perfdata is a set of tools to automate the recording and collection steps of sampling profile-guided optimization (SPGO).
This software is highly experimental.
It might not work how you expect.
It might change at any time.
It might never update again.
Proceed with caution.
Use for inspiration.
Not for production.
One great feature of sampling PGO is a reduction in overhead compared to instrumented PGO.
Rather than invoking extra injected code at runtime, the kernel uses CPU features to take samples.
And indeed it is true that perf record with properly tuned --mmap-flush= adds overhead of only about 2% CPU time.
But perf record --branch-filter 'any,u' makes massive dumps of data - about 10 GiB per hour or even more with --freq=max.
Compression can save space, but of course this adds more CPU overhead.
Even better, we can use an event filter like --event=branches to reduce the amount of data being gathered and processed.
Once the raw data is recorded it has to be summarized in a format a compiler can understand.
perf scriptreplays a simulation of the recordingllvm-profgentallies up branch counts in the scriptllvm-profdatacombines multiple profiles
I made some patches to perf, llvm-profdata, and llvm-profgen which allow them to pipe data around, avoiding intermediate temporary files completely.
Together with perf-record event filtering, this brings CPU overhead down below 5% and the IO overhead down to just a few MB per hour, even when recording with --freq=max.