Update #1: Use x..y//z
instead of x..y..z
.
This is a proposal to address some of the limitations we have in Elixir ranges today. They are:
- It is not possible to have ranges with custom steps
- It is not possible to have empty ranges
- Users may accidentally forget to check the range boundaries
The first limitation is clear: today our ranges are increasing (step of 1) or decreasing (step of -1), but we cannot set arbitrary steps as in most other languages with range. For example, we can't have a range from 1 to 9 by 2 (i.e. 1, 3, 5, 7, 9).
The second limitation is that, due to how we currently infer the direction of ranges, it is not possible to have empty ranges. Personally, I find this the biggest limitation of ranges. For example, take the function Macro.generate_arguments(n, context)
in Elixir. This is often used by macro implementations, such as defdelegate
, when it has to generate a list of n
arguments. One might try to implement this function as follows:
def generate_arguments(n, context) do
for i <- 1..n, do: Macro.var(:"arg#{n}", context)
end
However, because n
may be zero, the above won't work: for n = 0
, it will return a list with two elements! To workaround this issue, the current implementation works like this:
def generate_arguments(n, context) do
tl(for i <- 0..n, do: Macro.var(:"arg#{n}", context))
end
In other words, we have to start the range from 0 and always discard the first element which is unclear and wasteful.
Finally, another issue that may arise with ranges is that implementations may forget to check the range boundaries. For example, imagine you were to implement range_to_list/1
:
def range_to_list(x..y), do: range_to_list(x, y)
defp range_to_list(y, y), do: [y]
defp range_to_list(x, y), do: [x | range_to_list(x + 1, y)]
While the implementation above looks correct at first glance, it will loop forever if a decreasing range is given.
The proposed solution is to support steps in Elixir ranges by adding ..//
as a ternary operator:
start..stop//step
Where //step
is optional. The idea to use a separate operator for the step, instead of start..stop..step
is to make a visual distinction between the parts. Also, you think the step of a to split the range, which then makes sense to use a division-like operator.
The ternary operator solves the three problems above:
It is not possible to have ranges with steps
Now you can write 1..9//2
(from 1 to 9 by 2).
It is not possible to have empty ranges
This can be addressed by explicitly passing the step to be 1, instead of letting Elixir infer it. The generate_arguments
function may now be implemented as:
def generate_arguments(n, context) do
for i <- 1..n//1, do: Macro.var(:"arg#{n}", context)
end
For n = 0
, it will construct 1..0//1
, an empty range.
Note 1..0//1
is distinct from 1..0
: the latter is equal to 1..0//-1
, a decreasing range of two elements: 1
and 0
. To avoid confusion, we plan to deprecate inferred decreasing ranges in the future.
Users may accidentally forget to check the range boundaries
If we introduce ranges with step and the ternary operator, we can forbid users to write x..y
in patterns. Doing so will emit a warning and request them to write x..y//z
instead, forcing them to explicitly consider the step, even if they match on the step to be 1. In my opinion, this is the biggest reason to add the ternary operator: to provide a convenient and correct way for users to match on ranges with steps.
The implementation happens in three steps:
-
Add
..//
as a ternary operator.x..y//z
will have the AST of{:..//, meta, [x, y, z]}
, consequently:..//
will be a valid atom and both&..///3
and&Kernel."..//"/3
will be valid captures -
Add the
:step
to range structs and implementdef x..y//z do
inKernel
-
Update the formatter to handle
x..y//z
accordingly -
Add deprecations. To follow Elixir's deprecation policy, the deprecation warnings shall only be emitted 4 Elixir versions after ranges with steps are added (most likely on v1.16):
-
Deprecate
x..y
as a shortcut for a decreasing range in favor ofx..y//-1
. The reason for this deprecation is because a non-empty range is more common than a decreasing range, so we want to optimize for that. Furthermore, having a step with a default of 1 is clearer than having a step that varies based on the arguments. Of course, we can only effectively change the defaults on Elixir v2.0, which is still not scheduled or planned. -
Deprecate
x..y
in patterns, requirex..y//z
instead. This will become an error on Elixir v2.0. -
Deprecate
x..y
in guards unless the arguments are literals (i.e.1..3
is fine, but not1..y
orx..1
orx..y
). This is necessary becausex..y
may be a decreasing range and there is no way we can warn about said cases in guards, so we need to restrict at the syntax level. For non-literals, you should either remove the range or use an explicit step. On Elixir v2.0,x..y
in guards will always mean a range with step of 1.
-
(will move to the mailing list if preferred)
Does this open up the possibility now or in the future of generalising the range? For example:
range.first
andrange.last
required to be the same struct. This would open up the possibility of specifying, for example, date and time ranges ormoney
ranges orunit of measure
ranges especially when they have sigils defined.I appreciate this would require some std lib functions (like
String.splice/2
in your examples) to explicitly guard forinteger
ranges. But this generalised syntax for expressing a range and a step is powerful and could reduce the instance of certain classes of errors.