-
Notifications
You must be signed in to change notification settings - Fork 1.4k
Add %p to LLVM_PROFILE_FILE pattern when running tests with coverage #8894
New issue
Have a question about this project? # for a free GitHub account to open an issue and contact its maintainers and the community.
By clicking “#”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.
Already on GitHub? # to your account
Add %p to LLVM_PROFILE_FILE pattern when running tests with coverage #8894
Conversation
@swift-ci please test |
Thanks Si! Can you add some simple tests here to make sure the attributes in the path resolve correctly? |
@dschaefer2 sure thing. I couldn't see any tests that check for the coverage output, but I maybe missed them? Could you point me at them if they exist or, if they don't, suggest the best place to add the tests (don't super know my way around in SwiftPM). The PR description shows the impact of running with %p with an out of tree workaround. |
@dschaefer2 I added a test in 5290b2b. I couldn't see anywhere where there were already tests for how the coverage files got exported, merged, or validated, so I decided to put them in integration tests. |
Nice. Thanks! |
@swift-ci please test |
@swift-ci please test windows |
@dschaefer2 I notice the Linux jobs are still running with Swift 6.1, is that expected?
— https://ci.swift.org/job/swift-package-manager-self-hosted-Linux-smoke-test/6515/consoleText I will update the test to either skip on pre-6.2 (because it uses exit tests) or to try and validate some other way. But just checking that's what we expect. I have validated locally that they do pass with the 6.2 nightly Docker image. |
@swift-ci please test |
@simonjbeaumont : This is currently expected. The self-hosted pipeline run using a released version of the toolchain, while the "Some Test" execute, what I cal, a pseudo-toolchain build - that is, executes the |
@swift-ci please test windows |
Windows CI appears to be upset even getting going.
|
@swift-ci test self hosted windows |
@@ -58,4 +58,15 @@ extension Trait where Self == Testing.ConditionTrait { | |||
#endif | |||
} | |||
} | |||
|
|||
/// Skip test if compiler is older than 6.2. |
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
praise: Thanks for creating a trait for this.
let coverage = try JSONDecoder().decode(Coverage.self, from: Data(coverageJSON.contents)) | ||
|
||
// Check for 100% coverage for Subject.swift, which should happen because the per-PID files got merged. | ||
let subjectCoverage = coverage.data.first?.files.first(where: { $0.filename.hasSuffix("Subject.swift") }) |
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
suggestion: use #require()
Swift Testing api to ensure a value is set.
e.g.:
let firstCoverageData = try #require(coverage.data.first)
let subjectCoverage = try #require(firstCoverageData.files.first(where: { $0.filename.hasSuffix("Subject.swift") }))
#expect(subjectCoverage.summary.functions.count == 2)
#expect(subjectCoverage.summary.functions.covered == 2)
#expect(subjectCoverage.summary.functions.percent == 100)
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
Yeah, we could do that. If the values not there it'll result in the #expect
calls comparing against nil
so they'll fail gracefully enough.
Do you want me to make this change before we land the PR?
|
||
// Check for 100% coverage for Subject.swift, which should happen because the per-PID files got merged. | ||
let subjectCoverage = coverage.data.first?.files.first(where: { $0.filename.hasSuffix("Subject.swift") }) | ||
#expect(subjectCoverage?.summary.functions.count == 2) |
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
thought: I'm not certain we should be validating the contents of the coverage data in this test as SwiftPM has no control over this.
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
I actually think SwiftPM does control this. It's SwiftPM that takes the LLVM raw profile data, in .profraw
files, merges it into a .profdata
file, and then creates a JSON file to represent the coverage.
The JSON file has a known schema which is API so we can rely on it and the test has been constructed above specifically so we can test that there's 100% coverage of a file, when split across two tests that run in separate processes.
This fixes #8893 |
…wiftlang#8894) ## Motivation The current setting for `LLVM_PROFILE_PATH`, used for code coverage, leads to corrupt profile data when tests are run in parallel or when writing "exit tests" with Swift Testing. This also results in the `swift test --enable-code-coverage` command to fail. The `LLVM_PROFILE_PATH` environment variable is used by the runtime to write raw profile files, which are then processed when the test command finishes to produce the coverage results as JSON. The variable supports several pattern variables[^1], including `%Nm`, which is currently set, and is documented to create a pool of files that the runtime will handle synchronisation of. This is fine for parallelism within the process but will not work across different processes. SwiftPM uses multiple invocations of the same binary for parallel testing and users may also fork processes within their tests, which is now a required workflow when using _exit tests_ with Swift Testing, which will fork the process internally. Furthermore, the current setting for this variable uses only `%m` (which implies `N=1`), which makes it even more likely that processes will stomp over each other when writing the raw profile data. We can see a discussion of this happening in practice in swiftlang#8893. The variable also supports `%p`[^1], which will expand to produce a per-process path for the raw profile, which is probably what we want here, since Swift PM is combining all the profiles in the configured directory. ## Modifications Add %p to LLVM_PROFILE_FILE pattern when running tests with coverage. ## Result - Running tests write coverage raw profile data to their own per-process file pool. - Running tests in parallel with code coverage no longer risks corrupting coverage data. - Running exit tests no longer risks corrupting coverage data. - Fixes swiftlang#8893. --- ## Appendix: Demonstrating the merging of per-process profiles ```swift // file: ReproTests.swift import Testing import struct Foundation.URL #if canImport(Darwin) import Darwin #elseif canImport(Glibc) import Glibc #endif @suite(.serialized) struct Suite { static func updateLLVMProfilePath() { let key = "LLVM_PROFILE_FILE" let profrawExtension = "profraw" guard let previousValueCString = getenv(key) else { return } let previousValue = String(cString: previousValueCString) let previousPath = URL(filePath: previousValue) guard previousPath.pathExtension == profrawExtension else { return } guard !previousPath.lastPathComponent.contains("%p") else { return } let newPath = previousPath.deletingPathExtension().appendingPathExtension("%p").appendingPathExtension(profrawExtension) let newValue = newPath.path(percentEncoded: false) print("Replacing \(key)=\(previousValue) with \(key)=\(newValue)") setenv(key, newValue, 1) } @test func testA() async { Self.updateLLVMProfilePath() await #expect(processExitsWith: .success) { Subject.a() } } @test func testB() async { Self.updateLLVMProfilePath() await #expect(processExitsWith: .success) { Subject.b() } } } ``` ```swift // file: Subject.swift struct Subject { static func a() { _ = "a" } static func b() { _ = "a" } } ``` Running with just one test results in one per-process profile and 50% coverage, as expected. ```console % swift test --enable-code-coverage --filter Suite.testa ... ◇ Test run started. ↳ Testing Library Version: 6.2 (9ebfc4ebbb2840d) ↳ Target Platform: aarch64-unknown-linux-gnu ◇ Suite Suite started. ◇ Test testa() started. Replacing LLVM_PROFILE_FILE=/pwd/.build/aarch64-unknown-linux-gnu/debug/codecov/Swift Testing%m.profraw with LLVM_PROFILE_FILE=/pwd/.build/aarch64-unknown-linux-gnu/debug/codecov/Swift Testing%m.%p.profraw ✔ Test testa() passed after 0.018 seconds. ✔ Suite Suite passed after 0.018 seconds. ✔ Test run with 1 test in 1 suite passed after 0.018 seconds. ``` ```console % ls -1 .build/debug/codecov/ default.profdata repro-exit-tests-coverage-corruption.json 'Swift Testing12847901981426048528_0.15828.profraw' 'Swift Testing12847901981426048528_0.profraw' XCTest12847901981426048528_0.profraw % cat .build/debug/codecov/repro-exit-tests-coverage-corruption.json | jq '.data[].files[] | select(.filename == "/pwd/Tests/ReproTests/Subject.swift").summary.functions' { "count": 2, "covered": 1, "percent": 50 } ``` Running the other test also results in one per-process profile and 50% coverage, as expected. ```console % swift test --enable-code-coverage --filter Suite.testb ... ◇ Test run started. ↳ Testing Library Version: 6.2 (9ebfc4ebbb2840d) ↳ Target Platform: aarch64-unknown-linux-gnu ◇ Suite Suite started. ◇ Test testb() started. Replacing LLVM_PROFILE_FILE=/pwd/.build/aarch64-unknown-linux-gnu/debug/codecov/Swift Testing%m.profraw with LLVM_PROFILE_FILE=/pwd/.build/aarch64-unknown-linux-gnu/debug/codecov/Swift Testing%m.%p.profraw ✔ Test testb() passed after 0.017 seconds. ✔ Suite Suite passed after 0.017 seconds. ✔ Test run with 1 test in 1 suite passed after 0.017 seconds. ``` ```console % ls -1 .build/debug/codecov/ default.profdata repro-exit-tests-coverage-corruption.json 'Swift Testing12847901981426048528_0.15905.profraw' 'Swift Testing12847901981426048528_0.profraw' XCTest12847901981426048528_0.profraw % cat .build/debug/codecov/repro-exit-tests-coverage-corruption.json | jq '.data[].files[] | select(.filename == "/pwd/Tests/ReproTests/Subject.swift").summary.functions' { "count": 2, "covered": 1, "percent": 50 } ``` Running both tests results in two per-process profile and 100% coverage, after merge. ```console % swift test --enable-code-coverage --filter Suite.testa --filter Suite.testb ... ◇ Test run started. ↳ Testing Library Version: 6.2 (9ebfc4ebbb2840d) ↳ Target Platform: aarch64-unknown-linux-gnu ◇ Suite Suite started. ◇ Test testa() started. Replacing LLVM_PROFILE_FILE=/pwd/.build/aarch64-unknown-linux-gnu/debug/codecov/Swift Testing%m.profraw with LLVM_PROFILE_FILE=/pwd/.build/aarch64-unknown-linux-gnu/debug/codecov/Swift Testing%m.%p.profraw ✔ Test testa() passed after 0.016 seconds. ◇ Test testb() started. ✔ Test testb() passed after 0.015 seconds. ✔ Suite Suite passed after 0.033 seconds. ✔ Test run with 2 tests in 1 suite passed after 0.033 seconds. ``` ```console % ls -1 .build/debug/codecov/ default.profdata repro-exit-tests-coverage-corruption.json 'Swift Testing12847901981426048528_0.15981.profraw' 'Swift Testing12847901981426048528_0.15988.profraw' 'Swift Testing12847901981426048528_0.profraw' XCTest12847901981426048528_0.profraw % cat .build/debug/codecov/repro-exit-tests-coverage-corruption.json | jq '.data[].files[] | select(.filename == "/pwd/Tests/ReproTests/Subject.swift").summary.functions' { "count": 2, "covered": 2, "percent": 100 } ``` [^1]: https://clang.llvm.org/docs/SourceBasedCodeCoverage.html#running-the-instrumented-program (cherry picked from commit 5e566d4)
Motivation
The current setting for
LLVM_PROFILE_PATH
, used for code coverage, leads to corrupt profile data when tests are run in parallel or when writing "exit tests" with Swift Testing. This also results in theswift test --enable-code-coverage
command to fail.The
LLVM_PROFILE_PATH
environment variable is used by the runtime to write raw profile files, which are then processed when the test command finishes to produce the coverage results as JSON. The variable supports several pattern variables1, including%Nm
, which is currently set, and is documented to create a pool of files that the runtime will handle synchronisation of. This is fine for parallelism within the process but will not work across different processes. SwiftPM uses multiple invocations of the same binary for parallel testing and users may also fork processes within their tests, which is now a required workflow when using exit tests with Swift Testing, which will fork the process internally. Furthermore, the current setting for this variable uses only%m
(which impliesN=1
), which makes it even more likely that processes will stomp over each other when writing the raw profile data.We can see a discussion of this happening in practice in #8893.
The variable also supports
%p
1, which will expand to produce a per-process path for the raw profile, which is probably what we want here, since Swift PM is combining all the profiles in the configured directory.Modifications
Add %p to LLVM_PROFILE_FILE pattern when running tests with coverage.
Result
.serialized
#8893.Appendix: Demonstrating the merging of per-process profiles
Running with just one test results in one per-process profile and 50% coverage, as expected.
Running the other test also results in one per-process profile and 50% coverage, as expected.
Running both tests results in two per-process profile and 100% coverage, after merge.
Footnotes
https://clang.llvm.org/docs/SourceBasedCodeCoverage.html#running-the-instrumented-program ↩ ↩2