Discuss
  • Syntax sugar for partial application

Syntax sugar for partial application

# Jussi Kalliokoski (9 years ago)

Yesterday I came up with an idea for syntactic sugar for partial application, introducing two new operators: placeholder (?) and rest placeholder (???).

You can see the details in a proposal gist I made [1], but the gist of the gist is that you could do partial application like this:

foo(1, ?, 2);

or with the rest placeholder:

foo(?, 1, ???);

This allows for partial application at arbitrary argument indices.

There's a cowpath to be paved here as well, lodash introduced placeholders for _.partial as of 3.0.0 [2], where you can do something similar:

_.partial(foo, 1, _, 2); _.partial(foo, _, 1);

However, the proposed syntax is even more flexible than that of lodash because it also allows you to have rest placeholders at arbitrary argument indices instead of at the end, e.g.

foo(1, ???, 2);

would have to be expressed with lodash as

.partialRight(.partial(foo, 1), 2);

WDYT?

[1] gist.github.com/anonymous/5c4f6ea07ad3017d61be [2] lodash.com/docs#partial

# Salvador de la Puente González (9 years ago)

I like it but I think I would prefer to make it explicit. Something like:

foo = foo.partial(1, ?, 2, ???, 3)

Extending it to bind when I want to do something like:

obj.foo = foo.bind(obj, 1, ?, 2, ???, 3)

WDYT?

# Nick Krempel (9 years ago)

On 9 April 2015 at 08:46, Jussi Kalliokoski <jussi.kalliokoski at gmail.com>

wrote:

Yesterday I came up with an idea for syntactic sugar for partial application, introducing two new operators: placeholder (?) and rest placeholder (???).

While this looks nice, I'm not sure it's worth new syntax as it can be adequately implemented in a library, as you point out with lodash.

The library implementation could handle a rest placeholder too.

E.g.

foo.partial(1, _arg, 2, _rest, 3)

(or with whatever names you prefer, could be imports from a module).

What's the big advantage of syntactic sugar here?

# liorean (9 years ago)

Do we really need it? Your «foo(1, ?, 2);» is equivalent to «a=>foo(1,a,2)».

Your «foo(?, 1, ???);» is equivalent to «(a,...b)=>foo(a,1,...b)».

Your «foo(1, ???, 2);» is equivalent to «(...a)=>foo(...[1,...a,2])».

Also, the ? token is already taken by the ternary conditional operator. Do we really want to overload it here for a nullary operator/special form, when we have as low overhead syntax as we already do in fat arrows for doing the exact same thing?

# Andrea Giammarchi (9 years ago)

FWIW: agreed with others, it looks a pretty pointless sugar. It doesn't seem to bring anything new or "that needed" to the language.

-1 here

# Jussi Kalliokoski (9 years ago)

On Thu, Apr 9, 2015 at 4:04 PM, liorean <liorean at gmail.com> wrote:

Do we really need it? Your «foo(1, ?, 2);» is equivalent to «a=>foo(1,a,2)». Your «foo(?, 1, ???);» is equivalent to «(a,...b)=>foo(a,1,...b)». Your «foo(1, ???, 2);» is equivalent to «(...a)=>foo(...[1,...a,2])».

Not exactly. Using the placeholder syntax, this remains context dependent, whereas with your examples you get null as this.

This might not seem like such a big deal until you consider it in combination with the proposed bind syntax [1].

Also in your examples, redefining foo will lead to different results. The placeholder syntax has a lot more room for optimization in the JIT compiler (the partially applied result is guaranteed to have no side effects for example, so the compiler can create a version of the original function where it can inline the specified arguments; less moving parts, easier to optimize).

[1] strawman:bind_operator

# Alexander Jones (9 years ago)

Using an indiscriminate ? is quite inflexible as you wouldn't be able to reorder the arguments.

In Swift there is a nice feature whereby you can avoid naming arguments in closures and refer to the args by $1 $2 $3 etc., e.g.

val g = { f($1, foo, $2, bar, $3) }

which is equivalent to the more verbose

val g = { arg1, arg2, arg3 in f(arg1, foo, arg2, bar, arg3) }

which is nice as it trivially unifies partial application with closure syntax.

Given that the boat has sailed on arrow functions in ES6 now, something like

const g = ($1, $2, $3) => f($1, foo, $2, bar, $3)

could be shortened right now to

const g = (...$) => f($[0], foo, $[1], bar, $[2])

and I suppose a new syntax could shorten that to something like

const g = #f(#0, foo, #1, bar, #2)

which I naively believe wouldn't be too much of a problem for the grammar. It doesn't really save you that much typing though.

(Also, re your comment about this, arrow functions explicitly have lexical this!)

# Salvador de la Puente González (9 years ago)

I always saw partial application lacking from function prototype but its true that with arrow functions it is straightforward enough. Furthermore it behaves as I always expected, binding the new this to current this.

Works for me. :) Using an indiscriminate ? is quite inflexible as you wouldn't be able to reorder the arguments.

In Swift there is a nice feature whereby you can avoid naming arguments in closures and refer to the args by $1 $2 $3 etc., e.g.

val g = { f($1, foo, $2, bar, $3) }

which is equivalent to the more verbose

val g = { arg1, arg2, arg3 in f(arg1, foo, arg2, bar, arg3) }

which is nice as it trivially unifies partial application with closure syntax.

Given that the boat has sailed on arrow functions in ES6 now, something like

const g = ($1, $2, $3) => f($1, foo, $2, bar, $3)

could be shortened right now to

const g = (...$) => f($[0], foo, $[1], bar, $[2])

and I suppose a new syntax could shorten that to something like

const g = #f(#0, foo, #1, bar, #2)

which I naively believe wouldn't be too much of a problem for the grammar. It doesn't really save you that much typing though.

(Also, re your comment about this, arrow functions explicitly have lexical this!)

# liorean (9 years ago)

On 9 April 2015 at 16:11, Jussi Kalliokoski <jussi.kalliokoski at gmail.com> wrote:

On Thu, Apr 9, 2015 at 4:04 PM, liorean <liorean at gmail.com> wrote:

Do we really need it? Your «foo(1, ?, 2);» is equivalent to «a=>foo(1,a,2)». Your «foo(?, 1, ???);» is equivalent to «(a,...b)=>foo(a,1,...b)». Your «foo(1, ???, 2);» is equivalent to «(...a)=>foo(...[1,...a,2])».

Not exactly. Using the placeholder syntax, this remains context dependent, whereas with your examples you get null as this.

No, «this» is lexically bound to be that of the enclosing lexical scope in arrow functions, so it would be whatever that is. But that doesn't really matter as the function call to «foo» doesn't use the «this» of the arrow function.

Now, if we were to say your «foo» were actually «foo.bar», and you did the same replacement in the arrow function, the «this» value of the «bar» call would be «foo», so that's pretty much what is wanted as well. The case where this breaks is if you were to replace only the «bar» method with the arrow function, in which case it would use the lexical «this» instead of «foo», but that's obviously not the right transformation to use.

This might not seem like such a big deal until you consider it in combination with the proposed bind syntax [1].

Also in your examples, redefining foo will lead to different results. The placeholder syntax has a lot more room for optimization in the JIT compiler (the partially applied result is guaranteed to have no side effects for example, so the compiler can create a version of the original function where it can inline the specified arguments; less moving parts, easier to optimize).

Yeah, it's susceptible to that problem, yes. Do you want me to fix that for you if you really want it?

Your «foo(1, ?, 2);» is equivalent to «((f,a)=>f(1,a,2))(foo)».

Your «foo(?, 1, ???);» is equivalent to «((f,a,...b)=>f(a,1,...b))(foo)».

Your «foo(1, ???, 2);» is equivalent to «((f,...a)=>f(...[1,...a,2]))(foo)».

I guess I didn't think of these cases though, because I only use explicit arguments to my functions these days, I never use the «this» keyword. If I want a function to operate on an object, I pass that object into the function. I also try to not reuse my variables unless they are part of an iteration, in which case they are always local variables that are only handled in the iteration process itself. But that's a side issue, as it's about my code rather than precepts of the language.

# Jussi Kalliokoski (9 years ago)

On Sat, Apr 11, 2015 at 4:54 AM, liorean <liorean at gmail.com> wrote:

On 9 April 2015 at 16:11, Jussi Kalliokoski <jussi.kalliokoski at gmail.com> wrote:

On Thu, Apr 9, 2015 at 4:04 PM, liorean <liorean at gmail.com> wrote:

Do we really need it? Your «foo(1, ?, 2);» is equivalent to «a=>foo(1,a,2)». Your «foo(?, 1, ???);» is equivalent to «(a,...b)=>foo(a,1,...b)». Your «foo(1, ???, 2);» is equivalent to «(...a)=>foo(...[1,...a,2])».

Not exactly. Using the placeholder syntax, this remains context dependent, whereas with your examples you get null as this.

No, «this» is lexically bound to be that of the enclosing lexical scope in arrow functions, so it would be whatever that is. But that doesn't really matter as the function call to «foo» doesn't use the «this» of the arrow function.

Exactly why you get null as this.

Now, if we were to say your «foo» were actually «foo.bar», and you did the same replacement in the arrow function, the «this» value of the «bar» call would be «foo», so that's pretty much what is wanted as well. The case where this breaks is if you were to replace only the «bar» method with the arrow function, in which case it would use the lexical «this» instead of «foo», but that's obviously not the right transformation to use.

This might not seem like such a big deal until you consider it in combination with the proposed bind syntax [1].

Also in your examples, redefining foo will lead to different results. The placeholder syntax has a lot more room for optimization in the JIT compiler (the partially applied result is guaranteed to have no side effects for example, so the compiler can create a version of the original function where it can inline the specified arguments; less moving parts, easier to optimize).

Yeah, it's susceptible to that problem, yes. Do you want me to fix that for you if you really want it?

Your «foo(1, ?, 2);» is equivalent to «((f,a)=>f(1,a,2))(foo)».

Your «foo(?, 1, ???);» is equivalent to «((f,a,...b)=>f(a,1,...b))(foo)».

Your «foo(1, ???, 2);» is equivalent to «((f,...a)=>f(...[1,...a,2]))(foo)».

Your new examples directly execute the function instead of creating a new function. :) Which goes to show how it would be nice to have specific syntax for this to make it more obvious what's happening.

I guess I didn't think of these cases though, because I only use explicit arguments to my functions these days, I never use the «this» keyword. If I want a function to operate on an object, I pass that object into the function. I also try to not reuse my variables unless they are part of an iteration, in which case they are always local variables that are only handled in the iteration process itself. But that's a side issue, as it's about my code rather than precepts of the language.

I write my code pretty much the same way. However, it's hard for the compiler to trust that you're not changing things, regardless of style.

Also because most of the standard library of the language operates on this instead of a separate argument, combining standard library methods with methods that have their data as an explicit argument often lead to awkward reading order issues, e.g.

foo(x .filter(...) .map(...) .reduce(...) )

whereas with the bind operator you get

x .filter(...) .map(...) .reduce(...) ::foo()

Which is where this proposal shines, if foo is a partially applied function.

But anyway, seems that this is not something people want, at least yet, so I'll rest my case. :)

# liorean (9 years ago)

On 12 April 2015 at 17:39, Jussi Kalliokoski <jussi.kalliokoski at gmail.com> wrote:

No, «this» is lexically bound to be that of the enclosing lexical scope in arrow functions, so it would be whatever that is. But that doesn't really matter as the function call to «foo» doesn't use the «this» of the arrow function.

Exactly why you get null as this.

Which makes the behaviour identical to that of the code as you wrote it.

Now, if we were to say your «foo» were actually «foo.bar», and you did the same replacement in the arrow function, the «this» value of the «bar» call would be «foo», so that's pretty much what is wanted as well. The case where this breaks is if you were to replace only the «bar» method with the arrow function, in which case it would use the lexical «this» instead of «foo», but that's obviously not the right transformation to use.

This might not seem like such a big deal until you consider it in combination with the proposed bind syntax [1].

Also in your examples, redefining foo will lead to different results. The placeholder syntax has a lot more room for optimization in the JIT compiler (the partially applied result is guaranteed to have no side effects for example, so the compiler can create a version of the original function where it can inline the specified arguments; less moving parts, easier to optimize).

Yeah, it's susceptible to that problem, yes. Do you want me to fix that for you if you really want it?

Your «foo(1, ?, 2);» is equivalent to «((f,a)=>f(1,a,2))(foo)».

Your «foo(?, 1, ???);» is equivalent to «((f,a,...b)=>f(a,1,...b))(foo)». Your «foo(1, ???, 2);» is equivalent to «((f,...a)=>f(...[1,...a,2]))(foo)».

Your new examples directly execute the function instead of creating a new function. :) Which goes to show how it would be nice to have specific syntax for this to make it more obvious what's happening.

Oops. I needed to actually add that extra argument as a separate fat arrow, «(f=>(...a)=>f(...[1,...a,2]))(foo)» etc.

I write my code pretty much the same way. However, it's hard for the compiler to trust that you're not changing things, regardless of style.

Guess it'd be hard for it unless it has the knowledge of whether functions are pure or not, yes.

I'd love for a compiler that can tell that I don't modify my arguments and thus optimises code like « let map= // Usage: map(function)(...array) (f,...acc)=>(head,...tail)=>( undefined===head ?acc :map(f,...acc,f(head))(...tail)); »

So that it doesn't actually create the «tail» array every recursion, just a narrower and narrower subarray of the same actual array, and likewise that the only thing that is done with «acc» is the production of an array that is identical to it with an addition of one element at its end, so doesn't break it down and rebuild it every recursion. And of course tail call optimisation on it, because that code is horrid without those optimisations.

# Jussi Kalliokoski (9 years ago)

On Mon, Apr 13, 2015 at 12:15 AM, liorean <liorean at gmail.com> wrote:

On 12 April 2015 at 17:39, Jussi Kalliokoski <jussi.kalliokoski at gmail.com> wrote:

No, «this» is lexically bound to be that of the enclosing lexical scope in arrow functions, so it would be whatever that is. But that doesn't really matter as the function call to «foo» doesn't use the «this» of the arrow function.

Exactly why you get null as this.

Which makes the behaviour identical to that of the code as you wrote it.

If you looked at the gist I made, the placeholder syntax creates functions where this remains untouched, so it can be separately bound, unlike in your examples. See [1] for example of how the syntax desugars to ES6.

Now, if we were to say your «foo» were actually «foo.bar», and you did the same replacement in the arrow function, the «this» value of the «bar» call would be «foo», so that's pretty much what is wanted as well. The case where this breaks is if you were to replace only the «bar» method with the arrow function, in which case it would use the lexical «this» instead of «foo», but that's obviously not the right transformation to use.

This might not seem like such a big deal until you consider it in combination with the proposed bind syntax [1].

Also in your examples, redefining foo will lead to different results.

The placeholder syntax has a lot more room for optimization in the JIT compiler (the partially applied result is guaranteed to have no side effects for

example, so the compiler can create a version of the original function where it can inline the specified arguments; less moving parts, easier to optimize).

Yeah, it's susceptible to that problem, yes. Do you want me to fix that for you if you really want it?

Your «foo(1, ?, 2);» is equivalent to «((f,a)=>f(1,a,2))(foo)».

Your «foo(?, 1, ???);» is equivalent to «((f,a,...b)=>f(a,1,...b))(foo)».

Your «foo(1, ???, 2);» is equivalent to «((f,...a)=>f(...[1,...a,2]))(foo)».

Your new examples directly execute the function instead of creating a new function. :) Which goes to show how it would be nice to have specific syntax for this to make it more obvious what's happening.

Oops. I needed to actually add that extra argument as a separate fat arrow, «(f=>(...a)=>f(...[1,...a,2]))(foo)» etc.

I write my code pretty much the same way. However, it's hard for the compiler to trust that you're not changing things, regardless of style.

Guess it'd be hard for it unless it has the knowledge of whether functions are pure or not, yes.

I'd love for a compiler that can tell that I don't modify my arguments and thus optimises code like « let map= // Usage: map(function)(...array) (f,...acc)=>(head,...tail)=>( undefined===head ?acc :map(f,...acc,f(head))(...tail)); »

So that it doesn't actually create the «tail» array every recursion, just a narrower and narrower subarray of the same actual array, and likewise that the only thing that is done with «acc» is the production of an array that is identical to it with an addition of one element at its end, so doesn't break it down and rebuild it every recursion. And of course tail call optimisation on it, because that code is horrid without those optimisations.

Me too, yet while nothing is impossible, this sort of optimization (tail call optimization aside) would be super difficult to implement, and expensive too, because the compiler would have to check that the invariants weren't violated on every call. However, there is hope in the light of immutable collections proposals where the compiler has solid guarantee that the collection isn't morphed mid-iteration and can safely use a cursor to a sub-collection.

[1] gist.github.com/anonymous/5c4f6ea07ad3017d61be#leading