[U-Boot] [PATCH v8 0/38] Switch over to real Kbuild

[Masahiro Yamada]

Hello Simon and Tom,


This is my analysis of Kbuild performance.



[1] -j option is working

What I must say first is -j option is working correctly with Kbuild.

You can double-check by following the steps below.

Apply Kbuild series v8 on
commit 07e2822d158940a0e8ba45b6ab0344ffa1011a07.


First, build without -j option.

$ make mrproper
$ time make CROSS_COMPILE=arm-linux-gnueabi-  snow_config all
Configuring for snow board...
  GEN     include/autoconf.mk.dep
  [snip]

real	1m15.089s
user	0m44.092s
sys	0m32.513s


Next, build with -j8 option.
(Run "make mrproper" every time because we want to be sure
that there are no generated files before build.)

$ make mrproper
$ time make -j8 CROSS_COMPILE=arm-linux-gnueabi-  snow_config all 
  [snip]

real	0m17.223s
user	0m50.010s
sys	0m29.038s


It is much faster with -j8 option than without -j option.
(4.3 times faster on my box.)



You will easily notice another proof that -j option is working.


If you do not add -j option, the shorten log will be displayed
in the alphabetical order of output file name:

  LD      arch/arm/cpu/built-in.o
  CC      arch/arm/cpu/armv7/cache_v7.o
  CC      arch/arm/cpu/armv7/cpu.o
  CC      arch/arm/cpu/armv7/syslib.o
  CC      arch/arm/cpu/armv7/s5p-common/cpu_info.o
  CC      arch/arm/cpu/armv7/s5p-common/timer.o
  CC      arch/arm/cpu/armv7/s5p-common/sromc.o
  CC      arch/arm/cpu/armv7/s5p-common/pwm.o
  LD      arch/arm/cpu/armv7/s5p-common/built-in.o
  LD      arch/arm/cpu/armv7/built-in.o
  AS      arch/arm/cpu/armv7/start.o
  CC      arch/arm/cpu/armv7/exynos/clock.o
  CC      arch/arm/cpu/armv7/exynos/power.o
 

On the other hand, if you add -j option, the log will be shown
in a different order:

  LD      arch/arm/cpu/built-in.o
  CC      arch/arm/cpu/armv7/cache_v7.o
  CC      arch/arm/cpu/armv7/exynos/clock.o
  CC      disk/part.o
  AS      arch/arm/lib/crt0.o
  CC      board/samsung/common/board.o
  CC      board/samsung/smdk5250/smdk5250_spl.o
  CC      common/main.o
  LD      drivers/block/built-in.o
  LD      board/samsung/common/built-in.o
  CC      arch/arm/cpu/armv7/cpu.o
  AS      arch/arm/lib/relocate.o
  CC      disk/part_dos.o
  CC      arch/arm/cpu/armv7/exynos/power.o



[2] Is Kbuild slower than the old U-Boot build system?

Yes, Kbuild is definitely slower.
(But, as far as I tested,  I don't think it is 3 times slower.)

Let's compare the build time with the conventional build system.

Checkout master branch.
(commit 07e2822d158940a0e8ba45b6ab0344ffa1011a07)

$ make mrproper
$ time make CROSS_COMPILE=arm-linux-gnueabi-  snow

real	0m45.612s
user	0m28.367s
sys	0m18.969s


So, Kbuild(=real 1m15.089s) is 1.6 times slower
than the old U-Boot build system.


[3] Why is Kbuild slower?

One reason is "fixdep".

The helper program, fixdep, parses the source file
and all headers included from it to search all CONFIG
macro used there. This is a rather heavy task.

If you don't know the reason why "fixdep" is necessary,
please read the comment block of scripts/basic/fixdep.c
It is true that fixdep is meaningless for now,
but it will be a great help when switching to Kconfig.
We will get more return than we pay.
(And Kconfig series is almost ready.
I will test more and post version 1 within a couple of weeks.)


Let' check how big the impact of fixdep is.

I prepared a patch for you to disable fixdep:
http://patchwork.ozlabs.org/patch/316057/

Apply it on
commit 07e2822d158940a0e8ba45b6ab0344ffa1011a07 + Kbuild v8

And then, build.

$ make mrproper
$ time make CROSS_COMPILE=arm-linux-gnueabi-  snow_config all 

real	1m2.436s
user	0m38.281s
sys	0m25.871s

It is faster by 1.2 times faster without fixdep than it is with fixdep.


Another big factor is "arg-check".
This excellent routine is defined in scripts/Kbuild.include.

The dependency tracking of U-Boot old build system is
absolutely unreliable.
It compares the timestamp between object files and source files,
but never checks the arguments given to the compiler.

Kbuild checks both of them to precisely detect which objects must
be re-built.

To see how heavy "arg-check" task is,
build with KBUILD_NOCMDDEP=1 option to disable "arg-check".

$ make mrproper
$ time make CROSS_COMPILE=arm-linux-gnueabi- KBUILD_NOCMDDEP=1 snow_config all

real	0m41.882s
user	0m28.432s
sys	0m14.971s

See?
It is as fast as the old U-boot system(=0m45.612) .


Conclusion:
The main reasons of the slow down with Kbuild are "fixdep" and "arg-check".
Both of them are really important features for Kbuild and Kconfig.
    - "fixdep" is mandatory for our better life with Kconfig.
    - "arg-check" is for perfect dependency tracking.



> > What's your plan about this series?
> > Are we ready to switch to Kbuild, or need more review?
> 
> Lets get the performance problem Simon found figured out, but then
> otherwise, yes, I think we're about ready to merge.

Tom, are you satisfied with my analisys?

But, please hold merging Kbuild series.
I will post version 9 with a little minor update.



Best Regards
Masahiro Yamada