35% of performance going to `#available(..., *)` (`__isPlatformVersionAtLeast`)

[MahdiBM]

Motivation

I'm benchmarking and profiling my package, and apparently 35% of the time is spent in __isPlatformVersionAtLeast.

I have a bunch of #available(swiftIDNAApplePlatforms 26, *) / #available(swiftIDNAApplePlatforms 13, *)s in the code.
These are defined as:

.enableExperimentalFeature(
    "AvailabilityMacro=swiftIDNAApplePlatforms 26:macOS 26, iOS 26, tvOS 26, watchOS 26, visionOS 26"
),
.enableExperimentalFeature(
    "AvailabilityMacro=swiftIDNAApplePlatforms 13:macOS 13, iOS 16, tvOS 16, watchOS 9"
),

Proposed solution

Sounds to me implementing some caching mechanism can solve the issue?

Alternatives considered

No response

Additional information

No response

[MahdiBM]

To reproduce, clone swift-idna, checkout 1ab76dd148da31cf8b5cd0281e3cfb80ef1f3607, and add this as a new test in Tests/IDNATests, and run it with Xcode's "Profile test" functionality.

import SwiftIDNA
import Testing

@Suite
struct LocalProfilingTests {
    @available(swiftIDNAApplePlatforms 13, *)
    @Test
    func toASCII() throws {
        let idnaConfig = IDNA(configuration: .mostLax)
        for _ in 0..<500_000_000 {
            var domainName = "google.com"
            let converted: Void = try! idnaConfig.toASCII(domainName: &domainName)
            blackHole(converted)
        }
     }
}

/// A function to foil compiler optimizations that would otherwise optimize out code you want to benchmark.
@_optimize(none)  // Used after tip here: https://forums.swift.org/t/compiler-swallows-blackhole/64305/10 - see also https://github.com/apple/swift/commit/1fceeab71e79dc96f1b6f560bf745b016d7fcdcf
private func blackHole(_: some Any) {}

[tbkka]

CC: @al45tair @mikeash @Catfish-Man

[tbkka]

CC: @tshortli

[tshortli]

Thanks for the report. Your trace shows all of the overhead is under the __isPlatformVersionAtLeast function, which is what if #available(...) compiles down to when targeting Apple platforms. This function isn't part of the Swift project; it's a platform runtime function that is provided through the LLVM compiler-rt library with a vendor specific implementation provided by Apple. There isn't any specific change we can make to the Swift project to improve the performance of this, but I will ask folks at Apple responsible for this to take a look. I'm not sure what to do with this issue, though, since like I said it's not going to be resolved by a change to the Swift project. If you want to be notified about a fix, it would be better to report this through https://feedbackassistant.apple.com.

[tshortli]

When filing a feedback, you can reference rdar://162278384 as the original Radar tracking this issue.