Bug 15836 – [REG 2.071-b1] memory error when a class is not implicitly destructed and constructed with new

Created attachment 1592 source + dub description to reproduce the issue - download the attachement - unzip, cd to the dir - try the dub package from the attachement with - dub --config=strange - dub --config=ok1 - dub --config=ok2 - dub --config=ok3 without explicit destruction (strange config) the following error appears, at the end of the run-time: > core.exception.InvalidMemoryOperationError@src/core/exception.d(693): Invalid memory operation This error never appeared before and the class exists since 20 monthes, which leads me to think that there's a regression, maybe related to the GC or at least in druntime.

still present in 2.071-b2

The trouble is that the destructor of MemoryStream is calling GC.addrOf in the finalization stage of a collection, and there is a new check that explicitly disallows this. At the sweep stage, the GC is in an intermediate state that might not yield sensible results for GC.addrOf and friends. I think using this as an indicator whether a block of memory is allocated by the GC depends on its exact implementation (block immediately "freed" after finalization?), and it might restrict the inner workings of the GC too much if we make this a requirement.

Ok I close. The problem can be solved without using AddrOf.

Please post reduced test cases as inline code here, rather than providing zip projects with build files. With 2.071.0 calls to GC.free from a finalizer http://dlang.org/changelog/2.071.0.html are allowed (they will be ignored), so you can use that to replace the usage of addrOf.