1 /* Interprocedural constant propagation
2 Copyright (C) 2005-2013 Free Software Foundation, Inc.
4 Contributed by Razya Ladelsky <RAZYA@il.ibm.com> and Martin Jambor
7 This file is part of GCC.
9 GCC is free software; you can redistribute it and/or modify it under
10 the terms of the GNU General Public License as published by the Free
11 Software Foundation; either version 3, or (at your option) any later
14 GCC is distributed in the hope that it will be useful, but WITHOUT ANY
15 WARRANTY; without even the implied warranty of MERCHANTABILITY or
16 FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
19 You should have received a copy of the GNU General Public License
20 along with GCC; see the file COPYING3. If not see
21 <http://www.gnu.org/licenses/>. */
23 /* Interprocedural constant propagation (IPA-CP).
25 The goal of this transformation is to
27 1) discover functions which are always invoked with some arguments with the
28 same known constant values and modify the functions so that the
29 subsequent optimizations can take advantage of the knowledge, and
31 2) partial specialization - create specialized versions of functions
32 transformed in this way if some parameters are known constants only in
33 certain contexts but the estimated tradeoff between speedup and cost size
36 The algorithm also propagates types and attempts to perform type based
37 devirtualization. Types are propagated much like constants.
39 The algorithm basically consists of three stages. In the first, functions
40 are analyzed one at a time and jump functions are constructed for all known
41 call-sites. In the second phase, the pass propagates information from the
42 jump functions across the call to reveal what values are available at what
43 call sites, performs estimations of effects of known values on functions and
44 their callees, and finally decides what specialized extra versions should be
45 created. In the third, the special versions materialize and appropriate
48 The algorithm used is to a certain extent based on "Interprocedural Constant
49 Propagation", by David Callahan, Keith D Cooper, Ken Kennedy, Linda Torczon,
50 Comp86, pg 152-161 and "A Methodology for Procedure Cloning" by Keith D
51 Cooper, Mary W. Hall, and Ken Kennedy.
54 First stage - intraprocedural analysis
55 =======================================
57 This phase computes jump_function and modification flags.
59 A jump function for a call-site represents the values passed as an actual
60 arguments of a given call-site. In principle, there are three types of
63 Pass through - the caller's formal parameter is passed as an actual
64 argument, plus an operation on it can be performed.
65 Constant - a constant is passed as an actual argument.
66 Unknown - neither of the above.
68 All jump function types are described in detail in ipa-prop.h, together with
69 the data structures that represent them and methods of accessing them.
71 ipcp_generate_summary() is the main function of the first stage.
73 Second stage - interprocedural analysis
74 ========================================
76 This stage is itself divided into two phases. In the first, we propagate
77 known values over the call graph, in the second, we make cloning decisions.
78 It uses a different algorithm than the original Callahan's paper.
80 First, we traverse the functions topologically from callers to callees and,
81 for each strongly connected component (SCC), we propagate constants
82 according to previously computed jump functions. We also record what known
83 values depend on other known values and estimate local effects. Finally, we
84 propagate cumulative information about these effects from dependent values
85 to those on which they depend.
87 Second, we again traverse the call graph in the same topological order and
88 make clones for functions which we know are called with the same values in
89 all contexts and decide about extra specialized clones of functions just for
90 some contexts - these decisions are based on both local estimates and
91 cumulative estimates propagated from callees.
93 ipcp_propagate_stage() and ipcp_decision_stage() together constitute the
96 Third phase - materialization of clones, call statement updates.
97 ============================================
99 This stage is currently performed by call graph code (mainly in cgraphunit.c
100 and tree-inline.c) according to instructions inserted to the call graph by
105 #include "coretypes.h"
110 #include "ipa-prop.h"
111 #include "tree-flow.h"
112 #include "tree-pass.h"
114 #include "diagnostic.h"
115 #include "tree-pretty-print.h"
116 #include "tree-inline.h"
118 #include "ipa-inline.h"
119 #include "ipa-utils.h"
123 /* Describes a particular source for an IPA-CP value. */
125 struct ipcp_value_source
127 /* Aggregate offset of the source, negative if the source is scalar value of
128 the argument itself. */
129 HOST_WIDE_INT offset;
130 /* The incoming edge that brought the value. */
131 struct cgraph_edge *cs;
132 /* If the jump function that resulted into his value was a pass-through or an
133 ancestor, this is the ipcp_value of the caller from which the described
134 value has been derived. Otherwise it is NULL. */
135 struct ipcp_value *val;
136 /* Next pointer in a linked list of sources of a value. */
137 struct ipcp_value_source *next;
138 /* If the jump function that resulted into his value was a pass-through or an
139 ancestor, this is the index of the parameter of the caller the jump
140 function references. */
144 /* Describes one particular value stored in struct ipcp_lattice. */
148 /* The actual value for the given parameter. This is either an IPA invariant
149 or a TREE_BINFO describing a type that can be used for
152 /* The list of sources from which this value originates. */
153 struct ipcp_value_source *sources;
154 /* Next pointers in a linked list of all values in a lattice. */
155 struct ipcp_value *next;
156 /* Next pointers in a linked list of values in a strongly connected component
158 struct ipcp_value *scc_next;
159 /* Next pointers in a linked list of SCCs of values sorted topologically
160 according their sources. */
161 struct ipcp_value *topo_next;
162 /* A specialized node created for this value, NULL if none has been (so far)
164 struct cgraph_node *spec_node;
165 /* Depth first search number and low link for topological sorting of
168 /* Time benefit and size cost that specializing the function for this value
169 would bring about in this function alone. */
170 int local_time_benefit, local_size_cost;
171 /* Time benefit and size cost that specializing the function for this value
172 can bring about in it's callees (transitively). */
173 int prop_time_benefit, prop_size_cost;
174 /* True if this valye is currently on the topo-sort stack. */
178 /* Lattice describing potential values of a formal parameter of a function, or
179 a part of an aggreagate. TOP is represented by a lattice with zero values
180 and with contains_variable and bottom flags cleared. BOTTOM is represented
181 by a lattice with the bottom flag set. In that case, values and
182 contains_variable flag should be disregarded. */
186 /* The list of known values and types in this lattice. Note that values are
187 not deallocated if a lattice is set to bottom because there may be value
188 sources referencing them. */
189 struct ipcp_value *values;
190 /* Number of known values and types in this lattice. */
192 /* The lattice contains a variable component (in addition to values). */
193 bool contains_variable;
194 /* The value of the lattice is bottom (i.e. variable and unusable for any
199 /* Lattice with an offset to describe a part of an aggregate. */
201 struct ipcp_agg_lattice : public ipcp_lattice
203 /* Offset that is being described by this lattice. */
204 HOST_WIDE_INT offset;
205 /* Size so that we don't have to re-compute it every time we traverse the
206 list. Must correspond to TYPE_SIZE of all lat values. */
208 /* Next element of the linked list. */
209 struct ipcp_agg_lattice *next;
212 /* Structure containing lattices for a parameter itself and for pieces of
213 aggregates that are passed in the parameter or by a reference in a parameter
214 plus some other useful flags. */
216 struct ipcp_param_lattices
218 /* Lattice describing the value of the parameter itself. */
219 struct ipcp_lattice itself;
220 /* Lattices describing aggregate parts. */
221 struct ipcp_agg_lattice *aggs;
222 /* Number of aggregate lattices */
224 /* True if aggregate data were passed by reference (as opposed to by
227 /* All aggregate lattices contain a variable component (in addition to
229 bool aggs_contain_variable;
230 /* The value of all aggregate lattices is bottom (i.e. variable and unusable
231 for any propagation). */
234 /* There is a virtual call based on this parameter. */
238 /* Allocation pools for values and their sources in ipa-cp. */
240 alloc_pool ipcp_values_pool;
241 alloc_pool ipcp_sources_pool;
242 alloc_pool ipcp_agg_lattice_pool;
244 /* Maximal count found in program. */
246 static gcov_type max_count;
248 /* Original overall size of the program. */
250 static long overall_size, max_new_size;
252 /* Head of the linked list of topologically sorted values. */
254 static struct ipcp_value *values_topo;
256 /* Return the param lattices structure corresponding to the Ith formal
257 parameter of the function described by INFO. */
258 static inline struct ipcp_param_lattices *
259 ipa_get_parm_lattices (struct ipa_node_params *info, int i)
261 gcc_assert (i >= 0 && i < ipa_get_param_count (info));
262 gcc_checking_assert (!info->ipcp_orig_node);
263 gcc_checking_assert (info->lattices);
264 return &(info->lattices[i]);
267 /* Return the lattice corresponding to the scalar value of the Ith formal
268 parameter of the function described by INFO. */
269 static inline struct ipcp_lattice *
270 ipa_get_scalar_lat (struct ipa_node_params *info, int i)
272 struct ipcp_param_lattices *plats = ipa_get_parm_lattices (info, i);
273 return &plats->itself;
276 /* Return whether LAT is a lattice with a single constant and without an
280 ipa_lat_is_single_const (struct ipcp_lattice *lat)
283 || lat->contains_variable
284 || lat->values_count != 1)
290 /* Return true iff the CS is an edge within a strongly connected component as
291 computed by ipa_reduced_postorder. */
294 edge_within_scc (struct cgraph_edge *cs)
296 struct ipa_dfs_info *caller_dfs = (struct ipa_dfs_info *) cs->caller->symbol.aux;
297 struct ipa_dfs_info *callee_dfs;
298 struct cgraph_node *callee = cgraph_function_node (cs->callee, NULL);
300 callee_dfs = (struct ipa_dfs_info *) callee->symbol.aux;
303 && caller_dfs->scc_no == callee_dfs->scc_no);
306 /* Print V which is extracted from a value in a lattice to F. */
309 print_ipcp_constant_value (FILE * f, tree v)
311 if (TREE_CODE (v) == TREE_BINFO)
313 fprintf (f, "BINFO ");
314 print_generic_expr (f, BINFO_TYPE (v), 0);
316 else if (TREE_CODE (v) == ADDR_EXPR
317 && TREE_CODE (TREE_OPERAND (v, 0)) == CONST_DECL)
320 print_generic_expr (f, DECL_INITIAL (TREE_OPERAND (v, 0)), 0);
323 print_generic_expr (f, v, 0);
326 /* Print a lattice LAT to F. */
329 print_lattice (FILE * f, struct ipcp_lattice *lat,
330 bool dump_sources, bool dump_benefits)
332 struct ipcp_value *val;
337 fprintf (f, "BOTTOM\n");
341 if (!lat->values_count && !lat->contains_variable)
343 fprintf (f, "TOP\n");
347 if (lat->contains_variable)
349 fprintf (f, "VARIABLE");
355 for (val = lat->values; val; val = val->next)
357 if (dump_benefits && prev)
359 else if (!dump_benefits && prev)
364 print_ipcp_constant_value (f, val->value);
368 struct ipcp_value_source *s;
370 fprintf (f, " [from:");
371 for (s = val->sources; s; s = s->next)
372 fprintf (f, " %i(%i)", s->cs->caller->uid,s->cs->frequency);
377 fprintf (f, " [loc_time: %i, loc_size: %i, "
378 "prop_time: %i, prop_size: %i]\n",
379 val->local_time_benefit, val->local_size_cost,
380 val->prop_time_benefit, val->prop_size_cost);
386 /* Print all ipcp_lattices of all functions to F. */
389 print_all_lattices (FILE * f, bool dump_sources, bool dump_benefits)
391 struct cgraph_node *node;
394 fprintf (f, "\nLattices:\n");
395 FOR_EACH_FUNCTION_WITH_GIMPLE_BODY (node)
397 struct ipa_node_params *info;
399 info = IPA_NODE_REF (node);
400 fprintf (f, " Node: %s/%i:\n", cgraph_node_name (node), node->uid);
401 count = ipa_get_param_count (info);
402 for (i = 0; i < count; i++)
404 struct ipcp_agg_lattice *aglat;
405 struct ipcp_param_lattices *plats = ipa_get_parm_lattices (info, i);
406 fprintf (f, " param [%d]: ", i);
407 print_lattice (f, &plats->itself, dump_sources, dump_benefits);
409 if (plats->virt_call)
410 fprintf (f, " virt_call flag set\n");
412 if (plats->aggs_bottom)
414 fprintf (f, " AGGS BOTTOM\n");
417 if (plats->aggs_contain_variable)
418 fprintf (f, " AGGS VARIABLE\n");
419 for (aglat = plats->aggs; aglat; aglat = aglat->next)
421 fprintf (f, " %soffset " HOST_WIDE_INT_PRINT_DEC ": ",
422 plats->aggs_by_ref ? "ref " : "", aglat->offset);
423 print_lattice (f, aglat, dump_sources, dump_benefits);
429 /* Determine whether it is at all technically possible to create clones of NODE
430 and store this information in the ipa_node_params structure associated
434 determine_versionability (struct cgraph_node *node)
436 const char *reason = NULL;
438 /* There are a number of generic reasons functions cannot be versioned. We
439 also cannot remove parameters if there are type attributes such as fnspec
441 if (node->alias || node->thunk.thunk_p)
442 reason = "alias or thunk";
443 else if (!node->local.versionable)
444 reason = "not a tree_versionable_function";
445 else if (cgraph_function_body_availability (node) <= AVAIL_OVERWRITABLE)
446 reason = "insufficient body availability";
448 if (reason && dump_file && !node->alias && !node->thunk.thunk_p)
449 fprintf (dump_file, "Function %s/%i is not versionable, reason: %s.\n",
450 cgraph_node_name (node), node->uid, reason);
452 node->local.versionable = (reason == NULL);
455 /* Return true if it is at all technically possible to create clones of a
459 ipcp_versionable_function_p (struct cgraph_node *node)
461 return node->local.versionable;
464 /* Structure holding accumulated information about callers of a node. */
466 struct caller_statistics
469 int n_calls, n_hot_calls, freq_sum;
472 /* Initialize fields of STAT to zeroes. */
475 init_caller_stats (struct caller_statistics *stats)
477 stats->count_sum = 0;
479 stats->n_hot_calls = 0;
483 /* Worker callback of cgraph_for_node_and_aliases accumulating statistics of
484 non-thunk incoming edges to NODE. */
487 gather_caller_stats (struct cgraph_node *node, void *data)
489 struct caller_statistics *stats = (struct caller_statistics *) data;
490 struct cgraph_edge *cs;
492 for (cs = node->callers; cs; cs = cs->next_caller)
493 if (cs->caller->thunk.thunk_p)
494 cgraph_for_node_and_aliases (cs->caller, gather_caller_stats,
498 stats->count_sum += cs->count;
499 stats->freq_sum += cs->frequency;
501 if (cgraph_maybe_hot_edge_p (cs))
502 stats->n_hot_calls ++;
508 /* Return true if this NODE is viable candidate for cloning. */
511 ipcp_cloning_candidate_p (struct cgraph_node *node)
513 struct caller_statistics stats;
515 gcc_checking_assert (cgraph_function_with_gimple_body_p (node));
517 if (!flag_ipa_cp_clone)
520 fprintf (dump_file, "Not considering %s for cloning; "
521 "-fipa-cp-clone disabled.\n",
522 cgraph_node_name (node));
526 if (!optimize_function_for_speed_p (DECL_STRUCT_FUNCTION (node->symbol.decl)))
529 fprintf (dump_file, "Not considering %s for cloning; "
530 "optimizing it for size.\n",
531 cgraph_node_name (node));
535 init_caller_stats (&stats);
536 cgraph_for_node_and_aliases (node, gather_caller_stats, &stats, false);
538 if (inline_summary (node)->self_size < stats.n_calls)
541 fprintf (dump_file, "Considering %s for cloning; code might shrink.\n",
542 cgraph_node_name (node));
546 /* When profile is available and function is hot, propagate into it even if
547 calls seems cold; constant propagation can improve function's speed
551 if (stats.count_sum > node->count * 90 / 100)
554 fprintf (dump_file, "Considering %s for cloning; "
555 "usually called directly.\n",
556 cgraph_node_name (node));
560 if (!stats.n_hot_calls)
563 fprintf (dump_file, "Not considering %s for cloning; no hot calls.\n",
564 cgraph_node_name (node));
568 fprintf (dump_file, "Considering %s for cloning.\n",
569 cgraph_node_name (node));
573 /* Arrays representing a topological ordering of call graph nodes and a stack
574 of noes used during constant propagation. */
578 struct cgraph_node **order;
579 struct cgraph_node **stack;
580 int nnodes, stack_top;
583 /* Allocate the arrays in TOPO and topologically sort the nodes into order. */
586 build_toporder_info (struct topo_info *topo)
588 topo->order = XCNEWVEC (struct cgraph_node *, cgraph_n_nodes);
589 topo->stack = XCNEWVEC (struct cgraph_node *, cgraph_n_nodes);
591 topo->nnodes = ipa_reduced_postorder (topo->order, true, true, NULL);
594 /* Free information about strongly connected components and the arrays in
598 free_toporder_info (struct topo_info *topo)
600 ipa_free_postorder_info ();
605 /* Add NODE to the stack in TOPO, unless it is already there. */
608 push_node_to_stack (struct topo_info *topo, struct cgraph_node *node)
610 struct ipa_node_params *info = IPA_NODE_REF (node);
611 if (info->node_enqueued)
613 info->node_enqueued = 1;
614 topo->stack[topo->stack_top++] = node;
617 /* Pop a node from the stack in TOPO and return it or return NULL if the stack
620 static struct cgraph_node *
621 pop_node_from_stack (struct topo_info *topo)
625 struct cgraph_node *node;
627 node = topo->stack[topo->stack_top];
628 IPA_NODE_REF (node)->node_enqueued = 0;
635 /* Set lattice LAT to bottom and return true if it previously was not set as
639 set_lattice_to_bottom (struct ipcp_lattice *lat)
641 bool ret = !lat->bottom;
646 /* Mark lattice as containing an unknown value and return true if it previously
647 was not marked as such. */
650 set_lattice_contains_variable (struct ipcp_lattice *lat)
652 bool ret = !lat->contains_variable;
653 lat->contains_variable = true;
657 /* Set all aggegate lattices in PLATS to bottom and return true if they were
658 not previously set as such. */
661 set_agg_lats_to_bottom (struct ipcp_param_lattices *plats)
663 bool ret = !plats->aggs_bottom;
664 plats->aggs_bottom = true;
668 /* Mark all aggegate lattices in PLATS as containing an unknown value and
669 return true if they were not previously marked as such. */
672 set_agg_lats_contain_variable (struct ipcp_param_lattices *plats)
674 bool ret = !plats->aggs_contain_variable;
675 plats->aggs_contain_variable = true;
679 /* Mark bot aggregate and scalar lattices as containing an unknown variable,
680 return true is any of them has not been marked as such so far. */
683 set_all_contains_variable (struct ipcp_param_lattices *plats)
685 bool ret = !plats->itself.contains_variable || !plats->aggs_contain_variable;
686 plats->itself.contains_variable = true;
687 plats->aggs_contain_variable = true;
691 /* Initialize ipcp_lattices. */
694 initialize_node_lattices (struct cgraph_node *node)
696 struct ipa_node_params *info = IPA_NODE_REF (node);
697 struct cgraph_edge *ie;
698 bool disable = false, variable = false;
701 gcc_checking_assert (cgraph_function_with_gimple_body_p (node));
702 if (!node->local.local)
704 /* When cloning is allowed, we can assume that externally visible
705 functions are not called. We will compensate this by cloning
707 if (ipcp_versionable_function_p (node)
708 && ipcp_cloning_candidate_p (node))
714 if (disable || variable)
716 for (i = 0; i < ipa_get_param_count (info) ; i++)
718 struct ipcp_param_lattices *plats = ipa_get_parm_lattices (info, i);
721 set_lattice_to_bottom (&plats->itself);
722 set_agg_lats_to_bottom (plats);
725 set_all_contains_variable (plats);
727 if (dump_file && (dump_flags & TDF_DETAILS)
728 && !node->alias && !node->thunk.thunk_p)
729 fprintf (dump_file, "Marking all lattices of %s/%i as %s\n",
730 cgraph_node_name (node), node->uid,
731 disable ? "BOTTOM" : "VARIABLE");
734 for (i = 0; i < ipa_get_param_count (info) ; i++)
736 struct ipcp_param_lattices *plats = ipa_get_parm_lattices (info, i);
737 tree t = TREE_TYPE (ipa_get_param(info, i));
739 if (POINTER_TYPE_P (t) && TYPE_RESTRICT (t)
740 && TREE_CODE (TREE_TYPE (t)) == ARRAY_TYPE)
742 set_lattice_to_bottom (&plats->itself);
743 if (dump_file && (dump_flags & TDF_DETAILS)
744 && !node->alias && !node->thunk.thunk_p)
745 fprintf (dump_file, "Going to ignore param %i of of %s/%i.\n",
746 i, cgraph_node_name (node), node->uid);
750 for (ie = node->indirect_calls; ie; ie = ie->next_callee)
751 if (ie->indirect_info->polymorphic)
753 gcc_checking_assert (ie->indirect_info->param_index >= 0);
754 ipa_get_parm_lattices (info,
755 ie->indirect_info->param_index)->virt_call = 1;
759 /* Return the result of a (possibly arithmetic) pass through jump function
760 JFUNC on the constant value INPUT. Return NULL_TREE if that cannot be
761 determined or itself is considered an interprocedural invariant. */
764 ipa_get_jf_pass_through_result (struct ipa_jump_func *jfunc, tree input)
768 if (ipa_get_jf_pass_through_operation (jfunc) == NOP_EXPR)
770 else if (TREE_CODE (input) == TREE_BINFO)
773 gcc_checking_assert (is_gimple_ip_invariant (input));
774 if (TREE_CODE_CLASS (ipa_get_jf_pass_through_operation (jfunc))
776 restype = boolean_type_node;
778 restype = TREE_TYPE (input);
779 res = fold_binary (ipa_get_jf_pass_through_operation (jfunc), restype,
780 input, ipa_get_jf_pass_through_operand (jfunc));
782 if (res && !is_gimple_ip_invariant (res))
788 /* Return the result of an ancestor jump function JFUNC on the constant value
789 INPUT. Return NULL_TREE if that cannot be determined. */
792 ipa_get_jf_ancestor_result (struct ipa_jump_func *jfunc, tree input)
794 if (TREE_CODE (input) == TREE_BINFO)
795 return get_binfo_at_offset (input,
796 ipa_get_jf_ancestor_offset (jfunc),
797 ipa_get_jf_ancestor_type (jfunc));
798 else if (TREE_CODE (input) == ADDR_EXPR)
800 tree t = TREE_OPERAND (input, 0);
801 t = build_ref_for_offset (EXPR_LOCATION (t), t,
802 ipa_get_jf_ancestor_offset (jfunc),
803 ipa_get_jf_ancestor_type (jfunc), NULL, false);
804 return build_fold_addr_expr (t);
810 /* Extract the acual BINFO being described by JFUNC which must be a known type
814 ipa_value_from_known_type_jfunc (struct ipa_jump_func *jfunc)
816 tree base_binfo = TYPE_BINFO (ipa_get_jf_known_type_base_type (jfunc));
819 return get_binfo_at_offset (base_binfo,
820 ipa_get_jf_known_type_offset (jfunc),
821 ipa_get_jf_known_type_component_type (jfunc));
824 /* Determine whether JFUNC evaluates to a known value (that is either a
825 constant or a binfo) and if so, return it. Otherwise return NULL. INFO
826 describes the caller node so that pass-through jump functions can be
830 ipa_value_from_jfunc (struct ipa_node_params *info, struct ipa_jump_func *jfunc)
832 if (jfunc->type == IPA_JF_CONST)
833 return ipa_get_jf_constant (jfunc);
834 else if (jfunc->type == IPA_JF_KNOWN_TYPE)
835 return ipa_value_from_known_type_jfunc (jfunc);
836 else if (jfunc->type == IPA_JF_PASS_THROUGH
837 || jfunc->type == IPA_JF_ANCESTOR)
842 if (jfunc->type == IPA_JF_PASS_THROUGH)
843 idx = ipa_get_jf_pass_through_formal_id (jfunc);
845 idx = ipa_get_jf_ancestor_formal_id (jfunc);
847 if (info->ipcp_orig_node)
848 input = info->known_vals[idx];
851 struct ipcp_lattice *lat;
855 gcc_checking_assert (!flag_ipa_cp);
858 lat = ipa_get_scalar_lat (info, idx);
859 if (!ipa_lat_is_single_const (lat))
861 input = lat->values->value;
867 if (jfunc->type == IPA_JF_PASS_THROUGH)
868 return ipa_get_jf_pass_through_result (jfunc, input);
870 return ipa_get_jf_ancestor_result (jfunc, input);
877 /* If checking is enabled, verify that no lattice is in the TOP state, i.e. not
878 bottom, not containing a variable component and without any known value at
882 ipcp_verify_propagated_values (void)
884 struct cgraph_node *node;
886 FOR_EACH_FUNCTION_WITH_GIMPLE_BODY (node)
888 struct ipa_node_params *info = IPA_NODE_REF (node);
889 int i, count = ipa_get_param_count (info);
891 for (i = 0; i < count; i++)
893 struct ipcp_lattice *lat = ipa_get_scalar_lat (info, i);
896 && !lat->contains_variable
897 && lat->values_count == 0)
901 fprintf (dump_file, "\nIPA lattices after constant "
903 print_all_lattices (dump_file, true, false);
912 /* Return true iff X and Y should be considered equal values by IPA-CP. */
915 values_equal_for_ipcp_p (tree x, tree y)
917 gcc_checking_assert (x != NULL_TREE && y != NULL_TREE);
922 if (TREE_CODE (x) == TREE_BINFO || TREE_CODE (y) == TREE_BINFO)
925 if (TREE_CODE (x) == ADDR_EXPR
926 && TREE_CODE (y) == ADDR_EXPR
927 && TREE_CODE (TREE_OPERAND (x, 0)) == CONST_DECL
928 && TREE_CODE (TREE_OPERAND (y, 0)) == CONST_DECL)
929 return operand_equal_p (DECL_INITIAL (TREE_OPERAND (x, 0)),
930 DECL_INITIAL (TREE_OPERAND (y, 0)), 0);
932 return operand_equal_p (x, y, 0);
935 /* Add a new value source to VAL, marking that a value comes from edge CS and
936 (if the underlying jump function is a pass-through or an ancestor one) from
937 a caller value SRC_VAL of a caller parameter described by SRC_INDEX. OFFSET
938 is negative if the source was the scalar value of the parameter itself or
939 the offset within an aggregate. */
942 add_value_source (struct ipcp_value *val, struct cgraph_edge *cs,
943 struct ipcp_value *src_val, int src_idx, HOST_WIDE_INT offset)
945 struct ipcp_value_source *src;
947 src = (struct ipcp_value_source *) pool_alloc (ipcp_sources_pool);
948 src->offset = offset;
951 src->index = src_idx;
953 src->next = val->sources;
957 /* Try to add NEWVAL to LAT, potentially creating a new struct ipcp_value for
958 it. CS, SRC_VAL SRC_INDEX and OFFSET are meant for add_value_source and
959 have the same meaning. */
962 add_value_to_lattice (struct ipcp_lattice *lat, tree newval,
963 struct cgraph_edge *cs, struct ipcp_value *src_val,
964 int src_idx, HOST_WIDE_INT offset)
966 struct ipcp_value *val;
971 for (val = lat->values; val; val = val->next)
972 if (values_equal_for_ipcp_p (val->value, newval))
974 if (edge_within_scc (cs))
976 struct ipcp_value_source *s;
977 for (s = val->sources; s ; s = s->next)
984 add_value_source (val, cs, src_val, src_idx, offset);
988 if (lat->values_count == PARAM_VALUE (PARAM_IPA_CP_VALUE_LIST_SIZE))
990 /* We can only free sources, not the values themselves, because sources
991 of other values in this this SCC might point to them. */
992 for (val = lat->values; val; val = val->next)
996 struct ipcp_value_source *src = val->sources;
997 val->sources = src->next;
998 pool_free (ipcp_sources_pool, src);
1003 return set_lattice_to_bottom (lat);
1006 lat->values_count++;
1007 val = (struct ipcp_value *) pool_alloc (ipcp_values_pool);
1008 memset (val, 0, sizeof (*val));
1010 add_value_source (val, cs, src_val, src_idx, offset);
1011 val->value = newval;
1012 val->next = lat->values;
1017 /* Like above but passes a special value of offset to distinguish that the
1018 origin is the scalar value of the parameter rather than a part of an
1022 add_scalar_value_to_lattice (struct ipcp_lattice *lat, tree newval,
1023 struct cgraph_edge *cs,
1024 struct ipcp_value *src_val, int src_idx)
1026 return add_value_to_lattice (lat, newval, cs, src_val, src_idx, -1);
1029 /* Propagate values through a pass-through jump function JFUNC associated with
1030 edge CS, taking values from SRC_LAT and putting them into DEST_LAT. SRC_IDX
1031 is the index of the source parameter. */
1034 propagate_vals_accross_pass_through (struct cgraph_edge *cs,
1035 struct ipa_jump_func *jfunc,
1036 struct ipcp_lattice *src_lat,
1037 struct ipcp_lattice *dest_lat,
1040 struct ipcp_value *src_val;
1043 if (ipa_get_jf_pass_through_operation (jfunc) == NOP_EXPR)
1044 for (src_val = src_lat->values; src_val; src_val = src_val->next)
1045 ret |= add_scalar_value_to_lattice (dest_lat, src_val->value, cs,
1047 /* Do not create new values when propagating within an SCC because if there
1048 are arithmetic functions with circular dependencies, there is infinite
1049 number of them and we would just make lattices bottom. */
1050 else if (edge_within_scc (cs))
1051 ret = set_lattice_contains_variable (dest_lat);
1053 for (src_val = src_lat->values; src_val; src_val = src_val->next)
1055 tree cstval = src_val->value;
1057 if (TREE_CODE (cstval) == TREE_BINFO)
1059 ret |= set_lattice_contains_variable (dest_lat);
1062 cstval = ipa_get_jf_pass_through_result (jfunc, cstval);
1065 ret |= add_scalar_value_to_lattice (dest_lat, cstval, cs, src_val,
1068 ret |= set_lattice_contains_variable (dest_lat);
1074 /* Propagate values through an ancestor jump function JFUNC associated with
1075 edge CS, taking values from SRC_LAT and putting them into DEST_LAT. SRC_IDX
1076 is the index of the source parameter. */
1079 propagate_vals_accross_ancestor (struct cgraph_edge *cs,
1080 struct ipa_jump_func *jfunc,
1081 struct ipcp_lattice *src_lat,
1082 struct ipcp_lattice *dest_lat,
1085 struct ipcp_value *src_val;
1088 if (edge_within_scc (cs))
1089 return set_lattice_contains_variable (dest_lat);
1091 for (src_val = src_lat->values; src_val; src_val = src_val->next)
1093 tree t = ipa_get_jf_ancestor_result (jfunc, src_val->value);
1096 ret |= add_scalar_value_to_lattice (dest_lat, t, cs, src_val, src_idx);
1098 ret |= set_lattice_contains_variable (dest_lat);
1104 /* Propagate scalar values across jump function JFUNC that is associated with
1105 edge CS and put the values into DEST_LAT. */
1108 propagate_scalar_accross_jump_function (struct cgraph_edge *cs,
1109 struct ipa_jump_func *jfunc,
1110 struct ipcp_lattice *dest_lat)
1112 if (dest_lat->bottom)
1115 if (jfunc->type == IPA_JF_CONST
1116 || jfunc->type == IPA_JF_KNOWN_TYPE)
1120 if (jfunc->type == IPA_JF_KNOWN_TYPE)
1122 val = ipa_value_from_known_type_jfunc (jfunc);
1124 return set_lattice_contains_variable (dest_lat);
1127 val = ipa_get_jf_constant (jfunc);
1128 return add_scalar_value_to_lattice (dest_lat, val, cs, NULL, 0);
1130 else if (jfunc->type == IPA_JF_PASS_THROUGH
1131 || jfunc->type == IPA_JF_ANCESTOR)
1133 struct ipa_node_params *caller_info = IPA_NODE_REF (cs->caller);
1134 struct ipcp_lattice *src_lat;
1138 if (jfunc->type == IPA_JF_PASS_THROUGH)
1139 src_idx = ipa_get_jf_pass_through_formal_id (jfunc);
1141 src_idx = ipa_get_jf_ancestor_formal_id (jfunc);
1143 src_lat = ipa_get_scalar_lat (caller_info, src_idx);
1144 if (src_lat->bottom)
1145 return set_lattice_contains_variable (dest_lat);
1147 /* If we would need to clone the caller and cannot, do not propagate. */
1148 if (!ipcp_versionable_function_p (cs->caller)
1149 && (src_lat->contains_variable
1150 || (src_lat->values_count > 1)))
1151 return set_lattice_contains_variable (dest_lat);
1153 if (jfunc->type == IPA_JF_PASS_THROUGH)
1154 ret = propagate_vals_accross_pass_through (cs, jfunc, src_lat,
1157 ret = propagate_vals_accross_ancestor (cs, jfunc, src_lat, dest_lat,
1160 if (src_lat->contains_variable)
1161 ret |= set_lattice_contains_variable (dest_lat);
1166 /* TODO: We currently do not handle member method pointers in IPA-CP (we only
1167 use it for indirect inlining), we should propagate them too. */
1168 return set_lattice_contains_variable (dest_lat);
1171 /* If DEST_PLATS already has aggregate items, check that aggs_by_ref matches
1172 NEW_AGGS_BY_REF and if not, mark all aggs as bottoms and return true (in all
1173 other cases, return false). If there are no aggregate items, set
1174 aggs_by_ref to NEW_AGGS_BY_REF. */
1177 set_check_aggs_by_ref (struct ipcp_param_lattices *dest_plats,
1178 bool new_aggs_by_ref)
1180 if (dest_plats->aggs)
1182 if (dest_plats->aggs_by_ref != new_aggs_by_ref)
1184 set_agg_lats_to_bottom (dest_plats);
1189 dest_plats->aggs_by_ref = new_aggs_by_ref;
1193 /* Walk aggregate lattices in DEST_PLATS from ***AGLAT on, until ***aglat is an
1194 already existing lattice for the given OFFSET and SIZE, marking all skipped
1195 lattices as containing variable and checking for overlaps. If there is no
1196 already existing lattice for the OFFSET and VAL_SIZE, create one, initialize
1197 it with offset, size and contains_variable to PRE_EXISTING, and return true,
1198 unless there are too many already. If there are two many, return false. If
1199 there are overlaps turn whole DEST_PLATS to bottom and return false. If any
1200 skipped lattices were newly marked as containing variable, set *CHANGE to
1204 merge_agg_lats_step (struct ipcp_param_lattices *dest_plats,
1205 HOST_WIDE_INT offset, HOST_WIDE_INT val_size,
1206 struct ipcp_agg_lattice ***aglat,
1207 bool pre_existing, bool *change)
1209 gcc_checking_assert (offset >= 0);
1211 while (**aglat && (**aglat)->offset < offset)
1213 if ((**aglat)->offset + (**aglat)->size > offset)
1215 set_agg_lats_to_bottom (dest_plats);
1218 *change |= set_lattice_contains_variable (**aglat);
1219 *aglat = &(**aglat)->next;
1222 if (**aglat && (**aglat)->offset == offset)
1224 if ((**aglat)->size != val_size
1226 && (**aglat)->next->offset < offset + val_size))
1228 set_agg_lats_to_bottom (dest_plats);
1231 gcc_checking_assert (!(**aglat)->next
1232 || (**aglat)->next->offset >= offset + val_size);
1237 struct ipcp_agg_lattice *new_al;
1239 if (**aglat && (**aglat)->offset < offset + val_size)
1241 set_agg_lats_to_bottom (dest_plats);
1244 if (dest_plats->aggs_count == PARAM_VALUE (PARAM_IPA_MAX_AGG_ITEMS))
1246 dest_plats->aggs_count++;
1247 new_al = (struct ipcp_agg_lattice *) pool_alloc (ipcp_agg_lattice_pool);
1248 memset (new_al, 0, sizeof (*new_al));
1250 new_al->offset = offset;
1251 new_al->size = val_size;
1252 new_al->contains_variable = pre_existing;
1254 new_al->next = **aglat;
1260 /* Set all AGLAT and all other aggregate lattices reachable by next pointers as
1261 containing an unknown value. */
1264 set_chain_of_aglats_contains_variable (struct ipcp_agg_lattice *aglat)
1269 ret |= set_lattice_contains_variable (aglat);
1270 aglat = aglat->next;
1275 /* Merge existing aggregate lattices in SRC_PLATS to DEST_PLATS, subtracting
1276 DELTA_OFFSET. CS is the call graph edge and SRC_IDX the index of the source
1277 parameter used for lattice value sources. Return true if DEST_PLATS changed
1281 merge_aggregate_lattices (struct cgraph_edge *cs,
1282 struct ipcp_param_lattices *dest_plats,
1283 struct ipcp_param_lattices *src_plats,
1284 int src_idx, HOST_WIDE_INT offset_delta)
1286 bool pre_existing = dest_plats->aggs != NULL;
1287 struct ipcp_agg_lattice **dst_aglat;
1290 if (set_check_aggs_by_ref (dest_plats, src_plats->aggs_by_ref))
1292 if (src_plats->aggs_bottom)
1293 return set_agg_lats_contain_variable (dest_plats);
1294 if (src_plats->aggs_contain_variable)
1295 ret |= set_agg_lats_contain_variable (dest_plats);
1296 dst_aglat = &dest_plats->aggs;
1298 for (struct ipcp_agg_lattice *src_aglat = src_plats->aggs;
1300 src_aglat = src_aglat->next)
1302 HOST_WIDE_INT new_offset = src_aglat->offset - offset_delta;
1306 if (merge_agg_lats_step (dest_plats, new_offset, src_aglat->size,
1307 &dst_aglat, pre_existing, &ret))
1309 struct ipcp_agg_lattice *new_al = *dst_aglat;
1311 dst_aglat = &(*dst_aglat)->next;
1312 if (src_aglat->bottom)
1314 ret |= set_lattice_contains_variable (new_al);
1317 if (src_aglat->contains_variable)
1318 ret |= set_lattice_contains_variable (new_al);
1319 for (struct ipcp_value *val = src_aglat->values;
1322 ret |= add_value_to_lattice (new_al, val->value, cs, val, src_idx,
1325 else if (dest_plats->aggs_bottom)
1328 ret |= set_chain_of_aglats_contains_variable (*dst_aglat);
1332 /* Determine whether there is anything to propagate FROM SRC_PLATS through a
1333 pass-through JFUNC and if so, whether it has conform and conforms to the
1334 rules about propagating values passed by reference. */
1337 agg_pass_through_permissible_p (struct ipcp_param_lattices *src_plats,
1338 struct ipa_jump_func *jfunc)
1340 return src_plats->aggs
1341 && (!src_plats->aggs_by_ref
1342 || ipa_get_jf_pass_through_agg_preserved (jfunc));
1345 /* Propagate scalar values across jump function JFUNC that is associated with
1346 edge CS and put the values into DEST_LAT. */
1349 propagate_aggs_accross_jump_function (struct cgraph_edge *cs,
1350 struct ipa_jump_func *jfunc,
1351 struct ipcp_param_lattices *dest_plats)
1355 if (dest_plats->aggs_bottom)
1358 if (jfunc->type == IPA_JF_PASS_THROUGH
1359 && ipa_get_jf_pass_through_operation (jfunc) == NOP_EXPR)
1361 struct ipa_node_params *caller_info = IPA_NODE_REF (cs->caller);
1362 int src_idx = ipa_get_jf_pass_through_formal_id (jfunc);
1363 struct ipcp_param_lattices *src_plats;
1365 src_plats = ipa_get_parm_lattices (caller_info, src_idx);
1366 if (agg_pass_through_permissible_p (src_plats, jfunc))
1368 /* Currently we do not produce clobber aggregate jump
1369 functions, replace with merging when we do. */
1370 gcc_assert (!jfunc->agg.items);
1371 ret |= merge_aggregate_lattices (cs, dest_plats, src_plats,
1375 ret |= set_agg_lats_contain_variable (dest_plats);
1377 else if (jfunc->type == IPA_JF_ANCESTOR
1378 && ipa_get_jf_ancestor_agg_preserved (jfunc))
1380 struct ipa_node_params *caller_info = IPA_NODE_REF (cs->caller);
1381 int src_idx = ipa_get_jf_ancestor_formal_id (jfunc);
1382 struct ipcp_param_lattices *src_plats;
1384 src_plats = ipa_get_parm_lattices (caller_info, src_idx);
1385 if (src_plats->aggs && src_plats->aggs_by_ref)
1387 /* Currently we do not produce clobber aggregate jump
1388 functions, replace with merging when we do. */
1389 gcc_assert (!jfunc->agg.items);
1390 ret |= merge_aggregate_lattices (cs, dest_plats, src_plats, src_idx,
1391 ipa_get_jf_ancestor_offset (jfunc));
1393 else if (!src_plats->aggs_by_ref)
1394 ret |= set_agg_lats_to_bottom (dest_plats);
1396 ret |= set_agg_lats_contain_variable (dest_plats);
1398 else if (jfunc->agg.items)
1400 bool pre_existing = dest_plats->aggs != NULL;
1401 struct ipcp_agg_lattice **aglat = &dest_plats->aggs;
1402 struct ipa_agg_jf_item *item;
1405 if (set_check_aggs_by_ref (dest_plats, jfunc->agg.by_ref))
1408 FOR_EACH_VEC_ELT (*jfunc->agg.items, i, item)
1410 HOST_WIDE_INT val_size;
1412 if (item->offset < 0)
1414 gcc_checking_assert (is_gimple_ip_invariant (item->value));
1415 val_size = tree_low_cst (TYPE_SIZE (TREE_TYPE (item->value)), 1);
1417 if (merge_agg_lats_step (dest_plats, item->offset, val_size,
1418 &aglat, pre_existing, &ret))
1420 ret |= add_value_to_lattice (*aglat, item->value, cs, NULL, 0, 0);
1421 aglat = &(*aglat)->next;
1423 else if (dest_plats->aggs_bottom)
1427 ret |= set_chain_of_aglats_contains_variable (*aglat);
1430 ret |= set_agg_lats_contain_variable (dest_plats);
1435 /* Propagate constants from the caller to the callee of CS. INFO describes the
1439 propagate_constants_accross_call (struct cgraph_edge *cs)
1441 struct ipa_node_params *callee_info;
1442 enum availability availability;
1443 struct cgraph_node *callee, *alias_or_thunk;
1444 struct ipa_edge_args *args;
1446 int i, args_count, parms_count;
1448 callee = cgraph_function_node (cs->callee, &availability);
1449 if (!callee->analyzed)
1451 gcc_checking_assert (cgraph_function_with_gimple_body_p (callee));
1452 callee_info = IPA_NODE_REF (callee);
1454 args = IPA_EDGE_REF (cs);
1455 args_count = ipa_get_cs_argument_count (args);
1456 parms_count = ipa_get_param_count (callee_info);
1458 /* If this call goes through a thunk we must not propagate to the first (0th)
1459 parameter. However, we might need to uncover a thunk from below a series
1460 of aliases first. */
1461 alias_or_thunk = cs->callee;
1462 while (alias_or_thunk->alias)
1463 alias_or_thunk = cgraph_alias_aliased_node (alias_or_thunk);
1464 if (alias_or_thunk->thunk.thunk_p)
1466 ret |= set_all_contains_variable (ipa_get_parm_lattices (callee_info,
1473 for (; (i < args_count) && (i < parms_count); i++)
1475 struct ipa_jump_func *jump_func = ipa_get_ith_jump_func (args, i);
1476 struct ipcp_param_lattices *dest_plats;
1478 dest_plats = ipa_get_parm_lattices (callee_info, i);
1479 if (availability == AVAIL_OVERWRITABLE)
1480 ret |= set_all_contains_variable (dest_plats);
1483 ret |= propagate_scalar_accross_jump_function (cs, jump_func,
1484 &dest_plats->itself);
1485 ret |= propagate_aggs_accross_jump_function (cs, jump_func,
1489 for (; i < parms_count; i++)
1490 ret |= set_all_contains_variable (ipa_get_parm_lattices (callee_info, i));
1495 /* If an indirect edge IE can be turned into a direct one based on KNOWN_VALS
1496 (which can contain both constants and binfos) or KNOWN_BINFOS (which can be
1497 NULL) return the destination. */
1500 ipa_get_indirect_edge_target (struct cgraph_edge *ie,
1501 vec<tree> known_vals,
1502 vec<tree> known_binfos,
1503 vec<ipa_agg_jump_function_p> known_aggs)
1505 int param_index = ie->indirect_info->param_index;
1506 HOST_WIDE_INT token, anc_offset;
1510 if (param_index == -1
1511 || known_vals.length () <= (unsigned int) param_index)
1514 if (!ie->indirect_info->polymorphic)
1518 if (ie->indirect_info->agg_contents)
1520 if (known_aggs.length ()
1521 > (unsigned int) param_index)
1523 struct ipa_agg_jump_function *agg;
1524 agg = known_aggs[param_index];
1525 t = ipa_find_agg_cst_for_param (agg, ie->indirect_info->offset,
1526 ie->indirect_info->by_ref);
1532 t = known_vals[param_index];
1535 TREE_CODE (t) == ADDR_EXPR
1536 && TREE_CODE (TREE_OPERAND (t, 0)) == FUNCTION_DECL)
1537 return TREE_OPERAND (t, 0);
1542 gcc_assert (!ie->indirect_info->agg_contents);
1543 token = ie->indirect_info->otr_token;
1544 anc_offset = ie->indirect_info->offset;
1545 otr_type = ie->indirect_info->otr_type;
1547 t = known_vals[param_index];
1548 if (!t && known_binfos.length () > (unsigned int) param_index)
1549 t = known_binfos[param_index];
1553 if (TREE_CODE (t) != TREE_BINFO)
1556 binfo = gimple_extract_devirt_binfo_from_cst (t);
1559 binfo = get_binfo_at_offset (binfo, anc_offset, otr_type);
1562 return gimple_get_virt_method_for_binfo (token, binfo);
1568 binfo = get_binfo_at_offset (t, anc_offset, otr_type);
1571 return gimple_get_virt_method_for_binfo (token, binfo);
1575 /* Calculate devirtualization time bonus for NODE, assuming we know KNOWN_CSTS
1576 and KNOWN_BINFOS. */
1579 devirtualization_time_bonus (struct cgraph_node *node,
1580 vec<tree> known_csts,
1581 vec<tree> known_binfos)
1583 struct cgraph_edge *ie;
1586 for (ie = node->indirect_calls; ie; ie = ie->next_callee)
1588 struct cgraph_node *callee;
1589 struct inline_summary *isummary;
1592 target = ipa_get_indirect_edge_target (ie, known_csts, known_binfos,
1597 /* Only bare minimum benefit for clearly un-inlineable targets. */
1599 callee = cgraph_get_node (target);
1600 if (!callee || !callee->analyzed)
1602 isummary = inline_summary (callee);
1603 if (!isummary->inlinable)
1606 /* FIXME: The values below need re-considering and perhaps also
1607 integrating into the cost metrics, at lest in some very basic way. */
1608 if (isummary->size <= MAX_INLINE_INSNS_AUTO / 4)
1610 else if (isummary->size <= MAX_INLINE_INSNS_AUTO / 2)
1612 else if (isummary->size <= MAX_INLINE_INSNS_AUTO
1613 || DECL_DECLARED_INLINE_P (callee->symbol.decl))
1620 /* Return time bonus incurred because of HINTS. */
1623 hint_time_bonus (inline_hints hints)
1625 if (hints & (INLINE_HINT_loop_iterations | INLINE_HINT_loop_stride))
1626 return PARAM_VALUE (PARAM_IPA_CP_LOOP_HINT_BONUS);
1630 /* Return true if cloning NODE is a good idea, given the estimated TIME_BENEFIT
1631 and SIZE_COST and with the sum of frequencies of incoming edges to the
1632 potential new clone in FREQUENCIES. */
1635 good_cloning_opportunity_p (struct cgraph_node *node, int time_benefit,
1636 int freq_sum, gcov_type count_sum, int size_cost)
1638 if (time_benefit == 0
1639 || !flag_ipa_cp_clone
1640 || !optimize_function_for_speed_p (DECL_STRUCT_FUNCTION (node->symbol.decl)))
1643 gcc_assert (size_cost > 0);
1647 int factor = (count_sum * 1000) / max_count;
1648 HOST_WIDEST_INT evaluation = (((HOST_WIDEST_INT) time_benefit * factor)
1651 if (dump_file && (dump_flags & TDF_DETAILS))
1652 fprintf (dump_file, " good_cloning_opportunity_p (time: %i, "
1653 "size: %i, count_sum: " HOST_WIDE_INT_PRINT_DEC
1654 ") -> evaluation: " HOST_WIDEST_INT_PRINT_DEC
1655 ", threshold: %i\n",
1656 time_benefit, size_cost, (HOST_WIDE_INT) count_sum,
1657 evaluation, PARAM_VALUE (PARAM_IPA_CP_EVAL_THRESHOLD));
1659 return evaluation >= PARAM_VALUE (PARAM_IPA_CP_EVAL_THRESHOLD);
1663 HOST_WIDEST_INT evaluation = (((HOST_WIDEST_INT) time_benefit * freq_sum)
1666 if (dump_file && (dump_flags & TDF_DETAILS))
1667 fprintf (dump_file, " good_cloning_opportunity_p (time: %i, "
1668 "size: %i, freq_sum: %i) -> evaluation: "
1669 HOST_WIDEST_INT_PRINT_DEC ", threshold: %i\n",
1670 time_benefit, size_cost, freq_sum, evaluation,
1671 PARAM_VALUE (PARAM_IPA_CP_EVAL_THRESHOLD));
1673 return evaluation >= PARAM_VALUE (PARAM_IPA_CP_EVAL_THRESHOLD);
1677 /* Return all context independent values from aggregate lattices in PLATS in a
1678 vector. Return NULL if there are none. */
1680 static vec<ipa_agg_jf_item_t, va_gc> *
1681 context_independent_aggregate_values (struct ipcp_param_lattices *plats)
1683 vec<ipa_agg_jf_item_t, va_gc> *res = NULL;
1685 if (plats->aggs_bottom
1686 || plats->aggs_contain_variable
1687 || plats->aggs_count == 0)
1690 for (struct ipcp_agg_lattice *aglat = plats->aggs;
1692 aglat = aglat->next)
1693 if (ipa_lat_is_single_const (aglat))
1695 struct ipa_agg_jf_item item;
1696 item.offset = aglat->offset;
1697 item.value = aglat->values->value;
1698 vec_safe_push (res, item);
1703 /* Allocate KNOWN_CSTS, KNOWN_BINFOS and, if non-NULL, KNOWN_AGGS and populate
1704 them with values of parameters that are known independent of the context.
1705 INFO describes the function. If REMOVABLE_PARAMS_COST is non-NULL, the
1706 movement cost of all removable parameters will be stored in it. */
1709 gather_context_independent_values (struct ipa_node_params *info,
1710 vec<tree> *known_csts,
1711 vec<tree> *known_binfos,
1712 vec<ipa_agg_jump_function_t> *known_aggs,
1713 int *removable_params_cost)
1715 int i, count = ipa_get_param_count (info);
1718 known_csts->create (0);
1719 known_binfos->create (0);
1720 known_csts->safe_grow_cleared (count);
1721 known_binfos->safe_grow_cleared (count);
1724 known_aggs->create (0);
1725 known_aggs->safe_grow_cleared (count);
1728 if (removable_params_cost)
1729 *removable_params_cost = 0;
1731 for (i = 0; i < count ; i++)
1733 struct ipcp_param_lattices *plats = ipa_get_parm_lattices (info, i);
1734 struct ipcp_lattice *lat = &plats->itself;
1736 if (ipa_lat_is_single_const (lat))
1738 struct ipcp_value *val = lat->values;
1739 if (TREE_CODE (val->value) != TREE_BINFO)
1741 (*known_csts)[i] = val->value;
1742 if (removable_params_cost)
1743 *removable_params_cost
1744 += estimate_move_cost (TREE_TYPE (val->value));
1747 else if (plats->virt_call)
1749 (*known_binfos)[i] = val->value;
1752 else if (removable_params_cost
1753 && !ipa_is_param_used (info, i))
1754 *removable_params_cost
1755 += estimate_move_cost (TREE_TYPE (ipa_get_param (info, i)));
1757 else if (removable_params_cost
1758 && !ipa_is_param_used (info, i))
1759 *removable_params_cost
1760 += estimate_move_cost (TREE_TYPE (ipa_get_param (info, i)));
1764 vec<ipa_agg_jf_item_t, va_gc> *agg_items;
1765 struct ipa_agg_jump_function *ajf;
1767 agg_items = context_independent_aggregate_values (plats);
1768 ajf = &(*known_aggs)[i];
1769 ajf->items = agg_items;
1770 ajf->by_ref = plats->aggs_by_ref;
1771 ret |= agg_items != NULL;
1778 /* The current interface in ipa-inline-analysis requires a pointer vector.
1781 FIXME: That interface should be re-worked, this is slightly silly. Still,
1782 I'd like to discuss how to change it first and this demonstrates the
1785 static vec<ipa_agg_jump_function_p>
1786 agg_jmp_p_vec_for_t_vec (vec<ipa_agg_jump_function_t> known_aggs)
1788 vec<ipa_agg_jump_function_p> ret;
1789 struct ipa_agg_jump_function *ajf;
1792 ret.create (known_aggs.length ());
1793 FOR_EACH_VEC_ELT (known_aggs, i, ajf)
1794 ret.quick_push (ajf);
1798 /* Iterate over known values of parameters of NODE and estimate the local
1799 effects in terms of time and size they have. */
1802 estimate_local_effects (struct cgraph_node *node)
1804 struct ipa_node_params *info = IPA_NODE_REF (node);
1805 int i, count = ipa_get_param_count (info);
1806 vec<tree> known_csts, known_binfos;
1807 vec<ipa_agg_jump_function_t> known_aggs;
1808 vec<ipa_agg_jump_function_p> known_aggs_ptrs;
1810 int base_time = inline_summary (node)->time;
1811 int removable_params_cost;
1813 if (!count || !ipcp_versionable_function_p (node))
1816 if (dump_file && (dump_flags & TDF_DETAILS))
1817 fprintf (dump_file, "\nEstimating effects for %s/%i, base_time: %i.\n",
1818 cgraph_node_name (node), node->uid, base_time);
1820 always_const = gather_context_independent_values (info, &known_csts,
1821 &known_binfos, &known_aggs,
1822 &removable_params_cost);
1823 known_aggs_ptrs = agg_jmp_p_vec_for_t_vec (known_aggs);
1826 struct caller_statistics stats;
1830 init_caller_stats (&stats);
1831 cgraph_for_node_and_aliases (node, gather_caller_stats, &stats, false);
1832 estimate_ipcp_clone_size_and_time (node, known_csts, known_binfos,
1833 known_aggs_ptrs, &size, &time, &hints);
1834 time -= devirtualization_time_bonus (node, known_csts, known_binfos);
1835 time -= hint_time_bonus (hints);
1836 time -= removable_params_cost;
1837 size -= stats.n_calls * removable_params_cost;
1840 fprintf (dump_file, " - context independent values, size: %i, "
1841 "time_benefit: %i\n", size, base_time - time);
1844 || cgraph_will_be_removed_from_program_if_no_direct_calls (node))
1846 info->do_clone_for_all_contexts = true;
1850 fprintf (dump_file, " Decided to specialize for all "
1851 "known contexts, code not going to grow.\n");
1853 else if (good_cloning_opportunity_p (node, base_time - time,
1854 stats.freq_sum, stats.count_sum,
1857 if (size + overall_size <= max_new_size)
1859 info->do_clone_for_all_contexts = true;
1861 overall_size += size;
1864 fprintf (dump_file, " Decided to specialize for all "
1865 "known contexts, growth deemed beneficial.\n");
1867 else if (dump_file && (dump_flags & TDF_DETAILS))
1868 fprintf (dump_file, " Not cloning for all contexts because "
1869 "max_new_size would be reached with %li.\n",
1870 size + overall_size);
1874 for (i = 0; i < count ; i++)
1876 struct ipcp_param_lattices *plats = ipa_get_parm_lattices (info, i);
1877 struct ipcp_lattice *lat = &plats->itself;
1878 struct ipcp_value *val;
1887 for (val = lat->values; val; val = val->next)
1889 int time, size, time_benefit;
1892 if (TREE_CODE (val->value) != TREE_BINFO)
1894 known_csts[i] = val->value;
1895 known_binfos[i] = NULL_TREE;
1896 emc = estimate_move_cost (TREE_TYPE (val->value));
1898 else if (plats->virt_call)
1900 known_csts[i] = NULL_TREE;
1901 known_binfos[i] = val->value;
1907 estimate_ipcp_clone_size_and_time (node, known_csts, known_binfos,
1908 known_aggs_ptrs, &size, &time,
1910 time_benefit = base_time - time
1911 + devirtualization_time_bonus (node, known_csts, known_binfos)
1912 + hint_time_bonus (hints)
1913 + removable_params_cost + emc;
1915 gcc_checking_assert (size >=0);
1916 /* The inliner-heuristics based estimates may think that in certain
1917 contexts some functions do not have any size at all but we want
1918 all specializations to have at least a tiny cost, not least not to
1923 if (dump_file && (dump_flags & TDF_DETAILS))
1925 fprintf (dump_file, " - estimates for value ");
1926 print_ipcp_constant_value (dump_file, val->value);
1927 fprintf (dump_file, " for parameter ");
1928 print_generic_expr (dump_file, ipa_get_param (info, i), 0);
1929 fprintf (dump_file, ": time_benefit: %i, size: %i\n",
1930 time_benefit, size);
1933 val->local_time_benefit = time_benefit;
1934 val->local_size_cost = size;
1936 known_binfos[i] = NULL_TREE;
1937 known_csts[i] = NULL_TREE;
1940 for (i = 0; i < count ; i++)
1942 struct ipcp_param_lattices *plats = ipa_get_parm_lattices (info, i);
1943 struct ipa_agg_jump_function *ajf;
1944 struct ipcp_agg_lattice *aglat;
1946 if (plats->aggs_bottom || !plats->aggs)
1949 ajf = &known_aggs[i];
1950 for (aglat = plats->aggs; aglat; aglat = aglat->next)
1952 struct ipcp_value *val;
1953 if (aglat->bottom || !aglat->values
1954 /* If the following is true, the one value is in known_aggs. */
1955 || (!plats->aggs_contain_variable
1956 && ipa_lat_is_single_const (aglat)))
1959 for (val = aglat->values; val; val = val->next)
1961 int time, size, time_benefit;
1962 struct ipa_agg_jf_item item;
1965 item.offset = aglat->offset;
1966 item.value = val->value;
1967 vec_safe_push (ajf->items, item);
1969 estimate_ipcp_clone_size_and_time (node, known_csts, known_binfos,
1970 known_aggs_ptrs, &size, &time,
1972 time_benefit = base_time - time
1973 + devirtualization_time_bonus (node, known_csts, known_binfos)
1974 + hint_time_bonus (hints);
1975 gcc_checking_assert (size >=0);
1979 if (dump_file && (dump_flags & TDF_DETAILS))
1981 fprintf (dump_file, " - estimates for value ");
1982 print_ipcp_constant_value (dump_file, val->value);
1983 fprintf (dump_file, " for parameter ");
1984 print_generic_expr (dump_file, ipa_get_param (info, i), 0);
1985 fprintf (dump_file, "[%soffset: " HOST_WIDE_INT_PRINT_DEC
1986 "]: time_benefit: %i, size: %i\n",
1987 plats->aggs_by_ref ? "ref " : "",
1988 aglat->offset, time_benefit, size);
1991 val->local_time_benefit = time_benefit;
1992 val->local_size_cost = size;
1998 for (i = 0; i < count ; i++)
1999 vec_free (known_aggs[i].items);
2001 known_csts.release ();
2002 known_binfos.release ();
2003 known_aggs.release ();
2004 known_aggs_ptrs.release ();
2008 /* Add value CUR_VAL and all yet-unsorted values it is dependent on to the
2009 topological sort of values. */
2012 add_val_to_toposort (struct ipcp_value *cur_val)
2014 static int dfs_counter = 0;
2015 static struct ipcp_value *stack;
2016 struct ipcp_value_source *src;
2022 cur_val->dfs = dfs_counter;
2023 cur_val->low_link = dfs_counter;
2025 cur_val->topo_next = stack;
2027 cur_val->on_stack = true;
2029 for (src = cur_val->sources; src; src = src->next)
2032 if (src->val->dfs == 0)
2034 add_val_to_toposort (src->val);
2035 if (src->val->low_link < cur_val->low_link)
2036 cur_val->low_link = src->val->low_link;
2038 else if (src->val->on_stack
2039 && src->val->dfs < cur_val->low_link)
2040 cur_val->low_link = src->val->dfs;
2043 if (cur_val->dfs == cur_val->low_link)
2045 struct ipcp_value *v, *scc_list = NULL;
2050 stack = v->topo_next;
2051 v->on_stack = false;
2053 v->scc_next = scc_list;
2056 while (v != cur_val);
2058 cur_val->topo_next = values_topo;
2059 values_topo = cur_val;
2063 /* Add all values in lattices associated with NODE to the topological sort if
2064 they are not there yet. */
2067 add_all_node_vals_to_toposort (struct cgraph_node *node)
2069 struct ipa_node_params *info = IPA_NODE_REF (node);
2070 int i, count = ipa_get_param_count (info);
2072 for (i = 0; i < count ; i++)
2074 struct ipcp_param_lattices *plats = ipa_get_parm_lattices (info, i);
2075 struct ipcp_lattice *lat = &plats->itself;
2076 struct ipcp_agg_lattice *aglat;
2077 struct ipcp_value *val;
2080 for (val = lat->values; val; val = val->next)
2081 add_val_to_toposort (val);
2083 if (!plats->aggs_bottom)
2084 for (aglat = plats->aggs; aglat; aglat = aglat->next)
2086 for (val = aglat->values; val; val = val->next)
2087 add_val_to_toposort (val);
2091 /* One pass of constants propagation along the call graph edges, from callers
2092 to callees (requires topological ordering in TOPO), iterate over strongly
2093 connected components. */
2096 propagate_constants_topo (struct topo_info *topo)
2100 for (i = topo->nnodes - 1; i >= 0; i--)
2102 struct cgraph_node *v, *node = topo->order[i];
2103 struct ipa_dfs_info *node_dfs_info;
2105 if (!cgraph_function_with_gimple_body_p (node))
2108 node_dfs_info = (struct ipa_dfs_info *) node->symbol.aux;
2109 /* First, iteratively propagate within the strongly connected component
2110 until all lattices stabilize. */
2111 v = node_dfs_info->next_cycle;
2114 push_node_to_stack (topo, v);
2115 v = ((struct ipa_dfs_info *) v->symbol.aux)->next_cycle;
2121 struct cgraph_edge *cs;
2123 for (cs = v->callees; cs; cs = cs->next_callee)
2124 if (edge_within_scc (cs)
2125 && propagate_constants_accross_call (cs))
2126 push_node_to_stack (topo, cs->callee);
2127 v = pop_node_from_stack (topo);
2130 /* Afterwards, propagate along edges leading out of the SCC, calculates
2131 the local effects of the discovered constants and all valid values to
2132 their topological sort. */
2136 struct cgraph_edge *cs;
2138 estimate_local_effects (v);
2139 add_all_node_vals_to_toposort (v);
2140 for (cs = v->callees; cs; cs = cs->next_callee)
2141 if (!edge_within_scc (cs))
2142 propagate_constants_accross_call (cs);
2144 v = ((struct ipa_dfs_info *) v->symbol.aux)->next_cycle;
2150 /* Return the sum of A and B if none of them is bigger than INT_MAX/2, return
2151 the bigger one if otherwise. */
2154 safe_add (int a, int b)
2156 if (a > INT_MAX/2 || b > INT_MAX/2)
2157 return a > b ? a : b;
2163 /* Propagate the estimated effects of individual values along the topological
2164 from the dependent values to those they depend on. */
2167 propagate_effects (void)
2169 struct ipcp_value *base;
2171 for (base = values_topo; base; base = base->topo_next)
2173 struct ipcp_value_source *src;
2174 struct ipcp_value *val;
2175 int time = 0, size = 0;
2177 for (val = base; val; val = val->scc_next)
2179 time = safe_add (time,
2180 val->local_time_benefit + val->prop_time_benefit);
2181 size = safe_add (size, val->local_size_cost + val->prop_size_cost);
2184 for (val = base; val; val = val->scc_next)
2185 for (src = val->sources; src; src = src->next)
2187 && cgraph_maybe_hot_edge_p (src->cs))
2189 src->val->prop_time_benefit = safe_add (time,
2190 src->val->prop_time_benefit);
2191 src->val->prop_size_cost = safe_add (size,
2192 src->val->prop_size_cost);
2198 /* Propagate constants, binfos and their effects from the summaries
2199 interprocedurally. */
2202 ipcp_propagate_stage (struct topo_info *topo)
2204 struct cgraph_node *node;
2207 fprintf (dump_file, "\n Propagating constants:\n\n");
2210 ipa_update_after_lto_read ();
2213 FOR_EACH_DEFINED_FUNCTION (node)
2215 struct ipa_node_params *info = IPA_NODE_REF (node);
2217 determine_versionability (node);
2218 if (cgraph_function_with_gimple_body_p (node))
2220 info->lattices = XCNEWVEC (struct ipcp_param_lattices,
2221 ipa_get_param_count (info));
2222 initialize_node_lattices (node);
2224 if (node->count > max_count)
2225 max_count = node->count;
2226 overall_size += inline_summary (node)->self_size;
2229 max_new_size = overall_size;
2230 if (max_new_size < PARAM_VALUE (PARAM_LARGE_UNIT_INSNS))
2231 max_new_size = PARAM_VALUE (PARAM_LARGE_UNIT_INSNS);
2232 max_new_size += max_new_size * PARAM_VALUE (PARAM_IPCP_UNIT_GROWTH) / 100 + 1;
2235 fprintf (dump_file, "\noverall_size: %li, max_new_size: %li\n",
2236 overall_size, max_new_size);
2238 propagate_constants_topo (topo);
2239 #ifdef ENABLE_CHECKING
2240 ipcp_verify_propagated_values ();
2242 propagate_effects ();
2246 fprintf (dump_file, "\nIPA lattices after all propagation:\n");
2247 print_all_lattices (dump_file, (dump_flags & TDF_DETAILS), true);
2251 /* Discover newly direct outgoing edges from NODE which is a new clone with
2252 known KNOWN_VALS and make them direct. */
2255 ipcp_discover_new_direct_edges (struct cgraph_node *node,
2256 vec<tree> known_vals)
2258 struct cgraph_edge *ie, *next_ie;
2261 for (ie = node->indirect_calls; ie; ie = next_ie)
2265 next_ie = ie->next_callee;
2266 target = ipa_get_indirect_edge_target (ie, known_vals, vNULL, vNULL);
2269 ipa_make_edge_direct_to_target (ie, target);
2273 /* Turning calls to direct calls will improve overall summary. */
2275 inline_update_overall_summary (node);
2278 /* Vector of pointers which for linked lists of clones of an original crgaph
2281 static vec<cgraph_edge_p> next_edge_clone;
2284 grow_next_edge_clone_vector (void)
2286 if (next_edge_clone.length ()
2287 <= (unsigned) cgraph_edge_max_uid)
2288 next_edge_clone.safe_grow_cleared (cgraph_edge_max_uid + 1);
2291 /* Edge duplication hook to grow the appropriate linked list in
2295 ipcp_edge_duplication_hook (struct cgraph_edge *src, struct cgraph_edge *dst,
2296 __attribute__((unused)) void *data)
2298 grow_next_edge_clone_vector ();
2299 next_edge_clone[dst->uid] = next_edge_clone[src->uid];
2300 next_edge_clone[src->uid] = dst;
2303 /* See if NODE is a clone with a known aggregate value at a given OFFSET of a
2304 parameter with the given INDEX. */
2307 get_clone_agg_value (struct cgraph_node *node, HOST_WIDEST_INT offset,
2310 struct ipa_agg_replacement_value *aggval;
2312 aggval = ipa_get_agg_replacements_for_node (node);
2315 if (aggval->offset == offset
2316 && aggval->index == index)
2317 return aggval->value;
2318 aggval = aggval->next;
2323 /* Return true if edge CS does bring about the value described by SRC. */
2326 cgraph_edge_brings_value_p (struct cgraph_edge *cs,
2327 struct ipcp_value_source *src)
2329 struct ipa_node_params *caller_info = IPA_NODE_REF (cs->caller);
2330 struct ipa_node_params *dst_info = IPA_NODE_REF (cs->callee);
2332 if ((dst_info->ipcp_orig_node && !dst_info->is_all_contexts_clone)
2333 || caller_info->node_dead)
2338 if (caller_info->ipcp_orig_node)
2341 if (src->offset == -1)
2342 t = caller_info->known_vals[src->index];
2344 t = get_clone_agg_value (cs->caller, src->offset, src->index);
2345 return (t != NULL_TREE
2346 && values_equal_for_ipcp_p (src->val->value, t));
2350 struct ipcp_agg_lattice *aglat;
2351 struct ipcp_param_lattices *plats = ipa_get_parm_lattices (caller_info,
2353 if (src->offset == -1)
2354 return (ipa_lat_is_single_const (&plats->itself)
2355 && values_equal_for_ipcp_p (src->val->value,
2356 plats->itself.values->value));
2359 if (plats->aggs_bottom || plats->aggs_contain_variable)
2361 for (aglat = plats->aggs; aglat; aglat = aglat->next)
2362 if (aglat->offset == src->offset)
2363 return (ipa_lat_is_single_const (aglat)
2364 && values_equal_for_ipcp_p (src->val->value,
2365 aglat->values->value));
2371 /* Get the next clone in the linked list of clones of an edge. */
2373 static inline struct cgraph_edge *
2374 get_next_cgraph_edge_clone (struct cgraph_edge *cs)
2376 return next_edge_clone[cs->uid];
2379 /* Given VAL, iterate over all its sources and if they still hold, add their
2380 edge frequency and their number into *FREQUENCY and *CALLER_COUNT
2384 get_info_about_necessary_edges (struct ipcp_value *val, int *freq_sum,
2385 gcov_type *count_sum, int *caller_count)
2387 struct ipcp_value_source *src;
2388 int freq = 0, count = 0;
2392 for (src = val->sources; src; src = src->next)
2394 struct cgraph_edge *cs = src->cs;
2397 if (cgraph_edge_brings_value_p (cs, src))
2400 freq += cs->frequency;
2402 hot |= cgraph_maybe_hot_edge_p (cs);
2404 cs = get_next_cgraph_edge_clone (cs);
2410 *caller_count = count;
2414 /* Return a vector of incoming edges that do bring value VAL. It is assumed
2415 their number is known and equal to CALLER_COUNT. */
2417 static vec<cgraph_edge_p>
2418 gather_edges_for_value (struct ipcp_value *val, int caller_count)
2420 struct ipcp_value_source *src;
2421 vec<cgraph_edge_p> ret;
2423 ret.create (caller_count);
2424 for (src = val->sources; src; src = src->next)
2426 struct cgraph_edge *cs = src->cs;
2429 if (cgraph_edge_brings_value_p (cs, src))
2430 ret.quick_push (cs);
2431 cs = get_next_cgraph_edge_clone (cs);
2438 /* Construct a replacement map for a know VALUE for a formal parameter PARAM.
2439 Return it or NULL if for some reason it cannot be created. */
2441 static struct ipa_replace_map *
2442 get_replacement_map (tree value, tree parm)
2444 tree req_type = TREE_TYPE (parm);
2445 struct ipa_replace_map *replace_map;
2447 if (!useless_type_conversion_p (req_type, TREE_TYPE (value)))
2449 if (fold_convertible_p (req_type, value))
2450 value = fold_build1 (NOP_EXPR, req_type, value);
2451 else if (TYPE_SIZE (req_type) == TYPE_SIZE (TREE_TYPE (value)))
2452 value = fold_build1 (VIEW_CONVERT_EXPR, req_type, value);
2457 fprintf (dump_file, " const ");
2458 print_generic_expr (dump_file, value, 0);
2459 fprintf (dump_file, " can't be converted to param ");
2460 print_generic_expr (dump_file, parm, 0);
2461 fprintf (dump_file, "\n");
2467 replace_map = ggc_alloc_ipa_replace_map ();
2470 fprintf (dump_file, " replacing param ");
2471 print_generic_expr (dump_file, parm, 0);
2472 fprintf (dump_file, " with const ");
2473 print_generic_expr (dump_file, value, 0);
2474 fprintf (dump_file, "\n");
2476 replace_map->old_tree = parm;
2477 replace_map->new_tree = value;
2478 replace_map->replace_p = true;
2479 replace_map->ref_p = false;
2484 /* Dump new profiling counts */
2487 dump_profile_updates (struct cgraph_node *orig_node,
2488 struct cgraph_node *new_node)
2490 struct cgraph_edge *cs;
2492 fprintf (dump_file, " setting count of the specialized node to "
2493 HOST_WIDE_INT_PRINT_DEC "\n", (HOST_WIDE_INT) new_node->count);
2494 for (cs = new_node->callees; cs ; cs = cs->next_callee)
2495 fprintf (dump_file, " edge to %s has count "
2496 HOST_WIDE_INT_PRINT_DEC "\n",
2497 cgraph_node_name (cs->callee), (HOST_WIDE_INT) cs->count);
2499 fprintf (dump_file, " setting count of the original node to "
2500 HOST_WIDE_INT_PRINT_DEC "\n", (HOST_WIDE_INT) orig_node->count);
2501 for (cs = orig_node->callees; cs ; cs = cs->next_callee)
2502 fprintf (dump_file, " edge to %s is left with "
2503 HOST_WIDE_INT_PRINT_DEC "\n",
2504 cgraph_node_name (cs->callee), (HOST_WIDE_INT) cs->count);
2507 /* After a specialized NEW_NODE version of ORIG_NODE has been created, update
2508 their profile information to reflect this. */
2511 update_profiling_info (struct cgraph_node *orig_node,
2512 struct cgraph_node *new_node)
2514 struct cgraph_edge *cs;
2515 struct caller_statistics stats;
2516 gcov_type new_sum, orig_sum;
2517 gcov_type remainder, orig_node_count = orig_node->count;
2519 if (orig_node_count == 0)
2522 init_caller_stats (&stats);
2523 cgraph_for_node_and_aliases (orig_node, gather_caller_stats, &stats, false);
2524 orig_sum = stats.count_sum;
2525 init_caller_stats (&stats);
2526 cgraph_for_node_and_aliases (new_node, gather_caller_stats, &stats, false);
2527 new_sum = stats.count_sum;
2529 if (orig_node_count < orig_sum + new_sum)
2532 fprintf (dump_file, " Problem: node %s/%i has too low count "
2533 HOST_WIDE_INT_PRINT_DEC " while the sum of incoming "
2534 "counts is " HOST_WIDE_INT_PRINT_DEC "\n",
2535 cgraph_node_name (orig_node), orig_node->uid,
2536 (HOST_WIDE_INT) orig_node_count,
2537 (HOST_WIDE_INT) (orig_sum + new_sum));
2539 orig_node_count = (orig_sum + new_sum) * 12 / 10;
2541 fprintf (dump_file, " proceeding by pretending it was "
2542 HOST_WIDE_INT_PRINT_DEC "\n",
2543 (HOST_WIDE_INT) orig_node_count);
2546 new_node->count = new_sum;
2547 remainder = orig_node_count - new_sum;
2548 orig_node->count = remainder;
2550 for (cs = new_node->callees; cs ; cs = cs->next_callee)
2552 cs->count = cs->count * (new_sum * REG_BR_PROB_BASE
2553 / orig_node_count) / REG_BR_PROB_BASE;
2557 for (cs = orig_node->callees; cs ; cs = cs->next_callee)
2558 cs->count = cs->count * (remainder * REG_BR_PROB_BASE
2559 / orig_node_count) / REG_BR_PROB_BASE;
2562 dump_profile_updates (orig_node, new_node);
2565 /* Update the respective profile of specialized NEW_NODE and the original
2566 ORIG_NODE after additional edges with cumulative count sum REDIRECTED_SUM
2567 have been redirected to the specialized version. */
2570 update_specialized_profile (struct cgraph_node *new_node,
2571 struct cgraph_node *orig_node,
2572 gcov_type redirected_sum)
2574 struct cgraph_edge *cs;
2575 gcov_type new_node_count, orig_node_count = orig_node->count;
2578 fprintf (dump_file, " the sum of counts of redirected edges is "
2579 HOST_WIDE_INT_PRINT_DEC "\n", (HOST_WIDE_INT) redirected_sum);
2580 if (orig_node_count == 0)
2583 gcc_assert (orig_node_count >= redirected_sum);
2585 new_node_count = new_node->count;
2586 new_node->count += redirected_sum;
2587 orig_node->count -= redirected_sum;
2589 for (cs = new_node->callees; cs ; cs = cs->next_callee)
2591 cs->count += cs->count * redirected_sum / new_node_count;
2595 for (cs = orig_node->callees; cs ; cs = cs->next_callee)
2597 gcov_type dec = cs->count * (redirected_sum * REG_BR_PROB_BASE
2598 / orig_node_count) / REG_BR_PROB_BASE;
2599 if (dec < cs->count)
2606 dump_profile_updates (orig_node, new_node);
2609 /* Create a specialized version of NODE with known constants and types of
2610 parameters in KNOWN_VALS and redirect all edges in CALLERS to it. */
2612 static struct cgraph_node *
2613 create_specialized_node (struct cgraph_node *node,
2614 vec<tree> known_vals,
2615 struct ipa_agg_replacement_value *aggvals,
2616 vec<cgraph_edge_p> callers)
2618 struct ipa_node_params *new_info, *info = IPA_NODE_REF (node);
2619 vec<ipa_replace_map_p, va_gc> *replace_trees = NULL;
2620 struct cgraph_node *new_node;
2621 int i, count = ipa_get_param_count (info);
2622 bitmap args_to_skip;
2624 gcc_assert (!info->ipcp_orig_node);
2626 if (node->local.can_change_signature)
2628 args_to_skip = BITMAP_GGC_ALLOC ();
2629 for (i = 0; i < count; i++)
2631 tree t = known_vals[i];
2633 if ((t && TREE_CODE (t) != TREE_BINFO)
2634 || !ipa_is_param_used (info, i))
2635 bitmap_set_bit (args_to_skip, i);
2640 args_to_skip = NULL;
2641 if (dump_file && (dump_flags & TDF_DETAILS))
2642 fprintf (dump_file, " cannot change function signature\n");
2645 for (i = 0; i < count ; i++)
2647 tree t = known_vals[i];
2648 if (t && TREE_CODE (t) != TREE_BINFO)
2650 struct ipa_replace_map *replace_map;
2652 replace_map = get_replacement_map (t, ipa_get_param (info, i));
2654 vec_safe_push (replace_trees, replace_map);
2658 new_node = cgraph_create_virtual_clone (node, callers, replace_trees,
2659 args_to_skip, "constprop");
2660 ipa_set_node_agg_value_chain (new_node, aggvals);
2661 if (dump_file && (dump_flags & TDF_DETAILS))
2663 fprintf (dump_file, " the new node is %s/%i.\n",
2664 cgraph_node_name (new_node), new_node->uid);
2666 ipa_dump_agg_replacement_values (dump_file, aggvals);
2668 gcc_checking_assert (ipa_node_params_vector.exists ()
2669 && (ipa_node_params_vector.length ()
2670 > (unsigned) cgraph_max_uid));
2671 update_profiling_info (node, new_node);
2672 new_info = IPA_NODE_REF (new_node);
2673 new_info->ipcp_orig_node = node;
2674 new_info->known_vals = known_vals;
2676 ipcp_discover_new_direct_edges (new_node, known_vals);
2682 /* Given a NODE, and a subset of its CALLERS, try to populate blanks slots in
2683 KNOWN_VALS with constants and types that are also known for all of the
2687 find_more_scalar_values_for_callers_subset (struct cgraph_node *node,
2688 vec<tree> known_vals,
2689 vec<cgraph_edge_p> callers)
2691 struct ipa_node_params *info = IPA_NODE_REF (node);
2692 int i, count = ipa_get_param_count (info);
2694 for (i = 0; i < count ; i++)
2696 struct cgraph_edge *cs;
2697 tree newval = NULL_TREE;
2700 if (ipa_get_scalar_lat (info, i)->bottom || known_vals[i])
2703 FOR_EACH_VEC_ELT (callers, j, cs)
2705 struct ipa_jump_func *jump_func;
2708 if (i >= ipa_get_cs_argument_count (IPA_EDGE_REF (cs)))
2713 jump_func = ipa_get_ith_jump_func (IPA_EDGE_REF (cs), i);
2714 t = ipa_value_from_jfunc (IPA_NODE_REF (cs->caller), jump_func);
2717 && !values_equal_for_ipcp_p (t, newval)))
2728 if (dump_file && (dump_flags & TDF_DETAILS))
2730 fprintf (dump_file, " adding an extra known scalar value ");
2731 print_ipcp_constant_value (dump_file, newval);
2732 fprintf (dump_file, " for parameter ");
2733 print_generic_expr (dump_file, ipa_get_param (info, i), 0);
2734 fprintf (dump_file, "\n");
2737 known_vals[i] = newval;
2742 /* Go through PLATS and create a vector of values consisting of values and
2743 offsets (minus OFFSET) of lattices that contain only a single value. */
2745 static vec<ipa_agg_jf_item_t>
2746 copy_plats_to_inter (struct ipcp_param_lattices *plats, HOST_WIDE_INT offset)
2748 vec<ipa_agg_jf_item_t> res = vNULL;
2750 if (!plats->aggs || plats->aggs_contain_variable || plats->aggs_bottom)
2753 for (struct ipcp_agg_lattice *aglat = plats->aggs; aglat; aglat = aglat->next)
2754 if (ipa_lat_is_single_const (aglat))
2756 struct ipa_agg_jf_item ti;
2757 ti.offset = aglat->offset - offset;
2758 ti.value = aglat->values->value;
2764 /* Intersect all values in INTER with single value lattices in PLATS (while
2765 subtracting OFFSET). */
2768 intersect_with_plats (struct ipcp_param_lattices *plats,
2769 vec<ipa_agg_jf_item_t> *inter,
2770 HOST_WIDE_INT offset)
2772 struct ipcp_agg_lattice *aglat;
2773 struct ipa_agg_jf_item *item;
2776 if (!plats->aggs || plats->aggs_contain_variable || plats->aggs_bottom)
2782 aglat = plats->aggs;
2783 FOR_EACH_VEC_ELT (*inter, k, item)
2790 if (aglat->offset - offset > item->offset)
2792 if (aglat->offset - offset == item->offset)
2794 gcc_checking_assert (item->value);
2795 if (values_equal_for_ipcp_p (item->value, aglat->values->value))
2799 aglat = aglat->next;
2802 item->value = NULL_TREE;
2806 /* Copy agggregate replacement values of NODE (which is an IPA-CP clone) to the
2807 vector result while subtracting OFFSET from the individual value offsets. */
2809 static vec<ipa_agg_jf_item_t>
2810 agg_replacements_to_vector (struct cgraph_node *node, int index,
2811 HOST_WIDE_INT offset)
2813 struct ipa_agg_replacement_value *av;
2814 vec<ipa_agg_jf_item_t> res = vNULL;
2816 for (av = ipa_get_agg_replacements_for_node (node); av; av = av->next)
2817 if (av->index == index
2818 && (av->offset - offset) >= 0)
2820 struct ipa_agg_jf_item item;
2821 gcc_checking_assert (av->value);
2822 item.offset = av->offset - offset;
2823 item.value = av->value;
2824 res.safe_push (item);
2830 /* Intersect all values in INTER with those that we have already scheduled to
2831 be replaced in parameter number INDEX of NODE, which is an IPA-CP clone
2832 (while subtracting OFFSET). */
2835 intersect_with_agg_replacements (struct cgraph_node *node, int index,
2836 vec<ipa_agg_jf_item_t> *inter,
2837 HOST_WIDE_INT offset)
2839 struct ipa_agg_replacement_value *srcvals;
2840 struct ipa_agg_jf_item *item;
2843 srcvals = ipa_get_agg_replacements_for_node (node);
2850 FOR_EACH_VEC_ELT (*inter, i, item)
2852 struct ipa_agg_replacement_value *av;
2856 for (av = srcvals; av; av = av->next)
2858 gcc_checking_assert (av->value);
2859 if (av->index == index
2860 && av->offset - offset == item->offset)
2862 if (values_equal_for_ipcp_p (item->value, av->value))
2868 item->value = NULL_TREE;
2872 /* Intersect values in INTER with aggregate values that come along edge CS to
2873 parameter number INDEX and return it. If INTER does not actually exist yet,
2874 copy all incoming values to it. If we determine we ended up with no values
2875 whatsoever, return a released vector. */
2877 static vec<ipa_agg_jf_item_t>
2878 intersect_aggregates_with_edge (struct cgraph_edge *cs, int index,
2879 vec<ipa_agg_jf_item_t> inter)
2881 struct ipa_jump_func *jfunc;
2882 jfunc = ipa_get_ith_jump_func (IPA_EDGE_REF (cs), index);
2883 if (jfunc->type == IPA_JF_PASS_THROUGH
2884 && ipa_get_jf_pass_through_operation (jfunc) == NOP_EXPR)
2886 struct ipa_node_params *caller_info = IPA_NODE_REF (cs->caller);
2887 int src_idx = ipa_get_jf_pass_through_formal_id (jfunc);
2889 if (caller_info->ipcp_orig_node)
2891 struct cgraph_node *orig_node = caller_info->ipcp_orig_node;
2892 struct ipcp_param_lattices *orig_plats;
2893 orig_plats = ipa_get_parm_lattices (IPA_NODE_REF (orig_node),
2895 if (agg_pass_through_permissible_p (orig_plats, jfunc))
2897 if (!inter.exists ())
2898 inter = agg_replacements_to_vector (cs->caller, src_idx, 0);
2900 intersect_with_agg_replacements (cs->caller, src_idx,
2906 struct ipcp_param_lattices *src_plats;
2907 src_plats = ipa_get_parm_lattices (caller_info, src_idx);
2908 if (agg_pass_through_permissible_p (src_plats, jfunc))
2910 /* Currently we do not produce clobber aggregate jump
2911 functions, adjust when we do. */
2912 gcc_checking_assert (!jfunc->agg.items);
2913 if (!inter.exists ())
2914 inter = copy_plats_to_inter (src_plats, 0);
2916 intersect_with_plats (src_plats, &inter, 0);
2920 else if (jfunc->type == IPA_JF_ANCESTOR
2921 && ipa_get_jf_ancestor_agg_preserved (jfunc))
2923 struct ipa_node_params *caller_info = IPA_NODE_REF (cs->caller);
2924 int src_idx = ipa_get_jf_ancestor_formal_id (jfunc);
2925 struct ipcp_param_lattices *src_plats;
2926 HOST_WIDE_INT delta = ipa_get_jf_ancestor_offset (jfunc);
2928 if (caller_info->ipcp_orig_node)
2930 if (!inter.exists ())
2931 inter = agg_replacements_to_vector (cs->caller, src_idx, delta);
2933 intersect_with_agg_replacements (cs->caller, src_idx, &inter,
2938 src_plats = ipa_get_parm_lattices (caller_info, src_idx);;
2939 /* Currently we do not produce clobber aggregate jump
2940 functions, adjust when we do. */
2941 gcc_checking_assert (!src_plats->aggs || !jfunc->agg.items);
2942 if (!inter.exists ())
2943 inter = copy_plats_to_inter (src_plats, delta);
2945 intersect_with_plats (src_plats, &inter, delta);
2948 else if (jfunc->agg.items)
2950 struct ipa_agg_jf_item *item;
2953 if (!inter.exists ())
2954 for (unsigned i = 0; i < jfunc->agg.items->length (); i++)
2955 inter.safe_push ((*jfunc->agg.items)[i]);
2957 FOR_EACH_VEC_ELT (inter, k, item)
2960 bool found = false;;
2965 while ((unsigned) l < jfunc->agg.items->length ())
2967 struct ipa_agg_jf_item *ti;
2968 ti = &(*jfunc->agg.items)[l];
2969 if (ti->offset > item->offset)
2971 if (ti->offset == item->offset)
2973 gcc_checking_assert (ti->value);
2974 if (values_equal_for_ipcp_p (item->value,
2988 return vec<ipa_agg_jf_item_t>();
2993 /* Look at edges in CALLERS and collect all known aggregate values that arrive
2994 from all of them. */
2996 static struct ipa_agg_replacement_value *
2997 find_aggregate_values_for_callers_subset (struct cgraph_node *node,
2998 vec<cgraph_edge_p> callers)
3000 struct ipa_node_params *dest_info = IPA_NODE_REF (node);
3001 struct ipa_agg_replacement_value *res = NULL;
3002 struct cgraph_edge *cs;
3003 int i, j, count = ipa_get_param_count (dest_info);
3005 FOR_EACH_VEC_ELT (callers, j, cs)
3007 int c = ipa_get_cs_argument_count (IPA_EDGE_REF (cs));
3012 for (i = 0; i < count ; i++)
3014 struct cgraph_edge *cs;
3015 vec<ipa_agg_jf_item_t> inter = vNULL;
3016 struct ipa_agg_jf_item *item;
3017 struct ipcp_param_lattices *plats = ipa_get_parm_lattices (dest_info, i);
3020 /* Among other things, the following check should deal with all by_ref
3022 if (plats->aggs_bottom)
3025 FOR_EACH_VEC_ELT (callers, j, cs)
3027 inter = intersect_aggregates_with_edge (cs, i, inter);
3029 if (!inter.exists ())
3033 FOR_EACH_VEC_ELT (inter, j, item)
3035 struct ipa_agg_replacement_value *v;
3040 v = ggc_alloc_ipa_agg_replacement_value ();
3042 v->offset = item->offset;
3043 v->value = item->value;
3044 v->by_ref = plats->aggs_by_ref;
3050 if (inter.exists ())
3056 /* Turn KNOWN_AGGS into a list of aggreate replacement values. */
3058 static struct ipa_agg_replacement_value *
3059 known_aggs_to_agg_replacement_list (vec<ipa_agg_jump_function_t> known_aggs)
3061 struct ipa_agg_replacement_value *res = NULL;
3062 struct ipa_agg_jump_function *aggjf;
3063 struct ipa_agg_jf_item *item;
3066 FOR_EACH_VEC_ELT (known_aggs, i, aggjf)
3067 FOR_EACH_VEC_SAFE_ELT (aggjf->items, j, item)
3069 struct ipa_agg_replacement_value *v;
3070 v = ggc_alloc_ipa_agg_replacement_value ();
3072 v->offset = item->offset;
3073 v->value = item->value;
3074 v->by_ref = aggjf->by_ref;
3081 /* Determine whether CS also brings all scalar values that the NODE is
3085 cgraph_edge_brings_all_scalars_for_node (struct cgraph_edge *cs,
3086 struct cgraph_node *node)
3088 struct ipa_node_params *dest_info = IPA_NODE_REF (node);
3089 int count = ipa_get_param_count (dest_info);
3090 struct ipa_node_params *caller_info;
3091 struct ipa_edge_args *args;
3094 caller_info = IPA_NODE_REF (cs->caller);
3095 args = IPA_EDGE_REF (cs);
3096 for (i = 0; i < count; i++)
3098 struct ipa_jump_func *jump_func;
3101 val = dest_info->known_vals[i];
3105 if (i >= ipa_get_cs_argument_count (args))
3107 jump_func = ipa_get_ith_jump_func (args, i);
3108 t = ipa_value_from_jfunc (caller_info, jump_func);
3109 if (!t || !values_equal_for_ipcp_p (val, t))
3115 /* Determine whether CS also brings all aggregate values that NODE is
3118 cgraph_edge_brings_all_agg_vals_for_node (struct cgraph_edge *cs,
3119 struct cgraph_node *node)
3121 struct ipa_node_params *orig_caller_info = IPA_NODE_REF (cs->caller);
3122 struct ipa_agg_replacement_value *aggval;
3125 aggval = ipa_get_agg_replacements_for_node (node);
3129 count = ipa_get_param_count (IPA_NODE_REF (node));
3130 ec = ipa_get_cs_argument_count (IPA_EDGE_REF (cs));
3132 for (struct ipa_agg_replacement_value *av = aggval; av; av = av->next)
3133 if (aggval->index >= ec)
3136 if (orig_caller_info->ipcp_orig_node)
3137 orig_caller_info = IPA_NODE_REF (orig_caller_info->ipcp_orig_node);
3139 for (i = 0; i < count; i++)
3141 static vec<ipa_agg_jf_item_t> values = vec<ipa_agg_jf_item_t>();
3142 struct ipcp_param_lattices *plats;
3143 bool interesting = false;
3144 for (struct ipa_agg_replacement_value *av = aggval; av; av = av->next)
3145 if (aggval->index == i)
3153 plats = ipa_get_parm_lattices (orig_caller_info, aggval->index);
3154 if (plats->aggs_bottom)
3157 values = intersect_aggregates_with_edge (cs, i, values);
3158 if (!values.exists())
3161 for (struct ipa_agg_replacement_value *av = aggval; av; av = av->next)
3162 if (aggval->index == i)
3164 struct ipa_agg_jf_item *item;
3167 FOR_EACH_VEC_ELT (values, j, item)
3169 && item->offset == av->offset
3170 && values_equal_for_ipcp_p (item->value, av->value))
3182 /* Given an original NODE and a VAL for which we have already created a
3183 specialized clone, look whether there are incoming edges that still lead
3184 into the old node but now also bring the requested value and also conform to
3185 all other criteria such that they can be redirected the the special node.
3186 This function can therefore redirect the final edge in a SCC. */
3189 perhaps_add_new_callers (struct cgraph_node *node, struct ipcp_value *val)
3191 struct ipcp_value_source *src;
3192 gcov_type redirected_sum = 0;
3194 for (src = val->sources; src; src = src->next)
3196 struct cgraph_edge *cs = src->cs;
3199 enum availability availability;
3200 struct cgraph_node *dst = cgraph_function_node (cs->callee,
3202 if ((dst == node || IPA_NODE_REF (dst)->is_all_contexts_clone)
3203 && availability > AVAIL_OVERWRITABLE
3204 && cgraph_edge_brings_value_p (cs, src))
3206 if (cgraph_edge_brings_all_scalars_for_node (cs, val->spec_node)
3207 && cgraph_edge_brings_all_agg_vals_for_node (cs,
3211 fprintf (dump_file, " - adding an extra caller %s/%i"
3213 xstrdup (cgraph_node_name (cs->caller)),
3215 xstrdup (cgraph_node_name (val->spec_node)),
3216 val->spec_node->uid);
3218 cgraph_redirect_edge_callee (cs, val->spec_node);
3219 redirected_sum += cs->count;
3222 cs = get_next_cgraph_edge_clone (cs);
3227 update_specialized_profile (val->spec_node, node, redirected_sum);
3231 /* Copy KNOWN_BINFOS to KNOWN_VALS. */
3234 move_binfos_to_values (vec<tree> known_vals,
3235 vec<tree> known_binfos)
3240 for (i = 0; known_binfos.iterate (i, &t); i++)
3245 /* Return true if there is a replacement equivalent to VALUE, INDEX and OFFSET
3246 among those in the AGGVALS list. */
3249 ipcp_val_in_agg_replacements_p (struct ipa_agg_replacement_value *aggvals,
3250 int index, HOST_WIDE_INT offset, tree value)
3254 if (aggvals->index == index
3255 && aggvals->offset == offset
3256 && values_equal_for_ipcp_p (aggvals->value, value))
3258 aggvals = aggvals->next;
3263 /* Decide wheter to create a special version of NODE for value VAL of parameter
3264 at the given INDEX. If OFFSET is -1, the value is for the parameter itself,
3265 otherwise it is stored at the given OFFSET of the parameter. KNOWN_CSTS,
3266 KNOWN_BINFOS and KNOWN_AGGS describe the other already known values. */
3269 decide_about_value (struct cgraph_node *node, int index, HOST_WIDE_INT offset,
3270 struct ipcp_value *val, vec<tree> known_csts,
3271 vec<tree> known_binfos)
3273 struct ipa_agg_replacement_value *aggvals;
3274 int freq_sum, caller_count;
3275 gcov_type count_sum;
3276 vec<cgraph_edge_p> callers;
3281 perhaps_add_new_callers (node, val);
3284 else if (val->local_size_cost + overall_size > max_new_size)
3286 if (dump_file && (dump_flags & TDF_DETAILS))
3287 fprintf (dump_file, " Ignoring candidate value because "
3288 "max_new_size would be reached with %li.\n",
3289 val->local_size_cost + overall_size);
3292 else if (!get_info_about_necessary_edges (val, &freq_sum, &count_sum,
3296 if (dump_file && (dump_flags & TDF_DETAILS))
3298 fprintf (dump_file, " - considering value ");
3299 print_ipcp_constant_value (dump_file, val->value);
3300 fprintf (dump_file, " for parameter ");
3301 print_generic_expr (dump_file, ipa_get_param (IPA_NODE_REF (node),
3304 fprintf (dump_file, ", offset: " HOST_WIDE_INT_PRINT_DEC, offset);
3305 fprintf (dump_file, " (caller_count: %i)\n", caller_count);
3308 if (!good_cloning_opportunity_p (node, val->local_time_benefit,
3309 freq_sum, count_sum,
3310 val->local_size_cost)
3311 && !good_cloning_opportunity_p (node,
3312 val->local_time_benefit
3313 + val->prop_time_benefit,
3314 freq_sum, count_sum,
3315 val->local_size_cost
3316 + val->prop_size_cost))
3320 fprintf (dump_file, " Creating a specialized node of %s/%i.\n",
3321 cgraph_node_name (node), node->uid);
3323 callers = gather_edges_for_value (val, caller_count);
3324 kv = known_csts.copy ();
3325 move_binfos_to_values (kv, known_binfos);
3327 kv[index] = val->value;
3328 find_more_scalar_values_for_callers_subset (node, kv, callers);
3329 aggvals = find_aggregate_values_for_callers_subset (node, callers);
3330 gcc_checking_assert (offset == -1
3331 || ipcp_val_in_agg_replacements_p (aggvals, index,
3332 offset, val->value));
3333 val->spec_node = create_specialized_node (node, kv, aggvals, callers);
3334 overall_size += val->local_size_cost;
3336 /* TODO: If for some lattice there is only one other known value
3337 left, make a special node for it too. */
3342 /* Decide whether and what specialized clones of NODE should be created. */
3345 decide_whether_version_node (struct cgraph_node *node)
3347 struct ipa_node_params *info = IPA_NODE_REF (node);
3348 int i, count = ipa_get_param_count (info);
3349 vec<tree> known_csts, known_binfos;
3350 vec<ipa_agg_jump_function_t> known_aggs = vNULL;
3356 if (dump_file && (dump_flags & TDF_DETAILS))
3357 fprintf (dump_file, "\nEvaluating opportunities for %s/%i.\n",
3358 cgraph_node_name (node), node->uid);
3360 gather_context_independent_values (info, &known_csts, &known_binfos,
3361 info->do_clone_for_all_contexts ? &known_aggs
3364 for (i = 0; i < count ;i++)
3366 struct ipcp_param_lattices *plats = ipa_get_parm_lattices (info, i);
3367 struct ipcp_lattice *lat = &plats->itself;
3368 struct ipcp_value *val;
3372 && !known_binfos[i])
3373 for (val = lat->values; val; val = val->next)
3374 ret |= decide_about_value (node, i, -1, val, known_csts,
3377 if (!plats->aggs_bottom)
3379 struct ipcp_agg_lattice *aglat;
3380 struct ipcp_value *val;
3381 for (aglat = plats->aggs; aglat; aglat = aglat->next)
3382 if (!aglat->bottom && aglat->values
3383 /* If the following is false, the one value is in
3385 && (plats->aggs_contain_variable
3386 || !ipa_lat_is_single_const (aglat)))
3387 for (val = aglat->values; val; val = val->next)
3388 ret |= decide_about_value (node, i, aglat->offset, val,
3389 known_csts, known_binfos);
3391 info = IPA_NODE_REF (node);
3394 if (info->do_clone_for_all_contexts)
3396 struct cgraph_node *clone;
3397 vec<cgraph_edge_p> callers;
3400 fprintf (dump_file, " - Creating a specialized node of %s/%i "
3401 "for all known contexts.\n", cgraph_node_name (node),
3404 callers = collect_callers_of_node (node);
3405 move_binfos_to_values (known_csts, known_binfos);
3406 clone = create_specialized_node (node, known_csts,
3407 known_aggs_to_agg_replacement_list (known_aggs),
3409 info = IPA_NODE_REF (node);
3410 info->do_clone_for_all_contexts = false;
3411 IPA_NODE_REF (clone)->is_all_contexts_clone = true;
3412 for (i = 0; i < count ; i++)
3413 vec_free (known_aggs[i].items);
3414 known_aggs.release ();
3418 known_csts.release ();
3420 known_binfos.release ();
3424 /* Transitively mark all callees of NODE within the same SCC as not dead. */
3427 spread_undeadness (struct cgraph_node *node)
3429 struct cgraph_edge *cs;
3431 for (cs = node->callees; cs; cs = cs->next_callee)
3432 if (edge_within_scc (cs))
3434 struct cgraph_node *callee;
3435 struct ipa_node_params *info;
3437 callee = cgraph_function_node (cs->callee, NULL);
3438 info = IPA_NODE_REF (callee);
3440 if (info->node_dead)
3442 info->node_dead = 0;
3443 spread_undeadness (callee);
3448 /* Return true if NODE has a caller from outside of its SCC that is not
3449 dead. Worker callback for cgraph_for_node_and_aliases. */
3452 has_undead_caller_from_outside_scc_p (struct cgraph_node *node,
3453 void *data ATTRIBUTE_UNUSED)
3455 struct cgraph_edge *cs;
3457 for (cs = node->callers; cs; cs = cs->next_caller)
3458 if (cs->caller->thunk.thunk_p
3459 && cgraph_for_node_and_aliases (cs->caller,
3460 has_undead_caller_from_outside_scc_p,
3463 else if (!edge_within_scc (cs)
3464 && !IPA_NODE_REF (cs->caller)->node_dead)
3470 /* Identify nodes within the same SCC as NODE which are no longer needed
3471 because of new clones and will be removed as unreachable. */
3474 identify_dead_nodes (struct cgraph_node *node)
3476 struct cgraph_node *v;
3477 for (v = node; v ; v = ((struct ipa_dfs_info *) v->symbol.aux)->next_cycle)
3478 if (cgraph_will_be_removed_from_program_if_no_direct_calls (v)
3479 && !cgraph_for_node_and_aliases (v,
3480 has_undead_caller_from_outside_scc_p,
3482 IPA_NODE_REF (v)->node_dead = 1;
3484 for (v = node; v ; v = ((struct ipa_dfs_info *) v->symbol.aux)->next_cycle)
3485 if (!IPA_NODE_REF (v)->node_dead)
3486 spread_undeadness (v);
3488 if (dump_file && (dump_flags & TDF_DETAILS))
3490 for (v = node; v ; v = ((struct ipa_dfs_info *) v->symbol.aux)->next_cycle)
3491 if (IPA_NODE_REF (v)->node_dead)
3492 fprintf (dump_file, " Marking node as dead: %s/%i.\n",
3493 cgraph_node_name (v), v->uid);
3497 /* The decision stage. Iterate over the topological order of call graph nodes
3498 TOPO and make specialized clones if deemed beneficial. */
3501 ipcp_decision_stage (struct topo_info *topo)
3506 fprintf (dump_file, "\nIPA decision stage:\n\n");
3508 for (i = topo->nnodes - 1; i >= 0; i--)
3510 struct cgraph_node *node = topo->order[i];
3511 bool change = false, iterate = true;
3515 struct cgraph_node *v;
3517 for (v = node; v ; v = ((struct ipa_dfs_info *) v->symbol.aux)->next_cycle)
3518 if (cgraph_function_with_gimple_body_p (v)
3519 && ipcp_versionable_function_p (v))
3520 iterate |= decide_whether_version_node (v);
3525 identify_dead_nodes (node);
3529 /* The IPCP driver. */
3534 struct cgraph_2edge_hook_list *edge_duplication_hook_holder;
3535 struct topo_info topo;
3537 ipa_check_create_node_params ();
3538 ipa_check_create_edge_args ();
3539 grow_next_edge_clone_vector ();
3540 edge_duplication_hook_holder =
3541 cgraph_add_edge_duplication_hook (&ipcp_edge_duplication_hook, NULL);
3542 ipcp_values_pool = create_alloc_pool ("IPA-CP values",
3543 sizeof (struct ipcp_value), 32);
3544 ipcp_sources_pool = create_alloc_pool ("IPA-CP value sources",
3545 sizeof (struct ipcp_value_source), 64);
3546 ipcp_agg_lattice_pool = create_alloc_pool ("IPA_CP aggregate lattices",
3547 sizeof (struct ipcp_agg_lattice),
3551 fprintf (dump_file, "\nIPA structures before propagation:\n");
3552 if (dump_flags & TDF_DETAILS)
3553 ipa_print_all_params (dump_file);
3554 ipa_print_all_jump_functions (dump_file);
3557 /* Topological sort. */
3558 build_toporder_info (&topo);
3559 /* Do the interprocedural propagation. */
3560 ipcp_propagate_stage (&topo);
3561 /* Decide what constant propagation and cloning should be performed. */
3562 ipcp_decision_stage (&topo);
3564 /* Free all IPCP structures. */
3565 free_toporder_info (&topo);
3566 next_edge_clone.release ();
3567 cgraph_remove_edge_duplication_hook (edge_duplication_hook_holder);
3568 ipa_free_all_structures_after_ipa_cp ();
3570 fprintf (dump_file, "\nIPA constant propagation end\n");
3574 /* Initialization and computation of IPCP data structures. This is the initial
3575 intraprocedural analysis of functions, which gathers information to be
3576 propagated later on. */
3579 ipcp_generate_summary (void)
3581 struct cgraph_node *node;
3584 fprintf (dump_file, "\nIPA constant propagation start:\n");
3585 ipa_register_cgraph_hooks ();
3587 FOR_EACH_FUNCTION_WITH_GIMPLE_BODY (node)
3589 node->local.versionable
3590 = tree_versionable_function_p (node->symbol.decl);
3591 ipa_analyze_node (node);
3595 /* Write ipcp summary for nodes in SET. */
3598 ipcp_write_summary (void)
3600 ipa_prop_write_jump_functions ();
3603 /* Read ipcp summary. */
3606 ipcp_read_summary (void)
3608 ipa_prop_read_jump_functions ();
3611 /* Gate for IPCP optimization. */
3614 cgraph_gate_cp (void)
3616 /* FIXME: We should remove the optimize check after we ensure we never run
3617 IPA passes when not optimizing. */
3618 return flag_ipa_cp && optimize;
3621 struct ipa_opt_pass_d pass_ipa_cp =
3626 OPTGROUP_NONE, /* optinfo_flags */
3627 cgraph_gate_cp, /* gate */
3628 ipcp_driver, /* execute */
3631 0, /* static_pass_number */
3632 TV_IPA_CONSTANT_PROP, /* tv_id */
3633 0, /* properties_required */
3634 0, /* properties_provided */
3635 0, /* properties_destroyed */
3636 0, /* todo_flags_start */
3638 TODO_remove_functions | TODO_ggc_collect /* todo_flags_finish */
3640 ipcp_generate_summary, /* generate_summary */
3641 ipcp_write_summary, /* write_summary */
3642 ipcp_read_summary, /* read_summary */
3643 ipa_prop_write_all_agg_replacement, /* write_optimization_summary */
3644 ipa_prop_read_all_agg_replacement, /* read_optimization_summary */
3645 NULL, /* stmt_fixup */
3647 ipcp_transform_function, /* function_transform */
3648 NULL, /* variable_transform */