Intel 17.0.2 on UNIX introduced a regression from 17.0.1 in its
definition of `__cpp_constexpr` in `-std=c++14` mode. It incorrectly
defines it as `200704` instead of the expected `201304`. Fix our
feature detection table to account for this.
The `__cplusplus` macro is not defined to the language level reliably.
When simulating MSVC it is always set to `1`. When simulating GCC it is
set to the lower value of the current language level and the maximum
level supported by the simulated version of GCC (from observation).
For MSVC we already used a combination of `__INTEL_CXX11_MODE__` and
`__cpp_aggregate_nsdmi` to detect C++14 mode. Extend this for general
use by removing the condition on `_MSC_VER`.
This condition needs to follow the same pattern added in note `[1]` by
commit a5a3642f (Features: Port Intel CXX features to test macros where
possible, 2016-10-26). It was accidentally left out of that commit.
Since this compiler always defines `__cplusplus` to `1` we need to use
`_MSC_VER`, `__INTEL_CXX11_MODE__`, and the feature test macro named
`__cpp_aggregate_nsdmi` to detect C++11 and C++14 modes.
With no `-Qstd=` flag this compiler defaults to C++98 plus a subset of
C++11/C++14 features needed to be compatible with MSVC. We pretend it
is plain C++98 and add a `-Qstd=` flag whenever needed for C++11 or
above features even if they would happen to be available in MSVC-mode.
Closes: #16384
Versions below 12.1 do not provide enough information to properly detect
if compiling with c++98 or c++0x enabled so remove them from the
supported list.
