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

    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

  auto recv =
      builder.create<RecvOp>(loc, recv_result_type, token, channel_handle,
                             /*is_host_transfer=*/builder.getBoolAttr(true));

  SetFrontendAttributes(recv, index, key, result_type,
                        /*device_to_host=*/false, host_handler_name);

  SetOpSharding(recv, manual_sharding);

  result.replaceAllUsesWith(recv.getResult(0));

		}
		if len(t.AllMethods()) != 0 {
			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:

	// 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;
}

			// we copy directly. Note that we need to increment
			// the head/tail locations only when raceenabled.
			racenotify(c, c.recvx, nil)
			racenotify(c, c.recvx, sg)
			c.recvx++
			if c.recvx == c.dataqsiz {
				c.recvx = 0
			}
			c.sendx = c.recvx // c.sendx = (c.sendx+1) % c.dataqsiz
		}
	}
	if sg.elem != nil {
		sendDirect(c.elemtype, sg, ep)
		sg.elem = nil
	}
	gp := sg.g

			for p := 0; p < P; p++ {
				go func() {
					recv := make(map[int]int)
					for i := 0; i < L; i++ {
						v := <-c
						recv[v] = recv[v] + 1
					}
					done <- recv
				}()
			}
			recv := make(map[int]int)
			for p := 0; p < P; p++ {
				for k, v := range <-done {
					recv[k] = recv[k] + v
				}
			}
			if len(recv) != L {

	if nsends+nrecvs == 0 {
		return dflt, false
	}

	// Compact sel and orig if necessary.
	if nsends+nrecvs < len(cases) {
		copy(sel[nsends:], sel[len(cases)-nrecvs:])
		copy(orig[nsends:], orig[len(cases)-nrecvs:])
	}

	order := make([]uint16, 2*(nsends+nrecvs))
	var pc0 *uintptr
	if raceenabled {
		pcs := make([]uintptr, nsends+nrecvs)
		for i := range pcs {
			selectsetpc(&pcs[i])
		}

		apiRouter.ServeHTTP(recV4, reqV4)
		if recV4.Code != testCase.expectedRespStatus {
			t.Errorf("Test %d: %s: Expected the response status to be `%d`, but instead found `%d`", i+1, instanceType, testCase.expectedRespStatus, recV4.Code)
		}
		// initialize HTTP NewRecorder, this records any mutations to response writer inside the handler.
		recV2 := httptest.NewRecorder()