2010-09-02 05:46:25

by Miao Xie

[permalink] [raw]
Subject: [PATCH 3/3] x86_64/lib: improve the performance of memmove

This patch improved the performance of memmove on x86_64 by using mem_copy_bwd()
instead of byte copy. But x86_64's memcpy is quite fast, so I didn't change it.

I have tested this patchset by doing 500 bytes memory copy for 50000 times
on my x86_64 box:
memmove
2.6.36-rc1 2s 610445us
2.6.36-rc1 + patch 0s 257358us

After appling this patchset, the performance of the file creation and deletion
on some filesystem become better. I have tested it with the following benchmark
tool on my x86_64 box.
http://marc.info/?l=linux-btrfs&m=128212635122920&q=p3

Test steps:
# ./creat_unlink 50000

The result(Total time):
Ext4:
2.6.36-rc1 2.6.36-rc1 + patchset
file creation 0.771240 0.698983 9.4%UP
file deletion 0.459065 0.425530 7.3%UP


Btrfs:
2.6.36-rc1 2.6.36-rc1 + patchset
file creation 0.966807 0.947592 1.9%UP
file deletion 1.355671 1.217787 10.2%UP

Signed-off-by: Miao Xie <[email protected]>
---
arch/x86/lib/memmove_64.c | 10 +++++-----
1 files changed, 5 insertions(+), 5 deletions(-)

diff --git a/arch/x86/lib/memmove_64.c b/arch/x86/lib/memmove_64.c
index 0a33909..3640232 100644
--- a/arch/x86/lib/memmove_64.c
+++ b/arch/x86/lib/memmove_64.c
@@ -4,17 +4,17 @@
#define _STRING_C
#include <linux/string.h>
#include <linux/module.h>
+#include <linux/memcopy.h>

#undef memmove
void *memmove(void *dest, const void *src, size_t count)
{
- if (dest < src) {
+ if (dest - src >= count) { /* *Unsigned* compare! */
return memcpy(dest, src, count);
} else {
- char *p = dest + count;
- const char *s = src + count;
- while (count--)
- *--p = *--s;
+ unsigned long dstp = (unsigned long)dest;
+ unsigned long srcp = (unsigned long)src;
+ mem_copy_bwd(dstp, srcp, count);
}
return dest;
}
--
1.7.0.1