[m-users.] MR_GC_free ?

[Peter Wang]

On Tue, 07 Nov 2023 20:17:06 +0100 Volker Wysk <post at volker-wysk.de> wrote:
> Hi
> 
> Am Mittwoch, dem 08.11.2023 um 05:53 +1100 schrieb Zoltan Somogyi:
> > On 2023-11-08 05:48 +11:00 AEDT, "Volker Wysk" <post at volker-wysk.de> wrote:
> > > Hi.
> > > 
> > > In the ODBC library, there's space allocated by MR_GC_NEW and explicitly
> > > freed with MR_GC_free. This confuses me. Shouldn't space allocated by
> > > MR_CG_NEW be deallocated by the garbage collector? If it is to be
> > > deallocated explicitly, shouldn't the space be allocated by MR_NEW and be
> > > deallocated by MR_free?
> > 
> > Follow the definition of MR_GC_free in runtime/mercury_memory.h
> > to boehm_gc/include/gc.h.
> 
> This leads to this (I couldn't understand the definition of this function):
> 
> /* Explicitly deallocate an object.  Dangerous if used incorrectly.     */
> /* Requires a pointer to the base of an object.                         */
> /* An object should not be enabled for finalization (and it should not  */
> /* contain registered disappearing links of any kind) when it is        */
> /* explicitly deallocated.                                              */
> /* GC_free(0) is a no-op, as required by ANSI C for free.               */
> GC_API void GC_CALL GC_free(void *);
> 
> So the gc explicitly frees the object and remembers that it has done so, so
> it won't be garbage collected later.
> 
> Is it bad style to use MR_GC_NEW and MR_GC_free, since it should rather be
> done with MR_NEW/MR_free? (That's what those are there for, aren't they?)

They do different things.

MR_GC_NEW allocates memory that will be garbage collected when there are
no more references to it (as far as the GC can see). MR_GC_free
deallocates the memory immediately, before a GC cycle, even if there
might actually be references to it hanging around. Rarely, there MAY be
performance advantages to doing so, e.g. if you allocate a large object
and definitely don't need it any more, and don't want to wait for one or
more GC cycles for that memory to be reclaimed.

MR_NEW is just a wrapper around the C library's malloc routine.
The GC doesn't see inside inside malloc-allocated memory. You should
generally not use MR_NEW or any variant of malloc to allocate memory for
an object that may contain pointers to GC-allocated memory. If the last
reference(s) to a GC-allocated object only exist within memory allocated
by malloc then, as far as the GC is concerned, that GC-allocated object
is no longer referenced and may be reclaimed.

In a Mercury program, you'll want to allocate GC-able memory
unless there is a specific reason not to.

Peter