[PATCH] Fix PR28575.

[Rafael Espíndola via llvm-commits]

On 19 July 2016 at 13:34, Rui Ueyama <ruiu at google.com> wrote:
> On Tue, Jul 19, 2016 at 7:16 AM, Rafael Espíndola
> <rafael.espindola at gmail.com> wrote:
>>
>> > I think you want to remove "That means ..." part because it doesn't
>> > describe the behavior of this function.
>>
>> Done.
>>
>> > Why don't you store relocation indices to section pieces? It would be
>> > better than having parallel arrays.
>>
>> The common use is for SHF_MERGE sections that don't have relocations.
>> The existing code is trying to pack it to be cache friendly (uses
>> bitfield).
>>
>> I could just remove SplitInputSection and have MergeInputSection and
>> EhInputSection declare their own Pieces array. The one in
>> EhInputSection would have the first relocation of each piece. What do
>> you think?
>
>
> Because of my previous change to use a hash table rather than doing binary
> search, I think cache friendliness of SectionPiece is now less important. So
> you may just want to add a new field. You may want to benchmark?


Good point. I will give it a try.

Cheers,
Rafael