2006-12-03 15:42:59 +00:00
|
|
|
/*
|
|
|
|
* 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.
|
|
|
|
*
|
|
|
|
* Quick'n'dirty IP checksum ...
|
|
|
|
*
|
|
|
|
* Copyright (C) 1998, 1999 Ralf Baechle
|
|
|
|
* Copyright (C) 1999 Silicon Graphics, Inc.
|
2007-10-23 11:43:25 +00:00
|
|
|
* Copyright (C) 2007 Maciej W. Rozycki
|
2013-12-12 16:21:00 +00:00
|
|
|
* Copyright (C) 2014 Imagination Technologies Ltd.
|
2006-12-03 15:42:59 +00:00
|
|
|
*/
|
2006-12-12 16:22:06 +00:00
|
|
|
#include <linux/errno.h>
|
2006-12-03 15:42:59 +00:00
|
|
|
#include <asm/asm.h>
|
2006-12-12 16:22:06 +00:00
|
|
|
#include <asm/asm-offsets.h>
|
2016-11-07 11:14:13 +00:00
|
|
|
#include <asm/export.h>
|
2006-12-03 15:42:59 +00:00
|
|
|
#include <asm/regdef.h>
|
|
|
|
|
|
|
|
#ifdef CONFIG_64BIT
|
2006-12-07 16:04:31 +00:00
|
|
|
/*
|
|
|
|
* As we are sharing code base with the mips32 tree (which use the o32 ABI
|
|
|
|
* register definitions). We need to redefine the register definitions from
|
|
|
|
* the n64 ABI register naming to the o32 ABI register naming.
|
|
|
|
*/
|
|
|
|
#undef t0
|
|
|
|
#undef t1
|
|
|
|
#undef t2
|
|
|
|
#undef t3
|
|
|
|
#define t0 $8
|
|
|
|
#define t1 $9
|
|
|
|
#define t2 $10
|
|
|
|
#define t3 $11
|
|
|
|
#define t4 $12
|
|
|
|
#define t5 $13
|
|
|
|
#define t6 $14
|
|
|
|
#define t7 $15
|
2006-12-07 16:04:51 +00:00
|
|
|
|
|
|
|
#define USE_DOUBLE
|
2006-12-03 15:42:59 +00:00
|
|
|
#endif
|
|
|
|
|
2006-12-07 16:04:51 +00:00
|
|
|
#ifdef USE_DOUBLE
|
|
|
|
|
|
|
|
#define LOAD ld
|
2008-09-20 15:20:04 +00:00
|
|
|
#define LOAD32 lwu
|
2006-12-07 16:04:51 +00:00
|
|
|
#define ADD daddu
|
|
|
|
#define NBYTES 8
|
|
|
|
|
|
|
|
#else
|
|
|
|
|
|
|
|
#define LOAD lw
|
2008-09-20 15:20:04 +00:00
|
|
|
#define LOAD32 lw
|
2006-12-07 16:04:51 +00:00
|
|
|
#define ADD addu
|
|
|
|
#define NBYTES 4
|
|
|
|
|
|
|
|
#endif /* USE_DOUBLE */
|
|
|
|
|
|
|
|
#define UNIT(unit) ((unit)*NBYTES)
|
|
|
|
|
2006-12-03 15:42:59 +00:00
|
|
|
#define ADDC(sum,reg) \
|
2014-04-04 02:32:54 +00:00
|
|
|
.set push; \
|
|
|
|
.set noat; \
|
2006-12-07 16:04:51 +00:00
|
|
|
ADD sum, reg; \
|
2006-12-03 15:42:59 +00:00
|
|
|
sltu v1, sum, reg; \
|
2007-10-23 11:43:25 +00:00
|
|
|
ADD sum, v1; \
|
2014-04-04 02:32:54 +00:00
|
|
|
.set pop
|
2006-12-03 15:42:59 +00:00
|
|
|
|
2008-09-20 15:20:04 +00:00
|
|
|
#define ADDC32(sum,reg) \
|
2014-04-04 02:32:54 +00:00
|
|
|
.set push; \
|
|
|
|
.set noat; \
|
2008-09-20 15:20:04 +00:00
|
|
|
addu sum, reg; \
|
|
|
|
sltu v1, sum, reg; \
|
|
|
|
addu sum, v1; \
|
2014-04-04 02:32:54 +00:00
|
|
|
.set pop
|
2008-09-20 15:20:04 +00:00
|
|
|
|
2006-12-07 16:04:51 +00:00
|
|
|
#define CSUM_BIGCHUNK1(src, offset, sum, _t0, _t1, _t2, _t3) \
|
|
|
|
LOAD _t0, (offset + UNIT(0))(src); \
|
|
|
|
LOAD _t1, (offset + UNIT(1))(src); \
|
2013-01-22 11:59:30 +00:00
|
|
|
LOAD _t2, (offset + UNIT(2))(src); \
|
|
|
|
LOAD _t3, (offset + UNIT(3))(src); \
|
MIPS: csum_partial: Improve instruction parallelism.
Computing sum introduces true data dependency. This patch removes some
true data depdendencies, hence increases instruction level parallelism.
This patch brings up to 50% csum performance gain on Loongson 3a.
One example about how this patch works is in CSUM_BIGCHUNK1:
// ** original ** vs ** patch applied **
ADDC(sum, t0) ADDC(t0, t1)
ADDC(sum, t1) ADDC(t2, t3)
ADDC(sum, t2) ADDC(sum, t0)
ADDC(sum, t3) ADDC(sum, t2)
In the original implementation, each ADDC(sum, ...) depends on the sum
value updated by previous ADDC(as source operand).
With this patch applied, the first two ADDC operations are independent,
hence can be executed simultaneously if possible.
Another example is in the "copy and sum calculating chunk":
// ** original ** vs ** patch applied **
STORE(t0, UNIT(0) ... STORE(t0, UNIT(0) ...
ADDC(sum, t0) ADDC(t0, t1)
STORE(t1, UNIT(1) ... STORE(t1, UNIT(1) ...
ADDC(sum, t1) ADDC(sum, t0)
STORE(t2, UNIT(2) ... STORE(t2, UNIT(2) ...
ADDC(sum, t2) ADDC(t2, t3)
STORE(t3, UNIT(3) ... STORE(t3, UNIT(3) ...
ADDC(sum, t3) ADDC(sum, t2)
With this patch applied, ADDC and the **next next** ADDC are independent.
Signed-off-by: chenj <chenj@lemote.com>
Cc: linux-mips@linux-mips.org
Patchwork: https://patchwork.linux-mips.org/patch/9608/
Signed-off-by: Ralf Baechle <ralf@linux-mips.org>
2015-03-26 17:07:24 +00:00
|
|
|
ADDC(_t0, _t1); \
|
|
|
|
ADDC(_t2, _t3); \
|
2006-12-03 15:42:59 +00:00
|
|
|
ADDC(sum, _t0); \
|
MIPS: csum_partial: Improve instruction parallelism.
Computing sum introduces true data dependency. This patch removes some
true data depdendencies, hence increases instruction level parallelism.
This patch brings up to 50% csum performance gain on Loongson 3a.
One example about how this patch works is in CSUM_BIGCHUNK1:
// ** original ** vs ** patch applied **
ADDC(sum, t0) ADDC(t0, t1)
ADDC(sum, t1) ADDC(t2, t3)
ADDC(sum, t2) ADDC(sum, t0)
ADDC(sum, t3) ADDC(sum, t2)
In the original implementation, each ADDC(sum, ...) depends on the sum
value updated by previous ADDC(as source operand).
With this patch applied, the first two ADDC operations are independent,
hence can be executed simultaneously if possible.
Another example is in the "copy and sum calculating chunk":
// ** original ** vs ** patch applied **
STORE(t0, UNIT(0) ... STORE(t0, UNIT(0) ...
ADDC(sum, t0) ADDC(t0, t1)
STORE(t1, UNIT(1) ... STORE(t1, UNIT(1) ...
ADDC(sum, t1) ADDC(sum, t0)
STORE(t2, UNIT(2) ... STORE(t2, UNIT(2) ...
ADDC(sum, t2) ADDC(t2, t3)
STORE(t3, UNIT(3) ... STORE(t3, UNIT(3) ...
ADDC(sum, t3) ADDC(sum, t2)
With this patch applied, ADDC and the **next next** ADDC are independent.
Signed-off-by: chenj <chenj@lemote.com>
Cc: linux-mips@linux-mips.org
Patchwork: https://patchwork.linux-mips.org/patch/9608/
Signed-off-by: Ralf Baechle <ralf@linux-mips.org>
2015-03-26 17:07:24 +00:00
|
|
|
ADDC(sum, _t2)
|
2006-12-07 16:04:51 +00:00
|
|
|
|
|
|
|
#ifdef USE_DOUBLE
|
|
|
|
#define CSUM_BIGCHUNK(src, offset, sum, _t0, _t1, _t2, _t3) \
|
|
|
|
CSUM_BIGCHUNK1(src, offset, sum, _t0, _t1, _t2, _t3)
|
|
|
|
#else
|
|
|
|
#define CSUM_BIGCHUNK(src, offset, sum, _t0, _t1, _t2, _t3) \
|
|
|
|
CSUM_BIGCHUNK1(src, offset, sum, _t0, _t1, _t2, _t3); \
|
|
|
|
CSUM_BIGCHUNK1(src, offset + 0x10, sum, _t0, _t1, _t2, _t3)
|
|
|
|
#endif
|
2006-12-03 15:42:59 +00:00
|
|
|
|
|
|
|
/*
|
|
|
|
* a0: source address
|
|
|
|
* a1: length of the area to checksum
|
|
|
|
* a2: partial checksum
|
|
|
|
*/
|
|
|
|
|
|
|
|
#define src a0
|
|
|
|
#define sum v0
|
|
|
|
|
|
|
|
.text
|
|
|
|
.set noreorder
|
|
|
|
.align 5
|
|
|
|
LEAF(csum_partial)
|
2016-11-07 11:14:13 +00:00
|
|
|
EXPORT_SYMBOL(csum_partial)
|
2006-12-03 15:42:59 +00:00
|
|
|
move sum, zero
|
2006-12-07 16:04:31 +00:00
|
|
|
move t7, zero
|
2006-12-03 15:42:59 +00:00
|
|
|
|
|
|
|
sltiu t8, a1, 0x8
|
2008-01-29 10:14:59 +00:00
|
|
|
bnez t8, .Lsmall_csumcpy /* < 8 bytes to copy */
|
2006-12-07 16:04:31 +00:00
|
|
|
move t2, a1
|
2006-12-03 15:42:59 +00:00
|
|
|
|
2006-12-07 16:04:45 +00:00
|
|
|
andi t7, src, 0x1 /* odd buffer? */
|
2006-12-03 15:42:59 +00:00
|
|
|
|
2008-01-29 10:14:59 +00:00
|
|
|
.Lhword_align:
|
|
|
|
beqz t7, .Lword_align
|
2006-12-03 15:42:59 +00:00
|
|
|
andi t8, src, 0x2
|
|
|
|
|
2006-12-07 16:04:31 +00:00
|
|
|
lbu t0, (src)
|
2006-12-03 15:42:59 +00:00
|
|
|
LONG_SUBU a1, a1, 0x1
|
|
|
|
#ifdef __MIPSEL__
|
2006-12-07 16:04:31 +00:00
|
|
|
sll t0, t0, 8
|
2006-12-03 15:42:59 +00:00
|
|
|
#endif
|
2006-12-07 16:04:31 +00:00
|
|
|
ADDC(sum, t0)
|
2006-12-03 15:42:59 +00:00
|
|
|
PTR_ADDU src, src, 0x1
|
|
|
|
andi t8, src, 0x2
|
|
|
|
|
2008-01-29 10:14:59 +00:00
|
|
|
.Lword_align:
|
|
|
|
beqz t8, .Ldword_align
|
2006-12-03 15:42:59 +00:00
|
|
|
sltiu t8, a1, 56
|
|
|
|
|
2006-12-07 16:04:31 +00:00
|
|
|
lhu t0, (src)
|
2006-12-03 15:42:59 +00:00
|
|
|
LONG_SUBU a1, a1, 0x2
|
2006-12-07 16:04:31 +00:00
|
|
|
ADDC(sum, t0)
|
2006-12-03 15:42:59 +00:00
|
|
|
sltiu t8, a1, 56
|
|
|
|
PTR_ADDU src, src, 0x2
|
|
|
|
|
2008-01-29 10:14:59 +00:00
|
|
|
.Ldword_align:
|
|
|
|
bnez t8, .Ldo_end_words
|
2006-12-03 15:42:59 +00:00
|
|
|
move t8, a1
|
|
|
|
|
|
|
|
andi t8, src, 0x4
|
2008-01-29 10:14:59 +00:00
|
|
|
beqz t8, .Lqword_align
|
2006-12-03 15:42:59 +00:00
|
|
|
andi t8, src, 0x8
|
|
|
|
|
2008-09-20 15:20:04 +00:00
|
|
|
LOAD32 t0, 0x00(src)
|
2006-12-03 15:42:59 +00:00
|
|
|
LONG_SUBU a1, a1, 0x4
|
2006-12-07 16:04:31 +00:00
|
|
|
ADDC(sum, t0)
|
2006-12-03 15:42:59 +00:00
|
|
|
PTR_ADDU src, src, 0x4
|
|
|
|
andi t8, src, 0x8
|
|
|
|
|
2008-01-29 10:14:59 +00:00
|
|
|
.Lqword_align:
|
|
|
|
beqz t8, .Loword_align
|
2006-12-03 15:42:59 +00:00
|
|
|
andi t8, src, 0x10
|
|
|
|
|
2006-12-07 16:04:51 +00:00
|
|
|
#ifdef USE_DOUBLE
|
|
|
|
ld t0, 0x00(src)
|
|
|
|
LONG_SUBU a1, a1, 0x8
|
|
|
|
ADDC(sum, t0)
|
|
|
|
#else
|
2006-12-07 16:04:31 +00:00
|
|
|
lw t0, 0x00(src)
|
|
|
|
lw t1, 0x04(src)
|
2006-12-03 15:42:59 +00:00
|
|
|
LONG_SUBU a1, a1, 0x8
|
2006-12-07 16:04:31 +00:00
|
|
|
ADDC(sum, t0)
|
|
|
|
ADDC(sum, t1)
|
2006-12-07 16:04:51 +00:00
|
|
|
#endif
|
2006-12-03 15:42:59 +00:00
|
|
|
PTR_ADDU src, src, 0x8
|
|
|
|
andi t8, src, 0x10
|
|
|
|
|
2008-01-29 10:14:59 +00:00
|
|
|
.Loword_align:
|
|
|
|
beqz t8, .Lbegin_movement
|
2006-12-03 15:42:59 +00:00
|
|
|
LONG_SRL t8, a1, 0x7
|
|
|
|
|
2006-12-07 16:04:51 +00:00
|
|
|
#ifdef USE_DOUBLE
|
|
|
|
ld t0, 0x00(src)
|
|
|
|
ld t1, 0x08(src)
|
2006-12-07 16:04:31 +00:00
|
|
|
ADDC(sum, t0)
|
|
|
|
ADDC(sum, t1)
|
2006-12-07 16:04:51 +00:00
|
|
|
#else
|
|
|
|
CSUM_BIGCHUNK1(src, 0x00, sum, t0, t1, t3, t4)
|
|
|
|
#endif
|
2006-12-03 15:42:59 +00:00
|
|
|
LONG_SUBU a1, a1, 0x10
|
|
|
|
PTR_ADDU src, src, 0x10
|
|
|
|
LONG_SRL t8, a1, 0x7
|
|
|
|
|
2008-01-29 10:14:59 +00:00
|
|
|
.Lbegin_movement:
|
2006-12-03 15:42:59 +00:00
|
|
|
beqz t8, 1f
|
2006-12-07 16:04:31 +00:00
|
|
|
andi t2, a1, 0x40
|
2006-12-03 15:42:59 +00:00
|
|
|
|
2008-01-29 10:14:59 +00:00
|
|
|
.Lmove_128bytes:
|
2006-12-07 16:04:31 +00:00
|
|
|
CSUM_BIGCHUNK(src, 0x00, sum, t0, t1, t3, t4)
|
|
|
|
CSUM_BIGCHUNK(src, 0x20, sum, t0, t1, t3, t4)
|
|
|
|
CSUM_BIGCHUNK(src, 0x40, sum, t0, t1, t3, t4)
|
|
|
|
CSUM_BIGCHUNK(src, 0x60, sum, t0, t1, t3, t4)
|
2006-12-03 15:42:59 +00:00
|
|
|
LONG_SUBU t8, t8, 0x01
|
2007-10-23 11:43:25 +00:00
|
|
|
.set reorder /* DADDI_WAR */
|
|
|
|
PTR_ADDU src, src, 0x80
|
2008-01-29 10:14:59 +00:00
|
|
|
bnez t8, .Lmove_128bytes
|
2007-10-23 11:43:25 +00:00
|
|
|
.set noreorder
|
2006-12-03 15:42:59 +00:00
|
|
|
|
|
|
|
1:
|
2006-12-07 16:04:31 +00:00
|
|
|
beqz t2, 1f
|
|
|
|
andi t2, a1, 0x20
|
2006-12-03 15:42:59 +00:00
|
|
|
|
2008-01-29 10:14:59 +00:00
|
|
|
.Lmove_64bytes:
|
2006-12-07 16:04:31 +00:00
|
|
|
CSUM_BIGCHUNK(src, 0x00, sum, t0, t1, t3, t4)
|
|
|
|
CSUM_BIGCHUNK(src, 0x20, sum, t0, t1, t3, t4)
|
2006-12-03 15:42:59 +00:00
|
|
|
PTR_ADDU src, src, 0x40
|
|
|
|
|
|
|
|
1:
|
2008-01-29 10:14:59 +00:00
|
|
|
beqz t2, .Ldo_end_words
|
2006-12-03 15:42:59 +00:00
|
|
|
andi t8, a1, 0x1c
|
|
|
|
|
2008-01-29 10:14:59 +00:00
|
|
|
.Lmove_32bytes:
|
2006-12-07 16:04:31 +00:00
|
|
|
CSUM_BIGCHUNK(src, 0x00, sum, t0, t1, t3, t4)
|
2006-12-03 15:42:59 +00:00
|
|
|
andi t8, a1, 0x1c
|
|
|
|
PTR_ADDU src, src, 0x20
|
|
|
|
|
2008-01-29 10:14:59 +00:00
|
|
|
.Ldo_end_words:
|
|
|
|
beqz t8, .Lsmall_csumcpy
|
2006-12-07 16:04:45 +00:00
|
|
|
andi t2, a1, 0x3
|
|
|
|
LONG_SRL t8, t8, 0x2
|
2006-12-03 15:42:59 +00:00
|
|
|
|
2008-01-29 10:14:59 +00:00
|
|
|
.Lend_words:
|
2008-09-20 15:20:04 +00:00
|
|
|
LOAD32 t0, (src)
|
2006-12-03 15:42:59 +00:00
|
|
|
LONG_SUBU t8, t8, 0x1
|
2006-12-07 16:04:31 +00:00
|
|
|
ADDC(sum, t0)
|
2007-10-23 11:43:25 +00:00
|
|
|
.set reorder /* DADDI_WAR */
|
|
|
|
PTR_ADDU src, src, 0x4
|
2008-01-29 10:14:59 +00:00
|
|
|
bnez t8, .Lend_words
|
2007-10-23 11:43:25 +00:00
|
|
|
.set noreorder
|
2006-12-03 15:42:59 +00:00
|
|
|
|
2006-12-07 16:04:45 +00:00
|
|
|
/* unknown src alignment and < 8 bytes to go */
|
2008-01-29 10:14:59 +00:00
|
|
|
.Lsmall_csumcpy:
|
2006-12-07 16:04:45 +00:00
|
|
|
move a1, t2
|
2006-12-03 15:42:59 +00:00
|
|
|
|
2006-12-07 16:04:45 +00:00
|
|
|
andi t0, a1, 4
|
|
|
|
beqz t0, 1f
|
|
|
|
andi t0, a1, 2
|
2006-12-03 15:42:59 +00:00
|
|
|
|
2006-12-07 16:04:45 +00:00
|
|
|
/* Still a full word to go */
|
|
|
|
ulw t1, (src)
|
|
|
|
PTR_ADDIU src, 4
|
2008-09-20 15:20:04 +00:00
|
|
|
#ifdef USE_DOUBLE
|
|
|
|
dsll t1, t1, 32 /* clear lower 32bit */
|
|
|
|
#endif
|
2006-12-07 16:04:45 +00:00
|
|
|
ADDC(sum, t1)
|
|
|
|
|
|
|
|
1: move t1, zero
|
|
|
|
beqz t0, 1f
|
|
|
|
andi t0, a1, 1
|
|
|
|
|
|
|
|
/* Still a halfword to go */
|
|
|
|
ulhu t1, (src)
|
|
|
|
PTR_ADDIU src, 2
|
|
|
|
|
|
|
|
1: beqz t0, 1f
|
|
|
|
sll t1, t1, 16
|
|
|
|
|
|
|
|
lbu t2, (src)
|
|
|
|
nop
|
|
|
|
|
|
|
|
#ifdef __MIPSEB__
|
|
|
|
sll t2, t2, 8
|
|
|
|
#endif
|
|
|
|
or t1, t2
|
|
|
|
|
|
|
|
1: ADDC(sum, t1)
|
2006-12-03 15:42:59 +00:00
|
|
|
|
2006-12-07 16:04:45 +00:00
|
|
|
/* fold checksum */
|
2006-12-07 16:04:51 +00:00
|
|
|
#ifdef USE_DOUBLE
|
|
|
|
dsll32 v1, sum, 0
|
|
|
|
daddu sum, v1
|
|
|
|
sltu v1, sum, v1
|
|
|
|
dsra32 sum, sum, 0
|
|
|
|
addu sum, v1
|
|
|
|
#endif
|
2006-12-07 16:04:45 +00:00
|
|
|
|
|
|
|
/* odd buffer alignment? */
|
mips: Add MIPS Release 5 support
There are five MIPS32/64 architecture releases currently available:
from 1 to 6 except fourth one, which was intentionally skipped.
Three of them can be called as major: 1st, 2nd and 6th, that not only
have some system level alterations, but also introduced significant
core/ISA level updates. The rest of the MIPS architecture releases are
minor.
Even though they don't have as much ISA/system/core level changes
as the major ones with respect to the previous releases, they still
provide a set of updates (I'd say they were intended to be the
intermediate releases before a major one) that might be useful for the
kernel and user-level code, when activated by the kernel or compiler.
In particular the following features were introduced or ended up being
available at/after MIPS32/64 Release 5 architecture:
+ the last release of the misaligned memory access instructions,
+ virtualisation - VZ ASE - is optional component of the arch,
+ SIMD - MSA ASE - is optional component of the arch,
+ DSP ASE is optional component of the arch,
+ CP0.Status.FR=1 for CP1.FIR.F64=1 (pure 64-bit FPU general registers)
must be available if FPU is implemented,
+ CP1.FIR.Has2008 support is required so CP1.FCSR.{ABS2008,NAN2008} bits
are available.
+ UFR/UNFR aliases to access CP0.Status.FR from user-space by means of
ctc1/cfc1 instructions (enabled by CP0.Config5.UFR),
+ CP0.COnfig5.LLB=1 and eretnc instruction are implemented to without
accidentally clearing LL-bit when returning from an interrupt,
exception, or error trap,
+ XPA feature together with extended versions of CPx registers is
introduced, which needs to have mfhc0/mthc0 instructions available.
So due to these changes GNU GCC provides an extended instructions set
support for MIPS32/64 Release 5 by default like eretnc/mfhc0/mthc0. Even
though the architecture alteration isn't that big, it still worth to be
taken into account by the kernel software. Finally we can't deny that
some optimization/limitations might be found in future and implemented
on some level in kernel or compiler. In this case having even
intermediate MIPS architecture releases support would be more than
useful.
So the most of the changes provided by this commit can be split into
either compile- or runtime configs related. The compile-time related
changes are caused by adding the new CONFIG_CPU_MIPS32_R5/CONFIG_CPU_MIPSR5
configs and concern the code activating MIPSR2 or MIPSR6 already
implemented features (like eretnc/LLbit, mthc0/mfhc0). In addition
CPU_HAS_MSA can be now freely enabled for MIPS32/64 release 5 based
platforms as this is done for CPU_MIPS32_R6 CPUs. The runtime changes
concerns the features which are handled with respect to the MIPS ISA
revision detected at run-time by means of CP0.Config.{AT,AR} bits. Alas
these fields can be used to detect either r1 or r2 or r6 releases.
But since we know which CPUs in fact support the R5 arch, we can manually
set MIPS_CPU_ISA_M32R5/MIPS_CPU_ISA_M64R5 bit of c->isa_level and then
use cpu_has_mips32r5/cpu_has_mips64r5 where it's appropriate.
Since XPA/EVA provide too complex alterationss and to have them used with
MIPS32 Release 2 charged kernels (for compatibility with current platform
configs) they are left to be setup as a separate kernel configs.
Co-developed-by: Alexey Malahov <Alexey.Malahov@baikalelectronics.ru>
Signed-off-by: Alexey Malahov <Alexey.Malahov@baikalelectronics.ru>
Signed-off-by: Serge Semin <Sergey.Semin@baikalelectronics.ru>
Cc: Thomas Bogendoerfer <tsbogend@alpha.franken.de>
Cc: Paul Burton <paulburton@kernel.org>
Cc: Ralf Baechle <ralf@linux-mips.org>
Cc: Arnd Bergmann <arnd@arndb.de>
Cc: Rob Herring <robh+dt@kernel.org>
Cc: devicetree@vger.kernel.org
Signed-off-by: Thomas Bogendoerfer <tsbogend@alpha.franken.de>
2020-05-21 14:07:14 +00:00
|
|
|
#if defined(CONFIG_CPU_MIPSR2) || defined(CONFIG_CPU_MIPSR5) || \
|
|
|
|
defined(CONFIG_CPU_LOONGSON64)
|
2014-08-15 08:56:58 +00:00
|
|
|
.set push
|
|
|
|
.set arch=mips32r2
|
2008-10-11 15:18:53 +00:00
|
|
|
wsbh v1, sum
|
|
|
|
movn sum, v1, t7
|
2014-08-15 08:56:58 +00:00
|
|
|
.set pop
|
2008-10-11 15:18:53 +00:00
|
|
|
#else
|
|
|
|
beqz t7, 1f /* odd buffer alignment? */
|
|
|
|
lui v1, 0x00ff
|
|
|
|
addu v1, 0x00ff
|
|
|
|
and t0, sum, v1
|
|
|
|
sll t0, t0, 8
|
2006-12-07 16:04:45 +00:00
|
|
|
srl sum, sum, 8
|
2008-10-11 15:18:53 +00:00
|
|
|
and sum, sum, v1
|
|
|
|
or sum, sum, t0
|
2006-12-07 16:04:45 +00:00
|
|
|
1:
|
2008-10-11 15:18:53 +00:00
|
|
|
#endif
|
2006-12-07 16:04:45 +00:00
|
|
|
.set reorder
|
2013-01-22 11:59:30 +00:00
|
|
|
/* Add the passed partial csum. */
|
2008-09-20 15:20:04 +00:00
|
|
|
ADDC32(sum, a2)
|
2006-12-03 15:42:59 +00:00
|
|
|
jr ra
|
2006-12-07 16:04:45 +00:00
|
|
|
.set noreorder
|
2006-12-03 15:42:59 +00:00
|
|
|
END(csum_partial)
|
2006-12-12 16:22:06 +00:00
|
|
|
|
|
|
|
|
|
|
|
/*
|
|
|
|
* checksum and copy routines based on memcpy.S
|
|
|
|
*
|
2020-07-19 21:37:15 +00:00
|
|
|
* csum_partial_copy_nocheck(src, dst, len)
|
|
|
|
* __csum_partial_copy_kernel(src, dst, len)
|
2006-12-12 16:22:06 +00:00
|
|
|
*
|
2013-01-22 11:59:30 +00:00
|
|
|
* See "Spec" in memcpy.S for details. Unlike __copy_user, all
|
2006-12-12 16:22:06 +00:00
|
|
|
* function in this file use the standard calling convention.
|
|
|
|
*/
|
|
|
|
|
|
|
|
#define src a0
|
|
|
|
#define dst a1
|
|
|
|
#define len a2
|
|
|
|
#define sum v0
|
|
|
|
#define odd t8
|
|
|
|
|
|
|
|
/*
|
2020-07-19 21:37:15 +00:00
|
|
|
* All exception handlers simply return 0.
|
2006-12-12 16:22:06 +00:00
|
|
|
*/
|
|
|
|
|
2014-01-16 17:02:13 +00:00
|
|
|
/* Instruction type */
|
|
|
|
#define LD_INSN 1
|
|
|
|
#define ST_INSN 2
|
2014-01-17 10:48:46 +00:00
|
|
|
#define LEGACY_MODE 1
|
|
|
|
#define EVA_MODE 2
|
|
|
|
#define USEROP 1
|
|
|
|
#define KERNELOP 2
|
2014-01-16 17:02:13 +00:00
|
|
|
|
|
|
|
/*
|
|
|
|
* Wrapper to add an entry in the exception table
|
|
|
|
* in case the insn causes a memory exception.
|
|
|
|
* Arguments:
|
|
|
|
* insn : Load/store instruction
|
|
|
|
* type : Instruction type
|
|
|
|
* reg : Register
|
|
|
|
* addr : Address
|
|
|
|
* handler : Exception handler
|
|
|
|
*/
|
2020-07-19 21:37:15 +00:00
|
|
|
#define EXC(insn, type, reg, addr) \
|
2014-01-17 10:48:46 +00:00
|
|
|
.if \mode == LEGACY_MODE; \
|
|
|
|
9: insn reg, addr; \
|
|
|
|
.section __ex_table,"a"; \
|
2020-07-19 21:37:15 +00:00
|
|
|
PTR 9b, .L_exc; \
|
2014-01-17 10:48:46 +00:00
|
|
|
.previous; \
|
2014-01-17 11:36:16 +00:00
|
|
|
/* This is enabled in EVA mode */ \
|
|
|
|
.else; \
|
|
|
|
/* If loading from user or storing to user */ \
|
|
|
|
.if ((\from == USEROP) && (type == LD_INSN)) || \
|
|
|
|
((\to == USEROP) && (type == ST_INSN)); \
|
|
|
|
9: __BUILD_EVA_INSN(insn##e, reg, addr); \
|
|
|
|
.section __ex_table,"a"; \
|
2020-07-19 21:37:15 +00:00
|
|
|
PTR 9b, .L_exc; \
|
2014-01-17 11:36:16 +00:00
|
|
|
.previous; \
|
|
|
|
.else; \
|
|
|
|
/* EVA without exception */ \
|
|
|
|
insn reg, addr; \
|
|
|
|
.endif; \
|
2014-01-17 10:48:46 +00:00
|
|
|
.endif
|
2006-12-12 16:22:06 +00:00
|
|
|
|
2014-01-16 17:02:13 +00:00
|
|
|
#undef LOAD
|
|
|
|
|
2006-12-12 16:22:06 +00:00
|
|
|
#ifdef USE_DOUBLE
|
|
|
|
|
2014-01-16 17:02:13 +00:00
|
|
|
#define LOADK ld /* No exception */
|
2020-07-19 21:37:15 +00:00
|
|
|
#define LOAD(reg, addr) EXC(ld, LD_INSN, reg, addr)
|
|
|
|
#define LOADBU(reg, addr) EXC(lbu, LD_INSN, reg, addr)
|
|
|
|
#define LOADL(reg, addr) EXC(ldl, LD_INSN, reg, addr)
|
|
|
|
#define LOADR(reg, addr) EXC(ldr, LD_INSN, reg, addr)
|
|
|
|
#define STOREB(reg, addr) EXC(sb, ST_INSN, reg, addr)
|
|
|
|
#define STOREL(reg, addr) EXC(sdl, ST_INSN, reg, addr)
|
|
|
|
#define STORER(reg, addr) EXC(sdr, ST_INSN, reg, addr)
|
|
|
|
#define STORE(reg, addr) EXC(sd, ST_INSN, reg, addr)
|
2006-12-12 16:22:06 +00:00
|
|
|
#define ADD daddu
|
|
|
|
#define SUB dsubu
|
|
|
|
#define SRL dsrl
|
|
|
|
#define SLL dsll
|
|
|
|
#define SLLV dsllv
|
|
|
|
#define SRLV dsrlv
|
|
|
|
#define NBYTES 8
|
|
|
|
#define LOG_NBYTES 3
|
|
|
|
|
|
|
|
#else
|
|
|
|
|
2014-01-16 17:02:13 +00:00
|
|
|
#define LOADK lw /* No exception */
|
2020-07-19 21:37:15 +00:00
|
|
|
#define LOAD(reg, addr) EXC(lw, LD_INSN, reg, addr)
|
|
|
|
#define LOADBU(reg, addr) EXC(lbu, LD_INSN, reg, addr)
|
|
|
|
#define LOADL(reg, addr) EXC(lwl, LD_INSN, reg, addr)
|
|
|
|
#define LOADR(reg, addr) EXC(lwr, LD_INSN, reg, addr)
|
|
|
|
#define STOREB(reg, addr) EXC(sb, ST_INSN, reg, addr)
|
|
|
|
#define STOREL(reg, addr) EXC(swl, ST_INSN, reg, addr)
|
|
|
|
#define STORER(reg, addr) EXC(swr, ST_INSN, reg, addr)
|
|
|
|
#define STORE(reg, addr) EXC(sw, ST_INSN, reg, addr)
|
2006-12-12 16:22:06 +00:00
|
|
|
#define ADD addu
|
|
|
|
#define SUB subu
|
|
|
|
#define SRL srl
|
|
|
|
#define SLL sll
|
|
|
|
#define SLLV sllv
|
|
|
|
#define SRLV srlv
|
|
|
|
#define NBYTES 4
|
|
|
|
#define LOG_NBYTES 2
|
|
|
|
|
|
|
|
#endif /* USE_DOUBLE */
|
|
|
|
|
|
|
|
#ifdef CONFIG_CPU_LITTLE_ENDIAN
|
|
|
|
#define LDFIRST LOADR
|
2013-01-22 11:59:30 +00:00
|
|
|
#define LDREST LOADL
|
2006-12-12 16:22:06 +00:00
|
|
|
#define STFIRST STORER
|
2013-01-22 11:59:30 +00:00
|
|
|
#define STREST STOREL
|
2006-12-12 16:22:06 +00:00
|
|
|
#define SHIFT_DISCARD SLLV
|
|
|
|
#define SHIFT_DISCARD_REVERT SRLV
|
|
|
|
#else
|
|
|
|
#define LDFIRST LOADL
|
2013-01-22 11:59:30 +00:00
|
|
|
#define LDREST LOADR
|
2006-12-12 16:22:06 +00:00
|
|
|
#define STFIRST STOREL
|
2013-01-22 11:59:30 +00:00
|
|
|
#define STREST STORER
|
2006-12-12 16:22:06 +00:00
|
|
|
#define SHIFT_DISCARD SRLV
|
|
|
|
#define SHIFT_DISCARD_REVERT SLLV
|
|
|
|
#endif
|
|
|
|
|
|
|
|
#define FIRST(unit) ((unit)*NBYTES)
|
|
|
|
#define REST(unit) (FIRST(unit)+NBYTES-1)
|
|
|
|
|
|
|
|
#define ADDRMASK (NBYTES-1)
|
|
|
|
|
2007-10-23 11:43:25 +00:00
|
|
|
#ifndef CONFIG_CPU_DADDI_WORKAROUNDS
|
2006-12-12 16:22:06 +00:00
|
|
|
.set noat
|
2007-10-23 11:43:25 +00:00
|
|
|
#else
|
|
|
|
.set at=v1
|
|
|
|
#endif
|
2006-12-12 16:22:06 +00:00
|
|
|
|
2020-07-19 21:37:15 +00:00
|
|
|
.macro __BUILD_CSUM_PARTIAL_COPY_USER mode, from, to
|
2014-01-17 10:48:46 +00:00
|
|
|
|
2020-07-19 21:37:15 +00:00
|
|
|
li sum, -1
|
2006-12-12 16:22:06 +00:00
|
|
|
move odd, zero
|
|
|
|
/*
|
|
|
|
* Note: dst & src may be unaligned, len may be 0
|
|
|
|
* Temps
|
|
|
|
*/
|
|
|
|
/*
|
|
|
|
* The "issue break"s below are very approximate.
|
|
|
|
* Issue delays for dcache fills will perturb the schedule, as will
|
|
|
|
* load queue full replay traps, etc.
|
|
|
|
*
|
|
|
|
* If len < NBYTES use byte operations.
|
|
|
|
*/
|
|
|
|
sltu t2, len, NBYTES
|
|
|
|
and t1, dst, ADDRMASK
|
2014-01-17 10:48:46 +00:00
|
|
|
bnez t2, .Lcopy_bytes_checklen\@
|
2006-12-12 16:22:06 +00:00
|
|
|
and t0, src, ADDRMASK
|
|
|
|
andi odd, dst, 0x1 /* odd buffer? */
|
2014-01-17 10:48:46 +00:00
|
|
|
bnez t1, .Ldst_unaligned\@
|
2006-12-12 16:22:06 +00:00
|
|
|
nop
|
2014-01-17 10:48:46 +00:00
|
|
|
bnez t0, .Lsrc_unaligned_dst_aligned\@
|
2006-12-12 16:22:06 +00:00
|
|
|
/*
|
|
|
|
* use delay slot for fall-through
|
|
|
|
* src and dst are aligned; need to compute rem
|
|
|
|
*/
|
2014-01-17 10:48:46 +00:00
|
|
|
.Lboth_aligned\@:
|
2013-01-22 11:59:30 +00:00
|
|
|
SRL t0, len, LOG_NBYTES+3 # +3 for 8 units/iter
|
2014-01-17 10:48:46 +00:00
|
|
|
beqz t0, .Lcleanup_both_aligned\@ # len < 8*NBYTES
|
2006-12-12 16:22:06 +00:00
|
|
|
nop
|
|
|
|
SUB len, 8*NBYTES # subtract here for bgez loop
|
|
|
|
.align 4
|
|
|
|
1:
|
2020-07-19 21:37:15 +00:00
|
|
|
LOAD(t0, UNIT(0)(src))
|
|
|
|
LOAD(t1, UNIT(1)(src))
|
|
|
|
LOAD(t2, UNIT(2)(src))
|
|
|
|
LOAD(t3, UNIT(3)(src))
|
|
|
|
LOAD(t4, UNIT(4)(src))
|
|
|
|
LOAD(t5, UNIT(5)(src))
|
|
|
|
LOAD(t6, UNIT(6)(src))
|
|
|
|
LOAD(t7, UNIT(7)(src))
|
2006-12-12 16:22:06 +00:00
|
|
|
SUB len, len, 8*NBYTES
|
|
|
|
ADD src, src, 8*NBYTES
|
2020-07-19 21:37:15 +00:00
|
|
|
STORE(t0, UNIT(0)(dst))
|
MIPS: csum_partial: Improve instruction parallelism.
Computing sum introduces true data dependency. This patch removes some
true data depdendencies, hence increases instruction level parallelism.
This patch brings up to 50% csum performance gain on Loongson 3a.
One example about how this patch works is in CSUM_BIGCHUNK1:
// ** original ** vs ** patch applied **
ADDC(sum, t0) ADDC(t0, t1)
ADDC(sum, t1) ADDC(t2, t3)
ADDC(sum, t2) ADDC(sum, t0)
ADDC(sum, t3) ADDC(sum, t2)
In the original implementation, each ADDC(sum, ...) depends on the sum
value updated by previous ADDC(as source operand).
With this patch applied, the first two ADDC operations are independent,
hence can be executed simultaneously if possible.
Another example is in the "copy and sum calculating chunk":
// ** original ** vs ** patch applied **
STORE(t0, UNIT(0) ... STORE(t0, UNIT(0) ...
ADDC(sum, t0) ADDC(t0, t1)
STORE(t1, UNIT(1) ... STORE(t1, UNIT(1) ...
ADDC(sum, t1) ADDC(sum, t0)
STORE(t2, UNIT(2) ... STORE(t2, UNIT(2) ...
ADDC(sum, t2) ADDC(t2, t3)
STORE(t3, UNIT(3) ... STORE(t3, UNIT(3) ...
ADDC(sum, t3) ADDC(sum, t2)
With this patch applied, ADDC and the **next next** ADDC are independent.
Signed-off-by: chenj <chenj@lemote.com>
Cc: linux-mips@linux-mips.org
Patchwork: https://patchwork.linux-mips.org/patch/9608/
Signed-off-by: Ralf Baechle <ralf@linux-mips.org>
2015-03-26 17:07:24 +00:00
|
|
|
ADDC(t0, t1)
|
2020-07-19 21:37:15 +00:00
|
|
|
STORE(t1, UNIT(1)(dst))
|
MIPS: csum_partial: Improve instruction parallelism.
Computing sum introduces true data dependency. This patch removes some
true data depdendencies, hence increases instruction level parallelism.
This patch brings up to 50% csum performance gain on Loongson 3a.
One example about how this patch works is in CSUM_BIGCHUNK1:
// ** original ** vs ** patch applied **
ADDC(sum, t0) ADDC(t0, t1)
ADDC(sum, t1) ADDC(t2, t3)
ADDC(sum, t2) ADDC(sum, t0)
ADDC(sum, t3) ADDC(sum, t2)
In the original implementation, each ADDC(sum, ...) depends on the sum
value updated by previous ADDC(as source operand).
With this patch applied, the first two ADDC operations are independent,
hence can be executed simultaneously if possible.
Another example is in the "copy and sum calculating chunk":
// ** original ** vs ** patch applied **
STORE(t0, UNIT(0) ... STORE(t0, UNIT(0) ...
ADDC(sum, t0) ADDC(t0, t1)
STORE(t1, UNIT(1) ... STORE(t1, UNIT(1) ...
ADDC(sum, t1) ADDC(sum, t0)
STORE(t2, UNIT(2) ... STORE(t2, UNIT(2) ...
ADDC(sum, t2) ADDC(t2, t3)
STORE(t3, UNIT(3) ... STORE(t3, UNIT(3) ...
ADDC(sum, t3) ADDC(sum, t2)
With this patch applied, ADDC and the **next next** ADDC are independent.
Signed-off-by: chenj <chenj@lemote.com>
Cc: linux-mips@linux-mips.org
Patchwork: https://patchwork.linux-mips.org/patch/9608/
Signed-off-by: Ralf Baechle <ralf@linux-mips.org>
2015-03-26 17:07:24 +00:00
|
|
|
ADDC(sum, t0)
|
2020-07-19 21:37:15 +00:00
|
|
|
STORE(t2, UNIT(2)(dst))
|
MIPS: csum_partial: Improve instruction parallelism.
Computing sum introduces true data dependency. This patch removes some
true data depdendencies, hence increases instruction level parallelism.
This patch brings up to 50% csum performance gain on Loongson 3a.
One example about how this patch works is in CSUM_BIGCHUNK1:
// ** original ** vs ** patch applied **
ADDC(sum, t0) ADDC(t0, t1)
ADDC(sum, t1) ADDC(t2, t3)
ADDC(sum, t2) ADDC(sum, t0)
ADDC(sum, t3) ADDC(sum, t2)
In the original implementation, each ADDC(sum, ...) depends on the sum
value updated by previous ADDC(as source operand).
With this patch applied, the first two ADDC operations are independent,
hence can be executed simultaneously if possible.
Another example is in the "copy and sum calculating chunk":
// ** original ** vs ** patch applied **
STORE(t0, UNIT(0) ... STORE(t0, UNIT(0) ...
ADDC(sum, t0) ADDC(t0, t1)
STORE(t1, UNIT(1) ... STORE(t1, UNIT(1) ...
ADDC(sum, t1) ADDC(sum, t0)
STORE(t2, UNIT(2) ... STORE(t2, UNIT(2) ...
ADDC(sum, t2) ADDC(t2, t3)
STORE(t3, UNIT(3) ... STORE(t3, UNIT(3) ...
ADDC(sum, t3) ADDC(sum, t2)
With this patch applied, ADDC and the **next next** ADDC are independent.
Signed-off-by: chenj <chenj@lemote.com>
Cc: linux-mips@linux-mips.org
Patchwork: https://patchwork.linux-mips.org/patch/9608/
Signed-off-by: Ralf Baechle <ralf@linux-mips.org>
2015-03-26 17:07:24 +00:00
|
|
|
ADDC(t2, t3)
|
2020-07-19 21:37:15 +00:00
|
|
|
STORE(t3, UNIT(3)(dst))
|
MIPS: csum_partial: Improve instruction parallelism.
Computing sum introduces true data dependency. This patch removes some
true data depdendencies, hence increases instruction level parallelism.
This patch brings up to 50% csum performance gain on Loongson 3a.
One example about how this patch works is in CSUM_BIGCHUNK1:
// ** original ** vs ** patch applied **
ADDC(sum, t0) ADDC(t0, t1)
ADDC(sum, t1) ADDC(t2, t3)
ADDC(sum, t2) ADDC(sum, t0)
ADDC(sum, t3) ADDC(sum, t2)
In the original implementation, each ADDC(sum, ...) depends on the sum
value updated by previous ADDC(as source operand).
With this patch applied, the first two ADDC operations are independent,
hence can be executed simultaneously if possible.
Another example is in the "copy and sum calculating chunk":
// ** original ** vs ** patch applied **
STORE(t0, UNIT(0) ... STORE(t0, UNIT(0) ...
ADDC(sum, t0) ADDC(t0, t1)
STORE(t1, UNIT(1) ... STORE(t1, UNIT(1) ...
ADDC(sum, t1) ADDC(sum, t0)
STORE(t2, UNIT(2) ... STORE(t2, UNIT(2) ...
ADDC(sum, t2) ADDC(t2, t3)
STORE(t3, UNIT(3) ... STORE(t3, UNIT(3) ...
ADDC(sum, t3) ADDC(sum, t2)
With this patch applied, ADDC and the **next next** ADDC are independent.
Signed-off-by: chenj <chenj@lemote.com>
Cc: linux-mips@linux-mips.org
Patchwork: https://patchwork.linux-mips.org/patch/9608/
Signed-off-by: Ralf Baechle <ralf@linux-mips.org>
2015-03-26 17:07:24 +00:00
|
|
|
ADDC(sum, t2)
|
2020-07-19 21:37:15 +00:00
|
|
|
STORE(t4, UNIT(4)(dst))
|
MIPS: csum_partial: Improve instruction parallelism.
Computing sum introduces true data dependency. This patch removes some
true data depdendencies, hence increases instruction level parallelism.
This patch brings up to 50% csum performance gain on Loongson 3a.
One example about how this patch works is in CSUM_BIGCHUNK1:
// ** original ** vs ** patch applied **
ADDC(sum, t0) ADDC(t0, t1)
ADDC(sum, t1) ADDC(t2, t3)
ADDC(sum, t2) ADDC(sum, t0)
ADDC(sum, t3) ADDC(sum, t2)
In the original implementation, each ADDC(sum, ...) depends on the sum
value updated by previous ADDC(as source operand).
With this patch applied, the first two ADDC operations are independent,
hence can be executed simultaneously if possible.
Another example is in the "copy and sum calculating chunk":
// ** original ** vs ** patch applied **
STORE(t0, UNIT(0) ... STORE(t0, UNIT(0) ...
ADDC(sum, t0) ADDC(t0, t1)
STORE(t1, UNIT(1) ... STORE(t1, UNIT(1) ...
ADDC(sum, t1) ADDC(sum, t0)
STORE(t2, UNIT(2) ... STORE(t2, UNIT(2) ...
ADDC(sum, t2) ADDC(t2, t3)
STORE(t3, UNIT(3) ... STORE(t3, UNIT(3) ...
ADDC(sum, t3) ADDC(sum, t2)
With this patch applied, ADDC and the **next next** ADDC are independent.
Signed-off-by: chenj <chenj@lemote.com>
Cc: linux-mips@linux-mips.org
Patchwork: https://patchwork.linux-mips.org/patch/9608/
Signed-off-by: Ralf Baechle <ralf@linux-mips.org>
2015-03-26 17:07:24 +00:00
|
|
|
ADDC(t4, t5)
|
2020-07-19 21:37:15 +00:00
|
|
|
STORE(t5, UNIT(5)(dst))
|
MIPS: csum_partial: Improve instruction parallelism.
Computing sum introduces true data dependency. This patch removes some
true data depdendencies, hence increases instruction level parallelism.
This patch brings up to 50% csum performance gain on Loongson 3a.
One example about how this patch works is in CSUM_BIGCHUNK1:
// ** original ** vs ** patch applied **
ADDC(sum, t0) ADDC(t0, t1)
ADDC(sum, t1) ADDC(t2, t3)
ADDC(sum, t2) ADDC(sum, t0)
ADDC(sum, t3) ADDC(sum, t2)
In the original implementation, each ADDC(sum, ...) depends on the sum
value updated by previous ADDC(as source operand).
With this patch applied, the first two ADDC operations are independent,
hence can be executed simultaneously if possible.
Another example is in the "copy and sum calculating chunk":
// ** original ** vs ** patch applied **
STORE(t0, UNIT(0) ... STORE(t0, UNIT(0) ...
ADDC(sum, t0) ADDC(t0, t1)
STORE(t1, UNIT(1) ... STORE(t1, UNIT(1) ...
ADDC(sum, t1) ADDC(sum, t0)
STORE(t2, UNIT(2) ... STORE(t2, UNIT(2) ...
ADDC(sum, t2) ADDC(t2, t3)
STORE(t3, UNIT(3) ... STORE(t3, UNIT(3) ...
ADDC(sum, t3) ADDC(sum, t2)
With this patch applied, ADDC and the **next next** ADDC are independent.
Signed-off-by: chenj <chenj@lemote.com>
Cc: linux-mips@linux-mips.org
Patchwork: https://patchwork.linux-mips.org/patch/9608/
Signed-off-by: Ralf Baechle <ralf@linux-mips.org>
2015-03-26 17:07:24 +00:00
|
|
|
ADDC(sum, t4)
|
2020-07-19 21:37:15 +00:00
|
|
|
STORE(t6, UNIT(6)(dst))
|
MIPS: csum_partial: Improve instruction parallelism.
Computing sum introduces true data dependency. This patch removes some
true data depdendencies, hence increases instruction level parallelism.
This patch brings up to 50% csum performance gain on Loongson 3a.
One example about how this patch works is in CSUM_BIGCHUNK1:
// ** original ** vs ** patch applied **
ADDC(sum, t0) ADDC(t0, t1)
ADDC(sum, t1) ADDC(t2, t3)
ADDC(sum, t2) ADDC(sum, t0)
ADDC(sum, t3) ADDC(sum, t2)
In the original implementation, each ADDC(sum, ...) depends on the sum
value updated by previous ADDC(as source operand).
With this patch applied, the first two ADDC operations are independent,
hence can be executed simultaneously if possible.
Another example is in the "copy and sum calculating chunk":
// ** original ** vs ** patch applied **
STORE(t0, UNIT(0) ... STORE(t0, UNIT(0) ...
ADDC(sum, t0) ADDC(t0, t1)
STORE(t1, UNIT(1) ... STORE(t1, UNIT(1) ...
ADDC(sum, t1) ADDC(sum, t0)
STORE(t2, UNIT(2) ... STORE(t2, UNIT(2) ...
ADDC(sum, t2) ADDC(t2, t3)
STORE(t3, UNIT(3) ... STORE(t3, UNIT(3) ...
ADDC(sum, t3) ADDC(sum, t2)
With this patch applied, ADDC and the **next next** ADDC are independent.
Signed-off-by: chenj <chenj@lemote.com>
Cc: linux-mips@linux-mips.org
Patchwork: https://patchwork.linux-mips.org/patch/9608/
Signed-off-by: Ralf Baechle <ralf@linux-mips.org>
2015-03-26 17:07:24 +00:00
|
|
|
ADDC(t6, t7)
|
2020-07-19 21:37:15 +00:00
|
|
|
STORE(t7, UNIT(7)(dst))
|
MIPS: csum_partial: Improve instruction parallelism.
Computing sum introduces true data dependency. This patch removes some
true data depdendencies, hence increases instruction level parallelism.
This patch brings up to 50% csum performance gain on Loongson 3a.
One example about how this patch works is in CSUM_BIGCHUNK1:
// ** original ** vs ** patch applied **
ADDC(sum, t0) ADDC(t0, t1)
ADDC(sum, t1) ADDC(t2, t3)
ADDC(sum, t2) ADDC(sum, t0)
ADDC(sum, t3) ADDC(sum, t2)
In the original implementation, each ADDC(sum, ...) depends on the sum
value updated by previous ADDC(as source operand).
With this patch applied, the first two ADDC operations are independent,
hence can be executed simultaneously if possible.
Another example is in the "copy and sum calculating chunk":
// ** original ** vs ** patch applied **
STORE(t0, UNIT(0) ... STORE(t0, UNIT(0) ...
ADDC(sum, t0) ADDC(t0, t1)
STORE(t1, UNIT(1) ... STORE(t1, UNIT(1) ...
ADDC(sum, t1) ADDC(sum, t0)
STORE(t2, UNIT(2) ... STORE(t2, UNIT(2) ...
ADDC(sum, t2) ADDC(t2, t3)
STORE(t3, UNIT(3) ... STORE(t3, UNIT(3) ...
ADDC(sum, t3) ADDC(sum, t2)
With this patch applied, ADDC and the **next next** ADDC are independent.
Signed-off-by: chenj <chenj@lemote.com>
Cc: linux-mips@linux-mips.org
Patchwork: https://patchwork.linux-mips.org/patch/9608/
Signed-off-by: Ralf Baechle <ralf@linux-mips.org>
2015-03-26 17:07:24 +00:00
|
|
|
ADDC(sum, t6)
|
2007-10-23 11:43:25 +00:00
|
|
|
.set reorder /* DADDI_WAR */
|
|
|
|
ADD dst, dst, 8*NBYTES
|
2006-12-12 16:22:06 +00:00
|
|
|
bgez len, 1b
|
2007-10-23 11:43:25 +00:00
|
|
|
.set noreorder
|
2006-12-12 16:22:06 +00:00
|
|
|
ADD len, 8*NBYTES # revert len (see above)
|
|
|
|
|
|
|
|
/*
|
|
|
|
* len == the number of bytes left to copy < 8*NBYTES
|
|
|
|
*/
|
2014-01-17 10:48:46 +00:00
|
|
|
.Lcleanup_both_aligned\@:
|
2006-12-12 16:22:06 +00:00
|
|
|
#define rem t7
|
2014-01-17 10:48:46 +00:00
|
|
|
beqz len, .Ldone\@
|
2006-12-12 16:22:06 +00:00
|
|
|
sltu t0, len, 4*NBYTES
|
2014-01-17 10:48:46 +00:00
|
|
|
bnez t0, .Lless_than_4units\@
|
2006-12-12 16:22:06 +00:00
|
|
|
and rem, len, (NBYTES-1) # rem = len % NBYTES
|
|
|
|
/*
|
|
|
|
* len >= 4*NBYTES
|
|
|
|
*/
|
2020-07-19 21:37:15 +00:00
|
|
|
LOAD(t0, UNIT(0)(src))
|
|
|
|
LOAD(t1, UNIT(1)(src))
|
|
|
|
LOAD(t2, UNIT(2)(src))
|
|
|
|
LOAD(t3, UNIT(3)(src))
|
2006-12-12 16:22:06 +00:00
|
|
|
SUB len, len, 4*NBYTES
|
|
|
|
ADD src, src, 4*NBYTES
|
2020-07-19 21:37:15 +00:00
|
|
|
STORE(t0, UNIT(0)(dst))
|
MIPS: csum_partial: Improve instruction parallelism.
Computing sum introduces true data dependency. This patch removes some
true data depdendencies, hence increases instruction level parallelism.
This patch brings up to 50% csum performance gain on Loongson 3a.
One example about how this patch works is in CSUM_BIGCHUNK1:
// ** original ** vs ** patch applied **
ADDC(sum, t0) ADDC(t0, t1)
ADDC(sum, t1) ADDC(t2, t3)
ADDC(sum, t2) ADDC(sum, t0)
ADDC(sum, t3) ADDC(sum, t2)
In the original implementation, each ADDC(sum, ...) depends on the sum
value updated by previous ADDC(as source operand).
With this patch applied, the first two ADDC operations are independent,
hence can be executed simultaneously if possible.
Another example is in the "copy and sum calculating chunk":
// ** original ** vs ** patch applied **
STORE(t0, UNIT(0) ... STORE(t0, UNIT(0) ...
ADDC(sum, t0) ADDC(t0, t1)
STORE(t1, UNIT(1) ... STORE(t1, UNIT(1) ...
ADDC(sum, t1) ADDC(sum, t0)
STORE(t2, UNIT(2) ... STORE(t2, UNIT(2) ...
ADDC(sum, t2) ADDC(t2, t3)
STORE(t3, UNIT(3) ... STORE(t3, UNIT(3) ...
ADDC(sum, t3) ADDC(sum, t2)
With this patch applied, ADDC and the **next next** ADDC are independent.
Signed-off-by: chenj <chenj@lemote.com>
Cc: linux-mips@linux-mips.org
Patchwork: https://patchwork.linux-mips.org/patch/9608/
Signed-off-by: Ralf Baechle <ralf@linux-mips.org>
2015-03-26 17:07:24 +00:00
|
|
|
ADDC(t0, t1)
|
2020-07-19 21:37:15 +00:00
|
|
|
STORE(t1, UNIT(1)(dst))
|
MIPS: csum_partial: Improve instruction parallelism.
Computing sum introduces true data dependency. This patch removes some
true data depdendencies, hence increases instruction level parallelism.
This patch brings up to 50% csum performance gain on Loongson 3a.
One example about how this patch works is in CSUM_BIGCHUNK1:
// ** original ** vs ** patch applied **
ADDC(sum, t0) ADDC(t0, t1)
ADDC(sum, t1) ADDC(t2, t3)
ADDC(sum, t2) ADDC(sum, t0)
ADDC(sum, t3) ADDC(sum, t2)
In the original implementation, each ADDC(sum, ...) depends on the sum
value updated by previous ADDC(as source operand).
With this patch applied, the first two ADDC operations are independent,
hence can be executed simultaneously if possible.
Another example is in the "copy and sum calculating chunk":
// ** original ** vs ** patch applied **
STORE(t0, UNIT(0) ... STORE(t0, UNIT(0) ...
ADDC(sum, t0) ADDC(t0, t1)
STORE(t1, UNIT(1) ... STORE(t1, UNIT(1) ...
ADDC(sum, t1) ADDC(sum, t0)
STORE(t2, UNIT(2) ... STORE(t2, UNIT(2) ...
ADDC(sum, t2) ADDC(t2, t3)
STORE(t3, UNIT(3) ... STORE(t3, UNIT(3) ...
ADDC(sum, t3) ADDC(sum, t2)
With this patch applied, ADDC and the **next next** ADDC are independent.
Signed-off-by: chenj <chenj@lemote.com>
Cc: linux-mips@linux-mips.org
Patchwork: https://patchwork.linux-mips.org/patch/9608/
Signed-off-by: Ralf Baechle <ralf@linux-mips.org>
2015-03-26 17:07:24 +00:00
|
|
|
ADDC(sum, t0)
|
2020-07-19 21:37:15 +00:00
|
|
|
STORE(t2, UNIT(2)(dst))
|
MIPS: csum_partial: Improve instruction parallelism.
Computing sum introduces true data dependency. This patch removes some
true data depdendencies, hence increases instruction level parallelism.
This patch brings up to 50% csum performance gain on Loongson 3a.
One example about how this patch works is in CSUM_BIGCHUNK1:
// ** original ** vs ** patch applied **
ADDC(sum, t0) ADDC(t0, t1)
ADDC(sum, t1) ADDC(t2, t3)
ADDC(sum, t2) ADDC(sum, t0)
ADDC(sum, t3) ADDC(sum, t2)
In the original implementation, each ADDC(sum, ...) depends on the sum
value updated by previous ADDC(as source operand).
With this patch applied, the first two ADDC operations are independent,
hence can be executed simultaneously if possible.
Another example is in the "copy and sum calculating chunk":
// ** original ** vs ** patch applied **
STORE(t0, UNIT(0) ... STORE(t0, UNIT(0) ...
ADDC(sum, t0) ADDC(t0, t1)
STORE(t1, UNIT(1) ... STORE(t1, UNIT(1) ...
ADDC(sum, t1) ADDC(sum, t0)
STORE(t2, UNIT(2) ... STORE(t2, UNIT(2) ...
ADDC(sum, t2) ADDC(t2, t3)
STORE(t3, UNIT(3) ... STORE(t3, UNIT(3) ...
ADDC(sum, t3) ADDC(sum, t2)
With this patch applied, ADDC and the **next next** ADDC are independent.
Signed-off-by: chenj <chenj@lemote.com>
Cc: linux-mips@linux-mips.org
Patchwork: https://patchwork.linux-mips.org/patch/9608/
Signed-off-by: Ralf Baechle <ralf@linux-mips.org>
2015-03-26 17:07:24 +00:00
|
|
|
ADDC(t2, t3)
|
2020-07-19 21:37:15 +00:00
|
|
|
STORE(t3, UNIT(3)(dst))
|
MIPS: csum_partial: Improve instruction parallelism.
Computing sum introduces true data dependency. This patch removes some
true data depdendencies, hence increases instruction level parallelism.
This patch brings up to 50% csum performance gain on Loongson 3a.
One example about how this patch works is in CSUM_BIGCHUNK1:
// ** original ** vs ** patch applied **
ADDC(sum, t0) ADDC(t0, t1)
ADDC(sum, t1) ADDC(t2, t3)
ADDC(sum, t2) ADDC(sum, t0)
ADDC(sum, t3) ADDC(sum, t2)
In the original implementation, each ADDC(sum, ...) depends on the sum
value updated by previous ADDC(as source operand).
With this patch applied, the first two ADDC operations are independent,
hence can be executed simultaneously if possible.
Another example is in the "copy and sum calculating chunk":
// ** original ** vs ** patch applied **
STORE(t0, UNIT(0) ... STORE(t0, UNIT(0) ...
ADDC(sum, t0) ADDC(t0, t1)
STORE(t1, UNIT(1) ... STORE(t1, UNIT(1) ...
ADDC(sum, t1) ADDC(sum, t0)
STORE(t2, UNIT(2) ... STORE(t2, UNIT(2) ...
ADDC(sum, t2) ADDC(t2, t3)
STORE(t3, UNIT(3) ... STORE(t3, UNIT(3) ...
ADDC(sum, t3) ADDC(sum, t2)
With this patch applied, ADDC and the **next next** ADDC are independent.
Signed-off-by: chenj <chenj@lemote.com>
Cc: linux-mips@linux-mips.org
Patchwork: https://patchwork.linux-mips.org/patch/9608/
Signed-off-by: Ralf Baechle <ralf@linux-mips.org>
2015-03-26 17:07:24 +00:00
|
|
|
ADDC(sum, t2)
|
2007-10-23 11:43:25 +00:00
|
|
|
.set reorder /* DADDI_WAR */
|
|
|
|
ADD dst, dst, 4*NBYTES
|
2014-01-17 10:48:46 +00:00
|
|
|
beqz len, .Ldone\@
|
2007-10-23 11:43:25 +00:00
|
|
|
.set noreorder
|
2014-01-17 10:48:46 +00:00
|
|
|
.Lless_than_4units\@:
|
2006-12-12 16:22:06 +00:00
|
|
|
/*
|
|
|
|
* rem = len % NBYTES
|
|
|
|
*/
|
2014-01-17 10:48:46 +00:00
|
|
|
beq rem, len, .Lcopy_bytes\@
|
2006-12-12 16:22:06 +00:00
|
|
|
nop
|
|
|
|
1:
|
2020-07-19 21:37:15 +00:00
|
|
|
LOAD(t0, 0(src))
|
2006-12-12 16:22:06 +00:00
|
|
|
ADD src, src, NBYTES
|
|
|
|
SUB len, len, NBYTES
|
2020-07-19 21:37:15 +00:00
|
|
|
STORE(t0, 0(dst))
|
2006-12-12 16:22:06 +00:00
|
|
|
ADDC(sum, t0)
|
2007-10-23 11:43:25 +00:00
|
|
|
.set reorder /* DADDI_WAR */
|
|
|
|
ADD dst, dst, NBYTES
|
2006-12-12 16:22:06 +00:00
|
|
|
bne rem, len, 1b
|
2007-10-23 11:43:25 +00:00
|
|
|
.set noreorder
|
2006-12-12 16:22:06 +00:00
|
|
|
|
|
|
|
/*
|
|
|
|
* src and dst are aligned, need to copy rem bytes (rem < NBYTES)
|
|
|
|
* A loop would do only a byte at a time with possible branch
|
2013-01-22 11:59:30 +00:00
|
|
|
* mispredicts. Can't do an explicit LOAD dst,mask,or,STORE
|
2006-12-12 16:22:06 +00:00
|
|
|
* because can't assume read-access to dst. Instead, use
|
|
|
|
* STREST dst, which doesn't require read access to dst.
|
|
|
|
*
|
|
|
|
* This code should perform better than a simple loop on modern,
|
|
|
|
* wide-issue mips processors because the code has fewer branches and
|
|
|
|
* more instruction-level parallelism.
|
|
|
|
*/
|
|
|
|
#define bits t2
|
2014-01-17 10:48:46 +00:00
|
|
|
beqz len, .Ldone\@
|
2006-12-12 16:22:06 +00:00
|
|
|
ADD t1, dst, len # t1 is just past last byte of dst
|
|
|
|
li bits, 8*NBYTES
|
|
|
|
SLL rem, len, 3 # rem = number of bits to keep
|
2020-07-19 21:37:15 +00:00
|
|
|
LOAD(t0, 0(src))
|
2013-01-22 11:59:30 +00:00
|
|
|
SUB bits, bits, rem # bits = number of bits to discard
|
2006-12-12 16:22:06 +00:00
|
|
|
SHIFT_DISCARD t0, t0, bits
|
2020-07-19 21:37:15 +00:00
|
|
|
STREST(t0, -1(t1))
|
2006-12-12 16:22:06 +00:00
|
|
|
SHIFT_DISCARD_REVERT t0, t0, bits
|
|
|
|
.set reorder
|
|
|
|
ADDC(sum, t0)
|
2014-01-17 10:48:46 +00:00
|
|
|
b .Ldone\@
|
2006-12-12 16:22:06 +00:00
|
|
|
.set noreorder
|
2014-01-17 10:48:46 +00:00
|
|
|
.Ldst_unaligned\@:
|
2006-12-12 16:22:06 +00:00
|
|
|
/*
|
|
|
|
* dst is unaligned
|
|
|
|
* t0 = src & ADDRMASK
|
|
|
|
* t1 = dst & ADDRMASK; T1 > 0
|
|
|
|
* len >= NBYTES
|
|
|
|
*
|
|
|
|
* Copy enough bytes to align dst
|
|
|
|
* Set match = (src and dst have same alignment)
|
|
|
|
*/
|
|
|
|
#define match rem
|
2020-07-19 21:37:15 +00:00
|
|
|
LDFIRST(t3, FIRST(0)(src))
|
2006-12-12 16:22:06 +00:00
|
|
|
ADD t2, zero, NBYTES
|
2020-07-19 21:37:15 +00:00
|
|
|
LDREST(t3, REST(0)(src))
|
2006-12-12 16:22:06 +00:00
|
|
|
SUB t2, t2, t1 # t2 = number of bytes copied
|
|
|
|
xor match, t0, t1
|
2020-07-19 21:37:15 +00:00
|
|
|
STFIRST(t3, FIRST(0)(dst))
|
2006-12-12 16:22:06 +00:00
|
|
|
SLL t4, t1, 3 # t4 = number of bits to discard
|
|
|
|
SHIFT_DISCARD t3, t3, t4
|
|
|
|
/* no SHIFT_DISCARD_REVERT to handle odd buffer properly */
|
|
|
|
ADDC(sum, t3)
|
2014-01-17 10:48:46 +00:00
|
|
|
beq len, t2, .Ldone\@
|
2006-12-12 16:22:06 +00:00
|
|
|
SUB len, len, t2
|
|
|
|
ADD dst, dst, t2
|
2014-01-17 10:48:46 +00:00
|
|
|
beqz match, .Lboth_aligned\@
|
2006-12-12 16:22:06 +00:00
|
|
|
ADD src, src, t2
|
|
|
|
|
2014-01-17 10:48:46 +00:00
|
|
|
.Lsrc_unaligned_dst_aligned\@:
|
2013-01-22 11:59:30 +00:00
|
|
|
SRL t0, len, LOG_NBYTES+2 # +2 for 4 units/iter
|
2014-01-17 10:48:46 +00:00
|
|
|
beqz t0, .Lcleanup_src_unaligned\@
|
2013-01-22 11:59:30 +00:00
|
|
|
and rem, len, (4*NBYTES-1) # rem = len % 4*NBYTES
|
2006-12-12 16:22:06 +00:00
|
|
|
1:
|
|
|
|
/*
|
|
|
|
* Avoid consecutive LD*'s to the same register since some mips
|
|
|
|
* implementations can't issue them in the same cycle.
|
|
|
|
* It's OK to load FIRST(N+1) before REST(N) because the two addresses
|
|
|
|
* are to the same unit (unless src is aligned, but it's not).
|
|
|
|
*/
|
2020-07-19 21:37:15 +00:00
|
|
|
LDFIRST(t0, FIRST(0)(src))
|
|
|
|
LDFIRST(t1, FIRST(1)(src))
|
2013-01-22 11:59:30 +00:00
|
|
|
SUB len, len, 4*NBYTES
|
2020-07-19 21:37:15 +00:00
|
|
|
LDREST(t0, REST(0)(src))
|
|
|
|
LDREST(t1, REST(1)(src))
|
|
|
|
LDFIRST(t2, FIRST(2)(src))
|
|
|
|
LDFIRST(t3, FIRST(3)(src))
|
|
|
|
LDREST(t2, REST(2)(src))
|
|
|
|
LDREST(t3, REST(3)(src))
|
2006-12-12 16:22:06 +00:00
|
|
|
ADD src, src, 4*NBYTES
|
|
|
|
#ifdef CONFIG_CPU_SB1
|
|
|
|
nop # improves slotting
|
|
|
|
#endif
|
2020-07-19 21:37:15 +00:00
|
|
|
STORE(t0, UNIT(0)(dst))
|
MIPS: csum_partial: Improve instruction parallelism.
Computing sum introduces true data dependency. This patch removes some
true data depdendencies, hence increases instruction level parallelism.
This patch brings up to 50% csum performance gain on Loongson 3a.
One example about how this patch works is in CSUM_BIGCHUNK1:
// ** original ** vs ** patch applied **
ADDC(sum, t0) ADDC(t0, t1)
ADDC(sum, t1) ADDC(t2, t3)
ADDC(sum, t2) ADDC(sum, t0)
ADDC(sum, t3) ADDC(sum, t2)
In the original implementation, each ADDC(sum, ...) depends on the sum
value updated by previous ADDC(as source operand).
With this patch applied, the first two ADDC operations are independent,
hence can be executed simultaneously if possible.
Another example is in the "copy and sum calculating chunk":
// ** original ** vs ** patch applied **
STORE(t0, UNIT(0) ... STORE(t0, UNIT(0) ...
ADDC(sum, t0) ADDC(t0, t1)
STORE(t1, UNIT(1) ... STORE(t1, UNIT(1) ...
ADDC(sum, t1) ADDC(sum, t0)
STORE(t2, UNIT(2) ... STORE(t2, UNIT(2) ...
ADDC(sum, t2) ADDC(t2, t3)
STORE(t3, UNIT(3) ... STORE(t3, UNIT(3) ...
ADDC(sum, t3) ADDC(sum, t2)
With this patch applied, ADDC and the **next next** ADDC are independent.
Signed-off-by: chenj <chenj@lemote.com>
Cc: linux-mips@linux-mips.org
Patchwork: https://patchwork.linux-mips.org/patch/9608/
Signed-off-by: Ralf Baechle <ralf@linux-mips.org>
2015-03-26 17:07:24 +00:00
|
|
|
ADDC(t0, t1)
|
2020-07-19 21:37:15 +00:00
|
|
|
STORE(t1, UNIT(1)(dst))
|
MIPS: csum_partial: Improve instruction parallelism.
Computing sum introduces true data dependency. This patch removes some
true data depdendencies, hence increases instruction level parallelism.
This patch brings up to 50% csum performance gain on Loongson 3a.
One example about how this patch works is in CSUM_BIGCHUNK1:
// ** original ** vs ** patch applied **
ADDC(sum, t0) ADDC(t0, t1)
ADDC(sum, t1) ADDC(t2, t3)
ADDC(sum, t2) ADDC(sum, t0)
ADDC(sum, t3) ADDC(sum, t2)
In the original implementation, each ADDC(sum, ...) depends on the sum
value updated by previous ADDC(as source operand).
With this patch applied, the first two ADDC operations are independent,
hence can be executed simultaneously if possible.
Another example is in the "copy and sum calculating chunk":
// ** original ** vs ** patch applied **
STORE(t0, UNIT(0) ... STORE(t0, UNIT(0) ...
ADDC(sum, t0) ADDC(t0, t1)
STORE(t1, UNIT(1) ... STORE(t1, UNIT(1) ...
ADDC(sum, t1) ADDC(sum, t0)
STORE(t2, UNIT(2) ... STORE(t2, UNIT(2) ...
ADDC(sum, t2) ADDC(t2, t3)
STORE(t3, UNIT(3) ... STORE(t3, UNIT(3) ...
ADDC(sum, t3) ADDC(sum, t2)
With this patch applied, ADDC and the **next next** ADDC are independent.
Signed-off-by: chenj <chenj@lemote.com>
Cc: linux-mips@linux-mips.org
Patchwork: https://patchwork.linux-mips.org/patch/9608/
Signed-off-by: Ralf Baechle <ralf@linux-mips.org>
2015-03-26 17:07:24 +00:00
|
|
|
ADDC(sum, t0)
|
2020-07-19 21:37:15 +00:00
|
|
|
STORE(t2, UNIT(2)(dst))
|
MIPS: csum_partial: Improve instruction parallelism.
Computing sum introduces true data dependency. This patch removes some
true data depdendencies, hence increases instruction level parallelism.
This patch brings up to 50% csum performance gain on Loongson 3a.
One example about how this patch works is in CSUM_BIGCHUNK1:
// ** original ** vs ** patch applied **
ADDC(sum, t0) ADDC(t0, t1)
ADDC(sum, t1) ADDC(t2, t3)
ADDC(sum, t2) ADDC(sum, t0)
ADDC(sum, t3) ADDC(sum, t2)
In the original implementation, each ADDC(sum, ...) depends on the sum
value updated by previous ADDC(as source operand).
With this patch applied, the first two ADDC operations are independent,
hence can be executed simultaneously if possible.
Another example is in the "copy and sum calculating chunk":
// ** original ** vs ** patch applied **
STORE(t0, UNIT(0) ... STORE(t0, UNIT(0) ...
ADDC(sum, t0) ADDC(t0, t1)
STORE(t1, UNIT(1) ... STORE(t1, UNIT(1) ...
ADDC(sum, t1) ADDC(sum, t0)
STORE(t2, UNIT(2) ... STORE(t2, UNIT(2) ...
ADDC(sum, t2) ADDC(t2, t3)
STORE(t3, UNIT(3) ... STORE(t3, UNIT(3) ...
ADDC(sum, t3) ADDC(sum, t2)
With this patch applied, ADDC and the **next next** ADDC are independent.
Signed-off-by: chenj <chenj@lemote.com>
Cc: linux-mips@linux-mips.org
Patchwork: https://patchwork.linux-mips.org/patch/9608/
Signed-off-by: Ralf Baechle <ralf@linux-mips.org>
2015-03-26 17:07:24 +00:00
|
|
|
ADDC(t2, t3)
|
2020-07-19 21:37:15 +00:00
|
|
|
STORE(t3, UNIT(3)(dst))
|
MIPS: csum_partial: Improve instruction parallelism.
Computing sum introduces true data dependency. This patch removes some
true data depdendencies, hence increases instruction level parallelism.
This patch brings up to 50% csum performance gain on Loongson 3a.
One example about how this patch works is in CSUM_BIGCHUNK1:
// ** original ** vs ** patch applied **
ADDC(sum, t0) ADDC(t0, t1)
ADDC(sum, t1) ADDC(t2, t3)
ADDC(sum, t2) ADDC(sum, t0)
ADDC(sum, t3) ADDC(sum, t2)
In the original implementation, each ADDC(sum, ...) depends on the sum
value updated by previous ADDC(as source operand).
With this patch applied, the first two ADDC operations are independent,
hence can be executed simultaneously if possible.
Another example is in the "copy and sum calculating chunk":
// ** original ** vs ** patch applied **
STORE(t0, UNIT(0) ... STORE(t0, UNIT(0) ...
ADDC(sum, t0) ADDC(t0, t1)
STORE(t1, UNIT(1) ... STORE(t1, UNIT(1) ...
ADDC(sum, t1) ADDC(sum, t0)
STORE(t2, UNIT(2) ... STORE(t2, UNIT(2) ...
ADDC(sum, t2) ADDC(t2, t3)
STORE(t3, UNIT(3) ... STORE(t3, UNIT(3) ...
ADDC(sum, t3) ADDC(sum, t2)
With this patch applied, ADDC and the **next next** ADDC are independent.
Signed-off-by: chenj <chenj@lemote.com>
Cc: linux-mips@linux-mips.org
Patchwork: https://patchwork.linux-mips.org/patch/9608/
Signed-off-by: Ralf Baechle <ralf@linux-mips.org>
2015-03-26 17:07:24 +00:00
|
|
|
ADDC(sum, t2)
|
2007-10-23 11:43:25 +00:00
|
|
|
.set reorder /* DADDI_WAR */
|
|
|
|
ADD dst, dst, 4*NBYTES
|
2006-12-12 16:22:06 +00:00
|
|
|
bne len, rem, 1b
|
2007-10-23 11:43:25 +00:00
|
|
|
.set noreorder
|
2006-12-12 16:22:06 +00:00
|
|
|
|
2014-01-17 10:48:46 +00:00
|
|
|
.Lcleanup_src_unaligned\@:
|
|
|
|
beqz len, .Ldone\@
|
2006-12-12 16:22:06 +00:00
|
|
|
and rem, len, NBYTES-1 # rem = len % NBYTES
|
2014-01-17 10:48:46 +00:00
|
|
|
beq rem, len, .Lcopy_bytes\@
|
2006-12-12 16:22:06 +00:00
|
|
|
nop
|
|
|
|
1:
|
2020-07-19 21:37:15 +00:00
|
|
|
LDFIRST(t0, FIRST(0)(src))
|
|
|
|
LDREST(t0, REST(0)(src))
|
2006-12-12 16:22:06 +00:00
|
|
|
ADD src, src, NBYTES
|
|
|
|
SUB len, len, NBYTES
|
2020-07-19 21:37:15 +00:00
|
|
|
STORE(t0, 0(dst))
|
2006-12-12 16:22:06 +00:00
|
|
|
ADDC(sum, t0)
|
2007-10-23 11:43:25 +00:00
|
|
|
.set reorder /* DADDI_WAR */
|
|
|
|
ADD dst, dst, NBYTES
|
2006-12-12 16:22:06 +00:00
|
|
|
bne len, rem, 1b
|
2007-10-23 11:43:25 +00:00
|
|
|
.set noreorder
|
2006-12-12 16:22:06 +00:00
|
|
|
|
2014-01-17 10:48:46 +00:00
|
|
|
.Lcopy_bytes_checklen\@:
|
|
|
|
beqz len, .Ldone\@
|
2006-12-12 16:22:06 +00:00
|
|
|
nop
|
2014-01-17 10:48:46 +00:00
|
|
|
.Lcopy_bytes\@:
|
2006-12-12 16:22:06 +00:00
|
|
|
/* 0 < len < NBYTES */
|
|
|
|
#ifdef CONFIG_CPU_LITTLE_ENDIAN
|
|
|
|
#define SHIFT_START 0
|
|
|
|
#define SHIFT_INC 8
|
|
|
|
#else
|
|
|
|
#define SHIFT_START 8*(NBYTES-1)
|
|
|
|
#define SHIFT_INC -8
|
|
|
|
#endif
|
|
|
|
move t2, zero # partial word
|
2013-01-22 11:59:30 +00:00
|
|
|
li t3, SHIFT_START # shift
|
2006-12-12 16:22:06 +00:00
|
|
|
#define COPY_BYTE(N) \
|
2020-07-19 21:37:15 +00:00
|
|
|
LOADBU(t0, N(src)); \
|
2006-12-12 16:22:06 +00:00
|
|
|
SUB len, len, 1; \
|
2020-07-19 21:37:15 +00:00
|
|
|
STOREB(t0, N(dst)); \
|
2006-12-12 16:22:06 +00:00
|
|
|
SLLV t0, t0, t3; \
|
|
|
|
addu t3, SHIFT_INC; \
|
2014-01-17 10:48:46 +00:00
|
|
|
beqz len, .Lcopy_bytes_done\@; \
|
2006-12-12 16:22:06 +00:00
|
|
|
or t2, t0
|
|
|
|
|
|
|
|
COPY_BYTE(0)
|
|
|
|
COPY_BYTE(1)
|
|
|
|
#ifdef USE_DOUBLE
|
|
|
|
COPY_BYTE(2)
|
|
|
|
COPY_BYTE(3)
|
|
|
|
COPY_BYTE(4)
|
|
|
|
COPY_BYTE(5)
|
|
|
|
#endif
|
2020-07-19 21:37:15 +00:00
|
|
|
LOADBU(t0, NBYTES-2(src))
|
2006-12-12 16:22:06 +00:00
|
|
|
SUB len, len, 1
|
2020-07-19 21:37:15 +00:00
|
|
|
STOREB(t0, NBYTES-2(dst))
|
2006-12-12 16:22:06 +00:00
|
|
|
SLLV t0, t0, t3
|
|
|
|
or t2, t0
|
2014-01-17 10:48:46 +00:00
|
|
|
.Lcopy_bytes_done\@:
|
2006-12-12 16:22:06 +00:00
|
|
|
ADDC(sum, t2)
|
2014-01-17 10:48:46 +00:00
|
|
|
.Ldone\@:
|
2006-12-12 16:22:06 +00:00
|
|
|
/* fold checksum */
|
2014-04-04 02:32:54 +00:00
|
|
|
.set push
|
|
|
|
.set noat
|
2006-12-12 16:22:06 +00:00
|
|
|
#ifdef USE_DOUBLE
|
|
|
|
dsll32 v1, sum, 0
|
|
|
|
daddu sum, v1
|
|
|
|
sltu v1, sum, v1
|
|
|
|
dsra32 sum, sum, 0
|
|
|
|
addu sum, v1
|
|
|
|
#endif
|
|
|
|
|
mips: Add MIPS Release 5 support
There are five MIPS32/64 architecture releases currently available:
from 1 to 6 except fourth one, which was intentionally skipped.
Three of them can be called as major: 1st, 2nd and 6th, that not only
have some system level alterations, but also introduced significant
core/ISA level updates. The rest of the MIPS architecture releases are
minor.
Even though they don't have as much ISA/system/core level changes
as the major ones with respect to the previous releases, they still
provide a set of updates (I'd say they were intended to be the
intermediate releases before a major one) that might be useful for the
kernel and user-level code, when activated by the kernel or compiler.
In particular the following features were introduced or ended up being
available at/after MIPS32/64 Release 5 architecture:
+ the last release of the misaligned memory access instructions,
+ virtualisation - VZ ASE - is optional component of the arch,
+ SIMD - MSA ASE - is optional component of the arch,
+ DSP ASE is optional component of the arch,
+ CP0.Status.FR=1 for CP1.FIR.F64=1 (pure 64-bit FPU general registers)
must be available if FPU is implemented,
+ CP1.FIR.Has2008 support is required so CP1.FCSR.{ABS2008,NAN2008} bits
are available.
+ UFR/UNFR aliases to access CP0.Status.FR from user-space by means of
ctc1/cfc1 instructions (enabled by CP0.Config5.UFR),
+ CP0.COnfig5.LLB=1 and eretnc instruction are implemented to without
accidentally clearing LL-bit when returning from an interrupt,
exception, or error trap,
+ XPA feature together with extended versions of CPx registers is
introduced, which needs to have mfhc0/mthc0 instructions available.
So due to these changes GNU GCC provides an extended instructions set
support for MIPS32/64 Release 5 by default like eretnc/mfhc0/mthc0. Even
though the architecture alteration isn't that big, it still worth to be
taken into account by the kernel software. Finally we can't deny that
some optimization/limitations might be found in future and implemented
on some level in kernel or compiler. In this case having even
intermediate MIPS architecture releases support would be more than
useful.
So the most of the changes provided by this commit can be split into
either compile- or runtime configs related. The compile-time related
changes are caused by adding the new CONFIG_CPU_MIPS32_R5/CONFIG_CPU_MIPSR5
configs and concern the code activating MIPSR2 or MIPSR6 already
implemented features (like eretnc/LLbit, mthc0/mfhc0). In addition
CPU_HAS_MSA can be now freely enabled for MIPS32/64 release 5 based
platforms as this is done for CPU_MIPS32_R6 CPUs. The runtime changes
concerns the features which are handled with respect to the MIPS ISA
revision detected at run-time by means of CP0.Config.{AT,AR} bits. Alas
these fields can be used to detect either r1 or r2 or r6 releases.
But since we know which CPUs in fact support the R5 arch, we can manually
set MIPS_CPU_ISA_M32R5/MIPS_CPU_ISA_M64R5 bit of c->isa_level and then
use cpu_has_mips32r5/cpu_has_mips64r5 where it's appropriate.
Since XPA/EVA provide too complex alterationss and to have them used with
MIPS32 Release 2 charged kernels (for compatibility with current platform
configs) they are left to be setup as a separate kernel configs.
Co-developed-by: Alexey Malahov <Alexey.Malahov@baikalelectronics.ru>
Signed-off-by: Alexey Malahov <Alexey.Malahov@baikalelectronics.ru>
Signed-off-by: Serge Semin <Sergey.Semin@baikalelectronics.ru>
Cc: Thomas Bogendoerfer <tsbogend@alpha.franken.de>
Cc: Paul Burton <paulburton@kernel.org>
Cc: Ralf Baechle <ralf@linux-mips.org>
Cc: Arnd Bergmann <arnd@arndb.de>
Cc: Rob Herring <robh+dt@kernel.org>
Cc: devicetree@vger.kernel.org
Signed-off-by: Thomas Bogendoerfer <tsbogend@alpha.franken.de>
2020-05-21 14:07:14 +00:00
|
|
|
#if defined(CONFIG_CPU_MIPSR2) || defined(CONFIG_CPU_MIPSR5) || \
|
|
|
|
defined(CONFIG_CPU_LOONGSON64)
|
2014-08-15 08:56:58 +00:00
|
|
|
.set push
|
|
|
|
.set arch=mips32r2
|
2008-10-11 15:18:53 +00:00
|
|
|
wsbh v1, sum
|
|
|
|
movn sum, v1, odd
|
2014-08-15 08:56:58 +00:00
|
|
|
.set pop
|
2008-10-11 15:18:53 +00:00
|
|
|
#else
|
|
|
|
beqz odd, 1f /* odd buffer alignment? */
|
|
|
|
lui v1, 0x00ff
|
|
|
|
addu v1, 0x00ff
|
|
|
|
and t0, sum, v1
|
|
|
|
sll t0, t0, 8
|
2006-12-12 16:22:06 +00:00
|
|
|
srl sum, sum, 8
|
2008-10-11 15:18:53 +00:00
|
|
|
and sum, sum, v1
|
|
|
|
or sum, sum, t0
|
2006-12-12 16:22:06 +00:00
|
|
|
1:
|
2008-10-11 15:18:53 +00:00
|
|
|
#endif
|
2014-04-04 02:32:54 +00:00
|
|
|
.set pop
|
2006-12-12 16:22:06 +00:00
|
|
|
.set reorder
|
|
|
|
jr ra
|
|
|
|
.set noreorder
|
2020-07-19 21:37:15 +00:00
|
|
|
.endm
|
2006-12-12 16:22:06 +00:00
|
|
|
|
2020-07-19 21:37:15 +00:00
|
|
|
.set noreorder
|
|
|
|
.L_exc:
|
2006-12-12 16:22:06 +00:00
|
|
|
jr ra
|
2020-07-19 21:37:15 +00:00
|
|
|
li v0, 0
|
2014-01-17 10:48:46 +00:00
|
|
|
|
2020-07-19 21:37:15 +00:00
|
|
|
FEXPORT(__csum_partial_copy_nocheck)
|
|
|
|
EXPORT_SYMBOL(__csum_partial_copy_nocheck)
|
2014-01-17 11:36:16 +00:00
|
|
|
#ifndef CONFIG_EVA
|
2014-01-17 10:48:46 +00:00
|
|
|
FEXPORT(__csum_partial_copy_to_user)
|
2016-11-07 11:14:13 +00:00
|
|
|
EXPORT_SYMBOL(__csum_partial_copy_to_user)
|
2014-01-17 10:48:46 +00:00
|
|
|
FEXPORT(__csum_partial_copy_from_user)
|
2016-11-07 11:14:13 +00:00
|
|
|
EXPORT_SYMBOL(__csum_partial_copy_from_user)
|
2014-01-17 11:36:16 +00:00
|
|
|
#endif
|
2020-07-19 21:37:15 +00:00
|
|
|
__BUILD_CSUM_PARTIAL_COPY_USER LEGACY_MODE USEROP USEROP
|
2014-01-17 11:36:16 +00:00
|
|
|
|
|
|
|
#ifdef CONFIG_EVA
|
|
|
|
LEAF(__csum_partial_copy_to_user)
|
2020-07-19 21:37:15 +00:00
|
|
|
__BUILD_CSUM_PARTIAL_COPY_USER EVA_MODE KERNELOP USEROP
|
2014-01-17 11:36:16 +00:00
|
|
|
END(__csum_partial_copy_to_user)
|
|
|
|
|
|
|
|
LEAF(__csum_partial_copy_from_user)
|
2020-07-19 21:37:15 +00:00
|
|
|
__BUILD_CSUM_PARTIAL_COPY_USER EVA_MODE USEROP KERNELOP
|
2014-01-17 11:36:16 +00:00
|
|
|
END(__csum_partial_copy_from_user)
|
|
|
|
#endif
|