Re: reload bugfix

[Alexandre Oliva <aoliva at redhat dot com>]

On Aug 19, 2000, Bernd Schmidt <bernds@cygnus.co.uk> wrote:

> [or we could install my monster patch which gets rid of all the
> RELOAD_FOR_foo cruft :-)]

I can't say I'd mind that.  I've just spent a couple of hours fighting
another reload problem for SH.

Basically, I'm trying to fix a SH problem of loading SFmode constants
to pseudo registers that end up on the stack.  GCC fails because it
needs r0 as a scratch register with earlyclobber, but reload assigns
the frame-pointer offset to r0 (which is the only index register too),
and it loses.  An example:

(insn 23 21 24 (parallel[ 
            (set (subreg:SF (reg/v:DI 87) 0)
                (const_double:SF (cc0) 0 [0x0] 0 [0x0] [0]))
            (use (reg/v:PSI 48 fpscr))
            (clobber (scratch:SI))
        ] ) 91 {movsf_ie} (nil)
    (expr_list:REG_UNUSED (scratch:SI)
        (nil)))

reload replaces the subreg with
(mem:SF (plus:SI (reg:SI 14) (const_int 48))), then replaces
(const_int 48) with (reg:SI r0).

But then, when it gets to the scratch register, it needs r0 too,
because the constraint is `&z', and reload aborts.

Ok, so I introduced a new variant, that accepts a mem as the SET_DEST,
a floating-point constant as SET_SRC and any register `&r' as scratch,
then defined a split that copies SET_SRC to the scratch register, then
stores the scratch register in SET_DEST.

But GCC still insisted in allocating both the FP offset and the
scratch register to r0.  I seem to have just found a solution.  First,
I had to disable the optimization in find_reloads() that used
RELOAD_FOR_INSN for scratch earlyclobber registers that are noted as
unused.  Then, I had to add test reload_reg_used_in_op_addr_reload in
reload_reg_free_p(), for a RELOAD_OTHER reload type.

Does this make sense for you?

-- 
Alexandre Oliva   Enjoy Guarana', see http://www.ic.unicamp.br/~oliva/
Red Hat GCC Developer                  aoliva@{cygnus.com, redhat.com}
CS PhD student at IC-Unicamp        oliva@{lsd.ic.unicamp.br, gnu.org}
Free Software Evangelist    *Please* write to mailing lists, not to me