Faster string merging

* use power-of-two hash table
* use better hash function (hashing 32bits at once and with better
  mixing characteristics)
* use input-offset-to-entry maps instead of retaining full input
  contents for lookup time
* don't reread SEC_MERGE section multiple times
* care for cache behaviour for the hot lookup routine

The overall effect is less usage in libz and much faster string merging
itself.  On a debug-info-enabled cc1 the effect at the time of this
writing on the machine I used was going from 14400 perf samples to 9300
perf samples or from 3.7 seconds to 2.4 seconds, i.e. about 33% .

bfd/elflink.c

@@ -11361,6 +11361,13 @@ elf_link_input_bfd (struct elf_final_link_info *flinfo, bfd *input_bfd)
 	      contents = flinfo->contents;
 	    }
 	}
+      else if (!(o->flags & SEC_RELOC)
+	       && !bed->elf_backend_write_section
+	       && o->sec_info_type == SEC_INFO_TYPE_MERGE)
+	/* A MERGE section that has no relocations doesn't need the
+	   contents anymore, they have been recorded earlier.  Except
+	   if the backend has special provisions for writing sections.  */
+	contents = NULL;
       else
 	{
 	  contents = flinfo->contents;