[2.6 patch] add an -Os config option

[2.6 patch] add an -Os config option

[Adrian Bunk]

The patch below adds an option OPTIMIZE_FOR_SIZE (depending on EMBEDDED) 
that changes the optimization from -O2 to -Os.

cu
Adrian

--- linux-2.6.0-test3/init/Kconfig.old	2003-08-11 01:33:31.000000000 +0200
+++ linux-2.6.0-test3/init/Kconfig	2003-08-11 01:46:48.000000000 +0200
@@ -118,6 +118,17 @@
 	  a "non-standard" kernel.  Only use this if you really know what you
 	  are doing.
 
+config OPTIMIZE_FOR_SIZE
+	bool "Optimize for size" if EMBEDDED
+	default n
+	help
+	  Enabling this option will pass "-Os" instead of "-O2" to gcc
+	  resulting in a smaller kernel.
+
+	  The resulting kernel might be significantly slower.
+
+	  If unsure, say N.
+
 config KALLSYMS
 	 bool "Load all symbols for debugging/kksymoops" if EMBEDDED
 	 default y
--- linux-2.6.0-test3/Makefile.old	2003-08-11 01:34:41.000000000 +0200
+++ linux-2.6.0-test3/Makefile	2003-08-11 01:37:16.000000000 +0200
@@ -223,7 +223,7 @@
 NOSTDINC_FLAGS  = -nostdinc -iwithprefix include
 
 CPPFLAGS	:= -D__KERNEL__ -Iinclude
-CFLAGS 		:= -Wall -Wstrict-prototypes -Wno-trigraphs -O2 \
+CFLAGS 		:= -Wall -Wstrict-prototypes -Wno-trigraphs \
 	  	   -fno-strict-aliasing -fno-common
 AFLAGS		:= -D__ASSEMBLY__
 
@@ -318,6 +318,12 @@
 
 -include .config.cmd
 
+ifdef CONFIG_OPTIMIZE_FOR_SIZE
+CFLAGS		+= -Os
+else
+CFLAGS		+= -O2
+endif
+
 ifndef CONFIG_FRAME_POINTER
 CFLAGS		+= -fomit-frame-pointer
 endif

Re: [2.6 patch] add an -Os config option

[Michael Buesch]

-----BEGIN PGP SIGNED MESSAGE-----
Hash: SHA1

On Monday 11 August 2003 23:11, Adrian Bunk wrote:
> +	  The resulting kernel might be significantly slower.

With my poor english knowledge I would say it should be "significant".
Correct?

- -- 
Regards Michael Buesch  [ http://www.8ung.at/tuxsoft ]
Animals on this machine: some GNUs and Penguin 2.6.0-test3

-----BEGIN PGP SIGNATURE-----
Version: GnuPG v1.2.1 (GNU/Linux)

iD8DBQE/OAhroxoigfggmSgRAuZhAJ4kRQBY0+ryHIgFDLaRCq/QBJWoiwCfX/HT
HPIqp2wBp2yXM+mlwGQck+s=
=ZjQ4
-----END PGP SIGNATURE-----

Re: [2.6 patch] add an -Os config option

[Mike Fedyk]

On Mon, Aug 11, 2003 at 11:19:39PM +0200, Michael Buesch wrote:
> -----BEGIN PGP SIGNED MESSAGE-----
> Hash: SHA1
> 
> On Monday 11 August 2003 23:11, Adrian Bunk wrote:
> > +	  The resulting kernel might be significantly slower.
> 
> With my poor english knowledge I would say it should be "significant".
> Correct?

No, the english in the patch is correct.

Re: [2.6 patch] add an -Os config option

[Christophe Saout]

Am Mo, 2003-08-11 um 23.19 schrieb Michael Buesch:

> On Monday 11 August 2003 23:11, Adrian Bunk wrote:
> > +	  The resulting kernel might be significantly slower.
> 
> With my poor english knowledge I would say it should be "significant".
> Correct?

I don't think so. The sentence above looks fine to me... (and I'm not an
english genius either) :)

--
Christophe Saout <christophe@saout.de>
Please avoid sending me Word or PowerPoint attachments.
See http://www.fsf.org/philosophy/no-word-attachments.html

Re: [2.6 patch] add an -Os config option

[Dagfinn Ilmari Mannsåker]

Michael Buesch <fsdeveloper@yahoo.de> writes:

> On Monday 11 August 2003 23:11, Adrian Bunk wrote:
>> +	  The resulting kernel might be significantly slower.
>
> With my poor english knowledge I would say it should be "significant".
> Correct?

Nope, it's an adverb (modifying the adjective "slower"), hence the -ly
suffix.

-- 
ilmari

Re: [2.6 patch] add an -Os config option

[Russell King]

On Mon, Aug 11, 2003 at 11:11:45PM +0200, Adrian Bunk wrote:
> The patch below adds an option OPTIMIZE_FOR_SIZE (depending on EMBEDDED) 
> that changes the optimization from -O2 to -Os.

What about those of us who already build the kernel with -Os (eg, ARM) ?

This option will be confusing in that situation.

-- 
Russell King (rmk@arm.linux.org.uk)                The developer of ARM Linux
             http://www.arm.linux.org.uk/personal/aboutme.html

Re: [2.6 patch] add an -Os config option

[Gene Heskett]

On Monday 11 August 2003 17:19, Michael Buesch wrote:
>-----BEGIN PGP SIGNED MESSAGE-----
>Hash: SHA1
>
>On Monday 11 August 2003 23:11, Adrian Bunk wrote:
>> +	  The resulting kernel might be significantly slower.
>
>With my poor english knowledge I would say it should be
> "significant". Correct?
>
No, the quoted "significantly" version is the correct english useage 
in this apparently present tense.

-- 
Cheers, Gene
AMD K6-III@500mhz 320M
Athlon1600XP@1400mhz  512M
99.27% setiathome rank, not too shabby for a WV hillbilly
Yahoo.com attornies please note, additions to this message
by Gene Heskett are:
Copyright 2003 by Maurice Eugene Heskett, all rights reserved.

Re: [2.6 patch] add an -Os config option

[Warren Turkal]

Michael Buesch wrote:
> On Monday 11 August 2003 23:11, Adrian Bunk wrote:
>> +      The resulting kernel might be significantly slower.
> 
> With my poor english knowledge I would say it should be "significant".
> Correct?

"Significantly" is correct in this case as it is an adverb modifying the
adjective "slower."

Warren Turkal
-- 
Treasurer, GOLUM, Inc.
http://www.golum.org

Re: [2.6 patch] add an -Os config option

[Thomas Molina]

On Mon, 11 Aug 2003, Gene Heskett wrote:

> On Monday 11 August 2003 17:19, Michael Buesch wrote:
> >-----BEGIN PGP SIGNED MESSAGE-----
> >Hash: SHA1
> >
> >On Monday 11 August 2003 23:11, Adrian Bunk wrote:
> >> +	  The resulting kernel might be significantly slower.
> >
> >With my poor english knowledge I would say it should be
> > "significant". Correct?
> >
> No, the quoted "significantly" version is the correct english useage 
> in this apparently present tense.
> 
> 
Now you have hit one of my hot buttons :)

Your use of useage is an incorrect use.  Simply use use in the above 
sentence :)

sineage, useage, etc. are simply incorrect

[OT] Re: [2.6 patch] add an -Os config option

[Steven Cole]

On Mon, 2003-08-11 at 20:51, Thomas Molina wrote:
> On Mon, 11 Aug 2003, Gene Heskett wrote:
> 
> > On Monday 11 August 2003 17:19, Michael Buesch wrote:
> > >-----BEGIN PGP SIGNED MESSAGE-----
> > >Hash: SHA1
> > >
> > >On Monday 11 August 2003 23:11, Adrian Bunk wrote:
> > >> +	  The resulting kernel might be significantly slower.
> > >
> > >With my poor english knowledge I would say it should be
> > > "significant". Correct?
> > >
> > No, the quoted "significantly" version is the correct english useage 
> > in this apparently present tense.
> > 
> > 
> Now you have hit one of my hot buttons :)
> 
> Your use of useage is an incorrect use.  Simply use use in the above 
> sentence :)

Pedantry leakage causes no shrinkage of English use wreckage.

Steven

Re: [2.6 patch] add an -Os config option

[Christoph Hellwig]

On Mon, Aug 11, 2003 at 11:11:45PM +0200, Adrian Bunk wrote:
> The patch below adds an option OPTIMIZE_FOR_SIZE (depending on EMBEDDED) 
> that changes the optimization from -O2 to -Os.

Please dropt the if EMBEDDED - this makes perfecty sense for lots of
todays hardware..

In fact we should probably rather do some some benchmarking whether it
would be a good idea to make -Os the default.

Re: [2.6 patch] add an -Os config option

[Michael Buesch]

-----BEGIN PGP SIGNED MESSAGE-----
Hash: SHA1

On Tuesday 12 August 2003 03:29, Gene Heskett wrote:
> No, the quoted "significantly" version is the correct english useage
> in this apparently present tense.


Thank you all for the correction.
I'll now take my english-dictinoary to learn a bit. :)


- -- 
Regards Michael Buesch  [ http://www.8ung.at/tuxsoft ]
Animals on this machine: some GNUs and Penguin 2.6.0-test3

-----BEGIN PGP SIGNATURE-----
Version: GnuPG v1.2.1 (GNU/Linux)

iD8DBQE/OL1eoxoigfggmSgRAjFMAJkB2JUMgOal+TPPNje5N4RbgxdViQCfZsEI
PWSDjc5fyF+M1NpCV82hnhI=
=OyMT
-----END PGP SIGNATURE-----

Re: [2.6 patch] add an -Os config option

[Alan Cox]

On Llu, 2003-08-11 at 22:11, Adrian Bunk wrote:
> +config OPTIMIZE_FOR_SIZE
> +	bool "Optimize for size" if EMBEDDED
> +	default n
> +	help
> +	  Enabling this option will pass "-Os" instead of "-O2" to gcc
> +	  resulting in a smaller kernel.
> +
> +	  The resulting kernel might be significantly slower.

With most of the gcc's I tried -Os was faster.

Re: [2.6 patch] add an -Os config option

[Gene Heskett]

On Tuesday 12 August 2003 06:11, Michael Buesch wrote:
>-----BEGIN PGP SIGNED MESSAGE-----
>Hash: SHA1
>
>On Tuesday 12 August 2003 03:29, Gene Heskett wrote:
>> No, the quoted "significantly" version is the correct english
>> useage in this apparently present tense.
>
>Thank you all for the correction.
>I'll now take my english-dictinoary to learn a bit. :)

And I pushed somebodies hot button when I said 'useage' instead of 
'usage'.  You win some, and you lose some. :)

-- 
Cheers, Gene
AMD K6-III@500mhz 320M
Athlon1600XP@1400mhz  512M
99.27% setiathome rank, not too shabby for a WV hillbilly
Yahoo.com attornies please note, additions to this message
by Gene Heskett are:
Copyright 2003 by Maurice Eugene Heskett, all rights reserved.

Re: [2.6 patch] add an -Os config option

[Gene Heskett]

On Monday 11 August 2003 22:51, Thomas Molina wrote:
>On Mon, 11 Aug 2003, Gene Heskett wrote:
>> On Monday 11 August 2003 17:19, Michael Buesch wrote:
>> >-----BEGIN PGP SIGNED MESSAGE-----
>> >Hash: SHA1
>> >
>> >On Monday 11 August 2003 23:11, Adrian Bunk wrote:
>> >> +	  The resulting kernel might be significantly slower.
>> >
>> >With my poor english knowledge I would say it should be
>> > "significant". Correct?
>>
>> No, the quoted "significantly" version is the correct english
>> useage in this apparently present tense.
>
>Now you have hit one of my hot buttons :)
>
>Your use of useage is an incorrect use.  Simply use use in the above
>sentence :)
>
>sineage, useage, etc. are simply incorrect

Well, in the everyday use, they do seem to be miss-used quite 
frequently as it is observed by me.  And I am not an english 
professor, just a senior (68) citizen with an 8th grade education, a 
GED, and a diploma from the University of Hard Knocks.  There really 
is such a thing you see.

-- 
Cheers, Gene
AMD K6-III@500mhz 320M
Athlon1600XP@1400mhz  512M
99.27% setiathome rank, not too shabby for a WV hillbilly
Yahoo.com attornies please note, additions to this message
by Gene Heskett are:
Copyright 2003 by Maurice Eugene Heskett, all rights reserved.

Re: [OT] Re: [2.6 patch] add an -Os config option

[Timothy Miller]

Steven Cole wrote:

> 
> Pedantry leakage causes no shrinkage of English use wreckage.
>

Argh, to my head, this has caused much damage.  :)

Re: [2.6 patch] add an -Os config option

[Timothy Miller]

Christoph Hellwig wrote:
> On Mon, Aug 11, 2003 at 11:11:45PM +0200, Adrian Bunk wrote:
> 
>>The patch below adds an option OPTIMIZE_FOR_SIZE (depending on EMBEDDED) 
>>that changes the optimization from -O2 to -Os.
> 
> 
> Please dropt the if EMBEDDED - this makes perfecty sense for lots of
> todays hardware..
> 
> In fact we should probably rather do some some benchmarking whether it
> would be a good idea to make -Os the default.


Interesting thought...  In reality, we want the system to spend as 
little time in the kernel as possible.  If we've done that job right, 
then optimizing for size shouldn't matter as much.  We're still spending 
most of our time in user space.

Furthermore, it may be that we could benefit from compiling some source 
files with -Os and others with -O2, depending on how critical they are 
and how much they are ACTUALLY affected.

So... we need profiling to determine what is used and how much.  And we 
need benchmarking to determine what is affected and by how much.  Over 
the next couple of years, it might be good for Linus and others to adopt 
a policy where the default optimization for individual files is 
determined based on experience and need.

If only 5% of the source files are compiled with -O2, that's not going 
to bloat the kernel much, we'll get an optimal balance of performance 
and size, and everyone will be happy.

We may find it valuable to divide up the sources differently.  There may 
be source files with both critical and noncritical functions that needed 
to be separated out so they can be compiled differently.

And while we're at it, if there is ever any benefit to -O3, that could 
be used on a per-file basis as well.

Comments?

Re: [2.6 patch] add an -Os config option

[Joe Korty]

On Tue, Aug 12, 2003 at 02:35:41PM +0100, Alan Cox wrote:
> On Llu, 2003-08-11 at 22:11, Adrian Bunk wrote:
>> +	  Enabling this option will pass "-Os" instead of "-O2" to gcc
>> +	  resulting in a smaller kernel.
>> +
>> +	  The resulting kernel might be significantly slower.
> 
> With most of the gcc's I tried -Os was faster.

Because of a better L1-instruction cache footprint?

Joe

Re: [2.6 patch] add an -Os config option

[Adrian Bunk]

On Tue, Aug 12, 2003 at 11:43:50AM -0400, Timothy Miller wrote:
> 
> 
> Christoph Hellwig wrote:
> >On Mon, Aug 11, 2003 at 11:11:45PM +0200, Adrian Bunk wrote:
> >
> >>The patch below adds an option OPTIMIZE_FOR_SIZE (depending on EMBEDDED) 
> >>that changes the optimization from -O2 to -Os.
> >
> >
> >Please dropt the if EMBEDDED - this makes perfecty sense for lots of
> >todays hardware..
> >
> >In fact we should probably rather do some some benchmarking whether it
> >would be a good idea to make -Os the default.
> 
> 
> Interesting thought...  In reality, we want the system to spend as 
> little time in the kernel as possible.  If we've done that job right, 
> then optimizing for size shouldn't matter as much.  We're still spending 
> most of our time in user space.
> 
> Furthermore, it may be that we could benefit from compiling some source 
> files with -Os and others with -O2, depending on how critical they are 
> and how much they are ACTUALLY affected.
>...
> Comments?

First of all, compiling my own kernel with gcc 3.3.1 resulted in less 
than 10% difference in the size of the kernel image (and ACPI is not 
part of my kernel) - the mere difference in size isn't that important 
except for special needs (like boot _floppies_ or embedded systems 
with very limited memory).

The benchmarking Christoph was referring to was to check whether a 
kernel completely compiled with -Os is faster than a kernel compiled 
with -O2. This might sound strange, but if the code is smaller more fits 
into the cache and modern CPUs suffer much every time they need to 
access the incredibly slow RAM.

cu
Adrian

-- 

       "Is there not promise of rain?" Ling Tan asked suddenly out
        of the darkness. There had been need of rain for many days.
       "Only a promise," Lao Er said.
                                       Pearl S. Buck - Dragon Seed

Re: [2.6 patch] add an -Os config option

[Kevin P. Fleming]

Timothy Miller wrote:

> Furthermore, it may be that we could benefit from compiling some source 
> files with -Os and others with -O2, depending on how critical they are 
> and how much they are ACTUALLY affected.
> 

At one time last year, there was a brief discussion about modifying 
the kernel's build process to make it build whole directories of 
sources as a single "compilation unit". As I remember, this allowed 
the compiler to make better decisions about how to organize the 
resulting code, what to keep/discard, optimization, etc.

With the simplifications that have already occurred in the 2.6 
makefiles, it seems to me that this would be an easier thing to do now 
(as opposed to the old makefiles where the build process had a harder 
time figuring out what source files were going to end up in the same 
object file). Now that kbuild always builds "built-on.o" in each 
source directory, could it perform that step by actually feeding gcc 
all the relevant source files in some combined form?

Granted, this is not something you'd want to use during debugging of 
kernel code, it would only be relevant for those trying to minimize 
their kernel size and/or get the last possible bits of speed out of 
the running kernel.

Re: [2.6 patch] add an -Os config option

[Timothy Miller]

Alan Cox wrote:
> On Llu, 2003-08-11 at 22:11, Adrian Bunk wrote:
> 
>>+config OPTIMIZE_FOR_SIZE
>>+	bool "Optimize for size" if EMBEDDED
>>+	default n
>>+	help
>>+	  Enabling this option will pass "-Os" instead of "-O2" to gcc
>>+	  resulting in a smaller kernel.
>>+
>>+	  The resulting kernel might be significantly slower.
> 
> 
> With most of the gcc's I tried -Os was faster.


Why is -Os faster?  Fewer cache misses?

Wouldn't that make -O2 kinda pointless?  It seems kinda futile to 
optimize for speed just to have it come out slower.

Re: [2.6 patch] add an -Os config option

[John Bradford]

> >>+config OPTIMIZE_FOR_SIZE
> >>+	bool "Optimize for size" if EMBEDDED
> >>+	default n
> >>+	help
> >>+	  Enabling this option will pass "-Os" instead of "-O2" to gcc
> >>+	  resulting in a smaller kernel.
> >>+
> >>+	  The resulting kernel might be significantly slower.
> > 
> > 
> > With most of the gcc's I tried -Os was faster.
>
>
> Why is -Os faster?  Fewer cache misses?
>
> Wouldn't that make -O2 kinda pointless?  It seems kinda futile to 
> optimize for speed just to have it come out slower.

See the comments Linus made earlier this year on the same subject:

http://marc.theaimsgroup.com/?l=linux-kernel&m=104457390406050&w=2

Alan, could Valgrind help us to profile cache hits/misses in different
parts of the kernel?

John.