Nice TWiki > Dev > CheckedIntegerArithmetic (r1.3) TWiki webs:
Dev | Doc | Main | TWiki | Sandbox
Dev . { Changes | Index | Search | Go }
The current behaviour of numeric type is to silently overflow. In some cases this is what is needed, and in some cases you don't really care, because it is guaranteed no overflow will occur. However, there are also situations in which you would like a runtime exception to be thrown when overflows/underflows occur (hello, Isaac!).

The JVM does not directly support this arithmetic mode. However, at least for non-long types, it is possible to perform operations on a larger JVM type than the declared type, and then check for overflows. This has of course a small performance hit, but much less than using objects (BigInteger? for instance). So it might be possible to add new types that have overflow detection semantic and good performance. I'll focus on int, but this should be doable for shorter types too. What cases are useful in practice?

Since we want to retain the current semantics for int, we should add a new type, with the meaning "a 32 bit signed integer not modulo 2**32" (that is, which throws an exception when an overflow occurs). I'm not sure what the best name for it would be. integer is a possibility, but I was also considering it for infinite length integers (that is an alias for BigDecimal?). Other ideas: checkedInt (not pretty), checked-int (better, but a new convention for type names), cint (cryptic). Let's use cint in this document, it's short :-)

Regarding typing, int should be a subtype of cint, because you gain "safety" by considering an int as a cint. On the other hand, going from cint to int would later allow overflows to occur, so it should not be automatic. I would also think that cint should be a subtype of long, does that make sense?

Regarding implementation, one possibility is to represent cint as long in the JVM, and to check the result of each operation. Another possibility is to represent cint values as int, but operations on cint values would be performed on long, and the result checked before conversion back to int. The former might be more efficient (less conversions), while the later is more economous in space, which is especially interesting if you have an array of cint. Alternatively, is it possible to detect all/most overflows by working on int only (for instance, for addition, if both numbers are positive, there is overflow if the number is negative)?

-- DanielBonniot - 21 Dec 2003

Overflow detection is possible with int for addition and substraction by sign check. For multiplication, i am not quite shure if even long would be enough.

Topic CheckedIntegerArithmetic . { Edit | Attach | Ref-By | Printable | Diffs | r1.12 | > | r1.11 | > | r1.10 | More }
Revision r1.3 - 22 Dec 2003 - 12:32 GMT - TWikiGuest Copyright © 1999-2003 by the contributing authors. All material on this collaboration platform is the property of the contributing authors.
Ideas, requests, problems regarding TWiki? Send feedback.