// it's double-declared (it would cost extra in the common case to verify
// this). But the MethodSet computation will not find it due to the name
// collision caused by the double-declaration, leading to an internal
// inconsistency while we are verifying one computation against the other.
// var _ = T1c{}.Pointer

// T2's method declared before the type.
func (*T2) f /* ERROR "field and method" */ () {}

{{DECLS_FROM_OBJ_FILE}}

{{NS_START}}
// {{CLASS}} represents a computation previously specified in a
// TensorFlow graph, now compiled into executable code. This extends the generic
// XlaCompiledCpuFunction class with statically type-safe arg and result
// methods. Usage example:
//
//   {{CLASS}} computation;
//   // ...set args using computation.argN methods
//   CHECK(computation.Run());

// use_tuple_args: when this is true, always create a tuple argument for the
//   entry computation.
// enable_op_fallback: when this is true, prefer tf2xla fallback kernels over
// MLIR
//   native kernels for legalization to HLO.
// return_tuple: when this is true, always create a tuple result for the
//   entry computation.
// shape_determination_fns: Contains layout preference fn and shape

// supports float types. tf.round with integer input type will become an
// identity op, so we will never face an mhlo.floor with an integer input type.
// The pattern matched executes the following computation:
// frac = x - floor(x)
// to_even = (floor(x) - 2 * floor(0.5 * x)) == 1
// if frac > 0.5 || (frac == 0.5 && to_even)
//   return floor(x) + 1
// else
//   return floor(x)
def : Pat<(MHLO_SelectOp

// For case 3, we need to create a PjRtBuffer from the raw device mem pointer,
// and we need to ensure the PjRtBuffer persists till XLA computation is
// complete. Therefore we put the newly created PjRtBuffer into `owned_args`.
// Caller is responsible to ensure `owned_args` lives till the end of XLA
// computation.
Status PreparePjRtExecutableArguments(
    int num_missing_prefix_ctx_inputs, const std::vector<int>& input_mapping,

    }
  }
}

// Since we have the outputs from host and device computation after moving
// outside compiled ops, we can create the actual parallel_execute regions.
// Still, one region is for the host computation for outside compilation and
// the other one is for the original Device cluster computation.
mlir::tf_device::ParallelExecuteOp CreateFinalParallelExecuteOp(

* Introducing `tf.types.experimental.AtomicFunction` as the fastest way to perform TF computations in Python.

    * Can be accessed through `inference_fn` property of `ConcreteFunction`s
    * Does not support gradients.
    * See `tf.types.experimental.AtomicFunction` documentation for how to call and use it.

XlaCompiler::CompileOptions GetCompileOptions(bool for_pjrt = false) {
  XlaCompiler::CompileOptions compile_options;
  compile_options.is_entry_computation = true;
  // Optimization: where possible, have the computation return a naked array
  // rather than a one-element tuple.
  compile_options.always_return_tuple = false;
  if (for_pjrt) {
    compile_options.use_tuple_arg = false;
    compile_options.always_return_tuple = true;
  }

namespace tf2xla {
namespace internal {

// Legalize the given MLIR module to XLA HLO using a combination of the MLIR
// Bridge and XlaBuilder
absl::StatusOr<XlaCompilationResult> LegalizeTfToHlo(
    const tpu::MlirToHloArgs& computation,
    const tpu::TPUCompileMetadataProto& metadata, bool use_tuple_args,
    llvm::StringRef device_type,
    XlaShapeLayoutHelpers::ShapeDeterminationFns shape_determination_fns,

    $mlir_module is a serialized MLIR module with a `main` function that contains target computation.
    $metadata is a serialized TPUCompileMetadataProto describing the shapes and types of the inputs to the computation, as well as a mapping onto the TPU pod topology.