Reder

Probably not a big deal, if you have the time to write "A1" (not even mentioning small string optimization) You'll use `XLCell XLWorksheet::cell(uint32_t rowNumber, uint16_t columnNumber) const` everywhere else.

Side note: it has nothing to do with __builtin_prefetch. MSVC Clang accepts it, does not change results.

Incidental findings: defining `#define really_inline __declspec(noinline) ` MSVC perfomance jumps to 5.6GB/sec. Much better.

Yes, discovered this by chance. MSVC results are counter-intuitive, something's wrong in their inlining. __forceinline and inline give same results. I tried noinline to get a split of function time...

it is find_quote_mask + cmp_mask_against_input that I need to un-inline

I've been playing a while with it. It turns out that MSVC performance is worst with /Ob2 than /Ob1. With /Ob1 performance is 50% of Clang-CL at /Ob2 (and same...

Maybe I read that too quickly. I'll submit the issue to MSFT. /Ob2 worst than /Ob1 is not good news. Even though this is a corner case.

Clearly approximation is not acceptable for the task I'm pursuing. Not sure about what pandas do though. Thanks for the help, but it looks more complicated than this. I've stepped...

Is that different to USE_SYSTEM_TZ_DB (already ON) ? I have tried USE_OS_TZDB=1 but did not find any performance difference. But maybe the option is not really swallowed (not sure FetchContent_Declare...

> I don't use the CMake systems associated with this project and deeply regret their introduction. They introduce way too much complication for compiling a single source: tz.cpp. Their over...