warn_unused_result attribute on class definition

classic Classic list List threaded Threaded
1 message Options
Reply | Threaded
Open this post in threaded view

warn_unused_result attribute on class definition

James Y Knight via cfe-dev
Responses inline, note that Aaron is likely thebest one to :

I think this is because of the behavior outlined here.

  /// \brief Returns true if this function or its return type has the
  /// warn_unused_result attribute. If the return type has the attribute and
  /// this function is a method of the return type's class, then false will be
  /// returned to avoid spurious warnings on member methods such as assignment
  /// operators.
  bool hasUnusedResultAttr() const { return getUnusedResultAttr() != nullptr; }
[Keane, Erich] I don't really see a situation where the exception for spurious warnings is correct.  Assignment operators typically return a reference, thus this is suppressed anyway.  Either way, this is a much bigger hammer than it should use, I suspect we should just special-case the ones we wish to suppress.  What is particularly annoying about this suppression is that it will only work for MEMBER operators, not all operators, so it is wrong in the first place!! Looking through operators only the following would be subject to the warning anyway:

1- Post increment/decrement <-- warning is possibly invalid?
2- arithmetic operators  <-- warning is probably valid.

Looking through history, it seems that the suppression of the warning on references/pointers was added AFTER this, so I wonder if simply removing the "getCorrespondingMethodInClass" condition is the right thing?  If we decide situation 1 above should be excepted, we should specifically except IT, for both member and non-member.


On Mon, Apr 17, 2017 at 11:48 PM, Craig Topper <[hidden email]>

> It seems if warn_unused_result is put on a class definition it doesn't
> apply to any of the methods in that class that return the class type?
> This test should generate 2 warnings but only generates one.
> https://godbolt.org/g/kImv9k
> The ArrayRef and APInt classes both do this. I accidentally created a
> case where I didn't consume the result of APInt::trunc and received no warning.
> ~Craig

cfe-dev mailing list
[hidden email]