[gcc/Linux] std::numeric_limits<int>::traps = false when integer arithmetic traps

• Key: STDCXX-624
• Type: Bug
• Status: Closed
• Resolution: Fixed
• Priority: Trivial
• Assignee: Martin Sebor
• Reporter: Martin Sebor
• Component/s: 18. Language Support
• Affects Version/s: 4.2.0
• Fix Version/s: 4.2.1

Description

As an extension, stdcxx defines the value of std::numeric_limits<T>::traps for integer T to be true when there exists an arithmetic operation on type T that traps for some value of T. Based on this definition the program below should always run successfully to completion but when compiled with gcc 4.1.0 with optimization enabled on SUSE Linux Enterprise Server 10 (x86_64) it aborts at runtime. (This is also why the example program limits.cpp produces a DIFF result in some builds).

$ cat t.cpp && make t && ./t
#include <cassert>
#include <cstdio>
#include <cstdlib>
#include <limits>

char digits[] = " +0.?23456789e+01";

int main ()
{
    digits [4] = '0';
    const int zero = (int)std::strtod (digits, 0);

    digits [4] = '1';
    const int x = (int)std::strtod (digits, 0);

    if (!std::numeric_limits<int>::traps) {
        const int div = x / zero;
        const int mod = x % zero;

        std::printf ("%d, %d\n", div, mod);
    }
}
gcc -c -I/amd/devco/sebor/stdcxx-4.2.x/include/ansi   -pthread -I/amd/devco/sebor/stdcxx-4.2.x/include -I/build/sebor/stdcxx-4.2.x-gcc-4.1.0-12D/include -I/amd/devco/sebor/stdcxx-4.2.x/examples/include  -pedantic -nostdinc++ -O2   -W -Wall -Wcast-qual -Winline -Wshadow -Wwrite-strings -Wno-long-long -Wcast-align   t.cpp
gcc t.o -o t -pthread  -L/build/sebor/stdcxx-4.2.x-gcc-4.1.0-12D/lib  -Wl,-R/build/sebor/stdcxx-4.2.x-gcc-4.1.0-12D/lib -lstd12D -lsupc++ -lm 
Floating point exception

Comments

[Martin Sebor]

Reassigning to self after accidentally assigning it to Mark.

I can't reproduce this with gcc/Linux, either on AMD64 or EM64T (tried Fedora 8/EM64T, SuSE 9/AMD64 and SuSE 10/EM64T). I've added a regression test in r641091 to see if there are any platforms or configurations where this still an issue.

[Martin Sebor]

As to why the example program limits.cpp produces a DIFF result in some builds, I suspect it's because the expected output file limits.out hardcodes true as the value of numeric_limits<int>::traps while the actual value as determined by the NO_INT_TRAPS.cpp config test is false on Linux/x86_64.

[Martin Sebor]

Fixed in r589912.
Regression test committed in r641091.
Merged to 4.2.x in r589913 and r642845.