[PATCH] _Cilk_for for C and C++

Jakub Jelinek jakub@redhat.com
Fri Jan 24 19:41:00 GMT 2014

On Thu, Jan 23, 2014 at 04:38:53PM +0000, Iyer, Balaji V wrote:
> 	This is how I started to think of it at first, but then when I thought about it ... in _Cilk_for unlike the #pragma simd's for, the for statement - not the body - (e.g. "_Cilk_for (int ii = 0; ii < 10; ii++") doesn't really do anything nor does it belong in the child function. It is really mostly used to calculate the loop count and capture step-size and starting point.
> 	The child function has its own loop that will have a step size of 1 regardless of your step size. You use the step-size to find the correct spot. Let me give you an example:
> _Cilk_for (int ii = 0; ii < 10; ii = ii  + 2)
> {
> 	Array [ii] = 5;
> }
> This is translated to the following (assume grain is something that the user input):
> data_ptr.start = 0;
> data_ptr.end = 10;
> data_ptr.step_size = 2;
> __cilkrts_cilk_for_32 (child_function, &data_ptr, (10-0)/2, grain);
> Child_function (void *data_ptr, int high, int low)
> {
> 	for (xx = low; xx < high; xx++) 
> 	 {
> 		Tmp_var = (xx * data_ptr->step_size) + data_ptr->start;
> 		// Note: if the _Cilk_for was (ii = 9; ii >= 0; ii -= 2), we would have something like this:
> 		// Tmp_var = data_ptr->end - (xx * data_ptr->step_size)
> 		// The for-loop above won't change.  
> 		Array[Tmp_var] = 5;
> 	}
> }

This isn't really much different from
#pragma omp parallel for schedule(runtime, N)
(i.e. the combined construct), when it is combined, we also don't emit a
call to GOMP_parallel but to some other function to which we pass the
number of iterations and chunk size (== grain in Cilk+ terminology), the
only (minor) difference is that for OpenMP when you handle the whole low ...
high range the child function doesn't exit, but calls a function to give it
next pari of low/high and only when that function tells it there is no
further work to do, it returns.  But, the Cilk+ case is clearly the same
thing with just implicit telling there is no further work in the current

So, I'd strongly prefer if you swap the parallel with Cilk_for, just set
the flag that the two are combined like OpenMP already has for tons of
constructs, and during expansion you just treat it together.


More information about the Gcc-patches mailing list