The notion of using TLS callbacks is interesting as a means of attaching a stub without disrupting the EP, but be aware that they are ignored by the loader on 9x platforms. If you only care about NT, then they work.

It's also interesting the Microsoft claims (in the PE spec) that they can be used to call global constructors, though this is a bad idea because the CRT has not been initialized yet. I suspect this is why their own compilers don't use them.