Failing to build for Android in SPI

[grynspan]

Looks like Swift Testing is failing to build for Android on Swift Package Index's site: https://swiftpackageindex.com/builds/875FEA40-0F76-47AF-A7E3-54A39616BA9C

/host/spi-builder-workspace/Sources/Testing/SourceAttribution/Backtrace.swift:70:56: error: cannot convert value of type 'UnsafeMutablePointer<UnsafeMutableRawPointer>' to expected argument type 'UnsafeMutablePointer<UnsafeMutableRawPointer?>'
 68 | #elseif os(Android)
 69 |       initializedCount = addresses.withMemoryRebound(to: UnsafeMutableRawPointer.self) { addresses in
 70 |         .init(clamping: backtrace(addresses.baseAddress!, .init(clamping: addresses.count)))
    |                                                        |- error: cannot convert value of type 'UnsafeMutablePointer<UnsafeMutableRawPointer>' to expected argument type 'UnsafeMutablePointer<UnsafeMutableRawPointer?>'
    |                                                        `- note: arguments to generic parameter 'Pointee' ('UnsafeMutableRawPointer' and 'UnsafeMutableRawPointer?') are expected to be equal
 71 |       }
 72 | #elseif os(Linux) || os(FreeBSD)

We have special-casing to convert the pointers here to remove optionality because, at one point, the signature of backtrace() was annotated such that the pointers were non-optional. This has either changed or is inconsistent across different Android versions.

[grynspan]

See #653. @compnerd I believe we made this change for Android because you saw it failing to build when using the same code as Linux?

[grynspan]

Likely root cause: https://android.googlesource.com/platform/bionic/+/e951acd2caf26a820a1a89647c66ea10f2852b16%5E%21/#F0

[compnerd]

Right, the API differences between bionic and glibc IIRC. This is currently building fine (recent build at https://github.com/thebrowsercompany/swift-build/actions/runs/15399505072/job/43333535153). I think that this is likely an issue with SPI.

[grynspan]

@finestructure FYI we're running into a build failure on your end here.

That said: __INTRODUCED_IN(33) probably means we shouldn't use it or should dynamically link to it.

[finagolfin]

The issue is that the SPI is currently building against API 24 using my Android SDK, or at least something based on it, but as you note, this backtrace() API was not added to Bionic till API 33. To keep libTesting building against older Android APIs, I add a backported version of libexecinfo from Bionic to my SDK, taken from the Termux repository for native Android packages, termux/termux-packages@ac6b027b5, which doesn't have the recently added nullability annotations in Bionic.

I will look into adding those annotations to my Android SDK and see if it fixes this issue.

[finestructure]

Yes, we're building with --swift-sdk aarch64-unknown-linux-android24 as advised by @marcprux. I suspect if we were to lift this to 33 we'd break other packages?

What we could do is add an Android SDK parameter to .spi.yml (defaulting to 24) that allows an override and we prepare SDK images per SDK.

I'd be worried a little about how to maintain that though. That's potentially a lot of images 24...33 (what's the max?) and something new we'd need to track and update, potentially quite frequently.

What'd be the best way forward here?