Patchwork powerpc: Optimize __arch_swab32 and __arch_swab16

login
register
mail settings
Submitter Joakim Tjernlund
Date Sept. 9, 2011, 12:10 p.m.
Message ID <1315570258-12275-1-git-send-email-Joakim.Tjernlund@transmode.se>
Download mbox | patch
Permalink /patch/114091/
State New
Delegated to: Benjamin Herrenschmidt
Headers show

Comments

Joakim Tjernlund - Sept. 9, 2011, 12:10 p.m.
PPC __arch_swab32 and __arch_swab16 generates non optimal code.
They do not schedule very well, need to copy its input register
and swab16 needs an extra insn to clear its upper bits.
Fix this with better inline ASM.

Signed-off-by: Joakim Tjernlund <Joakim.Tjernlund@transmode.se>
---
 arch/powerpc/include/asm/swab.h |   28 ++++++++++++++--------------
 1 files changed, 14 insertions(+), 14 deletions(-)
Benjamin Herrenschmidt - Sept. 10, 2011, 9:24 a.m.
On Fri, 2011-09-09 at 14:10 +0200, Joakim Tjernlund wrote:
> PPC __arch_swab32 and __arch_swab16 generates non optimal code.
> They do not schedule very well, need to copy its input register
> and swab16 needs an extra insn to clear its upper bits.
> Fix this with better inline ASM.
> 
> Signed-off-by: Joakim Tjernlund <Joakim.Tjernlund@transmode.se>
> ---
>  arch/powerpc/include/asm/swab.h |   28 ++++++++++++++--------------
>  1 files changed, 14 insertions(+), 14 deletions(-)
> 
> diff --git a/arch/powerpc/include/asm/swab.h b/arch/powerpc/include/asm/swab.h
> index c581e3e..3b9a200 100644
> --- a/arch/powerpc/include/asm/swab.h
> +++ b/arch/powerpc/include/asm/swab.h
> @@ -61,25 +61,25 @@ static inline void __arch_swab32s(__u32 *addr)
>  
>  static inline __attribute_const__ __u16 __arch_swab16(__u16 value)
>  {
> -	__u16 result;
> -
> -	__asm__("rlwimi %0,%1,8,16,23"
> -	    : "=r" (result)
> -	    : "r" (value), "0" (value >> 8));
> -	return result;
> +	__asm__("rlwimi %0,%0,16,0x00ff0000\n\t"
> +		"rlwinm %0,%0,24,0x0000ffff"
> +		: "+r"(value));
> +	return value;
>  }
>  #define __arch_swab16 __arch_swab16

I don't quite get the thing about needing to clear the high bits.

Value is a u16 to start with, %0 is pre-filled with value >> 8 which
won't add anything to the upper bits, neither will rlwimi, so why would
you need to clear upper bits ?

Now I do see why gcc might generate something sub-optimal here, but can
you provide examples of asm output before/after in the patch commit ?

>  static inline __attribute_const__ __u32 __arch_swab32(__u32 value)
>  {
> -	__u32 result;
> -
> -	__asm__("rlwimi %0,%1,24,16,23\n\t"
> -	    "rlwimi %0,%1,8,8,15\n\t"
> -	    "rlwimi %0,%1,24,0,7"
> -	    : "=r" (result)
> -	    : "r" (value), "0" (value >> 24));
> -	return result;
> +	__u32 tmp;
> +
> +	__asm__("rlwimi %0,%1,24,0xffffffff"
> +		: "=r" (value) : "r" (value));
> +	tmp = value;
> +	__asm__("rlwimi %0,%1,16,0x00ff0000"
> +		: "+r" (value) : "r" (tmp));
> +	__asm__("rlwimi %0,%1,16,0x000000ff"
> +		: "+r" (value) : "r" (tmp));
> +	return value;
>  }
>  #define __arch_swab32 __arch_swab32

Cheers,
Ben.
Joakim Tjernlund - Sept. 11, 2011, 9:32 a.m.
Benjamin Herrenschmidt <benh@kernel.crashing.org> wrote on 2011/09/10 11:24:39:
>
> On Fri, 2011-09-09 at 14:10 +0200, Joakim Tjernlund wrote:
> > PPC __arch_swab32 and __arch_swab16 generates non optimal code.
> > They do not schedule very well, need to copy its input register
> > and swab16 needs an extra insn to clear its upper bits.
> > Fix this with better inline ASM.
> >
> > Signed-off-by: Joakim Tjernlund <Joakim.Tjernlund@transmode.se>
> > ---
> >  arch/powerpc/include/asm/swab.h |   28 ++++++++++++++--------------
> >  1 files changed, 14 insertions(+), 14 deletions(-)
> >
> > diff --git a/arch/powerpc/include/asm/swab.h b/arch/powerpc/include/asm/swab.h
> > index c581e3e..3b9a200 100644
> > --- a/arch/powerpc/include/asm/swab.h
> > +++ b/arch/powerpc/include/asm/swab.h
> > @@ -61,25 +61,25 @@ static inline void __arch_swab32s(__u32 *addr)
> >
> >  static inline __attribute_const__ __u16 __arch_swab16(__u16 value)
> >  {
> > -   __u16 result;
> > -
> > -   __asm__("rlwimi %0,%1,8,16,23"
> > -       : "=r" (result)
> > -       : "r" (value), "0" (value >> 8));
> > -   return result;
> > +   __asm__("rlwimi %0,%0,16,0x00ff0000\n\t"
> > +      "rlwinm %0,%0,24,0x0000ffff"
> > +      : "+r"(value));
> > +   return value;
> >  }
> >  #define __arch_swab16 __arch_swab16
>
> I don't quite get the thing about needing to clear the high bits.
>
> Value is a u16 to start with, %0 is pre-filled with value >> 8 which
> won't add anything to the upper bits, neither will rlwimi, so why would
> you need to clear upper bits ?
>
> Now I do see why gcc might generate something sub-optimal here, but can
> you provide examples of asm output before/after in the patch commit ?

I did a simple test earlier like so:

unsigned short __arch_swab16(unsigned short value)
{
		 unsigned short result;

		 __asm__("rlwimi %0,%1,8,16,23"
		     : "=r" (result)
		     : "r" (value), "0" (value >> 8));
		 return result;
}

unsigned short my__arch_swab16(unsigned short value)
{
		 __asm__("rlwimi %0,%0,16,0x00ff0000\n\t"
			 "rlwinm %0,%0,24,0x0000ffff"
			 : "+r"(value));
		 return value;
}

gcc -O2 -S outputs:

	.file	"my_swab.c"
	.section	".text"
	.align 2
	.globl __arch_swab16
	.type	__arch_swab16, @function
__arch_swab16:
	mr 0,3
	srwi 3,3,8
#APP
	rlwimi 3,0,8,16,23
#NO_APP
	rlwinm 3,3,0,0xffff
	blr
	.size	__arch_swab16, .-__arch_swab16
	.align 2
	.globl my__arch_swab16
	.type	my__arch_swab16, @function
my__arch_swab16:
#APP
	rlwimi 3,3,16,0x00ff0000
	rlwinm 3,3,24,0x0000ffff
#NO_APP
	blr
	.size	my__arch_swab16, .-my__arch_swab16
	.section	.note.GNU-stack,"",@progbits
	.ident	"GCC: (GNU) 3.4.6 (Gentoo 3.4.6-r2, ssp-3.4.6-1.0, pie-8.7.9)"

However I recently installed gcc 4.5.2 to do some tests and that gives
me this instead:
	.file	"my_swab.c"
	.gnu_attribute 4, 2
	.gnu_attribute 8, 1
	.gnu_attribute 12, 2
	.section	".text"
	.align 2
	.globl __arch_swab16
	.type	__arch_swab16, @function
__arch_swab16:
	srwi 0,3,8
#APP
 # 5 "my_swab.c" 1
	rlwimi 0,3,8,16,23
 # 0 "" 2
#NO_APP
	rlwinm 3,0,0,0xffff
	blr
	.size	__arch_swab16, .-__arch_swab16
	.align 2
	.globl my__arch_swab16
	.type	my__arch_swab16, @function
my__arch_swab16:
#APP
 # 13 "my_swab.c" 1
	rlwimi 3,3,16,0x00ff0000
	rlwinm 3,3,24,0x0000ffff
 # 0 "" 2
#NO_APP
	rlwinm 3,3,0,0xffff
	blr
	.size	my__arch_swab16, .-my__arch_swab16
	.ident	"GCC: (Gentoo 4.5.2 p1.1, pie-0.4.5) 4.5.2"
	.section	.note.GNU-stack,"",@progbits

So now both versions mask off the high bits, strange.
Not sure what is going on here?

 Jocke

Patch

diff --git a/arch/powerpc/include/asm/swab.h b/arch/powerpc/include/asm/swab.h
index c581e3e..3b9a200 100644
--- a/arch/powerpc/include/asm/swab.h
+++ b/arch/powerpc/include/asm/swab.h
@@ -61,25 +61,25 @@  static inline void __arch_swab32s(__u32 *addr)
 
 static inline __attribute_const__ __u16 __arch_swab16(__u16 value)
 {
-	__u16 result;
-
-	__asm__("rlwimi %0,%1,8,16,23"
-	    : "=r" (result)
-	    : "r" (value), "0" (value >> 8));
-	return result;
+	__asm__("rlwimi %0,%0,16,0x00ff0000\n\t"
+		"rlwinm %0,%0,24,0x0000ffff"
+		: "+r"(value));
+	return value;
 }
 #define __arch_swab16 __arch_swab16
 
 static inline __attribute_const__ __u32 __arch_swab32(__u32 value)
 {
-	__u32 result;
-
-	__asm__("rlwimi %0,%1,24,16,23\n\t"
-	    "rlwimi %0,%1,8,8,15\n\t"
-	    "rlwimi %0,%1,24,0,7"
-	    : "=r" (result)
-	    : "r" (value), "0" (value >> 24));
-	return result;
+	__u32 tmp;
+
+	__asm__("rlwimi %0,%1,24,0xffffffff"
+		: "=r" (value) : "r" (value));
+	tmp = value;
+	__asm__("rlwimi %0,%1,16,0x00ff0000"
+		: "+r" (value) : "r" (tmp));
+	__asm__("rlwimi %0,%1,16,0x000000ff"
+		: "+r" (value) : "r" (tmp));
+	return value;
 }
 #define __arch_swab32 __arch_swab32