[deliverable/linux.git] / arch / sh / lib64 / copy_page.S

/*
   Copyright 2003 Richard Curnow, SuperH (UK) Ltd.

   This file is subject to the terms and conditions of the GNU General Public
   License.  See the file "COPYING" in the main directory of this archive
   for more details.

   Tight version of mempy for the case of just copying a page.
   Prefetch strategy empirically optimised against RTL simulations
   of SH5-101 cut2 eval chip with Cayman board DDR memory.

   Parameters:
   r2 : destination effective address (start of page)
   r3 : source effective address (start of page)

   Always copies 4096 bytes.

   Points to review.
   * Currently the prefetch is 4 lines ahead and the alloco is 2 lines ahead.
     It seems like the prefetch needs to be at at least 4 lines ahead to get
     the data into the cache in time, and the allocos contend with outstanding
     prefetches for the same cache set, so it's better to have the numbers
     different.
   */

	.section .text..SHmedia32,"ax"
	.little

	.balign 8
	.global copy_page
copy_page:

	/* Copy 4096 bytes worth of data from r3 to r2.
	   Do prefetches 4 lines ahead.
	   Do alloco 2 lines ahead */

	pta 1f, tr1
	pta 2f, tr2
	pta 3f, tr3
	ptabs r18, tr0

#if 0
	/* TAKum03020 */
	ld.q r3, 0x00, r63
	ld.q r3, 0x20, r63
	ld.q r3, 0x40, r63
	ld.q r3, 0x60, r63
#endif
	alloco r2, 0x00
	synco		! TAKum03020
	alloco r2, 0x20
	synco		! TAKum03020

	movi 3968, r6
	add  r2, r6, r6
	addi r6, 64, r7
	addi r7, 64, r8
	sub r3, r2, r60
	addi r60, 8, r61
	addi r61, 8, r62
	addi r62, 8, r23
	addi r60, 0x80, r22

/* Minimal code size.  The extra branches inside the loop don't cost much
   because they overlap with the time spent waiting for prefetches to
   complete. */
1:
#if 0
	/* TAKum03020 */
	bge/u r2, r6, tr2  ! skip prefetch for last 4 lines
	ldx.q r2, r22, r63 ! prefetch 4 lines hence
#endif
2:
	bge/u r2, r7, tr3  ! skip alloco for last 2 lines
	alloco r2, 0x40    ! alloc destination line 2 lines ahead
	synco		! TAKum03020
3:
	ldx.q r2, r60, r36
	ldx.q r2, r61, r37
	ldx.q r2, r62, r38
	ldx.q r2, r23, r39
	st.q  r2,   0, r36
	st.q  r2,   8, r37
	st.q  r2,  16, r38
	st.q  r2,  24, r39
	addi r2, 32, r2
	bgt/l r8, r2, tr1

	blink tr0, r63	   ! return
Commit	Line	Data
1da177e4 LT	1	/*
	2	Copyright 2003 Richard Curnow, SuperH (UK) Ltd.
	3
	4	This file is subject to the terms and conditions of the GNU General Public
	5	License. See the file "COPYING" in the main directory of this archive
	6	for more details.
	7
	8	Tight version of mempy for the case of just copying a page.
	9	Prefetch strategy empirically optimised against RTL simulations
	10	of SH5-101 cut2 eval chip with Cayman board DDR memory.
	11
	12	Parameters:
379a95d1 PM	13	r2 : destination effective address (start of page)
379a95d1 PM	14	r3 : source effective address (start of page)
1da177e4 LT	15
	16	Always copies 4096 bytes.
	17
	18	Points to review.
	19	* Currently the prefetch is 4 lines ahead and the alloco is 2 lines ahead.
	20	It seems like the prefetch needs to be at at least 4 lines ahead to get
	21	the data into the cache in time, and the allocos contend with outstanding
	22	prefetches for the same cache set, so it's better to have the numbers
	23	different.
	24	*/
	25
	26	.section .text..SHmedia32,"ax"
	27	.little
	28
	29	.balign 8
379a95d1 PM	30	.global copy_page
379a95d1 PM	31	copy_page:
1da177e4	32
379a95d1	33	/* Copy 4096 bytes worth of data from r3 to r2.
1da177e4 LT	34	Do prefetches 4 lines ahead.
	35	Do alloco 2 lines ahead */
	36
	37	pta 1f, tr1
	38	pta 2f, tr2
	39	pta 3f, tr3
	40	ptabs r18, tr0
	41
	42	#if 0
	43	/* TAKum03020 */
379a95d1 PM	44	ld.q r3, 0x00, r63
	45	ld.q r3, 0x20, r63
	46	ld.q r3, 0x40, r63
	47	ld.q r3, 0x60, r63
1da177e4	48	#endif
379a95d1	49	alloco r2, 0x00
1da177e4	50	synco ! TAKum03020
379a95d1	51	alloco r2, 0x20
1da177e4 LT	52	synco ! TAKum03020
	53
	54	movi 3968, r6
379a95d1	55	add r2, r6, r6
1da177e4 LT	56	addi r6, 64, r7
1da177e4 LT	57	addi r7, 64, r8
379a95d1	58	sub r3, r2, r60
1da177e4 LT	59	addi r60, 8, r61
	60	addi r61, 8, r62
	61	addi r62, 8, r23
	62	addi r60, 0x80, r22
	63
	64	/* Minimal code size. The extra branches inside the loop don't cost much
	65	because they overlap with the time spent waiting for prefetches to
	66	complete. */
	67	1:
	68	#if 0
	69	/* TAKum03020 */
379a95d1 PM	70	bge/u r2, r6, tr2 ! skip prefetch for last 4 lines
379a95d1 PM	71	ldx.q r2, r22, r63 ! prefetch 4 lines hence
1da177e4 LT	72	#endif
1da177e4 LT	73	2:
379a95d1 PM	74	bge/u r2, r7, tr3 ! skip alloco for last 2 lines
379a95d1 PM	75	alloco r2, 0x40 ! alloc destination line 2 lines ahead
1da177e4 LT	76	synco ! TAKum03020
1da177e4 LT	77	3:
379a95d1 PM	78	ldx.q r2, r60, r36
	79	ldx.q r2, r61, r37
	80	ldx.q r2, r62, r38
	81	ldx.q r2, r23, r39
	82	st.q r2, 0, r36
	83	st.q r2, 8, r37
	84	st.q r2, 16, r38
	85	st.q r2, 24, r39
	86	addi r2, 32, r2
	87	bgt/l r8, r2, tr1
1da177e4 LT	88
1da177e4 LT	89	blink tr0, r63 ! return