// CHECK-NEXT:  %[[v10:.*]] = "mhlo.broadcast_in_dim"(%9) <{broadcast_dimensions = dense<> : tensor<0xi64>}> : (tensor<f32>) -> tensor<1x300x300x40xf32>
  // CHECK-NEXT:  %11 = mhlo.multiply %8, %10 : tensor<1x300x300x40xf32>
  // CHECK-NEXT:  %12 = mhlo.convert %arg2 : tensor<40xf32>

	var trunc uint32
	for mant >= 1<<32 {
		trunc |= uint32(mant) & 1
		mant >>= 1
		exp++
	}

	return fpack32(uint32(fs>>32), uint32(mant), exp, trunc)
}

// 64x64 -> 128 multiply.
// adapted from hacker's delight.
func mullu(u, v uint64) (lo, hi uint64) {
	const (
		s    = 32
		mask = 1<<s - 1
	)
	u0 := u & mask
	u1 := u >> s
	v0 := v & mask
	v1 := v >> s

  // These fields must satisfy the requirements of Stats' constructor as well as those of the stat
  // methods of this class.
  private long count = 0;
  private double mean = 0.0; // any finite value will do, we only use it to multiply by zero for sum
  private double sumOfSquaresOfDeltas = 0.0;
  private double min = NaN; // any value will do
  private double max = NaN; // any value will do

  /** Adds the given value to the dataset. */

     */
    open fun header(
      name: String,
      value: String,
    ) = commonHeader(name, value)

    /**
     * Adds a header with [name] to [value]. Prefer this method for multiply-valued
     * headers like "Set-Cookie".
     */
    open fun addHeader(
      name: String,
      value: String,
    ) = commonAddHeader(name, value)

  // These fields must satisfy the requirements of Stats' constructor as well as those of the stat
  // methods of this class.
  private long count = 0;
  private double mean = 0.0; // any finite value will do, we only use it to multiply by zero for sum
  private double sumOfSquaresOfDeltas = 0.0;
  private double min = NaN; // any value will do
  private double max = NaN; // any value will do

  /** Adds the given value to the dataset. */

	// and want to reduce it to the range [0,n) preserving exact uniformity.
	// We can simulate a scaling arbitrary precision x * (n/2⁶⁴) by
	// the high bits of a double-width multiply of x*n, meaning (x*n)/2⁶⁴.
	// Since there are 2⁶⁴ possible inputs x and only n possible outputs,
	// the output is necessarily biased if n does not divide 2⁶⁴.
	// In general (x*n)/2⁶⁴ = k for x*n in [k*2⁶⁴,(k+1)*2⁶⁴).

	case 2:
		exp2 += exp
	default:
		panic("unexpected exponent base")
	}
	// exp consumed - not needed anymore

	// apply exp5 contributions
	// (start with exp5 so the numbers to multiply are smaller)
	if exp5 != 0 {
		n := exp5
		if n < 0 {
			n = -n
			if n < 0 {
				// This can occur if -n overflows. -(-1 << 63) would become
				// -1 << 63, which is still negative.

        if (forward_tensor >= 0) AddUse(op, forward_tensor);
        forward_tensor = AddTensor(size);
        AddUse(op, forward_tensor);
      }

      // Backward pass: Right-multiply the jacobian from the outside in.
      // dLoss/df[n] * df[n]/df[n-1] * ...
      // The i-th term g[i] depends on g[i-1] and f[n-i] and has size |f[n-i]|.
      for (; forward_tensor >= 0; --forward_tensor) {

			out.montgomeryMul(out, out, m)

			// Select x^k in constant time from the table.
			k := uint((b >> j) & 0b1111)
			for i := range table {
				tmp.assign(ctEq(k, uint(i+1)), table[i])
			}

			// Multiply by x^k, discarding the result if k = 0.
			tmp.montgomeryMul(out, tmp, m)
			out.assign(not(ctEq(k, 0)), tmp)
		}
	}

	return out.montgomeryReduction(m)
}

	// Create command shorthand.
	if len(words) == 1 {
		g.errorf("no command specified for -command")
	}
	command := words[1]
	if g.commands[command] != nil {
		g.errorf("command %q multiply defined", command)
	}
	g.commands[command] = slices.Clip(words[2:])
}

// exec runs the command specified by the argument. The first word is
// the command name itself.