[Bug c++/43284] New: Explicit casting of double to long long causes value to overflow

[Bug c++/43284] New: Explicit casting of double to long long causes value to overflow

[ullner at gmail dot com]

When casting a double to a long long causes GCC to overflow the output value.
Imagine; 
#include <iostream>
double foo()  { return 9223372036854775807.0; }
int main() {
std::wcout << (long long)foo(); // Prints -9223372036854775808
}

Additionally, std::pow(double, int) and std::pow(double, double) seem to behave
differently, presumably for the same reason;
#include <iostream>
#include <cmath>
int main() {
std::wcout << (long long)std::pow(2.0, 64); // -9223372036854775808
std::wcout << (long long)std::pow(2.0, 64.0); // 9223372036854775807
}

Note that this behaviour is not the same if the return value of foo(),
std::pow() (in both cases) are stored in variables first.

Compiling with "g++ -Wall -ansi -std=gnu++0x -pedantic test.cpp", with GCC
4.4.1. Note that compiling with "g++ -Wall -ansi -std=gnu++0x -pedantic -O3
test.cpp" does NOT exhibit this behaviour; foo() and both std::pow() results in
9223372036854775807.

Note that (unsigned long long)std::pow(2.0, 64) == 0, and (unsigned long
long)std::pow(2.0, 64.0) == 18446744073709551615.


-- 
           Summary: Explicit casting of double to long long causes value to
                    overflow
           Product: gcc
           Version: 4.4.1
            Status: UNCONFIRMED
          Severity: normal
          Priority: P3
         Component: c++
        AssignedTo: unassigned at gcc dot gnu dot org
        ReportedBy: ullner at gmail dot com


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=43284

[Bug c++/43284] Explicit casting of double to long long causes value to overflow

[schwab at linux-m68k dot org]

------- Comment #1 from schwab at linux-m68k dot org  2010-03-07 22:07 -------
9223372036854775807 has more significant bits than what fits in a double, it is
rounded to 9223372036854775808.0, which then overflows when converted to long
long.


-- 

schwab at linux-m68k dot org changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
             Status|UNCONFIRMED                 |RESOLVED
         Resolution|                            |INVALID


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=43284

[Bug c++/43284] Explicit casting of double to long long causes value to overflow

[ullner at gmail dot com]

------- Comment #2 from ullner at gmail dot com  2010-03-07 22:10 -------
Hm? How does calling std::pow with different types behave differently? The
value can be stored fine if one does "double dValue = std::pow(2.0, 64);long
long llValue = dValue;" // OK


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=43284

[Bug c++/43284] Explicit casting of double to long long causes value to overflow

[rguenth at gcc dot gnu dot org]

------- Comment #3 from rguenth at gcc dot gnu dot org  2010-03-08 10:39 -------
(In reply to comment #2)
> Hm? How does calling std::pow with different types behave differently? The
> value can be stored fine if one does "double dValue = std::pow(2.0, 64);long
> long llValue = dValue;" // OK

Integer exponents have an overload that uses an implementation that does
multiple roundings.


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=43284

[Bug c++/43284] Explicit casting of double to long long causes value to overflow

[ullner at gmail dot com]

------- Comment #4 from ullner at gmail dot com  2010-03-08 17:59 -------
That still doesn't make sense.
1. Why does enabling -O3 (O1 and O2 does the same) remove this problem?
2. Why does storing the value in an intermediate variable make any change in
what the result is?

Consider without O3:
(long long)std::pow(2.0, 64); // Overflow.
(long long)std::pow(2.0, 64.0); // OK

double d1 = std::pow(2.0, 64);
(long long)d1; // Overflow
double d2 = std::pow(2.0, 64.0);
(long long)d2; // Overflow

long long l1 = d1; // l1 Overflows
long long l2 = d2; // l2 Overflows

Enabling -O3 causes every one of those that overflows to be OK. 


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=43284