Make sure all the relocations are filled in for partially cloned targ…

…et (#44262)

We collect the relocations (i.e. the GOT slots that is used in the code) for each target
in `tgt.relocs`. Needing a relocation, however, does not imply that the function is cloned
for this target within the group (It does mean that at least one target
in the group has it cloned). The previous version would miss the relocation in this case.

This was triggerred with the following cloning situation

    caller: clone_1
    callee: clone_1, clone_1.clone_3

Since caller.clone_1 may call either callee.clone_1 or callee.clone_1.clone_3 a relocation
for callee will be used and is required to be initialized.
In addition to target 1, target 2 (and in fact target 3) within group 1
will also use caller.clone_1. However, since callee isn't cloned for target 2
the previous version wouldn't have saved this slot in the relocation array.

src/llvm-multiversioning.cpp

@@ -1037,15 +1037,15 @@ void CloneCtx::emit_metadata()
                 idxs.push_back(baseidx);
                 for (uint32_t j = 0; j < nfvars; j++) {
                     auto base_f = grp->base_func(fvars[j]);
-                    if (shared_relocs.count(j)) {
+                    if (shared_relocs.count(j) || tgt->relocs.count(j)) {
                         count++;
                         idxs.push_back(jl_sysimg_tag_mask | j);
                         auto f = map_get(*tgt->vmap, base_f, base_f);
                         offsets.push_back(get_ptrdiff32(cast<Function>(f), fbase));
                     }
                     else if (auto f = map_get(*tgt->vmap, base_f)) {
                         count++;
-                        idxs.push_back(tgt->relocs.count(j) ? (jl_sysimg_tag_mask | j) : j);
+                        idxs.push_back(j);
                         offsets.push_back(get_ptrdiff32(cast<Function>(f), fbase));
                     }
                 }