2023-04-10 12:00:48

by WANG Rui

[permalink] [raw]
Subject: [PATCH] LoongArch: Improve memory ops

We use a jump table to dispatch cases for short data lengths, and for long data
lengths, we use unaligned memory access instruction to process the first and
last 8 bytes. The data in the middle is processed at an aligned destination
address.

We have tested micro-bench[1] on a Loongson-3C5000 16-core machine (2.2GHz):

1. memset

| length | src offset | dst offset | speed before | speed after | % |
|--------|------------|------------|--------------|-------------|---------|
| 8 | 0 | 0 | 696.191 | 1518.785 | 118.16% |
| 8 | 0 | 1 | 696.325 | 1518.937 | 118.14% |
| 50 | 0 | 0 | 969.976 | 8053.902 | 730.32% |
| 50 | 0 | 1 | 970.034 | 8058.475 | 730.74% |
| 300 | 0 | 0 | 5876.612 | 16544.703 | 181.53% |
| 300 | 0 | 1 | 5030.849 | 16549.011 | 228.95% |
| 1200 | 0 | 0 | 11797.077 | 16752.137 | 42.00% |
| 1200 | 0 | 1 | 5687.141 | 16645.233 | 192.68% |
| 4000 | 0 | 0 | 15723.27 | 16761.557 | 6.60% |
| 4000 | 0 | 1 | 5906.114 | 16732.316 | 183.30% |
| 8000 | 0 | 0 | 16751.403 | 16770.002 | 0.11% |
| 8000 | 0 | 1 | 5995.449 | 16754.07 | 179.45% |

2. memcpy

| length | src offset | dst offset | speed before | speed after | % |
|--------|------------|------------|--------------|-------------|---------|
| 8 | 0 | 0 | 696.2 | 1670.605 | 139.96% |
| 8 | 0 | 1 | 696.325 | 1671.138 | 139.99% |
| 50 | 0 | 0 | 969.974 | 8724.999 | 799.51% |
| 50 | 0 | 1 | 970.032 | 8730.138 | 799.98% |
| 300 | 0 | 0 | 5564.662 | 16272.652 | 192.43% |
| 300 | 0 | 1 | 4670.436 | 14972.842 | 220.59% |
| 1200 | 0 | 0 | 10740.23 | 16751.728 | 55.97% |
| 1200 | 0 | 1 | 5027.741 | 14874.564 | 195.85% |
| 4000 | 0 | 0 | 15122.367 | 16737.642 | 10.68% |
| 4000 | 0 | 1 | 5536.918 | 14890.397 | 168.93% |
| 8000 | 0 | 0 | 16505.453 | 16553.543 | 0.29% |
| 8000 | 0 | 1 | 5821.619 | 14841.804 | 154.94% |

3. memmove

| length | src offset | dst offset | speed before | speed after | % |
|--------|------------|------------|--------------|-------------|---------|
| 8 | 0 | 0 | 982.693 | 1670.568 | 70.00% |
| 8 | 0 | 1 | 983.023 | 1671.174 | 70.00% |
| 50 | 0 | 0 | 1230.87 | 8727.625 | 609.06% |
| 50 | 0 | 1 | 1232.515 | 8730.138 | 608.32% |
| 300 | 0 | 0 | 6490.375 | 16296.993 | 151.09% |
| 300 | 0 | 1 | 4282.687 | 14972.842 | 249.61% |
| 1200 | 0 | 0 | 11742.755 | 16752.546 | 42.66% |
| 1200 | 0 | 1 | 5039.338 | 14872.951 | 195.14% |
| 4000 | 0 | 0 | 15467.786 | 16737.09 | 8.21% |
| 4000 | 0 | 1 | 5009.905 | 14890.542 | 197.22% |
| 8000 | 0 | 0 | 16489.664 | 16553.273 | 0.39% |
| 8000 | 0 | 1 | 5823.786 | 14858.646 | 155.14% |

* speed: MB/s
* length: byte

[1] https://github.com/heiher/mem-bench

Signed-off-by: WANG rui <[email protected]>
---
arch/loongarch/lib/clear_user.S | 137 +++++++++++++++--
arch/loongarch/lib/copy_user.S | 251 +++++++++++++++++++++++++-------
arch/loongarch/lib/memcpy.S | 147 +++++++++++++++----
arch/loongarch/lib/memmove.S | 120 ++++++++-------
arch/loongarch/lib/memset.S | 117 ++++++++++++---
5 files changed, 605 insertions(+), 167 deletions(-)

diff --git a/arch/loongarch/lib/clear_user.S b/arch/loongarch/lib/clear_user.S
index 2dc48e61a2c8..02f112b265cd 100644
--- a/arch/loongarch/lib/clear_user.S
+++ b/arch/loongarch/lib/clear_user.S
@@ -13,7 +13,14 @@

.irp to, 0, 1, 2, 3, 4, 5, 6, 7
.L_fixup_handle_\to\():
- addi.d a0, a1, (\to) * (-8)
+ sub.d a0, a2, a0
+ addi.d a0, a0, (\to) * (-8)
+ jr ra
+.endr
+
+.irp to, 0, 2, 4
+.L_fixup_handle_s\to\():
+ addi.d a0, a1, -\to
jr ra
.endr

@@ -44,7 +51,7 @@ SYM_FUNC_START(__clear_user_generic)
2: move a0, a1
jr ra

- _asm_extable 1b, .L_fixup_handle_0
+ _asm_extable 1b, .L_fixup_handle_s0
SYM_FUNC_END(__clear_user_generic)

/*
@@ -54,12 +61,22 @@ SYM_FUNC_END(__clear_user_generic)
* a1: size
*/
SYM_FUNC_START(__clear_user_fast)
- beqz a1, 10f
+ sltui t0, a1, 9
+ bnez t0, .Lsmall

- ori a2, zero, 64
- blt a1, a2, 9f
+ add.d a2, a0, a1
+0: st.d zero, a0, 0
+
+ /* align up address */
+ andi t1, a0, 7
+ sub.d a0, a0, t1
+ addi.d a0, a0, 8
+
+ addi.d a3, a2, -64
+ bgeu a0, a3, .Llt64

/* set 64 bytes at a time */
+.Lloop64:
1: st.d zero, a0, 0
2: st.d zero, a0, 8
3: st.d zero, a0, 16
@@ -68,24 +85,95 @@ SYM_FUNC_START(__clear_user_fast)
6: st.d zero, a0, 40
7: st.d zero, a0, 48
8: st.d zero, a0, 56
-
addi.d a0, a0, 64
- addi.d a1, a1, -64
- bge a1, a2, 1b
-
- beqz a1, 10f
+ bltu a0, a3, .Lloop64

/* set the remaining bytes */
-9: st.b zero, a0, 0
- addi.d a0, a0, 1
- addi.d a1, a1, -1
- bgt a1, zero, 9b
+.Llt64:
+ addi.d a3, a2, -32
+ bgeu a0, a3, .Llt32
+9: st.d zero, a0, 0
+10: st.d zero, a0, 8
+11: st.d zero, a0, 16
+12: st.d zero, a0, 24
+ addi.d a0, a0, 32
+
+.Llt32:
+ addi.d a3, a2, -16
+ bgeu a0, a3, .Llt16
+13: st.d zero, a0, 0
+14: st.d zero, a0, 8
+ addi.d a0, a0, 16
+
+.Llt16:
+ addi.d a3, a2, -8
+ bgeu a0, a3, .Llt8
+15: st.d zero, a0, 0
+
+.Llt8:
+16: st.d zero, a2, -8

/* return */
-10: move a0, a1
+ move a0, zero
+ jr ra
+
+ .align 4
+.Lsmall:
+ pcaddi t0, 4
+ slli.d a2, a1, 4
+ add.d t0, t0, a2
+ jr t0
+
+ .align 4
+ move a0, zero
+ jr ra
+
+ .align 4
+17: st.b zero, a0, 0
+ move a0, zero
+ jr ra
+
+ .align 4
+18: st.h zero, a0, 0
+ move a0, zero
+ jr ra
+
+ .align 4
+19: st.h zero, a0, 0
+20: st.b zero, a0, 2
+ move a0, zero
+ jr ra
+
+ .align 4
+21: st.w zero, a0, 0
+ move a0, zero
+ jr ra
+
+ .align 4
+22: st.w zero, a0, 0
+23: st.b zero, a0, 4
+ move a0, zero
+ jr ra
+
+ .align 4
+24: st.w zero, a0, 0
+25: st.h zero, a0, 4
+ move a0, zero
+ jr ra
+
+ .align 4
+26: st.w zero, a0, 0
+27: st.w zero, a0, 3
+ move a0, zero
+ jr ra
+
+ .align 4
+28: st.d zero, a0, 0
+ move a0, zero
jr ra

/* fixup and ex_table */
+ _asm_extable 0b, .L_fixup_handle_0
_asm_extable 1b, .L_fixup_handle_0
_asm_extable 2b, .L_fixup_handle_1
_asm_extable 3b, .L_fixup_handle_2
@@ -95,4 +183,23 @@ SYM_FUNC_START(__clear_user_fast)
_asm_extable 7b, .L_fixup_handle_6
_asm_extable 8b, .L_fixup_handle_7
_asm_extable 9b, .L_fixup_handle_0
+ _asm_extable 10b, .L_fixup_handle_1
+ _asm_extable 11b, .L_fixup_handle_2
+ _asm_extable 12b, .L_fixup_handle_3
+ _asm_extable 13b, .L_fixup_handle_0
+ _asm_extable 14b, .L_fixup_handle_1
+ _asm_extable 15b, .L_fixup_handle_0
+ _asm_extable 16b, .L_fixup_handle_1
+ _asm_extable 17b, .L_fixup_handle_s0
+ _asm_extable 18b, .L_fixup_handle_s0
+ _asm_extable 19b, .L_fixup_handle_s0
+ _asm_extable 20b, .L_fixup_handle_s2
+ _asm_extable 21b, .L_fixup_handle_s0
+ _asm_extable 22b, .L_fixup_handle_s0
+ _asm_extable 23b, .L_fixup_handle_s4
+ _asm_extable 24b, .L_fixup_handle_s0
+ _asm_extable 25b, .L_fixup_handle_s4
+ _asm_extable 26b, .L_fixup_handle_s0
+ _asm_extable 27b, .L_fixup_handle_s4
+ _asm_extable 28b, .L_fixup_handle_s0
SYM_FUNC_END(__clear_user_fast)
diff --git a/arch/loongarch/lib/copy_user.S b/arch/loongarch/lib/copy_user.S
index 55ac6020a1ad..b21f6d5d38f5 100644
--- a/arch/loongarch/lib/copy_user.S
+++ b/arch/loongarch/lib/copy_user.S
@@ -13,7 +13,14 @@

.irp to, 0, 1, 2, 3, 4, 5, 6, 7
.L_fixup_handle_\to\():
- addi.d a0, a2, (\to) * (-8)
+ sub.d a0, a2, a0
+ addi.d a0, a0, (\to) * (-8)
+ jr ra
+.endr
+
+.irp to, 0, 2, 4
+.L_fixup_handle_s\to\():
+ addi.d a0, a2, -\to
jr ra
.endr

@@ -47,8 +54,8 @@ SYM_FUNC_START(__copy_user_generic)
3: move a0, a2
jr ra

- _asm_extable 1b, .L_fixup_handle_0
- _asm_extable 2b, .L_fixup_handle_0
+ _asm_extable 1b, .L_fixup_handle_s0
+ _asm_extable 2b, .L_fixup_handle_s0
SYM_FUNC_END(__copy_user_generic)

/*
@@ -59,65 +66,209 @@ SYM_FUNC_END(__copy_user_generic)
* a2: n
*/
SYM_FUNC_START(__copy_user_fast)
- beqz a2, 19f
+ sltui t0, a2, 9
+ bnez t0, .Lsmall

- ori a3, zero, 64
- blt a2, a3, 17f
+ add.d a3, a1, a2
+ add.d a2, a0, a2
+0: ld.d t0, a1, 0
+1: st.d t0, a0, 0

- /* copy 64 bytes at a time */
-1: ld.d t0, a1, 0
-2: ld.d t1, a1, 8
-3: ld.d t2, a1, 16
-4: ld.d t3, a1, 24
-5: ld.d t4, a1, 32
-6: ld.d t5, a1, 40
-7: ld.d t6, a1, 48
-8: ld.d t7, a1, 56
-9: st.d t0, a0, 0
-10: st.d t1, a0, 8
-11: st.d t2, a0, 16
-12: st.d t3, a0, 24
-13: st.d t4, a0, 32
-14: st.d t5, a0, 40
-15: st.d t6, a0, 48
-16: st.d t7, a0, 56
+ /* align up destination address */
+ andi t1, a0, 7
+ sub.d t0, zero, t1
+ addi.d t0, t0, 8
+ add.d a1, a1, t0
+ add.d a0, a0, t0

- addi.d a0, a0, 64
- addi.d a1, a1, 64
- addi.d a2, a2, -64
- bge a2, a3, 1b
+ addi.d a4, a3, -64
+ bgeu a1, a4, .Llt64

- beqz a2, 19f
+ /* copy 64 bytes at a time */
+.Lloop64:
+2: ld.d t0, a1, 0
+3: ld.d t1, a1, 8
+4: ld.d t2, a1, 16
+5: ld.d t3, a1, 24
+6: ld.d t4, a1, 32
+7: ld.d t5, a1, 40
+8: ld.d t6, a1, 48
+9: ld.d t7, a1, 56
+ addi.d a1, a1, 64
+10: st.d t0, a0, 0
+11: st.d t1, a0, 8
+12: st.d t2, a0, 16
+13: st.d t3, a0, 24
+14: st.d t4, a0, 32
+15: st.d t5, a0, 40
+16: st.d t6, a0, 48
+17: st.d t7, a0, 56
+ addi.d a0, a0, 64
+ bltu a1, a4, .Lloop64

/* copy the remaining bytes */
-17: ld.b t0, a1, 0
-18: st.b t0, a0, 0
- addi.d a0, a0, 1
- addi.d a1, a1, 1
- addi.d a2, a2, -1
- bgt a2, zero, 17b
+.Llt64:
+ addi.d a4, a3, -32
+ bgeu a1, a4, .Llt32
+18: ld.d t0, a1, 0
+19: ld.d t1, a1, 8
+20: ld.d t2, a1, 16
+21: ld.d t3, a1, 24
+ addi.d a1, a1, 32
+22: st.d t0, a0, 0
+23: st.d t1, a0, 8
+24: st.d t2, a0, 16
+25: st.d t3, a0, 24
+ addi.d a0, a0, 32
+
+.Llt32:
+ addi.d a4, a3, -16
+ bgeu a1, a4, .Llt16
+26: ld.d t0, a1, 0
+27: ld.d t1, a1, 8
+ addi.d a1, a1, 16
+28: st.d t0, a0, 0
+29: st.d t1, a0, 8
+ addi.d a0, a0, 16
+
+.Llt16:
+ addi.d a4, a3, -8
+ bgeu a1, a4, .Llt8
+30: ld.d t0, a1, 0
+31: st.d t0, a0, 0
+
+.Llt8:
+32: ld.d t0, a3, -8
+33: st.d t0, a2, -8

/* return */
-19: move a0, a2
+ move a0, zero
+ jr ra
+
+ .align 5
+.Lsmall:
+ pcaddi t0, 8
+ slli.d a3, a2, 5
+ add.d t0, t0, a3
+ jr t0
+
+ .align 5
+ move a0, zero
+ jr ra
+
+ .align 5
+34: ld.b t0, a1, 0
+35: st.b t0, a0, 0
+ move a0, zero
+ jr ra
+
+ .align 5
+36: ld.h t0, a1, 0
+37: st.h t0, a0, 0
+ move a0, zero
+ jr ra
+
+ .align 5
+38: ld.h t0, a1, 0
+39: ld.b t1, a1, 2
+40: st.h t0, a0, 0
+41: st.b t1, a0, 2
+ move a0, zero
+ jr ra
+
+ .align 5
+42: ld.w t0, a1, 0
+43: st.w t0, a0, 0
+ move a0, zero
+ jr ra
+
+ .align 5
+44: ld.w t0, a1, 0
+45: ld.b t1, a1, 4
+46: st.w t0, a0, 0
+47: st.b t1, a0, 4
+ move a0, zero
+ jr ra
+
+ .align 5
+48: ld.w t0, a1, 0
+49: ld.h t1, a1, 4
+50: st.w t0, a0, 0
+51: st.h t1, a0, 4
+ move a0, zero
+ jr ra
+
+ .align 5
+52: ld.w t0, a1, 0
+53: ld.w t1, a1, 3
+54: st.w t0, a0, 0
+55: st.w t1, a0, 3
+ move a0, zero
+ jr ra
+
+ .align 5
+56: ld.d t0, a1, 0
+57: st.d t0, a0, 0
+ move a0, zero
jr ra

/* fixup and ex_table */
+ _asm_extable 0b, .L_fixup_handle_0
_asm_extable 1b, .L_fixup_handle_0
- _asm_extable 2b, .L_fixup_handle_1
- _asm_extable 3b, .L_fixup_handle_2
- _asm_extable 4b, .L_fixup_handle_3
- _asm_extable 5b, .L_fixup_handle_4
- _asm_extable 6b, .L_fixup_handle_5
- _asm_extable 7b, .L_fixup_handle_6
- _asm_extable 8b, .L_fixup_handle_7
+ _asm_extable 2b, .L_fixup_handle_0
+ _asm_extable 3b, .L_fixup_handle_0
+ _asm_extable 4b, .L_fixup_handle_0
+ _asm_extable 5b, .L_fixup_handle_0
+ _asm_extable 6b, .L_fixup_handle_0
+ _asm_extable 7b, .L_fixup_handle_0
+ _asm_extable 8b, .L_fixup_handle_0
_asm_extable 9b, .L_fixup_handle_0
- _asm_extable 10b, .L_fixup_handle_1
- _asm_extable 11b, .L_fixup_handle_2
- _asm_extable 12b, .L_fixup_handle_3
- _asm_extable 13b, .L_fixup_handle_4
- _asm_extable 14b, .L_fixup_handle_5
- _asm_extable 15b, .L_fixup_handle_6
- _asm_extable 16b, .L_fixup_handle_7
- _asm_extable 17b, .L_fixup_handle_0
+ _asm_extable 10b, .L_fixup_handle_0
+ _asm_extable 11b, .L_fixup_handle_1
+ _asm_extable 12b, .L_fixup_handle_2
+ _asm_extable 13b, .L_fixup_handle_3
+ _asm_extable 14b, .L_fixup_handle_4
+ _asm_extable 15b, .L_fixup_handle_5
+ _asm_extable 16b, .L_fixup_handle_6
+ _asm_extable 17b, .L_fixup_handle_7
_asm_extable 18b, .L_fixup_handle_0
+ _asm_extable 19b, .L_fixup_handle_0
+ _asm_extable 20b, .L_fixup_handle_0
+ _asm_extable 21b, .L_fixup_handle_0
+ _asm_extable 22b, .L_fixup_handle_0
+ _asm_extable 23b, .L_fixup_handle_1
+ _asm_extable 24b, .L_fixup_handle_2
+ _asm_extable 25b, .L_fixup_handle_3
+ _asm_extable 26b, .L_fixup_handle_0
+ _asm_extable 27b, .L_fixup_handle_0
+ _asm_extable 28b, .L_fixup_handle_0
+ _asm_extable 29b, .L_fixup_handle_1
+ _asm_extable 30b, .L_fixup_handle_0
+ _asm_extable 31b, .L_fixup_handle_0
+ _asm_extable 32b, .L_fixup_handle_0
+ _asm_extable 33b, .L_fixup_handle_1
+ _asm_extable 34b, .L_fixup_handle_s0
+ _asm_extable 35b, .L_fixup_handle_s0
+ _asm_extable 36b, .L_fixup_handle_s0
+ _asm_extable 37b, .L_fixup_handle_s0
+ _asm_extable 38b, .L_fixup_handle_s0
+ _asm_extable 39b, .L_fixup_handle_s0
+ _asm_extable 40b, .L_fixup_handle_s0
+ _asm_extable 41b, .L_fixup_handle_s2
+ _asm_extable 42b, .L_fixup_handle_s0
+ _asm_extable 43b, .L_fixup_handle_s0
+ _asm_extable 44b, .L_fixup_handle_s0
+ _asm_extable 45b, .L_fixup_handle_s0
+ _asm_extable 46b, .L_fixup_handle_s0
+ _asm_extable 47b, .L_fixup_handle_s4
+ _asm_extable 48b, .L_fixup_handle_s0
+ _asm_extable 49b, .L_fixup_handle_s0
+ _asm_extable 50b, .L_fixup_handle_s0
+ _asm_extable 51b, .L_fixup_handle_s4
+ _asm_extable 52b, .L_fixup_handle_s0
+ _asm_extable 53b, .L_fixup_handle_s0
+ _asm_extable 54b, .L_fixup_handle_s0
+ _asm_extable 55b, .L_fixup_handle_s4
+ _asm_extable 56b, .L_fixup_handle_s0
+ _asm_extable 57b, .L_fixup_handle_s0
SYM_FUNC_END(__copy_user_fast)
diff --git a/arch/loongarch/lib/memcpy.S b/arch/loongarch/lib/memcpy.S
index 3b7e1dec7109..39ce6621c704 100644
--- a/arch/loongarch/lib/memcpy.S
+++ b/arch/loongarch/lib/memcpy.S
@@ -44,6 +44,66 @@ SYM_FUNC_START(__memcpy_generic)
SYM_FUNC_END(__memcpy_generic)
_ASM_NOKPROBE(__memcpy_generic)

+ .align 5
+SYM_FUNC_START_NOALIGN(__memcpy_small)
+ pcaddi t0, 8
+ slli.d a2, a2, 5
+ add.d t0, t0, a2
+ jr t0
+
+ .align 5
+0: jr ra
+
+ .align 5
+1: ld.b t0, a1, 0
+ st.b t0, a0, 0
+ jr ra
+
+ .align 5
+2: ld.h t0, a1, 0
+ st.h t0, a0, 0
+ jr ra
+
+ .align 5
+3: ld.h t0, a1, 0
+ ld.b t1, a1, 2
+ st.h t0, a0, 0
+ st.b t1, a0, 2
+ jr ra
+
+ .align 5
+4: ld.w t0, a1, 0
+ st.w t0, a0, 0
+ jr ra
+
+ .align 5
+5: ld.w t0, a1, 0
+ ld.b t1, a1, 4
+ st.w t0, a0, 0
+ st.b t1, a0, 4
+ jr ra
+
+ .align 5
+6: ld.w t0, a1, 0
+ ld.h t1, a1, 4
+ st.w t0, a0, 0
+ st.h t1, a0, 4
+ jr ra
+
+ .align 5
+7: ld.w t0, a1, 0
+ ld.w t1, a1, 3
+ st.w t0, a0, 0
+ st.w t1, a0, 3
+ jr ra
+
+ .align 5
+8: ld.d t0, a1, 0
+ st.d t0, a0, 0
+ jr ra
+SYM_FUNC_END(__memcpy_small)
+_ASM_NOKPROBE(__memcpy_small)
+
/*
* void *__memcpy_fast(void *dst, const void *src, size_t n)
*
@@ -52,14 +112,27 @@ _ASM_NOKPROBE(__memcpy_generic)
* a2: n
*/
SYM_FUNC_START(__memcpy_fast)
- move a3, a0
- beqz a2, 3f
+ sltui t0, a2, 9
+ bnez t0, __memcpy_small
+
+ add.d a3, a1, a2
+ add.d a2, a0, a2
+ ld.d a6, a1, 0
+ ld.d a7, a3, -8
+
+ /* align up destination address */
+ andi t1, a0, 7
+ sub.d t0, zero, t1
+ addi.d t0, t0, 8
+ add.d a1, a1, t0
+ add.d a5, a0, t0

- ori a4, zero, 64
- blt a2, a4, 2f
+ addi.d a4, a3, -64
+ bgeu a1, a4, .Llt64

/* copy 64 bytes at a time */
-1: ld.d t0, a1, 0
+.Lloop64:
+ ld.d t0, a1, 0
ld.d t1, a1, 8
ld.d t2, a1, 16
ld.d t3, a1, 24
@@ -67,32 +140,54 @@ SYM_FUNC_START(__memcpy_fast)
ld.d t5, a1, 40
ld.d t6, a1, 48
ld.d t7, a1, 56
- st.d t0, a0, 0
- st.d t1, a0, 8
- st.d t2, a0, 16
- st.d t3, a0, 24
- st.d t4, a0, 32
- st.d t5, a0, 40
- st.d t6, a0, 48
- st.d t7, a0, 56
-
- addi.d a0, a0, 64
addi.d a1, a1, 64
- addi.d a2, a2, -64
- bge a2, a4, 1b
-
- beqz a2, 3f
+ st.d t0, a5, 0
+ st.d t1, a5, 8
+ st.d t2, a5, 16
+ st.d t3, a5, 24
+ st.d t4, a5, 32
+ st.d t5, a5, 40
+ st.d t6, a5, 48
+ st.d t7, a5, 56
+ addi.d a5, a5, 64
+ bltu a1, a4, .Lloop64

/* copy the remaining bytes */
-2: ld.b t0, a1, 0
- st.b t0, a0, 0
- addi.d a0, a0, 1
- addi.d a1, a1, 1
- addi.d a2, a2, -1
- bgt a2, zero, 2b
+.Llt64:
+ addi.d a4, a3, -32
+ bgeu a1, a4, .Llt32
+ ld.d t0, a1, 0
+ ld.d t1, a1, 8
+ ld.d t2, a1, 16
+ ld.d t3, a1, 24
+ addi.d a1, a1, 32
+ st.d t0, a5, 0
+ st.d t1, a5, 8
+ st.d t2, a5, 16
+ st.d t3, a5, 24
+ addi.d a5, a5, 32
+
+.Llt32:
+ addi.d a4, a3, -16
+ bgeu a1, a4, .Llt16
+ ld.d t0, a1, 0
+ ld.d t1, a1, 8
+ addi.d a1, a1, 16
+ st.d t0, a5, 0
+ st.d t1, a5, 8
+ addi.d a5, a5, 16
+
+.Llt16:
+ addi.d a4, a3, -8
+ bgeu a1, a4, .Llt8
+ ld.d t0, a1, 0
+ st.d t0, a5, 0
+
+.Llt8:
+ st.d a6, a0, 0
+ st.d a7, a2, -8

/* return */
-3: move a0, a3
jr ra
SYM_FUNC_END(__memcpy_fast)
_ASM_NOKPROBE(__memcpy_fast)
diff --git a/arch/loongarch/lib/memmove.S b/arch/loongarch/lib/memmove.S
index b796c3d6da05..45b725ba7867 100644
--- a/arch/loongarch/lib/memmove.S
+++ b/arch/loongarch/lib/memmove.S
@@ -11,23 +11,9 @@
#include <asm/regdef.h>

SYM_FUNC_START(memmove)
- blt a0, a1, 1f /* dst < src, memcpy */
- blt a1, a0, 3f /* src < dst, rmemcpy */
+ blt a0, a1, memcpy /* dst < src, memcpy */
+ blt a1, a0, rmemcpy /* src < dst, rmemcpy */
jr ra /* dst == src, return */
-
- /* if (src - dst) < 64, copy 1 byte at a time */
-1: ori a3, zero, 64
- sub.d t0, a1, a0
- blt t0, a3, 2f
- b memcpy
-2: b __memcpy_generic
-
- /* if (dst - src) < 64, copy 1 byte at a time */
-3: ori a3, zero, 64
- sub.d t0, a0, a1
- blt t0, a3, 4f
- b rmemcpy
-4: b __rmemcpy_generic
SYM_FUNC_END(memmove)
_ASM_NOKPROBE(memmove)

@@ -76,50 +62,80 @@ _ASM_NOKPROBE(__rmemcpy_generic)
* a2: n
*/
SYM_FUNC_START(__rmemcpy_fast)
- move a3, a0
- beqz a2, 3f
+ sltui t0, a2, 9
+ bnez t0, __memcpy_small

- add.d a0, a0, a2
- add.d a1, a1, a2
+ add.d a3, a1, a2
+ add.d a2, a0, a2
+ ld.d a6, a1, 0
+ ld.d a7, a3, -8
+
+ /* align up destination address */
+ andi t1, a2, 7
+ sub.d a3, a3, t1
+ sub.d a5, a2, t1

- ori a4, zero, 64
- blt a2, a4, 2f
+ addi.d a4, a1, 64
+ bgeu a4, a3, .Llt64

/* copy 64 bytes at a time */
-1: ld.d t0, a1, -8
- ld.d t1, a1, -16
- ld.d t2, a1, -24
- ld.d t3, a1, -32
- ld.d t4, a1, -40
- ld.d t5, a1, -48
- ld.d t6, a1, -56
- ld.d t7, a1, -64
- st.d t0, a0, -8
- st.d t1, a0, -16
- st.d t2, a0, -24
- st.d t3, a0, -32
- st.d t4, a0, -40
- st.d t5, a0, -48
- st.d t6, a0, -56
- st.d t7, a0, -64
-
- addi.d a0, a0, -64
- addi.d a1, a1, -64
- addi.d a2, a2, -64
- bge a2, a4, 1b
-
- beqz a2, 3f
+.Lloop64:
+ ld.d t0, a3, -8
+ ld.d t1, a3, -16
+ ld.d t2, a3, -24
+ ld.d t3, a3, -32
+ ld.d t4, a3, -40
+ ld.d t5, a3, -48
+ ld.d t6, a3, -56
+ ld.d t7, a3, -64
+ addi.d a3, a3, -64
+ st.d t0, a5, -8
+ st.d t1, a5, -16
+ st.d t2, a5, -24
+ st.d t3, a5, -32
+ st.d t4, a5, -40
+ st.d t5, a5, -48
+ st.d t6, a5, -56
+ st.d t7, a5, -64
+ addi.d a5, a5, -64
+ bltu a4, a3, .Lloop64

/* copy the remaining bytes */
-2: ld.b t0, a1, -1
- st.b t0, a0, -1
- addi.d a0, a0, -1
- addi.d a1, a1, -1
- addi.d a2, a2, -1
- bgt a2, zero, 2b
+.Llt64:
+ addi.d a4, a1, 32
+ bgeu a4, a3, .Llt32
+ ld.d t0, a3, -8
+ ld.d t1, a3, -16
+ ld.d t2, a3, -24
+ ld.d t3, a3, -32
+ addi.d a3, a3, -32
+ st.d t0, a5, -8
+ st.d t1, a5, -16
+ st.d t2, a5, -24
+ st.d t3, a5, -32
+ addi.d a5, a5, -32
+
+.Llt32:
+ addi.d a4, a1, 16
+ bgeu a4, a3, .Llt16
+ ld.d t0, a3, -8
+ ld.d t1, a3, -16
+ addi.d a3, a3, -16
+ st.d t0, a5, -8
+ st.d t1, a5, -16
+ addi.d a5, a5, -16
+
+.Llt16:
+ addi.d a4, a1, 8
+ bgeu a4, a3, .Llt8
+ ld.d t0, a3, -8
+ st.d t0, a5, -8
+
+.Llt8:
+ st.d a6, a0, 0
+ st.d a7, a2, -8

/* return */
-3: move a0, a3
jr ra
SYM_FUNC_END(__rmemcpy_fast)
_ASM_NOKPROBE(__rmemcpy_fast)
diff --git a/arch/loongarch/lib/memset.S b/arch/loongarch/lib/memset.S
index a9eb732ab2ad..c5836d916736 100644
--- a/arch/loongarch/lib/memset.S
+++ b/arch/loongarch/lib/memset.S
@@ -56,39 +56,108 @@ _ASM_NOKPROBE(__memset_generic)
* a2: n
*/
SYM_FUNC_START(__memset_fast)
- move a3, a0
- beqz a2, 3f
-
- ori a4, zero, 64
- blt a2, a4, 2f
-
/* fill a1 to 64 bits */
fill_to_64 a1

- /* set 64 bytes at a time */
-1: st.d a1, a0, 0
- st.d a1, a0, 8
- st.d a1, a0, 16
- st.d a1, a0, 24
- st.d a1, a0, 32
- st.d a1, a0, 40
- st.d a1, a0, 48
- st.d a1, a0, 56
+ sltui t0, a2, 9
+ bnez t0, .Lsmall

- addi.d a0, a0, 64
- addi.d a2, a2, -64
- bge a2, a4, 1b
+ add.d a2, a0, a2
+ st.d a1, a0, 0

- beqz a2, 3f
+ /* align up address */
+ andi t1, a0, 7
+ sub.d a3, a0, t1
+ addi.d a3, a3, 8
+
+ addi.d a4, a2, -64
+ bgeu a3, a4, .Llt64
+
+ /* set 64 bytes at a time */
+.Lloop64:
+ st.d a1, a3, 0
+ st.d a1, a3, 8
+ st.d a1, a3, 16
+ st.d a1, a3, 24
+ st.d a1, a3, 32
+ st.d a1, a3, 40
+ st.d a1, a3, 48
+ st.d a1, a3, 56
+ addi.d a3, a3, 64
+ bltu a3, a4, .Lloop64

/* set the remaining bytes */
-2: st.b a1, a0, 0
- addi.d a0, a0, 1
- addi.d a2, a2, -1
- bgt a2, zero, 2b
+.Llt64:
+ addi.d a4, a2, -32
+ bgeu a3, a4, .Llt32
+ st.d a1, a3, 0
+ st.d a1, a3, 8
+ st.d a1, a3, 16
+ st.d a1, a3, 24
+ addi.d a3, a3, 32
+
+.Llt32:
+ addi.d a4, a2, -16
+ bgeu a3, a4, .Llt16
+ st.d a1, a3, 0
+ st.d a1, a3, 8
+ addi.d a3, a3, 16
+
+.Llt16:
+ addi.d a4, a2, -8
+ bgeu a3, a4, .Llt8
+ st.d a1, a3, 0
+
+.Llt8:
+ st.d a1, a2, -8

/* return */
-3: move a0, a3
+ jr ra
+
+ .align 4
+.Lsmall:
+ pcaddi t0, 4
+ slli.d a2, a2, 4
+ add.d t0, t0, a2
+ jr t0
+
+ .align 4
+0: jr ra
+
+ .align 4
+1: st.b a1, a0, 0
+ jr ra
+
+ .align 4
+2: st.h a1, a0, 0
+ jr ra
+
+ .align 4
+3: st.h a1, a0, 0
+ st.b a1, a0, 2
+ jr ra
+
+ .align 4
+4: st.w a1, a0, 0
+ jr ra
+
+ .align 4
+5: st.w a1, a0, 0
+ st.b a1, a0, 4
+ jr ra
+
+ .align 4
+6: st.w a1, a0, 0
+ st.h a1, a0, 4
+ jr ra
+
+ .align 4
+7: st.w a1, a0, 0
+ st.w a1, a0, 3
+ jr ra
+
+ .align 4
+8: st.d a1, a0, 0
jr ra
SYM_FUNC_END(__memset_fast)
_ASM_NOKPROBE(__memset_fast)
--
2.40.0


2023-04-10 12:26:11

by Xi Ruoyao

[permalink] [raw]
Subject: Re: [PATCH] LoongArch: Improve memory ops

On Mon, 2023-04-10 at 19:57 +0800, WANG rui wrote:
> +       /* align up address */
> +       andi    t1, a0, 7
> +       sub.d   a0, a0, t1

bstrins.d a0, zero, 2, 0

Likewise for other aligning operations if the temporary is not used.

> +       addi.d  a0, a0, 8

--
Xi Ruoyao <[email protected]>
School of Aerospace Science and Technology, Xidian University

2023-04-10 12:55:02

by Xi Ruoyao

[permalink] [raw]
Subject: Re: [PATCH] LoongArch: Improve memory ops

On Mon, 2023-04-10 at 19:57 +0800, WANG rui wrote:
> We use a jump table to dispatch cases for short data lengths, and for long data
> lengths, we use unaligned memory access instruction to process the first and
> last 8 bytes. The data in the middle is processed at an aligned destination
> address.

Regarding these functions: we have -ffreestanding which is preventing
the compiler from optimizing for e.g. "memcpy(a, b, 8);" into a simple
ld.d/st.d pair. A explicit compiler built-in usage like

#define memcpy(a, b, c) __builtin_memcpy(a, b, c)

would allow the compiler to do such kind of optimization. Will this
improve the performance?

>
> We have tested micro-bench[1] on a Loongson-3C5000 16-core machine (2.2GHz):
>
> 1. memset
>
> > length | src offset | dst offset | speed before | speed after |
> > %       |
> > --------|------------|------------|--------------|-------------|----
> > -----|
> > 8      | 0          | 0          | 696.191      | 1518.785    |
> > 118.16% |
> > 8      | 0          | 1          | 696.325      | 1518.937    |
> > 118.14% |
> > 50     | 0          | 0          | 969.976      | 8053.902    |
> > 730.32% |
> > 50     | 0          | 1          | 970.034      | 8058.475    |
> > 730.74% |
> > 300    | 0          | 0          | 5876.612     | 16544.703   |
> > 181.53% |
> > 300    | 0          | 1          | 5030.849     | 16549.011   |
> > 228.95% |
> > 1200   | 0          | 0          | 11797.077    | 16752.137   |
> > 42.00%  |
> > 1200   | 0          | 1          | 5687.141     | 16645.233   |
> > 192.68% |
> > 4000   | 0          | 0          | 15723.27     | 16761.557   |
> > 6.60%   |
> > 4000   | 0          | 1          | 5906.114     | 16732.316   |
> > 183.30% |
> > 8000   | 0          | 0          | 16751.403    | 16770.002   |
> > 0.11%   |
> > 8000   | 0          | 1          | 5995.449     | 16754.07    |
> > 179.45% |
>
> 2. memcpy
>
> > length | src offset | dst offset | speed before | speed after |
> > %       |
> > --------|------------|------------|--------------|-------------|----
> > -----|
> > 8      | 0          | 0          | 696.2        | 1670.605    |
> > 139.96% |
> > 8      | 0          | 1          | 696.325      | 1671.138    |
> > 139.99% |
> > 50     | 0          | 0          | 969.974      | 8724.999    |
> > 799.51% |
> > 50     | 0          | 1          | 970.032      | 8730.138    |
> > 799.98% |
> > 300    | 0          | 0          | 5564.662     | 16272.652   |
> > 192.43% |
> > 300    | 0          | 1          | 4670.436     | 14972.842   |
> > 220.59% |
> > 1200   | 0          | 0          | 10740.23     | 16751.728   |
> > 55.97%  |
> > 1200   | 0          | 1          | 5027.741     | 14874.564   |
> > 195.85% |
> > 4000   | 0          | 0          | 15122.367    | 16737.642   |
> > 10.68%  |
> > 4000   | 0          | 1          | 5536.918     | 14890.397   |
> > 168.93% |
> > 8000   | 0          | 0          | 16505.453    | 16553.543   |
> > 0.29%   |
> > 8000   | 0          | 1          | 5821.619     | 14841.804   |
> > 154.94% |
>
> 3. memmove
>
> > length | src offset | dst offset | speed before | speed after |
> > %       |
> > --------|------------|------------|--------------|-------------|----
> > -----|
> > 8      | 0          | 0          | 982.693      | 1670.568    |
> > 70.00%  |
> > 8      | 0          | 1          | 983.023      | 1671.174    |
> > 70.00%  |
> > 50     | 0          | 0          | 1230.87      | 8727.625    |
> > 609.06% |
> > 50     | 0          | 1          | 1232.515     | 8730.138    |
> > 608.32% |
> > 300    | 0          | 0          | 6490.375     | 16296.993   |
> > 151.09% |
> > 300    | 0          | 1          | 4282.687     | 14972.842   |
> > 249.61% |
> > 1200   | 0          | 0          | 11742.755    | 16752.546   |
> > 42.66%  |
> > 1200   | 0          | 1          | 5039.338     | 14872.951   |
> > 195.14% |
> > 4000   | 0          | 0          | 15467.786    | 16737.09    |
> > 8.21%   |
> > 4000   | 0          | 1          | 5009.905     | 14890.542   |
> > 197.22% |
> > 8000   | 0          | 0          | 16489.664    | 16553.273   |
> > 0.39%   |
> > 8000   | 0          | 1          | 5823.786     | 14858.646   |
> > 155.14% |
>
> * speed: MB/s
> * length: byte
>
> [1] https://github.com/heiher/mem-bench
>
> Signed-off-by: WANG rui <[email protected]>
> ---
>  arch/loongarch/lib/clear_user.S | 137 +++++++++++++++--
>  arch/loongarch/lib/copy_user.S  | 251 +++++++++++++++++++++++++------
> -
>  arch/loongarch/lib/memcpy.S     | 147 +++++++++++++++----
>  arch/loongarch/lib/memmove.S    | 120 ++++++++-------
>  arch/loongarch/lib/memset.S     | 117 ++++++++++++---
>  5 files changed, 605 insertions(+), 167 deletions(-)
>
> diff --git a/arch/loongarch/lib/clear_user.S
> b/arch/loongarch/lib/clear_user.S
> index 2dc48e61a2c8..02f112b265cd 100644
> --- a/arch/loongarch/lib/clear_user.S
> +++ b/arch/loongarch/lib/clear_user.S
> @@ -13,7 +13,14 @@
>  
>  .irp to, 0, 1, 2, 3, 4, 5, 6, 7
>  .L_fixup_handle_\to\():
> -       addi.d  a0, a1, (\to) * (-8)
> +       sub.d   a0, a2, a0
> +       addi.d  a0, a0, (\to) * (-8)
> +       jr      ra
> +.endr
> +
> +.irp to, 0, 2, 4
> +.L_fixup_handle_s\to\():
> +       addi.d  a0, a1, -\to
>         jr      ra
>  .endr
>  
> @@ -44,7 +51,7 @@ SYM_FUNC_START(__clear_user_generic)
>  2:     move    a0, a1
>         jr      ra
>  
> -       _asm_extable 1b, .L_fixup_handle_0
> +       _asm_extable 1b, .L_fixup_handle_s0
>  SYM_FUNC_END(__clear_user_generic)
>  
>  /*
> @@ -54,12 +61,22 @@ SYM_FUNC_END(__clear_user_generic)
>   * a1: size
>   */
>  SYM_FUNC_START(__clear_user_fast)
> -       beqz    a1, 10f
> +       sltui   t0, a1, 9
> +       bnez    t0, .Lsmall
>  
> -       ori     a2, zero, 64
> -       blt     a1, a2, 9f
> +       add.d   a2, a0, a1
> +0:     st.d    zero, a0, 0
> +
> +       /* align up address */
> +       andi    t1, a0, 7
> +       sub.d   a0, a0, t1
> +       addi.d  a0, a0, 8
> +
> +       addi.d  a3, a2, -64
> +       bgeu    a0, a3, .Llt64
>  
>         /* set 64 bytes at a time */
> +.Lloop64:
>  1:     st.d    zero, a0, 0
>  2:     st.d    zero, a0, 8
>  3:     st.d    zero, a0, 16
> @@ -68,24 +85,95 @@ SYM_FUNC_START(__clear_user_fast)
>  6:     st.d    zero, a0, 40
>  7:     st.d    zero, a0, 48
>  8:     st.d    zero, a0, 56
> -
>         addi.d  a0, a0, 64
> -       addi.d  a1, a1, -64
> -       bge     a1, a2, 1b
> -
> -       beqz    a1, 10f
> +       bltu    a0, a3, .Lloop64
>  
>         /* set the remaining bytes */
> -9:     st.b    zero, a0, 0
> -       addi.d  a0, a0, 1
> -       addi.d  a1, a1, -1
> -       bgt     a1, zero, 9b
> +.Llt64:
> +       addi.d  a3, a2, -32
> +       bgeu    a0, a3, .Llt32
> +9:     st.d    zero, a0, 0
> +10:    st.d    zero, a0, 8
> +11:    st.d    zero, a0, 16
> +12:    st.d    zero, a0, 24
> +       addi.d  a0, a0, 32
> +
> +.Llt32:
> +       addi.d  a3, a2, -16
> +       bgeu    a0, a3, .Llt16
> +13:    st.d    zero, a0, 0
> +14:    st.d    zero, a0, 8
> +       addi.d  a0, a0, 16
> +
> +.Llt16:
> +       addi.d  a3, a2, -8
> +       bgeu    a0, a3, .Llt8
> +15:    st.d    zero, a0, 0
> +
> +.Llt8:
> +16:    st.d    zero, a2, -8
>  
>         /* return */
> -10:    move    a0, a1
> +       move    a0, zero
> +       jr      ra
> +
> +       .align  4
> +.Lsmall:
> +       pcaddi  t0, 4
> +       slli.d  a2, a1, 4
> +       add.d   t0, t0, a2
> +       jr      t0
> +
> +       .align  4
> +       move    a0, zero
> +       jr      ra
> +
> +       .align  4
> +17:    st.b    zero, a0, 0
> +       move    a0, zero
> +       jr      ra
> +
> +       .align  4
> +18:    st.h    zero, a0, 0
> +       move    a0, zero
> +       jr      ra
> +
> +       .align  4
> +19:    st.h    zero, a0, 0
> +20:    st.b    zero, a0, 2
> +       move    a0, zero
> +       jr      ra
> +
> +       .align  4
> +21:    st.w    zero, a0, 0
> +       move    a0, zero
> +       jr      ra
> +
> +       .align  4
> +22:    st.w    zero, a0, 0
> +23:    st.b    zero, a0, 4
> +       move    a0, zero
> +       jr      ra
> +
> +       .align  4
> +24:    st.w    zero, a0, 0
> +25:    st.h    zero, a0, 4
> +       move    a0, zero
> +       jr      ra
> +
> +       .align  4
> +26:    st.w    zero, a0, 0
> +27:    st.w    zero, a0, 3
> +       move    a0, zero
> +       jr      ra
> +
> +       .align  4
> +28:    st.d    zero, a0, 0
> +       move    a0, zero
>         jr      ra
>  
>         /* fixup and ex_table */
> +       _asm_extable 0b, .L_fixup_handle_0
>         _asm_extable 1b, .L_fixup_handle_0
>         _asm_extable 2b, .L_fixup_handle_1
>         _asm_extable 3b, .L_fixup_handle_2
> @@ -95,4 +183,23 @@ SYM_FUNC_START(__clear_user_fast)
>         _asm_extable 7b, .L_fixup_handle_6
>         _asm_extable 8b, .L_fixup_handle_7
>         _asm_extable 9b, .L_fixup_handle_0
> +       _asm_extable 10b, .L_fixup_handle_1
> +       _asm_extable 11b, .L_fixup_handle_2
> +       _asm_extable 12b, .L_fixup_handle_3
> +       _asm_extable 13b, .L_fixup_handle_0
> +       _asm_extable 14b, .L_fixup_handle_1
> +       _asm_extable 15b, .L_fixup_handle_0
> +       _asm_extable 16b, .L_fixup_handle_1
> +       _asm_extable 17b, .L_fixup_handle_s0
> +       _asm_extable 18b, .L_fixup_handle_s0
> +       _asm_extable 19b, .L_fixup_handle_s0
> +       _asm_extable 20b, .L_fixup_handle_s2
> +       _asm_extable 21b, .L_fixup_handle_s0
> +       _asm_extable 22b, .L_fixup_handle_s0
> +       _asm_extable 23b, .L_fixup_handle_s4
> +       _asm_extable 24b, .L_fixup_handle_s0
> +       _asm_extable 25b, .L_fixup_handle_s4
> +       _asm_extable 26b, .L_fixup_handle_s0
> +       _asm_extable 27b, .L_fixup_handle_s4
> +       _asm_extable 28b, .L_fixup_handle_s0
>  SYM_FUNC_END(__clear_user_fast)
> diff --git a/arch/loongarch/lib/copy_user.S
> b/arch/loongarch/lib/copy_user.S
> index 55ac6020a1ad..b21f6d5d38f5 100644
> --- a/arch/loongarch/lib/copy_user.S
> +++ b/arch/loongarch/lib/copy_user.S
> @@ -13,7 +13,14 @@
>  
>  .irp to, 0, 1, 2, 3, 4, 5, 6, 7
>  .L_fixup_handle_\to\():
> -       addi.d  a0, a2, (\to) * (-8)
> +       sub.d   a0, a2, a0
> +       addi.d  a0, a0, (\to) * (-8)
> +       jr      ra
> +.endr
> +
> +.irp to, 0, 2, 4
> +.L_fixup_handle_s\to\():
> +       addi.d  a0, a2, -\to
>         jr      ra
>  .endr
>  
> @@ -47,8 +54,8 @@ SYM_FUNC_START(__copy_user_generic)
>  3:     move    a0, a2
>         jr      ra
>  
> -       _asm_extable 1b, .L_fixup_handle_0
> -       _asm_extable 2b, .L_fixup_handle_0
> +       _asm_extable 1b, .L_fixup_handle_s0
> +       _asm_extable 2b, .L_fixup_handle_s0
>  SYM_FUNC_END(__copy_user_generic)
>  
>  /*
> @@ -59,65 +66,209 @@ SYM_FUNC_END(__copy_user_generic)
>   * a2: n
>   */
>  SYM_FUNC_START(__copy_user_fast)
> -       beqz    a2, 19f
> +       sltui   t0, a2, 9
> +       bnez    t0, .Lsmall
>  
> -       ori     a3, zero, 64
> -       blt     a2, a3, 17f
> +       add.d   a3, a1, a2
> +       add.d   a2, a0, a2
> +0:     ld.d    t0, a1, 0
> +1:     st.d    t0, a0, 0
>  
> -       /* copy 64 bytes at a time */
> -1:     ld.d    t0, a1, 0
> -2:     ld.d    t1, a1, 8
> -3:     ld.d    t2, a1, 16
> -4:     ld.d    t3, a1, 24
> -5:     ld.d    t4, a1, 32
> -6:     ld.d    t5, a1, 40
> -7:     ld.d    t6, a1, 48
> -8:     ld.d    t7, a1, 56
> -9:     st.d    t0, a0, 0
> -10:    st.d    t1, a0, 8
> -11:    st.d    t2, a0, 16
> -12:    st.d    t3, a0, 24
> -13:    st.d    t4, a0, 32
> -14:    st.d    t5, a0, 40
> -15:    st.d    t6, a0, 48
> -16:    st.d    t7, a0, 56
> +       /* align up destination address */
> +       andi    t1, a0, 7
> +       sub.d   t0, zero, t1
> +       addi.d  t0, t0, 8
> +       add.d   a1, a1, t0
> +       add.d   a0, a0, t0
>  
> -       addi.d  a0, a0, 64
> -       addi.d  a1, a1, 64
> -       addi.d  a2, a2, -64
> -       bge     a2, a3, 1b
> +       addi.d  a4, a3, -64
> +       bgeu    a1, a4, .Llt64
>  
> -       beqz    a2, 19f
> +       /* copy 64 bytes at a time */
> +.Lloop64:
> +2:     ld.d    t0, a1, 0
> +3:     ld.d    t1, a1, 8
> +4:     ld.d    t2, a1, 16
> +5:     ld.d    t3, a1, 24
> +6:     ld.d    t4, a1, 32
> +7:     ld.d    t5, a1, 40
> +8:     ld.d    t6, a1, 48
> +9:     ld.d    t7, a1, 56
> +       addi.d  a1, a1, 64
> +10:    st.d    t0, a0, 0
> +11:    st.d    t1, a0, 8
> +12:    st.d    t2, a0, 16
> +13:    st.d    t3, a0, 24
> +14:    st.d    t4, a0, 32
> +15:    st.d    t5, a0, 40
> +16:    st.d    t6, a0, 48
> +17:    st.d    t7, a0, 56
> +       addi.d  a0, a0, 64
> +       bltu    a1, a4, .Lloop64
>  
>         /* copy the remaining bytes */
> -17:    ld.b    t0, a1, 0
> -18:    st.b    t0, a0, 0
> -       addi.d  a0, a0, 1
> -       addi.d  a1, a1, 1
> -       addi.d  a2, a2, -1
> -       bgt     a2, zero, 17b
> +.Llt64:
> +       addi.d  a4, a3, -32
> +       bgeu    a1, a4, .Llt32
> +18:    ld.d    t0, a1, 0
> +19:    ld.d    t1, a1, 8
> +20:    ld.d    t2, a1, 16
> +21:    ld.d    t3, a1, 24
> +       addi.d  a1, a1, 32
> +22:    st.d    t0, a0, 0
> +23:    st.d    t1, a0, 8
> +24:    st.d    t2, a0, 16
> +25:    st.d    t3, a0, 24
> +       addi.d  a0, a0, 32
> +
> +.Llt32:
> +       addi.d  a4, a3, -16
> +       bgeu    a1, a4, .Llt16
> +26:    ld.d    t0, a1, 0
> +27:    ld.d    t1, a1, 8
> +       addi.d  a1, a1, 16
> +28:    st.d    t0, a0, 0
> +29:    st.d    t1, a0, 8
> +       addi.d  a0, a0, 16
> +
> +.Llt16:
> +       addi.d  a4, a3, -8
> +       bgeu    a1, a4, .Llt8
> +30:    ld.d    t0, a1, 0
> +31:    st.d    t0, a0, 0
> +
> +.Llt8:
> +32:    ld.d    t0, a3, -8
> +33:    st.d    t0, a2, -8
>  
>         /* return */
> -19:    move    a0, a2
> +       move    a0, zero
> +       jr      ra
> +
> +       .align  5
> +.Lsmall:
> +       pcaddi  t0, 8
> +       slli.d  a3, a2, 5
> +       add.d   t0, t0, a3
> +       jr      t0
> +
> +       .align  5
> +       move    a0, zero
> +       jr      ra
> +
> +       .align  5
> +34:    ld.b    t0, a1, 0
> +35:    st.b    t0, a0, 0
> +       move    a0, zero
> +       jr      ra
> +
> +       .align  5
> +36:    ld.h    t0, a1, 0
> +37:    st.h    t0, a0, 0
> +       move    a0, zero
> +       jr      ra
> +
> +       .align  5
> +38:    ld.h    t0, a1, 0
> +39:    ld.b    t1, a1, 2
> +40:    st.h    t0, a0, 0
> +41:    st.b    t1, a0, 2
> +       move    a0, zero
> +       jr      ra
> +
> +       .align  5
> +42:    ld.w    t0, a1, 0
> +43:    st.w    t0, a0, 0
> +       move    a0, zero
> +       jr      ra
> +
> +       .align  5
> +44:    ld.w    t0, a1, 0
> +45:    ld.b    t1, a1, 4
> +46:    st.w    t0, a0, 0
> +47:    st.b    t1, a0, 4
> +       move    a0, zero
> +       jr      ra
> +
> +       .align  5
> +48:    ld.w    t0, a1, 0
> +49:    ld.h    t1, a1, 4
> +50:    st.w    t0, a0, 0
> +51:    st.h    t1, a0, 4
> +       move    a0, zero
> +       jr      ra
> +
> +       .align  5
> +52:    ld.w    t0, a1, 0
> +53:    ld.w    t1, a1, 3
> +54:    st.w    t0, a0, 0
> +55:    st.w    t1, a0, 3
> +       move    a0, zero
> +       jr      ra
> +
> +       .align  5
> +56:    ld.d    t0, a1, 0
> +57:    st.d    t0, a0, 0
> +       move    a0, zero
>         jr      ra
>  
>         /* fixup and ex_table */
> +       _asm_extable 0b, .L_fixup_handle_0
>         _asm_extable 1b, .L_fixup_handle_0
> -       _asm_extable 2b, .L_fixup_handle_1
> -       _asm_extable 3b, .L_fixup_handle_2
> -       _asm_extable 4b, .L_fixup_handle_3
> -       _asm_extable 5b, .L_fixup_handle_4
> -       _asm_extable 6b, .L_fixup_handle_5
> -       _asm_extable 7b, .L_fixup_handle_6
> -       _asm_extable 8b, .L_fixup_handle_7
> +       _asm_extable 2b, .L_fixup_handle_0
> +       _asm_extable 3b, .L_fixup_handle_0
> +       _asm_extable 4b, .L_fixup_handle_0
> +       _asm_extable 5b, .L_fixup_handle_0
> +       _asm_extable 6b, .L_fixup_handle_0
> +       _asm_extable 7b, .L_fixup_handle_0
> +       _asm_extable 8b, .L_fixup_handle_0
>         _asm_extable 9b, .L_fixup_handle_0
> -       _asm_extable 10b, .L_fixup_handle_1
> -       _asm_extable 11b, .L_fixup_handle_2
> -       _asm_extable 12b, .L_fixup_handle_3
> -       _asm_extable 13b, .L_fixup_handle_4
> -       _asm_extable 14b, .L_fixup_handle_5
> -       _asm_extable 15b, .L_fixup_handle_6
> -       _asm_extable 16b, .L_fixup_handle_7
> -       _asm_extable 17b, .L_fixup_handle_0
> +       _asm_extable 10b, .L_fixup_handle_0
> +       _asm_extable 11b, .L_fixup_handle_1
> +       _asm_extable 12b, .L_fixup_handle_2
> +       _asm_extable 13b, .L_fixup_handle_3
> +       _asm_extable 14b, .L_fixup_handle_4
> +       _asm_extable 15b, .L_fixup_handle_5
> +       _asm_extable 16b, .L_fixup_handle_6
> +       _asm_extable 17b, .L_fixup_handle_7
>         _asm_extable 18b, .L_fixup_handle_0
> +       _asm_extable 19b, .L_fixup_handle_0
> +       _asm_extable 20b, .L_fixup_handle_0
> +       _asm_extable 21b, .L_fixup_handle_0
> +       _asm_extable 22b, .L_fixup_handle_0
> +       _asm_extable 23b, .L_fixup_handle_1
> +       _asm_extable 24b, .L_fixup_handle_2
> +       _asm_extable 25b, .L_fixup_handle_3
> +       _asm_extable 26b, .L_fixup_handle_0
> +       _asm_extable 27b, .L_fixup_handle_0
> +       _asm_extable 28b, .L_fixup_handle_0
> +       _asm_extable 29b, .L_fixup_handle_1
> +       _asm_extable 30b, .L_fixup_handle_0
> +       _asm_extable 31b, .L_fixup_handle_0
> +       _asm_extable 32b, .L_fixup_handle_0
> +       _asm_extable 33b, .L_fixup_handle_1
> +       _asm_extable 34b, .L_fixup_handle_s0
> +       _asm_extable 35b, .L_fixup_handle_s0
> +       _asm_extable 36b, .L_fixup_handle_s0
> +       _asm_extable 37b, .L_fixup_handle_s0
> +       _asm_extable 38b, .L_fixup_handle_s0
> +       _asm_extable 39b, .L_fixup_handle_s0
> +       _asm_extable 40b, .L_fixup_handle_s0
> +       _asm_extable 41b, .L_fixup_handle_s2
> +       _asm_extable 42b, .L_fixup_handle_s0
> +       _asm_extable 43b, .L_fixup_handle_s0
> +       _asm_extable 44b, .L_fixup_handle_s0
> +       _asm_extable 45b, .L_fixup_handle_s0
> +       _asm_extable 46b, .L_fixup_handle_s0
> +       _asm_extable 47b, .L_fixup_handle_s4
> +       _asm_extable 48b, .L_fixup_handle_s0
> +       _asm_extable 49b, .L_fixup_handle_s0
> +       _asm_extable 50b, .L_fixup_handle_s0
> +       _asm_extable 51b, .L_fixup_handle_s4
> +       _asm_extable 52b, .L_fixup_handle_s0
> +       _asm_extable 53b, .L_fixup_handle_s0
> +       _asm_extable 54b, .L_fixup_handle_s0
> +       _asm_extable 55b, .L_fixup_handle_s4
> +       _asm_extable 56b, .L_fixup_handle_s0
> +       _asm_extable 57b, .L_fixup_handle_s0
>  SYM_FUNC_END(__copy_user_fast)
> diff --git a/arch/loongarch/lib/memcpy.S b/arch/loongarch/lib/memcpy.S
> index 3b7e1dec7109..39ce6621c704 100644
> --- a/arch/loongarch/lib/memcpy.S
> +++ b/arch/loongarch/lib/memcpy.S
> @@ -44,6 +44,66 @@ SYM_FUNC_START(__memcpy_generic)
>  SYM_FUNC_END(__memcpy_generic)
>  _ASM_NOKPROBE(__memcpy_generic)
>  
> +       .align  5
> +SYM_FUNC_START_NOALIGN(__memcpy_small)
> +       pcaddi  t0, 8
> +       slli.d  a2, a2, 5
> +       add.d   t0, t0, a2
> +       jr      t0
> +
> +       .align  5
> +0:     jr      ra
> +
> +       .align  5
> +1:     ld.b    t0, a1, 0
> +       st.b    t0, a0, 0
> +       jr      ra
> +
> +       .align  5
> +2:     ld.h    t0, a1, 0
> +       st.h    t0, a0, 0
> +       jr      ra
> +
> +       .align  5
> +3:     ld.h    t0, a1, 0
> +       ld.b    t1, a1, 2
> +       st.h    t0, a0, 0
> +       st.b    t1, a0, 2
> +       jr      ra
> +
> +       .align  5
> +4:     ld.w    t0, a1, 0
> +       st.w    t0, a0, 0
> +       jr      ra
> +
> +       .align  5
> +5:     ld.w    t0, a1, 0
> +       ld.b    t1, a1, 4
> +       st.w    t0, a0, 0
> +       st.b    t1, a0, 4
> +       jr      ra
> +
> +       .align  5
> +6:     ld.w    t0, a1, 0
> +       ld.h    t1, a1, 4
> +       st.w    t0, a0, 0
> +       st.h    t1, a0, 4
> +       jr      ra
> +
> +       .align  5
> +7:     ld.w    t0, a1, 0
> +       ld.w    t1, a1, 3
> +       st.w    t0, a0, 0
> +       st.w    t1, a0, 3
> +       jr      ra
> +
> +       .align  5
> +8:     ld.d    t0, a1, 0
> +       st.d    t0, a0, 0
> +       jr      ra
> +SYM_FUNC_END(__memcpy_small)
> +_ASM_NOKPROBE(__memcpy_small)
> +
>  /*
>   * void *__memcpy_fast(void *dst, const void *src, size_t n)
>   *
> @@ -52,14 +112,27 @@ _ASM_NOKPROBE(__memcpy_generic)
>   * a2: n
>   */
>  SYM_FUNC_START(__memcpy_fast)
> -       move    a3, a0
> -       beqz    a2, 3f
> +       sltui   t0, a2, 9
> +       bnez    t0, __memcpy_small
> +
> +       add.d   a3, a1, a2
> +       add.d   a2, a0, a2
> +       ld.d    a6, a1, 0
> +       ld.d    a7, a3, -8
> +
> +       /* align up destination address */
> +       andi    t1, a0, 7
> +       sub.d   t0, zero, t1
> +       addi.d  t0, t0, 8
> +       add.d   a1, a1, t0
> +       add.d   a5, a0, t0
>  
> -       ori     a4, zero, 64
> -       blt     a2, a4, 2f
> +       addi.d  a4, a3, -64
> +       bgeu    a1, a4, .Llt64
>  
>         /* copy 64 bytes at a time */
> -1:     ld.d    t0, a1, 0
> +.Lloop64:
> +       ld.d    t0, a1, 0
>         ld.d    t1, a1, 8
>         ld.d    t2, a1, 16
>         ld.d    t3, a1, 24
> @@ -67,32 +140,54 @@ SYM_FUNC_START(__memcpy_fast)
>         ld.d    t5, a1, 40
>         ld.d    t6, a1, 48
>         ld.d    t7, a1, 56
> -       st.d    t0, a0, 0
> -       st.d    t1, a0, 8
> -       st.d    t2, a0, 16
> -       st.d    t3, a0, 24
> -       st.d    t4, a0, 32
> -       st.d    t5, a0, 40
> -       st.d    t6, a0, 48
> -       st.d    t7, a0, 56
> -
> -       addi.d  a0, a0, 64
>         addi.d  a1, a1, 64
> -       addi.d  a2, a2, -64
> -       bge     a2, a4, 1b
> -
> -       beqz    a2, 3f
> +       st.d    t0, a5, 0
> +       st.d    t1, a5, 8
> +       st.d    t2, a5, 16
> +       st.d    t3, a5, 24
> +       st.d    t4, a5, 32
> +       st.d    t5, a5, 40
> +       st.d    t6, a5, 48
> +       st.d    t7, a5, 56
> +       addi.d  a5, a5, 64
> +       bltu    a1, a4, .Lloop64
>  
>         /* copy the remaining bytes */
> -2:     ld.b    t0, a1, 0
> -       st.b    t0, a0, 0
> -       addi.d  a0, a0, 1
> -       addi.d  a1, a1, 1
> -       addi.d  a2, a2, -1
> -       bgt     a2, zero, 2b
> +.Llt64:
> +       addi.d  a4, a3, -32
> +       bgeu    a1, a4, .Llt32
> +       ld.d    t0, a1, 0
> +       ld.d    t1, a1, 8
> +       ld.d    t2, a1, 16
> +       ld.d    t3, a1, 24
> +       addi.d  a1, a1, 32
> +       st.d    t0, a5, 0
> +       st.d    t1, a5, 8
> +       st.d    t2, a5, 16
> +       st.d    t3, a5, 24
> +       addi.d  a5, a5, 32
> +
> +.Llt32:
> +       addi.d  a4, a3, -16
> +       bgeu    a1, a4, .Llt16
> +       ld.d    t0, a1, 0
> +       ld.d    t1, a1, 8
> +       addi.d  a1, a1, 16
> +       st.d    t0, a5, 0
> +       st.d    t1, a5, 8
> +       addi.d  a5, a5, 16
> +
> +.Llt16:
> +       addi.d  a4, a3, -8
> +       bgeu    a1, a4, .Llt8
> +       ld.d    t0, a1, 0
> +       st.d    t0, a5, 0
> +
> +.Llt8:
> +       st.d    a6, a0, 0
> +       st.d    a7, a2, -8
>  
>         /* return */
> -3:     move    a0, a3
>         jr      ra
>  SYM_FUNC_END(__memcpy_fast)
>  _ASM_NOKPROBE(__memcpy_fast)
> diff --git a/arch/loongarch/lib/memmove.S
> b/arch/loongarch/lib/memmove.S
> index b796c3d6da05..45b725ba7867 100644
> --- a/arch/loongarch/lib/memmove.S
> +++ b/arch/loongarch/lib/memmove.S
> @@ -11,23 +11,9 @@
>  #include <asm/regdef.h>
>  
>  SYM_FUNC_START(memmove)
> -       blt     a0, a1, 1f      /* dst < src, memcpy */
> -       blt     a1, a0, 3f      /* src < dst, rmemcpy */
> +       blt     a0, a1, memcpy  /* dst < src, memcpy */
> +       blt     a1, a0, rmemcpy /* src < dst, rmemcpy */
>         jr      ra              /* dst == src, return */
> -
> -       /* if (src - dst) < 64, copy 1 byte at a time */
> -1:     ori     a3, zero, 64
> -       sub.d   t0, a1, a0
> -       blt     t0, a3, 2f
> -       b       memcpy
> -2:     b       __memcpy_generic
> -
> -       /* if (dst - src) < 64, copy 1 byte at a time */
> -3:     ori     a3, zero, 64
> -       sub.d   t0, a0, a1
> -       blt     t0, a3, 4f
> -       b       rmemcpy
> -4:     b       __rmemcpy_generic
>  SYM_FUNC_END(memmove)
>  _ASM_NOKPROBE(memmove)
>  
> @@ -76,50 +62,80 @@ _ASM_NOKPROBE(__rmemcpy_generic)
>   * a2: n
>   */
>  SYM_FUNC_START(__rmemcpy_fast)
> -       move    a3, a0
> -       beqz    a2, 3f
> +       sltui   t0, a2, 9
> +       bnez    t0, __memcpy_small
>  
> -       add.d   a0, a0, a2
> -       add.d   a1, a1, a2
> +       add.d   a3, a1, a2
> +       add.d   a2, a0, a2
> +       ld.d    a6, a1, 0
> +       ld.d    a7, a3, -8
> +
> +       /* align up destination address */
> +       andi    t1, a2, 7
> +       sub.d   a3, a3, t1
> +       sub.d   a5, a2, t1
>  
> -       ori     a4, zero, 64
> -       blt     a2, a4, 2f
> +       addi.d  a4, a1, 64
> +       bgeu    a4, a3, .Llt64
>  
>         /* copy 64 bytes at a time */
> -1:     ld.d    t0, a1, -8
> -       ld.d    t1, a1, -16
> -       ld.d    t2, a1, -24
> -       ld.d    t3, a1, -32
> -       ld.d    t4, a1, -40
> -       ld.d    t5, a1, -48
> -       ld.d    t6, a1, -56
> -       ld.d    t7, a1, -64
> -       st.d    t0, a0, -8
> -       st.d    t1, a0, -16
> -       st.d    t2, a0, -24
> -       st.d    t3, a0, -32
> -       st.d    t4, a0, -40
> -       st.d    t5, a0, -48
> -       st.d    t6, a0, -56
> -       st.d    t7, a0, -64
> -
> -       addi.d  a0, a0, -64
> -       addi.d  a1, a1, -64
> -       addi.d  a2, a2, -64
> -       bge     a2, a4, 1b
> -
> -       beqz    a2, 3f
> +.Lloop64:
> +       ld.d    t0, a3, -8
> +       ld.d    t1, a3, -16
> +       ld.d    t2, a3, -24
> +       ld.d    t3, a3, -32
> +       ld.d    t4, a3, -40
> +       ld.d    t5, a3, -48
> +       ld.d    t6, a3, -56
> +       ld.d    t7, a3, -64
> +       addi.d  a3, a3, -64
> +       st.d    t0, a5, -8
> +       st.d    t1, a5, -16
> +       st.d    t2, a5, -24
> +       st.d    t3, a5, -32
> +       st.d    t4, a5, -40
> +       st.d    t5, a5, -48
> +       st.d    t6, a5, -56
> +       st.d    t7, a5, -64
> +       addi.d  a5, a5, -64
> +       bltu    a4, a3, .Lloop64
>  
>         /* copy the remaining bytes */
> -2:     ld.b    t0, a1, -1
> -       st.b    t0, a0, -1
> -       addi.d  a0, a0, -1
> -       addi.d  a1, a1, -1
> -       addi.d  a2, a2, -1
> -       bgt     a2, zero, 2b
> +.Llt64:
> +       addi.d  a4, a1, 32
> +       bgeu    a4, a3, .Llt32
> +       ld.d    t0, a3, -8
> +       ld.d    t1, a3, -16
> +       ld.d    t2, a3, -24
> +       ld.d    t3, a3, -32
> +       addi.d  a3, a3, -32
> +       st.d    t0, a5, -8
> +       st.d    t1, a5, -16
> +       st.d    t2, a5, -24
> +       st.d    t3, a5, -32
> +       addi.d  a5, a5, -32
> +
> +.Llt32:
> +       addi.d  a4, a1, 16
> +       bgeu    a4, a3, .Llt16
> +       ld.d    t0, a3, -8
> +       ld.d    t1, a3, -16
> +       addi.d  a3, a3, -16
> +       st.d    t0, a5, -8
> +       st.d    t1, a5, -16
> +       addi.d  a5, a5, -16
> +
> +.Llt16:
> +       addi.d  a4, a1, 8
> +       bgeu    a4, a3, .Llt8
> +       ld.d    t0, a3, -8
> +       st.d    t0, a5, -8
> +
> +.Llt8:
> +       st.d    a6, a0, 0
> +       st.d    a7, a2, -8
>  
>         /* return */
> -3:     move    a0, a3
>         jr      ra
>  SYM_FUNC_END(__rmemcpy_fast)
>  _ASM_NOKPROBE(__rmemcpy_fast)
> diff --git a/arch/loongarch/lib/memset.S b/arch/loongarch/lib/memset.S
> index a9eb732ab2ad..c5836d916736 100644
> --- a/arch/loongarch/lib/memset.S
> +++ b/arch/loongarch/lib/memset.S
> @@ -56,39 +56,108 @@ _ASM_NOKPROBE(__memset_generic)
>   * a2: n
>   */
>  SYM_FUNC_START(__memset_fast)
> -       move    a3, a0
> -       beqz    a2, 3f
> -
> -       ori     a4, zero, 64
> -       blt     a2, a4, 2f
> -
>         /* fill a1 to 64 bits */
>         fill_to_64 a1
>  
> -       /* set 64 bytes at a time */
> -1:     st.d    a1, a0, 0
> -       st.d    a1, a0, 8
> -       st.d    a1, a0, 16
> -       st.d    a1, a0, 24
> -       st.d    a1, a0, 32
> -       st.d    a1, a0, 40
> -       st.d    a1, a0, 48
> -       st.d    a1, a0, 56
> +       sltui   t0, a2, 9
> +       bnez    t0, .Lsmall
>  
> -       addi.d  a0, a0, 64
> -       addi.d  a2, a2, -64
> -       bge     a2, a4, 1b
> +       add.d   a2, a0, a2
> +       st.d    a1, a0, 0
>  
> -       beqz    a2, 3f
> +       /* align up address */
> +       andi    t1, a0, 7
> +       sub.d   a3, a0, t1
> +       addi.d  a3, a3, 8
> +
> +       addi.d  a4, a2, -64
> +       bgeu    a3, a4, .Llt64
> +
> +       /* set 64 bytes at a time */
> +.Lloop64:
> +       st.d    a1, a3, 0
> +       st.d    a1, a3, 8
> +       st.d    a1, a3, 16
> +       st.d    a1, a3, 24
> +       st.d    a1, a3, 32
> +       st.d    a1, a3, 40
> +       st.d    a1, a3, 48
> +       st.d    a1, a3, 56
> +       addi.d  a3, a3, 64
> +       bltu    a3, a4, .Lloop64
>  
>         /* set the remaining bytes */
> -2:     st.b    a1, a0, 0
> -       addi.d  a0, a0, 1
> -       addi.d  a2, a2, -1
> -       bgt     a2, zero, 2b
> +.Llt64:
> +       addi.d  a4, a2, -32
> +       bgeu    a3, a4, .Llt32
> +       st.d    a1, a3, 0
> +       st.d    a1, a3, 8
> +       st.d    a1, a3, 16
> +       st.d    a1, a3, 24
> +       addi.d  a3, a3, 32
> +
> +.Llt32:
> +       addi.d  a4, a2, -16
> +       bgeu    a3, a4, .Llt16
> +       st.d    a1, a3, 0
> +       st.d    a1, a3, 8
> +       addi.d  a3, a3, 16
> +
> +.Llt16:
> +       addi.d  a4, a2, -8
> +       bgeu    a3, a4, .Llt8
> +       st.d    a1, a3, 0
> +
> +.Llt8:
> +       st.d    a1, a2, -8
>  
>         /* return */
> -3:     move    a0, a3
> +       jr      ra
> +
> +       .align  4
> +.Lsmall:
> +       pcaddi  t0, 4
> +       slli.d  a2, a2, 4
> +       add.d   t0, t0, a2
> +       jr      t0
> +
> +       .align  4
> +0:     jr      ra
> +
> +       .align  4
> +1:     st.b    a1, a0, 0
> +       jr      ra
> +
> +       .align  4
> +2:     st.h    a1, a0, 0
> +       jr      ra
> +
> +       .align  4
> +3:     st.h    a1, a0, 0
> +       st.b    a1, a0, 2
> +       jr      ra
> +
> +       .align  4
> +4:     st.w    a1, a0, 0
> +       jr      ra
> +
> +       .align  4
> +5:     st.w    a1, a0, 0
> +       st.b    a1, a0, 4
> +       jr      ra
> +
> +       .align  4
> +6:     st.w    a1, a0, 0
> +       st.h    a1, a0, 4
> +       jr      ra
> +
> +       .align  4
> +7:     st.w    a1, a0, 0
> +       st.w    a1, a0, 3
> +       jr      ra
> +
> +       .align  4
> +8:     st.d    a1, a0, 0
>         jr      ra
>  SYM_FUNC_END(__memset_fast)
>  _ASM_NOKPROBE(__memset_fast)

--
Xi Ruoyao <[email protected]>
School of Aerospace Science and Technology, Xidian University

2023-04-10 14:24:39

by WANG Rui

[permalink] [raw]
Subject: Re: [PATCH] LoongArch: Improve memory ops

On Mon, Apr 10, 2023 at 8:20 PM Xi Ruoyao <[email protected]> wrote:
>
> On Mon, 2023-04-10 at 19:57 +0800, WANG rui wrote:
> > + /* align up address */
> > + andi t1, a0, 7
> > + sub.d a0, a0, t1
>
> bstrins.d a0, zero, 2, 0
>
> Likewise for other aligning operations if the temporary is not used.

I think we're on the same page. I had previously tested this on the
user-space version[1], but it's not a performance-critical area.

[1] https://github.com/heiher/mem-bench/blob/0083d4e5a82e57939517413da3bcad81e01adbea/memset-int.S#L35-L37

Regards,
Rui

2023-04-10 14:32:11

by WANG Rui

[permalink] [raw]
Subject: Re: [PATCH] LoongArch: Improve memory ops

On Mon, Apr 10, 2023 at 8:54 PM Xi Ruoyao <[email protected]> wrote:
> Regarding these functions: we have -ffreestanding which is preventing
> the compiler from optimizing for e.g. "memcpy(a, b, 8);" into a simple
> ld.d/st.d pair. A explicit compiler built-in usage like
>
> #define memcpy(a, b, c) __builtin_memcpy(a, b, c)
>
> would allow the compiler to do such kind of optimization. Will this
> improve the performance?

That's a good question. IIUC, the current compiler generates
inefficient code for constant-length memcpy, which may sacrifice
performance for compatibility, as not all hardware supports unaligned
memory access. We need a runtime CPU feature dispatch mechanism
similar to alternatives to improve the compiler. This is indeed a
problem that needs to be addressed.

Regards,
Rui

2023-04-10 14:44:08

by Xi Ruoyao

[permalink] [raw]
Subject: Re: [PATCH] LoongArch: Improve memory ops

On Mon, 2023-04-10 at 22:22 +0800, Rui Wang wrote:
> On Mon, Apr 10, 2023 at 8:54 PM Xi Ruoyao <[email protected]> wrote:
> > Regarding these functions: we have -ffreestanding which is
> > preventing
> > the compiler from optimizing for e.g. "memcpy(a, b, 8);" into a
> > simple
> > ld.d/st.d pair.  A explicit compiler built-in usage like
> >
> >   #define memcpy(a, b, c) __builtin_memcpy(a, b, c)
> >
> > would allow the compiler to do such kind of optimization.   Will
> > this
> > improve the performance?
>
> That's a good question. IIUC, the current compiler generates
> inefficient code for constant-length memcpy, which may sacrifice
> performance for compatibility, as not all hardware supports unaligned
> memory access. We need a runtime CPU feature dispatch mechanism
> similar to alternatives to improve the compiler. This is indeed a
> problem that needs to be addressed.

Hmm, yep, GCC does not take the advantage of -mstrict-align expanding
memcpy etc. I'll add it into my TODO list for GCC 14.


--
Xi Ruoyao <[email protected]>
School of Aerospace Science and Technology, Xidian University

2023-04-10 14:46:06

by Xi Ruoyao

[permalink] [raw]
Subject: Re: [PATCH] LoongArch: Improve memory ops

On Mon, 2023-04-10 at 22:19 +0800, Rui Wang wrote:
> On Mon, Apr 10, 2023 at 8:20 PM Xi Ruoyao <[email protected]> wrote:
> >
> > On Mon, 2023-04-10 at 19:57 +0800, WANG rui wrote:
> > > +       /* align up address */
> > > +       andi    t1, a0, 7
> > > +       sub.d   a0, a0, t1
> >
> > bstrins.d a0, zero, 2, 0
> >
> > Likewise for other aligning operations if the temporary is not used.
>
> I think we're on the same page. I had previously tested this on the
> user-space version[1], but it's not a performance-critical area.

The point is "if you can reduce one instruction and one register usage
with no harm, why not do it?" :)

AFAIK bstrins.d should be available on all 64-bit LoongArch CPUs.


--
Xi Ruoyao <[email protected]>
School of Aerospace Science and Technology, Xidian University

2023-04-10 15:56:00

by WANG Rui

[permalink] [raw]
Subject: Re: [PATCH] LoongArch: Improve memory ops

On Mon, Apr 10, 2023 at 10:38 PM Xi Ruoyao <[email protected]> wrote:
>
> The point is "if you can reduce one instruction and one register usage
> with no harm, why not do it?" :)
>
> AFAIK bstrins.d should be available on all 64-bit LoongArch CPUs.

Hmm, I will make the change in the next version. :)

Rui

2023-04-10 15:56:27

by WANG Rui

[permalink] [raw]
Subject: Re: [PATCH] LoongArch: Improve memory ops

On Mon, Apr 10, 2023 at 10:36 PM Xi Ruoyao <[email protected]> wrote:
>
> Hmm, yep, GCC does not take the advantage of -mstrict-align expanding
> memcpy etc. I'll add it into my TODO list for GCC 14.

Great! It looks like you already have some good ideas, and I look
forward to hearing your good news.

Rui

2023-04-11 06:24:49

by Xi Ruoyao

[permalink] [raw]
Subject: Re: [PATCH] LoongArch: Improve memory ops

On Mon, 2023-04-10 at 23:52 +0800, WANG Rui wrote:
> On Mon, Apr 10, 2023 at 10:36 PM Xi Ruoyao <[email protected]> wrote:
> >
> > Hmm, yep, GCC does not take the advantage of -mstrict-align expanding
> > memcpy etc.  I'll add it into my TODO list for GCC 14.

Technically, it takes the advantage but not very well. It's now
https://gcc.gnu.org/PR109465.

> Great! It looks like you already have some good ideas, and I look
> forward to hearing your good news.
>
> Rui
>

--
Xi Ruoyao <[email protected]>
School of Aerospace Science and Technology, Xidian University