integer arithmetic overflow

[StefanKarpinski]

To deal with integer arithmetic overflow we do the following:

1. add checked integer arithmetic intrinsics named checked_add, checked_mul, etc.
2. have a compiler switch that uses these everywhere.
3. write a @check_overflow macro that transforms an expression to turn * into checked_mul and + into checked_add, etc., but only in the expression — called functions are on their own.

[StefanKarpinski]

Note that by "compiler switch", I just mean a command-line option that choses between different sets of definitions for core integer arithmetic. Kind of awesome that that's all this takes to change the way core arithmetic works.

[timholy]

Is this still relevant, or does the commit fix it?

FYI, I recently added a "morebits" function to make it easier to write code that avoids integer multiplication overflow (used in nextprod and prevprod), not sure whether this was a good idea or not.

[pao]

The commit 54e40e3 appears to address (1), but not (2) or (3).

[ViralBShah]

There are also those who would prefer saturation arithmetic instead of overflow. At least that's what Matlab users get.

[ViralBShah]

Can we do this without a compiler switch? At the risk of code bloat, how about a SafeInt type?

[JeffBezanson]

There are multiple features here; the purpose of the switch is to instrument all code for debugging. If you know ahead of time that certain numbers or certain operations need to handle overflow somehow, then you want a separate mechanism for that.

[werediver]

Please, clarify whether the following behaviour is related to this issue or should be reported as a new one:

julia> 2^50, (2^50)*(2^50), 2^100
(1125899906842624,0,0)

But

julia> 1/0
Inf

It's rather confusing, isn't it?

julia> versioninfo()
Julia Version 0.2.1
Commit e44b593* (2014-02-11 06:30 UTC)
Platform Info:
  System: Windows (x86_64-w64-mingw32)
  WORD_SIZE: 64
  BLAS: libopenblas (USE64BITINT DYNAMIC_ARCH NO_AFFINITY)
  LAPACK: libopenblas
  LIBM: libopenlibm

Thanks.

[andreasnoack]

Int/Int isn't integer division. It promotes to floats, 1/1=1.0.

[mbauman]

That behavior is very much intentional and working as intended. Julia currently always uses native machine arithmetic. This issue is to allow some sort of command-line or compilation switch to change that behavior, but native arithmetic will almost certainly always be the default.

There's a very good explanation in the documentation on why Julia uses native machine integer arithmetic. It discusses this very clearly.

The behavior of 1/0 is orthogonal to this. Inf is infinity in a floating point representation, but 2^100 is not a floating point number; integer exponentiation always returns an integer.