Improve benchmark harness and compare Swift 6.3 performance

[gh123man]

Summary

• replace the benchmark executable with a configurable harness that supports warmups, configurable workloads, and markdown or JSON output
• add helper scripts to run the suite per toolchain and compare two benchmark reports
• document the benchmark workflow and ignore local benchmark result files

Benchmark findings

Full release benchmark runs were collected on the same machine with 1 warmup round and 5 measured rounds for each toolchain.

AsyncChannels vs Swift 6.2.3:

• 47 of 48 scenarios improved on Swift 6.3.0
• geometric-mean runtime change: about -37.7%
• only regression observed: MPMC Buffered(100) | Int at +0.92%

Largest AsyncChannels wins:

• MPSC Buffered(100) | RefData: -68.81%
• SPMC Buffered(100) | RefData: -57.98%
• MPSC Write Contention Buffered(100) | RefData: -54.79%
• SPSC | String: -52.82%
• MPSC | ValueData: -50.81%

AsyncAlgorithms also improved materially under Swift 6.3.0, with a geometric-mean runtime change of about -32.9% on the comparable scenarios.

Raw outputs

• Benchmarks/results/swift-6.2.3-20260326-185306.json
• Benchmarks/results/swift-6.3.0-20260326-191022.json

Reproduction

• BENCHMARK_NAME=swift-6.2.3 Benchmarks/run_swift_benchmarks.sh
• BENCHMARK_NAME=swift-6.3.0 SWIFT_BIN=$HOME/.swiftly/bin/swift Benchmarks/run_swift_benchmarks.sh
• Benchmarks/compare_swift_benchmarks.py <baseline.json> <candidate.json>