From mboxrd@z Thu Jan  1 00:00:00 1970
Return-Path: <gcc-patches-return-290409-listarch-gcc-patches=gcc.gnu.org@gcc.gnu.org>
Received: (qmail 453 invoked by alias); 25 Apr 2011 23:39:08 -0000
Received: (qmail 31142 invoked by uid 22791); 25 Apr 2011 23:38:31 -0000
X-SWARE-Spam-Status: No, hits=1.1 required=5.0	tests=BAYES_50,DKIM_SIGNED,DKIM_VALID,DKIM_VALID_AU,SPF_HELO_PASS,TW_BJ,TW_CP,TW_JC,TW_JL,TW_VP,T_RP_MATCHES_RCVD
X-Spam-Check-By: sourceware.org
Received: from smtp-out.google.com (HELO smtp-out.google.com) (74.125.121.67)    by sourceware.org (qpsmtpd/0.43rc1) with ESMTP; Mon, 25 Apr 2011 23:37:43 +0000
Received: from kpbe12.cbf.corp.google.com (kpbe12.cbf.corp.google.com [172.25.105.76])	by smtp-out.google.com with ESMTP id p3PNbemH002726	for <gcc-patches@gcc.gnu.org>; Mon, 25 Apr 2011 16:37:40 -0700
Received: from pvg11 (pvg11.prod.google.com [10.241.210.139])	by kpbe12.cbf.corp.google.com with ESMTP id p3PNbZPm007376	(version=TLSv1/SSLv3 cipher=RC4-SHA bits=128 verify=NOT)	for <gcc-patches@gcc.gnu.org>; Mon, 25 Apr 2011 16:37:38 -0700
Received: by pvg11 with SMTP id 11so90580pvg.13        for <gcc-patches@gcc.gnu.org>; Mon, 25 Apr 2011 16:37:38 -0700 (PDT)
MIME-Version: 1.0
Received: by 10.143.177.6 with SMTP id e6mr46097wfp.223.1303774657920; Mon, 25 Apr 2011 16:37:37 -0700 (PDT)
Received: by 10.142.202.10 with HTTP; Mon, 25 Apr 2011 16:37:37 -0700 (PDT)
In-Reply-To: <BANLkTimFO5r7qW891YT2jXfWUdTd+whX0w@mail.gmail.com>
References: <20110425232927.75B241B40B6@aples.mtv.corp.google.com>	<BANLkTimFO5r7qW891YT2jXfWUdTd+whX0w@mail.gmail.com>
Date: Tue, 26 Apr 2011 02:15:00 -0000
Message-ID: <BANLkTikZYND8meqYW+wH02EvR+YXV2MYrg@mail.gmail.com>
Subject: Re: [google] Port LIPO support to google/main (issue4431065)
From: Rong Xu <xur@google.com>
To: Xinliang David Li <davidxl@google.com>
Cc: reply@codereview.appspotmail.com, gcc-patches@gcc.gnu.org
Content-Type: text/plain; charset=ISO-8859-1
Content-Transfer-Encoding: quoted-printable
X-System-Of-Record: true
Mailing-List: contact gcc-patches-help@gcc.gnu.org; run by ezmlm
Precedence: bulk
List-Id: <gcc-patches.gcc.gnu.org>
List-Archive: <http://gcc.gnu.org/ml/gcc-patches/>
List-Post: <mailto:gcc-patches@gcc.gnu.org>
List-Help: <mailto:gcc-patches-help@gcc.gnu.org>
Sender: gcc-patches-owner@gcc.gnu.org
X-SW-Source: 2011-04/txt/msg02013.txt.bz2

LGTM

On Mon, Apr 25, 2011 at 4:32 PM, Xinliang David Li <davidxl@google.com> wro=
te:
> Bootstrapped on x86_64/linux with all regression tests passing.
>
> Additional tests: SPEC2k, SPEC06 FDO and LIPO testing.
>
> Note -- there will be work in move LIPO away from FE based solution,
> this patch is needed to make the base for forward progress.
>
> David
>
> On Mon, Apr 25, 2011 at 4:29 PM, David Li <davidxl@google.com> wrote:
>> 2011-04-25 =A0David Li =A0<davidxl@google.com>
>>
>> =A0 =A0 =A0 =A0* cgraphbuild.c (rebuild_cgraph_edges): Create edges based
>> =A0 =A0 =A0 =A0using resolved nodes in LIPO mode after ipa_tree_profilin=
g.
>> =A0 =A0 =A0 =A0* dyn-ipa.c: =A0New file.
>> =A0 =A0 =A0 =A0* dbgcnt.def: New debug counters.
>> =A0 =A0 =A0 =A0* c-family/c-opts.c (c_common_post_options): New option.
>> =A0 =A0 =A0 =A0(c_common_parse_file): LIPO mode driver change.
>> =A0 =A0 =A0 =A0(push_command_line_include): New function.
>> =A0 =A0 =A0 =A0* cgraph.c (cgraph_mark_reachable_node): Handle aux modul=
e nodes.
>> =A0 =A0 =A0 =A0(cgraph_clone_node): New field initialization.
>> =A0 =A0 =A0 =A0(cgraph_create_virtual_clone): New field initialization.
>> =A0 =A0 =A0 =A0* value-prof.c (dump_histogram_value): New hist kind.
>> =A0 =A0 =A0 =A0(check_counter): New function.
>> =A0 =A0 =A0 =A0(gimple_value_profile_transformations): New ic kind.
>> =A0 =A0 =A0 =A0(gimple_ic_transform): Changed interface.
>> =A0 =A0 =A0 =A0(gimple_indirect_call_to_profile): New vpt kind.
>> =A0 =A0 =A0 =A0(gimple_find_values_to_profile): New vpt kind.
>> =A0 =A0 =A0 =A0* cgraphunit.c (cgraph_finalize_function): Handle LIPO mo=
de.
>> =A0 =A0 =A0 =A0(verify_edge_count_and_frequency): Ditto.
>> =A0 =A0 =A0 =A0(verify_cgraph_node): Ditto.
>> =A0 =A0 =A0 =A0(cgraph_finalize_compilation_unit): Ditto.
>> =A0 =A0 =A0 =A0(cgraph_mark_functions_to_output): Ditto.
>> =A0 =A0 =A0 =A0(cgraph_expand_function): Ditto.
>> =A0 =A0 =A0 =A0(cgraph_output_in_order): Ditto.
>> =A0 =A0 =A0 =A0(cgraph_optimize): Ditto.
>> =A0 =A0 =A0 =A0(cgraph_copy_node_for_versioning): Ditto.
>> =A0 =A0 =A0 =A0(cgraph_materialize_clone): Ditto.
>> =A0 =A0 =A0 =A0(cgraph_redirect_edge_call_stmt_to_callee): Ditto.
>> =A0 =A0 =A0 =A0* cp/decl.c (walk_namespaces): Ditto.
>> =A0 =A0 =A0 =A0(duplicate_decls): Ditto.
>> =A0 =A0 =A0 =A0(make_rtl_for_nonlocal_decl): Ditto.
>> =A0 =A0 =A0 =A0* cp/rtti.c (get_tinfo_decl): Ditto.
>> =A0 =A0 =A0 =A0(create_pseudo_type_info): Ditto.
>> =A0 =A0 =A0 =A0(create_tinfo_types): Ditto.
>> =A0 =A0 =A0 =A0* cp/cp-lang.c (cp_classify_record): Ditto.
>> =A0 =A0 =A0 =A0* cp/pt.c (instantiate_decl): Ditto.
>> =A0 =A0 =A0 =A0* cp/semantics.c (emit_associated_thunks): Ditto.
>> =A0 =A0 =A0 =A0* cp/decl2.c (start_static_storage_duration_function): Di=
tto.
>> =A0 =A0 =A0 =A0(prune_vars_needing_no_initialization): =A0Ditto.
>> =A0 =A0 =A0 =A0(cxx_callgraph_analyze_expr): Ditto.
>> =A0 =A0 =A0 =A0(no_linkage_error): Ditto.
>> =A0 =A0 =A0 =A0(collect_all_refs): Ditto.
>> =A0 =A0 =A0 =A0(cp_write_global_declarations): Ditto.
>> =A0 =A0 =A0 =A0* cp/parser.c (pragma_lex): Ditto.
>> =A0 =A0 =A0 =A0* cp/cp-objcp-common.c (cp_function_decl_explicit_p): Dit=
to.
>> =A0 =A0 =A0 =A0* cp/mangle.c (mangle_conv_op_name_for_type): Ditto.
>> =A0 =A0 =A0 =A0* cp/name-lookup.c (add_decl_to_level): Ditto.
>> =A0 =A0 =A0 =A0* gcov-io.c (gcov_read_summary): Ditto.
>> =A0 =A0 =A0 =A0* tree-ssa-alias.c (same_type_for_tbaa): Ditto.
>> =A0 =A0 =A0 =A0(refs_may_alias_p_1): Ditto.
>> =A0 =A0 =A0 =A0(walk_aliased_vdefs): Ditto.
>> =A0 =A0 =A0 =A0* ipa-inline.c (cgraph_mark_inline_edge): Ditto.
>> =A0 =A0 =A0 =A0(cgraph_recursive_inlining_p): Ditto.
>> =A0 =A0 =A0 =A0(cgraph_edge_badness): Ditto.
>> =A0 =A0 =A0 =A0(update_caller_keys): Ditto.
>> =A0 =A0 =A0 =A0(cgraph_decide_recursive_inlining): Ditto.
>> =A0 =A0 =A0 =A0(cgraph_decide_inlining_of_small_function): Ditto.
>> =A0 =A0 =A0 =A0(cgraph_decide_inlining): Ditto.
>> =A0 =A0 =A0 =A0(cgraph_decide_inlining_incrementally): Ditto.
>> =A0 =A0 =A0 =A0(estimate_function_body_sizes): Ditto.
>> =A0 =A0 =A0 =A0* dwarf2out.c (dwarf2out_do_cfi_startproc): Ditto.
>> =A0 =A0 =A0 =A0(dwarf2out_begin_prologue): Ditto.
>> =A0 =A0 =A0 =A0(dwarf2out_end_epilogue): Ditto.
>> =A0 =A0 =A0 =A0(dw_loc_list): Ditto.
>> =A0 =A0 =A0 =A0(gen_subprogram_die): Ditto.
>> =A0 =A0 =A0 =A0(dwarf2out_source_line): Ditto.
>> =A0 =A0 =A0 =A0* c-decl.c (bind): Ditto.
>> =A0 =A0 =A0 =A0(pop_scope): Ditto. Ditto.
>> =A0 =A0 =A0 =A0(push_file_scope): Ditto.
>> =A0 =A0 =A0 =A0(pop_file_scope): Ditto.
>> =A0 =A0 =A0 =A0(finish_decl): Ditto.
>> =A0 =A0 =A0 =A0(c_write_global_declarations): Ditto.
>> =A0 =A0 =A0 =A0* langhooks.c (lhd_do_nothing_f): Ditto.
>> =A0 =A0 =A0 =A0(lhd_do_nothing_t_t_return_null_tree): Ditto.
>> =A0 =A0 =A0 =A0(lhd_return_null_tree_v): Ditto.
>> =A0 =A0 =A0 =A0(lhd_builtin_function): Ditto.
>> =A0 =A0 =A0 =A0* function.c (pop_cfun): Ditto.
>> =A0 =A0 =A0 =A0(allocate_struct_function): Ditto.
>> =A0 =A0 =A0 =A0* profile.c (instrument_values): Ditto.
>> =A0 =A0 =A0 =A0(is_edge_inconsistent): Ditto.
>> =A0 =A0 =A0 =A0* ipa.c (process_references): Ditto.
>> =A0 =A0 =A0 =A0(cgraph_remove_unreachable_nodes): Ditto.
>> =A0 =A0 =A0 =A0(function_and_variable_visibility):
>> =A0 =A0 =A0 =A0* c-typeck.c (tagged_types_tu_compatible_p): Ditto.
>> =A0 =A0 =A0 =A0* gimplify.c (gimplify_addr_expr): Ditto.
>> =A0 =A0 =A0 =A0* except.c (sjlj_emit_function_enter): Ditto.
>> =A0 =A0 =A0 =A0(output_one_function_exception_table): Ditto.
>> =A0 =A0 =A0 =A0* coverage.c (get_gcov_type): Ditto.
>> =A0 =A0 =A0 =A0(htab_counts_entry_del): Ditto.
>> =A0 =A0 =A0 =A0(read_counts_file): Ditto.
>> =A0 =A0 =A0 =A0(get_coverage_counts): Ditto.
>> =A0 =A0 =A0 =A0(coverage_checksum_string): Ditto.
>> =A0 =A0 =A0 =A0(coverage_begin_output): Ditto.
>> =A0 =A0 =A0 =A0(coverage_end_function): Ditto.
>> =A0 =A0 =A0 =A0(build_fn_info_type): Ditto.
>> =A0 =A0 =A0 =A0(build_fn_info_value): Ditto.
>> =A0 =A0 =A0 =A0(build_ctr_info_value): Ditto.
>> =A0 =A0 =A0 =A0(build_gcov_info): Ditto.
>> =A0 =A0 =A0 =A0(create_coverage): Ditto.
>> =A0 =A0 =A0 =A0(coverage_finish): Ditto.
>> =A0 =A0 =A0 =A0* tree-sra.c (convert_callers): Ditto.
>> =A0 =A0 =A0 =A0(modify_function): Ditto.
>> =A0 =A0 =A0 =A0(ipa_early_sra): Ditto.
>> =A0 =A0 =A0 =A0* varasm.c (notice_global_symbol): Ditto.
>> =A0 =A0 =A0 =A0(assemble_external): Ditto.
>> =A0 =A0 =A0 =A0(mark_decl_referenced): Ditto.
>> =A0 =A0 =A0 =A0(find_decl_and_mark_needed): Ditto.
>> =A0 =A0 =A0 =A0(finish_aliases_1): Ditto.
>> =A0 =A0 =A0 =A0(assemble_alias): Ditto.
>> =A0 =A0 =A0 =A0* tree-ssa.c (useless_type_conversion_p): Ditto.
>> =A0 =A0 =A0 =A0* tree-inline.c (copy_bb): Ditto.
>> =A0 =A0 =A0 =A0(copy_edges_for_bb): Ditto.
>> =A0 =A0 =A0 =A0(initialize_cfun): Ditto.
>> =A0 =A0 =A0 =A0(copy_cfg_body): Ditto.
>> =A0 =A0 =A0 =A0(tree_inlinable_function_p): Ditto.
>> =A0 =A0 =A0 =A0(optimize_inline_calls): Ditto.
>> =A0 =A0 =A0 =A0(copy_tree_r): Ditto.
>> =A0 =A0 =A0 =A0* tree-profile.c (init_ic_make_global_vars): Ditto.
>> =A0 =A0 =A0 =A0(gimple_init_edge_profiler): Ditto.
>> =A0 =A0 =A0 =A0(gimple_gen_ic_profiler): Ditto.
>> =A0 =A0 =A0 =A0(gimple_gen_ic_func_profiler): Ditto.
>> =A0 =A0 =A0 =A0(tree_profiling): Ditto.
>> =A0 =A0 =A0 =A0* opts-global.c (lang_handle_option): Ditto.
>> =A0 =A0 =A0 =A0(add_input_filename): Ditto.
>> =A0 =A0 =A0 =A0(read_cmdline_options): Ditto.
>> =A0 =A0 =A0 =A0* l-ipo.c: Ditto.
>> =A0 =A0 =A0 =A0* libgcov.c (gcov_version): Ditto.
>> =A0 =A0 =A0 =A0(gcov_exit): Ditto.
>> =A0 =A0 =A0 =A0(__gcov_init): Ditto.
>> =A0 =A0 =A0 =A0(__gcov_flush): Ditto.
>> =A0 =A0 =A0 =A0(__gcov_merge_ior): Ditto.
>> =A0 =A0 =A0 =A0(__gcov_one_value_profiler_body): Ditto.
>> =A0 =A0 =A0 =A0(__gcov_indirect_call_profiler): Ditto.
>> =A0 =A0 =A0 =A0* l-ipo.h: Ditto.
>> =A0 =A0 =A0 =A0* tree-cfg.c: (verify_types_in_gimple_reference): Ditto.
>> =A0 =A0 =A0 =A0(verify_gimple_call): Ditto.
>> =A0 =A0 =A0 =A0(verify_gimple_assign_single): Ditto.
>> =A0 =A0 =A0 =A0(verify_gimple_return): Ditto.
>> =A0 =A0 =A0 =A0(gimple_verify_flow_info): Ditto.
>> =A0 =A0 =A0 =A0* passes.c: (rest_of_decl_compilation): Ditto.
>> =A0 =A0 =A0 =A0(init_optimization_passes): Ditto.
>> =A0 =A0 =A0 =A0(pass_init_dump_file): Ditto.
>> =A0 =A0 =A0 =A0* varpool.c (varpool_node): Ditto.
>> =A0 =A0 =A0 =A0(debug_varpool): Ditto.
>> =A0 =A0 =A0 =A0(varpool_node_for_asm): Ditto.
>> =A0 =A0 =A0 =A0(varpool_enqueue_needed_node): Ditto.
>> =A0 =A0 =A0 =A0* stmt.c (expand_asm_operands): Ditto.
>> =A0 =A0 =A0 =A0* gcov-dump.c (main): Ditto.
>> =A0 =A0 =A0 =A0(print_usage): Ditto.
>> =A0 =A0 =A0 =A0(print_prefix): Ditto.
>> =A0 =A0 =A0 =A0(dump_file): Ditto.
>> =A0 =A0 =A0 =A0(tag_summary): Ditto.
>>
>> Index: doc/invoke.texi
>> =3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=
=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=
=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D
>> --- doc/invoke.texi =A0 =A0 (revision 172880)
>> +++ doc/invoke.texi =A0 =A0 (working copy)
>> @@ -258,7 +258,7 @@ Objective-C and Objective-C++ Dialects}.
>> =A0-Wparentheses =A0-Wpedantic-ms-format -Wno-pedantic-ms-format @gol
>> =A0-Wpointer-arith =A0-Wno-pointer-to-int-cast @gol
>> =A0-Wredundant-decls @gol
>> --Wreturn-type =A0-Wsequence-point =A0-Wshadow @gol
>> +-Wreturn-type =A0-Wripa-opt-mismatch =A0-Wsequence-point =A0-Wshadow @g=
ol
>> =A0-Wsign-compare =A0-Wsign-conversion =A0-Wstack-protector @gol
>> =A0-Wstrict-aliasing -Wstrict-aliasing=3Dn @gol
>> =A0-Wstrict-overflow -Wstrict-overflow=3D@var{n} @gol
>> @@ -380,7 +380,9 @@ Objective-C and Objective-C++ Dialects}.
>> =A0-freciprocal-math -fregmove -frename-registers -freorder-blocks @gol
>> =A0-freorder-blocks-and-partition -freorder-functions @gol
>> =A0-frerun-cse-after-loop -freschedule-modulo-scheduled-loops @gol
>> --frounding-math -fsched2-use-superblocks -fsched-pressure @gol
>> +-fripa -fripa-disallow-asm-modules -fripa-disallow-opt-mismatch @gol
>> +-fripa-no-promote-always-inline-func -fripa-verbose -frounding-math @gol
>> +-fsched2-use-superblocks -fsched-pressure @gol
>> =A0-fsched-spec-load -fsched-spec-load-dangerous @gol
>> =A0-fsched-stalled-insns-dep[=3D@var{n}] -fsched-stalled-insns[=3D@var{n=
}] @gol
>> =A0-fsched-group-heuristic -fsched-critical-path-heuristic @gol
>> @@ -1364,8 +1366,8 @@ prefix.
>> =A0@opindex no-canonical-prefixes
>> =A0Never expand any symbolic links, resolve references to @samp{/../}
>> =A0or @samp{/./}, or make the path absolute when generating a relative
>> -prefix. =A0If neither @option{-canonical-prefixes} nor
>> -@option{-canonical-prefixes} is given, GCC tries to set an appropriate
>> +prefix. If neither @option{-canonical-prefixes} nor
>> +@option{-nocanonical-prefixes} is given, GCC tries to set an appropriate
>> =A0default by looking for a target-specific subdirectory alongside the
>> =A0directory containing the compiler driver.
>>
>> @@ -2969,6 +2971,7 @@ Options} and @ref{Objective-C and Object
>> =A0-Wpointer-sign =A0@gol
>> =A0-Wreorder =A0 @gol
>> =A0-Wreturn-type =A0@gol
>> +-Wripa-opt-mismatch @gol
>> =A0-Wsequence-point =A0@gol
>> =A0-Wsign-compare @r{(only in C++)} =A0@gol
>> =A0-Wstrict-aliasing =A0@gol
>> @@ -3395,6 +3398,16 @@ exceptions are @samp{main} and functions
>>
>> =A0This warning is enabled by @option{-Wall}.
>>
>> +@item -Wripa-opt-mismatch
>> +@opindex Wripa-opt-mismatch
>> +@opindex Wno-ripa-opt-mismatch
>> +When doing an FDO build with @option{-fprofile-use} and @option{-fripa},
>> +warn if importing an axuiliary module that was built with a different
>> +GCC command line during the profile-generate phase than the primary
>> +module.
>> +
>> +This warning is enabled by @option{-Wall}.
>> +
>> =A0@item -Wswitch
>> =A0@opindex Wswitch
>> =A0@opindex Wno-switch
>> @@ -4493,6 +4506,12 @@ Suppress warnings from casts to pointer
>> =A0different size. In C++, casting to a pointer type of smaller size is
>> =A0an error. @option{Wint-to-pointer-cast} is enabled by default.
>>
>> +@item max-lipo-mem
>> +When importing auxiliary modules during profile-use, check current
>> +memory consumption after parsing each auxiliary module. If it exceeds
>> +this limit (specified in kb), don't import any more auxiliary modules.
>> +Specifying a value of 0 means don't enforce this limit. This parameter
>> +is only useful when using @option{-fprofile-use} and @option{-fripa}.
>>
>> =A0@item -Wno-pointer-to-int-cast @r{(C and Objective-C only)}
>> =A0@opindex Wno-pointer-to-int-cast
>> @@ -7864,6 +7883,40 @@ code.
>>
>> =A0If @var{path} is specified, GCC will look at the @var{path} to find
>> =A0the profile feedback data files. See @option{-fprofile-dir}.
>> +
>> +@item -fripa
>> +@opindex fripa
>> +Perform dynamic inter-procedural analysis. This is used in conjunction =
with
>> +the @option{-fprofile-generate} and @option{-fprofile-use} options.
>> +During the @option{-fprofile-generate} phase, this flag turns on some a=
dditional
>> +instrumentation code that enables dynamic call-graph analysis.
>> +During the @option{-fprofile-use} phase, this flag enables cross-module
>> +optimizations such as inlining.
>> +
>> +@item -fripa-disallow-asm-modules
>> +@opindex fripa-disallow-asm-modules
>> +During profile-gen, if this flag is enabled, and the module has asm sta=
tements,
>> +arrange so that a bit recording this information will be set in the pro=
file
>> +feedback data file.
>> +During profile-use, if this flag is enabled, and the same bit in auxili=
ary
>> +module's profile feedback data is set, don't import this auxiliary modu=
le.
>> +If this is the primary module, don't export it.
>> +
>> +@item -fripa-disallow-opt-mismatch
>> +@opindex fripa-disallow-opt-mismatch
>> +Don't import an auxiliary module, if the GCC command line options used =
for this
>> +auxiliary module during the profile-generate stage were different from =
those used
>> +for the primary module. Note that any mismatches in warning-related opt=
ions are
>> +ignored for this comparison.
>> +
>> +@item -fripa-no-promote-always-inline-func
>> +@opindex fripa-no-promote-always-inline-func
>> +Do not promote static functions with always inline attribute in LIPO co=
mpilation.
>> +
>> +@item -fripa-verbose
>> +@opindex fripa-verbose
>> +Enable printing of verbose information about dynamic inter-procedural o=
ptimizations.
>> +This is used in conjunction with the @option{-fripa}.
>> =A0@end table
>>
>> =A0The following options control compiler behavior regarding floating
>> @@ -9954,6 +10007,12 @@ given to GCC, substitutes @code{Y}; else
>> =A0be as many clauses as you need. =A0This may be combined with @code{.},
>> =A0@code{,}, @code{!}, @code{|}, and @code{*} as needed.
>>
>> +@item max-lipo-mem
>> +When importing auxiliary modules during profile-use, check current
>> +memory consumption after parsing each auxiliary module. If it exceeds
>> +this limit (specified in kb), don't import any more auxiliary modules.
>> +Specifying a value of 0 means don't enforce this limit. This parameter
>> +is only useful when using @option{-fprofile-use} and @option{-fripa}.
>>
>> =A0@end table
>>
>> Index: cgraphbuild.c
>> =3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=
=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=
=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D
>> --- cgraphbuild.c =A0 =A0 =A0 (revision 172880)
>> +++ cgraphbuild.c =A0 =A0 =A0 (working copy)
>> @@ -30,9 +30,13 @@ along with GCC; see the file COPYING3.
>> =A0#include "cgraph.h"
>> =A0#include "intl.h"
>> =A0#include "gimple.h"
>> +#include "toplev.h"
>> +#include "gcov-io.h"
>> +#include "coverage.h"
>> =A0#include "tree-pass.h"
>> =A0#include "ipa-utils.h"
>> =A0#include "except.h"
>> +#include "l-ipo.h"
>>
>> =A0/* Context of record_reference. =A0*/
>> =A0struct record_reference_ctx
>> @@ -236,6 +240,140 @@ compute_call_stmt_bb_frequency (tree dec
>> =A0 return freq;
>> =A0}
>>
>> +
>> +bool cgraph_pre_profiling_inlining_done =3D false;
>> +
>> +/* Return true if E is a fake indirect call edge. =A0*/
>> +
>> +bool
>> +cgraph_is_fake_indirect_call_edge (struct cgraph_edge *e)
>> +{
>> + =A0return !e->call_stmt;
>> +}
>> +
>> +
>> +/* Add fake cgraph edges from NODE to its indirect call callees
>> + =A0 using profile data. =A0*/
>> +
>> +static void
>> +add_fake_indirect_call_edges (struct cgraph_node *node)
>> +{
>> + =A0unsigned n_counts, i;
>> + =A0gcov_type *ic_counts;
>> +
>> + =A0/* Enable this only for LIPO for now. =A0*/
>> + =A0if (!L_IPO_COMP_MODE)
>> + =A0 =A0return;
>> +
>> + =A0if (cgraph_pre_profiling_inlining_done)
>> + =A0 =A0return;
>> +
>> + =A0ic_counts
>> + =A0 =A0 =A0=3D get_coverage_counts_no_warn (DECL_STRUCT_FUNCTION (node=
->decl),
>> + =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =
=A0 GCOV_COUNTER_ICALL_TOPNV, &n_counts);
>> +
>> + =A0if (!ic_counts)
>> + =A0 =A0return;
>> +
>> + =A0gcc_assert ((n_counts % GCOV_ICALL_TOPN_NCOUNTS) =3D=3D 0);
>> +
>> +/* After the early_inline_1 before value profile transformation,
>> + =A0 functions that are indirect call targets may have their bodies
>> + =A0 removed (extern inline functions or functions from aux modules,
>> + =A0 functions in comdat etc) if all direct callsites are inlined. This
>> + =A0 will lead to missing inline opportunities after profile based
>> + =A0 indirect call promotion. The solution is to add fake edges to
>> + =A0 indirect call targets. Note that such edges are not associated
>> + =A0 with actual indirect call sites because it is not possible to
>> + =A0 reliably match pre-early-inline indirect callsites with indirect
>> + =A0 call profile counters which are from post-early inline function bo=
dy. =A0*/
>> +
>> + =A0for (i =3D 0; i < n_counts;
>> + =A0 =A0 =A0 i +=3D GCOV_ICALL_TOPN_NCOUNTS, ic_counts +=3D GCOV_ICALL_=
TOPN_NCOUNTS)
>> + =A0 =A0{
>> + =A0 =A0 =A0gcov_type val1, val2, count1, count2;
>> + =A0 =A0 =A0struct cgraph_node *direct_call1 =3D 0, *direct_call2 =3D 0;
>> +
>> + =A0 =A0 =A0val1 =3D ic_counts[1];
>> + =A0 =A0 =A0count1 =3D ic_counts[2];
>> + =A0 =A0 =A0val2 =3D ic_counts[3];
>> + =A0 =A0 =A0count2 =3D ic_counts[4];
>> +
>> + =A0 =A0 =A0if (val1 =3D=3D 0 || count1 =3D=3D 0)
>> + =A0 =A0 =A0 =A0continue;
>> +
>> + =A0 =A0 =A0direct_call1 =3D find_func_by_global_id (val1);
>> + =A0 =A0 =A0if (direct_call1)
>> + =A0 =A0 =A0 =A0{
>> + =A0 =A0 =A0 =A0 =A0tree decl =3D direct_call1->decl;
>> + =A0 =A0 =A0 =A0 =A0cgraph_create_edge (node,
>> + =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 cgraph_node (d=
ecl),
>> + =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 NULL,
>> + =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0count1, 0, =
0);
>> + =A0 =A0 =A0 =A0}
>> +
>> + =A0 =A0 =A0if (val2 =3D=3D 0 || count2 =3D=3D 0)
>> + =A0 =A0 =A0 =A0continue;
>> + =A0 =A0 =A0direct_call2 =3D find_func_by_global_id (val2);
>> + =A0 =A0 =A0if (direct_call2)
>> + =A0 =A0 =A0 =A0{
>> + =A0 =A0 =A0 =A0 =A0tree decl =3D direct_call2->decl;
>> + =A0 =A0 =A0 =A0 =A0cgraph_create_edge (node,
>> + =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 cgraph_node (d=
ecl),
>> + =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0NULL,
>> + =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0count2, 0, =
0);
>> + =A0 =A0 =A0 =A0}
>> + =A0 =A0}
>> +}
>> +
>> +
>> +/* This can be implemented as an IPA pass that must be first one
>> + =A0 before any unreachable node elimination. */
>> +
>> +void
>> +cgraph_add_fake_indirect_call_edges (void)
>> +{
>> + =A0struct cgraph_node *node;
>> +
>> + =A0/* Enable this only for LIPO for now. =A0*/
>> + =A0if (!L_IPO_COMP_MODE)
>> + =A0 =A0return;
>> +
>> + =A0for (node =3D cgraph_nodes; node; node =3D node->next)
>> + =A0 =A0{
>> + =A0 =A0 =A0if (node->analyzed && (node->needed || node->reachable))
>> + =A0 =A0 =A0 =A0add_fake_indirect_call_edges (node);
>> + =A0 =A0}
>> +}
>> +
>> +/* Remove zero count fake edges added for the purpose of ensuring
>> + =A0 the right processing order. =A0This should be called after all
>> + =A0 small ipa passes. =A0*/
>> +void
>> +cgraph_remove_zero_count_fake_edges (void)
>> +{
>> + =A0struct cgraph_node *node;
>> +
>> + =A0/* Enable this only for LIPO for now. =A0*/
>> + =A0if (!L_IPO_COMP_MODE)
>> + =A0 =A0return;
>> +
>> + =A0for (node =3D cgraph_nodes; node; node =3D node->next)
>> + =A0 =A0{
>> + =A0 =A0 =A0if (node->analyzed && (node->needed || node->reachable))
>> + =A0 =A0 =A0 =A0{
>> + =A0 =A0 =A0 =A0 =A0struct cgraph_edge *e, *f;
>> + =A0 =A0 =A0 =A0 for (e =3D node->callees; e; e =3D f)
>> + =A0 =A0 =A0 =A0 =A0 {
>> + =A0 =A0 =A0 =A0 =A0 =A0 f =3D e->next_callee;
>> + =A0 =A0 =A0 =A0 =A0 =A0 if (!e->call_stmt && !e->count && !e->loop_nest
>> + =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 && !e->frequency)
>> + =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0cgraph_remove_edge (e);
>> + =A0 =A0 =A0 =A0 =A0 =A0}
>> + =A0 =A0 =A0 }
>> + =A0 =A0}
>> +}
>> +
>> =A0/* Mark address taken in STMT. =A0*/
>>
>> =A0static bool
>> @@ -390,6 +528,7 @@ build_cgraph_edges (void)
>> =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =
=A0 mark_load, mark_store, mark_address);
>> =A0 =A0}
>>
>> +
>> =A0 /* Look for initializers of constant variables and private statics. =
=A0*/
>> =A0 FOR_EACH_LOCAL_DECL (cfun, ix, decl)
>> =A0 =A0 if (TREE_CODE (decl) =3D=3D VAR_DECL
>> @@ -466,9 +605,24 @@ rebuild_cgraph_edges (void)
>> =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =
=A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 bb);
>> =A0 =A0 =A0 =A0 =A0 =A0 =A0decl =3D gimple_call_fndecl (stmt);
>> =A0 =A0 =A0 =A0 =A0 =A0 =A0if (decl)
>> - =A0 =A0 =A0 =A0 =A0 =A0 =A0 cgraph_create_edge (node, cgraph_node (dec=
l), stmt,
>> - =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 bb=
->count, freq,
>> - =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 bb=
->loop_depth);
>> + =A0 =A0 =A0 =A0 =A0 =A0 =A0 {
>> + =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 struct cgraph_node *callee;
>> + =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 /* In LIPO mode, before tree_profiling=
, the call graph edge
>> + =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0needs to be built with the orig=
inal target node to make
>> + =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0sure consistent early inline de=
cisions between profile generate
>> + =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0and profile use. After tree-pro=
filing, the target needs to be
>> + =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0set to the resolved node so tha=
t ipa-inline sees the definitions. =A0*/
>> + =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 if (L_IPO_COMP_MODE && cgraph_pre_prof=
iling_inlining_done)
>> + =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 callee =3D cgraph_lipo_get_resolve=
d_node (decl);
>> + =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0else
>> + =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 callee =3D cgraph_node (decl);
>> + =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0cgraph_create_edge (node, callee, s=
tmt,
>> + =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =
=A0 =A0bb->count, freq,
>> + =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =
=A0 =A0bb->loop_depth);
>> + =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0if (L_IPO_COMP_MODE && cgraph_pre_p=
rofiling_inlining_done
>> + =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 && decl !=3D callee->decl)
>> + =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 gimple_call_set_fndecl (stmt, call=
ee->decl);
>> + =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0}
>> =A0 =A0 =A0 =A0 =A0 =A0 =A0else
>> =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0cgraph_create_indirect_edge (node, stmt,
>> =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =
=A0 =A0 =A0 =A0 gimple_call_flags (stmt),
>> @@ -483,6 +637,7 @@ rebuild_cgraph_edges (void)
>> =A0 =A0 =A0 =A0walk_stmt_load_store_addr_ops (gsi_stmt (gsi), node,
>> =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =
=A0 mark_load, mark_store, mark_address);
>> =A0 =A0 }
>> + =A0add_fake_indirect_call_edges (node);
>> =A0 record_eh_tables (node, cfun);
>> =A0 gcc_assert (!node->global.inlined_to);
>>
>> Index: dyn-ipa.c
>> =3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=
=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=
=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D
>> --- dyn-ipa.c =A0 (revision 0)
>> +++ dyn-ipa.c =A0 (revision 0)
>> @@ -0,0 +1,1476 @@
>> +/* Compile this one with gcc. =A0*/
>> +/* Copyright (C) 2009. Free Software Foundation, Inc.
>> + =A0 Contributed by Xinliang David Li (davidxl@google.com) and
>> + =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0Raksit Ashok =A0(raksit@google.com)
>> +
>> +This file is part of GCC.
>> +
>> +GCC is free software; you can redistribute it and/or modify it under
>> +the terms of the GNU General Public License as published by the Free
>> +Software Foundation; either version 3, or (at your option) any later
>> +version.
>> +
>> +GCC is distributed in the hope that it will be useful, but WITHOUT ANY
>> +WARRANTY; without even the implied warranty of MERCHANTABILITY or
>> +FITNESS FOR A PARTICULAR PURPOSE. =A0See the GNU General Public License
>> +for more details.
>> +
>> +Under Section 7 of GPL version 3, you are granted additional
>> +permissions described in the GCC Runtime Library Exception, version
>> +3.1, as published by the Free Software Foundation.
>> +
>> +You should have received a copy of the GNU General Public License and
>> +a copy of the GCC Runtime Library Exception along with this program;
>> +see the files COPYING3 and COPYING.RUNTIME respectively. =A0If not, see
>> +<http://www.gnu.org/licenses/>. =A0*/
>> +
>> +#include "tconfig.h"
>> +#include "tsystem.h"
>> +#include "coretypes.h"
>> +#include "tm.h"
>> +
>> +#if defined(inhibit_libc)
>> +#define IN_LIBGCOV (-1)
>> +#else
>> +#undef NULL /* Avoid errors if stdio.h and our stddef.h mismatch. =A0*/
>> +#include <stdio.h>
>> +#include <stdlib.h>
>> +#define IN_LIBGCOV 1
>> +#if defined(L_gcov)
>> +#define GCOV_LINKAGE /* nothing */
>> +#endif
>> +#endif
>> +#include "gcov-io.h"
>> +
>> +struct dyn_pointer_set;
>> +
>> +#define XNEWVEC(type,ne) (type *)malloc(sizeof(type) * (ne))
>> +#define XNEW(type) (type *)malloc(sizeof(type))
>> +#define XDELETEVEC(p) free(p)
>> +#define XDELETE(p) free(p)
>> +
>> +struct dyn_cgraph_node
>> +{
>> + =A0struct dyn_cgraph_edge *callees;
>> + =A0struct dyn_cgraph_edge *callers;
>> + =A0struct dyn_pointer_set *imported_modules;
>> +
>> + =A0gcov_type guid;
>> + =A0gcov_type sum_in_count;
>> + =A0gcov_unsigned_t visited;
>> +};
>> +
>> +struct dyn_cgraph_edge
>> +{
>> + =A0struct dyn_cgraph_node *caller;
>> + =A0struct dyn_cgraph_node *callee;
>> + =A0struct dyn_cgraph_edge *next_caller;
>> + =A0struct dyn_cgraph_edge *next_callee;
>> + =A0gcov_type count;
>> +};
>> +
>> +struct dyn_module_info
>> +{
>> + =A0struct dyn_pointer_set *imported_modules;
>> + =A0gcov_unsigned_t max_func_ident;
>> +};
>> +
>> +struct dyn_cgraph
>> +{
>> + =A0struct dyn_pointer_set **call_graph_nodes;
>> + =A0struct gcov_info **modules;
>> + =A0/* supplement module information =A0*/
>> + =A0struct dyn_module_info *sup_modules;
>> + =A0const struct gcov_fn_info ***functions;
>> + =A0unsigned num_modules;
>> + =A0unsigned num_nodes_executed;
>> +};
>> +
>> +struct dyn_pointer_set
>> +{
>> + =A0size_t log_slots;
>> + =A0size_t n_slots; =A0 =A0 =A0 =A0 =A0 =A0 =A0/* n_slots =3D 2^log_slo=
ts */
>> + =A0size_t n_elements;
>> +
>> + =A0void **slots;
>> + =A0unsigned (*get_key) (const void *);
>> +};
>> +
>> +
>> +#if defined(inhibit_libc)
>> +__gcov_build_callgraph (void) {}
>> +#else
>> +
>> +void __gcov_compute_module_groups (void) ATTRIBUTE_HIDDEN;
>> +void __gcov_finalize_dyn_callgraph (void) ATTRIBUTE_HIDDEN;
>> +static void gcov_dump_callgraph (gcov_type);
>> +static void gcov_dump_cgraph_node_short (struct dyn_cgraph_node *node);
>> +static void gcov_dump_cgraph_node (struct dyn_cgraph_node *node,
>> + =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0uns=
igned m, unsigned f);
>> +static void
>> +gcov_dump_cgraph_node_dot (struct dyn_cgraph_node *node,
>> + =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 unsigned m, unsign=
ed f,
>> + =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 gcov_type cutoff_c=
ount);
>> +static void
>> +pointer_set_destroy (struct dyn_pointer_set *pset);
>> +static void **
>> +pointer_set_find_or_insert (struct dyn_pointer_set *pset, unsigned key);
>> +static struct dyn_pointer_set *
>> +pointer_set_create (unsigned (*get_key) (const void *));
>> +
>> +static struct dyn_cgraph the_dyn_call_graph;
>> +static int total_zero_count =3D 0;
>> +static int total_insane_count =3D 0;
>> +
>> +static void
>> +init_dyn_cgraph_node (struct dyn_cgraph_node *node, gcov_type guid)
>> +{
>> + =A0node->callees =3D 0;
>> + =A0node->callers =3D 0;
>> + =A0node->imported_modules =3D 0;
>> + =A0node->guid =3D guid;
>> + =A0node->visited =3D 0;
>> +}
>> +
>> +/* Return (module_id - 1). FUNC_GUID is the global unique id. =A0*/
>> +
>> +static inline gcov_unsigned_t
>> +get_module_idx_from_func_glob_uid (gcov_type func_guid)
>> +{
>> + =A0return EXTRACT_MODULE_ID_FROM_GLOBAL_ID (func_guid) - 1;
>> +}
>> +
>> +/* Return (module_id - 1) for MODULE_INFO. =A0*/
>> +
>> +static inline gcov_unsigned_t
>> +get_module_idx (const struct gcov_info *module_info)
>> +{
>> + =A0return module_info->mod_info->ident - 1;
>> +}
>> +
>> +/* Return intra-module function id given function global unique id
>> + =A0 FUNC_GUID. =A0*/
>> +
>> +static inline gcov_unsigned_t
>> +get_intra_module_func_id (gcov_type func_guid)
>> +{
>> + =A0return EXTRACT_FUNC_ID_FROM_GLOBAL_ID (func_guid);
>> +}
>> +
>> +/* Return the pointer to the dynamic call graph node for FUNC_GUID. =A0=
*/
>> +
>> +static inline struct dyn_cgraph_node *
>> +get_cgraph_node (gcov_type func_guid)
>> +{
>> + =A0gcov_unsigned_t mod_id, func_id;
>> +
>> + =A0mod_id =3D get_module_idx_from_func_glob_uid (func_guid);
>> +
>> + =A0/* This is to workaround: calls in __static_initialization_and_dest=
ruction
>> + =A0 =A0 should not be instrumented as the module id context for the ca=
llees have
>> + =A0 =A0 not setup yet -- this leads to mod_id =3D=3D (unsigned) (0 - 1=
). Multithreaded
>> + =A0 =A0 programs may also produce insane func_guid in the profile coun=
ter. =A0*/
>> + =A0if (mod_id >=3D the_dyn_call_graph.num_modules)
>> + =A0 =A0return 0;
>> +
>> + =A0func_id =3D get_intra_module_func_id (func_guid);
>> + =A0if (func_id > the_dyn_call_graph.sup_modules[mod_id].max_func_ident)
>> + =A0 =A0return 0;
>> +
>> + =A0return *(pointer_set_find_or_insert
>> + =A0 =A0 =A0 =A0 =A0(the_dyn_call_graph.call_graph_nodes[mod_id], func_=
id));
>> +}
>> +
>> +/* Return the gcov_info pointer for module with id MODULE_ID. =A0*/
>> +
>> +static inline struct gcov_info *
>> +get_module_info (gcov_unsigned_t module_id)
>> +{
>> + =A0return the_dyn_call_graph.modules[module_id];
>> +}
>> +
>> +struct gcov_info *__gcov_list ATTRIBUTE_HIDDEN;
>> +
>> +static inline unsigned
>> +cgraph_node_get_key (const void *p)
>> +{
>> + =A0return get_intra_module_func_id (((const struct dyn_cgraph_node *) =
p)->guid);
>> +}
>> +
>> +/* Initialize dynamic call graph. =A0*/
>> +
>> +static void
>> +init_dyn_call_graph (void)
>> +{
>> + =A0unsigned num_modules =3D 0;
>> + =A0struct gcov_info *gi_ptr;
>> +
>> + =A0the_dyn_call_graph.call_graph_nodes =3D 0;
>> + =A0the_dyn_call_graph.modules =3D 0;
>> + =A0the_dyn_call_graph.functions =3D 0;
>> + =A0the_dyn_call_graph.num_nodes_executed =3D 0;
>> +
>> + =A0gi_ptr =3D __gcov_list;
>> +
>> + =A0for (; gi_ptr; gi_ptr =3D gi_ptr->next)
>> + =A0 =A0num_modules++;
>> +
>> + =A0the_dyn_call_graph.num_modules =3D num_modules;
>> +
>> + =A0the_dyn_call_graph.modules
>> + =A0 =A0=3D XNEWVEC (struct gcov_info *, num_modules);
>> +
>> + =A0the_dyn_call_graph.sup_modules
>> + =A0 =A0=3D XNEWVEC (struct dyn_module_info, num_modules);
>> + =A0memset (the_dyn_call_graph.sup_modules, 0,
>> + =A0 =A0 =A0 =A0 =A0num_modules * sizeof (struct dyn_module_info));
>> +
>> + =A0the_dyn_call_graph.functions
>> + =A0 =A0=3D XNEWVEC (const struct gcov_fn_info **, num_modules);
>> +
>> + =A0the_dyn_call_graph.call_graph_nodes
>> + =A0 =A0=3D XNEWVEC (struct dyn_pointer_set *, num_modules);
>> +
>> + =A0gi_ptr =3D __gcov_list;
>> +
>> + =A0for (; gi_ptr; gi_ptr =3D gi_ptr->next)
>> + =A0 =A0{
>> + =A0 =A0 =A0unsigned c_ix =3D 0, t_ix, j, mod_id, fi_stride, max_func_i=
dent =3D 0;
>> + =A0 =A0 =A0struct dyn_cgraph_node *node;
>> +
>> + =A0 =A0 =A0mod_id =3D get_module_idx (gi_ptr);
>> +
>> + =A0 =A0 =A0the_dyn_call_graph.modules[mod_id] =3D gi_ptr;
>> +
>> + =A0 =A0 =A0the_dyn_call_graph.functions[mod_id]
>> + =A0 =A0 =A0 =A0 =A0=3D XNEWVEC (const struct gcov_fn_info *, gi_ptr->n=
_functions);
>> +
>> + =A0 =A0 =A0for (t_ix =3D 0; t_ix < GCOV_COUNTERS; t_ix++)
>> + =A0 =A0 =A0 if ((1 << t_ix) & gi_ptr->ctr_mask)
>> + =A0 =A0 =A0 =A0 =A0 c_ix++;
>> +
>> + =A0 =A0 =A0fi_stride =3D offsetof (struct gcov_fn_info, n_ctrs) + c_ix=
 * sizeof (unsigned);
>> + =A0 =A0 =A0if (__alignof__ (struct gcov_fn_info) > sizeof (unsigned))
>> + =A0 =A0 =A0 {
>> + =A0 =A0 =A0 =A0 fi_stride +=3D __alignof__ (struct gcov_fn_info) - 1;
>> + =A0 =A0 =A0 =A0 fi_stride &=3D ~(__alignof__ (struct gcov_fn_info) - 1=
);
>> + =A0 =A0 =A0 }
>> +
>> + =A0 =A0 =A0for (j =3D 0; j < gi_ptr->n_functions; j++)
>> + =A0 =A0 =A0 {
>> + =A0 =A0 =A0 =A0 =A0const struct gcov_fn_info *fi_ptr =3D (const struct=
 gcov_fn_info *)
>> + =A0 =A0 =A0 =A0 =A0 ((const char *) gi_ptr->functions + j * fi_stride);
>> + =A0 =A0 =A0 =A0 =A0the_dyn_call_graph.functions[mod_id][j] =3D fi_ptr;
>> + =A0 =A0 =A0 =A0 =A0if (fi_ptr->ident > max_func_ident)
>> + =A0 =A0 =A0 =A0 =A0 =A0max_func_ident =3D fi_ptr->ident;
>> + =A0 =A0 =A0 =A0}
>> +
>> +
>> + =A0 =A0 =A0the_dyn_call_graph.call_graph_nodes[mod_id]
>> + =A0 =A0 =A0 =3D pointer_set_create (cgraph_node_get_key);
>> +
>> + =A0 =A0 =A0the_dyn_call_graph.sup_modules[mod_id].max_func_ident =3D m=
ax_func_ident;
>> +
>> + =A0 =A0 =A0for (j =3D 0; j < gi_ptr->n_functions; j++)
>> + =A0 =A0 =A0 {
>> + =A0 =A0 =A0 =A0 =A0const struct gcov_fn_info *fi_ptr
>> + =A0 =A0 =A0 =A0 =A0 =A0 =A0=3D the_dyn_call_graph.functions[mod_id][j];
>> + =A0 =A0 =A0 =A0 *(pointer_set_find_or_insert
>> + =A0 =A0 =A0 =A0 =A0 (the_dyn_call_graph.call_graph_nodes[mod_id], fi_p=
tr->ident))
>> + =A0 =A0 =A0 =A0 =A0 =3D node =3D XNEW (struct dyn_cgraph_node);
>> + =A0 =A0 =A0 =A0 the_dyn_call_graph.call_graph_nodes[mod_id]->n_element=
s++;
>> + =A0 =A0 =A0 =A0 init_dyn_cgraph_node (node, GEN_FUNC_GLOBAL_ID (gi_ptr=
->mod_info->ident,
>> + =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =
=A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 fi_ptr->ident));
>> + =A0 =A0 =A0 }
>> + =A0 =A0}
>> +}
>> +
>> +/* Free up memory allocated for dynamic call graph. =A0*/
>> +
>> +void
>> +__gcov_finalize_dyn_callgraph (void)
>> +{
>> + =A0unsigned i;
>> + =A0struct gcov_info *gi_ptr;
>> +
>> + =A0for (i =3D 0; i < the_dyn_call_graph.num_modules; i++)
>> + =A0 =A0{
>> + =A0 =A0 =A0gi_ptr =3D the_dyn_call_graph.modules[i];
>> + =A0 =A0 =A0const struct gcov_fn_info *fi_ptr;
>> + =A0 =A0 =A0unsigned f_ix;
>> + =A0 =A0 =A0for (f_ix =3D 0; f_ix < gi_ptr->n_functions; f_ix++)
>> + =A0 =A0 =A0 {
>> + =A0 =A0 =A0 =A0 struct dyn_cgraph_node *node;
>> + =A0 =A0 =A0 =A0 =A0struct dyn_cgraph_edge *callees, *next_callee;
>> + =A0 =A0 =A0 =A0 =A0fi_ptr =3D the_dyn_call_graph.functions[i][f_ix];
>> + =A0 =A0 =A0 =A0 =A0node =3D *(pointer_set_find_or_insert
>> + =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0(the_dyn_call_graph.call_graph_node=
s[i], fi_ptr->ident));
>> + =A0 =A0 =A0 =A0 gcc_assert (node);
>> + =A0 =A0 =A0 =A0 =A0callees =3D node->callees;
>> +
>> + =A0 =A0 =A0 =A0 =A0if (!callees)
>> + =A0 =A0 =A0 =A0 =A0 =A0continue;
>> + =A0 =A0 =A0 =A0 =A0while (callees !=3D 0)
>> + =A0 =A0 =A0 =A0 =A0 =A0{
>> + =A0 =A0 =A0 =A0 =A0 =A0 =A0next_callee =3D callees->next_callee;
>> + =A0 =A0 =A0 =A0 =A0 =A0 =A0XDELETE (callees);
>> + =A0 =A0 =A0 =A0 =A0 =A0 =A0callees =3D next_callee;
>> + =A0 =A0 =A0 =A0 =A0 =A0}
>> + =A0 =A0 =A0 =A0 if (node->imported_modules)
>> + =A0 =A0 =A0 =A0 =A0 pointer_set_destroy (node->imported_modules);
>> + =A0 =A0 =A0 =A0}
>> + =A0 =A0 =A0if (the_dyn_call_graph.call_graph_nodes[i])
>> + =A0 =A0 =A0 =A0pointer_set_destroy (the_dyn_call_graph.call_graph_node=
s[i]);
>> + =A0 =A0 =A0if (the_dyn_call_graph.functions[i])
>> + =A0 =A0 =A0 =A0XDELETEVEC (the_dyn_call_graph.functions[i]);
>> + =A0 =A0 =A0/* Now delete sup modules */
>> + =A0 =A0 =A0if (the_dyn_call_graph.sup_modules[i].imported_modules)
>> + =A0 =A0 =A0 =A0pointer_set_destroy (the_dyn_call_graph.sup_modules[i].=
imported_modules);
>> + =A0 =A0}
>> + =A0XDELETEVEC (the_dyn_call_graph.call_graph_nodes);
>> + =A0XDELETEVEC (the_dyn_call_graph.functions);
>> + =A0XDELETEVEC (the_dyn_call_graph.sup_modules);
>> + =A0XDELETEVEC (the_dyn_call_graph.modules);
>> +}
>> +
>> +/* Add outgoing edge OUT_EDGE for caller node CALLER. =A0*/
>> +
>> +static void
>> +gcov_add_out_edge (struct dyn_cgraph_node *caller,
>> + =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0struct dyn_cgraph_edge *out_edge)
>> +{
>> + =A0if (!caller->callees)
>> + =A0 =A0caller->callees =3D out_edge;
>> + =A0else
>> + =A0 =A0{
>> + =A0 =A0 =A0out_edge->next_callee =3D caller->callees;
>> + =A0 =A0 =A0caller->callees =3D out_edge;
>> + =A0 =A0}
>> +}
>> +
>> +/* Add incoming edge IN_EDGE for callee node CALLEE. =A0*/
>> +
>> +static void
>> +gcov_add_in_edge (struct dyn_cgraph_node *callee,
>> + =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 struct dyn_cgraph_edge *in_edge)
>> +{
>> + =A0if (!callee->callers)
>> + =A0 =A0callee->callers =3D in_edge;
>> + =A0else
>> + =A0 =A0{
>> + =A0 =A0 =A0in_edge->next_caller =3D callee->callers;
>> + =A0 =A0 =A0callee->callers =3D in_edge;
>> + =A0 =A0}
>> +}
>> +
>> +/* Add a call graph edge between caller CALLER and callee CALLEE.
>> + =A0 The edge count is COUNT. =A0*/
>> +
>> +static void
>> +gcov_add_cgraph_edge (struct dyn_cgraph_node *caller,
>> + =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 struct dyn_cgraph_node *callee,
>> + =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 gcov_type count)
>> +{
>> + =A0struct dyn_cgraph_edge *new_edge =3D XNEW (struct dyn_cgraph_edge);
>> + =A0new_edge->caller =3D caller;
>> + =A0new_edge->callee =3D callee;
>> + =A0new_edge->count =3D count;
>> + =A0new_edge->next_caller =3D 0;
>> + =A0new_edge->next_callee =3D 0;
>> +
>> + =A0gcov_add_out_edge (caller, new_edge);
>> + =A0gcov_add_in_edge (callee, new_edge);
>> +}
>> +
>> +/* Add call graph edges from direct calls for caller CALLER. DIR_CALL_C=
OUNTERS
>> + =A0 is the array of call counters. N_COUNTS is the number of counters.=
 =A0*/
>> +
>> +static void
>> +gcov_build_callgraph_dc_fn (struct dyn_cgraph_node *caller,
>> + =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0gcov_type *dir_=
call_counters,
>> + =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0unsigned n_coun=
ts)
>> +{
>> + =A0unsigned i;
>> +
>> + =A0for (i =3D 0; i < n_counts; i +=3D 2)
>> + =A0 =A0{
>> + =A0 =A0 =A0struct dyn_cgraph_node *callee;
>> + =A0 =A0 =A0gcov_type count;
>> + =A0 =A0 =A0gcov_type callee_guid =3D dir_call_counters[i];
>> +
>> + =A0 =A0 =A0count =3D dir_call_counters[i + 1];
>> + =A0 =A0 =A0if (count =3D=3D 0)
>> + =A0 =A0 =A0 =A0{
>> + =A0 =A0 =A0 =A0 =A0total_zero_count++;
>> + =A0 =A0 =A0 =A0 =A0continue;
>> + =A0 =A0 =A0 =A0}
>> + =A0 =A0 =A0callee =3D get_cgraph_node (callee_guid);
>> + =A0 =A0 =A0if (!callee)
>> + =A0 =A0 =A0 =A0{
>> + =A0 =A0 =A0 =A0 =A0total_insane_count++;
>> + =A0 =A0 =A0 =A0 =A0continue;
>> + =A0 =A0 =A0 =A0}
>> + =A0 =A0 =A0gcov_add_cgraph_edge (caller, callee, count);
>> + =A0 =A0}
>> +}
>> +
>> +/* Add call graph edges from indirect calls for caller CALLER. ICALL_CO=
UNTERS
>> + =A0 is the array of icall counters. N_COUNTS is the number of counters=
. =A0*/
>> +
>> +static void
>> +gcov_build_callgraph_ic_fn (struct dyn_cgraph_node *caller,
>> + =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0gcov_type *ical=
l_counters,
>> + =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0unsigned n_coun=
ts)
>> +{
>> + =A0unsigned i, j;
>> +
>> + =A0for (i =3D 0; i < n_counts; i +=3D GCOV_ICALL_TOPN_NCOUNTS)
>> + =A0 =A0{
>> + =A0 =A0 =A0gcov_type *value_array =3D &icall_counters[i + 1];
>> + =A0 =A0 =A0for (j =3D 0; j < GCOV_ICALL_TOPN_NCOUNTS - 1; j +=3D 2)
>> + =A0 =A0 =A0 =A0{
>> + =A0 =A0 =A0 =A0 =A0struct dyn_cgraph_node *callee;
>> + =A0 =A0 =A0 =A0 =A0gcov_type count;
>> + =A0 =A0 =A0 =A0 =A0gcov_type callee_guid =3D value_array[j];
>> +
>> + =A0 =A0 =A0 =A0 =A0count =3D value_array[j + 1];
>> + =A0 =A0 =A0 =A0 /* Do not update zero count edge count
>> + =A0 =A0 =A0 =A0 =A0* as it means there is no target in this entry. =A0=
*/
>> + =A0 =A0 =A0 =A0 =A0if (count =3D=3D 0)
>> + =A0 =A0 =A0 =A0 =A0 =A0continue;
>> + =A0 =A0 =A0 =A0 =A0callee =3D get_cgraph_node (callee_guid);
>> + =A0 =A0 =A0 =A0 =A0if (!callee)
>> + =A0 =A0 =A0 =A0 =A0 {
>> + =A0 =A0 =A0 =A0 =A0 =A0 =A0total_insane_count++;
>> + =A0 =A0 =A0 =A0 =A0 =A0 =A0continue;
>> + =A0 =A0 =A0 =A0 =A0 }
>> + =A0 =A0 =A0 =A0 =A0gcov_add_cgraph_edge (caller, callee, count);
>> + =A0 =A0 =A0 =A0}
>> + =A0 =A0}
>> +}
>> +
>> +/* Build the dynamic call graph. =A0*/
>> +
>> +static void
>> +gcov_build_callgraph (void)
>> +{
>> + =A0struct gcov_info *gi_ptr;
>> + =A0unsigned t_ix, m_ix;
>> +
>> + =A0init_dyn_call_graph ();
>> +
>> + =A0for (m_ix =3D 0; m_ix < the_dyn_call_graph.num_modules; m_ix++)
>> + =A0 =A0{
>> + =A0 =A0 =A0const struct gcov_fn_info *fi_ptr;
>> + =A0 =A0 =A0unsigned c_ix, f_ix, n_counts, dp_cix =3D 0, ip_cix =3D 0;
>> + =A0 =A0 =A0gcov_type *dcall_profile_values, *icall_profile_values;
>> + =A0 =A0 =A0gcov_type *arcs_values =3D 0; unsigned arcs_cix;
>> +
>> + =A0 =A0 =A0gi_ptr =3D the_dyn_call_graph.modules[m_ix];
>> +
>> + =A0 =A0 =A0dcall_profile_values =3D 0;
>> + =A0 =A0 =A0icall_profile_values =3D 0;
>> + =A0 =A0 =A0c_ix =3D 0;
>> + =A0 =A0 =A0for (t_ix =3D 0; t_ix < GCOV_COUNTERS; t_ix++)
>> + =A0 =A0 =A0 if ((1 << t_ix) & gi_ptr->ctr_mask)
>> + =A0 =A0 =A0 =A0 {
>> + =A0 =A0 =A0 =A0 =A0 if (t_ix =3D=3D GCOV_COUNTER_DIRECT_CALL)
>> + =A0 =A0 =A0 =A0 =A0 =A0 {
>> + =A0 =A0 =A0 =A0 =A0 =A0 =A0 dcall_profile_values =3D gi_ptr->counts[c_=
ix].values;
>> + =A0 =A0 =A0 =A0 =A0 =A0 =A0 dp_cix =3D c_ix;
>> + =A0 =A0 =A0 =A0 =A0 =A0 }
>> + =A0 =A0 =A0 =A0 =A0 if (t_ix =3D=3D GCOV_COUNTER_ICALL_TOPNV)
>> + =A0 =A0 =A0 =A0 =A0 =A0 {
>> + =A0 =A0 =A0 =A0 =A0 =A0 =A0 icall_profile_values =3D gi_ptr->counts[c_=
ix].values;
>> + =A0 =A0 =A0 =A0 =A0 =A0 =A0 ip_cix =3D c_ix;
>> + =A0 =A0 =A0 =A0 =A0 =A0 }
>> + =A0 =A0 =A0 =A0 =A0 if (t_ix =3D=3D GCOV_COUNTER_ARCS)
>> + =A0 =A0 =A0 =A0 =A0 =A0 {
>> + =A0 =A0 =A0 =A0 =A0 =A0 =A0 arcs_values =3D gi_ptr->counts[c_ix].value=
s;
>> + =A0 =A0 =A0 =A0 =A0 =A0 =A0 arcs_cix =3D c_ix;
>> + =A0 =A0 =A0 =A0 =A0 =A0 =A0}
>> + =A0 =A0 =A0 =A0 =A0 c_ix++;
>> + =A0 =A0 =A0 =A0 }
>> +
>> + =A0 =A0 =A0if (dcall_profile_values =3D=3D 0 && icall_profile_values =
=3D=3D 0)
>> + =A0 =A0 =A0 =A0continue;
>> +
>> + =A0 =A0 =A0for (f_ix =3D 0; f_ix < gi_ptr->n_functions; f_ix++)
>> + =A0 =A0 =A0 =A0{
>> + =A0 =A0 =A0 =A0 =A0struct dyn_cgraph_node *caller;
>> + =A0 =A0 =A0 =A0 =A0fi_ptr =3D the_dyn_call_graph.functions[m_ix][f_ix];
>> + =A0 =A0 =A0 =A0 =A0caller =3D *(pointer_set_find_or_insert
>> + =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0(the_dyn_call_graph.call_graph_=
nodes[m_ix],
>> + =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 fi_ptr->ident));
>> + =A0 =A0 =A0 =A0 gcc_assert (caller);
>> + =A0 =A0 =A0 =A0 =A0if (dcall_profile_values)
>> + =A0 =A0 =A0 =A0 =A0 =A0{
>> + =A0 =A0 =A0 =A0 =A0 =A0 =A0unsigned offset;
>> + =A0 =A0 =A0 =A0 =A0 =A0 =A0n_counts =3D fi_ptr->n_ctrs[dp_cix];
>> + =A0 =A0 =A0 =A0 =A0 =A0 =A0offset =3D fi_ptr->dc_offset;
>> + =A0 =A0 =A0 =A0 =A0 =A0 =A0gcov_build_callgraph_dc_fn (caller,
>> + =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =
=A0 =A0 =A0 =A0dcall_profile_values + offset,
>> + =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =
=A0 =A0 =A0 =A0n_counts);
>> + =A0 =A0 =A0 =A0 =A0 =A0}
>> + =A0 =A0 =A0 =A0 =A0if (icall_profile_values)
>> + =A0 =A0 =A0 =A0 =A0 =A0{
>> + =A0 =A0 =A0 =A0 =A0 =A0 =A0n_counts =3D fi_ptr->n_ctrs[ip_cix];
>> + =A0 =A0 =A0 =A0 =A0 =A0 =A0gcov_build_callgraph_ic_fn (caller, icall_p=
rofile_values, n_counts);
>> + =A0 =A0 =A0 =A0 =A0 =A0 =A0icall_profile_values +=3D n_counts;
>> + =A0 =A0 =A0 =A0 =A0 =A0}
>> + =A0 =A0 =A0 =A0 =A0if (arcs_values && 0)
>> + =A0 =A0 =A0 =A0 =A0 =A0{
>> + =A0 =A0 =A0 =A0 =A0 =A0 =A0gcov_type total_arc_count =3D 0;
>> + =A0 =A0 =A0 =A0 =A0 =A0 =A0unsigned arc;
>> + =A0 =A0 =A0 =A0 =A0 =A0 =A0n_counts =3D fi_ptr->n_ctrs[arcs_cix];
>> + =A0 =A0 =A0 =A0 =A0 =A0 =A0for (arc =3D 0; arc < n_counts; arc++)
>> + =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0total_arc_count +=3D arcs_values[arc];
>> + =A0 =A0 =A0 =A0 =A0 =A0 =A0if (total_arc_count !=3D 0)
>> + =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0the_dyn_call_graph.num_nodes_executed++;
>> + =A0 =A0 =A0 =A0 =A0 =A0 =A0arcs_values +=3D n_counts;
>> + =A0 =A0 =A0 =A0 =A0 =A0}
>> + =A0 =A0 =A0 =A0}
>> + =A0 =A0}
>> +
>> +}
>> +
>> +static inline size_t
>> +hash1 (unsigned p, unsigned long max, unsigned long logmax)
>> +{
>> + =A0const unsigned long long A =3D 0x9e3779b97f4a7c16ull;
>> + =A0const unsigned long long shift =3D 64 - logmax;
>> +
>> + =A0return ((A * (unsigned long) p) >> shift) & (max - 1);
>> +}
>> +
>> +/* Allocate an empty imported-modules set. =A0*/
>> +
>> +static struct dyn_pointer_set *
>> +pointer_set_create (unsigned (*get_key) (const void *))
>> +{
>> + =A0struct dyn_pointer_set *result =3D XNEW (struct dyn_pointer_set);
>> +
>> + =A0result->n_elements =3D 0;
>> + =A0result->log_slots =3D 8;
>> + =A0result->n_slots =3D (size_t) 1 << result->log_slots;
>> +
>> + =A0result->slots =3D XNEWVEC (void *, result->n_slots);
>> + =A0memset (result->slots, 0, sizeof (void *) * result->n_slots);
>> + =A0result->get_key =3D get_key;
>> +
>> + =A0return result;
>> +}
>> +
>> +/* Reclaim all memory associated with PSET. =A0*/
>> +
>> +static void
>> +pointer_set_destroy (struct dyn_pointer_set *pset)
>> +{
>> + =A0size_t i;
>> + =A0for (i =3D 0; i < pset->n_slots; i++)
>> + =A0 =A0if (pset->slots[i])
>> + =A0 =A0 =A0XDELETE (pset->slots[i]);
>> + =A0XDELETEVEC (pset->slots);
>> + =A0XDELETE (pset);
>> +}
>> +
>> +/* Subroutine of pointer_set_find_or_insert. =A0Return the insertion sl=
ot for KEY
>> + =A0 into an empty element of SLOTS, an array of length N_SLOTS. =A0*/
>> +static inline size_t
>> +insert_aux (unsigned key, void **slots,
>> + =A0 =A0 =A0 =A0 =A0 size_t n_slots, size_t log_slots,
>> + =A0 =A0 =A0 =A0 =A0 unsigned (*get_key) (const void *))
>> +{
>> + =A0size_t n =3D hash1 (key, n_slots, log_slots);
>> + =A0while (1)
>> + =A0 =A0{
>> + =A0 =A0 =A0if (slots[n] =3D=3D 0 || get_key (slots[n]) =3D=3D key)
>> + =A0 =A0 =A0 return n;
>> + =A0 =A0 =A0else
>> + =A0 =A0 =A0 {
>> + =A0 =A0 =A0 =A0 ++n;
>> + =A0 =A0 =A0 =A0 if (n =3D=3D n_slots)
>> + =A0 =A0 =A0 =A0 =A0 n =3D 0;
>> + =A0 =A0 =A0 }
>> + =A0 =A0}
>> +}
>> +
>> +/* Find slot for KEY. KEY must be nonnull. =A0*/
>> +
>> +static void **
>> +pointer_set_find_or_insert (struct dyn_pointer_set *pset, unsigned key)
>> +{
>> + =A0size_t n;
>> +
>> + =A0/* For simplicity, expand the set even if KEY is already there. =A0=
This can be
>> + =A0 =A0 superfluous but can happen at most once. =A0*/
>> + =A0if (pset->n_elements > pset->n_slots / 4)
>> + =A0 =A0{
>> + =A0 =A0 =A0size_t new_log_slots =3D pset->log_slots + 1;
>> + =A0 =A0 =A0size_t new_n_slots =3D pset->n_slots * 2;
>> + =A0 =A0 =A0void **new_slots =3D XNEWVEC (void *, new_n_slots);
>> + =A0 =A0 =A0memset (new_slots, 0, sizeof (void *) * new_n_slots);
>> + =A0 =A0 =A0size_t i;
>> +
>> + =A0 =A0 =A0for (i =3D 0; i < pset->n_slots; ++i)
>> + =A0 =A0 =A0 =A0{
>> + =A0 =A0 =A0 =A0 void *value =3D pset->slots[i];
>> + =A0 =A0 =A0 =A0 if (!value)
>> + =A0 =A0 =A0 =A0 =A0 continue;
>> + =A0 =A0 =A0 =A0 n =3D insert_aux (pset->get_key (value), new_slots, ne=
w_n_slots,
>> + =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 new_log_slots, pset->g=
et_key);
>> + =A0 =A0 =A0 =A0 new_slots[n] =3D value;
>> + =A0 =A0 =A0 }
>> +
>> + =A0 =A0 =A0XDELETEVEC (pset->slots);
>> + =A0 =A0 =A0pset->n_slots =3D new_n_slots;
>> + =A0 =A0 =A0pset->log_slots =3D new_log_slots;
>> + =A0 =A0 =A0pset->slots =3D new_slots;
>> + =A0 =A0}
>> +
>> + =A0n =3D insert_aux (key, pset->slots, pset->n_slots, pset->log_slots,
>> + =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 pset->get_key);
>> + =A0return &pset->slots[n];
>> +}
>> +
>> +/* Pass each pointer in PSET to the function in FN, together with the f=
ixed
>> + =A0 parameters DATA1, DATA2, DATA3. =A0If FN returns false, the iterat=
ion stops. =A0*/
>> +
>> +static void
>> +pointer_set_traverse (const struct dyn_pointer_set *pset,
>> + =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0int (*fn) (const void *, vo=
id *, void *, void *),
>> + =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 void *data1, void *data2, void=
 *data3)
>> +{
>> + =A0size_t i;
>> + =A0for (i =3D 0; i < pset->n_slots; ++i)
>> + =A0 =A0if (pset->slots[i] && !fn (pset->slots[i], data1, data2, data3))
>> + =A0 =A0 =A0break;
>> +}
>> +
>> +static int
>> +imp_mod_set_insert (struct dyn_pointer_set *p, const struct gcov_info *=
imp_mod,
>> + =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 double wt)
>> +{
>> + =A0struct dyn_imp_mod **m =3D (struct dyn_imp_mod **)
>> + =A0 =A0pointer_set_find_or_insert (p, imp_mod->mod_info->ident);
>> + =A0if (*m)
>> + =A0 =A0{
>> + =A0 =A0 =A0(*m)->weight +=3D wt;
>> + =A0 =A0 =A0return 1;
>> + =A0 =A0}
>> + =A0else
>> + =A0 =A0{
>> + =A0 =A0 =A0*m =3D XNEW (struct dyn_imp_mod);
>> + =A0 =A0 =A0(*m)->imp_mod =3D imp_mod;
>> + =A0 =A0 =A0(*m)->weight =3D wt;
>> + =A0 =A0 =A0p->n_elements++;
>> + =A0 =A0 =A0return 0;
>> + =A0 =A0}
>> +}
>> +
>> +/* Callback function to propagate import module (VALUE) from callee to
>> + =A0 caller's imported-module-set (DATA1).
>> + =A0 The weight is scaled by the scaling-factor (DATA2) before propagat=
ion,
>> + =A0 and accumulated into DATA3. =A0*/
>> +static int
>> +gcov_propagate_imp_modules (const void *value, void *data1, void *data2,
>> + =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 void *data3)
>> +{
>> + =A0const struct dyn_imp_mod *m =3D (const struct dyn_imp_mod *) value;
>> + =A0struct dyn_pointer_set *receiving_set =3D (struct dyn_pointer_set *=
) data1;
>> + =A0double *scale =3D (double *) data2;
>> + =A0double *sum =3D (double *) data3;
>> + =A0double wt =3D m->weight;
>> + =A0if (scale)
>> + =A0 =A0wt *=3D *scale;
>> + =A0if (sum)
>> + =A0 =A0(*sum) +=3D wt;
>> + =A0imp_mod_set_insert (receiving_set, m->imp_mod, wt);
>> + =A0return 1;
>> +}
>> +
>> +static int
>> +sort_by_count (const void *pa, const void *pb)
>> +{
>> + =A0const struct dyn_cgraph_edge *edge_a =3D *(struct dyn_cgraph_edge *=
 const *)pa;
>> + =A0const struct dyn_cgraph_edge *edge_b =3D *(struct dyn_cgraph_edge *=
 const *)pb;
>> +
>> + =A0/* This can overvlow. =A0*/
>> + =A0/* return edge_b->count - edge_a->count; =A0*/
>> + =A0if (edge_b->count > edge_a->count)
>> + =A0 =A0return 1;
>> + =A0else if (edge_b->count =3D=3D edge_a->count)
>> + =A0 =A0return 0;
>> + =A0else
>> + =A0 =A0return -1;
>> +}
>> +
>> +/* Compute the hot callgraph edge threhold. =A0*/
>> +
>> +static gcov_type
>> +gcov_compute_cutoff_count (void)
>> +{
>> + =A0unsigned m_ix, capacity, i;
>> + =A0unsigned num_edges =3D 0;
>> + =A0gcov_type cutoff_count;
>> + =A0double total, cum, cum_cutoff;
>> + =A0struct dyn_cgraph_edge **edges;
>> + =A0struct gcov_info *gi_ptr;
>> + =A0char *cutoff_str;
>> + =A0char *num_perc_str;
>> + =A0unsigned cutoff_perc;
>> + =A0unsigned num_perc;
>> + =A0int do_dump;
>> +
>> + =A0capacity =3D 100;
>> + =A0/* allocate an edge array */
>> + =A0edges =3D XNEWVEC (struct dyn_cgraph_edge*, capacity);
>> + =A0/* First count the number of edges. =A0*/
>> + =A0for (m_ix =3D 0; m_ix < the_dyn_call_graph.num_modules; m_ix++)
>> + =A0 =A0{
>> + =A0 =A0 =A0const struct gcov_fn_info *fi_ptr;
>> + =A0 =A0 =A0unsigned f_ix;
>> +
>> + =A0 =A0 =A0gi_ptr =3D the_dyn_call_graph.modules[m_ix];
>> +
>> + =A0 =A0 =A0for (f_ix =3D 0; f_ix < gi_ptr->n_functions; f_ix++)
>> + =A0 =A0 =A0 {
>> + =A0 =A0 =A0 =A0 struct dyn_cgraph_node *node;
>> + =A0 =A0 =A0 =A0 =A0struct dyn_cgraph_edge *callees;
>> +
>> + =A0 =A0 =A0 =A0 fi_ptr =3D the_dyn_call_graph.functions[m_ix][f_ix];
>> +
>> + =A0 =A0 =A0 =A0 node =3D *(pointer_set_find_or_insert
>> + =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0(the_dyn_call_graph.call_graph_node=
s[m_ix], fi_ptr->ident));
>> + =A0 =A0 =A0 =A0 gcc_assert (node);
>> +
>> + =A0 =A0 =A0 =A0 =A0callees =3D node->callees;
>> + =A0 =A0 =A0 =A0 =A0while (callees !=3D 0)
>> + =A0 =A0 =A0 =A0 =A0 =A0{
>> + =A0 =A0 =A0 =A0 =A0 =A0 =A0num_edges++;
>> + =A0 =A0 =A0 =A0 =A0 =A0 =A0if (num_edges < capacity)
>> + =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0edges[num_edges - 1] =3D callees;
>> + =A0 =A0 =A0 =A0 =A0 =A0 =A0else
>> + =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0{
>> + =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0capacity =3D capacity + (capacity >=
> 1);
>> + =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0edges =3D (struct dyn_cgraph_edge *=
*)realloc (edges, sizeof (void*) * capacity);
>> + =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0edges[num_edges - 1] =3D callees;
>> + =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0}
>> + =A0 =A0 =A0 =A0 =A0 =A0 =A0callees =3D callees->next_callee;
>> + =A0 =A0 =A0 =A0 =A0 =A0}
>> + =A0 =A0 =A0 }
>> + =A0 =A0}
>> +
>> + =A0/* Now sort */
>> + qsort (edges, num_edges, sizeof (void *), sort_by_count);
>> +#define CUM_CUTOFF_PERCENT 80
>> +#define MIN_NUM_EDGE_PERCENT 0
>> + =A0cutoff_str =3D getenv ("GCOV_DYN_CGRAPH_CUTOFF");
>> + =A0if (cutoff_str && strlen (cutoff_str))
>> + =A0 =A0{
>> + =A0 =A0 =A0if ((num_perc_str =3D strchr (cutoff_str, ':')))
>> + =A0 =A0 =A0 =A0{
>> + =A0 =A0 =A0 =A0 =A0*num_perc_str =3D '\0';
>> + =A0 =A0 =A0 =A0 =A0num_perc_str++;
>> + =A0 =A0 =A0 =A0}
>> + =A0 =A0 =A0cutoff_perc =3D atoi (cutoff_str);
>> + =A0 =A0 =A0if (num_perc_str)
>> + =A0 =A0 =A0 =A0num_perc =3D atoi (num_perc_str);
>> + =A0 =A0 =A0else
>> + =A0 =A0 =A0 =A0num_perc =3D MIN_NUM_EDGE_PERCENT;
>> + =A0 =A0}
>> + =A0else
>> + =A0 =A0{
>> + =A0 =A0 =A0cutoff_perc =3D CUM_CUTOFF_PERCENT;
>> + =A0 =A0 =A0num_perc =3D MIN_NUM_EDGE_PERCENT;
>> + =A0 =A0}
>> +
>> + =A0total =3D 0;
>> + =A0cum =3D 0;
>> + =A0for (i =3D 0; i < num_edges; i++)
>> + =A0 =A0total +=3D edges[i]->count;
>> +
>> + =A0cum_cutoff =3D (total * cutoff_perc)/100;
>> + =A0do_dump =3D (getenv ("GCOV_DYN_CGRAPH_DUMP") !=3D 0);
>> + =A0for (i =3D 0; i < num_edges; i++)
>> + =A0 =A0{
>> + =A0 =A0 =A0cum +=3D edges[i]->count;
>> + =A0 =A0 =A0if (do_dump)
>> + =A0 =A0 =A0 =A0fprintf (stderr, "// edge[%d] count =3D %.0f [%llx --> =
%llx]\n",
>> + =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 i, (double) edges[i]->count,
>> + =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 (long long) edges[i]->caller->guid,
>> + =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 (long long) edges[i]->callee->guid);
>> + =A0 =A0 =A0if (cum >=3D cum_cutoff && (i * 100 >=3D num_edges * num_pe=
rc))
>> + =A0 =A0 =A0 =A0{
>> + =A0 =A0 =A0 =A0 =A0cutoff_count =3D edges[i]->count;
>> + =A0 =A0 =A0 =A0 =A0break;
>> + =A0 =A0 =A0 =A0}
>> + =A0 =A0}
>> +
>> + =A0if (do_dump)
>> + =A0 =A0fprintf (stderr, "// total =3D %.0f cum =3D %.0f cum/total =3D =
%.0f%%"
>> + =A0 =A0 =A0 =A0 =A0 =A0 " cutoff_count =3D %lld [total edges: %d hot e=
dges: %d perc: %d%%]\n"
>> + =A0 =A0 =A0 =A0 =A0 =A0" total_zero_count_edges =3D %d total_insane_co=
unt_edgess =3D %d\n"
>> + =A0 =A0 =A0 =A0 =A0 =A0 " total_nodes_executed =3D %d\n",
>> + =A0 =A0 =A0 =A0 =A0 =A0 total, cum, (cum * 100)/total, (long long) cut=
off_count,
>> + =A0 =A0 =A0 =A0 =A0 =A0 num_edges, i, (i * 100)/num_edges, total_zero_=
count,
>> + =A0 =A0 =A0 =A0 =A0 =A0 total_insane_count, the_dyn_call_graph.num_nod=
es_executed);
>> +
>> + =A0XDELETEVEC (edges);
>> + =A0return cutoff_count;
>> +}
>> +
>> +static inline unsigned
>> +imp_mod_get_key (const void *p)
>> +{
>> + =A0return ((const struct dyn_imp_mod *) p)->imp_mod->mod_info->ident;
>> +}
>> +
>> +/* Return the imported module set for NODE. =A0*/
>> +
>> +static struct dyn_pointer_set *
>> +gcov_get_imp_module_set (struct dyn_cgraph_node *node)
>> +{
>> + =A0if (!node->imported_modules)
>> + =A0 =A0node->imported_modules =3D pointer_set_create (imp_mod_get_key);
>> +
>> + =A0return node->imported_modules;
>> +}
>> +
>> +/* Return the imported module set for MODULE MI. =A0*/
>> +
>> +static struct dyn_pointer_set *
>> +gcov_get_module_imp_module_set (struct dyn_module_info *mi)
>> +{
>> + =A0if (!mi->imported_modules)
>> + =A0 =A0mi->imported_modules =3D pointer_set_create (imp_mod_get_key);
>> +
>> + =A0return mi->imported_modules;
>> +}
>> +
>> +/* Callback function to mark if a module needs to be exported. =A0*/
>> +
>> +static int
>> +gcov_mark_export_modules (const void *value,
>> + =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 void *data1 ATTRIBUTE_=
UNUSED,
>> + =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 void *data2 ATTRIBUTE_=
UNUSED,
>> + =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 void *data3 ATTRIBUTE_=
UNUSED)
>> +{
>> + =A0const struct gcov_info *module_info
>> + =A0 =A0=3D ((const struct dyn_imp_mod *) value)->imp_mod;
>> +
>> + =A0module_info->mod_info->is_exported =3D 1;
>> + =A0return 1;
>> +}
>> +
>> +struct gcov_import_mod_array
>> +{
>> + =A0const struct dyn_imp_mod **imported_modules;
>> + =A0struct gcov_info *importing_module;
>> + =A0unsigned len;
>> +};
>> +
>> +/* Callback function to compute pointer set size. =A0*/
>> +
>> +static int
>> +gcov_compute_mset_size (const void *value ATTRIBUTE_UNUSED,
>> + =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0void *data1,
>> + =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 void *data2 ATTRIBUTE_UNUS=
ED,
>> + =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 void *data3 ATTRIBUTE_UNUS=
ED)
>> +{
>> + =A0unsigned *len =3D (unsigned *) data1;
>> + =A0(*len)++;
>> + =A0return 1;
>> +}
>> +
>> +/* Callback function to collect imported modules. =A0*/
>> +
>> +static int
>> +gcov_collect_imported_modules (const void *value,
>> + =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0void *data1,
>> + =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0void *data2=
 ATTRIBUTE_UNUSED,
>> + =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0void *data3=
 ATTRIBUTE_UNUSED)
>> +{
>> + =A0struct gcov_import_mod_array *out_array;
>> + =A0const struct dyn_imp_mod *m
>> + =A0 =A0=3D (const struct dyn_imp_mod *) value;
>> +
>> + =A0out_array =3D (struct gcov_import_mod_array *) data1;
>> +
>> + =A0if (m->imp_mod !=3D out_array->importing_module)
>> + =A0 =A0out_array->imported_modules[out_array->len++] =3D m;
>> +
>> + =A0return 1;
>> +}
>> +
>> +/* Comparator for sorting imported modules using weights. =A0*/
>> +
>> +static int
>> +sort_by_module_wt (const void *pa, const void *pb)
>> +{
>> + =A0const struct dyn_imp_mod *m_a =3D *((const struct dyn_imp_mod * con=
st *) pa);
>> + =A0const struct dyn_imp_mod *m_b =3D *((const struct dyn_imp_mod * con=
st *) pb);
>> +
>> + =A0/* We want to sort in descending order of weights. =A0*/
>> + =A0if (m_a->weight < m_b->weight)
>> + =A0 =A0return +1;
>> + =A0if (m_a->weight > m_b->weight)
>> + =A0 =A0return -1;
>> + =A0return get_module_idx (m_a->imp_mod) - get_module_idx (m_b->imp_mod=
);
>> +}
>> +
>> +/* Return a dynamic array of imported modules that is sorted for
>> + =A0 the importing module MOD_INFO. The length of the array is returned
>> + =A0 in *LEN. =A0*/
>> +
>> +const struct dyn_imp_mod **
>> +gcov_get_sorted_import_module_array (struct gcov_info *mod_info,
>> + =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =
=A0 unsigned *len)
>> +{
>> + =A0unsigned mod_id;
>> + =A0struct dyn_module_info *sup_mod_info;
>> + =A0unsigned array_len =3D 0;
>> + =A0struct gcov_import_mod_array imp_array;
>> +
>> + =A0mod_id =3D get_module_idx (mod_info);
>> + =A0sup_mod_info =3D &the_dyn_call_graph.sup_modules[mod_id];
>> +
>> + =A0if (sup_mod_info->imported_modules =3D=3D 0)
>> + =A0 =A0return 0;
>> +
>> + =A0pointer_set_traverse (sup_mod_info->imported_modules,
>> + =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0gcov_compute_mset_size,=
 &array_len, 0, 0);
>> + =A0imp_array.imported_modules =3D XNEWVEC (const struct dyn_imp_mod *,=
 array_len);
>> + =A0imp_array.len =3D 0;
>> + =A0imp_array.importing_module =3D mod_info;
>> + =A0pointer_set_traverse (sup_mod_info->imported_modules,
>> + =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0gcov_collect_imported_m=
odules, &imp_array, 0, 0);
>> + =A0*len =3D imp_array.len;
>> + =A0qsort (imp_array.imported_modules, imp_array.len,
>> + =A0 =A0 =A0 =A0 sizeof (void *), sort_by_module_wt);
>> + =A0return imp_array.imported_modules;
>> +}
>> +
>> +/* Compute modules that are needed for NODE (for cross module inlining).
>> + =A0 CUTTOFF_COUNT is the call graph edge count cutoff value.
>> + =A0 IMPORT_SCALE is the scaling-factor (percent) by which to scale the
>> + =A0 weights of imported modules of a callee before propagating them to
>> + =A0 the caller, if the callee and caller are in different modules.
>> +
>> + =A0 Each imported module is assigned a weight that corresponds to the
>> + =A0 expected benefit due to cross-module inlining. When the imported m=
odules
>> + =A0 are written out, they are sorted with highest weight first.
>> +
>> + =A0 The following example illustrates how the weight is computed:
>> +
>> + =A0 Suppose we are processing call-graph node A. It calls function B 5=
0 times,
>> + =A0 which calls function C 1000 times, and function E 800 times. Lets =
say B has
>> + =A0 another in-edge from function D, with edge-count of 50. Say all the
>> + =A0 functions are in separate modules (modules a, b, c, d, e, respecti=
vely):
>> +
>> + =A0 =A0 =A0 =A0 =A0 =A0 =A0D
>> + =A0 =A0 =A0 =A0 =A0 =A0 =A0|
>> + =A0 =A0 =A0 =A0 =A0 =A0 =A0| 50
>> + =A0 =A0 =A0 =A0 =A0 =A0 =A0|
>> + =A0 =A0 =A0 50 =A0 =A0 v =A0 =A0 1000
>> + =A0A --------> B ----------> C
>> + =A0 =A0 =A0 =A0 =A0 =A0 =A0|
>> + =A0 =A0 =A0 =A0 =A0 =A0 =A0| 800
>> + =A0 =A0 =A0 =A0 =A0 =A0 =A0|
>> + =A0 =A0 =A0 =A0 =A0 =A0 =A0v
>> + =A0 =A0 =A0 =A0 =A0 =A0 =A0E
>> +
>> + =A0Nodes are processed in depth-first order, so when processing A, we =
first
>> + =A0process B. For node B, we are going to add module c to the imported=
-module
>> + =A0set, with weight 1000 (edge-count), and module e with weight 800.
>> + =A0Coming back to A, we are going to add the imported-module-set of B =
to A,
>> + =A0after doing some scaling.
>> + =A0The first scaling factor comes from the fact that A calls B 50 time=
s, but B
>> + =A0has in-edge-count total of 100. So this scaling factor is 50/100 =
=3D 0.5
>> + =A0The second scaling factor is that since B is in a different module =
than A,
>> + =A0we want to slightly downgrade imported modules of B, before adding =
to the
>> + =A0imported-modules set of A. This scaling factor has a default value =
of 50%
>> + =A0(can be set via env variable GCOV_DYN_IMPORT_SCALE).
>> + =A0So we end up adding modules c and e to the imported-set of A, with =
weights
>> + =A00.5*0.5*1000=3D250 and 0.5*0.5*800=3D200, respectively.
>> +
>> + =A0Next, we have to add module b itself to A. The weight is computed a=
s the
>> + =A0edge-count plus the sum of scaled-weights of all modules in the
>> + =A0imported-module set of B, i.e., 50 + 250 + 200 =3D 500.
>> +
>> + =A0In computing the weight of module b, we add the sum of scaled-weigh=
ts of
>> + =A0imported modules of b, because it doesn't make sense to import c, e=
 in
>> + =A0module a, until module b is imported. =A0*/
>> +
>> +static void
>> +gcov_process_cgraph_node (struct dyn_cgraph_node *node,
>> + =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0gcov_type cutoff_co=
unt,
>> + =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 unsigned import_scale)
>> +{
>> + =A0unsigned mod_id;
>> + =A0struct dyn_cgraph_edge *callees;
>> + =A0struct dyn_cgraph_edge *callers;
>> + =A0node->visited =3D 1;
>> + =A0node->sum_in_count =3D 0;
>> +
>> + =A0callers =3D node->callers;
>> + =A0while (callers)
>> + =A0 =A0{
>> + =A0 =A0 =A0node->sum_in_count +=3D callers->count;
>> + =A0 =A0 =A0callers =3D callers->next_caller;
>> + =A0 =A0}
>> +
>> + =A0callees =3D node->callees;
>> + =A0mod_id =3D get_module_idx_from_func_glob_uid (node->guid);
>> +
>> + =A0while (callees)
>> + =A0 =A0{
>> + =A0 =A0 =A0if (!callees->callee->visited)
>> + =A0 =A0 =A0 =A0gcov_process_cgraph_node (callees->callee,
>> + =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0cut=
off_count,
>> + =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 import=
_scale);
>> + =A0 =A0 =A0callees =3D callees->next_callee;
>> + =A0 =A0}
>> +
>> + =A0callees =3D node->callees;
>> + =A0while (callees)
>> + =A0 =A0{
>> + =A0 =A0 =A0if (callees->count >=3D cutoff_count)
>> + =A0 =A0 =A0 =A0{
>> + =A0 =A0 =A0 =A0 =A0unsigned callee_mod_id;
>> + =A0 =A0 =A0 =A0 =A0struct dyn_pointer_set *imp_modules
>> + =A0 =A0 =A0 =A0 =A0 =A0 =A0=3D gcov_get_imp_module_set (node);
>> +
>> + =A0 =A0 =A0 =A0 =A0callee_mod_id
>> + =A0 =A0 =A0 =A0 =A0 =A0 =A0=3D get_module_idx_from_func_glob_uid (call=
ees->callee->guid);
>> +
>> + =A0 =A0 =A0 =A0 double callee_mod_wt =3D (double) callees->count;
>> + =A0 =A0 =A0 =A0 =A0if (callees->callee->imported_modules)
>> + =A0 =A0 =A0 =A0 =A0 {
>> + =A0 =A0 =A0 =A0 =A0 =A0 double scale =3D ((double) callees->count) /
>> + =A0 =A0 =A0 =A0 =A0 =A0 =A0 ((double) callees->callee->sum_in_count);
>> + =A0 =A0 =A0 =A0 =A0 =A0 /* Reduce weight if callee is in different mod=
ule. =A0*/
>> + =A0 =A0 =A0 =A0 =A0 =A0 if (mod_id !=3D callee_mod_id)
>> + =A0 =A0 =A0 =A0 =A0 =A0 =A0 scale =3D (scale * import_scale) / 100.0;
>> + =A0 =A0 =A0 =A0 =A0 =A0 pointer_set_traverse (callees->callee->importe=
d_modules,
>> + =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 gc=
ov_propagate_imp_modules,
>> + =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 im=
p_modules, &scale, &callee_mod_wt);
>> + =A0 =A0 =A0 =A0 =A0 }
>> + =A0 =A0 =A0 =A0 =A0if (mod_id !=3D callee_mod_id)
>> + =A0 =A0 =A0 =A0 =A0 =A0{
>> + =A0 =A0 =A0 =A0 =A0 =A0 =A0struct gcov_info *callee_mod_info
>> + =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0=3D get_module_info (callee_mod_id);
>> + =A0 =A0 =A0 =A0 =A0 =A0 =A0imp_mod_set_insert (imp_modules, callee_mod=
_info, callee_mod_wt);
>> + =A0 =A0 =A0 =A0 =A0 =A0}
>> + =A0 =A0 =A0 =A0}
>> +
>> + =A0 =A0 =A0callees =3D callees->next_callee;
>> + =A0 =A0}
>> +}
>> +
>> +/* Compute module grouping using CUTOFF_COUNT as the hot edge
>> + =A0 threshold. =A0*/
>> +
>> +#define DEFAULT_IMPORT_SCALE 100
>> +static void
>> +gcov_compute_module_groups (gcov_type cutoff_count)
>> +{
>> + =A0unsigned m_ix;
>> + =A0struct gcov_info *gi_ptr;
>> + =A0const char *import_scale_str;
>> + =A0unsigned import_scale =3D DEFAULT_IMPORT_SCALE;
>> +
>> + =A0import_scale_str =3D getenv ("GCOV_DYN_IMPORT_SCALE");
>> + =A0if (import_scale_str && strlen (import_scale_str))
>> + =A0 =A0import_scale =3D atoi (import_scale_str);
>> +
>> + =A0for (m_ix =3D 0; m_ix < the_dyn_call_graph.num_modules; m_ix++)
>> + =A0 =A0{
>> + =A0 =A0 =A0const struct gcov_fn_info *fi_ptr;
>> + =A0 =A0 =A0unsigned f_ix;
>> +
>> + =A0 =A0 =A0gi_ptr =3D the_dyn_call_graph.modules[m_ix];
>> +
>> + =A0 =A0 =A0for (f_ix =3D 0; f_ix < gi_ptr->n_functions; f_ix++)
>> + =A0 =A0 =A0 {
>> + =A0 =A0 =A0 =A0 struct dyn_cgraph_node *node;
>> +
>> + =A0 =A0 =A0 =A0 fi_ptr =3D the_dyn_call_graph.functions[m_ix][f_ix];
>> + =A0 =A0 =A0 =A0 node =3D *(pointer_set_find_or_insert
>> + =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0(the_dyn_call_graph.call_graph_node=
s[m_ix], fi_ptr->ident));
>> + =A0 =A0 =A0 =A0 gcc_assert (node);
>> + =A0 =A0 =A0 =A0 =A0if (node->visited)
>> + =A0 =A0 =A0 =A0 =A0 =A0continue;
>> +
>> + =A0 =A0 =A0 =A0 =A0gcov_process_cgraph_node (node, cutoff_count, impor=
t_scale);
>> + =A0 =A0 =A0 }
>> + =A0 =A0}
>> +
>> + =A0for (m_ix =3D 0; m_ix < the_dyn_call_graph.num_modules; m_ix++)
>> + =A0 =A0{
>> + =A0 =A0 =A0const struct gcov_fn_info *fi_ptr;
>> + =A0 =A0 =A0unsigned f_ix;
>> +
>> + =A0 =A0 =A0gi_ptr =3D the_dyn_call_graph.modules[m_ix];
>> +
>> + =A0 =A0 =A0for (f_ix =3D 0; f_ix < gi_ptr->n_functions; f_ix++)
>> + =A0 =A0 =A0 {
>> + =A0 =A0 =A0 =A0 struct dyn_cgraph_node *node;
>> + =A0 =A0 =A0 =A0 =A0unsigned mod_id;
>> + =A0 =A0 =A0 =A0 =A0struct dyn_pointer_set *imp_modules;
>> +
>> + =A0 =A0 =A0 =A0 fi_ptr =3D the_dyn_call_graph.functions[m_ix][f_ix];
>> + =A0 =A0 =A0 =A0 node =3D *(pointer_set_find_or_insert
>> + =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0(the_dyn_call_graph.call_graph_node=
s[m_ix], fi_ptr->ident));
>> + =A0 =A0 =A0 =A0 gcc_assert (node);
>> +
>> + =A0 =A0 =A0 =A0 =A0if (!node->imported_modules)
>> + =A0 =A0 =A0 =A0 =A0 =A0continue;
>> +
>> + =A0 =A0 =A0 =A0 =A0mod_id =3D get_module_idx_from_func_glob_uid (node-=
>guid);
>> + =A0 =A0 =A0 =A0 =A0gcc_assert (mod_id =3D=3D m_ix);
>> +
>> + =A0 =A0 =A0 =A0 =A0imp_modules
>> + =A0 =A0 =A0 =A0 =A0 =A0 =A0=3D gcov_get_module_imp_module_set (
>> + =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0&the_dyn_call_graph.sup_modules[mod=
_id]);
>> +
>> + =A0 =A0 =A0 =A0 =A0pointer_set_traverse (node->imported_modules,
>> + =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0gcov_pr=
opagate_imp_modules,
>> + =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0imp_mod=
ules, 0, 0);
>> + =A0 =A0 =A0 }
>> + =A0 =A0}
>> +
>> + =A0/* Now compute the export attribute =A0*/
>> + =A0for (m_ix =3D 0; m_ix < the_dyn_call_graph.num_modules; m_ix++)
>> + =A0 =A0{
>> + =A0 =A0 =A0struct dyn_module_info *mi
>> + =A0 =A0 =A0 =A0 =A0=3D &the_dyn_call_graph.sup_modules[m_ix];
>> + =A0 =A0 =A0if (mi->imported_modules)
>> + =A0 =A0 =A0 =A0pointer_set_traverse (mi->imported_modules,
>> + =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0gcov_mark_e=
xport_modules, 0, 0, 0);
>> + =A0 =A0}
>> +}
>> +
>> +/* For each module, compute at random, the group of imported modules,
>> + =A0 that is of size at most MAX_GROUP_SIZE. =A0*/
>> +
>> +static void
>> +gcov_compute_random_module_groups (unsigned max_group_size)
>> +{
>> + =A0unsigned m_ix;
>> +
>> + =A0if (max_group_size > the_dyn_call_graph.num_modules)
>> + =A0 =A0max_group_size =3D the_dyn_call_graph.num_modules;
>> +
>> + =A0for (m_ix =3D 0; m_ix < the_dyn_call_graph.num_modules; m_ix++)
>> + =A0 =A0{
>> + =A0 =A0 =A0struct dyn_pointer_set *imp_modules =3D
>> + =A0 =A0 =A0 gcov_get_module_imp_module_set (&the_dyn_call_graph.sup_mo=
dules[m_ix]);
>> + =A0 =A0 =A0int cur_group_size =3D random () % max_group_size;
>> + =A0 =A0 =A0int i =3D 0;
>> + =A0 =A0 =A0while (i < cur_group_size)
>> + =A0 =A0 =A0 {
>> + =A0 =A0 =A0 =A0 struct gcov_info *imp_mod_info;
>> + =A0 =A0 =A0 =A0 unsigned mod_id =3D random () % the_dyn_call_graph.num=
_modules;
>> + =A0 =A0 =A0 =A0 if (mod_id =3D=3D m_ix)
>> + =A0 =A0 =A0 =A0 =A0 continue;
>> + =A0 =A0 =A0 =A0 imp_mod_info =3D get_module_info (mod_id);
>> + =A0 =A0 =A0 =A0 if (!imp_mod_set_insert (imp_modules, imp_mod_info, 1.=
0))
>> + =A0 =A0 =A0 =A0 =A0 i++;
>> + =A0 =A0 =A0 }
>> + =A0 =A0}
>> +
>> + =A0/* Now compute the export attribute =A0*/
>> + =A0for (m_ix =3D 0; m_ix < the_dyn_call_graph.num_modules; m_ix++)
>> + =A0 =A0{
>> + =A0 =A0 =A0struct dyn_module_info *mi
>> + =A0 =A0 =A0 =3D &the_dyn_call_graph.sup_modules[m_ix];
>> + =A0 =A0 =A0if (mi->imported_modules)
>> + =A0 =A0 =A0 =A0pointer_set_traverse (mi->imported_modules,
>> + =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0gcov_mark_e=
xport_modules, 0, 0, 0);
>> + =A0 =A0}
>> +}
>> +
>> +/* Write out MOD_INFO into the gcda file. IS_PRIMARY is a flag
>> + =A0 indicating if the module is the primary module in the group. =A0*/
>> +
>> +static void
>> +gcov_write_module_info (const struct gcov_info *mod_info,
>> + =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0unsigned is_primary)
>> +{
>> + =A0gcov_unsigned_t len =3D 0, filename_len =3D 0, src_filename_len =3D=
 0, i, j;
>> + =A0gcov_unsigned_t num_strings;
>> + =A0gcov_unsigned_t *aligned_fname;
>> + =A0struct gcov_module_info =A0*module_info =3D mod_info->mod_info;
>> + =A0filename_len =3D (strlen (module_info->da_filename) +
>> + =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 sizeof (gcov_unsigned_t)) / sizeof (gc=
ov_unsigned_t);
>> + =A0src_filename_len =3D (strlen (module_info->source_filename) +
>> + =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 sizeof (gcov_unsigned_t)) / si=
zeof (gcov_unsigned_t);
>> + =A0len =3D filename_len + src_filename_len;
>> + =A0len +=3D 2; /* each name string is led by a length. =A0*/
>> +
>> + =A0num_strings =3D module_info->num_quote_paths + module_info->num_bra=
cket_paths +
>> + =A0 =A0module_info->num_cpp_defines + module_info->num_cpp_includes +
>> + =A0 =A0module_info->num_cl_args;
>> + =A0for (i =3D 0; i < num_strings; i++)
>> + =A0 =A0{
>> + =A0 =A0 =A0gcov_unsigned_t string_len
>> + =A0 =A0 =A0 =A0 =A0=3D (strlen (module_info->string_array[i]) + sizeof=
 (gcov_unsigned_t))
>> + =A0 =A0 =A0 =A0 =A0/ sizeof (gcov_unsigned_t);
>> + =A0 =A0 =A0len +=3D string_len;
>> + =A0 =A0 =A0len +=3D 1; /* Each string is lead by a length. =A0*/
>> + =A0 =A0}
>> +
>> + =A0len +=3D 9; /* 9 more fields */
>> +
>> + =A0gcov_write_tag_length (GCOV_TAG_MODULE_INFO, len);
>> + =A0gcov_write_unsigned (module_info->ident);
>> + =A0gcov_write_unsigned (is_primary);
>> + =A0gcov_write_unsigned (module_info->is_exported);
>> + =A0gcov_write_unsigned (module_info->lang);
>> + =A0gcov_write_unsigned (module_info->num_quote_paths);
>> + =A0gcov_write_unsigned (module_info->num_bracket_paths);
>> + =A0gcov_write_unsigned (module_info->num_cpp_defines);
>> + =A0gcov_write_unsigned (module_info->num_cpp_includes);
>> + =A0gcov_write_unsigned (module_info->num_cl_args);
>> +
>> + =A0/* Now write the filenames */
>> + =A0aligned_fname =3D (gcov_unsigned_t *) alloca ((filename_len + src_f=
ilename_len + 2) *
>> + =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =
=A0 =A0 =A0 =A0 =A0 sizeof (gcov_unsigned_t));
>> + =A0memset (aligned_fname, 0,
>> + =A0 =A0 =A0 =A0 =A0(filename_len + src_filename_len + 2) * sizeof (gco=
v_unsigned_t));
>> + =A0aligned_fname[0] =3D filename_len;
>> + =A0strcpy ((char*) (aligned_fname + 1), module_info->da_filename);
>> + =A0aligned_fname[filename_len + 1] =3D src_filename_len;
>> + =A0strcpy ((char*) (aligned_fname + filename_len + 2), module_info->so=
urce_filename);
>> +
>> + =A0for (i =3D 0; i < (filename_len + src_filename_len + 2); i++)
>> + =A0 =A0gcov_write_unsigned (aligned_fname[i]);
>> +
>> + =A0/* Now write the string array. =A0*/
>> + =A0for (j =3D 0; j < num_strings; j++)
>> + =A0 =A0{
>> + =A0 =A0 =A0gcov_unsigned_t *aligned_string;
>> + =A0 =A0 =A0gcov_unsigned_t string_len =3D
>> + =A0 =A0 =A0 (strlen (module_info->string_array[j]) + sizeof (gcov_unsi=
gned_t)) /
>> + =A0 =A0 =A0 sizeof (gcov_unsigned_t);
>> + =A0 =A0 =A0aligned_string =3D (gcov_unsigned_t *)
>> + =A0 =A0 =A0 alloca ((string_len + 1) * sizeof (gcov_unsigned_t));
>> + =A0 =A0 =A0memset (aligned_string, 0, (string_len + 1) * sizeof (gcov_=
unsigned_t));
>> + =A0 =A0 =A0aligned_string[0] =3D string_len;
>> + =A0 =A0 =A0strcpy ((char*) (aligned_string + 1), module_info->string_a=
rray[j]);
>> + =A0 =A0 =A0for (i =3D 0; i < (string_len + 1); i++)
>> + =A0 =A0 =A0 =A0gcov_write_unsigned (aligned_string[i]);
>> + =A0 =A0}
>> +}
>> +
>> +/* Write out MOD_INFO and its imported modules into gcda file. =A0*/
>> +
>> +void
>> +gcov_write_module_infos (struct gcov_info *mod_info)
>> +{
>> + =A0unsigned mod_id, imp_len =3D 0;
>> + =A0const struct dyn_imp_mod **imp_mods;
>> +
>> + =A0mod_id =3D get_module_idx (mod_info);
>> + =A0gcov_write_module_info (mod_info, 1);
>> +
>> + =A0imp_mods =3D gcov_get_sorted_import_module_array (mod_info, &imp_le=
n);
>> + =A0if (imp_mods)
>> + =A0 =A0{
>> + =A0 =A0 =A0unsigned i;
>> +
>> + =A0 =A0 =A0for (i =3D 0; i < imp_len; i++)
>> + =A0 =A0 =A0 =A0{
>> + =A0 =A0 =A0 =A0 =A0const struct gcov_info *imp_mod =3D imp_mods[i]->im=
p_mod;
>> + =A0 =A0 =A0 =A0 =A0gcov_write_module_info (imp_mod, 0);
>> + =A0 =A0 =A0 =A0}
>> + =A0 =A0 =A0free (imp_mods);
>> + =A0 =A0}
>> +}
>> +
>> +/* Compute module groups needed for L-IPO compilation. =A0*/
>> +
>> +void
>> +__gcov_compute_module_groups (void)
>> +{
>> + =A0gcov_type cut_off_count;
>> + =A0char *seed =3D getenv ("LIPO_RANDOM_GROUPING");
>> + =A0char *max_group_size =3D seed ? strchr (seed, ':') : 0;
>> +
>> + =A0if (seed && max_group_size)
>> + =A0 =A0{
>> + =A0 =A0 =A0*max_group_size =3D '\0';
>> + =A0 =A0 =A0max_group_size++;
>> + =A0 =A0 =A0srandom (atoi (seed));
>> + =A0 =A0 =A0init_dyn_call_graph ();
>> + =A0 =A0 =A0gcov_compute_random_module_groups (atoi (max_group_size));
>> + =A0 =A0 =A0return;
>> + =A0 =A0}
>> +
>> + =A0/* First compute dynamic call graph. =A0*/
>> + =A0gcov_build_callgraph ();
>> +
>> + =A0cut_off_count =3D gcov_compute_cutoff_count ();
>> +
>> + =A0gcov_compute_module_groups (cut_off_count);
>> +
>> + =A0gcov_dump_callgraph (cut_off_count);
>> +
>> +}
>> +
>> +/* Dumper function for NODE. =A0*/
>> +static void
>> +gcov_dump_cgraph_node_short (struct dyn_cgraph_node *node)
>> +{
>> + =A0unsigned mod_id, func_id;
>> + =A0struct gcov_info *mod_info;
>> + =A0mod_id =3D get_module_idx_from_func_glob_uid (node->guid);
>> + =A0func_id =3D get_intra_module_func_id (node->guid);
>> +
>> + =A0mod_info =3D the_dyn_call_graph.modules[mod_id];
>> +
>> + =A0fprintf (stderr, "NODE(%llx) module(%s) func(%u)",
>> + =A0 =A0 =A0 =A0 =A0 (long long)node->guid,
>> + =A0 =A0 =A0 =A0 =A0 mod_info->mod_info->source_filename, func_id);
>> +}
>> +
>> +/* Dumper function for NODE. =A0 M is the module id and F is the functi=
on id. =A0*/
>> +
>> +static void
>> +gcov_dump_cgraph_node (struct dyn_cgraph_node *node, unsigned m, unsign=
ed f)
>> +{
>> + =A0unsigned mod_id, func_id;
>> + =A0struct gcov_info *mod_info;
>> + =A0struct dyn_cgraph_edge *callers;
>> + =A0struct dyn_cgraph_edge *callees;
>> +
>> + =A0mod_id =3D get_module_idx_from_func_glob_uid (node->guid);
>> + =A0func_id =3D get_intra_module_func_id (node->guid);
>> + =A0gcc_assert (mod_id =3D=3D m && func_id =3D=3D f);
>> +
>> + =A0mod_info =3D the_dyn_call_graph.modules[mod_id];
>> +
>> + =A0fprintf (stderr, "NODE(%llx) module(%s) func(%x)\n",
>> + =A0 =A0 =A0 =A0 =A0 (long long) node->guid,
>> + =A0 =A0 =A0 =A0 =A0 mod_info->mod_info->source_filename, f);
>> +
>> + =A0/* Now dump callers. =A0*/
>> + =A0callers =3D node->callers;
>> + =A0fprintf (stderr, "\t[CALLERS]\n");
>> + =A0while (callers !=3D 0)
>> + =A0 =A0{
>> + =A0 =A0 =A0fprintf (stderr,"\t\t[count=3D%ld] ", (long) =A0callers->co=
unt);
>> + =A0 =A0 =A0gcov_dump_cgraph_node_short (callers->caller);
>> + =A0 =A0 =A0fprintf (stderr,"\n");
>> + =A0 =A0 =A0callers =3D callers->next_caller;
>> + =A0 =A0}
>> +
>> + =A0callees =3D node->callees;
>> + =A0fprintf (stderr, "\t[CALLEES]\n");
>> + =A0while (callees !=3D 0)
>> + =A0 =A0{
>> + =A0 =A0 =A0fprintf (stderr,"\t\t[count=3D%ld] ", (long) =A0callees->co=
unt);
>> + =A0 =A0 =A0gcov_dump_cgraph_node_short (callees->callee);
>> + =A0 =A0 =A0fprintf (stderr,"\n");
>> + =A0 =A0 =A0callees =3D callees->next_callee;
>> + =A0 =A0}
>> +}
>> +
>> +/* Dumper function for NODE. =A0 M is the module id and F is the functi=
on id. =A0*/
>> +
>> +static void
>> +gcov_dump_cgraph_node_dot (struct dyn_cgraph_node *node,
>> + =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 unsigned m, unsign=
ed f,
>> + =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 gcov_type cutoff_c=
ount)
>> +{
>> + =A0unsigned mod_id, func_id, imp_len =3D 0, i;
>> + =A0struct gcov_info *mod_info;
>> + =A0const struct dyn_imp_mod **imp_mods;
>> + =A0struct dyn_cgraph_edge *callees;
>> +
>> + =A0mod_id =3D get_module_idx_from_func_glob_uid (node->guid);
>> + =A0func_id =3D get_intra_module_func_id (node->guid);
>> + =A0gcc_assert (mod_id =3D=3D m && func_id =3D=3D f);
>> +
>> + =A0mod_info =3D the_dyn_call_graph.modules[mod_id];
>> +
>> + =A0fprintf (stderr, "NODE_%llx[label=3D\"MODULE\\n(%s)\\n FUNC(%x)\\n",
>> + =A0 =A0 =A0 =A0 =A0 (long long) node->guid, mod_info->mod_info->source=
_filename, f);
>> +
>> + =A0imp_mods =3D gcov_get_sorted_import_module_array (mod_info, &imp_le=
n);
>> + =A0fprintf (stderr, "IMPORTS:\\n");
>> + =A0if (imp_mods)
>> + =A0 =A0{
>> + =A0 =A0 =A0for (i =3D 0; i < imp_len; i++)
>> + =A0 =A0 =A0 =A0fprintf (stderr, "%s\\n", imp_mods[i]->imp_mod->mod_inf=
o->source_filename);
>> + =A0 =A0 =A0fprintf (stderr, "\"]\n");
>> + =A0 =A0 =A0free (imp_mods);
>> + =A0 =A0}
>> + =A0else
>> + =A0 =A0fprintf (stderr, "\"]\n");
>> +
>> + =A0callees =3D node->callees;
>> + =A0while (callees !=3D 0)
>> + =A0 =A0{
>> + =A0 =A0 =A0if (callees->count >=3D cutoff_count)
>> + =A0 =A0 =A0 =A0fprintf (stderr, "NODE_%llx -> NODE_%llx[label=3D%lld c=
olor=3Dred]\n",
>> + =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 (long long) node->guid, (long long) ca=
llees->callee->guid,
>> + =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 (long long) callees->count);
>> + =A0 =A0 =A0else
>> + =A0 =A0 =A0 =A0fprintf (stderr, "NODE_%llx -> NODE_%llx[label=3D%lld c=
olor=3Dblue]\n",
>> + =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 (long long) node->guid, (long long) ca=
llees->callee->guid,
>> + =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 (long long) callees->count);
>> + =A0 =A0 =A0callees =3D callees->next_callee;
>> + =A0 =A0}
>> +}
>> +
>> +/* Dump dynamic call graph. =A0CUTOFF_COUNT is the computed hot edge th=
reshold. =A0*/
>> +
>> +static void
>> +gcov_dump_callgraph (gcov_type cutoff_count)
>> +{
>> + =A0struct gcov_info *gi_ptr;
>> + =A0unsigned m_ix;
>> + =A0const char *dyn_cgraph_dump =3D 0;
>> +
>> + =A0dyn_cgraph_dump =3D getenv ("GCOV_DYN_CGRAPH_DUMP");
>> +
>> + =A0if (!dyn_cgraph_dump || !strlen (dyn_cgraph_dump))
>> + =A0 =A0 =A0return;
>> +
>> + =A0fprintf (stderr,"digraph dyn_call_graph {\n");
>> + =A0fprintf (stderr,"node[shape=3Dbox]\nsize=3D\"11,8.5\"\n");
>> +
>> + =A0for (m_ix =3D 0; m_ix < the_dyn_call_graph.num_modules; m_ix++)
>> + =A0 =A0{
>> + =A0 =A0 =A0const struct gcov_fn_info *fi_ptr;
>> + =A0 =A0 =A0unsigned f_ix;
>> +
>> + =A0 =A0 =A0gi_ptr =3D the_dyn_call_graph.modules[m_ix];
>> +
>> + =A0 =A0 =A0for (f_ix =3D 0; f_ix < gi_ptr->n_functions; f_ix++)
>> + =A0 =A0 =A0 {
>> + =A0 =A0 =A0 =A0 struct dyn_cgraph_node *node;
>> + =A0 =A0 =A0 =A0 fi_ptr =3D the_dyn_call_graph.functions[m_ix][f_ix];
>> +
>> + =A0 =A0 =A0 =A0 node =3D *(pointer_set_find_or_insert
>> + =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0(the_dyn_call_graph.call_graph_node=
s[m_ix], fi_ptr->ident));
>> + =A0 =A0 =A0 =A0 gcc_assert (node);
>> +
>> + =A0 =A0 =A0 =A0 =A0/* skip dead functions =A0*/
>> + =A0 =A0 =A0 =A0 =A0if (!node->callees && !node->callers)
>> + =A0 =A0 =A0 =A0 =A0 =A0continue;
>> +
>> + =A0 =A0 =A0 =A0 =A0if (dyn_cgraph_dump[0] =3D=3D '1')
>> + =A0 =A0 =A0 =A0 =A0 =A0gcov_dump_cgraph_node (node, m_ix, fi_ptr->iden=
t);
>> + =A0 =A0 =A0 =A0 =A0else
>> + =A0 =A0 =A0 =A0 =A0 =A0gcov_dump_cgraph_node_dot (node, m_ix, fi_ptr->=
ident,
>> + =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =
=A0 =A0 cutoff_count);
>> + =A0 =A0 =A0 }
>> + =A0 =A0}
>> + =A0fprintf (stderr,"}\n");
>> +}
>> +
>> +
>> +#endif
>> Index: dbgcnt.def
>> =3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=
=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=
=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D
>> --- dbgcnt.def =A0(revision 172880)
>> +++ dbgcnt.def =A0(working copy)
>> @@ -141,10 +141,12 @@ echo ubound: $ub
>> =A0*/
>>
>> =A0/* Debug counter definitions. =A0*/
>> +DEBUG_COUNTER (alias)
>> =A0DEBUG_COUNTER (auto_inc_dec)
>> =A0DEBUG_COUNTER (ccp)
>> =A0DEBUG_COUNTER (cfg_cleanup)
>> =A0DEBUG_COUNTER (cse2_move2add)
>> +DEBUG_COUNTER (clone)
>> =A0DEBUG_COUNTER (cprop)
>> =A0DEBUG_COUNTER (dce)
>> =A0DEBUG_COUNTER (dce_fast)
>> @@ -165,6 +167,7 @@ DEBUG_COUNTER (if_conversion)
>> =A0DEBUG_COUNTER (if_conversion_tree)
>> =A0DEBUG_COUNTER (if_after_combine)
>> =A0DEBUG_COUNTER (if_after_reload)
>> +DEBUG_COUNTER (inl)
>> =A0DEBUG_COUNTER (local_alloc_for_sched)
>> =A0DEBUG_COUNTER (postreload_cse)
>> =A0DEBUG_COUNTER (pre)
>> Index: c-family/c-opts.c
>> =3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=
=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=
=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D
>> --- c-family/c-opts.c =A0 (revision 172880)
>> +++ c-family/c-opts.c =A0 (working copy)
>> @@ -38,6 +38,9 @@ along with GCC; see the file COPYING3.
>> =A0#include "mkdeps.h"
>> =A0#include "target.h" =A0 =A0 =A0 =A0 =A0 =A0/* For gcc_targetcm. =A0*/
>> =A0#include "tm_p.h" =A0 =A0 =A0 =A0 =A0 =A0 =A0/* For C_COMMON_OVERRIDE=
_OPTIONS. =A0*/
>> +#include "function.h"
>> +#include "params.h"
>> +#include "l-ipo.h"
>>
>> =A0#ifndef DOLLARS_IN_IDENTIFIERS
>> =A0# define DOLLARS_IN_IDENTIFIERS true
>> @@ -364,6 +367,7 @@ c_common_handle_option (size_t scode, co
>> =A0 =A0 =A0 warn_missing_braces =3D value;
>> =A0 =A0 =A0 warn_parentheses =3D value;
>> =A0 =A0 =A0 warn_return_type =3D value;
>> + =A0 =A0 =A0warn_ripa_opt_mismatch =3D value;
>> =A0 =A0 =A0 warn_sequence_point =3D value; =A0 =A0 /* Was C only. =A0*/
>> =A0 =A0 =A0 warn_switch =3D value;
>> =A0 =A0 =A0 if (warn_strict_aliasing =3D=3D -1)
>> @@ -864,6 +868,10 @@ c_common_post_options (const char **pfil
>> =A0 else if (!flag_gnu89_inline && !flag_isoc99)
>> =A0 =A0 error ("-fno-gnu89-inline is only supported in GNU99 or C99 mode=
");
>>
>> + =A0if (flag_dyn_ipa && cpp_opts->preprocessed)
>> + =A0 =A0error ("-fpreprocessed/-save-temps are not supported with -frip=
a");
>> +
>> +
>> =A0 /* Default to ObjC sjlj exception handling if NeXT runtime. =A0*/
>> =A0 if (flag_objc_sjlj_exceptions < 0)
>> =A0 =A0 flag_objc_sjlj_exceptions =3D flag_next_runtime;
>> @@ -1061,6 +1069,28 @@ c_common_init (void)
>> =A0 return true;
>> =A0}
>>
>> +/* Return TRUE if the lipo maximum memory consumption limit is reached,=
 and
>> + =A0 we should not import any further auxiliary modules. Check after pa=
rsing
>> + =A0 each module, the Ith module being the just parsed module. =A0*/
>> +static bool
>> +lipo_max_mem_reached (unsigned int i)
>> +{
>> + =A0if (L_IPO_COMP_MODE && PARAM_VALUE (PARAM_MAX_LIPO_MEMORY)
>> + =A0 =A0 =A0&& i < (num_in_fnames - 1)
>> + =A0 =A0 =A0&& ((ggc_total_allocated () >> 10)
>> + =A0 =A0 =A0 =A0 =A0> (size_t) PARAM_VALUE (PARAM_MAX_LIPO_MEMORY))) {
>> + =A0 =A0i++;
>> + =A0 =A0do {
>> + =A0 =A0 =A0inform (input_location, "Not importing %s: maximum memory "
>> + =A0 =A0 =A0 =A0 =A0 =A0 "consumption reached", in_fnames[i]);
>> + =A0 =A0 =A0i++;
>> + =A0 =A0} while (i < num_in_fnames);
>> + =A0 =A0return true;
>> + =A0}
>> + =A0return false;
>> +}
>> +
>> +
>> =A0/* Initialize the integrated preprocessor after debug output has been
>> =A0 =A0initialized; loop over each input file. =A0*/
>> =A0void
>> @@ -1073,8 +1103,15 @@ c_common_parse_file (void)
>> =A0 =A0 {
>> =A0 =A0 =A0 c_finish_options ();
>> =A0 =A0 =A0 pch_init ();
>> + =A0 =A0 =A0set_lipo_c_parsing_context (parse_in, i, verbose);
>> =A0 =A0 =A0 push_file_scope ();
>> =A0 =A0 =A0 c_parse_file ();
>> + =A0 =A0 =A0/* In lipo mode, processing too many auxiliary files will c=
ause us
>> + =A0 =A0 =A0 =A0to hit memory limits, and cause thrashing -- prevent th=
is by not
>> + =A0 =A0 =A0 =A0processing any further auxiliary modules if we reach a =
certain
>> + =A0 =A0 =A0 =A0memory limit. =A0*/
>> + =A0 =A0 =A0if (lipo_max_mem_reached (i))
>> + =A0 =A0 =A0 num_in_fnames =3D i + 1;
>> =A0 =A0 =A0 pop_file_scope ();
>> =A0 =A0 =A0 /* And end the main input file, if the debug writer wants it=
 =A0*/
>> =A0 =A0 =A0 if (debug_hooks->start_end_main_source_file)
>> @@ -1083,6 +1120,7 @@ c_common_parse_file (void)
>> =A0 =A0 =A0 =A0break;
>> =A0 =A0 =A0 cpp_undef_all (parse_in);
>> =A0 =A0 =A0 cpp_clear_file_cache (parse_in);
>> + =A0 =A0 =A0deferred_count =3D 0;
>> =A0 =A0 =A0 this_input_filename
>> =A0 =A0 =A0 =A0=3D cpp_read_main_file (parse_in, in_fnames[i]);
>> =A0 =A0 =A0 /* If an input file is missing, abandon further compilation.
>> @@ -1320,9 +1358,15 @@ c_finish_options (void)
>> =A0 =A0 =A0 =A0 =A0struct deferred_opt *opt =3D &deferred_opts[i];
>>
>> =A0 =A0 =A0 =A0 =A0if (opt->code =3D=3D OPT_D)
>> - =A0 =A0 =A0 =A0 =A0 cpp_define (parse_in, opt->arg);
>> + =A0 =A0 =A0 =A0 =A0 {
>> + =A0 =A0 =A0 =A0 =A0 =A0 cpp_define (parse_in, opt->arg);
>> + =A0 =A0 =A0 =A0 =A0 =A0 coverage_note_define (opt->arg, true);
>> + =A0 =A0 =A0 =A0 =A0 }
>> =A0 =A0 =A0 =A0 =A0else if (opt->code =3D=3D OPT_U)
>> - =A0 =A0 =A0 =A0 =A0 cpp_undef (parse_in, opt->arg);
>> + =A0 =A0 =A0 =A0 =A0 {
>> + =A0 =A0 =A0 =A0 =A0 =A0 cpp_undef (parse_in, opt->arg);
>> + =A0 =A0 =A0 =A0 =A0 =A0 coverage_note_define (opt->arg, false);
>> + =A0 =A0 =A0 =A0 =A0 }
>> =A0 =A0 =A0 =A0 =A0else if (opt->code =3D=3D OPT_A)
>> =A0 =A0 =A0 =A0 =A0 =A0{
>> =A0 =A0 =A0 =A0 =A0 =A0 =A0if (opt->arg[0] =3D=3D '-')
>> @@ -1345,6 +1389,7 @@ c_finish_options (void)
>> =A0 =A0 =A0 =A0 =A0if (opt->code =3D=3D OPT_imacros
>> =A0 =A0 =A0 =A0 =A0 =A0 =A0&& cpp_push_include (parse_in, opt->arg))
>> =A0 =A0 =A0 =A0 =A0 =A0{
>> + =A0 =A0 =A0 =A0 =A0 =A0 coverage_note_include (opt->arg);
>> =A0 =A0 =A0 =A0 =A0 =A0 =A0/* Disable push_command_line_include callback=
 for now. =A0*/
>> =A0 =A0 =A0 =A0 =A0 =A0 =A0include_cursor =3D deferred_count + 1;
>> =A0 =A0 =A0 =A0 =A0 =A0 =A0cpp_scan_nooutput (parse_in);
>> @@ -1376,7 +1421,10 @@ push_command_line_include (void)
>>
>> =A0 =A0 =A0 if (!cpp_opts->preprocessed && opt->code =3D=3D OPT_include
>> =A0 =A0 =A0 =A0 =A0&& cpp_push_include (parse_in, opt->arg))
>> - =A0 =A0 =A0 return;
>> + =A0 =A0 =A0 {
>> + =A0 =A0 =A0 =A0 coverage_note_include (opt->arg);
>> + =A0 =A0 =A0 =A0 return;
>> + =A0 =A0 =A0 }
>> =A0 =A0 }
>>
>> =A0 if (include_cursor =3D=3D deferred_count)
>> Index: cgraph.c
>> =3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=
=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=
=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D
>> --- cgraph.c =A0 =A0(revision 172880)
>> +++ cgraph.c =A0 =A0(working copy)
>> @@ -98,6 +98,7 @@ The callgraph:
>> =A0#include "rtl.h"
>> =A0#include "ipa-utils.h"
>> =A0#include "lto-streamer.h"
>> +#include "l-ipo.h"
>>
>> =A0const char * const ld_plugin_symbol_resolution_names[]=3D
>> =A0{
>> @@ -485,6 +486,44 @@ cgraph_create_node (void)
>> =A0 return node;
>> =A0}
>>
>> +/* Add an entry in assembler hash for NODE. =A0*/
>> +void
>> +cgraph_add_assembler_hash_node (struct cgraph_node *node)
>> +{
>> + =A0if (assembler_name_hash)
>> + =A0 =A0{
>> + =A0 =A0 =A0void **aslot;
>> + =A0 =A0 =A0tree name =3D DECL_ASSEMBLER_NAME (node->decl);
>> +
>> + =A0 =A0 =A0aslot =3D htab_find_slot_with_hash (assembler_name_hash, na=
me,
>> + =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =
=A0 =A0 decl_assembler_name_hash (name),
>> + =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =
=A0 =A0 INSERT);
>> + =A0 =A0 =A0/* We can have multiple declarations with same assembler na=
me. For C++
>> + =A0 =A0 =A0 =A0it is __builtin_strlen and strlen, for instance. =A0Do =
we need to
>> + =A0 =A0 =A0 =A0record them all? =A0Original implementation marked just=
 first one
>> + =A0 =A0 =A0 =A0so lets hope for the best. =A0*/
>> + =A0 =A0 =A0if (*aslot =3D=3D NULL)
>> + =A0 =A0 =A0 *aslot =3D node;
>> + =A0 =A0}
>> +}
>> +
>> +/* Remove from assembler hash for NODE. =A0*/
>> +void
>> +cgraph_remove_assembler_hash_node (struct cgraph_node *node)
>> +{
>> + =A0void **slot;
>> + =A0if (assembler_name_hash)
>> + =A0 =A0{
>> + =A0 =A0 =A0tree name =3D DECL_ASSEMBLER_NAME (node->decl);
>> + =A0 =A0 =A0slot =3D htab_find_slot_with_hash (assembler_name_hash, nam=
e,
>> + =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =
=A0 =A0decl_assembler_name_hash (name),
>> + =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =
=A0 =A0NO_INSERT);
>> + =A0 =A0 =A0/* Inline clones are not hashed. =A0*/
>> + =A0 =A0 =A0if (slot && *slot =3D=3D node)
>> + =A0 =A0 =A0 =A0htab_clear_slot (assembler_name_hash, slot);
>> + =A0 =A0}
>> +}
>> +
>> =A0/* Return cgraph node assigned to DECL. =A0Create new one when needed=
. =A0*/
>>
>> =A0struct cgraph_node *
>> @@ -518,21 +557,7 @@ cgraph_node (tree decl)
>> =A0 =A0 =A0 node->next_nested =3D node->origin->nested;
>> =A0 =A0 =A0 node->origin->nested =3D node;
>> =A0 =A0 }
>> - =A0if (assembler_name_hash)
>> - =A0 =A0{
>> - =A0 =A0 =A0void **aslot;
>> - =A0 =A0 =A0tree name =3D DECL_ASSEMBLER_NAME (decl);
>> -
>> - =A0 =A0 =A0aslot =3D htab_find_slot_with_hash (assembler_name_hash, na=
me,
>> - =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =
=A0 =A0 decl_assembler_name_hash (name),
>> - =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =
=A0 =A0 INSERT);
>> - =A0 =A0 =A0/* We can have multiple declarations with same assembler na=
me. For C++
>> - =A0 =A0 =A0 =A0it is __builtin_strlen and strlen, for instance. =A0Do =
we need to
>> - =A0 =A0 =A0 =A0record them all? =A0Original implementation marked just=
 first one
>> - =A0 =A0 =A0 =A0so lets hope for the best. =A0*/
>> - =A0 =A0 =A0if (*aslot =3D=3D NULL)
>> - =A0 =A0 =A0 *aslot =3D node;
>> - =A0 =A0}
>> + =A0cgraph_add_assembler_hash_node (node);
>> =A0 return node;
>> =A0}
>>
>> @@ -830,7 +855,9 @@ cgraph_edge (struct cgraph_node *node, g
>> =A0 =A0 {
>> =A0 =A0 =A0 node->call_site_hash =3D htab_create_ggc (120, edge_hash, ed=
ge_eq, NULL);
>> =A0 =A0 =A0 for (e2 =3D node->callees; e2; e2 =3D e2->next_callee)
>> - =A0 =A0 =A0 cgraph_add_edge_to_call_site_hash (e2);
>> + =A0 =A0 =A0 /* Skip fake edges. =A0*/
>> + =A0 =A0 =A0 if (e2->call_stmt)
>> + =A0 =A0 =A0 =A0 cgraph_add_edge_to_call_site_hash (e2);
>> =A0 =A0 =A0 for (e2 =3D node->indirect_calls; e2; e2 =3D e2->next_callee)
>> =A0 =A0 =A0 =A0cgraph_add_edge_to_call_site_hash (e2);
>> =A0 =A0 }
>> @@ -1139,7 +1166,7 @@ cgraph_edge_remove_caller (struct cgraph
>> =A0 =A0 =A0 else
>> =A0 =A0 =A0 =A0e->caller->callees =3D e->next_callee;
>> =A0 =A0 }
>> - =A0if (e->caller->call_site_hash)
>> + =A0if (e->caller->call_site_hash && e->call_stmt)
>> =A0 =A0 htab_remove_elt_with_hash (e->caller->call_site_hash,
>> =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 e->call_stmt,
>> =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 htab_hash_po=
inter (e->call_stmt));
>> @@ -1178,6 +1205,26 @@ cgraph_remove_edge (struct cgraph_edge *
>> =A0 cgraph_free_edge (e);
>> =A0}
>>
>> +/* Remove fake cgraph edges for indirect calls. NODE is the callee
>> + =A0 of the edges. =A0*/
>> +
>> +void
>> +cgraph_remove_fake_indirect_call_in_edges (struct cgraph_node *node)
>> +{
>> + =A0struct cgraph_edge *f, *e;
>> +
>> + =A0if (!L_IPO_COMP_MODE)
>> + =A0 =A0return;
>> +
>> + =A0for (e =3D node->callers; e; e =3D f)
>> + =A0 =A0{
>> + =A0 =A0 =A0f =3D e->next_caller;
>> + =A0 =A0 =A0if (!e->call_stmt)
>> + =A0 =A0 =A0 =A0cgraph_remove_edge (e);
>> + =A0 =A0}
>> +}
>> +
>> +
>> =A0/* Set callee of call graph edge E and add it to the corresponding se=
t of
>> =A0 =A0callers. */
>>
>> @@ -1400,6 +1447,8 @@ cgraph_node_remove_callers (struct cgrap
>> =A0void
>> =A0cgraph_release_function_body (struct cgraph_node *node)
>> =A0{
>> + =A0if (cgraph_is_aux_decl_external (node))
>> + =A0 =A0DECL_EXTERNAL (node->decl) =3D 1;
>> =A0 if (DECL_STRUCT_FUNCTION (node->decl))
>> =A0 =A0 {
>> =A0 =A0 =A0 tree old_decl =3D current_function_decl;
>> @@ -1667,19 +1716,14 @@ cgraph_remove_node (struct cgraph_node *
>> =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0|| n->in_other_partition)))
>> =A0 =A0 =A0 =A0kill_body =3D true;
>> =A0 =A0 }
>> - =A0if (assembler_name_hash)
>> - =A0 =A0{
>> - =A0 =A0 =A0tree name =3D DECL_ASSEMBLER_NAME (node->decl);
>> - =A0 =A0 =A0slot =3D htab_find_slot_with_hash (assembler_name_hash, nam=
e,
>> - =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =
=A0 =A0decl_assembler_name_hash (name),
>> - =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =
=A0 =A0NO_INSERT);
>> - =A0 =A0 =A0/* Inline clones are not hashed. =A0*/
>> - =A0 =A0 =A0if (slot && *slot =3D=3D node)
>> - =A0 =A0 =A0 =A0htab_clear_slot (assembler_name_hash, slot);
>> - =A0 =A0}
>> +
>> + =A0cgraph_remove_assembler_hash_node (node);
>>
>> =A0 if (kill_body)
>> =A0 =A0 cgraph_release_function_body (node);
>> +
>> + =A0cgraph_remove_link_node (node);
>> +
>> =A0 node->decl =3D NULL;
>> =A0 if (node->call_site_hash)
>> =A0 =A0 {
>> @@ -1724,7 +1768,8 @@ cgraph_mark_reachable_node (struct cgrap
>> =A0 =A0 =A0 =A0 =A0 =A0 during the optimization process. =A0This can hap=
pen for extern
>> =A0 =A0 =A0 =A0 =A0 =A0 inlines when bodies was removed after inlining. =
=A0*/
>> =A0 =A0 =A0 =A0 =A0gcc_assert ((node->analyzed || node->in_other_partiti=
on
>> - =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0|| DECL_EXTERNAL (node->dec=
l)));
>> + =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0|| DECL_EXTERNAL (node->dec=
l)
>> + =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 || cgraph_is_aux_decl_exte=
rnal (node)));
>> =A0 =A0 =A0 =A0}
>> =A0 =A0 =A0 else
>> =A0 =A0 =A0 =A0 notice_global_symbol (node->decl);
>> @@ -2190,6 +2235,7 @@ cgraph_clone_node (struct cgraph_node *n
>> =A0 new_node->global =3D n->global;
>> =A0 new_node->rtl =3D n->rtl;
>> =A0 new_node->count =3D count;
>> + =A0new_node->is_versioned_clone =3D n->is_versioned_clone;
>> =A0 new_node->frequency =3D n->frequency;
>> =A0 new_node->clone =3D n->clone;
>> =A0 new_node->clone.tree_map =3D 0;
>> @@ -2239,17 +2285,8 @@ cgraph_clone_node (struct cgraph_node *n
>> =A0 =A0 =A0 slot =3D (struct cgraph_node **) htab_find_slot (cgraph_hash=
, new_node, INSERT);
>> =A0 =A0 =A0 gcc_assert (!*slot);
>> =A0 =A0 =A0 *slot =3D new_node;
>> - =A0 =A0 =A0if (assembler_name_hash)
>> - =A0 =A0 =A0 {
>> - =A0 =A0 =A0 =A0 void **aslot;
>> - =A0 =A0 =A0 =A0 tree name =3D DECL_ASSEMBLER_NAME (decl);
>> -
>> - =A0 =A0 =A0 =A0 aslot =3D htab_find_slot_with_hash (assembler_name_has=
h, name,
>> - =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =
=A0 =A0 =A0 =A0 decl_assembler_name_hash (name),
>> - =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =
=A0 =A0 =A0 =A0 INSERT);
>> - =A0 =A0 =A0 =A0 gcc_assert (!*aslot);
>> - =A0 =A0 =A0 =A0 *aslot =3D new_node;
>> - =A0 =A0 =A0 }
>> + =A0 =A0 =A0cgraph_add_assembler_hash_node (new_node);
>> + =A0 =A0 =A0cgraph_link_node (new_node);
>> =A0 =A0 }
>> =A0 return new_node;
>> =A0}
>> @@ -2379,6 +2416,7 @@ cgraph_create_virtual_clone (struct cgra
>> =A0 =A0 new_node->clone.combined_args_to_skip =3D args_to_skip;
>> =A0 new_node->local.externally_visible =3D 0;
>> =A0 new_node->local.local =3D 1;
>> + =A0new_node->is_versioned_clone =3D 1;
>> =A0 new_node->lowered =3D true;
>> =A0 new_node->reachable =3D true;
>>
>> Index: cgraph.h
>> =3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=
=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=
=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D
>> --- cgraph.h =A0 =A0(revision 172880)
>> +++ cgraph.h =A0 =A0(working copy)
>> @@ -286,6 +286,8 @@ struct GTY((chain_next ("%h.next"), chai
>> =A0 unsigned alias : 1;
>> =A0 /* Set for nodes that was constructed and finalized by frontend. =A0=
*/
>> =A0 unsigned finalized_by_frontend : 1;
>> + =A0/* Is this function cloned during versioning ? =A0*/
>> + =A0unsigned is_versioned_clone : 1;
>> =A0 /* Set for alias and thunk nodes, same_body points to the node they =
are alias
>> =A0 =A0 =A0of and they are linked through the next/previous pointers. =
=A0*/
>> =A0 unsigned same_body_alias : 1;
>> @@ -475,6 +477,8 @@ struct GTY((chain_next ("%h.next"), chai
>> =A0 int order;
>> =A0 enum ld_plugin_symbol_resolution resolution;
>>
>> + =A0/* The module in which it is first declared. =A0*/
>> + =A0unsigned module_id;
>> =A0 /* Set when function must be output - it is externally visible
>> =A0 =A0 =A0or its address is taken. =A0*/
>> =A0 unsigned needed : 1;
>> @@ -547,6 +551,11 @@ void debug_cgraph_node (struct cgraph_no
>> =A0void cgraph_insert_node_to_hashtable (struct cgraph_node *node);
>> =A0void cgraph_remove_edge (struct cgraph_edge *);
>> =A0void cgraph_remove_node (struct cgraph_node *);
>> +void cgraph_add_assembler_hash_node (struct cgraph_node *);
>> +void cgraph_remove_assembler_hash_node (struct cgraph_node *);
>> +void cgraph_remove_fake_indirect_call_in_edges (struct cgraph_node *);
>> +extern bool cgraph_pre_profiling_inlining_done;
>> +extern bool cgraph_is_fake_indirect_call_edge (struct cgraph_edge *e);
>> =A0void cgraph_remove_node_and_inline_clones (struct cgraph_node *);
>> =A0void cgraph_release_function_body (struct cgraph_node *);
>> =A0void cgraph_node_remove_callees (struct cgraph_node *node);
>> @@ -641,6 +650,48 @@ struct cgraph_node *save_inline_function
>> =A0void record_references_in_initializer (tree, bool);
>> =A0bool cgraph_process_new_functions (void);
>>
>> +/* Module info structure. =A0*/
>> +struct GTY (()) cgraph_mod_info
>> +{
>> + =A0unsigned module_id;
>> +};
>> +
>> +/* LIPO linker symbol table entry for function symbols. =A0*/
>> +struct GTY (()) cgraph_sym
>> +{
>> + =A0tree assembler_name;
>> + =A0struct cgraph_node *rep_node;
>> + =A0tree rep_decl;
>> + =A0htab_t GTY ((param_is (struct cgraph_mod_info))) def_module_hash;
>> + =A0bool is_promoted_static;
>> +};
>> +
>> +void cgraph_init_gid_map (void);
>> +void cgraph_add_fake_indirect_call_edges (void);
>> +void cgraph_remove_zero_count_fake_edges (void);
>> +void cgraph_do_link (void);
>> +struct cgraph_sym *cgraph_link_node (struct cgraph_node *);
>> +tree cgraph_find_decl (tree asm_name);
>> +void cgraph_remove_link_node (struct cgraph_node *node);
>> +struct cgraph_node *cgraph_lipo_get_resolved_node (tree decl);
>> +struct cgraph_node *cgraph_lipo_get_resolved_node_1 (tree decl, bool);
>> +unsigned =A0cgraph_get_module_id (tree fndecl);
>> +bool cgraph_is_auxiliary (tree fndecl);
>> +void cgraph_process_module_scope_statics (void);
>> +bool cgraph_is_promoted_static_func (tree fndecl);
>> +bool cgraph_is_inline_body_available_in_module (tree fndecl, unsigned m=
odule_id);
>> +bool cgraph_is_aux_decl_external (struct cgraph_node *);
>> +void cgraph_unify_type_alias_sets (void);
>> +void varpool_do_link (void);
>> +void varpool_link_node (struct varpool_node *);
>> +void varpool_remove_link_node (struct varpool_node *node);
>> +struct varpool_node *real_varpool_node (tree decl);
>> +bool varpool_is_auxiliary (struct varpool_node *node);
>> +void varpool_get_referenced_asm_ids (VEC(tree, gc) **);
>> +void varpool_clear_asm_id_reference_bit (void);
>> +void varpool_reset_queue (void);
>> +void varpool_remove_duplicate_weak_decls (void);
>> +
>> =A0bool cgraph_decide_is_function_needed (struct cgraph_node *, tree);
>>
>> =A0typedef void (*cgraph_edge_hook)(struct cgraph_edge *, void *);
>> Index: value-prof.c
>> =3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=
=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=
=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D
>> --- value-prof.c =A0 =A0 =A0 =A0(revision 172880)
>> +++ value-prof.c =A0 =A0 =A0 =A0(working copy)
>> @@ -46,6 +46,9 @@ along with GCC; see the file COPYING3.
>> =A0#include "timevar.h"
>> =A0#include "tree-pass.h"
>> =A0#include "pointer-set.h"
>> +#include "langhooks.h"
>> +#include "params.h"
>> +#include "l-ipo.h"
>>
>> =A0/* In this file value profile based optimizations are placed. =A0Curr=
ently the
>> =A0 =A0following optimizations are implemented (for more detailed descri=
ptions
>> @@ -300,6 +303,10 @@ dump_histogram_value (FILE *dump_file, h
>> =A0 =A0 =A0 =A0}
>> =A0 =A0 =A0 fprintf (dump_file, ".\n");
>> =A0 =A0 =A0 break;
>> + =A0 =A0case HIST_TYPE_INDIR_CALL_TOPN:
>> + =A0 =A0 =A0fprintf (dump_file, "Indirect call -- top N\n");
>> + =A0 =A0 =A0/* TODO add more elaborate dumping code. =A0*/
>> + =A0 =A0 =A0break;
>> =A0 =A0}
>> =A0}
>>
>> @@ -487,6 +494,70 @@ check_counter (gimple stmt, const char *
>> =A0 return false;
>> =A0}
>>
>> +/* The overall number of invocations of the counter should match
>> + =A0 execution count of basic block. =A0Report it as error rather than
>> + =A0 internal error as it might mean that user has misused the profile
>> + =A0 somehow. =A0STMT is the indiret call, COUNT1 and COUNT2 are counts
>> + =A0 of two top targets, and ALL is the enclosing basic block execution
>> + =A0 count. =A0*/
>> +
>> +static bool
>> +check_ic_counter (gimple stmt, gcov_type *count1, gcov_type *count2,
>> + =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0gcov_type all)
>> +{
>> + =A0location_t locus;
>> + =A0if (*count1 > all && flag_profile_correction)
>> + =A0 =A0{
>> + =A0 =A0 =A0locus =3D (stmt !=3D NULL)
>> + =A0 =A0 =A0 =A0 =A0 =A0 =A0? gimple_location (stmt)
>> + =A0 =A0 =A0 =A0 =A0 =A0 =A0: DECL_SOURCE_LOCATION (current_function_de=
cl);
>> + =A0 =A0 =A0inform (locus, "Correcting inconsistent value profile: "
>> + =A0 =A0 =A0 =A0 =A0 =A0 =A0"ic (topn) profiler top target count (%ld) =
exceeds "
>> + =A0 =A0 =A0 =A0 =A0 =A0 "BB count (%ld)", (long)*count1, (long)all);
>> + =A0 =A0 =A0*count1 =3D all;
>> + =A0 =A0}
>> + =A0if (*count2 > all && flag_profile_correction)
>> + =A0 =A0{
>> + =A0 =A0 =A0locus =3D (stmt !=3D NULL)
>> + =A0 =A0 =A0 =A0 =A0 =A0 =A0? gimple_location (stmt)
>> + =A0 =A0 =A0 =A0 =A0 =A0 =A0: DECL_SOURCE_LOCATION (current_function_de=
cl);
>> + =A0 =A0 =A0inform (locus, "Correcting inconsistent value profile: "
>> + =A0 =A0 =A0 =A0 =A0 =A0 =A0"ic (topn) profiler second target count (%l=
d) exceeds "
>> + =A0 =A0 =A0 =A0 =A0 =A0 "BB count (%ld)", (long)*count2, (long)all);
>> + =A0 =A0 =A0*count2 =3D all;
>> + =A0 =A0}
>> +
>> + =A0if (*count2 > *count1)
>> + =A0 =A0{
>> + =A0 =A0 =A0locus =3D (stmt !=3D NULL)
>> + =A0 =A0 =A0 =A0 =A0 =A0 =A0? gimple_location (stmt)
>> + =A0 =A0 =A0 =A0 =A0 =A0 =A0: DECL_SOURCE_LOCATION (current_function_de=
cl);
>> + =A0 =A0 =A0inform (locus, "Corrupted topn ic value profile: "
>> + =A0 =A0 =A0 =A0 =A0 =A0 "first target count (%ld) is less than the sec=
ond "
>> + =A0 =A0 =A0 =A0 =A0 =A0 "target count (%ld)", (long)*count1, (long)*co=
unt2);
>> + =A0 =A0 =A0return true;
>> + =A0 =A0}
>> +
>> + =A0if (*count1 + *count2 > all)
>> + =A0 =A0{
>> + =A0 =A0 =A0/* If (COUNT1 + COUNT2) is greater than ALL by less than ar=
ound 10% then
>> + =A0 =A0 =A0 =A0just fix COUNT2 up so that (COUNT1 + COUNT2) equals ALL=
. =A0*/
>> + =A0 =A0 =A0if ((*count1 + *count2 - all) < (all >> 3))
>> + =A0 =A0 =A0 *count2 =3D all - *count1;
>> + =A0 =A0 =A0else
>> + =A0 =A0 =A0 {
>> + =A0 =A0 =A0 =A0 locus =3D (stmt !=3D NULL)
>> + =A0 =A0 =A0 =A0 =A0 ? gimple_location (stmt)
>> + =A0 =A0 =A0 =A0 =A0 : DECL_SOURCE_LOCATION (current_function_decl);
>> + =A0 =A0 =A0 =A0 inform (locus, "Corrupted topn ic value profile: top t=
wo targets's"
>> + =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 " total count (%ld) exceeds bb count (=
%ld)",
>> + =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 (long)(*count1 + *count2), (long)all);
>> + =A0 =A0 =A0 =A0 return true;
>> + =A0 =A0 =A0 }
>> + =A0 =A0}
>> + =A0return false;
>> +}
>> +
>>
>> =A0/* GIMPLE based transformations. */
>>
>> @@ -542,6 +613,14 @@ gimple_value_profile_transformations (vo
>> =A0 if (changed)
>> =A0 =A0 {
>> =A0 =A0 =A0 counts_to_freqs ();
>> + =A0 =A0 =A0/* Value profile transformations may change inline paramete=
rs
>> + =A0 =A0 =A0 =A0 a lot (e.g., indirect call promotion introduces new di=
rect calls).
>> + =A0 =A0 =A0 =A0 The update is also needed to avoid compiler ICE -- whe=
n MULTI
>> + =A0 =A0 =A0 =A0 target icall promotion happens, the caller's size may =
become
>> + =A0 =A0 =A0 =A0 negative when the promoted direct calls get promoted. =
=A0*/
>> + =A0 =A0 =A0/* Guard this for LIPO for now. =A0*/
>> + =A0 =A0 =A0if (L_IPO_COMP_MODE)
>> + =A0 =A0 =A0 =A0compute_inline_parameters (cgraph_node (current_functio=
n_decl));
>> =A0 =A0 }
>>
>> =A0 return changed;
>> @@ -1090,6 +1169,111 @@ find_func_by_pid (int =A0 pid)
>> =A0 return pid_map [pid];
>> =A0}
>>
>> +/* Initialize map of gids (gid -> cgraph node) */
>> +
>> +static htab_t gid_map =3D NULL;
>> +
>> +typedef struct func_gid_entry
>> +{
>> + =A0struct cgraph_node *node;
>> + =A0unsigned HOST_WIDEST_INT gid;
>> +} func_gid_entry_t;
>> +
>> +/* Hash function for function global unique ids. =A0*/
>> +
>> +static hashval_t
>> +htab_gid_hash (const void * ent)
>> +{
>> + =A0const func_gid_entry_t *const entry =3D (const func_gid_entry_t *) =
ent;
>> + =A0return entry->gid;
>> +}
>> +
>> +/* Hash table equality function for function global unique ids. =A0*/
>> +
>> +static int
>> +htab_gid_eq (const void *ent1, const void * ent2)
>> +{
>> + =A0const func_gid_entry_t *const entry1 =3D (const func_gid_entry_t *)=
 ent1;
>> + =A0const func_gid_entry_t *const entry2 =3D (const func_gid_entry_t *)=
 ent2;
>> + =A0return entry1->gid =3D=3D entry2->gid;
>> +}
>> +
>> +static void
>> +htab_gid_del (void *ent)
>> +{
>> + =A0func_gid_entry_t *const entry =3D (func_gid_entry_t *) ent;
>> + =A0free (entry);
>> +}
>> +
>> +/* Initialize the global unique id map for functions. =A0*/
>> +
>> +static void
>> +init_gid_map (void)
>> +{
>> + =A0struct cgraph_node *n;
>> +
>> + =A0gcc_assert (!gid_map);
>> +
>> + =A0gid_map
>> + =A0 =A0 =A0=3D htab_create (10, htab_gid_hash, htab_gid_eq, htab_gid_d=
el);
>> +
>> + =A0for (n =3D cgraph_nodes; n; n =3D n->next)
>> + =A0 =A0{
>> + =A0 =A0 =A0func_gid_entry_t ent, *entp;
>> + =A0 =A0 =A0func_gid_entry_t **slot;
>> + =A0 =A0 =A0struct function *f;
>> + =A0 =A0 =A0ent.node =3D n;
>> + =A0 =A0 =A0f =3D DECL_STRUCT_FUNCTION (n->decl);
>> + =A0 =A0 =A0/* Do not care to indirect call promote a function with id.=
 =A0*/
>> + =A0 =A0 =A0if (!f || DECL_ABSTRACT (n->decl))
>> + =A0 =A0 =A0 =A0continue;
>> + =A0 =A0 =A0/* The global function id computed at profile-use time
>> + =A0 =A0 =A0 =A0 is slightly different from the one computed in
>> + =A0 =A0 =A0 =A0 instrumentation runtime -- for the latter, the intra-
>> + =A0 =A0 =A0 =A0 module function ident is 1 based while in profile-use
>> + =A0 =A0 =A0 =A0 phase, it is zero based. See get_next_funcdef_no in
>> + =A0 =A0 =A0 =A0 function.c. =A0*/
>> + =A0 =A0 =A0ent.gid =3D FUNC_DECL_GLOBAL_ID (DECL_STRUCT_FUNCTION (n->d=
ecl));
>> + =A0 =A0 =A0slot =3D (func_gid_entry_t **) htab_find_slot (gid_map, &en=
t, INSERT);
>> +
>> + =A0 =A0 =A0gcc_assert (!*slot || ((*slot)->gid =3D=3D ent.gid && (*slo=
t)->node =3D=3D n));
>> + =A0 =A0 =A0if (!*slot)
>> + =A0 =A0 =A0 =A0{
>> + =A0 =A0 =A0 =A0 =A0*slot =3D entp =3D XCNEW (func_gid_entry_t);
>> + =A0 =A0 =A0 =A0 =A0entp->node =3D n;
>> + =A0 =A0 =A0 =A0 =A0entp->gid =3D ent.gid;
>> + =A0 =A0 =A0 =A0}
>> + =A0 =A0}
>> +}
>> +
>> +/* Initialize the global unique id map for functions. =A0*/
>> +
>> +void
>> +cgraph_init_gid_map (void)
>> +{
>> + =A0if (!L_IPO_COMP_MODE)
>> + =A0 =A0return;
>> +
>> + =A0init_gid_map ();
>> +}
>> +
>> +/* Return cgraph node for function with global id. =A0*/
>> +
>> +struct cgraph_node *
>> +find_func_by_global_id (unsigned HOST_WIDE_INT gid)
>> +{
>> + =A0func_gid_entry_t ent, *entp;
>> +
>> + =A0gcc_assert (gid_map);
>> +
>> + =A0ent.node =3D NULL;
>> + =A0ent.gid =3D gid;
>> + =A0entp =3D (func_gid_entry_t *)htab_find (gid_map, &ent);
>> + =A0if (entp)
>> + =A0 =A0return entp->node;
>> + =A0return NULL;
>> +}
>> +
>> =A0/* Perform sanity check on the indirect call target. Due to race cond=
itions,
>> =A0 =A0false function target may be attributed to an indirect call site.=
 If the
>> =A0 =A0call expression type mismatches with the target function's type, =
expand_call
>> @@ -1244,27 +1428,13 @@ gimple_ic (gimple icall_stmt, struct cgr
>> =A0*/
>>
>> =A0static bool
>> -gimple_ic_transform (gimple stmt)
>> +gimple_ic_transform_single_targ (gimple stmt, histogram_value histogram)
>> =A0{
>> - =A0histogram_value histogram;
>> =A0 gcov_type val, count, all, bb_all;
>> =A0 gcov_type prob;
>> - =A0tree callee;
>> =A0 gimple modify;
>> =A0 struct cgraph_node *direct_call;
>>
>> - =A0if (gimple_code (stmt) !=3D GIMPLE_CALL)
>> - =A0 =A0return false;
>> -
>> - =A0callee =3D gimple_call_fn (stmt);
>> -
>> - =A0if (TREE_CODE (callee) =3D=3D FUNCTION_DECL)
>> - =A0 =A0return false;
>> -
>> - =A0histogram =3D gimple_histogram_value_of_type (cfun, stmt, HIST_TYPE=
_INDIR_CALL);
>> - =A0if (!histogram)
>> - =A0 =A0return false;
>> -
>> =A0 val =3D histogram->hvalue.counters [0];
>> =A0 count =3D histogram->hvalue.counters [1];
>> =A0 all =3D histogram->hvalue.counters [2];
>> @@ -1312,6 +1482,195 @@ gimple_ic_transform (gimple stmt)
>> =A0 return true;
>> =A0}
>>
>> +/* Convert indirect function call STMT into guarded direct function
>> + =A0 calls. Multiple indirect call targets are supported. HISTOGRAM
>> + =A0 is the target distribution for the callsite. =A0*/
>> +
>> +static bool
>> +gimple_ic_transform_mult_targ (gimple stmt, histogram_value histogram)
>> +{
>> + =A0gcov_type val1, val2, count1, count2, all, bb_all;
>> + =A0gcov_type prob1, prob2;
>> + =A0gimple modify1, modify2;
>> + =A0struct cgraph_node *direct_call1 =3D 0, *direct_call2 =3D 0;
>> + =A0int perc_threshold, count_threshold, always_inline;
>> + =A0location_t locus;
>> +
>> + =A0val1 =3D histogram->hvalue.counters [1];
>> + =A0count1 =3D histogram->hvalue.counters [2];
>> + =A0val2 =3D histogram->hvalue.counters [3];
>> + =A0count2 =3D histogram->hvalue.counters [4];
>> + =A0bb_all =3D gimple_bb (stmt)->count;
>> + =A0all =3D bb_all;
>> +
>> + =A0gimple_remove_histogram_value (cfun, stmt, histogram);
>> +
>> + =A0if (count1 =3D=3D 0)
>> + =A0 =A0return false;
>> +
>> + =A0perc_threshold =3D PARAM_VALUE (PARAM_ICALL_PROMOTE_PERCENT_THRESHO=
LD);
>> + =A0count_threshold =3D PARAM_VALUE (PARAM_ICALL_PROMOTE_COUNT_THRESHOL=
D);
>> + =A0always_inline =3D PARAM_VALUE (PARAM_ALWAYS_INLINE_ICALL_TARGET);
>> +
>> + =A0if (100 * count1 < all * perc_threshold || count1 < count_threshold)
>> + =A0 =A0return false;
>> +
>> + =A0if (check_ic_counter (stmt, &count1, &count2, all))
>> + =A0 =A0return false;
>> +
>> + =A0if (all > 0)
>> + =A0 =A0{
>> + =A0 =A0 =A0prob1 =3D (count1 * REG_BR_PROB_BASE + all / 2) / all;
>> + =A0 =A0 =A0if (all - count1 > 0)
>> + =A0 =A0 =A0 =A0prob2 =3D (count2 * REG_BR_PROB_BASE
>> + =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 + (all - count1) / 2) / (all - count1);
>> + =A0 =A0 =A0else
>> + =A0 =A0 =A0 =A0prob2 =3D 0;
>> + =A0 =A0}
>> + =A0else
>> + =A0 =A0prob1 =3D prob2 =3D 0;
>> +
>> + =A0direct_call1 =3D find_func_by_global_id (val1);
>> +
>> + =A0if (val2 && (100 * count2 >=3D all * perc_threshold)
>> + =A0 =A0 =A0&& count2 > count_threshold)
>> + =A0 =A0direct_call2 =3D find_func_by_global_id (val2);
>> +
>> + =A0locus =3D (stmt !=3D NULL) ? gimple_location (stmt)
>> + =A0 =A0 =A0: DECL_SOURCE_LOCATION (current_function_decl);
>> + =A0if (direct_call1 =3D=3D NULL
>> + =A0 =A0 =A0|| !check_ic_target (stmt, direct_call1))
>> + =A0 =A0{
>> + =A0 =A0 =A0if (flag_ripa_verbose)
>> + =A0 =A0 =A0 =A0{
>> + =A0 =A0 =A0 =A0 =A0if (!direct_call1)
>> + =A0 =A0 =A0 =A0 =A0 =A0inform (locus, "Can not find indirect call targ=
et decl "
>> + =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0"(%d:%d)[cnt:%u] in current mod=
ule",
>> + =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0EXTRACT_MODULE_ID_FROM_GLOBAL_I=
D (val1),
>> + =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0EXTRACT_FUNC_ID_FROM_GLOBAL_ID =
(val1), (unsigned) count1);
>> + =A0 =A0 =A0 =A0 =A0else
>> + =A0 =A0 =A0 =A0 =A0 =A0inform (locus,
>> + =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0"Can not find promote indirect =
call target decl -- type mismatch "
>> + =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0"(%d:%d)[cnt:%u] in current mod=
ule",
>> + =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0EXTRACT_MODULE_ID_FROM_GLOBAL_I=
D (val1),
>> + =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0EXTRACT_FUNC_ID_FROM_GLOBAL_ID =
(val1), (unsigned) count1);
>> + =A0 =A0 =A0 =A0}
>> + =A0 =A0 =A0return false;
>> + =A0 =A0}
>> +
>> + =A0/* Don't indirect-call promote if the target is in auxiliary module=
 and
>> + =A0 =A0 DECL_ARTIFICIAL and not TREE_PUBLIC, because we don't static-p=
romote
>> + =A0 =A0 DECL_ARTIFICIALs yet. =A0*/
>> + =A0if (cgraph_is_auxiliary (direct_call1->decl)
>> + =A0 =A0 =A0&& DECL_ARTIFICIAL (direct_call1->decl)
>> + =A0 =A0 =A0&& ! TREE_PUBLIC (direct_call1->decl))
>> + =A0 =A0return false;
>> +
>> + =A0modify1 =3D gimple_ic (stmt, direct_call1, prob1, count