[claytonwramsey]

I would argue the opposite: the problem is that our languages give us only rudimentary tools to represent data in SoA styles, which makes them difficult to apply for efficient processing. In other words, it's not SIMD's fault but rather the C-like approach to data structures. In fact, SoA is also great for non-SIMD parallel implementations because it yields improvements to memory density and cache performance.

[titzer]

This is much nicer in functional languages that have tuples. Then an array of tuples (if normalized or flattened) is both stored efficiently and ready for vectorization. You really don't want to model mathematical objects like vectors as mutable, heap-allocated objects. OO isn't the right fit here.

[mgaunard]

It's not just C, but also any object-oriented language.

Basically the most popular imperative paradigms are just ill-suited to making the most of SIMD hardware. They're designed with other goals in mind.