thread-local data access

[Richard Henderson]

On Wed, Sep 27, 2000 at 05:25:21PM +0200, schwidefsky@de.ibm.com wrote:
> Access register 0 is unused anywhere so we could just use this one.
> We can't address data directly with this register but we can load it to a
> normal register in one cycle. I will take a look at the pthread
> implementation and give it a try.

This is very similar to Alpha's situation (though it takes more
like 5-10 cycles to get into PALmode and back out again with 
the value.  Certainly it's faster than playing stack games
(which wouldn't even work with the stuff we're now discussing)).

To be sure, ideally you'd teach gcc about this register, and how to
move data to and from.  You'd mark it fixed, since you're reserving
it for the thread ABI.  Then you can implement pthreads with

  register struct _pthread_descr_struct *__thread_self __asm__("acc0");
  #define THREAD_SELF  __thread_self

Now gcc can properly schedule the code, which would not be true
with inline assembly.


r~