2013-05-03 Richard Biener <rguenther@suse.de>
[official-gcc.git] / gcc / ipa-cp.c
blobf16bd1aa5b51f95637495a096870971da6f2a084
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->uid,s->cs->frequency);
373 fprintf (f, "]");
376 if (dump_benefits)
377 fprintf (f, " [loc_time: %i, loc_size: %i, "
378 "prop_time: %i, prop_size: %i]\n",
379 val->local_time_benefit, val->local_size_cost,
380 val->prop_time_benefit, val->prop_size_cost);
382 if (!dump_benefits)
383 fprintf (f, "\n");
386 /* Print all ipcp_lattices of all functions to F. */
388 static void
389 print_all_lattices (FILE * f, bool dump_sources, bool dump_benefits)
391 struct cgraph_node *node;
392 int i, count;
394 fprintf (f, "\nLattices:\n");
395 FOR_EACH_FUNCTION_WITH_GIMPLE_BODY (node)
397 struct ipa_node_params *info;
399 info = IPA_NODE_REF (node);
400 fprintf (f, " Node: %s/%i:\n", cgraph_node_name (node), node->uid);
401 count = ipa_get_param_count (info);
402 for (i = 0; i < count; i++)
404 struct ipcp_agg_lattice *aglat;
405 struct ipcp_param_lattices *plats = ipa_get_parm_lattices (info, i);
406 fprintf (f, " param [%d]: ", i);
407 print_lattice (f, &plats->itself, dump_sources, dump_benefits);
409 if (plats->virt_call)
410 fprintf (f, " virt_call flag set\n");
412 if (plats->aggs_bottom)
414 fprintf (f, " AGGS BOTTOM\n");
415 continue;
417 if (plats->aggs_contain_variable)
418 fprintf (f, " AGGS VARIABLE\n");
419 for (aglat = plats->aggs; aglat; aglat = aglat->next)
421 fprintf (f, " %soffset " HOST_WIDE_INT_PRINT_DEC ": ",
422 plats->aggs_by_ref ? "ref " : "", aglat->offset);
423 print_lattice (f, aglat, dump_sources, dump_benefits);
429 /* Determine whether it is at all technically possible to create clones of NODE
430 and store this information in the ipa_node_params structure associated
431 with NODE. */
433 static void
434 determine_versionability (struct cgraph_node *node)
436 const char *reason = NULL;
438 /* There are a number of generic reasons functions cannot be versioned. We
439 also cannot remove parameters if there are type attributes such as fnspec
440 present. */
441 if (node->alias || node->thunk.thunk_p)
442 reason = "alias or thunk";
443 else if (!node->local.versionable)
444 reason = "not a tree_versionable_function";
445 else if (cgraph_function_body_availability (node) <= AVAIL_OVERWRITABLE)
446 reason = "insufficient body availability";
448 if (reason && dump_file && !node->alias && !node->thunk.thunk_p)
449 fprintf (dump_file, "Function %s/%i is not versionable, reason: %s.\n",
450 cgraph_node_name (node), node->uid, reason);
452 node->local.versionable = (reason == NULL);
455 /* Return true if it is at all technically possible to create clones of a
456 NODE. */
458 static bool
459 ipcp_versionable_function_p (struct cgraph_node *node)
461 return node->local.versionable;
464 /* Structure holding accumulated information about callers of a node. */
466 struct caller_statistics
468 gcov_type count_sum;
469 int n_calls, n_hot_calls, freq_sum;
472 /* Initialize fields of STAT to zeroes. */
474 static inline void
475 init_caller_stats (struct caller_statistics *stats)
477 stats->count_sum = 0;
478 stats->n_calls = 0;
479 stats->n_hot_calls = 0;
480 stats->freq_sum = 0;
483 /* Worker callback of cgraph_for_node_and_aliases accumulating statistics of
484 non-thunk incoming edges to NODE. */
486 static bool
487 gather_caller_stats (struct cgraph_node *node, void *data)
489 struct caller_statistics *stats = (struct caller_statistics *) data;
490 struct cgraph_edge *cs;
492 for (cs = node->callers; cs; cs = cs->next_caller)
493 if (cs->caller->thunk.thunk_p)
494 cgraph_for_node_and_aliases (cs->caller, gather_caller_stats,
495 stats, false);
496 else
498 stats->count_sum += cs->count;
499 stats->freq_sum += cs->frequency;
500 stats->n_calls++;
501 if (cgraph_maybe_hot_edge_p (cs))
502 stats->n_hot_calls ++;
504 return false;
508 /* Return true if this NODE is viable candidate for cloning. */
510 static bool
511 ipcp_cloning_candidate_p (struct cgraph_node *node)
513 struct caller_statistics stats;
515 gcc_checking_assert (cgraph_function_with_gimple_body_p (node));
517 if (!flag_ipa_cp_clone)
519 if (dump_file)
520 fprintf (dump_file, "Not considering %s for cloning; "
521 "-fipa-cp-clone disabled.\n",
522 cgraph_node_name (node));
523 return false;
526 if (!optimize_function_for_speed_p (DECL_STRUCT_FUNCTION (node->symbol.decl)))
528 if (dump_file)
529 fprintf (dump_file, "Not considering %s for cloning; "
530 "optimizing it for size.\n",
531 cgraph_node_name (node));
532 return false;
535 init_caller_stats (&stats);
536 cgraph_for_node_and_aliases (node, gather_caller_stats, &stats, false);
538 if (inline_summary (node)->self_size < stats.n_calls)
540 if (dump_file)
541 fprintf (dump_file, "Considering %s for cloning; code might shrink.\n",
542 cgraph_node_name (node));
543 return true;
546 /* When profile is available and function is hot, propagate into it even if
547 calls seems cold; constant propagation can improve function's speed
548 significantly. */
549 if (max_count)
551 if (stats.count_sum > node->count * 90 / 100)
553 if (dump_file)
554 fprintf (dump_file, "Considering %s for cloning; "
555 "usually called directly.\n",
556 cgraph_node_name (node));
557 return true;
560 if (!stats.n_hot_calls)
562 if (dump_file)
563 fprintf (dump_file, "Not considering %s for cloning; no hot calls.\n",
564 cgraph_node_name (node));
565 return false;
567 if (dump_file)
568 fprintf (dump_file, "Considering %s for cloning.\n",
569 cgraph_node_name (node));
570 return true;
573 /* Arrays representing a topological ordering of call graph nodes and a stack
574 of noes used during constant propagation. */
576 struct topo_info
578 struct cgraph_node **order;
579 struct cgraph_node **stack;
580 int nnodes, stack_top;
583 /* Allocate the arrays in TOPO and topologically sort the nodes into order. */
585 static void
586 build_toporder_info (struct topo_info *topo)
588 topo->order = XCNEWVEC (struct cgraph_node *, cgraph_n_nodes);
589 topo->stack = XCNEWVEC (struct cgraph_node *, cgraph_n_nodes);
590 topo->stack_top = 0;
591 topo->nnodes = ipa_reduced_postorder (topo->order, true, true, NULL);
594 /* Free information about strongly connected components and the arrays in
595 TOPO. */
597 static void
598 free_toporder_info (struct topo_info *topo)
600 ipa_free_postorder_info ();
601 free (topo->order);
602 free (topo->stack);
605 /* Add NODE to the stack in TOPO, unless it is already there. */
607 static inline void
608 push_node_to_stack (struct topo_info *topo, struct cgraph_node *node)
610 struct ipa_node_params *info = IPA_NODE_REF (node);
611 if (info->node_enqueued)
612 return;
613 info->node_enqueued = 1;
614 topo->stack[topo->stack_top++] = node;
617 /* Pop a node from the stack in TOPO and return it or return NULL if the stack
618 is empty. */
620 static struct cgraph_node *
621 pop_node_from_stack (struct topo_info *topo)
623 if (topo->stack_top)
625 struct cgraph_node *node;
626 topo->stack_top--;
627 node = topo->stack[topo->stack_top];
628 IPA_NODE_REF (node)->node_enqueued = 0;
629 return node;
631 else
632 return NULL;
635 /* Set lattice LAT to bottom and return true if it previously was not set as
636 such. */
638 static inline bool
639 set_lattice_to_bottom (struct ipcp_lattice *lat)
641 bool ret = !lat->bottom;
642 lat->bottom = true;
643 return ret;
646 /* Mark lattice as containing an unknown value and return true if it previously
647 was not marked as such. */
649 static inline bool
650 set_lattice_contains_variable (struct ipcp_lattice *lat)
652 bool ret = !lat->contains_variable;
653 lat->contains_variable = true;
654 return ret;
657 /* Set all aggegate lattices in PLATS to bottom and return true if they were
658 not previously set as such. */
660 static inline bool
661 set_agg_lats_to_bottom (struct ipcp_param_lattices *plats)
663 bool ret = !plats->aggs_bottom;
664 plats->aggs_bottom = true;
665 return ret;
668 /* Mark all aggegate lattices in PLATS as containing an unknown value and
669 return true if they were not previously marked as such. */
671 static inline bool
672 set_agg_lats_contain_variable (struct ipcp_param_lattices *plats)
674 bool ret = !plats->aggs_contain_variable;
675 plats->aggs_contain_variable = true;
676 return ret;
679 /* Mark bot aggregate and scalar lattices as containing an unknown variable,
680 return true is any of them has not been marked as such so far. */
682 static inline bool
683 set_all_contains_variable (struct ipcp_param_lattices *plats)
685 bool ret = !plats->itself.contains_variable || !plats->aggs_contain_variable;
686 plats->itself.contains_variable = true;
687 plats->aggs_contain_variable = true;
688 return ret;
691 /* Initialize ipcp_lattices. */
693 static void
694 initialize_node_lattices (struct cgraph_node *node)
696 struct ipa_node_params *info = IPA_NODE_REF (node);
697 struct cgraph_edge *ie;
698 bool disable = false, variable = false;
699 int i;
701 gcc_checking_assert (cgraph_function_with_gimple_body_p (node));
702 if (!node->local.local)
704 /* When cloning is allowed, we can assume that externally visible
705 functions are not called. We will compensate this by cloning
706 later. */
707 if (ipcp_versionable_function_p (node)
708 && ipcp_cloning_candidate_p (node))
709 variable = true;
710 else
711 disable = true;
714 if (disable || variable)
716 for (i = 0; i < ipa_get_param_count (info) ; i++)
718 struct ipcp_param_lattices *plats = ipa_get_parm_lattices (info, i);
719 if (disable)
721 set_lattice_to_bottom (&plats->itself);
722 set_agg_lats_to_bottom (plats);
724 else
725 set_all_contains_variable (plats);
727 if (dump_file && (dump_flags & TDF_DETAILS)
728 && !node->alias && !node->thunk.thunk_p)
729 fprintf (dump_file, "Marking all lattices of %s/%i as %s\n",
730 cgraph_node_name (node), node->uid,
731 disable ? "BOTTOM" : "VARIABLE");
734 for (ie = node->indirect_calls; ie; ie = ie->next_callee)
735 if (ie->indirect_info->polymorphic)
737 gcc_checking_assert (ie->indirect_info->param_index >= 0);
738 ipa_get_parm_lattices (info,
739 ie->indirect_info->param_index)->virt_call = 1;
743 /* Return the result of a (possibly arithmetic) pass through jump function
744 JFUNC on the constant value INPUT. Return NULL_TREE if that cannot be
745 determined or itself is considered an interprocedural invariant. */
747 static tree
748 ipa_get_jf_pass_through_result (struct ipa_jump_func *jfunc, tree input)
750 tree restype, res;
752 if (ipa_get_jf_pass_through_operation (jfunc) == NOP_EXPR)
753 return input;
754 else if (TREE_CODE (input) == TREE_BINFO)
755 return NULL_TREE;
757 gcc_checking_assert (is_gimple_ip_invariant (input));
758 if (TREE_CODE_CLASS (ipa_get_jf_pass_through_operation (jfunc))
759 == tcc_comparison)
760 restype = boolean_type_node;
761 else
762 restype = TREE_TYPE (input);
763 res = fold_binary (ipa_get_jf_pass_through_operation (jfunc), restype,
764 input, ipa_get_jf_pass_through_operand (jfunc));
766 if (res && !is_gimple_ip_invariant (res))
767 return NULL_TREE;
769 return res;
772 /* Return the result of an ancestor jump function JFUNC on the constant value
773 INPUT. Return NULL_TREE if that cannot be determined. */
775 static tree
776 ipa_get_jf_ancestor_result (struct ipa_jump_func *jfunc, tree input)
778 if (TREE_CODE (input) == TREE_BINFO)
779 return get_binfo_at_offset (input,
780 ipa_get_jf_ancestor_offset (jfunc),
781 ipa_get_jf_ancestor_type (jfunc));
782 else if (TREE_CODE (input) == ADDR_EXPR)
784 tree t = TREE_OPERAND (input, 0);
785 t = build_ref_for_offset (EXPR_LOCATION (t), t,
786 ipa_get_jf_ancestor_offset (jfunc),
787 ipa_get_jf_ancestor_type (jfunc), NULL, false);
788 return build_fold_addr_expr (t);
790 else
791 return NULL_TREE;
794 /* Determine whether JFUNC evaluates to a known value (that is either a
795 constant or a binfo) and if so, return it. Otherwise return NULL. INFO
796 describes the caller node so that pass-through jump functions can be
797 evaluated. */
799 tree
800 ipa_value_from_jfunc (struct ipa_node_params *info, struct ipa_jump_func *jfunc)
802 if (jfunc->type == IPA_JF_CONST)
803 return ipa_get_jf_constant (jfunc);
804 else if (jfunc->type == IPA_JF_KNOWN_TYPE)
805 return ipa_binfo_from_known_type_jfunc (jfunc);
806 else if (jfunc->type == IPA_JF_PASS_THROUGH
807 || jfunc->type == IPA_JF_ANCESTOR)
809 tree input;
810 int idx;
812 if (jfunc->type == IPA_JF_PASS_THROUGH)
813 idx = ipa_get_jf_pass_through_formal_id (jfunc);
814 else
815 idx = ipa_get_jf_ancestor_formal_id (jfunc);
817 if (info->ipcp_orig_node)
818 input = info->known_vals[idx];
819 else
821 struct ipcp_lattice *lat;
823 if (!info->lattices)
825 gcc_checking_assert (!flag_ipa_cp);
826 return NULL_TREE;
828 lat = ipa_get_scalar_lat (info, idx);
829 if (!ipa_lat_is_single_const (lat))
830 return NULL_TREE;
831 input = lat->values->value;
834 if (!input)
835 return NULL_TREE;
837 if (jfunc->type == IPA_JF_PASS_THROUGH)
838 return ipa_get_jf_pass_through_result (jfunc, input);
839 else
840 return ipa_get_jf_ancestor_result (jfunc, input);
842 else
843 return NULL_TREE;
847 /* If checking is enabled, verify that no lattice is in the TOP state, i.e. not
848 bottom, not containing a variable component and without any known value at
849 the same time. */
851 DEBUG_FUNCTION void
852 ipcp_verify_propagated_values (void)
854 struct cgraph_node *node;
856 FOR_EACH_FUNCTION_WITH_GIMPLE_BODY (node)
858 struct ipa_node_params *info = IPA_NODE_REF (node);
859 int i, count = ipa_get_param_count (info);
861 for (i = 0; i < count; i++)
863 struct ipcp_lattice *lat = ipa_get_scalar_lat (info, i);
865 if (!lat->bottom
866 && !lat->contains_variable
867 && lat->values_count == 0)
869 if (dump_file)
871 fprintf (dump_file, "\nIPA lattices after constant "
872 "propagation:\n");
873 print_all_lattices (dump_file, true, false);
876 gcc_unreachable ();
882 /* Return true iff X and Y should be considered equal values by IPA-CP. */
884 static bool
885 values_equal_for_ipcp_p (tree x, tree y)
887 gcc_checking_assert (x != NULL_TREE && y != NULL_TREE);
889 if (x == y)
890 return true;
892 if (TREE_CODE (x) == TREE_BINFO || TREE_CODE (y) == TREE_BINFO)
893 return false;
895 if (TREE_CODE (x) == ADDR_EXPR
896 && TREE_CODE (y) == ADDR_EXPR
897 && TREE_CODE (TREE_OPERAND (x, 0)) == CONST_DECL
898 && TREE_CODE (TREE_OPERAND (y, 0)) == CONST_DECL)
899 return operand_equal_p (DECL_INITIAL (TREE_OPERAND (x, 0)),
900 DECL_INITIAL (TREE_OPERAND (y, 0)), 0);
901 else
902 return operand_equal_p (x, y, 0);
905 /* Add a new value source to VAL, marking that a value comes from edge CS and
906 (if the underlying jump function is a pass-through or an ancestor one) from
907 a caller value SRC_VAL of a caller parameter described by SRC_INDEX. OFFSET
908 is negative if the source was the scalar value of the parameter itself or
909 the offset within an aggregate. */
911 static void
912 add_value_source (struct ipcp_value *val, struct cgraph_edge *cs,
913 struct ipcp_value *src_val, int src_idx, HOST_WIDE_INT offset)
915 struct ipcp_value_source *src;
917 src = (struct ipcp_value_source *) pool_alloc (ipcp_sources_pool);
918 src->offset = offset;
919 src->cs = cs;
920 src->val = src_val;
921 src->index = src_idx;
923 src->next = val->sources;
924 val->sources = src;
927 /* Try to add NEWVAL to LAT, potentially creating a new struct ipcp_value for
928 it. CS, SRC_VAL SRC_INDEX and OFFSET are meant for add_value_source and
929 have the same meaning. */
931 static bool
932 add_value_to_lattice (struct ipcp_lattice *lat, tree newval,
933 struct cgraph_edge *cs, struct ipcp_value *src_val,
934 int src_idx, HOST_WIDE_INT offset)
936 struct ipcp_value *val;
938 if (lat->bottom)
939 return false;
941 for (val = lat->values; val; val = val->next)
942 if (values_equal_for_ipcp_p (val->value, newval))
944 if (edge_within_scc (cs))
946 struct ipcp_value_source *s;
947 for (s = val->sources; s ; s = s->next)
948 if (s->cs == cs)
949 break;
950 if (s)
951 return false;
954 add_value_source (val, cs, src_val, src_idx, offset);
955 return false;
958 if (lat->values_count == PARAM_VALUE (PARAM_IPA_CP_VALUE_LIST_SIZE))
960 /* We can only free sources, not the values themselves, because sources
961 of other values in this this SCC might point to them. */
962 for (val = lat->values; val; val = val->next)
964 while (val->sources)
966 struct ipcp_value_source *src = val->sources;
967 val->sources = src->next;
968 pool_free (ipcp_sources_pool, src);
972 lat->values = NULL;
973 return set_lattice_to_bottom (lat);
976 lat->values_count++;
977 val = (struct ipcp_value *) pool_alloc (ipcp_values_pool);
978 memset (val, 0, sizeof (*val));
980 add_value_source (val, cs, src_val, src_idx, offset);
981 val->value = newval;
982 val->next = lat->values;
983 lat->values = val;
984 return true;
987 /* Like above but passes a special value of offset to distinguish that the
988 origin is the scalar value of the parameter rather than a part of an
989 aggregate. */
991 static inline bool
992 add_scalar_value_to_lattice (struct ipcp_lattice *lat, tree newval,
993 struct cgraph_edge *cs,
994 struct ipcp_value *src_val, int src_idx)
996 return add_value_to_lattice (lat, newval, cs, src_val, src_idx, -1);
999 /* Propagate values through a pass-through jump function JFUNC associated with
1000 edge CS, taking values from SRC_LAT and putting them into DEST_LAT. SRC_IDX
1001 is the index of the source parameter. */
1003 static bool
1004 propagate_vals_accross_pass_through (struct cgraph_edge *cs,
1005 struct ipa_jump_func *jfunc,
1006 struct ipcp_lattice *src_lat,
1007 struct ipcp_lattice *dest_lat,
1008 int src_idx)
1010 struct ipcp_value *src_val;
1011 bool ret = false;
1013 if (ipa_get_jf_pass_through_operation (jfunc) == NOP_EXPR)
1014 for (src_val = src_lat->values; src_val; src_val = src_val->next)
1015 ret |= add_scalar_value_to_lattice (dest_lat, src_val->value, cs,
1016 src_val, src_idx);
1017 /* Do not create new values when propagating within an SCC because if there
1018 are arithmetic functions with circular dependencies, there is infinite
1019 number of them and we would just make lattices bottom. */
1020 else if (edge_within_scc (cs))
1021 ret = set_lattice_contains_variable (dest_lat);
1022 else
1023 for (src_val = src_lat->values; src_val; src_val = src_val->next)
1025 tree cstval = src_val->value;
1027 if (TREE_CODE (cstval) == TREE_BINFO)
1029 ret |= set_lattice_contains_variable (dest_lat);
1030 continue;
1032 cstval = ipa_get_jf_pass_through_result (jfunc, cstval);
1034 if (cstval)
1035 ret |= add_scalar_value_to_lattice (dest_lat, cstval, cs, src_val,
1036 src_idx);
1037 else
1038 ret |= set_lattice_contains_variable (dest_lat);
1041 return ret;
1044 /* Propagate values through an ancestor jump function JFUNC associated with
1045 edge CS, taking values from SRC_LAT and putting them into DEST_LAT. SRC_IDX
1046 is the index of the source parameter. */
1048 static bool
1049 propagate_vals_accross_ancestor (struct cgraph_edge *cs,
1050 struct ipa_jump_func *jfunc,
1051 struct ipcp_lattice *src_lat,
1052 struct ipcp_lattice *dest_lat,
1053 int src_idx)
1055 struct ipcp_value *src_val;
1056 bool ret = false;
1058 if (edge_within_scc (cs))
1059 return set_lattice_contains_variable (dest_lat);
1061 for (src_val = src_lat->values; src_val; src_val = src_val->next)
1063 tree t = ipa_get_jf_ancestor_result (jfunc, src_val->value);
1065 if (t)
1066 ret |= add_scalar_value_to_lattice (dest_lat, t, cs, src_val, src_idx);
1067 else
1068 ret |= set_lattice_contains_variable (dest_lat);
1071 return ret;
1074 /* Propagate scalar values across jump function JFUNC that is associated with
1075 edge CS and put the values into DEST_LAT. */
1077 static bool
1078 propagate_scalar_accross_jump_function (struct cgraph_edge *cs,
1079 struct ipa_jump_func *jfunc,
1080 struct ipcp_lattice *dest_lat)
1082 if (dest_lat->bottom)
1083 return false;
1085 if (jfunc->type == IPA_JF_CONST
1086 || jfunc->type == IPA_JF_KNOWN_TYPE)
1088 tree val;
1090 if (jfunc->type == IPA_JF_KNOWN_TYPE)
1092 val = ipa_binfo_from_known_type_jfunc (jfunc);
1093 if (!val)
1094 return set_lattice_contains_variable (dest_lat);
1096 else
1097 val = ipa_get_jf_constant (jfunc);
1098 return add_scalar_value_to_lattice (dest_lat, val, cs, NULL, 0);
1100 else if (jfunc->type == IPA_JF_PASS_THROUGH
1101 || jfunc->type == IPA_JF_ANCESTOR)
1103 struct ipa_node_params *caller_info = IPA_NODE_REF (cs->caller);
1104 struct ipcp_lattice *src_lat;
1105 int src_idx;
1106 bool ret;
1108 if (jfunc->type == IPA_JF_PASS_THROUGH)
1109 src_idx = ipa_get_jf_pass_through_formal_id (jfunc);
1110 else
1111 src_idx = ipa_get_jf_ancestor_formal_id (jfunc);
1113 src_lat = ipa_get_scalar_lat (caller_info, src_idx);
1114 if (src_lat->bottom)
1115 return set_lattice_contains_variable (dest_lat);
1117 /* If we would need to clone the caller and cannot, do not propagate. */
1118 if (!ipcp_versionable_function_p (cs->caller)
1119 && (src_lat->contains_variable
1120 || (src_lat->values_count > 1)))
1121 return set_lattice_contains_variable (dest_lat);
1123 if (jfunc->type == IPA_JF_PASS_THROUGH)
1124 ret = propagate_vals_accross_pass_through (cs, jfunc, src_lat,
1125 dest_lat, src_idx);
1126 else
1127 ret = propagate_vals_accross_ancestor (cs, jfunc, src_lat, dest_lat,
1128 src_idx);
1130 if (src_lat->contains_variable)
1131 ret |= set_lattice_contains_variable (dest_lat);
1133 return ret;
1136 /* TODO: We currently do not handle member method pointers in IPA-CP (we only
1137 use it for indirect inlining), we should propagate them too. */
1138 return set_lattice_contains_variable (dest_lat);
1141 /* If DEST_PLATS already has aggregate items, check that aggs_by_ref matches
1142 NEW_AGGS_BY_REF and if not, mark all aggs as bottoms and return true (in all
1143 other cases, return false). If there are no aggregate items, set
1144 aggs_by_ref to NEW_AGGS_BY_REF. */
1146 static bool
1147 set_check_aggs_by_ref (struct ipcp_param_lattices *dest_plats,
1148 bool new_aggs_by_ref)
1150 if (dest_plats->aggs)
1152 if (dest_plats->aggs_by_ref != new_aggs_by_ref)
1154 set_agg_lats_to_bottom (dest_plats);
1155 return true;
1158 else
1159 dest_plats->aggs_by_ref = new_aggs_by_ref;
1160 return false;
1163 /* Walk aggregate lattices in DEST_PLATS from ***AGLAT on, until ***aglat is an
1164 already existing lattice for the given OFFSET and SIZE, marking all skipped
1165 lattices as containing variable and checking for overlaps. If there is no
1166 already existing lattice for the OFFSET and VAL_SIZE, create one, initialize
1167 it with offset, size and contains_variable to PRE_EXISTING, and return true,
1168 unless there are too many already. If there are two many, return false. If
1169 there are overlaps turn whole DEST_PLATS to bottom and return false. If any
1170 skipped lattices were newly marked as containing variable, set *CHANGE to
1171 true. */
1173 static bool
1174 merge_agg_lats_step (struct ipcp_param_lattices *dest_plats,
1175 HOST_WIDE_INT offset, HOST_WIDE_INT val_size,
1176 struct ipcp_agg_lattice ***aglat,
1177 bool pre_existing, bool *change)
1179 gcc_checking_assert (offset >= 0);
1181 while (**aglat && (**aglat)->offset < offset)
1183 if ((**aglat)->offset + (**aglat)->size > offset)
1185 set_agg_lats_to_bottom (dest_plats);
1186 return false;
1188 *change |= set_lattice_contains_variable (**aglat);
1189 *aglat = &(**aglat)->next;
1192 if (**aglat && (**aglat)->offset == offset)
1194 if ((**aglat)->size != val_size
1195 || ((**aglat)->next
1196 && (**aglat)->next->offset < offset + val_size))
1198 set_agg_lats_to_bottom (dest_plats);
1199 return false;
1201 gcc_checking_assert (!(**aglat)->next
1202 || (**aglat)->next->offset >= offset + val_size);
1203 return true;
1205 else
1207 struct ipcp_agg_lattice *new_al;
1209 if (**aglat && (**aglat)->offset < offset + val_size)
1211 set_agg_lats_to_bottom (dest_plats);
1212 return false;
1214 if (dest_plats->aggs_count == PARAM_VALUE (PARAM_IPA_MAX_AGG_ITEMS))
1215 return false;
1216 dest_plats->aggs_count++;
1217 new_al = (struct ipcp_agg_lattice *) pool_alloc (ipcp_agg_lattice_pool);
1218 memset (new_al, 0, sizeof (*new_al));
1220 new_al->offset = offset;
1221 new_al->size = val_size;
1222 new_al->contains_variable = pre_existing;
1224 new_al->next = **aglat;
1225 **aglat = new_al;
1226 return true;
1230 /* Set all AGLAT and all other aggregate lattices reachable by next pointers as
1231 containing an unknown value. */
1233 static bool
1234 set_chain_of_aglats_contains_variable (struct ipcp_agg_lattice *aglat)
1236 bool ret = false;
1237 while (aglat)
1239 ret |= set_lattice_contains_variable (aglat);
1240 aglat = aglat->next;
1242 return ret;
1245 /* Merge existing aggregate lattices in SRC_PLATS to DEST_PLATS, subtracting
1246 DELTA_OFFSET. CS is the call graph edge and SRC_IDX the index of the source
1247 parameter used for lattice value sources. Return true if DEST_PLATS changed
1248 in any way. */
1250 static bool
1251 merge_aggregate_lattices (struct cgraph_edge *cs,
1252 struct ipcp_param_lattices *dest_plats,
1253 struct ipcp_param_lattices *src_plats,
1254 int src_idx, HOST_WIDE_INT offset_delta)
1256 bool pre_existing = dest_plats->aggs != NULL;
1257 struct ipcp_agg_lattice **dst_aglat;
1258 bool ret = false;
1260 if (set_check_aggs_by_ref (dest_plats, src_plats->aggs_by_ref))
1261 return true;
1262 if (src_plats->aggs_bottom)
1263 return set_agg_lats_contain_variable (dest_plats);
1264 if (src_plats->aggs_contain_variable)
1265 ret |= set_agg_lats_contain_variable (dest_plats);
1266 dst_aglat = &dest_plats->aggs;
1268 for (struct ipcp_agg_lattice *src_aglat = src_plats->aggs;
1269 src_aglat;
1270 src_aglat = src_aglat->next)
1272 HOST_WIDE_INT new_offset = src_aglat->offset - offset_delta;
1274 if (new_offset < 0)
1275 continue;
1276 if (merge_agg_lats_step (dest_plats, new_offset, src_aglat->size,
1277 &dst_aglat, pre_existing, &ret))
1279 struct ipcp_agg_lattice *new_al = *dst_aglat;
1281 dst_aglat = &(*dst_aglat)->next;
1282 if (src_aglat->bottom)
1284 ret |= set_lattice_contains_variable (new_al);
1285 continue;
1287 if (src_aglat->contains_variable)
1288 ret |= set_lattice_contains_variable (new_al);
1289 for (struct ipcp_value *val = src_aglat->values;
1290 val;
1291 val = val->next)
1292 ret |= add_value_to_lattice (new_al, val->value, cs, val, src_idx,
1293 src_aglat->offset);
1295 else if (dest_plats->aggs_bottom)
1296 return true;
1298 ret |= set_chain_of_aglats_contains_variable (*dst_aglat);
1299 return ret;
1302 /* Determine whether there is anything to propagate FROM SRC_PLATS through a
1303 pass-through JFUNC and if so, whether it has conform and conforms to the
1304 rules about propagating values passed by reference. */
1306 static bool
1307 agg_pass_through_permissible_p (struct ipcp_param_lattices *src_plats,
1308 struct ipa_jump_func *jfunc)
1310 return src_plats->aggs
1311 && (!src_plats->aggs_by_ref
1312 || ipa_get_jf_pass_through_agg_preserved (jfunc));
1315 /* Propagate scalar values across jump function JFUNC that is associated with
1316 edge CS and put the values into DEST_LAT. */
1318 static bool
1319 propagate_aggs_accross_jump_function (struct cgraph_edge *cs,
1320 struct ipa_jump_func *jfunc,
1321 struct ipcp_param_lattices *dest_plats)
1323 bool ret = false;
1325 if (dest_plats->aggs_bottom)
1326 return false;
1328 if (jfunc->type == IPA_JF_PASS_THROUGH
1329 && ipa_get_jf_pass_through_operation (jfunc) == NOP_EXPR)
1331 struct ipa_node_params *caller_info = IPA_NODE_REF (cs->caller);
1332 int src_idx = ipa_get_jf_pass_through_formal_id (jfunc);
1333 struct ipcp_param_lattices *src_plats;
1335 src_plats = ipa_get_parm_lattices (caller_info, src_idx);
1336 if (agg_pass_through_permissible_p (src_plats, jfunc))
1338 /* Currently we do not produce clobber aggregate jump
1339 functions, replace with merging when we do. */
1340 gcc_assert (!jfunc->agg.items);
1341 ret |= merge_aggregate_lattices (cs, dest_plats, src_plats,
1342 src_idx, 0);
1344 else
1345 ret |= set_agg_lats_contain_variable (dest_plats);
1347 else if (jfunc->type == IPA_JF_ANCESTOR
1348 && ipa_get_jf_ancestor_agg_preserved (jfunc))
1350 struct ipa_node_params *caller_info = IPA_NODE_REF (cs->caller);
1351 int src_idx = ipa_get_jf_ancestor_formal_id (jfunc);
1352 struct ipcp_param_lattices *src_plats;
1354 src_plats = ipa_get_parm_lattices (caller_info, src_idx);
1355 if (src_plats->aggs && src_plats->aggs_by_ref)
1357 /* Currently we do not produce clobber aggregate jump
1358 functions, replace with merging when we do. */
1359 gcc_assert (!jfunc->agg.items);
1360 ret |= merge_aggregate_lattices (cs, dest_plats, src_plats, src_idx,
1361 ipa_get_jf_ancestor_offset (jfunc));
1363 else if (!src_plats->aggs_by_ref)
1364 ret |= set_agg_lats_to_bottom (dest_plats);
1365 else
1366 ret |= set_agg_lats_contain_variable (dest_plats);
1368 else if (jfunc->agg.items)
1370 bool pre_existing = dest_plats->aggs != NULL;
1371 struct ipcp_agg_lattice **aglat = &dest_plats->aggs;
1372 struct ipa_agg_jf_item *item;
1373 int i;
1375 if (set_check_aggs_by_ref (dest_plats, jfunc->agg.by_ref))
1376 return true;
1378 FOR_EACH_VEC_ELT (*jfunc->agg.items, i, item)
1380 HOST_WIDE_INT val_size;
1382 if (item->offset < 0)
1383 continue;
1384 gcc_checking_assert (is_gimple_ip_invariant (item->value));
1385 val_size = tree_low_cst (TYPE_SIZE (TREE_TYPE (item->value)), 1);
1387 if (merge_agg_lats_step (dest_plats, item->offset, val_size,
1388 &aglat, pre_existing, &ret))
1390 ret |= add_value_to_lattice (*aglat, item->value, cs, NULL, 0, 0);
1391 aglat = &(*aglat)->next;
1393 else if (dest_plats->aggs_bottom)
1394 return true;
1397 ret |= set_chain_of_aglats_contains_variable (*aglat);
1399 else
1400 ret |= set_agg_lats_contain_variable (dest_plats);
1402 return ret;
1405 /* Propagate constants from the caller to the callee of CS. INFO describes the
1406 caller. */
1408 static bool
1409 propagate_constants_accross_call (struct cgraph_edge *cs)
1411 struct ipa_node_params *callee_info;
1412 enum availability availability;
1413 struct cgraph_node *callee, *alias_or_thunk;
1414 struct ipa_edge_args *args;
1415 bool ret = false;
1416 int i, args_count, parms_count;
1418 callee = cgraph_function_node (cs->callee, &availability);
1419 if (!callee->analyzed)
1420 return false;
1421 gcc_checking_assert (cgraph_function_with_gimple_body_p (callee));
1422 callee_info = IPA_NODE_REF (callee);
1424 args = IPA_EDGE_REF (cs);
1425 args_count = ipa_get_cs_argument_count (args);
1426 parms_count = ipa_get_param_count (callee_info);
1428 /* If this call goes through a thunk we must not propagate to the first (0th)
1429 parameter. However, we might need to uncover a thunk from below a series
1430 of aliases first. */
1431 alias_or_thunk = cs->callee;
1432 while (alias_or_thunk->alias)
1433 alias_or_thunk = cgraph_alias_aliased_node (alias_or_thunk);
1434 if (alias_or_thunk->thunk.thunk_p)
1436 ret |= set_all_contains_variable (ipa_get_parm_lattices (callee_info,
1437 0));
1438 i = 1;
1440 else
1441 i = 0;
1443 for (; (i < args_count) && (i < parms_count); i++)
1445 struct ipa_jump_func *jump_func = ipa_get_ith_jump_func (args, i);
1446 struct ipcp_param_lattices *dest_plats;
1448 dest_plats = ipa_get_parm_lattices (callee_info, i);
1449 if (availability == AVAIL_OVERWRITABLE)
1450 ret |= set_all_contains_variable (dest_plats);
1451 else
1453 ret |= propagate_scalar_accross_jump_function (cs, jump_func,
1454 &dest_plats->itself);
1455 ret |= propagate_aggs_accross_jump_function (cs, jump_func,
1456 dest_plats);
1459 for (; i < parms_count; i++)
1460 ret |= set_all_contains_variable (ipa_get_parm_lattices (callee_info, i));
1462 return ret;
1465 /* If an indirect edge IE can be turned into a direct one based on KNOWN_VALS
1466 (which can contain both constants and binfos), KNOWN_BINFOS, KNOWN_AGGS or
1467 AGG_REPS return the destination. The latter three can be NULL. If AGG_REPS
1468 is not NULL, KNOWN_AGGS is ignored. */
1470 static tree
1471 ipa_get_indirect_edge_target_1 (struct cgraph_edge *ie,
1472 vec<tree> known_vals,
1473 vec<tree> known_binfos,
1474 vec<ipa_agg_jump_function_p> known_aggs,
1475 struct ipa_agg_replacement_value *agg_reps)
1477 int param_index = ie->indirect_info->param_index;
1478 HOST_WIDE_INT token, anc_offset;
1479 tree otr_type;
1480 tree t;
1482 if (param_index == -1)
1483 return NULL_TREE;
1485 if (!ie->indirect_info->polymorphic)
1487 tree t;
1489 if (ie->indirect_info->agg_contents)
1491 if (agg_reps)
1493 t = NULL;
1494 while (agg_reps)
1496 if (agg_reps->index == param_index
1497 && agg_reps->offset == ie->indirect_info->offset
1498 && agg_reps->by_ref == ie->indirect_info->by_ref)
1500 t = agg_reps->value;
1501 break;
1503 agg_reps = agg_reps->next;
1506 else if (known_aggs.length () > (unsigned int) param_index)
1508 struct ipa_agg_jump_function *agg;
1509 agg = known_aggs[param_index];
1510 t = ipa_find_agg_cst_for_param (agg, ie->indirect_info->offset,
1511 ie->indirect_info->by_ref);
1513 else
1514 t = NULL;
1516 else
1517 t = (known_vals.length () > (unsigned int) param_index
1518 ? known_vals[param_index] : NULL);
1520 if (t &&
1521 TREE_CODE (t) == ADDR_EXPR
1522 && TREE_CODE (TREE_OPERAND (t, 0)) == FUNCTION_DECL)
1523 return TREE_OPERAND (t, 0);
1524 else
1525 return NULL_TREE;
1528 gcc_assert (!ie->indirect_info->agg_contents);
1529 token = ie->indirect_info->otr_token;
1530 anc_offset = ie->indirect_info->offset;
1531 otr_type = ie->indirect_info->otr_type;
1533 t = known_vals[param_index];
1534 if (!t && known_binfos.length () > (unsigned int) param_index)
1535 t = known_binfos[param_index];
1536 if (!t)
1537 return NULL_TREE;
1539 if (TREE_CODE (t) != TREE_BINFO)
1541 tree binfo;
1542 binfo = gimple_extract_devirt_binfo_from_cst (t);
1543 if (!binfo)
1544 return NULL_TREE;
1545 binfo = get_binfo_at_offset (binfo, anc_offset, otr_type);
1546 if (!binfo)
1547 return NULL_TREE;
1548 return gimple_get_virt_method_for_binfo (token, binfo);
1550 else
1552 tree binfo;
1554 binfo = get_binfo_at_offset (t, anc_offset, otr_type);
1555 if (!binfo)
1556 return NULL_TREE;
1557 return gimple_get_virt_method_for_binfo (token, binfo);
1562 /* If an indirect edge IE can be turned into a direct one based on KNOWN_VALS
1563 (which can contain both constants and binfos), KNOWN_BINFOS (which can be
1564 NULL) or KNOWN_AGGS (which also can be NULL) return the destination. */
1566 tree
1567 ipa_get_indirect_edge_target (struct cgraph_edge *ie,
1568 vec<tree> known_vals,
1569 vec<tree> known_binfos,
1570 vec<ipa_agg_jump_function_p> known_aggs)
1572 return ipa_get_indirect_edge_target_1 (ie, known_vals, known_binfos,
1573 known_aggs, NULL);
1576 /* Calculate devirtualization time bonus for NODE, assuming we know KNOWN_CSTS
1577 and KNOWN_BINFOS. */
1579 static int
1580 devirtualization_time_bonus (struct cgraph_node *node,
1581 vec<tree> known_csts,
1582 vec<tree> known_binfos,
1583 vec<ipa_agg_jump_function_p> known_aggs)
1585 struct cgraph_edge *ie;
1586 int res = 0;
1588 for (ie = node->indirect_calls; ie; ie = ie->next_callee)
1590 struct cgraph_node *callee;
1591 struct inline_summary *isummary;
1592 tree target;
1594 target = ipa_get_indirect_edge_target (ie, known_csts, known_binfos,
1595 known_aggs);
1596 if (!target)
1597 continue;
1599 /* Only bare minimum benefit for clearly un-inlineable targets. */
1600 res += 1;
1601 callee = cgraph_get_node (target);
1602 if (!callee || !callee->analyzed)
1603 continue;
1604 isummary = inline_summary (callee);
1605 if (!isummary->inlinable)
1606 continue;
1608 /* FIXME: The values below need re-considering and perhaps also
1609 integrating into the cost metrics, at lest in some very basic way. */
1610 if (isummary->size <= MAX_INLINE_INSNS_AUTO / 4)
1611 res += 31;
1612 else if (isummary->size <= MAX_INLINE_INSNS_AUTO / 2)
1613 res += 15;
1614 else if (isummary->size <= MAX_INLINE_INSNS_AUTO
1615 || DECL_DECLARED_INLINE_P (callee->symbol.decl))
1616 res += 7;
1619 return res;
1622 /* Return time bonus incurred because of HINTS. */
1624 static int
1625 hint_time_bonus (inline_hints hints)
1627 int result = 0;
1628 if (hints & (INLINE_HINT_loop_iterations | INLINE_HINT_loop_stride))
1629 result += PARAM_VALUE (PARAM_IPA_CP_LOOP_HINT_BONUS);
1630 if (hints & INLINE_HINT_array_index)
1631 result += PARAM_VALUE (PARAM_IPA_CP_ARRAY_INDEX_HINT_BONUS);
1632 return result;
1635 /* Return true if cloning NODE is a good idea, given the estimated TIME_BENEFIT
1636 and SIZE_COST and with the sum of frequencies of incoming edges to the
1637 potential new clone in FREQUENCIES. */
1639 static bool
1640 good_cloning_opportunity_p (struct cgraph_node *node, int time_benefit,
1641 int freq_sum, gcov_type count_sum, int size_cost)
1643 if (time_benefit == 0
1644 || !flag_ipa_cp_clone
1645 || !optimize_function_for_speed_p (DECL_STRUCT_FUNCTION (node->symbol.decl)))
1646 return false;
1648 gcc_assert (size_cost > 0);
1650 if (max_count)
1652 int factor = (count_sum * 1000) / max_count;
1653 HOST_WIDEST_INT evaluation = (((HOST_WIDEST_INT) time_benefit * factor)
1654 / size_cost);
1656 if (dump_file && (dump_flags & TDF_DETAILS))
1657 fprintf (dump_file, " good_cloning_opportunity_p (time: %i, "
1658 "size: %i, count_sum: " HOST_WIDE_INT_PRINT_DEC
1659 ") -> evaluation: " HOST_WIDEST_INT_PRINT_DEC
1660 ", threshold: %i\n",
1661 time_benefit, size_cost, (HOST_WIDE_INT) count_sum,
1662 evaluation, PARAM_VALUE (PARAM_IPA_CP_EVAL_THRESHOLD));
1664 return evaluation >= PARAM_VALUE (PARAM_IPA_CP_EVAL_THRESHOLD);
1666 else
1668 HOST_WIDEST_INT evaluation = (((HOST_WIDEST_INT) time_benefit * freq_sum)
1669 / size_cost);
1671 if (dump_file && (dump_flags & TDF_DETAILS))
1672 fprintf (dump_file, " good_cloning_opportunity_p (time: %i, "
1673 "size: %i, freq_sum: %i) -> evaluation: "
1674 HOST_WIDEST_INT_PRINT_DEC ", threshold: %i\n",
1675 time_benefit, size_cost, freq_sum, evaluation,
1676 PARAM_VALUE (PARAM_IPA_CP_EVAL_THRESHOLD));
1678 return evaluation >= PARAM_VALUE (PARAM_IPA_CP_EVAL_THRESHOLD);
1682 /* Return all context independent values from aggregate lattices in PLATS in a
1683 vector. Return NULL if there are none. */
1685 static vec<ipa_agg_jf_item_t, va_gc> *
1686 context_independent_aggregate_values (struct ipcp_param_lattices *plats)
1688 vec<ipa_agg_jf_item_t, va_gc> *res = NULL;
1690 if (plats->aggs_bottom
1691 || plats->aggs_contain_variable
1692 || plats->aggs_count == 0)
1693 return NULL;
1695 for (struct ipcp_agg_lattice *aglat = plats->aggs;
1696 aglat;
1697 aglat = aglat->next)
1698 if (ipa_lat_is_single_const (aglat))
1700 struct ipa_agg_jf_item item;
1701 item.offset = aglat->offset;
1702 item.value = aglat->values->value;
1703 vec_safe_push (res, item);
1705 return res;
1708 /* Allocate KNOWN_CSTS, KNOWN_BINFOS and, if non-NULL, KNOWN_AGGS and populate
1709 them with values of parameters that are known independent of the context.
1710 INFO describes the function. If REMOVABLE_PARAMS_COST is non-NULL, the
1711 movement cost of all removable parameters will be stored in it. */
1713 static bool
1714 gather_context_independent_values (struct ipa_node_params *info,
1715 vec<tree> *known_csts,
1716 vec<tree> *known_binfos,
1717 vec<ipa_agg_jump_function_t> *known_aggs,
1718 int *removable_params_cost)
1720 int i, count = ipa_get_param_count (info);
1721 bool ret = false;
1723 known_csts->create (0);
1724 known_binfos->create (0);
1725 known_csts->safe_grow_cleared (count);
1726 known_binfos->safe_grow_cleared (count);
1727 if (known_aggs)
1729 known_aggs->create (0);
1730 known_aggs->safe_grow_cleared (count);
1733 if (removable_params_cost)
1734 *removable_params_cost = 0;
1736 for (i = 0; i < count ; i++)
1738 struct ipcp_param_lattices *plats = ipa_get_parm_lattices (info, i);
1739 struct ipcp_lattice *lat = &plats->itself;
1741 if (ipa_lat_is_single_const (lat))
1743 struct ipcp_value *val = lat->values;
1744 if (TREE_CODE (val->value) != TREE_BINFO)
1746 (*known_csts)[i] = val->value;
1747 if (removable_params_cost)
1748 *removable_params_cost
1749 += estimate_move_cost (TREE_TYPE (val->value));
1750 ret = true;
1752 else if (plats->virt_call)
1754 (*known_binfos)[i] = val->value;
1755 ret = true;
1757 else if (removable_params_cost
1758 && !ipa_is_param_used (info, i))
1759 *removable_params_cost
1760 += estimate_move_cost (TREE_TYPE (ipa_get_param (info, i)));
1762 else if (removable_params_cost
1763 && !ipa_is_param_used (info, i))
1764 *removable_params_cost
1765 += estimate_move_cost (TREE_TYPE (ipa_get_param (info, i)));
1767 if (known_aggs)
1769 vec<ipa_agg_jf_item_t, va_gc> *agg_items;
1770 struct ipa_agg_jump_function *ajf;
1772 agg_items = context_independent_aggregate_values (plats);
1773 ajf = &(*known_aggs)[i];
1774 ajf->items = agg_items;
1775 ajf->by_ref = plats->aggs_by_ref;
1776 ret |= agg_items != NULL;
1780 return ret;
1783 /* The current interface in ipa-inline-analysis requires a pointer vector.
1784 Create it.
1786 FIXME: That interface should be re-worked, this is slightly silly. Still,
1787 I'd like to discuss how to change it first and this demonstrates the
1788 issue. */
1790 static vec<ipa_agg_jump_function_p>
1791 agg_jmp_p_vec_for_t_vec (vec<ipa_agg_jump_function_t> known_aggs)
1793 vec<ipa_agg_jump_function_p> ret;
1794 struct ipa_agg_jump_function *ajf;
1795 int i;
1797 ret.create (known_aggs.length ());
1798 FOR_EACH_VEC_ELT (known_aggs, i, ajf)
1799 ret.quick_push (ajf);
1800 return ret;
1803 /* Iterate over known values of parameters of NODE and estimate the local
1804 effects in terms of time and size they have. */
1806 static void
1807 estimate_local_effects (struct cgraph_node *node)
1809 struct ipa_node_params *info = IPA_NODE_REF (node);
1810 int i, count = ipa_get_param_count (info);
1811 vec<tree> known_csts, known_binfos;
1812 vec<ipa_agg_jump_function_t> known_aggs;
1813 vec<ipa_agg_jump_function_p> known_aggs_ptrs;
1814 bool always_const;
1815 int base_time = inline_summary (node)->time;
1816 int removable_params_cost;
1818 if (!count || !ipcp_versionable_function_p (node))
1819 return;
1821 if (dump_file && (dump_flags & TDF_DETAILS))
1822 fprintf (dump_file, "\nEstimating effects for %s/%i, base_time: %i.\n",
1823 cgraph_node_name (node), node->uid, base_time);
1825 always_const = gather_context_independent_values (info, &known_csts,
1826 &known_binfos, &known_aggs,
1827 &removable_params_cost);
1828 known_aggs_ptrs = agg_jmp_p_vec_for_t_vec (known_aggs);
1829 if (always_const)
1831 struct caller_statistics stats;
1832 inline_hints hints;
1833 int time, size;
1835 init_caller_stats (&stats);
1836 cgraph_for_node_and_aliases (node, gather_caller_stats, &stats, false);
1837 estimate_ipcp_clone_size_and_time (node, known_csts, known_binfos,
1838 known_aggs_ptrs, &size, &time, &hints);
1839 time -= devirtualization_time_bonus (node, known_csts, known_binfos,
1840 known_aggs_ptrs);
1841 time -= hint_time_bonus (hints);
1842 time -= removable_params_cost;
1843 size -= stats.n_calls * removable_params_cost;
1845 if (dump_file)
1846 fprintf (dump_file, " - context independent values, size: %i, "
1847 "time_benefit: %i\n", size, base_time - time);
1849 if (size <= 0
1850 || cgraph_will_be_removed_from_program_if_no_direct_calls (node))
1852 info->do_clone_for_all_contexts = true;
1853 base_time = time;
1855 if (dump_file)
1856 fprintf (dump_file, " Decided to specialize for all "
1857 "known contexts, code not going to grow.\n");
1859 else if (good_cloning_opportunity_p (node, base_time - time,
1860 stats.freq_sum, stats.count_sum,
1861 size))
1863 if (size + overall_size <= max_new_size)
1865 info->do_clone_for_all_contexts = true;
1866 base_time = time;
1867 overall_size += size;
1869 if (dump_file)
1870 fprintf (dump_file, " Decided to specialize for all "
1871 "known contexts, growth deemed beneficial.\n");
1873 else if (dump_file && (dump_flags & TDF_DETAILS))
1874 fprintf (dump_file, " Not cloning for all contexts because "
1875 "max_new_size would be reached with %li.\n",
1876 size + overall_size);
1880 for (i = 0; i < count ; i++)
1882 struct ipcp_param_lattices *plats = ipa_get_parm_lattices (info, i);
1883 struct ipcp_lattice *lat = &plats->itself;
1884 struct ipcp_value *val;
1885 int emc;
1887 if (lat->bottom
1888 || !lat->values
1889 || known_csts[i]
1890 || known_binfos[i])
1891 continue;
1893 for (val = lat->values; val; val = val->next)
1895 int time, size, time_benefit;
1896 inline_hints hints;
1898 if (TREE_CODE (val->value) != TREE_BINFO)
1900 known_csts[i] = val->value;
1901 known_binfos[i] = NULL_TREE;
1902 emc = estimate_move_cost (TREE_TYPE (val->value));
1904 else if (plats->virt_call)
1906 known_csts[i] = NULL_TREE;
1907 known_binfos[i] = val->value;
1908 emc = 0;
1910 else
1911 continue;
1913 estimate_ipcp_clone_size_and_time (node, known_csts, known_binfos,
1914 known_aggs_ptrs, &size, &time,
1915 &hints);
1916 time_benefit = base_time - time
1917 + devirtualization_time_bonus (node, known_csts, known_binfos,
1918 known_aggs_ptrs)
1919 + hint_time_bonus (hints)
1920 + removable_params_cost + emc;
1922 gcc_checking_assert (size >=0);
1923 /* The inliner-heuristics based estimates may think that in certain
1924 contexts some functions do not have any size at all but we want
1925 all specializations to have at least a tiny cost, not least not to
1926 divide by zero. */
1927 if (size == 0)
1928 size = 1;
1930 if (dump_file && (dump_flags & TDF_DETAILS))
1932 fprintf (dump_file, " - estimates for value ");
1933 print_ipcp_constant_value (dump_file, val->value);
1934 fprintf (dump_file, " for parameter ");
1935 print_generic_expr (dump_file, ipa_get_param (info, i), 0);
1936 fprintf (dump_file, ": time_benefit: %i, size: %i\n",
1937 time_benefit, size);
1940 val->local_time_benefit = time_benefit;
1941 val->local_size_cost = size;
1943 known_binfos[i] = NULL_TREE;
1944 known_csts[i] = NULL_TREE;
1947 for (i = 0; i < count ; i++)
1949 struct ipcp_param_lattices *plats = ipa_get_parm_lattices (info, i);
1950 struct ipa_agg_jump_function *ajf;
1951 struct ipcp_agg_lattice *aglat;
1953 if (plats->aggs_bottom || !plats->aggs)
1954 continue;
1956 ajf = &known_aggs[i];
1957 for (aglat = plats->aggs; aglat; aglat = aglat->next)
1959 struct ipcp_value *val;
1960 if (aglat->bottom || !aglat->values
1961 /* If the following is true, the one value is in known_aggs. */
1962 || (!plats->aggs_contain_variable
1963 && ipa_lat_is_single_const (aglat)))
1964 continue;
1966 for (val = aglat->values; val; val = val->next)
1968 int time, size, time_benefit;
1969 struct ipa_agg_jf_item item;
1970 inline_hints hints;
1972 item.offset = aglat->offset;
1973 item.value = val->value;
1974 vec_safe_push (ajf->items, item);
1976 estimate_ipcp_clone_size_and_time (node, known_csts, known_binfos,
1977 known_aggs_ptrs, &size, &time,
1978 &hints);
1979 time_benefit = base_time - time
1980 + devirtualization_time_bonus (node, known_csts, known_binfos,
1981 known_aggs_ptrs)
1982 + hint_time_bonus (hints);
1983 gcc_checking_assert (size >=0);
1984 if (size == 0)
1985 size = 1;
1987 if (dump_file && (dump_flags & TDF_DETAILS))
1989 fprintf (dump_file, " - estimates for value ");
1990 print_ipcp_constant_value (dump_file, val->value);
1991 fprintf (dump_file, " for parameter ");
1992 print_generic_expr (dump_file, ipa_get_param (info, i), 0);
1993 fprintf (dump_file, "[%soffset: " HOST_WIDE_INT_PRINT_DEC
1994 "]: time_benefit: %i, size: %i\n",
1995 plats->aggs_by_ref ? "ref " : "",
1996 aglat->offset, time_benefit, size);
1999 val->local_time_benefit = time_benefit;
2000 val->local_size_cost = size;
2001 ajf->items->pop ();
2006 for (i = 0; i < count ; i++)
2007 vec_free (known_aggs[i].items);
2009 known_csts.release ();
2010 known_binfos.release ();
2011 known_aggs.release ();
2012 known_aggs_ptrs.release ();
2016 /* Add value CUR_VAL and all yet-unsorted values it is dependent on to the
2017 topological sort of values. */
2019 static void
2020 add_val_to_toposort (struct ipcp_value *cur_val)
2022 static int dfs_counter = 0;
2023 static struct ipcp_value *stack;
2024 struct ipcp_value_source *src;
2026 if (cur_val->dfs)
2027 return;
2029 dfs_counter++;
2030 cur_val->dfs = dfs_counter;
2031 cur_val->low_link = dfs_counter;
2033 cur_val->topo_next = stack;
2034 stack = cur_val;
2035 cur_val->on_stack = true;
2037 for (src = cur_val->sources; src; src = src->next)
2038 if (src->val)
2040 if (src->val->dfs == 0)
2042 add_val_to_toposort (src->val);
2043 if (src->val->low_link < cur_val->low_link)
2044 cur_val->low_link = src->val->low_link;
2046 else if (src->val->on_stack
2047 && src->val->dfs < cur_val->low_link)
2048 cur_val->low_link = src->val->dfs;
2051 if (cur_val->dfs == cur_val->low_link)
2053 struct ipcp_value *v, *scc_list = NULL;
2057 v = stack;
2058 stack = v->topo_next;
2059 v->on_stack = false;
2061 v->scc_next = scc_list;
2062 scc_list = v;
2064 while (v != cur_val);
2066 cur_val->topo_next = values_topo;
2067 values_topo = cur_val;
2071 /* Add all values in lattices associated with NODE to the topological sort if
2072 they are not there yet. */
2074 static void
2075 add_all_node_vals_to_toposort (struct cgraph_node *node)
2077 struct ipa_node_params *info = IPA_NODE_REF (node);
2078 int i, count = ipa_get_param_count (info);
2080 for (i = 0; i < count ; i++)
2082 struct ipcp_param_lattices *plats = ipa_get_parm_lattices (info, i);
2083 struct ipcp_lattice *lat = &plats->itself;
2084 struct ipcp_agg_lattice *aglat;
2085 struct ipcp_value *val;
2087 if (!lat->bottom)
2088 for (val = lat->values; val; val = val->next)
2089 add_val_to_toposort (val);
2091 if (!plats->aggs_bottom)
2092 for (aglat = plats->aggs; aglat; aglat = aglat->next)
2093 if (!aglat->bottom)
2094 for (val = aglat->values; val; val = val->next)
2095 add_val_to_toposort (val);
2099 /* One pass of constants propagation along the call graph edges, from callers
2100 to callees (requires topological ordering in TOPO), iterate over strongly
2101 connected components. */
2103 static void
2104 propagate_constants_topo (struct topo_info *topo)
2106 int i;
2108 for (i = topo->nnodes - 1; i >= 0; i--)
2110 struct cgraph_node *v, *node = topo->order[i];
2111 struct ipa_dfs_info *node_dfs_info;
2113 if (!cgraph_function_with_gimple_body_p (node))
2114 continue;
2116 node_dfs_info = (struct ipa_dfs_info *) node->symbol.aux;
2117 /* First, iteratively propagate within the strongly connected component
2118 until all lattices stabilize. */
2119 v = node_dfs_info->next_cycle;
2120 while (v)
2122 push_node_to_stack (topo, v);
2123 v = ((struct ipa_dfs_info *) v->symbol.aux)->next_cycle;
2126 v = node;
2127 while (v)
2129 struct cgraph_edge *cs;
2131 for (cs = v->callees; cs; cs = cs->next_callee)
2132 if (edge_within_scc (cs)
2133 && propagate_constants_accross_call (cs))
2134 push_node_to_stack (topo, cs->callee);
2135 v = pop_node_from_stack (topo);
2138 /* Afterwards, propagate along edges leading out of the SCC, calculates
2139 the local effects of the discovered constants and all valid values to
2140 their topological sort. */
2141 v = node;
2142 while (v)
2144 struct cgraph_edge *cs;
2146 estimate_local_effects (v);
2147 add_all_node_vals_to_toposort (v);
2148 for (cs = v->callees; cs; cs = cs->next_callee)
2149 if (!edge_within_scc (cs))
2150 propagate_constants_accross_call (cs);
2152 v = ((struct ipa_dfs_info *) v->symbol.aux)->next_cycle;
2158 /* Return the sum of A and B if none of them is bigger than INT_MAX/2, return
2159 the bigger one if otherwise. */
2161 static int
2162 safe_add (int a, int b)
2164 if (a > INT_MAX/2 || b > INT_MAX/2)
2165 return a > b ? a : b;
2166 else
2167 return a + b;
2171 /* Propagate the estimated effects of individual values along the topological
2172 from the dependent values to those they depend on. */
2174 static void
2175 propagate_effects (void)
2177 struct ipcp_value *base;
2179 for (base = values_topo; base; base = base->topo_next)
2181 struct ipcp_value_source *src;
2182 struct ipcp_value *val;
2183 int time = 0, size = 0;
2185 for (val = base; val; val = val->scc_next)
2187 time = safe_add (time,
2188 val->local_time_benefit + val->prop_time_benefit);
2189 size = safe_add (size, val->local_size_cost + val->prop_size_cost);
2192 for (val = base; val; val = val->scc_next)
2193 for (src = val->sources; src; src = src->next)
2194 if (src->val
2195 && cgraph_maybe_hot_edge_p (src->cs))
2197 src->val->prop_time_benefit = safe_add (time,
2198 src->val->prop_time_benefit);
2199 src->val->prop_size_cost = safe_add (size,
2200 src->val->prop_size_cost);
2206 /* Propagate constants, binfos and their effects from the summaries
2207 interprocedurally. */
2209 static void
2210 ipcp_propagate_stage (struct topo_info *topo)
2212 struct cgraph_node *node;
2214 if (dump_file)
2215 fprintf (dump_file, "\n Propagating constants:\n\n");
2217 if (in_lto_p)
2218 ipa_update_after_lto_read ();
2221 FOR_EACH_DEFINED_FUNCTION (node)
2223 struct ipa_node_params *info = IPA_NODE_REF (node);
2225 determine_versionability (node);
2226 if (cgraph_function_with_gimple_body_p (node))
2228 info->lattices = XCNEWVEC (struct ipcp_param_lattices,
2229 ipa_get_param_count (info));
2230 initialize_node_lattices (node);
2232 if (node->count > max_count)
2233 max_count = node->count;
2234 overall_size += inline_summary (node)->self_size;
2237 max_new_size = overall_size;
2238 if (max_new_size < PARAM_VALUE (PARAM_LARGE_UNIT_INSNS))
2239 max_new_size = PARAM_VALUE (PARAM_LARGE_UNIT_INSNS);
2240 max_new_size += max_new_size * PARAM_VALUE (PARAM_IPCP_UNIT_GROWTH) / 100 + 1;
2242 if (dump_file)
2243 fprintf (dump_file, "\noverall_size: %li, max_new_size: %li\n",
2244 overall_size, max_new_size);
2246 propagate_constants_topo (topo);
2247 #ifdef ENABLE_CHECKING
2248 ipcp_verify_propagated_values ();
2249 #endif
2250 propagate_effects ();
2252 if (dump_file)
2254 fprintf (dump_file, "\nIPA lattices after all propagation:\n");
2255 print_all_lattices (dump_file, (dump_flags & TDF_DETAILS), true);
2259 /* Discover newly direct outgoing edges from NODE which is a new clone with
2260 known KNOWN_VALS and make them direct. */
2262 static void
2263 ipcp_discover_new_direct_edges (struct cgraph_node *node,
2264 vec<tree> known_vals,
2265 struct ipa_agg_replacement_value *aggvals)
2267 struct cgraph_edge *ie, *next_ie;
2268 bool found = false;
2270 for (ie = node->indirect_calls; ie; ie = next_ie)
2272 tree target;
2274 next_ie = ie->next_callee;
2275 target = ipa_get_indirect_edge_target_1 (ie, known_vals, vNULL, vNULL,
2276 aggvals);
2277 if (target)
2279 ipa_make_edge_direct_to_target (ie, target);
2280 found = true;
2283 /* Turning calls to direct calls will improve overall summary. */
2284 if (found)
2285 inline_update_overall_summary (node);
2288 /* Vector of pointers which for linked lists of clones of an original crgaph
2289 edge. */
2291 static vec<cgraph_edge_p> next_edge_clone;
2293 static inline void
2294 grow_next_edge_clone_vector (void)
2296 if (next_edge_clone.length ()
2297 <= (unsigned) cgraph_edge_max_uid)
2298 next_edge_clone.safe_grow_cleared (cgraph_edge_max_uid + 1);
2301 /* Edge duplication hook to grow the appropriate linked list in
2302 next_edge_clone. */
2304 static void
2305 ipcp_edge_duplication_hook (struct cgraph_edge *src, struct cgraph_edge *dst,
2306 __attribute__((unused)) void *data)
2308 grow_next_edge_clone_vector ();
2309 next_edge_clone[dst->uid] = next_edge_clone[src->uid];
2310 next_edge_clone[src->uid] = dst;
2313 /* See if NODE is a clone with a known aggregate value at a given OFFSET of a
2314 parameter with the given INDEX. */
2316 static tree
2317 get_clone_agg_value (struct cgraph_node *node, HOST_WIDEST_INT offset,
2318 int index)
2320 struct ipa_agg_replacement_value *aggval;
2322 aggval = ipa_get_agg_replacements_for_node (node);
2323 while (aggval)
2325 if (aggval->offset == offset
2326 && aggval->index == index)
2327 return aggval->value;
2328 aggval = aggval->next;
2330 return NULL_TREE;
2333 /* Return true if edge CS does bring about the value described by SRC. */
2335 static bool
2336 cgraph_edge_brings_value_p (struct cgraph_edge *cs,
2337 struct ipcp_value_source *src)
2339 struct ipa_node_params *caller_info = IPA_NODE_REF (cs->caller);
2340 struct ipa_node_params *dst_info = IPA_NODE_REF (cs->callee);
2342 if ((dst_info->ipcp_orig_node && !dst_info->is_all_contexts_clone)
2343 || caller_info->node_dead)
2344 return false;
2345 if (!src->val)
2346 return true;
2348 if (caller_info->ipcp_orig_node)
2350 tree t;
2351 if (src->offset == -1)
2352 t = caller_info->known_vals[src->index];
2353 else
2354 t = get_clone_agg_value (cs->caller, src->offset, src->index);
2355 return (t != NULL_TREE
2356 && values_equal_for_ipcp_p (src->val->value, t));
2358 else
2360 struct ipcp_agg_lattice *aglat;
2361 struct ipcp_param_lattices *plats = ipa_get_parm_lattices (caller_info,
2362 src->index);
2363 if (src->offset == -1)
2364 return (ipa_lat_is_single_const (&plats->itself)
2365 && values_equal_for_ipcp_p (src->val->value,
2366 plats->itself.values->value));
2367 else
2369 if (plats->aggs_bottom || plats->aggs_contain_variable)
2370 return false;
2371 for (aglat = plats->aggs; aglat; aglat = aglat->next)
2372 if (aglat->offset == src->offset)
2373 return (ipa_lat_is_single_const (aglat)
2374 && values_equal_for_ipcp_p (src->val->value,
2375 aglat->values->value));
2377 return false;
2381 /* Get the next clone in the linked list of clones of an edge. */
2383 static inline struct cgraph_edge *
2384 get_next_cgraph_edge_clone (struct cgraph_edge *cs)
2386 return next_edge_clone[cs->uid];
2389 /* Given VAL, iterate over all its sources and if they still hold, add their
2390 edge frequency and their number into *FREQUENCY and *CALLER_COUNT
2391 respectively. */
2393 static bool
2394 get_info_about_necessary_edges (struct ipcp_value *val, int *freq_sum,
2395 gcov_type *count_sum, int *caller_count)
2397 struct ipcp_value_source *src;
2398 int freq = 0, count = 0;
2399 gcov_type cnt = 0;
2400 bool hot = false;
2402 for (src = val->sources; src; src = src->next)
2404 struct cgraph_edge *cs = src->cs;
2405 while (cs)
2407 if (cgraph_edge_brings_value_p (cs, src))
2409 count++;
2410 freq += cs->frequency;
2411 cnt += cs->count;
2412 hot |= cgraph_maybe_hot_edge_p (cs);
2414 cs = get_next_cgraph_edge_clone (cs);
2418 *freq_sum = freq;
2419 *count_sum = cnt;
2420 *caller_count = count;
2421 return hot;
2424 /* Return a vector of incoming edges that do bring value VAL. It is assumed
2425 their number is known and equal to CALLER_COUNT. */
2427 static vec<cgraph_edge_p>
2428 gather_edges_for_value (struct ipcp_value *val, int caller_count)
2430 struct ipcp_value_source *src;
2431 vec<cgraph_edge_p> ret;
2433 ret.create (caller_count);
2434 for (src = val->sources; src; src = src->next)
2436 struct cgraph_edge *cs = src->cs;
2437 while (cs)
2439 if (cgraph_edge_brings_value_p (cs, src))
2440 ret.quick_push (cs);
2441 cs = get_next_cgraph_edge_clone (cs);
2445 return ret;
2448 /* Construct a replacement map for a know VALUE for a formal parameter PARAM.
2449 Return it or NULL if for some reason it cannot be created. */
2451 static struct ipa_replace_map *
2452 get_replacement_map (tree value, tree parm)
2454 tree req_type = TREE_TYPE (parm);
2455 struct ipa_replace_map *replace_map;
2457 if (!useless_type_conversion_p (req_type, TREE_TYPE (value)))
2459 if (fold_convertible_p (req_type, value))
2460 value = fold_build1 (NOP_EXPR, req_type, value);
2461 else if (TYPE_SIZE (req_type) == TYPE_SIZE (TREE_TYPE (value)))
2462 value = fold_build1 (VIEW_CONVERT_EXPR, req_type, value);
2463 else
2465 if (dump_file)
2467 fprintf (dump_file, " const ");
2468 print_generic_expr (dump_file, value, 0);
2469 fprintf (dump_file, " can't be converted to param ");
2470 print_generic_expr (dump_file, parm, 0);
2471 fprintf (dump_file, "\n");
2473 return NULL;
2477 replace_map = ggc_alloc_ipa_replace_map ();
2478 if (dump_file)
2480 fprintf (dump_file, " replacing param ");
2481 print_generic_expr (dump_file, parm, 0);
2482 fprintf (dump_file, " with const ");
2483 print_generic_expr (dump_file, value, 0);
2484 fprintf (dump_file, "\n");
2486 replace_map->old_tree = parm;
2487 replace_map->new_tree = value;
2488 replace_map->replace_p = true;
2489 replace_map->ref_p = false;
2491 return replace_map;
2494 /* Dump new profiling counts */
2496 static void
2497 dump_profile_updates (struct cgraph_node *orig_node,
2498 struct cgraph_node *new_node)
2500 struct cgraph_edge *cs;
2502 fprintf (dump_file, " setting count of the specialized node to "
2503 HOST_WIDE_INT_PRINT_DEC "\n", (HOST_WIDE_INT) new_node->count);
2504 for (cs = new_node->callees; cs ; cs = cs->next_callee)
2505 fprintf (dump_file, " edge to %s has count "
2506 HOST_WIDE_INT_PRINT_DEC "\n",
2507 cgraph_node_name (cs->callee), (HOST_WIDE_INT) cs->count);
2509 fprintf (dump_file, " setting count of the original node to "
2510 HOST_WIDE_INT_PRINT_DEC "\n", (HOST_WIDE_INT) orig_node->count);
2511 for (cs = orig_node->callees; cs ; cs = cs->next_callee)
2512 fprintf (dump_file, " edge to %s is left with "
2513 HOST_WIDE_INT_PRINT_DEC "\n",
2514 cgraph_node_name (cs->callee), (HOST_WIDE_INT) cs->count);
2517 /* After a specialized NEW_NODE version of ORIG_NODE has been created, update
2518 their profile information to reflect this. */
2520 static void
2521 update_profiling_info (struct cgraph_node *orig_node,
2522 struct cgraph_node *new_node)
2524 struct cgraph_edge *cs;
2525 struct caller_statistics stats;
2526 gcov_type new_sum, orig_sum;
2527 gcov_type remainder, orig_node_count = orig_node->count;
2529 if (orig_node_count == 0)
2530 return;
2532 init_caller_stats (&stats);
2533 cgraph_for_node_and_aliases (orig_node, gather_caller_stats, &stats, false);
2534 orig_sum = stats.count_sum;
2535 init_caller_stats (&stats);
2536 cgraph_for_node_and_aliases (new_node, gather_caller_stats, &stats, false);
2537 new_sum = stats.count_sum;
2539 if (orig_node_count < orig_sum + new_sum)
2541 if (dump_file)
2542 fprintf (dump_file, " Problem: node %s/%i has too low count "
2543 HOST_WIDE_INT_PRINT_DEC " while the sum of incoming "
2544 "counts is " HOST_WIDE_INT_PRINT_DEC "\n",
2545 cgraph_node_name (orig_node), orig_node->uid,
2546 (HOST_WIDE_INT) orig_node_count,
2547 (HOST_WIDE_INT) (orig_sum + new_sum));
2549 orig_node_count = (orig_sum + new_sum) * 12 / 10;
2550 if (dump_file)
2551 fprintf (dump_file, " proceeding by pretending it was "
2552 HOST_WIDE_INT_PRINT_DEC "\n",
2553 (HOST_WIDE_INT) orig_node_count);
2556 new_node->count = new_sum;
2557 remainder = orig_node_count - new_sum;
2558 orig_node->count = remainder;
2560 for (cs = new_node->callees; cs ; cs = cs->next_callee)
2561 if (cs->frequency)
2562 cs->count = apply_probability (cs->count,
2563 GCOV_COMPUTE_SCALE (new_sum,
2564 orig_node_count));
2565 else
2566 cs->count = 0;
2568 for (cs = orig_node->callees; cs ; cs = cs->next_callee)
2569 cs->count = apply_probability (cs->count,
2570 GCOV_COMPUTE_SCALE (remainder,
2571 orig_node_count));
2573 if (dump_file)
2574 dump_profile_updates (orig_node, new_node);
2577 /* Update the respective profile of specialized NEW_NODE and the original
2578 ORIG_NODE after additional edges with cumulative count sum REDIRECTED_SUM
2579 have been redirected to the specialized version. */
2581 static void
2582 update_specialized_profile (struct cgraph_node *new_node,
2583 struct cgraph_node *orig_node,
2584 gcov_type redirected_sum)
2586 struct cgraph_edge *cs;
2587 gcov_type new_node_count, orig_node_count = orig_node->count;
2589 if (dump_file)
2590 fprintf (dump_file, " the sum of counts of redirected edges is "
2591 HOST_WIDE_INT_PRINT_DEC "\n", (HOST_WIDE_INT) redirected_sum);
2592 if (orig_node_count == 0)
2593 return;
2595 gcc_assert (orig_node_count >= redirected_sum);
2597 new_node_count = new_node->count;
2598 new_node->count += redirected_sum;
2599 orig_node->count -= redirected_sum;
2601 for (cs = new_node->callees; cs ; cs = cs->next_callee)
2602 if (cs->frequency)
2603 cs->count += apply_probability (cs->count,
2604 GCOV_COMPUTE_SCALE (redirected_sum,
2605 new_node_count));
2606 else
2607 cs->count = 0;
2609 for (cs = orig_node->callees; cs ; cs = cs->next_callee)
2611 gcov_type dec = apply_probability (cs->count,
2612 GCOV_COMPUTE_SCALE (redirected_sum,
2613 orig_node_count));
2614 if (dec < cs->count)
2615 cs->count -= dec;
2616 else
2617 cs->count = 0;
2620 if (dump_file)
2621 dump_profile_updates (orig_node, new_node);
2624 /* Create a specialized version of NODE with known constants and types of
2625 parameters in KNOWN_VALS and redirect all edges in CALLERS to it. */
2627 static struct cgraph_node *
2628 create_specialized_node (struct cgraph_node *node,
2629 vec<tree> known_vals,
2630 struct ipa_agg_replacement_value *aggvals,
2631 vec<cgraph_edge_p> callers)
2633 struct ipa_node_params *new_info, *info = IPA_NODE_REF (node);
2634 vec<ipa_replace_map_p, va_gc> *replace_trees = NULL;
2635 struct cgraph_node *new_node;
2636 int i, count = ipa_get_param_count (info);
2637 bitmap args_to_skip;
2639 gcc_assert (!info->ipcp_orig_node);
2641 if (node->local.can_change_signature)
2643 args_to_skip = BITMAP_GGC_ALLOC ();
2644 for (i = 0; i < count; i++)
2646 tree t = known_vals[i];
2648 if ((t && TREE_CODE (t) != TREE_BINFO)
2649 || !ipa_is_param_used (info, i))
2650 bitmap_set_bit (args_to_skip, i);
2653 else
2655 args_to_skip = NULL;
2656 if (dump_file && (dump_flags & TDF_DETAILS))
2657 fprintf (dump_file, " cannot change function signature\n");
2660 for (i = 0; i < count ; i++)
2662 tree t = known_vals[i];
2663 if (t && TREE_CODE (t) != TREE_BINFO)
2665 struct ipa_replace_map *replace_map;
2667 replace_map = get_replacement_map (t, ipa_get_param (info, i));
2668 if (replace_map)
2669 vec_safe_push (replace_trees, replace_map);
2673 new_node = cgraph_create_virtual_clone (node, callers, replace_trees,
2674 args_to_skip, "constprop");
2675 ipa_set_node_agg_value_chain (new_node, aggvals);
2676 if (dump_file && (dump_flags & TDF_DETAILS))
2678 fprintf (dump_file, " the new node is %s/%i.\n",
2679 cgraph_node_name (new_node), new_node->uid);
2680 if (aggvals)
2681 ipa_dump_agg_replacement_values (dump_file, aggvals);
2683 gcc_checking_assert (ipa_node_params_vector.exists ()
2684 && (ipa_node_params_vector.length ()
2685 > (unsigned) cgraph_max_uid));
2686 update_profiling_info (node, new_node);
2687 new_info = IPA_NODE_REF (new_node);
2688 new_info->ipcp_orig_node = node;
2689 new_info->known_vals = known_vals;
2691 ipcp_discover_new_direct_edges (new_node, known_vals, aggvals);
2693 callers.release ();
2694 return new_node;
2697 /* Given a NODE, and a subset of its CALLERS, try to populate blanks slots in
2698 KNOWN_VALS with constants and types that are also known for all of the
2699 CALLERS. */
2701 static void
2702 find_more_scalar_values_for_callers_subset (struct cgraph_node *node,
2703 vec<tree> known_vals,
2704 vec<cgraph_edge_p> callers)
2706 struct ipa_node_params *info = IPA_NODE_REF (node);
2707 int i, count = ipa_get_param_count (info);
2709 for (i = 0; i < count ; i++)
2711 struct cgraph_edge *cs;
2712 tree newval = NULL_TREE;
2713 int j;
2715 if (ipa_get_scalar_lat (info, i)->bottom || known_vals[i])
2716 continue;
2718 FOR_EACH_VEC_ELT (callers, j, cs)
2720 struct ipa_jump_func *jump_func;
2721 tree t;
2723 if (i >= ipa_get_cs_argument_count (IPA_EDGE_REF (cs)))
2725 newval = NULL_TREE;
2726 break;
2728 jump_func = ipa_get_ith_jump_func (IPA_EDGE_REF (cs), i);
2729 t = ipa_value_from_jfunc (IPA_NODE_REF (cs->caller), jump_func);
2730 if (!t
2731 || (newval
2732 && !values_equal_for_ipcp_p (t, newval)))
2734 newval = NULL_TREE;
2735 break;
2737 else
2738 newval = t;
2741 if (newval)
2743 if (dump_file && (dump_flags & TDF_DETAILS))
2745 fprintf (dump_file, " adding an extra known scalar value ");
2746 print_ipcp_constant_value (dump_file, newval);
2747 fprintf (dump_file, " for parameter ");
2748 print_generic_expr (dump_file, ipa_get_param (info, i), 0);
2749 fprintf (dump_file, "\n");
2752 known_vals[i] = newval;
2757 /* Go through PLATS and create a vector of values consisting of values and
2758 offsets (minus OFFSET) of lattices that contain only a single value. */
2760 static vec<ipa_agg_jf_item_t>
2761 copy_plats_to_inter (struct ipcp_param_lattices *plats, HOST_WIDE_INT offset)
2763 vec<ipa_agg_jf_item_t> res = vNULL;
2765 if (!plats->aggs || plats->aggs_contain_variable || plats->aggs_bottom)
2766 return vNULL;
2768 for (struct ipcp_agg_lattice *aglat = plats->aggs; aglat; aglat = aglat->next)
2769 if (ipa_lat_is_single_const (aglat))
2771 struct ipa_agg_jf_item ti;
2772 ti.offset = aglat->offset - offset;
2773 ti.value = aglat->values->value;
2774 res.safe_push (ti);
2776 return res;
2779 /* Intersect all values in INTER with single value lattices in PLATS (while
2780 subtracting OFFSET). */
2782 static void
2783 intersect_with_plats (struct ipcp_param_lattices *plats,
2784 vec<ipa_agg_jf_item_t> *inter,
2785 HOST_WIDE_INT offset)
2787 struct ipcp_agg_lattice *aglat;
2788 struct ipa_agg_jf_item *item;
2789 int k;
2791 if (!plats->aggs || plats->aggs_contain_variable || plats->aggs_bottom)
2793 inter->release ();
2794 return;
2797 aglat = plats->aggs;
2798 FOR_EACH_VEC_ELT (*inter, k, item)
2800 bool found = false;
2801 if (!item->value)
2802 continue;
2803 while (aglat)
2805 if (aglat->offset - offset > item->offset)
2806 break;
2807 if (aglat->offset - offset == item->offset)
2809 gcc_checking_assert (item->value);
2810 if (values_equal_for_ipcp_p (item->value, aglat->values->value))
2811 found = true;
2812 break;
2814 aglat = aglat->next;
2816 if (!found)
2817 item->value = NULL_TREE;
2821 /* Copy agggregate replacement values of NODE (which is an IPA-CP clone) to the
2822 vector result while subtracting OFFSET from the individual value offsets. */
2824 static vec<ipa_agg_jf_item_t>
2825 agg_replacements_to_vector (struct cgraph_node *node, int index,
2826 HOST_WIDE_INT offset)
2828 struct ipa_agg_replacement_value *av;
2829 vec<ipa_agg_jf_item_t> res = vNULL;
2831 for (av = ipa_get_agg_replacements_for_node (node); av; av = av->next)
2832 if (av->index == index
2833 && (av->offset - offset) >= 0)
2835 struct ipa_agg_jf_item item;
2836 gcc_checking_assert (av->value);
2837 item.offset = av->offset - offset;
2838 item.value = av->value;
2839 res.safe_push (item);
2842 return res;
2845 /* Intersect all values in INTER with those that we have already scheduled to
2846 be replaced in parameter number INDEX of NODE, which is an IPA-CP clone
2847 (while subtracting OFFSET). */
2849 static void
2850 intersect_with_agg_replacements (struct cgraph_node *node, int index,
2851 vec<ipa_agg_jf_item_t> *inter,
2852 HOST_WIDE_INT offset)
2854 struct ipa_agg_replacement_value *srcvals;
2855 struct ipa_agg_jf_item *item;
2856 int i;
2858 srcvals = ipa_get_agg_replacements_for_node (node);
2859 if (!srcvals)
2861 inter->release ();
2862 return;
2865 FOR_EACH_VEC_ELT (*inter, i, item)
2867 struct ipa_agg_replacement_value *av;
2868 bool found = false;
2869 if (!item->value)
2870 continue;
2871 for (av = srcvals; av; av = av->next)
2873 gcc_checking_assert (av->value);
2874 if (av->index == index
2875 && av->offset - offset == item->offset)
2877 if (values_equal_for_ipcp_p (item->value, av->value))
2878 found = true;
2879 break;
2882 if (!found)
2883 item->value = NULL_TREE;
2887 /* Intersect values in INTER with aggregate values that come along edge CS to
2888 parameter number INDEX and return it. If INTER does not actually exist yet,
2889 copy all incoming values to it. If we determine we ended up with no values
2890 whatsoever, return a released vector. */
2892 static vec<ipa_agg_jf_item_t>
2893 intersect_aggregates_with_edge (struct cgraph_edge *cs, int index,
2894 vec<ipa_agg_jf_item_t> inter)
2896 struct ipa_jump_func *jfunc;
2897 jfunc = ipa_get_ith_jump_func (IPA_EDGE_REF (cs), index);
2898 if (jfunc->type == IPA_JF_PASS_THROUGH
2899 && ipa_get_jf_pass_through_operation (jfunc) == NOP_EXPR)
2901 struct ipa_node_params *caller_info = IPA_NODE_REF (cs->caller);
2902 int src_idx = ipa_get_jf_pass_through_formal_id (jfunc);
2904 if (caller_info->ipcp_orig_node)
2906 struct cgraph_node *orig_node = caller_info->ipcp_orig_node;
2907 struct ipcp_param_lattices *orig_plats;
2908 orig_plats = ipa_get_parm_lattices (IPA_NODE_REF (orig_node),
2909 src_idx);
2910 if (agg_pass_through_permissible_p (orig_plats, jfunc))
2912 if (!inter.exists ())
2913 inter = agg_replacements_to_vector (cs->caller, src_idx, 0);
2914 else
2915 intersect_with_agg_replacements (cs->caller, src_idx,
2916 &inter, 0);
2919 else
2921 struct ipcp_param_lattices *src_plats;
2922 src_plats = ipa_get_parm_lattices (caller_info, src_idx);
2923 if (agg_pass_through_permissible_p (src_plats, jfunc))
2925 /* Currently we do not produce clobber aggregate jump
2926 functions, adjust when we do. */
2927 gcc_checking_assert (!jfunc->agg.items);
2928 if (!inter.exists ())
2929 inter = copy_plats_to_inter (src_plats, 0);
2930 else
2931 intersect_with_plats (src_plats, &inter, 0);
2935 else if (jfunc->type == IPA_JF_ANCESTOR
2936 && ipa_get_jf_ancestor_agg_preserved (jfunc))
2938 struct ipa_node_params *caller_info = IPA_NODE_REF (cs->caller);
2939 int src_idx = ipa_get_jf_ancestor_formal_id (jfunc);
2940 struct ipcp_param_lattices *src_plats;
2941 HOST_WIDE_INT delta = ipa_get_jf_ancestor_offset (jfunc);
2943 if (caller_info->ipcp_orig_node)
2945 if (!inter.exists ())
2946 inter = agg_replacements_to_vector (cs->caller, src_idx, delta);
2947 else
2948 intersect_with_agg_replacements (cs->caller, src_idx, &inter,
2949 delta);
2951 else
2953 src_plats = ipa_get_parm_lattices (caller_info, src_idx);;
2954 /* Currently we do not produce clobber aggregate jump
2955 functions, adjust when we do. */
2956 gcc_checking_assert (!src_plats->aggs || !jfunc->agg.items);
2957 if (!inter.exists ())
2958 inter = copy_plats_to_inter (src_plats, delta);
2959 else
2960 intersect_with_plats (src_plats, &inter, delta);
2963 else if (jfunc->agg.items)
2965 struct ipa_agg_jf_item *item;
2966 int k;
2968 if (!inter.exists ())
2969 for (unsigned i = 0; i < jfunc->agg.items->length (); i++)
2970 inter.safe_push ((*jfunc->agg.items)[i]);
2971 else
2972 FOR_EACH_VEC_ELT (inter, k, item)
2974 int l = 0;
2975 bool found = false;;
2977 if (!item->value)
2978 continue;
2980 while ((unsigned) l < jfunc->agg.items->length ())
2982 struct ipa_agg_jf_item *ti;
2983 ti = &(*jfunc->agg.items)[l];
2984 if (ti->offset > item->offset)
2985 break;
2986 if (ti->offset == item->offset)
2988 gcc_checking_assert (ti->value);
2989 if (values_equal_for_ipcp_p (item->value,
2990 ti->value))
2991 found = true;
2992 break;
2994 l++;
2996 if (!found)
2997 item->value = NULL;
3000 else
3002 inter.release();
3003 return vec<ipa_agg_jf_item_t>();
3005 return inter;
3008 /* Look at edges in CALLERS and collect all known aggregate values that arrive
3009 from all of them. */
3011 static struct ipa_agg_replacement_value *
3012 find_aggregate_values_for_callers_subset (struct cgraph_node *node,
3013 vec<cgraph_edge_p> callers)
3015 struct ipa_node_params *dest_info = IPA_NODE_REF (node);
3016 struct ipa_agg_replacement_value *res = NULL;
3017 struct cgraph_edge *cs;
3018 int i, j, count = ipa_get_param_count (dest_info);
3020 FOR_EACH_VEC_ELT (callers, j, cs)
3022 int c = ipa_get_cs_argument_count (IPA_EDGE_REF (cs));
3023 if (c < count)
3024 count = c;
3027 for (i = 0; i < count ; i++)
3029 struct cgraph_edge *cs;
3030 vec<ipa_agg_jf_item_t> inter = vNULL;
3031 struct ipa_agg_jf_item *item;
3032 struct ipcp_param_lattices *plats = ipa_get_parm_lattices (dest_info, i);
3033 int j;
3035 /* Among other things, the following check should deal with all by_ref
3036 mismatches. */
3037 if (plats->aggs_bottom)
3038 continue;
3040 FOR_EACH_VEC_ELT (callers, j, cs)
3042 inter = intersect_aggregates_with_edge (cs, i, inter);
3044 if (!inter.exists ())
3045 goto next_param;
3048 FOR_EACH_VEC_ELT (inter, j, item)
3050 struct ipa_agg_replacement_value *v;
3052 if (!item->value)
3053 continue;
3055 v = ggc_alloc_ipa_agg_replacement_value ();
3056 v->index = i;
3057 v->offset = item->offset;
3058 v->value = item->value;
3059 v->by_ref = plats->aggs_by_ref;
3060 v->next = res;
3061 res = v;
3064 next_param:
3065 if (inter.exists ())
3066 inter.release ();
3068 return res;
3071 /* Turn KNOWN_AGGS into a list of aggreate replacement values. */
3073 static struct ipa_agg_replacement_value *
3074 known_aggs_to_agg_replacement_list (vec<ipa_agg_jump_function_t> known_aggs)
3076 struct ipa_agg_replacement_value *res = NULL;
3077 struct ipa_agg_jump_function *aggjf;
3078 struct ipa_agg_jf_item *item;
3079 int i, j;
3081 FOR_EACH_VEC_ELT (known_aggs, i, aggjf)
3082 FOR_EACH_VEC_SAFE_ELT (aggjf->items, j, item)
3084 struct ipa_agg_replacement_value *v;
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 = aggjf->by_ref;
3090 v->next = res;
3091 res = v;
3093 return res;
3096 /* Determine whether CS also brings all scalar values that the NODE is
3097 specialized for. */
3099 static bool
3100 cgraph_edge_brings_all_scalars_for_node (struct cgraph_edge *cs,
3101 struct cgraph_node *node)
3103 struct ipa_node_params *dest_info = IPA_NODE_REF (node);
3104 int count = ipa_get_param_count (dest_info);
3105 struct ipa_node_params *caller_info;
3106 struct ipa_edge_args *args;
3107 int i;
3109 caller_info = IPA_NODE_REF (cs->caller);
3110 args = IPA_EDGE_REF (cs);
3111 for (i = 0; i < count; i++)
3113 struct ipa_jump_func *jump_func;
3114 tree val, t;
3116 val = dest_info->known_vals[i];
3117 if (!val)
3118 continue;
3120 if (i >= ipa_get_cs_argument_count (args))
3121 return false;
3122 jump_func = ipa_get_ith_jump_func (args, i);
3123 t = ipa_value_from_jfunc (caller_info, jump_func);
3124 if (!t || !values_equal_for_ipcp_p (val, t))
3125 return false;
3127 return true;
3130 /* Determine whether CS also brings all aggregate values that NODE is
3131 specialized for. */
3132 static bool
3133 cgraph_edge_brings_all_agg_vals_for_node (struct cgraph_edge *cs,
3134 struct cgraph_node *node)
3136 struct ipa_node_params *orig_caller_info = IPA_NODE_REF (cs->caller);
3137 struct ipa_agg_replacement_value *aggval;
3138 int i, ec, count;
3140 aggval = ipa_get_agg_replacements_for_node (node);
3141 if (!aggval)
3142 return true;
3144 count = ipa_get_param_count (IPA_NODE_REF (node));
3145 ec = ipa_get_cs_argument_count (IPA_EDGE_REF (cs));
3146 if (ec < count)
3147 for (struct ipa_agg_replacement_value *av = aggval; av; av = av->next)
3148 if (aggval->index >= ec)
3149 return false;
3151 if (orig_caller_info->ipcp_orig_node)
3152 orig_caller_info = IPA_NODE_REF (orig_caller_info->ipcp_orig_node);
3154 for (i = 0; i < count; i++)
3156 static vec<ipa_agg_jf_item_t> values = vec<ipa_agg_jf_item_t>();
3157 struct ipcp_param_lattices *plats;
3158 bool interesting = false;
3159 for (struct ipa_agg_replacement_value *av = aggval; av; av = av->next)
3160 if (aggval->index == i)
3162 interesting = true;
3163 break;
3165 if (!interesting)
3166 continue;
3168 plats = ipa_get_parm_lattices (orig_caller_info, aggval->index);
3169 if (plats->aggs_bottom)
3170 return false;
3172 values = intersect_aggregates_with_edge (cs, i, values);
3173 if (!values.exists())
3174 return false;
3176 for (struct ipa_agg_replacement_value *av = aggval; av; av = av->next)
3177 if (aggval->index == i)
3179 struct ipa_agg_jf_item *item;
3180 int j;
3181 bool found = false;
3182 FOR_EACH_VEC_ELT (values, j, item)
3183 if (item->value
3184 && item->offset == av->offset
3185 && values_equal_for_ipcp_p (item->value, av->value))
3186 found = true;
3187 if (!found)
3189 values.release();
3190 return false;
3194 return true;
3197 /* Given an original NODE and a VAL for which we have already created a
3198 specialized clone, look whether there are incoming edges that still lead
3199 into the old node but now also bring the requested value and also conform to
3200 all other criteria such that they can be redirected the the special node.
3201 This function can therefore redirect the final edge in a SCC. */
3203 static void
3204 perhaps_add_new_callers (struct cgraph_node *node, struct ipcp_value *val)
3206 struct ipcp_value_source *src;
3207 gcov_type redirected_sum = 0;
3209 for (src = val->sources; src; src = src->next)
3211 struct cgraph_edge *cs = src->cs;
3212 while (cs)
3214 enum availability availability;
3215 struct cgraph_node *dst = cgraph_function_node (cs->callee,
3216 &availability);
3217 if ((dst == node || IPA_NODE_REF (dst)->is_all_contexts_clone)
3218 && availability > AVAIL_OVERWRITABLE
3219 && cgraph_edge_brings_value_p (cs, src))
3221 if (cgraph_edge_brings_all_scalars_for_node (cs, val->spec_node)
3222 && cgraph_edge_brings_all_agg_vals_for_node (cs,
3223 val->spec_node))
3225 if (dump_file)
3226 fprintf (dump_file, " - adding an extra caller %s/%i"
3227 " of %s/%i\n",
3228 xstrdup (cgraph_node_name (cs->caller)),
3229 cs->caller->uid,
3230 xstrdup (cgraph_node_name (val->spec_node)),
3231 val->spec_node->uid);
3233 cgraph_redirect_edge_callee (cs, val->spec_node);
3234 redirected_sum += cs->count;
3237 cs = get_next_cgraph_edge_clone (cs);
3241 if (redirected_sum)
3242 update_specialized_profile (val->spec_node, node, redirected_sum);
3246 /* Copy KNOWN_BINFOS to KNOWN_VALS. */
3248 static void
3249 move_binfos_to_values (vec<tree> known_vals,
3250 vec<tree> known_binfos)
3252 tree t;
3253 int i;
3255 for (i = 0; known_binfos.iterate (i, &t); i++)
3256 if (t)
3257 known_vals[i] = t;
3260 /* Return true if there is a replacement equivalent to VALUE, INDEX and OFFSET
3261 among those in the AGGVALS list. */
3263 DEBUG_FUNCTION bool
3264 ipcp_val_in_agg_replacements_p (struct ipa_agg_replacement_value *aggvals,
3265 int index, HOST_WIDE_INT offset, tree value)
3267 while (aggvals)
3269 if (aggvals->index == index
3270 && aggvals->offset == offset
3271 && values_equal_for_ipcp_p (aggvals->value, value))
3272 return true;
3273 aggvals = aggvals->next;
3275 return false;
3278 /* Decide wheter to create a special version of NODE for value VAL of parameter
3279 at the given INDEX. If OFFSET is -1, the value is for the parameter itself,
3280 otherwise it is stored at the given OFFSET of the parameter. KNOWN_CSTS,
3281 KNOWN_BINFOS and KNOWN_AGGS describe the other already known values. */
3283 static bool
3284 decide_about_value (struct cgraph_node *node, int index, HOST_WIDE_INT offset,
3285 struct ipcp_value *val, vec<tree> known_csts,
3286 vec<tree> known_binfos)
3288 struct ipa_agg_replacement_value *aggvals;
3289 int freq_sum, caller_count;
3290 gcov_type count_sum;
3291 vec<cgraph_edge_p> callers;
3292 vec<tree> kv;
3294 if (val->spec_node)
3296 perhaps_add_new_callers (node, val);
3297 return false;
3299 else if (val->local_size_cost + overall_size > max_new_size)
3301 if (dump_file && (dump_flags & TDF_DETAILS))
3302 fprintf (dump_file, " Ignoring candidate value because "
3303 "max_new_size would be reached with %li.\n",
3304 val->local_size_cost + overall_size);
3305 return false;
3307 else if (!get_info_about_necessary_edges (val, &freq_sum, &count_sum,
3308 &caller_count))
3309 return false;
3311 if (dump_file && (dump_flags & TDF_DETAILS))
3313 fprintf (dump_file, " - considering value ");
3314 print_ipcp_constant_value (dump_file, val->value);
3315 fprintf (dump_file, " for parameter ");
3316 print_generic_expr (dump_file, ipa_get_param (IPA_NODE_REF (node),
3317 index), 0);
3318 if (offset != -1)
3319 fprintf (dump_file, ", offset: " HOST_WIDE_INT_PRINT_DEC, offset);
3320 fprintf (dump_file, " (caller_count: %i)\n", caller_count);
3323 if (!good_cloning_opportunity_p (node, val->local_time_benefit,
3324 freq_sum, count_sum,
3325 val->local_size_cost)
3326 && !good_cloning_opportunity_p (node,
3327 val->local_time_benefit
3328 + val->prop_time_benefit,
3329 freq_sum, count_sum,
3330 val->local_size_cost
3331 + val->prop_size_cost))
3332 return false;
3334 if (dump_file)
3335 fprintf (dump_file, " Creating a specialized node of %s/%i.\n",
3336 cgraph_node_name (node), node->uid);
3338 callers = gather_edges_for_value (val, caller_count);
3339 kv = known_csts.copy ();
3340 move_binfos_to_values (kv, known_binfos);
3341 if (offset == -1)
3342 kv[index] = val->value;
3343 find_more_scalar_values_for_callers_subset (node, kv, callers);
3344 aggvals = find_aggregate_values_for_callers_subset (node, callers);
3345 gcc_checking_assert (offset == -1
3346 || ipcp_val_in_agg_replacements_p (aggvals, index,
3347 offset, val->value));
3348 val->spec_node = create_specialized_node (node, kv, aggvals, callers);
3349 overall_size += val->local_size_cost;
3351 /* TODO: If for some lattice there is only one other known value
3352 left, make a special node for it too. */
3354 return true;
3357 /* Decide whether and what specialized clones of NODE should be created. */
3359 static bool
3360 decide_whether_version_node (struct cgraph_node *node)
3362 struct ipa_node_params *info = IPA_NODE_REF (node);
3363 int i, count = ipa_get_param_count (info);
3364 vec<tree> known_csts, known_binfos;
3365 vec<ipa_agg_jump_function_t> known_aggs = vNULL;
3366 bool ret = false;
3368 if (count == 0)
3369 return false;
3371 if (dump_file && (dump_flags & TDF_DETAILS))
3372 fprintf (dump_file, "\nEvaluating opportunities for %s/%i.\n",
3373 cgraph_node_name (node), node->uid);
3375 gather_context_independent_values (info, &known_csts, &known_binfos,
3376 info->do_clone_for_all_contexts ? &known_aggs
3377 : NULL, NULL);
3379 for (i = 0; i < count ;i++)
3381 struct ipcp_param_lattices *plats = ipa_get_parm_lattices (info, i);
3382 struct ipcp_lattice *lat = &plats->itself;
3383 struct ipcp_value *val;
3385 if (!lat->bottom
3386 && !known_csts[i]
3387 && !known_binfos[i])
3388 for (val = lat->values; val; val = val->next)
3389 ret |= decide_about_value (node, i, -1, val, known_csts,
3390 known_binfos);
3392 if (!plats->aggs_bottom)
3394 struct ipcp_agg_lattice *aglat;
3395 struct ipcp_value *val;
3396 for (aglat = plats->aggs; aglat; aglat = aglat->next)
3397 if (!aglat->bottom && aglat->values
3398 /* If the following is false, the one value is in
3399 known_aggs. */
3400 && (plats->aggs_contain_variable
3401 || !ipa_lat_is_single_const (aglat)))
3402 for (val = aglat->values; val; val = val->next)
3403 ret |= decide_about_value (node, i, aglat->offset, val,
3404 known_csts, known_binfos);
3406 info = IPA_NODE_REF (node);
3409 if (info->do_clone_for_all_contexts)
3411 struct cgraph_node *clone;
3412 vec<cgraph_edge_p> callers;
3414 if (dump_file)
3415 fprintf (dump_file, " - Creating a specialized node of %s/%i "
3416 "for all known contexts.\n", cgraph_node_name (node),
3417 node->uid);
3419 callers = collect_callers_of_node (node);
3420 move_binfos_to_values (known_csts, known_binfos);
3421 clone = create_specialized_node (node, known_csts,
3422 known_aggs_to_agg_replacement_list (known_aggs),
3423 callers);
3424 info = IPA_NODE_REF (node);
3425 info->do_clone_for_all_contexts = false;
3426 IPA_NODE_REF (clone)->is_all_contexts_clone = true;
3427 for (i = 0; i < count ; i++)
3428 vec_free (known_aggs[i].items);
3429 known_aggs.release ();
3430 ret = true;
3432 else
3433 known_csts.release ();
3435 known_binfos.release ();
3436 return ret;
3439 /* Transitively mark all callees of NODE within the same SCC as not dead. */
3441 static void
3442 spread_undeadness (struct cgraph_node *node)
3444 struct cgraph_edge *cs;
3446 for (cs = node->callees; cs; cs = cs->next_callee)
3447 if (edge_within_scc (cs))
3449 struct cgraph_node *callee;
3450 struct ipa_node_params *info;
3452 callee = cgraph_function_node (cs->callee, NULL);
3453 info = IPA_NODE_REF (callee);
3455 if (info->node_dead)
3457 info->node_dead = 0;
3458 spread_undeadness (callee);
3463 /* Return true if NODE has a caller from outside of its SCC that is not
3464 dead. Worker callback for cgraph_for_node_and_aliases. */
3466 static bool
3467 has_undead_caller_from_outside_scc_p (struct cgraph_node *node,
3468 void *data ATTRIBUTE_UNUSED)
3470 struct cgraph_edge *cs;
3472 for (cs = node->callers; cs; cs = cs->next_caller)
3473 if (cs->caller->thunk.thunk_p
3474 && cgraph_for_node_and_aliases (cs->caller,
3475 has_undead_caller_from_outside_scc_p,
3476 NULL, true))
3477 return true;
3478 else if (!edge_within_scc (cs)
3479 && !IPA_NODE_REF (cs->caller)->node_dead)
3480 return true;
3481 return false;
3485 /* Identify nodes within the same SCC as NODE which are no longer needed
3486 because of new clones and will be removed as unreachable. */
3488 static void
3489 identify_dead_nodes (struct cgraph_node *node)
3491 struct cgraph_node *v;
3492 for (v = node; v ; v = ((struct ipa_dfs_info *) v->symbol.aux)->next_cycle)
3493 if (cgraph_will_be_removed_from_program_if_no_direct_calls (v)
3494 && !cgraph_for_node_and_aliases (v,
3495 has_undead_caller_from_outside_scc_p,
3496 NULL, true))
3497 IPA_NODE_REF (v)->node_dead = 1;
3499 for (v = node; v ; v = ((struct ipa_dfs_info *) v->symbol.aux)->next_cycle)
3500 if (!IPA_NODE_REF (v)->node_dead)
3501 spread_undeadness (v);
3503 if (dump_file && (dump_flags & TDF_DETAILS))
3505 for (v = node; v ; v = ((struct ipa_dfs_info *) v->symbol.aux)->next_cycle)
3506 if (IPA_NODE_REF (v)->node_dead)
3507 fprintf (dump_file, " Marking node as dead: %s/%i.\n",
3508 cgraph_node_name (v), v->uid);
3512 /* The decision stage. Iterate over the topological order of call graph nodes
3513 TOPO and make specialized clones if deemed beneficial. */
3515 static void
3516 ipcp_decision_stage (struct topo_info *topo)
3518 int i;
3520 if (dump_file)
3521 fprintf (dump_file, "\nIPA decision stage:\n\n");
3523 for (i = topo->nnodes - 1; i >= 0; i--)
3525 struct cgraph_node *node = topo->order[i];
3526 bool change = false, iterate = true;
3528 while (iterate)
3530 struct cgraph_node *v;
3531 iterate = false;
3532 for (v = node; v ; v = ((struct ipa_dfs_info *) v->symbol.aux)->next_cycle)
3533 if (cgraph_function_with_gimple_body_p (v)
3534 && ipcp_versionable_function_p (v))
3535 iterate |= decide_whether_version_node (v);
3537 change |= iterate;
3539 if (change)
3540 identify_dead_nodes (node);
3544 /* The IPCP driver. */
3546 static unsigned int
3547 ipcp_driver (void)
3549 struct cgraph_2edge_hook_list *edge_duplication_hook_holder;
3550 struct topo_info topo;
3552 ipa_check_create_node_params ();
3553 ipa_check_create_edge_args ();
3554 grow_next_edge_clone_vector ();
3555 edge_duplication_hook_holder =
3556 cgraph_add_edge_duplication_hook (&ipcp_edge_duplication_hook, NULL);
3557 ipcp_values_pool = create_alloc_pool ("IPA-CP values",
3558 sizeof (struct ipcp_value), 32);
3559 ipcp_sources_pool = create_alloc_pool ("IPA-CP value sources",
3560 sizeof (struct ipcp_value_source), 64);
3561 ipcp_agg_lattice_pool = create_alloc_pool ("IPA_CP aggregate lattices",
3562 sizeof (struct ipcp_agg_lattice),
3563 32);
3564 if (dump_file)
3566 fprintf (dump_file, "\nIPA structures before propagation:\n");
3567 if (dump_flags & TDF_DETAILS)
3568 ipa_print_all_params (dump_file);
3569 ipa_print_all_jump_functions (dump_file);
3572 /* Topological sort. */
3573 build_toporder_info (&topo);
3574 /* Do the interprocedural propagation. */
3575 ipcp_propagate_stage (&topo);
3576 /* Decide what constant propagation and cloning should be performed. */
3577 ipcp_decision_stage (&topo);
3579 /* Free all IPCP structures. */
3580 free_toporder_info (&topo);
3581 next_edge_clone.release ();
3582 cgraph_remove_edge_duplication_hook (edge_duplication_hook_holder);
3583 ipa_free_all_structures_after_ipa_cp ();
3584 if (dump_file)
3585 fprintf (dump_file, "\nIPA constant propagation end\n");
3586 return 0;
3589 /* Initialization and computation of IPCP data structures. This is the initial
3590 intraprocedural analysis of functions, which gathers information to be
3591 propagated later on. */
3593 static void
3594 ipcp_generate_summary (void)
3596 struct cgraph_node *node;
3598 if (dump_file)
3599 fprintf (dump_file, "\nIPA constant propagation start:\n");
3600 ipa_register_cgraph_hooks ();
3602 FOR_EACH_FUNCTION_WITH_GIMPLE_BODY (node)
3604 node->local.versionable
3605 = tree_versionable_function_p (node->symbol.decl);
3606 ipa_analyze_node (node);
3610 /* Write ipcp summary for nodes in SET. */
3612 static void
3613 ipcp_write_summary (void)
3615 ipa_prop_write_jump_functions ();
3618 /* Read ipcp summary. */
3620 static void
3621 ipcp_read_summary (void)
3623 ipa_prop_read_jump_functions ();
3626 /* Gate for IPCP optimization. */
3628 static bool
3629 cgraph_gate_cp (void)
3631 /* FIXME: We should remove the optimize check after we ensure we never run
3632 IPA passes when not optimizing. */
3633 return flag_ipa_cp && optimize;
3636 struct ipa_opt_pass_d pass_ipa_cp =
3639 IPA_PASS,
3640 "cp", /* name */
3641 OPTGROUP_NONE, /* optinfo_flags */
3642 cgraph_gate_cp, /* gate */
3643 ipcp_driver, /* execute */
3644 NULL, /* sub */
3645 NULL, /* next */
3646 0, /* static_pass_number */
3647 TV_IPA_CONSTANT_PROP, /* tv_id */
3648 0, /* properties_required */
3649 0, /* properties_provided */
3650 0, /* properties_destroyed */
3651 0, /* todo_flags_start */
3652 TODO_dump_symtab |
3653 TODO_remove_functions /* todo_flags_finish */
3655 ipcp_generate_summary, /* generate_summary */
3656 ipcp_write_summary, /* write_summary */
3657 ipcp_read_summary, /* read_summary */
3658 ipa_prop_write_all_agg_replacement, /* write_optimization_summary */
3659 ipa_prop_read_all_agg_replacement, /* read_optimization_summary */
3660 NULL, /* stmt_fixup */
3661 0, /* TODOs */
3662 ipcp_transform_function, /* function_transform */
3663 NULL, /* variable_transform */