coverage: 91.279% (+0.005%) from 91.274%
when pulling 5c84388 on Rohde-Schwarz:rene/simd_load_store
into 1499274 on randombit:master.

GCC 13 is doing an equally good job.

In my experience GCC only unrolls loops with a constexpr or equivalent loop count with -O3 - with -O2 or lower it fails to unroll. That's fine, cause we use -O3 ... except that most Linux distros compile us with their "approved" set of compiler flags which uses -O2 😞

with -O3 - with -O2 or lower it fails to unroll.

We could force the unrolling by using an index sequence and fold expressions. We're using this trick already for the bswap fallback implementation and IIRC compilers even recognized this as an actual bswap.

Wrapping that into an "okay-to-use" template, doesn't even lead to unreadable code, and unrolls regardless of the optimization settings (godbolt.org).

template <size_t begin, size_t end, std::invocable<size_t> FnT>
    requires(end >= begin)
constexpr void unrolled_for(FnT&& fn) {
    [&]<size_t... indices>(std::index_sequence<indices...>) {
        (fn(indices + begin), ...);
    }(std::make_index_sequence<end - begin>());
}

int main() {
  unrolled_for<2, 10>([](size_t i) {std::cout << i << '\n'; });
}

Frankly, I'm much more concerned about MSVC here. And mostly because the mentioned slowdown seemed to come from the poor optimization of the load/store abstractions. I toyed with its flags a bit but no luck. 😞

CI failures are relevant. Easiest fix is probably to move the test to test_simd.cpp

I tried adding the same load/store helpers to SIMD_8x32 and SIMD_16x32. Though, this isn't as straight-forward due to the inline ISA annotations. Namely, the load_any wrappers aren't marked with those ISA flags and therefore cannot inline the implementations. I'll leave this as future work here.