Support dead-stripping in ELF objects

[Robinson, Paul]

From: Rafael Espíndola [mailto:rafael.espindola at gmail.com]
> > == Deadstripping ==
> >
> > The deadstrip roots is the set of all symbols defined in sections not
> > marked
> > SHF_DEADSTRIP, symbols with STF_NO_DEADSTRIP, and any implementation
> > defined
> > symbols.
> >
> > A symbol is live if it is a deadstrip root, or it is referenced from a
> > deadstrip root or another live symbol, or it is in a section not
> > marked
> > SHF_SUBSECTION_VIA_SYMBOLS that has other live symbols.
> >
> > A symbol X is referenced by Z if:
> > * If Z and X are not in a section marked SHF_VALID_SIZE.
> >   * There is a relocation in the section containing Z with a resolved
> > symbol
> > which comes from the same file and has the same st_shndx as X (the
> > section
> > containing Z references the section containing X).
> > * If Z is not in a section marked SHF_VALID_SIZE but X is.
> >   * There is a relocation in the section containing Z with the
> > resolved
> > symbol X.
> > * If both Z and X are in sections marked SHF_VALID_SIZE.
> >   * There is a relocation with an r_offset within the range [st_value,
> > st_value + st_size) of Z with the resolved symbol X.
> >
> > Any symbol that is not live may be removed if allowed by section
> > flags.
> 
> Do we need to compute this? We should only need to know if a symbol is
> referenced, not who is referencing it. With that we can just say that
> a symbol is referenced if
> 
> * A relocation points to it.
> * It is in a section without SHF_SUBSECTION_VIA_SYMBOLS and a
> relocation points to that section.
> 
> My concern is that in the non  SHF_SUBSECTION_VIA_SYMBOLS case the
> full symbol reference graph as defined above would be n^2 in cases
> like .text and .eh_frame. There is (approximately) one symbol in
> .eh_frame for each symbol in .text and it has a relocation to the
> corresponding .text symbol. We would get n^2 edges without
> SHF_SUBSECTION_VIA_SYMBOLS.

For dead-stripping purposes, it seems to me you don't care about symbols
in their own right, you care about [sub]sections.  You also don't care about
the full graph of subsection references, you care about the liveness property.
That would be a lot cheaper to compute.
 
> > - Michael Spencer
> >
> 
> Cheers,
> Rafael

--paulr