objcopy: restore behaviour required by GCC's build

[emaste]

Summary

In rS339350 (D17519) filter_reloc() was removed, to fix the case of stripping statically linked binaries with relocations (which may come from ifunc use, for example). As a side effect this changed the behaviour when stripping object files - the output was broken both before and after rS339350 in different ways. Unfortunately GCC's build process relies on the previous behaviour, so:

• Revert rS339350, restoring filter_reloc().
• Fix an unitialized variable use (commited as r3638 in ELF Tool Chain).
• Change filter_reloc() to omit relocations referencing removed symbols, while retaining relocations with no symbol reference.

(Note this review represents only the third change.)

Test Plan

• strip a .o file, observe relocations have been removed
• strip a statically linked binary containing ifunc, observe sym-less relocation remains

[emaste]

This comment has been deleted.

[kaiw]

Ed, patch looks good. Just need a small change. When symbol table does not exist in the output, we need to check
and skip relocation sections for dynamic symbols. Otherwise strip will segfault when called on dynamic linked executables
that's already stripped.

[kaiw]

I've emailed you a modfied patch.

[emaste]

Dynamic symtab fix from @kaiw

[kaiw]

I applied the new patch and all the test cases passed!

[gerald]

Thanks for tackling this, Ed.

Do I read this correctly that once this is applied, the lang/gcc* ports should be bootstrapping fine again? Or do we need to look into further changes related to those?

[emaste]

In D17596#376063, @gerald wrote:

Do I read this correctly that once this is applied, the lang/gcc* ports should be bootstrapping fine again?

Correct, they should.