Restrict not honored for non-intel compilers

[rfbird]

Currently when using the restrict memory trait, KOKKOS_RESTRICT is defined as empty when KOKKOS_COMPILER_INTEL is not found (i.e when using GCC).

I propose that despite other compilers not having great restrict behavior, we change the code to generate the restrict keyword any way and let the compiler have the chance of making the optimization (better than just not generating it)

Impending PR will implement a change which always defines KOKKOS_RESTRICT as __restrict__ regardless of compiler

[crtrott]

It can't be independent of compiler, since not all compilers know restrict. We need an enumerated list of compilers who know it.

[rfbird]

Ah, my mistake. I was thinking that restrict is standard C so all compilers know it but most will ignore it.

Out of interest which compilers don't know it?

I'll update the PR to just do intel and GCC as those are the only ones I've personally tested

[guangyechen]

I found that nvcc, ibm should also use __restrict__, as well as gcc and intel.
See
https://devblogs.nvidia.com/cuda-pro-tip-optimize-pointer-aliasing/
https://www.ibm.com/support/knowledgecenter/en/SSLTBW_2.3.0/com.ibm.zos.v2r3.cbclx01/restrict_type_qualifier.htm

[ndellingwood]

Cross-referencing PR #1923

[dhollman]

The fix for this is now in master, so I'm closing this one.