David Miller authored Feb 26, 2009 and Brian Paul committed Feb 26, 2009

This rewrites the sparc GLAPI code so that it's PIC friendly and works
with all of the TLS/PTHREADS/64-bit/32-bit combinations properly.

As a result we can turn SPARC asm back on.  Currently it's only
enabled on Linux, as that's the only place where I can test this
stuff out.

For the moment the cliptest SPARC asm routines are disabled as they
are non-working.  The problem is that they use register %g7 as a
temporary which is where the threading libraries store the thread
pointer on SPARC.  I will fix that code up in a future change as it's
a pretty important routine to optimize.

Like x86 we do the runtime patch as a pthread once-invoked initializer
in init_glapi_relocs().

Unlike x86, however, our GLAPI stubs on SPARC are just two instruction
sequences that branch to a trampoline and put the GLAPI offset into a
register.  The trampoline is what we run-time patch.  The stubs thus
all look like:

glFoo:
	ba		__glapi_sparc_foo_stub
	 sethi		GLAPI_OFFSET(glFOO) * PTR_SIZE, %g3

This actually makes generate_entrypoint() a lot simpler on SPARC.  For
this case in generate_entrypoint() we generate stubs using a 'call'
instead of the 'ba' above to make sure it can reach.

In order to get a proper tail call going here, in the unpatched case,
we do several tricks.  To get the current PC, for example, we save the
return address register into a temporary, do a call, save the return
address register written by the call to another temporary, then
restore the original return address register value.  This is to
avoid having to allocate a stack frame.

This is necessary for PIC address formation.

This new GLAPI scheme lets us get rid of the ugly SPARC GLAPI hacks in
__glXInitialize() and one_time_init().

Signed-off-by: David S. Miller <davem@davemloft.net>