Inconsistency between documentation and compiler for _subborrow_u64().

On the intrinsics reference site, the _subborrow_u64() intrinsic says it computes "b - a". But in both ICC and MSVC, the intrinsic does "a - b".

So the documentation contradicts what the intrinsic actually does in ICC and MSVC. Unfortunately, GCC's implementation of the intrinsic follows the documentation rather than ICC/MSVC's behavior. And this is causing a bit of a headache when porting some Windows code to Linux/GCC.

The following code prints 23 in ICC/MSVC. But it prints 18446744073709551593 in GCC. I don't know who's "right" or "wrong", but this inconsistency should probably be fixed.
 

#include <stdint.h>
#include <iostream>
using namespace std;

#if __GNUC__
#include <x86intrin.h>
#else
#include <intrin.h>
#endif

int main(){
    unsigned char carry = 0;

    uint64_t a = 123;
    uint64_t b = 100;
    uint64_t c;

    carry = _subborrow_u64(carry, a, b, (unsigned long long*)&c);
    cout << c << endl;

    system("pause");
}