Counting to infinity at compile time
Higher better faster longer
We saw in a previous blog post of mine an introduction to counting at compile time in scala, using the conventional
Nat is slow. Really, really slow, and there are hard limits on what we can make the scala compiler do, as it currently stands.
For example, here are some rough and ready and entirely-unscientific benchmarks I’ve just run on approximate compile times for the
Sum typeclass we defined last time. The numbers would be completely different on your machine, but probably to the right order of magnitude:
Sum | Approx Time
3+6 | 0s
8+8 | 0s
10+7 | 0s
12+12 | 2s
15+15 | <does not complete>
17+4 | 1s
20+5 | 7s
20+8 | 104s
22+11 | <does not complete>
22+22 | <does not complete>
So, as you can see, it’s less than ideal for real, practical uses. It’s true we’re asking the compiler to do a lot, and the scala compiler is famously slow, but we’d still like to have
Nat computations complete in a reasonable time. I don’t want the scala compiler to be the limiting factor, I want the JVM to be the limiting factor.
But first, why do we want that?
What is it good for
If you’re a very dedicated reader of my blog, you’ll no doubt remember that a long time ago I wrote a blog post about preventing data-too-wide exceptions in database-based code.
The overall concept is a type
FString which is parameterised by a flavour of
The idea being that an
FString[_7] contains a
String which you have proven to be of length ≤7, by going through the safe constructor
from. If these types are present in your class which represents your database table, and the lengths mimic the column widths in your database, then you can never have a data-too-wide exception again. The length evidence will stretch all the way from your API to your database, and any invalid data will be kicked out correctly at the boundaries to your program.
You can write methods on
FString which use
Nat typeclasses like the following:
So you can see why in this case it’s important to have a speedy
Sum in particular. Every time you add an
FString to another we add their lengths. This may never occur in your code, or it could be very frequent such as concatenating first and last names, or address components.
Another use would be the encoding of matrix dimensions. When you multiply two matrices together, the width of one must match the height of the other. But this information is traditionally not represented at compile time, meaning there’s a (very frequent) exception case just waiting to jump out.
But if you defined your matrices parameterised by their dimensions as:
then you can set things up so multiplying two invalid matrices is a compile-time impossibility.
In particular, in data science and machine learning very large matrices are necessary. Very, very large. Billions of columns and rows large. Clearly using traditional
Nat for such a case is a non-starter.
And so on. Any time you have bounded data or size limits, or exceptions caused by arithmetic problems,
Nat can remove entire swathes of run time problems from your code.
Why is it so slow
Recall the algorithm for
Sum. It’s recursive — it calls itself after it has taken one layer of
Succ from one parameter and put it on top of the other. The base case, where the recursion terminates, is when the side it is deconstructing reaches zero, at which point we return whatever we have calculated on the other.
The first thing that jumps out is that the algorithm itself is pretty inefficient. The complexity for calculating
An immediate ‘fix’ for this is to consider that we needn’t unwrap one
Succ layer at a time: we could do two if we want, and have the single
Succ case as a ‘back-up’ low-priority implicit for the final case.
Now when you add
11+5, the following flow would be observed:
11 + 5 =
9 + 7 =
7 + 9 =
5 + 11 =
3 + 13 =
1 + 15 =
0 + 16 =
So we’ve improved the complexity of the operation. But why stop at 2? If we can do 2 we can do 4, and 8, and 16.
By this point we’re encoding the operation of
Sum by deconstructing our
Nat into its binary form. If we’re operating on
_27 then we would deconstruct it as
16+8+2+1 in our implicit searching whilst summing.
But this idea is not satisfying, because there’s no way in this approach to encode an arbitrary binary solution. We have to encode the 8 and 16 and etc cases manually.
Also, it turns out, it’s not that efficient. Here is an implementation with the first extra layer, handling two
Succ layers at once:
Again with the entirely unscientific benchmarks, but it sped up my local calculation compile time by about 40%.
Not as fast as we would expect. We should expect a halving of the compile time based purely on the algorithm. What’s the problem?
The problem is the encoding of the number itself:
Succ. It’s too many nested types. Each time you add one you increase the nesting and the entire thing becomes more cumbersome for the compiler to handle and compare.
Imagine we strip four layers of
_11: the compiler has to examine
_11, count four layers of type nesting, unwrap it, then apply four layers of type nesting to the other number.
While we decreased the complexity of the numerical algorithm, we didn’t do much to the actual job the compiler has to do. Wouldn’t it be great if we had just one operation to do at every stage, instead of four or more like in the above example?
Succ is the problem. It’s elegant, but not fit for purpose. We know a binary form of
Sum, by some measure, more efficient. We should try and move this binary encoding into the very definition of
This becomes extraordinarily simple if you realise the the
Succ encoding is just the unary encoding of integers. It’s just the repetition of a symbol to represent the magnitude of the number, like a tally chart. Usually in unary that symbol is
1 and here it happens to be
So we just have to replace
Succ with two wrappers, one representing
0 and one representing
1 for the binary expansion of integers. I chose
We have a choice to make though — which direction our encoding runs in. For example, 4 could be
One[Zero[Zero[b0]] or it could be
Zero[Zero[One[b0]]]. We could encode the binary expansion
100 forwards or backwards. (Here
b0 is the ‘initial’ binary number representing 0, like
Nat._0 represents the initial unary number (also 0)).
If you imagine you were doing a long sum or multiplication by hand, like one of these from primary school:
You’ll remember that we operate from the right inwards, from the lowest denomination. For that reason, we should choose a reverse binary expansion for our new
Nat, so all our encodings start at the lowest denomination on the outer layer.
One[X] becomes a simple arithmetic operation, and represents the number
2 * X + 1.
Zero[X] similarly represents the number
2 * X.
Let’s call our new version
BNat to avoid any confusion.
Here’s what a basic setup might look like:
But before we go any further, we have to ask ourselves something.
Why stop at Binary?
Binary is the natural thing software developers reach for, for obvious reasons. But is it the best choice?
We saw above that we’re trying to minimise the length of the expansion of the number
N when encoded in some flavour of
Nat, some base, unary being base 1 and binary being base 2.
The ‘most efficient’ encoding in these terms is when the base used to represent our
N is infinite — give each number its own unique symbol and suddenly every type is 1 deep and there’s no recursion at all. Solved it!
Obviously that’s absurd. We’re also limiting the number of symbols themselves. That’s what
Succ is optimised for: there is only one symbol and the typeclasses as a result are very simple and easy to reason about.
We’d like to be more balanced. We’re trying to minimise the number of symbols used — our base — while also minimising the length of the number
N in our base. We’re trying to balance having a high base so things compile faster, and keeping the complexity of our typeclasses as low as possible.
So we’re trying to minimise
L = Length of expansion of N in base B and
B = Chosen base, with respect to
L = log_b(N) so we have to find the minimum of
B * log_b(N). This is the same as finding the minimum of
B / log(B), where
log is the natural logarithm.
Do the maths yourself, plug it into wolfram alpha, whatever you like. Turns out the function is minimised when
B = e, the natural logarithm base, roughly
2.71828... Which is sadly no good to us, since we definitely can’t operate with a non-integer base in the scala compiler.
B=3 gives a lower value in the above function than
B=2. So we’re abandoning binary, and ploughing full steam ahead on to the provably more efficient base ternary.
Same rules apply as binary. We’re going to produce a reverse ternary encoding of the natural numbers using three symbols instead of two this time:
The arithmetic operations these represent are as follows:
Zero[X] = 3X
One[X] = 3X + 1
Two[X] = 3X + 2
Here is the basic setup:
We’re going to be defining
TNat, but we need a couple of utilities first. They serve as good examples to see what ternary algorithms look like:
An easy way of increasing a
1 would be nice. In unary’s case this operation was just to wrap a
Succ round it.
Incr implementation for
Four cases, telling us how to increment each possible form of a
0 + 1 = 1
(3x) + 1 = (3x + 1)which is
(3x + 1) + 1 = (3x + 2)which is
(3x + 2) + 1 = 3(x + 1)which is
Hopefully by now such typelevel functions are becoming easy to deconstruct.
Now we’re working with a more complex structure, there is one thing it makes no sense to express:
Zero[t0]. That’s a different way of just saying
t0 — since
Zero represents multiplying by 3. We can’t allow different representations of the same number as it would make any automatic calculation very difficult indeed. So we need a safe way of wrapping a
Zero, of trebling it:
and now we can safely apply
Zero to any number we like.
Sum is… not nice to implement in ternary. That’s one of the reasons we try to minimise the number of symbols used in the above efficiency calculation. The
Sum typeclass gets really big really fast. There are 11 different cases, compared to
Succ‘s 2 cases.
I prevaricated over whether to include it here or not and in the end decided not to, in case it scared you off. This article is already too long. But it is available for you linked at the end of the article. It’s mostly grunt work anyway, there’s nothing ground-breaking in there. If you understand the approach above in
Treble then figuring out
Sum is just elbow grease.
But, it does exist and the definition looks like this:
Let’s take it for a spin.
Let’s dump some random numbers in to our new addition typeclass and see what falls out.
We will be using the following method to test:
Succ encoding could not compile
22+22=44 for us. Well:
So far so good, we’ve broken our previous record! Let’s kick things up to 11:
Those are billions. Compiling in negligible time.
Quintillions? This is in the realm of long overflow:
We’ve finally hit something that takes a noticeable amount of time to compile (it also breaks the github syntax highlighter). Two random numbers in the order of 10¹⁵⁰.
Succ was hitting this time calculating
20+5. Let’s see how far we can push it: 10³⁰⁰?
So we’ve pretty much hit the limit here — 40s is not suitable for use in a project, and any longer and the compiler will start crashing.
10³⁰⁰ though is a 10²⁹⁹ times improvement over
Succ, so it’s a pretty good improvement¹. And it is definitely large enough to encode your string lengths or matrix dimensions, no matter how much data you’re processing.
TNat (including the promised
Sum typeclass) is available in my experimental library Numerology. Also included is
BNat (the binary version we skipped over), many more typeclasses than were demonstrated here and various other experiments to do with compile-time arithmetic.
It’s not yet in stable release (experiments are still on-going), but the code as it is is more than ready to be used in production (and is used, in TypeChecked’s private repos) and the functionality currently available will not be regressed at any point. (UPDATE: Numerology is now in stable release and is available on Maven)
The above code may break your compiler. In the Numerology repo there is a
.jvmopts file including the following line:
-Xss50m. This extravagant value provides plenty of stack space to compile every example in this post and in the repo.
 Fun fact: If we expressed a number in the order of 10³⁰⁰ in terms of unary
Succ, and we did something wrong and the compiler gave us an error message about that number, the error message would be (approximately) 10³³ gigabytes in size. That is many orders of magnitude more than all the information on the internet.