Implementing TLS: step 1

[Terry Lambert]

Marcel Moolenaar wrote:
> On Thu, Jun 19, 2003 at 11:39:42PM -0400, Daniel Eischen wrote:
> > > Q2: Given that libthr has been ported to ia64, what needs to be done
> > >     to port libkse to ia64 (roughtly)?
> >
> > libpthread/arch/i386/include/ksd.h:
> > libpthread/arch/i386/include/atomic_ops.h
> > libpthread/arch/i386/include/pthread_md.h
> > libpthread/arch/i386/i386:
> 
> No kernel code? Hmmm, looks like something that's worth giving a
> shot. If we have libkse on ia64, we can more easily work on the
> issues related to TLS...


You guys are missing something.

Daniel's earlier point about the IA64 documentation not
specifying that the TP point to TLS, rather than a data
structure that contains a pointer to TLS is correct (from
my reading of both references posted by Marcel).

On the other hand, zeroing the TP register in user space
to indicate that upcalls are disabled is probably illegal,
from the point of vie of that same documentation, so there
needs to be another way to do the deed.

The problem I forsee here is that it's possible that the
SMT cores assign special significance to this register when
doing instruction reordering inside the pipeline, and this
would definitely break any benefits you would otherwise have
obtained from using the register for real threads, if that's
the case.

Note that this predominantly applies only on UP machines; any
MP machines will have a statistical chance of 1 in <num CPU>
of breaking the pipelining anyway (IMO, Intel is amining SMT
machines at the UP market, more than the SMP market, and SMP
users should not expect much benefit: negaffinity will win
you more cycles than 1/2-a-CPU-core-when-the-other-1/2-is-idle).

-- Terry