1 /* Interprocedural constant propagation
2 Copyright (C) 2005, 2006, 2007, 2008 Free Software Foundation, Inc.
3 Contributed by Razya Ladelsky <RAZYA@il.ibm.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/>. */
21 /* Interprocedural constant propagation. The aim of interprocedural constant
22 propagation (IPCP) is to find which function's argument has the same
23 constant value in each invocation throughout the whole program. For example,
24 consider the following program:
28 printf ("value is %d",y);
48 The IPCP algorithm will find that g's formal argument y is always called
51 The algorithm used is based on "Interprocedural Constant Propagation", by
52 Challahan David, Keith D Cooper, Ken Kennedy, Linda Torczon, Comp86, pg
55 The optimization is divided into three stages:
57 First stage - intraprocedural analysis
58 =======================================
59 This phase computes jump_function and modification flags.
61 A jump function for a callsite represents the values passed as an actual
62 arguments of a given callsite. There are three types of values:
63 Pass through - the caller's formal parameter is passed as an actual argument.
64 Constant - a constant is passed as an actual argument.
65 Unknown - neither of the above.
67 The jump function info, ipa_jump_func, is stored in ipa_edge_args
68 structure (defined in ipa_prop.h and pointed to by cgraph_node->aux)
69 modified_flags are defined in ipa_node_params structure
70 (defined in ipa_prop.h and pointed to by cgraph_edge->aux).
72 -ipcp_init_stage() is the first stage driver.
74 Second stage - interprocedural analysis
75 ========================================
76 This phase does the interprocedural constant propagation.
77 It computes lattices for all formal parameters in the program
78 and their value that may be:
80 BOTTOM - non constant.
81 CONSTANT - constant value.
83 Lattice describing a formal parameter p will have a constant value if all
84 callsites invoking this function have the same constant value passed to p.
86 The lattices are stored in ipcp_lattice which is itself in ipa_node_params
87 structure (defined in ipa_prop.h and pointed to by cgraph_edge->aux).
89 -ipcp_iterate_stage() is the second stage driver.
91 Third phase - transformation of function code
92 ============================================
93 Propagates the constant-valued formals into the function.
94 For each function whose parameters are constants, we create its clone.
96 Then we process the clone in two ways:
97 1. We insert an assignment statement 'parameter = const' at the beginning
98 of the cloned function.
99 2. For read-only parameters that do not live in memory, we replace all their
100 uses with the constant.
102 We also need to modify some callsites to call the cloned functions instead
103 of the original ones. For a callsite passing an argument found to be a
104 constant by IPCP, there are two different cases to handle:
105 1. A constant is passed as an argument. In this case the callsite in the
106 should be redirected to call the cloned callee.
107 2. A parameter (of the caller) passed as an argument (pass through
108 argument). In such cases both the caller and the callee have clones and
109 only the callsite in the cloned caller is redirected to call to the
112 This update is done in two steps: First all cloned functions are created
113 during a traversal of the call graph, during which all callsites are
114 redirected to call the cloned function. Then the callsites are traversed
115 and many calls redirected back to fit the description above.
117 -ipcp_insert_stage() is the third phase driver.
123 #include "coretypes.h"
127 #include "ipa-prop.h"
128 #include "tree-flow.h"
129 #include "tree-pass.h"
132 #include "diagnostic.h"
133 #include "tree-dump.h"
134 #include "tree-inline.h"
138 /* Number of functions identified as candidates for cloning. When not cloning
139 we can simplify iterate stage not forcing it to go through the decision
140 on what is profitable and what not. */
141 static int n_cloning_candidates;
143 /* Maximal count found in program. */
144 static gcov_type max_count;
146 /* Cgraph nodes that has been completely replaced by cloning during iterate
147 * stage and will be removed after ipcp is finished. */
148 static bitmap dead_nodes;
150 static void ipcp_print_profile_data (FILE *);
151 static void ipcp_function_scale_print (FILE *);
153 /* Get the original node field of ipa_node_params associated with node NODE. */
154 static inline struct cgraph_node *
155 ipcp_get_orig_node (struct cgraph_node *node)
157 return IPA_NODE_REF (node)->ipcp_orig_node;
160 /* Return true if NODE describes a cloned/versioned function. */
162 ipcp_node_is_clone (struct cgraph_node *node)
164 return (ipcp_get_orig_node (node) != NULL);
167 /* Create ipa_node_params and its data structures for NEW_NODE. Set ORIG_NODE
168 as the ipcp_orig_node field in ipa_node_params. */
170 ipcp_init_cloned_node (struct cgraph_node *orig_node,
171 struct cgraph_node *new_node)
173 ipa_check_create_node_params ();
174 ipa_initialize_node_params (new_node);
175 IPA_NODE_REF (new_node)->ipcp_orig_node = orig_node;
178 /* Perform intraprocedrual analysis needed for ipcp. */
180 ipcp_analyze_node (struct cgraph_node *node)
182 /* Unreachable nodes should have been eliminated before ipcp. */
183 gcc_assert (node->needed || node->reachable);
185 ipa_initialize_node_params (node);
186 ipa_detect_param_modifications (node);
189 /* Recompute all local information since node might've got new
190 direct calls after cloning. */
192 ipcp_update_cloned_node (struct cgraph_node *new_node)
195 gimple_stmt_iterator gsi;
197 /* We might've introduced new direct calls. */
198 push_cfun (DECL_STRUCT_FUNCTION (new_node->decl));
199 current_function_decl = new_node->decl;
202 for (gsi = gsi_start_bb (bb); !gsi_end_p (gsi); gsi_next (&gsi))
204 gimple stmt = gsi_stmt (gsi);
207 if (is_gimple_call (stmt)
208 && (decl = gimple_call_fndecl (stmt))
209 && !cgraph_edge (new_node, stmt))
211 struct cgraph_edge *new_edge;
213 new_edge = cgraph_create_edge (new_node, cgraph_node (decl), stmt,
215 compute_call_stmt_bb_frequency (bb),
217 new_edge->indirect_call = 1;
221 /* Indirect inlinng rely on fact that we've already analyzed
223 if (flag_indirect_inlining)
225 struct cgraph_edge *cs;
227 ipcp_analyze_node (new_node);
229 for (cs = new_node->callees; cs; cs = cs->next_callee)
231 ipa_count_arguments (cs);
232 ipa_compute_jump_functions (cs);
236 current_function_decl = NULL;
239 /* Return scale for NODE. */
240 static inline gcov_type
241 ipcp_get_node_scale (struct cgraph_node *node)
243 return IPA_NODE_REF (node)->count_scale;
246 /* Set COUNT as scale for NODE. */
248 ipcp_set_node_scale (struct cgraph_node *node, gcov_type count)
250 IPA_NODE_REF (node)->count_scale = count;
253 /* Return whether LAT is a constant lattice. */
255 ipcp_lat_is_const (struct ipcp_lattice *lat)
257 if (lat->type == IPA_CONST_VALUE)
263 /* Return whether LAT is a constant lattice that ipa-cp can actually insert
264 into the code (i.e. constants excluding member pointers and pointers). */
266 ipcp_lat_is_insertable (struct ipcp_lattice *lat)
268 return lat->type == IPA_CONST_VALUE;
271 /* Return true if LAT1 and LAT2 are equal. */
273 ipcp_lats_are_equal (struct ipcp_lattice *lat1, struct ipcp_lattice *lat2)
275 gcc_assert (ipcp_lat_is_const (lat1) && ipcp_lat_is_const (lat2));
276 if (lat1->type != lat2->type)
279 if (operand_equal_p (lat1->constant, lat2->constant, 0))
285 /* Compute Meet arithmetics:
286 Meet (IPA_BOTTOM, x) = IPA_BOTTOM
288 Meet (const_a,const_b) = IPA_BOTTOM, if const_a != const_b.
289 MEET (const_a,const_b) = const_a, if const_a == const_b.*/
291 ipa_lattice_meet (struct ipcp_lattice *res, struct ipcp_lattice *lat1,
292 struct ipcp_lattice *lat2)
294 if (lat1->type == IPA_BOTTOM || lat2->type == IPA_BOTTOM)
296 res->type = IPA_BOTTOM;
299 if (lat1->type == IPA_TOP)
301 res->type = lat2->type;
302 res->constant = lat2->constant;
305 if (lat2->type == IPA_TOP)
307 res->type = lat1->type;
308 res->constant = lat1->constant;
311 if (!ipcp_lats_are_equal (lat1, lat2))
313 res->type = IPA_BOTTOM;
316 res->type = lat1->type;
317 res->constant = lat1->constant;
320 /* Return the lattice corresponding to the Ith formal parameter of the function
321 described by INFO. */
322 static inline struct ipcp_lattice *
323 ipcp_get_lattice (struct ipa_node_params *info, int i)
325 return &(info->params[i].ipcp_lattice);
328 /* Given the jump function JFUNC, compute the lattice LAT that describes the
329 value coming down the callsite. INFO describes the caller node so that
330 pass-through jump functions can be evaluated. */
332 ipcp_lattice_from_jfunc (struct ipa_node_params *info, struct ipcp_lattice *lat,
333 struct ipa_jump_func *jfunc)
335 if (jfunc->type == IPA_CONST)
337 lat->type = IPA_CONST_VALUE;
338 lat->constant = jfunc->value.constant;
340 else if (jfunc->type == IPA_PASS_THROUGH)
342 struct ipcp_lattice *caller_lat;
344 caller_lat = ipcp_get_lattice (info, jfunc->value.formal_id);
345 lat->type = caller_lat->type;
346 lat->constant = caller_lat->constant;
349 lat->type = IPA_BOTTOM;
352 /* True when OLD_LAT and NEW_LAT values are not the same. */
355 ipcp_lattice_changed (struct ipcp_lattice *old_lat,
356 struct ipcp_lattice *new_lat)
358 if (old_lat->type == new_lat->type)
360 if (!ipcp_lat_is_const (old_lat))
362 if (ipcp_lats_are_equal (old_lat, new_lat))
368 /* Print all ipcp_lattices of all functions to F. */
370 ipcp_print_all_lattices (FILE * f)
372 struct cgraph_node *node;
375 fprintf (f, "\nLattice:\n");
376 for (node = cgraph_nodes; node; node = node->next)
378 struct ipa_node_params *info;
382 info = IPA_NODE_REF (node);
383 fprintf (f, " Node: %s:\n", cgraph_node_name (node));
384 count = ipa_get_param_count (info);
385 for (i = 0; i < count; i++)
387 struct ipcp_lattice *lat = ipcp_get_lattice (info, i);
389 fprintf (f, " param [%d]: ", i);
390 if (lat->type == IPA_CONST_VALUE)
392 fprintf (f, "type is CONST ");
393 print_generic_expr (f, lat->constant, 0);
396 else if (lat->type == IPA_TOP)
397 fprintf (f, "type is TOP\n");
399 fprintf (f, "type is BOTTOM\n");
404 /* Return true if this NODE is viable candidate for cloning. */
406 ipcp_cloning_candidate_p (struct cgraph_node *node)
410 gcov_type direct_call_sum = 0;
411 struct cgraph_edge *e;
413 /* We never clone functions that are not visible from outside.
414 FIXME: in future we should clone such functions when they are called with
415 different constants, but current ipcp implementation is not good on this.
417 if (!node->needed || !node->analyzed)
420 if (cgraph_function_body_availability (node) <= AVAIL_OVERWRITABLE)
423 fprintf (dump_file, "Not considering %s for cloning; body is overwrittable.\n",
424 cgraph_node_name (node));
427 if (!tree_versionable_function_p (node->decl))
430 fprintf (dump_file, "Not considering %s for cloning; body is not versionable.\n",
431 cgraph_node_name (node));
434 for (e = node->callers; e; e = e->next_caller)
436 direct_call_sum += e->count;
438 if (cgraph_maybe_hot_edge_p (e))
445 fprintf (dump_file, "Not considering %s for cloning; no direct calls.\n",
446 cgraph_node_name (node));
449 if (node->local.inline_summary.self_insns < n_calls)
452 fprintf (dump_file, "Considering %s for cloning; code would shrink.\n",
453 cgraph_node_name (node));
457 if (!flag_ipa_cp_clone)
460 fprintf (dump_file, "Not considering %s for cloning; -fipa-cp-clone disabled.\n",
461 cgraph_node_name (node));
465 if (!optimize_function_for_speed_p (DECL_STRUCT_FUNCTION (node->decl)))
468 fprintf (dump_file, "Not considering %s for cloning; optimizing it for size.\n",
469 cgraph_node_name (node));
473 /* When profile is available and function is hot, propagate into it even if
474 calls seems cold; constant propagation can improve function's speed
478 if (direct_call_sum > node->count * 90 / 100)
481 fprintf (dump_file, "Considering %s for cloning; usually called directly.\n",
482 cgraph_node_name (node));
489 fprintf (dump_file, "Not considering %s for cloning; no hot calls.\n",
490 cgraph_node_name (node));
494 fprintf (dump_file, "Considering %s for cloning.\n",
495 cgraph_node_name (node));
499 /* Initialize ipcp_lattices array. The lattices corresponding to supported
500 types (integers, real types and Fortran constants defined as const_decls)
501 are initialized to IPA_TOP, the rest of them to IPA_BOTTOM. */
503 ipcp_initialize_node_lattices (struct cgraph_node *node)
506 struct ipa_node_params *info = IPA_NODE_REF (node);
507 enum ipa_lattice_type type;
509 if (ipa_is_called_with_var_arguments (info))
511 else if (!node->needed && !node->local.externally_visible)
513 /* When cloning is allowed, we can assume that externally visible functions
514 are not called. We will compensate this by cloning later. */
515 else if (ipcp_cloning_candidate_p (node))
516 type = IPA_TOP, n_cloning_candidates ++;
520 for (i = 0; i < ipa_get_param_count (info) ; i++)
521 ipcp_get_lattice (info, i)->type = type;
524 /* build INTEGER_CST tree with type TREE_TYPE and value according to LAT.
527 build_const_val (struct ipcp_lattice *lat, tree tree_type)
531 gcc_assert (ipcp_lat_is_const (lat));
534 if (!useless_type_conversion_p (tree_type, TREE_TYPE (val)))
536 if (fold_convertible_p (tree_type, val))
537 return fold_build1 (NOP_EXPR, tree_type, val);
539 return fold_build1 (VIEW_CONVERT_EXPR, tree_type, val);
544 /* Compute the proper scale for NODE. It is the ratio between the number of
545 direct calls (represented on the incoming cgraph_edges) and sum of all
546 invocations of NODE (represented as count in cgraph_node). */
548 ipcp_compute_node_scale (struct cgraph_node *node)
551 struct cgraph_edge *cs;
554 /* Compute sum of all counts of callers. */
555 for (cs = node->callers; cs != NULL; cs = cs->next_caller)
557 if (node->count == 0)
558 ipcp_set_node_scale (node, 0);
560 ipcp_set_node_scale (node, sum * REG_BR_PROB_BASE / node->count);
563 /* Initialization and computation of IPCP data structures. This is the initial
564 intraprocedural analysis of functions, which gathers information to be
565 propagated later on. */
567 ipcp_init_stage (void)
569 struct cgraph_node *node;
570 struct cgraph_edge *cs;
572 for (node = cgraph_nodes; node; node = node->next)
574 ipcp_analyze_node (node);
575 for (node = cgraph_nodes; node; node = node->next)
579 /* building jump functions */
580 for (cs = node->callees; cs; cs = cs->next_callee)
582 if (!cs->callee->analyzed)
584 ipa_count_arguments (cs);
585 if (ipa_get_cs_argument_count (IPA_EDGE_REF (cs))
586 != ipa_get_param_count (IPA_NODE_REF (cs->callee)))
588 /* Handle cases of functions with
589 a variable number of parameters. */
590 ipa_set_called_with_variable_arg (IPA_NODE_REF (cs->callee));
591 if (flag_indirect_inlining)
592 ipa_compute_jump_functions (cs);
595 ipa_compute_jump_functions (cs);
600 /* Return true if there are some formal parameters whose value is IPA_TOP (in
601 the whole compilation unit). Change their values to IPA_BOTTOM, since they
602 most probably get their values from outside of this compilation unit. */
604 ipcp_change_tops_to_bottom (void)
607 struct cgraph_node *node;
611 for (node = cgraph_nodes; node; node = node->next)
613 struct ipa_node_params *info = IPA_NODE_REF (node);
614 count = ipa_get_param_count (info);
615 for (i = 0; i < count; i++)
617 struct ipcp_lattice *lat = ipcp_get_lattice (info, i);
618 if (lat->type == IPA_TOP)
623 fprintf (dump_file, "Forcing param ");
624 print_generic_expr (dump_file, ipa_get_param (info, i), 0);
625 fprintf (dump_file, " of node %s to bottom.\n",
626 cgraph_node_name (node));
628 lat->type = IPA_BOTTOM;
635 /* Interprocedural analysis. The algorithm propagates constants from the
636 caller's parameters to the callee's arguments. */
638 ipcp_propagate_stage (void)
641 struct ipcp_lattice inc_lat = { IPA_BOTTOM, NULL };
642 struct ipcp_lattice new_lat = { IPA_BOTTOM, NULL };
643 struct ipcp_lattice *dest_lat;
644 struct cgraph_edge *cs;
645 struct ipa_jump_func *jump_func;
646 struct ipa_func_list *wl;
649 ipa_check_create_node_params ();
650 ipa_check_create_edge_args ();
652 /* Initialize worklist to contain all functions. */
653 wl = ipa_init_func_list ();
656 struct cgraph_node *node = ipa_pop_func_from_list (&wl);
657 struct ipa_node_params *info = IPA_NODE_REF (node);
659 for (cs = node->callees; cs; cs = cs->next_callee)
661 struct ipa_node_params *callee_info = IPA_NODE_REF (cs->callee);
662 struct ipa_edge_args *args = IPA_EDGE_REF (cs);
664 if (ipa_is_called_with_var_arguments (callee_info))
667 count = ipa_get_cs_argument_count (args);
668 for (i = 0; i < count; i++)
670 jump_func = ipa_get_ith_jump_func (args, i);
671 ipcp_lattice_from_jfunc (info, &inc_lat, jump_func);
672 dest_lat = ipcp_get_lattice (callee_info, i);
673 ipa_lattice_meet (&new_lat, &inc_lat, dest_lat);
674 if (ipcp_lattice_changed (&new_lat, dest_lat))
676 dest_lat->type = new_lat.type;
677 dest_lat->constant = new_lat.constant;
678 ipa_push_func_to_list (&wl, cs->callee);
685 /* Call the constant propagation algorithm and re-call it if necessary
686 (if there are undetermined values left). */
688 ipcp_iterate_stage (void)
690 struct cgraph_node *node;
691 n_cloning_candidates = 0;
694 fprintf (dump_file, "\nIPA iterate stage:\n\n");
695 for (node = cgraph_nodes; node; node = node->next)
697 ipcp_initialize_node_lattices (node);
698 ipcp_compute_node_scale (node);
700 if (dump_file && (dump_flags & TDF_DETAILS))
702 ipcp_print_all_lattices (dump_file);
703 ipcp_function_scale_print (dump_file);
706 ipcp_propagate_stage ();
707 if (ipcp_change_tops_to_bottom ())
708 /* Some lattices have changed from IPA_TOP to IPA_BOTTOM.
709 This change should be propagated. */
711 gcc_assert (n_cloning_candidates);
712 ipcp_propagate_stage ();
716 fprintf (dump_file, "\nIPA lattices after propagation:\n");
717 ipcp_print_all_lattices (dump_file);
718 if (dump_flags & TDF_DETAILS)
719 ipcp_print_profile_data (dump_file);
723 /* Check conditions to forbid constant insertion to function described by
726 ipcp_node_modifiable_p (struct cgraph_node *node)
728 /* Once we will be able to do in-place replacement, we can be more
730 return tree_versionable_function_p (node->decl);
733 /* Print count scale data structures. */
735 ipcp_function_scale_print (FILE * f)
737 struct cgraph_node *node;
739 for (node = cgraph_nodes; node; node = node->next)
743 fprintf (f, "printing scale for %s: ", cgraph_node_name (node));
744 fprintf (f, "value is " HOST_WIDE_INT_PRINT_DEC
745 " \n", (HOST_WIDE_INT) ipcp_get_node_scale (node));
749 /* Print counts of all cgraph nodes. */
751 ipcp_print_func_profile_counts (FILE * f)
753 struct cgraph_node *node;
755 for (node = cgraph_nodes; node; node = node->next)
757 fprintf (f, "function %s: ", cgraph_node_name (node));
758 fprintf (f, "count is " HOST_WIDE_INT_PRINT_DEC
759 " \n", (HOST_WIDE_INT) node->count);
763 /* Print counts of all cgraph edges. */
765 ipcp_print_call_profile_counts (FILE * f)
767 struct cgraph_node *node;
768 struct cgraph_edge *cs;
770 for (node = cgraph_nodes; node; node = node->next)
772 for (cs = node->callees; cs; cs = cs->next_callee)
774 fprintf (f, "%s -> %s ", cgraph_node_name (cs->caller),
775 cgraph_node_name (cs->callee));
776 fprintf (f, "count is " HOST_WIDE_INT_PRINT_DEC " \n",
777 (HOST_WIDE_INT) cs->count);
782 /* Print all counts and probabilities of cfg edges of all functions. */
784 ipcp_print_edge_profiles (FILE * f)
786 struct cgraph_node *node;
791 for (node = cgraph_nodes; node; node = node->next)
793 fprintf (f, "function %s: \n", cgraph_node_name (node));
797 ENTRY_BLOCK_PTR_FOR_FUNCTION (DECL_STRUCT_FUNCTION (node->decl));
798 fprintf (f, "ENTRY: ");
799 fprintf (f, " " HOST_WIDE_INT_PRINT_DEC
800 " %d\n", (HOST_WIDE_INT) bb->count, bb->frequency);
803 FOR_EACH_EDGE (e, ei, bb->succs)
806 EXIT_BLOCK_PTR_FOR_FUNCTION (DECL_STRUCT_FUNCTION
808 fprintf (f, "edge ENTRY -> EXIT, Count");
810 fprintf (f, "edge ENTRY -> %d, Count", e->dest->index);
811 fprintf (f, " " HOST_WIDE_INT_PRINT_DEC
812 " Prob %d\n", (HOST_WIDE_INT) e->count,
815 FOR_EACH_BB_FN (bb, DECL_STRUCT_FUNCTION (node->decl))
817 fprintf (f, "bb[%d]: ", bb->index);
818 fprintf (f, " " HOST_WIDE_INT_PRINT_DEC
819 " %d\n", (HOST_WIDE_INT) bb->count, bb->frequency);
820 FOR_EACH_EDGE (e, ei, bb->succs)
823 EXIT_BLOCK_PTR_FOR_FUNCTION (DECL_STRUCT_FUNCTION
825 fprintf (f, "edge %d -> EXIT, Count", e->src->index);
827 fprintf (f, "edge %d -> %d, Count", e->src->index,
829 fprintf (f, " " HOST_WIDE_INT_PRINT_DEC " Prob %d\n",
830 (HOST_WIDE_INT) e->count, e->probability);
837 /* Print counts and frequencies for all basic blocks of all functions. */
839 ipcp_print_bb_profiles (FILE * f)
842 struct cgraph_node *node;
844 for (node = cgraph_nodes; node; node = node->next)
846 fprintf (f, "function %s: \n", cgraph_node_name (node));
850 ENTRY_BLOCK_PTR_FOR_FUNCTION (DECL_STRUCT_FUNCTION (node->decl));
851 fprintf (f, "ENTRY: Count");
852 fprintf (f, " " HOST_WIDE_INT_PRINT_DEC
853 " Frequency %d\n", (HOST_WIDE_INT) bb->count,
856 FOR_EACH_BB_FN (bb, DECL_STRUCT_FUNCTION (node->decl))
858 fprintf (f, "bb[%d]: Count", bb->index);
859 fprintf (f, " " HOST_WIDE_INT_PRINT_DEC
860 " Frequency %d\n", (HOST_WIDE_INT) bb->count,
864 EXIT_BLOCK_PTR_FOR_FUNCTION (DECL_STRUCT_FUNCTION (node->decl));
865 fprintf (f, "EXIT: Count");
866 fprintf (f, " " HOST_WIDE_INT_PRINT_DEC
867 " Frequency %d\n", (HOST_WIDE_INT) bb->count,
874 /* Print profile info for all functions. */
876 ipcp_print_profile_data (FILE * f)
878 fprintf (f, "\nNODE COUNTS :\n");
879 ipcp_print_func_profile_counts (f);
880 fprintf (f, "\nCS COUNTS stage:\n");
881 ipcp_print_call_profile_counts (f);
882 fprintf (f, "\nBB COUNTS and FREQUENCIES :\n");
883 ipcp_print_bb_profiles (f);
884 fprintf (f, "\nCFG EDGES COUNTS and PROBABILITIES :\n");
885 ipcp_print_edge_profiles (f);
888 /* Build and initialize ipa_replace_map struct according to LAT. This struct is
889 processed by versioning, which operates according to the flags set.
890 PARM_TREE is the formal parameter found to be constant. LAT represents the
892 static struct ipa_replace_map *
893 ipcp_create_replace_map (tree parm_tree, struct ipcp_lattice *lat)
895 struct ipa_replace_map *replace_map;
898 replace_map = XCNEW (struct ipa_replace_map);
899 const_val = build_const_val (lat, TREE_TYPE (parm_tree));
902 fprintf (dump_file, " replacing param ");
903 print_generic_expr (dump_file, parm_tree, 0);
904 fprintf (dump_file, " with const ");
905 print_generic_expr (dump_file, const_val, 0);
906 fprintf (dump_file, "\n");
908 replace_map->old_tree = parm_tree;
909 replace_map->new_tree = const_val;
910 replace_map->replace_p = true;
911 replace_map->ref_p = false;
916 /* Return true if this callsite should be redirected to the original callee
917 (instead of the cloned one). */
919 ipcp_need_redirect_p (struct cgraph_edge *cs)
921 struct ipa_node_params *orig_callee_info;
923 struct ipa_jump_func *jump_func;
924 struct cgraph_node *node = cs->callee, *orig;
926 if (!n_cloning_candidates)
929 if ((orig = ipcp_get_orig_node (node)) != NULL)
931 if (ipcp_get_orig_node (cs->caller))
934 orig_callee_info = IPA_NODE_REF (node);
935 count = ipa_get_param_count (orig_callee_info);
936 for (i = 0; i < count; i++)
938 struct ipcp_lattice *lat = ipcp_get_lattice (orig_callee_info, i);
939 if (ipcp_lat_is_const (lat))
941 jump_func = ipa_get_ith_jump_func (IPA_EDGE_REF (cs), i);
942 if (jump_func->type != IPA_CONST)
950 /* Fix the callsites and the call graph after function cloning was done. */
952 ipcp_update_callgraph (void)
954 struct cgraph_node *node;
956 for (node = cgraph_nodes; node; node = node->next)
957 if (node->analyzed && ipcp_node_is_clone (node))
959 bitmap args_to_skip = BITMAP_ALLOC (NULL);
960 struct cgraph_node *orig_node = ipcp_get_orig_node (node);
961 struct ipa_node_params *info = IPA_NODE_REF (orig_node);
962 int i, count = ipa_get_param_count (info);
963 struct cgraph_edge *cs, *next;
965 for (i = 0; i < count; i++)
967 struct ipcp_lattice *lat = ipcp_get_lattice (info, i);
968 tree parm_tree = ipa_get_param (info, i);
970 /* We can proactively remove obviously unused arguments. */
971 if (is_gimple_reg (parm_tree)
972 && !gimple_default_def (DECL_STRUCT_FUNCTION (orig_node->decl),
975 bitmap_set_bit (args_to_skip, i);
979 if (lat->type == IPA_CONST_VALUE)
980 bitmap_set_bit (args_to_skip, i);
982 for (cs = node->callers; cs; cs = next)
984 next = cs->next_caller;
985 if (!cs->indirect_call
986 && (ipcp_node_is_clone (cs->caller)
987 || !ipcp_need_redirect_p (cs)))
990 gimple_stmt_iterator gsi;
992 current_function_decl = cs->caller->decl;
993 push_cfun (DECL_STRUCT_FUNCTION (cs->caller->decl));
995 new_stmt = gimple_call_copy_skip_args (cs->call_stmt,
997 gsi = gsi_for_stmt (cs->call_stmt);
998 gsi_replace (&gsi, new_stmt, true);
999 cgraph_set_call_stmt (cs, new_stmt);
1001 current_function_decl = NULL;
1005 cgraph_redirect_edge_callee (cs, orig_node);
1006 gimple_call_set_fndecl (cs->call_stmt, orig_node->decl);
1012 /* Update all cfg basic blocks in NODE according to SCALE. */
1014 ipcp_update_bb_counts (struct cgraph_node *node, gcov_type scale)
1018 FOR_ALL_BB_FN (bb, DECL_STRUCT_FUNCTION (node->decl))
1019 bb->count = bb->count * scale / REG_BR_PROB_BASE;
1022 /* Update all cfg edges in NODE according to SCALE. */
1024 ipcp_update_edges_counts (struct cgraph_node *node, gcov_type scale)
1030 FOR_ALL_BB_FN (bb, DECL_STRUCT_FUNCTION (node->decl))
1031 FOR_EACH_EDGE (e, ei, bb->succs)
1032 e->count = e->count * scale / REG_BR_PROB_BASE;
1035 /* Update profiling info for versioned functions and the functions they were
1038 ipcp_update_profiling (void)
1040 struct cgraph_node *node, *orig_node;
1041 gcov_type scale, scale_complement;
1042 struct cgraph_edge *cs;
1044 for (node = cgraph_nodes; node; node = node->next)
1046 if (ipcp_node_is_clone (node))
1048 orig_node = ipcp_get_orig_node (node);
1049 scale = ipcp_get_node_scale (orig_node);
1050 node->count = orig_node->count * scale / REG_BR_PROB_BASE;
1051 scale_complement = REG_BR_PROB_BASE - scale;
1053 orig_node->count * scale_complement / REG_BR_PROB_BASE;
1054 for (cs = node->callees; cs; cs = cs->next_callee)
1055 cs->count = cs->count * scale / REG_BR_PROB_BASE;
1056 for (cs = orig_node->callees; cs; cs = cs->next_callee)
1057 cs->count = cs->count * scale_complement / REG_BR_PROB_BASE;
1058 ipcp_update_bb_counts (node, scale);
1059 ipcp_update_bb_counts (orig_node, scale_complement);
1060 ipcp_update_edges_counts (node, scale);
1061 ipcp_update_edges_counts (orig_node, scale_complement);
1066 /* If NODE was cloned, how much would program grow? */
1068 ipcp_estimate_growth (struct cgraph_node *node)
1070 struct cgraph_edge *cs;
1071 int redirectable_node_callers = 0;
1072 int removable_args = 0;
1073 bool need_original = node->needed;
1074 struct ipa_node_params *info;
1078 for (cs = node->callers; cs != NULL; cs = cs->next_caller)
1079 if (cs->caller == node || !ipcp_need_redirect_p (cs))
1080 redirectable_node_callers++;
1082 need_original = true;
1084 /* If we will be able to fully replace orignal node, we never increase
1089 info = IPA_NODE_REF (node);
1090 count = ipa_get_param_count (info);
1091 for (i = 0; i < count; i++)
1093 struct ipcp_lattice *lat = ipcp_get_lattice (info, i);
1094 tree parm_tree = ipa_get_param (info, i);
1096 /* We can proactively remove obviously unused arguments. */
1097 if (is_gimple_reg (parm_tree)
1098 && !gimple_default_def (DECL_STRUCT_FUNCTION (node->decl),
1102 if (lat->type == IPA_CONST_VALUE)
1106 /* We make just very simple estimate of savings for removal of operand from
1107 call site. Precise cost is dificult to get, as our size metric counts
1108 constants and moves as free. Generally we are looking for cases that
1109 small function is called very many times. */
1110 growth = node->local.inline_summary.self_insns
1111 - removable_args * redirectable_node_callers;
1118 /* Estimate cost of cloning NODE. */
1120 ipcp_estimate_cloning_cost (struct cgraph_node *node)
1123 gcov_type count_sum = 1;
1124 struct cgraph_edge *e;
1127 cost = ipcp_estimate_growth (node) * 1000;
1131 fprintf (dump_file, "Versioning of %s will save code size\n",
1132 cgraph_node_name (node));
1136 for (e = node->callers; e; e = e->next_caller)
1137 if (!bitmap_bit_p (dead_nodes, e->caller->uid)
1138 && !ipcp_need_redirect_p (e))
1140 count_sum += e->count;
1141 freq_sum += e->frequency + 1;
1145 cost /= count_sum * 1000 / max_count + 1;
1147 cost /= freq_sum * 1000 / REG_BR_PROB_BASE + 1;
1149 fprintf (dump_file, "Cost of versioning %s is %i, (size: %i, freq: %i)\n",
1150 cgraph_node_name (node), cost, node->local.inline_summary.self_insns,
1155 /* Return number of live constant parameters. */
1157 ipcp_const_param_count (struct cgraph_node *node)
1159 int const_param = 0;
1160 struct ipa_node_params *info = IPA_NODE_REF (node);
1161 int count = ipa_get_param_count (info);
1164 for (i = 0; i < count; i++)
1166 struct ipcp_lattice *lat = ipcp_get_lattice (info, i);
1167 tree parm_tree = ipa_get_param (info, i);
1168 if (ipcp_lat_is_insertable (lat)
1169 /* Do not count obviously unused arguments. */
1170 && (!is_gimple_reg (parm_tree)
1171 || gimple_default_def (DECL_STRUCT_FUNCTION (node->decl),
1178 /* Propagate the constant parameters found by ipcp_iterate_stage()
1179 to the function's code. */
1181 ipcp_insert_stage (void)
1183 struct cgraph_node *node, *node1 = NULL;
1185 VEC (cgraph_edge_p, heap) * redirect_callers;
1186 varray_type replace_trees;
1187 int node_callers, count;
1189 struct ipa_replace_map *replace_param;
1191 long overall_insns = 0, new_insns = 0;
1194 ipa_check_create_node_params ();
1195 ipa_check_create_edge_args ();
1197 fprintf (dump_file, "\nIPA insert stage:\n\n");
1199 dead_nodes = BITMAP_ALLOC (NULL);
1201 for (node = cgraph_nodes; node; node = node->next)
1204 if (node->count > max_count)
1205 max_count = node->count;
1206 overall_insns += node->local.inline_summary.self_insns;
1209 max_new_insns = overall_insns;
1210 if (max_new_insns < PARAM_VALUE (PARAM_LARGE_UNIT_INSNS))
1211 max_new_insns = PARAM_VALUE (PARAM_LARGE_UNIT_INSNS);
1212 max_new_insns = max_new_insns * PARAM_VALUE (PARAM_IPCP_UNIT_GROWTH) / 100 + 1;
1214 /* First collect all functions we proved to have constant arguments to heap. */
1215 heap = fibheap_new ();
1216 for (node = cgraph_nodes; node; node = node->next)
1218 struct ipa_node_params *info;
1219 /* Propagation of the constant is forbidden in certain conditions. */
1220 if (!node->analyzed || !ipcp_node_modifiable_p (node))
1222 info = IPA_NODE_REF (node);
1223 if (ipa_is_called_with_var_arguments (info))
1225 if (ipcp_const_param_count (node))
1226 node->aux = fibheap_insert (heap, ipcp_estimate_cloning_cost (node), node);
1229 /* Now clone in priority order until code size growth limits are met or
1231 while (!fibheap_empty (heap))
1233 struct ipa_node_params *info;
1235 bitmap args_to_skip;
1236 struct cgraph_edge *cs;
1238 node = (struct cgraph_node *)fibheap_extract_min (heap);
1241 fprintf (dump_file, "considering function %s\n",
1242 cgraph_node_name (node));
1244 growth = ipcp_estimate_growth (node);
1246 if (new_insns + growth > max_new_insns)
1249 && optimize_function_for_size_p (DECL_STRUCT_FUNCTION (node->decl)))
1252 fprintf (dump_file, "Not versioning, cold code would grow");
1256 new_insns += growth;
1258 /* Look if original function becomes dead after clonning. */
1259 for (cs = node->callers; cs != NULL; cs = cs->next_caller)
1260 if (cs->caller == node || ipcp_need_redirect_p (cs))
1262 if (!cs && !node->needed)
1263 bitmap_set_bit (dead_nodes, node->uid);
1265 info = IPA_NODE_REF (node);
1266 count = ipa_get_param_count (info);
1268 VARRAY_GENERIC_PTR_INIT (replace_trees, ipcp_const_param_count (node),
1270 args_to_skip = BITMAP_ALLOC (NULL);
1271 for (i = 0; i < count; i++)
1273 struct ipcp_lattice *lat = ipcp_get_lattice (info, i);
1274 parm_tree = ipa_get_param (info, i);
1276 /* We can proactively remove obviously unused arguments. */
1277 if (is_gimple_reg (parm_tree)
1278 && !gimple_default_def (DECL_STRUCT_FUNCTION (node->decl),
1281 bitmap_set_bit (args_to_skip, i);
1285 if (lat->type == IPA_CONST_VALUE)
1288 ipcp_create_replace_map (parm_tree, lat);
1289 VARRAY_PUSH_GENERIC_PTR (replace_trees, replace_param);
1290 bitmap_set_bit (args_to_skip, i);
1294 /* Compute how many callers node has. */
1296 for (cs = node->callers; cs != NULL; cs = cs->next_caller)
1298 redirect_callers = VEC_alloc (cgraph_edge_p, heap, node_callers);
1299 for (cs = node->callers; cs != NULL; cs = cs->next_caller)
1300 VEC_quick_push (cgraph_edge_p, redirect_callers, cs);
1302 /* Redirecting all the callers of the node to the
1303 new versioned node. */
1305 cgraph_function_versioning (node, redirect_callers, replace_trees,
1307 BITMAP_FREE (args_to_skip);
1308 VEC_free (cgraph_edge_p, heap, redirect_callers);
1309 VARRAY_CLEAR (replace_trees);
1313 fprintf (dump_file, "versioned function %s with growth %i, overall %i\n",
1314 cgraph_node_name (node), (int)growth, (int)new_insns);
1315 ipcp_init_cloned_node (node, node1);
1317 /* We've possibly introduced direct calls. */
1318 ipcp_update_cloned_node (node1);
1321 dump_function_to_file (node1->decl, dump_file, dump_flags);
1323 for (cs = node->callees; cs; cs = cs->next_callee)
1324 if (cs->callee->aux)
1326 fibheap_delete_node (heap, (fibnode_t) cs->callee->aux);
1327 cs->callee->aux = fibheap_insert (heap,
1328 ipcp_estimate_cloning_cost (cs->callee),
1333 while (!fibheap_empty (heap))
1336 fprintf (dump_file, "skipping function %s\n",
1337 cgraph_node_name (node));
1338 node = (struct cgraph_node *) fibheap_extract_min (heap);
1341 fibheap_delete (heap);
1342 BITMAP_FREE (dead_nodes);
1343 ipcp_update_callgraph ();
1344 ipcp_update_profiling ();
1347 /* The IPCP driver. */
1351 cgraph_remove_unreachable_nodes (true,dump_file);
1354 fprintf (dump_file, "\nIPA structures before propagation:\n");
1355 if (dump_flags & TDF_DETAILS)
1356 ipa_print_all_params (dump_file);
1357 ipa_print_all_jump_functions (dump_file);
1359 /* 2. Do the interprocedural propagation. */
1360 ipcp_iterate_stage ();
1361 /* 3. Insert the constants found to the functions. */
1362 ipcp_insert_stage ();
1363 if (dump_file && (dump_flags & TDF_DETAILS))
1365 fprintf (dump_file, "\nProfiling info after insert stage:\n");
1366 ipcp_print_profile_data (dump_file);
1368 /* Free all IPCP structures. */
1369 free_all_ipa_structures_after_ipa_cp ();
1371 fprintf (dump_file, "\nIPA constant propagation end\n");
1375 /* Note function body size. */
1377 ipcp_generate_summary (void)
1380 fprintf (dump_file, "\nIPA constant propagation start:\n");
1381 ipa_check_create_node_params ();
1382 ipa_check_create_edge_args ();
1383 ipa_register_cgraph_hooks ();
1384 /* 1. Call the init stage to initialize
1385 the ipa_node_params and ipa_edge_args structures. */
1389 /* Gate for IPCP optimization. */
1391 cgraph_gate_cp (void)
1396 struct ipa_opt_pass pass_ipa_cp =
1401 cgraph_gate_cp, /* gate */
1402 ipcp_driver, /* execute */
1405 0, /* static_pass_number */
1406 TV_IPA_CONSTANT_PROP, /* tv_id */
1407 0, /* properties_required */
1408 PROP_trees, /* properties_provided */
1409 0, /* properties_destroyed */
1410 0, /* todo_flags_start */
1411 TODO_dump_cgraph | TODO_dump_func |
1412 TODO_remove_functions /* todo_flags_finish */
1414 ipcp_generate_summary, /* generate_summary */
1415 NULL, /* write_summary */
1416 NULL, /* read_summary */
1417 NULL, /* function_read_summary */
1419 NULL, /* function_transform */
1420 NULL, /* variable_transform */