https://bugs.webkit.org/show_bug.cgi?id=115525

Reviewed by Geoffrey Garen.

The structure transition table basically maps string to structure. The string is
always also stored, and ref'd, in the structure in Structure::m_nameInPrevious.
m_nameInPrevious is never mutated, and never cleared. The string cannot die unless
the structure dies. If the structure dies, then that entry in the transition map
becomes a zombie anyway and we will detect this separately.

So, we don't need to use RefPtr<StringImpl>. We can just use StringImpl*.

This also fixes a goof where we were getting the StringImpl's hash rather than
using a pointer hash. Not only is the latter faster, but it prevents my change
from leading to crashes: with my change we can have zombie keys, not just zombie
values. They will exist only until the next map mutation, which will clear them.
Lookups will work fine because the lookup routine will reject zombies. But it
does mean that the HashMap will have to deal with dangling StringImpl*'s; all it
takes to make this work is to ensure that the HashMap itself never dereferences
them. Using a pointer hash rather than StringImpl::existingHash() accomplishes
this.

This also ensures that we don't accidentally call ref() or deref() from the
compilation thread, if the compilation thread inspects the transition table.

And no, we wouldn't have been able to use the HashMap<RefPtr<...>, ...>
specialization, because the transition table is actually
HashMap<pair<RefPtr<StringImpl>, unsigned>, ...>: hence that specialization
doesn't kick in. We could have written a new specialization or something, but
that seemed like a lot of work given that we don't need the table to be ref'ing
the strings anyways.

* runtime/Structure.cpp:
(JSC::StructureTransitionTable::add):
* runtime/StructureTransitionTable.h:
(StructureTransitionTable):
(Hash):
(JSC::StructureTransitionTable::Hash::hash):

git-svn-id: http://svn.webkit.org/repository/webkit/trunk@153141 268f45cc-cd09-0410-ab3c-d52691b4dbfc