Dmitry Babokin

I think structure and function templates should go together. I've changed the title of the issue to address both. The feature is targeted for the next release.

Dropping ``const`` in ``kISPC8Permutations`` definition fixes the problem. Looks like LLVM considers assignment to ``const`` variable as undefined behavior and silently abuses it.

@pbrubaker thanks for trying clang. I think we need to have a closer look.

This comes from ``-mavx2`` used to compile C++ portion of the benchmark. Need to add warning if ``sse`` or ``avx1`` targets are passed to benchmarks.

`goto` is definitely a bad idea in this case, as logically it should apply to a single lane where it is active, while not affecting other iterations, even though which...

`unmasked` regions is a pain on GPUs, that's probably my primary concern for allowing that. We'd like to evolve the language in GPU-friendly manner.

Adding reproducer source code in case Godbolt link will expires (does it expire?): ```c unmasked uniform float dotProductSoA_Unroll(uniform float a[], uniform float b[], uniform size_t N) { varying float dot...

These are glibc functions to convert `float``float16`. I assume on macOS and Windows they will be different functions. Are they introduced by LLVM code gen? What operations are triggering these...

Could you investigate what operation are triggering generation of these function? If these are explicit converts, then we can avoid that by doing converts with our own conversion routines. If...

[Here](https://github.com/llvm/llvm-project/blob/54cc7de4bc01e6178213e4487d6ab49b809ba2b0/llvm/lib/CodeGen/TargetLoweringBase.cpp#L116) the code gen is initialized / configured with specific routine to use for certain operations. Could you check if we can change specific library function though LLVM API (i.e....