func main() {

	tests(0)
	tests(1)
	tests(10)
	tests(100)

	t := 4 * // buffer sizes
		(4*4 + // tests 1,2,3,4 channels
			8 + // test 5 channels
			12) * // test 6 channels
		76 // sends/recvs on a channel

	if tots != t || totr != t {
		print("tots=", tots, " totr=", totr, " sb=", t, "\n")
		os.Exit(1)
	}
	os.Exit(0)

    communication via send/recvs). In this case, users of ParallelExecute op
    must provide correct control dependencies between regions to guarantee
    correctness. Regions in ParallelExecute may include Resource ops.

    In the case where different regions include ops access the same resource,
    the users of the ParallelExecute op must provide mechanism (via send/recvs

// This can be used as the type of multi-valued call expressions.
func (t *Type) ResultsTuple() *Type { return t.funcType().resultsTuple }

// Recvs returns a slice of receiver parameters of signature type t.
// The returned slice always has length 0 or 1.
func (t *Type) Recvs() []*Field { return t.funcType().recvs() }

// Params returns a slice of regular parameters of signature type t.
func (t *Type) Params() []*Field { return t.funcType().params() }

    it always produces a valid output even when inputs are dead.
  }];

  let description = [{
    Its primary use so far is in the scheduling of recvs, where we add
    ControlTrigger nodes and use them to trigger recvs. We allow ControlTrigger
    nodes to be enabled by dead nodes.
  }];

  let arguments = (ins
    Variadic<TfeControlType>:$controlInputs
  );
  let results = (outs

      func.getSecond().clone = clone;
    }
  }

  return success();
}

// Assigns either MAXIMAL or MANUAL sharding. The MAXIMAL sharding sends/recvs
// one message from core `kShardingTpuCore` with the full tensor. MANUAL
// sharding sends/recvs one message for each core with the core's shard.
void SetOpSharding(Operation* op, bool manual_sharding) {
  xla::OpSharding sharding =

			b.WriteByte(' ')
		}
		b.WriteByte('}')

	case TFUNC:
		if verb == 'S' {
			// no leading func
		} else {
			if t.Recv() != nil {
				b.WriteString("method")
				formatParams(b, t.Recvs(), mode, visited)
				b.WriteByte(' ')
			}
			b.WriteString("func")
		}
		formatParams(b, t.Params(), mode, visited)

		switch t.NumResults() {
		case 0:
			// nothing to do

		case 1:

	// function is 3 cated structures;
	// compute their widths as side-effect.
	case TFUNCARGS:
		t1 := t.FuncArgs()
		// TODO(mdempsky): Should package abi be responsible for computing argwid?
		w = calcStructOffset(t1, t1.Recvs(), 0)
		w = calcStructOffset(t1, t1.Params(), w)
		w = RoundUp(w, int64(RegSize))
		w = calcStructOffset(t1, t1.Results(), w)
		w = RoundUp(w, int64(RegSize))
		t1.extra.(*Func).Argwid = w
		t.align = 1
	}

  // CPU tensor to GPU).
  // Setting a larger thread pool does not help with the Swift caller, as we use
  // a different TFE context for each thread of execution (for running graph
  // functions, and their send/recvs corountines).
  config.set_inter_op_parallelism_threads(1);

  TF_Buffer* ret = TF_NewBuffer();
  TF_CHECK_OK(MessageToBuffer(config, ret));
  return ret;
}

      tf_device.return %2 : tensor<2xi32>
    }

    func.return %1 : tensor<2xi32>
  }

  // Ensures that separate send/recvs are added for values that are used by ops inside of multiple IfRegions.

  // CHECK-LABEL: func @outside_compiled_ops_multiple_tf_if
  func.func @outside_compiled_ops_multiple_tf_if(%arg0: tensor<2xi32>) -> tensor<2xi32> {

				continue
			}
			s.newHeapaddr(n)
			if n.Class == ir.PPARAM {
				s.move(n.Type(), s.expr(n.Heapaddr), s.decladdrs[n])
			}
		}
	}

	typ := s.curfn.Type()
	do(typ.Recvs())
	do(typ.Params())
	do(typ.Results())
}

// newHeapaddr allocates heap memory for n and sets its heap address.
func (s *state) newHeapaddr(n *ir.Name) {
	s.setHeapaddr(n.Pos(), n, s.newObject(n.Type(), nil))