%0 = "tfl.pseudo_sparse_const"() {compressed_data = dense<[1.0, 2.0, 0.5, 0.25,...

  %0 = "tfl.pseudo_sparse_const"() {compressed_data = dense<[1.0, 2.0, 0.5, 0.25,...

//
// Eg 1. An update in canonical form:
//  * indices shape(A,B,C)
//  * updates shape(A,B,D,E,F)
// Then:
//  * D,E,F are the update window dims [2,3,4]
//  * C is the index vector dimension
//  * A,B iterate over the updates and indices
//
// If `update_window_dims` are not the trailing dimensions then updates must be
// transposed.
//
// Eg 2. An update in non-canonical form:
//  * indices shape(a,b,c)
//  * updates shape(d,e,f,a,b)

		// indices in the first place (if it returned fewer than that then
		// there cannot be more).
		for n1 := n; ; n1 += 2 * (n - len(result)) /* overflow ok */ {
			indices := x.Lookup(lit, n1)
			if len(indices) == 0 {
				return
			}
			slices.Sort(indices)
			pairs := make([]int, 2*len(indices))
			result = make([][]int, len(indices))
			count := 0
			prev := 0

		return e.X, e.Lbrack, e.Indices, e.Rbrack
	}
	return nil, token.NoPos, nil, token.NoPos
}

// PackIndexExpr returns an *ast.IndexExpr or *ast.IndexListExpr, depending on
// the cardinality of indices. Calling PackIndexExpr with len(indices) == 0
// will panic.
func PackIndexExpr(x ast.Expr, lbrack token.Pos, indices []ast.Expr, rbrack token.Pos) ast.Expr {
	switch len(indices) {
	case 0:
		panic("empty indices")

def LowerScatterNdOp :
  Pat<(TF_ScatterNdOp $indices,
       TensorOf<[AnyInteger, AnyFloat, AnyComplex]>:$updates, $shape, ConstantStrAttr<StrAttr, "">),
      (TF_TensorScatterAddOp
       (TF_FillOp $shape, (TF_ConstOp (GetScalarOfType<0> $updates))),
       $indices, $updates)>;

def LowerScatterNdOpDefaultBadIndicesPolicy :
  Pat<(TF_ScatterNdOp $indices,

namespace {

// Returns contiguous offset (address) of the position represented by `indices`
// in a `shape` shaped tensor. Assumes row-major order. `indices` and `shape`
// should have the same size.
// Example: Index (2, 3) of a (4, 5)-shaped tensor has the contiguous offset of
// 2 * 5 + 3 = 13.
int64_t GetContiguousOffset(const ArrayRef<int64_t> indices,
                            const ArrayRef<int64_t> shape) {

      indices += absl::StrCat("[dim", dim, "]");
      count *= shape.dimensions(dim);
    }
  }
  rewrites->push_back({"{{I}}", absl::StrCat(i)});
  rewrites->push_back({"{{TYPE}}", type});
  rewrites->push_back({"{{DIM_VARS}}", absl::StrJoin(dim_vars, ", ")});
  rewrites->push_back({"{{DIM_SIZES}}", dim_sizes});
  rewrites->push_back({"{{INDICES}}", indices});

                                     ArrayRef<Value>{a, b});
}

namespace {

int64_t GetFirstIfIndicesAreContiguous(Value indices) {
  auto type = mlir::dyn_cast<RankedTensorType>(indices.getType());
  if (!type) return -1;
  auto indices_op = indices.getDefiningOp();
  if (!indices_op) return -1;
  auto const_op = llvm::dyn_cast<TF::ConstOp>(indices_op);
  if (!const_op) return -1;

  let description = [{
This operation computes

    # Scalar indices
    ref[indices, ...] = max(ref[indices, ...], updates[...])

    # Vector indices (for each i)
    ref[indices[i], ...] = max(ref[indices[i], ...], updates[i, ...])

    # High rank indices (for each i, ..., j)
    ref[indices[i, ..., j], ...] = max(ref[indices[i, ..., j], ...], updates[i, ..., j, ...])