[-Wunsafe-buffer-usage] Check standalone count-attributed assignments

[patrykstefanski]

This is an initial part of an analysis of count-attributed assignment groups. This commit adds an AST visitor that is responsible for finding bounds-attributed assignment groups and standalone assignments to bounds-attributed objects (pointers and dependent counts).

As a PoC, this commit adds checks for standalone assignments, which are assignments that are not directly inside of a compound statement (like other assignment groups) and modify the pointer or count in some way. Our model rejects those and requires the user to simplify their code. For example:

  void foo(int *__counted_by(count) p, int count) {
    q = p = ...;
          ^ this is rejected
    n = count = ...;
              ^ this is rejected
    // the following is fine:
    p = ...;
    count = ...;
  }

rdar://161607826

[patrykstefanski]

This is taken from the original PR: #11490

[patrykstefanski]

Copy-paste from the original PR to continue the discussion:

q = p = q;     // warn: assignment to bounds-attributed pointer 'p' must be inside of a bounds-attributed group in a compound statement

I feel like the diagnostic message may confuse users. What is a bounds-attributed group? How about just say

assignment to bounds-attributed pointer 'p' must be a simple statement 'p = RHS;' followed by a simple statement 'count = RHS;'. 

When the count expression is a constant, we hide followed by a simple statement 'count = RHS;'.

I agree that my initial diagnostic needs rework. I think your diagnostic is better. Though, does 'p' must be a simple statement 'p = RHS;' imply that { p = RHS; ... } (assignment directly in compound) is fine but q = p = RHS; is not? I feel like we should provide some info that the assignment cannot be nested.

[patrykstefanski]

I updated the warning to:

assignment to count-attributed pointer 'p' must be a simple statement 'p = ...'