[U-Boot] [PATCH] armv7: fix linker file for newer ld support

[Albert ARIBAUD]

Le 02/11/2010 18:00, Måns Rullgård a écrit :
> Albert ARIBAUD<albert.aribaud at free.fr>  writes:
>
>> Le 02/11/2010 14:08, Steve Sakoman a écrit :
>>> On Tue, Nov 2, 2010 at 12:48 AM, Albert ARIBAUD<albert.aribaud at free.fr>   wrote:
>>>> Le 02/11/2010 05:05, Steve Sakoman a écrit :
>>>>
>>>>> I've been using gcc 4.3.3, so I haven't run into the issue that this
>>>>> patch is attempting to fix.
>>>>>
>>>>> I tested this patch using gcc 4.3.3, and while it produces a usable
>>>>> image, it causes the size of the image to grow from 227K to 433K!
>>>>>
>>>>> So perhaps we need a patch that uses a more restrictive wildcard.
>>>>>
>>>>> Steve
>>>>
>>>> Thanks for pointing this out, Steve. That'll go into V3 of my patch set I
>>>> guess, as Alexander's patch derives from mine and I most probably hit the
>>>> same size increase issue as he does.
>>>>
>>>> Wolfgang: that may mean the tx25 config file patch is unneeded. Stay tuned.
>>>
>>> I should have stated that I was using gcc 4.3.3 and binutils 2.18.50,
>>> since indeed it does seem to be binutils related.
>>>
>>> I have had a couple of reports that the following patch works with
>>> recent gcc/binutils without increasing the size of the binary.
>>>
>>> Could others check and report results?
>>>
>>> Steve
>>>
>>>
>>> ARMV7: Fix build issue with recent versions of gcc/binutils
>>> ---
>>>
>>> diff --git a/arch/arm/cpu/armv7/u-boot.lds b/arch/arm/cpu/armv7/u-boot.lds
>>> index 88a0fec..e690b58 100644
>>> --- a/arch/arm/cpu/armv7/u-boot.lds
>>> +++ b/arch/arm/cpu/armv7/u-boot.lds
>>> @@ -55,7 +55,7 @@ SECTIONS
>>>    	}
>>>    	. = ALIGN(4);
>>>    	__rel_dyn_start = .;
>>> -	.rel.dyn : { *(.rel.dyn) }
>>> +	.rel.dyn : { *(.rel.*) }
>>>    	__rel_dyn_end = .;
>>>
>>>    	__dynsym_start = .;
>>
>> Weird... This patch seems indeed more restrictive than mine, but the
>> input sections collected in .rel.dyn are actually the same for both.
>>
>> We can gain a bit less than 60k by overlapping .bss and .rel.dyn, but
>> certainly not 200k!
>>
>> OTOH, it happened during my 'cross-build' tests (using one chain's gcc
>> with another chain's ld) that the resulting u-boot was drastically
>> reduced; but I chalked that to my misusing of the build tools.
>>
>> Steve, can you indicate which toolchain excatly exhibits the
>> 230k-to-430k issue, and which board I should build?
>
> The problem is that the first patch posted here places .rel.dyn after
> ..bss, so when dumping u-boot.img, all of .bss is included in the file.
> Keeping it before .bss allows the flat binary to omit .bss as usual.

Indeed, but that accounts only for about 60k, as I said. Steve talks 
about a 200k delta; I can't find how this much can be added just by 
moving the .rel.dyn section around.

The idea of having .text, .*data and .bss contiguous is that it will 
reduce RAM footprint. In order to keep this and not increase FLASH/NAND 
footprint, I think we can overlap .bss and .rel.dyn.

> Using the patch quoted above, I get a working u-boot from all
> combinations of gcc and binutils I cared to try, including gcc
> 4.5-linaro with binutils 2.20.1.

Thanks for the feedback.

Amicalement,
-- 
Albert.