First appearance

• ARM SME2: Accelerate MatMul FP16 #457

Symptom

• cargo test / cargo build fails at the final link step on Apple Silicon with undefined symbols:
  • ___arm_tpidr2_save
  • ___arm_tpidr2_restore
  • (shown as ___arm_tpidr2_* by ld)
• This shows up most often in CI (e.g. M1 runners) but can affect any macOS build that compiles the SME2 kernel.

Root Cause

• The SME2 kernel (kernel_sme2.cpp) uses SME2 attributes/ABIs that reference __arm_tpidr2_* helper functions.
• Those helpers are provided by Apple clang’s runtime archive libclang_rt.osx.a (inside the clang resource dir), not reliably by libSystem in a way Rust picks up automatically.
• When linking through rustc, the final link invocation uses -nodefaultlibs, so the implicit “clang driver adds its runtime” behavior does not apply.
• Result:
  • The object file(s) referencing __arm_tpidr2_* are present.
  • The defining archive is never linked.
  • The final link therefore fails.

What This PR Does

• In rust/cactus-sys/build.rs, on macOS only:
  • runs clang --print-resource-dir
  • adds .../lib/darwin to the native search path
  • links libclang_rt.osx.a via
    cargo:rustc-link-lib=static=clang_rt.osx
• This makes __arm_tpidr2_* resolution deterministic under Rust’s -nodefaultlibs linking.

Why Not “Allow Undefined” (-Wl,-U)

• -Wl,-U,<sym> can make the link step succeed even if the symbol remains unresolved.
• That risks a runtime crash if the SME2 path executes on hardware where cpu_has_sme2() is true but the symbols aren’t actually resolved in the final binary image.
• This PR links the real definitions instead of deferring resolution.

Correct Conditions for SME2 to Actually Run

• Compile-time (build) condition

  • The build must compile the SME2 kernel into the library (CMake BUILD_SME2 ends up true).
  • This is driven by the build toolchain’s SME2 support (headers/flags), not strictly “build machine is M4”.

• Link-time condition (this change)

  • The final Rust link must include the clang runtime archive that defines __arm_tpidr2_* (libclang_rt.osx.a).

• Runtime condition

  • The program must run on SME2-capable hardware and the runtime gate must allow it (cpu_has_sme2() true).
  • Otherwise the non-SME2 path is used.