------- Additional Comments From law at redhat dot com 2004-11-01 20:04 -------
Subject: Re: [4.0 Regression] jump threading
on trees is slow with switch statements with large # of cases
On Mon, 2004-11-01 at 19:35 +0000, kazu at cs dot umass dot edu wrote:
> ------- Additional Comments From kazu at cs dot umass dot edu 2004-11-01 19:34
> -------
> Subject: Re: [4.0 Regression] jump threading
> on trees is slow with switch statements with large # of cases
>
> Hi Steven,
>
> > OK, then can you see if this hack helps...?
>
> + /* Step 4: Update the case label vector. */
> + TREE_VEC_LENGTH (label_vec) = n;
> + for (i = 0; i < n; ++i)
> + {
> + tree elt = TREE_VEC_ELT (label_vec, i);
> + e = case_to_edge[i];
> + CASE_LABEL (elt) = tree_block_label (e->dest);
> + }
>
> Wow, this is barely safe. You are assuming that the edge redirection
> is done by changing e->dest and that cfg.c:redirect_edge_succ_nodup
> does not remove e, which is actually the case because there is no
> existing edge from the block containing SWITCH_EXPR to the new basic
> block created by split_edge.
>
> In any case, I agree this should work. I'll be the second one to
> admit this is gross. :-)
I think it'll ultimately be cleaner to simply drop the labels after
we've built the CFG and associate an edge with each of the cases.
We'd still have the assumption that edge redirection happens by
changing e->dest, but I think that approach is otherwise reasonably
clean.
Probably the ugliest part of that approach is the need to have edges
in the switch statement.
The nice thing is this approach matches pretty well with what we've
done with simple gotos.
Jeff
--
http://gcc.gnu.org/bugzilla/show_bug.cgi?id=15524
