how to force reloader(?) to respect clobber directive?

how to force reloader(?) to respect clobber directive?

[Dmitry]

Fellows,

I define:


(define_insn "tstsi"
  [(set (cc0) (match_operand:SI 0 "register_operand" "r"))
        (clobber (match_dup 0))]
  ""
  "* return msp430_emit_tstsi(insn, operands, NULL);"
[(set_attr "length" "2")
 (set_attr "cc" "compare")])

Compiler produces a code in which 0 operand clobbered and used after this insn as being saved.

Should I define something extra to say to gcc to threat 0 operand as clobbered after insn?

Thanks,
~d


*********************************************************************
   ("`-''-/").___..--''"`-._     (\       Dimmy the Wild      UA1ACZ
    `6_ 6  )   `-.  (     ).`-.__.`)      Enterprise Information Sys 
    (_Y_.)'  ._   )  `._ `. ``-..-'       Nevsky prospekt,   20 / 44
  _..`--'_..-_/  /--'_.' ,'               Saint Petersburg,   Russia
 (il),-''  (li),'  ((!.-'                 +7 (812) 314-8860, 5585314
*********************************************************************

Re: how to force reloader(?) to respect clobber directive?

[Richard Henderson]

On Tue, Feb 26, 2002 at 12:13:43PM +0300, Dmitry wrote:
> (define_insn "tstsi"
>   [(set (cc0) (match_operand:SI 0 "register_operand" "r"))
>         (clobber (match_dup 0))]

The constraint must be "+r" in this case.


r~

Re: how to force reloader(?) to respect clobber directive?

[Dimmy]

Thanks Richard,
"+r" did not help much.

Actually, I transform a condition code of the insn followed by.
So, for the code given:

long foo(long a, long b)
{
        if(a>0) goto ee;
        return a;
ee:
        return b;
}

gcc wants to use LE condition code. Whithin tstsi operation I'm setting cond code for the next insn to LT.

Then I get: ('a' being passed in r14,r15, 'b' - in r12,r13; returns in r14,r15)
----------------
aa:
/* prologue: frame size = 0 */
.L__FrameSize_aa=0x0
/* prologue end (size=0) */
/* UID:11 ** ADDR:0 ** COST:16 *
(insn:QI 11 34 12 (parallel[ 
            (set (cc0)
                (reg/v:SI 14 r14 [22]))
            (clobber (reg/v:SI 14 r14 [22]))
        ] ) 19 {tstsi} (insn_list 4 (nil))
    (expr_list:REG_DEAD (reg/v:SI 14 r14 [22])
        (nil)))
*****************/
        sub     #1, r14  ;  11  tstsi   [length = 2]
        subc    #0, r15
/* UID:12 ** ADDR:2 ** COST:26 *
(jump_insn 12 11 22 (set (pc)
        (if_then_else (lt (cc0)
                (const_int 0 [0x0]))
            (label_ref 32)
            (pc))) 132 {blt} (nil)
    (expr_list:REG_BR_PROB (const_int 3999 [0xf9f])
        (nil)))
*****************/
        jl      .L1     ;       .L1      ;  12  blt     [length = 2]
.L3:
/* UID:26 ** ADDR:4 ** COST:4 *
(insn 26 35 32 (set (reg/i:SI 14 r14)
        (reg/v:SI 12 r12 [23])) 48 {*movsi3} (nil)
    (expr_list:REG_DEAD (reg/v:SI 12 r12 [23])
        (nil)))
*****************/
        mov     r12, r14         ;  26  *movsi3 [length = 2]
        mov     r13, r15
.L1:
/* epilogue: frame size=0 */
        ret
/* epilogue end (size=1) */
/* function aa size 7 (6) */
.Lfe1:
        .size   aa,.Lfe1-aa

----------------

So, as follows from the RTL dump reg:SI 14 dead, but gcc uses it as return operand.

Did I miss something here?



On Tue, 26 Feb 2002 09:44:56 -0800
Richard Henderson <rth@redhat.com> wrote:

> On Tue, Feb 26, 2002 at 12:13:43PM +0300, Dmitry wrote:
> > (define_insn "tstsi"
> >   [(set (cc0) (match_operand:SI 0 "register_operand" "r"))
> >         (clobber (match_dup 0))]
> 
> The constraint must be "+r" in this case.
> 
> 
> r~
> 

Re: how to force reloader(?) to respect clobber directive?

[Jan Hubicka]

> > On Tue, Feb 26, 2002 at 12:13:43PM +0300, Dmitry wrote:
> > > (define_insn "tstsi"
> > >   [(set (cc0) (match_operand:SI 0 "register_operand" "r"))
> > >         (clobber (match_dup 0))]
> > 
> > The constraint must be "+r" in this case.
My guess is that when expanding the insn you overwrite the input operand,
you should not.
In such situations you should probably use match_scratch construct to
get the output operand.

Honza
> > 
> > 
> > r~
> > 

Re: how to force reloader(?) to respect clobber directive?

[Dimmy]

Jan,
yes, match_scratch does work...
But this is not that simple:

I have to decide do I or do I not need a scratch register upon next insn condition code.
So, if I can pass 'insn' to the preparation statement in expander (i need something like:
rtx nextinsn = NEXT_INSN(insn);) - everything fine!
But how?

Thanks,
Dmitry.

On Tue, 26 Feb 2002 20:34:16 +0100
Jan Hubicka <jh@suse.cz> wrote:

> > > On Tue, Feb 26, 2002 at 12:13:43PM +0300, Dmitry wrote:
> > > > (define_insn "tstsi"
> > > >   [(set (cc0) (match_operand:SI 0 "register_operand" "r"))
> > > >         (clobber (match_dup 0))]
> > > 
> > > The constraint must be "+r" in this case.
> My guess is that when expanding the insn you overwrite the input operand,
> you should not.
> In such situations you should probably use match_scratch construct to
> get the output operand.
> 
> Honza
> > > 
> > > 
> > > r~
> > > 
>