This page tracks the workarounds for the various compiler issues that we encountered in the development. This is mostly of interest for developers interested in contributing to xtensor.
Visual Studio 2015 and
With Visual Studio,
std::enable_if evaluates its second argument, even if
the condition is false. This is the reason for the presence of the indirection
in the implementation of the
Visual Studio 2017 and alias templates with non-class template parameters and multiple aliasing levels¶
Alias template with non-class parameters only, and multiple levels of aliasing
are not properly considered as types by Visual Studio 2017. The base
xcontainer template class underlying xtensor container types has such alias
templates defined. We avoid the multiple levels of aliasing in the case of Visual
Visual Studio and
Visual Studio defines
max macros causing calls to e.g.
std::max to be interpreted as syntax errors. The
NOMINMAX definition may be used to disable these macros.
In xtensor, to prevent macro replacements of
max functions, we
wrap them with parentheses, so that client code does not need the
Visual Studio 2017 (15.7.1) seeing declarations as extra overloads¶
xvectorize.hpp, Visual Studio 15.7.1 sees the forward declaration of
vectorize(E&&) as a separate ovarload.
GCC-4.9 and Clang < 3.8 and constexpr
std::max are not constexpr in these compilers. In
xio.hpp, we locally define a
XTENSOR_MIN macro used instead of
std::min. The macro is undefined right after it is used.
Clang < 3.8 not matching
initializer_list with static arrays¶
Old versions of Clang don’t handle overload resolution with braced initializer lists correctly: braced initializer lists are not properly matched to static arrays. This prevent compile-time detection of the length of a braced initializer list.
A consequence is that we need to use stack-allocated shape types in these cases.
Workarounds for this compiler bug arise in various files of the code base.
Everywhere, the handling of Clang < 3.8 is wrapped with checks for the
GCC < 5.1 and
The versions of the STL shipped with versions of GCC older than 5.1 are missing
a number of type traits, such as
However, for some of them, equivalent type traits with different names are
provided, such as
In this case, we polyfill the proper standard names using the deprecated
std::has_trivial_default_constructor. This must also be done when the
compiler is clang when it makes use of the GCC implementation of the STL,
which is the default behavior on linux. Properly detecting the version of the
GCC STL used by clang cannot be done with the
__GNUC__ macro, which are
overridden by clang. Instead, we check for the definition of the macro
_GLIBCXX_USE_CXX11_ABI which is only defined with GCC versions greater than
GCC-6 and the signature of
We are not directly using
std::isinf for the
xt::isinf, as a workaround to the
following bug in GCC-6 for the following reason.
- C++11 requires that the
- C99 requires that the
These two definitions would clash when importing both headers and using namespace std.
As of version 6, GCC detects whether the obsolete functions are present in the
<math.h> and uses them if they are, avoiding the clash. However,
this means that the function might return int instead of bool as C++11
requires, which is a bug.