Fix ppc64 stub generation for non-ABI-compliant objects

Fix ppc64 stub generation for non-ABI-compliant objects

[Alan Modra]

Someone sent me a 200M powerpc64-linux testcase, that had (at least) one
function lacking a function descriptor.  This isn't exactly
ABI-compliant, but ld tries to cope as best it may, creating fake
function descriptors for these functions.  However, it so happened that
some call sites to the function in question could not be reached by a
simple branch, but ld was confused by the fake function descriptor into
not building a long branch stub.

	* elf64-ppc.c (ppc_type_of_stub): Check both func desc and func
	entry sym before deciding no stub is needed.
	(ppc64_elf_size_stubs): When calculating branch destination,
	don't use func desc sym for old ABI objects unless func entry
	is undefined.

Index: bfd/elf64-ppc.c
===================================================================
RCS file: /cvs/src/src/bfd/elf64-ppc.c,v
retrieving revision 1.221
diff -u -p -r1.221 elf64-ppc.c
--- bfd/elf64-ppc.c	19 Sep 2005 09:35:26 -0000	1.221
+++ bfd/elf64-ppc.c	30 Sep 2005 07:17:11 -0000
@@ -7888,26 +7888,34 @@ ppc_type_of_stub (asection *input_sec,
 
   if (h != NULL)
     {
-      if (h->oh != NULL
-	  && h->oh->is_func_descriptor)
-	h = h->oh;
+      struct ppc_link_hash_entry *fdh = h;
+      if (fdh->oh != NULL
+	  && fdh->oh->is_func_descriptor)
+	fdh = fdh->oh;
 
-      if (h->elf.dynindx != -1)
+      if (fdh->elf.dynindx != -1)
 	{
 	  struct plt_entry *ent;
 
-	  for (ent = h->elf.plt.plist; ent != NULL; ent = ent->next)
+	  for (ent = fdh->elf.plt.plist; ent != NULL; ent = ent->next)
 	    if (ent->addend == rel->r_addend
 		&& ent->plt.offset != (bfd_vma) -1)
 	      {
-		*hash = h;
+		*hash = fdh;
 		return ppc_stub_plt_call;
 	      }
 	}
 
-      if (!(h->elf.root.type == bfd_link_hash_defined
-	    || h->elf.root.type == bfd_link_hash_defweak)
-	  || h->elf.root.u.def.section->output_section == NULL)
+      /* Here, we know we don't have a plt entry.  If we don't have a
+	 either a defined function descriptor or a defined entry symbol
+	 in a regular object file, then it is pointless trying to make
+	 any other type of stub.  */
+      if (!((fdh->elf.root.type == bfd_link_hash_defined
+	    || fdh->elf.root.type == bfd_link_hash_defweak)
+	    && fdh->elf.root.u.def.section->output_section != NULL)
+	  && !((h->elf.root.type == bfd_link_hash_defined
+		|| h->elf.root.type == bfd_link_hash_defweak)
+	       && h->elf.root.u.def.section->output_section != NULL))
 	return ppc_stub_none;
     }
 
@@ -8951,16 +8959,25 @@ ppc64_elf_size_stubs (bfd *output_bfd,
 
 		  ok_dest = FALSE;
 		  fdh = NULL;
+		  sym_value = 0;
 		  if (hash == NULL)
 		    {
 		      sym_value = sym->st_value;
 		      ok_dest = TRUE;
 		    }
-		  else
+		  else if (hash->elf.root.type == bfd_link_hash_defined
+			   || hash->elf.root.type == bfd_link_hash_defweak)
+		    {
+		      sym_value = hash->elf.root.u.def.value;
+		      if (sym_sec->output_section != NULL)
+			ok_dest = TRUE;
+		    }
+		  else if (hash->elf.root.type == bfd_link_hash_undefweak
+			   || hash->elf.root.type == bfd_link_hash_undefined)
 		    {
-		      sym_value = 0;
 		      /* Recognise an old ABI func code entry sym, and
-			 use the func descriptor sym instead.  */
+			 use the func descriptor sym instead if it is
+			 defined.  */
 		      if (hash->elf.root.root.string[0] == '.'
 			  && (fdh = get_fdh (hash, htab)) != NULL)
 			{
@@ -8975,22 +8992,11 @@ ppc64_elf_size_stubs (bfd *output_bfd,
 			  else
 			    fdh = NULL;
 			}
-		      else if (hash->elf.root.type == bfd_link_hash_defined
-			       || hash->elf.root.type == bfd_link_hash_defweak)
-			{
-			  sym_value = hash->elf.root.u.def.value;
-			  if (sym_sec->output_section != NULL)
-			    ok_dest = TRUE;
-			}
-		      else if (hash->elf.root.type == bfd_link_hash_undefweak)
-			;
-		      else if (hash->elf.root.type == bfd_link_hash_undefined)
-			;
-		      else
-			{
-			  bfd_set_error (bfd_error_bad_value);
-			  goto error_ret_free_internal;
-			}
+		    }
+		  else
+		    {
+		      bfd_set_error (bfd_error_bad_value);
+		      goto error_ret_free_internal;
 		    }
 
 		  destination = 0;

-- 
Alan Modra
IBM OzLabs - Linux Technology Centre