[libcxx-commits] [PATCH] D58201: Make std::memory_order an enum class (P0439R0)

[Zoe Carver via Phabricator via libcxx-commits]

zoecarver marked an inline comment as done.
zoecarver added a comment.

@EricWF I have resolved all your comments except for making `memory_order` scoped (which I can do after more discussion if still necessary).



================
Comment at: include/atomic:603
+
+#endif
 
----------------
zoecarver wrote:
> jfb wrote:
> > EricWF wrote:
> > > jwakely wrote:
> > > > zoecarver wrote:
> > > > > jfb wrote:
> > > > > > jwakely wrote:
> > > > > > > jfb wrote:
> > > > > > > > I think you want to keep the old `typedef enum memory_order` before C++20, and only enable the new thing in C++20 and later.
> > > > > > > That's what we did for libstdc++.
> > > > > > You did what I'm suggesting, or what @zoecarver did? 🤔
> > > > > Changing how `memory_order ` is defined changes how some of the code below is written -- if it is defined differently for different version then the code below also needs to be changed for different versions (actually a `static_cast` would work in both cases but it is only necessary in the latter). 
> > > > We did what you suggested: keeping it unchanged for C++11/14/17 and only making it a scoped enumeration (and adding the new enumerators) for C++20.
> > > > 
> > > > And we cast to `int` unconditionally, because as zoecarver said, that always works, even if the cast is redundant for C++11/14/17.
> > > Although I know it's not conforming. I would prefer to make it a scoped enum retroactively. In the past 3 months, I've had 3 separate users get burned by the implicit conversion to int. 
> > > 
> > I'd rather not do this: WG21 would have made the paper a defect and applied its resolution retroactively. I think we should conform here, and we could instead teach clang-tidy to catch this issue.
> For what its worth, I would also rather not break the standard. What if we sort of compromised and left `memory_order`  strongly typed but made all of `memory_order_relaxed` (etc.) ints so they could be backward compatible. I know it is not directly solving the issue you brought up but it would give a source of truth people could use and it would allow pre C++20 code to compile unmodified. 
I guess the issue with that is things which are defined as `memory_order` still couldn't accept ints. 


CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D58201/new/

https://reviews.llvm.org/D58201