Optimize sorted() for primitive-type arrays

[pubiqq]

See KT-76423

[fzhinkin]

@pubiqq, nice catch! Thanks for opening the PR.

I slightly modified your benchmark to both validate results and check how the updated version of sorted works for unsigned array types: https://github.com/fzhinkin/kt-76423-benchmarks/tree/main
Results are inside results directory.

For regular primitive array types the change improves sorted performance across all benchmarked targets (JVM, JS, Wasm/JS, native linuxX64, native macosAarch64).

For unsigned array types, however, the change does not affect performance, so it's better to abstain from updating corresponding implementations (or try to find a way to speed them up too).

Also, it is worth updating sortedDescending implementations the same way.

[pubiqq]

For unsigned array types, however, the change does not affect performance, so it's better to abstain from updating corresponding implementations (or try to find a way to speed them up too).

The changes to the unsigned arrays were added only to make the sorted() implementations uniform. They did not imply any performance changes, the copyOf().apply { sort() }.asList() (current sorted()) and sortedArray().asList() (sorted() from the PR) are almost equivalent (see sortedArray() implementation).

However, okay, I'll roll it back.

Also, it is worth updating sortedDescending implementations the same way.

sortedDescending does not have the same problem as the sorted method. However, I've created another issue related to sortedDescending, see KT-76458.

[fzhinkin]

Previously, asList() function was invoked on a Array<T>, now it is invoked on a corresponding primitive array.
The change is subtle, but it results in different underlying types of the resulting list. Further analysis is required to determine how this change may affect users. For example, some codebases may rely on the fact that the resulting list is effectively mutable and explicitly cast it to MutableList<T>. Changed underlying type may also affect call sites where a list produced by the sorted() call is used. This need to be investigated further.
Besides these potential issues, there is a less ephemeral one: PrimitiveArray.asList() returns a wrapper around the primitive array (for example,

kotlin/libraries/stdlib/jvm/src/generated/_ArraysJvm.kt

Line 167 in bf92d5d

public actual fun IntArray.asList(): List<Int> {

). As a result, access to individual elements requires boxing, which results in worse performance in general. It's not clear if performance gained by optimizing sorting will always offset performance regression in element-access scenarios.

As a middle ground, sorted may convert sorted primitive array into a typed array first and then invoke asList on it (sortedArray().toTypedArray().asList()). Such an implementation outperforms the existing one (by saving time on sorting itself), but returns exactly the same list type as before.

[pubiqq]

Besides these potential issues, there is a less ephemeral one: PrimitiveArray.asList() returns a wrapper around the primitive array ... As a result, access to individual elements requires boxing, which results in worse performance in general.

The benchmark results (https://github.com/pubiqq/benchmarks-kmp/tree/list-get-benchmarks) do not allow us to state this unequivocally, in most cases accessing a primitive from PrimitiveArray.asList() is faster than from PrimitiveArray.toTypedArray().asList().

It should also be noted that toTypedArray() boxes primitives too. Yes, it does this preemptively, but it does this for all elements, and this is not always better than using a list that converts primitives on demand.

As a middle ground, sorted may convert sorted primitive array into a typed array first and then invoke asList on it (sortedArray().toTypedArray().asList()). Such an implementation outperforms the existing one (by saving time on sorting itself), but returns exactly the same list type as before.

Such an implementation also performs two array copy operations instead of one. I'm not sure it's worth it.