[PATCH -mm] i386 usercopy.c opcode reordering (pipelining)

[PATCH -mm] i386 usercopy.c opcode reordering (pipelining)

[Andreas Mohr]

Hi all,

this one wrote to register %0 and immediately right-shifted it,
so let's andl register %3 first for better parallelism (rrright?).

Signed-off-by: Andreas Mohr <andi@lisas.de>


diff -urN linux-2.6.17-rc6-mm2.orig/arch/i386/lib/usercopy.c linux-2.6.17-rc6-mm2.my/arch/i386/lib/usercopy.c
--- linux-2.6.17-rc6-mm2.orig/arch/i386/lib/usercopy.c	2006-06-13 19:28:09.000000000 +0200
+++ linux-2.6.17-rc6-mm2.my/arch/i386/lib/usercopy.c	2006-06-13 19:38:11.000000000 +0200
@@ -646,8 +646,8 @@
 		"	subl %0,%3\n"					\
 		"4:	rep; movsb\n"					\
 		"	movl %3,%0\n"					\
-		"	shrl $2,%0\n"					\
 		"	andl $3,%3\n"					\
+		"	shrl $2,%0\n"					\
 		"	.align 2,0x90\n"				\
 		"0:	rep; movsl\n"					\
 		"	movl %3,%0\n"					\
@@ -682,8 +682,8 @@
 		"	subl %0,%3\n"					\
 		"4:	rep; movsb\n"					\
 		"	movl %3,%0\n"					\
-		"	shrl $2,%0\n"					\
 		"	andl $3,%3\n"					\
+		"	shrl $2,%0\n"					\
 		"	.align 2,0x90\n"				\
 		"0:	rep; movsl\n"					\
 		"	movl %3,%0\n"					\

Re: [PATCH -mm] i386 usercopy.c opcode reordering (pipelining)

[Arjan van de Ven]

Andreas Mohr wrote:
> Hi all,
> 
> this one wrote to register %0 and immediately right-shifted it,
> so let's andl register %3 first for better parallelism (rrright?).

have you benchmarked this?
I would actually expect no difference since a reg-reg move is basically
a thing for the register rename engine which then... well it's supposed
to be near free like this.