Notes to users

1. The GC is a conservative mark-and-sweep collector. It only runs a collection cycle when an allocation is requested of it, never otherwise. Hence, if the program is not doing allocations, there will be no GC collection pauses. The pauses occur because all threads the GC knows about are halted so the threads' stacks and registers can be scanned for references to GC allocated data.
2. The GC does not know about threads that were created by directly calling the OS/C runtime thread creation APIs and D threads that were detached from the D runtime after creation. Such threads will not be paused for a GC collection, and the GC might not detect references to GC allocated data held by them. This can cause memory corruption. There are several ways to resolve this issue:
   1. Do not hold references to GC allocated data in such threads.
   2. Register/unregister such data with calls to addRoot/removeRoot and addRange/removeRange.
   3. Maintain another reference to that same data in another thread that the GC does know about.
   4. Disable GC collection cycles while that thread is active with disable/enable.
   5. Register the thread with the GC using core.thread.osthread.thread_attachThis/core.thread.threadbase.thread_detachThis.

Notes to implementors

• On POSIX systems, the signals SIGRTMIN and SIGRTMIN + 1 are reserved by this module for use in the garbage collector implementation. Typically, they will be used to stop and resume other threads when performing a collection, but an implementation may choose not to use this mechanism (or not stop the world at all, in the case of concurrent garbage collectors).
• Registers, the stack, and any other memory locations added through the GC.$(LREF addRange) function are always scanned conservatively. This means that even if a variable is e.g. of type float, it will still be scanned for possible GC pointers. And, if the word-interpreted representation of the variable matches a GC-managed memory block's address, that memory block is considered live.
• Implementations are free to scan the non-root heap in a precise manner, so that fields of types like float will not be considered relevant when scanning the heap. Thus, casting a GC pointer to an integral type (e.g. size_t) and storing it in a field of that type inside the GC heap may mean that it will not be recognized if the memory block was allocated with precise type info or with the GC.BlkAttr.$(LREF NO_SCAN) attribute.
• Destructors will always be executed while other threads are active; that is, an implementation that stops the world must not execute destructors until the world has been resumed.
• A destructor of an object must not access object references within the object. This means that an implementation is free to optimize based on this rule.
• An implementation is free to perform heap compaction and copying so long as no valid GC pointers are invalidated in the process. However, memory allocated with GC.BlkAttr.$(LREF NO_MOVE) must not be moved/copied.
• Implementations must support interior pointers. That is, if the only reference to a GC-managed memory block points into the middle of the block rather than the beginning (for example), the GC must consider the memory block live. The exception to this rule is when a memory block is allocated with the GC.BlkAttr.$(LREF NO_INTERIOR) attribute; it is the user's responsibility to make sure such memory blocks have a proper pointer to them when they should be considered live.
• It is acceptable for an implementation to store bit flags into pointer values and GC-managed memory blocks, so long as such a trick is not visible to the application. In practice, this means that only a stop-the-world collector can do this.
• Implementations are free to assume that GC pointers are only stored on word boundaries. Unaligned pointers may be ignored entirely.
• Implementations are free to run collections at any point. It is, however, recommendable to only do so when an allocation attempt happens and there is insufficient memory available.