Extend -Wdangling to handle assignments, not just initializations

[higher-performance]

Currently, -Wdangling/-Wdangling-gsl detect the following:

std::string foo() { return ""; }
int main() {
    std::string_view v = foo();  // error: object backing the pointer will be destroyed at the end of the full-expression [-Werror,-Wdangling-gsl]
}

However, they fail to detect this case:

std::string foo() { return ""; }
int main() {
    std::string_view v;
    v = foo();  // no error
}

It seems the compiler currently only handles initializations, not regular assignments, despite them posing the same issue. So it would be great to handle the second case as well.

More generally, if a lifetimebound object is assigned to a non-temporary expression, then I believe it would be correct to issue the diagnostic.

[shafik]

CC @Xazax-hun

[Xazax-hun]

+1, 100% agreed. In fact, MSVC's /analyze can warn about the code in the example. Unfortunately, I might not be able to work on this in the near future, but I am always happy to review PRs.

[shafik]

+1, 100% agreed. In fact, MSVC's /analyze can warn about the code in the example. Unfortunately, I might not be able to work on this in the near future, but I am always happy to review PRs.

Do you think you could give pointers to what changes would be needed to support this?

[Xazax-hun]

Sure!

The current implementation is in clang/lib/Sema/SemaInit.cpp, one can follow the uses of IndirectLocalPathEntry around
https://github.com/llvm/llvm-project/blob/main/clang/lib/Sema/SemaInit.cpp#L7207 to get an idea what is going on.

Unfortunately, the code is somewhat convoluted in the name of performance. The check is done while the AST is being built, so lifetime extension, handling of lifetimebound, and the -Wdangling-gsl all interleaved with some other Sema stuff making the code somewhat harder to follow, and definitely harder to share between initialization and assignment.

I am not convinced that the potential performance benefits would worth the cost of complexity here, it might be better to factor the lifetimebound and -Wdanlging-gsl logic out to some common place and reuse it both for initialization and assignment.

As for where to put the check for the assignments, one could take a look at Sema::CheckAssignmentOperands, but I am not 100% confident whether that is the best place to do this sort of checking.

[frederick-vs-ja]

Duplicate of #54492?

[Xazax-hun]

#46984 also seems to be a duplicate of this.