[PATCH] CHECK_IRQ_PER_CPU() to avoid dead code in __do_IRQ()

[PATCH] CHECK_IRQ_PER_CPU() to avoid dead code in __do_IRQ()

[Karsten Wiese]

Hi Ingo,

This version of the patch has better coding style
thanks to comments from Ingo Oeser.
Please comment or pass it on as appropriate.
Patch is for mainline 2.6.13-rc6.

   Thanks,
   Karsten 

------

From: Karsten Wiese <annabellesgarden@yahoo.de>

IRQ_PER_CPU is not used by all architectures.
This patch introduces the macros
ARCH_HAS_IRQ_PER_CPU and CHECK_IRQ_PER_CPU() to avoid the generation of
dead code in __do_IRQ().

ARCH_HAS_IRQ_PER_CPU is defined by architectures using
IRQ_PER_CPU in their
        include/asm_ARCH/irq.h
file.

Through grepping the tree I found the following
architectures currently use IRQ_PER_CPU:

        cris, ia64, ppc, ppc64 and parisc. 

Signed-off-by: Karsten Wiese <annabellesgarden@yahoo.de>



diff -upr linux-2.6.13-rc6/include/asm-cris/irq.h linux-2.6.13/include/asm-cris/irq.h
--- linux-2.6.13-rc6/include/asm-cris/irq.h	2005-08-08 11:46:10.000000000 +0200
+++ linux-2.6.13/include/asm-cris/irq.h	2005-08-08 11:41:12.000000000 +0200
@@ -1,6 +1,11 @@
 #ifndef _ASM_IRQ_H
 #define _ASM_IRQ_H
 
+/*
+ * IRQ line status macro IRQ_PER_CPU is used
+ */
+#define ARCH_HAS_IRQ_PER_CPU
+
 #include <asm/arch/irq.h>
 
 extern __inline__ int irq_canonicalize(int irq)
diff -upr linux-2.6.13-rc6/include/asm-ia64/irq.h linux-2.6.13/include/asm-ia64/irq.h
--- linux-2.6.13-rc6/include/asm-ia64/irq.h	2005-03-02 08:38:33.000000000 +0100
+++ linux-2.6.13/include/asm-ia64/irq.h	2005-08-06 18:06:53.000000000 +0200
@@ -14,6 +14,11 @@
 #define NR_IRQS		256
 #define NR_IRQ_VECTORS	NR_IRQS
 
+/*
+ * IRQ line status macro IRQ_PER_CPU is used
+ */
+#define ARCH_HAS_IRQ_PER_CPU
+
 static __inline__ int
 irq_canonicalize (int irq)
 {
diff -upr linux-2.6.13-rc6/include/asm-parisc/irq.h linux-2.6.13/include/asm-parisc/irq.h
--- linux-2.6.13-rc6/include/asm-parisc/irq.h	2005-08-08 11:45:26.000000000 +0200
+++ linux-2.6.13/include/asm-parisc/irq.h	2005-08-06 18:05:22.000000000 +0200
@@ -26,6 +26,11 @@
 
 #define NR_IRQS		(CPU_IRQ_MAX + 1)
 
+/*
+ * IRQ line status macro IRQ_PER_CPU is used
+ */
+#define ARCH_HAS_IRQ_PER_CPU
+
 static __inline__ int irq_canonicalize(int irq)
 {
 	return (irq == 2) ? 9 : irq;
diff -upr linux-2.6.13-rc6/include/asm-ppc/irq.h linux-2.6.13/include/asm-ppc/irq.h
--- linux-2.6.13-rc6/include/asm-ppc/irq.h	2005-08-08 11:46:10.000000000 +0200
+++ linux-2.6.13/include/asm-ppc/irq.h	2005-08-08 11:41:14.000000000 +0200
@@ -19,6 +19,11 @@
 #define IRQ_POLARITY_POSITIVE	0x2	/* high level or low->high edge */
 #define IRQ_POLARITY_NEGATIVE	0x0	/* low level or high->low edge */
 
+/*
+ * IRQ line status macro IRQ_PER_CPU is used
+ */
+#define ARCH_HAS_IRQ_PER_CPU
+
 #if defined(CONFIG_40x)
 #include <asm/ibm4xx.h>
 
diff -upr linux-2.6.13-rc6/include/asm-ppc64/irq.h linux-2.6.13/include/asm-ppc64/irq.h
--- linux-2.6.13-rc6/include/asm-ppc64/irq.h	2005-03-02 08:38:33.000000000 +0100
+++ linux-2.6.13/include/asm-ppc64/irq.h	2005-08-06 18:06:58.000000000 +0200
@@ -33,6 +33,11 @@
 #define IRQ_POLARITY_POSITIVE	0x2	/* high level or low->high edge */
 #define IRQ_POLARITY_NEGATIVE	0x0	/* low level or high->low edge */
 
+/*
+ * IRQ line status macro IRQ_PER_CPU is used
+ */
+#define ARCH_HAS_IRQ_PER_CPU
+
 #define get_irq_desc(irq) (&irq_desc[(irq)])
 
 /* Define a way to iterate across irqs. */
diff -upr linux-2.6.13-rc6/include/linux/irq.h linux-2.6.13/include/linux/irq.h
--- linux-2.6.13-rc6/include/linux/irq.h	2005-08-08 11:46:10.000000000 +0200
+++ linux-2.6.13/include/linux/irq.h	2005-08-08 11:55:11.000000000 +0200
@@ -32,7 +32,12 @@
 #define IRQ_WAITING	32	/* IRQ not yet seen - for autodetection */
 #define IRQ_LEVEL	64	/* IRQ level triggered */
 #define IRQ_MASKED	128	/* IRQ masked - shouldn't be seen again */
-#define IRQ_PER_CPU	256	/* IRQ is per CPU */
+#if defined(ARCH_HAS_IRQ_PER_CPU)
+# define IRQ_PER_CPU	256	/* IRQ is per CPU */
+# define CHECK_IRQ_PER_CPU(var) ((var) & IRQ_PER_CPU)
+#else
+# define CHECK_IRQ_PER_CPU(var) 0
+#endif
 
 /*
  * Interrupt controller descriptor. This is all we need
diff -upr linux-2.6.13-rc6/kernel/irq/handle.c linux-2.6.13/kernel/irq/handle.c
--- linux-2.6.13-rc6/kernel/irq/handle.c	2005-08-08 11:46:11.000000000 +0200
+++ linux-2.6.13/kernel/irq/handle.c	2005-08-08 11:53:00.000000000 +0200
@@ -111,7 +111,7 @@ fastcall unsigned int __do_IRQ(unsigned 
 	unsigned int status;
 
 	kstat_this_cpu.irqs[irq]++;
-	if (desc->status & IRQ_PER_CPU) {
+	if (CHECK_IRQ_PER_CPU(desc->status)) {
 		irqreturn_t action_ret;
 
 		/*

	

	
		
___________________________________________________________ 
Gesendet von Yahoo! Mail - Jetzt mit 1GB Speicher kostenlos - Hier anmelden: http://mail.yahoo.de

Re: [PATCH] CHECK_IRQ_PER_CPU() to avoid dead code in __do_IRQ()

[Alexander Nyberg]

> 
> IRQ_PER_CPU is not used by all architectures.
> This patch introduces the macros
> ARCH_HAS_IRQ_PER_CPU and CHECK_IRQ_PER_CPU() to avoid the generation of
> dead code in __do_IRQ().
> 
> ARCH_HAS_IRQ_PER_CPU is defined by architectures using
> IRQ_PER_CPU in their
>         include/asm_ARCH/irq.h
> file.
> 
> Through grepping the tree I found the following
> architectures currently use IRQ_PER_CPU:
> 
>         cris, ia64, ppc, ppc64 and parisc. 
> 

There are many places where one could replace run-time tests with 
#ifdef's but it makes reading more difficult (and in longer terms
maintainence). Have you benchmarked any workload that benefits 
from this?

> 
> diff -upr linux-2.6.13-rc6/include/asm-cris/irq.h linux-2.6.13/include/asm-cris/irq.h
> --- linux-2.6.13-rc6/include/asm-cris/irq.h	2005-08-08 11:46:10.000000000 +0200
> +++ linux-2.6.13/include/asm-cris/irq.h	2005-08-08 11:41:12.000000000 +0200
> @@ -1,6 +1,11 @@
>  #ifndef _ASM_IRQ_H
>  #define _ASM_IRQ_H
>  
> +/*
> + * IRQ line status macro IRQ_PER_CPU is used
> + */
> +#define ARCH_HAS_IRQ_PER_CPU
> +
>  #include <asm/arch/irq.h>
>  
>  extern __inline__ int irq_canonicalize(int irq)
> diff -upr linux-2.6.13-rc6/include/asm-ia64/irq.h linux-2.6.13/include/asm-ia64/irq.h
> --- linux-2.6.13-rc6/include/asm-ia64/irq.h	2005-03-02 08:38:33.000000000 +0100
> +++ linux-2.6.13/include/asm-ia64/irq.h	2005-08-06 18:06:53.000000000 +0200
> @@ -14,6 +14,11 @@
>  #define NR_IRQS		256
>  #define NR_IRQ_VECTORS	NR_IRQS
>  
> +/*
> + * IRQ line status macro IRQ_PER_CPU is used
> + */
> +#define ARCH_HAS_IRQ_PER_CPU
> +
>  static __inline__ int
>  irq_canonicalize (int irq)
>  {
> diff -upr linux-2.6.13-rc6/include/asm-parisc/irq.h linux-2.6.13/include/asm-parisc/irq.h
> --- linux-2.6.13-rc6/include/asm-parisc/irq.h	2005-08-08 11:45:26.000000000 +0200
> +++ linux-2.6.13/include/asm-parisc/irq.h	2005-08-06 18:05:22.000000000 +0200
> @@ -26,6 +26,11 @@
>  
>  #define NR_IRQS		(CPU_IRQ_MAX + 1)
>  
> +/*
> + * IRQ line status macro IRQ_PER_CPU is used
> + */
> +#define ARCH_HAS_IRQ_PER_CPU
> +
>  static __inline__ int irq_canonicalize(int irq)
>  {
>  	return (irq == 2) ? 9 : irq;
> diff -upr linux-2.6.13-rc6/include/asm-ppc/irq.h linux-2.6.13/include/asm-ppc/irq.h
> --- linux-2.6.13-rc6/include/asm-ppc/irq.h	2005-08-08 11:46:10.000000000 +0200
> +++ linux-2.6.13/include/asm-ppc/irq.h	2005-08-08 11:41:14.000000000 +0200
> @@ -19,6 +19,11 @@
>  #define IRQ_POLARITY_POSITIVE	0x2	/* high level or low->high edge */
>  #define IRQ_POLARITY_NEGATIVE	0x0	/* low level or high->low edge */
>  
> +/*
> + * IRQ line status macro IRQ_PER_CPU is used
> + */
> +#define ARCH_HAS_IRQ_PER_CPU
> +
>  #if defined(CONFIG_40x)
>  #include <asm/ibm4xx.h>
>  
> diff -upr linux-2.6.13-rc6/include/asm-ppc64/irq.h linux-2.6.13/include/asm-ppc64/irq.h
> --- linux-2.6.13-rc6/include/asm-ppc64/irq.h	2005-03-02 08:38:33.000000000 +0100
> +++ linux-2.6.13/include/asm-ppc64/irq.h	2005-08-06 18:06:58.000000000 +0200
> @@ -33,6 +33,11 @@
>  #define IRQ_POLARITY_POSITIVE	0x2	/* high level or low->high edge */
>  #define IRQ_POLARITY_NEGATIVE	0x0	/* low level or high->low edge */
>  
> +/*
> + * IRQ line status macro IRQ_PER_CPU is used
> + */
> +#define ARCH_HAS_IRQ_PER_CPU
> +
>  #define get_irq_desc(irq) (&irq_desc[(irq)])
>  
>  /* Define a way to iterate across irqs. */
> diff -upr linux-2.6.13-rc6/include/linux/irq.h linux-2.6.13/include/linux/irq.h
> --- linux-2.6.13-rc6/include/linux/irq.h	2005-08-08 11:46:10.000000000 +0200
> +++ linux-2.6.13/include/linux/irq.h	2005-08-08 11:55:11.000000000 +0200
> @@ -32,7 +32,12 @@
>  #define IRQ_WAITING	32	/* IRQ not yet seen - for autodetection */
>  #define IRQ_LEVEL	64	/* IRQ level triggered */
>  #define IRQ_MASKED	128	/* IRQ masked - shouldn't be seen again */
> -#define IRQ_PER_CPU	256	/* IRQ is per CPU */
> +#if defined(ARCH_HAS_IRQ_PER_CPU)
> +# define IRQ_PER_CPU	256	/* IRQ is per CPU */
> +# define CHECK_IRQ_PER_CPU(var) ((var) & IRQ_PER_CPU)
> +#else
> +# define CHECK_IRQ_PER_CPU(var) 0
> +#endif
>  
>  /*
>   * Interrupt controller descriptor. This is all we need
> diff -upr linux-2.6.13-rc6/kernel/irq/handle.c linux-2.6.13/kernel/irq/handle.c
> --- linux-2.6.13-rc6/kernel/irq/handle.c	2005-08-08 11:46:11.000000000 +0200
> +++ linux-2.6.13/kernel/irq/handle.c	2005-08-08 11:53:00.000000000 +0200
> @@ -111,7 +111,7 @@ fastcall unsigned int __do_IRQ(unsigned 
>  	unsigned int status;
>  
>  	kstat_this_cpu.irqs[irq]++;
> -	if (desc->status & IRQ_PER_CPU) {
> +	if (CHECK_IRQ_PER_CPU(desc->status)) {
>  		irqreturn_t action_ret;
>  
>  		/*

Re: [PATCH] CHECK_IRQ_PER_CPU() to avoid dead code in __do_IRQ()

[Karsten Wiese]

Am Montag, 8. August 2005 13:19 schrieb Alexander Nyberg:
> 
> There are many places where one could replace run-time tests with 
> #ifdef's but it makes reading more difficult (and in longer terms
> maintainence). Have you benchmarked any workload that benefits 
> from this?

Performance gain is small, but measurable: 0,02%
Tested on an Atlon64 running at 1000MHz.
I took this value from 9 runs (each with/without the patch) of 
	$ time lame x.wav
where each took about 1 minute.
3000 Interrupts/s were generated at the time by running
	$ jackd -R -dalsa -p64 -n2

0,02% really isn't that much....but Athlon64 is better than P4
with branch predictions, I think.

Erm... ok, I won't insist on having this patch applied ;-) 

   Karsten

	

	
		
___________________________________________________________ 
Gesendet von Yahoo! Mail - Jetzt mit 1GB Speicher kostenlos - Hier anmelden: http://mail.yahoo.de

Re: [PATCH] CHECK_IRQ_PER_CPU() to avoid dead code in __do_IRQ()

[Zachary Amsden]

Karsten Wiese wrote:

>Am Montag, 8. August 2005 13:19 schrieb Alexander Nyberg:
>  
>
>>There are many places where one could replace run-time tests with 
>>#ifdef's but it makes reading more difficult (and in longer terms
>>maintainence). Have you benchmarked any workload that benefits 
>>from this?
>>    
>>
>
>Performance gain is small, but measurable: 0,02%
>Tested on an Atlon64 running at 1000MHz.
>I took this value from 9 runs (each with/without the patch) of 
>	$ time lame x.wav
>where each took about 1 minute.
>3000 Interrupts/s were generated at the time by running
>	$ jackd -R -dalsa -p64 -n2
>
>0,02% really isn't that much....but Athlon64 is better than P4
>with branch predictions, I think.
>
>Erm... ok, I won't insist on having this patch applied ;-) 
>
>   Karsten
>  
>

Removing dead code is always good - 0.02% is small, but if 100 kernel 
developers all did the same, that adds up to 2% rather quickly, and that 
is nothing to sneeze at.  I like your patch, but you should add some 
comments for maintainability about what CHECK_IRQ_PER_CPU does - see 
include/asm-generic/pgtable.h for similar styling.  If also probably 
doesn't hurt to leave IRQ_PER_CPU defined even when 
ARCH_HAS_CHECK_IRQ_PER_CPU is not, since it looks cleaner and prevents 
future collisions with bits defined inside of an #ifdef.

Zach

Re: [PATCH] CHECK_IRQ_PER_CPU() to avoid dead code in __do_IRQ()

[Zwane Mwaikambo]

On Mon, 8 Aug 2005, Alexander Nyberg wrote:

> > IRQ_PER_CPU is not used by all architectures.
> > This patch introduces the macros
> > ARCH_HAS_IRQ_PER_CPU and CHECK_IRQ_PER_CPU() to avoid the generation of
> > dead code in __do_IRQ().
> > 
> > ARCH_HAS_IRQ_PER_CPU is defined by architectures using
> > IRQ_PER_CPU in their
> >         include/asm_ARCH/irq.h
> > file.
> > 
> > Through grepping the tree I found the following
> > architectures currently use IRQ_PER_CPU:
> > 
> >         cris, ia64, ppc, ppc64 and parisc. 
> > 
> 
> There are many places where one could replace run-time tests with 
> #ifdef's but it makes reading more difficult (and in longer terms
> maintainence). Have you benchmarked any workload that benefits 
> from this?

I doubt you'd be able to collect convincing benchmark data, but skipping a 
branch (possibly mispredicted) is worth it IMO.