Re: Using SCRATCH Register for a mov op,

Re: Using SCRATCH Register for a mov op,

[Pierre Mallard]

Hi and thanks for paying attention to my problem,
I still got a problem around those scratch regs :
Here is how I define a mohi (I 'm gonna define
define_split later):
(define_expand "movhi"
  [(parallel[(set (match_operand:HI 0
"general_operand" "=r,r,m,r,r")
        (match_operand:HI 1 "general_operand"
"r,m,r,I,i"))
   (clobber (match_scratch:QI 2	      ""))])]
  ""
  "{
  rtx temp;
  /* One of the ops has to be in a register */
  if (!register_operand(operand0, HImode)
      && !register_operand(operand1, HImode))
    {
      operands[1] = copy_to_mode_reg(HImode,
operand1);
    }
}")

(define_insn "*movhi"
  [(set	(match_operand:HI 0 "nonimmediate_operand" 
"=r,r,m,r,r")
       	(match_operand:HI 1 "general_operand"       
"r,m,r,I,i"))
   (clobber (match_scratch:QI 2				"=X,&r,X,X,X"))]
  "(register_operand (operands[0],HImode)||
register_operand (operands[1],HImode))"
  "* return output_movhi (insn, operands, NULL);"
  [(set_attr "length" "1,2,2,1,2")
   (set_attr "cc" "none,none,none,none,none")
   (set_attr "type" "none,load,none,none,none")])

But It looks like the reload pass still generate movhi
without using scratch regs : 
../../gcc-3.0.4/gcc/libgcc2.c: In function `__modhi3':
../../gcc-3.0.4/gcc/libgcc2.c:773: Insn does not
satisfy its constraints:
(insn 1064 1061 1055 (parallel[ 
            (set (reg:HI 22 r22)
                (mem:HI (reg:HI 22 r22) 8))
            (clobber (scratch:QI))
        ] ) 1 {*movhi} (nil)
    (nil))
../../gcc-3.0.4/gcc/libgcc2.c:773: Internal compiler
error in reload_cse_simplify_operands, at
reload1.c:8368

--- Richard Henderson <rth@redhat.com> a écrit : > On
Wed, Oct 30, 2002 at 05:45:57PM +0100, Pierre
> Mallard wrote:
> > I'd like to use scratch reg for operation such as 
> > reg X = mem(reg X) generated during the RELOAD
> PASS.
> > with X a number of a hard reg same for both side
> of
> > expression
> 
> You'll have to use a PARALLEL with a (clobber
> (match_scratch)).
> 
> 
> r~ 

___________________________________________________________
Do You Yahoo!? -- Une adresse @yahoo.fr gratuite et en français !
Yahoo! Mail : http://fr.mail.yahoo.com

Re: Using SCRATCH Register for a mov op,

[Richard Henderson]

On Mon, Nov 04, 2002 at 11:03:12AM +0100, Pierre Mallard wrote:
> (define_insn "*movhi"
>   [(set	(match_operand:HI 0 "nonimmediate_operand" 
> "=r,r,m,r,r")
>        	(match_operand:HI 1 "general_operand"       
> "r,m,r,I,i"))
>    (clobber (match_scratch:QI 2				"=X,&r,X,X,X"))]

Looks ok; no idea what's going wrong off the top of my head.


r~

Re: Using SCRATCH Register for a mov op,

[Pierre Mallard]

I'm wondering if it has sthg to do with the Macro
defined in .h ...

#define PREFERRED_RELOAD_CLASS(X,CLASS)	CLASS

#define CLASS_LIKELY_SPILLED_P(c) GENERAL_REGS

#define LEGITIMIZE_ADDRESS(X, OLDX, MODE, WIN)				\
{									\
  (X) = legitimize_address (X, OLDX, MODE);				\
  if (memory_address_p (MODE, X))					\
    goto WIN;								\
}

#define GO_IF_MODE_DEPENDENT_ADDRESS(ADDR,LABEL) \
{}

Is there other macro I got to define to help reload
pass generate a SCRATCH register ?
 --- Richard Henderson <rth@redhat.com> a écrit : > On
Mon, Nov 04, 2002 at 11:03:12AM +0100, Pierre
> Mallard wrote:
> > (define_insn "*movhi"
> >   [(set	(match_operand:HI 0 "nonimmediate_operand"
> 
> > "=r,r,m,r,r")
> >        	(match_operand:HI 1 "general_operand"     
>  
> > "r,m,r,I,i"))
> >    (clobber (match_scratch:QI 2			
> "=X,&r,X,X,X"))]
> 
> Looks ok; no idea what's going wrong off the top of
> my head.
> 
> 
> r~ 

___________________________________________________________
Do You Yahoo!? -- Une adresse @yahoo.fr gratuite et en français !
Yahoo! Mail : http://fr.mail.yahoo.com

Re: Using SCRATCH Register for a mov op,

[David Edelsohn]

(define_expand "movhi"
  [(parallel[(set (match_operand:HI 0 "general_operand" "=r,r,m,r,r")
        (match_operand:HI 1 "general_operand" "r,m,r,I,i"))
   (clobber (match_scratch:QI 2 ""))])]
  ""
  "{
  rtx temp;
  /* One of the ops has to be in a register */
  if (!register_operand(operand0, HImode)
      && !register_operand(operand1, HImode))
    {
      operands[1] = copy_to_mode_reg(HImode, operand1);
    }
}")

	Shouldn't the define_expand use

(clobber (scratch:QI))

not

(clobber (match_scratch:QI 2 ""))

	The define_insn needs to match the scratch, but the define_expand
needs to create the scratch register.

David

Re: Using SCRATCH Register for a mov op,

[Peter Barada]

You could look at SECONDARY_RELOAD_CLASS and its cousins, the
reload_inXX and reload_outXX patterns(section 17.6 of the manual).
I've used them in gcc-3.0.4 (and working on gcc-3.2) with great effect
to generate a scratch address register to hold the address of an
abolute symbol while trying to load/store a float/double (for the
ColdFire v4e architecture).

I lifted my code form the PA architecture(foind in
gcc/config/gcc/pa/pa.{h,c,md}) which has some severe restructicions on
its instructiions and addressing modes so it ends up(for some
cases)loading the resultant operand address into a register and then
dloing the load/store indirect from that(see emite_move_sequence for
the code).  I'll have to take the same approach for some of the
ColdFire load/store pairs since certain addressing modes are
not allowed in conjunction with other addressing modes.  For example:

	   move d8(Ax,Xi),Dy
	   move Dy,d8(Ax,Xi)

Are both allowed, but:

	   move d8(Ax,Xi),d8(Ay,Xj)

is not allowed.  And since GO_IF_LEGITIMATE_ADDRESS/LEGITIMIZE_ADDRESS
have no knowledge of the instruction its trying to validate/fix, nor
which side of the instruction(src/dst) its on, so in this case I will
have to force *everything* through the reload_in/reload_out patterns
and not use the scratch register that is proffered by the pattern if
it is not needed...

I hope this might shed some light on your problem...

-- 
Peter Barada                                   Peter.Barada@motorola.com
Wizard                                         781-852-2768 (direct)
WaveMark Solutions(wholly owned by Motorola)   781-270-0193 (fax)

Re: Using SCRATCH Register for a mov op,

[Peter Barada]

>dloing the load/store indirect from that(see emite_move_sequence for

That should be:

>doing the load/store indirect from that(see emit_move_sequence for

-- 
Peter Barada                                   Peter.Barada@motorola.com
Wizard                                         781-852-2768 (direct)
WaveMark Solutions(wholly owned by Motorola)   781-270-0193 (fax)

Using SCRATCH Register for a mov op,

[Pierre Mallard]

Hi,
I'd like to use scratch reg for operation such as 
reg X = mem(reg X) generated during the RELOAD PASS.
with X a number of a hard reg same for both side of
expression

Unfortunately it appears that a good means to obtain
the SCRATCH register from gcc is to use reload_in.
But this doesn't work in my proc cause I have no
different classes of regs (for me they are all
equivalent), I can't be able to define
SECONDARY_RELOAD_CLASS or things like this cause it
will abort if using same CLASS of regs ().
So I have no idea how to have a SCRATCH reg generated
with the op seen above.

Please Help Me
Regards
Pierre Mallard

___________________________________________________________
Do You Yahoo!? -- Une adresse @yahoo.fr gratuite et en français !
Yahoo! Mail : http://fr.mail.yahoo.com