On March 26, 2015 3:41:12 PM GMT+01:00, Jan Hubicka <hubi...@ucw.cz> wrote:
>>
>> Does that work on callgraph cycles?
>
>Yes. Because ...
>> > +
>> > + node->set_nothrow_flag (true);
>> > + if (dump_file)
>> > + fprintf (dump_file, "Function found to be nothrow: %s\n",
>> > + current_function_name ());
>> > + return execute_fixup_cfg ();
>>
>> Err - this doesn't make sense - you'd need to call this for
>> all _callers_. That is, this is dealt with by
>>
>> PUSH_INSERT_PASSES_WITHIN (pass_build_ssa_passes)
>> NEXT_PASS (pass_fixup_cfg);
>> ^^^^^^^
>>
>> ? If not then this may paper over the issue I saw (dangling
>> call_stmt on a cgraph edge?).
>
>... this one is unnecesary. Fixup happens in the following
>local optimization pass queue, so we do not need fixing here.
But the fixup is required to benefit from the notgrow discovery in the callers.
Just doing it on the current function body does not help into-SSA
>I remember I removed this at one point, but apparently accidentally
>reverted
>the change. Will re-test with return 0.
>
>OK with these changes (and commiting separately the checking bits)?
See above.
>>
>> So please re-check with the above execute_fixup_cfg removed
>> (your simple patch doesn't handle direct recursion optimistically).
>
>Ah, you are right. I will add it - that is easy enoug to check.
But then you need to do the fixup or remember the direct recursion sites and
fix them up manually.
>Thanks,
>Honza
>>
>> Thanks,
>> Richard.
>>
>> > +}
>> > +
>> > +} // anon namespace
>> > +
>> > +gimple_opt_pass *
>> > +make_pass_nothrow (gcc::context *ctxt)
>> > +{
>> > + return new pass_nothrow (ctxt);
>> > +}
>> > Index: tree-pass.h
>> > ===================================================================
>> > --- tree-pass.h (revision 221682)
>> > +++ tree-pass.h (working copy)
>> > @@ -436,6 +436,7 @@ extern gimple_opt_pass *make_pass_remove
>> > *ctxt);
>> > extern gimple_opt_pass *make_pass_build_cgraph_edges (gcc::context
>*ctxt);
>> > extern gimple_opt_pass *make_pass_local_pure_const (gcc::context
>*ctxt);
>> > +extern gimple_opt_pass *make_pass_nothrow (gcc::context *ctxt);
>> > extern gimple_opt_pass *make_pass_tracer (gcc::context *ctxt);
>> > extern gimple_opt_pass *make_pass_warn_unused_result (gcc::context
>*ctxt);
>> > extern gimple_opt_pass *make_pass_diagnose_tm_blocks (gcc::context
>*ctxt);
>> > Index: ipa-inline-analysis.c
>> > ===================================================================
>> > --- ipa-inline-analysis.c (revision 221682)
>> > +++ ipa-inline-analysis.c (working copy)
>> > @@ -769,11 +769,16 @@ edge_set_predicate (struct cgraph_edge *
>> >
>> > /* If the edge is determined to be never executed, redirect it
>> > to BUILTIN_UNREACHABLE to save inliner from inlining into it.
> */
>> > - if (predicate && false_predicate_p (predicate) && e->callee)
>> > + if (predicate && false_predicate_p (predicate))
>> > {
>> > struct cgraph_node *callee = !e->inline_failed ? e->callee :
>NULL;
>> > -
>> > - e->redirect_callee (cgraph_node::get_create
>> > + if (e->speculative)
>> > + e->resolve_speculation (builtin_decl_implicit
>(BUILT_IN_UNREACHABLE));
>> > + if (!e->callee)
>> > + e->make_direct (cgraph_node::get_create
>> > + (builtin_decl_implicit (BUILT_IN_UNREACHABLE)));
>> > + else
>> > + e->redirect_callee (cgraph_node::get_create
>> > (builtin_decl_implicit (BUILT_IN_UNREACHABLE)));
>> > e->inline_failed = CIF_UNREACHABLE;
>> > es->call_stmt_size = 0;
>> >
>> >
>>
>> --
>> Richard Biener <rguent...@suse.de>
>> SUSE LINUX GmbH, GF: Felix Imendoerffer, Jane Smithard, Jennifer
>Guild,
>> Dilip Upmanyu, Graham Norton HRB 21284 (AG Nuernberg)
