TFP'18 Trip Report

Posted on June 13, 2018

These are my notes from my first visit to Trends in Functional Programming. Overall I had a great time - the audience was friendly, and the talks more accessible than is usually the case at the larger conferences. I suspect this is due to most of the presentations being given by fairly new researchers who are still themselves new to their subjects, rather than senior academics who are reporting the latest news in a decades-old research agenda.

Also, Gothenburg is one big hill. Ugh.

Day One

The invited talk was Refactoring Reflected by Simon Thompson. This was a general overview of the difficulties and motivation behind refactoring (renaming, function extraction, etc). The speaker went over a lot of various quirks in Haskell, Erlang, and OCaml that make refactoring troublesome in these languages. While there has been some progress in automatic refactoring for fixing bugs and such, he seemed quite sceptical about fully automated refactoring. A human must be kept in the loop. This seems sensible, since much refactoring is about making the code more readable to a human. He also believes that a refactoring tool need not be 100% automatic; as long as it can do 95% (and clearly label the remaining 5%, not just screw them up), that’s fine.

The speaker used one very striking catchphrase: “The more precise the types, the more fragile the structure”. The context was a point about how usually refactoring is done without changing the types of anything, but programs in dependently types languages are often so rigid that refactorings will have to change the types.

I was a bit disappointed that the presentation skipped what I consider the most difficult aspect of refactoring tools and similar, namely retaining original formatting in the untouched parts of the code. This is because normal parsing techniques all produce abstract syntax trees, where irrelevant details about whitespace (and comments!) are gone. The speaker suggested that this problem can be ameliorated via the use of code formatters in the style of gofmt, but I am not so sure.

It was mentioned that they debug their refactoring tools by generating random programs and then firing random refactorings at them. That sounds fun. I wonder what the resulting programs look like.

The first ordinary talk was Liquid: A concurrent calculus with declaring first-order asynchronous functions. This was an attempt to provide some syntactic sugar on the usual await/async mechanisms by making calls to asynchronous functions syntactically the same as calls to synchronous functions, then automatically also turning dependent statements into futures. For example, assuming request is an asynchronous function:

var r = request("http://...")
print(r)       // Not executed until r is available
print("hello") // Executed immediately

This seems very messy and unpredictable. I doubt this is superior to a bit of boilerplate.

The second talk was Shifting and Resetting in the Calculus of Constructions, and while it seemed solid work, I did not have the background necessary to understand it. I think it was fundamentally about supporting call/cc in dependently typed languages, but I’m not really familiar with the shift/reset mechanism.

Next up was Colocation of Potential Parallelism in a Distributed Adaptive Run-time System for Parallel Haskell. I’m generally sceptical of parallel Haskell because I find it difficult to use efficiently and I see very few real-world applications in practice (concurrent Haskell, on the other hand, is a big success). This presentation talked about a new work-stealing scheduler for distributed parallel Haskell that used a notion of “ancestry” to decide which sparks should be stolen by idle processors. Essentially, the runtime keeps track of how sparks are related by their call stack (simplifying a bit here), and processors that finish a spark will go on to try to execute the most related one. The hope is that this spark will already be close by. The main weakness of the presentation was that the benchmarks did not contain any sparks that required large quantities of data, so migration was only latency-bound, not bandwidth-bound. Also, as is unfortunately common with parallel Haskell benchmarks, only scaling was shown, and not performance compared to non-Haskell parallel implementations.

Then was Reversible Choreographies in Haskell, which is an implementation of a Haskell DSL for specifying protocol implementations, where the actions of the protocol implementations are (dynamically) checked via session types. The work was the practical implementation of the theory paper Causally consistent reversible choreographies, which I have not read, so I had a bit of trouble appreciating the finer details.

The penultimate presentation had the most frightening title in all of TFP: Intrinsic Currying for C++ Template Metaprograms. It was about the effort to support traditional functional language features in C++ template metaprogramming, this time focusing on currying. It was, as expected, horrifying, but quite well done. I still have a hard time abiding a programming methodology that uses compile errors as control flow.

The final presentation was On Optimizing Bignum Multiplication, and talked about an effort to implement a new library for arbitrary-size integers. Most of the talk was about experimentally determining the optimal size for the bignum digits (make them as big as possible) and the threshold for switching from Karatsuba multiplication to conventional grade school multiplication (do essentially what GMP does.

Day Two

This day only had two sessions for some reason.

The first talk was Active-Code Reloading in the OODIDA Platform, and essentially about supporting hot code (re)loading for on-board car analysis platforms by sending them Python code (as strings) to be evaluated on-device. While this is certainly flexible, I can’t help but feel that the potential issues were glossed over.

The second talk was Towards Optic-Based Algebraic Theories: the Case of Lenses, and went a bit over my head. It seemed to focus on constructing a generalisation of lenses that also works for impure structures. In particular, it was observed that MonadState a (State s) can be seen as a generalisation of Lens s a.

Next up was Type Safe Interpreters for Free, a pretty nifty library, called Saint, which “connects your Haskell eDSL to the cloud”. Fundamentally, Saint provides a semi-automated away to expose an untyped interface (a necessity when it goes over the web) to a typed API, where the type checking is obtained for free. Cool stuff.

The final talk was yet another one I had a hard time following: Handling Recursion in Generic Programming Using Closed Type Families. I have done only a little generic programming in the Haskell sense of the word, and only enough to decide that I prefer the boilerplate to the complex techniques you otherwise end up using.

In the evening we went on a lovely combined four-hour boat ride and dinner on the rivers surrounding Gothenburg. Major kudos to the hosts for that one.

Day Three

The invited talk on day three was one of my favourites of the entire event. It was titled Retrofitting purity with comonads, and that was exactly what it was about. In a pure language, we add support for effects with monads. So in an effectful language, how do we add support for guaranteed-pure computation? Comonads of course! Comonads are a wonderful dual to monads that are constructed by “reversing” the arrows from a category theoretical point of view. In practice, it means that for a comonad m you have functions

coreturn :: m a -> a

cojoin :: m a -> m (m a)

(Usually called extract and duplicate.) It makes logical sense, too: in an impure language, you don’t want the programmer to be able to promote the result of an impure computation to a pure value, but it is fine to inject a pure value into the impure (non-comonadic) world. The use case seems to be to help retrofit existing impure functional languages, like OCaml, with purity for optimisation and reasoning purposes. I talked a bit with the presenter, and he did not seem to think there was a good reason to use this kind of design for a fresh language (just make it pure instead).

The first normal research talk was Inductive type refinement by conjugate hylomorphisms. It seemed to be about using some category theory-ish diagram to explain types defined via something called “ornaments”, but I had a really hard time following this one. I guess that the purpose is explaining the principles behind identities such as

length (xs ++ ys) = length xs + length y

Intuitively, we can see an integer as a list that has “forgotten” some extra information (its values) and how contains only the lists. There was also a connection to refinement types that I completely failed to grasp. Somewhat frustrating that the post-proceedings are months away!

The next talk was Folds, Unfolds, and Metaheuristics: Towards Automatic Rewriting and Derivation of Metaheuristics. Heuristics are algorithms we use to approximately solve really hard problems (especially the NP-hard ones). Meta-heuristics are the building blocks for heuristic approximation algorithms. This talk was on an in-progress framework for expressing (in Idris) a dependently typed representation of these building blocks, permitting programming transformation of the algorithm itself. Unfortunately, the only example use that was hinted at was adding timeouts to various parts. I will definitely be keeping an eye on this one, to see where it goes.

Next up was a position paper (well, talk): Graph Reduction Hardware Revisited. The presenter argued that while graph reduction hardware was out-competed by commodity hardware in the 80s, the death of sequential speedups and (especially) the rise of FPGAs means we should reconsider this old idea. There seems to be two ways of going about it: either compile a functional (core) language directly to a hardware description language, or compile it to a set of supercombinators (think SKI), an interpreter for which is then manually implemented on an FPGA. I think the latter is what the presenter was most enthusiastic about. I am less enthusiastic about the idea as a whole; or rather I believe it is premature. FPGAs are not some perfectly flexible piece of reprogrammable hardware, while more flexible than traditional CPUs (or GPUs), they still need to be used just right to reach peak performance. In particular, the hardwired functional units (ALUs etc) are much more efficient than what can be built using the reprogrammable gates. I think we may have to wait for either cheap ASIC fabrication to become possible (which is decades away), or use functional hardware description languages, that can map more cleanly to the FPGA domain. There’s still no free lunch.

Next was Improving Haskell, which was not about improving Haskell, but rather about improving Haskell programs. Reasoning about performance in a lazy language is tricky, because it depends on the context in which the expression is used. Consider an expression (fast x, slow x) versus (slow x, fast x). The former is faster if they are passed to fst, but the second is faster if passed to snd. Improvement theory is a theory for reasoning about performance in all possible contexts. Apart from a basic introduction to improvement theory, this talk was a demo of Unie - the University of Nottingham Improvement Engine - which is an interactive tool for helping write proofs in improvement theory. It looked quite polished as the presenter took us through part of a proof of why xs ++ (ys ++ zx) is an improvement on (xs ++ ys) ++ zs.

The penultimate talk was High-performance defunctionalization in Futhark, given by the DIKU student Anders Kiel Hovgaard, who implemented higher-order functions for Futhark for his masters project. This was a very successful project and Anders’s work was very polished (including both a robust implementation and a formal proof of correctness of the defunctionalisation algorithm). The talk was good, too.

The final talk was Strategic Skeleton Composition with Location Aware Remote Data. The context was the distributed parallel Haskell dialect Eden, which I had some exposure to earlier in my studies via Jost Berthold. It works quite well, but is hampered by requiring the use of a fork of the GHC runtime system. Eden is an elaborate fork-join model, and this talk was about trying to avoid having to frequently send data back to the master node, by parameterising parallelisation skeletons with a policy of where to send the resulting data. The example given was a bitonic sorting network, where nodes should interchange data with each other in a very structured form that can be predicted in advance. Unfortunately there were no experimental results reported, so it is very hard to judge how well it works.