[gclist] The ML Kit with Regions?

[Hao-yang Wang]

I am trying to understand The ML Kit with Regions 
<http://www.diku.dk/research-groups/topps/activities/kit2/>.

In Oberon or Java, programmers have to make the "artificial" distinction 
between the objects on stack and the objects on heap. However, this 
effort on the programmers' part can reduce the heap usage, to the extent 
that we can achieve a good overall performance even with a classic 
mark-n-sweep garbage collector. So I think it is a great idea to do some 
compile-time analyses to the ML programs, to achieve the performance 
comparable to those stack-based languages, in terms of both speed and 
space.

However, the compiler cannot perform the "compile-time garbage 
collection" in all cases, so sometimes you end up with memory leaks. The 
ML Kit with Regions comes with a profiler, so you can monitor your 
program, catch these cases, and explicitly insert the region annotations 
to fix the leaks.

A profiler is like a debugger. We make mistakes and we need the debugger 
to help us catch bugs, but still we have to do our best to aim at a good 
design that is ideally bug-less. So I am not quite comfortable with the 
idea of catching all the memory leaks solely with the profiler: Some 
leaks may escape from the profiling sessions. You may have a library that 
works perfectly with one program, but leaks with another program.

On the other hand, tail-recursion elimination can be viewed as a kind of 
compile-time (stack space) garbage collection, but I am quite comfortable 
with tail-recursion elimination. When I code a program, I can tell myself 
"This recursive call is a tail recursion, and the compiler will optimize 
it out, so it's ok".

So my question is: Is it possible to obtain some "intuitive feelings" 
toward the region management, just like I have had with tail recursion?

Thanks,
Hao-yang Wang