CCLS doesn't scale well to much higher thread counts

[lanza]

My work machine is an AMD Epyc 9634 with 166 threads. While indexing an LLVM build with clang and mlir enabled and ~10k items to be indexed, ccls often hovers between 1300% and 4000% CPU usage with index.threads set to either 0 or 166.

Any idea what the bottleneck in CCLS is?