[clang] Thread Safety Analysis: Drop most call-based alias invalidation (PR #190154)

[Marco Elver via cfe-commits]

melver wrote:

> Your reasoning is interesting, but we don't really know that the reason for a cast is some kind of reallocation. It could be anything. Another problem is that this makes the analysis less predictable. If we invalidate on the basis of casts, users will have no idea what's going on when they change the code and the warning pops up (or disappears).
> 
> Not saying no, but I'm not quite convinced that this is the best way.

Replied here: https://github.com/llvm/llvm-project/pull/187691#issuecomment-4235756125

TLDR; because passing `Foo**` to another incompatible type (e.g. `void**`) requires an explicit cast, this rule is self-documenting. We aren't guessing at the callee's behavior; we are respecting the developer's explicit signal that the pointer's type safety - and thus its tracked identity - is being suspended at that boundary. IMHO, it's a clean, predictable rule for the analysis.

If you're still not convinced, we can narrow it to the concrete examples I was using: at the very least I'd invalidate aliases after passing a pointer-to-alias with`void` casts.

https://github.com/llvm/llvm-project/pull/190154