mips/cavium-octeon/octeon-memcpy.S

18 #include <asm/asm-offsets.h>
30  *   - src and dst don't overlap
31  *   - src is readable
32  *   - dst is writable
40  *     - src is readable  (no exceptions when reading src)
42  *     - dst is writable  (no exceptions when writing dst)
43  * __copy_user uses a non-standard calling convention; see
57  *  1- AT contain the address of the byte just past the end of the source
59  *  2- src_entry <= src < AT, and
60  *  3- (dst - src) == (dst_entry - src_entry),
81  * Only on the 64-bit kernel we can made use of 64-bit registers.
91 #define SUB    dsubu  macro
133 #define REST(unit)  (FIRST(unit)+NBYTES-1)
136 #define ADDRMASK (NBYTES-1)
191 	SUB	len, len, 16*NBYTES
206 EXC(	LOAD	t0, UNIT(-8)(src),	l_exc_copy_rewind16)
207 EXC(	LOAD	t1, UNIT(-7)(src),	l_exc_copy_rewind16)
208 EXC(	LOAD	t2, UNIT(-6)(src),	l_exc_copy_rewind16)
209 EXC(	LOAD	t3, UNIT(-5)(src),	l_exc_copy_rewind16)
210 EXC(	STORE	t0, UNIT(-8)(dst),	s_exc_p8u)
211 EXC(	STORE	t1, UNIT(-7)(dst),	s_exc_p7u)
212 EXC(	STORE	t2, UNIT(-6)(dst),	s_exc_p6u)
213 EXC(	STORE	t3, UNIT(-5)(dst),	s_exc_p5u)
214 EXC(	LOAD	t0, UNIT(-4)(src),	l_exc_copy_rewind16)
215 EXC(	LOAD	t1, UNIT(-3)(src),	l_exc_copy_rewind16)
216 EXC(	LOAD	t2, UNIT(-2)(src),	l_exc_copy_rewind16)
217 EXC(	LOAD	t3, UNIT(-1)(src),	l_exc_copy_rewind16)
218 EXC(	STORE	t0, UNIT(-4)(dst),	s_exc_p4u)
219 EXC(	STORE	t1, UNIT(-3)(dst),	s_exc_p3u)
220 EXC(	STORE	t2, UNIT(-2)(dst),	s_exc_p2u)
221 EXC(	STORE	t3, UNIT(-1)(dst),	s_exc_p1u)
239 	SUB	len, len, 8*NBYTES
266 	SUB	len, len, 4*NBYTES
286 	SUB	len, len, NBYTES
296 	SUB	len, len, NBYTES
306 	SUB	len, len, NBYTES
310 EXC(	 STORE	t0, -8(dst),		s_exc_p1u)
314 	SRL	t0, len, LOG_NBYTES+2	 # +2 for 4 units/iter
316 	 and	rem, len, (4*NBYTES-1)	 # rem = len % 4*NBYTES
326 	SUB	len, len, 4*NBYTES
343 	 and	rem, len, NBYTES-1  # rem = len % NBYTES
349 	SUB	len, len, NBYTES
362 	SUB	len, len, 1;		\
372 EXC(	lb	t0, NBYTES-2(src), l_exc)
373 	SUB	len, len, 1
375 EXC(	 sb	t0, NBYTES-2(dst), s_exc_p1)
383 	SUB	src, src, 16*NBYTES
384 	SUB	dst, dst, 16*NBYTES
401 	sb	t1, 0(dst)	# can't fault -- we're copy_from_user
407 	SUB	len, AT, t0		# len number of uncopied bytes
446 	sltu	t0, a1, t0			# dst + len <= src -> memcpy
447 	sltu	t1, a0, t1			# dst >= src + len -> memcpy
463 	lb	t0, -1(a1)
464 	SUB	a2, a2, 0x1
465 	sb	t0, -1(a0)
466 	SUB	a1, a1, 0x1
468 	 SUB	a0, a0, 0x1
476 	SUB	a2, a2, 0x1