Interpreters typically use polymorphic locations to hold values for the interpreted language. Technically almost any type of polymorphism is enough, but the fewer indirections you can get away with, the faster you can make a simple interpreter. So that would usually mean a sum type in a functional programming language with algebraic data types, or a simple object in an OO language, or a tagged union in C and Pascal.
But for better performance, approaches like tagged pointers (often used in Lisp implementations) are useful, as it eliminates a level of indirection for most integer operations. If you can afford to use 64 bits for your values, you might consider using doubles everywhere, with invalid exponents for stuffing a shortened pointer inside the float (I believe luajit uses, or used to use, this technique; it would probably be a good match for JS as well, as JS doesn't have integers).
But for better performance, approaches like tagged pointers (often used in Lisp implementations) are useful, as it eliminates a level of indirection for most integer operations. If you can afford to use 64 bits for your values, you might consider using doubles everywhere, with invalid exponents for stuffing a shortened pointer inside the float (I believe luajit uses, or used to use, this technique; it would probably be a good match for JS as well, as JS doesn't have integers).