c++ app crashing on non AVX CPUs

Question

I have an AVX optimized app which I do not need to make compatible with non AVX CPUs. However, I would like to display a clean error dialog on these older CPUs, rather than having the app crashing, as that causes customer confusion.

In my main() I create the QApplication instance (I'm using the Qt framework), then test for AVX using gcc __builtin_cpu_supports ("avx") . If it fails, I display an error dialog. This proved to work on non-AVX CPUs on a simple test app.

However, our (large) application crashes before displaying the dialog box on non AVX CPUs.

I have suspected:

• Global variable initialization somewhere which uses AVX intrinsics. I'm not 100% sure to have checked everywhere, but it seems this is not the case.
• gcc optimizer uses AVX instructions on some code called before the check.

Problem is, I have no non-AVX system at work for debugging and I would prefer avoiding purchasing one if possible.

1. Is it possible to disable AVX when debugging on my CPU so that debugger stops on any AVX instruction?
2. Any alternative ideas?

Answer 1

You could create a (non-optimised) wrapper program that performs the feature test, then either exits with a friendly message or execs your application.

The QApplication constructor is allowed to modify the passed arguments, so you'll need to take a copy (or not use any Qt in the path that goes through to exec() ).

Something like (from my head):

int main(int argc, char **argv)
{
    if (__builtin_cpu_supports ("avx")) {
        execv("/the/real/program", argv);
        perror("exec");
        exit 1;
    } else {
        QApplication app(argc, argv);
        QDialog d;
        d.show();
        return 1;
    }
}

Answer 2

VMWare can fake the CPUID of the guest OS, and you could just turn off all AVX bits.

However, this is just lying to the application : it won't cause the app to break into the debugger if it still executes that AVX instruction.