Re: [RFC][PATCH] Check ELF relocs after opening all all input files

["H.J. Lu" <hjl dot tools at gmail dot com>]

On Thu, Apr 28, 2016 at 6:40 AM, H.J. Lu <hjl.tools@gmail.com> wrote:
> On Thu, Apr 28, 2016 at 6:09 AM, H.J. Lu <hjl.tools@gmail.com> wrote:
>> On Thu, Apr 28, 2016 at 6:07 AM, Alan Modra <amodra@gmail.com> wrote:
>>> On Thu, Apr 28, 2016 at 05:49:57AM -0700, H.J. Lu wrote:
>>>> On Wed, Apr 27, 2016 at 6:11 PM, Alan Modra <amodra@gmail.com> wrote:
>>>> > On Wed, Apr 27, 2016 at 03:24:48PM -0700, H.J. Lu wrote:
>>>> >> bfd_elf_record_link_assignment
>>>> >> is called after check_relocs.  bfd_elf_record_link_assignment sets non_elf,
>>>> >> def_regular and  forced_local.   For PROVIDE, it also updates root.type. They
>>>> >> are needed in reloc_checks.
>>>> >
>>>> > My guess is that symbol twiddling done in before_allocation should be
>>>> > moved to a new ldemul hook called at the start of lang_do_assignments.
>>>> > The idea being to stabilize symbols earlier.
>>>> >
>>>> > The hook would twiddle __ehdr_start and call find_statement_assignment
>>>> > when lang_mark_phase_enum.  Reversing the __ehdr_start twiddle stays
>>>> > in before_allocation.
>>>> >
>>>>
>>>> I tried this.  But it doesn't work with __start/__stop symbols.  I
>>>> need to know if they are defined and referenced local in check_relocs.
>>> [snip]
>>>
>>>> --- a/ld/ldlang.c
>>>> +++ b/ld/ldlang.c
>>>> @@ -6930,6 +6930,8 @@ lang_process (void)
>>>>       collection in order to make sure that all symbol aliases are resolved.  */
>>>>    lang_do_assignments (lang_mark_phase_enum);
>>>>
>>>> +  ldemul_record_link_assignments (lang_mark_phase_enum);
>>>> +
>>>>    lang_do_memory_regions();
>>>>    expld.phase = lang_first_phase_enum;
>>>
>>> You'll need to run ldemul_record_link_assignments before
>>> lang_do_assignments if you want provided symbols to be defined.
>>
>> I got many more failures when I did that since many assignments
>> haven't been processed yet.
>>
>>> I suggest renaming to ldemul_do_assignments and putting
>>>   ldemul_do_assignments (phase);
>>> inside lang_do_assignments just before the call to
>>> lang_do_assignments_1.
>>>
>
> With this patch, I only got 2 failures on x86-64.
>

This one works for both i386 and x86-64.


-- 
H.J.

From b6809eeed88f48dcbe71c1e4ce1213748b0bfe22 Mon Sep 17 00:00:00 2001
From: "H.J. Lu" <hjl.tools@gmail.com>
Date: Thu, 28 Apr 2016 06:38:41 -0700
Subject: [PATCH] Add _bfd_elf_record_start_stop

---
 bfd/elf-bfd.h         |  3 ++
 bfd/elflink.c         | 87 ++++++++++++++++++++++++++++++++++++++++-----------
 ld/emultempl/elf32.em |  4 +++
 3 files changed, 75 insertions(+), 19 deletions(-)

diff --git a/bfd/elf-bfd.h b/bfd/elf-bfd.h
index 6c05b55..49dc70b 100644
--- a/bfd/elf-bfd.h
+++ b/bfd/elf-bfd.h
@@ -2339,6 +2339,9 @@ extern bfd_boolean bfd_elf_gc_common_final_link
 extern asection *_bfd_elf_is_start_stop
   (const struct bfd_link_info *, struct elf_link_hash_entry *);
 
+extern void _bfd_elf_record_start_stop
+  (struct bfd_link_info *);
+
 extern bfd_boolean bfd_elf_reloc_symbol_deleted_p
   (bfd_vma, void *);
 
diff --git a/bfd/elflink.c b/bfd/elflink.c
index b6ff6b6..9395052 100644
--- a/bfd/elflink.c
+++ b/bfd/elflink.c
@@ -12230,8 +12230,9 @@ _bfd_elf_gc_mark_hook (asection *sec,
   return NULL;
 }
 
-/* For undefined __start_<name> and __stop_<name> symbols, return the
-   first input section matching <name>.  Return NULL otherwise.  */
+/* For __start_<name> and __stop_<name> symbols, return the first
+   input section matching <name> in a regular object.  Return NULL
+   otherwise.  */
 
 asection *
 _bfd_elf_is_start_stop (const struct bfd_link_info *info,
@@ -12239,17 +12240,28 @@ _bfd_elf_is_start_stop (const struct bfd_link_info *info,
 {
   asection *s;
   const char *sec_name;
+  bfd_boolean undefined = (h->root.type == bfd_link_hash_undefined
+			   || h->root.type == bfd_link_hash_undefweak);
 
-  if (h->root.type != bfd_link_hash_undefined
-      && h->root.type != bfd_link_hash_undefweak)
-    return NULL;
-
-  s = h->root.u.undef.section;
-  if (s != NULL)
+  if (undefined)
+    {
+      s = h->root.u.undef.section;
+      if (s != NULL)
+	{
+	  if (s == (asection *) 0 - 1)
+	    return NULL;
+	  return s;
+	}
+    }
+  else
     {
-      if (s == (asection *) 0 - 1)
+      /* Symbol is defined.  Check if it is also defined in a regular
+	 input file only when it is currently defined in a dynamic
+	 object, since otherwise, it can't be a __start_<name> nor
+	 __stop_<name> symbol.  */
+      if (!h->def_dynamic)
 	return NULL;
-      return s;
+      s = NULL;
     }
 
   sec_name = NULL;
@@ -12262,18 +12274,21 @@ _bfd_elf_is_start_stop (const struct bfd_link_info *info,
     {
       bfd *i;
 
+      /* Only check regular input files.  */
       for (i = info->input_bfds; i != NULL; i = i->link.next)
-	{
-	  s = bfd_get_section_by_name (i, sec_name);
-	  if (s != NULL)
-	    {
-	      h->root.u.undef.section = s;
-	      break;
-	    }
-	}
+	if ((i->flags
+	     & (DYNAMIC | BFD_LINKER_CREATED | BFD_PLUGIN)) == 0)
+	  {
+	    s = bfd_get_section_by_name (i, sec_name);
+	    if (s != NULL)
+	      {
+		h->root.u.undef.section = s;
+		break;
+	      }
+	  }
     }
 
-  if (s == NULL)
+  if (undefined && s == NULL)
     h->root.u.undef.section = (asection *) 0 - 1;
 
   return s;
@@ -13791,3 +13806,37 @@ elf_append_rel (bfd *abfd, asection *s, Elf_Internal_Rela *rel)
   BFD_ASSERT (loc + bed->s->sizeof_rel <= s->contents + s->size);
   bed->s->swap_reloc_out (abfd, rel, loc);
 }
+
+/* Record __start_<name> and __stop_<name> symbols referenced by
+   regular objects.  This is called via elf_link_hash_traverse.  */
+
+static bfd_boolean
+elf_link_record_start_stop (struct elf_link_hash_entry *h, void *data)
+{
+  if (!h->def_regular && h->ref_regular)
+    {
+      asection *sec
+	= _bfd_elf_is_start_stop ((struct bfd_link_info *) data, h);
+      if (sec != NULL)
+	{
+	  h->def_regular = 1;
+	  h->type = STT_OBJECT;
+	  /* If it is currently defined by a dynamic object, but not
+	     by a regular object, then mark it as undefined so that
+	     the generic linker will force the correct value.  */
+	  if (h->def_dynamic)
+	    h->root.type = bfd_link_hash_undefined;
+	}
+    }
+  return TRUE;
+}
+
+/* Record all __start_<name> and __stop_<name> symbols referenced by
+   regular objects.  */
+
+void
+_bfd_elf_record_start_stop (struct bfd_link_info * info)
+{
+  elf_link_hash_traverse (elf_hash_table (info),
+			  elf_link_record_start_stop, info);
+}
diff --git a/ld/emultempl/elf32.em b/ld/emultempl/elf32.em
index cc2f056..86105fe 100644
--- a/ld/emultempl/elf32.em
+++ b/ld/emultempl/elf32.em
@@ -1445,6 +1445,10 @@ gld${EMULATION_NAME}_record_link_assignments (lang_phase_type phase)
 	 let the ELF backend know about them in case the variables are
 	 referred to by dynamic objects.  */
       lang_for_each_statement (gld${EMULATION_NAME}_find_statement_assignment);
+
+      /* Also let the ELF backend know about __start_<name> and
+	 __stop_<name> symbols.  */
+      _bfd_elf_record_start_stop (&link_info);
     }
 }
 
-- 
2.5.5