1 /* Read and annotate call graph profile from the auto profile data file.
2 Copyright (C) 2014-2021 Free Software Foundation, Inc.
3 Contributed by Dehao Chen (dehao@google.com)
5 This file is part of GCC.
7 GCC is free software; you can redistribute it and/or modify it under
8 the terms of the GNU General Public License as published by the Free
9 Software Foundation; either version 3, or (at your option) any later
12 GCC is distributed in the hope that it will be useful, but WITHOUT ANY
13 WARRANTY; without even the implied warranty of MERCHANTABILITY or
14 FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
17 You should have received a copy of the GNU General Public License
18 along with GCC; see the file COPYING3. If not see
19 <http://www.gnu.org/licenses/>. */
25 #include "coretypes.h"
30 #include "alloc-pool.h"
31 #include "tree-pass.h"
35 #include "diagnostic-core.h"
37 #include "langhooks.h"
40 #include "tree-cfgcleanup.h"
41 #include "tree-into-ssa.h"
42 #include "gimple-iterator.h"
43 #include "value-prof.h"
44 #include "symbol-summary.h"
46 #include "ipa-fnsummary.h"
47 #include "ipa-inline.h"
48 #include "tree-inline.h"
49 #include "auto-profile.h"
50 #include "tree-pretty-print.h"
51 #include "gimple-pretty-print.h"
53 /* The following routines implements AutoFDO optimization.
55 This optimization uses sampling profiles to annotate basic block counts
56 and uses heuristics to estimate branch probabilities.
58 There are three phases in AutoFDO:
60 Phase 1: Read profile from the profile data file.
61 The following info is read from the profile datafile:
62 * string_table: a map between function name and its index.
63 * autofdo_source_profile: a map from function_instance name to
64 function_instance. This is represented as a forest of
66 * WorkingSet: a histogram of how many instructions are covered for a
67 given percentage of total cycles. This is describing the binary
68 level information (not source level). This info is used to help
69 decide if we want aggressive optimizations that could increase
70 code footprint (e.g. loop unroll etc.)
71 A function instance is an instance of function that could either be a
72 standalone symbol, or a clone of a function that is inlined into another
75 Phase 2: Early inline + value profile transformation.
76 Early inline uses autofdo_source_profile to find if a callsite is:
77 * inlined in the profiled binary.
78 * callee body is hot in the profiling run.
79 If both condition satisfies, early inline will inline the callsite
80 regardless of the code growth.
81 Phase 2 is an iterative process. During each iteration, we also check
82 if an indirect callsite is promoted and inlined in the profiling run.
83 If yes, vpt will happen to force promote it and in the next iteration,
84 einline will inline the promoted callsite in the next iteration.
86 Phase 3: Annotate control flow graph.
87 AutoFDO uses a separate pass to:
88 * Annotate basic block count
89 * Estimate branch probability
91 After the above 3 phases, all profile is readily annotated on the GCC IR.
92 AutoFDO tries to reuse all FDO infrastructure as much as possible to make
93 use of the profile. E.g. it uses existing mechanism to calculate the basic
94 block/edge frequency, as well as the cgraph node/edge count.
97 #define DEFAULT_AUTO_PROFILE_FILE "fbdata.afdo"
98 #define AUTO_PROFILE_VERSION 2
103 /* Intermediate edge info used when propagating AutoFDO profile information.
104 We can't edge->count() directly since it's computed from edge's probability
105 while probability is yet not decided during propagation. */
106 #define AFDO_EINFO(e) ((class edge_info *) e->aux)
110 edge_info () : count_ (profile_count::zero ().afdo ()), annotated_ (false) {}
111 bool is_annotated () const { return annotated_
; }
112 void set_annotated () { annotated_
= true; }
113 profile_count
get_count () const { return count_
; }
114 void set_count (profile_count count
) { count_
= count
; }
116 profile_count count_
;
120 /* Represent a source location: (function_decl, lineno). */
121 typedef std::pair
<tree
, unsigned> decl_lineno
;
123 /* Represent an inline stack. vector[0] is the leaf node. */
124 typedef auto_vec
<decl_lineno
> inline_stack
;
126 /* String array that stores function names. */
127 typedef auto_vec
<char *> string_vector
;
129 /* Map from function name's index in string_table to target's
131 typedef std::map
<unsigned, gcov_type
> icall_target_map
;
133 /* Set of gimple stmts. Used to track if the stmt has already been promoted
135 typedef std::set
<gimple
*> stmt_set
;
137 /* Represent count info of an inline stack. */
141 /* Sampled count of the inline stack. */
144 /* Map from indirect call target to its sample count. */
145 icall_target_map targets
;
147 /* Whether this inline stack is already used in annotation.
149 Each inline stack should only be used to annotate IR once.
150 This will be enforced when instruction-level discriminator
155 /* operator< for "const char *". */
156 struct string_compare
158 bool operator()(const char *a
, const char *b
) const
160 return strcmp (a
, b
) < 0;
164 /* Store a string array, indexed by string position in the array. */
173 /* For a given string, returns its index. */
174 int get_index (const char *name
) const;
176 /* For a given decl, returns the index of the decl name. */
177 int get_index_by_decl (tree decl
) const;
179 /* For a given index, returns the string. */
180 const char *get_name (int index
) const;
182 /* Read profile, return TRUE on success. */
186 typedef std::map
<const char *, unsigned, string_compare
> string_index_map
;
187 string_vector vector_
;
188 string_index_map map_
;
191 /* Profile of a function instance:
192 1. total_count of the function.
193 2. head_count (entry basic block count) of the function (only valid when
194 function is a top-level function_instance, i.e. it is the original copy
195 instead of the inlined copy).
196 3. map from source location (decl_lineno) to profile (count_info).
197 4. map from callsite to callee function_instance. */
198 class function_instance
201 typedef auto_vec
<function_instance
*> function_instance_stack
;
203 /* Read the profile and return a function_instance with head count as
204 HEAD_COUNT. Recursively read callsites to create nested function_instances
205 too. STACK is used to track the recursive creation process. */
206 static function_instance
*
207 read_function_instance (function_instance_stack
*stack
,
208 gcov_type head_count
);
210 /* Recursively deallocate all callsites (nested function_instances). */
211 ~function_instance ();
230 /* Traverse callsites of the current function_instance to find one at the
231 location of LINENO and callee name represented in DECL. */
232 function_instance
*get_function_instance_by_decl (unsigned lineno
,
235 /* Store the profile info for LOC in INFO. Return TRUE if profile info
237 bool get_count_info (location_t loc
, count_info
*info
) const;
239 /* Read the inlined indirect call target profile for STMT and store it in
240 MAP, return the total count for all inlined indirect calls. */
241 gcov_type
find_icall_target_map (gcall
*stmt
, icall_target_map
*map
) const;
243 /* Sum of counts that is used during annotation. */
244 gcov_type
total_annotated_count () const;
246 /* Mark LOC as annotated. */
247 void mark_annotated (location_t loc
);
250 /* Callsite, represented as (decl_lineno, callee_function_name_index). */
251 typedef std::pair
<unsigned, unsigned> callsite
;
253 /* Map from callsite to callee function_instance. */
254 typedef std::map
<callsite
, function_instance
*> callsite_map
;
256 function_instance (unsigned name
, gcov_type head_count
)
257 : name_ (name
), total_count_ (0), head_count_ (head_count
)
261 /* Map from source location (decl_lineno) to profile (count_info). */
262 typedef std::map
<unsigned, count_info
> position_count_map
;
264 /* function_instance name index in the string_table. */
267 /* Total sample count. */
268 gcov_type total_count_
;
270 /* Entry BB's sample count. */
271 gcov_type head_count_
;
273 /* Map from callsite location to callee function_instance. */
274 callsite_map callsites
;
276 /* Map from source location to count_info. */
277 position_count_map pos_counts
;
280 /* Profile for all functions. */
281 class autofdo_source_profile
284 static autofdo_source_profile
*
287 autofdo_source_profile
*map
= new autofdo_source_profile ();
295 ~autofdo_source_profile ();
297 /* For a given DECL, returns the top-level function_instance. */
298 function_instance
*get_function_instance_by_decl (tree decl
) const;
300 /* Find count_info for a given gimple STMT. If found, store the count_info
301 in INFO and return true; otherwise return false. */
302 bool get_count_info (gimple
*stmt
, count_info
*info
) const;
304 /* Find total count of the callee of EDGE. */
305 gcov_type
get_callsite_total_count (struct cgraph_edge
*edge
) const;
307 /* Update value profile INFO for STMT from the inlined indirect callsite.
308 Return true if INFO is updated. */
309 bool update_inlined_ind_target (gcall
*stmt
, count_info
*info
);
311 /* Mark LOC as annotated. */
312 void mark_annotated (location_t loc
);
315 /* Map from function_instance name index (in string_table) to
316 function_instance. */
317 typedef std::map
<unsigned, function_instance
*> name_function_instance_map
;
319 autofdo_source_profile () {}
321 /* Read AutoFDO profile and returns TRUE on success. */
324 /* Return the function_instance in the profile that correspond to the
327 get_function_instance_by_inline_stack (const inline_stack
&stack
) const;
329 name_function_instance_map map_
;
332 /* Store the strings read from the profile data file. */
333 static string_table
*afdo_string_table
;
335 /* Store the AutoFDO source profile. */
336 static autofdo_source_profile
*afdo_source_profile
;
338 /* gcov_summary structure to store the profile_info. */
339 static gcov_summary
*afdo_profile_info
;
341 /* Helper functions. */
343 /* Return the original name of NAME: strip the suffix that starts
344 with '.' Caller is responsible for freeing RET. */
347 get_original_name (const char *name
)
349 char *ret
= xstrdup (name
);
350 char *find
= strchr (ret
, '.');
356 /* Return the combined location, which is a 32bit integer in which
357 higher 16 bits stores the line offset of LOC to the start lineno
358 of DECL, The lower 16 bits stores the discriminator. */
361 get_combined_location (location_t loc
, tree decl
)
363 /* TODO: allow more bits for line and less bits for discriminator. */
364 if (LOCATION_LINE (loc
) - DECL_SOURCE_LINE (decl
) >= (1<<16))
365 warning_at (loc
, OPT_Woverflow
, "offset exceeds 16 bytes");
366 return ((LOCATION_LINE (loc
) - DECL_SOURCE_LINE (decl
)) << 16);
369 /* Return the function decl of a given lexical BLOCK. */
372 get_function_decl_from_block (tree block
)
374 if (!inlined_function_outer_scope_p (block
))
377 return BLOCK_ABSTRACT_ORIGIN (block
);
380 /* Store inline stack for STMT in STACK. */
383 get_inline_stack (location_t locus
, inline_stack
*stack
)
385 if (LOCATION_LOCUS (locus
) == UNKNOWN_LOCATION
)
388 tree block
= LOCATION_BLOCK (locus
);
389 if (block
&& TREE_CODE (block
) == BLOCK
)
392 for (block
= BLOCK_SUPERCONTEXT (block
);
393 block
&& (TREE_CODE (block
) == BLOCK
);
394 block
= BLOCK_SUPERCONTEXT (block
))
396 location_t tmp_locus
= BLOCK_SOURCE_LOCATION (block
);
397 if (LOCATION_LOCUS (tmp_locus
) == UNKNOWN_LOCATION
)
400 tree decl
= get_function_decl_from_block (block
);
402 std::make_pair (decl
, get_combined_location (locus
, decl
)));
408 std::make_pair (current_function_decl
,
409 get_combined_location (locus
, current_function_decl
)));
412 /* Return STMT's combined location, which is a 32bit integer in which
413 higher 16 bits stores the line offset of LOC to the start lineno
414 of DECL, The lower 16 bits stores the discriminator. */
417 get_relative_location_for_stmt (gimple
*stmt
)
419 location_t locus
= gimple_location (stmt
);
420 if (LOCATION_LOCUS (locus
) == UNKNOWN_LOCATION
)
421 return UNKNOWN_LOCATION
;
423 for (tree block
= gimple_block (stmt
); block
&& (TREE_CODE (block
) == BLOCK
);
424 block
= BLOCK_SUPERCONTEXT (block
))
425 if (LOCATION_LOCUS (BLOCK_SOURCE_LOCATION (block
)) != UNKNOWN_LOCATION
)
426 return get_combined_location (locus
,
427 get_function_decl_from_block (block
));
428 return get_combined_location (locus
, current_function_decl
);
431 /* Return true if BB contains indirect call. */
434 has_indirect_call (basic_block bb
)
436 gimple_stmt_iterator gsi
;
438 for (gsi
= gsi_start_bb (bb
); !gsi_end_p (gsi
); gsi_next (&gsi
))
440 gimple
*stmt
= gsi_stmt (gsi
);
441 if (gimple_code (stmt
) == GIMPLE_CALL
&& !gimple_call_internal_p (stmt
)
442 && (gimple_call_fn (stmt
) == NULL
443 || TREE_CODE (gimple_call_fn (stmt
)) != FUNCTION_DECL
))
449 /* Member functions for string_table. */
453 string_table::~string_table ()
455 for (unsigned i
= 0; i
< vector_
.length (); i
++)
460 /* Return the index of a given function NAME. Return -1 if NAME is not
461 found in string table. */
464 string_table::get_index (const char *name
) const
468 string_index_map::const_iterator iter
= map_
.find (name
);
469 if (iter
== map_
.end ())
475 /* Return the index of a given function DECL. Return -1 if DECL is not
476 found in string table. */
479 string_table::get_index_by_decl (tree decl
) const
482 = get_original_name (IDENTIFIER_POINTER (DECL_ASSEMBLER_NAME (decl
)));
483 int ret
= get_index (name
);
487 ret
= get_index (lang_hooks
.dwarf_name (decl
, 0));
490 if (DECL_FROM_INLINE (decl
))
491 return get_index_by_decl (DECL_ABSTRACT_ORIGIN (decl
));
496 /* Return the function name of a given INDEX. */
499 string_table::get_name (int index
) const
501 gcc_assert (index
> 0 && index
< (int)vector_
.length ());
502 return vector_
[index
];
505 /* Read the string table. Return TRUE if reading is successful. */
508 string_table::read ()
510 if (gcov_read_unsigned () != GCOV_TAG_AFDO_FILE_NAMES
)
512 /* Skip the length of the section. */
513 gcov_read_unsigned ();
514 /* Read in the file name table. */
515 unsigned string_num
= gcov_read_unsigned ();
516 for (unsigned i
= 0; i
< string_num
; i
++)
518 vector_
.safe_push (get_original_name (gcov_read_string ()));
519 map_
[vector_
.last ()] = i
;
524 /* Member functions for function_instance. */
526 function_instance::~function_instance ()
528 for (callsite_map::iterator iter
= callsites
.begin ();
529 iter
!= callsites
.end (); ++iter
)
533 /* Traverse callsites of the current function_instance to find one at the
534 location of LINENO and callee name represented in DECL. */
537 function_instance::get_function_instance_by_decl (unsigned lineno
,
540 int func_name_idx
= afdo_string_table
->get_index_by_decl (decl
);
541 if (func_name_idx
!= -1)
543 callsite_map::const_iterator ret
544 = callsites
.find (std::make_pair (lineno
, func_name_idx
));
545 if (ret
!= callsites
.end ())
549 = afdo_string_table
->get_index (lang_hooks
.dwarf_name (decl
, 0));
550 if (func_name_idx
!= -1)
552 callsite_map::const_iterator ret
553 = callsites
.find (std::make_pair (lineno
, func_name_idx
));
554 if (ret
!= callsites
.end ())
557 if (DECL_FROM_INLINE (decl
))
558 return get_function_instance_by_decl (lineno
, DECL_ABSTRACT_ORIGIN (decl
));
563 /* Store the profile info for LOC in INFO. Return TRUE if profile info
567 function_instance::get_count_info (location_t loc
, count_info
*info
) const
569 position_count_map::const_iterator iter
= pos_counts
.find (loc
);
570 if (iter
== pos_counts
.end ())
572 *info
= iter
->second
;
576 /* Mark LOC as annotated. */
579 function_instance::mark_annotated (location_t loc
)
581 position_count_map::iterator iter
= pos_counts
.find (loc
);
582 if (iter
== pos_counts
.end ())
584 iter
->second
.annotated
= true;
587 /* Read the inlined indirect call target profile for STMT and store it in
588 MAP, return the total count for all inlined indirect calls. */
591 function_instance::find_icall_target_map (gcall
*stmt
,
592 icall_target_map
*map
) const
595 unsigned stmt_offset
= get_relative_location_for_stmt (stmt
);
597 for (callsite_map::const_iterator iter
= callsites
.begin ();
598 iter
!= callsites
.end (); ++iter
)
600 unsigned callee
= iter
->second
->name ();
601 /* Check if callsite location match the stmt. */
602 if (iter
->first
.first
!= stmt_offset
)
604 struct cgraph_node
*node
= cgraph_node::get_for_asmname (
605 get_identifier (afdo_string_table
->get_name (callee
)));
608 (*map
)[callee
] = iter
->second
->total_count ();
609 ret
+= iter
->second
->total_count ();
614 /* Read the profile and create a function_instance with head count as
615 HEAD_COUNT. Recursively read callsites to create nested function_instances
616 too. STACK is used to track the recursive creation process. */
618 /* function instance profile format:
622 NUM_POS_COUNTS: 4 bytes
623 NUM_CALLSITES: 4 byte
625 POS_1_OFFSET: 4 bytes
629 VALUE_PROFILE_TYPE: 4 bytes
639 CALLSITE_1_OFFSET: 4 bytes
640 FUNCTION_INSTANCE_PROFILE (nested)
646 function_instance::read_function_instance (function_instance_stack
*stack
,
647 gcov_type head_count
)
649 unsigned name
= gcov_read_unsigned ();
650 unsigned num_pos_counts
= gcov_read_unsigned ();
651 unsigned num_callsites
= gcov_read_unsigned ();
652 function_instance
*s
= new function_instance (name
, head_count
);
653 stack
->safe_push (s
);
655 for (unsigned i
= 0; i
< num_pos_counts
; i
++)
657 unsigned offset
= gcov_read_unsigned () & 0xffff0000;
658 unsigned num_targets
= gcov_read_unsigned ();
659 gcov_type count
= gcov_read_counter ();
660 s
->pos_counts
[offset
].count
= count
;
661 for (unsigned j
= 0; j
< stack
->length (); j
++)
662 (*stack
)[j
]->total_count_
+= count
;
663 for (unsigned j
= 0; j
< num_targets
; j
++)
665 /* Only indirect call target histogram is supported now. */
666 gcov_read_unsigned ();
667 gcov_type target_idx
= gcov_read_counter ();
668 s
->pos_counts
[offset
].targets
[target_idx
] = gcov_read_counter ();
671 for (unsigned i
= 0; i
< num_callsites
; i
++)
673 unsigned offset
= gcov_read_unsigned ();
674 function_instance
*callee_function_instance
675 = read_function_instance (stack
, 0);
676 s
->callsites
[std::make_pair (offset
, callee_function_instance
->name ())]
677 = callee_function_instance
;
683 /* Sum of counts that is used during annotation. */
686 function_instance::total_annotated_count () const
689 for (callsite_map::const_iterator iter
= callsites
.begin ();
690 iter
!= callsites
.end (); ++iter
)
691 ret
+= iter
->second
->total_annotated_count ();
692 for (position_count_map::const_iterator iter
= pos_counts
.begin ();
693 iter
!= pos_counts
.end (); ++iter
)
694 if (iter
->second
.annotated
)
695 ret
+= iter
->second
.count
;
699 /* Member functions for autofdo_source_profile. */
701 autofdo_source_profile::~autofdo_source_profile ()
703 for (name_function_instance_map::const_iterator iter
= map_
.begin ();
704 iter
!= map_
.end (); ++iter
)
708 /* For a given DECL, returns the top-level function_instance. */
711 autofdo_source_profile::get_function_instance_by_decl (tree decl
) const
713 int index
= afdo_string_table
->get_index_by_decl (decl
);
716 name_function_instance_map::const_iterator ret
= map_
.find (index
);
717 return ret
== map_
.end () ? NULL
: ret
->second
;
720 /* Find count_info for a given gimple STMT. If found, store the count_info
721 in INFO and return true; otherwise return false. */
724 autofdo_source_profile::get_count_info (gimple
*stmt
, count_info
*info
) const
726 if (LOCATION_LOCUS (gimple_location (stmt
)) == cfun
->function_end_locus
)
730 get_inline_stack (gimple_location (stmt
), &stack
);
731 if (stack
.length () == 0)
733 function_instance
*s
= get_function_instance_by_inline_stack (stack
);
736 return s
->get_count_info (stack
[0].second
, info
);
739 /* Mark LOC as annotated. */
742 autofdo_source_profile::mark_annotated (location_t loc
)
745 get_inline_stack (loc
, &stack
);
746 if (stack
.length () == 0)
748 function_instance
*s
= get_function_instance_by_inline_stack (stack
);
751 s
->mark_annotated (stack
[0].second
);
754 /* Update value profile INFO for STMT from the inlined indirect callsite.
755 Return true if INFO is updated. */
758 autofdo_source_profile::update_inlined_ind_target (gcall
*stmt
,
763 fprintf (dump_file
, "Checking indirect call -> direct call ");
764 print_gimple_stmt (dump_file
, stmt
, 0, TDF_SLIM
);
767 if (LOCATION_LOCUS (gimple_location (stmt
)) == cfun
->function_end_locus
)
770 fprintf (dump_file
, " good locus\n");
775 get_count_info (stmt
, &old_info
);
777 for (icall_target_map::const_iterator iter
= old_info
.targets
.begin ();
778 iter
!= old_info
.targets
.end (); ++iter
)
779 total
+= iter
->second
;
781 /* Program behavior changed, original promoted (and inlined) target is not
782 hot any more. Will avoid promote the original target.
784 To check if original promoted target is still hot, we check the total
785 count of the unpromoted targets (stored in TOTAL). If a callsite count
786 (stored in INFO) is smaller than half of the total count, the original
787 promoted target is considered not hot any more. */
788 if (info
->count
< total
/ 2)
791 fprintf (dump_file
, " not hot anymore %ld < %ld",
798 get_inline_stack (gimple_location (stmt
), &stack
);
799 if (stack
.length () == 0)
802 fprintf (dump_file
, " no inline stack\n");
805 function_instance
*s
= get_function_instance_by_inline_stack (stack
);
809 fprintf (dump_file
, " function not found in inline stack\n");
812 icall_target_map map
;
813 if (s
->find_icall_target_map (stmt
, &map
) == 0)
816 fprintf (dump_file
, " no target map\n");
819 for (icall_target_map::const_iterator iter
= map
.begin ();
820 iter
!= map
.end (); ++iter
)
821 info
->targets
[iter
->first
] = iter
->second
;
823 fprintf (dump_file
, " looks good\n");
827 /* Find total count of the callee of EDGE. */
830 autofdo_source_profile::get_callsite_total_count (
831 struct cgraph_edge
*edge
) const
834 stack
.safe_push (std::make_pair (edge
->callee
->decl
, 0));
835 get_inline_stack (gimple_location (edge
->call_stmt
), &stack
);
837 function_instance
*s
= get_function_instance_by_inline_stack (stack
);
839 || afdo_string_table
->get_index (IDENTIFIER_POINTER (
840 DECL_ASSEMBLER_NAME (edge
->callee
->decl
))) != s
->name ())
843 return s
->total_count ();
846 /* Read AutoFDO profile and returns TRUE on success. */
848 /* source profile format:
850 GCOV_TAG_AFDO_FUNCTION: 4 bytes
852 NUM_FUNCTIONS: 4 bytes
856 FUNCTION_INSTANCE_N. */
859 autofdo_source_profile::read ()
861 if (gcov_read_unsigned () != GCOV_TAG_AFDO_FUNCTION
)
863 inform (UNKNOWN_LOCATION
, "Not expected TAG.");
867 /* Skip the length of the section. */
868 gcov_read_unsigned ();
870 /* Read in the function/callsite profile, and store it in local
872 unsigned function_num
= gcov_read_unsigned ();
873 for (unsigned i
= 0; i
< function_num
; i
++)
875 function_instance::function_instance_stack stack
;
876 function_instance
*s
= function_instance::read_function_instance (
877 &stack
, gcov_read_counter ());
878 map_
[s
->name ()] = s
;
883 /* Return the function_instance in the profile that correspond to the
887 autofdo_source_profile::get_function_instance_by_inline_stack (
888 const inline_stack
&stack
) const
890 name_function_instance_map::const_iterator iter
= map_
.find (
891 afdo_string_table
->get_index_by_decl (stack
[stack
.length () - 1].first
));
892 if (iter
== map_
.end())
894 function_instance
*s
= iter
->second
;
895 for (unsigned i
= stack
.length() - 1; i
> 0; i
--)
897 s
= s
->get_function_instance_by_decl (
898 stack
[i
].second
, stack
[i
- 1].first
);
905 /* Module profile is only used by LIPO. Here we simply ignore it. */
908 fake_read_autofdo_module_profile ()
910 /* Read in the module info. */
911 gcov_read_unsigned ();
913 /* Skip the length of the section. */
914 gcov_read_unsigned ();
916 /* Read in the file name table. */
917 unsigned total_module_num
= gcov_read_unsigned ();
918 gcc_assert (total_module_num
== 0);
921 /* Read data from profile data file. */
926 if (gcov_open (auto_profile_file
, 1) == 0)
928 error ("cannot open profile file %s", auto_profile_file
);
932 if (gcov_read_unsigned () != GCOV_DATA_MAGIC
)
934 error ("AutoFDO profile magic number does not match");
938 /* Skip the version number. */
939 unsigned version
= gcov_read_unsigned ();
940 if (version
!= AUTO_PROFILE_VERSION
)
942 error ("AutoFDO profile version %u does not match %u",
943 version
, AUTO_PROFILE_VERSION
);
947 /* Skip the empty integer. */
948 gcov_read_unsigned ();
951 afdo_string_table
= new string_table ();
952 if (!afdo_string_table
->read())
954 error ("cannot read string table from %s", auto_profile_file
);
958 /* autofdo_source_profile. */
959 afdo_source_profile
= autofdo_source_profile::create ();
960 if (afdo_source_profile
== NULL
)
962 error ("cannot read function profile from %s", auto_profile_file
);
966 /* autofdo_module_profile. */
967 fake_read_autofdo_module_profile ();
970 /* From AutoFDO profiles, find values inside STMT for that we want to measure
971 histograms for indirect-call optimization.
973 This function is actually served for 2 purposes:
974 * before annotation, we need to mark histogram, promote and inline
975 * after annotation, we just need to mark, and let follow-up logic to
976 decide if it needs to promote and inline. */
979 afdo_indirect_call (gimple_stmt_iterator
*gsi
, const icall_target_map
&map
,
982 gimple
*gs
= gsi_stmt (*gsi
);
985 if (map
.size () == 0)
987 gcall
*stmt
= dyn_cast
<gcall
*> (gs
);
989 || gimple_call_internal_p (stmt
)
990 || gimple_call_fndecl (stmt
) != NULL_TREE
)
994 icall_target_map::const_iterator max_iter
= map
.end ();
996 for (icall_target_map::const_iterator iter
= map
.begin ();
997 iter
!= map
.end (); ++iter
)
999 total
+= iter
->second
;
1000 if (max_iter
== map
.end () || max_iter
->second
< iter
->second
)
1003 struct cgraph_node
*direct_call
= cgraph_node::get_for_asmname (
1004 get_identifier (afdo_string_table
->get_name (max_iter
->first
)));
1005 if (direct_call
== NULL
|| !direct_call
->profile_id
)
1008 callee
= gimple_call_fn (stmt
);
1010 histogram_value hist
= gimple_alloc_histogram_value (
1011 cfun
, HIST_TYPE_INDIR_CALL
, stmt
, callee
);
1012 hist
->n_counters
= 4;
1013 hist
->hvalue
.counters
= XNEWVEC (gcov_type
, hist
->n_counters
);
1014 gimple_add_histogram_value (cfun
, stmt
, hist
);
1017 hist
->hvalue
.counters
[0] = total
;
1018 // Number of value/counter pairs
1019 hist
->hvalue
.counters
[1] = 1;
1021 hist
->hvalue
.counters
[2] = direct_call
->profile_id
;
1023 hist
->hvalue
.counters
[3] = max_iter
->second
;
1028 struct cgraph_edge
*indirect_edge
1029 = cgraph_node::get (current_function_decl
)->get_edge (stmt
);
1033 fprintf (dump_file
, "Indirect call -> direct call ");
1034 print_generic_expr (dump_file
, callee
, TDF_SLIM
);
1035 fprintf (dump_file
, " => ");
1036 print_generic_expr (dump_file
, direct_call
->decl
, TDF_SLIM
);
1039 if (direct_call
== NULL
)
1042 fprintf (dump_file
, " not transforming\n");
1045 if (DECL_STRUCT_FUNCTION (direct_call
->decl
) == NULL
)
1048 fprintf (dump_file
, " no declaration\n");
1054 fprintf (dump_file
, " transformation on insn ");
1055 print_gimple_stmt (dump_file
, stmt
, 0, TDF_SLIM
);
1056 fprintf (dump_file
, "\n");
1059 /* FIXME: Count should be initialized. */
1060 struct cgraph_edge
*new_edge
1061 = indirect_edge
->make_speculative (direct_call
,
1062 profile_count::uninitialized ());
1063 cgraph_edge::redirect_call_stmt_to_callee (new_edge
);
1064 gimple_remove_histogram_value (cfun
, stmt
, hist
);
1065 inline_call (new_edge
, true, NULL
, NULL
, false);
1068 /* From AutoFDO profiles, find values inside STMT for that we want to measure
1069 histograms and adds them to list VALUES. */
1072 afdo_vpt (gimple_stmt_iterator
*gsi
, const icall_target_map
&map
,
1075 afdo_indirect_call (gsi
, map
, transform
);
1078 typedef std::set
<basic_block
> bb_set
;
1079 typedef std::set
<edge
> edge_set
;
1082 is_bb_annotated (const basic_block bb
, const bb_set
&annotated
)
1084 return annotated
.find (bb
) != annotated
.end ();
1088 set_bb_annotated (basic_block bb
, bb_set
*annotated
)
1090 annotated
->insert (bb
);
1093 /* For a given BB, set its execution count. Attach value profile if a stmt
1094 is not in PROMOTED, because we only want to promote an indirect call once.
1095 Return TRUE if BB is annotated. */
1098 afdo_set_bb_count (basic_block bb
, const stmt_set
&promoted
)
1100 gimple_stmt_iterator gsi
;
1103 gcov_type max_count
= 0;
1104 bool has_annotated
= false;
1106 for (gsi
= gsi_start_bb (bb
); !gsi_end_p (gsi
); gsi_next (&gsi
))
1109 gimple
*stmt
= gsi_stmt (gsi
);
1110 if (gimple_clobber_p (stmt
) || is_gimple_debug (stmt
))
1112 if (afdo_source_profile
->get_count_info (stmt
, &info
))
1114 if (info
.count
> max_count
)
1115 max_count
= info
.count
;
1116 has_annotated
= true;
1117 if (info
.targets
.size () > 0
1118 && promoted
.find (stmt
) == promoted
.end ())
1119 afdo_vpt (&gsi
, info
.targets
, false);
1126 for (gsi
= gsi_start_bb (bb
); !gsi_end_p (gsi
); gsi_next (&gsi
))
1127 afdo_source_profile
->mark_annotated (gimple_location (gsi_stmt (gsi
)));
1128 for (gphi_iterator gpi
= gsi_start_phis (bb
);
1132 gphi
*phi
= gpi
.phi ();
1134 for (i
= 0; i
< gimple_phi_num_args (phi
); i
++)
1135 afdo_source_profile
->mark_annotated (gimple_phi_arg_location (phi
, i
));
1137 FOR_EACH_EDGE (e
, ei
, bb
->succs
)
1138 afdo_source_profile
->mark_annotated (e
->goto_locus
);
1140 bb
->count
= profile_count::from_gcov_type (max_count
).afdo ();
1144 /* BB1 and BB2 are in an equivalent class iff:
1145 1. BB1 dominates BB2.
1146 2. BB2 post-dominates BB1.
1147 3. BB1 and BB2 are in the same loop nest.
1148 This function finds the equivalent class for each basic block, and
1149 stores a pointer to the first BB in its equivalent class. Meanwhile,
1150 set bb counts for the same equivalent class to be idenical. Update
1151 ANNOTATED_BB for the first BB in its equivalent class. */
1154 afdo_find_equiv_class (bb_set
*annotated_bb
)
1158 FOR_ALL_BB_FN (bb
, cfun
)
1161 FOR_ALL_BB_FN (bb
, cfun
)
1163 if (bb
->aux
!= NULL
)
1166 for (basic_block bb1
: get_dominated_by (CDI_DOMINATORS
, bb
))
1167 if (bb1
->aux
== NULL
&& dominated_by_p (CDI_POST_DOMINATORS
, bb
, bb1
)
1168 && bb1
->loop_father
== bb
->loop_father
)
1171 if (bb1
->count
> bb
->count
&& is_bb_annotated (bb1
, *annotated_bb
))
1173 bb
->count
= bb1
->count
;
1174 set_bb_annotated (bb
, annotated_bb
);
1178 for (basic_block bb1
: get_dominated_by (CDI_POST_DOMINATORS
, bb
))
1179 if (bb1
->aux
== NULL
&& dominated_by_p (CDI_DOMINATORS
, bb
, bb1
)
1180 && bb1
->loop_father
== bb
->loop_father
)
1183 if (bb1
->count
> bb
->count
&& is_bb_annotated (bb1
, *annotated_bb
))
1185 bb
->count
= bb1
->count
;
1186 set_bb_annotated (bb
, annotated_bb
);
1192 /* If a basic block's count is known, and only one of its in/out edges' count
1193 is unknown, its count can be calculated. Meanwhile, if all of the in/out
1194 edges' counts are known, then the basic block's unknown count can also be
1195 calculated. Also, if a block has a single predecessor or successor, the block's
1196 count can be propagated to that predecessor or successor.
1197 IS_SUCC is true if out edges of a basic blocks are examined.
1198 Update ANNOTATED_BB accordingly.
1199 Return TRUE if any basic block/edge count is changed. */
1202 afdo_propagate_edge (bool is_succ
, bb_set
*annotated_bb
)
1205 bool changed
= false;
1207 FOR_EACH_BB_FN (bb
, cfun
)
1209 edge e
, unknown_edge
= NULL
;
1211 int num_unknown_edge
= 0;
1213 profile_count total_known_count
= profile_count::zero ().afdo ();
1215 FOR_EACH_EDGE (e
, ei
, is_succ
? bb
->succs
: bb
->preds
)
1217 gcc_assert (AFDO_EINFO (e
) != NULL
);
1218 if (! AFDO_EINFO (e
)->is_annotated ())
1219 num_unknown_edge
++, unknown_edge
= e
;
1221 total_known_count
+= AFDO_EINFO (e
)->get_count ();
1225 /* Be careful not to annotate block with no successor in special cases. */
1226 if (num_unknown_edge
== 0 && total_known_count
> bb
->count
)
1228 bb
->count
= total_known_count
;
1229 if (!is_bb_annotated (bb
, *annotated_bb
))
1230 set_bb_annotated (bb
, annotated_bb
);
1233 else if (num_unknown_edge
== 1 && is_bb_annotated (bb
, *annotated_bb
))
1235 if (bb
->count
> total_known_count
)
1237 profile_count new_count
= bb
->count
- total_known_count
;
1238 AFDO_EINFO(unknown_edge
)->set_count(new_count
);
1241 basic_block succ_or_pred_bb
= is_succ
? unknown_edge
->dest
: unknown_edge
->src
;
1242 if (new_count
> succ_or_pred_bb
->count
)
1244 succ_or_pred_bb
->count
= new_count
;
1245 if (!is_bb_annotated (succ_or_pred_bb
, *annotated_bb
))
1246 set_bb_annotated (succ_or_pred_bb
, annotated_bb
);
1251 AFDO_EINFO (unknown_edge
)->set_count (profile_count::zero().afdo ());
1252 AFDO_EINFO (unknown_edge
)->set_annotated ();
1259 /* Special propagation for circuit expressions. Because GCC translates
1260 control flow into data flow for circuit expressions. E.g.
1267 will be translated into:
1282 tmp = PHI (0 (BB1), 0 (BB.t1), 1 (BB.t2)
1288 In this case, we need to propagate through PHI to determine the edge
1289 count of BB1->BB.t1, BB.t1->BB.t2. */
1292 afdo_propagate_circuit (const bb_set
&annotated_bb
)
1295 FOR_ALL_BB_FN (bb
, cfun
)
1298 tree cmp_rhs
, cmp_lhs
;
1299 gimple
*cmp_stmt
= last_stmt (bb
);
1303 if (!cmp_stmt
|| gimple_code (cmp_stmt
) != GIMPLE_COND
)
1305 cmp_rhs
= gimple_cond_rhs (cmp_stmt
);
1306 cmp_lhs
= gimple_cond_lhs (cmp_stmt
);
1307 if (!TREE_CONSTANT (cmp_rhs
)
1308 || !(integer_zerop (cmp_rhs
) || integer_onep (cmp_rhs
)))
1310 if (TREE_CODE (cmp_lhs
) != SSA_NAME
)
1312 if (!is_bb_annotated (bb
, annotated_bb
))
1314 def_stmt
= SSA_NAME_DEF_STMT (cmp_lhs
);
1315 while (def_stmt
&& gimple_code (def_stmt
) == GIMPLE_ASSIGN
1316 && gimple_assign_single_p (def_stmt
)
1317 && TREE_CODE (gimple_assign_rhs1 (def_stmt
)) == SSA_NAME
)
1318 def_stmt
= SSA_NAME_DEF_STMT (gimple_assign_rhs1 (def_stmt
));
1321 gphi
*phi_stmt
= dyn_cast
<gphi
*> (def_stmt
);
1324 FOR_EACH_EDGE (e
, ei
, bb
->succs
)
1326 unsigned i
, total
= 0;
1328 bool check_value_one
= (((integer_onep (cmp_rhs
))
1329 ^ (gimple_cond_code (cmp_stmt
) == EQ_EXPR
))
1330 ^ ((e
->flags
& EDGE_TRUE_VALUE
) != 0));
1331 if (! AFDO_EINFO (e
)->is_annotated ())
1333 for (i
= 0; i
< gimple_phi_num_args (phi_stmt
); i
++)
1335 tree val
= gimple_phi_arg_def (phi_stmt
, i
);
1336 edge ep
= gimple_phi_arg_edge (phi_stmt
, i
);
1338 if (!TREE_CONSTANT (val
)
1339 || !(integer_zerop (val
) || integer_onep (val
)))
1341 if (check_value_one
^ integer_onep (val
))
1345 if (! (AFDO_EINFO (e
)->get_count ()).nonzero_p ()
1346 && ! AFDO_EINFO (ep
)->is_annotated ())
1348 AFDO_EINFO (ep
)->set_count (profile_count::zero ().afdo ());
1349 AFDO_EINFO (ep
)->set_annotated ();
1352 if (total
== 1 && ! AFDO_EINFO (only_one
)->is_annotated ())
1354 AFDO_EINFO (only_one
)->set_count (AFDO_EINFO (e
)->get_count ());
1355 AFDO_EINFO (only_one
)->set_annotated ();
1361 /* Propagate the basic block count and edge count on the control flow
1362 graph. We do the propagation iteratively until stablize. */
1365 afdo_propagate (bb_set
*annotated_bb
)
1368 bool changed
= true;
1371 FOR_ALL_BB_FN (bb
, cfun
)
1373 bb
->count
= ((basic_block
)bb
->aux
)->count
;
1374 if (is_bb_annotated ((basic_block
)bb
->aux
, *annotated_bb
))
1375 set_bb_annotated (bb
, annotated_bb
);
1378 while (changed
&& i
++ < 10)
1382 if (afdo_propagate_edge (true, annotated_bb
))
1384 if (afdo_propagate_edge (false, annotated_bb
))
1386 afdo_propagate_circuit (*annotated_bb
);
1390 /* Propagate counts on control flow graph and calculate branch
1394 afdo_calculate_branch_prob (bb_set
*annotated_bb
)
1400 calculate_dominance_info (CDI_POST_DOMINATORS
);
1401 calculate_dominance_info (CDI_DOMINATORS
);
1402 loop_optimizer_init (0);
1404 FOR_ALL_BB_FN (bb
, cfun
)
1406 gcc_assert (bb
->aux
== NULL
);
1407 FOR_EACH_EDGE (e
, ei
, bb
->succs
)
1409 gcc_assert (e
->aux
== NULL
);
1410 e
->aux
= new edge_info ();
1414 afdo_find_equiv_class (annotated_bb
);
1415 afdo_propagate (annotated_bb
);
1417 FOR_EACH_BB_FN (bb
, cfun
)
1419 int num_unknown_succ
= 0;
1420 profile_count total_count
= profile_count::zero ().afdo ();
1422 FOR_EACH_EDGE (e
, ei
, bb
->succs
)
1424 gcc_assert (AFDO_EINFO (e
) != NULL
);
1425 if (! AFDO_EINFO (e
)->is_annotated ())
1428 total_count
+= AFDO_EINFO (e
)->get_count ();
1430 if (num_unknown_succ
== 0 && total_count
> profile_count::zero ())
1432 FOR_EACH_EDGE (e
, ei
, bb
->succs
)
1434 = AFDO_EINFO (e
)->get_count ().probability_in (total_count
);
1437 FOR_ALL_BB_FN (bb
, cfun
)
1440 FOR_EACH_EDGE (e
, ei
, bb
->succs
)
1441 if (AFDO_EINFO (e
) != NULL
)
1443 delete AFDO_EINFO (e
);
1448 loop_optimizer_finalize ();
1449 free_dominance_info (CDI_DOMINATORS
);
1450 free_dominance_info (CDI_POST_DOMINATORS
);
1453 /* Perform value profile transformation using AutoFDO profile. Add the
1454 promoted stmts to PROMOTED_STMTS. Return TRUE if there is any
1455 indirect call promoted. */
1458 afdo_vpt_for_early_inline (stmt_set
*promoted_stmts
)
1461 if (afdo_source_profile
->get_function_instance_by_decl (
1462 current_function_decl
) == NULL
)
1465 compute_fn_summary (cgraph_node::get (current_function_decl
), true);
1467 bool has_vpt
= false;
1468 FOR_EACH_BB_FN (bb
, cfun
)
1470 if (!has_indirect_call (bb
))
1472 gimple_stmt_iterator gsi
;
1474 gcov_type bb_count
= 0;
1475 for (gsi
= gsi_start_bb (bb
); !gsi_end_p (gsi
); gsi_next (&gsi
))
1478 gimple
*stmt
= gsi_stmt (gsi
);
1479 if (afdo_source_profile
->get_count_info (stmt
, &info
))
1480 bb_count
= MAX (bb_count
, info
.count
);
1483 for (gsi
= gsi_start_bb (bb
); !gsi_end_p (gsi
); gsi_next (&gsi
))
1485 gcall
*stmt
= dyn_cast
<gcall
*> (gsi_stmt (gsi
));
1486 /* IC_promotion and early_inline_2 is done in multiple iterations.
1487 No need to promoted the stmt if its in promoted_stmts (means
1488 it is already been promoted in the previous iterations). */
1489 if ((!stmt
) || gimple_call_fn (stmt
) == NULL
1490 || TREE_CODE (gimple_call_fn (stmt
)) == FUNCTION_DECL
1491 || promoted_stmts
->find (stmt
) != promoted_stmts
->end ())
1495 afdo_source_profile
->get_count_info (stmt
, &info
);
1496 info
.count
= bb_count
;
1497 if (afdo_source_profile
->update_inlined_ind_target (stmt
, &info
))
1499 /* Promote the indirect call and update the promoted_stmts. */
1500 promoted_stmts
->insert (stmt
);
1501 afdo_vpt (&gsi
, info
.targets
, true);
1509 unsigned todo
= optimize_inline_calls (current_function_decl
);
1510 if (todo
& TODO_update_ssa_any
)
1511 update_ssa (TODO_update_ssa
);
1518 /* Annotate auto profile to the control flow graph. Do not annotate value
1519 profile for stmts in PROMOTED_STMTS. */
1522 afdo_annotate_cfg (const stmt_set
&promoted_stmts
)
1525 bb_set annotated_bb
;
1526 const function_instance
*s
1527 = afdo_source_profile
->get_function_instance_by_decl (
1528 current_function_decl
);
1532 cgraph_node::get (current_function_decl
)->count
1533 = profile_count::from_gcov_type (s
->head_count ()).afdo ();
1534 ENTRY_BLOCK_PTR_FOR_FN (cfun
)->count
1535 = profile_count::from_gcov_type (s
->head_count ()).afdo ();
1536 EXIT_BLOCK_PTR_FOR_FN (cfun
)->count
= profile_count::zero ().afdo ();
1537 profile_count max_count
= ENTRY_BLOCK_PTR_FOR_FN (cfun
)->count
;
1539 FOR_EACH_BB_FN (bb
, cfun
)
1541 /* As autoFDO uses sampling approach, we have to assume that all
1542 counters are zero when not seen by autoFDO. */
1543 bb
->count
= profile_count::zero ().afdo ();
1544 if (afdo_set_bb_count (bb
, promoted_stmts
))
1545 set_bb_annotated (bb
, &annotated_bb
);
1546 if (bb
->count
> max_count
)
1547 max_count
= bb
->count
;
1549 if (ENTRY_BLOCK_PTR_FOR_FN (cfun
)->count
1550 > ENTRY_BLOCK_PTR_FOR_FN (cfun
)->next_bb
->count
)
1552 ENTRY_BLOCK_PTR_FOR_FN (cfun
)->next_bb
->count
1553 = ENTRY_BLOCK_PTR_FOR_FN (cfun
)->count
;
1554 set_bb_annotated (ENTRY_BLOCK_PTR_FOR_FN (cfun
)->next_bb
, &annotated_bb
);
1556 if (ENTRY_BLOCK_PTR_FOR_FN (cfun
)->count
1557 > EXIT_BLOCK_PTR_FOR_FN (cfun
)->prev_bb
->count
)
1559 EXIT_BLOCK_PTR_FOR_FN (cfun
)->prev_bb
->count
1560 = ENTRY_BLOCK_PTR_FOR_FN (cfun
)->count
;
1561 set_bb_annotated (EXIT_BLOCK_PTR_FOR_FN (cfun
)->prev_bb
, &annotated_bb
);
1563 afdo_source_profile
->mark_annotated (
1564 DECL_SOURCE_LOCATION (current_function_decl
));
1565 afdo_source_profile
->mark_annotated (cfun
->function_start_locus
);
1566 afdo_source_profile
->mark_annotated (cfun
->function_end_locus
);
1567 if (max_count
> profile_count::zero ())
1569 /* Calculate, propagate count and probability information on CFG. */
1570 afdo_calculate_branch_prob (&annotated_bb
);
1572 update_max_bb_count ();
1573 profile_status_for_fn (cfun
) = PROFILE_READ
;
1574 if (flag_value_profile_transformations
)
1576 gimple_value_profile_transformations ();
1577 free_dominance_info (CDI_DOMINATORS
);
1578 free_dominance_info (CDI_POST_DOMINATORS
);
1579 update_ssa (TODO_update_ssa
);
1583 /* Wrapper function to invoke early inliner. */
1588 compute_fn_summary (cgraph_node::get (current_function_decl
), true);
1589 unsigned todo
= early_inliner (cfun
);
1590 if (todo
& TODO_update_ssa_any
)
1591 update_ssa (TODO_update_ssa
);
1594 /* Use AutoFDO profile to annoate the control flow graph.
1595 Return the todo flag. */
1600 struct cgraph_node
*node
;
1602 if (symtab
->state
== FINISHED
)
1605 init_node_map (true);
1606 profile_info
= autofdo::afdo_profile_info
;
1608 FOR_EACH_FUNCTION (node
)
1610 if (!gimple_has_body_p (node
->decl
))
1613 /* Don't profile functions produced for builtin stuff. */
1614 if (DECL_SOURCE_LOCATION (node
->decl
) == BUILTINS_LOCATION
)
1617 push_cfun (DECL_STRUCT_FUNCTION (node
->decl
));
1619 /* First do indirect call promotion and early inline to make the
1620 IR match the profiled binary before actual annotation.
1622 This is needed because an indirect call might have been promoted
1623 and inlined in the profiled binary. If we do not promote and
1624 inline these indirect calls before annotation, the profile for
1625 these promoted functions will be lost.
1627 e.g. foo() --indirect_call--> bar()
1628 In profiled binary, the callsite is promoted and inlined, making
1629 the profile look like:
1639 Before AutoFDO pass, loc_foo_2 is not promoted thus not inlined.
1640 If we perform annotation on it, the profile inside bar@loc_foo2
1643 To avoid this, we promote loc_foo_2 and inline the promoted bar
1644 function before annotation, so the profile inside bar@loc_foo2
1646 autofdo::stmt_set promoted_stmts
;
1647 for (int i
= 0; i
< opt_for_fn (node
->decl
,
1648 param_early_inliner_max_iterations
); i
++)
1650 if (!flag_value_profile_transformations
1651 || !autofdo::afdo_vpt_for_early_inline (&promoted_stmts
))
1657 autofdo::afdo_annotate_cfg (promoted_stmts
);
1658 compute_function_frequency ();
1660 /* Local pure-const may imply need to fixup the cfg. */
1661 if (execute_fixup_cfg () & TODO_cleanup_cfg
)
1662 cleanup_tree_cfg ();
1664 free_dominance_info (CDI_DOMINATORS
);
1665 free_dominance_info (CDI_POST_DOMINATORS
);
1666 cgraph_edge::rebuild_edges ();
1667 compute_fn_summary (cgraph_node::get (current_function_decl
), true);
1671 return TODO_rebuild_cgraph_edges
;
1673 } /* namespace autofdo. */
1675 /* Read the profile from the profile data file. */
1678 read_autofdo_file (void)
1680 if (auto_profile_file
== NULL
)
1681 auto_profile_file
= DEFAULT_AUTO_PROFILE_FILE
;
1683 autofdo::afdo_profile_info
= XNEW (gcov_summary
);
1684 autofdo::afdo_profile_info
->runs
= 1;
1685 autofdo::afdo_profile_info
->sum_max
= 0;
1687 /* Read the profile from the profile file. */
1688 autofdo::read_profile ();
1691 /* Free the resources. */
1694 end_auto_profile (void)
1696 delete autofdo::afdo_source_profile
;
1697 delete autofdo::afdo_string_table
;
1698 profile_info
= NULL
;
1701 /* Returns TRUE if EDGE is hot enough to be inlined early. */
1704 afdo_callsite_hot_enough_for_early_inline (struct cgraph_edge
*edge
)
1707 = autofdo::afdo_source_profile
->get_callsite_total_count (edge
);
1712 profile_count pcount
= profile_count::from_gcov_type (count
).afdo ();
1713 gcov_summary
*saved_profile_info
= profile_info
;
1714 /* At early inline stage, profile_info is not set yet. We need to
1715 temporarily set it to afdo_profile_info to calculate hotness. */
1716 profile_info
= autofdo::afdo_profile_info
;
1717 is_hot
= maybe_hot_count_p (NULL
, pcount
);
1718 profile_info
= saved_profile_info
;
1728 const pass_data pass_data_ipa_auto_profile
= {
1729 SIMPLE_IPA_PASS
, "afdo", /* name */
1730 OPTGROUP_NONE
, /* optinfo_flags */
1731 TV_IPA_AUTOFDO
, /* tv_id */
1732 0, /* properties_required */
1733 0, /* properties_provided */
1734 0, /* properties_destroyed */
1735 0, /* todo_flags_start */
1736 0, /* todo_flags_finish */
1739 class pass_ipa_auto_profile
: public simple_ipa_opt_pass
1742 pass_ipa_auto_profile (gcc::context
*ctxt
)
1743 : simple_ipa_opt_pass (pass_data_ipa_auto_profile
, ctxt
)
1747 /* opt_pass methods: */
1751 return flag_auto_profile
;
1753 virtual unsigned int
1754 execute (function
*)
1756 return autofdo::auto_profile ();
1758 }; // class pass_ipa_auto_profile
1762 simple_ipa_opt_pass
*
1763 make_pass_ipa_auto_profile (gcc::context
*ctxt
)
1765 return new pass_ipa_auto_profile (ctxt
);