Hi, On Wed, Sep 12, 2012 at 04:47:11PM +0200, Richard Guenther wrote: > On Wed, Sep 12, 2012 at 4:37 PM, Martin Jambor <mjam...@suse.cz> wrote: > > On Wed, Sep 12, 2012 at 04:17:45PM +0200, Michael Matz wrote: > >> On Wed, 12 Sep 2012, Michael Matz wrote: > >> > >> > > Hm, but we shouldn't end up streaming any BLOCKs at this point (nor > >> > > local TYPE_DECLs). Those are supposed to be in the local function > >> > > sections only where no fixup for prevailing decls happens. > >> > > >> > That's true, something is fishy with the patch, will try to investigate. > >> > >> ipa-prop creates the problem. Its tree mapping can contain expressions, > >> expressions can have locations, locations now have blocks. The tree maps > >> are stored as part of jump functions, and hence as part of node summaries. > >> Node summaries are global, hence blocks, and therefore block vars can be > >> placed in the global blob. > >> > >> That's not supposed to happen. The patch below fixes this instance of the > >> problem and makes the testcase work with Dehaos patch with the > >> LTO_NO_PREVAIL call added back in. > >> > >> > >> Ciao, > >> Michael. > >> ------------ > >> Index: lto-cgraph.c > >> =================================================================== > >> --- lto-cgraph.c (revision 190803) > >> +++ lto-cgraph.c (working copy) > >> @@ -1373,6 +1373,7 @@ output_node_opt_summary (struct output_b > >> mechanism to store function local declarations into summaries. > >> */ > >> gcc_assert (parm); > >> streamer_write_uhwi (ob, parm_num); > >> + gcc_assert (IS_UNKNOWN_LOCATION (EXPR_LOCATION (map->new_tree))); > >> stream_write_tree (ob, map->new_tree, true); > >> bp = bitpack_create (ob->main_stream); > >> bp_pack_value (&bp, map->replace_p, 1); > >> Index: ipa-prop.c > >> =================================================================== > >> --- ipa-prop.c (revision 190803) > >> +++ ipa-prop.c (working copy) > >> @@ -1378,7 +1378,11 @@ ipa_compute_jump_functions_for_edge (str > >> tree arg = gimple_call_arg (call, n); > >> > >> if (is_gimple_ip_invariant (arg)) > >> - ipa_set_jf_constant (jfunc, arg); > >> + { > >> + arg = unshare_expr (arg); > >> + SET_EXPR_LOCATION (arg, UNKNOWN_LOCATION); > >> + ipa_set_jf_constant (jfunc, arg); > >> + } > >> else if (!is_gimple_reg_type (TREE_TYPE (arg)) > >> && TREE_CODE (arg) == PARM_DECL) > > > > Perhaps it would be better if ipa_set_jf_constant did that, just in > > case we ever add another caller? Note that arithmetic functions also > > have their second operand tree stored in them and so perhaps > > ipa_set_jf_arith_pass_through should do the same. > > > > And I it is also necessary to do the same thing at the end of > > determine_known_aggregate_parts, i.e. before assignment to > > item->value. I can post a separate patch if necessary. > > > > I wasn't following this thread but I hope that streaming types does > > not cause this problem. If they do, there are quite a few in various > > jump functions and indirect call graph edges. > > Well, or if jump functions would not use trees ... >
Eventually, yes, but it is not a short-term goal. At least I'd like to replace the type-based devirtualization with pointer tracking first, which will avoid the need to store (hopefully) all types and I will have a better idea how to represent the jump functions in a different way. I will still need access to VMT initializers to keep indirect inlining, indirect call inlining hints and IPA-CP devirtualization working, though. But as I said, this is all a long-term plan. Martin