quoted

quoted

On 13.08.12 at 13:43, "Kirill A. Shutemov" [off-list ref] wrote:

On Thu, Aug 09, 2012 at 04:22:04PM +0100, Jan Beulich wrote:

quoted

quoted

quoted

quoted

On 09.08.12 at 17:03, "Kirill A. Shutemov" [off-list ref]  wrote:

...

quoted

quoted

---
 arch/x86/include/asm/page.h          |    2 ++
 arch/x86/include/asm/string_32.h     |    5 +++++
 arch/x86/include/asm/string_64.h     |    5 +++++
 arch/x86/lib/Makefile                |    1 +
 arch/x86/lib/clear_page_nocache_32.S |   30 ++++++++++++++++++++++++++++++
 arch/x86/lib/clear_page_nocache_64.S |   29 +++++++++++++++++++++++++++++

Couldn't this more reasonably go into clear_page_{32,64}.S?

We don't have clear_page_32.S.

Sure, but you're introducing a file anyway. Fold the new code into
the existing file for 64-bit, and create a new, similarly named one
for 32-bit.

quoted

quoted

+	xorl   %eax,%eax
+	movl   $4096/64,%ecx
+	.p2align 4
+.Lloop:
+	decl	%ecx
+#define PUT(x) movnti %eax,x*8(%edi) ; movnti %eax,x*8+4(%edi)

Is doing twice as much unrolling as on 64-bit really worth it?

Moving 64 bytes per cycle is faster on Sandy Bridge, but slower on
Westmere. Any preference? ;)

If it's not a clear win, I'd favor the 8-stores-per-cycle variant,
matching x86-64.

Jan