If you're forced to use a platform that requires manual memory management, I'm all for this approach -- either explicitly, or implicitly on a fixed-size stack.

Or, use a platform that handles dynamic allocation for you, or that lets you perform "object" based memory management instead of size-based memory management.

A similar principle applies to locking: do as little as possible. If you can use a development platform that lets you solve your parallel/concurrent problem naturally without locking (like folks have applied functional programming to do so successfully), do it!