| Message ID | 53283E3B.4050700@suse.cz |
|---|---|
| State | New |
| Headers | show |
On Tue, Mar 18, 2014 at 1:38 PM, Martin Liška <mliska@suse.cz> wrote: > Hello, > I found ICE in Chromium compiled with LTO. There's a call that is proved > by ipa-devirt as __builtin_unreachable; same decision is done by gimple-fold > and this call is replaced by GIMPLE_CALL and GIMPLE_ASSIGN (in this order). > After that condition for cgraph_update_edges_for_call_stmt_node is not > satisfied and corresponding cgraph_edge is not updated. Thus a verifier > reports a wrong edge. You should be able to simply do update_call_from_tree (gsi, def); gsi_insert_before (gsi, new_stmt, GSI_NEW_STMT); also cgraph_edge (node, old_stmt) is already computed in 'e' AFAICS. Richard. > Bootstrapped and tested on a x86_64 machine. > > > Changelog: > 2014-03-18 Martin Liska <mliska@suse.cz> > > * cgraph.c (cgraph_update_edges_for_call_stmt_node): added case when > gimple call statement is updated. > * gimple-fold.c (gimple_fold_call): changed order for GIMPLE_ASSIGN > and > GIMPLE_CALL, where gsi iterator still points to GIMPLE CALL. > > OK for trunk? > > Thank you, > Martin > >
Hi! > > 2014-03-18 Martin Liska <mliska@suse.cz> > > > > * cgraph.c (cgraph_update_edges_for_call_stmt_node): added case when > > gimple call statement is updated. Capital letter after : > > * gimple-fold.c (gimple_fold_call): changed order for GIMPLE_ASSIGN and Likewise here. Jakub
diff --git a/gcc/cgraph.c b/gcc/cgraph.c index a15b6bc..cd68894 100644 --- a/gcc/cgraph.c +++ b/gcc/cgraph.c @@ -1519,7 +1519,11 @@ cgraph_update_edges_for_call_stmt_node (struct cgraph_node *node, { if (callee->decl == new_call || callee->former_clone_of == new_call) - return; + { + cgraph_set_call_stmt (cgraph_edge (node, old_stmt), + new_stmt); + return; + } callee = callee->clone_of; } } diff --git a/gcc/gimple-fold.c b/gcc/gimple-fold.c index eafdb2d..a033fbc 100644 --- a/gcc/gimple-fold.c +++ b/gcc/gimple-fold.c @@ -1153,8 +1153,14 @@ gimple_fold_call (gimple_stmt_iterator *gsi, bool inplace) { tree var = create_tmp_var (TREE_TYPE (lhs), NULL); tree def = get_or_create_ssa_default_def (cfun, var); - gsi_insert_before (gsi, new_stmt, GSI_SAME_STMT); - update_call_from_tree (gsi, def); + + /* To satisfy condition for + cgraph_update_edges_for_call_stmt_node, + we need to preserve GIMPLE_CALL statement + at position of GSI iterator. */ + gimple_stmt_iterator oldgsi = *gsi; + gsi_insert_before (gsi, new_stmt, GSI_NEW_STMT); + update_call_from_tree (&oldgsi, def); } else gsi_replace (gsi, new_stmt, true);
Hello, I found ICE in Chromium compiled with LTO. There's a call that is proved by ipa-devirt as __builtin_unreachable; same decision is done by gimple-fold and this call is replaced by GIMPLE_CALL and GIMPLE_ASSIGN (in this order). After that condition for cgraph_update_edges_for_call_stmt_node is not satisfied and corresponding cgraph_edge is not updated. Thus a verifier reports a wrong edge. Bootstrapped and tested on a x86_64 machine. Changelog: 2014-03-18 Martin Liska <mliska@suse.cz> * cgraph.c (cgraph_update_edges_for_call_stmt_node): added case when gimple call statement is updated. * gimple-fold.c (gimple_fold_call): changed order for GIMPLE_ASSIGN and GIMPLE_CALL, where gsi iterator still points to GIMPLE CALL. OK for trunk? Thank you, Martin