gcc-vec, glibc, LFS, and 2.4.0t9


Subject: gcc-vec, glibc, LFS, and 2.4.0t9
From: Chris Bednar (cjb@AdvancedDataSolutions.com)
Date: Sun Sep 24 2000 - 09:35:12 MDT


Here's my situation:

  1) I'm running a 2.4.0-test9 kernel on my G4.

  2) I would like to make full use of LFS support.

  3) I believe this would be easier with the latest glibc.

  4) I seem to need the latest gcc to build the latest glibc.

  5) I don't want to give up AltiVec for all this.

Here's what I have to date:

  1) Kernel works OK (better than on Intel, as far as I can tell;
    kudos to Paul Makerras) despite some hangs with heavy memory
    use, and panics during shutdown when trying to stop NFS (which
    I can usually stop by hand just fine?!?!?).

  2) I can create Large Files, and I can write my own code to
    cat them, but things like `ls' and `tar' will insist on
    being able to fstat() them, and fail; I think this is
    because they never hit fstat64() in the kernel (at least
    that's what I remember from some time ago).

  3) Haven't built glibc-2.1.9X yet.

  4) I managed to patch up gcc-2.96 to compile, and built ppc RPMs,
    but haven't installed them yet, because....

  5) The altivec patch will not be trivial (for me) to apply to 2.96,
    and although I imagine I could do it, I would rather put effort
    into getting Motorola off its butt to reassign copyright. Doesn't
    it seem like they ought to want to do whatever they can to encourage
    people to BUY THEIR HARDWARE? I would love to put in an order for
    1000 PowerPC's instead of 1000 Pentium3's, but it's just so damned
    hard to develop for them.

Let me put it as clearly as I can. Our next-gen software is very
CPU hungry. We're going to need to buy a lot of hardware to keep
it fed, and our clients will, too. Without AltiVec, it runs almost twice
as fast clock-for-clock on a G4 as on a p3. With an AV-enhanced FFT,
we might get a factor of 4, which I think would make it worth getting
into the game (why isn't ``almost 2'' good enough? not quite as stable,
{Price + devel effort}/performance, etc.).

    Anyone have any info on recent developments? Of course, an
easier path to LFS+AltiVec would also be appreicated.

----
Chris J. Bednar   <http://optics.tamu.edu/~bednar/>
Director, Distributed Computing Product Group
http://AdvancedDataSolutions.com/



This archive was generated by hypermail 2a24 : Sun Sep 24 2000 - 09:38:07 MDT