One thing that makes discussions between programmers so lively is that we lack firm objective definitions of terms such as “strongly typed” or “low level”. A common recourse is to look for some snappy quote by a dead computer scientist. While Dijkstra tends to be my favourite arms supplier when fighting about terminology, here we find Alan Perlis ready to be of service:
“A programming language is low level when it requires attention to the irrelevant.” - Alan Perlis, 1982
Of course, pithy quotes such as these tend to raise more questions: what is irrelevant? I think all programming languages require the user to care about things that are not central to whatever problem they are trying to solve, and so by Perlis’ quote they would all be low level. Maybe so and that makes the definition less useful. (Except to motivate constant progress towards ever-more high level languages, which Perlis would likely strongly agree with.) The definition of “irrelevant” probably depends on the context, and even languages that are by most standards considered quite high level may have aspects that are of a low level nature.
Which brings us to Futhark, the most prominent among all functional array language named after the Runic alphabet. Futhark was originally conceived as a very high level language for parallel programming, although I’m not sure there was a shared understanding among the initial participants about what that meant exactly. As for me, I was intoxicated by the experience of working on an optimising compiler at all, and I was infused with naive ideas about the compiler being able to do anything! Just write the code and the compiler will figure out how to make it run fast.
I still remember the precise situation where I was disabused of that notion and became perhaps not that much wiser, but at least a good bit more cynical. During my master’s studies I was in a meeting with my advisers Fritz Henglein and Cosmin Oancea, and we ended up discussing the problem of multiplying multiple matrices. Suppose we have three matrices A, B, C. Since matrix multiplication is associative, we can multiply in any order we wish and still get the right result: (AB)C = A(BC). But one order of operations may be asymptotically much better than the other, depending on the sizes of A, B, C. In principle, a compiler could generate code that at runtime uses to the most efficient ordering based on the actual sizes observed.
Now, Futhark has neither matrices nor matrix multiplication as a
language concept. You can model them using two-dimensional arrays and
reduce compositions (and it will run quite fast too), but the
compiler has no idea about the algebraic properties of the resulting
code, so there is no way it can start doing re-association
optimisations. I recognised this at the time (I was naive, not
ignorant!), but mumbled something about optimisations such as these
perhaps being merely a special case of some more fundamental principle
that could be applied at the level of
reduce. I was
grasping at straws, and I knew it. This was the first of many times
that I encountered useful optimisations that Futhark could never do,
but that might be possible in a more high level language.
While unpleasant at first, saying “no, we cannot do that and probably never will” became quite a foundational principle for Futhark, and helped conceptualise in my own mind what the language is fundamentally about, and what the compiler will do for you.
So what are those principles? What is the nature of Futhark? At the most basic level, Futhark is a language of array values, coupled with a handful of primitive functions that transform arrays. Futhark will store these arrays in memory in a relatively straightforward way, and it will execute the code you write more or less in the way that you write it. In particular, the compiler will not change your algorithms, and except for pathological cases involving dead or redundantly duplicated computation, it will not fiddle with the asymptotic complexity of your program.
Futhark will also not mess around with data types: if you want an integer, you have to decide how many bits you need. There is no bignum type that tries to model mathematical integers (unless you write one as a library). There is also no type of rational numbers, computational reals, or other exotic types that model mathematical objects and allow you to avoid worrying about floating-point roundoff or cancellation. When you ask for an array, you get an array on the heap, even if it has a small constant size and it might really be used as a tuple. If your array is mostly zeroes, too bad - they still get stored individually in memory. If you want a fancy sparse representation, you have to write the code yourself, in terms of the dense arrays that Futhark makes available. The language provides useful abstraction mechanisms for doing so (modules, polymorphism, higher-order functions), but the compiler will ultimately not understand the algebraic properties of your abstraction, and cannot optimise based on them.
In this way, Futhark is definitely low level. Many programmers probably feel that fiddling with the number of bits in a float is well inside the realm of Perlis’s notion of irrelevant.
But it gets worse: unusually among functional languages, Futhark does not support recursion. When you want to work with a structure such as trees, you have to model them as arrays. There are pretty cool algorithms for doing so, but it definitely leads to significant “attention to the irrelevant” if you just want to build a tree. Similarly, irregular arrays - arrays where the elements have different sizes - must also be encoded using various techniques. Terribly irrelevant!
So at this point, what does Futhark even do for you? Well, if you
tell it which arrays you want, it will figure out where and how to put
them in memory. If you tell it about parallel operations such as
reduce, it will figure out how to execute them
efficiently. And perhaps most importantly, if you write a bunch of
small functions using these concepts and then put them together, the
compiler will work very hard to remove the overhead of all those
intermediate results. When you nest parallel constructs, it will work
very hard to figure out a good way to map the levels of application
parallelism to the available levels of hardware
parallelism. And the compiler will
look at how you actually traverse those arrays in memory, and ensure
they are stored in whatever order leads to the most efficient
Obviously, describing Futhark as a low-level language is a provocation and not a serious attempt at placing it in some taxonomy. But if you compare the code that a Futhark programmer writes with the code that ultimately gets executed, you will find that most of what Futhark does is boil away abstractions such as higher-order functions, and generate a huge amount of tedious boilerplate for you. The scalar computation in the original program remains more or less unmolested. Any array-typed variable can probably be found as a heap-allocated slab of memory, except for the ones that are fused away entirely. One might even say that the heart of the Futhark compiler is little more than a fancy macro expander with delusions of grandeur.
Ultimately, while I don’t expect or hope that users will have to know about everything the compiler does, they should have an understanding about what it does not do. The compiler was not made by a sorcerer; it is not magical. It is not even particularly smart. It will take care of array layouts and exploiting parallelism, but you are still responsible for writing good code. It will not touch your algorithms.
I started writing this post over a month ago, but got sidetracked on account of Putin’s invasion of Ukraine. I am agonised to see a country with a long and difficult past struggle to emerge into liberty and democracy, only to be dragged backwards into tragedy by the vain ambitions of one man and the people he has misled. I encourage all readers to put pressure on their political leaders to ensure constant and thorough support for the people of Ukraine in all ways, as they fight for the hope of a free future. I also suggest donating to the humanitarian organisations active in Ukraine, or even directly to the Ukrainian defence forces. Few conflicts are truly black-and-white, and neither is this one, but the utter blackness of Putin’s naked aggression renders any Ukrainian blemishes invisible in contrast. Слава Україні!