fix memory errors with demangled name hash

fix memory errors with demangled name hash

This fixes a regression that Jan pointed out.

The bug is that some names were allocated by dwarf2read on the objfile
obstack, but then passed to SYMBOL_SET_NAMES with copy_name=0.  This
violates the invariant that the names must have a lifetime tied to the
lifetime of the BFD.

The fix is to allocate names on the per-BFD obstack.

I looked at all callers, direct or indirect, of SYMBOL_SET_NAMES that
pass copy_name=0.  Note that only the ELF and DWARF readers do this;
other symbol readers were never updated (and perhaps cannot be,
depending on the details of the formats).  This is why the patch is
relatively small.

Built and regtested on x86-64 Fedora 20.

2014-06-26  Tom Tromey  <tromey@redhat.com>

	PR symtab/16902:
	* dwarf2read.c (fixup_go_packaging, dwarf2_compute_name)
	(dwarf2_physname, read_partial_die)
	(guess_partial_die_structure_name, fixup_partial_die)
	(guess_full_die_structure_name, anonymous_struct_prefix)
	(dwarf2_name): Use per-BFD obstack.