ld bug when self-referential variables are used as output section addresses ?

[Samuel Jones]

On 20/01/14 12:12, Alan Modra wrote:
> On Thu, Jan 16, 2014 at 02:55:37PM +0100, Samuel Jones wrote:
>> I have the following linker script:
>>
>>
>> TOTO = 1024;
>> TOTO += 1024;
>>
>> SECTIONS
>> {
>>    .rodata    TOTO :
>>    {
>>    __toto_symbol_abs = ABSOLUTE(TOTO);
>>      *(.rodata)
>>      *(.rodata.*)
>>    }
>>
>> }
>>
>>
>> When I link this example with the latest ld, the section ".rodata"
>> is placed at 1024, whereas I would have expected 2048.
>> __toto_symbol_abs gets the value 2048.
>>
>>
>> With ld 2.22.52 .rodata is placed at 2048. I believe this commit by
>> Alan Modra changed ld's behaviour:
>> 4194268f43623a5f893b9a92b0456d3cb43ab915. My understanding of the
>> patch is that it delays the evaluation of self-referential
>> expressions until section sizing has been completed. We depended on
>> the previous behaviour for correct placement of output sections
>> (mostly calculating alignment constraints).
>>
>>
>> Is what I observe expected behaviour? If so is there any way to
>> achieve what I want apart from using static single assignment? And
>> why the change?
>
> See https://sourceware.org/ml/binutils/2012-12/msg00155.html
>
> Clearly, your linker script is quite reasonable.  No matter how many
> times the script is evaluated, you should get the same result.  The
> tricky part is differentiating expressions involving self-assignment
> to absolute symbols with a previous linker script definition from all
> other symbols..
>
> Fortunately we already have some machinery in ld to track such
> symbols, but it needs some modification.
 >
><snip>

Thanks for this patch. I confirm that it fixes my problem.

Regards,
Samuel Jones