Constexpr: floating-point representation errors?

Question

I was hoping to convert the constant from degrees to radians (at compile time), so I decided to use constexpr. However, my program will not compile, so I tried to debug the problem with several tests. These tests continue to cause compile-time errors.

The problem seems to correlate with floating point arithmetic when many significant numbers are involved.

I tried a quick Google search and I read section 10.4 (Constant Expressions) in the Stroustrup book. Any help would be greatly appreciated. I have to miss something obvious.

Test code:

void testConstantExpressions() { constexpr double x0 = 1.0; constexpr double y0 = 2.0; constexpr double z0 = 4.0; constexpr double w0 = x0 / (y0 / z0); std::cout << w0 << std::endl; constexpr double x1 = 1.0; constexpr double y1 = 2.2; constexpr double z1 = 4.0; constexpr double w1 = x1 / (y1 / z1); std::cout << w1 << std::endl; constexpr double x2 = 1.0; constexpr double y2 = 4.0; constexpr double z2 = 2.3; constexpr double w2 = x2 / (y2 / z2); std::cout << w2 << std::endl; } 

Compiler:

 g++ -Wall -c -g -O2 -std=c++11 -frounding-math main.cpp -o main.o main.cpp: In function 'void testConstantExpressions()': main.cpp:30:32: error: '(1.0e+0 / 5.5000000000000004e-1)' is not a constant expression constexpr double w1 = x1 / (y1 / z1); ^ main.cpp:36:38: error: '(4.0e+0 / 2.2999999999999998e+0)' is not a constant expression constexpr double w2 = x2 / (y2 / z2); ^ make: *** [main.o] Error 1 

Answer 1

This is because you specified -frounding-math. You told the compiler that you can change the rounding mode at runtime, so it cannot round at compile time. Did you really mean it?