RE: libgcj and incremental gc

[Andrew Haley <aph at redhat dot com>]

Boehm, Hans writes:
 > Thanks for trying this.  This was on X86?  I wonder if the trampolines
 > just lose execute permission?

I don't think so.  What seems to happen is that when I jump to a
trampoline, it has turned into a string.  It could be, of course, that
some insane piece of code has written a string to the wrong place in
memory, but I reckon that it's a good guess the trampoline has been
reclaimed.

 > I would not in general turn this on by default.  But it might be a
 > useful option.  The vague plan I had was to add some mechanism for
 > turning this on only if both:
 > 
 > - the user asks, and
 > - the page size is equal to HBLKSIZE
 > 
 > It may be that the GC_ENABLE_INCREMENTAL environment variable should
 > acquire this meaning.  It currently doesn't check the second
 > condition.
 > 
 > The second condition is necessary to ensure that pointer-free objects
 > are never write protected.

Yes, I can understand that one.

 > I think the hard part in making this work is that we would have to
 > ensure that system calls only write to pointer-free objects.  I
 > think that's hard only because I'm not sure there's an easy way to
 > find and check all of them.

I can't think of any case in Java where a system call writing to a
Java object that contains pointers would not be a bug.

My interest here is to get some proper numbers about performance
issues.

Andrew.