http://gcc.gnu.org/bugzilla/show_bug.cgi?id=46760
--- Comment #23 from Jan Hubicka <hubicka at ucw dot cz> 2011-01-10 00:21:35
UTC ---
> 5386
> 5387 /* Allow the backend to decide if inlining is ok. */
> 5388 if (!targetm.target_option.can_inline_p (caller, callee))
> 5389 {
> 5390 e->inline_failed = CIF_TARGET_OPTION_MISMATCH;
> 5391 gimple_call_set_cannot_inline (e->call_stmt, true);
> 5392 e->call_stmt_cannot_inline_p = true;
> 5393 return false;
> 5394 }
> 5395
> (gdb) p e
> $1 = (struct cgraph_edge *) 0x7fffdaf9fd68
> (gdb) p e->call_stmt
> $2 = (union gimple_statement_d *) 0x0
> (gdb)
Obviously we can not set gimple_call_set_cannot_inline when call_stmt is
missing.
I guess just guardining the set with
if (e->call_stmt)
gimple_call_set_cannot_inline (e->call_stmt, true);
in all occurences in the function should fix the issue. Consider it preaproved
if it helps. In 4.7 we probably should drop the gimple_call_set_cannot_inline
calls completely, I don't see any value in them.
Can you do dump_cgraph_node (stderr, e->caller) and dump_cgraph_node (stderr,
e->callee)?? I think we should decide function to be uninlinable only in
presence of target optimization attributes and I can not think of any during
the bootstrap....
Honza
