[compiler-rt] [llvm] [clang] [Profile] Refactor profile correlation. (PR #70856)

[Petr Hosek via cfe-commits]

petrhosek wrote:

> > What I think would be a better design is to just omit the respective sections altogether when -debug-info-correlate/-profile-correlate= is enabled.
> 
> I don't understand how this solves the problem you described above. Can you elaborate a bit more? We still need to decide if we need to omit data and names sections at runtime right?

If you omit the data and names sections from the object files, then `__llvm_profile_end_data() - __llvm_profile_begin_data()` and `__llvm_profile_end_names() - __llvm_profile_begin_names()` are both `0` and there's no need for [special casing this in the runtime](https://github.com/llvm/llvm-project/blob/24b11ba24da3e65f718391ccc85d4d22a081e893/compiler-rt/lib/profile/InstrProfilingBuffer.c#L59). When linking together objects complied with different flags, sizes of those sections would correspond only to object files compiled without correlation. For example, if linking together `a.o` compiled with debug info correlation and `b.o` compiled without, the profile would only contain data and names from `b.o`.

This approach would require slightly more complex logic in the correlator. Currently, we either read the data and names from the profile or from DWARF (depending on the flag), there's no in-between state. With the approach I'm proposing, we would need to combine data from different sources: read the data and names in the profile and combine them with the data and names in the DWARF and/or the unstripped file.

https://github.com/llvm/llvm-project/pull/70856