}
			nbyte := n / 8
			for i := uintptr(0); i < nbyte; i++ {
				bits |= uintptr(*p) << nbits
				p = add1(p)
				*dst = uint8(bits)
				dst = add1(dst)
				bits >>= 8
			}
			if n %= 8; n > 0 {
				bits |= uintptr(*p) << nbits
				p = add1(p)
				nbits += n
			}
			continue Run
		}

		// Repeat. If n == 0, it is encoded in a varint in the next bytes.
		if n == 0 {

// CHECK-DAG:  %[[CONSTANT1:.*]] = arith.constant dense<0.000000e+00> : tensor<4xf32>
//              variance + epsilon
// CHECK:  %[[ADD1:.*]] = "tf.Add"(%[[ARG4:.*]], %[[CONSTANT]])
//              rsqrt(variance + epsilon)
// CHECK:  %[[RSQRT:.*]] = "tf.Rsqrt"(%[[ADD1]])
//              scale * rsqrt(variance + epsilon)
// CHECK:  %[[MUL1:.*]] = "tf.Mul"(%[[ARG1:.*]], %[[RSQRT]])