[2/n] OpenMP 4.0 offloading infrastructure: LTO streaming

On 29 Sep 03:10, Jan Hubicka wrote:
> dump for me implied debug dump. LTO is usually called streaming, so prehaps
> need_lto_stremaing?

Fixed.

> > +initialize_offload (void)
> Perhaps have_offload_p? Nothing is initialized here...

The next patch will add some initialization to this function.  And they'll be
committed in a series.  So, I'd prefer to keep this name.

> How does LTO combine with offloading?

Both .gnu.lto_ and .gnu.target_lto_ sections are created.  LTO just ignores
target sections, and offload compiler ignores .gnu.lto_ sections.  Everything
works fine on my testcases.

> > @@ -4325,11 +4325,6 @@ void
> >  inline_free_summary (void)
> >  {
> >    struct cgraph_node *node;
> > -  if (!inline_edge_summary_vec.exists ())
> > -    return;
> > -  FOR_EACH_DEFINED_FUNCTION (node)
> > -    if (!node->alias)
> > -      reset_inline_summary (node);
> >    if (function_insertion_hook_holder)
> >      symtab->remove_cgraph_insertion_hook (function_insertion_hook_holder);
> >    function_insertion_hook_holder = NULL;
> > @@ -4345,6 +4340,11 @@ inline_free_summary (void)
> >    if (edge_duplication_hook_holder)
> >      symtab->remove_edge_duplication_hook (edge_duplication_hook_holder);
> >    edge_duplication_hook_holder = NULL;
> > +  if (!inline_edge_summary_vec.exists ())
> > +    return;
> > +  FOR_EACH_DEFINED_FUNCTION (node)
> > +    if (!node->alias)
> > +      reset_inline_summary (node);
> 
> Why this is needed?

Without this change gcc/testsuite/g++.dg/gomp/declare-simd-1.C will fail at -O0,
since inline_generate_summary adds add_new_function hook, but at -O0 the
inline_edge_summary_vec is empty, and we don't call remove_cgraph_insertion_hook
( https://gcc.gnu.org/ml/gcc-patches/2014-02/msg00055.html )

> >  lto_set_symtab_encoder_in_partition (lto_symtab_encoder_t encoder,
> >  				     symtab_node *node)
> >  {
> > +  /* Ignore not needed nodes.  */
> > +  if (!node->need_dump)
> > +    return;
> 
> I think it should be rather done at caller side (in the loop setting what to output)
> rather than in this simple datastructure accestor.

Done.

> > +      /* Ignore references from non-target functions in offload lto mode.  */
> > +      if (offload_lto_mode
> > +	  && !lookup_attribute ("omp declare target",
> > +				DECL_ATTRIBUTES (ref->referring->decl)))
> > +	continue;
> 
> Those are quite busy loops, you may consder making offload a flag.  Why you can't test
> need_dump here?

Definitely.  I have no idea why I did not used this flag here :)  Fixed.

> I think you also need to run free lang data when you decide to stream something.

When I compile a file with offloading, but without -flto, I see free lang data,
executed during all_small_ipa_passes:

#0  free_lang_data () at gcc/tree.c:5655
#1  in (anonymous namespace)::pass_ipa_free_lang_data::execute (this=0x20ce470) at gcc/tree.c:5708
#2  in execute_one_pass (pass=0x20ce470) at gcc/passes.c:2151
#3  in execute_ipa_pass_list (pass=0x20ce470) at gcc/passes.c:2543
#4  in ipa_passes () at gcc/cgraphunit.c:2055
#5  in symbol_table::compile (this=0x7ffff19fd000) at gcc/cgraphunit.c:2187
#6  in symbol_table::finalize_compilation_unit (this=0x7ffff19fd000) at gcc/cgraphunit.c:2340
#7  in c_write_global_declarations () at gcc/c/c-decl.c:10431
#8  in compile_file () at gcc/toplev.c:566
#9  in do_compile () at gcc/toplev.c:1949
#10 in toplev_main (argc=17, argv=0x7fffffffe3a8) at gcc/toplev.c:2025
#11 in main (argc=17, argv=0x7fffffffe3a8) at gcc/main.c:36

> Otherwise the cgraph bits seems resonable. I think Richi will want to comment on LTO
> part.

Here is updated patch.  Bootstrapped and regtested.
OK for trunk (after all patches from the series will be approved)?

Thanks,
  -- Ilya

gcc/
	* cgraph.h (symtab_node): Add need_lto_streaming flag.
	* cgraphunit.c: Include lto-section-names.h.
	(initialize_offload): New function.
	(ipa_passes): Initialize offload and call ipa_write_summaries if there
	is something to write to OMP_SECTION_NAME_PREFIX sections.
	(symbol_table::compile): Call lto_streamer_hooks_init under flag_openmp.
	* ipa-inline-analysis.c (inline_generate_summary): Do not exit under
	flag_openmp.
	(inline_free_summary): Always remove hooks.
	* lto-cgraph.c (referenced_from_other_partition_p): Ignore references
	from non-target functions to target functions if we are streaming out
	target-side bytecode (offload lto mode).
	(reachable_from_other_partition_p): Likewise.
	(select_what_to_stream): New function.
	(compute_ltrans_boundary): Do not call
	lto_set_symtab_encoder_in_partition if the node should not be streamed.
	* lto-section-names.h (OMP_SECTION_NAME_PREFIX): Define.
	(section_name_prefix): Declare.
	* lto-streamer.c (section_name_prefix): New variable.
	(lto_get_section_name): Use section_name_prefix instead of
	LTO_SECTION_NAME_PREFIX.
	* lto-streamer.h (select_what_to_stream): Declare.
	* omp-low.c (is_targetreg_ctx): New function.
	(create_omp_child_function, check_omp_nesting_restrictions): Use it.
	(expand_omp_target): Set mark_force_output for the target functions.
	(lower_omp_critical): Add target attribute for omp critical symbol.
	* passes.c (ipa_write_summaries): New argument offload_lto_mode.  Call
	select_what_to_stream.  Do not call lto_set_symtab_encoder_in_partition
	if the node should not be streamed out.
	* tree-pass.h (ipa_write_summaries): New bool argument.
gcc/lto/
	* lto-object.c (lto_obj_add_section): Use section_name_prefix instead of
	LTO_SECTION_NAME_PREFIX.
	* lto-partition.c (lto_promote_cross_file_statics): Call
	select_what_to_stream.
	* lto.c (lto_section_with_id): Use section_name_prefix instead of
	LTO_SECTION_NAME_PREFIX.
	(read_cgraph_and_symbols): Read OMP_SECTION_NAME_PREFIX sections, if
	being built as an offload compiler.

---

[2/n] OpenMP 4.0 offloading infrastructure: LTO streaming

Commit Message

Comments

Patch