Performances on Windows

[JohanMabille]

We should consider the following options to improve performances on Windows:

• use FORCE_INLINE when possible
• reintroduce the const member in const_array; this gives an Internal Compiler Error, that can be removed by removing the noexcept qualifier of the methods of const_array.

[SylvainCorlay]

I think that the changes in const_array with the conditionals on MSVC can be introduced as a compiler workaround. I am more hesitant about having a XTENSOR_INLINE macro....

[JohanMabille]

This is required if we want to reduce the overhead we have on Windows (on my machine, for small containers, we are 50% slower than Eigen without it, and "only" 20% with it).

[SylvainCorlay]

Yeah, maybe in two different PRs 😒

[JohanMabille]

I agree, replacing inline with XTENSOR_INLINE everywhere should be done in a dedicated PR.

[SylvainCorlay]

Also, do we want to replace all the inline with XTENSOR_INLINE?

[wolfv]

does the const qualifier make a difference in performance?

I think XTENSOR_INLINE makes sense if we can measure it.