------- Additional Comments From law at redhat dot com 2005-03-30 18:57 -------
Subject: Re: [PR tree-optimization/20460] add phi args to dests of
dce-redirected edges
On Wed, 2005-03-30 at 02:56 -0300, Alexandre Oliva wrote:
> When remove_dead_stmt() redirects a control stmt, the edge redirection
> reserves space for the phi arg for the new incoming edge in all phi
> nodes, but, instead of filling them in with information obtained from
> the edge redirection, we simply discard this information. This leaves
> NULL in the phi args, which may cause crashes later on.
>
> This patch fixes the problem by filling in the phi args using the
> PENDING_STMT list created during edge redirection. This appears to be
> the intended use for this information, and it is used similarly in
> e.g. loop unrolling.
>
> Bootstrapping mainline and 4.0 branch on amd64-linux-gnu, and mainline
> on i686-pc-linux-gnu. Ok to install if bootstrap and regtesting pass?
>
> The patch below is for the 4.0 branch, but it applies cleanly and
> correctly in mainline as well, since it's just a few lines off.
/* Redirect the first edge out of BB to reach POST_DOM_BB. */
redirect_edge_and_branch (EDGE_SUCC (bb, 0), post_dom_bb);
- PENDING_STMT (EDGE_SUCC (bb, 0)) = NULL;
+ flush_pending_stmts (EDGE_SUCC (bb, 0));
I'm having trouble seeing how this can be correct.
AFAICT this assumes that EDGE_SUCC (bb, 0)->dest before the redirection
has similar PHI as post_dom_bb and that the PHIs appear in the same
order in both blocks. I'm not sure you can make that assumption.
This code is triggered rarely, I would expect it to be even rarer still
for POST_DOM_BB to have PHI nodes. You could probably just ignore dead
control statements where the post dominator has PHI nodes and I doubt
it would ever be noticed.
Jeff
--
http://gcc.gnu.org/bugzilla/show_bug.cgi?id=20460
