[clang][dataflow] Weaken abstract comparison to enable loop termination.
Currently, when the framework is used with an analysis that does not override `compareEquivalent`, it does not terminate for most loops. The root cause is the interaction of (the default implementation of) environment comparison (`compareEquivalent`) and the means by which locations and values are allocated. Specifically, the creation of certain values (including: reference and pointer values; merged values) results in allocations of fresh locations in the environment. As a result, analysis of even trivial loop bodies produces different (if isomorphic) environments, on identical inputs. At the same time, the default analysis relies on strict equality (versus some relaxed notion of equivalence). Together, when the analysis compares these isomorphic, yet unequal, environments, to determine whether the successors of the given block need to be (re)processed, the result is invariably "yes", thus preventing loop analysis from reaching a fixed point. There are many possible solutions to this problem, including equivalence that is less than strict pointer equality (like structural equivalence) and/or the introduction of an explicit widening operation. However, these solutions will require care to be implemented correctly. While a high priority, it seems more urgent that we fix the current default implentation to allow termination. Therefore, this patch proposes, essentially, to change the default comparison to trivally equate any two values. As a result, we can say precisely that the analysis will process the loop exactly twice -- once to establish an initial result state and the second to produce an updated result which will (always) compare equal to the previous. While clearly unsound -- we are not reaching a fix point of the transfer function, in practice, this level of analysis will find many practical issues where a single iteration of the loop impacts abstract program state. Note, however, that the change to the default `merge` operation does not affect soundness, because the framework already produces a fresh (sound) abstraction of the value when the two values are distinct. The previous setting was overly conservative. Differential Revision: https://reviews.llvm.org/D123586
Loading
Please sign in to comment