Edit: (Not waiting for reply link to appear...) When I say "the memory barriers could also be skipped" I am referring to any other memory barriers past those associated with initialization, not those associated with initialization. There would be memory barriers at initialization, but that is the only need for memory barriers and I believe that is provided here by the synchronized keyword. The object presented in the article then looks like a cache-friendlier version of a single global variable with the necessary barriers limited to the initialization of the global instance pointer (which might also be done as in the Java example given in the comments on the page). The TLS does not eliminate the need for any barriers, only helps eliminate non-local reads. That might allow a performance gain depending on the cache behavior (or, for example, in the presence of NUMA), but it hasn't reduced the need for locking IMHO.
There is no way around this. Your statement that memory barriers could be skipped is exactly the source of the double checked locking bug. I know this is hard to understand, which is why it has tripped up a looong list of very smart people over the years.
David's article solves the problem by having one of the globals not be global at all, but thread local.
As for performance, David's article shows with benchmarks that his solution is a huge win.