Also I hope it opens the way to stop leaking TEBs on force terminating threads and to stop leaking tls vectors after reallocation (although I didn't try to guess native structures for the latter, while it is probably possible to use our own layout for the time being). But none of that is in this MR.