} else if nrgba != nil {
				stride, pix = nrgba.Stride, nrgba.Pix
			}
			if stride != 0 {
				j0 := (y - b.Min.Y) * stride
				j1 := j0 + b.Dx()*4
				for j := j0; j < j1; j += 4 {
					cr0[i+0] = pix[j+0]
					cr0[i+1] = pix[j+1]
					cr0[i+2] = pix[j+2]
					i += 3
				}
			} else {
				for x := b.Min.X; x < b.Max.X; x++ {
					r, g, b, _ := m.At(x, y).RGBA()

  //
  // `axis` should be a valid dimension in `shape`.
  static int64_t StrideForAxis(ArrayRef<int64_t> shape, int64_t axis) {
    int64_t stride = 1;  // Start with the trailing dimension.
    for (int64_t dim = shape.size() - 1; dim > axis; --dim) {
      stride *= shape[dim];
    }
    return stride;
  }

  // Calculates how many values to skip across a 1-D contiguous array that holds

// CHECK-NEXT:   func.func @main(%arg0: tensor<8x8x1x207xf32>, %arg1: tensor<3x3x16x207xf32>) -> tensor<16x8x8x1xf32> {

  int kh;  // Kernel height.
  int kw;  // Kernel width.

  int ph;  // Padding on "height" dimension (both sides).
  int pw;  // Padding on "width" dimension (both sides).

  int sh;  // Stride on "height" dimension.
  int sw;  // Stride on "width" dimension.

  bool ceil_mode;  // Rounding strategy (ceil or floor).
};

// Rounds the dimension based on the ceil mode.
int RoundDim(float dim, bool ceil_mode) {
  if (ceil_mode) {

      }

      @Override
      public int compare(byte[] left, byte[] right) {
        int stride = 8;
        int minLength = Math.min(left.length, right.length);
        int strideLimit = minLength & ~(stride - 1);
        int i;

        /*
         * Compare 8 bytes at a time. Benchmarking on x86 shows a stride of 8 bytes is no slower

// For n < t.cutoff, the block is a simple lookup table. Otherwise, the block
// is a list of ranges with an accompanying value. Given a matching range r,
// the value for b is by r.value + (b - r.lo) * stride.
func (t *sparseBlocks) lookup(n uint32, b byte) uint16 {
	offset := t.offset[n]
	header := t.values[offset]
	lo := offset + 1
	hi := lo + uint16(header.lo)
	for lo < hi {
		m := lo + (hi-lo)/2
		r := t.values[m]

// For n < t.cutoff, the block is a simple lookup table. Otherwise, the block
// is a list of ranges with an accompanying value. Given a matching range r,
// the value for b is by r.value + (b - r.lo) * stride.
func (t *sparseBlocks) lookup(n uint32, b byte) uint16 {
	offset := t.offset[n]
	header := t.values[offset]
	lo := offset + 1
	hi := lo + uint16(header.lo)
	for lo < hi {
		m := lo + (hi-lo)/2
		r := t.values[m]

// For n < t.cutoff, the block is a simple lookup table. Otherwise, the block
// is a list of ranges with an accompanying value. Given a matching range r,
// the value for b is by r.value + (b - r.lo) * stride.
func (t *sparseBlocks) lookup(n uint32, b byte) uint16 {
	offset := t.offset[n]
	header := t.values[offset]
	lo := offset + 1
	hi := lo + uint16(header.lo)
	for lo < hi {
		m := lo + (hi-lo)/2
		r := t.values[m]

      }

      @Override
      public int compare(byte[] left, byte[] right) {
        int stride = 8;
        int minLength = Math.min(left.length, right.length);
        int strideLimit = minLength & ~(stride - 1);
        int i;

        /*
         * Compare 8 bytes at a time. Benchmarking on x86 shows a stride of 8 bytes is no slower

// To trigger issue21655, the index offset needs to be small
// enough to fit into an int32 (to get rewritten to an ADDQconst)
// but large enough to overflow an int32 after multiplying by the stride.

package main

func f1(a []int64, i int64) int64 {
	return a[i+1<<30]
}
func f2(a []int32, i int64) int32 {
	return a[i+1<<30]
}
func f3(a []int16, i int64) int16 {
	return a[i+1<<30]
}