elm-optimize-level-2/notes/transformations.md

Overview of Transformations

This is an overview of the transformations for elm-optimize-level-2.

Not all of them made the cut, but seeing that a transformation is not as effective as initially thought is really good information.

We got a huge head start because of Robin's article.

Each transformation also has a rough summary of impact.

Applying Functions Directly

Elm wraps functions in an object that tracks how many arguments the function takes(also known as 'arity').

This is so that functions can be partially applied, meaning you can apply a few arguments and get a new function that has those arguments "built in".

The most significant speedups we've seen is in finding places where we can skip the wrapper and call the actual function directly. This happens when you call a function with exactly the number of arguments it needs.

In order to do this, we need to adjust function declarations so that the original function can be called either in the standard 'wrapped' way, or directly.

before

var MyFunction = F2(function (tag, value) {
  return value;
});

after

var MyFunction_fn = function (tag, value) {
    return value;
  },
  MyFunction = F2(MyFunction_fn);

Then, if this function is called with A2, we can unwrap the wrapper and call the function directly.

before

A2(MyFunction, one two)

after

MyFunction_fn(one two)

Results Summary

• Included in elm-optimize-level-2 tool**
• Potentially large positive effect on speed
• Likley small but positive effect on asset size

This has lead to dramatic speedups in some cases, especially when a large number of smaller functions are called and the overhead of calling twice as many functions is significant.

As well, it has a really interesting characteristic in that it makes the initial size of the generated JS larger, but usually results in a smaller minified asset size.

We generate two definitions for a function, but in most cases a function is either always partially applied, or always called with the full number of arguments.

If a function is always called with the full number of arguments, the minifier can eliminate our wrapped version (F2(MyFunction_fn)) and also eliminate the A2 call, which is explicitly smaller than before.

Passing unwrapped functions and calling them directly

Let's say we have some elm code that produces the following js.

var f = function (func, a, b) {
  return A2(func, a, b);
};

f(
  F2(function (a, b) {
    return a + b;
  }),
  1,
  2
);

we can transform it to

var f = function (func, a, b) {
    return A2(func, a, b);
  },
  f_unwrapped = function (func, a, b) {
    return func(a, b); // <-- direct function call!
  };

// note that the lambda is unwrapped as well
f_unwrapped(
  function (a, b) {
    return a + b;
  },
  1,
  2
);

This transformation works with separately defined functions too.

Passing in Unwrappable Functions to Higher Order Functions

Future Work

Higher order functions like List.map have a hard time taking advantage of the direct function calls because we don't know the arity of the function within the List.map call.

This is a challenging case, but worth exploring!

Making type representation isomorphic

Currently the Elm compiler will generate objects that match the shape of a given type.

Maybe looks like this:

var elm$core$Maybe$Just = function (a) {
  return { $: 0, a: a };
};

var elm$core$Maybe$Nothing = { $: 1 };

However, the V8 engine is likely better able to optimize these objects if they have the same shape.

So, this transformation fills out the rest of the variants with field: null so that they have the same shape.

var elm$core$Maybe$Just = function (a) {
  return { $: 0, a: a };
};

var elm$core$Maybe$Nothing = { $: 1, a: null };

This does require information from the Elm code itself, which we're currently getting through elm-tree-sitter.

Results Summary

• Partially Included - We stub in shapes for Maybe and List, but that's it.
• Has an effect in certain circumstances in browsers using V8(Chrome and Edge). Nothing observable otherwise.
  • Most prominently observed in the Elm Core - sum 300 list of custom types benchmark. Otherwise I didn't notice it.
• No noticable effect on asset size.
• More work is needed to make parsing an existing project more robust and also to discover what approach for representing the shapes actually produces the best benefit.

Inlining literal list constructors

Before

_List_fromArray(['a', 'b', 'c']);

After, using InlineMode.UsingConsFunc

_List_cons('a', _List_cons('b', _List_cons('c', _List_Nil)));

with InlineMode.UsingLiteralObjects

({ $: 1, a: 'a', b: { $: 1, a: 'b', b: { $: 1, a: 'c', b: _List_Nil } } });

Note - Elm actually had this originally (the literal objects version)! But there's an issue in Chrome with more than 1000 elements.

There's also tradeoff between asset size and speed.

Also of note, because _List_fromArray is used for lists of anything, that it's likely being deoptimized by the JavaScript compiler.

There may be a nice trade-off here of using InlineMode.UsingConsFunc, but only inlining at most 20 elements or something, and then using List_fromArray after that.

Results Summary

• Not included in the elm-optimize-level-2 tool because it was hard to find a benchmark that reported numbers to justify it.
• Though maybe we just need to be better at benchmarking it.

Object Update

When updating a record in Elm via { record | field = new }, Elm runs the following function:

function _Utils_update(oldRecord, updatedFields) {
  var newRecord = {};
  for (var key in oldRecord) {
    newRecord[key] = oldRecord[key];
  }
  for (var key in updatedFields) {
    newRecord[key] = updatedFields[key];
  }
  return newRecord;
}

We tried a few different variations in order to see if we could speed this up.

The trick here is that we need to copy the entire record so that it has a new reference.

So, we can't just do record.field = new in the js.

All of these tricks rely on either the spread operator or Object.assign, both of which are not supported in IE.

Replacing the implementation of _Util_update:

Spread operator

const _Utils_update = (oldRecord, updatedFields) => {
  var newRecord = { ...oldRecord };

  for (var key in updatedFields) {
    newRecord[key] = updatedFields[key];
  }
  return newRecord;
};

Spread for both

const _Utils_update = (oldRecord, updatedFields) => ({
  ...oldRecord,
  ...updatedFields,
});

Use Object.assign

const _Utils_update = (oldRecord, updatedFields) =>
  Object.assign({}, oldRecord, updatedFields);

Inline the call altogether

At the call site, replace

_Utils_update(old, newFields);

with

Object.assign({}, old, newFields);

Or we can use the spread operator inline:

{...old, field: new}

Result Summary

• Not included in elm-optimize-level-2 tool
• Again, all of these tricks rely on either the spread operator or Object.assign, both of which are not supported in IE.
• The most promising approach was inlining the call completely with {...old, field: newValue}.
  • Gave a 501% boost in chrome!
  • And caused safari to reduce performance by 50% 😅

Simply creating a new record and copying each field manually is significantly faster than any of these transformations (~9x in chrome, and ~6.5x in firefox). You can do this directly in Elm.

updateSingleRecordManually record =
    { one = 87
    , two = record.two
    , three = record.three
    }

It's worth exploring automating this transformation, though of course there's a question of how much this affects asset size on larger projects.

However, it's hard to explore further without knowing the actual shape of the records being updated.

Future work Explore more approaches. Next on TODO list:

_Utils_update(old, {a: newA})

to

{...old, a: newA}

Inline Equality

If Elm's == is applied to any primitive such as:

• Int
• Float
• String
• Bool

Then we can inline the definition directly as JS strict equality: ===.

Right now elm-optimize-level-2 will infer if something is a primitive if a literal is used.

Results Summary

• Included in elm-optimize-level-2 tool.
• Looks to have the some impact on code that does a lot of equality comparisons, like parsing.

The _Utils_eq function is very likely deoptimized because it can take any two values and either do a reference check, or do structural equality, which we also know takes a while.

So, my guess is the benefit here is from avoiding the call to a deoptimized function completely.

Chrome doesn't really see a speedup here though, so it's likely smart enough to do that already.

Inline String.fromFloat/Int

Before

String$fromFloat(val);

After:

val + '';

Results Summary

• Not included in the tool

This hasn't shown any measureable benefit. Likely because this is such a simple function that all js compilers are already optimizing the intermedaite calls.

Arrowizing Functions

Before

var x = function (x) {};

After

var x = (x) => {};

This was done for asset size. The nuance being that it's done to potentially optimize the minified size of code, but not necessarily the gzipped version.

This is still a benefit because the minified code is what ultimately needs to be parsed and parsing is one of the larger steps on the way to getting a page running.

Results Summary

• Not included in the elm-optimize-level-2 tool
• Comes with the caveat that the code will not work on IE

We weren't able to pin down a benchmark where this reported a benefit in the numbers, though likely to explore this we need (1) A larger codebase, and (2)

We didn't include this in the first version of the tool because the effect seems to be so modest and carries the risk of breaking things on IE.

We would have to add something like a --modernize or --no-ie flag to the tool, and I really like this tool having no configurability.

Lifting Constants

Future Work

This transformation hasn't been attempted yet, but the idea is that if a constant is detected in a let statement, it can be declared moved to top-level instead of recalculated every function run.

This is risky! You do less computation, but you are (1) moving a bunch of computation to happen on start-up and (2) the results are allocated but can never be freed. This could be worthwhile in HTML though, where there is a x === y check on nodes: https://github.com/elm/virtual-dom/blob/master/src/Elm/Kernel/VirtualDom.js#L706-L709 So, if two nodes were reference equal, you wouldn't have to ever diff them. I imagine this could be a big benefit if there was a long list where each element contained a somewhat large "constant" node for some UI thing.

Eta Conversion

This is when you add or remove anonymous functions: map (f x y) zs to map (\z -> f x y z) zs

Because of our previous optimizations where we can call a function directly, this can make sure we're getting the fast version of f!

Tail Recursion Modulo Cons

Whew, what a name!

In Elm, if you have a recursive function that calls itself at the top level, then it will be compiled into a while loop. Here's an example:

sum : Int -> List Int ->  Int
sum current list =
  case list of
    [] ->
        current

    (x :: remaining) ->
        sum (x + current) remaining

The critical part here is that sum is called as the first thing on that branch. Because that's easy to detect, we can reliably convert the above code into a while loop. (Note, all this stuff is called Tail Call Optimization).

However, it's a pretty common case where we can't quite do that.

Let's take a look at an implementation of List.map. (Note, this isn't how Elm currently does it, this is just for illustrative purposes.)

We could implement List.map like this:

type List a = Nil | Cons a (List a)

map : (a -> b) -> List a -> List b
map func list =
  case list of
    Nil -> Nil
    Cons x xs ->
      Cons (func x) (map func xs)

But our recursive function(map) is not the first thing being called in that branch! It's the Cons constructor! (Heyy, Tail Recursion Modulo Cons)

The idea for this transformation is that we could take the above code, and generate the following JS code.

function map(func, list) {
  var first;
  var prev;
  while (true) {
    if (list.$ === 'Nil') {
      if (prev) {
        prev.b = Nil;
      } else {
        first = Nil;
      }
      return first;
    } else {
      var node = Cons(func(list.a), null);
      if (prev) {
        prev.b = node;
      } else {
        first = node;
      }
      prev = node;
      list = list.b;
    }
  }
}

Skip Allocating Tuples in Case Statements

It's pretty common to put things in a tuple (or threeple) to start a case statement.

    case (a, b) of
       (ThingOne, ThingTwo) ->
          --...

We could skip allocating the tuple though.