[RFC] Use target filename in BUG_ON and friends

[RFC] Use target filename in BUG_ON and friends

[Sam Ravnborg]

When building the kernel using make O=.. all uses of __FILE__ becomes
filenames with absolute path resulting in increased text size.
Following patch supply the target filename as a commandline define
KBUILD_TARGET_FILE="mmslab.o"

With this doing a defconfig build I saw follow size reduction of
vmlinux (for x86_64):

-rc1:
   text	   data	    bss	    dec	    hex	filename
4727595	1174844	 605224	6507663	 634c8f	vmlinux

With patch applied:
   text	   data	    bss	    dec	    hex	filename
4733637	1174844	 605224	6513705	 636429	vmlinux.with-full-filenames

So a net saving of 6042 bytes. Corresponding to BUG or WARN being used
215 times with this configuration. (My path is 28 bytes up to the kernel
src dir).
With this config before applying the patch I have 344 strings in vmlinux
that starts with /home/sam/... so there is 129 more uses of __FILE__ 
that is not addressed by this patch.

The drawback of this patch is that the provided filename is no longer the
file that used the macro but the target filename. So it would be
"mm/slab.o" and not as before "mm/slab.c" that is identified. I played
no tricks to find the extension since identifying the .o file give a
hint that this may be in another file than just the corresponding .c
file (think included .c file, usage in other .h files etc).

If gcc could be teached not to use full path for __FILE__ this would be
an even better fix, but with current make O=.. support I have not found a
way to do so.

Patch below only modify x86_64, but if this is accepted all arch's bug.h
will be fixed.

	Sam

diffstat:
 include/asm-generic/bug.h |    6 ++++--
 include/asm-x86_64/bug.h  |    2 +-
 scripts/Makefile.lib      |    5 +++--
 3 files changed, 8 insertions(+), 5 deletions(-)


diff --git a/include/asm-generic/bug.h b/include/asm-generic/bug.h
index 8ceab7b..7811afa 100644
--- a/include/asm-generic/bug.h
+++ b/include/asm-generic/bug.h
@@ -6,7 +6,8 @@ #include <linux/compiler.h>
 #ifdef CONFIG_BUG
 #ifndef HAVE_ARCH_BUG
 #define BUG() do { \
-	printk("BUG: failure at %s:%d/%s()!\n", __FILE__, __LINE__, __FUNCTION__); \
+	printk("BUG: failure at %s:%d/%s()!\n", \
+	       KBUILD_TARGET_FILE, __LINE__, __FUNCTION__); \
 	panic("BUG!"); \
 } while (0)
 #endif
@@ -18,7 +19,8 @@ #endif
 #ifndef HAVE_ARCH_WARN_ON
 #define WARN_ON(condition) do { \
 	if (unlikely((condition)!=0)) { \
-		printk("BUG: warning at %s:%d/%s()\n", __FILE__, __LINE__, __FUNCTION__); \
+		printk("BUG: warning at %s:%d/%s()\n", \
+		       KBUILD_TARGET_FILE, __LINE__, __FUNCTION__); \
 		dump_stack(); \
 	} \
 } while (0)
diff --git a/include/asm-x86_64/bug.h b/include/asm-x86_64/bug.h
index 80ac1fe..355fac4 100644
--- a/include/asm-x86_64/bug.h
+++ b/include/asm-x86_64/bug.h
@@ -24,7 +24,7 @@ #define HAVE_ARCH_BUG
 #define BUG() 								\
 	asm volatile(							\
 	"ud2 ; pushq $%c1 ; ret $%c0" :: 				\
-		     "i"(__LINE__), "i" (__FILE__))
+		     "i"(__LINE__), "i" (KBUILD_TARGET_FILE))
 void out_of_line_bug(void);
 #else
 static inline void out_of_line_bug(void) { }
diff --git a/scripts/Makefile.lib b/scripts/Makefile.lib
index fc498fe..64753d0 100644
--- a/scripts/Makefile.lib
+++ b/scripts/Makefile.lib
@@ -85,6 +85,7 @@ name-fix = $(subst $(comma),_,$(subst -,
 basename_flags = -D"KBUILD_BASENAME=KBUILD_STR($(call name-fix,$(basetarget)))"
 modname_flags  = $(if $(filter 1,$(words $(modname))),\
                  -D"KBUILD_MODNAME=KBUILD_STR($(call name-fix,$(modname)))")
+filename_flags = -D"KBUILD_TARGET_FILE=KBUILD_STR($(src)/$(notdir $@))"
 
 _c_flags       = $(CFLAGS) $(EXTRA_CFLAGS) $(CFLAGS_$(basetarget).o)
 _a_flags       = $(AFLAGS) $(EXTRA_AFLAGS) $(AFLAGS_$(basetarget).o)
@@ -109,8 +110,8 @@ __cpp_flags     =                       
 endif
 
 c_flags        = -Wp,-MD,$(depfile) $(NOSTDINC_FLAGS) $(CPPFLAGS) \
-		 $(__c_flags) $(modkern_cflags) \
-		 -D"KBUILD_STR(s)=\#s" $(basename_flags) $(modname_flags)
+		 $(__c_flags) $(modkern_cflags) -D"KBUILD_STR(s)=\#s" \
+		 $(basename_flags) $(modname_flags) $(filename_flags)
 
 a_flags        = -Wp,-MD,$(depfile) $(NOSTDINC_FLAGS) $(CPPFLAGS) \
 		 $(__a_flags) $(modkern_aflags)

Re: [RFC] Use target filename in BUG_ON and friends

[Andi Kleen]

Sam Ravnborg <sam@ravnborg.org> writes:
> 
> If gcc could be teached not to use full path for __FILE__ this would be
> an even better fix, but with current make O=.. support I have not found a
> way to do so.

I suppose you could ask the gcc people? 

> 
> Patch below only modify x86_64, but if this is accepted all arch's bug.h
> will be fixed.

Looks good to me.

-Andi

Re: [RFC] Use target filename in BUG_ON and friends

[Milton Miller]

			
On Jul 8, 2006, at 04:45:54 EST, Sam Ravnborg wrote:
>  When building the kernel using make O=.. all uses of __FILE__ becomes
>  filenames with absolute path resulting in increased text size.
>  Following patch supply the target filename as a commandline define
>  KBUILD_TARGET_FILE="mmslab.o"

Unfortunately this ignores the fact that __LINE__ is meaningless
without __FILE__ because there are way too many BUGs in header
files.

Well, it does give us a hint as to which user of the header is
the problem one without going to System.map or objdump.

Even though it is hard on ccache, it is nice to be able to cut
and paste the file name.

milton

Re: [RFC] Use target filename in BUG_ON and friends

[Sam Ravnborg]

On Sat, Jul 08, 2006 at 04:00:55PM +0200, Andi Kleen wrote:
> Sam Ravnborg <sam@ravnborg.org> writes:
> > 
> > If gcc could be teached not to use full path for __FILE__ this would be
> > an even better fix, but with current make O=.. support I have not found a
> > way to do so.
> 
> I suppose you could ask the gcc people? 
This is more a kbuild issue. gcc supply the filename given on the
commandline and when using vpath support in make the input files have
absolute path resulting in gcc using absolute paths to __FILE__.
The only real fix I think would be to teach kbuild to use
top-og-source-tree as root when building the kernel. And that will break
a lot if this is changed.

	Sam

Re: [RFC] Use target filename in BUG_ON and friends

[Sam Ravnborg]

On Sat, Jul 08, 2006 at 11:45:49AM -0500, Milton Miller wrote:
> 			
> On Jul 8, 2006, at 04:45:54 EST, Sam Ravnborg wrote:
> > When building the kernel using make O=.. all uses of __FILE__ becomes
> > filenames with absolute path resulting in increased text size.
> > Following patch supply the target filename as a commandline define
> > KBUILD_TARGET_FILE="mmslab.o"
> 
> Unfortunately this ignores the fact that __LINE__ is meaningless
> without __FILE__ because there are way too many BUGs in header
> files.

__LINE__ gives a very precise hint of the offending .h file.
For x86_64 there are only one line-number clash in include/ for uses of
__FILE__.

"git grep -n __FILE__ | grep line-number" is your friend.

	Sam

Re: [RFC] Use target filename in BUG_ON and friends

[Andi Kleen]

On Saturday 08 July 2006 22:57, Sam Ravnborg wrote:
> On Sat, Jul 08, 2006 at 11:45:49AM -0500, Milton Miller wrote:
> > 			
> > On Jul 8, 2006, at 04:45:54 EST, Sam Ravnborg wrote:
> > > When building the kernel using make O=.. all uses of __FILE__ becomes
> > > filenames with absolute path resulting in increased text size.
> > > Following patch supply the target filename as a commandline define
> > > KBUILD_TARGET_FILE="mmslab.o"
> > 
> > Unfortunately this ignores the fact that __LINE__ is meaningless
> > without __FILE__ because there are way too many BUGs in header
> > files.
> 
> __LINE__ gives a very precise hint of the offending .h file.
> For x86_64 there are only one line-number clash in include/ for uses of
> __FILE__.


It's a nasty encoding. Maybe you could add a script to resolve them? 

-Andi

Re: [RFC] Use target filename in BUG_ON and friends

[Sam Ravnborg]

On Mon, Jul 10, 2006 at 02:03:07AM +0200, Andi Kleen wrote:
> On Saturday 08 July 2006 22:57, Sam Ravnborg wrote:
> > On Sat, Jul 08, 2006 at 11:45:49AM -0500, Milton Miller wrote:
> > > 			
> > > On Jul 8, 2006, at 04:45:54 EST, Sam Ravnborg wrote:
> > > > When building the kernel using make O=.. all uses of __FILE__ becomes
> > > > filenames with absolute path resulting in increased text size.
> > > > Following patch supply the target filename as a commandline define
> > > > KBUILD_TARGET_FILE="mmslab.o"
> > > 
> > > Unfortunately this ignores the fact that __LINE__ is meaningless
> > > without __FILE__ because there are way too many BUGs in header
> > > files.
> > 
> > __LINE__ gives a very precise hint of the offending .h file.
> > For x86_64 there are only one line-number clash in include/ for uses of
> > __FILE__.
> 
> 
> It's a nasty encoding. Maybe you could add a script to resolve them?
The patch was dropped. The benefit was too small compared to less
readable bug's.

	Sam