C++ is bad: structs aren't the same as structs

[big_bad_al]

Here's a fun little problem. Suppose you've got a system that makes use of the following struct:

struct User {
unsigned int user_id;
unsigned short access_level;
float account_balance; // stored as US dollars
}Your system is ridiculously well-tested, all the tests pass, everything works totally fine. I'll be so bold as to say it's (kinda)

Comments

Re: Heh ISO floating point...

[big_bad_al]

Precisely! Using floating point numbers for money can land you in federal pound-me-in-the-ass prison. Addition and subtraction are noncommutative, which can lead to catastrophic cancellation (add a million dollars, add one cent, then subtract a million dollars, and you're left with nothing) and round-off errors when calculating exchange rates can cause similar woes (convert from dollars to euros to dollars again, and the balance is different). On top of that, doubles take up 64 bits in RAM but 80 bits in the FPU (see "extended precision"), which means that floating point comparison can be nondeterministic (depending on what else needs the FPU and what optimizations the compiler is trying, there isn't a good way to know how many bits of precision a given double will have).

If you need to work with money, store it as an integer to ensure that addition, multiplication, and comparisons work correctly. If you need to deal with non-integer amounts of money, just make your units sufficiently small (for instance, you could store everything in microcents). The important part is to ensure that you will neither create nor destroy money as you process it.

Re: Heh ISO floating point...

[big_bad_al]

Wait, I think floating point operations are still commutative. However, they're not associative, which is what leads to the problems.

Re: Heh ISO floating point...

[big_bad_al]

You may be right about that. I believe ISO actually specifies commutativity for floating point ops, though associativity is definitely out.

Re: Heh ISO floating point...

[ahhunter]

I'm pretty sure floating point subtraction isn't commutative. :)