Treat memory allocation as effect?

[cblake]

PR-wise, it is often easier to put some new, questionable feature behind an 
`experimental` or `when defined` barrier to make it easier to remove/overhaul 
if it turns out to be undesirable. Unsure if that is possible/warranted in this 
case, but seemed worth mentioning (and also mentioning that it if is **_not_** 
possible then even more caution about adding it may be warranted).