On Thu, Sep 17, 2026 at 10:39:20AM +1000, Kugan Vivekanandarajah wrote: > Bucket normalized perf ranges, branches, and address counts onto containing > ELF functions. Functions are selected using grouped profile mass before > expensive DWARF symbolization begins.
This seems very complicated. The symbolization cannot just be completely lazy when processing samples? (perf-gcov does it this way and it seems to work fine with some caching). Or is the idea to not dwarf process rarely used functions at all? It's a bit risky to completely ignore cold stuff because it might be still important when the workload changes. In this case the profile is not very good, but usually any profile is better than nothing. -Andi
