On Sun, Apr 17, 2011 at 7:14 AM, Jan Hubicka <hubi...@ucw.cz> wrote: > Hi, > there is now tracking PR48636 for fortran inlining issues. I would suggest > reporting problems like this there or PR that blocks it instead of sending it > to different threads so we have track of them. It is quite important to get > some understanding what sort of issues we have here. We might want to collect > some sort of inlining benchmark like we have for C++. For the moment I think > the polyhedron we have is working well, but just having a summary what > benchmarks needs what --param tweaks and why (i.e. what is important to > inline) > would be useful. > > While final testing before comming I noticed that last minue gcc_assert added > to can_inline checking that gimple stmt and edge cannot_inline_flags are in > sync > (they must be, otherwise we would do invalid optimizations at IPA) is failing > on one of C++ testcases. > > The problem is code in tree-inline that incorrectly handles the flag on > functions that do have mismatching EH personalities. > > This is final version of patch I comitted that fixes this by moving all the > logic > scattered across tree-inline.c into can_inline_edge_p where it belongs and > adding > corresponding CIF codes. It also turned out that our all_set_cannot_inline > are > not quite up to date after devirtualization and other transforms. As first > cut > i decided to move the computation into inline summary computation and I think > i can remove it from gimplifier and value profiler in followup patch. > > Bootstrapped/regtested x86_64-linux, comitted. > > Index: ChangeLog > =================================================================== > *** ChangeLog (revision 172608) > --- ChangeLog (working copy) > *************** > *** 1,3 **** > --- 1,74 ---- > + 2011-04-17 Jan Hubicka <j...@suse.cz> > + > + * lto-symtab.c (lto_cgraph_replace_node): When call statement is > + present, also set gimple_call_set_cannot_inline. > + * ipa-inline.c: Update toplevel comment. > + (MAX_TIME): Remove. > + (cgraph_clone_inlined_nodes): Fix linebreaks. > + (cgraph_check_inline_limits): Restructure to ... > + (caller_growth_limits): ... this one; be more tolerant > + on growth in nested inline chains; add explanatory comment; > + fix stack accounting thinko introduced by previous patch. > + (cgraph_default_inline_p): Remove. > + (report_inline_failed_reason): New function. > + (can_inline_edge_p): New function. > + (can_early_inline_edge_p): New function. > + (leaf_node_p): Move upwards in file. > + (want_early_inline_function_p): New function. > + (want_inline_small_function_p): New function. > + (want_inline_self_recursive_call_p): New function. > + (cgraph_edge_badness): Rename to ... > + (edge_badness) ... this one; fix linebreaks. > + (update_edge_key): Update call of edge_baddness; add > + detailed dump about queue updates. > + (update_caller_keys): Use can_inline_edge_p and > + want_inline_small_function_p. > + (cgraph_decide_recursive_inlining): Rename to... > + (recursive_inlining): Use can_inline_edge_p and > + want_inline_self_recursive_call_p; simplify and > + remove no longer valid FIXME. > + (cgraph_set_inline_failed): Remove. > + (add_new_edges_to_heap): Use can_inline_edge_p and > + want_inline_small_function_p. > + (cgraph_decide_inlining_of_small_functions): Rename to ... > + (inline_small_functions): ... this one; cleanup; use > + can/want predicates; cleanup debug ouput; work edges > + till fibheap is exhausted and do not stop once unit > + growth is reached; remove later loop processing remaining > + edges. > + (cgraph_flatten): Rename to ... > + (flatten_function): ... this one; use can_inline_edge_p > + and can_early_inline_edge_p predicates. > + (cgraph_decide_inlining): Rename to ... > + (ipa_inline): ... this one; remove unreachable nodes before > + inlining functions called once; simplify the pass. > + (cgraph_perform_always_inlining): Rename to ... > + (inline_always_inline_functions): ... this one; use > + DECL_DISREGARD_INLINE_LIMITS; use can_inline_edge_p > + predicate > + (cgraph_decide_inlining_incrementally): Rename to ... > + (early_inline_small_functions): ... this one; simplify > + using new predicates; cleanup; make dumps prettier. > + (cgraph_early_inlining): Rename to ... > + (early_inliner): newer inline regular functions into always-inlines; > + fix updating of call stmt summaries. > + (pass_early_inline): Update for new names. > + (inline_transform): Fix formating. > + (gate_cgraph_decide_inlining): Rename to ... > + (pass_ipa_inline): ... this one. > + * ipa-inline.h (inline_summary): Remove disregard_inline_limits. > + * ipa-inline-analysis.c (dump_inline_summary): Update. > + (compute_inline_parameters): Do not compute disregard_inline_limits; > + look for mismatching arguments. > + (estimate_growth): Fix handlig of non-trivial self recursion. > + (inline_read_summary): Do not read info->disregard_inline_limits. > + (inline_write_summary): Do not write info->disregard_inline_limits. > + * tree-inline.c (inline_forbidden_into_p, tree_can_inline_p): Remove > and > + move all checks into can_inline_edge_p predicate; re-enable code > comparing > + optimization levels. > + (expand_call_inline): Do not test inline_forbidden_into_p. > + * Makefile.in (ipa-inline.o): Update arguments. > +
This caused: http://gcc.gnu.org/bugzilla/show_bug.cgi?id=48683 -- H.J.
