As I understand it, MLKit is entirely region-based memory management and the
Stalin Scheme compiler uses similar techniques and this is fundamentally
different from both reference counting and mark and sweep GC.

What are the implications of this technique and how well do the compilers
perform in practice? Does it suffer from stalls or keeping values alive too
long?

Jon Harrop ffconsultancy.com> writes:

Region inference can do a lot, but I don't see how it's possible to
have no GC at all, and support typical functional programming style.
E.g., using functional maps as association tables that are "updated" a
lot by recursion, creates tons of garbage whose structure is entirely
data dependent and can't be predicted by region inference.

JHC (experimental Haskell compiler that is beating GHC at some
benchmarks) also uses region inference and is only now in the process
of getting a GC.

On Sat, 05 Apr 2008 07:55:14 +0100, Jon Harrop ffconsultancy.com>
wrote:

Region management is "mark-release" ... essentially a stack of heaps.
The unit of reclamation is the heap - objects are (usually) allocated
as needed but all the objects within a heap are deallocated all at
once. An object's lifetime determines to which heap it is assigned
and good management depends crucially on the compiler being able to
determine close bounds on the object's lifetime.

Jon Harrop ffconsultancy.com> writes:

The ML-kit later got GC added on, but you can switch this off and use
regions exclusively.

There are definitely space-leak issues with pure region-based
management. In my experience, you need to profile your code and
insert region annotations and a few explicit deep copies to get rid of
space leaks. This isn't as difficult as it may sound, and you
typically only have to do it in a few places, but you certainly can't
expect off-the-shelf ML programs to have good memory performance with
regions.

That said, once you get the annotations in place, ML programs with
regions usually require less memory than with GC (unless you keep the
heap so small that you trash it).

torbenm@app-2.diku.dk (Torben Ægidius Mogensen) writes:

Does that impose a particular style? What if you use functional data
structures extensively?

Paul Rubin <http://phr.cx@NOSPAM.invalid> writes:

I haven't seen the source code for the ITU administrative system, so I
can't really comment on its style. But I have tried some small
examples on the ML-kit with regions (without GC) and found that
functional data structures are not a problem, but that you sometimes
have to make a deep copy of a result of a lengthy computation to put
it in another region than that of the now-dead intermediate values for
the computation, so the latter can be reclaimed.

Torben