#include <cofor.hfa>

//////////////////////////////////////////////////////////////////////////////////////////
// cofor ( uC++ COFOR )

thread co_runner {
	ssize_t low, high;
	__cofor_body_t loop_body;
};

static void ?{}( co_runner & this, ssize_t low, ssize_t high, __cofor_body_t loop_body ) {
	this.low = low;
	this.high = high;
	this.loop_body = loop_body;
}

void main( co_runner & this ) with( this ) {
	for ( ssize_t i = low; i < high; i++ )
		loop_body(i);
}

void cofor( ssize_t low, ssize_t high, __cofor_body_t loop_body ) libcfa_public {
	ssize_t range = high - low;
  if ( range <= 0 ) return;
	ssize_t nprocs = get_proc_count( *active_cluster() );
  if ( nprocs == 0 ) return;
	ssize_t threads = range < nprocs ? range : nprocs;
	ssize_t stride = range / threads + 1, extras = range % threads;
	ssize_t i = 0;
	ssize_t stride_iter = low;
	co_runner * runners[ threads ];
	for ( i; threads ) {
		runners[i] = alloc();
	}
	for ( i = 0; i < extras; i += 1, stride_iter += stride ) {
		(*runners[i]){ stride_iter, stride_iter + stride, loop_body };
	}
	stride -= 1;
	for ( ; i < threads; i += 1, stride_iter += stride ) {
		(*runners[i]){ stride_iter, stride_iter + stride, loop_body };
	}
	for ( i; threads ) {
		delete( runners[i] );
	}
}

//////////////////////////////////////////////////////////////////////////////////////////
// parallel (COBEGIN/COEND)

thread para_runner {
	parallel_stmt_t body;
	void * arg;
};

static void ?{}( para_runner & this, parallel_stmt_t body, void * arg ) { 
	this.body = body;
	this.arg = arg;
}

void main( para_runner & this ) with( this ) { body( arg ); }

void parallel( parallel_stmt_t * stmts, void ** args, size_t num ) libcfa_public {
	para_runner * runners[ num ];
	for ( i; num )
		(*(runners[i] = malloc())){ stmts[i], args[i] };
	for ( i; num )
		delete( runners[i] );
}