Re: Checking out PR4483 (constant overflow on PPC, problems compiling Linux kernel)

[Richard Henderson <rth at redhat dot com>]

On Thu, Nov 29, 2001 at 01:08:02PM -0600, Corey Minyard wrote:
> RCS file: /cvs/gcc/gcc/gcc/recog.c,v
> retrieving revision 1.133
> diff -u -p -r1.133 recog.c
> --- recog.c	2001/11/02 10:52:08	1.133
> +++ recog.c	2001/11/29 19:03:05
> @@ -512,7 +512,8 @@ validate_replace_rtx_1 (loc, from, to, o
>           separated from this function.  */
>        if (GET_CODE (XEXP (x, 1)) == CONST_INT)
>  	validate_change (object, loc,
> -			 plus_constant (XEXP (x, 0), INTVAL (XEXP (x, 1))), 1);
> +			 simplify_gen_binary
> +			 (PLUS, GET_MODE (x), XEXP (x, 0), XEXP (x, 1)), 1);

Ok.

> diff -u -p -r1.86 simplify-rtx.c
> --- simplify-rtx.c	2001/11/14 13:51:10	1.86
> +++ simplify-rtx.c	2001/11/29 19:03:12
> @@ -1276,7 +1276,8 @@ simplify_binary_operation (code, mode, o
>  
>  	  /* Don't let a relocatable value get a negative coeff.  */
>  	  if (GET_CODE (op1) == CONST_INT && GET_MODE (op0) != VOIDmode)
> -	    return plus_constant (op0, - INTVAL (op1));
> +	    return plus_constant (op0,
> +				  trunc_int_for_mode (- INTVAL (op1), mode));

On cursory glance, I see at least 3 other examples of this bug
just in simplify-rtx.c.  I suspect this is all over the place.  :-(

There was a suggestion at some point to add a plus_constant_mode,
but I guess nothing ever happened there.  I don't know if it's
better or worse with an implicit truncation.  In some other case
we could just as well use

	simplify_gen_binary (MINUS, mode, op0, op1);

but here, that's exactly what we're trying to simplify.  ;-)

I think I'd rather you used

	simplify_gen_binary (PLUS, mode, op0, neg_const_int (mode, op1))

with

rtx
neg_const_int (mode, i)
     enum machine_mode mode;
     rtx i;
{
  return GEN_INT (trunc_int_for_mode (- INTVAL (i), mode));
}

defined somewhere in simplify-rtx.c and prototyped with the rest
of the functions from that file.  That sequence is common enough
I think a function-let will turn out go be worthwhile in the end.
(Fixing these in simplify_plus_minus at the same time encouraged.)

My reasoning here is that if you've got

	code = PLUS
	mode = SImode
	op0 = (plus:SI (reg:SI 10101) (const_int 1))
	op1 = (const_int 0x7fffffff)

then plus_constant isn't going to do the right thing.  I don't
know for sure if simplify_gen_binary will at the moment either,
but at least it has the proper information (i.e. the mode) and
can be fixed relatively easily.


r~