// Returns TypeId for the provided hardware.
  // Usually should be something like mlir::TypeID::get<MyType>()
  virtual mlir::TypeID GetTypeId() const = 0;

  virtual void GetDependentDialects(mlir::DialectRegistry& registry) const {}

 protected:
  // All registered hardware ops.
  std::vector<std::unique_ptr<TargetHardwareOperation>> hardware_ops_;
};

// Returns pointer to the Hardware identified by 'hardware_name'.

    // It is a terrible idea to have local MLIR contexts so we need to
    // register extensions here, again.
    mlir::DialectRegistry registry;
    registry.insert<mlir::func::FuncDialect>();
    mlir::func::registerAllExtensions(registry);
    xla_call_module_context_.appendDialectRegistry(registry);

    auto l = tensorflow::XlaCallModuleLoader::Create(
        &xla_call_module_context_, op.getVersion(), op.getModule().str(),

//
// As we need these URI schemes to instantiate the test suite when
// `testing::InitGoogleTest` gets called, here we just store them to an
// internal scheme registry. See `URISchemeRegister` above.
static bool GetURIScheme(const std::string& scheme) {
  tensorflow::SchemeVector()->push_back(scheme);
  return true;
}

// This function is used for cloud filesystem

  }
  func.return
  // expected-remark@above {{ID: 25}}
  // expected-remark@above {{Sinks: {24}}}
}

// -----

// Tests that the pass tracks control dependencies based on TF op registry
// statefulness flag, for ops not yet defined in ODS.

// CHECK-LABEL: func @tf_registry_ops
func.func @tf_registry_ops(
  // expected-remark@above {{ID: 8}}

Fixed point shape refinement pass that utilizes the shape functions
registered on ops using the InferTypeOpInterface as well as by bridging to
the TensorFlow op registry's shape functions. This is an interprocedural
pass that propagates information across function calls/control flow
operations where possible (the GuaranteeAllFuncsOneUsePass is often run