Note: this is a public draft. It contains a lot of errors and may be too meme-ish and handholding in some parts. I know it needs improvements. I'll review and finish in the future. Corrections and feedbacks are welcome!
- TL;DR
- Core Language Overview
- What makes it fast
- Rewrite Rules
- Low-level Implementation
- Bonus: Copatterns
- Bonus: Abusing Beta-Optimality
- Bonus: Abusing Parallelism
Since this became a long, in-depth overview, here is the TL;DR for the lazy:
HVM doesn't need a global, stop-the-world garbage collector because every "object" only exists in one place, exactly like in Rust; i.e., HVM is linear. The catch is that, when an object needs to be referenced in multiple places, instead of a complex borrow system, HVM has an elegant, pervasive lazy clone primitive that works very similarly to Haskell's evaluation model. This makes cloning essentially free, because the copy of any object isn't made in a single, expensive pass, but in a layer-by-layer, on-demand fashion. And the nicest part is that this clone primitive works for not only data, but also for lambdas, which explains why HVM has better asymptotics than GHC: it is capable of sharing computations inside lambdas, which GHC can't. That was only possible due to a key insight that comes from Lamping's Abstract Algorithm for optimal evaluation of λ-calculus terms. Finally, the fact that objects only exist in one place greatly simplifies parallelism.
This was all known and possible since years ago (see other implementations of optimal reduction), but all implementations of this algorithm, until now, represented terms as graphs. This demanded a lot of pointer indirection, making it slow in practice. A new memory format, based on the Interaction Calculus, takes advantage of the fact that inputs are known to be λ-terms, allowing for a 50% lower memory usage, and letting us avoid several impossible cases. This made the runtime 50x (!) faster, which finally allowed it to compete with GHC and similar. And this is just a prototype I wrote in about a month. I don't even consider myself proficient in C, so I have expectations for the long-term potential of HVM.
HVM's optimality and complexity reasoning comes from the vast literature on the optimal evaluation of functional programming languages. This book, by Andrea Asperti and Stefano Guerrini, has a great overview. HVM is merely a practical, efficient implementation of the bookkeeping-free reduction machine depicted in the book (pages 14-39). Its higher-order machinery has a 1-to-1 relationship to the theoretical model, and the same complexity bounds, and respective proofs (chapter 10) apply. HVM has additional features (machine integers, datatypes) that do not affect complexity.
That's about it. Now, onto the long, in-depth explanation.
HVM is, in essence, just a virtual machine that evaluates terms in its core language. So, before we dig deeper, let's review that language. HVM's Core is a very simple language that resembles untyped Haskell. It features lambdas (eliminated by applications), constructors (eliminated by user-defined rewrite rules) and machine integers (eliminated by operators).
term ::=
| λvar. term # a lambda
| (term term) # an application
| (ctr term term ...) # a constructor
| num # a machine int
| (op2 term term) # an operator
| let var = term; term # a local definition
rule ::=
| term = term
file ::= list<rule>
A constructor begins with a name and is followed by up to 16 fields. Constructor names must start uppercased. For example, below is a pair with 2 numbers:
(Pair 42 7)HVM files consist of a series of rules, each with a left-hand term and a
right-hand term. These rules enact a computational behavior where every
occurrence of the left-hand term is replaced by its corresponding right-hand
term. For example, below is a rule that gets the first element of a pair:
(Fst (Pair x y)) = xOnce that rule is enacted, the (Fst (Pair 42 7)) term will be reduced to
42. Note that (Fst ...) is itself just a constructor, even though it is
used like a function. This has important consequences which will be elaborated
later on. From this, the remaining syntax should be pretty easy to guess. As an
example, we define and use the Map function on List as follows:
(Map f Nil) = Nil
(Map f (Cons x xs)) = (Cons (f x) (Map f xs))
(Main) =
let list = (Cons 1 (Cons 2 Nil))
let add1 = λx (+ x 1)
(Map add1 list)By running this file (with hvm r main), HVM outputs (Cons 2 (Cons 3 Nil)),
having incremented each number in list by 1. Notes:
-
Application is distinguished from constructors by case (
(f x)vs(F x)). -
The parentheses can be omitted from unary constructors (
Nil==(Nil)) -
You can abbreviate applications (
(((f a) b) c ...)==(f a b c ...)). -
You may write
@instead ofλ. -
Check this issue about how constructors, applications and currying work.
What makes HVM special, though, is how it evaluates its programs. HVM has one simple trick that hackers don't want you to know. This trick is responsible for HVM's major features: beta-optimality, parallelism, and no garbage collection. But before we get too technical, we must first talk about clones, and how their work obsession ruins everything, for everyone. This section should provide more context and a better intuition about why things are the way they are.
By clones, I mean when a value is copied, shared, replicated, or whatever else you call it. For example, consider the JavaScript program below:
function foo(x, y) {
return x + x;
}To compute foo(2, 3), the number 2 must be cloned before adding it
to itself. This seemingly innocent operation has made a lot of people very
confused and has been widely regarded as the hardest problem of the 21st
century.
The main issue with clones is how they interact with the order of
evaluation. For example, consider the expression foo(2 * 2, 3 * 3). In a
strict language, it is evaluated as such:
foo(2 * 2, 3 * 3) // the input
foo(4 , 9 ) // arguments are reduced
4 + 4 // foo is applied
8 // the outputBut this computation has a silly issue: the 3 * 3 value is not necessary to
produce the output, so the 3 * 3 multiplication was wasted work. This led to
the idea of lazy evaluation. An initial implementation of this idea would
operate as follows:
foo(2 * 2, 3 * 3) // the input
(2 * 2) + (2 * 2) // foo is applied
4 + 4 // arguments are reduced
8 // the outputNotice how the 3 * 3 expression was never computed, thereby saving work?
Instead, the other 2 * 2 expression has been computed twice, yet again
resulting in wasted work! That's because x was used two times in the body
of foo, which caused the 2 * 2 expression to be cloned and, thus,
computed twice. In other words, clones ruined the virtue of laziness.
Imagine a language without clones. Such a language would be computationally perfect. Lazy evaluators wouldn't waste work, since expressions can't be cloned. Garbage collection would be cheap, as every object would only have one reference. Parallelism would be trivial, because there would be no simultaneous accesses to the same object. Sadly, such a language wouldn't be practical. Imagine never being able to copy anything! Therefore, real languages must find a way to let their users replicate values, without impacting the features they desire, all while avoiding these expensive clones. In previous languages, the solution has almost always been the use of references of some sort.
For example, Haskell is lazy. To avoid "cloning computations", it implements
thunks, which are nothing but memoized references to shared expressions,
allowing the (2 * 2) example above to be cached. This solution, though, breaks
down when there are lambdas. Similarly, Rust is GC-free, so every object has
only one "owner". To avoid too much cloning, it implements a complex borrowed
references system, allowing the same object to be accessed from multiple places,
when the compiler can prove it is safe. Finally, parallel languages require
mutexes and atomics to synchronize accesses to shared references. In other
words, references saved the world by letting us avoid these clones, and that's
great... right?
clone wasn't the impostor
References. They ruin everything. They're the reason Rust is so hard to use. They're the reason parallel programming is so complex. They're the reason Haskell isn't optimal, since thunks can't share computations that have free variables (i.e., any expression inside lambdas). They're why a 1-month-old prototype beats GHC in the same class of programs it should thrive in. It isn't GHC's fault. References are the real culprits.
Clones aren't bad. They just need to relax.
Once you understand the context above, grasping how HVM can be optimally lazy, non-GC'd and inherently parallel is easy. At its base, it has the same "linear" core that both Haskell and Rust share in common (which, as we've just established, already exhibit these properties). The difference is that, instead of adding some kind of clever reference system to circumvent the cost of cloning... HVM introduces a pervasive, lazy clone primitive.
HVM's runtime has no references. Instead, it features a .clone() primitive
that has zero cost, until the cloned value needs to be read. Once it does,
instead of being copied whole, it's done layer by layer, on-demand.
For the purpose of lazy evaluation, HVM's lazy clones works like Haskell's
thunks, except they do not break down on lambdas. In the context of garbage
collection, since the data is actually copied, there are no shared references,
so memory can be freed when values go out of scope. For the same reason,
parallelism becomes trivial, and the runtime's reduce() procedure is almost
entirely thread safe, requiring minimal synchronization.
In other words, think of HVM as Rust, except replacing the borrow system by a
very cheap .clone() operation that can be used and abused with no mercy. This
is the secret sauce! Easy, right? Well, no. There is still a big problem to be
solved: how do we incrementally clone a lambda? There is a beautiful answer
to this problem that made this all possible. Let's get technical!
HVM is, in essence, a graph-rewrite system, which means that all it does is repeatedly rewrite terms in memory until there is no more work left to do. These rewrites come in two forms: user-defined and primitive rules.
User-defined rules are generated from equations in a file. For example, the following equation:
(Foo (Tic a) (Tac b)) = (+ a b)Generates the following rewrite rule:
(Foo (Tic a) (Tac b))
--------------------- Foo-Rule-0
(+ a b)It should be read as "the expression above reduces to the expression below".
So, for example, Foo-rule-0 dictates that (Foo (Tic 42) (Tac 7)) reduces to
(+ 42 7). As for the primitive rules, they deal with lambdas, native numbers
and the duplication primitive. Let's start with numeric operations.
(<op> x y)
------------------- Op2-U32
x + y if <op> is +
x - y if <op> is -
x * y if <op> is *
x / y if <op> is /
x % y if <op> is %
x & y if <op> is &
x | y if <op> is |
x ^ y if <op> is ^
x >> y if <op> is >>
x << y if <op> is <<
x < y if <op> is <
x <= y if <op> is <=
x == y if <op> is ==
x >= y if <op> is >=
x > y if <op> is >
x != y if <op> is !=
This should be read as: "the addition of x and y reduces to x + y".
This just says that we can perform numeric operations on HVM. For example,
(+ 2 3) is reduced to 5, (* 5 10) is reduced to 50, and so on. HVM
numbers are 32-bit unsigned integers, but more numeric types will be added in
the future.
dup x y = N
-----------
x <- N
y <- NThis should be read as: "the duplication of the number N as x and y
reduces to the substitution of x by a copy of N, and of y by another copy
of N". Before explaining what is going on here, let me also present the
constructor duplication rule below.
dup x y = (Foo a b ...)
----------------------- Dup-Ctr
dup a0 a1 = a
dup b0 b1 = b
...
x <- (Foo a0 b0 ...)
y <- (Foo a1 b1 ...)This should be read as: "the duplication of the constructor (Foo a b ...) as
x and y reduces to the duplication of a as a0 and a1, the duplication
of b as b0 and b1, and the substitution of x by (Foo a0 b0 ...) and
the substitution of y by (Foo a1 b1 ...)".
There is a lot of new information here, so, before moving on, let's dissect it all one by one.
1. What the hell is dup? That is an internal duplication node. You
can't write it directly on the user-facing language; instead, it is inserted by
the pre-processor whenever you use a variable more than once. For example, at
compile time, the equation below:
(Foo a) = (+ a a)Is actually replaced by:
(Foo a) =
dup x y = a
(+ x y)Because of that transformation, every runtime variable only occurs once.
The effect of dup is that of cloning an expression, and moving it to two
locations. For example, the program below:
dup x y = 42
(Pair x y)Is reduced to:
(Pair 42 42)2. By "substitution", we mean "replacing a variable by a value". For
example, the substitution of x by 7 in [1, x, 8] would be [1, 7, 8].
Since every variable only occurs once in the runtime, substitution is a fast,
constant time operation that performs either 1 or 2 array writes.
3. dups aren't stored inside the expressions. Instead, they "float" on
the global scope. That's why they're always written on top.
4. Remember that dup (like all other rules) is only triggered when it is
needed, due to lazy evaluation. That's what makes it ultra-cheap. In a way, it
is as if HVM has added a .clone() to every variable used more than once. And
that's fine.
5. Even though the user-facing language makes no distinction between constructors and functions, the runtime does, for optimality purposes. Specifically, a duplication is only applied for constructors that are not used as functions. This equal treatment means we can write Copatterns easily in HVM (more on this in the bonus section).
Now that you know all that, let's watch dup in action, by visualizing how the
[1 + 1, 2 + 2, 3 + 3] list is cloned. Lines separate reduction steps.
dup x y = (Cons (+ 1 1) (Cons (+ 2 2) (Cons (+ 3 3) Nil)))
(Pair x y)
------------------------------------------- Dup-Ctr
dup x y = (Cons (+ 2 2) (Cons (+ 3 3) Nil))
dup a b = (+ 1 1)
(Pair
(Cons a x)
(Cons b y)
)
------------------------------------------- Op2-U32
dup x y = (Cons (+ 2 2) (Cons (+ 3 3) Nil))
dup a b = 2
(Pair
(Cons a x)
(Cons b y)
)
------------------------------------------- Dup-U32
dup x y = (Cons (+ 2 2) (Cons (+ 3 3) Nil))
(Pair
(Cons 2 x)
(Cons 2 y)
)
------------------------------------------- Dup-Ctr
dup x y = (Cons (+ 3 3) Nil)
dup a b = (+ 2 2)
(Pair
(Cons 2 (Cons a x))
(Cons 2 (Cons b y))
)
------------------------------------------- Op2-U32
dup x y = (Cons (+ 3 3) Nil)
dup a b = 4
(Pair
(Cons 2 (Cons a x))
(Cons 2 (Cons b y))
)
------------------------------------------- Dup-U32
dup x y = (Cons (+ 3 3) Nil)
(Pair
(Cons 2 (Cons 4 x))
(Cons 2 (Cons 4 y))
)
------------------------------------------- Dup-Ctr
dup x y = Nil
dup a b = (+ 3 3)
(Pair
(Cons 2 (Cons 4 (Cons a x)))
(Cons 2 (Cons 4 (Cons b y)))
)
------------------------------------------- Op2-U32
dup x y = Nil
dup a b = 6
(Pair
(Cons 2 (Cons 4 (Cons a x)))
(Cons 2 (Cons 4 (Cons b y)))
)
------------------------------------------- Dup-U32
dup x y = Nil
(Pair
(Cons 2 (Cons 4 (Cons 6 x)))
(Cons 2 (Cons 4 (Cons 6 y)))
)
------------------------------------------- Dup-Ctr
(Pair
(Cons 2 (Cons 4 (Cons 6 Nil)))
(Cons 2 (Cons 4 (Cons 6 Nil)))
)In the end, we made two copies of the list. Note how the (+ 1 1) expression,
was NOT "cloned". It only happened once, even though we evaluated the program
lazily. And, of course, since the cloning itself is lazy, if we only needed
parts of the list, we wouldn't need to make two full copies. For example,
consider the following program instead:
dup x y = (Cons (+ 1 1) (Cons (+ 2 2) (Cons (+ 3 3) Nil)))
(Pair (Head x) (Head (Tail y)))
------------------------------------------- Dup-Ctr
dup x y = (Cons (+ 2 2) (Cons (+ 3 3) Nil))
dup a b = (+ 1 1)
(Pair (Head (Cons a x)) (Head (Tail (Cons b y))))
------------------------------------------- Head
dup x y = (Cons (+ 2 2) (Cons (+ 3 3) Nil))
dup a b = (+ 1 1)
(Pair a (Head (Tail (Cons b y))))
------------------------------------------- Op2-U32
dup x y = (Cons (+ 2 2) (Cons (+ 3 3) Nil))
dup a b = 2
(Pair a (Head (Tail (Cons b y))))
------------------------------------------- Dup-U32
dup x y = (Cons (+ 2 2) (Cons (+ 3 3) Nil))
(Pair 2 (Head (Tail (Cons 2 y))))
------------------------------------------- Tail
dup x y = (Cons (+ 2 2) (Cons (+ 3 3) Nil))
(Pair 2 (Head y))
------------------------------------------- Dup-Ctr
dup x y = (Cons (+ 3 3) Nil)
dup a b = (+ 2 2)
(Pair 2 (Head (Cons b y)))
------------------------------------------- Head
dup x y = (Cons (+ 3 3) Nil)
dup a b = (+ 2 2)
(Pair 2 b)
------------------------------------------- Op2-U32
dup x y = (Cons (+ 3 3) Nil)
dup a b = 4
(Pair 2 b)
------------------------------------------- Dup-U32
dup x y = (Cons (+ 3 3) Nil)
(Pair 2 4)
------------------------------------------- Collect
(Pair 2 4)Notice how only the minimal amount of copying was performed. The first part of
the list ((Cons (+ 1 1) ...)) was copied twice, the second part
((Cons (+ 2 2) ...)) was copied once, and the rest ((Cons (+ 3 3) Nil)) was
simply collected as garbage. Collection is orchestrated by variables
that go out of scope. For example, in the last lines, x and y both aren't
referenced anywhere. That triggers the collection of the remaining list.
That was a lot of info. Hopefully, by now, you have an intuition about how the lazy duplication primitive works. Moving on.
(λx(body) arg)
-------------- App-Lam
x <- arg
bodyThis is the famous beta-reduction rule. This must be read as: "the application
of the lambda λx(body) to the argument arg reduces to body, and the
substitution of x by arg". For example, (λx(Single x) 42) is reduced to
(Single 42). Remember that variables only occur once. Because of that,
beta-reduction is a very fast operation. A modern CPU can perform more than 200
million beta-reductions per second, in a single core. Here's an example of it
in action:
(λxλy(Pair x y) 2 3)
-------------------- App-Lam
(λy(Pair 2 y) 3)
-------------------- App-Lam
(Pair 2 3)Simple, right? This rule is beautiful, but the next one is special, as it is responsible for making all of HVM possible.
Incrementally cloning datatypes is a neat idea. But there is nothing special to it. In fact, that is exactly how Haskell's thunks behave! But, now, take a moment and ask yourself: how the hell do we incrementally clone a lambda?
dup a b = λx(body)
------------------ Dup-Lam
a <- λx0(b0)
b <- λx1(b1)
x <- {x0 x1}
dup b0 b1 = body
dup a b = {r s}
--------------- Dup-Sup
a <- r
b <- sHere is how. This may be a bit overwhelming. A good place to start is by writing
this in plain English. It reads as: "the duplication of a lambda λx(body) as
a and b reduces in the duplication of its body as b0 and b1, and the
substitution of a by λx0(b0), b by λx1(b1) and x by the superposition
{x0 x1}".
What this is saying is that, in order to duplicate a lambda, we must duplicate its body; then we must create two lambdas. Then, weird things happen with its variable. And then there is a brand new construct, the superposition, that I haven't explained yet. But, this is fine. Let's try to do it with an example:
dup a b = λx λy (Pair x y)
(Pair a b)This program just makes two copies of the λx λy (Pair x y) lambda. But, to
get there, we are not allowed to copy the entire lambda whole. Instead, we
must go through a series of incremental lazy steps. Let's try it, and copy the
outermost lambda (λx):
dup a b = λy (Pair x y)
(Pair λx(a) λx(b))Can you spot the issue? As soon as the lambda is copied, it is moved to another
location of the program, which means it gets detached from its own body.
Because of that, the variable x gets unbound on the first line, and the body
of each copied λx has no reference to x. That makes no sense at all! How do
we solve this?
First, we must let go of material goods and accept a cruel reality of HVM:
lambdas don't have scopes. That is, a variable bound by a lambda can occur
outside of its body. So, for example, (Pair x (λx(8) 7)) would reduce to
(Pair 7 8). Please, take a moment to make sense out of this... even if looks
like it doesn't.
Once you accept that in your heart, you'll find that the program above will
make a little more sense, because we can say that the λx binder on the second
line is "connected" to the x variable on the first line, even if it's
outside. But there is still a problem: there are two lambdas bound to
the same variable. If the left lambda gets applied to an argument, it should
NOT affect the second one. But with the way it is written, that's what would
happen. To work around this issue, we need a new construct: the
superposition. Written as {r s}, a superposition stands for an
expression that is part of two partially copied lambdas. So, for example,
(Pair {1 2} 3) can represent either (Pair 1 3) or (Pair 2 3), depending
on the context.
This gives us the tools we need to incrementally copy these lambdas. Here is how that would work:
dup a b = λx(λy(Pair x y))
(Pair a b)
------------------------------------------------ Dup-Lam
dup a b = λy(Pair {x0 x1} y)
(Pair λx0(a) λx1(b))
------------------------------------------------ Dup-Lam
dup a b = (Pair {x0 x1} {y0 y1})
(Pair λx0(λy0(a)) λx1(λy1(b)))
------------------------------------------------ Dup-Ctr
dup a b = {x0 x1}
dup c d = {y0 y1}
(Pair λx0(λy0(Pair a c)) λx1(λy1(Pair b d)))
------------------------------------------------ Dup-Sup
dup c d = {y0 y1}
(Pair λx0(λy0(Pair x0 c)) λx1(λy1(Pair x1 d)))
------------------------------------------------ Dup-Sup
(Pair λx0(λy0(Pair x0 y0)) λx1(λy1(Pair x1 y1)))Wow, did it actually work? Yes, it did. Notice that, despite the fact that "weird things" happened during the intermediate steps (specifically, variables got out of their own lambda bodies, and parts of the program got temporarily superposed), in the end, it all worked out, and the result was proper copies of the original lambdas. This allows us to share computations inside lambdas, something that GHC isn't capable of. For example, consider the following reduction:
dup f g = ((λx λy (Pair (+ x x) y)) 2)
(Pair (f 10) (g 20))
-------------------------------------- App-Lam
dup f g = λy (Pair (+ 2 2) y)
(Pair (f 10) (g 20))
-------------------------------------- Dup-Lam
dup f g = (Pair (+ 2 2) {y0 y1})
(Pair (λy0(f) 10) (λy1(g) 20))
-------------------------------------- App-Lam
dup f g = (Pair (+ 2 2) {10 y1})
(Pair f (λy1(g) 20))
-------------------------------------- App-Lam
dup f g = (Pair (+ 2 2) {10 20})
(Pair f g)
-------------------------------------- Dup-Ctr
dup a b = (+ 2 2)
dup c d = {10 20}
(Pair (Pair a c) (Pair b d))
-------------------------------------- Op2-U32
dup a b = 4
dup c d = {10 20}
(Pair (Pair a c) (Pair b d))
-------------------------------------- Dup-U32
dup c d = {10 20}
(Pair (Pair 4 c) (Pair 4 d))
-------------------------------------- Dup-Sup
(Pair (Pair 4 10) (Pair 4 20))Notice that the (+ 2 2) addition only happened once, even though it was
nested inside two copied lambda binders! In Haskell, this situation would lead
to the un-sharing of the lambdas, and (+ 2 2) would happen twice. Notice also
how, in some steps, lambdas were applied to arguments that appeared outside of
their bodies. This is all fine, and, in the end, the result is correct.
Uff. That was hard, wasn't it? The good news is the worst part is done. From now on, nothing too surprising will happen.
Since duplications are pervasive, what may happen is that a superposition will end up in the function position of an application. For example, the situation below can happen at runtime:
({λx(x) λy(y)} 10)This represents two superposed lambdas, applied to an argument 10. If we
leave this expression as is, certain programs would get stuck, and we wouldn't
be able to evaluate them. We need a way out. Because of that, there is a
superposed application rule that deals with that situation:
({a b} c)
--------------- App-Sup
dup x0 x1 = c
{(a x0) (b x1)}In English, this rule says that: "the application of a superposition {a b} to
c is the superposition of the application of a and b to copies of c".
Makes sense, doesn't it? That rule also applies to user-defined functions. The
logic is the same, only adapted depending on the arity. I won't show it here.
(+ {a0 a1} b)
--------------------- Op2-Sup-A
dup b0 b1 = b
{(+ a0 b0) (+ a1 b1)}
(+ a {b0 b1})
--------------------- Op2-Sup-B
dup a0 a1 = a
{(+ a0 b0) (+ a1 b1)}This, too, follows the same logic of superposed application, except operators are strict on both arguments.
There is one last rule that is worth discussing.
dup x y = {a b}
--------------- Dup-Sup (different)
x <- {xA xB}
y <- {yA yB}
dup xA yA = a
dup xB yB = bThis rule handles the duplication of a superposition. In English, it says that:
"the duplication of a superposition {a b} as x and y reduces to the
duplication of a as xA and yA, b as xB and yB, and the substitution
of x by the superposition {xA xB}, and the substitution of y by {yA yB}". At that point, the formal notation is probably doing a better job than
English at conveying this information.
If you've paid close attention, though, you may have noticed the Dup-Sup has already been defined, on the Lambda Application section. So, what is going on here? Well, it turns out that Dup-Sup is a special case that has two different reduction rules. If this Dup-Sup represents the end of a duplication process, it must go with the former rule. However, if you're duplicating a term, which itself duplicates something, then this rule must be used. Due to the extremely local nature of HVM reductions though, determining when each rule should be used in general would require an expensive book-keeping machinery. To avoid that extra cost, HVM instead placed a limitation that allowed for a much faster decision procedure. That limitation is:
If a lambda that clones its argument is itself cloned, then its clones aren't allowed to clone each other.
For example, this term is not allowed:
let g = λf(λx(f (f x)))
(g g)That's because g is a function that clones its argument (since f is used
twice). It is then cloned, so each g on the second line is a clone. Then the
first clone attempts to clone the second clone. That is considered undefined
behavior, and a typed language that compiles to HVM must check that this kind of
situation won't happen.
How common is this? Well, unless you like multiplying Church-encoded natural numbers in a loop, you've probably never seen a program that reaches this limitation in your entire career. Even if you're a fan of λ encodings, you're fine. For example, the program above can be fixed by just avoiding one clone:
let g = λf(λx(f (f x)))
let h = λf(λx(f (f x)))
(g h)And all the other "hardcore" functional programming tools are compatible. Y-Combinators, Church encodings, nested maps of maps, all work just fine. If you think you'll reach this limitation in practice, you're probably misunderstanding how esoteric a program must be for it to happen. It is a common (and annoying) misconception that this limit has much relevance in practice. C programmers survived without closures, for decades. Rust programmers live well with far more restrictive limitations on what shapes of programs they're allowed to write. HVM has all sorts of extremely high-level closures you can use. You just can't have a clone clone its own clone. Without this limitation, which is almost irrelevant in practice, it wouldn't be possible for HVM to achieve its current performance, so we believe it is justified.
To learn more about HVM's low level implementation and memory layout, check the comments on runtime/base/memory.rs.
Since functions and constructors are treated the same, this means there is nothing preventing us from writing copatterns, by just swapping the roles of eliminators and introducers. As an example, consider the program below:
// List Map function
(Map f Nil) = Nil
(Map f (Cons x xs)) = (Cons (f x) (Map f xs))
// List projectors
(Head (Cons x xs)) = x
(Tail (Cons x xs)) = xs
// The infinite list: 0, 1, 2, 3 ...
Nats = (Cons 0 (Map λx(+ x 1) Nats))
// Just a test (returns 2)
Main = (Head (Tail (Tail Nats)))This is the usual recursive Map applied to an infinite List. Here, Map is
used in the function position, and the List constructors (Nil and Cons) are
matched. The same program can be written in a corecursive fashion, by inverting
everything: the List destructors (Head/Tail) are used in the function
position, and the function Map is matched:
// CoList Map function
(Head (Map f xs)) = (f (Head xs))
(Tail (Map f xs)) = (Map f (Tail xs))
// The infinite colist: 0, 1, 2, 3 ...
(Head Nats) = 0
(Tail Nats) = (Map λx(+ x 1) Nats)
// Just a test (returns 2)
(Main n) = (Head (Tail (Tail Nats)))By abusing beta-optimality, we're able to turn some exponential-time algorithms
into linear-time ones. That is why we're able to implement Add on BitStrings
as repeated applications of increment:
// Addition is just "increment N times"
(Add xs ys) = (App xs λx(Inc x) ys)This small, elegant and mathematical one-liner is as efficient as the
manually-crafted add-with-carry operation, which is an 8-cases, low-level and
error-prone definition. In order for this to be possible, we must apply some
techniques to make sure the self-composition (λx (f (f x))) of the function
remains as small as possible. First, we must use λ-encoded algorithms. If we
don't, then the normal form will not be small. For example:
(Not True) = False
(Not False) = TrueThis is easy to read, but then λx (Not (Not x)) will not have a small normal
form. If we use λ encodings, we can write Not as:
True = λt λf t
False = λt λf f
Not = λb (b False True)This correctly negates a λ-encoded boolean. But λx (Not (Not x)) still has a
large normal form: λx (x λtλf(f) λtλf(t) λtλf(f) λtλf(t)). Now, if we inline
the definition of Not, we get:
True = λt λf t
False = λt λf f
Not = λb (b λtλf(f) λtλf(t))Notice how both branches start with the same lambdas? We can lift them up and share them:
True = λt λf t
False = λt λf f
Not = λb λt λf (b f t)This will make the normal form of λx (Not (Not x)) small: i.e., it becomes
λx λt λf (x t f). This makes Not^(2^N) linear time in N!
The same technique also applies for Inc. We start with the usual definition:
(Inc E) = E
(Inc (O x)) = (I x)
(Inc (I x)) = (O (Inc x))Then we make it λ-encoded:
(Inc x) =
let case_e = λe λo λi e
let case_o = λx λe λo λi (i x)
let case_i = λx λe λo λi (o (Inc x))
(x case_e case_o case_i)Then we lift the shared lambdas up:
(Inc x) = λe λo λi
let case_e = e
let case_o = λx (i x)
let case_i = λx (o (Inc x))
(x case_e case_o case_i)This makes λx (Inc (Inc x)) have a constant-space normal form, which in turn
makes the composition of Inc fast, allowing Add to be efficiently
implemented as repeated increment.
Similar uses of this idea can greatly speed up functional algorithms. For
example, a clever way to implement a Data.List would be to let all algorithms
operate on Church-encoded Lists under the hoods, converting as needed. This has
the same "deforestation" effect of Haskell's rewrite pragmas, without any hard-
coded compile-time rewriting, and in a more flexible way. For example, using
map in a loop is "deforested" in HVM. GHC can't do that, because the number
of applications is not known statically.
Note that too much cloning will often make your normal forms large, so avoid these by keeping your programs linear. For example, instead of:
Add = λa λb
let case_zero = b
let case_succ = λa_pred (Add a_pred b)
(a case_succ case_zero)Write:
Add = λa
let case_zero = λb b
let case_succ = λa_pred λb (Add a_pred b)
(a case_succ case_zero b)Notice how the latter avoids cloning b entirely.
[TODO]