https://gcc.gnu.org/bugzilla/show_bug.cgi?id=83388
Richard Biener <rguenth at gcc dot gnu.org> changed:
What |Removed |Added
----------------------------------------------------------------------------
Status|UNCONFIRMED |NEW
Last reconfirmed| |2017-12-12
CC| |hubicka at gcc dot gnu.org
Ever confirmed|0 |1
--- Comment #6 from Richard Biener <rguenth at gcc dot gnu.org> ---
Index: gcc/lto-streamer-in.c
===================================================================
--- gcc/lto-streamer-in.c (revision 255578)
+++ gcc/lto-streamer-in.c (working copy)
@@ -1175,6 +1175,16 @@ input_function (tree fn_decl, struct dat
}
if (remove)
{
+ /* If we are asked to remove a non-debug stmt replace it
+ with a nop instead to not confuse IPA references that
+ might refer to the original. */
+ if (! is_gimple_debug (stmt))
+ {
+ gimple *nop = gimple_build_nop ();
+ gimple_set_uid (nop, gimple_uid (stmt));
+ gsi_insert_before (&bsi, nop, GSI_SAME_STMT);
+ stmts[gimple_uid (nop)] = nop;
+ }
gimple_stmt_iterator gsi = bsi;
gsi_next (&bsi);
unlink_stmt_vdef (stmt);
works to fix the ICE. We'll end up with a "dangling" IPA reference but
hopefully that's safe...
The other option would be to remove those IPA references on-the-fly.
Honza, would it be safe to simply do
Index: gcc/lto-streamer-in.c
===================================================================
--- gcc/lto-streamer-in.c (revision 255578)
+++ gcc/lto-streamer-in.c (working copy)
@@ -921,8 +921,6 @@ fixup_call_stmt_edges_1 (struct cgraph_n
fatal_error (input_location,
"Reference statement index out of range");
ref->stmt = stmts[ref->lto_stmt_uid - 1];
- if (!ref->stmt)
- fatal_error (input_location, "Reference statement index not found");
}
}
instead? Thus leave ref->stmt == NULL? Or would it be safe to remove
the IPA reference at this stage (I suppose IPA transform phase might get
confused if they stream references to such references?)