As with basically everything, there are tradeoffs involved. Sometimes restrictions can be helpful for keeping things understandable, which can in turn make optimizations easier to implement. As a rather hamfisted example: completely unrestricted goto. Very general, debatably powerful, but relatively easy to use in a way that makes comprehension difficult. That same generality can also make it difficult to verify that optimizations don't change observable program semantics compared to something more restricted.

Price paid: a) you need a GC (ok, whatever, sure, you have one), b) well, the performance hit of having to GC activation records (you don't care; others do), c) whoops, thread-safety got much harder and maybe you can't even have threads (imagine two threads executing in the same activation record at the same time!).

That's a very very steep price.