Use generic scheduling on Apple Silicon - #1294

Summary

Make the macOS arm64 target's CPU policy explicit: retain the Apple Silicon-compatible M1 instruction baseline, while requesting generic scheduling instead of tuning the universal artifact to a particular M generation.

This keeps the existing -arch arm64 and macOS 11 deployment target. PGO and LTO settings are unchanged.

Measurements

On an M1 Pro, I built CPython 3.14.6 with PGO and ThinLTO using LLVM 23.1.1 and compared the proposed flags against both plain -arch arm64 and explicit M1 tuning. Across five small workloads, the proposed flags were within measurement noise of plain -arch arm64; explicit M1 tuning was about 3–11% faster than the proposed generic scheduling policy. That is an intentional universal-target tradeoff, measured on one machine, not a broad hardware claim.

Validation

• LLVM 22.1.3, the currently pinned PBS toolchain, accepts the proposed flags.
• A full CPython 3.14.6 PGO + ThinLTO build using these flags succeeded with LLVM 23.1.1 in an independent build harness.
• No PBS build or test harness was run for this change; the upstream CI should validate the target configuration.