Trying to understand ggc-page vs. ggc-simple

[Richard Henderson]

On Sun, Jan 28, 2001 at 01:05:50PM -0800, Michael Sokolov wrote:
> Could someone please explain to me the difference between ggc-page and
> ggc-simple and the logic behind them?

ggc-page uses something akin to vm page tables in order to track
pages allocated by the collector.  We have a bit less overhead
per object that way, and it's also a bit faster to mark the 
objects during collection.

ggc-simple has 3 or 4 words of overhead per object (depending on
required alignment), plus whatever overhead your malloc needs.
During collection it must look up each object in an approximately
ballanced tree.

> Back in December when I was looking into this, ggc-page required either
> mmap or valloc, otherwise ggc-simple was chosen. 4.3BSD does have valloc,
> though, which is why ggc-page got chosen.

ggc-page no longer uses valloc.  We now use malloc directly for
systems that do not cannot mmap anonymous memory.  This is quite
a bit more efficient than valloc, since in this scheme we waste
no more than 1 page in 16, whereas valloc tends to waste 1 in 2.

> ... but I got a rude awakening around this New Year when ggc-page,
> which configure picked by default, overflowed my 64 MB systemwide VM
> space while compiling fold-const.c in stage 2.

Try adding a call to release_pages in ggc_collect after sweep_pages.
Normally we keep hold of the "freed" pages under the assumption that
we're going to reuse them again shortly.  It's possible that is not
a winning strategy for malloc, since it might badly fragment the heap.


r~