[CMake] C++ standard version fallbacks.

[Roger Leigh]

On 05/06/17 23:50, Stephen Kelly wrote:
> Craig Scott wrote:
> 
>> On Tue, Jun 6, 2017 at 7:50 AM, Stephen Kelly
>> <steveire at gmail.com> wrote:
>>
>>> Roger Leigh wrote:
>>>
>>>> Hi folks,
>>>>
>>>> I'm currently using this logic to use C++14 with a fallback to C++11
>>>> when C++14 is unavailable:
>>>>
>>>>     if(NOT CMAKE_CXX_STANDARD)
>>>>       set(CMAKE_CXX_STANDARD 14)
>>>>     endif()
>>>>     if(NOT CMAKE_CXX_STANDARD_REQUIRED)
>>>>       set(CMAKE_CXX_STANDARD_REQUIRED 11)
>>>>     endif()
>>>>
>>>> which seems to work OK.
>>>>
>>>> However, for some new stuff, I'd like to use C++17 when available, but
>>>> fall back to C++14, C++11 or C++98.  Is it possible to do this?
>>>
>>> Probably set CMAKE_CXX_STANDARD
>>>
>>> without CMAKE_CXX_STANDARD_REQUIRED (That variable doesn't really make
>>> sense
>>> to me and I think it is overused when not needed).
>>>
>>
>> If you don't set CMAKE_CXX_STANDARD_REQUIRED, then there's no guarantee
>> you get any particular minimum standard.
> 
> He wants to fall back all the way to C++98. Am I missing something?

No, that's exactly what I want.  I'd like to have the compiler put into 
the highest mode possible and then do feature tests for various things. 
For example, try std::thread and fall back to boost or POSIX/Win32 
threads if not available.  Not setting CMAKE_CXX_STANDARD_REQUIRED seems 
to be all that's required, so that solves my problem nicely.

>> Roger's example (sorry Roger!)
>> highlights part of the confusion about this latter variable (and the
>> target property it ultimately controls). He appears to be setting it
>> expecting it to specify a minimum version, but that's not how it works. It
>> is expected to be a boolean which says whether CMAKE_CXX_STANDARD must be
>> honoured or not, which most developers (myself included) tend to find
>> unintuitive.
> 
> Ok.

Yes, I misread the documentation here, and I think I also picked it up 
by reading someone else's (broken) example.  I've now fixed up my code 
to work according to the documented behaviour.  Thanks everyone for the 
corrections and suggestions in this thread.

> I remember I was opposed to introducing CMAKE_CXX_STANDARD_REQUIRED in the
> first place as I think it is redundant. I recommend populating compile
> features for whatever you absolutely need and let cmake populate the std
> flag. If your code can benefit from a more-recent std flag than the
> requirement, then set CMAKE_CXX_STANDARD to that.
> 
>> Roger's use would actually make it a bit better, if that was
>> how it worked, but unfortunately there's currently no way to set a
>> *minimum* standard version,
> 
> If you have a minimum, then you must be relying on some language features
> existing and you can list those.

This would make sense, and I'll look at doing that.

I'm still not 100% sure that the behaviour is as good as it could be 
though.  Some scenarios:

1) I set CMAKE_CXX_STANDARD to 17.  This is unsupported and it falls 
back to 14.  Is the fallback introspectable?  Can I query that the 
fallback happened/is going to happen?

2) I do a check for <cstdint> which succeeds, and I use it in a 
configuration header, but this later fails to build because it's only 
usable in C++11 mode, and if I set CMAKE_CXX_STANDARD to 98 it fails.

I know the *requested* standard version.  But I don't know the 
*effective* standard version, and I need to know that at cmake time.

Much of the standard version selection is based upon setting the minimum 
set of required compile features and having cmake put the compiler in a 
mode appropriate to support that featureset.  What I'd like to do here 
is the opposite: I want to dynamically adjust the features I use by 
gracefully falling back to alternative implementations based upon my own 
feature testing.  But I can't see how I can ask that question of CMake 
even by getting the appropriate target properties; these look like they 
are the requested version as well, do they ever get turned into the 
effective version?

I'm sure I can work around the lack with some more sophisticated feature 
tests if required, e.g. compile checks rather than header checks.


Thanks,
Roger