5

Prehistory:

We have just switched our developing environment to VC++ 2015 from VC++ 2008. Right after that we have found the issue: program raised division by 0 despite testing divider in C++ code.

Test code:

#include <cstdlib>

int test(int n, int test_for_zero)
{
    int result = 0;
    for (int i = 0; i < n; ++i) {

        if (test_for_zero)
            result += rand() >> ((8 % test_for_zero) * test_for_zero);
    }

    return result;
}

int main()
{
    return test(rand(), rand() & 0x80000000);
}

Compiled by VC++ 2015 Update 3 or VC++ 2017 with default Release options, on run it raises division by zero. Compiled by VC 2008 runs just fine.

Analysis:

; 6    :    for (int i = 0; i < n; ++i) {

    test    edi, edi
    jle SHORT $LN15@main

; 7    : 
; 8    :        if (test_for_zero)
; 9    :            result += rand() >> ((8 % test_for_zero) * test_for_zero);

    mov ecx, DWORD PTR _test_for_zero$1$[ebp]
    mov eax, 8
    cdq
    idiv    ecx  ; <== IT'S HERE, IDIV BEFORE CHECKING FOR ZERO
    mov eax, edx
    imul    eax, ecx
    mov DWORD PTR tv147[ebp], eax
    test    ecx, ecx
    je  SHORT $LN15@main
$LL11@main:
    call    ebx
    mov ecx, DWORD PTR tv147[ebp]
    sar eax, cl
    add esi, eax
    sub edi, 1
    jne SHORT $LL11@main

The compiler takes constant part ((8 % test_for_zero) * test_for_zero) out of loop body and just forget for testing test_for_zero before division. Obviously it can be fixed easily at place just by doing the compiler job but correctly.

I've played with several compiler options, like -d2SSAOptimizer- and -Oxx, but the only option to fix this is -Od.

Questions:

  1. Is it a bug or not? The C++ standard was heavily changed since VC 2008 so may it be affected like this?
  2. The main question, is there any workaround to fix the issue thru compiler options, except -Ob?
Ilya Ivanov
  • 51
  • 2
  • This doesn't answer your question, but wouldn't it make more sense to skip the loop altogether if that variable is 0? You never alter it, so if it was 0 in one iteration, it will be 0 for all of them, in which case the loop doesn't do anything. – Carcigenicate Nov 13 '17 at 16:21
  • Also, what happens if you change the condition to check against 0 instead of just checking if the number is truthy? Shouldn't matter, but this error doesn't make sense. – Carcigenicate Nov 13 '17 at 16:24
  • Could be the lack of braces too. Idk how exactly c++ parses statements like that. Try adding braces. – Carcigenicate Nov 13 '17 at 16:25
  • 2
    It is a bug. Hard to provide a workaround beyond #pragma optimize, too synthetic. In VS2017 use Help > Send Feedback > Report a Problem. – Hans Passant Nov 13 '17 at 17:02

1 Answers1

1

Thanks for the report and the small repro (especially feedback from within Visual Studio!), they were very helpful in isolating the problem.

This is indeed a bug in the compiler itself, specifically a safety check in an "Invariant Code Motion" optimization pass. I've made the fix to this, which should appear in VS 15.7.

For now, the easiest workaround is to disable optimizations for functions which have this code pattern via https://learn.microsoft.com/en-us/cpp/preprocessor/optimize:

    #pragma optimize( "", off )  
    <Function containing code with this loop>
    #pragma optimize( "", on )

Unfortunately this optimization pass is tied directly to the optimizer, so there's no compiler option outside of completely disabling the optimizer (-Od) to work around this as you've discovered.

Di Mo
  • 11
  • 1