Why is compiler optimisation breaking my code?

Question

I was having some issues with this code. It was working in debug mode but would erroneously output 0.0 for some of the csv lines in release mode. It was only when I added the preprocessor commands that turn off optimisation around line 19 that everything started to work in release mode. Does anyone have any insight into why this is the case, I've never come across this sort of behavior before.

#include <vector>
#include <complex>
#include <fstream>

size_t constexpr kBins = 64;
double constexpr kFrequency = 16.0;

double Triangle(double const bin)
{
    return abs(bin - floor(bin + 1.0 / 2.0));
}

int main()
{
    std::vector<std::complex<double>> input{kBins, {0.0, 0.0}};
    for (size_t i = 0; i < kBins; ++i)
    {
#pragma optimize("" off)
        double value = sin(2.0 * M_PI * kFrequency * Triangle(static_cast<double>(i) / kBins)) / (2.0 * M_PI * kFrequency * Triangle(static_cast<double>(i) / kBins));
#pragma optimize("" on)
        input[i] = fpclassify(value) == FP_NAN ? 1.0 : value;
    }

    std::ofstream output_file{"output.csv"};
    if (output_file.is_open())
    {
        for (size_t i = 0; i < kBins; ++i)
        {
            output_file << (static_cast<double>(i) / kBins) << ", " << input[i].real() << ", " << input[i].imag() << std::endl;
        }
        output_file.close();
    }
}

Please provide a minimal reproducible example. When you feel like the optimizer is breaking your code, 99% of the time you have UB, 0.9% of the time is when you have misconstrued assumptions on floating point accuracy. The remaining is a compiler bug. — Passer By
– Passer By, Commented Dec 2, 2020 at 22:05
Clang. The minimal reproducible example is just when the #pragma lines are removed. — Joshua Williams
– Joshua Williams, Commented Dec 2, 2020 at 22:08
@JoshuaWilliams The ofstream stuff is irrelevant (unless you are being rounded down by the formatting), and you don't need 64 values to see one is broken. — Passer By
– Passer By, Commented Dec 2, 2020 at 22:10
Simplify your giant expression on line 19 into sub-expressions, and add checks for each of them to ensure they're correct. Turn on the strictest compiler warning level as well. Usually, "my code doesn't work when optimized" types of errors mean you've used an undefined value somewhere, or modified a value within an expression (between sequence points) — Tumbleweed53
– Tumbleweed53, Commented Dec 2, 2020 at 22:13

alter_igel · Accepted Answer · 2020-12-02 22:27:12Z

7

Your input vector is not being initialized as you expect. It always has size two, and all the strange results you observe are a result of Undefined Behaviour due to reading the vector out of bounds. You could have found this with your debugger, or by checking against input.size() or using input.at(i).

It turns out that size_t is convertible to std::complex<double>, and that this line:

std::vector<std::complex<double>> input{kBins, {0.0, 0.0}};

Is constructing a vector with two elements, even though kBins = 64! It appears the vector constructor taking a std::initializer_list<std::complex<double>> is being chosen and that kBins has been implicitly converted to std::complex<double>.

One way to get around this is to initialize your vector differently, using the auto keyword to avoid the Most Vexing Parse and using parentheses instead of curly braces to avoid accidentally passing a std::initializer_list:

auto input = std::vector<std::complex<double>>(kBins, {0.0, 0.0});`

With this change, the (range-checked) code runs without errors.

edited Dec 2, 2020 at 22:27

answered Dec 2, 2020 at 22:12

alter_igel

7,2293 gold badges24 silver badges43 bronze badges

Sign up to request clarification or add additional context in comments.

2 Comments

Passer By Over a year ago

You don't get the vexing parse because of {0.0, 0.0}.

alter_igel Over a year ago

@PasserBy good catch, although it's a good habit to be aware of MVP and to know how to avoid it

Collectives™ on Stack Overflow

Why is compiler optimisation breaking my code?

1 Answer 1

2 Comments

Your Answer

Hot Network Questions

Collectives™ on Stack Overflow

1 Answer 1

2 Comments

Your Answer

Sign up or log in

Post as a guest

Related