2006-12-03 18:42:59 +03:00
|
|
|
/*
|
|
|
|
* 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 15:43:25 +04:00
|
|
|
* Copyright (C) 2007 Maciej W. Rozycki
|
2013-12-12 20:21:00 +04:00
|
|
|
* Copyright (C) 2014 Imagination Technologies Ltd.
|
2006-12-03 18:42:59 +03:00
|
|
|
*/
|
2006-12-12 19:22:06 +03:00
|
|
|
#include <linux/errno.h>
|
2006-12-03 18:42:59 +03:00
|
|
|
#include <asm/asm.h>
|
2006-12-12 19:22:06 +03:00
|
|
|
#include <asm/asm-offsets.h>
|
2016-11-07 14:14:13 +03:00
|
|
|
#include <asm/export.h>
|
2006-12-03 18:42:59 +03:00
|
|
|
#include <asm/regdef.h>
|
|
|
|
|
|
|
|
#ifdef CONFIG_64BIT
|
2006-12-07 19:04:31 +03:00
|
|
|
/*
|
|
|
|
* 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 19:04:51 +03:00
|
|
|
|
|
|
|
#define USE_DOUBLE
|
2006-12-03 18:42:59 +03:00
|
|
|
#endif
|
|
|
|
|
2006-12-07 19:04:51 +03:00
|
|
|
#ifdef USE_DOUBLE
|
|
|
|
|
|
|
|
#define LOAD ld
|
2008-09-20 19:20:04 +04:00
|
|
|
#define LOAD32 lwu
|
2006-12-07 19:04:51 +03:00
|
|
|
#define ADD daddu
|
|
|
|
#define NBYTES 8
|
|
|
|
|
|
|
|
#else
|
|
|
|
|
|
|
|
#define LOAD lw
|
2008-09-20 19:20:04 +04:00
|
|
|
#define LOAD32 lw
|
2006-12-07 19:04:51 +03:00
|
|
|
#define ADD addu
|
|
|
|
#define NBYTES 4
|
|
|
|
|
|
|
|
#endif /* USE_DOUBLE */
|
|
|
|
|
|
|
|
#define UNIT(unit) ((unit)*NBYTES)
|
|
|
|
|
2006-12-03 18:42:59 +03:00
|
|
|
#define ADDC(sum,reg) \
|
2014-04-04 06:32:54 +04:00
|
|
|
.set push; \
|
|
|
|
.set noat; \
|
2006-12-07 19:04:51 +03:00
|
|
|
ADD sum, reg; \
|
2006-12-03 18:42:59 +03:00
|
|
|
sltu v1, sum, reg; \
|
2007-10-23 15:43:25 +04:00
|
|
|
ADD sum, v1; \
|
2014-04-04 06:32:54 +04:00
|
|
|
.set pop
|
2006-12-03 18:42:59 +03:00
|
|
|
|
2008-09-20 19:20:04 +04:00
|
|
|
#define ADDC32(sum,reg) \
|
2014-04-04 06:32:54 +04:00
|
|
|
.set push; \
|
|
|
|
.set noat; \
|
2008-09-20 19:20:04 +04:00
|
|
|
addu sum, reg; \
|
|
|
|
sltu v1, sum, reg; \
|
|
|
|
addu sum, v1; \
|
2014-04-04 06:32:54 +04:00
|
|
|
.set pop
|
2008-09-20 19:20:04 +04:00
|
|
|
|
2006-12-07 19:04:51 +03:00
|
|
|
#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 15:59:30 +04:00
|
|
|
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 20:07:24 +03:00
|
|
|
ADDC(_t0, _t1); \
|
|
|
|
ADDC(_t2, _t3); \
|
2006-12-03 18:42:59 +03:00
|
|
|
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 20:07:24 +03:00
|
|
|
ADDC(sum, _t2)
|
2006-12-07 19:04:51 +03:00
|
|
|
|
|
|
|
#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 18:42:59 +03:00
|
|
|
|
|
|
|
/*
|
|
|
|
* 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 14:14:13 +03:00
|
|
|
EXPORT_SYMBOL(csum_partial)
|
2006-12-03 18:42:59 +03:00
|
|
|
move sum, zero
|
2006-12-07 19:04:31 +03:00
|
|
|
move t7, zero
|
2006-12-03 18:42:59 +03:00
|
|
|
|
|
|
|
sltiu t8, a1, 0x8
|
2008-01-29 13:14:59 +03:00
|
|
|
bnez t8, .Lsmall_csumcpy /* < 8 bytes to copy */
|
2006-12-07 19:04:31 +03:00
|
|
|
move t2, a1
|
2006-12-03 18:42:59 +03:00
|
|
|
|
2006-12-07 19:04:45 +03:00
|
|
|
andi t7, src, 0x1 /* odd buffer? */
|
2006-12-03 18:42:59 +03:00
|
|
|
|
2008-01-29 13:14:59 +03:00
|
|
|
.Lhword_align:
|
|
|
|
beqz t7, .Lword_align
|
2006-12-03 18:42:59 +03:00
|
|
|
andi t8, src, 0x2
|
|
|
|
|
2006-12-07 19:04:31 +03:00
|
|
|
lbu t0, (src)
|
2006-12-03 18:42:59 +03:00
|
|
|
LONG_SUBU a1, a1, 0x1
|
|
|
|
#ifdef __MIPSEL__
|
2006-12-07 19:04:31 +03:00
|
|
|
sll t0, t0, 8
|
2006-12-03 18:42:59 +03:00
|
|
|
#endif
|
2006-12-07 19:04:31 +03:00
|
|
|
ADDC(sum, t0)
|
2006-12-03 18:42:59 +03:00
|
|
|
PTR_ADDU src, src, 0x1
|
|
|
|
andi t8, src, 0x2
|
|
|
|
|
2008-01-29 13:14:59 +03:00
|
|
|
.Lword_align:
|
|
|
|
beqz t8, .Ldword_align
|
2006-12-03 18:42:59 +03:00
|
|
|
sltiu t8, a1, 56
|
|
|
|
|
2006-12-07 19:04:31 +03:00
|
|
|
lhu t0, (src)
|
2006-12-03 18:42:59 +03:00
|
|
|
LONG_SUBU a1, a1, 0x2
|
2006-12-07 19:04:31 +03:00
|
|
|
ADDC(sum, t0)
|
2006-12-03 18:42:59 +03:00
|
|
|
sltiu t8, a1, 56
|
|
|
|
PTR_ADDU src, src, 0x2
|
|
|
|
|
2008-01-29 13:14:59 +03:00
|
|
|
.Ldword_align:
|
|
|
|
bnez t8, .Ldo_end_words
|
2006-12-03 18:42:59 +03:00
|
|
|
move t8, a1
|
|
|
|
|
|
|
|
andi t8, src, 0x4
|
2008-01-29 13:14:59 +03:00
|
|
|
beqz t8, .Lqword_align
|
2006-12-03 18:42:59 +03:00
|
|
|
andi t8, src, 0x8
|
|
|
|
|
2008-09-20 19:20:04 +04:00
|
|
|
LOAD32 t0, 0x00(src)
|
2006-12-03 18:42:59 +03:00
|
|
|
LONG_SUBU a1, a1, 0x4
|
2006-12-07 19:04:31 +03:00
|
|
|
ADDC(sum, t0)
|
2006-12-03 18:42:59 +03:00
|
|
|
PTR_ADDU src, src, 0x4
|
|
|
|
andi t8, src, 0x8
|
|
|
|
|
2008-01-29 13:14:59 +03:00
|
|
|
.Lqword_align:
|
|
|
|
beqz t8, .Loword_align
|
2006-12-03 18:42:59 +03:00
|
|
|
andi t8, src, 0x10
|
|
|
|
|
2006-12-07 19:04:51 +03:00
|
|
|
#ifdef USE_DOUBLE
|
|
|
|
ld t0, 0x00(src)
|
|
|
|
LONG_SUBU a1, a1, 0x8
|
|
|
|
ADDC(sum, t0)
|
|
|
|
#else
|
2006-12-07 19:04:31 +03:00
|
|
|
lw t0, 0x00(src)
|
|
|
|
lw t1, 0x04(src)
|
2006-12-03 18:42:59 +03:00
|
|
|
LONG_SUBU a1, a1, 0x8
|
2006-12-07 19:04:31 +03:00
|
|
|
ADDC(sum, t0)
|
|
|
|
ADDC(sum, t1)
|
2006-12-07 19:04:51 +03:00
|
|
|
#endif
|
2006-12-03 18:42:59 +03:00
|
|
|
PTR_ADDU src, src, 0x8
|
|
|
|
andi t8, src, 0x10
|
|
|
|
|
2008-01-29 13:14:59 +03:00
|
|
|
.Loword_align:
|
|
|
|
beqz t8, .Lbegin_movement
|
2006-12-03 18:42:59 +03:00
|
|
|
LONG_SRL t8, a1, 0x7
|
|
|
|
|
2006-12-07 19:04:51 +03:00
|
|
|
#ifdef USE_DOUBLE
|
|
|
|
ld t0, 0x00(src)
|
|
|
|
ld t1, 0x08(src)
|
2006-12-07 19:04:31 +03:00
|
|
|
ADDC(sum, t0)
|
|
|
|
ADDC(sum, t1)
|
2006-12-07 19:04:51 +03:00
|
|
|
#else
|
|
|
|
CSUM_BIGCHUNK1(src, 0x00, sum, t0, t1, t3, t4)
|
|
|
|
#endif
|
2006-12-03 18:42:59 +03:00
|
|
|
LONG_SUBU a1, a1, 0x10
|
|
|
|
PTR_ADDU src, src, 0x10
|
|
|
|
LONG_SRL t8, a1, 0x7
|
|
|
|
|
2008-01-29 13:14:59 +03:00
|
|
|
.Lbegin_movement:
|
2006-12-03 18:42:59 +03:00
|
|
|
beqz t8, 1f
|
2006-12-07 19:04:31 +03:00
|
|
|
andi t2, a1, 0x40
|
2006-12-03 18:42:59 +03:00
|
|
|
|
2008-01-29 13:14:59 +03:00
|
|
|
.Lmove_128bytes:
|
2006-12-07 19:04:31 +03:00
|
|
|
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 18:42:59 +03:00
|
|
|
LONG_SUBU t8, t8, 0x01
|
2007-10-23 15:43:25 +04:00
|
|
|
.set reorder /* DADDI_WAR */
|
|
|
|
PTR_ADDU src, src, 0x80
|
2008-01-29 13:14:59 +03:00
|
|
|
bnez t8, .Lmove_128bytes
|
2007-10-23 15:43:25 +04:00
|
|
|
.set noreorder
|
2006-12-03 18:42:59 +03:00
|
|
|
|
|
|
|
1:
|
2006-12-07 19:04:31 +03:00
|
|
|
beqz t2, 1f
|
|
|
|
andi t2, a1, 0x20
|
2006-12-03 18:42:59 +03:00
|
|
|
|
2008-01-29 13:14:59 +03:00
|
|
|
.Lmove_64bytes:
|
2006-12-07 19:04:31 +03:00
|
|
|
CSUM_BIGCHUNK(src, 0x00, sum, t0, t1, t3, t4)
|
|
|
|
CSUM_BIGCHUNK(src, 0x20, sum, t0, t1, t3, t4)
|
2006-12-03 18:42:59 +03:00
|
|
|
PTR_ADDU src, src, 0x40
|
|
|
|
|
|
|
|
1:
|
2008-01-29 13:14:59 +03:00
|
|
|
beqz t2, .Ldo_end_words
|
2006-12-03 18:42:59 +03:00
|
|
|
andi t8, a1, 0x1c
|
|
|
|
|
2008-01-29 13:14:59 +03:00
|
|
|
.Lmove_32bytes:
|
2006-12-07 19:04:31 +03:00
|
|
|
CSUM_BIGCHUNK(src, 0x00, sum, t0, t1, t3, t4)
|
2006-12-03 18:42:59 +03:00
|
|
|
andi t8, a1, 0x1c
|
|
|
|
PTR_ADDU src, src, 0x20
|
|
|
|
|
2008-01-29 13:14:59 +03:00
|
|
|
.Ldo_end_words:
|
|
|
|
beqz t8, .Lsmall_csumcpy
|
2006-12-07 19:04:45 +03:00
|
|
|
andi t2, a1, 0x3
|
|
|
|
LONG_SRL t8, t8, 0x2
|
2006-12-03 18:42:59 +03:00
|
|
|
|
2008-01-29 13:14:59 +03:00
|
|
|
.Lend_words:
|
2008-09-20 19:20:04 +04:00
|
|
|
LOAD32 t0, (src)
|
2006-12-03 18:42:59 +03:00
|
|
|
LONG_SUBU t8, t8, 0x1
|
2006-12-07 19:04:31 +03:00
|
|
|
ADDC(sum, t0)
|
2007-10-23 15:43:25 +04:00
|
|
|
.set reorder /* DADDI_WAR */
|
|
|
|
PTR_ADDU src, src, 0x4
|
2008-01-29 13:14:59 +03:00
|
|
|
bnez t8, .Lend_words
|
2007-10-23 15:43:25 +04:00
|
|
|
.set noreorder
|
2006-12-03 18:42:59 +03:00
|
|
|
|
2006-12-07 19:04:45 +03:00
|
|
|
/* unknown src alignment and < 8 bytes to go */
|
2008-01-29 13:14:59 +03:00
|
|
|
.Lsmall_csumcpy:
|
2006-12-07 19:04:45 +03:00
|
|
|
move a1, t2
|
2006-12-03 18:42:59 +03:00
|
|
|
|
2006-12-07 19:04:45 +03:00
|
|
|
andi t0, a1, 4
|
|
|
|
beqz t0, 1f
|
|
|
|
andi t0, a1, 2
|
2006-12-03 18:42:59 +03:00
|
|
|
|
2006-12-07 19:04:45 +03:00
|
|
|
/* Still a full word to go */
|
|
|
|
ulw t1, (src)
|
|
|
|
PTR_ADDIU src, 4
|
2008-09-20 19:20:04 +04:00
|
|
|
#ifdef USE_DOUBLE
|
|
|
|
dsll t1, t1, 32 /* clear lower 32bit */
|
|
|
|
#endif
|
2006-12-07 19:04:45 +03:00
|
|
|
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 18:42:59 +03:00
|
|
|
|
2006-12-07 19:04:45 +03:00
|
|
|
/* fold checksum */
|
2006-12-07 19:04:51 +03:00
|
|
|
#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 19:04:45 +03:00
|
|
|
|
|
|
|
/* odd buffer alignment? */
|
2019-10-20 17:43:13 +03:00
|
|
|
#if defined(CONFIG_CPU_MIPSR2) || defined(CONFIG_CPU_LOONGSON64)
|
2014-08-15 12:56:58 +04:00
|
|
|
.set push
|
|
|
|
.set arch=mips32r2
|
2008-10-11 19:18:53 +04:00
|
|
|
wsbh v1, sum
|
|
|
|
movn sum, v1, t7
|
2014-08-15 12:56:58 +04:00
|
|
|
.set pop
|
2008-10-11 19:18:53 +04:00
|
|
|
#else
|
|
|
|
beqz t7, 1f /* odd buffer alignment? */
|
|
|
|
lui v1, 0x00ff
|
|
|
|
addu v1, 0x00ff
|
|
|
|
and t0, sum, v1
|
|
|
|
sll t0, t0, 8
|
2006-12-07 19:04:45 +03:00
|
|
|
srl sum, sum, 8
|
2008-10-11 19:18:53 +04:00
|
|
|
and sum, sum, v1
|
|
|
|
or sum, sum, t0
|
2006-12-07 19:04:45 +03:00
|
|
|
1:
|
2008-10-11 19:18:53 +04:00
|
|
|
#endif
|
2006-12-07 19:04:45 +03:00
|
|
|
.set reorder
|
2013-01-22 15:59:30 +04:00
|
|
|
/* Add the passed partial csum. */
|
2008-09-20 19:20:04 +04:00
|
|
|
ADDC32(sum, a2)
|
2006-12-03 18:42:59 +03:00
|
|
|
jr ra
|
2006-12-07 19:04:45 +03:00
|
|
|
.set noreorder
|
2006-12-03 18:42:59 +03:00
|
|
|
END(csum_partial)
|
2006-12-12 19:22:06 +03:00
|
|
|
|
|
|
|
|
|
|
|
/*
|
|
|
|
* checksum and copy routines based on memcpy.S
|
|
|
|
*
|
|
|
|
* csum_partial_copy_nocheck(src, dst, len, sum)
|
2013-12-12 20:21:00 +04:00
|
|
|
* __csum_partial_copy_kernel(src, dst, len, sum, errp)
|
2006-12-12 19:22:06 +03:00
|
|
|
*
|
2013-01-22 15:59:30 +04:00
|
|
|
* See "Spec" in memcpy.S for details. Unlike __copy_user, all
|
2006-12-12 19:22:06 +03:00
|
|
|
* function in this file use the standard calling convention.
|
|
|
|
*/
|
|
|
|
|
|
|
|
#define src a0
|
|
|
|
#define dst a1
|
|
|
|
#define len a2
|
|
|
|
#define psum a3
|
|
|
|
#define sum v0
|
|
|
|
#define odd t8
|
|
|
|
#define errptr t9
|
|
|
|
|
|
|
|
/*
|
|
|
|
* The exception handler for loads requires that:
|
|
|
|
* 1- AT contain the address of the byte just past the end of the source
|
|
|
|
* of the copy,
|
|
|
|
* 2- src_entry <= src < AT, and
|
|
|
|
* 3- (dst - src) == (dst_entry - src_entry),
|
|
|
|
* The _entry suffix denotes values when __copy_user was called.
|
|
|
|
*
|
|
|
|
* (1) is set up up by __csum_partial_copy_from_user and maintained by
|
|
|
|
* not writing AT in __csum_partial_copy
|
|
|
|
* (2) is met by incrementing src by the number of bytes copied
|
|
|
|
* (3) is met by not doing loads between a pair of increments of dst and src
|
|
|
|
*
|
|
|
|
* The exception handlers for stores stores -EFAULT to errptr and return.
|
|
|
|
* These handlers do not need to overwrite any data.
|
|
|
|
*/
|
|
|
|
|
2014-01-16 21:02:13 +04:00
|
|
|
/* Instruction type */
|
|
|
|
#define LD_INSN 1
|
|
|
|
#define ST_INSN 2
|
2014-01-17 14:48:46 +04:00
|
|
|
#define LEGACY_MODE 1
|
|
|
|
#define EVA_MODE 2
|
|
|
|
#define USEROP 1
|
|
|
|
#define KERNELOP 2
|
2014-01-16 21:02:13 +04:00
|
|
|
|
|
|
|
/*
|
|
|
|
* 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
|
|
|
|
*/
|
|
|
|
#define EXC(insn, type, reg, addr, handler) \
|
2014-01-17 14:48:46 +04:00
|
|
|
.if \mode == LEGACY_MODE; \
|
|
|
|
9: insn reg, addr; \
|
|
|
|
.section __ex_table,"a"; \
|
|
|
|
PTR 9b, handler; \
|
|
|
|
.previous; \
|
2014-01-17 15:36:16 +04:00
|
|
|
/* 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"; \
|
|
|
|
PTR 9b, handler; \
|
|
|
|
.previous; \
|
|
|
|
.else; \
|
|
|
|
/* EVA without exception */ \
|
|
|
|
insn reg, addr; \
|
|
|
|
.endif; \
|
2014-01-17 14:48:46 +04:00
|
|
|
.endif
|
2006-12-12 19:22:06 +03:00
|
|
|
|
2014-01-16 21:02:13 +04:00
|
|
|
#undef LOAD
|
|
|
|
|
2006-12-12 19:22:06 +03:00
|
|
|
#ifdef USE_DOUBLE
|
|
|
|
|
2014-01-16 21:02:13 +04:00
|
|
|
#define LOADK ld /* No exception */
|
|
|
|
#define LOAD(reg, addr, handler) EXC(ld, LD_INSN, reg, addr, handler)
|
|
|
|
#define LOADBU(reg, addr, handler) EXC(lbu, LD_INSN, reg, addr, handler)
|
|
|
|
#define LOADL(reg, addr, handler) EXC(ldl, LD_INSN, reg, addr, handler)
|
|
|
|
#define LOADR(reg, addr, handler) EXC(ldr, LD_INSN, reg, addr, handler)
|
|
|
|
#define STOREB(reg, addr, handler) EXC(sb, ST_INSN, reg, addr, handler)
|
|
|
|
#define STOREL(reg, addr, handler) EXC(sdl, ST_INSN, reg, addr, handler)
|
|
|
|
#define STORER(reg, addr, handler) EXC(sdr, ST_INSN, reg, addr, handler)
|
|
|
|
#define STORE(reg, addr, handler) EXC(sd, ST_INSN, reg, addr, handler)
|
2006-12-12 19:22:06 +03:00
|
|
|
#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 21:02:13 +04:00
|
|
|
#define LOADK lw /* No exception */
|
|
|
|
#define LOAD(reg, addr, handler) EXC(lw, LD_INSN, reg, addr, handler)
|
|
|
|
#define LOADBU(reg, addr, handler) EXC(lbu, LD_INSN, reg, addr, handler)
|
|
|
|
#define LOADL(reg, addr, handler) EXC(lwl, LD_INSN, reg, addr, handler)
|
|
|
|
#define LOADR(reg, addr, handler) EXC(lwr, LD_INSN, reg, addr, handler)
|
|
|
|
#define STOREB(reg, addr, handler) EXC(sb, ST_INSN, reg, addr, handler)
|
|
|
|
#define STOREL(reg, addr, handler) EXC(swl, ST_INSN, reg, addr, handler)
|
|
|
|
#define STORER(reg, addr, handler) EXC(swr, ST_INSN, reg, addr, handler)
|
|
|
|
#define STORE(reg, addr, handler) EXC(sw, ST_INSN, reg, addr, handler)
|
2006-12-12 19:22:06 +03:00
|
|
|
#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 15:59:30 +04:00
|
|
|
#define LDREST LOADL
|
2006-12-12 19:22:06 +03:00
|
|
|
#define STFIRST STORER
|
2013-01-22 15:59:30 +04:00
|
|
|
#define STREST STOREL
|
2006-12-12 19:22:06 +03:00
|
|
|
#define SHIFT_DISCARD SLLV
|
|
|
|
#define SHIFT_DISCARD_REVERT SRLV
|
|
|
|
#else
|
|
|
|
#define LDFIRST LOADL
|
2013-01-22 15:59:30 +04:00
|
|
|
#define LDREST LOADR
|
2006-12-12 19:22:06 +03:00
|
|
|
#define STFIRST STOREL
|
2013-01-22 15:59:30 +04:00
|
|
|
#define STREST STORER
|
2006-12-12 19:22:06 +03:00
|
|
|
#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 15:43:25 +04:00
|
|
|
#ifndef CONFIG_CPU_DADDI_WORKAROUNDS
|
2006-12-12 19:22:06 +03:00
|
|
|
.set noat
|
2007-10-23 15:43:25 +04:00
|
|
|
#else
|
|
|
|
.set at=v1
|
|
|
|
#endif
|
2006-12-12 19:22:06 +03:00
|
|
|
|
2014-01-17 14:48:46 +04:00
|
|
|
.macro __BUILD_CSUM_PARTIAL_COPY_USER mode, from, to, __nocheck
|
|
|
|
|
2006-12-12 19:22:06 +03:00
|
|
|
PTR_ADDU AT, src, len /* See (1) above. */
|
2014-01-17 14:48:46 +04:00
|
|
|
/* initialize __nocheck if this the first time we execute this
|
|
|
|
* macro
|
|
|
|
*/
|
2006-12-12 19:22:06 +03:00
|
|
|
#ifdef CONFIG_64BIT
|
|
|
|
move errptr, a4
|
|
|
|
#else
|
|
|
|
lw errptr, 16(sp)
|
|
|
|
#endif
|
2014-01-17 14:48:46 +04:00
|
|
|
.if \__nocheck == 1
|
|
|
|
FEXPORT(csum_partial_copy_nocheck)
|
2016-11-07 14:14:13 +03:00
|
|
|
EXPORT_SYMBOL(csum_partial_copy_nocheck)
|
2014-01-17 14:48:46 +04:00
|
|
|
.endif
|
2006-12-12 19:22:06 +03:00
|
|
|
move sum, zero
|
|
|
|
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 14:48:46 +04:00
|
|
|
bnez t2, .Lcopy_bytes_checklen\@
|
2006-12-12 19:22:06 +03:00
|
|
|
and t0, src, ADDRMASK
|
|
|
|
andi odd, dst, 0x1 /* odd buffer? */
|
2014-01-17 14:48:46 +04:00
|
|
|
bnez t1, .Ldst_unaligned\@
|
2006-12-12 19:22:06 +03:00
|
|
|
nop
|
2014-01-17 14:48:46 +04:00
|
|
|
bnez t0, .Lsrc_unaligned_dst_aligned\@
|
2006-12-12 19:22:06 +03:00
|
|
|
/*
|
|
|
|
* use delay slot for fall-through
|
|
|
|
* src and dst are aligned; need to compute rem
|
|
|
|
*/
|
2014-01-17 14:48:46 +04:00
|
|
|
.Lboth_aligned\@:
|
2013-01-22 15:59:30 +04:00
|
|
|
SRL t0, len, LOG_NBYTES+3 # +3 for 8 units/iter
|
2014-01-17 14:48:46 +04:00
|
|
|
beqz t0, .Lcleanup_both_aligned\@ # len < 8*NBYTES
|
2006-12-12 19:22:06 +03:00
|
|
|
nop
|
|
|
|
SUB len, 8*NBYTES # subtract here for bgez loop
|
|
|
|
.align 4
|
|
|
|
1:
|
2014-01-17 14:48:46 +04:00
|
|
|
LOAD(t0, UNIT(0)(src), .Ll_exc\@)
|
|
|
|
LOAD(t1, UNIT(1)(src), .Ll_exc_copy\@)
|
|
|
|
LOAD(t2, UNIT(2)(src), .Ll_exc_copy\@)
|
|
|
|
LOAD(t3, UNIT(3)(src), .Ll_exc_copy\@)
|
|
|
|
LOAD(t4, UNIT(4)(src), .Ll_exc_copy\@)
|
|
|
|
LOAD(t5, UNIT(5)(src), .Ll_exc_copy\@)
|
|
|
|
LOAD(t6, UNIT(6)(src), .Ll_exc_copy\@)
|
|
|
|
LOAD(t7, UNIT(7)(src), .Ll_exc_copy\@)
|
2006-12-12 19:22:06 +03:00
|
|
|
SUB len, len, 8*NBYTES
|
|
|
|
ADD src, src, 8*NBYTES
|
2014-01-17 14:48:46 +04:00
|
|
|
STORE(t0, UNIT(0)(dst), .Ls_exc\@)
|
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 20:07:24 +03:00
|
|
|
ADDC(t0, t1)
|
2014-01-17 14:48:46 +04:00
|
|
|
STORE(t1, UNIT(1)(dst), .Ls_exc\@)
|
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 20:07:24 +03:00
|
|
|
ADDC(sum, t0)
|
2014-01-17 14:48:46 +04:00
|
|
|
STORE(t2, UNIT(2)(dst), .Ls_exc\@)
|
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 20:07:24 +03:00
|
|
|
ADDC(t2, t3)
|
2014-01-17 14:48:46 +04:00
|
|
|
STORE(t3, UNIT(3)(dst), .Ls_exc\@)
|
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 20:07:24 +03:00
|
|
|
ADDC(sum, t2)
|
2014-01-17 14:48:46 +04:00
|
|
|
STORE(t4, UNIT(4)(dst), .Ls_exc\@)
|
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 20:07:24 +03:00
|
|
|
ADDC(t4, t5)
|
2014-01-17 14:48:46 +04:00
|
|
|
STORE(t5, UNIT(5)(dst), .Ls_exc\@)
|
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 20:07:24 +03:00
|
|
|
ADDC(sum, t4)
|
2014-01-17 14:48:46 +04:00
|
|
|
STORE(t6, UNIT(6)(dst), .Ls_exc\@)
|
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 20:07:24 +03:00
|
|
|
ADDC(t6, t7)
|
2014-01-17 14:48:46 +04:00
|
|
|
STORE(t7, UNIT(7)(dst), .Ls_exc\@)
|
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 20:07:24 +03:00
|
|
|
ADDC(sum, t6)
|
2007-10-23 15:43:25 +04:00
|
|
|
.set reorder /* DADDI_WAR */
|
|
|
|
ADD dst, dst, 8*NBYTES
|
2006-12-12 19:22:06 +03:00
|
|
|
bgez len, 1b
|
2007-10-23 15:43:25 +04:00
|
|
|
.set noreorder
|
2006-12-12 19:22:06 +03:00
|
|
|
ADD len, 8*NBYTES # revert len (see above)
|
|
|
|
|
|
|
|
/*
|
|
|
|
* len == the number of bytes left to copy < 8*NBYTES
|
|
|
|
*/
|
2014-01-17 14:48:46 +04:00
|
|
|
.Lcleanup_both_aligned\@:
|
2006-12-12 19:22:06 +03:00
|
|
|
#define rem t7
|
2014-01-17 14:48:46 +04:00
|
|
|
beqz len, .Ldone\@
|
2006-12-12 19:22:06 +03:00
|
|
|
sltu t0, len, 4*NBYTES
|
2014-01-17 14:48:46 +04:00
|
|
|
bnez t0, .Lless_than_4units\@
|
2006-12-12 19:22:06 +03:00
|
|
|
and rem, len, (NBYTES-1) # rem = len % NBYTES
|
|
|
|
/*
|
|
|
|
* len >= 4*NBYTES
|
|
|
|
*/
|
2014-01-17 14:48:46 +04:00
|
|
|
LOAD(t0, UNIT(0)(src), .Ll_exc\@)
|
|
|
|
LOAD(t1, UNIT(1)(src), .Ll_exc_copy\@)
|
|
|
|
LOAD(t2, UNIT(2)(src), .Ll_exc_copy\@)
|
|
|
|
LOAD(t3, UNIT(3)(src), .Ll_exc_copy\@)
|
2006-12-12 19:22:06 +03:00
|
|
|
SUB len, len, 4*NBYTES
|
|
|
|
ADD src, src, 4*NBYTES
|
2014-01-17 14:48:46 +04:00
|
|
|
STORE(t0, UNIT(0)(dst), .Ls_exc\@)
|
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 20:07:24 +03:00
|
|
|
ADDC(t0, t1)
|
2014-01-17 14:48:46 +04:00
|
|
|
STORE(t1, UNIT(1)(dst), .Ls_exc\@)
|
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 20:07:24 +03:00
|
|
|
ADDC(sum, t0)
|
2014-01-17 14:48:46 +04:00
|
|
|
STORE(t2, UNIT(2)(dst), .Ls_exc\@)
|
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 20:07:24 +03:00
|
|
|
ADDC(t2, t3)
|
2014-01-17 14:48:46 +04:00
|
|
|
STORE(t3, UNIT(3)(dst), .Ls_exc\@)
|
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 20:07:24 +03:00
|
|
|
ADDC(sum, t2)
|
2007-10-23 15:43:25 +04:00
|
|
|
.set reorder /* DADDI_WAR */
|
|
|
|
ADD dst, dst, 4*NBYTES
|
2014-01-17 14:48:46 +04:00
|
|
|
beqz len, .Ldone\@
|
2007-10-23 15:43:25 +04:00
|
|
|
.set noreorder
|
2014-01-17 14:48:46 +04:00
|
|
|
.Lless_than_4units\@:
|
2006-12-12 19:22:06 +03:00
|
|
|
/*
|
|
|
|
* rem = len % NBYTES
|
|
|
|
*/
|
2014-01-17 14:48:46 +04:00
|
|
|
beq rem, len, .Lcopy_bytes\@
|
2006-12-12 19:22:06 +03:00
|
|
|
nop
|
|
|
|
1:
|
2014-01-17 14:48:46 +04:00
|
|
|
LOAD(t0, 0(src), .Ll_exc\@)
|
2006-12-12 19:22:06 +03:00
|
|
|
ADD src, src, NBYTES
|
|
|
|
SUB len, len, NBYTES
|
2014-01-17 14:48:46 +04:00
|
|
|
STORE(t0, 0(dst), .Ls_exc\@)
|
2006-12-12 19:22:06 +03:00
|
|
|
ADDC(sum, t0)
|
2007-10-23 15:43:25 +04:00
|
|
|
.set reorder /* DADDI_WAR */
|
|
|
|
ADD dst, dst, NBYTES
|
2006-12-12 19:22:06 +03:00
|
|
|
bne rem, len, 1b
|
2007-10-23 15:43:25 +04:00
|
|
|
.set noreorder
|
2006-12-12 19:22:06 +03:00
|
|
|
|
|
|
|
/*
|
|
|
|
* 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 15:59:30 +04:00
|
|
|
* mispredicts. Can't do an explicit LOAD dst,mask,or,STORE
|
2006-12-12 19:22:06 +03:00
|
|
|
* 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 14:48:46 +04:00
|
|
|
beqz len, .Ldone\@
|
2006-12-12 19:22:06 +03:00
|
|
|
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
|
2014-01-17 14:48:46 +04:00
|
|
|
LOAD(t0, 0(src), .Ll_exc\@)
|
2013-01-22 15:59:30 +04:00
|
|
|
SUB bits, bits, rem # bits = number of bits to discard
|
2006-12-12 19:22:06 +03:00
|
|
|
SHIFT_DISCARD t0, t0, bits
|
2014-01-17 14:48:46 +04:00
|
|
|
STREST(t0, -1(t1), .Ls_exc\@)
|
2006-12-12 19:22:06 +03:00
|
|
|
SHIFT_DISCARD_REVERT t0, t0, bits
|
|
|
|
.set reorder
|
|
|
|
ADDC(sum, t0)
|
2014-01-17 14:48:46 +04:00
|
|
|
b .Ldone\@
|
2006-12-12 19:22:06 +03:00
|
|
|
.set noreorder
|
2014-01-17 14:48:46 +04:00
|
|
|
.Ldst_unaligned\@:
|
2006-12-12 19:22:06 +03:00
|
|
|
/*
|
|
|
|
* 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
|
2014-01-17 14:48:46 +04:00
|
|
|
LDFIRST(t3, FIRST(0)(src), .Ll_exc\@)
|
2006-12-12 19:22:06 +03:00
|
|
|
ADD t2, zero, NBYTES
|
2014-01-17 14:48:46 +04:00
|
|
|
LDREST(t3, REST(0)(src), .Ll_exc_copy\@)
|
2006-12-12 19:22:06 +03:00
|
|
|
SUB t2, t2, t1 # t2 = number of bytes copied
|
|
|
|
xor match, t0, t1
|
2014-01-17 14:48:46 +04:00
|
|
|
STFIRST(t3, FIRST(0)(dst), .Ls_exc\@)
|
2006-12-12 19:22:06 +03:00
|
|
|
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 14:48:46 +04:00
|
|
|
beq len, t2, .Ldone\@
|
2006-12-12 19:22:06 +03:00
|
|
|
SUB len, len, t2
|
|
|
|
ADD dst, dst, t2
|
2014-01-17 14:48:46 +04:00
|
|
|
beqz match, .Lboth_aligned\@
|
2006-12-12 19:22:06 +03:00
|
|
|
ADD src, src, t2
|
|
|
|
|
2014-01-17 14:48:46 +04:00
|
|
|
.Lsrc_unaligned_dst_aligned\@:
|
2013-01-22 15:59:30 +04:00
|
|
|
SRL t0, len, LOG_NBYTES+2 # +2 for 4 units/iter
|
2014-01-17 14:48:46 +04:00
|
|
|
beqz t0, .Lcleanup_src_unaligned\@
|
2013-01-22 15:59:30 +04:00
|
|
|
and rem, len, (4*NBYTES-1) # rem = len % 4*NBYTES
|
2006-12-12 19:22:06 +03:00
|
|
|
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).
|
|
|
|
*/
|
2014-01-17 14:48:46 +04:00
|
|
|
LDFIRST(t0, FIRST(0)(src), .Ll_exc\@)
|
|
|
|
LDFIRST(t1, FIRST(1)(src), .Ll_exc_copy\@)
|
2013-01-22 15:59:30 +04:00
|
|
|
SUB len, len, 4*NBYTES
|
2014-01-17 14:48:46 +04:00
|
|
|
LDREST(t0, REST(0)(src), .Ll_exc_copy\@)
|
|
|
|
LDREST(t1, REST(1)(src), .Ll_exc_copy\@)
|
|
|
|
LDFIRST(t2, FIRST(2)(src), .Ll_exc_copy\@)
|
|
|
|
LDFIRST(t3, FIRST(3)(src), .Ll_exc_copy\@)
|
|
|
|
LDREST(t2, REST(2)(src), .Ll_exc_copy\@)
|
|
|
|
LDREST(t3, REST(3)(src), .Ll_exc_copy\@)
|
2006-12-12 19:22:06 +03:00
|
|
|
ADD src, src, 4*NBYTES
|
|
|
|
#ifdef CONFIG_CPU_SB1
|
|
|
|
nop # improves slotting
|
|
|
|
#endif
|
2014-01-17 14:48:46 +04:00
|
|
|
STORE(t0, UNIT(0)(dst), .Ls_exc\@)
|
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 20:07:24 +03:00
|
|
|
ADDC(t0, t1)
|
2014-01-17 14:48:46 +04:00
|
|
|
STORE(t1, UNIT(1)(dst), .Ls_exc\@)
|
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 20:07:24 +03:00
|
|
|
ADDC(sum, t0)
|
2014-01-17 14:48:46 +04:00
|
|
|
STORE(t2, UNIT(2)(dst), .Ls_exc\@)
|
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 20:07:24 +03:00
|
|
|
ADDC(t2, t3)
|
2014-01-17 14:48:46 +04:00
|
|
|
STORE(t3, UNIT(3)(dst), .Ls_exc\@)
|
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 20:07:24 +03:00
|
|
|
ADDC(sum, t2)
|
2007-10-23 15:43:25 +04:00
|
|
|
.set reorder /* DADDI_WAR */
|
|
|
|
ADD dst, dst, 4*NBYTES
|
2006-12-12 19:22:06 +03:00
|
|
|
bne len, rem, 1b
|
2007-10-23 15:43:25 +04:00
|
|
|
.set noreorder
|
2006-12-12 19:22:06 +03:00
|
|
|
|
2014-01-17 14:48:46 +04:00
|
|
|
.Lcleanup_src_unaligned\@:
|
|
|
|
beqz len, .Ldone\@
|
2006-12-12 19:22:06 +03:00
|
|
|
and rem, len, NBYTES-1 # rem = len % NBYTES
|
2014-01-17 14:48:46 +04:00
|
|
|
beq rem, len, .Lcopy_bytes\@
|
2006-12-12 19:22:06 +03:00
|
|
|
nop
|
|
|
|
1:
|
2014-01-17 14:48:46 +04:00
|
|
|
LDFIRST(t0, FIRST(0)(src), .Ll_exc\@)
|
|
|
|
LDREST(t0, REST(0)(src), .Ll_exc_copy\@)
|
2006-12-12 19:22:06 +03:00
|
|
|
ADD src, src, NBYTES
|
|
|
|
SUB len, len, NBYTES
|
2014-01-17 14:48:46 +04:00
|
|
|
STORE(t0, 0(dst), .Ls_exc\@)
|
2006-12-12 19:22:06 +03:00
|
|
|
ADDC(sum, t0)
|
2007-10-23 15:43:25 +04:00
|
|
|
.set reorder /* DADDI_WAR */
|
|
|
|
ADD dst, dst, NBYTES
|
2006-12-12 19:22:06 +03:00
|
|
|
bne len, rem, 1b
|
2007-10-23 15:43:25 +04:00
|
|
|
.set noreorder
|
2006-12-12 19:22:06 +03:00
|
|
|
|
2014-01-17 14:48:46 +04:00
|
|
|
.Lcopy_bytes_checklen\@:
|
|
|
|
beqz len, .Ldone\@
|
2006-12-12 19:22:06 +03:00
|
|
|
nop
|
2014-01-17 14:48:46 +04:00
|
|
|
.Lcopy_bytes\@:
|
2006-12-12 19:22:06 +03:00
|
|
|
/* 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 15:59:30 +04:00
|
|
|
li t3, SHIFT_START # shift
|
2008-01-29 13:14:59 +03:00
|
|
|
/* use .Ll_exc_copy here to return correct sum on fault */
|
2006-12-12 19:22:06 +03:00
|
|
|
#define COPY_BYTE(N) \
|
2014-01-17 14:48:46 +04:00
|
|
|
LOADBU(t0, N(src), .Ll_exc_copy\@); \
|
2006-12-12 19:22:06 +03:00
|
|
|
SUB len, len, 1; \
|
2014-01-17 14:48:46 +04:00
|
|
|
STOREB(t0, N(dst), .Ls_exc\@); \
|
2006-12-12 19:22:06 +03:00
|
|
|
SLLV t0, t0, t3; \
|
|
|
|
addu t3, SHIFT_INC; \
|
2014-01-17 14:48:46 +04:00
|
|
|
beqz len, .Lcopy_bytes_done\@; \
|
2006-12-12 19:22:06 +03:00
|
|
|
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
|
2014-01-17 14:48:46 +04:00
|
|
|
LOADBU(t0, NBYTES-2(src), .Ll_exc_copy\@)
|
2006-12-12 19:22:06 +03:00
|
|
|
SUB len, len, 1
|
2014-01-17 14:48:46 +04:00
|
|
|
STOREB(t0, NBYTES-2(dst), .Ls_exc\@)
|
2006-12-12 19:22:06 +03:00
|
|
|
SLLV t0, t0, t3
|
|
|
|
or t2, t0
|
2014-01-17 14:48:46 +04:00
|
|
|
.Lcopy_bytes_done\@:
|
2006-12-12 19:22:06 +03:00
|
|
|
ADDC(sum, t2)
|
2014-01-17 14:48:46 +04:00
|
|
|
.Ldone\@:
|
2006-12-12 19:22:06 +03:00
|
|
|
/* fold checksum */
|
2014-04-04 06:32:54 +04:00
|
|
|
.set push
|
|
|
|
.set noat
|
2006-12-12 19:22:06 +03:00
|
|
|
#ifdef USE_DOUBLE
|
|
|
|
dsll32 v1, sum, 0
|
|
|
|
daddu sum, v1
|
|
|
|
sltu v1, sum, v1
|
|
|
|
dsra32 sum, sum, 0
|
|
|
|
addu sum, v1
|
|
|
|
#endif
|
|
|
|
|
2019-10-20 17:43:13 +03:00
|
|
|
#if defined(CONFIG_CPU_MIPSR2) || defined(CONFIG_CPU_LOONGSON64)
|
2014-08-15 12:56:58 +04:00
|
|
|
.set push
|
|
|
|
.set arch=mips32r2
|
2008-10-11 19:18:53 +04:00
|
|
|
wsbh v1, sum
|
|
|
|
movn sum, v1, odd
|
2014-08-15 12:56:58 +04:00
|
|
|
.set pop
|
2008-10-11 19:18:53 +04:00
|
|
|
#else
|
|
|
|
beqz odd, 1f /* odd buffer alignment? */
|
|
|
|
lui v1, 0x00ff
|
|
|
|
addu v1, 0x00ff
|
|
|
|
and t0, sum, v1
|
|
|
|
sll t0, t0, 8
|
2006-12-12 19:22:06 +03:00
|
|
|
srl sum, sum, 8
|
2008-10-11 19:18:53 +04:00
|
|
|
and sum, sum, v1
|
|
|
|
or sum, sum, t0
|
2006-12-12 19:22:06 +03:00
|
|
|
1:
|
2008-10-11 19:18:53 +04:00
|
|
|
#endif
|
2014-04-04 06:32:54 +04:00
|
|
|
.set pop
|
2006-12-12 19:22:06 +03:00
|
|
|
.set reorder
|
2008-09-20 19:20:04 +04:00
|
|
|
ADDC32(sum, psum)
|
2006-12-12 19:22:06 +03:00
|
|
|
jr ra
|
|
|
|
.set noreorder
|
|
|
|
|
2014-01-17 14:48:46 +04:00
|
|
|
.Ll_exc_copy\@:
|
2006-12-12 19:22:06 +03:00
|
|
|
/*
|
|
|
|
* Copy bytes from src until faulting load address (or until a
|
|
|
|
* lb faults)
|
|
|
|
*
|
|
|
|
* When reached by a faulting LDFIRST/LDREST, THREAD_BUADDR($28)
|
|
|
|
* may be more than a byte beyond the last address.
|
|
|
|
* Hence, the lb below may get an exception.
|
|
|
|
*
|
|
|
|
* Assumes src < THREAD_BUADDR($28)
|
|
|
|
*/
|
2014-01-16 21:02:13 +04:00
|
|
|
LOADK t0, TI_TASK($28)
|
2006-12-12 19:22:06 +03:00
|
|
|
li t2, SHIFT_START
|
2014-01-16 21:02:13 +04:00
|
|
|
LOADK t0, THREAD_BUADDR(t0)
|
2006-12-12 19:22:06 +03:00
|
|
|
1:
|
2014-01-17 14:48:46 +04:00
|
|
|
LOADBU(t1, 0(src), .Ll_exc\@)
|
2006-12-12 19:22:06 +03:00
|
|
|
ADD src, src, 1
|
|
|
|
sb t1, 0(dst) # can't fault -- we're copy_from_user
|
|
|
|
SLLV t1, t1, t2
|
|
|
|
addu t2, SHIFT_INC
|
|
|
|
ADDC(sum, t1)
|
2007-10-23 15:43:25 +04:00
|
|
|
.set reorder /* DADDI_WAR */
|
|
|
|
ADD dst, dst, 1
|
2006-12-12 19:22:06 +03:00
|
|
|
bne src, t0, 1b
|
2007-10-23 15:43:25 +04:00
|
|
|
.set noreorder
|
2014-01-17 14:48:46 +04:00
|
|
|
.Ll_exc\@:
|
2014-01-16 21:02:13 +04:00
|
|
|
LOADK t0, TI_TASK($28)
|
2006-12-12 19:22:06 +03:00
|
|
|
nop
|
2014-01-16 21:02:13 +04:00
|
|
|
LOADK t0, THREAD_BUADDR(t0) # t0 is just past last good address
|
2006-12-12 19:22:06 +03:00
|
|
|
nop
|
|
|
|
SUB len, AT, t0 # len number of uncopied bytes
|
|
|
|
/*
|
|
|
|
* Here's where we rely on src and dst being incremented in tandem,
|
|
|
|
* See (3) above.
|
|
|
|
* dst += (fault addr - src) to put dst at first byte to clear
|
|
|
|
*/
|
|
|
|
ADD dst, t0 # compute start address in a1
|
|
|
|
SUB dst, src
|
|
|
|
/*
|
|
|
|
* Clear len bytes starting at dst. Can't call __bzero because it
|
|
|
|
* might modify len. An inefficient loop for these rare times...
|
|
|
|
*/
|
2007-10-23 15:43:25 +04:00
|
|
|
.set reorder /* DADDI_WAR */
|
|
|
|
SUB src, len, 1
|
2014-01-17 14:48:46 +04:00
|
|
|
beqz len, .Ldone\@
|
2007-10-23 15:43:25 +04:00
|
|
|
.set noreorder
|
2006-12-12 19:22:06 +03:00
|
|
|
1: sb zero, 0(dst)
|
|
|
|
ADD dst, dst, 1
|
2007-10-23 15:43:25 +04:00
|
|
|
.set push
|
|
|
|
.set noat
|
|
|
|
#ifndef CONFIG_CPU_DADDI_WORKAROUNDS
|
2006-12-12 19:22:06 +03:00
|
|
|
bnez src, 1b
|
|
|
|
SUB src, src, 1
|
2007-10-23 15:43:25 +04:00
|
|
|
#else
|
|
|
|
li v1, 1
|
|
|
|
bnez src, 1b
|
|
|
|
SUB src, src, v1
|
|
|
|
#endif
|
2006-12-12 19:22:06 +03:00
|
|
|
li v1, -EFAULT
|
2014-01-17 14:48:46 +04:00
|
|
|
b .Ldone\@
|
2006-12-12 19:22:06 +03:00
|
|
|
sw v1, (errptr)
|
|
|
|
|
2014-01-17 14:48:46 +04:00
|
|
|
.Ls_exc\@:
|
2006-12-12 19:22:06 +03:00
|
|
|
li v0, -1 /* invalid checksum */
|
|
|
|
li v1, -EFAULT
|
|
|
|
jr ra
|
|
|
|
sw v1, (errptr)
|
2007-10-23 15:43:25 +04:00
|
|
|
.set pop
|
2014-01-17 14:48:46 +04:00
|
|
|
.endm
|
|
|
|
|
|
|
|
LEAF(__csum_partial_copy_kernel)
|
2016-11-07 14:14:13 +03:00
|
|
|
EXPORT_SYMBOL(__csum_partial_copy_kernel)
|
2014-01-17 15:36:16 +04:00
|
|
|
#ifndef CONFIG_EVA
|
2014-01-17 14:48:46 +04:00
|
|
|
FEXPORT(__csum_partial_copy_to_user)
|
2016-11-07 14:14:13 +03:00
|
|
|
EXPORT_SYMBOL(__csum_partial_copy_to_user)
|
2014-01-17 14:48:46 +04:00
|
|
|
FEXPORT(__csum_partial_copy_from_user)
|
2016-11-07 14:14:13 +03:00
|
|
|
EXPORT_SYMBOL(__csum_partial_copy_from_user)
|
2014-01-17 15:36:16 +04:00
|
|
|
#endif
|
2014-01-17 14:48:46 +04:00
|
|
|
__BUILD_CSUM_PARTIAL_COPY_USER LEGACY_MODE USEROP USEROP 1
|
|
|
|
END(__csum_partial_copy_kernel)
|
2014-01-17 15:36:16 +04:00
|
|
|
|
|
|
|
#ifdef CONFIG_EVA
|
|
|
|
LEAF(__csum_partial_copy_to_user)
|
|
|
|
__BUILD_CSUM_PARTIAL_COPY_USER EVA_MODE KERNELOP USEROP 0
|
|
|
|
END(__csum_partial_copy_to_user)
|
|
|
|
|
|
|
|
LEAF(__csum_partial_copy_from_user)
|
|
|
|
__BUILD_CSUM_PARTIAL_COPY_USER EVA_MODE USEROP KERNELOP 0
|
|
|
|
END(__csum_partial_copy_from_user)
|
|
|
|
#endif
|