Fix missing ifdef HPX_SMT_PAUSE

[jnvance]

Due to the missing definition of HPX_SMT_PAUSE in compiler_fence.hpp, source file spinlock.cpp cannot be compiled with the Intel compiler.

hpx/libs/core/config/include/hpx/config/compiler_fence.hpp

Lines 22 to 26 in c320e11

	#if defined(__INTEL_COMPILER)
	#define HPX_COMPILER_FENCE __memory_barrier()
	#elif defined(_MSC_VER) && _MSC_VER >= 1310

[jenkins-cscs]

Can one of the admins verify this patch?

[hkaiser]

retest

[hkaiser]

Wouldn't it be better to actually define the macro instead of this conditional block?

[jnvance]

I agree. I guess _mm_pause would be the one for intel.

I just followed how it's currently done in other parts of the code:

hpx/libs/core/threading_base/src/execution_agent.cpp

Lines 65 to 70 in c320e11

	#if defined(HPX_SMT_PAUSE)
	else if (k < 16)
	{
	HPX_SMT_PAUSE;
	}
	#endif

hpx/libs/core/execution_base/src/this_thread.cpp

Lines 95 to 97 in c320e11

	#if defined(HPX_SMT_PAUSE)
	HPX_SMT_PAUSE;
	#else

[msimberg]

Please see https://github.com/STEllAR-GROUP/hpx/pull/5199/files#diff-94191e4d2264ae001231f8f358e44f0a855b33a427e48e41e8df445e5309fe5c as well. That defines HPX_SMT_PAUSE to be _mm_pause.

However, while doing those changes I was considering dropping support for ICC. Now that we have someone who seems to use ICC, I'd be curious to hear how hard of a requirement you have to use ICC? Do you find that ICC does a better job in whatever application you're using HPX?

[jnvance]

Then that PR should be able to fix this issue as well. The main reason for using ICC in our application (independent of HPX) is better automatic vectorization of some complex for-loops compared to GCC which leads to 2x speedup for that part of the code.

[msimberg]

Fair enough, thanks! That's a reasonable reason for using ICC. I'll see how we can keep the ICC CI going then (probably in a reduced form).