[RFC PATCH] Don't put local or undefined symbols into .hash

[Jakub Jelinek <jakub at redhat dot com>]

Hi!

It seems an average shared library has roughly 25% of undefined symbols
in its .dynsym section.
Currently, the linker puts all dynamic symbols (except the special ones like
section symbols or symbol 0) into the .hash section, but the dynamic linker
will always skip STB_LOCAL symbols or undefined symbols (except those with
st_value != 0) - it is always looking for symbol definitions.
Is there any reason why we put even the undefined or local symbols into
.hash?  It just makes the .hash buckets array larger and/or the hash chains
longer than necessary.
Or is there something that relies on all symbols being there?
So far I have noticed just readelf -Ds (but in that case it is questionable
if readelf -Ds would need to change or stay as is).

2006-06-26  Jakub Jelinek  <jakub@redhat.com>

	* elflink.c (elf_collect_hash_codes): Set u.elf_hash_value to ~0
	for local or undefined symbols, don't include it into hashcodes
	array.
	(elf_link_output_extsym): Don't put symbols with u.elf_hash_value
	~0 into the .hash section.

--- bfd/elflink.c.jj	2006-06-20 18:34:53.000000000 +0200
+++ bfd/elflink.c	2006-06-26 14:04:49.000000000 +0200
@@ -4785,6 +4785,17 @@ elf_collect_hash_codes (struct elf_link_
   if (h->dynindx == -1)
     return TRUE;
 
+  /* Ignore also local symbols and undefined symbols.  */
+  if (h->forced_local
+      || h->root.type == bfd_link_hash_undefined
+      || h->root.type == bfd_link_hash_undefweak)
+    {
+      /* bfd_elf_hash values never have topmost 4 bits set,
+	 so ~0 don't collide with any valid hash value.  */
+      h->u.elf_hash_value = ~(unsigned long) 0;
+      return TRUE;
+    }
+
   name = h->root.root.string;
   p = strchr (name, ELF_VER_CHR);
   if (p != NULL)
@@ -6678,16 +6689,24 @@ elf_link_output_extsym (struct elf_link_
       bed->s->swap_symbol_out (finfo->output_bfd, &sym, esym, 0);
 
       bucketcount = elf_hash_table (finfo->info)->bucketcount;
-      bucket = h->u.elf_hash_value % bucketcount;
-      hash_entry_size
-	= elf_section_data (finfo->hash_sec)->this_hdr.sh_entsize;
-      bucketpos = ((bfd_byte *) finfo->hash_sec->contents
-		   + (bucket + 2) * hash_entry_size);
-      chain = bfd_get (8 * hash_entry_size, finfo->output_bfd, bucketpos);
-      bfd_put (8 * hash_entry_size, finfo->output_bfd, h->dynindx, bucketpos);
-      bfd_put (8 * hash_entry_size, finfo->output_bfd, chain,
-	       ((bfd_byte *) finfo->hash_sec->contents
-		+ (bucketcount + 2 + h->dynindx) * hash_entry_size));
+      if (h->u.elf_hash_value == ~(unsigned long) 0)
+	BFD_ASSERT (h->forced_local
+		    || h->root.type == bfd_link_hash_undefined
+		    || h->root.type == bfd_link_hash_undefweak);
+      else
+	{
+	  bucket = h->u.elf_hash_value % bucketcount;
+	  hash_entry_size
+	    = elf_section_data (finfo->hash_sec)->this_hdr.sh_entsize;
+	  bucketpos = ((bfd_byte *) finfo->hash_sec->contents
+		      + (bucket + 2) * hash_entry_size);
+	  chain = bfd_get (8 * hash_entry_size, finfo->output_bfd, bucketpos);
+	  bfd_put (8 * hash_entry_size, finfo->output_bfd, h->dynindx,
+		   bucketpos);
+	  bfd_put (8 * hash_entry_size, finfo->output_bfd, chain,
+		   ((bfd_byte *) finfo->hash_sec->contents
+		    + (bucketcount + 2 + h->dynindx) * hash_entry_size));
+	}
 
       if (finfo->symver_sec != NULL && finfo->symver_sec->contents != NULL)
 	{

	Jakub