-fsingle-precision-constant breaks STL

[Marc Singer]

The upgrade to gcc-arm-none-eabi-4_9-2015q1 has changed the way that a portion of the STL compiles s.t. the -fsingle-precision-constant switch breaks compilation.

The attached file includes a compiler command line. When compiled with the 2015q1 compiler, an error is generated that includes this:

/opt/gcc-arm-none-eabi-4_9-2015q1/arm-none-eabi/include/c++/4.9.3/bits/random.tcc:1408:55: error: no matching function for call to 'max(float, const double&)'
    _M_d = std::round(std::max(6.0, std::min(__m, __dx)));

The 2014q4 compiler builds the source file without errors.

[Tony Liu (mrtoniliu)]

Hi,

The compiler actually did right thing. Two parameters of std::max() should be the same type. There are no definition of std::max() can take one float and one double.

This case is probably caused by the header file changes. The difference between 2014q4 and 2015q1 is that during pre-processing 2015q1 calls std::max() but 2014q4 didn't.

Regards,
Tony

[strongly-typed (stronglytyp3d)]

Hi,

the compiler fails with an empty main function as soon as <algorithm> is included and -fsingle-precision-constant is set.

Regards,

[Andre Vieira (andre-simoesdiasvieira)]

Hi,

As explained by Tony before, this bug is due to the STL *header* containing:

std::max (1.0, __d1x)

__d1x is a double and 1.0 is a float because of -fsingle-precision-constant. However, std::max is only defined for (float, float) or (double, double). This code is guarded by the _GLIBCXX_USE_C99_MATH_TR1 macro which was not set in our 4.8 based toolchain releases. This explains why it fails with only an include, why only since the 4.9 releases and why with -fsingle-precision-constant. STL developers might be convinced to add a (double) cast in front of the few constants there but they might as well refuse given that -fsingle-precision is not standard compliant [1]. Prefixing the 3 constants in arm-none-eabi/include/c++/4.9.3/bits/random.tcc with (double) fixes the issue.

[1] section 6.4.4.2 paragraph 4 of C11 (and similar text exist in earlier versions) explicitly states that unsuffixed floating constant have double precision.

Kind Regards,

[Xo Wang (xo)]

This looks to be fixed upstream: https://gcc.gnu.org/bugzilla/show_bug.cgi?id=67309