I saw a nondeterministic test failure of `AsyncUtilsTests.testWithTimeoutEscalatesPriority` and I believe this was the cause.
Also refactor a few test helper functions on the way.
The original idea behind this test was that only opening `Crash.swift` would retrieve build settings for `Crash.swift`. But that wasn’t actually true: Scanning for tests also retrieved build settings (to get the test’s module name). This caused test failures if we set `SOURCEKIT_LSP_TEST_TIMEOUT: 10`, like we do in `settings.json`.
Improve the test to actually check that the BSP server has crashed and increase timeout to be bigger than the 10s restart delay.
When we receive build settings after hitting the timeout, we call `fileBuildSettingsChanged` on the delegate, which should cause the document to get re-opened in sourcekitd and diagnostics to get refreshed.
rdar://136332685
If a BSP server crashes, we should try to relaunch it. Use a similar relaunch logic as the one we use for `clangd`: Try restarting immediately unless the BSP server has crashed within the last 30 seconds, in which case we delay the restart by 10s.
rdar://136311009
Fixes#1686
The interaction to an out-of-process BSP server still went through the `BuildServerBuildSystem`, which doesn’t forward all messages to the build system and uses the old push-based model for build settings.
If we discover that the BSP server supports the new pull-based build settings model, we now forward all methods to it directly, without going through `BuiltInBuildSystemAdapter`, which has been renamed to `LegacyBuildServerBuildSystem`.
rdar://136106323
rdar://127606323
rdar://126493405
Fixes#1226Fixes#1173
When you had a package at `/pkg_real` and a symlink `/pkg` pointing to `/pkg_real`, then opened a workspace at `/pkg`, we wouldn’t get any build settings for any of the files. This was masked in tests because they still called `SwiftPMBuildSystem.targets(for:)`, which handled symlink resolution but wasn’t called in production.
Handle symlink resolution `BuildSystemManager`, remove `SwiftPMBuildSystem.targets(for:)` and its related members/methods, and migrate the tests to go through `BuildSystemManager`, which is what production code does. A nice side effect of this is that the tests will log the requests sent to the build system.
Xcode 16 with Swift 6 has been released, we can drop support for building and testing SourceKit-LSP using a Swift 5.10 toolchain. This allows us to remove a number of workarounds.
`buildTarget/inverseSources` is not required to be implemented by BSP servers and we have almost all information needed for it in `BuildSystemManager`.
This also makes sure that `buildTarget/sources` and `buildTarget/inverseSources` actually match each other. Before this change, we had source files like `Package.swift` for which we returned a target from `buildTarget/inverseSources` but that weren’t part of that target’s sources retrieved using `buildTarget/sources`.
The implementation of which file’s dependencies have been updated is common across all build systems and thus build systems shouldn’t need to implement this logic. This also allows us to remove `BuildSystemDelegate`.
This finalizes the move of `BuiltInBuildSystem` creation into `BuiltInBuildSystemAdapter` and means that we can set the message handler of the `BuiltInBuildSystem` during initialization instead of using a setter method.
This allows us to create the build system from a `BuiltInBuildSystemAdapter` when it receives an `InitializeRequest`, which will be done in a follow-up commit.
We were making the initial `generateBuildGraph` call in `SourceKitLSPServer` from a `Task`. This means that `generateBuildGraph` could be executed after `waitForUpToDateBuildGraph` was called by `SemanticIndexManager`. Thus `waitForUpToDateBuildGraph` returned immediately and no files were background indexed.
Make `generateBuildGraph` immediately schedule a package reload for SwiftPM build systems instead of doing the hop through a `Task`, fixing the race condition.
rdar://135551812
We currently load the entire package before generating a `SwiftPMBuildSystem`. That means that the initialize request to SourceKit-LSP is blocked until the package has been loaded, preventing us from offering any sort of functionality, including syntactic functionality like formatting.
Decouple build system creation and build graph generation (aka. package loading for SwiftPM). We can operate with fallback build settings until the build graph has been loaded and reopen the document once the proper build settings are available.
rdar://126644596
Using `SwiftSDK.deriveTargetSwiftSDK`, which is the same method that SwiftPM's own CLI tools use to determine the SDK from passed-in info (target `--triple`, `--swift-sdk`, and host sdk). This allows us to better uphold the contract in the [Configuration File](d11c101ce2/Documentation/Configuration%20File.md (structure)) docs, namely that the `swiftSDK` param is "Equivalent to SwiftPM's `--swift-sdk` option" and similarly for `triple`.
As concrete examples of where (AFAICT) the current implementation diverges:
- Passing a `--triple` of `wasm32-unknown-wasi` to `swift-build` will use the toolchain-integrated Wasm SDK if one exists. Passing the same value to sourcekit-lsp does not do this.
- Perhaps more relevant: after landing https://github.com/swiftlang/swift-package-manager/pull/6828, this change will make it so that building for iOS is as simple as setting `"triple": "arm64-apple-ios"` in the config! Currently, it's necessary to set C/Swift flags and hardcode the sysroot. Should close https://github.com/swiftlang/sourcekit-lsp/issues/1587.
This PR depends on:
- https://github.com/swiftlang/swift-package-manager/pull/7925
