[enh] Add intel compiler support

[icfaust]

Macro definition for X86_SIMD_SORT_UNROLL_LOOP will break x86-simd-sort when compiling with icc as they define __GNUC__. This introduces a check for an Intel compiler by looking if __INTEL_COMPILER has been set, and sets the appropriate pragma (not done for __INTEL_LLVM_COMPILER due to compiling performance issues).

There is also an issue with avx2 where icc will complain that there isn't a return type for avx2_vector for 32 and 64 bit. The compile time evaluation and static assert are placed ahead of a generalized return, which shouldn't change the logic or runtime. I definitely recommend the maintainers to evaluate those changes if they make sense (should the static assert require a previous constexpr check?)

Intel LLVM C++ compilers do not define __builtin_cpu_supports("avx512fp16") which requires certain code not to be run when using icpx on sapphirerapids ISA-capable machines.

Testing uses the intel apt repository for getting the icpx compiler. The GPG key can be added, but that will add something that must be updated in the future/ requires future maintenance. Suggestions are welcome.

Meson from apt is out of date enough that it doesn't know of the icpx compiler. The pip version is up to date and is instead used in the new additional CI job.

Tasks

• Implement icc-specific changes
• Implement icpx-specific changes
• Add tests for Intel LLVM C++ compiler
• Pass tests

[icfaust]

I will reach out to compiler development about __builtin_cpu_supports("avx512fp16") issues with icpx

[icfaust]

So I had erroneously set the intel compiler check within an if statement for __GNUC__ which wouldn't actually be reached for icc and icpx. When made available to those compilers, the compile times for icpx skyrocketed both locally and in CI (to timeout). It will compile but takes significant time. I have removed loop unrolling for icpx, and have returned the new CI to the previous state. This will allow for CI to properly run, and will be investigated if it is a compiler issue on intel's side.

[icfaust]

Local testing shows that icpx/icc do not support __builtin_cpu_supports("avx512fp16"), even if icpx supports __FLT16_MAX__/ _Float16 types. ICC just doesn't support __FLT16_MAX__ at all, so isn't a problem. I will recondition the requisite check. This is a compiler problem (not even available in most recent builds).

[r-devulap]

LGTM. Thank you for adding CI as well.