Re: [PATCH 2/7] PowerPC: add unlikely() to BUG_ON()

4 messages, 3 authors, 2011-01-28 · open the first message on its own page

Re: [PATCH 2/7] PowerPC: add unlikely() to BUG_ON()

From: David Daney <hidden>
Date: 2011-01-27 17:57:41

Why not also CC the PPC maintainers as well?  I am not certain, but I 
think they may be reached at:

linuxppc-dev@lists.ozlabs.org


On 01/27/2011 04:12 AM, Coly Li wrote:
quoted hunk
Current BUG_ON() arch/powerpc/include/asm/bug.h does not use unlikely(),
in order to get better branch predict result, source code may have to call
BUG_ON() with unlikely() explicitly. This is not a suggested method
to use BUG_ON().

This patch adds unlikely() inside BUG_ON implementation on PPC
code, callers can use BUG_ON without explicit unlikely() now.

I don't have any PPC hardware to compile and test this fix, any feedback
of this patch is welcome.

Signed-off-by: Coly Li<redacted>
Cc: Jeremy Fitzhardinge<redacted>
Cc: David Daney<redacted>
Cc: Wang Cong<redacted>
Cc: Yong Zhang<redacted>
diff --git a/arch/powerpc/include/asm/bug.h b/arch/powerpc/include/asm/bug.h
index 065c590..10889a6 100644
--- a/arch/powerpc/include/asm/bug.h
+++ b/arch/powerpc/include/asm/bug.h
@@ -2,6 +2,7 @@
  #define _ASM_POWERPC_BUG_H
  #ifdef __KERNEL__

+#include<linux/compiler.h>
  #include<asm/asm-compat.h>

  /*
@@ -71,7 +72,7 @@
  	unreachable();						\
  } while (0)

-#define BUG_ON(x) do {						\
+#define __BUG_ON(x) do {					\
  	if (__builtin_constant_p(x)) {				\
  		if (x)						\
  			BUG();					\
@@ -85,6 +86,8 @@
  	}							\
  } while (0)

+#define BUG_ON(x) __BUG_ON(unlikely(x))
+
This is the same type of frobbing you were trying to do to MIPS.

I will let the powerpc maintainers weigh in on it, but my opinion is 
that, as with MIPS, BUG_ON() is expanded to a single machine 
instruction, and this unlikely() business will not change the generated 
code in any useful way.  It is thus gratuitous code churn and 
complexification.

David Daney
  #define __WARN_TAINT(taint) do {				\
  	__asm__ __volatile__(					\
  		"1:	twi 31,0,0\n"				\

Re: [PATCH 2/7] PowerPC: add unlikely() to BUG_ON()

From: Scott Wood <hidden>
Date: 2011-01-27 20:04:38

On Thu, 27 Jan 2011 09:57:39 -0800
David Daney [off-list ref] wrote:
On 01/27/2011 04:12 AM, Coly Li wrote:
quoted
diff --git a/arch/powerpc/include/asm/bug.h b/arch/powerpc/include/asm/bug.h
index 065c590..10889a6 100644
--- a/arch/powerpc/include/asm/bug.h
+++ b/arch/powerpc/include/asm/bug.h
@@ -2,6 +2,7 @@
  #define _ASM_POWERPC_BUG_H
  #ifdef __KERNEL__

+#include<linux/compiler.h>
  #include<asm/asm-compat.h>

  /*
@@ -71,7 +72,7 @@
  	unreachable();						\
  } while (0)

-#define BUG_ON(x) do {						\
+#define __BUG_ON(x) do {					\
  	if (__builtin_constant_p(x)) {				\
  		if (x)						\
  			BUG();					\
@@ -85,6 +86,8 @@
  	}							\
  } while (0)

+#define BUG_ON(x) __BUG_ON(unlikely(x))
+
This is the same type of frobbing you were trying to do to MIPS.

I will let the powerpc maintainers weigh in on it, but my opinion is 
that, as with MIPS, BUG_ON() is expanded to a single machine 
instruction, and this unlikely() business will not change the generated 
code in any useful way.  It is thus gratuitous code churn and 
complexification.
What about just doing this:

#define BUG() __builtin_trap()

#define BUG_ON(x) do {	\
	if (x) \
		BUG(); \
} while (0)

GCC should produce better code than forcing it into twnei.  A simple
BUG_ON(x != y) currently generates something like this (GCC 4.3):

xor     r0,r11,r0
addic   r10,r0,-1
subfe   r9,r10,r0
twnei   r9,0

Or this (GCC 4.5):

xor     r0,r11,r0
cntlzw	r0,r0
srwi	r0,r0,5
xori	r0,r0,1
twnei   r0,0

Instead of:

twne	r0,r11

And if GCC doesn't treat code paths that lead to __builtin_trap() as
unlikely (which could make a difference with complex expressions,
even with a conditional trap instruction), that's something that could
and should be fixed in GCC.

The downside is that GCC says, "The mechanism used may vary from
release to release so you should not rely on any particular
implementation."  However, some architectures (sparc, m68k, ia64)
already use __builtin_trap() for this, it seems unlikely that future GCC
versions would switch away from using the trap instruction[1], and there
doesn't seem to be a better-defined way to make GCC generate trap
instructions intelligently.

-Scott

[1] A more likely possibility is that an older compiler just generates a
call to abort() or similar, and later versions implemented trap -- need
to check what the oldest supported GCC does.

Re: [PATCH 2/7] PowerPC: add unlikely() to BUG_ON()

From: David Daney <hidden>
Date: 2011-01-27 20:32:08

On 01/27/2011 12:04 PM, Scott Wood wrote:
On Thu, 27 Jan 2011 09:57:39 -0800
David Daney[off-list ref]  wrote:
quoted
On 01/27/2011 04:12 AM, Coly Li wrote:
quoted
diff --git a/arch/powerpc/include/asm/bug.h b/arch/powerpc/include/asm/bug.h
index 065c590..10889a6 100644
--- a/arch/powerpc/include/asm/bug.h
+++ b/arch/powerpc/include/asm/bug.h
@@ -2,6 +2,7 @@
   #define _ASM_POWERPC_BUG_H
   #ifdef __KERNEL__

+#include<linux/compiler.h>
   #include<asm/asm-compat.h>

   /*
@@ -71,7 +72,7 @@
   	unreachable();						\
   } while (0)

-#define BUG_ON(x) do {						\
+#define __BUG_ON(x) do {					\
   	if (__builtin_constant_p(x)) {				\
   		if (x)						\
   			BUG();					\
@@ -85,6 +86,8 @@
   	}							\
   } while (0)

+#define BUG_ON(x) __BUG_ON(unlikely(x))
+
This is the same type of frobbing you were trying to do to MIPS.

I will let the powerpc maintainers weigh in on it, but my opinion is
that, as with MIPS, BUG_ON() is expanded to a single machine
instruction, and this unlikely() business will not change the generated
code in any useful way.  It is thus gratuitous code churn and
complexification.
What about just doing this:

#define BUG() __builtin_trap()

#define BUG_ON(x) do {	\
	if (x) \
		BUG(); \
} while (0)

GCC should produce better code than forcing it into twnei.  A simple
BUG_ON(x != y) currently generates something like this (GCC 4.3):

xor     r0,r11,r0
addic   r10,r0,-1
subfe   r9,r10,r0
twnei   r9,0

Or this (GCC 4.5):

xor     r0,r11,r0
cntlzw	r0,r0
srwi	r0,r0,5
xori	r0,r0,1
twnei   r0,0

Instead of:

twne	r0,r11

And if GCC doesn't treat code paths that lead to __builtin_trap() as
unlikely (which could make a difference with complex expressions,
even with a conditional trap instruction), that's something that could
and should be fixed in GCC.

The downside is that GCC says, "The mechanism used may vary from
release to release so you should not rely on any particular
implementation."  However, some architectures (sparc, m68k, ia64)
already use __builtin_trap() for this, it seems unlikely that future GCC
versions would switch away from using the trap instruction[1], and there
doesn't seem to be a better-defined way to make GCC generate trap
instructions intelligently.
This is good in theory, however powerpc has this _EMIT_BUG_ENTRY 
business that wouldn't work with __builtin_trap().

David Daney
-Scott

[1] A more likely possibility is that an older compiler just generates a
call to abort() or similar, and later versions implemented trap -- need
to check what the oldest supported GCC does.

RE: [PATCH 2/7] PowerPC: add unlikely() to BUG_ON()

From: David Laight <hidden>
Date: 2011-01-28 09:14:57

=20
quoted
+#define __BUG_ON(x) do {					\
  	if (__builtin_constant_p(x)) {				\
  		if (x)						\
  			BUG();					\
@@ -85,6 +86,8 @@
  	}							\
  } while (0)

+#define BUG_ON(x) __BUG_ON(unlikely(x))
+
From my experiments, adding an 'unlikely' at that point isn't
enough for non-trivial conditions - so its presence will
give a false sense the the optimisation is present!
In particular 'if (unlikely(x && y))' needs to be
'if (unlikely(x) && unlikely(y))' in order to avoid
mispredicted branches when 'x' is false.

Also, as (I think) in some of the generated code quoted,
use of __builtin_expect() with a boolean expression can
force some versions of gcc to generate the integer
value of the expression - rather than just selecting the
branch instructions that statically predict the
normal code path.

Sometimes I've also also had to add an asm() statement
that generates no code in order to actually force a
forwards branch (so it has something to place at the
target).

(I've been counting clocks ....)

	David
Keyboard shortcuts
hback out one level
jnext message in thread
kprevious message in thread
ldrill in
Escclose help / fold thread tree
?toggle this help