3 ms·

Don't be ridiculous. It's perfectly fine if you know what you're doing. For example, when you have a limited selection of floats (e.g. hashtable keys, or only n

by tomp 11y ago

Don't be ridiculous. It's perfectly fine if you know what you're doing. For example, when you have a limited selection of floats (e.g. hashtable keys, or only numbers between 0 and 100, accurate to 1 decimal) - in these cases, strings would work just as well, but floats are more efficient (and compare correctly).

jacquesm 11y ago
You've just redefined your problem to using floats as the underlying representation. In that case you don't even need floats... In most other cases the correct way is to check if the absolute value of the difference between one value and another is smaller than some arbitrary cut-off, a small number typically denoted with 'epsilon'.
- tomp 11y ago
  I don't need floats, but they're useful. E.g. for comparison, as I wrote above.
  - jacquesm 11y ago
    Until your code gets ported to a platform where float == double and it turns out the algo feeding your clever efficient comparison delivers slightly different values that previously collapsed into the same bucket. Relying on the particulars of an underlying representation requires deep knowledge of how those particulars are implemented (and I would happily agree with you that such knowledge is a requirement but more often than not when you come across a float comparison with equality you're looking at a bug. That you're the exception to the rule is fine with me, you seem to know what you're doing so don't take it personal). The real problem is: most people don't have deep knowledge of how those details are implemented and take 'equality' for granted, especially if the visual representation of the numbers they are comparing tells them they should be equal which is not always the case. The old adage goes: if you need floating point you don't understand your problem :) The comment that sparked this sub-thread is a nice illustration of what can and does go wrong.

by tomp 11y ago

jacquesm 11y ago

You've just redefined your problem to using floats as the underlying representation. In that case you don't even need floats... In most other cases the correct way is to check if the absolute value of the difference between one value and another is smaller than some arbitrary cut-off, a small number typically denoted with 'epsilon'.

tomp 11y ago
I don't need floats, but they're useful. E.g. for comparison, as I wrote above.
- jacquesm 11y ago
  Until your code gets ported to a platform where float == double and it turns out the algo feeding your clever efficient comparison delivers slightly different values that previously collapsed into the same bucket. Relying on the particulars of an underlying representation requires deep knowledge of how those particulars are implemented (and I would happily agree with you that such knowledge is a requirement but more often than not when you come across a float comparison with equality you're looking at a bug. That you're the exception to the rule is fine with me, you seem to know what you're doing so don't take it personal). The real problem is: most people don't have deep knowledge of how those details are implemented and take 'equality' for granted, especially if the visual representation of the numbers they are comparing tells them they should be equal which is not always the case. The old adage goes: if you need floating point you don't understand your problem :) The comment that sparked this sub-thread is a nice illustration of what can and does go wrong.