Re: [GOLD] Relayout?

[Ian Lance Taylor <iant at google dot com>]

"Doug Kwan (éæå)" <dougkwan@google.com> writes:

> 2009/8/24 Ian Lance Taylor <iant@google.com>:
>
>> This does not make sense conceptually, at least in my concept of layout.
>> When we reach Layout::finalize, the linker is supposed to know the final
>> sizes of all normal sections. ÂLayout::finalize then handles creating
>> the symbol table, the dynamic section and so forth. ÂRunning that code
>> again does not make sense.
>
> Unfortunately, stubs have symbols and they get exported by currently
> by gnu ld.  The ARM ELF specs say that veneer symbols have the same
> visibility as the symbols they veneered.  So the we need to rebuild
> the dynamic symbol table at least.

Well, I think that all we need to change is the size of the dynamic
symbol table.  But that might not really be any simpler.


>> So I don't think unfinalize is conceptually the way to go. ÂI think the
>> right approach is going to be along the lines of doing the first step of
>> finalization (compute the dynamic symbol table, et. al.), then a loop
>> which sets section addresses and calls a target specific relaxation
>> routine, then the last step of finalization (compute the regular symbol
>> table, et. al.).
>
> That sound's simpler.  Maybe I should pursue along this line and add
> capability to rebuild the dynamic symbol table incrementally?  We
> always add relocation stubs during relaxation so the dynamic symbol
> table can only grow.

OK.

Ian