Parse Int from Substring slower than String

Using Swift
1

I want to parse numbers from a string that contains multiple numbers separated by a space. After splitting the string, however, I found that rather than converting the split Substring to Int directly, converting it to String at first and then converting to Int is much faster.

I did a benchmark using Google's swift-benchmark:

// inputline is like "86 47 85 19 37 38 73"
let inputline: String = (1...10000)
    .map{ _ in String(Int.random(in: 0...1000)) }
    .joined(separator: " ")

let substrings = inputline.split(separator: " ")

benchmark("Substring -> Int") {
    let _ = substrings.map { Int($0)! }
}

benchmark("Substring -> String -> Int") {
    let _ = substrings.map { Int(String($0))! }
}

name                       time           std        iterations
---------------------------------------------------------------
Substring -> Int           6714750.000 ns ±  11.37 %        188
Substring -> String -> Int  450150.000 ns ±  11.97 %       2755

Theoretically, the Int initializer should just read the (sub)string's memory and won't modify it in any way. So there shouldn't be any additional memory tasks. I checked the source code of IntegerParser.swift but couldn't find any information.

Couldn't understand why it happens. Any thoughts would be appreciated.

1 Like
2

I'm going to guess that you're measuring performance of a debug build. Here's what I get running your code with a release build:

name                       time          std        iterations
--------------------------------------------------------------
Substring -> Int           278842.000 ns ±   9.61 %       4875
Substring -> String -> Int 462670.500 ns ±   5.66 %       2976
3

I ran it on Xcode 12.5.1 and CLI (swift 5.4.2), both in release mode, and got the same result indicating the latter one is faster. Strange.

4

Ah, you're using Swift 5.4! The code you're reading in IntegerParsing.swift was rewritten (by me) in March, so you'll need macOS 12/Swift 5.5 to see the results I posted. The previous implementation was indeed over 10 times slower for Substring, one of the reasons I rewrote it.

My main preoccupation during the rewrite was making sure code size didn't regress; improving the performance of Substring was so obviously happening that I neglected to include a benchmark. However, it'd probably be wise to add one to the standard library repository so that future revisions don't cause an unintended regression--that's a good starter issue I think. I'm happy to point you in the right direction if you're interested; if not, perhaps you could file a starter bug at bugs.swift.org?

4 Likes
5

Aha, yes! I ran it on Xcode13-beta and got the correct result, much faster. Great job!

Pretty sure I have the interest to contribute to the standard library, but not sure what you're suggesting to me. Are you saying post an issue to the repository that suggests adding a benchmark? Or even further, to write some benchmark code then PR…?

6

However much you're comfortable with! Issues are posted at bugs.swift.org, not on GitHub, but if you have a little experience with git (or even if you don't), creating a PR to add a benchmark is a pretty nice way to dip your toes into contributing to the standard library--the biggest barrier will be setting up the project to build locally, and the instructions are now pretty good.

7

Okay. I'm familiar with git but have no experience about contributing to Swift (nor even compiling it). Took a look at Contributing and thought there'll be a long way to go but I'm glad to try it.

Thanks!

4 Likes