编译器的不同结果：（gcc 4.8.1）和gcc（4.3.2）

Question

I submitted same solution for problem on an online judge on different compilers for C++.On gcc(4.3.2),I got WA, while same solution when submitted on gcc(4.8.1) got TLE.

Is it that 4.3.2 is faster but I think performance wise latest versions should outperform previous ones or is it the floating point anomalies in two compilers because problem require calculation of nth root of a 64-bit number and I am using long double and long long data types with pow function.I used something like:

    long long root,n;
    long double rad,rcnd;
    root = (long long)pow(rad,rcnd); where rcnd = 1.0/n;

Answer 1

Two versions of a same elementary function such as pow() only need each to be accurate to >0.5ULP to sometimes produce different results with the same arguments.

The C and C++ standards do not place any constraints on the accuracy on pow() . A proper implementation will try to be accurate to 1ULP , but that still leaves the possibility of the answer not being the best one and being different from the answer of another 1-ULP-accurate pow() function. Actually, several questions on this site are caused by pow() functions that are (in)accurate to more than 1ULP , however ugly it may be and despite the ginormous means available to the provider of the inaccurate function.

So, in short: if you use pow() , the computations may differ between two compilers, although the compilers have the same implementation-defined characteristics (size of int, endianness, ...). If you use pow() in a numerically instable computation, the end results may differ arbitrarily.

The usual solution for reproducible results is to provide your own pow() function. However, if the difference is caused by numerically instable computations, this does not fix the root problem: the result, now reproducible, may still be meaningless.