pertifiedcobra's blog

Strange behavior of GNU C++ 17
By pertifiedcobra, history, 3 years ago, In English

1. WA Submission

2. AC Submission

The only difference between the two submissions is the cout<<"", which is there in the AC Submission and not there in WA Submission. Can someone please help me understand why printing an empty string gives a different answer than when not printing print it?

Thank you!

  • Vote: I like it
  • +5
  • Vote: I do not like it

»
3 years ago, # |
  Vote: I like it +15 Vote: I do not like it

simple answer is precision errors. never use built in ceil and floor and double. Instead if you use custom built ceil and floor functions it will pass.

→ Reply
»
3 years ago, # |
  Vote: I like it 0 Vote: I do not like it

I'd say its more looks like GCC bug

Standard not says anything about any unspecified/undefined behaviour in floating point functions. Moreover there is not any float math error as I see. (y-mid*a+0.00)/(ba+0.00) = -1e-9 and (x-mid*b+0.00)/(ab+0.00) = 1 in that test case. Godbolt says that gcc produce quite different assembly code which is not seem right:

Without cout
With cout

→ Reply
»
3 years ago, # |
  Vote: I like it +4 Vote: I do not like it

google excess precision

→ Reply
  • »
    »
    3 years ago, # ^ |
      Vote: I like it 0 Vote: I do not like it

    Could you please add a link to a good blog so I could read about it?

    → Reply
»
3 years ago, # |
← Rev. 2   Vote: I like it +13 Vote: I do not like it

Thou Shalt Not Use Floating Point Numbers, gcc's bug 323 is even mentioned in their "do not report" section.

Basically, nobody implements the IEEE-754 standard for floating point numbers to the letter. If you do, you lose lots of arithmetic properties (e.g. addition is no longer associative) and optimization opportunities. Hence, compilers deviate in different ways. Another typical example is excess precision.

As a practical consequence, you should always assume that your floating point numbers are not precise. In my original post (the first link in the comment), I incorrectly assumed that $$$\varepsilon = 10^{-9}$$$ is ok when working with numbers around $$$2\cdot10^7$$$. It is not: there is not enough precision in double that $$$N+\varepsilon \neq N$$$. I requested slightly more than 16 digits of precision, and double is less than that.

→ Reply
  • »
    »
    3 years ago, # ^ |
      Vote: I like it 0 Vote: I do not like it

    There is no excess precision: ceiling/flooring 1 / 10^9 and 1 in double type correct. The problem is that printing empty string somehow affects order of float calculations.

    → Reply