slow links due to merge_strings

[Kevin Nomura <nomura at netapp dot com>]

binutils 2.13.2.1

On a link of 4000 object files, link times went from 13 seconds to 300
when switching from stabs to dwarf2 debug format.  gprof showed this
time to be coming from merge_strings() in bfd/merge.c and descendants 
doing hash lookups.  The hash tables were lightly loaded but experiencing 
90% collision rates, with 600 million calls to last4_eq.

The following code in merge_strings looked odd:

      s = e->root.string + e->len - e->u.entsize;
      hash = 0;
      for (i = 0; i < e->u.entsize; i++)
        {
          c = *--s;
          hash += c + (c << 17);
          hash ^= hash >> 2;
        }

e->u.entsize == 1, I guess because strings are being hashed, and the
"entsize" of a character string is one character.

The code seems to be trying to set s to the end of the string and 
work backwards.  But it won't get very far with e->u.entsize == 1.

As an experiment I made this change:

      for (i = 0; i < e->len; i++)

Link time decreased from 300 -> 60 seconds with a saner-looking profile.

Each sample counts as 0.01 seconds.
  %   cumulative   self              self     total           
 time   seconds   seconds    calls   s/call   s/call  name    
 36.79     14.10    14.10 23787199     0.00     0.00  sec_merge_hash_lookup
 30.34     25.73    11.63 27330752     0.00     0.00  last4_eq
 12.26     30.43     4.70   304477     0.00     0.00  htab_find_slot_with_hash
  3.44     31.75     1.32 12919123     0.00     0.00  _bfd_merged_section_offset
  3.03     32.91     1.16    19092     0.00     0.00  elf_i386_relocate_section
...

The executable output file was bitwise identical.