Utilities for working with partial functions.
spago install partial
Every now and then, you will want to use partial functions; that is,
functions which don't handle every possible case of their inputs. For example,
there is a function fromJust :: ∀ a. Partial ⇒ Maybe a → a
in Data.Maybe
,
which gives you the value inside a Just
value, or throws an error if given
Nothing
.
It's important that types tell the truth wherever possible, because this is a
large part of what allows us to understand PureScript code easily and refactor
it fearlessly. However, in certain contexts, you know that e.g. an Either
value is always going to be Right
, but you can't prove that to the type
checker, and so you want an escape hatch so that you can write a function that
doesn't have to deal with the Left
case. This is often the case when
performance is important, for instance.
Previously, partial functions have been indicated by putting the word "unsafe"
at the start of their names, or by putting them in an "Unsafe" module. For
instance, there was previously an unsafeIndex
function in
Data.Array.Unsafe
, and fromJust
used to be in Data.Maybe.Unsafe
. However,
this is not ideal, because the fact that these functions are partial, and
therefore unsafe if used carelessly, does not appear in the type. Consequently,
there is little to stop you from using it in an inappropriate manner by
accident.
The Partial type class allows us to put this information back into the types, and thereby allows us to clearly demarcate which parts of your code are responsible for making sure that unsafe functions are used in a safe manner.
If you try to just use a partial function, you'll most likely get an error
about no instance being found for the Partial
class. Take this program, for
instance:
module Main where
import Prelude
import Data.Maybe (Maybe(..), fromJust)
import Effect (Effect)
import Effect.Console (logShow)
main :: Effect Unit
main = logShow (fromJust (Just 3))
Because fromJust
is partial, and because the partiality hasn't been
explicitly handled, you'll get an error:
at src/Main.purs line 8, column 1 - line 8, column 56
No type class instance was found for
Prim.Partial
Aside: Yes, this is not a fantastic error. It's going to get better soon.
The solution is usually to add an application of unsafePartial
somewhere,
like this:
module Main where
import Prelude
import Data.Maybe (Maybe(..), fromJust)
import Effect (Effect)
import Effect.Console (logShow)
import Partial.Unsafe (unsafePartial)
main :: Effect Unit
main = logShow (unsafePartial (fromJust (Just 3)))
The rule of thumb is to put unsafePartial
at the level of your program such
that the types tell the truth, and the part of your program responsible for
making sure a use of a partial function is safe is also the part where the
unsafePartial
is. This is perhaps best demonstrated with an example.
Imagine that we want to represent vectors in 3D with an array containing
exactly 3 values (perhaps we want to use them with some other API that expects
this representation, and we don't want to be converting back and forth all the
time). In this case, we would usually use a newtype
and avoid exporting the
constructor:
module Data.V3
( V3()
, makeV3
, runV3
) where
newtype V3 = V3 (Array Number)
makeV3 :: Number -> Number -> Number -> V3
makeV3 x y z = V3 [x, y, z]
runV3 :: V3 -> Array Number
runV3 (V3 v) = v
This way, all of the functions are safe; the code will guarantee that any V3
does contain exactly 3 values (although the type checker is not aware of this).
Now imagine we want to write a dot product function:
dot :: V3 -> V3 -> Number
dot (V3 [x1, x2, x3]) (V3 [y1, y2, y3]) = x1*y1 + x2*y2 + x3*y3
We know this is ok, but the compiler disallows it:
A case expression could not be determined to cover all inputs.
The following additional cases are required to cover all inputs:
(V3 _) _
_ (V3 _)
Alternatively, add a Partial constraint to the type of the enclosing value.
in value declaration dot
In this case, we can use unsafePartial
to explicitly say that we don't
actually need to worry about those other cases, and therefore we don't want to
propagate a Partial
constraint; users of this dot
function should not have
to worry about this partiality. For example:
dot :: V3 -> V3 -> Number
dot x y = Partial.Unsafe.unsafePartial (go x y)
where
go :: Partial => V3 -> V3 -> Number
go (V3 [x1, x2, x3]) (V3 [y1, y2, y3]) = x1*y1 + x2*y2 + x3*y3
-- This second pattern can be omitted, but provides a better error message
-- in case we do get an invalid argument at runtime.
go _ _ = Partial.crash "Bad argument: expected exactly 3 elements."
The unsafePartial
function comes from the Partial.Unsafe
module, in the
purescript-partial
package.
In this case, we could also use Partial.Unsafe.unsafeCrashWith
:
dot :: V3 -> V3 -> Number
dot (V3 [x1, x2, x3]) (V3 [y1, y2, y3]) = x1*y1 + x2*y2 + x3*y3
dot _ _ = unsafeCrashWith "Bad argument: expected exactly 3 elements."
Both implementations will behave in the same way.
In this case, we know our dot
implementation is fine, and so users of it
should not have to worry about its partiality, so it makes sense to avoid
propagating the constraint. Now, we will see another case where a Partial
constraint should be propagated.
Let us suppose we want a foldr1
function, which works in a very similar way
to foldr
on Lists, except that it doesn't require an initial value to be
passed, and instead requires that the list argument contains at least one
element.
We can implement it like this:
foldr1 f (Cons x xs) = foldr f x xs
The compiler infers the correct type here, which is:
foldr1 :: forall a. Partial => (a -> a -> a) -> List a -> a
Now imagine we want a version of Data.Foldable.minimum
which returns an a
instead of a Maybe a
, and is therefore partial. We can implement it in terms
of our new foldr1
function:
minimumP = foldr1 min
Again, the compiler infers the correct type:
minimumP :: forall a. (Partial, Ord a) => List a -> a
Notice that the Partial
constraint is automatically propagated to the
minimumP
function because of the use of another partial function in its
definition, namely foldr1
. In this case, this is what we want; we should
propagate the Partial
constraint, because it is still the caller's
responsibility to make sure they supply a non-empty list.
So hopefully it is now clear why this partiality checking is implemented in terms of a type class: it allows us to elegantly reuse existing machinery in the type checker in order to check that a Partial constraint is either explictly handled or propagated. This should help ensure that when you're reading the code a few months later, it remains clear which part of the code is responsible for ensuring that any assumed invariants which cannot be encoded in the type system do hold.
- API documentation is published on Pursuit.