On Thu, Sep 06, 2018 at 04:24:59PM +0200, Arnd Bergmann wrote:

On Wed, Sep 5, 2018 at 2:08 PM Guo Ren [off-list ref] wrote:

quoted

--- /dev/null
+++ b/arch/csky/abiv1/memset.c

@@ -0,0 +1,38 @@
+// SPDX-License-Identifier: GPL-2.0
+// Copyright (C) 2018 Hangzhou C-SKY Microsystems co.,ltd.
+#include <linux/types.h>
+
+void *memset(void *dest, int c, size_t l)
+{
+       char *d = dest;
+       int ch = c;
+       int tmp;
+
+       if ((long)d & 0x3)
+               while (l--) *d++ = ch;
+       else {
+               ch &= 0xff;
+               tmp = (ch | ch << 8 | ch << 16 | ch << 24);
+
+               while (l >= 16) {
+                       *(((long *)d)) = tmp;
+                       *(((long *)d)+1) = tmp;
+                       *(((long *)d)+2) = tmp;
+                       *(((long *)d)+3) = tmp;
+                       l -= 16;
+                       d += 16;
+               }
+
+               while (l > 3) {
+                       *(((long *)d)) = tmp;
+                       d = d + 4;
+                       l -= 4;
+               }
+
+               while (l) {
+                       *d++ = ch;
+                       l--;
+               }
+       }
+       return dest;
+}

I see that we have a trivial memset() implementation in lib/string.c, but yours
seems to be better optimized. Where did you get it from?

We write it for our ck610 to improve the performance, but I think a lot
of other arch done it in asm style.

Is this a version
that works particularly well on C-Sky, or is this a generic optimized memset
that others could use as well?

We only test it on C-SKY, but I think it will also work better on other
arch CPU than current lib/string.c memset implement.

I see that in lib/string.c:
void *memset(void *s, int c, size_t count)
{
	char *xs = s;

	while (count--)
		*xs++ = c;
	return s;
}
The most problem is "char *xs;" and it will cause "st.b" in asm.
"st.b" is very slow.

Our key improvement is:

quoted

+                       *(((long *)d)) = tmp;
+                       *(((long *)d)+1) = tmp;
+                       *(((long *)d)+2) = tmp;
+                       *(((long *)d)+3) = tmp;

It will cause SOC AXI burst transfer.

In the latter case, we could add it to
lib/string.c and let architectures select it in place of the triivial version.

Good idea.

 Guo Ren