MLton

In region-based memory management, the heap is divided into a collection of regions into which objects are allocated. At compile time, either in the source program or through automatic inference, allocation points are annotated with the region in which the allocation will occur. Typically, although not always, the regions are allocated and deallocated according to a stack discipline.

MLton does not use region-based memory management; it uses traditional GarbageCollection. We have considered integrating regions with MLton, but in our opinion it is far from clear that regions would provide MLton with improved performance, while they would certainly add a lot of complexity to the compiler and complicate reasoning about and achieving SpaceSafety. Region-based memory management and garbage collection have different strengths and weaknesses; it’s pretty easy to come up with programs that do significantly better under regions than under GC, and vice versa. We believe that it is the case that common SML idioms tend to work better under GC than under regions.

One common argument for regions is that the region operations can all be done in (approximately) constant time; therefore, you eliminate GC pause times, leading to a real-time GC. However, because of space safety concerns (see below), we believe that region-based memory management for SML must also include a traditional garbage collector. Hence, to achieve real-time memory management for MLton/SML, we believe that it would be both easier and more efficient to implement a traditional real-time garbage collector than it would be to implement a region system.

Regions, the ML Kit, and space safety

The ML Kit pioneered the use of regions for compiling Standard ML. The ML Kit maintains a stack of regions at run time. At compile time, it uses region inference to decide when data can be allocated in a stack-like manner, assigning it to an appropriate region. The ML Kit has put a lot of effort into improving the supporting analyses and representations of regions, which are all necessary to improve the performance.

Unfortunately, under a pure stack-based region system, space leaks are inevitable in theory, and costly in practice. Data for which region inference can not determine the lifetime is moved into the "global region" whose lifetime is the entire program. There are two ways in which region inference will place an object to the global region.

  • When the inference is too conservative, that is, when the data is used in a stack-like manner but the region inference can’t figure it out.

  • When data is not used in a stack-like manner. In this case, correctness requires region inference to place the object

This global region is a source of space leaks. No matter what region system you use, there are some programs such that the global region must exist, and its size will grow to an unbounded multiple of the live data size. For these programs one must have a GC to achieve space safety.

To solve this problem, the ML Kit has undergone work to combine garbage collection with region-based memory management. HallenbergEtAl02 and Elsman03 describe the addition of a garbage collector to the ML Kit’s region-based system. These papers provide convincing evidence for space leaks in the global region. They show a number of benchmarks where the memory usage of the program running with just regions is a large multiple (2, 10, 50, even 150) of the program running with regions plus GC.

These papers also give some numbers to show the ML Kit with just regions does better than either a system with just GC or a combined system. Unfortunately, a pure region system isn’t practical because of the lack of space safety. And the other performance numbers are not so convincing, because they compare to an old version of SML/NJ and not at all with MLton. It would be interesting to see a comparison with a more serious collector.

Regions, Garbage Collection, and Cyclone

One possibility is to take Cyclone’s approach, and provide both region-based memory management and garbage collection, but at the programmer’s option (GrossmanEtAl02, HicksEtAl03).

One might ask whether we might do the same thing — i.e., provide a MLton.Regions structure with explicit region based memory management operations, so that the programmer could use them when appropriate. MatthewFluet has thought about this question

Unfortunately, his conclusion is that the SML type system is too weak to support this option, although there might be a "poor-man’s" version with dynamic checks.