[PATCH] D14090: [ELF2] R_X86_64_COPY relocation implemented

[Rui Ueyama via llvm-commits]

ruiu added inline comments.

================
Comment at: ELF/Writer.cpp:203-209
@@ -202,2 +202,9 @@
     if (Body) {
+      bool NeedsCopy = false;
+      if (auto *E = dyn_cast<SharedSymbol<ELFT>>(Body)) {
+        NeedsCopy = Target->relocNeedsCopy(Type, *Body);
+        if (NeedsCopy && E->NeedsCopy)
+          continue;
+        E->NeedsCopy = NeedsCopy;
+      }
       NeedsPlt = Target->relocNeedsPlt(Type, *Body);
----------------
grimar wrote:
> grimar wrote:
> > ruiu wrote:
> > > I'm trying to understand this. So, if two or more relocations exist for a shared symbol, the last relocation's NeedsCopy may override the previous results? Is this correct?
> > What kind of relocations ? According to current logic plt relocations will never happen because it is not function and so the only possible case is relocNeedsGot() which implemented as
> > 
> > 
> > bool X86_64TargetInfo::relocNeedsGot(uint32_t Type, const SymbolBody &S) const {
> >   return Type == R_X86_64_GOTPCREL || relocNeedsPlt(Type, S);
> > }
> > 
> > 
> > that gives us the only case is R_X86_64_GOTPCREL.
> > Is it possible that symbol will require both R_X86_64_GOTPCREL and copy relocation ? If I understand correctly that case - we can use copy relocation here. So I see nothing wrong, please correct me if I am mistaken.
> I am wrong here I think, I need to investigate that case, pelase ignore my comment.
Is that the same as this?

  if (auto *E = dyn_cast<SharedSymbol<ELFT>>(Body)) {
    E->NeedsCopy = E->NeedsCopy || Target->relocNeedsCopy(Type, *Body);
    if (E->NeedsCopy)
      continue;
  }


http://reviews.llvm.org/D14090