[PATCH] fix gcc4 warning in asm-ppc/time.h

A function must not return a const value, instead use
__attribute_const__.


Index: linux-2.6/include/asm/time.h
===================================================================

--- linux-2.6.orig/include/asm/time.h	2005-04-30 10:17:19.000000000 +0200
+++ linux-2.6/include/asm/time.h	2005-08-01 11:57:12.000000000 +0200

@@ -10,6 +10,7 @@
 #define __ASM_TIME_H__
 
 #include <linux/config.h>
+#include <linux/compiler.h>
 #include <linux/types.h>
 #include <linux/rtc.h>
 #include <linux/threads.h>

@@ -58,7 +59,7 @@
 /* Accessor functions for the timebase (RTC on 601) registers. */
 /* If one day CONFIG_POWER is added just define __USE_RTC as 1 */
 #ifdef CONFIG_6xx
-extern __inline__ int const __USE_RTC(void) {
+extern __inline__ int __attribute_const__ __USE_RTC(void) {
 	return (mfspr(SPRN_PVR)>>16) == 1;
 }
 #else

Christoph Hellwig [off-list ref] wrote:

Do we really want to do this?  The above implies that mfspr(SPRN_PVR) will
return the same value across the entire uptime of the kernel.  Is that
true?

Why is this function paying with const anyway?

Christoph Hellwig [off-list ref] wrote:

 A function must not return a const value, instead use
 __attribute_const__.


 Index: linux-2.6/include/asm/time.h
 ===================================================================
 --- linux-2.6.orig/include/asm/time.h	2005-04-30 10:17:19.000000000 +0200
 +++ linux-2.6/include/asm/time.h	2005-08-01 11:57:12.000000000 +0200

asm-ppc, please.  That didn't apply very well on an x86 tree...

On Wed, Aug 03, 2005 at 10:15:17PM -0700, Andrew Morton wrote:

Christoph Hellwig [off-list ref] wrote:

quoted

A function must not return a const value, instead use
__attribute_const__.


Index: linux-2.6/include/asm/time.h
===================================================================

--- linux-2.6.orig/include/asm/time.h	2005-04-30 10:17:19.000000000 +0200
+++ linux-2.6/include/asm/time.h	2005-08-01 11:57:12.000000000 +0200

@@ -10,6 +10,7 @@
 #define __ASM_TIME_H__
 
 #include <linux/config.h>
+#include <linux/compiler.h>
 #include <linux/types.h>
 #include <linux/rtc.h>
 #include <linux/threads.h>

@@ -58,7 +59,7 @@
 /* Accessor functions for the timebase (RTC on 601) registers. */
 /* If one day CONFIG_POWER is added just define __USE_RTC as 1 */
 #ifdef CONFIG_6xx
-extern __inline__ int const __USE_RTC(void) {
+extern __inline__ int __attribute_const__ __USE_RTC(void) {
 	return (mfspr(SPRN_PVR)>>16) == 1;
 }
 #else

Do we really want to do this?  The above implies that mfspr(SPRN_PVR) will
return the same value across the entire uptime of the kernel.  Is that
true?

Yes, it is a read-only register. PVR means processor version register
and is linked to the revision of the silicon. Besides that it is
only used in this context to check for processors of the 601 family
(the very first PPC processors). Nobody expects them to be hotpluggable,
nor even used in SMP systems for the matter.

Why is this function paying with const anyway?

I believe that it was originally to tell the compiler to be
able to cache it across several tests (a good compiler
on PPC would even cache the result of the test in a condition
register to allow branch prediction to work optimally).

Otherwise the compiler would never cache it since mfspr
is a volatile asm.

However this was written in pre gcc2.95 days so my 
memory may be failing.

	Gabriel

On Thu, 2005-08-04 at 07:54 +0200, Gabriel Paubert wrote:

On Wed, Aug 03, 2005 at 10:15:17PM -0700, Andrew Morton wrote:

quoted

Christoph Hellwig [off-list ref] wrote:

quoted

A function must not return a const value, instead use
__attribute_const__.


Index: linux-2.6/include/asm/time.h
===================================================================

--- linux-2.6.orig/include/asm/time.h	2005-04-30 10:17:19.000000000 +0200
+++ linux-2.6/include/asm/time.h	2005-08-01 11:57:12.000000000 +0200

@@ -10,6 +10,7 @@
 #define __ASM_TIME_H__
 
 #include <linux/config.h>
+#include <linux/compiler.h>
 #include <linux/types.h>
 #include <linux/rtc.h>
 #include <linux/threads.h>

@@ -58,7 +59,7 @@
 /* Accessor functions for the timebase (RTC on 601) registers. */
 /* If one day CONFIG_POWER is added just define __USE_RTC as 1 */
 #ifdef CONFIG_6xx
-extern __inline__ int const __USE_RTC(void) {
+extern __inline__ int __attribute_const__ __USE_RTC(void) {
 	return (mfspr(SPRN_PVR)>>16) == 1;
 }
 #else

Do we really want to do this?  The above implies that mfspr(SPRN_PVR) will
return the same value across the entire uptime of the kernel.  Is that
true?

Yes, it is a read-only register. PVR means processor version register
and is linked to the revision of the silicon. Besides that it is
only used in this context to check for processors of the 601 family
(the very first PPC processors). Nobody expects them to be hotpluggable,
nor even used in SMP systems for the matter.

Yes ... However, we want to get rid of PVR accesses. They are a problem
for Xen-like virtualization without HW support (that is running the
kernel actually in user space) among other things.

The above should be replaced by a test of CPU features.

Ben.

On Thu, Aug 04, 2005 at 11:46:38AM +0200, Benjamin Herrenschmidt wrote:

On Thu, 2005-08-04 at 07:54 +0200, Gabriel Paubert wrote:

quoted

On Wed, Aug 03, 2005 at 10:15:17PM -0700, Andrew Morton wrote:

quoted

Christoph Hellwig [off-list ref] wrote:

quoted

A function must not return a const value, instead use
__attribute_const__.


Index: linux-2.6/include/asm/time.h
===================================================================

--- linux-2.6.orig/include/asm/time.h	2005-04-30 10:17:19.000000000 +0200
+++ linux-2.6/include/asm/time.h	2005-08-01 11:57:12.000000000 +0200

@@ -10,6 +10,7 @@
 #define __ASM_TIME_H__
 
 #include <linux/config.h>
+#include <linux/compiler.h>
 #include <linux/types.h>
 #include <linux/rtc.h>
 #include <linux/threads.h>

@@ -58,7 +59,7 @@
 /* Accessor functions for the timebase (RTC on 601) registers. */
 /* If one day CONFIG_POWER is added just define __USE_RTC as 1 */
 #ifdef CONFIG_6xx
-extern __inline__ int const __USE_RTC(void) {
+extern __inline__ int __attribute_const__ __USE_RTC(void) {
 	return (mfspr(SPRN_PVR)>>16) == 1;
 }
 #else

Do we really want to do this?  The above implies that mfspr(SPRN_PVR) will
return the same value across the entire uptime of the kernel.  Is that
true?

Yes, it is a read-only register. PVR means processor version register
and is linked to the revision of the silicon. Besides that it is
only used in this context to check for processors of the 601 family
(the very first PPC processors). Nobody expects them to be hotpluggable,
nor even used in SMP systems for the matter.

Yes ... However, we want to get rid of PVR accesses. They are a problem
for Xen-like virtualization without HW support (that is running the
kernel actually in user space) among other things.

The above should be replaced by a test of CPU features.

I agree, this is old code that predates CPU features
(the first version was written in 1998), when the things
were simple: no PM, no Altivec, no really available 64 bit
machines.

Now I wonder whether current kernels really boot reliably on
601 (after all even something as stupid as the RI bit in the 
MSR is marked as non implemented on my 601 doc). 

	Gabriel

Andrew Morton writes:

Do we really want to do this?  The above implies that mfspr(SPRN_PVR) will
return the same value across the entire uptime of the kernel.  Is that
true?

Yes, the PVR is the processor version register, and it would be quite
disturbing if the silicon changed underneath us. :)

Why is this function paying with const anyway?

It seemed like a good idea at the time? :)

Paul.