[patch] Fix pr25919

[Michael Zolotukhin via llvm-commits]

Hi Rafael,

I temporarily reverted r256105 to unbreak our bots in r257751. Please feel free to recommit with a fix.

Thanks,
Michael

> On Jan 13, 2016, at 7:11 PM, Michael Zolotukhin <mzolotukhin at apple.com> wrote:
> 
> Hi Rafael,
> 
> We’ve started to hit this assert in our LLVM bootstraps. Do you expect this fix to be committed anytime soon, or is it better just to revert for now?
> 
> Thanks,
> Michael
>> On Jan 8, 2016, at 3:20 PM, Mikhail Zolotukhin <mzolotukhin at apple.com> wrote:
>> 
>> Hi Rafael,
>> 
>> I recently hit the same bug while building 483.xalanc from SPEC2006. Do you mind reverting the original patch while the fix is under review?
>> 
>> Thanks,
>> Michael
>> 
>>> On Dec 22, 2015, at 7:46 AM, Rafael Espíndola via llvm-commits <llvm-commits at lists.llvm.org> wrote:
>>> 
>>> The asserts for checking all uses while the module is being
>>> materialized found the first problem.
>>> 
>>> The issues happens when libLTO is trying to set
>>> LTO_SYMBOL_SCOPE_DEFAULT_CAN_BE_HIDDEN. It needs all uses in order to
>>> produce a valid answer.
>>> 
>>> The attached patch adds a graceful degradation.
>>> 
>>> What is done instead in the gold plugin is to check the uses in the
>>> merged module.
>>> 
>>> If this degrades quality too much, a solution would be to check the
>>> uses and add unnamed_addr marks at the end of IRGen. That way
>>> canBeOmittedFromSymbolTable would not need to run analyzeGlobal at
>>> all.
>>> 
>>> Cheers,
>>> Rafael
>>> <t.diff>_______________________________________________
>>> llvm-commits mailing list
>>> llvm-commits at lists.llvm.org
>>> http://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-commits
>> 
>