This is the mail archive of the gcc@gcc.gnu.org mailing list for the GCC project.


Index Nav: [Date Index] [Subject Index] [Author Index] [Thread Index]
Message Nav: [Date Prev] [Date Next] [Thread Prev] [Thread Next]

Re: Your change to jump.c breaks conditional move?


> 
> On Thu, Jun 04, 1998 at 07:40:12AM -0700, H.J. Lu wrote:
> > > I'll be happy to fix any problems if you can show me an incorrect
> > > transformation in any jump pass.  But I really suspect things go wrong
> > > for you in reload.
> 
> Well, I was wrong, the problem isn't in reload.  But it isn't in
> jump.c either.  The lossage is occurring in force_operand.
> 
> Given
> 
> (plus:SI (reg:SI 3 %ebx)
>     (symbol_ref/v:SI ("*.LC0")))
> 
> force_operand is generating
> 
> (insn 169 0 170 (set (reg:SI 55)
>         (plus:SI (reg:SI 3 %ebx)
>             (symbol_ref/v:SI ("*.LC0")))) -1 (nil)
>     (nil))
> 
> (insn 170 169 172 (set (reg:SI 54)
>         (reg:SI 55)) -1 (nil)
>     (nil))
> 
> (insn 172 170 0 (set (reg:SI 53)
>         (plus:SI (reg:SI 3 %ebx)
>             (reg:SI 54))) -1 (nil)
>     (nil))
> 
> I'm not sure what the proper thing to do to curtail this lossage, 
> but the following seems to work.
> 
> The generated code is actually pretty ugly.  But I think there's
> an easy way to fix that up.  
> 
> 
> r~
> 
> 
> Index: expr.c
> ===================================================================
> RCS file: /egcs/carton/cvsfiles/egcs/gcc/expr.c,v
> retrieving revision 1.64
> diff -c -p -d -r1.64 expr.c
> *** expr.c	1998/06/02 22:48:51	1.64
> --- expr.c	1998/06/04 18:12:36
> *************** force_operand (value, target)
> *** 4526,4532 ****
>     register rtx subtarget = (target != 0 && GET_CODE (target) == REG ? target : 0);
>   
>     if (GET_CODE (value) == PLUS)
> !     binoptab = add_optab;
>     else if (GET_CODE (value) == MINUS)
>       binoptab = sub_optab;
>     else if (GET_CODE (value) == MULT)
> --- 4526,4543 ----
>     register rtx subtarget = (target != 0 && GET_CODE (target) == REG ? target : 0);
>   
>     if (GET_CODE (value) == PLUS)
> !     {
> !       /* Check for a PIC address load.  */
> !       if (flag_pic && XEXP (value, 0) == pic_offset_table_rtx)
> !         {
> ! 	  if (!subtarget)
> ! 	    subtarget = gen_reg_rtx (GET_MODE (value));
> ! 	  emit_move_insn (subtarget, value);
> ! 	  return subtarget;
> ! 	}
> ! 
> !       binoptab = add_optab;
> !     }
>     else if (GET_CODE (value) == MINUS)
>       binoptab = sub_optab;
>     else if (GET_CODE (value) == MULT)
> 

Is that remotely possible to have flag_pic and
GET_CODE (value) == MINUS?


H.J.


Index Nav: [Date Index] [Subject Index] [Author Index] [Thread Index]
Message Nav: [Date Prev] [Date Next] [Thread Prev] [Thread Next]