assertEquals(1, future.get(-1, SECONDS).intValue());
  }

  @J2ktIncompatible
  @GwtIncompatible // threads
  public void testOverflowTimeout() throws Exception {
    // First, sanity check that naive multiplication would really overflow to a negative number:
    long nanosPerSecond = NANOSECONDS.convert(1, SECONDS);
    assertThat(nanosPerSecond * Long.MAX_VALUE).isLessThan(0L);

}

def TF_XlaSparseDenseMatmulWithStaticBufferSizeOp : TF_Op<"XlaSparseDenseMatmulWithStaticBufferSize", [Pure]> {
  let summary = "A XLA op which performs the dense-sparse matrix multiplication.";

  let arguments = (ins
    TF_Int32Tensor:$row_pointers,
    TF_Int32Tensor:$sorted_sample_ids,
    TF_Int32Tensor:$sorted_token_ids,
    TF_Float32Tensor:$sorted_gains,

	size := uintptr(class_to_size[c.spanclass.sizeclass()])

	s := mheap_.alloc(npages, c.spanclass)
	if s == nil {
		return nil
	}

	// Use division by multiplication and shifts to quickly compute:
	// n := (npages << _PageShift) / size
	n := s.divideByElemSize(npages << _PageShift)
	s.limit = s.base() + size*n
	s.initHeapBits(false)
	return s

  }

  const int64_t rows = lhs_shape[lhs_dims - 2];
  const int64_t cols = rhs_shape[rhs_dims - 1];

  if (lhs_shape[lhs_dims - 1] != rhs_shape[rhs_dims - 2]) {
    // Input dimensions must be compatible for multiplication.
    return failure();
  }

  const auto matmul_type = RankedTensorType::get({rows, cols}, element_type);

  if (lhs_dims == 2 && rhs_dims == 2) {

// The following assembly functions are implemented in p256_asm_*.s

// Montgomery multiplication. Sets res = in1 * in2 * R⁻¹ mod p.
//
//go:noescape
func p256Mul(res, in1, in2 *p256Element)

// Montgomery square, repeated n times (n >= 1).
//
//go:noescape
func p256Sqr(res, in *p256Element, n int)

// Montgomery multiplication by R⁻¹, or 1 outside the domain.

	"F0",
	"F2",
	"F4",
	"F6",
	"F8",
	"F10",
	"F12",
	"F14",
	"F16",
	"F18",
	"F20",
	"F22",
	"F24",
	"F26",
	"F28",
	"F30",

	"HI", // high bits of multiplication
	"LO", // low bits of multiplication

	// If you add registers, update asyncPreempt in runtime.

	// pseudo-registers
	"SB",
}

func init() {
	// Make map from reg names to reg integers.
	if len(regNamesMIPS) > 64 {

	"F16",
	"F17",
	"F18",
	"F19",
	"F20",
	"F21",
	"F22",
	"F23",
	"F24",
	"F25",
	"F26",
	"F27",
	"F28",
	"F29",
	"F30",
	"F31",

	"HI", // high bits of multiplication
	"LO", // low bits of multiplication

	// If you add registers, update asyncPreempt in runtime.

	// pseudo-registers
	"SB",
}

func init() {
	// Make map from reg names to reg integers.

		computeDivMagic(&classes[i])
	}

	return classes
}

// computeDivMagic checks that the division required to compute object
// index from span offset can be computed using 32-bit multiplication.
// n / c.size is implemented as (n * (^uint32(0)/uint32(c.size) + 1)) >> 32
// for all 0 <= n <= c.npages * pageSize
func computeDivMagic(c *class) {
	// divisor
	d := c.size
	if d == 0 {
		return
	}

    double ySumOfSquaresOfDeltas = yStats.sumOfSquaresOfDeltas();
    checkState(xSumOfSquaresOfDeltas > 0.0);
    checkState(ySumOfSquaresOfDeltas > 0.0);
    // The product of two positive numbers can be zero if the multiplication underflowed. We
    // force a positive value by effectively rounding up to MIN_VALUE.
    double productOfSumsOfSquaresOfDeltas =
        ensurePositive(xSumOfSquaresOfDeltas * ySumOfSquaresOfDeltas);

    double ySumOfSquaresOfDeltas = yStats().sumOfSquaresOfDeltas();
    checkState(xSumOfSquaresOfDeltas > 0.0);
    checkState(ySumOfSquaresOfDeltas > 0.0);
    // The product of two positive numbers can be zero if the multiplication underflowed. We
    // force a positive value by effectively rounding up to MIN_VALUE.
    double productOfSumsOfSquaresOfDeltas =
        ensurePositive(xSumOfSquaresOfDeltas * ySumOfSquaresOfDeltas);