The "5-7%" improvement is in terms of build time, but no mention is made of any potential performance impact to users at runtime.
The reference blog entry that the article refers to has further explanation - see https://ubuntu.com//blog/profile-guided-optimization-a-case-...
Summary (for building OpenSSL, Emacs, GDB, and Python) - "QEMU will be profiled only when building OpenSSL, and a new QEMU will be built using profiling data from this build"
The 5-7% improvement is seen when using QEMU on an x64 CPU to emulate a RISC-V CPU to build the 4 previously mentioned software projects.
So if you're building software that uses C, you'll probably see 5-7% improvement in RISC-V build times on an AMD x64 CPU.
I recall that Meta's BOLT optimizer can extract further reduction by optimizing software that has already been through PGO/FDO. ...assuming your workload matches the profiled workload. see https://github.com/llvm/llvm-project/blob/main/bolt/docs/Opt...