[PATCH] Fix PR tree-opt/32527: ICE in build2_stat (aff_combination_scale)

[Ian Lance Taylor]

Andrew_Pinski@PlayStation.Sony.Com writes:

>         * tree-affine.c (aff_combination_scale): Use sizetype for
>         mutiplication when the affine is a pointer type.
> 
> 
>         * gfortran.fortran-torture/compile/pr32527.f90: New test.
> 
> 
> Index: testsuite/gfortran.fortran-torture/compile/pr32527.f90
> ===================================================================
> --- testsuite/gfortran.fortran-torture/compile/pr32527.f90	(revision 0)
> +++ testsuite/gfortran.fortran-torture/compile/pr32527.f90	(revision 0)
> @@ -0,0 +1,18 @@
> +! This code used to crash in aff_combination_scale while trying to generate a
> +! multiply in a pointer type
> +
> +function nf90_put_var_7D_FourByteInt(ncid, varid, values, start, count, stride, map)
> +  integer,               intent(in)     :: ncid, varid
> +  integer, dimension(:, :, :, :, :, :, :), intent(in) :: values
> +  integer, dimension(:), intent(in)     :: start, count, stride, map
> +  integer                               :: nf90_put_var_7D_FourByteInt
> +
> +  integer, dimension(1024) :: localStart, localCount, localStride, localMap
> + 
> +  integer, external :: nf_put_varm_int
> +
> +  nf90_put_var_7D_FourByteInt = &
> +        nf_put_varm_int (ncid, varid, localStart, localCount, localStride, localMap, &
> +                         int (values))
> +end function nf90_put_var_7D_FourByteInt
> +
> Index: tree-affine.c
> ===================================================================
> --- tree-affine.c	(revision 126131)
> +++ tree-affine.c	(working copy)
> @@ -119,8 +119,15 @@ aff_combination_scale (aff_tree *comb, d
>  	  comb->n++;
>  	}
>        else
> -	comb->rest = fold_build2 (MULT_EXPR, comb->type, comb->rest, 
> -				  double_int_to_tree (comb->type, scale));
> +	{
> +	  tree type = comb->type;
> +	  /* The type of rest is sizetype for pointer types.  */
> +	  if (POINTER_TYPE_P (type))
> +	    type = sizetype;
> +
> +	  comb->rest = fold_build2 (MULT_EXPR, type, comb->rest, 
> +				    double_int_to_tree (type, scale));
> +	}
>      }
>  }
>  


This looks incomplete to me.  The comment in tree-affine.h says that
the type of comb->rest must be comb->type, but this change breaks that
invariant.  Is the comment wrong?

Also, the patch appears to assume that if comb->type is a pointer,
then the type of comb->rest will be sizetype.  But I don't see
anything which enforces that.

Ian