Measure through a benchmark package of its own

[sinoru]

The performance suites move out of the test targets into Benchmarks, a package of its own on package-benchmark, so that nothing a client resolves includes the harness: the library's manifest still has no dependencies. swift test no longer measures anything, and no XCTest case is left in the package.

The benchmark package

Every case comes across, 69 of them, each with its comparison: the standard library's Mutex, DispatchSemaphore, pthread_rwlock_t, a concurrent DispatchQueue, and an actor. The pointer chase, the shared turn budget, and the pinned threads carry over from the old harness; what the XCTest meter and the corelibs clock did is the harness's job now.

• The library is measured as it ships, with no -enable-testing.
• A sample is a million turns, so the scaled output reads per turn; --scale prints the raw sample.
• The default metrics are the wall clock and instructions retired, with CPU time and context switches under contention. Allocation and retain counts are a separate pass (--metric mallocCountTotal --metric retainCount --metric releaseCount): the harness counts them by hooking every allocation, retain, and release, which put about an eighth on the clock of an asynchronous handoff (AsyncMutex on a short queue, 2,029 → 2,324 ns) and instructions of its own on every path that retains.
• Baselines: baseline update before a change and baseline compare after, in place of running each side fifteen times by hand.

Workers live on a pool of threads kept for the whole run, because the allocation counter sets up a per-thread record on a thread's first free and traps when that free is inside the thread's exit. A contended sample runs in stretches of at most a tenth of a second, every worker parked between them.

The harness runs on macOS and Linux only, so Windows, Android, and WASI are no longer measured. They still run the stress suites in release.

Tests and CI

• The XCTest performance suites and Measurement.swift are gone, with the test utilities' Atomic dependency and AddressSanitizer flag that only they used.
• Both workflows' release run is the stress suites alone (SUITES: StressTests).
• The tests build and run with --disable-xctest on every row, as the WebAssembly row alone did before: the corelibs runner traps on an empty list. Android's runner loop takes Swift Testing only.
• Run benchmarks runs the package on the macOS and Ubuntu x86_64 rows of each Swift minor and puts the table in the job summary. Nothing fails on a number. Benchmarks/** joins the path filter, and the job timeout goes from 30 to 45 minutes.
• The arm64 Windows (Windows arm64: an actor counter ends one increment short in the measurements' comparison #6) and arm64 Linux (Linux arm64: a libdispatch worker crashes on a root queue whose head is DISPATCH_OBJECT_LISTLESS while tasks take turns on an actor #11) pairs stay experimental. Their failures showed in the measurements, which neither row runs now, so each stays marked until a run without them shows it clear.

README

The Performance tables are measured again, as the median of seven runs of the whole suite, each run's figure the median of its samples. A single run was not enough: one had concurrent reads at 5.8 ns on RWLock against 4.7 on Mutex, where five reruns of that pair and the seven-run median (4.8 against 5.0) have RWLock ahead. The prose is checked against the new figures. Two sentences change: the read-mostly async figures, and "readers at once cost no more than a Mutex", where the old harness measured less. Its contended Mutex was half again slower than now, with fresh threads each sample. How to run the benchmarks, keep a baseline, and take the counts gets a section of its own.

The CHANGELOG is untouched: nothing a client receives changes.

Checked

• macOS 26.7.1, Swift 6.4: the benchmark package and the library build in release without warnings. The debug unit tests (--disable-xctest --skip StressTests) and the release stress suites (--disable-xctest --disable-testable-imports --filter StressTests) pass.
• The full benchmark suite, seven times over, with no failure and no stall.
• Not checked locally: the Android row's test step with --disable-xctest, and Run benchmarks on a runner. This PR's CI is where both are seen first.