Noticed while working #4495 . When I decided to use sized if constexpr dispatch, instead of using the same version for all element sizes, I observed significant perf degradation for small element sizes. A part of it is due to not inlining the dispatcher.

The benchmark is built with /Ob1. Looks like it is implied due to CMake RelWithDebugInfo configuration, as opposed to Release.

What are our takeaways?

I see the following options:

• Mark vector algorithms dispatchers inline, consider making other STL functions inline.
  • This helps other projects with RelWithDebugInfo to inline STL, though it would obfuscate the debugger
• Override the option in the benchmark
• Make the benchmark Release by default, instead of RelWithDebugInfo
  • I wouldn't like that. RelWithDebugInfo is convenient for profiling
• Accept the cost of dispatching as a penalty for vector algorithms that use dispatching
  • I don't think this is fair
• Use specializations instead of if constexpr
  • Throughput?
• Manually inline the dispatch like for __std_reverse_copy_trivially_copyable...
  • Copypasta