1 /* Interprocedural constant propagation
2 Copyright (C) 2005, 2006, 2007, 2008, 2009, 2010
3 Free Software Foundation, Inc.
4 Contributed by Razya Ladelsky <RAZYA@il.ibm.com>
6 This file is part of GCC.
8 GCC is free software; you can redistribute it and/or modify it under
9 the terms of the GNU General Public License as published by the Free
10 Software Foundation; either version 3, or (at your option) any later
13 GCC is distributed in the hope that it will be useful, but WITHOUT ANY
14 WARRANTY; without even the implied warranty of MERCHANTABILITY or
15 FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
18 You should have received a copy of the GNU General Public License
19 along with GCC; see the file COPYING3. If not see
20 <http://www.gnu.org/licenses/>. */
22 /* Interprocedural constant propagation. The aim of interprocedural constant
23 propagation (IPCP) is to find which function's argument has the same
24 constant value in each invocation throughout the whole program. For example,
25 consider the following program:
29 printf ("value is %d",y);
49 The IPCP algorithm will find that g's formal argument y is always called
52 The algorithm used is based on "Interprocedural Constant Propagation", by
53 Challahan David, Keith D Cooper, Ken Kennedy, Linda Torczon, Comp86, pg
56 The optimization is divided into three stages:
58 First stage - intraprocedural analysis
59 =======================================
60 This phase computes jump_function and modification flags.
62 A jump function for a callsite represents the values passed as an actual
63 arguments of a given callsite. There are three types of values:
64 Pass through - the caller's formal parameter is passed as an actual argument.
65 Constant - a constant is passed as an actual argument.
66 Unknown - neither of the above.
68 The jump function info, ipa_jump_func, is stored in ipa_edge_args
69 structure (defined in ipa_prop.h and pointed to by cgraph_node->aux)
70 modified_flags are defined in ipa_node_params structure
71 (defined in ipa_prop.h and pointed to by cgraph_edge->aux).
73 -ipcp_init_stage() is the first stage driver.
75 Second stage - interprocedural analysis
76 ========================================
77 This phase does the interprocedural constant propagation.
78 It computes lattices for all formal parameters in the program
79 and their value that may be:
81 BOTTOM - non constant.
82 CONSTANT - constant value.
84 Lattice describing a formal parameter p will have a constant value if all
85 callsites invoking this function have the same constant value passed to p.
87 The lattices are stored in ipcp_lattice which is itself in ipa_node_params
88 structure (defined in ipa_prop.h and pointed to by cgraph_edge->aux).
90 -ipcp_iterate_stage() is the second stage driver.
92 Third phase - transformation of function code
93 ============================================
94 Propagates the constant-valued formals into the function.
95 For each function whose parameters are constants, we create its clone.
97 Then we process the clone in two ways:
98 1. We insert an assignment statement 'parameter = const' at the beginning
99 of the cloned function.
100 2. For read-only parameters that do not live in memory, we replace all their
101 uses with the constant.
103 We also need to modify some callsites to call the cloned functions instead
104 of the original ones. For a callsite passing an argument found to be a
105 constant by IPCP, there are two different cases to handle:
106 1. A constant is passed as an argument. In this case the callsite in the
107 should be redirected to call the cloned callee.
108 2. A parameter (of the caller) passed as an argument (pass through
109 argument). In such cases both the caller and the callee have clones and
110 only the callsite in the cloned caller is redirected to call to the
113 This update is done in two steps: First all cloned functions are created
114 during a traversal of the call graph, during which all callsites are
115 redirected to call the cloned function. Then the callsites are traversed
116 and many calls redirected back to fit the description above.
118 -ipcp_insert_stage() is the third phase driver.
124 #include "coretypes.h"
128 #include "ipa-prop.h"
129 #include "tree-flow.h"
130 #include "tree-pass.h"
133 #include "diagnostic.h"
134 #include "tree-dump.h"
135 #include "tree-inline.h"
139 /* Number of functions identified as candidates for cloning. When not cloning
140 we can simplify iterate stage not forcing it to go through the decision
141 on what is profitable and what not. */
142 static int n_cloning_candidates;
144 /* Maximal count found in program. */
145 static gcov_type max_count;
147 /* Cgraph nodes that has been completely replaced by cloning during iterate
148 * stage and will be removed after ipcp is finished. */
149 static bitmap dead_nodes;
151 static void ipcp_print_profile_data (FILE *);
152 static void ipcp_function_scale_print (FILE *);
154 /* Get the original node field of ipa_node_params associated with node NODE. */
155 static inline struct cgraph_node *
156 ipcp_get_orig_node (struct cgraph_node *node)
158 return IPA_NODE_REF (node)->ipcp_orig_node;
161 /* Return true if NODE describes a cloned/versioned function. */
163 ipcp_node_is_clone (struct cgraph_node *node)
165 return (ipcp_get_orig_node (node) != NULL);
168 /* Create ipa_node_params and its data structures for NEW_NODE. Set ORIG_NODE
169 as the ipcp_orig_node field in ipa_node_params. */
171 ipcp_init_cloned_node (struct cgraph_node *orig_node,
172 struct cgraph_node *new_node)
174 ipa_check_create_node_params ();
175 ipa_initialize_node_params (new_node);
176 IPA_NODE_REF (new_node)->ipcp_orig_node = orig_node;
179 /* Perform intraprocedrual analysis needed for ipcp. */
181 ipcp_analyze_node (struct cgraph_node *node)
183 /* Unreachable nodes should have been eliminated before ipcp. */
184 gcc_assert (node->needed || node->reachable);
186 node->local.versionable = tree_versionable_function_p (node->decl);
187 ipa_initialize_node_params (node);
188 ipa_detect_param_modifications (node);
191 /* Return scale for NODE. */
192 static inline gcov_type
193 ipcp_get_node_scale (struct cgraph_node *node)
195 return IPA_NODE_REF (node)->count_scale;
198 /* Set COUNT as scale for NODE. */
200 ipcp_set_node_scale (struct cgraph_node *node, gcov_type count)
202 IPA_NODE_REF (node)->count_scale = count;
205 /* Return whether LAT is a constant lattice. */
207 ipcp_lat_is_const (struct ipcp_lattice *lat)
209 if (lat->type == IPA_CONST_VALUE)
215 /* Return whether LAT is a constant lattice that ipa-cp can actually insert
216 into the code (i.e. constants excluding member pointers and pointers). */
218 ipcp_lat_is_insertable (struct ipcp_lattice *lat)
220 return lat->type == IPA_CONST_VALUE;
223 /* Return true if LAT1 and LAT2 are equal. */
225 ipcp_lats_are_equal (struct ipcp_lattice *lat1, struct ipcp_lattice *lat2)
227 gcc_assert (ipcp_lat_is_const (lat1) && ipcp_lat_is_const (lat2));
228 if (lat1->type != lat2->type)
231 if (TREE_CODE (lat1->constant) == ADDR_EXPR
232 && TREE_CODE (lat2->constant) == ADDR_EXPR
233 && TREE_CODE (TREE_OPERAND (lat1->constant, 0)) == CONST_DECL
234 && TREE_CODE (TREE_OPERAND (lat2->constant, 0)) == CONST_DECL)
235 return operand_equal_p (DECL_INITIAL (TREE_OPERAND (lat1->constant, 0)),
236 DECL_INITIAL (TREE_OPERAND (lat2->constant, 0)), 0);
238 return operand_equal_p (lat1->constant, lat2->constant, 0);
241 /* Compute Meet arithmetics:
242 Meet (IPA_BOTTOM, x) = IPA_BOTTOM
244 Meet (const_a,const_b) = IPA_BOTTOM, if const_a != const_b.
245 MEET (const_a,const_b) = const_a, if const_a == const_b.*/
247 ipa_lattice_meet (struct ipcp_lattice *res, struct ipcp_lattice *lat1,
248 struct ipcp_lattice *lat2)
250 if (lat1->type == IPA_BOTTOM || lat2->type == IPA_BOTTOM)
252 res->type = IPA_BOTTOM;
255 if (lat1->type == IPA_TOP)
257 res->type = lat2->type;
258 res->constant = lat2->constant;
261 if (lat2->type == IPA_TOP)
263 res->type = lat1->type;
264 res->constant = lat1->constant;
267 if (!ipcp_lats_are_equal (lat1, lat2))
269 res->type = IPA_BOTTOM;
272 res->type = lat1->type;
273 res->constant = lat1->constant;
276 /* Return the lattice corresponding to the Ith formal parameter of the function
277 described by INFO. */
278 static inline struct ipcp_lattice *
279 ipcp_get_lattice (struct ipa_node_params *info, int i)
281 return &(info->params[i].ipcp_lattice);
284 /* Given the jump function JFUNC, compute the lattice LAT that describes the
285 value coming down the callsite. INFO describes the caller node so that
286 pass-through jump functions can be evaluated. */
288 ipcp_lattice_from_jfunc (struct ipa_node_params *info, struct ipcp_lattice *lat,
289 struct ipa_jump_func *jfunc)
291 if (jfunc->type == IPA_JF_CONST)
293 lat->type = IPA_CONST_VALUE;
294 lat->constant = jfunc->value.constant;
296 else if (jfunc->type == IPA_JF_PASS_THROUGH)
298 struct ipcp_lattice *caller_lat;
301 caller_lat = ipcp_get_lattice (info, jfunc->value.pass_through.formal_id);
302 lat->type = caller_lat->type;
303 if (caller_lat->type != IPA_CONST_VALUE)
305 cst = caller_lat->constant;
307 if (jfunc->value.pass_through.operation != NOP_EXPR)
310 if (TREE_CODE_CLASS (jfunc->value.pass_through.operation)
312 restype = boolean_type_node;
314 restype = TREE_TYPE (cst);
315 cst = fold_binary (jfunc->value.pass_through.operation,
316 restype, cst, jfunc->value.pass_through.operand);
318 if (!cst || !is_gimple_ip_invariant (cst))
319 lat->type = IPA_BOTTOM;
322 else if (jfunc->type == IPA_JF_ANCESTOR)
324 struct ipcp_lattice *caller_lat;
328 caller_lat = ipcp_get_lattice (info, jfunc->value.ancestor.formal_id);
329 lat->type = caller_lat->type;
330 if (caller_lat->type != IPA_CONST_VALUE)
332 if (TREE_CODE (caller_lat->constant) != ADDR_EXPR)
334 /* This can happen when the constant is a NULL pointer. */
335 lat->type = IPA_BOTTOM;
338 t = TREE_OPERAND (caller_lat->constant, 0);
339 ok = build_ref_for_offset (&t, TREE_TYPE (t),
340 jfunc->value.ancestor.offset,
341 jfunc->value.ancestor.type, false);
344 lat->type = IPA_BOTTOM;
345 lat->constant = NULL_TREE;
348 lat->constant = build_fold_addr_expr (t);
351 lat->type = IPA_BOTTOM;
354 /* True when OLD_LAT and NEW_LAT values are not the same. */
357 ipcp_lattice_changed (struct ipcp_lattice *old_lat,
358 struct ipcp_lattice *new_lat)
360 if (old_lat->type == new_lat->type)
362 if (!ipcp_lat_is_const (old_lat))
364 if (ipcp_lats_are_equal (old_lat, new_lat))
370 /* Print all ipcp_lattices of all functions to F. */
372 ipcp_print_all_lattices (FILE * f)
374 struct cgraph_node *node;
377 fprintf (f, "\nLattice:\n");
378 for (node = cgraph_nodes; node; node = node->next)
380 struct ipa_node_params *info;
384 info = IPA_NODE_REF (node);
385 fprintf (f, " Node: %s:\n", cgraph_node_name (node));
386 count = ipa_get_param_count (info);
387 for (i = 0; i < count; i++)
389 struct ipcp_lattice *lat = ipcp_get_lattice (info, i);
391 fprintf (f, " param [%d]: ", i);
392 if (lat->type == IPA_CONST_VALUE)
394 tree cst = lat->constant;
395 fprintf (f, "type is CONST ");
396 print_generic_expr (f, cst, 0);
397 if (TREE_CODE (cst) == ADDR_EXPR
398 && TREE_CODE (TREE_OPERAND (cst, 0)) == CONST_DECL)
401 print_generic_expr (f, DECL_INITIAL (TREE_OPERAND (cst, 0)),
406 else if (lat->type == IPA_TOP)
407 fprintf (f, "type is TOP\n");
409 fprintf (f, "type is BOTTOM\n");
414 /* Return true if ipcp algorithms would allow cloning NODE. */
417 ipcp_versionable_function_p (struct cgraph_node *node)
419 struct cgraph_edge *edge;
421 /* There are a number of generic reasons functions cannot be versioned. */
422 if (!node->local.versionable)
425 /* Removing arguments doesn't work if the function takes varargs
426 or use __builtin_apply_args. */
427 for (edge = node->callees; edge; edge = edge->next_callee)
429 tree t = edge->callee->decl;
430 if (DECL_BUILT_IN_CLASS (t) == BUILT_IN_NORMAL
431 && (DECL_FUNCTION_CODE (t) == BUILT_IN_APPLY_ARGS
432 || DECL_FUNCTION_CODE (t) == BUILT_IN_VA_START))
439 /* Return true if this NODE is viable candidate for cloning. */
441 ipcp_cloning_candidate_p (struct cgraph_node *node)
445 gcov_type direct_call_sum = 0;
446 struct cgraph_edge *e;
448 /* We never clone functions that are not visible from outside.
449 FIXME: in future we should clone such functions when they are called with
450 different constants, but current ipcp implementation is not good on this.
452 if (cgraph_only_called_directly_p (node) || !node->analyzed)
455 if (cgraph_function_body_availability (node) <= AVAIL_OVERWRITABLE)
458 fprintf (dump_file, "Not considering %s for cloning; body is overwrittable.\n",
459 cgraph_node_name (node));
462 if (!ipcp_versionable_function_p (node))
465 fprintf (dump_file, "Not considering %s for cloning; body is not versionable.\n",
466 cgraph_node_name (node));
469 for (e = node->callers; e; e = e->next_caller)
471 direct_call_sum += e->count;
473 if (cgraph_maybe_hot_edge_p (e))
480 fprintf (dump_file, "Not considering %s for cloning; no direct calls.\n",
481 cgraph_node_name (node));
484 if (node->local.inline_summary.self_size < n_calls)
487 fprintf (dump_file, "Considering %s for cloning; code would shrink.\n",
488 cgraph_node_name (node));
492 if (!flag_ipa_cp_clone)
495 fprintf (dump_file, "Not considering %s for cloning; -fipa-cp-clone disabled.\n",
496 cgraph_node_name (node));
500 if (!optimize_function_for_speed_p (DECL_STRUCT_FUNCTION (node->decl)))
503 fprintf (dump_file, "Not considering %s for cloning; optimizing it for size.\n",
504 cgraph_node_name (node));
508 /* When profile is available and function is hot, propagate into it even if
509 calls seems cold; constant propagation can improve function's speed
513 if (direct_call_sum > node->count * 90 / 100)
516 fprintf (dump_file, "Considering %s for cloning; usually called directly.\n",
517 cgraph_node_name (node));
524 fprintf (dump_file, "Not considering %s for cloning; no hot calls.\n",
525 cgraph_node_name (node));
529 fprintf (dump_file, "Considering %s for cloning.\n",
530 cgraph_node_name (node));
534 /* Initialize ipcp_lattices array. The lattices corresponding to supported
535 types (integers, real types and Fortran constants defined as const_decls)
536 are initialized to IPA_TOP, the rest of them to IPA_BOTTOM. */
538 ipcp_initialize_node_lattices (struct cgraph_node *node)
541 struct ipa_node_params *info = IPA_NODE_REF (node);
542 enum ipa_lattice_type type;
544 if (ipa_is_called_with_var_arguments (info))
546 else if (cgraph_only_called_directly_p (node))
548 /* When cloning is allowed, we can assume that externally visible functions
549 are not called. We will compensate this by cloning later. */
550 else if (ipcp_cloning_candidate_p (node))
551 type = IPA_TOP, n_cloning_candidates ++;
555 for (i = 0; i < ipa_get_param_count (info) ; i++)
556 ipcp_get_lattice (info, i)->type = type;
559 /* build INTEGER_CST tree with type TREE_TYPE and value according to LAT.
562 build_const_val (struct ipcp_lattice *lat, tree tree_type)
566 gcc_assert (ipcp_lat_is_const (lat));
569 if (!useless_type_conversion_p (tree_type, TREE_TYPE (val)))
571 if (fold_convertible_p (tree_type, val))
572 return fold_build1 (NOP_EXPR, tree_type, val);
574 return fold_build1 (VIEW_CONVERT_EXPR, tree_type, val);
579 /* Compute the proper scale for NODE. It is the ratio between the number of
580 direct calls (represented on the incoming cgraph_edges) and sum of all
581 invocations of NODE (represented as count in cgraph_node).
583 FIXME: This code is wrong. Since the callers can be also clones and
584 the clones are not scaled yet, the sums gets unrealistically high.
585 To properly compute the counts, we would need to do propagation across
586 callgraph (as external call to A might imply call to non-clonned B
587 if A's clone calls clonned B). */
589 ipcp_compute_node_scale (struct cgraph_node *node)
592 struct cgraph_edge *cs;
595 /* Compute sum of all counts of callers. */
596 for (cs = node->callers; cs != NULL; cs = cs->next_caller)
598 /* Work around the unrealistically high sum problem. We just don't want
599 the non-cloned body to have negative or very low frequency. Since
600 majority of execution time will be spent in clones anyway, this should
601 give good enough profile. */
602 if (sum > node->count * 9 / 10)
603 sum = node->count * 9 / 10;
604 if (node->count == 0)
605 ipcp_set_node_scale (node, 0);
607 ipcp_set_node_scale (node, sum * REG_BR_PROB_BASE / node->count);
610 /* Initialization and computation of IPCP data structures. This is the initial
611 intraprocedural analysis of functions, which gathers information to be
612 propagated later on. */
614 ipcp_init_stage (void)
616 struct cgraph_node *node;
617 struct cgraph_edge *cs;
619 for (node = cgraph_nodes; node; node = node->next)
621 ipcp_analyze_node (node);
622 for (node = cgraph_nodes; node; node = node->next)
626 /* building jump functions */
627 for (cs = node->callees; cs; cs = cs->next_callee)
629 /* We do not need to bother analyzing calls to unknown
630 functions unless they may become known during lto/whopr. */
631 if (!cs->callee->analyzed && !flag_lto && !flag_whopr)
633 ipa_count_arguments (cs);
634 if (ipa_get_cs_argument_count (IPA_EDGE_REF (cs))
635 != ipa_get_param_count (IPA_NODE_REF (cs->callee)))
636 ipa_set_called_with_variable_arg (IPA_NODE_REF (cs->callee));
637 ipa_compute_jump_functions (cs);
642 /* Return true if there are some formal parameters whose value is IPA_TOP (in
643 the whole compilation unit). Change their values to IPA_BOTTOM, since they
644 most probably get their values from outside of this compilation unit. */
646 ipcp_change_tops_to_bottom (void)
649 struct cgraph_node *node;
653 for (node = cgraph_nodes; node; node = node->next)
655 struct ipa_node_params *info = IPA_NODE_REF (node);
656 count = ipa_get_param_count (info);
657 for (i = 0; i < count; i++)
659 struct ipcp_lattice *lat = ipcp_get_lattice (info, i);
660 if (lat->type == IPA_TOP)
665 fprintf (dump_file, "Forcing param ");
666 print_generic_expr (dump_file, ipa_get_param (info, i), 0);
667 fprintf (dump_file, " of node %s to bottom.\n",
668 cgraph_node_name (node));
670 lat->type = IPA_BOTTOM;
677 /* Interprocedural analysis. The algorithm propagates constants from the
678 caller's parameters to the callee's arguments. */
680 ipcp_propagate_stage (void)
683 struct ipcp_lattice inc_lat = { IPA_BOTTOM, NULL };
684 struct ipcp_lattice new_lat = { IPA_BOTTOM, NULL };
685 struct ipcp_lattice *dest_lat;
686 struct cgraph_edge *cs;
687 struct ipa_jump_func *jump_func;
688 struct ipa_func_list *wl;
691 ipa_check_create_node_params ();
692 ipa_check_create_edge_args ();
694 /* Initialize worklist to contain all functions. */
695 wl = ipa_init_func_list ();
698 struct cgraph_node *node = ipa_pop_func_from_list (&wl);
699 struct ipa_node_params *info = IPA_NODE_REF (node);
701 for (cs = node->callees; cs; cs = cs->next_callee)
703 struct ipa_node_params *callee_info = IPA_NODE_REF (cs->callee);
704 struct ipa_edge_args *args = IPA_EDGE_REF (cs);
706 if (ipa_is_called_with_var_arguments (callee_info)
707 || !cs->callee->analyzed
708 || ipa_is_called_with_var_arguments (callee_info))
711 count = ipa_get_cs_argument_count (args);
712 for (i = 0; i < count; i++)
714 jump_func = ipa_get_ith_jump_func (args, i);
715 ipcp_lattice_from_jfunc (info, &inc_lat, jump_func);
716 dest_lat = ipcp_get_lattice (callee_info, i);
717 ipa_lattice_meet (&new_lat, &inc_lat, dest_lat);
718 if (ipcp_lattice_changed (&new_lat, dest_lat))
720 dest_lat->type = new_lat.type;
721 dest_lat->constant = new_lat.constant;
722 ipa_push_func_to_list (&wl, cs->callee);
729 /* Call the constant propagation algorithm and re-call it if necessary
730 (if there are undetermined values left). */
732 ipcp_iterate_stage (void)
734 struct cgraph_node *node;
735 n_cloning_candidates = 0;
738 fprintf (dump_file, "\nIPA iterate stage:\n\n");
741 ipa_update_after_lto_read ();
743 for (node = cgraph_nodes; node; node = node->next)
745 ipcp_initialize_node_lattices (node);
746 ipcp_compute_node_scale (node);
748 if (dump_file && (dump_flags & TDF_DETAILS))
750 ipcp_print_all_lattices (dump_file);
751 ipcp_function_scale_print (dump_file);
754 ipcp_propagate_stage ();
755 if (ipcp_change_tops_to_bottom ())
756 /* Some lattices have changed from IPA_TOP to IPA_BOTTOM.
757 This change should be propagated. */
759 gcc_assert (n_cloning_candidates);
760 ipcp_propagate_stage ();
764 fprintf (dump_file, "\nIPA lattices after propagation:\n");
765 ipcp_print_all_lattices (dump_file);
766 if (dump_flags & TDF_DETAILS)
767 ipcp_print_profile_data (dump_file);
771 /* Check conditions to forbid constant insertion to function described by
774 ipcp_node_modifiable_p (struct cgraph_node *node)
776 /* Once we will be able to do in-place replacement, we can be more
778 return ipcp_versionable_function_p (node);
781 /* Print count scale data structures. */
783 ipcp_function_scale_print (FILE * f)
785 struct cgraph_node *node;
787 for (node = cgraph_nodes; node; node = node->next)
791 fprintf (f, "printing scale for %s: ", cgraph_node_name (node));
792 fprintf (f, "value is " HOST_WIDE_INT_PRINT_DEC
793 " \n", (HOST_WIDE_INT) ipcp_get_node_scale (node));
797 /* Print counts of all cgraph nodes. */
799 ipcp_print_func_profile_counts (FILE * f)
801 struct cgraph_node *node;
803 for (node = cgraph_nodes; node; node = node->next)
805 fprintf (f, "function %s: ", cgraph_node_name (node));
806 fprintf (f, "count is " HOST_WIDE_INT_PRINT_DEC
807 " \n", (HOST_WIDE_INT) node->count);
811 /* Print counts of all cgraph edges. */
813 ipcp_print_call_profile_counts (FILE * f)
815 struct cgraph_node *node;
816 struct cgraph_edge *cs;
818 for (node = cgraph_nodes; node; node = node->next)
820 for (cs = node->callees; cs; cs = cs->next_callee)
822 fprintf (f, "%s -> %s ", cgraph_node_name (cs->caller),
823 cgraph_node_name (cs->callee));
824 fprintf (f, "count is " HOST_WIDE_INT_PRINT_DEC " \n",
825 (HOST_WIDE_INT) cs->count);
830 /* Print profile info for all functions. */
832 ipcp_print_profile_data (FILE * f)
834 fprintf (f, "\nNODE COUNTS :\n");
835 ipcp_print_func_profile_counts (f);
836 fprintf (f, "\nCS COUNTS stage:\n");
837 ipcp_print_call_profile_counts (f);
840 /* Build and initialize ipa_replace_map struct according to LAT. This struct is
841 processed by versioning, which operates according to the flags set.
842 PARM_TREE is the formal parameter found to be constant. LAT represents the
844 static struct ipa_replace_map *
845 ipcp_create_replace_map (tree parm_tree, struct ipcp_lattice *lat)
847 struct ipa_replace_map *replace_map;
850 replace_map = GGC_NEW (struct ipa_replace_map);
851 const_val = build_const_val (lat, TREE_TYPE (parm_tree));
854 fprintf (dump_file, " replacing param ");
855 print_generic_expr (dump_file, parm_tree, 0);
856 fprintf (dump_file, " with const ");
857 print_generic_expr (dump_file, const_val, 0);
858 fprintf (dump_file, "\n");
860 replace_map->old_tree = parm_tree;
861 replace_map->new_tree = const_val;
862 replace_map->replace_p = true;
863 replace_map->ref_p = false;
868 /* Return true if this callsite should be redirected to the original callee
869 (instead of the cloned one). */
871 ipcp_need_redirect_p (struct cgraph_edge *cs)
873 struct ipa_node_params *orig_callee_info;
875 struct ipa_jump_func *jump_func;
876 struct cgraph_node *node = cs->callee, *orig;
878 if (!n_cloning_candidates)
881 if ((orig = ipcp_get_orig_node (node)) != NULL)
883 if (ipcp_get_orig_node (cs->caller))
886 orig_callee_info = IPA_NODE_REF (node);
887 count = ipa_get_param_count (orig_callee_info);
888 for (i = 0; i < count; i++)
890 struct ipcp_lattice *lat = ipcp_get_lattice (orig_callee_info, i);
891 if (ipcp_lat_is_const (lat))
893 jump_func = ipa_get_ith_jump_func (IPA_EDGE_REF (cs), i);
894 if (jump_func->type != IPA_JF_CONST)
902 /* Fix the callsites and the call graph after function cloning was done. */
904 ipcp_update_callgraph (void)
906 struct cgraph_node *node;
908 for (node = cgraph_nodes; node; node = node->next)
909 if (node->analyzed && ipcp_node_is_clone (node))
911 bitmap args_to_skip = BITMAP_ALLOC (NULL);
912 struct cgraph_node *orig_node = ipcp_get_orig_node (node);
913 struct ipa_node_params *info = IPA_NODE_REF (orig_node);
914 int i, count = ipa_get_param_count (info);
915 struct cgraph_edge *cs, *next;
917 for (i = 0; i < count; i++)
919 struct ipcp_lattice *lat = ipcp_get_lattice (info, i);
920 tree parm_tree = ipa_get_param (info, i);
922 /* We can proactively remove obviously unused arguments. */
923 if (is_gimple_reg (parm_tree)
924 && !gimple_default_def (DECL_STRUCT_FUNCTION (orig_node->decl),
927 bitmap_set_bit (args_to_skip, i);
931 if (lat->type == IPA_CONST_VALUE)
932 bitmap_set_bit (args_to_skip, i);
934 for (cs = node->callers; cs; cs = next)
936 next = cs->next_caller;
937 if (!ipcp_node_is_clone (cs->caller) && ipcp_need_redirect_p (cs))
938 cgraph_redirect_edge_callee (cs, orig_node);
943 /* Update profiling info for versioned functions and the functions they were
946 ipcp_update_profiling (void)
948 struct cgraph_node *node, *orig_node;
949 gcov_type scale, scale_complement;
950 struct cgraph_edge *cs;
952 for (node = cgraph_nodes; node; node = node->next)
954 if (ipcp_node_is_clone (node))
956 orig_node = ipcp_get_orig_node (node);
957 scale = ipcp_get_node_scale (orig_node);
958 node->count = orig_node->count * scale / REG_BR_PROB_BASE;
959 scale_complement = REG_BR_PROB_BASE - scale;
961 orig_node->count * scale_complement / REG_BR_PROB_BASE;
962 for (cs = node->callees; cs; cs = cs->next_callee)
963 cs->count = cs->count * scale / REG_BR_PROB_BASE;
964 for (cs = orig_node->callees; cs; cs = cs->next_callee)
965 cs->count = cs->count * scale_complement / REG_BR_PROB_BASE;
970 /* If NODE was cloned, how much would program grow? */
972 ipcp_estimate_growth (struct cgraph_node *node)
974 struct cgraph_edge *cs;
975 int redirectable_node_callers = 0;
976 int removable_args = 0;
977 bool need_original = !cgraph_only_called_directly_p (node);
978 struct ipa_node_params *info;
982 for (cs = node->callers; cs != NULL; cs = cs->next_caller)
983 if (cs->caller == node || !ipcp_need_redirect_p (cs))
984 redirectable_node_callers++;
986 need_original = true;
988 /* If we will be able to fully replace orignal node, we never increase
993 info = IPA_NODE_REF (node);
994 count = ipa_get_param_count (info);
995 for (i = 0; i < count; i++)
997 struct ipcp_lattice *lat = ipcp_get_lattice (info, i);
998 tree parm_tree = ipa_get_param (info, i);
1000 /* We can proactively remove obviously unused arguments. */
1001 if (is_gimple_reg (parm_tree)
1002 && !gimple_default_def (DECL_STRUCT_FUNCTION (node->decl),
1006 if (lat->type == IPA_CONST_VALUE)
1010 /* We make just very simple estimate of savings for removal of operand from
1011 call site. Precise cost is dificult to get, as our size metric counts
1012 constants and moves as free. Generally we are looking for cases that
1013 small function is called very many times. */
1014 growth = node->local.inline_summary.self_size
1015 - removable_args * redirectable_node_callers;
1022 /* Estimate cost of cloning NODE. */
1024 ipcp_estimate_cloning_cost (struct cgraph_node *node)
1027 gcov_type count_sum = 1;
1028 struct cgraph_edge *e;
1031 cost = ipcp_estimate_growth (node) * 1000;
1035 fprintf (dump_file, "Versioning of %s will save code size\n",
1036 cgraph_node_name (node));
1040 for (e = node->callers; e; e = e->next_caller)
1041 if (!bitmap_bit_p (dead_nodes, e->caller->uid)
1042 && !ipcp_need_redirect_p (e))
1044 count_sum += e->count;
1045 freq_sum += e->frequency + 1;
1049 cost /= count_sum * 1000 / max_count + 1;
1051 cost /= freq_sum * 1000 / REG_BR_PROB_BASE + 1;
1053 fprintf (dump_file, "Cost of versioning %s is %i, (size: %i, freq: %i)\n",
1054 cgraph_node_name (node), cost, node->local.inline_summary.self_size,
1059 /* Return number of live constant parameters. */
1061 ipcp_const_param_count (struct cgraph_node *node)
1063 int const_param = 0;
1064 struct ipa_node_params *info = IPA_NODE_REF (node);
1065 int count = ipa_get_param_count (info);
1068 for (i = 0; i < count; i++)
1070 struct ipcp_lattice *lat = ipcp_get_lattice (info, i);
1071 tree parm_tree = ipa_get_param (info, i);
1072 if (ipcp_lat_is_insertable (lat)
1073 /* Do not count obviously unused arguments. */
1074 && (!is_gimple_reg (parm_tree)
1075 || gimple_default_def (DECL_STRUCT_FUNCTION (node->decl),
1082 /* Propagate the constant parameters found by ipcp_iterate_stage()
1083 to the function's code. */
1085 ipcp_insert_stage (void)
1087 struct cgraph_node *node, *node1 = NULL;
1089 VEC (cgraph_edge_p, heap) * redirect_callers;
1090 VEC (ipa_replace_map_p,gc)* replace_trees;
1091 int node_callers, count;
1093 struct ipa_replace_map *replace_param;
1095 long overall_size = 0, new_size = 0;
1098 ipa_check_create_node_params ();
1099 ipa_check_create_edge_args ();
1101 fprintf (dump_file, "\nIPA insert stage:\n\n");
1103 dead_nodes = BITMAP_ALLOC (NULL);
1105 for (node = cgraph_nodes; node; node = node->next)
1108 if (node->count > max_count)
1109 max_count = node->count;
1110 overall_size += node->local.inline_summary.self_size;
1113 max_new_size = overall_size;
1114 if (max_new_size < PARAM_VALUE (PARAM_LARGE_UNIT_INSNS))
1115 max_new_size = PARAM_VALUE (PARAM_LARGE_UNIT_INSNS);
1116 max_new_size = max_new_size * PARAM_VALUE (PARAM_IPCP_UNIT_GROWTH) / 100 + 1;
1118 /* First collect all functions we proved to have constant arguments to heap. */
1119 heap = fibheap_new ();
1120 for (node = cgraph_nodes; node; node = node->next)
1122 struct ipa_node_params *info;
1123 /* Propagation of the constant is forbidden in certain conditions. */
1124 if (!node->analyzed || !ipcp_node_modifiable_p (node))
1126 info = IPA_NODE_REF (node);
1127 if (ipa_is_called_with_var_arguments (info))
1129 if (ipcp_const_param_count (node))
1130 node->aux = fibheap_insert (heap, ipcp_estimate_cloning_cost (node), node);
1133 /* Now clone in priority order until code size growth limits are met or
1135 while (!fibheap_empty (heap))
1137 struct ipa_node_params *info;
1139 bitmap args_to_skip;
1140 struct cgraph_edge *cs;
1142 node = (struct cgraph_node *)fibheap_extract_min (heap);
1145 fprintf (dump_file, "considering function %s\n",
1146 cgraph_node_name (node));
1148 growth = ipcp_estimate_growth (node);
1150 if (new_size + growth > max_new_size)
1153 && optimize_function_for_size_p (DECL_STRUCT_FUNCTION (node->decl)))
1156 fprintf (dump_file, "Not versioning, cold code would grow");
1162 /* Look if original function becomes dead after clonning. */
1163 for (cs = node->callers; cs != NULL; cs = cs->next_caller)
1164 if (cs->caller == node || ipcp_need_redirect_p (cs))
1166 if (!cs && cgraph_only_called_directly_p (node))
1167 bitmap_set_bit (dead_nodes, node->uid);
1169 info = IPA_NODE_REF (node);
1170 count = ipa_get_param_count (info);
1172 replace_trees = VEC_alloc (ipa_replace_map_p, gc, 1);
1173 args_to_skip = BITMAP_GGC_ALLOC ();
1174 for (i = 0; i < count; i++)
1176 struct ipcp_lattice *lat = ipcp_get_lattice (info, i);
1177 parm_tree = ipa_get_param (info, i);
1179 /* We can proactively remove obviously unused arguments. */
1180 if (is_gimple_reg (parm_tree)
1181 && !gimple_default_def (DECL_STRUCT_FUNCTION (node->decl),
1184 bitmap_set_bit (args_to_skip, i);
1188 if (lat->type == IPA_CONST_VALUE)
1191 ipcp_create_replace_map (parm_tree, lat);
1192 VEC_safe_push (ipa_replace_map_p, gc, replace_trees, replace_param);
1193 bitmap_set_bit (args_to_skip, i);
1197 /* Compute how many callers node has. */
1199 for (cs = node->callers; cs != NULL; cs = cs->next_caller)
1201 redirect_callers = VEC_alloc (cgraph_edge_p, heap, node_callers);
1202 for (cs = node->callers; cs != NULL; cs = cs->next_caller)
1203 VEC_quick_push (cgraph_edge_p, redirect_callers, cs);
1205 /* Redirecting all the callers of the node to the
1206 new versioned node. */
1208 cgraph_create_virtual_clone (node, redirect_callers, replace_trees,
1210 args_to_skip = NULL;
1211 VEC_free (cgraph_edge_p, heap, redirect_callers);
1212 replace_trees = NULL;
1217 fprintf (dump_file, "versioned function %s with growth %i, overall %i\n",
1218 cgraph_node_name (node), (int)growth, (int)new_size);
1219 ipcp_init_cloned_node (node, node1);
1221 /* TODO: We can use indirect inlning info to produce new calls. */
1224 dump_function_to_file (node1->decl, dump_file, dump_flags);
1226 for (cs = node->callees; cs; cs = cs->next_callee)
1227 if (cs->callee->aux)
1229 fibheap_delete_node (heap, (fibnode_t) cs->callee->aux);
1230 cs->callee->aux = fibheap_insert (heap,
1231 ipcp_estimate_cloning_cost (cs->callee),
1236 while (!fibheap_empty (heap))
1239 fprintf (dump_file, "skipping function %s\n",
1240 cgraph_node_name (node));
1241 node = (struct cgraph_node *) fibheap_extract_min (heap);
1244 fibheap_delete (heap);
1245 BITMAP_FREE (dead_nodes);
1246 ipcp_update_callgraph ();
1247 ipcp_update_profiling ();
1250 /* The IPCP driver. */
1254 cgraph_remove_unreachable_nodes (true,dump_file);
1257 fprintf (dump_file, "\nIPA structures before propagation:\n");
1258 if (dump_flags & TDF_DETAILS)
1259 ipa_print_all_params (dump_file);
1260 ipa_print_all_jump_functions (dump_file);
1262 /* 2. Do the interprocedural propagation. */
1263 ipcp_iterate_stage ();
1264 /* 3. Insert the constants found to the functions. */
1265 ipcp_insert_stage ();
1266 if (dump_file && (dump_flags & TDF_DETAILS))
1268 fprintf (dump_file, "\nProfiling info after insert stage:\n");
1269 ipcp_print_profile_data (dump_file);
1271 /* Free all IPCP structures. */
1272 ipa_free_all_structures_after_ipa_cp ();
1274 fprintf (dump_file, "\nIPA constant propagation end\n");
1278 /* Note function body size. */
1280 ipcp_generate_summary (void)
1283 fprintf (dump_file, "\nIPA constant propagation start:\n");
1284 ipa_check_create_node_params ();
1285 ipa_check_create_edge_args ();
1286 ipa_register_cgraph_hooks ();
1287 /* 1. Call the init stage to initialize
1288 the ipa_node_params and ipa_edge_args structures. */
1292 /* Write ipcp summary for nodes in SET. */
1294 ipcp_write_summary (cgraph_node_set set,
1295 varpool_node_set vset ATTRIBUTE_UNUSED)
1297 ipa_prop_write_jump_functions (set);
1300 /* Read ipcp summary. */
1302 ipcp_read_summary (void)
1304 ipa_prop_read_jump_functions ();
1307 /* Gate for IPCP optimization. */
1309 cgraph_gate_cp (void)
1314 struct ipa_opt_pass_d pass_ipa_cp =
1319 cgraph_gate_cp, /* gate */
1320 ipcp_driver, /* execute */
1323 0, /* static_pass_number */
1324 TV_IPA_CONSTANT_PROP, /* tv_id */
1325 0, /* properties_required */
1326 0, /* properties_provided */
1327 0, /* properties_destroyed */
1328 0, /* todo_flags_start */
1329 TODO_dump_cgraph | TODO_dump_func |
1330 TODO_remove_functions | TODO_ggc_collect /* todo_flags_finish */
1332 ipcp_generate_summary, /* generate_summary */
1333 ipcp_write_summary, /* write_summary */
1334 ipcp_read_summary, /* read_summary */
1335 NULL, /* write_optimization_summary */
1336 NULL, /* read_optimization_summary */
1337 NULL, /* stmt_fixup */
1339 NULL, /* function_transform */
1340 NULL, /* variable_transform */