benchmark: it is built with /Ob1, so vector algorithm dispatcher is noticeable

[AlexGuteniev]

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

[StephanTLavavej]

We talked about this at the weekly maintainer meeting and @barcharcraz had a suggestion - build as Release, but then add compiler options (e.g. /Zi) to generate debug info.

STL/benchmarks/CMakeLists.txt

Lines 62 to 65 in 8e2d724

	set(CMAKE_BUILD_TYPE RelWithDebInfo)
	# /utf-8 affects <format>.
	add_compile_options("$<$<COMPILE_LANGUAGE:CXX>:/diagnostics:caret;/W4;/WX;/w14265;/w15038;/w15262;/utf-8>")

Charlie was additionally concerned that we shouldn't be trying to force the build type here - it would be better to provide a default that can be overridden when building the benchmarks, if a contributor temporarily wants something different. We don't immediately know the proper incantations for that.