[tree-ssa] fold-const.c merge from mainline

[Richard Henderson]

On Mon, Oct 13, 2003 at 07:44:18PM -0600, law@redhat.com wrote:
>  >(1) Whilst it is true that CCP only cares about cases when we can fold
>  >down to a constant, it is frequently the case that fold reduces an
>  >tree expression to a constant but in a number of steps.  For example,
>  >"(x + 3) - x" can be reduced by "fold" via the recursive call to
>  >"(x - x) + 3".  Unfortunately, nondestructive_fold_binary_to_constant
>  >on the tree-ssa branch isn't able to do this.
> But you don't get those kinds of expressions during CCP.  They simply
> don't happen.

Hmm.  How about

	int x[10];
	int *p, *q, t;
	p = &x[1];
	q = &x[5];
	t = q - p;

gimplifies to

	p = &x + 4B;
	q = &x + 20B;
	t = q - p;

ccp's to

	t = (&x + 20b) - (&x + 4b)

which is indeed gimple, because &v+c matches is_gimple_val.

This is about the most extreme case we can create, however.


r~