2013-05-16 Richard Biener <rguenther@suse.de>
[official-gcc.git] / gcc / ipa-cp.c
blob40c946dc1f110d8c979ddff424da8b9428de6056
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
5 <mjambor@suse.cz>
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
12 version.
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
17 for more details.
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
34 is deemed good.
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
46 calls are redirected.
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
61 values:
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
94 third stage.
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
101 the second stage. */
103 #include "config.h"
104 #include "system.h"
105 #include "coretypes.h"
106 #include "tree.h"
107 #include "target.h"
108 #include "gimple.h"
109 #include "cgraph.h"
110 #include "ipa-prop.h"
111 #include "tree-flow.h"
112 #include "tree-pass.h"
113 #include "flags.h"
114 #include "diagnostic.h"
115 #include "tree-pretty-print.h"
116 #include "tree-inline.h"
117 #include "params.h"
118 #include "ipa-inline.h"
119 #include "ipa-utils.h"
121 struct ipcp_value;
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. */
141 int index;
144 /* Describes one particular value stored in struct ipcp_lattice. */
146 struct ipcp_value
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
150 devirtualization. */
151 tree value;
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
157 of values. */
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)
163 created. */
164 struct cgraph_node *spec_node;
165 /* Depth first search number and low link for topological sorting of
166 values. */
167 int dfs, low_link;
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. */
175 bool on_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. */
184 struct ipcp_lattice
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. */
191 int values_count;
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
195 propagation). */
196 bool bottom;
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. */
207 HOST_WIDE_INT size;
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 */
223 int aggs_count;
224 /* True if aggregate data were passed by reference (as opposed to by
225 value). */
226 bool aggs_by_ref;
227 /* All aggregate lattices contain a variable component (in addition to
228 values). */
229 bool aggs_contain_variable;
230 /* The value of all aggregate lattices is bottom (i.e. variable and unusable
231 for any propagation). */
232 bool aggs_bottom;
234 /* There is a virtual call based on this parameter. */
235 bool virt_call;
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
277 undefined value. */
279 static inline bool
280 ipa_lat_is_single_const (struct ipcp_lattice *lat)
282 if (lat->bottom
283 || lat->contains_variable
284 || lat->values_count != 1)
285 return false;
286 else
287 return true;
290 /* Return true iff the CS is an edge within a strongly connected component as
291 computed by ipa_reduced_postorder. */
293 static inline bool
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;
301 return (caller_dfs
302 && callee_dfs
303 && caller_dfs->scc_no == callee_dfs->scc_no);
306 /* Print V which is extracted from a value in a lattice to F. */
308 static void
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)
319 fprintf (f, "& ");
320 print_generic_expr (f, DECL_INITIAL (TREE_OPERAND (v, 0)), 0);
322 else
323 print_generic_expr (f, v, 0);
326 /* Print a lattice LAT to F. */
328 static void
329 print_lattice (FILE * f, struct ipcp_lattice *lat,
330 bool dump_sources, bool dump_benefits)
332 struct ipcp_value *val;
333 bool prev = false;
335 if (lat->bottom)
337 fprintf (f, "BOTTOM\n");
338 return;
341 if (!lat->values_count && !lat->contains_variable)
343 fprintf (f, "TOP\n");
344 return;
347 if (lat->contains_variable)
349 fprintf (f, "VARIABLE");
350 prev = true;
351 if (dump_benefits)
352 fprintf (f, "\n");
355 for (val = lat->values; val; val = val->next)
357 if (dump_benefits && prev)
358 fprintf (f, " ");
359 else if (!dump_benefits && prev)
360 fprintf (f, ", ");
361 else
362 prev = true;
364 print_ipcp_constant_value (f, val->value);
366 if (dump_sources)
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->symbol.order,
373 s->cs->frequency);
374 fprintf (f, "]");
377 if (dump_benefits)
378 fprintf (f, " [loc_time: %i, loc_size: %i, "
379 "prop_time: %i, prop_size: %i]\n",
380 val->local_time_benefit, val->local_size_cost,
381 val->prop_time_benefit, val->prop_size_cost);
383 if (!dump_benefits)
384 fprintf (f, "\n");
387 /* Print all ipcp_lattices of all functions to F. */
389 static void
390 print_all_lattices (FILE * f, bool dump_sources, bool dump_benefits)
392 struct cgraph_node *node;
393 int i, count;
395 fprintf (f, "\nLattices:\n");
396 FOR_EACH_FUNCTION_WITH_GIMPLE_BODY (node)
398 struct ipa_node_params *info;
400 info = IPA_NODE_REF (node);
401 fprintf (f, " Node: %s/%i:\n", cgraph_node_name (node),
402 node->symbol.order);
403 count = ipa_get_param_count (info);
404 for (i = 0; i < count; i++)
406 struct ipcp_agg_lattice *aglat;
407 struct ipcp_param_lattices *plats = ipa_get_parm_lattices (info, i);
408 fprintf (f, " param [%d]: ", i);
409 print_lattice (f, &plats->itself, dump_sources, dump_benefits);
411 if (plats->virt_call)
412 fprintf (f, " virt_call flag set\n");
414 if (plats->aggs_bottom)
416 fprintf (f, " AGGS BOTTOM\n");
417 continue;
419 if (plats->aggs_contain_variable)
420 fprintf (f, " AGGS VARIABLE\n");
421 for (aglat = plats->aggs; aglat; aglat = aglat->next)
423 fprintf (f, " %soffset " HOST_WIDE_INT_PRINT_DEC ": ",
424 plats->aggs_by_ref ? "ref " : "", aglat->offset);
425 print_lattice (f, aglat, dump_sources, dump_benefits);
431 /* Determine whether it is at all technically possible to create clones of NODE
432 and store this information in the ipa_node_params structure associated
433 with NODE. */
435 static void
436 determine_versionability (struct cgraph_node *node)
438 const char *reason = NULL;
440 /* There are a number of generic reasons functions cannot be versioned. We
441 also cannot remove parameters if there are type attributes such as fnspec
442 present. */
443 if (node->alias || node->thunk.thunk_p)
444 reason = "alias or thunk";
445 else if (!node->local.versionable)
446 reason = "not a tree_versionable_function";
447 else if (cgraph_function_body_availability (node) <= AVAIL_OVERWRITABLE)
448 reason = "insufficient body availability";
450 if (reason && dump_file && !node->alias && !node->thunk.thunk_p)
451 fprintf (dump_file, "Function %s/%i is not versionable, reason: %s.\n",
452 cgraph_node_name (node), node->symbol.order, reason);
454 node->local.versionable = (reason == NULL);
457 /* Return true if it is at all technically possible to create clones of a
458 NODE. */
460 static bool
461 ipcp_versionable_function_p (struct cgraph_node *node)
463 return node->local.versionable;
466 /* Structure holding accumulated information about callers of a node. */
468 struct caller_statistics
470 gcov_type count_sum;
471 int n_calls, n_hot_calls, freq_sum;
474 /* Initialize fields of STAT to zeroes. */
476 static inline void
477 init_caller_stats (struct caller_statistics *stats)
479 stats->count_sum = 0;
480 stats->n_calls = 0;
481 stats->n_hot_calls = 0;
482 stats->freq_sum = 0;
485 /* Worker callback of cgraph_for_node_and_aliases accumulating statistics of
486 non-thunk incoming edges to NODE. */
488 static bool
489 gather_caller_stats (struct cgraph_node *node, void *data)
491 struct caller_statistics *stats = (struct caller_statistics *) data;
492 struct cgraph_edge *cs;
494 for (cs = node->callers; cs; cs = cs->next_caller)
495 if (cs->caller->thunk.thunk_p)
496 cgraph_for_node_and_aliases (cs->caller, gather_caller_stats,
497 stats, false);
498 else
500 stats->count_sum += cs->count;
501 stats->freq_sum += cs->frequency;
502 stats->n_calls++;
503 if (cgraph_maybe_hot_edge_p (cs))
504 stats->n_hot_calls ++;
506 return false;
510 /* Return true if this NODE is viable candidate for cloning. */
512 static bool
513 ipcp_cloning_candidate_p (struct cgraph_node *node)
515 struct caller_statistics stats;
517 gcc_checking_assert (cgraph_function_with_gimple_body_p (node));
519 if (!flag_ipa_cp_clone)
521 if (dump_file)
522 fprintf (dump_file, "Not considering %s for cloning; "
523 "-fipa-cp-clone disabled.\n",
524 cgraph_node_name (node));
525 return false;
528 if (!optimize_function_for_speed_p (DECL_STRUCT_FUNCTION (node->symbol.decl)))
530 if (dump_file)
531 fprintf (dump_file, "Not considering %s for cloning; "
532 "optimizing it for size.\n",
533 cgraph_node_name (node));
534 return false;
537 init_caller_stats (&stats);
538 cgraph_for_node_and_aliases (node, gather_caller_stats, &stats, false);
540 if (inline_summary (node)->self_size < stats.n_calls)
542 if (dump_file)
543 fprintf (dump_file, "Considering %s for cloning; code might shrink.\n",
544 cgraph_node_name (node));
545 return true;
548 /* When profile is available and function is hot, propagate into it even if
549 calls seems cold; constant propagation can improve function's speed
550 significantly. */
551 if (max_count)
553 if (stats.count_sum > node->count * 90 / 100)
555 if (dump_file)
556 fprintf (dump_file, "Considering %s for cloning; "
557 "usually called directly.\n",
558 cgraph_node_name (node));
559 return true;
562 if (!stats.n_hot_calls)
564 if (dump_file)
565 fprintf (dump_file, "Not considering %s for cloning; no hot calls.\n",
566 cgraph_node_name (node));
567 return false;
569 if (dump_file)
570 fprintf (dump_file, "Considering %s for cloning.\n",
571 cgraph_node_name (node));
572 return true;
575 /* Arrays representing a topological ordering of call graph nodes and a stack
576 of noes used during constant propagation. */
578 struct topo_info
580 struct cgraph_node **order;
581 struct cgraph_node **stack;
582 int nnodes, stack_top;
585 /* Allocate the arrays in TOPO and topologically sort the nodes into order. */
587 static void
588 build_toporder_info (struct topo_info *topo)
590 topo->order = XCNEWVEC (struct cgraph_node *, cgraph_n_nodes);
591 topo->stack = XCNEWVEC (struct cgraph_node *, cgraph_n_nodes);
592 topo->stack_top = 0;
593 topo->nnodes = ipa_reduced_postorder (topo->order, true, true, NULL);
596 /* Free information about strongly connected components and the arrays in
597 TOPO. */
599 static void
600 free_toporder_info (struct topo_info *topo)
602 ipa_free_postorder_info ();
603 free (topo->order);
604 free (topo->stack);
607 /* Add NODE to the stack in TOPO, unless it is already there. */
609 static inline void
610 push_node_to_stack (struct topo_info *topo, struct cgraph_node *node)
612 struct ipa_node_params *info = IPA_NODE_REF (node);
613 if (info->node_enqueued)
614 return;
615 info->node_enqueued = 1;
616 topo->stack[topo->stack_top++] = node;
619 /* Pop a node from the stack in TOPO and return it or return NULL if the stack
620 is empty. */
622 static struct cgraph_node *
623 pop_node_from_stack (struct topo_info *topo)
625 if (topo->stack_top)
627 struct cgraph_node *node;
628 topo->stack_top--;
629 node = topo->stack[topo->stack_top];
630 IPA_NODE_REF (node)->node_enqueued = 0;
631 return node;
633 else
634 return NULL;
637 /* Set lattice LAT to bottom and return true if it previously was not set as
638 such. */
640 static inline bool
641 set_lattice_to_bottom (struct ipcp_lattice *lat)
643 bool ret = !lat->bottom;
644 lat->bottom = true;
645 return ret;
648 /* Mark lattice as containing an unknown value and return true if it previously
649 was not marked as such. */
651 static inline bool
652 set_lattice_contains_variable (struct ipcp_lattice *lat)
654 bool ret = !lat->contains_variable;
655 lat->contains_variable = true;
656 return ret;
659 /* Set all aggegate lattices in PLATS to bottom and return true if they were
660 not previously set as such. */
662 static inline bool
663 set_agg_lats_to_bottom (struct ipcp_param_lattices *plats)
665 bool ret = !plats->aggs_bottom;
666 plats->aggs_bottom = true;
667 return ret;
670 /* Mark all aggegate lattices in PLATS as containing an unknown value and
671 return true if they were not previously marked as such. */
673 static inline bool
674 set_agg_lats_contain_variable (struct ipcp_param_lattices *plats)
676 bool ret = !plats->aggs_contain_variable;
677 plats->aggs_contain_variable = true;
678 return ret;
681 /* Mark bot aggregate and scalar lattices as containing an unknown variable,
682 return true is any of them has not been marked as such so far. */
684 static inline bool
685 set_all_contains_variable (struct ipcp_param_lattices *plats)
687 bool ret = !plats->itself.contains_variable || !plats->aggs_contain_variable;
688 plats->itself.contains_variable = true;
689 plats->aggs_contain_variable = true;
690 return ret;
693 /* Initialize ipcp_lattices. */
695 static void
696 initialize_node_lattices (struct cgraph_node *node)
698 struct ipa_node_params *info = IPA_NODE_REF (node);
699 struct cgraph_edge *ie;
700 bool disable = false, variable = false;
701 int i;
703 gcc_checking_assert (cgraph_function_with_gimple_body_p (node));
704 if (!node->local.local)
706 /* When cloning is allowed, we can assume that externally visible
707 functions are not called. We will compensate this by cloning
708 later. */
709 if (ipcp_versionable_function_p (node)
710 && ipcp_cloning_candidate_p (node))
711 variable = true;
712 else
713 disable = true;
716 if (disable || variable)
718 for (i = 0; i < ipa_get_param_count (info) ; i++)
720 struct ipcp_param_lattices *plats = ipa_get_parm_lattices (info, i);
721 if (disable)
723 set_lattice_to_bottom (&plats->itself);
724 set_agg_lats_to_bottom (plats);
726 else
727 set_all_contains_variable (plats);
729 if (dump_file && (dump_flags & TDF_DETAILS)
730 && !node->alias && !node->thunk.thunk_p)
731 fprintf (dump_file, "Marking all lattices of %s/%i as %s\n",
732 cgraph_node_name (node), node->symbol.order,
733 disable ? "BOTTOM" : "VARIABLE");
736 for (ie = node->indirect_calls; ie; ie = ie->next_callee)
737 if (ie->indirect_info->polymorphic)
739 gcc_checking_assert (ie->indirect_info->param_index >= 0);
740 ipa_get_parm_lattices (info,
741 ie->indirect_info->param_index)->virt_call = 1;
745 /* Return the result of a (possibly arithmetic) pass through jump function
746 JFUNC on the constant value INPUT. Return NULL_TREE if that cannot be
747 determined or itself is considered an interprocedural invariant. */
749 static tree
750 ipa_get_jf_pass_through_result (struct ipa_jump_func *jfunc, tree input)
752 tree restype, res;
754 if (ipa_get_jf_pass_through_operation (jfunc) == NOP_EXPR)
755 return input;
756 else if (TREE_CODE (input) == TREE_BINFO)
757 return NULL_TREE;
759 gcc_checking_assert (is_gimple_ip_invariant (input));
760 if (TREE_CODE_CLASS (ipa_get_jf_pass_through_operation (jfunc))
761 == tcc_comparison)
762 restype = boolean_type_node;
763 else
764 restype = TREE_TYPE (input);
765 res = fold_binary (ipa_get_jf_pass_through_operation (jfunc), restype,
766 input, ipa_get_jf_pass_through_operand (jfunc));
768 if (res && !is_gimple_ip_invariant (res))
769 return NULL_TREE;
771 return res;
774 /* Return the result of an ancestor jump function JFUNC on the constant value
775 INPUT. Return NULL_TREE if that cannot be determined. */
777 static tree
778 ipa_get_jf_ancestor_result (struct ipa_jump_func *jfunc, tree input)
780 if (TREE_CODE (input) == TREE_BINFO)
781 return get_binfo_at_offset (input,
782 ipa_get_jf_ancestor_offset (jfunc),
783 ipa_get_jf_ancestor_type (jfunc));
784 else if (TREE_CODE (input) == ADDR_EXPR)
786 tree t = TREE_OPERAND (input, 0);
787 t = build_ref_for_offset (EXPR_LOCATION (t), t,
788 ipa_get_jf_ancestor_offset (jfunc),
789 ipa_get_jf_ancestor_type (jfunc), NULL, false);
790 return build_fold_addr_expr (t);
792 else
793 return NULL_TREE;
796 /* Determine whether JFUNC evaluates to a known value (that is either a
797 constant or a binfo) and if so, return it. Otherwise return NULL. INFO
798 describes the caller node so that pass-through jump functions can be
799 evaluated. */
801 tree
802 ipa_value_from_jfunc (struct ipa_node_params *info, struct ipa_jump_func *jfunc)
804 if (jfunc->type == IPA_JF_CONST)
805 return ipa_get_jf_constant (jfunc);
806 else if (jfunc->type == IPA_JF_KNOWN_TYPE)
807 return ipa_binfo_from_known_type_jfunc (jfunc);
808 else if (jfunc->type == IPA_JF_PASS_THROUGH
809 || jfunc->type == IPA_JF_ANCESTOR)
811 tree input;
812 int idx;
814 if (jfunc->type == IPA_JF_PASS_THROUGH)
815 idx = ipa_get_jf_pass_through_formal_id (jfunc);
816 else
817 idx = ipa_get_jf_ancestor_formal_id (jfunc);
819 if (info->ipcp_orig_node)
820 input = info->known_vals[idx];
821 else
823 struct ipcp_lattice *lat;
825 if (!info->lattices)
827 gcc_checking_assert (!flag_ipa_cp);
828 return NULL_TREE;
830 lat = ipa_get_scalar_lat (info, idx);
831 if (!ipa_lat_is_single_const (lat))
832 return NULL_TREE;
833 input = lat->values->value;
836 if (!input)
837 return NULL_TREE;
839 if (jfunc->type == IPA_JF_PASS_THROUGH)
840 return ipa_get_jf_pass_through_result (jfunc, input);
841 else
842 return ipa_get_jf_ancestor_result (jfunc, input);
844 else
845 return NULL_TREE;
849 /* If checking is enabled, verify that no lattice is in the TOP state, i.e. not
850 bottom, not containing a variable component and without any known value at
851 the same time. */
853 DEBUG_FUNCTION void
854 ipcp_verify_propagated_values (void)
856 struct cgraph_node *node;
858 FOR_EACH_FUNCTION_WITH_GIMPLE_BODY (node)
860 struct ipa_node_params *info = IPA_NODE_REF (node);
861 int i, count = ipa_get_param_count (info);
863 for (i = 0; i < count; i++)
865 struct ipcp_lattice *lat = ipa_get_scalar_lat (info, i);
867 if (!lat->bottom
868 && !lat->contains_variable
869 && lat->values_count == 0)
871 if (dump_file)
873 fprintf (dump_file, "\nIPA lattices after constant "
874 "propagation:\n");
875 print_all_lattices (dump_file, true, false);
878 gcc_unreachable ();
884 /* Return true iff X and Y should be considered equal values by IPA-CP. */
886 static bool
887 values_equal_for_ipcp_p (tree x, tree y)
889 gcc_checking_assert (x != NULL_TREE && y != NULL_TREE);
891 if (x == y)
892 return true;
894 if (TREE_CODE (x) == TREE_BINFO || TREE_CODE (y) == TREE_BINFO)
895 return false;
897 if (TREE_CODE (x) == ADDR_EXPR
898 && TREE_CODE (y) == ADDR_EXPR
899 && TREE_CODE (TREE_OPERAND (x, 0)) == CONST_DECL
900 && TREE_CODE (TREE_OPERAND (y, 0)) == CONST_DECL)
901 return operand_equal_p (DECL_INITIAL (TREE_OPERAND (x, 0)),
902 DECL_INITIAL (TREE_OPERAND (y, 0)), 0);
903 else
904 return operand_equal_p (x, y, 0);
907 /* Add a new value source to VAL, marking that a value comes from edge CS and
908 (if the underlying jump function is a pass-through or an ancestor one) from
909 a caller value SRC_VAL of a caller parameter described by SRC_INDEX. OFFSET
910 is negative if the source was the scalar value of the parameter itself or
911 the offset within an aggregate. */
913 static void
914 add_value_source (struct ipcp_value *val, struct cgraph_edge *cs,
915 struct ipcp_value *src_val, int src_idx, HOST_WIDE_INT offset)
917 struct ipcp_value_source *src;
919 src = (struct ipcp_value_source *) pool_alloc (ipcp_sources_pool);
920 src->offset = offset;
921 src->cs = cs;
922 src->val = src_val;
923 src->index = src_idx;
925 src->next = val->sources;
926 val->sources = src;
929 /* Try to add NEWVAL to LAT, potentially creating a new struct ipcp_value for
930 it. CS, SRC_VAL SRC_INDEX and OFFSET are meant for add_value_source and
931 have the same meaning. */
933 static bool
934 add_value_to_lattice (struct ipcp_lattice *lat, tree newval,
935 struct cgraph_edge *cs, struct ipcp_value *src_val,
936 int src_idx, HOST_WIDE_INT offset)
938 struct ipcp_value *val;
940 if (lat->bottom)
941 return false;
943 for (val = lat->values; val; val = val->next)
944 if (values_equal_for_ipcp_p (val->value, newval))
946 if (edge_within_scc (cs))
948 struct ipcp_value_source *s;
949 for (s = val->sources; s ; s = s->next)
950 if (s->cs == cs)
951 break;
952 if (s)
953 return false;
956 add_value_source (val, cs, src_val, src_idx, offset);
957 return false;
960 if (lat->values_count == PARAM_VALUE (PARAM_IPA_CP_VALUE_LIST_SIZE))
962 /* We can only free sources, not the values themselves, because sources
963 of other values in this this SCC might point to them. */
964 for (val = lat->values; val; val = val->next)
966 while (val->sources)
968 struct ipcp_value_source *src = val->sources;
969 val->sources = src->next;
970 pool_free (ipcp_sources_pool, src);
974 lat->values = NULL;
975 return set_lattice_to_bottom (lat);
978 lat->values_count++;
979 val = (struct ipcp_value *) pool_alloc (ipcp_values_pool);
980 memset (val, 0, sizeof (*val));
982 add_value_source (val, cs, src_val, src_idx, offset);
983 val->value = newval;
984 val->next = lat->values;
985 lat->values = val;
986 return true;
989 /* Like above but passes a special value of offset to distinguish that the
990 origin is the scalar value of the parameter rather than a part of an
991 aggregate. */
993 static inline bool
994 add_scalar_value_to_lattice (struct ipcp_lattice *lat, tree newval,
995 struct cgraph_edge *cs,
996 struct ipcp_value *src_val, int src_idx)
998 return add_value_to_lattice (lat, newval, cs, src_val, src_idx, -1);
1001 /* Propagate values through a pass-through jump function JFUNC associated with
1002 edge CS, taking values from SRC_LAT and putting them into DEST_LAT. SRC_IDX
1003 is the index of the source parameter. */
1005 static bool
1006 propagate_vals_accross_pass_through (struct cgraph_edge *cs,
1007 struct ipa_jump_func *jfunc,
1008 struct ipcp_lattice *src_lat,
1009 struct ipcp_lattice *dest_lat,
1010 int src_idx)
1012 struct ipcp_value *src_val;
1013 bool ret = false;
1015 if (ipa_get_jf_pass_through_operation (jfunc) == NOP_EXPR)
1016 for (src_val = src_lat->values; src_val; src_val = src_val->next)
1017 ret |= add_scalar_value_to_lattice (dest_lat, src_val->value, cs,
1018 src_val, src_idx);
1019 /* Do not create new values when propagating within an SCC because if there
1020 are arithmetic functions with circular dependencies, there is infinite
1021 number of them and we would just make lattices bottom. */
1022 else if (edge_within_scc (cs))
1023 ret = set_lattice_contains_variable (dest_lat);
1024 else
1025 for (src_val = src_lat->values; src_val; src_val = src_val->next)
1027 tree cstval = src_val->value;
1029 if (TREE_CODE (cstval) == TREE_BINFO)
1031 ret |= set_lattice_contains_variable (dest_lat);
1032 continue;
1034 cstval = ipa_get_jf_pass_through_result (jfunc, cstval);
1036 if (cstval)
1037 ret |= add_scalar_value_to_lattice (dest_lat, cstval, cs, src_val,
1038 src_idx);
1039 else
1040 ret |= set_lattice_contains_variable (dest_lat);
1043 return ret;
1046 /* Propagate values through an ancestor jump function JFUNC associated with
1047 edge CS, taking values from SRC_LAT and putting them into DEST_LAT. SRC_IDX
1048 is the index of the source parameter. */
1050 static bool
1051 propagate_vals_accross_ancestor (struct cgraph_edge *cs,
1052 struct ipa_jump_func *jfunc,
1053 struct ipcp_lattice *src_lat,
1054 struct ipcp_lattice *dest_lat,
1055 int src_idx)
1057 struct ipcp_value *src_val;
1058 bool ret = false;
1060 if (edge_within_scc (cs))
1061 return set_lattice_contains_variable (dest_lat);
1063 for (src_val = src_lat->values; src_val; src_val = src_val->next)
1065 tree t = ipa_get_jf_ancestor_result (jfunc, src_val->value);
1067 if (t)
1068 ret |= add_scalar_value_to_lattice (dest_lat, t, cs, src_val, src_idx);
1069 else
1070 ret |= set_lattice_contains_variable (dest_lat);
1073 return ret;
1076 /* Propagate scalar values across jump function JFUNC that is associated with
1077 edge CS and put the values into DEST_LAT. */
1079 static bool
1080 propagate_scalar_accross_jump_function (struct cgraph_edge *cs,
1081 struct ipa_jump_func *jfunc,
1082 struct ipcp_lattice *dest_lat)
1084 if (dest_lat->bottom)
1085 return false;
1087 if (jfunc->type == IPA_JF_CONST
1088 || jfunc->type == IPA_JF_KNOWN_TYPE)
1090 tree val;
1092 if (jfunc->type == IPA_JF_KNOWN_TYPE)
1094 val = ipa_binfo_from_known_type_jfunc (jfunc);
1095 if (!val)
1096 return set_lattice_contains_variable (dest_lat);
1098 else
1099 val = ipa_get_jf_constant (jfunc);
1100 return add_scalar_value_to_lattice (dest_lat, val, cs, NULL, 0);
1102 else if (jfunc->type == IPA_JF_PASS_THROUGH
1103 || jfunc->type == IPA_JF_ANCESTOR)
1105 struct ipa_node_params *caller_info = IPA_NODE_REF (cs->caller);
1106 struct ipcp_lattice *src_lat;
1107 int src_idx;
1108 bool ret;
1110 if (jfunc->type == IPA_JF_PASS_THROUGH)
1111 src_idx = ipa_get_jf_pass_through_formal_id (jfunc);
1112 else
1113 src_idx = ipa_get_jf_ancestor_formal_id (jfunc);
1115 src_lat = ipa_get_scalar_lat (caller_info, src_idx);
1116 if (src_lat->bottom)
1117 return set_lattice_contains_variable (dest_lat);
1119 /* If we would need to clone the caller and cannot, do not propagate. */
1120 if (!ipcp_versionable_function_p (cs->caller)
1121 && (src_lat->contains_variable
1122 || (src_lat->values_count > 1)))
1123 return set_lattice_contains_variable (dest_lat);
1125 if (jfunc->type == IPA_JF_PASS_THROUGH)
1126 ret = propagate_vals_accross_pass_through (cs, jfunc, src_lat,
1127 dest_lat, src_idx);
1128 else
1129 ret = propagate_vals_accross_ancestor (cs, jfunc, src_lat, dest_lat,
1130 src_idx);
1132 if (src_lat->contains_variable)
1133 ret |= set_lattice_contains_variable (dest_lat);
1135 return ret;
1138 /* TODO: We currently do not handle member method pointers in IPA-CP (we only
1139 use it for indirect inlining), we should propagate them too. */
1140 return set_lattice_contains_variable (dest_lat);
1143 /* If DEST_PLATS already has aggregate items, check that aggs_by_ref matches
1144 NEW_AGGS_BY_REF and if not, mark all aggs as bottoms and return true (in all
1145 other cases, return false). If there are no aggregate items, set
1146 aggs_by_ref to NEW_AGGS_BY_REF. */
1148 static bool
1149 set_check_aggs_by_ref (struct ipcp_param_lattices *dest_plats,
1150 bool new_aggs_by_ref)
1152 if (dest_plats->aggs)
1154 if (dest_plats->aggs_by_ref != new_aggs_by_ref)
1156 set_agg_lats_to_bottom (dest_plats);
1157 return true;
1160 else
1161 dest_plats->aggs_by_ref = new_aggs_by_ref;
1162 return false;
1165 /* Walk aggregate lattices in DEST_PLATS from ***AGLAT on, until ***aglat is an
1166 already existing lattice for the given OFFSET and SIZE, marking all skipped
1167 lattices as containing variable and checking for overlaps. If there is no
1168 already existing lattice for the OFFSET and VAL_SIZE, create one, initialize
1169 it with offset, size and contains_variable to PRE_EXISTING, and return true,
1170 unless there are too many already. If there are two many, return false. If
1171 there are overlaps turn whole DEST_PLATS to bottom and return false. If any
1172 skipped lattices were newly marked as containing variable, set *CHANGE to
1173 true. */
1175 static bool
1176 merge_agg_lats_step (struct ipcp_param_lattices *dest_plats,
1177 HOST_WIDE_INT offset, HOST_WIDE_INT val_size,
1178 struct ipcp_agg_lattice ***aglat,
1179 bool pre_existing, bool *change)
1181 gcc_checking_assert (offset >= 0);
1183 while (**aglat && (**aglat)->offset < offset)
1185 if ((**aglat)->offset + (**aglat)->size > offset)
1187 set_agg_lats_to_bottom (dest_plats);
1188 return false;
1190 *change |= set_lattice_contains_variable (**aglat);
1191 *aglat = &(**aglat)->next;
1194 if (**aglat && (**aglat)->offset == offset)
1196 if ((**aglat)->size != val_size
1197 || ((**aglat)->next
1198 && (**aglat)->next->offset < offset + val_size))
1200 set_agg_lats_to_bottom (dest_plats);
1201 return false;
1203 gcc_checking_assert (!(**aglat)->next
1204 || (**aglat)->next->offset >= offset + val_size);
1205 return true;
1207 else
1209 struct ipcp_agg_lattice *new_al;
1211 if (**aglat && (**aglat)->offset < offset + val_size)
1213 set_agg_lats_to_bottom (dest_plats);
1214 return false;
1216 if (dest_plats->aggs_count == PARAM_VALUE (PARAM_IPA_MAX_AGG_ITEMS))
1217 return false;
1218 dest_plats->aggs_count++;
1219 new_al = (struct ipcp_agg_lattice *) pool_alloc (ipcp_agg_lattice_pool);
1220 memset (new_al, 0, sizeof (*new_al));
1222 new_al->offset = offset;
1223 new_al->size = val_size;
1224 new_al->contains_variable = pre_existing;
1226 new_al->next = **aglat;
1227 **aglat = new_al;
1228 return true;
1232 /* Set all AGLAT and all other aggregate lattices reachable by next pointers as
1233 containing an unknown value. */
1235 static bool
1236 set_chain_of_aglats_contains_variable (struct ipcp_agg_lattice *aglat)
1238 bool ret = false;
1239 while (aglat)
1241 ret |= set_lattice_contains_variable (aglat);
1242 aglat = aglat->next;
1244 return ret;
1247 /* Merge existing aggregate lattices in SRC_PLATS to DEST_PLATS, subtracting
1248 DELTA_OFFSET. CS is the call graph edge and SRC_IDX the index of the source
1249 parameter used for lattice value sources. Return true if DEST_PLATS changed
1250 in any way. */
1252 static bool
1253 merge_aggregate_lattices (struct cgraph_edge *cs,
1254 struct ipcp_param_lattices *dest_plats,
1255 struct ipcp_param_lattices *src_plats,
1256 int src_idx, HOST_WIDE_INT offset_delta)
1258 bool pre_existing = dest_plats->aggs != NULL;
1259 struct ipcp_agg_lattice **dst_aglat;
1260 bool ret = false;
1262 if (set_check_aggs_by_ref (dest_plats, src_plats->aggs_by_ref))
1263 return true;
1264 if (src_plats->aggs_bottom)
1265 return set_agg_lats_contain_variable (dest_plats);
1266 if (src_plats->aggs_contain_variable)
1267 ret |= set_agg_lats_contain_variable (dest_plats);
1268 dst_aglat = &dest_plats->aggs;
1270 for (struct ipcp_agg_lattice *src_aglat = src_plats->aggs;
1271 src_aglat;
1272 src_aglat = src_aglat->next)
1274 HOST_WIDE_INT new_offset = src_aglat->offset - offset_delta;
1276 if (new_offset < 0)
1277 continue;
1278 if (merge_agg_lats_step (dest_plats, new_offset, src_aglat->size,
1279 &dst_aglat, pre_existing, &ret))
1281 struct ipcp_agg_lattice *new_al = *dst_aglat;
1283 dst_aglat = &(*dst_aglat)->next;
1284 if (src_aglat->bottom)
1286 ret |= set_lattice_contains_variable (new_al);
1287 continue;
1289 if (src_aglat->contains_variable)
1290 ret |= set_lattice_contains_variable (new_al);
1291 for (struct ipcp_value *val = src_aglat->values;
1292 val;
1293 val = val->next)
1294 ret |= add_value_to_lattice (new_al, val->value, cs, val, src_idx,
1295 src_aglat->offset);
1297 else if (dest_plats->aggs_bottom)
1298 return true;
1300 ret |= set_chain_of_aglats_contains_variable (*dst_aglat);
1301 return ret;
1304 /* Determine whether there is anything to propagate FROM SRC_PLATS through a
1305 pass-through JFUNC and if so, whether it has conform and conforms to the
1306 rules about propagating values passed by reference. */
1308 static bool
1309 agg_pass_through_permissible_p (struct ipcp_param_lattices *src_plats,
1310 struct ipa_jump_func *jfunc)
1312 return src_plats->aggs
1313 && (!src_plats->aggs_by_ref
1314 || ipa_get_jf_pass_through_agg_preserved (jfunc));
1317 /* Propagate scalar values across jump function JFUNC that is associated with
1318 edge CS and put the values into DEST_LAT. */
1320 static bool
1321 propagate_aggs_accross_jump_function (struct cgraph_edge *cs,
1322 struct ipa_jump_func *jfunc,
1323 struct ipcp_param_lattices *dest_plats)
1325 bool ret = false;
1327 if (dest_plats->aggs_bottom)
1328 return false;
1330 if (jfunc->type == IPA_JF_PASS_THROUGH
1331 && ipa_get_jf_pass_through_operation (jfunc) == NOP_EXPR)
1333 struct ipa_node_params *caller_info = IPA_NODE_REF (cs->caller);
1334 int src_idx = ipa_get_jf_pass_through_formal_id (jfunc);
1335 struct ipcp_param_lattices *src_plats;
1337 src_plats = ipa_get_parm_lattices (caller_info, src_idx);
1338 if (agg_pass_through_permissible_p (src_plats, jfunc))
1340 /* Currently we do not produce clobber aggregate jump
1341 functions, replace with merging when we do. */
1342 gcc_assert (!jfunc->agg.items);
1343 ret |= merge_aggregate_lattices (cs, dest_plats, src_plats,
1344 src_idx, 0);
1346 else
1347 ret |= set_agg_lats_contain_variable (dest_plats);
1349 else if (jfunc->type == IPA_JF_ANCESTOR
1350 && ipa_get_jf_ancestor_agg_preserved (jfunc))
1352 struct ipa_node_params *caller_info = IPA_NODE_REF (cs->caller);
1353 int src_idx = ipa_get_jf_ancestor_formal_id (jfunc);
1354 struct ipcp_param_lattices *src_plats;
1356 src_plats = ipa_get_parm_lattices (caller_info, src_idx);
1357 if (src_plats->aggs && src_plats->aggs_by_ref)
1359 /* Currently we do not produce clobber aggregate jump
1360 functions, replace with merging when we do. */
1361 gcc_assert (!jfunc->agg.items);
1362 ret |= merge_aggregate_lattices (cs, dest_plats, src_plats, src_idx,
1363 ipa_get_jf_ancestor_offset (jfunc));
1365 else if (!src_plats->aggs_by_ref)
1366 ret |= set_agg_lats_to_bottom (dest_plats);
1367 else
1368 ret |= set_agg_lats_contain_variable (dest_plats);
1370 else if (jfunc->agg.items)
1372 bool pre_existing = dest_plats->aggs != NULL;
1373 struct ipcp_agg_lattice **aglat = &dest_plats->aggs;
1374 struct ipa_agg_jf_item *item;
1375 int i;
1377 if (set_check_aggs_by_ref (dest_plats, jfunc->agg.by_ref))
1378 return true;
1380 FOR_EACH_VEC_ELT (*jfunc->agg.items, i, item)
1382 HOST_WIDE_INT val_size;
1384 if (item->offset < 0)
1385 continue;
1386 gcc_checking_assert (is_gimple_ip_invariant (item->value));
1387 val_size = tree_low_cst (TYPE_SIZE (TREE_TYPE (item->value)), 1);
1389 if (merge_agg_lats_step (dest_plats, item->offset, val_size,
1390 &aglat, pre_existing, &ret))
1392 ret |= add_value_to_lattice (*aglat, item->value, cs, NULL, 0, 0);
1393 aglat = &(*aglat)->next;
1395 else if (dest_plats->aggs_bottom)
1396 return true;
1399 ret |= set_chain_of_aglats_contains_variable (*aglat);
1401 else
1402 ret |= set_agg_lats_contain_variable (dest_plats);
1404 return ret;
1407 /* Propagate constants from the caller to the callee of CS. INFO describes the
1408 caller. */
1410 static bool
1411 propagate_constants_accross_call (struct cgraph_edge *cs)
1413 struct ipa_node_params *callee_info;
1414 enum availability availability;
1415 struct cgraph_node *callee, *alias_or_thunk;
1416 struct ipa_edge_args *args;
1417 bool ret = false;
1418 int i, args_count, parms_count;
1420 callee = cgraph_function_node (cs->callee, &availability);
1421 if (!callee->analyzed)
1422 return false;
1423 gcc_checking_assert (cgraph_function_with_gimple_body_p (callee));
1424 callee_info = IPA_NODE_REF (callee);
1426 args = IPA_EDGE_REF (cs);
1427 args_count = ipa_get_cs_argument_count (args);
1428 parms_count = ipa_get_param_count (callee_info);
1430 /* If this call goes through a thunk we must not propagate to the first (0th)
1431 parameter. However, we might need to uncover a thunk from below a series
1432 of aliases first. */
1433 alias_or_thunk = cs->callee;
1434 while (alias_or_thunk->alias)
1435 alias_or_thunk = cgraph_alias_aliased_node (alias_or_thunk);
1436 if (alias_or_thunk->thunk.thunk_p)
1438 ret |= set_all_contains_variable (ipa_get_parm_lattices (callee_info,
1439 0));
1440 i = 1;
1442 else
1443 i = 0;
1445 for (; (i < args_count) && (i < parms_count); i++)
1447 struct ipa_jump_func *jump_func = ipa_get_ith_jump_func (args, i);
1448 struct ipcp_param_lattices *dest_plats;
1450 dest_plats = ipa_get_parm_lattices (callee_info, i);
1451 if (availability == AVAIL_OVERWRITABLE)
1452 ret |= set_all_contains_variable (dest_plats);
1453 else
1455 ret |= propagate_scalar_accross_jump_function (cs, jump_func,
1456 &dest_plats->itself);
1457 ret |= propagate_aggs_accross_jump_function (cs, jump_func,
1458 dest_plats);
1461 for (; i < parms_count; i++)
1462 ret |= set_all_contains_variable (ipa_get_parm_lattices (callee_info, i));
1464 return ret;
1467 /* If an indirect edge IE can be turned into a direct one based on KNOWN_VALS
1468 (which can contain both constants and binfos), KNOWN_BINFOS, KNOWN_AGGS or
1469 AGG_REPS return the destination. The latter three can be NULL. If AGG_REPS
1470 is not NULL, KNOWN_AGGS is ignored. */
1472 static tree
1473 ipa_get_indirect_edge_target_1 (struct cgraph_edge *ie,
1474 vec<tree> known_vals,
1475 vec<tree> known_binfos,
1476 vec<ipa_agg_jump_function_p> known_aggs,
1477 struct ipa_agg_replacement_value *agg_reps)
1479 int param_index = ie->indirect_info->param_index;
1480 HOST_WIDE_INT token, anc_offset;
1481 tree otr_type;
1482 tree t;
1484 if (param_index == -1)
1485 return NULL_TREE;
1487 if (!ie->indirect_info->polymorphic)
1489 tree t;
1491 if (ie->indirect_info->agg_contents)
1493 if (agg_reps)
1495 t = NULL;
1496 while (agg_reps)
1498 if (agg_reps->index == param_index
1499 && agg_reps->offset == ie->indirect_info->offset
1500 && agg_reps->by_ref == ie->indirect_info->by_ref)
1502 t = agg_reps->value;
1503 break;
1505 agg_reps = agg_reps->next;
1508 else if (known_aggs.length () > (unsigned int) param_index)
1510 struct ipa_agg_jump_function *agg;
1511 agg = known_aggs[param_index];
1512 t = ipa_find_agg_cst_for_param (agg, ie->indirect_info->offset,
1513 ie->indirect_info->by_ref);
1515 else
1516 t = NULL;
1518 else
1519 t = (known_vals.length () > (unsigned int) param_index
1520 ? known_vals[param_index] : NULL);
1522 if (t &&
1523 TREE_CODE (t) == ADDR_EXPR
1524 && TREE_CODE (TREE_OPERAND (t, 0)) == FUNCTION_DECL)
1525 return TREE_OPERAND (t, 0);
1526 else
1527 return NULL_TREE;
1530 gcc_assert (!ie->indirect_info->agg_contents);
1531 token = ie->indirect_info->otr_token;
1532 anc_offset = ie->indirect_info->offset;
1533 otr_type = ie->indirect_info->otr_type;
1535 t = known_vals[param_index];
1536 if (!t && known_binfos.length () > (unsigned int) param_index)
1537 t = known_binfos[param_index];
1538 if (!t)
1539 return NULL_TREE;
1541 if (TREE_CODE (t) != TREE_BINFO)
1543 tree binfo;
1544 binfo = gimple_extract_devirt_binfo_from_cst (t);
1545 if (!binfo)
1546 return NULL_TREE;
1547 binfo = get_binfo_at_offset (binfo, anc_offset, otr_type);
1548 if (!binfo)
1549 return NULL_TREE;
1550 return gimple_get_virt_method_for_binfo (token, binfo);
1552 else
1554 tree binfo;
1556 binfo = get_binfo_at_offset (t, anc_offset, otr_type);
1557 if (!binfo)
1558 return NULL_TREE;
1559 return gimple_get_virt_method_for_binfo (token, binfo);
1564 /* If an indirect edge IE can be turned into a direct one based on KNOWN_VALS
1565 (which can contain both constants and binfos), KNOWN_BINFOS (which can be
1566 NULL) or KNOWN_AGGS (which also can be NULL) return the destination. */
1568 tree
1569 ipa_get_indirect_edge_target (struct cgraph_edge *ie,
1570 vec<tree> known_vals,
1571 vec<tree> known_binfos,
1572 vec<ipa_agg_jump_function_p> known_aggs)
1574 return ipa_get_indirect_edge_target_1 (ie, known_vals, known_binfos,
1575 known_aggs, NULL);
1578 /* Calculate devirtualization time bonus for NODE, assuming we know KNOWN_CSTS
1579 and KNOWN_BINFOS. */
1581 static int
1582 devirtualization_time_bonus (struct cgraph_node *node,
1583 vec<tree> known_csts,
1584 vec<tree> known_binfos,
1585 vec<ipa_agg_jump_function_p> known_aggs)
1587 struct cgraph_edge *ie;
1588 int res = 0;
1590 for (ie = node->indirect_calls; ie; ie = ie->next_callee)
1592 struct cgraph_node *callee;
1593 struct inline_summary *isummary;
1594 tree target;
1596 target = ipa_get_indirect_edge_target (ie, known_csts, known_binfos,
1597 known_aggs);
1598 if (!target)
1599 continue;
1601 /* Only bare minimum benefit for clearly un-inlineable targets. */
1602 res += 1;
1603 callee = cgraph_get_node (target);
1604 if (!callee || !callee->analyzed)
1605 continue;
1606 isummary = inline_summary (callee);
1607 if (!isummary->inlinable)
1608 continue;
1610 /* FIXME: The values below need re-considering and perhaps also
1611 integrating into the cost metrics, at lest in some very basic way. */
1612 if (isummary->size <= MAX_INLINE_INSNS_AUTO / 4)
1613 res += 31;
1614 else if (isummary->size <= MAX_INLINE_INSNS_AUTO / 2)
1615 res += 15;
1616 else if (isummary->size <= MAX_INLINE_INSNS_AUTO
1617 || DECL_DECLARED_INLINE_P (callee->symbol.decl))
1618 res += 7;
1621 return res;
1624 /* Return time bonus incurred because of HINTS. */
1626 static int
1627 hint_time_bonus (inline_hints hints)
1629 int result = 0;
1630 if (hints & (INLINE_HINT_loop_iterations | INLINE_HINT_loop_stride))
1631 result += PARAM_VALUE (PARAM_IPA_CP_LOOP_HINT_BONUS);
1632 if (hints & INLINE_HINT_array_index)
1633 result += PARAM_VALUE (PARAM_IPA_CP_ARRAY_INDEX_HINT_BONUS);
1634 return result;
1637 /* Return true if cloning NODE is a good idea, given the estimated TIME_BENEFIT
1638 and SIZE_COST and with the sum of frequencies of incoming edges to the
1639 potential new clone in FREQUENCIES. */
1641 static bool
1642 good_cloning_opportunity_p (struct cgraph_node *node, int time_benefit,
1643 int freq_sum, gcov_type count_sum, int size_cost)
1645 if (time_benefit == 0
1646 || !flag_ipa_cp_clone
1647 || !optimize_function_for_speed_p (DECL_STRUCT_FUNCTION (node->symbol.decl)))
1648 return false;
1650 gcc_assert (size_cost > 0);
1652 if (max_count)
1654 int factor = (count_sum * 1000) / max_count;
1655 HOST_WIDEST_INT evaluation = (((HOST_WIDEST_INT) time_benefit * factor)
1656 / size_cost);
1658 if (dump_file && (dump_flags & TDF_DETAILS))
1659 fprintf (dump_file, " good_cloning_opportunity_p (time: %i, "
1660 "size: %i, count_sum: " HOST_WIDE_INT_PRINT_DEC
1661 ") -> evaluation: " HOST_WIDEST_INT_PRINT_DEC
1662 ", threshold: %i\n",
1663 time_benefit, size_cost, (HOST_WIDE_INT) count_sum,
1664 evaluation, PARAM_VALUE (PARAM_IPA_CP_EVAL_THRESHOLD));
1666 return evaluation >= PARAM_VALUE (PARAM_IPA_CP_EVAL_THRESHOLD);
1668 else
1670 HOST_WIDEST_INT evaluation = (((HOST_WIDEST_INT) time_benefit * freq_sum)
1671 / size_cost);
1673 if (dump_file && (dump_flags & TDF_DETAILS))
1674 fprintf (dump_file, " good_cloning_opportunity_p (time: %i, "
1675 "size: %i, freq_sum: %i) -> evaluation: "
1676 HOST_WIDEST_INT_PRINT_DEC ", threshold: %i\n",
1677 time_benefit, size_cost, freq_sum, evaluation,
1678 PARAM_VALUE (PARAM_IPA_CP_EVAL_THRESHOLD));
1680 return evaluation >= PARAM_VALUE (PARAM_IPA_CP_EVAL_THRESHOLD);
1684 /* Return all context independent values from aggregate lattices in PLATS in a
1685 vector. Return NULL if there are none. */
1687 static vec<ipa_agg_jf_item_t, va_gc> *
1688 context_independent_aggregate_values (struct ipcp_param_lattices *plats)
1690 vec<ipa_agg_jf_item_t, va_gc> *res = NULL;
1692 if (plats->aggs_bottom
1693 || plats->aggs_contain_variable
1694 || plats->aggs_count == 0)
1695 return NULL;
1697 for (struct ipcp_agg_lattice *aglat = plats->aggs;
1698 aglat;
1699 aglat = aglat->next)
1700 if (ipa_lat_is_single_const (aglat))
1702 struct ipa_agg_jf_item item;
1703 item.offset = aglat->offset;
1704 item.value = aglat->values->value;
1705 vec_safe_push (res, item);
1707 return res;
1710 /* Allocate KNOWN_CSTS, KNOWN_BINFOS and, if non-NULL, KNOWN_AGGS and populate
1711 them with values of parameters that are known independent of the context.
1712 INFO describes the function. If REMOVABLE_PARAMS_COST is non-NULL, the
1713 movement cost of all removable parameters will be stored in it. */
1715 static bool
1716 gather_context_independent_values (struct ipa_node_params *info,
1717 vec<tree> *known_csts,
1718 vec<tree> *known_binfos,
1719 vec<ipa_agg_jump_function_t> *known_aggs,
1720 int *removable_params_cost)
1722 int i, count = ipa_get_param_count (info);
1723 bool ret = false;
1725 known_csts->create (0);
1726 known_binfos->create (0);
1727 known_csts->safe_grow_cleared (count);
1728 known_binfos->safe_grow_cleared (count);
1729 if (known_aggs)
1731 known_aggs->create (0);
1732 known_aggs->safe_grow_cleared (count);
1735 if (removable_params_cost)
1736 *removable_params_cost = 0;
1738 for (i = 0; i < count ; i++)
1740 struct ipcp_param_lattices *plats = ipa_get_parm_lattices (info, i);
1741 struct ipcp_lattice *lat = &plats->itself;
1743 if (ipa_lat_is_single_const (lat))
1745 struct ipcp_value *val = lat->values;
1746 if (TREE_CODE (val->value) != TREE_BINFO)
1748 (*known_csts)[i] = val->value;
1749 if (removable_params_cost)
1750 *removable_params_cost
1751 += estimate_move_cost (TREE_TYPE (val->value));
1752 ret = true;
1754 else if (plats->virt_call)
1756 (*known_binfos)[i] = val->value;
1757 ret = true;
1759 else if (removable_params_cost
1760 && !ipa_is_param_used (info, i))
1761 *removable_params_cost
1762 += estimate_move_cost (TREE_TYPE (ipa_get_param (info, i)));
1764 else if (removable_params_cost
1765 && !ipa_is_param_used (info, i))
1766 *removable_params_cost
1767 += estimate_move_cost (TREE_TYPE (ipa_get_param (info, i)));
1769 if (known_aggs)
1771 vec<ipa_agg_jf_item_t, va_gc> *agg_items;
1772 struct ipa_agg_jump_function *ajf;
1774 agg_items = context_independent_aggregate_values (plats);
1775 ajf = &(*known_aggs)[i];
1776 ajf->items = agg_items;
1777 ajf->by_ref = plats->aggs_by_ref;
1778 ret |= agg_items != NULL;
1782 return ret;
1785 /* The current interface in ipa-inline-analysis requires a pointer vector.
1786 Create it.
1788 FIXME: That interface should be re-worked, this is slightly silly. Still,
1789 I'd like to discuss how to change it first and this demonstrates the
1790 issue. */
1792 static vec<ipa_agg_jump_function_p>
1793 agg_jmp_p_vec_for_t_vec (vec<ipa_agg_jump_function_t> known_aggs)
1795 vec<ipa_agg_jump_function_p> ret;
1796 struct ipa_agg_jump_function *ajf;
1797 int i;
1799 ret.create (known_aggs.length ());
1800 FOR_EACH_VEC_ELT (known_aggs, i, ajf)
1801 ret.quick_push (ajf);
1802 return ret;
1805 /* Iterate over known values of parameters of NODE and estimate the local
1806 effects in terms of time and size they have. */
1808 static void
1809 estimate_local_effects (struct cgraph_node *node)
1811 struct ipa_node_params *info = IPA_NODE_REF (node);
1812 int i, count = ipa_get_param_count (info);
1813 vec<tree> known_csts, known_binfos;
1814 vec<ipa_agg_jump_function_t> known_aggs;
1815 vec<ipa_agg_jump_function_p> known_aggs_ptrs;
1816 bool always_const;
1817 int base_time = inline_summary (node)->time;
1818 int removable_params_cost;
1820 if (!count || !ipcp_versionable_function_p (node))
1821 return;
1823 if (dump_file && (dump_flags & TDF_DETAILS))
1824 fprintf (dump_file, "\nEstimating effects for %s/%i, base_time: %i.\n",
1825 cgraph_node_name (node), node->symbol.order, base_time);
1827 always_const = gather_context_independent_values (info, &known_csts,
1828 &known_binfos, &known_aggs,
1829 &removable_params_cost);
1830 known_aggs_ptrs = agg_jmp_p_vec_for_t_vec (known_aggs);
1831 if (always_const)
1833 struct caller_statistics stats;
1834 inline_hints hints;
1835 int time, size;
1837 init_caller_stats (&stats);
1838 cgraph_for_node_and_aliases (node, gather_caller_stats, &stats, false);
1839 estimate_ipcp_clone_size_and_time (node, known_csts, known_binfos,
1840 known_aggs_ptrs, &size, &time, &hints);
1841 time -= devirtualization_time_bonus (node, known_csts, known_binfos,
1842 known_aggs_ptrs);
1843 time -= hint_time_bonus (hints);
1844 time -= removable_params_cost;
1845 size -= stats.n_calls * removable_params_cost;
1847 if (dump_file)
1848 fprintf (dump_file, " - context independent values, size: %i, "
1849 "time_benefit: %i\n", size, base_time - time);
1851 if (size <= 0
1852 || cgraph_will_be_removed_from_program_if_no_direct_calls (node))
1854 info->do_clone_for_all_contexts = true;
1855 base_time = time;
1857 if (dump_file)
1858 fprintf (dump_file, " Decided to specialize for all "
1859 "known contexts, code not going to grow.\n");
1861 else if (good_cloning_opportunity_p (node, base_time - time,
1862 stats.freq_sum, stats.count_sum,
1863 size))
1865 if (size + overall_size <= max_new_size)
1867 info->do_clone_for_all_contexts = true;
1868 base_time = time;
1869 overall_size += size;
1871 if (dump_file)
1872 fprintf (dump_file, " Decided to specialize for all "
1873 "known contexts, growth deemed beneficial.\n");
1875 else if (dump_file && (dump_flags & TDF_DETAILS))
1876 fprintf (dump_file, " Not cloning for all contexts because "
1877 "max_new_size would be reached with %li.\n",
1878 size + overall_size);
1882 for (i = 0; i < count ; i++)
1884 struct ipcp_param_lattices *plats = ipa_get_parm_lattices (info, i);
1885 struct ipcp_lattice *lat = &plats->itself;
1886 struct ipcp_value *val;
1887 int emc;
1889 if (lat->bottom
1890 || !lat->values
1891 || known_csts[i]
1892 || known_binfos[i])
1893 continue;
1895 for (val = lat->values; val; val = val->next)
1897 int time, size, time_benefit;
1898 inline_hints hints;
1900 if (TREE_CODE (val->value) != TREE_BINFO)
1902 known_csts[i] = val->value;
1903 known_binfos[i] = NULL_TREE;
1904 emc = estimate_move_cost (TREE_TYPE (val->value));
1906 else if (plats->virt_call)
1908 known_csts[i] = NULL_TREE;
1909 known_binfos[i] = val->value;
1910 emc = 0;
1912 else
1913 continue;
1915 estimate_ipcp_clone_size_and_time (node, known_csts, known_binfos,
1916 known_aggs_ptrs, &size, &time,
1917 &hints);
1918 time_benefit = base_time - time
1919 + devirtualization_time_bonus (node, known_csts, known_binfos,
1920 known_aggs_ptrs)
1921 + hint_time_bonus (hints)
1922 + removable_params_cost + emc;
1924 gcc_checking_assert (size >=0);
1925 /* The inliner-heuristics based estimates may think that in certain
1926 contexts some functions do not have any size at all but we want
1927 all specializations to have at least a tiny cost, not least not to
1928 divide by zero. */
1929 if (size == 0)
1930 size = 1;
1932 if (dump_file && (dump_flags & TDF_DETAILS))
1934 fprintf (dump_file, " - estimates for value ");
1935 print_ipcp_constant_value (dump_file, val->value);
1936 fprintf (dump_file, " for parameter ");
1937 print_generic_expr (dump_file, ipa_get_param (info, i), 0);
1938 fprintf (dump_file, ": time_benefit: %i, size: %i\n",
1939 time_benefit, size);
1942 val->local_time_benefit = time_benefit;
1943 val->local_size_cost = size;
1945 known_binfos[i] = NULL_TREE;
1946 known_csts[i] = NULL_TREE;
1949 for (i = 0; i < count ; i++)
1951 struct ipcp_param_lattices *plats = ipa_get_parm_lattices (info, i);
1952 struct ipa_agg_jump_function *ajf;
1953 struct ipcp_agg_lattice *aglat;
1955 if (plats->aggs_bottom || !plats->aggs)
1956 continue;
1958 ajf = &known_aggs[i];
1959 for (aglat = plats->aggs; aglat; aglat = aglat->next)
1961 struct ipcp_value *val;
1962 if (aglat->bottom || !aglat->values
1963 /* If the following is true, the one value is in known_aggs. */
1964 || (!plats->aggs_contain_variable
1965 && ipa_lat_is_single_const (aglat)))
1966 continue;
1968 for (val = aglat->values; val; val = val->next)
1970 int time, size, time_benefit;
1971 struct ipa_agg_jf_item item;
1972 inline_hints hints;
1974 item.offset = aglat->offset;
1975 item.value = val->value;
1976 vec_safe_push (ajf->items, item);
1978 estimate_ipcp_clone_size_and_time (node, known_csts, known_binfos,
1979 known_aggs_ptrs, &size, &time,
1980 &hints);
1981 time_benefit = base_time - time
1982 + devirtualization_time_bonus (node, known_csts, known_binfos,
1983 known_aggs_ptrs)
1984 + hint_time_bonus (hints);
1985 gcc_checking_assert (size >=0);
1986 if (size == 0)
1987 size = 1;
1989 if (dump_file && (dump_flags & TDF_DETAILS))
1991 fprintf (dump_file, " - estimates for value ");
1992 print_ipcp_constant_value (dump_file, val->value);
1993 fprintf (dump_file, " for parameter ");
1994 print_generic_expr (dump_file, ipa_get_param (info, i), 0);
1995 fprintf (dump_file, "[%soffset: " HOST_WIDE_INT_PRINT_DEC
1996 "]: time_benefit: %i, size: %i\n",
1997 plats->aggs_by_ref ? "ref " : "",
1998 aglat->offset, time_benefit, size);
2001 val->local_time_benefit = time_benefit;
2002 val->local_size_cost = size;
2003 ajf->items->pop ();
2008 for (i = 0; i < count ; i++)
2009 vec_free (known_aggs[i].items);
2011 known_csts.release ();
2012 known_binfos.release ();
2013 known_aggs.release ();
2014 known_aggs_ptrs.release ();
2018 /* Add value CUR_VAL and all yet-unsorted values it is dependent on to the
2019 topological sort of values. */
2021 static void
2022 add_val_to_toposort (struct ipcp_value *cur_val)
2024 static int dfs_counter = 0;
2025 static struct ipcp_value *stack;
2026 struct ipcp_value_source *src;
2028 if (cur_val->dfs)
2029 return;
2031 dfs_counter++;
2032 cur_val->dfs = dfs_counter;
2033 cur_val->low_link = dfs_counter;
2035 cur_val->topo_next = stack;
2036 stack = cur_val;
2037 cur_val->on_stack = true;
2039 for (src = cur_val->sources; src; src = src->next)
2040 if (src->val)
2042 if (src->val->dfs == 0)
2044 add_val_to_toposort (src->val);
2045 if (src->val->low_link < cur_val->low_link)
2046 cur_val->low_link = src->val->low_link;
2048 else if (src->val->on_stack
2049 && src->val->dfs < cur_val->low_link)
2050 cur_val->low_link = src->val->dfs;
2053 if (cur_val->dfs == cur_val->low_link)
2055 struct ipcp_value *v, *scc_list = NULL;
2059 v = stack;
2060 stack = v->topo_next;
2061 v->on_stack = false;
2063 v->scc_next = scc_list;
2064 scc_list = v;
2066 while (v != cur_val);
2068 cur_val->topo_next = values_topo;
2069 values_topo = cur_val;
2073 /* Add all values in lattices associated with NODE to the topological sort if
2074 they are not there yet. */
2076 static void
2077 add_all_node_vals_to_toposort (struct cgraph_node *node)
2079 struct ipa_node_params *info = IPA_NODE_REF (node);
2080 int i, count = ipa_get_param_count (info);
2082 for (i = 0; i < count ; i++)
2084 struct ipcp_param_lattices *plats = ipa_get_parm_lattices (info, i);
2085 struct ipcp_lattice *lat = &plats->itself;
2086 struct ipcp_agg_lattice *aglat;
2087 struct ipcp_value *val;
2089 if (!lat->bottom)
2090 for (val = lat->values; val; val = val->next)
2091 add_val_to_toposort (val);
2093 if (!plats->aggs_bottom)
2094 for (aglat = plats->aggs; aglat; aglat = aglat->next)
2095 if (!aglat->bottom)
2096 for (val = aglat->values; val; val = val->next)
2097 add_val_to_toposort (val);
2101 /* One pass of constants propagation along the call graph edges, from callers
2102 to callees (requires topological ordering in TOPO), iterate over strongly
2103 connected components. */
2105 static void
2106 propagate_constants_topo (struct topo_info *topo)
2108 int i;
2110 for (i = topo->nnodes - 1; i >= 0; i--)
2112 struct cgraph_node *v, *node = topo->order[i];
2113 struct ipa_dfs_info *node_dfs_info;
2115 if (!cgraph_function_with_gimple_body_p (node))
2116 continue;
2118 node_dfs_info = (struct ipa_dfs_info *) node->symbol.aux;
2119 /* First, iteratively propagate within the strongly connected component
2120 until all lattices stabilize. */
2121 v = node_dfs_info->next_cycle;
2122 while (v)
2124 push_node_to_stack (topo, v);
2125 v = ((struct ipa_dfs_info *) v->symbol.aux)->next_cycle;
2128 v = node;
2129 while (v)
2131 struct cgraph_edge *cs;
2133 for (cs = v->callees; cs; cs = cs->next_callee)
2134 if (edge_within_scc (cs)
2135 && propagate_constants_accross_call (cs))
2136 push_node_to_stack (topo, cs->callee);
2137 v = pop_node_from_stack (topo);
2140 /* Afterwards, propagate along edges leading out of the SCC, calculates
2141 the local effects of the discovered constants and all valid values to
2142 their topological sort. */
2143 v = node;
2144 while (v)
2146 struct cgraph_edge *cs;
2148 estimate_local_effects (v);
2149 add_all_node_vals_to_toposort (v);
2150 for (cs = v->callees; cs; cs = cs->next_callee)
2151 if (!edge_within_scc (cs))
2152 propagate_constants_accross_call (cs);
2154 v = ((struct ipa_dfs_info *) v->symbol.aux)->next_cycle;
2160 /* Return the sum of A and B if none of them is bigger than INT_MAX/2, return
2161 the bigger one if otherwise. */
2163 static int
2164 safe_add (int a, int b)
2166 if (a > INT_MAX/2 || b > INT_MAX/2)
2167 return a > b ? a : b;
2168 else
2169 return a + b;
2173 /* Propagate the estimated effects of individual values along the topological
2174 from the dependent values to those they depend on. */
2176 static void
2177 propagate_effects (void)
2179 struct ipcp_value *base;
2181 for (base = values_topo; base; base = base->topo_next)
2183 struct ipcp_value_source *src;
2184 struct ipcp_value *val;
2185 int time = 0, size = 0;
2187 for (val = base; val; val = val->scc_next)
2189 time = safe_add (time,
2190 val->local_time_benefit + val->prop_time_benefit);
2191 size = safe_add (size, val->local_size_cost + val->prop_size_cost);
2194 for (val = base; val; val = val->scc_next)
2195 for (src = val->sources; src; src = src->next)
2196 if (src->val
2197 && cgraph_maybe_hot_edge_p (src->cs))
2199 src->val->prop_time_benefit = safe_add (time,
2200 src->val->prop_time_benefit);
2201 src->val->prop_size_cost = safe_add (size,
2202 src->val->prop_size_cost);
2208 /* Propagate constants, binfos and their effects from the summaries
2209 interprocedurally. */
2211 static void
2212 ipcp_propagate_stage (struct topo_info *topo)
2214 struct cgraph_node *node;
2216 if (dump_file)
2217 fprintf (dump_file, "\n Propagating constants:\n\n");
2219 if (in_lto_p)
2220 ipa_update_after_lto_read ();
2223 FOR_EACH_DEFINED_FUNCTION (node)
2225 struct ipa_node_params *info = IPA_NODE_REF (node);
2227 determine_versionability (node);
2228 if (cgraph_function_with_gimple_body_p (node))
2230 info->lattices = XCNEWVEC (struct ipcp_param_lattices,
2231 ipa_get_param_count (info));
2232 initialize_node_lattices (node);
2234 if (node->count > max_count)
2235 max_count = node->count;
2236 overall_size += inline_summary (node)->self_size;
2239 max_new_size = overall_size;
2240 if (max_new_size < PARAM_VALUE (PARAM_LARGE_UNIT_INSNS))
2241 max_new_size = PARAM_VALUE (PARAM_LARGE_UNIT_INSNS);
2242 max_new_size += max_new_size * PARAM_VALUE (PARAM_IPCP_UNIT_GROWTH) / 100 + 1;
2244 if (dump_file)
2245 fprintf (dump_file, "\noverall_size: %li, max_new_size: %li\n",
2246 overall_size, max_new_size);
2248 propagate_constants_topo (topo);
2249 #ifdef ENABLE_CHECKING
2250 ipcp_verify_propagated_values ();
2251 #endif
2252 propagate_effects ();
2254 if (dump_file)
2256 fprintf (dump_file, "\nIPA lattices after all propagation:\n");
2257 print_all_lattices (dump_file, (dump_flags & TDF_DETAILS), true);
2261 /* Discover newly direct outgoing edges from NODE which is a new clone with
2262 known KNOWN_VALS and make them direct. */
2264 static void
2265 ipcp_discover_new_direct_edges (struct cgraph_node *node,
2266 vec<tree> known_vals,
2267 struct ipa_agg_replacement_value *aggvals)
2269 struct cgraph_edge *ie, *next_ie;
2270 bool found = false;
2272 for (ie = node->indirect_calls; ie; ie = next_ie)
2274 tree target;
2276 next_ie = ie->next_callee;
2277 target = ipa_get_indirect_edge_target_1 (ie, known_vals, vNULL, vNULL,
2278 aggvals);
2279 if (target)
2281 struct cgraph_edge *cs = ipa_make_edge_direct_to_target (ie, target);
2282 found = true;
2284 if (cs && !ie->indirect_info->agg_contents
2285 && !ie->indirect_info->polymorphic)
2287 struct ipa_node_params *info = IPA_NODE_REF (node);
2288 int param_index = ie->indirect_info->param_index;
2289 int c = ipa_get_controlled_uses (info, param_index);
2290 if (c != IPA_UNDESCRIBED_USE)
2292 struct ipa_ref *to_del;
2294 c--;
2295 ipa_set_controlled_uses (info, param_index, c);
2296 if (dump_file && (dump_flags & TDF_DETAILS))
2297 fprintf (dump_file, " controlled uses count of param "
2298 "%i bumped down to %i\n", param_index, c);
2299 if (c == 0
2300 && (to_del = ipa_find_reference ((symtab_node) node,
2301 (symtab_node) cs->callee,
2302 NULL)))
2304 if (dump_file && (dump_flags & TDF_DETAILS))
2305 fprintf (dump_file, " and even removing its "
2306 "cloning-created reference\n");
2307 ipa_remove_reference (to_del);
2313 /* Turning calls to direct calls will improve overall summary. */
2314 if (found)
2315 inline_update_overall_summary (node);
2318 /* Vector of pointers which for linked lists of clones of an original crgaph
2319 edge. */
2321 static vec<cgraph_edge_p> next_edge_clone;
2323 static inline void
2324 grow_next_edge_clone_vector (void)
2326 if (next_edge_clone.length ()
2327 <= (unsigned) cgraph_edge_max_uid)
2328 next_edge_clone.safe_grow_cleared (cgraph_edge_max_uid + 1);
2331 /* Edge duplication hook to grow the appropriate linked list in
2332 next_edge_clone. */
2334 static void
2335 ipcp_edge_duplication_hook (struct cgraph_edge *src, struct cgraph_edge *dst,
2336 __attribute__((unused)) void *data)
2338 grow_next_edge_clone_vector ();
2339 next_edge_clone[dst->uid] = next_edge_clone[src->uid];
2340 next_edge_clone[src->uid] = dst;
2343 /* See if NODE is a clone with a known aggregate value at a given OFFSET of a
2344 parameter with the given INDEX. */
2346 static tree
2347 get_clone_agg_value (struct cgraph_node *node, HOST_WIDEST_INT offset,
2348 int index)
2350 struct ipa_agg_replacement_value *aggval;
2352 aggval = ipa_get_agg_replacements_for_node (node);
2353 while (aggval)
2355 if (aggval->offset == offset
2356 && aggval->index == index)
2357 return aggval->value;
2358 aggval = aggval->next;
2360 return NULL_TREE;
2363 /* Return true if edge CS does bring about the value described by SRC. */
2365 static bool
2366 cgraph_edge_brings_value_p (struct cgraph_edge *cs,
2367 struct ipcp_value_source *src)
2369 struct ipa_node_params *caller_info = IPA_NODE_REF (cs->caller);
2370 struct ipa_node_params *dst_info = IPA_NODE_REF (cs->callee);
2372 if ((dst_info->ipcp_orig_node && !dst_info->is_all_contexts_clone)
2373 || caller_info->node_dead)
2374 return false;
2375 if (!src->val)
2376 return true;
2378 if (caller_info->ipcp_orig_node)
2380 tree t;
2381 if (src->offset == -1)
2382 t = caller_info->known_vals[src->index];
2383 else
2384 t = get_clone_agg_value (cs->caller, src->offset, src->index);
2385 return (t != NULL_TREE
2386 && values_equal_for_ipcp_p (src->val->value, t));
2388 else
2390 struct ipcp_agg_lattice *aglat;
2391 struct ipcp_param_lattices *plats = ipa_get_parm_lattices (caller_info,
2392 src->index);
2393 if (src->offset == -1)
2394 return (ipa_lat_is_single_const (&plats->itself)
2395 && values_equal_for_ipcp_p (src->val->value,
2396 plats->itself.values->value));
2397 else
2399 if (plats->aggs_bottom || plats->aggs_contain_variable)
2400 return false;
2401 for (aglat = plats->aggs; aglat; aglat = aglat->next)
2402 if (aglat->offset == src->offset)
2403 return (ipa_lat_is_single_const (aglat)
2404 && values_equal_for_ipcp_p (src->val->value,
2405 aglat->values->value));
2407 return false;
2411 /* Get the next clone in the linked list of clones of an edge. */
2413 static inline struct cgraph_edge *
2414 get_next_cgraph_edge_clone (struct cgraph_edge *cs)
2416 return next_edge_clone[cs->uid];
2419 /* Given VAL, iterate over all its sources and if they still hold, add their
2420 edge frequency and their number into *FREQUENCY and *CALLER_COUNT
2421 respectively. */
2423 static bool
2424 get_info_about_necessary_edges (struct ipcp_value *val, int *freq_sum,
2425 gcov_type *count_sum, int *caller_count)
2427 struct ipcp_value_source *src;
2428 int freq = 0, count = 0;
2429 gcov_type cnt = 0;
2430 bool hot = false;
2432 for (src = val->sources; src; src = src->next)
2434 struct cgraph_edge *cs = src->cs;
2435 while (cs)
2437 if (cgraph_edge_brings_value_p (cs, src))
2439 count++;
2440 freq += cs->frequency;
2441 cnt += cs->count;
2442 hot |= cgraph_maybe_hot_edge_p (cs);
2444 cs = get_next_cgraph_edge_clone (cs);
2448 *freq_sum = freq;
2449 *count_sum = cnt;
2450 *caller_count = count;
2451 return hot;
2454 /* Return a vector of incoming edges that do bring value VAL. It is assumed
2455 their number is known and equal to CALLER_COUNT. */
2457 static vec<cgraph_edge_p>
2458 gather_edges_for_value (struct ipcp_value *val, int caller_count)
2460 struct ipcp_value_source *src;
2461 vec<cgraph_edge_p> ret;
2463 ret.create (caller_count);
2464 for (src = val->sources; src; src = src->next)
2466 struct cgraph_edge *cs = src->cs;
2467 while (cs)
2469 if (cgraph_edge_brings_value_p (cs, src))
2470 ret.quick_push (cs);
2471 cs = get_next_cgraph_edge_clone (cs);
2475 return ret;
2478 /* Construct a replacement map for a know VALUE for a formal parameter PARAM.
2479 Return it or NULL if for some reason it cannot be created. */
2481 static struct ipa_replace_map *
2482 get_replacement_map (tree value, tree parm)
2484 tree req_type = TREE_TYPE (parm);
2485 struct ipa_replace_map *replace_map;
2487 if (!useless_type_conversion_p (req_type, TREE_TYPE (value)))
2489 if (fold_convertible_p (req_type, value))
2490 value = fold_build1 (NOP_EXPR, req_type, value);
2491 else if (TYPE_SIZE (req_type) == TYPE_SIZE (TREE_TYPE (value)))
2492 value = fold_build1 (VIEW_CONVERT_EXPR, req_type, value);
2493 else
2495 if (dump_file)
2497 fprintf (dump_file, " const ");
2498 print_generic_expr (dump_file, value, 0);
2499 fprintf (dump_file, " can't be converted to param ");
2500 print_generic_expr (dump_file, parm, 0);
2501 fprintf (dump_file, "\n");
2503 return NULL;
2507 replace_map = ggc_alloc_ipa_replace_map ();
2508 if (dump_file)
2510 fprintf (dump_file, " replacing param ");
2511 print_generic_expr (dump_file, parm, 0);
2512 fprintf (dump_file, " with const ");
2513 print_generic_expr (dump_file, value, 0);
2514 fprintf (dump_file, "\n");
2516 replace_map->old_tree = parm;
2517 replace_map->new_tree = value;
2518 replace_map->replace_p = true;
2519 replace_map->ref_p = false;
2521 return replace_map;
2524 /* Dump new profiling counts */
2526 static void
2527 dump_profile_updates (struct cgraph_node *orig_node,
2528 struct cgraph_node *new_node)
2530 struct cgraph_edge *cs;
2532 fprintf (dump_file, " setting count of the specialized node to "
2533 HOST_WIDE_INT_PRINT_DEC "\n", (HOST_WIDE_INT) new_node->count);
2534 for (cs = new_node->callees; cs ; cs = cs->next_callee)
2535 fprintf (dump_file, " edge to %s has count "
2536 HOST_WIDE_INT_PRINT_DEC "\n",
2537 cgraph_node_name (cs->callee), (HOST_WIDE_INT) cs->count);
2539 fprintf (dump_file, " setting count of the original node to "
2540 HOST_WIDE_INT_PRINT_DEC "\n", (HOST_WIDE_INT) orig_node->count);
2541 for (cs = orig_node->callees; cs ; cs = cs->next_callee)
2542 fprintf (dump_file, " edge to %s is left with "
2543 HOST_WIDE_INT_PRINT_DEC "\n",
2544 cgraph_node_name (cs->callee), (HOST_WIDE_INT) cs->count);
2547 /* After a specialized NEW_NODE version of ORIG_NODE has been created, update
2548 their profile information to reflect this. */
2550 static void
2551 update_profiling_info (struct cgraph_node *orig_node,
2552 struct cgraph_node *new_node)
2554 struct cgraph_edge *cs;
2555 struct caller_statistics stats;
2556 gcov_type new_sum, orig_sum;
2557 gcov_type remainder, orig_node_count = orig_node->count;
2559 if (orig_node_count == 0)
2560 return;
2562 init_caller_stats (&stats);
2563 cgraph_for_node_and_aliases (orig_node, gather_caller_stats, &stats, false);
2564 orig_sum = stats.count_sum;
2565 init_caller_stats (&stats);
2566 cgraph_for_node_and_aliases (new_node, gather_caller_stats, &stats, false);
2567 new_sum = stats.count_sum;
2569 if (orig_node_count < orig_sum + new_sum)
2571 if (dump_file)
2572 fprintf (dump_file, " Problem: node %s/%i has too low count "
2573 HOST_WIDE_INT_PRINT_DEC " while the sum of incoming "
2574 "counts is " HOST_WIDE_INT_PRINT_DEC "\n",
2575 cgraph_node_name (orig_node), orig_node->symbol.order,
2576 (HOST_WIDE_INT) orig_node_count,
2577 (HOST_WIDE_INT) (orig_sum + new_sum));
2579 orig_node_count = (orig_sum + new_sum) * 12 / 10;
2580 if (dump_file)
2581 fprintf (dump_file, " proceeding by pretending it was "
2582 HOST_WIDE_INT_PRINT_DEC "\n",
2583 (HOST_WIDE_INT) orig_node_count);
2586 new_node->count = new_sum;
2587 remainder = orig_node_count - new_sum;
2588 orig_node->count = remainder;
2590 for (cs = new_node->callees; cs ; cs = cs->next_callee)
2591 if (cs->frequency)
2592 cs->count = apply_probability (cs->count,
2593 GCOV_COMPUTE_SCALE (new_sum,
2594 orig_node_count));
2595 else
2596 cs->count = 0;
2598 for (cs = orig_node->callees; cs ; cs = cs->next_callee)
2599 cs->count = apply_probability (cs->count,
2600 GCOV_COMPUTE_SCALE (remainder,
2601 orig_node_count));
2603 if (dump_file)
2604 dump_profile_updates (orig_node, new_node);
2607 /* Update the respective profile of specialized NEW_NODE and the original
2608 ORIG_NODE after additional edges with cumulative count sum REDIRECTED_SUM
2609 have been redirected to the specialized version. */
2611 static void
2612 update_specialized_profile (struct cgraph_node *new_node,
2613 struct cgraph_node *orig_node,
2614 gcov_type redirected_sum)
2616 struct cgraph_edge *cs;
2617 gcov_type new_node_count, orig_node_count = orig_node->count;
2619 if (dump_file)
2620 fprintf (dump_file, " the sum of counts of redirected edges is "
2621 HOST_WIDE_INT_PRINT_DEC "\n", (HOST_WIDE_INT) redirected_sum);
2622 if (orig_node_count == 0)
2623 return;
2625 gcc_assert (orig_node_count >= redirected_sum);
2627 new_node_count = new_node->count;
2628 new_node->count += redirected_sum;
2629 orig_node->count -= redirected_sum;
2631 for (cs = new_node->callees; cs ; cs = cs->next_callee)
2632 if (cs->frequency)
2633 cs->count += apply_probability (cs->count,
2634 GCOV_COMPUTE_SCALE (redirected_sum,
2635 new_node_count));
2636 else
2637 cs->count = 0;
2639 for (cs = orig_node->callees; cs ; cs = cs->next_callee)
2641 gcov_type dec = apply_probability (cs->count,
2642 GCOV_COMPUTE_SCALE (redirected_sum,
2643 orig_node_count));
2644 if (dec < cs->count)
2645 cs->count -= dec;
2646 else
2647 cs->count = 0;
2650 if (dump_file)
2651 dump_profile_updates (orig_node, new_node);
2654 /* Create a specialized version of NODE with known constants and types of
2655 parameters in KNOWN_VALS and redirect all edges in CALLERS to it. */
2657 static struct cgraph_node *
2658 create_specialized_node (struct cgraph_node *node,
2659 vec<tree> known_vals,
2660 struct ipa_agg_replacement_value *aggvals,
2661 vec<cgraph_edge_p> callers)
2663 struct ipa_node_params *new_info, *info = IPA_NODE_REF (node);
2664 vec<ipa_replace_map_p, va_gc> *replace_trees = NULL;
2665 struct cgraph_node *new_node;
2666 int i, count = ipa_get_param_count (info);
2667 bitmap args_to_skip;
2669 gcc_assert (!info->ipcp_orig_node);
2671 if (node->local.can_change_signature)
2673 args_to_skip = BITMAP_GGC_ALLOC ();
2674 for (i = 0; i < count; i++)
2676 tree t = known_vals[i];
2678 if ((t && TREE_CODE (t) != TREE_BINFO)
2679 || !ipa_is_param_used (info, i))
2680 bitmap_set_bit (args_to_skip, i);
2683 else
2685 args_to_skip = NULL;
2686 if (dump_file && (dump_flags & TDF_DETAILS))
2687 fprintf (dump_file, " cannot change function signature\n");
2690 for (i = 0; i < count ; i++)
2692 tree t = known_vals[i];
2693 if (t && TREE_CODE (t) != TREE_BINFO)
2695 struct ipa_replace_map *replace_map;
2697 replace_map = get_replacement_map (t, ipa_get_param (info, i));
2698 if (replace_map)
2699 vec_safe_push (replace_trees, replace_map);
2703 new_node = cgraph_create_virtual_clone (node, callers, replace_trees,
2704 args_to_skip, "constprop");
2705 ipa_set_node_agg_value_chain (new_node, aggvals);
2706 if (dump_file && (dump_flags & TDF_DETAILS))
2708 fprintf (dump_file, " the new node is %s/%i.\n",
2709 cgraph_node_name (new_node), new_node->symbol.order);
2710 if (aggvals)
2711 ipa_dump_agg_replacement_values (dump_file, aggvals);
2713 gcc_checking_assert (ipa_node_params_vector.exists ()
2714 && (ipa_node_params_vector.length ()
2715 > (unsigned) cgraph_max_uid));
2716 update_profiling_info (node, new_node);
2717 new_info = IPA_NODE_REF (new_node);
2718 new_info->ipcp_orig_node = node;
2719 new_info->known_vals = known_vals;
2721 ipcp_discover_new_direct_edges (new_node, known_vals, aggvals);
2723 callers.release ();
2724 return new_node;
2727 /* Given a NODE, and a subset of its CALLERS, try to populate blanks slots in
2728 KNOWN_VALS with constants and types that are also known for all of the
2729 CALLERS. */
2731 static void
2732 find_more_scalar_values_for_callers_subset (struct cgraph_node *node,
2733 vec<tree> known_vals,
2734 vec<cgraph_edge_p> callers)
2736 struct ipa_node_params *info = IPA_NODE_REF (node);
2737 int i, count = ipa_get_param_count (info);
2739 for (i = 0; i < count ; i++)
2741 struct cgraph_edge *cs;
2742 tree newval = NULL_TREE;
2743 int j;
2745 if (ipa_get_scalar_lat (info, i)->bottom || known_vals[i])
2746 continue;
2748 FOR_EACH_VEC_ELT (callers, j, cs)
2750 struct ipa_jump_func *jump_func;
2751 tree t;
2753 if (i >= ipa_get_cs_argument_count (IPA_EDGE_REF (cs)))
2755 newval = NULL_TREE;
2756 break;
2758 jump_func = ipa_get_ith_jump_func (IPA_EDGE_REF (cs), i);
2759 t = ipa_value_from_jfunc (IPA_NODE_REF (cs->caller), jump_func);
2760 if (!t
2761 || (newval
2762 && !values_equal_for_ipcp_p (t, newval)))
2764 newval = NULL_TREE;
2765 break;
2767 else
2768 newval = t;
2771 if (newval)
2773 if (dump_file && (dump_flags & TDF_DETAILS))
2775 fprintf (dump_file, " adding an extra known scalar value ");
2776 print_ipcp_constant_value (dump_file, newval);
2777 fprintf (dump_file, " for parameter ");
2778 print_generic_expr (dump_file, ipa_get_param (info, i), 0);
2779 fprintf (dump_file, "\n");
2782 known_vals[i] = newval;
2787 /* Go through PLATS and create a vector of values consisting of values and
2788 offsets (minus OFFSET) of lattices that contain only a single value. */
2790 static vec<ipa_agg_jf_item_t>
2791 copy_plats_to_inter (struct ipcp_param_lattices *plats, HOST_WIDE_INT offset)
2793 vec<ipa_agg_jf_item_t> res = vNULL;
2795 if (!plats->aggs || plats->aggs_contain_variable || plats->aggs_bottom)
2796 return vNULL;
2798 for (struct ipcp_agg_lattice *aglat = plats->aggs; aglat; aglat = aglat->next)
2799 if (ipa_lat_is_single_const (aglat))
2801 struct ipa_agg_jf_item ti;
2802 ti.offset = aglat->offset - offset;
2803 ti.value = aglat->values->value;
2804 res.safe_push (ti);
2806 return res;
2809 /* Intersect all values in INTER with single value lattices in PLATS (while
2810 subtracting OFFSET). */
2812 static void
2813 intersect_with_plats (struct ipcp_param_lattices *plats,
2814 vec<ipa_agg_jf_item_t> *inter,
2815 HOST_WIDE_INT offset)
2817 struct ipcp_agg_lattice *aglat;
2818 struct ipa_agg_jf_item *item;
2819 int k;
2821 if (!plats->aggs || plats->aggs_contain_variable || plats->aggs_bottom)
2823 inter->release ();
2824 return;
2827 aglat = plats->aggs;
2828 FOR_EACH_VEC_ELT (*inter, k, item)
2830 bool found = false;
2831 if (!item->value)
2832 continue;
2833 while (aglat)
2835 if (aglat->offset - offset > item->offset)
2836 break;
2837 if (aglat->offset - offset == item->offset)
2839 gcc_checking_assert (item->value);
2840 if (values_equal_for_ipcp_p (item->value, aglat->values->value))
2841 found = true;
2842 break;
2844 aglat = aglat->next;
2846 if (!found)
2847 item->value = NULL_TREE;
2851 /* Copy agggregate replacement values of NODE (which is an IPA-CP clone) to the
2852 vector result while subtracting OFFSET from the individual value offsets. */
2854 static vec<ipa_agg_jf_item_t>
2855 agg_replacements_to_vector (struct cgraph_node *node, int index,
2856 HOST_WIDE_INT offset)
2858 struct ipa_agg_replacement_value *av;
2859 vec<ipa_agg_jf_item_t> res = vNULL;
2861 for (av = ipa_get_agg_replacements_for_node (node); av; av = av->next)
2862 if (av->index == index
2863 && (av->offset - offset) >= 0)
2865 struct ipa_agg_jf_item item;
2866 gcc_checking_assert (av->value);
2867 item.offset = av->offset - offset;
2868 item.value = av->value;
2869 res.safe_push (item);
2872 return res;
2875 /* Intersect all values in INTER with those that we have already scheduled to
2876 be replaced in parameter number INDEX of NODE, which is an IPA-CP clone
2877 (while subtracting OFFSET). */
2879 static void
2880 intersect_with_agg_replacements (struct cgraph_node *node, int index,
2881 vec<ipa_agg_jf_item_t> *inter,
2882 HOST_WIDE_INT offset)
2884 struct ipa_agg_replacement_value *srcvals;
2885 struct ipa_agg_jf_item *item;
2886 int i;
2888 srcvals = ipa_get_agg_replacements_for_node (node);
2889 if (!srcvals)
2891 inter->release ();
2892 return;
2895 FOR_EACH_VEC_ELT (*inter, i, item)
2897 struct ipa_agg_replacement_value *av;
2898 bool found = false;
2899 if (!item->value)
2900 continue;
2901 for (av = srcvals; av; av = av->next)
2903 gcc_checking_assert (av->value);
2904 if (av->index == index
2905 && av->offset - offset == item->offset)
2907 if (values_equal_for_ipcp_p (item->value, av->value))
2908 found = true;
2909 break;
2912 if (!found)
2913 item->value = NULL_TREE;
2917 /* Intersect values in INTER with aggregate values that come along edge CS to
2918 parameter number INDEX and return it. If INTER does not actually exist yet,
2919 copy all incoming values to it. If we determine we ended up with no values
2920 whatsoever, return a released vector. */
2922 static vec<ipa_agg_jf_item_t>
2923 intersect_aggregates_with_edge (struct cgraph_edge *cs, int index,
2924 vec<ipa_agg_jf_item_t> inter)
2926 struct ipa_jump_func *jfunc;
2927 jfunc = ipa_get_ith_jump_func (IPA_EDGE_REF (cs), index);
2928 if (jfunc->type == IPA_JF_PASS_THROUGH
2929 && ipa_get_jf_pass_through_operation (jfunc) == NOP_EXPR)
2931 struct ipa_node_params *caller_info = IPA_NODE_REF (cs->caller);
2932 int src_idx = ipa_get_jf_pass_through_formal_id (jfunc);
2934 if (caller_info->ipcp_orig_node)
2936 struct cgraph_node *orig_node = caller_info->ipcp_orig_node;
2937 struct ipcp_param_lattices *orig_plats;
2938 orig_plats = ipa_get_parm_lattices (IPA_NODE_REF (orig_node),
2939 src_idx);
2940 if (agg_pass_through_permissible_p (orig_plats, jfunc))
2942 if (!inter.exists ())
2943 inter = agg_replacements_to_vector (cs->caller, src_idx, 0);
2944 else
2945 intersect_with_agg_replacements (cs->caller, src_idx,
2946 &inter, 0);
2949 else
2951 struct ipcp_param_lattices *src_plats;
2952 src_plats = ipa_get_parm_lattices (caller_info, src_idx);
2953 if (agg_pass_through_permissible_p (src_plats, jfunc))
2955 /* Currently we do not produce clobber aggregate jump
2956 functions, adjust when we do. */
2957 gcc_checking_assert (!jfunc->agg.items);
2958 if (!inter.exists ())
2959 inter = copy_plats_to_inter (src_plats, 0);
2960 else
2961 intersect_with_plats (src_plats, &inter, 0);
2965 else if (jfunc->type == IPA_JF_ANCESTOR
2966 && ipa_get_jf_ancestor_agg_preserved (jfunc))
2968 struct ipa_node_params *caller_info = IPA_NODE_REF (cs->caller);
2969 int src_idx = ipa_get_jf_ancestor_formal_id (jfunc);
2970 struct ipcp_param_lattices *src_plats;
2971 HOST_WIDE_INT delta = ipa_get_jf_ancestor_offset (jfunc);
2973 if (caller_info->ipcp_orig_node)
2975 if (!inter.exists ())
2976 inter = agg_replacements_to_vector (cs->caller, src_idx, delta);
2977 else
2978 intersect_with_agg_replacements (cs->caller, src_idx, &inter,
2979 delta);
2981 else
2983 src_plats = ipa_get_parm_lattices (caller_info, src_idx);;
2984 /* Currently we do not produce clobber aggregate jump
2985 functions, adjust when we do. */
2986 gcc_checking_assert (!src_plats->aggs || !jfunc->agg.items);
2987 if (!inter.exists ())
2988 inter = copy_plats_to_inter (src_plats, delta);
2989 else
2990 intersect_with_plats (src_plats, &inter, delta);
2993 else if (jfunc->agg.items)
2995 struct ipa_agg_jf_item *item;
2996 int k;
2998 if (!inter.exists ())
2999 for (unsigned i = 0; i < jfunc->agg.items->length (); i++)
3000 inter.safe_push ((*jfunc->agg.items)[i]);
3001 else
3002 FOR_EACH_VEC_ELT (inter, k, item)
3004 int l = 0;
3005 bool found = false;;
3007 if (!item->value)
3008 continue;
3010 while ((unsigned) l < jfunc->agg.items->length ())
3012 struct ipa_agg_jf_item *ti;
3013 ti = &(*jfunc->agg.items)[l];
3014 if (ti->offset > item->offset)
3015 break;
3016 if (ti->offset == item->offset)
3018 gcc_checking_assert (ti->value);
3019 if (values_equal_for_ipcp_p (item->value,
3020 ti->value))
3021 found = true;
3022 break;
3024 l++;
3026 if (!found)
3027 item->value = NULL;
3030 else
3032 inter.release();
3033 return vec<ipa_agg_jf_item_t>();
3035 return inter;
3038 /* Look at edges in CALLERS and collect all known aggregate values that arrive
3039 from all of them. */
3041 static struct ipa_agg_replacement_value *
3042 find_aggregate_values_for_callers_subset (struct cgraph_node *node,
3043 vec<cgraph_edge_p> callers)
3045 struct ipa_node_params *dest_info = IPA_NODE_REF (node);
3046 struct ipa_agg_replacement_value *res = NULL;
3047 struct cgraph_edge *cs;
3048 int i, j, count = ipa_get_param_count (dest_info);
3050 FOR_EACH_VEC_ELT (callers, j, cs)
3052 int c = ipa_get_cs_argument_count (IPA_EDGE_REF (cs));
3053 if (c < count)
3054 count = c;
3057 for (i = 0; i < count ; i++)
3059 struct cgraph_edge *cs;
3060 vec<ipa_agg_jf_item_t> inter = vNULL;
3061 struct ipa_agg_jf_item *item;
3062 struct ipcp_param_lattices *plats = ipa_get_parm_lattices (dest_info, i);
3063 int j;
3065 /* Among other things, the following check should deal with all by_ref
3066 mismatches. */
3067 if (plats->aggs_bottom)
3068 continue;
3070 FOR_EACH_VEC_ELT (callers, j, cs)
3072 inter = intersect_aggregates_with_edge (cs, i, inter);
3074 if (!inter.exists ())
3075 goto next_param;
3078 FOR_EACH_VEC_ELT (inter, j, item)
3080 struct ipa_agg_replacement_value *v;
3082 if (!item->value)
3083 continue;
3085 v = ggc_alloc_ipa_agg_replacement_value ();
3086 v->index = i;
3087 v->offset = item->offset;
3088 v->value = item->value;
3089 v->by_ref = plats->aggs_by_ref;
3090 v->next = res;
3091 res = v;
3094 next_param:
3095 if (inter.exists ())
3096 inter.release ();
3098 return res;
3101 /* Turn KNOWN_AGGS into a list of aggreate replacement values. */
3103 static struct ipa_agg_replacement_value *
3104 known_aggs_to_agg_replacement_list (vec<ipa_agg_jump_function_t> known_aggs)
3106 struct ipa_agg_replacement_value *res = NULL;
3107 struct ipa_agg_jump_function *aggjf;
3108 struct ipa_agg_jf_item *item;
3109 int i, j;
3111 FOR_EACH_VEC_ELT (known_aggs, i, aggjf)
3112 FOR_EACH_VEC_SAFE_ELT (aggjf->items, j, item)
3114 struct ipa_agg_replacement_value *v;
3115 v = ggc_alloc_ipa_agg_replacement_value ();
3116 v->index = i;
3117 v->offset = item->offset;
3118 v->value = item->value;
3119 v->by_ref = aggjf->by_ref;
3120 v->next = res;
3121 res = v;
3123 return res;
3126 /* Determine whether CS also brings all scalar values that the NODE is
3127 specialized for. */
3129 static bool
3130 cgraph_edge_brings_all_scalars_for_node (struct cgraph_edge *cs,
3131 struct cgraph_node *node)
3133 struct ipa_node_params *dest_info = IPA_NODE_REF (node);
3134 int count = ipa_get_param_count (dest_info);
3135 struct ipa_node_params *caller_info;
3136 struct ipa_edge_args *args;
3137 int i;
3139 caller_info = IPA_NODE_REF (cs->caller);
3140 args = IPA_EDGE_REF (cs);
3141 for (i = 0; i < count; i++)
3143 struct ipa_jump_func *jump_func;
3144 tree val, t;
3146 val = dest_info->known_vals[i];
3147 if (!val)
3148 continue;
3150 if (i >= ipa_get_cs_argument_count (args))
3151 return false;
3152 jump_func = ipa_get_ith_jump_func (args, i);
3153 t = ipa_value_from_jfunc (caller_info, jump_func);
3154 if (!t || !values_equal_for_ipcp_p (val, t))
3155 return false;
3157 return true;
3160 /* Determine whether CS also brings all aggregate values that NODE is
3161 specialized for. */
3162 static bool
3163 cgraph_edge_brings_all_agg_vals_for_node (struct cgraph_edge *cs,
3164 struct cgraph_node *node)
3166 struct ipa_node_params *orig_caller_info = IPA_NODE_REF (cs->caller);
3167 struct ipa_agg_replacement_value *aggval;
3168 int i, ec, count;
3170 aggval = ipa_get_agg_replacements_for_node (node);
3171 if (!aggval)
3172 return true;
3174 count = ipa_get_param_count (IPA_NODE_REF (node));
3175 ec = ipa_get_cs_argument_count (IPA_EDGE_REF (cs));
3176 if (ec < count)
3177 for (struct ipa_agg_replacement_value *av = aggval; av; av = av->next)
3178 if (aggval->index >= ec)
3179 return false;
3181 if (orig_caller_info->ipcp_orig_node)
3182 orig_caller_info = IPA_NODE_REF (orig_caller_info->ipcp_orig_node);
3184 for (i = 0; i < count; i++)
3186 static vec<ipa_agg_jf_item_t> values = vec<ipa_agg_jf_item_t>();
3187 struct ipcp_param_lattices *plats;
3188 bool interesting = false;
3189 for (struct ipa_agg_replacement_value *av = aggval; av; av = av->next)
3190 if (aggval->index == i)
3192 interesting = true;
3193 break;
3195 if (!interesting)
3196 continue;
3198 plats = ipa_get_parm_lattices (orig_caller_info, aggval->index);
3199 if (plats->aggs_bottom)
3200 return false;
3202 values = intersect_aggregates_with_edge (cs, i, values);
3203 if (!values.exists())
3204 return false;
3206 for (struct ipa_agg_replacement_value *av = aggval; av; av = av->next)
3207 if (aggval->index == i)
3209 struct ipa_agg_jf_item *item;
3210 int j;
3211 bool found = false;
3212 FOR_EACH_VEC_ELT (values, j, item)
3213 if (item->value
3214 && item->offset == av->offset
3215 && values_equal_for_ipcp_p (item->value, av->value))
3216 found = true;
3217 if (!found)
3219 values.release();
3220 return false;
3224 return true;
3227 /* Given an original NODE and a VAL for which we have already created a
3228 specialized clone, look whether there are incoming edges that still lead
3229 into the old node but now also bring the requested value and also conform to
3230 all other criteria such that they can be redirected the the special node.
3231 This function can therefore redirect the final edge in a SCC. */
3233 static void
3234 perhaps_add_new_callers (struct cgraph_node *node, struct ipcp_value *val)
3236 struct ipcp_value_source *src;
3237 gcov_type redirected_sum = 0;
3239 for (src = val->sources; src; src = src->next)
3241 struct cgraph_edge *cs = src->cs;
3242 while (cs)
3244 enum availability availability;
3245 struct cgraph_node *dst = cgraph_function_node (cs->callee,
3246 &availability);
3247 if ((dst == node || IPA_NODE_REF (dst)->is_all_contexts_clone)
3248 && availability > AVAIL_OVERWRITABLE
3249 && cgraph_edge_brings_value_p (cs, src))
3251 if (cgraph_edge_brings_all_scalars_for_node (cs, val->spec_node)
3252 && cgraph_edge_brings_all_agg_vals_for_node (cs,
3253 val->spec_node))
3255 if (dump_file)
3256 fprintf (dump_file, " - adding an extra caller %s/%i"
3257 " of %s/%i\n",
3258 xstrdup (cgraph_node_name (cs->caller)),
3259 cs->caller->symbol.order,
3260 xstrdup (cgraph_node_name (val->spec_node)),
3261 val->spec_node->symbol.order);
3263 cgraph_redirect_edge_callee (cs, val->spec_node);
3264 redirected_sum += cs->count;
3267 cs = get_next_cgraph_edge_clone (cs);
3271 if (redirected_sum)
3272 update_specialized_profile (val->spec_node, node, redirected_sum);
3276 /* Copy KNOWN_BINFOS to KNOWN_VALS. */
3278 static void
3279 move_binfos_to_values (vec<tree> known_vals,
3280 vec<tree> known_binfos)
3282 tree t;
3283 int i;
3285 for (i = 0; known_binfos.iterate (i, &t); i++)
3286 if (t)
3287 known_vals[i] = t;
3290 /* Return true if there is a replacement equivalent to VALUE, INDEX and OFFSET
3291 among those in the AGGVALS list. */
3293 DEBUG_FUNCTION bool
3294 ipcp_val_in_agg_replacements_p (struct ipa_agg_replacement_value *aggvals,
3295 int index, HOST_WIDE_INT offset, tree value)
3297 while (aggvals)
3299 if (aggvals->index == index
3300 && aggvals->offset == offset
3301 && values_equal_for_ipcp_p (aggvals->value, value))
3302 return true;
3303 aggvals = aggvals->next;
3305 return false;
3308 /* Decide wheter to create a special version of NODE for value VAL of parameter
3309 at the given INDEX. If OFFSET is -1, the value is for the parameter itself,
3310 otherwise it is stored at the given OFFSET of the parameter. KNOWN_CSTS,
3311 KNOWN_BINFOS and KNOWN_AGGS describe the other already known values. */
3313 static bool
3314 decide_about_value (struct cgraph_node *node, int index, HOST_WIDE_INT offset,
3315 struct ipcp_value *val, vec<tree> known_csts,
3316 vec<tree> known_binfos)
3318 struct ipa_agg_replacement_value *aggvals;
3319 int freq_sum, caller_count;
3320 gcov_type count_sum;
3321 vec<cgraph_edge_p> callers;
3322 vec<tree> kv;
3324 if (val->spec_node)
3326 perhaps_add_new_callers (node, val);
3327 return false;
3329 else if (val->local_size_cost + overall_size > max_new_size)
3331 if (dump_file && (dump_flags & TDF_DETAILS))
3332 fprintf (dump_file, " Ignoring candidate value because "
3333 "max_new_size would be reached with %li.\n",
3334 val->local_size_cost + overall_size);
3335 return false;
3337 else if (!get_info_about_necessary_edges (val, &freq_sum, &count_sum,
3338 &caller_count))
3339 return false;
3341 if (dump_file && (dump_flags & TDF_DETAILS))
3343 fprintf (dump_file, " - considering value ");
3344 print_ipcp_constant_value (dump_file, val->value);
3345 fprintf (dump_file, " for parameter ");
3346 print_generic_expr (dump_file, ipa_get_param (IPA_NODE_REF (node),
3347 index), 0);
3348 if (offset != -1)
3349 fprintf (dump_file, ", offset: " HOST_WIDE_INT_PRINT_DEC, offset);
3350 fprintf (dump_file, " (caller_count: %i)\n", caller_count);
3353 if (!good_cloning_opportunity_p (node, val->local_time_benefit,
3354 freq_sum, count_sum,
3355 val->local_size_cost)
3356 && !good_cloning_opportunity_p (node,
3357 val->local_time_benefit
3358 + val->prop_time_benefit,
3359 freq_sum, count_sum,
3360 val->local_size_cost
3361 + val->prop_size_cost))
3362 return false;
3364 if (dump_file)
3365 fprintf (dump_file, " Creating a specialized node of %s/%i.\n",
3366 cgraph_node_name (node), node->symbol.order);
3368 callers = gather_edges_for_value (val, caller_count);
3369 kv = known_csts.copy ();
3370 move_binfos_to_values (kv, known_binfos);
3371 if (offset == -1)
3372 kv[index] = val->value;
3373 find_more_scalar_values_for_callers_subset (node, kv, callers);
3374 aggvals = find_aggregate_values_for_callers_subset (node, callers);
3375 gcc_checking_assert (offset == -1
3376 || ipcp_val_in_agg_replacements_p (aggvals, index,
3377 offset, val->value));
3378 val->spec_node = create_specialized_node (node, kv, aggvals, callers);
3379 overall_size += val->local_size_cost;
3381 /* TODO: If for some lattice there is only one other known value
3382 left, make a special node for it too. */
3384 return true;
3387 /* Decide whether and what specialized clones of NODE should be created. */
3389 static bool
3390 decide_whether_version_node (struct cgraph_node *node)
3392 struct ipa_node_params *info = IPA_NODE_REF (node);
3393 int i, count = ipa_get_param_count (info);
3394 vec<tree> known_csts, known_binfos;
3395 vec<ipa_agg_jump_function_t> known_aggs = vNULL;
3396 bool ret = false;
3398 if (count == 0)
3399 return false;
3401 if (dump_file && (dump_flags & TDF_DETAILS))
3402 fprintf (dump_file, "\nEvaluating opportunities for %s/%i.\n",
3403 cgraph_node_name (node), node->symbol.order);
3405 gather_context_independent_values (info, &known_csts, &known_binfos,
3406 info->do_clone_for_all_contexts ? &known_aggs
3407 : NULL, NULL);
3409 for (i = 0; i < count ;i++)
3411 struct ipcp_param_lattices *plats = ipa_get_parm_lattices (info, i);
3412 struct ipcp_lattice *lat = &plats->itself;
3413 struct ipcp_value *val;
3415 if (!lat->bottom
3416 && !known_csts[i]
3417 && !known_binfos[i])
3418 for (val = lat->values; val; val = val->next)
3419 ret |= decide_about_value (node, i, -1, val, known_csts,
3420 known_binfos);
3422 if (!plats->aggs_bottom)
3424 struct ipcp_agg_lattice *aglat;
3425 struct ipcp_value *val;
3426 for (aglat = plats->aggs; aglat; aglat = aglat->next)
3427 if (!aglat->bottom && aglat->values
3428 /* If the following is false, the one value is in
3429 known_aggs. */
3430 && (plats->aggs_contain_variable
3431 || !ipa_lat_is_single_const (aglat)))
3432 for (val = aglat->values; val; val = val->next)
3433 ret |= decide_about_value (node, i, aglat->offset, val,
3434 known_csts, known_binfos);
3436 info = IPA_NODE_REF (node);
3439 if (info->do_clone_for_all_contexts)
3441 struct cgraph_node *clone;
3442 vec<cgraph_edge_p> callers;
3444 if (dump_file)
3445 fprintf (dump_file, " - Creating a specialized node of %s/%i "
3446 "for all known contexts.\n", cgraph_node_name (node),
3447 node->symbol.order);
3449 callers = collect_callers_of_node (node);
3450 move_binfos_to_values (known_csts, known_binfos);
3451 clone = create_specialized_node (node, known_csts,
3452 known_aggs_to_agg_replacement_list (known_aggs),
3453 callers);
3454 info = IPA_NODE_REF (node);
3455 info->do_clone_for_all_contexts = false;
3456 IPA_NODE_REF (clone)->is_all_contexts_clone = true;
3457 for (i = 0; i < count ; i++)
3458 vec_free (known_aggs[i].items);
3459 known_aggs.release ();
3460 ret = true;
3462 else
3463 known_csts.release ();
3465 known_binfos.release ();
3466 return ret;
3469 /* Transitively mark all callees of NODE within the same SCC as not dead. */
3471 static void
3472 spread_undeadness (struct cgraph_node *node)
3474 struct cgraph_edge *cs;
3476 for (cs = node->callees; cs; cs = cs->next_callee)
3477 if (edge_within_scc (cs))
3479 struct cgraph_node *callee;
3480 struct ipa_node_params *info;
3482 callee = cgraph_function_node (cs->callee, NULL);
3483 info = IPA_NODE_REF (callee);
3485 if (info->node_dead)
3487 info->node_dead = 0;
3488 spread_undeadness (callee);
3493 /* Return true if NODE has a caller from outside of its SCC that is not
3494 dead. Worker callback for cgraph_for_node_and_aliases. */
3496 static bool
3497 has_undead_caller_from_outside_scc_p (struct cgraph_node *node,
3498 void *data ATTRIBUTE_UNUSED)
3500 struct cgraph_edge *cs;
3502 for (cs = node->callers; cs; cs = cs->next_caller)
3503 if (cs->caller->thunk.thunk_p
3504 && cgraph_for_node_and_aliases (cs->caller,
3505 has_undead_caller_from_outside_scc_p,
3506 NULL, true))
3507 return true;
3508 else if (!edge_within_scc (cs)
3509 && !IPA_NODE_REF (cs->caller)->node_dead)
3510 return true;
3511 return false;
3515 /* Identify nodes within the same SCC as NODE which are no longer needed
3516 because of new clones and will be removed as unreachable. */
3518 static void
3519 identify_dead_nodes (struct cgraph_node *node)
3521 struct cgraph_node *v;
3522 for (v = node; v ; v = ((struct ipa_dfs_info *) v->symbol.aux)->next_cycle)
3523 if (cgraph_will_be_removed_from_program_if_no_direct_calls (v)
3524 && !cgraph_for_node_and_aliases (v,
3525 has_undead_caller_from_outside_scc_p,
3526 NULL, true))
3527 IPA_NODE_REF (v)->node_dead = 1;
3529 for (v = node; v ; v = ((struct ipa_dfs_info *) v->symbol.aux)->next_cycle)
3530 if (!IPA_NODE_REF (v)->node_dead)
3531 spread_undeadness (v);
3533 if (dump_file && (dump_flags & TDF_DETAILS))
3535 for (v = node; v ; v = ((struct ipa_dfs_info *) v->symbol.aux)->next_cycle)
3536 if (IPA_NODE_REF (v)->node_dead)
3537 fprintf (dump_file, " Marking node as dead: %s/%i.\n",
3538 cgraph_node_name (v), v->symbol.order);
3542 /* The decision stage. Iterate over the topological order of call graph nodes
3543 TOPO and make specialized clones if deemed beneficial. */
3545 static void
3546 ipcp_decision_stage (struct topo_info *topo)
3548 int i;
3550 if (dump_file)
3551 fprintf (dump_file, "\nIPA decision stage:\n\n");
3553 for (i = topo->nnodes - 1; i >= 0; i--)
3555 struct cgraph_node *node = topo->order[i];
3556 bool change = false, iterate = true;
3558 while (iterate)
3560 struct cgraph_node *v;
3561 iterate = false;
3562 for (v = node; v ; v = ((struct ipa_dfs_info *) v->symbol.aux)->next_cycle)
3563 if (cgraph_function_with_gimple_body_p (v)
3564 && ipcp_versionable_function_p (v))
3565 iterate |= decide_whether_version_node (v);
3567 change |= iterate;
3569 if (change)
3570 identify_dead_nodes (node);
3574 /* The IPCP driver. */
3576 static unsigned int
3577 ipcp_driver (void)
3579 struct cgraph_2edge_hook_list *edge_duplication_hook_holder;
3580 struct topo_info topo;
3582 ipa_check_create_node_params ();
3583 ipa_check_create_edge_args ();
3584 grow_next_edge_clone_vector ();
3585 edge_duplication_hook_holder =
3586 cgraph_add_edge_duplication_hook (&ipcp_edge_duplication_hook, NULL);
3587 ipcp_values_pool = create_alloc_pool ("IPA-CP values",
3588 sizeof (struct ipcp_value), 32);
3589 ipcp_sources_pool = create_alloc_pool ("IPA-CP value sources",
3590 sizeof (struct ipcp_value_source), 64);
3591 ipcp_agg_lattice_pool = create_alloc_pool ("IPA_CP aggregate lattices",
3592 sizeof (struct ipcp_agg_lattice),
3593 32);
3594 if (dump_file)
3596 fprintf (dump_file, "\nIPA structures before propagation:\n");
3597 if (dump_flags & TDF_DETAILS)
3598 ipa_print_all_params (dump_file);
3599 ipa_print_all_jump_functions (dump_file);
3602 /* Topological sort. */
3603 build_toporder_info (&topo);
3604 /* Do the interprocedural propagation. */
3605 ipcp_propagate_stage (&topo);
3606 /* Decide what constant propagation and cloning should be performed. */
3607 ipcp_decision_stage (&topo);
3609 /* Free all IPCP structures. */
3610 free_toporder_info (&topo);
3611 next_edge_clone.release ();
3612 cgraph_remove_edge_duplication_hook (edge_duplication_hook_holder);
3613 ipa_free_all_structures_after_ipa_cp ();
3614 if (dump_file)
3615 fprintf (dump_file, "\nIPA constant propagation end\n");
3616 return 0;
3619 /* Initialization and computation of IPCP data structures. This is the initial
3620 intraprocedural analysis of functions, which gathers information to be
3621 propagated later on. */
3623 static void
3624 ipcp_generate_summary (void)
3626 struct cgraph_node *node;
3628 if (dump_file)
3629 fprintf (dump_file, "\nIPA constant propagation start:\n");
3630 ipa_register_cgraph_hooks ();
3632 FOR_EACH_FUNCTION_WITH_GIMPLE_BODY (node)
3634 node->local.versionable
3635 = tree_versionable_function_p (node->symbol.decl);
3636 ipa_analyze_node (node);
3640 /* Write ipcp summary for nodes in SET. */
3642 static void
3643 ipcp_write_summary (void)
3645 ipa_prop_write_jump_functions ();
3648 /* Read ipcp summary. */
3650 static void
3651 ipcp_read_summary (void)
3653 ipa_prop_read_jump_functions ();
3656 /* Gate for IPCP optimization. */
3658 static bool
3659 cgraph_gate_cp (void)
3661 /* FIXME: We should remove the optimize check after we ensure we never run
3662 IPA passes when not optimizing. */
3663 return flag_ipa_cp && optimize;
3666 struct ipa_opt_pass_d pass_ipa_cp =
3669 IPA_PASS,
3670 "cp", /* name */
3671 OPTGROUP_NONE, /* optinfo_flags */
3672 cgraph_gate_cp, /* gate */
3673 ipcp_driver, /* execute */
3674 NULL, /* sub */
3675 NULL, /* next */
3676 0, /* static_pass_number */
3677 TV_IPA_CONSTANT_PROP, /* tv_id */
3678 0, /* properties_required */
3679 0, /* properties_provided */
3680 0, /* properties_destroyed */
3681 0, /* todo_flags_start */
3682 TODO_dump_symtab |
3683 TODO_remove_functions /* todo_flags_finish */
3685 ipcp_generate_summary, /* generate_summary */
3686 ipcp_write_summary, /* write_summary */
3687 ipcp_read_summary, /* read_summary */
3688 ipa_prop_write_all_agg_replacement, /* write_optimization_summary */
3689 ipa_prop_read_all_agg_replacement, /* read_optimization_summary */
3690 NULL, /* stmt_fixup */
3691 0, /* TODOs */
3692 ipcp_transform_function, /* function_transform */
3693 NULL, /* variable_transform */