[lld] [lld] Merge GOT entries for symbols that have been ICFed (PR #131630)

[Peter Smith via llvm-commits]

================
@@ -667,6 +667,20 @@ GotSection::GotSection(Ctx &ctx)
 void GotSection::addConstant(const Relocation &r) { relocations.push_back(r); }
 void GotSection::addEntry(const Symbol &sym) {
   assert(sym.auxIdx == ctx.symAux.size() - 1);
+  auto *d = dyn_cast<Defined>(&sym);
+  if (d && ctx.arg.icf != ICFLevel::None) {
----------------
smithp35 wrote:

I think the `if (!inserted && d->folded)` is an attempt to prevent that case, which I think would work for a single section, but on reflection it isn't sufficient as there could always be ICF'ed sections that `d->folded` would be true.

With respect to
> That prevents ICF on the sections containing the relocation but not the target sections.

Just to check my understanding, this property will prevent sections containing a GOT generating ADRP, or ADD/LDR relocation to a preemptible symbol from being folded. Even if the ADRP is separated into a different section, each "pair" of relocations will refer to the same folded symbol, so the GOT calculations will be correct.

We're left with non-preemptible symbols which no-one should be able to tell the difference if we merge the entries.

One corner case that I haven't thought through yet is CHERI capabilities which write information on the capability to be created at run-time into the GOT slot. These aren't strictly an upstream concern though. I expect that the CHERI fork of LLD already has some code in ICF to avoid merging sections with different derived capabilities (symbol size, type etc.) 

https://github.com/llvm/llvm-project/pull/131630