alpha/lib/ev6-memcpy.S

9  *	- uses bcmpge to compare 8 bytes in parallel
18  *	L	- lower subcluster; L0 - subcluster L0; L1 - subcluster L1
37 	and	$1, 7, $1		# E : are they the same mod 8?
40 	/* source and dest are same mod 8 address */
52 	ldbu	$1, 0($17)		# L : grab a byte
55 	stb	$1, 0($16)		# L :
68 	ldq	$1, 0($17)		# L : get 8 bytes
69 	subq	$18, 8, $18		# E : count -= 8
70 	addq	$17, 8, $17		# E : src += 8
73 	stq	$1, 0($16)		# L : store
74 	addq	$16, 8, $16		# E : dest += 8
91 	ldq	$4, 8($17)		# L : bytes 8..15
92 	ldq	$5, 16($17)		# L : bytes 16..23
96 	ldq	$3, 24($17)		# L : bytes 24..31
102 	stq	$6, 0($16)		# L : bytes 0..7
106 	stq	$4, 8($16)		# L : bytes 8..15
107 	stq	$5, 16($16)		# L : bytes 16..23
111 	stq	$3, 24($16)		# L : bytes 24..31
116 	ldq	$6, 0($17)		# L : bytes 0..7
117 	ldq	$4, 8($17)		# L : bytes 8..15
122 	ldq	$5, 16($17)		# L : bytes 16..23
123 	ldq	$3, 24($17)		# L : bytes 24..31
128 	stq	$6, -32($16)		# L : bytes 0..7
129 	stq	$4, -24($16)		# L : bytes 8..15
132 	stq	$5, -16($16)		# L : bytes 16..23
133 	stq	$3, -8($16)		# L : bytes 24..31
140 	subq	$18, 8, $18		# E : At least a quad left?
146 	ldq	$1, 0($17)		# L : fetch 8
147 	subq	$18, 8, $18		# E : count -= 8
148 	addq	$17, 8, $17		# E : src += 8
151 	stq	$1, 0($16)		# L : store 8
152 	addq	$16, 8, $16		# E : dest += 8
158 	addq	$18, 8, $18		# E : add back for trailing bytes
166 	ldbu	$1, 0($17)		# L : fetch a byte
170 	stb	$1, 0($16)		# L : store a byte
189 	ldbu	$1, 0($17)		# L : fetch a byte
193 	stb	$1, 0($4)		# L : store it
200 	subq	$18, 8, $18		# E : At least a quad left?
202 	ldq_u	$3, 0($17)		# L : seed (rotating load) of 8 bytes
206 	ldq_u	$16, 8($17)		# L : Fetch next 8
211 	subq	$18, 8, $18		# E : count -= 8
212 	addq	$17, 8, $17		# E : src += 8
213 	stq	$1, 0($4)		# L : store 8 (aligned)
216 	addq	$4, 8, $4		# E : dest += 8
222 	addq	$18, 8, $18		# E : account for tail stuff
228 	ldbu	$1, 0($17)		# L : fetch 1
233 	stb	$1, 0($4)		# L : store