//
	// This also allows scanning into user defined types such as "type Int int64".
	// For symmetry, also check for string destination types.
	switch dv.Kind() {
	case reflect.Pointer:
		if src == nil {
			dv.SetZero()
			return nil
		}
		dv.Set(reflect.New(dv.Type().Elem()))
		return convertAssignRows(dv.Interface(), src, rows)
	case reflect.Int, reflect.Int8, reflect.Int16, reflect.Int32, reflect.Int64:

	partLen := c.Rand.Intn(64) // len is [0, 63]
	if !canBeEmpty {
		partLen = c.Rand.Intn(63) + 1 // len is [1, 63]
	}

	runes := make([]rune, partLen)
	if partLen == 0 {
		return string(runes)
	}

	runes[0] = validStartEnd[c.Rand.Intn(len(validStartEnd))].choose(c.Rand)
	for i := range runes[1:] {
		runes[i+1] = validMiddle[c.Rand.Intn(len(validMiddle))].choose(c.Rand)
	}

namespace {

#define GEN_PASS_DEF_REDUCETYPEPRECISIONPASS
#include "tensorflow/compiler/mlir/lite/transforms/passes.h.inc"

// This pattern checks if an i8 arith::ConstantOp tensor has all values within
// the INT4 range, i.e. [-8,7] and converts it into i4 if so. This assumes that
// the input is sign-extended two's complement.
class CheckRangeAndConvertI8ToI4 : public OpRewritePattern<arith::ConstantOp> {
 public:

    const bool is_signed = quant_specs_.IsSignedInferenceType();
    const int bit_width = quant_specs_.GetQuantizationTypeWidth();

    std::unique_ptr<OpQuantSpec> spec = GetTFOpQuantSpec(quantized_op);
    const int quant_dim = spec->coeff_op_quant_dim[weight_idx];
    const bool is_per_channel_quantization =
        enable_per_channel_quantization_ && quant_dim != -1;

    QuantizedType quant_type;

// alias basic type named "byte" (and analogous for "rune").
var Typ = [...]*Basic{
	Invalid: {Invalid, 0, "invalid type"},

	Bool:          {Bool, IsBoolean, "bool"},
	Int:           {Int, IsInteger, "int"},
	Int8:          {Int8, IsInteger, "int8"},
	Int16:         {Int16, IsInteger, "int16"},
	Int32:         {Int32, IsInteger, "int32"},
	Int64:         {Int64, IsInteger, "int64"},

// alias basic type named "byte" (and analogous for "rune").
var Typ = []*Basic{
	Invalid: {Invalid, 0, "invalid type"},

	Bool:          {Bool, IsBoolean, "bool"},
	Int:           {Int, IsInteger, "int"},
	Int8:          {Int8, IsInteger, "int8"},
	Int16:         {Int16, IsInteger, "int16"},
	Int32:         {Int32, IsInteger, "int32"},
	Int64:         {Int64, IsInteger, "int64"},

# CHECK:     version: 1
# CHECK:     builtin_code: RESHAPE
# CHECK:   } ],
# CHECK:   subgraphs: [ {
# CHECK:     tensors: [ {
# CHECK:       shape: [ 1, 1, 1, 256 ],
# CHECK:       type: INT8,
# CHECK:       buffer: 1,
# CHECK:       name: "input",
# CHECK:       quantization: {
# CHECK:         scale: [ 0.216328 ],
# CHECK:         zero_point: [ 27 ]
# CHECK:       }
# CHECK:     }, {

std::vector<Value> ConvertTFBatchMatMulOp<BatchMatMulOpType>::sliceInput(
    Value value, int batch_size, Location loc, PatternRewriter& rewriter) {
  RankedTensorType tensorType = mlir::cast<RankedTensorType>(value.getType());
  Type element_type = tensorType.getElementType();

  int rank = tensorType.getShape().size();
  int num_rows = tensorType.getShape()[rank - 2];
  int num_cols = tensorType.getShape()[rank - 1];

  std::vector<Value> sliced;

          /*used_tensors=*/used_tensors, /*quantized=*/true);
      // If the tensor is a weight, it should have type INT8.
      // If the tensor is a bias, it should have type FLOAT32.
      // If the tensor is an input or output it should have type FLOAT32.
      // The input to dequantize should be INT8, and all other tensors should be
      // FLOAT32.
      if (i == dequant_input_idx) {

}

var ignoreTests = []ignoreTest{
	// Decode normal struct into an empty struct
	{&struct{ A int }{23}, &struct{}{}},
	// Decode normal struct into a nil.
	{&struct{ A int }{23}, nil},
	// Decode singleton string into a nil.
	{"hello, world", nil},
	// Decode singleton slice into a nil.
	{[]int{1, 2, 3, 4}, nil},
	// Decode struct containing an interface into a nil.
	{&Struct0{&NewType0{"value0"}}, nil},