[google] Patch to add compiler flag to dump callgraph edge profiles in special .note sections (issue4591045)

Sriraman Tallam tmsriram@google.com
Wed Jun 8 16:41:00 GMT 2011


On Wed, Jun 8, 2011 at 9:16 AM, Xinliang David Li <davidxl@google.com> wrote:
> ok for google/main.

Thanks, the patch is now committed.

>
> David
>
> On Wed, Jun 8, 2011 at 9:13 AM, Sriraman Tallam <tmsriram@google.com> wrote:
>> +davidxl
>>
>> On Tue, Jun 7, 2011 at 7:05 PM, Sriraman Tallam <tmsriram@google.com> wrote:
>>> Patch Description:
>>> =================
>>>
>>> I am working on a project to do global function layout in the linker where the linker reads the callgraph edge profile information, generated by FDO, and uses that to find a ordering of functions that will place functions calling each other frequently closer, like the Pettis-Hansen code ordering algorithm described in the paper "Profile-guided Code Poisitioning" in PLDI 1990.
>>>
>>> This patch adds a flag that allows the callgraph edge profile information to be stored .note sections called ".note.callgraph.text". The new compiler flag -fcallgraph-profiles-sections generates these sections and must be used along with -fprofile-use. I have added a PARAM to only output callgraph edges greater than a specified threshold. Once this is available, the linker can read these sections and generate a global callgraph which can be used to determine a global function ordering.
>>>
>>> I am adding plugin support in the gold linker to allow linker plugins to be able to read the contents of sections and also adding plugin hooks to specify a desired ordering of functions to the linker. The linker patch is available here : http://sourceware.org/ml/binutils/2011-03/msg00043.html. Once this is available, linker plugins can be used to determine the function layout, like the Pettis-Hansen algorithm, of the final binary.
>>>
>>> Example: The new .note.callgraph.text sections looks like this for a function foo that calls bar 100 times and zap 50 times:
>>> ****************************
>>> .section        .note.callgraph.text._Z3foov,"",@progbits
>>>        .string "Function _Z3foov"
>>>        .string "_Z3barv"
>>>        .string "100"
>>>        .string "_Z3zapv"
>>>        .string "50"
>>> ***************************
>>>
>>> For now, this is for google/main. I will re-submit for review to trunk along with data layout.
>>>
>>> Google ref 41940
>>>
>>> 2011-06-07  Sriraman Tallam  <tmsriram@google.com>
>>>
>>>        * doc/invoke.texi: document option -fcallgraph-profiles-sections.
>>>        * final.c  (dump_cgraph_profiles): New function.
>>>        (rest_of_handle_final): Create new section '.note.callgraph.text'
>>>        with compiler flag -fcallgraph-profiles-sections
>>>        * common.opt: New option -fcallgraph-profiles-sections.
>>>        * params.def (DEFPARAM): New param
>>>        PARAM_NOTE_CGRAPH_SECTION_EDGE_THRESHOLD.
>>>
>>> Index: doc/invoke.texi
>>> ===================================================================
>>> --- doc/invoke.texi     (revision 174789)
>>> +++ doc/invoke.texi     (working copy)
>>> @@ -351,7 +351,7 @@ Objective-C and Objective-C++ Dialects}.
>>>  -falign-labels[=@var{n}] -falign-loops[=@var{n}] -fassociative-math @gol
>>>  -fauto-inc-dec -fbranch-probabilities -fbranch-target-load-optimize @gol
>>>  -fbranch-target-load-optimize2 -fbtr-bb-exclusive -fcaller-saves @gol
>>> --fcheck-data-deps -fclone-hot-version-paths @gol
>>> +-fcallgraph-profiles-sections -fcheck-data-deps -fclone-hot-version-paths @gol
>>>  -fcombine-stack-adjustments -fconserve-stack @gol
>>>  -fcompare-elim -fcprop-registers -fcrossjumping @gol
>>>  -fcse-follow-jumps -fcse-skip-blocks -fcx-fortran-rules @gol
>>> @@ -8114,6 +8114,15 @@ Do not promote static functions with always inline
>>>  @opindex fripa-verbose
>>>  Enable printing of verbose information about dynamic inter-procedural optimizations.
>>>  This is used in conjunction with the @option{-fripa}.
>>> +
>>> +@item -fcallgraph-profiles-sections
>>> +@opindex fcallgraph-profiles-sections
>>> +Emit call graph edge profile counts in .note.callgraph.text sections. This is
>>> +used in conjunction with @option{-fprofile-use}. A new .note.callgraph.text
>>> +section is created for each function. This section lists every callee and the
>>> +number of times it is called. The params variable
>>> +"note-cgraph-section-edge-threshold" can be used to only list edges above a
>>> +certain threshold.
>>>  @end table
>>>
>>>  The following options control compiler behavior regarding floating
>>> Index: final.c
>>> ===================================================================
>>> --- final.c     (revision 174789)
>>> +++ final.c     (working copy)
>>> @@ -4321,13 +4321,37 @@ debug_free_queue (void)
>>>       symbol_queue_size = 0;
>>>     }
>>>  }
>>> -
>>> +
>>> +/* List the call graph profiled edges whise value is greater than
>>> +   PARAM_NOTE_CGRAPH_SECTION_EDGE_THRESHOLD in the
>>> +   ".note.callgraph.text" section. */
>>> +static void
>>> +dump_cgraph_profiles (void)
>>> +{
>>> +  struct cgraph_node *node = cgraph_node (current_function_decl);
>>> +  struct cgraph_edge *e;
>>> +  struct cgraph_node *callee;
>>> +
>>> +  for (e = node->callees; e != NULL; e = e->next_callee)
>>> +    {
>>> +      if (e->count <= PARAM_VALUE (PARAM_NOTE_CGRAPH_SECTION_EDGE_THRESHOLD))
>>> +        continue;
>>> +      callee = e->callee;
>>> +      fprintf (asm_out_file, "\t.string \"%s\"\n",
>>> +               IDENTIFIER_POINTER (decl_assembler_name (callee->decl)));
>>> +      fprintf (asm_out_file, "\t.string \"" HOST_WIDEST_INT_PRINT_DEC "\"\n",
>>> +               e->count);
>>> +    }
>>> +}
>>> +
>>>  /* Turn the RTL into assembly.  */
>>>  static unsigned int
>>>  rest_of_handle_final (void)
>>>  {
>>>   rtx x;
>>>   const char *fnname;
>>> +  char *profile_fnname;
>>> +  unsigned int flags;
>>>
>>>   /* Get the function's name, as described by its RTL.  This may be
>>>      different from the DECL_NAME name used in the source file.  */
>>> @@ -4387,6 +4411,21 @@ rest_of_handle_final (void)
>>>     targetm.asm_out.destructor (XEXP (DECL_RTL (current_function_decl), 0),
>>>                                decl_fini_priority_lookup
>>>                                  (current_function_decl));
>>> +
>>> +  /* With -fcgraph-section, add ".note.callgraph.text" section for storing
>>> +     profiling information. */
>>> +  if (flag_callgraph_profiles_sections
>>> +      && flag_profile_use
>>> +      && cgraph_node (current_function_decl) != NULL)
>>> +    {
>>> +      flags = SECTION_DEBUG;
>>> +      asprintf (&profile_fnname, ".note.callgraph.text.%s", fnname);
>>> +      switch_to_section (get_section (profile_fnname, flags, NULL));
>>> +      fprintf (asm_out_file, "\t.string \"Function %s\"\n", fnname);
>>> +      dump_cgraph_profiles ();
>>> +      free (profile_fnname);
>>> +    }
>>> +
>>>   return 0;
>>>  }
>>>
>>> Index: common.opt
>>> ===================================================================
>>> --- common.opt  (revision 174789)
>>> +++ common.opt  (working copy)
>>> @@ -907,6 +907,10 @@ fcaller-saves
>>>  Common Report Var(flag_caller_saves) Optimization
>>>  Save registers around function calls
>>>
>>> +fcallgraph-profiles-sections
>>> +Common Report Var(flag_callgraph_profiles_sections) Init(0)
>>> +Generate .note.callgraph.text sections listing callees and edge counts.
>>> +
>>>  fcheck-data-deps
>>>  Common Report Var(flag_check_data_deps)
>>>  Compare the results of several data dependence analyzers.
>>> Index: params.def
>>> ===================================================================
>>> --- params.def  (revision 174789)
>>> +++ params.def  (working copy)
>>> @@ -1002,6 +1002,15 @@ DEFPARAM (PARAM_MVERSN_CLONE_CGRAPH_DEPTH,
>>>          "maximum length of the call graph path to be cloned "
>>>           "while doing multiversioning",
>>>          2, 0, 5)
>>> +
>>> +/* Only output those call graph edges in .note.callgraph.text sections
>>> +   whose count is greater than this value. */
>>> +DEFPARAM (PARAM_NOTE_CGRAPH_SECTION_EDGE_THRESHOLD,
>>> +         "note-cgraph-section-edge-threshold",
>>> +         "minimum call graph edge count for inclusion in "
>>> +          ".note.callgraph.text section",
>>> +         0, 0, 0)
>>> +
>>>  /*
>>>  Local variables:
>>>  mode:c
>>>
>>> --
>>> This patch is available for review at http://codereview.appspot.com/4591045
>>>
>>
>



More information about the Gcc-patches mailing list