clear(extraLimbs)
	x.limbs = x.limbs[:n]
	return x
}

// reset returns a zero nat of n limbs, reusing x's storage if n <= cap(x.limbs).
func (x *Nat) reset(n int) *Nat {
	if cap(x.limbs) < n {
		x.limbs = make([]uint, n)
		return x
	}
	clear(x.limbs)
	x.limbs = x.limbs[:n]
	return x
}

// set assigns x = y, optionally resizing x to the appropriate size.
func (x *Nat) set(y *Nat) *Nat {

)

func (n *Nat) String() string {
	var limbs []string
	for i := range n.limbs {
		limbs = append(limbs, fmt.Sprintf("%016X", n.limbs[len(n.limbs)-1-i]))
	}
	return "{" + strings.Join(limbs, " ") + "}"
}

// Generate generates an even nat. It's used by testing/quick to produce random
// *nat values for quick.Check invocations.
func (*Nat) Generate(r *rand.Rand, size int) reflect.Value {
	limbs := make([]uint, size)

			h2 += c

			msg = nil
		}

		// Multiplication of big number limbs is similar to elementary school
		// columnar multiplication. Instead of digits, there are 64-bit limbs.
		//
		// We are multiplying a 3 limbs number, h, by a 2 limbs number, r.
		//
		//                        h2    h1    h0  x
		//                              r1    r0  =

// https://cryptojedi.org/papers/neoncrypto-20120320.pdf
//
// This algorithm uses 5 26-bit limbs to represent a 130-bit
// value. These limbs are, for the most part, zero extended and
// placed into 64-bit vector register elements. Each vector
// register is 128-bits wide and so holds 2 of these elements.
// Using 26-bit limbs allows us plenty of headroom to accommodate
// accumulations before and after multiplication without

// Use of this source code is governed by a BSD-style
// license that can be found in the LICENSE file.

package field

import "math/bits"

// uint128 holds a 128-bit number as two 64-bit limbs, for use with the
// bits.Mul64 and bits.Add64 intrinsics.
type uint128 struct {
	lo, hi uint64
}

// mul64 returns a * b.
func mul64(a, b uint64) uint128 {
	hi, lo := bits.Mul64(a, b)
	return uint128{lo, hi}

  constexpr int64_t dims[] = {100, 100};
  constexpr int num_elements = dims[0] * dims[1];
  float data[num_elements];
  for (int i = 0; i < num_elements; ++i) {
    data[i] = 1.0f;
  }
  TF_Status* status = TF_NewStatus();
  TF_Tensor* t = TFE_AllocateHostTensor(ctx, TF_FLOAT, &dims[0],
                                        sizeof(dims) / sizeof(int64_t), status);

  %1 = stablehlo.transpose %0, dims = [1, 0] : (tensor<2x3xi32>) -> tensor<3x2xi32>
  return %1 : tensor<3x2xi32>
}
// CHECK: transpose

// -----

// Tests that transposing an argument cannot be folded.

// CHECK-LABEL: transpose_arg
func.func @transpose_arg(%arg0: tensor<2x3xf32>) -> tensor<3x2xf32> {
  %0 = stablehlo.transpose %arg0, dims = [1, 0] : (tensor<2x3xf32>) -> tensor<3x2xf32>

    - mikedanese
  sig-auth-authorizers-reviewers:
    - deads2k
    - dims
    - enj
    - liggitt
    - mikedanese
    - ncdc
    - smarterclayton
    - sttts
    - thockin
    - wojtek-t
  sig-auth-certificates-approvers:
    - liggitt
    - mikedanese
    - smarterclayton
  sig-auth-certificates-reviewers:
    - deads2k
    - dims
    - enj
    - liggitt
    - mikedanese

  static char empty;
  int64_t nelems = 1;
  std::vector<int64_t> dims;
  auto shape_dims = shape.dims();
  dims.reserve(shape_dims);
  for (int i = 0; i < shape_dims; ++i) {
    dims.push_back(shape.dim_size(i));
    nelems *= shape.dim_size(i);
  }
  CHECK_EQ(nelems, 0);
  return TF_NewTensor(
      dtype, reinterpret_cast<const int64_t*>(dims.data()), shape.dims(),

  %1 = stablehlo.transpose %arg0, dims = [0, 3, 1, 2] : (tensor<1x3x3x4xf32>) -> tensor<1x4x3x3xf32>
  %2 = stablehlo.add %1, %0 : tensor<1x4x3x3xf32>
  return %2 : tensor<1x4x3x3xf32>
}
// CHECK-SAME: (%[[ARG_0:.+]]: tensor<1x3x3x4xf32>) -> tensor<1x4x3x3xf32>
// CHECK-DAG: %[[CONST_0:.+]] = stablehlo.constant