[PATCH] D46368: [analyzer] pr18953: Zeroing constructors, store binding extents, ASTRecordLayouts.

[Artem Dergachev via Phabricator via llvm-commits]

NoQ added a comment.

> We could work around that by invalidating object contents every time we detect something fishy. The only reasonable detection i can come up with would be to detect if we have a coinciding binding (i.e. what the assertion was previously protecting us from), eg.:
> 
>   if (B.getDefaultBinding(R) || B.getDirectBinding(R))
>     V = UnknownVal();
> 
> 
> This detection is imperfect - there are tests that would still fail. In particular it won't fix the `ASTRecordLayout` problem. And i also suspect that it really destroys a lot more good data than bad data. So i'll be experimenting with that now. In the current version of the patch i remove the special case handling and simply bind the default value blindly even if i know that `RegionStore` can't handle it precisely.

Committed as is for now. Didn't notice any difference between the three approaches on my C++ testing grounds.


Repository:
  rL LLVM

https://reviews.llvm.org/D46368