static
void cilk_for_recursive(count_t low, count_t high,
F body, void *data, int grain,
__cilkrts_worker *w,
__cilkrts_pedigree *loop_root_pedigree)
{
tail_recurse:
// Cilkscreen should not report this call in a stack trace
// This needs to be done everytime the worker resumes
NOTIFY_ZC_INTRINSIC((char *)"cilkscreen_hide_call", 0);
count_t count = high - low;
// Invariant: count > 0, grain >= 1
if (count > grain)
{
// Invariant: count >= 2
count_t mid = low + count / 2;
// The worker is valid only until the first spawn and is expensive to
// retrieve (using '__cilkrts_get_tls_worker') after the spawn. The
// '__cilkrts_stack_frame' is more stable, but isn't initialized until
// the first spawn. Thus, we want to grab the address of the
// '__cilkrts_stack_frame' after it is initialized but before the
// spawn detaches. The only place we can do that is within the
// argument list of the spawned function, hence the call to
// capture_spawn_arg_stack_frame().
__cilkrts_stack_frame *sf;
#if defined(__GNUC__) && ! defined(__INTEL_COMPILER) && ! defined(__clang__)
// The current version of gcc initializes the sf structure eagerly.
// We can take advantage of this fact to avoid calling
// `capture_spawn_arg_stack_frame` when compiling with gcc.
// Remove this if the "shrink-wrap" optimization is implemented.
sf = w->current_stack_frame;
_Cilk_spawn cilk_for_recursive(low, mid, body, data, grain, w,
loop_root_pedigree);
#else
_Cilk_spawn cilk_for_recursive(low, mid, body, data, grain,
capture_spawn_arg_stack_frame(sf, w),
loop_root_pedigree);
#endif
w = sf->worker;
low = mid;
goto tail_recurse;
}
// Call the cilk_for loop body lambda function passed in by the compiler to
// execute one grain
call_cilk_for_loop_body(low, high, body, data, w, loop_root_pedigree);
}
static
void cilk_for_recursive(count_t low, count_t high,
F body, void *data, int grain,
__cilkrts_worker *w,
__cilkrts_pedigree *loop_root_pedigree)
{
tail_recurse:
// Cilkscreen should not report this call in a stack trace
// This needs to be done everytime the worker resumes
NOTIFY_ZC_INTRINSIC((char *)"cilkscreen_hide_call", 0);
count_t count = high - low;
// Invariant: count > 0, grain >= 1
if (count > grain)
{
// Invariant: count >= 2
count_t mid = low + count / 2;
// The worker is valid only until the first spawn and is expensive to
// retrieve (using '__cilkrts_get_tls_worker') after the spawn. The
// '__cilkrts_stack_frame' is more stable, but isn't initialized until
// the first spawn. Thus, we want to grab the address of the
// '__cilkrts_stack_frame' after it is initialized but before the
// spawn detaches. The only place we can do that is within the
// argument list of the spawned function, hence the call to
// capture_spawn_arg_stack_frame().
__cilkrts_stack_frame *sf;
_Cilk_spawn cilk_for_recursive(low, mid, body, data, grain,
capture_spawn_arg_stack_frame(sf, w),
loop_root_pedigree);
w = sf->worker;
low = mid;
goto tail_recurse;
}
// Call the cilk_for loop body lambda function passed in by the compiler to
// execute one grain
call_cilk_for_loop_body(low, high, body, data, w, loop_root_pedigree);
}
