Vectorize minmax_element.

[hiraditya]

As per #84663 it isn't vectorized yet.
The microsoft STL directly uses avx2 instructions microsoft/STL#4659 for this.

Testcase: https://godbolt.org/z/PYabMjxa3

[VedantParanjape]

I want to work on this issue.

[wsehjk]

Hi, @hiraditya. I have been trying to work on this and have gone through the MSVC/STL minmax_element. But I have some questions. Vectorize optimization only works for the continuous contianer, e.g vector, array, T[] right? I think we shoud check whether the inputiterator points to the continuous range. But I didn't find this check in the std::mismatch optimization

[hiraditya]

I think we shoud check whether the inputiterator points to the continuous range. But I didn't find this check in the std::mismatch optimization

agreed. please create an issue with that one if you don't mind.

[mordante]

As per #84663 it isn't vectorized yet. The microsoft STL directly uses avx2 instructions microsoft/STL#4659 for this.

AFAIK MSVC STL only supports amd64 (AVX2) and ARM (NEON). I don't know whether all our supported platforms have vector instructions. If they do not all support AVX2. So just using AVX will not work. Clang also has builtin vector types that are platform agnostic. This is what we use in the experimental SIMD library.

I think it would be a good idea to discuss the approach before taking it.

[hiraditya]

maybe we should vectorize it in llvm then?

[wsehjk]

simd_utils.h has used the builtin vector types and provides some convinence.