AutoFDO profile capture and apply

AutoFDO is sample-based profile-guided optimization. It captures perf (or RISC-V PMU equivalent) samples on a representative workload, converts to LLVM's sample profile format, and feeds the result back to clang via -fprofile-sample-use.

Measured uplift on similar pipelines:

• AutoFDO alone: ~10.5% geomean on warehouse-scale, ~85% of traditional FDO.
• AutoFDO + Propeller stacked: ~10% throughput uplift in Linux kernel builds.

This harness produces the AutoFDO .prof artifact consumed by autofdo-propeller-bolt.md.

Scripts

• capture.sh — wraps perf record + create_llvm_prof to produce a sample profile from a workload run.
• apply.sh — feeds the profile into a clang rebuild.

Status

The recipe is committed. End-to-end capture is BLOCKED until:

1. LLVM stage-2 toolchain is built (scripts/build_llvm_riscv.sh).
2. A real RISC-V target is available for perf record (QEMU virt / Verilator / FPGA / silicon).
3. create_llvm_prof is installed (LLVM ships it as llvm-profdata for sample profiles plus the autofdo project at https://github.com/google/autofdo).