Consider enhancing sampling data for use in GCC/LLVM AutoFDO/PGO workflows

[cpaelzer]

Hi,
as pre-discussed with @brancz this is about enhancing the sampling data with mechanisms to provide better data for AutoFDO/PGO workflows of GCC/LLVM toolchains.

I'm aware that the details of this are still in flux, hence I'm not outlining the topics and papers we discussed as it might be something completely different by tomorrow. Or on the other hand, some experiments might work, but not fit architecturally into the existing profiler (and hence would need wider changes in the ecosystem to become useful).

But while in flux, I wanted to file this for the community to have something to track any related improvement, and for you it is a place you might update with the new insights whenever you are starting or have completed any of your experiments in that regard.

As I see it whatever you try will go one of two ways:

• Any approach that failed, would at least be visible and documented here
• Any approach that worked be mentioned here after a POC and then become a reference to a proper issue tracking the addition of the feature.