Re: Poky & armv5te

[Gary Thomas <gary@mlbassoc.com>]

On 02/03/2011 07:32 AM, Bruce Ashfield wrote:
> On 11-02-02 07:21 PM, Richard Purdie wrote:
>> On Wed, 2011-02-02 at 15:17 -0700, Gary Thomas wrote:
>>> On 02/02/2011 11:27 AM, Gary Thomas wrote:
>>>> On 02/02/2011 11:00 AM, Richard Purdie wrote:
>>>>> On Wed, 2011-02-02 at 09:42 -0800, Darren Hart wrote:
>>>>>> On 02/02/2011 09:24 AM, Gary Thomas wrote:
>>>>>>> I'd like to use Poky with my OMAP-L138 which is armv5te
>>>>>>> Sadly today, this just isn't a go because of GCC issues.
>>>>>>> I've tried all the combinations which are in the trees
>>>>>>> (main& contrib) with no luck:
>>>>>>> * 4.3.3 - fails to build GCC
>>>>>>> * 4.5.1 - everything builds but kernel crashes
>>>>>>> * 4.5.2 - fails to build GCC
>>>>>>>
>>>>>>> Am I totally out of luck?
>>>>>>
>>>>>> What are you running into? I've recently backported some changes in
>>>>>> support of the Beagleboard (armv7-a) which were necessary to build with
>>>>>> 4.5.1 and our 2.21 binutils.
>>>>>
>>>>> Those wouldn't affect armv5. What puzzles me is that qemuarm is
>>>>> effectively an armv5 like system and those do build and boot in qemu.
>>>>>
>>>>> When you managed images above for 4.5.1 did you try using a known good
>>>>> kernel with our userspace? I'm wondering if we can narrow it to a kernel
>>>>> issue?
>>>>
>>>> Not sure I got that far, but I'll check it out.
>>>
>>> A simple test indicates that the user-space tools built in Poky
>>> do run fine on this hardware when built with GCC/4.5.1 and BINUTILS/2.21
>>> It's only the kernel that crashes hard with this combo.
>>
>> Ok, thats positive in many ways :)
>>
>> When you say the kernel crashes hard can you be any more specific about
>> where in the process its doing that? Any chance its using -Os anywhere
>> in the kernel build and that is causing the problem?
>
> For ARM .. it definitely is using Os. Unless gcc 4.5.1 has
> some changes in this area, ARM builds require Os, as they
> have for a while (see below).
>
> Typically if you build an ARM kernel without Os you get
> indeterminate errors (in the network stack, scheduler
> optimizations, etc). At least in kernels 2.6.34 and older
> this is true, I've shared in the marathon debug sessions
> to prove it :) In fact, we explicitly force on the
> CC_OPTIMIZE_FOR_SIZE in our configuration of th ARM kernels
> to make sure this doesn't creep back in.
>
> A quick check shows that out of ~170 ARM defconfigs 151
> turn on CONFIG_CC_OPTIMIZE_FOR_SIZE, and the ones with it
> off are typically older ones.
>
> I just doubled checked qemuarm, and it also turns on
> CC_OPTIMIZE_FOR_SIZE (which it was supposed to do). I started
> a build for it and will do a boot to see what happens
> shortly.
>
> Maybe this has changed recently (I haven't had to have a
> deeper look in 2.6.37 (yet)), so I am happy to be corrected :)

The kernel I'm using is 2.6.32 and it does have CONFIG_CC_OPTIMIZE_FOR_SIZE=y

I should be able to investigate what's crashing shortly .

-- 
------------------------------------------------------------
Gary Thomas                 |  Consulting for the
MLB Associates              |    Embedded world
------------------------------------------------------------