//
// Therefore when a shift is used as the input to a rotate then insert
// selected bits instruction we can merge the two together. We just have
// to be careful that the resultant mask is representable (non-zero and
// contiguous). For example, assuming that x is variable and c, y and m
// are constants, a shift followed by a rotate then insert selected bits
// could be represented as:
//

This op determines the maximum scale_factor that would map the initial
[input_min, input_max] range to a range that lies within the representable
quantized range.

It determines the scale from one of input_min and input_max, then updates the
other one to maximize the representable range.

e.g.

*   if the output is signed, num_bits = 8, [input_min, input_max] = [-10.0,

// Using MOVZX instead of AND is cheaper.
(AND(Q|L)const [  0xFF] x) => (MOVBQZX x)
(AND(Q|L)const [0xFFFF] x) => (MOVWQZX x)
// This rule is currently invalid because 0xFFFFFFFF is not representable by a signed int32.
// Commenting out for now, because it also can't trigger because of the is32bit guard on the
// ANDQconst lowering-rule, above, prevents 0xFFFFFFFF from matching (for the same reason)

        `tf.GradientTape` inside a `tf.function`.
    *   Changed the default step size in `gradient_checker_v2.compute_gradients`
        to be exactly representable as a binary floating point numbers. This
        avoids poluting gradient approximations needlessly, which is some cases
        leads to false negatives in op gradient tests.