Both GCC and clang offer the option of generating code that is optimized
with CPU programming counter data by tools like Linux perf or FreeBSD
pmc(3).

With Linux perf, the recommendation is to use the -b flag to use the
Last Branch Record (LBR) to record call chains: on FreeBSD, enhance the
pmcstat(8) tool with a new -B option that generates a similar profile.

hwpmc sample

│
├── timestamp ───────→ perf-script timestamp
├── PID ──────────────→ PID
├── CPU ──────────────→ CPU
├── PMC ──────────────→ event name
├── PC ───────────────→ IP/callchain
└── process mappings ─→ MMAP2

pmcstat -B profile.perf
llvm-profgen --perfscript=profile.perf ...

If llvm-profgen accepts this and llvm-profdata show demonstrates meaningful function/inline-context sample counts, then the proposal has achieved its essential goal with a surprisingly small FreeBSD change