PWR025: Consider annotating pure function with OpenMP ‘declare simd’

Issue

A SIMD version of the function can most likely be generated by the compiler.

Relevance

A loop invoking functions tends to make it difficult  for the compiler to vectorize. However, calls to some functions can be vectorized. Calls to a pure function (function whose return value depends only on function parameters) can be vectorized if the compiler is instructed to do so with a compiler pragma.

Actions

Annotate the pure function with #pragma omp declare simd. 

Notes
If the compiler manages to inline the function, then vectorization pragma is not needed. To see the performance benefit of this approach, the caller loop and called functions must reside in different compilation units.

Also, make sure OpenMP support in your compiler is enabled using compiler switches (typically -fopenmp-simd or -fopenmp).

Code example

The following loop invokes a pure function foo:

int foo(int a) { return 2 * a; }
 
void example(int *A, int n) {
   for (int i = 0; i < n; i++) {
       A[i] = foo(i);
   }
}

By adding the #pragma omp declare simd the compiler will create a vectorizable version of foo:

#pragma omp declare simd
int foo(int a) { return 2 * a; }
 
void example(int *A, int n) {
   for (int i = 0; i < n; i++) {
       A[i] = foo(i);
   }
}

References

SIMD Directives – OPENMP API Specification: Version 5.0 November 2018