Deliberately limiting the power of your (domain specific) language can be a useful thing to do, too.
As an example, SQL (ignoring the crazy bits) is so amenable to query optimization precisely because it is not Turing complete.
This has been codified in the Rule of Least Power: http://en.wikipedia.org/wiki/Rule_of_least_power
Indeed. Bitcoin transactions all contain validation scripts that are run through a stack-based interpreter. This Forth-like language does lack loops but, even after being there for years, the current client still only accepts a small set of standard templates.
This shows up all the time in the myriad Haskell eDSLs. Monad-based interfaces are obvious but throw away information that is useful for optimization (see http://homepages.inf.ed.ac.uk/slindley/papers/unembedding.pd... for how to get some of it back). Applicative interfaces are really amenable to optimization, though (see Facebook's Haxl project: http://www.reddit.com/r/haskell/comments/1le4y5/the_haxl_pro...)
Applicatives were the example that I had in mind. See also http://gergo.erdi.hu/blog/2012-12-01-static_analysis_with_ap..., or think about how regular expressions' limited power makes parsing with them a linear affair. (Incidentally, regular language parsers are Applicatives, too.)
[1] http://en.wikipedia.org/wiki/Greenspun's_tenth_rule