Smart pointer constructors, assignments, and `reset()` can actually have fairly large side effects, especially with custom deleters, destructors for objects being destroyed, and so on. I've re-introduced `{AllAliased}` side effects for these functions. There was no immediate effect on analysis results.
`DataFlowFunction` models treat references a pointers - an explicit level of indirection. The AST dataflow library generally treats references as if they were the referred-to object. This commit removes a workaround in the dataflow model for unary `operator*` on smart pointers, and makes the AST dataflow library adjust the results of querying the model so that a returned reference only gets flow that was modeled as going to the dereference of the return value.
This fixes some missing flow in IR dataflow, and recovers some (presumably) missing reverse taint flow in AST taint tracking as well.
**Revert the revert** of the workaround for CFG issues when a
`FunctionCall` has a `getTarget` that does not exist. While we've fixed
the main cause of the problem, it can apparently still happen in rare
cases as a result of extractor crashes.
This reverts commit ee5eaef5e4.
We have special code to handle field flow for single-field structs, but that special case was too specific. Some `Store`s to single-field structs have no `Chi` instruction, which is the case that we handled already. However, it is possible for the `Store` to have a `Chi` instruction (e.g. for `{AllAliased}`), but still have a use of the result of the `Store` directly. We now add a `PostUpdateNode` for the result of the `Store` itself in those cases, just like we already did if the `Store` had no `Chi`.
This change was necessary for my upcoming changes to introduce side effect instructions for indirections of smart pointers. The code to decide which parameters have which side effects appeared in both the IPA constructor for `TTranslatedSideEffect` and in `TranslatedCall`. These two versions didn't quite agree, especially once the `SideEffectFunction` model provides its own side effects instead of the defaults.
The relevant code has now been factored out into `SideEffects.qll`. This queries the model if one exists, and provides default side effects if no model exists. This fixes at least one existing issue, where we were emitting a buffer read side effect for `*this` instead of an indirect read side effect. This accounts for all of the IR diffs in the tests.
