// Output the state machine for each function remaining.
	for _, s := range unit.Textp {
		fnSym := loader.Sym(s)
		_, _, _, lines := d.ldr.GetFuncDwarfAuxSyms(fnSym)

		// Chain the line symbol onto the list.
		if lines != 0 {
			syms = append(syms, lines)
			unitlen += int64(len(d.ldr.Data(lines)))
		}
	}

	if d.linkctxt.HeadType == objabi.Haix {

              edge->src()->name(), ":", edge->src_output());
        }
      }

      Subgraph& dst_subgraph = subgraphs_[dst_func_id];
      // Ignore control edges entering the subgraph. We will lift them onto
      // the enclosing call operators in BuildOutputGraph().
      if (!edge->IsControlEdge()) {
        TF_RETURN_IF_ERROR(
            dst_subgraph.RecordArg(edge, node_images, src_arg_pairs));
      }
    }

		hsub := p.calcHeight(sub, false)
		if h < 1+hsub {
			h = 1 + hsub
		}
	}
	p.height[re] = h
	return h
}

// Parse stack manipulation.

// push pushes the regexp re onto the parse stack and returns the regexp.
func (p *parser) push(re *Regexp) *Regexp {
	p.numRunes += len(re.Rune)
	if re.Op == OpCharClass && len(re.Rune) == 2 && re.Rune[0] == re.Rune[1] {
		// Single rune.

    OpsTestBase::SetDevice(DEVICE_GPU, std::move(device));
#endif
    TF_ASSERT_OK(NodeDefBuilder(op_name, op_name).Finalize(node_def()));
    TF_ASSERT_OK(InitOp());
  }

#if GOOGLE_CUDA || TENSORFLOW_USE_ROCM
  const char* device_name_ = tensorflow::DEVICE_GPU;
#else
  const char* device_name_ = tensorflow::DEVICE_CPU;
#endif
};

         func.getRegion().getOps<TF::GlobalIterIdOp>()) {
      auto loc = global_iter_id_op->getLoc();
      builder.setInsertionPointAfter(global_iter_id_op);
      auto offset = builder.create<TF::ConstOp>(
          loc, builder.getI64IntegerAttr(offset_value));
      auto new_global_iter_id = builder.create<TF::AddV2Op>(
          loc, global_iter_id_op->getResultTypes(),

  // The mode defines the action the controller will take when a pod is detected as broken.
  // If deletePods is true, the controller will delete the broken pod. The pod will then be rescheduled, hopefully onto a node that is fully ready.
  // Note this gives the DaemonSet a relatively high privilege, as it can delete any Pod.
  bool deletePods = 7;

            // variant and here we are in the case the "including" component said that transitive
            // should be false so we need to arbitrarily carry that onto the dependency metadata
            assert dependencies.size() == 1;
            dependencies = Collections.singletonList(makeNonTransitive(dependencies.get(0)));
        }

                                          /*allow_float=*/false, tensor_type_,
                                          output_buffer_);
  ASSERT_THAT(status, Eq(kTfLiteOk));

  // Verify ConstOp is quantized.
  const auto& subgraph = model_.subgraphs[0];
  auto op = subgraph->operators[0].get();
  ASSERT_THAT(GetBuiltinCode(model_.operator_codes[op->opcode_index].get()),
              Eq(BuiltinOperator_ADD));

              device_cluster.getLoc(),
              RankedTensorType::get({}, builder.getI64Type()),
              builder.getI64IntegerAttr(core));
    } else {
      device_ordinal_op = builder.create<mlir::TF::ConstOp>(
          device_cluster.getLoc(),
          DenseIntElementsAttr::get(
              RankedTensorType::get({}, builder.getI64Type()),
              static_cast<int64_t>(0)));
    }

	// Set m->sched.sp = SP, so that if a panic happens
	// during the function we are about to execute, it will
	// have a valid SP to run on the g0 stack.
	// The next few lines (after the havem label)
	// will save this SP onto the stack and then write
	// the same SP back to m->sched.sp. That seems redundant,
	// but if an unrecovered panic happens, unwindm will
	// restore the g->sched.sp from the stack location