30 ms·

Can you give an example of math that would be exact with 32-bit integers, but is inexact in JavaScript? Floating-point math on values that happen to be integers

Can you give an example of math that would be exact with 32-bit integers, but is inexact in JavaScript? Floating-point math on values that happen to be integers is exact, if an exact answer exists and fits within the mantissa.

slededit 8y ago
I think we're approaching this from philosophically different directions. On your end if the programmer controls everything correctly then its safe because doubles have a 53 bit significand. The problem happens when you treat numbers one way and they are treated differently elsewhere. For example -0 when undergoing a bitwise +/- check (val & (1U<<31)) will appear positive when its negative by definition. The cast to an integer prior to the bitwise operator looses information. You can use the operators safely if you know there is no information to be lost, but it is not type checked like with C. I will say at least JavaScript guarantees twos compliment. You never know when your favourite compiler will decide to break your code in the name of a faster benchmark - correctness of existing programs be damned. Implementation defined used to mean do something sane, but I digress.
- tripletao 8y ago
  The trick is to let your "double" only take on values that are also valid 32-bit integers, like with lots of gratuitous (x|0) and stuff. That excludes your (indeed problematic) -0. JavaScript's semantics give you enough to implement 32-bit integer math this way pretty easily. The speed may vary dramatically with small changes to the compiler, but correctness won't.
  - slededit 8y ago
    I still stick by my original point that: ((int)dblval) & flag would require serious justification in a C code review. In JS its all you have. Under those restrictions I'd use it as rarely as possible. But yes you can use it correctly if you try.
    - tripletao 8y ago
      That's certainly true. The only reason to play these games is that JavaScript gives you no better option.

slededit 8y ago

I think we're approaching this from philosophically different directions. On your end if the programmer controls everything correctly then its safe because doubles have a 53 bit significand. The problem happens when you treat numbers one way and they are treated differently elsewhere. For example -0 when undergoing a bitwise +/- check (val & (1U<<31)) will appear positive when its negative by definition. The cast to an integer prior to the bitwise operator looses information. You can use the operators safely if you know there is no information to be lost, but it is not type checked like with C. I will say at least JavaScript guarantees twos compliment. You never know when your favourite compiler will decide to break your code in the name of a faster benchmark - correctness of existing programs be damned. Implementation defined used to mean do something sane, but I digress.

tripletao 8y ago
The trick is to let your "double" only take on values that are also valid 32-bit integers, like with lots of gratuitous (x|0) and stuff. That excludes your (indeed problematic) -0. JavaScript's semantics give you enough to implement 32-bit integer math this way pretty easily. The speed may vary dramatically with small changes to the compiler, but correctness won't.
- slededit 8y ago
  I still stick by my original point that: ((int)dblval) & flag would require serious justification in a C code review. In JS its all you have. Under those restrictions I'd use it as rarely as possible. But yes you can use it correctly if you try.
  - tripletao 8y ago
    That's certainly true. The only reason to play these games is that JavaScript gives you no better option.