mlir::IntegerType::get(permutation1.getContext(), 32)),
      llvm::ArrayRef(new_permutation));
}

// Utility function to map final permutation to initial permutation
// initial -> permutation1 -> permutation2 -> final
inline DenseElementsAttr RemapPermutation(Value permutation1,
                                          Value permutation2) {
  DenseElementsAttr perm2_const;

  }
}

// Revert the permutation applied in `type`.
static mlir::ShapedType ReversePermuteShapedType(
    mlir::ShapedType type, ArrayRef<int64_t> permutation) {
  if (!type.hasRank()) return type;

  auto shape = type.getShape();
  SmallVector<int64_t, 4> new_shape(shape.size());

  for (int i = 0; i < permutation.size(); ++i) {
    int64_t index = permutation[i];
    assert(index < shape.size());

}

// Convenience function to create a `TransposeOp` with a given `permutation`.
// The Location is set as `input`'s loc.
TransposeOp CreateTransposeOp(Value input, const ArrayRef<int64_t> permutation,
                              PatternRewriter& rewriter) {
  return rewriter.create<TransposeOp>(
      input.getLoc(), input, rewriter.getDenseI64ArrayAttr(permutation));
}

  }
  auto transposed_type =
      RankedTensorType::get(transposed_shape, input_type.getElementType());
  DenseIntElementsAttr permutation = DenseIntElementsAttr::get(
      RankedTensorType::get(permutation_array.size(), rewriter.getI64Type()),
      permutation_array);
  return {permutation, transposed_type};
}

Value BuildIntConstOp(ImplicitLocOpBuilder& builder,

inline constexpr StringRef kQuantizationMethodAttr = "_quantization_method";

// Permutation from the NHWC tensor format to NCHW. This is an inverse
// permutation of `kNchwToNhwcPermutation`.
inline constexpr std::array<int64_t, 4> kNhwcToNchwPermutation = {0, 3, 1, 2};

// Permutation from the NCHW tensor format to NHWC. This is an inverse
// permutation of `kNchwToNhwcPermutation`.

	permutation(t, h, []uint32{0, 1}, 20)
	permutation(t, h, []uint32{0, 1 << 31}, 20)
	permutation(t, h, []uint32{0, 1, 2, 3, 4, 5, 6, 7, 1 << 29, 2 << 29, 3 << 29, 4 << 29, 5 << 29, 6 << 29, 7 << 29}, 6)
}
func permutation(t *testing.T, h *hashSet, s []uint32, n int) {
	b := make([]byte, n*4)
	genPerm(h, b, s, 0)
	h.check(t)
}

    ],
)

cc_library(
    name = "permutation",
    hdrs = ["permutation.h"],
    compatible_with = get_compatible_with_portable(),
    deps = [
        "@llvm-project//llvm:Support",
        "@llvm-project//mlir:Support",
    ],
)

tf_cc_test(
    name = "permutation_test",
    srcs = ["permutation_test.cc"],
    deps = [
        ":permutation",
        "@com_google_googletest//:gtest_main",

		t.Skip("Too long for race mode")
	}
	testenv.ParallelOn64Bit(t)
	h := newHashSet()
	permutation(t, h, []uint32{0, 1, 2, 3, 4, 5, 6, 7}, 8)
	permutation(t, h, []uint32{0, 1 << 29, 2 << 29, 3 << 29, 4 << 29, 5 << 29, 6 << 29, 7 << 29}, 8)
	permutation(t, h, []uint32{0, 1}, 20)
	permutation(t, h, []uint32{0, 1 << 31}, 20)
	permutation(t, h, []uint32{0, 1, 2, 3, 4, 5, 6, 7, 1 << 29, 2 << 29, 3 << 29, 4 << 29, 5 << 29, 6 << 29, 7 << 29}, 6)
}

//
//  The update needs permuting to be in the form (a,b,d,e,f) so that the update
//  window dims are the trailing dimensions.
//
// To canonicalize the updates above, replace the updates with:
//   transpose(updates, permutation={3,4,0,1,2})
//
// Note: NormalizeIndexVector is assumed to have run on the indices already so
// that the index_vector_dim is the trailing dimension in `indices`.
LogicalResult CanonicalizeScatterUpdates(

      redux_op);
}

TF::TransposeOp createTransposeOp(Value value, Location loc,
                                  llvm::ArrayRef<int32_t> permutation,
                                  PatternRewriter* rewriter) {
  auto perm_op = createI32ConstantOp(permutation, loc, rewriter);
  auto value_type = mlir::cast<RankedTensorType>(value.getType());
  auto shape = value_type.getShape();