On Tue, Aug 15, 2017 at 1:30 AM, Richard Biener <rguent...@suse.de> wrote:
> On Mon, 14 Aug 2017, Ian Lance Taylor wrote:
>
>> It seems to me that if you keep any section of an SHT_GROUP, you need
>> to keep all the sections. Otherwise the section indexes in the group
>> will be left dangling.
>
> Note that all sections are "kept", removed sections are just marked
> as SHT_NULL. Which means that indeed some group members might
> be SHT_NULL.
>
> Note that I don't remember if I actually run into any SHT_GROUP
> member being necessary when another is not (IIRC this was with
> debug types sections or so). I'll double-check and add a comment
> where this matters for early-debug (the simple-object code tries
> to be more generic but obviously all testing was done with
> early-debug section copying in mind).
In general they are SHT_GROUP for a reason, that reason being that if
you keep one of the group sections you need to keep all of them. If
that didn't matter, there wouldn't be a SHT_GROUP section at all. So
I think it would be appropriate that if you keep one of a group you
keep all of the group.
>> The symbol table handling is pretty awful. Can't you just remove the
>> symbols you don't want? You will have to update the sh_info field of
>> the SHT_SYMTAB section, which holds the index of the first STB_GLOBAL
>> symbol in the table.
>
> Maybe that's what I was missing, let me see if I can make it work.
> It looks like if there's no STB_GLOBAL symbol the index is one after
> the last symbol and it seems it is the first ! STB_LOCAL symbol
> ("One greater than the symbol table index of the last local symbol
> (binding STB_LOCAL)").
>
> Ah, and implementing this now I remember why I chose the "awkward"
> way. This was to preserve symtab indices to not need to rewrite
> relocation sections... There isn't a documented way to make
> an "none" symtab entry (eventually duplicating UND might work
> but IIRC it messes up ordering as UND is STB_LOCAL and thus may
> not appear after the first ! STB_LOCAL symbol). As comments in
> the code indicate I tried to mangle things in a way that makes
> both GNU ld and gold happy...
Bother.
> Any suggestion how to improve things when not removing symbols?
> (I could of course remove things from the end of the symtab)
> Like replacing local symbols with UND and globals with
> a COMMON of zero size (and no name)?
You can have a UND global symbol, after all. You just shouldn't
simply zero out the entire symbol, as that will indeed change the
binding to STB_LOCAL. So probably the best approach is to change
every unwanted symbol to SHN_UNDEF, keep the binding as is, and zero
all the other fields. That should let the linker simply discard the
symbol, I hope.
> I suppose rewriting relocation sections is possible but I tried
> to do as little as necessary (like not actually removing sections).
The patch is OK with the above suggestions, assuming they work.
Ian
