Given that we're talking about token concatenation in macros, I doubt that's going to show up in that much day-to-day code, so the small gains come from benchma

Given that we're talking about token concatenation in macros, I doubt that's going to show up in that much day-to-day code, so the small gains come from benchmarks using stdlib functionality that uses this feature. The bit about benchmarks using tokens sounds to me like he's talking about harness code, rather than the test subject code. You can speed up running the benchmarks without necessarily affecting the benchmark results themselves.