}
      (*types)[index] = n->output_type(0);
    }
  }
  return absl::OkStatus();
}

// Renumber the indices of _Arg nodes in a graph, according to
// 'permutation' that maps old indices to new indices.
static Status RenumberArguments(Graph* graph,
                                const std::vector<int>& permutation) {
  for (Node* n : graph->op_nodes()) {

	tparams      *TypeParamList // local type parameters
	paramNames   bool           // if set, write function parameter names, otherwise, write types only
	tpSubscripts bool           // if set, write type parameter indices as subscripts
	pkgInfo      bool           // package-annotate first unexported-type field to avoid confusing type description
}

func newTypeWriter(buf *bytes.Buffer, qf Qualifier) *typeWriter {

// new argument index that corresponds to each original index (-1 means it is
// removed). If remaining_resource_data_types is provided, it will store the
// data types of the remaining resource arguments, where the indices are after
// removing unused ones.
void RemoveUnusedResourceArgumentsAndForwardedRetvals(
    const llvm::SmallDenseMap<int64_t, ResourceArgUseInfo>& infos,
    func::FuncOp func_op,

		}
	}
}

type subexpIndex struct {
	name  string
	index int
}

type subexpCase struct {
	input   string
	num     int
	names   []string
	indices []subexpIndex
}

var emptySubexpIndices = []subexpIndex{{"", -1}, {"missing", -1}}

var subexpCases = []subexpCase{
	{``, 0, nil, emptySubexpIndices},
	{`.*`, 0, nil, emptySubexpIndices},

			for ; p < q; p += SUBBUCKETSIZE {
				i = int((p - min) / SUBBUCKETSIZE)
				if indexes[i] > idx {
					indexes[i] = idx
				}
			}

			i = int((q - 1 - min) / SUBBUCKETSIZE)
			if indexes[i] > idx {
				indexes[i] = idx
			}
			idx++
		}

		// fill in table
		for i := int32(0); i < nbuckets; i++ {
			base := indexes[i*SUBBUCKETS]
			if base == NOIDX {
				Errorf(nil, "hole in findfunctab")
			}

	}

	// The Huffman tree can switch every 50 symbols so there's a list of
	// tree indexes telling us which tree to use for each 50 symbol block.
	numSelectors := br.ReadBits(15)
	treeIndexes := make([]uint8, numSelectors)

	// The tree indexes are move-to-front transformed and stored as unary
	// numbers.
	mtfTreeDecoder := newMTFDecoderWithRange(numHuffmanTrees)

	tparams      *TypeParamList // local type parameters
	paramNames   bool           // if set, write function parameter names, otherwise, write types only
	tpSubscripts bool           // if set, write type parameter indices as subscripts
	pkgInfo      bool           // package-annotate first unexported-type field to avoid confusing type description
}

func newTypeWriter(buf *bytes.Buffer, qf Qualifier) *typeWriter {

  if (lstm_func.getNumArguments() == 7) return failure();

  // We should know the batch size in advance for the lstm fusion.
  // A good indicator of batch size is both cell state and input state (indices
  // 1 & 2) have fixed shape and other input tenors should have ranked tensor
  // types.
  for (int i = 0; i < 6; ++i) {
    auto input = lstm_func.getArgument(i);

    assert(output_axis < num_dimensions);
    const int input_axis = perm[output_axis];
    for (int i = 0; i < output_shape[output_axis]; ++i) {
      // Update the input indices on `input_axis`.
      assert(input_axis < input_indices->size());
      input_indices->operator[](input_axis) = static_cast<uint64_t>(i);
      // Write the value from `input_tensor` if it is the last axis or

      padded_shape[i] += padding;
    }

    return padded_shape;
  }

  std::pair<int64_t, int64_t> GetDimSize(
      const ArrayRef<int64_t> shape, const ArrayRef<int64_t> indexes) const {
    return {shape[indexes[0]], shape[indexes[1]]};
  }

  bool IsTransposeConv(
      stablehlo::ConvolutionOp op,
      stablehlo::ConvDimensionNumbersAttr dimension_numbers) const {