In r229405 I removed a call to redirect_edge_var_map_destroy from
delete_tree_ssa which I thought cannot be necessary because that
map isn't GCed and thus stale data in it should have caused quite
some havoc otherwise. Turns out I was wrong ;) We were lucky
instead. The following patch amends the code in remove_edge
which clears data for removed edges (when in gimple!) in that it
now destroys the map at RTL expansion time.
Now there is still a latent issue I believe as somebody is making
use of the map for an edge it didn't previously push to
(pushing always first clears the entry for the edge).
The comment in redirect_edge_var_map_vector is also revealing in
that context:
redirect_edge_var_map_vector (edge e)
{
/* Hey, what kind of idiot would... you'd be surprised. */
if (!edge_var_maps)
return NULL;
In particular we probably have (very many?) redirect_edge_and_branch
calls not paired with a redirect_edge_var_map_clear call.
In all it looks like a very fragile thing, that SSA edge redirect
hook stuff.
Bootstrap / regtest running on x86_64-unknown-linux-gnu. Will
apply if that succeeds and hope for the best.
Richard.
2015-11-16 Richard Biener <rguent...@suse.de>
PR middle-end/68117
* cfgexpand.c (pass_expand::execute): Destroy the edge
redirection var map before setting RTL CFG hooks.
Index: gcc/cfgexpand.c
===================================================================
--- gcc/cfgexpand.c (revision 230404)
+++ gcc/cfgexpand.c (working copy)
@@ -6275,6 +6278,9 @@ pass_expand::execute (function *fun)
expand_phi_nodes (&SA);
+ /* Release any stale SSA redirection data. */
+ redirect_edge_var_map_destroy ();
+
/* Register rtl specific functions for cfg. */
rtl_register_cfg_hooks ();
