ARM Linux GCC 4.1.0 broken optimization?

[Dirk Behme <dirk dot behme at googlemail dot com>]

Hi,

I use ARM Linux GCC 4.1.0 built with crosstool 0.42 with

eval `cat arm.dat gcc-4.1.0-glibc-2.3.2.dat`  sh all.sh --notest

to build Linux kernel 2.6.17-rc5 for embedded OMAP target. I 
use it since several weeks and everything looks quite good, 
kernel is booting, everything like expected.

However, only thing which doesn't work is playing mp3 using 
madplay. With help of OMAP mailing list I tried using 
madplay binary which works for others, using other kernel 
config, using other mp3s etc. With this, I found that 
switching back to my old GCC 3.4.1 and using it for kernel 
compilation, mp3 works. So looks like a toolchain issue 
compiling the sound modules.

Now, I took again gcc 4.1.0 and recompiled the sound modules 
using different optimization levels. Everything else stays 
the same (kernel config, kernel itself, mp3, madplay etc.). 
Result:

-Os -> mp3 playback fails
-O2 -> mp3 playback fails
-O0 -> building of some modules fails
-O1 -> mp3 playback works!

Looks to me that there is an issue with optimization in 
kernel's sound code. Really not sure, but may be anything 
(with inlining?) in include/sound/pcm_params.h?. Adding some 
printks it looked to me that I get  -EINVAL error in 
snd_mask_refine() in this file while compiling with -O1 
everything is okay.

Is this a known issue? Any idea regarding this? Do I missed 
a patch to apply to gcc while building?

Many thanks

Dirk



--
For unsubscribe information see http://sourceware.org/lists.html#faq