[Top][All Lists]

[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

Re: build fails on alphaev67-dec-osf5.1

From: Dave Love
Subject: Re: build fails on alphaev67-dec-osf5.1
Date: Thu, 14 Aug 2003 15:23:36 +0100
User-agent: Gnus/5.1003 (Gnus v5.10.3) Emacs/21.2 (gnu/linux)

"Stefan Monnier" <monnier+gnu/emacs/address@hidden> writes:

> Also can you show us a copy of src/config.h to see what code
> of alloc.c gets compiled ?

This may not be pristine against current sources.

Attachment: config.h
Description: Text document

By the way, Irix in 64-bit mode fails similarly.

> They are intended to get rid of the markbit so as to reduce the
> number of bits eaten by tags from 4 to 3 so that they can then
> be moved from the top 4 bits to the bottom 3 bits (which are always
> 0 on 64bit systems and almost always 0 on 32bit systems).

> IIUC this is mostly orthogonal to the use of something like the Boehm GC.

No, it dispenses with mark bits and uses a separate bitmap instead.
README.GC on the branch says:

  The mark bit in objects should be dispensed with under Boehm GC.
  (Doing that doubles the range of Lisp integers/heap pointers).

The conventional way of tagging in the low-order bits ups the space
taken by conses, which would be pretty significant in my sessions.  As
far as I remember, it also caused bad problems for XEmacs.  I would
expect to experiment with it when the Boehm GC was basically working.
Now the GC is modified to cope with Emacs's current tagging scheme,
you can play about with this if the codebase is made to work properly.

How does using 3 low tag bits work when you can't guarantee that

> I personally don't know enough about the Boehm GC to know what benefit
> it could provide.

Generations, incremental collection, conservative scanning on all
systems of interest as far as I know, probably better marking (WRT
paging, especially), probably better use of free pages, possibly
running out of heap address space slower by allowing some mallocing to
be done from the top of address space with mmap, other people
(experts) maintaining it ...

reply via email to

[Prev in Thread] Current Thread [Next in Thread]