transpose op code behavior mismatch with comments in xla

[mingting]

tensorflow 1.15.5
https://github.com/tensorflow/tensorflow/blob/v1.15.5/tensorflow/compiler/xla/service/shape_inference.cc#L2740

  // Permute(dimensions,input) computes output[dimensions[i]]=input[i]. However,
  // we need output[i]=input[dimensions[i]] which is
  // Permute(Inverse(dimensions),input).
  return ShapeUtil::PermuteDimensions(InversePermutation(dimensions), operand);

https://github.com/tensorflow/tensorflow/blob/v1.15.5/tensorflow/compiler/xla/util.h#L340

// Applies `permutation` on `input` and returns the permuted array.
// For each i, output[permutation[i]] = input[i].
//
// Precondition:
// 1. `permutation` is a permutation of 0..permutation.size()-1.
// 2. permutation.size() == input.size().
template <typename Container>
std::vector<typename Container::value_type> Permute(
    absl::Span<const int64> permutation, const Container& input) {
  using T = typename Container::value_type;
  absl::Span<const T> data(input);
  CHECK(IsPermutation(permutation, data.size()));
  std::vector<T> output(data.size());
  for (size_t i = 0; i < permutation.size(); ++i) {
    output[permutation[i]] = data[i];
  }
  return output;
}

my question is
accoding to the comments, if using inverse permutation, we should use output[i]=input[dimensions[i]].
but, code still use output[dimensions[i]]=input[i].

seems like a bug or mismatch?
I checked latest code, this is still there, but inverse moved to other place.
so I used 1.15.5, as the inverse code is together with permute.

BR.
Mingting

[mingting]

@hawkinsp can you help to check this?