On Tuesday, 30 August 2016 at 03:37:06 UTC, Chris Wright wrote:
On Mon, 29 Aug 2016 18:20:05 +0000, tsbockman wrote:
They should. Generally speaking, if that doesn't produce
reasonable bounds (leaving aside rounding errors) at the end
of the computation, it means that the logic of the computation
itself is wrong.
The ranges expand very fast. Addition and subtraction double
the range, multiplication squares it, exponentiation is n^^n.
Larger bounds are usually not useful bounds.
Ranges don't always grow. Some operations will also cause them to
shrink, if they're really being tracked correctly: bitwise AND,
integer division, and modulus all tend to do so.
If the ranges are intrinsically significant (it's logically
impossible for a creature to have -7 eyes) and the calculation is
meaningful and correct, things should balance out at the end to
produce a useful range for the final result.
On the other hand, arbitrary sanity check ranges (it's merely
extremely unlikely for a creature to have 2147483647 eyes)
shouldn't be passed on to intermediate values at all; they must
be assigned manually based on human intuition, and this is only
really worth doing for inputs and outputs. In such cases, the
best return value for binary operations would be normal
(conceptually) unbounded CheckedInt.
Give me mathematically correct bounds, or no bounds at all.
Anything else will cause problems.
