Was there an issue to implement loop merging for c...

# language-proposals

[voddan]

Was there an issue to implement loop merging for chains of collection transformations (map, filter, etc)? I remember some discussions going, but I can't find the issue.

[voddan]

A.k.a loop fusion. But I still can't find an issue or a forum discussion

[voddan]

I think I remember @abreslav explaining why it was hard to implement. Was it documented somewhere?

[ilya.gorbunov]

A chain of operations cannot be fused into a loop readily because it would change the operation execution order.

[ilya.gorbunov]

However recently it was proposed that the chain of sequence operations like coll.asSequence().map{...}.filter{...}.toList() could be fused as it won't change the order of operations.

[voddan]

@ilya.gorbunov my idea is to fuse normal coll.map{}.filter{} by re-useing the intermediate list that they create. I am running some benchmarks now, but preliminary results show 30% reduction in allocated memory, and this is for relatively uncomplicated examples.

[dmitry.petrov]

`xs.map(f).filter(g)`: f f f g g g
For-loop: f g f g f g
Unless we can prove that such transformation is equivalent (or at least that f is "pure"), we can't do that automatically.

[dmitry.petrov]

On HotSpot, call chain in fact has mostly similar performance with the 'for', because intermediate containers are rather short-lived, and gc cleans them fast enough.
Unfortunately, that's not true on Android. We know that, and we are considering some options (loop fusion, collection comprehensions, intrinsics for sequences,...). Not in a short-term, though.

[voddan]

@dmitry.petrov my solution preserves the order. Looping is kept the same, but instead of several intermediate lists it creates only one, then mutates it.

[dmitry.petrov]

Sounds good. Feel free to submit a pr.

[voddan]

Some preliminary results https://youtrack.jetbrains.com/issue/KT-21997