LoongArch: Optimize memory ops (memset/memcpy/memmove)

To optimize memset()/memcpy()/memmove() and so on, we use a jump table
to dispatch cases for short data lengths; and for long data lengths, we
split the destination into head part (first 8 bytes), tail part (last 8
bytes) and middle part. The head part and tail part may be at unaligned
addresses, while the middle part is always aligned (the middle part is
allowed to overlap the head/tail part). In this way, the first and last
8 bytes may be unaligned accesses, but we can make sure 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 <wangrui@loongson.cn>
Signed-off-by: Huacai Chen <chenhuacai@loongson.cn>

1 files changed, 68 insertions, 52 deletions

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)