A new version of the Futhark compiler has been released (source tarballs and full changelog here). We have continued work on compiler optimisations and code generation - notably a significant improvement to the kernel extraction algorithm, although it is not yet enabled by default. However, the main changes in this release are in the source language. Specifically, the work of Anders Kiel Hovgaard, a master’s student at DIKU, has resulted in Futhark now supporting higher-order functions and Hindley-Milner-style type inference. This is by far the largest extension we have ever made to the source language, and it has completely transformed how we write Futhark programs. Considering the complexity of these features, it is likely that they are still somewhat buggy, but we have been using them quite intensively for the past few weeks to ensure they are in good shape. Hopefully, most straightforward uses will not trigger bugs. While one goal of Futhark is to be similar to existing functional languages, there are some quirks and limitations that I will discuss below.
To a first approximation, Futhark’s higher-order functions work as one might expect from Haskell or an ML language. However, the focus on performance has led us to introduce some unusual restrictions:
- A function may not be returned from a branch.
- A function may not be put in an array.
- A function may not be used as a loop parameter.
This means that while Futhark supports higher-order functions, functions are not first class values. The justification for these restrictions is that they permit efficient defunctionalisation - a process that converts a higher-order program to a first-order program. Specifically, the Futhark compiler guarantees that use higher-order functions carry no run-time overhead what-so-ever. This is in contrast to conventional defunctionalisation, which may produce large amounts of control flow, and most other functional languages, which simply use function pointers. These techniques are inefficient for the kind of high-performance execution intended for Futhark, especially on restricted devices such as GPUs. Thus, as long as your program type-checks, you should not worry that your use of higher-order functions has a negative performance impact. To try it out, we have produced a Futhark port of Conal Elliott’s Functional Images, which represents images as functions from coordinates to colours.
One nice consequence of this design is that the Futhark core language (which is monomorphic and first-order) remains entirely unchanged (we also perform monomorphisation). The programmer can write advanced polymorphic and higher-order code, but it will all be compiled away by the time the optimisation passes are run.
Our implementation of type inference is quite conventional, and perhaps even a little old-fashioned. The error messages are quite crude, but the Futhark type system is simple enough that they tend to be comprehensible.
Some limitations remain that will be addressed in the next release. The main such limitation is that type inference is limited in the presence of record and field projection. For example, the following will not type check:
The compiler will complain that
x is not a record with a field
foo. However, it will work if we add a type annotation:
Or we could simply use a record pattern:
Standard ML already has a clean solution to this problem, so it will definitely be fixed eventually. For now, a few type annotations may be necessary for code that makes heavy use of records.