tree 63b5ec60a4cb5f5c16c7ecc2bc7ad0c7764ba18a
parent eb8a62477938290a3734e3e53939ffeac9f44600
author Jez Ng <jezng@fb.com> 1618535669 -0400
committer Copybara-Service <copybara-worker@google.com> 1618601098 -0700

[lld-macho] LTO: Unset VisibleToRegularObj where possible

This allows LLVM's LTO to internalize symbols that are not referenced
directly by regular objects. Naturally, this means we need to track
which symbols are referenced by regular objects. The approach taken here
is similar to LLD-COFF's: like the COFF port, we extend
`SymbolTable::insert()` to set the isVisibleToRegularObj bit. (LLD-ELF
relies on the Symbol constructor and `Symbol::mergeProperties()`, but
the Mach-O port does not have a `mergeProperties()` equivalent.)

From what I can tell, ld64 (which uses libLTO) doesn't do this
optimization at all. I'm not even sure libLTO provides a way to do this.
Not having ld64's behavior as a reference implementation is unfortunate;
instead, I am relying on LLD-ELF/COFF's behavior as references while
erring on the conservative side. In particular, LLD-MachO will only do
this optimization for executables right now.

We also don't attempt it when `-flat_namespace` is used -- otherwise
we'd need scan the symbol table to find matches for every un-namespaced
symbol reference, which is expensive.

internalize.ll is based off the LLD-ELF tests `internalize-basic.ll` and
`internalize-undef.ll`. Looks like @davide added some of LLD-ELF's internalize
tests, so adding him as a reviewer...

Reviewed By: #lld-macho, gkm

Differential Revision: https://reviews.llvm.org/D99105

GitOrigin-RevId: eb5b7d4497e323cf6214eb3e008dc37bc5ed1fd7