testcase for hppa64 gcc bug

[Alan Modra]

On Mon, 30 Oct 2000, John David Anglin wrote:

> > 	* rtl.h (ARG_POINTER_INVARIANT): Define as 1 if undef.
> > 	* rtlanal.c (rtx_unstable_p): Qualify arg_pointer_rtx match with
> > 	ARG_POINTER_INVARIANT.
> > 	(rtx_varies_p): Likewise.
> > 	(rtx_addr_can_trap_p): Likewise.
> > 	* local-alloc.c (function_invariant_p): Likewise.
> > 	* loop.c (loop_invariant_p): Likewise.
> > 	* tm.texi (ARG_POINTER_INVARIANT): Describe.
> 
> I don't think this is the correct fix.  Isn't r29 used for 128 bit
> return values?

Yes, r29 is used for 128 bit return values, but how is that relevant to
use of r29 as the arg pointer?

I tried setting up ELIMINATE_REGS and so forth (see puffin CVS
gcc/config/pa/pa-linux64.h) but that failed badly with what looked like,
from a very cursory inspection, a use of r29 as a temporary reg being
"eliminated" to r30+const.  Oops.

Anyway, nothing in the above patch should make the situation worse.  All
it really does is let the compiler see places where arg_pointer changes,
instead of blindly assuming that arg_pointer will not change.  Take a peek
at the .lreg rtl for the test case I posted.

Alan Modra
-- 
Linuxcare.  Support for the Revolution.