word complete
58ed2f9c IK	44	extql $1,$16,$1 # e0 :
1da177e4 LT	45	addq $20,$0,$20 # .. e1 : begin summing the words
1da177e4 LT	46
58ed2f9c	47	extqh $22,$16,$22 # e0 :
1da177e4	48	cmpult $20,$0,$0 # .. e1 :
58ed2f9c IK	49	sll $6,8,$6 # e0 :
58ed2f9c IK	50	or $1,$22,$1 # .. e1 : 2nd src word complete
1da177e4	51
58ed2f9c IK	52	extql $2,$17,$2 # e0 :
	53	or $4,$18,$18 # .. e1 :
	54	extqh $23,$17,$23 # e0 :
1da177e4	55	or $5,$6,$5 # .. e1 :
1da177e4	56
58ed2f9c IK	57	extql $3,$17,$3 # e0 :
	58	or $2,$23,$2 # .. e1 : 1st dst word complete
	59	extqh $24,$17,$24 # e0 :
	60	or $18,$5,$18 # .. e1 : len complete
	61
	62	extwh $19,7,$7 # e0 :
	63	or $3,$24,$3 # .. e1 : 2nd dst word complete
	64	extbl $19,1,$19 # e0 :
1da177e4	65	addq $20,$1,$20 # .. e1 :
58ed2f9c IK	66
58ed2f9c IK	67	or $19,$7,$19 # e0 :
1da177e4	68	cmpult $20,$1,$1 # .. e1 :
58ed2f9c IK	69	sll $19,48,$19 # e0 :
58ed2f9c IK	70	nop # .. e0 :
1da177e4	71
58ed2f9c	72	sra $19,32,$19 # e0 : proto complete
1da177e4 LT	73	addq $20,$2,$20 # .. e1 :
	74	cmpult $20,$2,$2 # e0 :
	75	addq $20,$3,$20 # .. e1 :
	76
	77	cmpult $20,$3,$3 # e0 :
	78	addq $20,$18,$20 # .. e1 :
	79	cmpult $20,$18,$18 # e0 :
	80	addq $20,$19,$20 # .. e1 :
	81
	82	cmpult $20,$19,$19 # e0 :
	83	addq $0,$1,$0 # .. e1 : merge the carries back into the csum
	84	addq $2,$3,$2 # e0 :
	85	addq $18,$19,$18 # .. e1 :
	86
	87	addq $0,$2,$0 # e0 :
	88	addq $20,$18,$20 # .. e1 :
	89	addq $0,$20,$0 # e0 :
	90	unop # :
	91
	92	extwl $0,2,$2 # e0 : begin folding the 64-bit value
	93	zapnot $0,3,$3 # .. e1 :
	94	extwl $0,4,$1 # e0 :
	95	addq $2,$3,$3 # .. e1 :
	96
	97	extwl $0,6,$0 # e0 :
	98	addq $3,$1,$3 # .. e1 :
	99	addq $0,$3,$0 # e0 :
	100	unop # :
	101
	102	extwl $0,2,$1 # e0 : fold 18-bit value
	103	zapnot $0,3,$0 # .. e1 :
	104	addq $0,$1,$0 # e0 :
	105	unop # :
	106
	107	extwl $0,2,$1 # e0 : fold 17-bit value
	108	zapnot $0,3,$0 # .. e1 :
	109	addq $0,$1,$0 # e0 :
58ed2f9c	110	not $0,$0 # .. e1 : and complement.
1da177e4 LT	111
	112	zapnot $0,3,$0 # e0 :
	113	ret # .. e1 :
	114
	115	.end csum_ipv6_magic