This patch uses the reference to MaterializeTemporaryExpr stored in the construction context since D43477 in order to model that expression correctly, following the plan described in http://lists.llvm.org/pipermail/cfe-dev/2018-February/056898.html
MaterializeTemporaryExpr is an expression that takes an rvalue object (or one of its xvalue sub-objects) and transforms it into a reference to that object. From the analyzer's point of view, the value of the sub-expression would be a NonLoc, and the value of MaterializeTemporaryExpr would be a Loc that this NonLoc is stored in.
For now this behavior has been achieved by creating a Loc out of thin air and re-binding the NonLoc value into the newly created location. This, however, is incorrect, because the constructor that has constructed the respective NonLoc value was operating on a different memory region, so all of the subsequent method calls, including the destructor, should have the exact same "this" region that the constructor was using.
The problem with behaving correctly is that we've already forgot the original storage. It is technically impossible to reliably recover it from the NonLoc.
The patch addresses it by relying on the ConstructionContext to inform the constructor that lifetime extension is taking place, so that the constructor could store the current "this" region in the program state for later use. The region is uniquely determined by the MaterializeTemporaryExpr itself. Then when time comes to model MaterializeTemporaryExpr, it looks up itself in the program state to find the respective region, and avoids the hassle of creating a new region and copying the value in case it does indeed find the existing correct region.
The temporary region's liveness (in the sense of removeDeadBindings) is extended until the MaterializeTemporaryExpr is resolved, in order to keep the store bindings around, because it wouldn't be referenced from anywhere else in the program state.
Situations where correct lifetime extension behavior is needed require relatively awkward test cases, because most of the time you don't care about the object's address as long as its contents are correct.
This patch is enough to assert that for every initialized temporary (lifetime-extended or not), a destructor is called (with the correct address) on all existing test cases(!). I added a new test case (with BinaryConditionalOperator) on which this assertion still fails (due to lack of construction context and other problems) - i've actually seen this operator used on C++ objects. There are more cases that we don't handle yet; i'd see what i can do about them.