[patch] Fix pr24486

[Rafael Espíndola via llvm-commits]

r249303.

On 5 October 2015 at 04:09, Maxim Ostapenko
<m.ostapenko at partner.samsung.com> wrote:
> On 03/10/15 21:50, Rafael Espíndola wrote:
>>>
>>> Thank you for your patch.
>>>
>>> I've performed some investigation for failures occurred and found two
>>> issues:
>>>
>>> - We should assign a dummy fragment in current section in EmitLabel
>>> function.
>>> - In MCObjectStreamer::getCurrentFragment(), we should skip dummy
>>> fragments
>>> (or perhaps just not push them into appropriate section's fragments list
>>> at
>>> all). If we don't do this, MCExpr::evaluateAsAbsolute() will fail to
>>> evaluate .La - .Lb expressions.
>>
>> Thanks for all the work on this.
>>
>> There were still some things that needed improvement:
>>
>> * There should really only be one dummy fragment per section.
>> * Ideally they would only be used by the asm streamer and not show up
>> in during relaxation.
>>
>> The attached patch is an extension of yours.The only exception to
>> second goal is that the dummy fragment is still used as a placeholder
>> in the pending labels optimization. The entire area of bundles is a
>> bit fuzzy to me. In any case, it is a local enough issue that it is
>> probably not worth making this patch bigger to clean it up.
>>
>> So, the summary is that the attached patch fixes the bug and LGTM.
>> Does it work for you? Do you have commit access?
>>
>> Cheers,
>> Rafael
>
>
> The patch works perfectly fine for me, thank you!
>
> Actually I haven't a commit access to LLVM repository, so could you please
> land it?
>
> Thanks,
> -Maxim