Citation Context

...ators [4, 10] lets us reify a probabilistic program as a weighted search tree without the run-time overhead of stepping through deterministic code. 2. The finally-tagless embedding of object programs =-=[1]-=- using combinator functions [33, 47] allows multiple inference algorithms to operate on the same model without the run-time overhead of tagging values with their types. We apply these general techniqu...

Citation Context

...preter. We have not yet investigated in depth how to deal with this overhead, or indeed if it is a serious problem in practice, but we expect that partial evaluation [8] or a finally tagless approach =-=[10]-=- would be sufficient. Another weakness which we hope to address is that mixing control and resource effects is a challenge. For example, we cannot currently thread state through all branches of a non-...

Citation Context

...] presented only the pure front-end, single pass instantiation. The use of Rep[T] types to define binding times in LMS is inspired by earlier work on finally tagless or polymorphic language embedding =-=[8, 18]-=-. Tobin-Hochstadt et al. [50] propose languages as libraries in Racket and make pervasive use of macros for translation. Our approach of using staging is similar in spirit. Earlier work on realistic c...

Citation Context

...arrays results in compact and readable types. Using parallel , a program that computes the first 10 powers of two is defined as follows: powersOfTwo = parallel 10 (λi → 2ˆ i ) ∗Main> eval powersOfTwo =-=[1,2,4,8,16,32,64,128,256,512]-=- The purpose of parallel is to capture the fact that the array elements are independent. This means that they can be computed in any order, or even in parallel. However, our C backend does not yet gen...

Citation Context

... to a higher-order polymorphism encoding of Leibniz equality [Baars and Swierstra 2002; Cheney and Hinze 2002; Weirich 2004], to overlapping type classes [Kiselyov et al. 2004], to a tagless algebra [=-=Carette et al. 2009-=-], to functional dependencies [Guillemette and Monnier 2008a; McBride 2002]. The flexibility of the Haskell type system and the ingenuity of Haskell programmers have been demonstrated beyond doubt. Ho...

Citation Context

...sT a) Here CpsT maps a type a to its CPS-converted version [34]. Guillemette and Monnier express CpsT as a type-level function [18], whereas Carette et al. show how to do without type-level functions =-=[4]-=-. 6 Related work and reflections The goal of type families is to build on the success of static type systems, by extending their power and expressiveness without losing their brevity and comprehensibi...

Citation Context

...on problem, in which an evaluator aims to move type checking of intermediate results in the interpreter from runtime to compile-time, has been extensively studied in the partial evaluation literature =-=[10, 21, 39, 40]-=-. The presence of dependent types simplifies the implementation of a tagless interpreter greatly, in that we are able to write it in a natural style (avoiding, for example, continuation passing style)...

Citation Context

... were later shown to be related to the Visitor pattern and used to provide modular and generic visitors [36, 41]. They have also been successfully used to represent embedded DSLs in more modular ways =-=[10,25]-=-. However, most existing work considers only the creation of objects with single features: not much attention is paid to creating objects composed of multiple features. Hofer et al. [25] use delegatio...

Citation Context

...ode can be generated. In the DSL program, wrapped types such as Rep[Int] represent staged computations while expressions of plain unwrapped types (Int, Bool, etc.) are evaluated at staging time as in =-=[5,18]-=-. Consider the difference between these two programs: def prog1(b: Bool, x: Rep[Int]) = if (b) x else x+1 def prog2(b: Rep[Bool], x: Rep[Int]) = if (b) x else x+1 The only difference in these two prog...