re PR tree-optimization/37021 (Fortran Complex reduction / multiplication not vectorized)
[official-gcc.git] / gcc / ipa-cp.c
blobaac21ceb6fa5a8267ef525653928deefb1058c23
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 /* Extract the acual BINFO being described by JFUNC which must be a known type
795 jump function. */
797 static tree
798 ipa_value_from_known_type_jfunc (struct ipa_jump_func *jfunc)
800 tree base_binfo = TYPE_BINFO (ipa_get_jf_known_type_base_type (jfunc));
801 if (!base_binfo)
802 return NULL_TREE;
803 return get_binfo_at_offset (base_binfo,
804 ipa_get_jf_known_type_offset (jfunc),
805 ipa_get_jf_known_type_component_type (jfunc));
808 /* Determine whether JFUNC evaluates to a known value (that is either a
809 constant or a binfo) and if so, return it. Otherwise return NULL. INFO
810 describes the caller node so that pass-through jump functions can be
811 evaluated. */
813 tree
814 ipa_value_from_jfunc (struct ipa_node_params *info, struct ipa_jump_func *jfunc)
816 if (jfunc->type == IPA_JF_CONST)
817 return ipa_get_jf_constant (jfunc);
818 else if (jfunc->type == IPA_JF_KNOWN_TYPE)
819 return ipa_value_from_known_type_jfunc (jfunc);
820 else if (jfunc->type == IPA_JF_PASS_THROUGH
821 || jfunc->type == IPA_JF_ANCESTOR)
823 tree input;
824 int idx;
826 if (jfunc->type == IPA_JF_PASS_THROUGH)
827 idx = ipa_get_jf_pass_through_formal_id (jfunc);
828 else
829 idx = ipa_get_jf_ancestor_formal_id (jfunc);
831 if (info->ipcp_orig_node)
832 input = info->known_vals[idx];
833 else
835 struct ipcp_lattice *lat;
837 if (!info->lattices)
839 gcc_checking_assert (!flag_ipa_cp);
840 return NULL_TREE;
842 lat = ipa_get_scalar_lat (info, idx);
843 if (!ipa_lat_is_single_const (lat))
844 return NULL_TREE;
845 input = lat->values->value;
848 if (!input)
849 return NULL_TREE;
851 if (jfunc->type == IPA_JF_PASS_THROUGH)
852 return ipa_get_jf_pass_through_result (jfunc, input);
853 else
854 return ipa_get_jf_ancestor_result (jfunc, input);
856 else
857 return NULL_TREE;
861 /* If checking is enabled, verify that no lattice is in the TOP state, i.e. not
862 bottom, not containing a variable component and without any known value at
863 the same time. */
865 DEBUG_FUNCTION void
866 ipcp_verify_propagated_values (void)
868 struct cgraph_node *node;
870 FOR_EACH_FUNCTION_WITH_GIMPLE_BODY (node)
872 struct ipa_node_params *info = IPA_NODE_REF (node);
873 int i, count = ipa_get_param_count (info);
875 for (i = 0; i < count; i++)
877 struct ipcp_lattice *lat = ipa_get_scalar_lat (info, i);
879 if (!lat->bottom
880 && !lat->contains_variable
881 && lat->values_count == 0)
883 if (dump_file)
885 fprintf (dump_file, "\nIPA lattices after constant "
886 "propagation:\n");
887 print_all_lattices (dump_file, true, false);
890 gcc_unreachable ();
896 /* Return true iff X and Y should be considered equal values by IPA-CP. */
898 static bool
899 values_equal_for_ipcp_p (tree x, tree y)
901 gcc_checking_assert (x != NULL_TREE && y != NULL_TREE);
903 if (x == y)
904 return true;
906 if (TREE_CODE (x) == TREE_BINFO || TREE_CODE (y) == TREE_BINFO)
907 return false;
909 if (TREE_CODE (x) == ADDR_EXPR
910 && TREE_CODE (y) == ADDR_EXPR
911 && TREE_CODE (TREE_OPERAND (x, 0)) == CONST_DECL
912 && TREE_CODE (TREE_OPERAND (y, 0)) == CONST_DECL)
913 return operand_equal_p (DECL_INITIAL (TREE_OPERAND (x, 0)),
914 DECL_INITIAL (TREE_OPERAND (y, 0)), 0);
915 else
916 return operand_equal_p (x, y, 0);
919 /* Add a new value source to VAL, marking that a value comes from edge CS and
920 (if the underlying jump function is a pass-through or an ancestor one) from
921 a caller value SRC_VAL of a caller parameter described by SRC_INDEX. OFFSET
922 is negative if the source was the scalar value of the parameter itself or
923 the offset within an aggregate. */
925 static void
926 add_value_source (struct ipcp_value *val, struct cgraph_edge *cs,
927 struct ipcp_value *src_val, int src_idx, HOST_WIDE_INT offset)
929 struct ipcp_value_source *src;
931 src = (struct ipcp_value_source *) pool_alloc (ipcp_sources_pool);
932 src->offset = offset;
933 src->cs = cs;
934 src->val = src_val;
935 src->index = src_idx;
937 src->next = val->sources;
938 val->sources = src;
941 /* Try to add NEWVAL to LAT, potentially creating a new struct ipcp_value for
942 it. CS, SRC_VAL SRC_INDEX and OFFSET are meant for add_value_source and
943 have the same meaning. */
945 static bool
946 add_value_to_lattice (struct ipcp_lattice *lat, tree newval,
947 struct cgraph_edge *cs, struct ipcp_value *src_val,
948 int src_idx, HOST_WIDE_INT offset)
950 struct ipcp_value *val;
952 if (lat->bottom)
953 return false;
955 for (val = lat->values; val; val = val->next)
956 if (values_equal_for_ipcp_p (val->value, newval))
958 if (edge_within_scc (cs))
960 struct ipcp_value_source *s;
961 for (s = val->sources; s ; s = s->next)
962 if (s->cs == cs)
963 break;
964 if (s)
965 return false;
968 add_value_source (val, cs, src_val, src_idx, offset);
969 return false;
972 if (lat->values_count == PARAM_VALUE (PARAM_IPA_CP_VALUE_LIST_SIZE))
974 /* We can only free sources, not the values themselves, because sources
975 of other values in this this SCC might point to them. */
976 for (val = lat->values; val; val = val->next)
978 while (val->sources)
980 struct ipcp_value_source *src = val->sources;
981 val->sources = src->next;
982 pool_free (ipcp_sources_pool, src);
986 lat->values = NULL;
987 return set_lattice_to_bottom (lat);
990 lat->values_count++;
991 val = (struct ipcp_value *) pool_alloc (ipcp_values_pool);
992 memset (val, 0, sizeof (*val));
994 add_value_source (val, cs, src_val, src_idx, offset);
995 val->value = newval;
996 val->next = lat->values;
997 lat->values = val;
998 return true;
1001 /* Like above but passes a special value of offset to distinguish that the
1002 origin is the scalar value of the parameter rather than a part of an
1003 aggregate. */
1005 static inline bool
1006 add_scalar_value_to_lattice (struct ipcp_lattice *lat, tree newval,
1007 struct cgraph_edge *cs,
1008 struct ipcp_value *src_val, int src_idx)
1010 return add_value_to_lattice (lat, newval, cs, src_val, src_idx, -1);
1013 /* Propagate values through a pass-through jump function JFUNC associated with
1014 edge CS, taking values from SRC_LAT and putting them into DEST_LAT. SRC_IDX
1015 is the index of the source parameter. */
1017 static bool
1018 propagate_vals_accross_pass_through (struct cgraph_edge *cs,
1019 struct ipa_jump_func *jfunc,
1020 struct ipcp_lattice *src_lat,
1021 struct ipcp_lattice *dest_lat,
1022 int src_idx)
1024 struct ipcp_value *src_val;
1025 bool ret = false;
1027 if (ipa_get_jf_pass_through_operation (jfunc) == NOP_EXPR)
1028 for (src_val = src_lat->values; src_val; src_val = src_val->next)
1029 ret |= add_scalar_value_to_lattice (dest_lat, src_val->value, cs,
1030 src_val, src_idx);
1031 /* Do not create new values when propagating within an SCC because if there
1032 are arithmetic functions with circular dependencies, there is infinite
1033 number of them and we would just make lattices bottom. */
1034 else if (edge_within_scc (cs))
1035 ret = set_lattice_contains_variable (dest_lat);
1036 else
1037 for (src_val = src_lat->values; src_val; src_val = src_val->next)
1039 tree cstval = src_val->value;
1041 if (TREE_CODE (cstval) == TREE_BINFO)
1043 ret |= set_lattice_contains_variable (dest_lat);
1044 continue;
1046 cstval = ipa_get_jf_pass_through_result (jfunc, cstval);
1048 if (cstval)
1049 ret |= add_scalar_value_to_lattice (dest_lat, cstval, cs, src_val,
1050 src_idx);
1051 else
1052 ret |= set_lattice_contains_variable (dest_lat);
1055 return ret;
1058 /* Propagate values through an ancestor jump function JFUNC associated with
1059 edge CS, taking values from SRC_LAT and putting them into DEST_LAT. SRC_IDX
1060 is the index of the source parameter. */
1062 static bool
1063 propagate_vals_accross_ancestor (struct cgraph_edge *cs,
1064 struct ipa_jump_func *jfunc,
1065 struct ipcp_lattice *src_lat,
1066 struct ipcp_lattice *dest_lat,
1067 int src_idx)
1069 struct ipcp_value *src_val;
1070 bool ret = false;
1072 if (edge_within_scc (cs))
1073 return set_lattice_contains_variable (dest_lat);
1075 for (src_val = src_lat->values; src_val; src_val = src_val->next)
1077 tree t = ipa_get_jf_ancestor_result (jfunc, src_val->value);
1079 if (t)
1080 ret |= add_scalar_value_to_lattice (dest_lat, t, cs, src_val, src_idx);
1081 else
1082 ret |= set_lattice_contains_variable (dest_lat);
1085 return ret;
1088 /* Propagate scalar values across jump function JFUNC that is associated with
1089 edge CS and put the values into DEST_LAT. */
1091 static bool
1092 propagate_scalar_accross_jump_function (struct cgraph_edge *cs,
1093 struct ipa_jump_func *jfunc,
1094 struct ipcp_lattice *dest_lat)
1096 if (dest_lat->bottom)
1097 return false;
1099 if (jfunc->type == IPA_JF_CONST
1100 || jfunc->type == IPA_JF_KNOWN_TYPE)
1102 tree val;
1104 if (jfunc->type == IPA_JF_KNOWN_TYPE)
1106 val = ipa_value_from_known_type_jfunc (jfunc);
1107 if (!val)
1108 return set_lattice_contains_variable (dest_lat);
1110 else
1111 val = ipa_get_jf_constant (jfunc);
1112 return add_scalar_value_to_lattice (dest_lat, val, cs, NULL, 0);
1114 else if (jfunc->type == IPA_JF_PASS_THROUGH
1115 || jfunc->type == IPA_JF_ANCESTOR)
1117 struct ipa_node_params *caller_info = IPA_NODE_REF (cs->caller);
1118 struct ipcp_lattice *src_lat;
1119 int src_idx;
1120 bool ret;
1122 if (jfunc->type == IPA_JF_PASS_THROUGH)
1123 src_idx = ipa_get_jf_pass_through_formal_id (jfunc);
1124 else
1125 src_idx = ipa_get_jf_ancestor_formal_id (jfunc);
1127 src_lat = ipa_get_scalar_lat (caller_info, src_idx);
1128 if (src_lat->bottom)
1129 return set_lattice_contains_variable (dest_lat);
1131 /* If we would need to clone the caller and cannot, do not propagate. */
1132 if (!ipcp_versionable_function_p (cs->caller)
1133 && (src_lat->contains_variable
1134 || (src_lat->values_count > 1)))
1135 return set_lattice_contains_variable (dest_lat);
1137 if (jfunc->type == IPA_JF_PASS_THROUGH)
1138 ret = propagate_vals_accross_pass_through (cs, jfunc, src_lat,
1139 dest_lat, src_idx);
1140 else
1141 ret = propagate_vals_accross_ancestor (cs, jfunc, src_lat, dest_lat,
1142 src_idx);
1144 if (src_lat->contains_variable)
1145 ret |= set_lattice_contains_variable (dest_lat);
1147 return ret;
1150 /* TODO: We currently do not handle member method pointers in IPA-CP (we only
1151 use it for indirect inlining), we should propagate them too. */
1152 return set_lattice_contains_variable (dest_lat);
1155 /* If DEST_PLATS already has aggregate items, check that aggs_by_ref matches
1156 NEW_AGGS_BY_REF and if not, mark all aggs as bottoms and return true (in all
1157 other cases, return false). If there are no aggregate items, set
1158 aggs_by_ref to NEW_AGGS_BY_REF. */
1160 static bool
1161 set_check_aggs_by_ref (struct ipcp_param_lattices *dest_plats,
1162 bool new_aggs_by_ref)
1164 if (dest_plats->aggs)
1166 if (dest_plats->aggs_by_ref != new_aggs_by_ref)
1168 set_agg_lats_to_bottom (dest_plats);
1169 return true;
1172 else
1173 dest_plats->aggs_by_ref = new_aggs_by_ref;
1174 return false;
1177 /* Walk aggregate lattices in DEST_PLATS from ***AGLAT on, until ***aglat is an
1178 already existing lattice for the given OFFSET and SIZE, marking all skipped
1179 lattices as containing variable and checking for overlaps. If there is no
1180 already existing lattice for the OFFSET and VAL_SIZE, create one, initialize
1181 it with offset, size and contains_variable to PRE_EXISTING, and return true,
1182 unless there are too many already. If there are two many, return false. If
1183 there are overlaps turn whole DEST_PLATS to bottom and return false. If any
1184 skipped lattices were newly marked as containing variable, set *CHANGE to
1185 true. */
1187 static bool
1188 merge_agg_lats_step (struct ipcp_param_lattices *dest_plats,
1189 HOST_WIDE_INT offset, HOST_WIDE_INT val_size,
1190 struct ipcp_agg_lattice ***aglat,
1191 bool pre_existing, bool *change)
1193 gcc_checking_assert (offset >= 0);
1195 while (**aglat && (**aglat)->offset < offset)
1197 if ((**aglat)->offset + (**aglat)->size > offset)
1199 set_agg_lats_to_bottom (dest_plats);
1200 return false;
1202 *change |= set_lattice_contains_variable (**aglat);
1203 *aglat = &(**aglat)->next;
1206 if (**aglat && (**aglat)->offset == offset)
1208 if ((**aglat)->size != val_size
1209 || ((**aglat)->next
1210 && (**aglat)->next->offset < offset + val_size))
1212 set_agg_lats_to_bottom (dest_plats);
1213 return false;
1215 gcc_checking_assert (!(**aglat)->next
1216 || (**aglat)->next->offset >= offset + val_size);
1217 return true;
1219 else
1221 struct ipcp_agg_lattice *new_al;
1223 if (**aglat && (**aglat)->offset < offset + val_size)
1225 set_agg_lats_to_bottom (dest_plats);
1226 return false;
1228 if (dest_plats->aggs_count == PARAM_VALUE (PARAM_IPA_MAX_AGG_ITEMS))
1229 return false;
1230 dest_plats->aggs_count++;
1231 new_al = (struct ipcp_agg_lattice *) pool_alloc (ipcp_agg_lattice_pool);
1232 memset (new_al, 0, sizeof (*new_al));
1234 new_al->offset = offset;
1235 new_al->size = val_size;
1236 new_al->contains_variable = pre_existing;
1238 new_al->next = **aglat;
1239 **aglat = new_al;
1240 return true;
1244 /* Set all AGLAT and all other aggregate lattices reachable by next pointers as
1245 containing an unknown value. */
1247 static bool
1248 set_chain_of_aglats_contains_variable (struct ipcp_agg_lattice *aglat)
1250 bool ret = false;
1251 while (aglat)
1253 ret |= set_lattice_contains_variable (aglat);
1254 aglat = aglat->next;
1256 return ret;
1259 /* Merge existing aggregate lattices in SRC_PLATS to DEST_PLATS, subtracting
1260 DELTA_OFFSET. CS is the call graph edge and SRC_IDX the index of the source
1261 parameter used for lattice value sources. Return true if DEST_PLATS changed
1262 in any way. */
1264 static bool
1265 merge_aggregate_lattices (struct cgraph_edge *cs,
1266 struct ipcp_param_lattices *dest_plats,
1267 struct ipcp_param_lattices *src_plats,
1268 int src_idx, HOST_WIDE_INT offset_delta)
1270 bool pre_existing = dest_plats->aggs != NULL;
1271 struct ipcp_agg_lattice **dst_aglat;
1272 bool ret = false;
1274 if (set_check_aggs_by_ref (dest_plats, src_plats->aggs_by_ref))
1275 return true;
1276 if (src_plats->aggs_bottom)
1277 return set_agg_lats_contain_variable (dest_plats);
1278 if (src_plats->aggs_contain_variable)
1279 ret |= set_agg_lats_contain_variable (dest_plats);
1280 dst_aglat = &dest_plats->aggs;
1282 for (struct ipcp_agg_lattice *src_aglat = src_plats->aggs;
1283 src_aglat;
1284 src_aglat = src_aglat->next)
1286 HOST_WIDE_INT new_offset = src_aglat->offset - offset_delta;
1288 if (new_offset < 0)
1289 continue;
1290 if (merge_agg_lats_step (dest_plats, new_offset, src_aglat->size,
1291 &dst_aglat, pre_existing, &ret))
1293 struct ipcp_agg_lattice *new_al = *dst_aglat;
1295 dst_aglat = &(*dst_aglat)->next;
1296 if (src_aglat->bottom)
1298 ret |= set_lattice_contains_variable (new_al);
1299 continue;
1301 if (src_aglat->contains_variable)
1302 ret |= set_lattice_contains_variable (new_al);
1303 for (struct ipcp_value *val = src_aglat->values;
1304 val;
1305 val = val->next)
1306 ret |= add_value_to_lattice (new_al, val->value, cs, val, src_idx,
1307 src_aglat->offset);
1309 else if (dest_plats->aggs_bottom)
1310 return true;
1312 ret |= set_chain_of_aglats_contains_variable (*dst_aglat);
1313 return ret;
1316 /* Determine whether there is anything to propagate FROM SRC_PLATS through a
1317 pass-through JFUNC and if so, whether it has conform and conforms to the
1318 rules about propagating values passed by reference. */
1320 static bool
1321 agg_pass_through_permissible_p (struct ipcp_param_lattices *src_plats,
1322 struct ipa_jump_func *jfunc)
1324 return src_plats->aggs
1325 && (!src_plats->aggs_by_ref
1326 || ipa_get_jf_pass_through_agg_preserved (jfunc));
1329 /* Propagate scalar values across jump function JFUNC that is associated with
1330 edge CS and put the values into DEST_LAT. */
1332 static bool
1333 propagate_aggs_accross_jump_function (struct cgraph_edge *cs,
1334 struct ipa_jump_func *jfunc,
1335 struct ipcp_param_lattices *dest_plats)
1337 bool ret = false;
1339 if (dest_plats->aggs_bottom)
1340 return false;
1342 if (jfunc->type == IPA_JF_PASS_THROUGH
1343 && ipa_get_jf_pass_through_operation (jfunc) == NOP_EXPR)
1345 struct ipa_node_params *caller_info = IPA_NODE_REF (cs->caller);
1346 int src_idx = ipa_get_jf_pass_through_formal_id (jfunc);
1347 struct ipcp_param_lattices *src_plats;
1349 src_plats = ipa_get_parm_lattices (caller_info, src_idx);
1350 if (agg_pass_through_permissible_p (src_plats, jfunc))
1352 /* Currently we do not produce clobber aggregate jump
1353 functions, replace with merging when we do. */
1354 gcc_assert (!jfunc->agg.items);
1355 ret |= merge_aggregate_lattices (cs, dest_plats, src_plats,
1356 src_idx, 0);
1358 else
1359 ret |= set_agg_lats_contain_variable (dest_plats);
1361 else if (jfunc->type == IPA_JF_ANCESTOR
1362 && ipa_get_jf_ancestor_agg_preserved (jfunc))
1364 struct ipa_node_params *caller_info = IPA_NODE_REF (cs->caller);
1365 int src_idx = ipa_get_jf_ancestor_formal_id (jfunc);
1366 struct ipcp_param_lattices *src_plats;
1368 src_plats = ipa_get_parm_lattices (caller_info, src_idx);
1369 if (src_plats->aggs && src_plats->aggs_by_ref)
1371 /* Currently we do not produce clobber aggregate jump
1372 functions, replace with merging when we do. */
1373 gcc_assert (!jfunc->agg.items);
1374 ret |= merge_aggregate_lattices (cs, dest_plats, src_plats, src_idx,
1375 ipa_get_jf_ancestor_offset (jfunc));
1377 else if (!src_plats->aggs_by_ref)
1378 ret |= set_agg_lats_to_bottom (dest_plats);
1379 else
1380 ret |= set_agg_lats_contain_variable (dest_plats);
1382 else if (jfunc->agg.items)
1384 bool pre_existing = dest_plats->aggs != NULL;
1385 struct ipcp_agg_lattice **aglat = &dest_plats->aggs;
1386 struct ipa_agg_jf_item *item;
1387 int i;
1389 if (set_check_aggs_by_ref (dest_plats, jfunc->agg.by_ref))
1390 return true;
1392 FOR_EACH_VEC_ELT (*jfunc->agg.items, i, item)
1394 HOST_WIDE_INT val_size;
1396 if (item->offset < 0)
1397 continue;
1398 gcc_checking_assert (is_gimple_ip_invariant (item->value));
1399 val_size = tree_low_cst (TYPE_SIZE (TREE_TYPE (item->value)), 1);
1401 if (merge_agg_lats_step (dest_plats, item->offset, val_size,
1402 &aglat, pre_existing, &ret))
1404 ret |= add_value_to_lattice (*aglat, item->value, cs, NULL, 0, 0);
1405 aglat = &(*aglat)->next;
1407 else if (dest_plats->aggs_bottom)
1408 return true;
1411 ret |= set_chain_of_aglats_contains_variable (*aglat);
1413 else
1414 ret |= set_agg_lats_contain_variable (dest_plats);
1416 return ret;
1419 /* Propagate constants from the caller to the callee of CS. INFO describes the
1420 caller. */
1422 static bool
1423 propagate_constants_accross_call (struct cgraph_edge *cs)
1425 struct ipa_node_params *callee_info;
1426 enum availability availability;
1427 struct cgraph_node *callee, *alias_or_thunk;
1428 struct ipa_edge_args *args;
1429 bool ret = false;
1430 int i, args_count, parms_count;
1432 callee = cgraph_function_node (cs->callee, &availability);
1433 if (!callee->analyzed)
1434 return false;
1435 gcc_checking_assert (cgraph_function_with_gimple_body_p (callee));
1436 callee_info = IPA_NODE_REF (callee);
1438 args = IPA_EDGE_REF (cs);
1439 args_count = ipa_get_cs_argument_count (args);
1440 parms_count = ipa_get_param_count (callee_info);
1442 /* If this call goes through a thunk we must not propagate to the first (0th)
1443 parameter. However, we might need to uncover a thunk from below a series
1444 of aliases first. */
1445 alias_or_thunk = cs->callee;
1446 while (alias_or_thunk->alias)
1447 alias_or_thunk = cgraph_alias_aliased_node (alias_or_thunk);
1448 if (alias_or_thunk->thunk.thunk_p)
1450 ret |= set_all_contains_variable (ipa_get_parm_lattices (callee_info,
1451 0));
1452 i = 1;
1454 else
1455 i = 0;
1457 for (; (i < args_count) && (i < parms_count); i++)
1459 struct ipa_jump_func *jump_func = ipa_get_ith_jump_func (args, i);
1460 struct ipcp_param_lattices *dest_plats;
1462 dest_plats = ipa_get_parm_lattices (callee_info, i);
1463 if (availability == AVAIL_OVERWRITABLE)
1464 ret |= set_all_contains_variable (dest_plats);
1465 else
1467 ret |= propagate_scalar_accross_jump_function (cs, jump_func,
1468 &dest_plats->itself);
1469 ret |= propagate_aggs_accross_jump_function (cs, jump_func,
1470 dest_plats);
1473 for (; i < parms_count; i++)
1474 ret |= set_all_contains_variable (ipa_get_parm_lattices (callee_info, i));
1476 return ret;
1479 /* If an indirect edge IE can be turned into a direct one based on KNOWN_VALS
1480 (which can contain both constants and binfos), KNOWN_BINFOS, KNOWN_AGGS or
1481 AGG_REPS return the destination. The latter three can be NULL. If AGG_REPS
1482 is not NULL, KNOWN_AGGS is ignored. */
1484 static tree
1485 ipa_get_indirect_edge_target_1 (struct cgraph_edge *ie,
1486 vec<tree> known_vals,
1487 vec<tree> known_binfos,
1488 vec<ipa_agg_jump_function_p> known_aggs,
1489 struct ipa_agg_replacement_value *agg_reps)
1491 int param_index = ie->indirect_info->param_index;
1492 HOST_WIDE_INT token, anc_offset;
1493 tree otr_type;
1494 tree t;
1496 if (param_index == -1)
1497 return NULL_TREE;
1499 if (!ie->indirect_info->polymorphic)
1501 tree t;
1503 if (ie->indirect_info->agg_contents)
1505 if (agg_reps)
1507 t = NULL;
1508 while (agg_reps)
1510 if (agg_reps->index == param_index
1511 && agg_reps->offset == ie->indirect_info->offset)
1513 t = agg_reps->value;
1514 break;
1516 agg_reps = agg_reps->next;
1519 else if (known_aggs.length () > (unsigned int) param_index)
1521 struct ipa_agg_jump_function *agg;
1522 agg = known_aggs[param_index];
1523 t = ipa_find_agg_cst_for_param (agg, ie->indirect_info->offset,
1524 ie->indirect_info->by_ref);
1526 else
1527 t = NULL;
1529 else
1530 t = (known_vals.length () > (unsigned int) param_index
1531 ? known_vals[param_index] : NULL);
1533 if (t &&
1534 TREE_CODE (t) == ADDR_EXPR
1535 && TREE_CODE (TREE_OPERAND (t, 0)) == FUNCTION_DECL)
1536 return TREE_OPERAND (t, 0);
1537 else
1538 return NULL_TREE;
1541 gcc_assert (!ie->indirect_info->agg_contents);
1542 token = ie->indirect_info->otr_token;
1543 anc_offset = ie->indirect_info->offset;
1544 otr_type = ie->indirect_info->otr_type;
1546 t = known_vals[param_index];
1547 if (!t && known_binfos.length () > (unsigned int) param_index)
1548 t = known_binfos[param_index];
1549 if (!t)
1550 return NULL_TREE;
1552 if (TREE_CODE (t) != TREE_BINFO)
1554 tree binfo;
1555 binfo = gimple_extract_devirt_binfo_from_cst (t);
1556 if (!binfo)
1557 return NULL_TREE;
1558 binfo = get_binfo_at_offset (binfo, anc_offset, otr_type);
1559 if (!binfo)
1560 return NULL_TREE;
1561 return gimple_get_virt_method_for_binfo (token, binfo);
1563 else
1565 tree binfo;
1567 binfo = get_binfo_at_offset (t, anc_offset, otr_type);
1568 if (!binfo)
1569 return NULL_TREE;
1570 return gimple_get_virt_method_for_binfo (token, binfo);
1575 /* If an indirect edge IE can be turned into a direct one based on KNOWN_VALS
1576 (which can contain both constants and binfos), KNOWN_BINFOS (which can be
1577 NULL) or KNOWN_AGGS (which also can be NULL) return the destination. */
1579 tree
1580 ipa_get_indirect_edge_target (struct cgraph_edge *ie,
1581 vec<tree> known_vals,
1582 vec<tree> known_binfos,
1583 vec<ipa_agg_jump_function_p> known_aggs)
1585 return ipa_get_indirect_edge_target_1 (ie, known_vals, known_binfos,
1586 known_aggs, NULL);
1589 /* Calculate devirtualization time bonus for NODE, assuming we know KNOWN_CSTS
1590 and KNOWN_BINFOS. */
1592 static int
1593 devirtualization_time_bonus (struct cgraph_node *node,
1594 vec<tree> known_csts,
1595 vec<tree> known_binfos,
1596 vec<ipa_agg_jump_function_p> known_aggs)
1598 struct cgraph_edge *ie;
1599 int res = 0;
1601 for (ie = node->indirect_calls; ie; ie = ie->next_callee)
1603 struct cgraph_node *callee;
1604 struct inline_summary *isummary;
1605 tree target;
1607 target = ipa_get_indirect_edge_target (ie, known_csts, known_binfos,
1608 known_aggs);
1609 if (!target)
1610 continue;
1612 /* Only bare minimum benefit for clearly un-inlineable targets. */
1613 res += 1;
1614 callee = cgraph_get_node (target);
1615 if (!callee || !callee->analyzed)
1616 continue;
1617 isummary = inline_summary (callee);
1618 if (!isummary->inlinable)
1619 continue;
1621 /* FIXME: The values below need re-considering and perhaps also
1622 integrating into the cost metrics, at lest in some very basic way. */
1623 if (isummary->size <= MAX_INLINE_INSNS_AUTO / 4)
1624 res += 31;
1625 else if (isummary->size <= MAX_INLINE_INSNS_AUTO / 2)
1626 res += 15;
1627 else if (isummary->size <= MAX_INLINE_INSNS_AUTO
1628 || DECL_DECLARED_INLINE_P (callee->symbol.decl))
1629 res += 7;
1632 return res;
1635 /* Return time bonus incurred because of HINTS. */
1637 static int
1638 hint_time_bonus (inline_hints hints)
1640 int result = 0;
1641 if (hints & (INLINE_HINT_loop_iterations | INLINE_HINT_loop_stride))
1642 result += PARAM_VALUE (PARAM_IPA_CP_LOOP_HINT_BONUS);
1643 if (hints & INLINE_HINT_array_index)
1644 result += PARAM_VALUE (PARAM_IPA_CP_ARRAY_INDEX_HINT_BONUS);
1645 return result;
1648 /* Return true if cloning NODE is a good idea, given the estimated TIME_BENEFIT
1649 and SIZE_COST and with the sum of frequencies of incoming edges to the
1650 potential new clone in FREQUENCIES. */
1652 static bool
1653 good_cloning_opportunity_p (struct cgraph_node *node, int time_benefit,
1654 int freq_sum, gcov_type count_sum, int size_cost)
1656 if (time_benefit == 0
1657 || !flag_ipa_cp_clone
1658 || !optimize_function_for_speed_p (DECL_STRUCT_FUNCTION (node->symbol.decl)))
1659 return false;
1661 gcc_assert (size_cost > 0);
1663 if (max_count)
1665 int factor = (count_sum * 1000) / max_count;
1666 HOST_WIDEST_INT evaluation = (((HOST_WIDEST_INT) time_benefit * factor)
1667 / size_cost);
1669 if (dump_file && (dump_flags & TDF_DETAILS))
1670 fprintf (dump_file, " good_cloning_opportunity_p (time: %i, "
1671 "size: %i, count_sum: " HOST_WIDE_INT_PRINT_DEC
1672 ") -> evaluation: " HOST_WIDEST_INT_PRINT_DEC
1673 ", threshold: %i\n",
1674 time_benefit, size_cost, (HOST_WIDE_INT) count_sum,
1675 evaluation, PARAM_VALUE (PARAM_IPA_CP_EVAL_THRESHOLD));
1677 return evaluation >= PARAM_VALUE (PARAM_IPA_CP_EVAL_THRESHOLD);
1679 else
1681 HOST_WIDEST_INT evaluation = (((HOST_WIDEST_INT) time_benefit * freq_sum)
1682 / size_cost);
1684 if (dump_file && (dump_flags & TDF_DETAILS))
1685 fprintf (dump_file, " good_cloning_opportunity_p (time: %i, "
1686 "size: %i, freq_sum: %i) -> evaluation: "
1687 HOST_WIDEST_INT_PRINT_DEC ", threshold: %i\n",
1688 time_benefit, size_cost, freq_sum, evaluation,
1689 PARAM_VALUE (PARAM_IPA_CP_EVAL_THRESHOLD));
1691 return evaluation >= PARAM_VALUE (PARAM_IPA_CP_EVAL_THRESHOLD);
1695 /* Return all context independent values from aggregate lattices in PLATS in a
1696 vector. Return NULL if there are none. */
1698 static vec<ipa_agg_jf_item_t, va_gc> *
1699 context_independent_aggregate_values (struct ipcp_param_lattices *plats)
1701 vec<ipa_agg_jf_item_t, va_gc> *res = NULL;
1703 if (plats->aggs_bottom
1704 || plats->aggs_contain_variable
1705 || plats->aggs_count == 0)
1706 return NULL;
1708 for (struct ipcp_agg_lattice *aglat = plats->aggs;
1709 aglat;
1710 aglat = aglat->next)
1711 if (ipa_lat_is_single_const (aglat))
1713 struct ipa_agg_jf_item item;
1714 item.offset = aglat->offset;
1715 item.value = aglat->values->value;
1716 vec_safe_push (res, item);
1718 return res;
1721 /* Allocate KNOWN_CSTS, KNOWN_BINFOS and, if non-NULL, KNOWN_AGGS and populate
1722 them with values of parameters that are known independent of the context.
1723 INFO describes the function. If REMOVABLE_PARAMS_COST is non-NULL, the
1724 movement cost of all removable parameters will be stored in it. */
1726 static bool
1727 gather_context_independent_values (struct ipa_node_params *info,
1728 vec<tree> *known_csts,
1729 vec<tree> *known_binfos,
1730 vec<ipa_agg_jump_function_t> *known_aggs,
1731 int *removable_params_cost)
1733 int i, count = ipa_get_param_count (info);
1734 bool ret = false;
1736 known_csts->create (0);
1737 known_binfos->create (0);
1738 known_csts->safe_grow_cleared (count);
1739 known_binfos->safe_grow_cleared (count);
1740 if (known_aggs)
1742 known_aggs->create (0);
1743 known_aggs->safe_grow_cleared (count);
1746 if (removable_params_cost)
1747 *removable_params_cost = 0;
1749 for (i = 0; i < count ; i++)
1751 struct ipcp_param_lattices *plats = ipa_get_parm_lattices (info, i);
1752 struct ipcp_lattice *lat = &plats->itself;
1754 if (ipa_lat_is_single_const (lat))
1756 struct ipcp_value *val = lat->values;
1757 if (TREE_CODE (val->value) != TREE_BINFO)
1759 (*known_csts)[i] = val->value;
1760 if (removable_params_cost)
1761 *removable_params_cost
1762 += estimate_move_cost (TREE_TYPE (val->value));
1763 ret = true;
1765 else if (plats->virt_call)
1767 (*known_binfos)[i] = val->value;
1768 ret = true;
1770 else if (removable_params_cost
1771 && !ipa_is_param_used (info, i))
1772 *removable_params_cost
1773 += estimate_move_cost (TREE_TYPE (ipa_get_param (info, i)));
1775 else if (removable_params_cost
1776 && !ipa_is_param_used (info, i))
1777 *removable_params_cost
1778 += estimate_move_cost (TREE_TYPE (ipa_get_param (info, i)));
1780 if (known_aggs)
1782 vec<ipa_agg_jf_item_t, va_gc> *agg_items;
1783 struct ipa_agg_jump_function *ajf;
1785 agg_items = context_independent_aggregate_values (plats);
1786 ajf = &(*known_aggs)[i];
1787 ajf->items = agg_items;
1788 ajf->by_ref = plats->aggs_by_ref;
1789 ret |= agg_items != NULL;
1793 return ret;
1796 /* The current interface in ipa-inline-analysis requires a pointer vector.
1797 Create it.
1799 FIXME: That interface should be re-worked, this is slightly silly. Still,
1800 I'd like to discuss how to change it first and this demonstrates the
1801 issue. */
1803 static vec<ipa_agg_jump_function_p>
1804 agg_jmp_p_vec_for_t_vec (vec<ipa_agg_jump_function_t> known_aggs)
1806 vec<ipa_agg_jump_function_p> ret;
1807 struct ipa_agg_jump_function *ajf;
1808 int i;
1810 ret.create (known_aggs.length ());
1811 FOR_EACH_VEC_ELT (known_aggs, i, ajf)
1812 ret.quick_push (ajf);
1813 return ret;
1816 /* Iterate over known values of parameters of NODE and estimate the local
1817 effects in terms of time and size they have. */
1819 static void
1820 estimate_local_effects (struct cgraph_node *node)
1822 struct ipa_node_params *info = IPA_NODE_REF (node);
1823 int i, count = ipa_get_param_count (info);
1824 vec<tree> known_csts, known_binfos;
1825 vec<ipa_agg_jump_function_t> known_aggs;
1826 vec<ipa_agg_jump_function_p> known_aggs_ptrs;
1827 bool always_const;
1828 int base_time = inline_summary (node)->time;
1829 int removable_params_cost;
1831 if (!count || !ipcp_versionable_function_p (node))
1832 return;
1834 if (dump_file && (dump_flags & TDF_DETAILS))
1835 fprintf (dump_file, "\nEstimating effects for %s/%i, base_time: %i.\n",
1836 cgraph_node_name (node), node->uid, base_time);
1838 always_const = gather_context_independent_values (info, &known_csts,
1839 &known_binfos, &known_aggs,
1840 &removable_params_cost);
1841 known_aggs_ptrs = agg_jmp_p_vec_for_t_vec (known_aggs);
1842 if (always_const)
1844 struct caller_statistics stats;
1845 inline_hints hints;
1846 int time, size;
1848 init_caller_stats (&stats);
1849 cgraph_for_node_and_aliases (node, gather_caller_stats, &stats, false);
1850 estimate_ipcp_clone_size_and_time (node, known_csts, known_binfos,
1851 known_aggs_ptrs, &size, &time, &hints);
1852 time -= devirtualization_time_bonus (node, known_csts, known_binfos,
1853 known_aggs_ptrs);
1854 time -= hint_time_bonus (hints);
1855 time -= removable_params_cost;
1856 size -= stats.n_calls * removable_params_cost;
1858 if (dump_file)
1859 fprintf (dump_file, " - context independent values, size: %i, "
1860 "time_benefit: %i\n", size, base_time - time);
1862 if (size <= 0
1863 || cgraph_will_be_removed_from_program_if_no_direct_calls (node))
1865 info->do_clone_for_all_contexts = true;
1866 base_time = time;
1868 if (dump_file)
1869 fprintf (dump_file, " Decided to specialize for all "
1870 "known contexts, code not going to grow.\n");
1872 else if (good_cloning_opportunity_p (node, base_time - time,
1873 stats.freq_sum, stats.count_sum,
1874 size))
1876 if (size + overall_size <= max_new_size)
1878 info->do_clone_for_all_contexts = true;
1879 base_time = time;
1880 overall_size += size;
1882 if (dump_file)
1883 fprintf (dump_file, " Decided to specialize for all "
1884 "known contexts, growth deemed beneficial.\n");
1886 else if (dump_file && (dump_flags & TDF_DETAILS))
1887 fprintf (dump_file, " Not cloning for all contexts because "
1888 "max_new_size would be reached with %li.\n",
1889 size + overall_size);
1893 for (i = 0; i < count ; i++)
1895 struct ipcp_param_lattices *plats = ipa_get_parm_lattices (info, i);
1896 struct ipcp_lattice *lat = &plats->itself;
1897 struct ipcp_value *val;
1898 int emc;
1900 if (lat->bottom
1901 || !lat->values
1902 || known_csts[i]
1903 || known_binfos[i])
1904 continue;
1906 for (val = lat->values; val; val = val->next)
1908 int time, size, time_benefit;
1909 inline_hints hints;
1911 if (TREE_CODE (val->value) != TREE_BINFO)
1913 known_csts[i] = val->value;
1914 known_binfos[i] = NULL_TREE;
1915 emc = estimate_move_cost (TREE_TYPE (val->value));
1917 else if (plats->virt_call)
1919 known_csts[i] = NULL_TREE;
1920 known_binfos[i] = val->value;
1921 emc = 0;
1923 else
1924 continue;
1926 estimate_ipcp_clone_size_and_time (node, known_csts, known_binfos,
1927 known_aggs_ptrs, &size, &time,
1928 &hints);
1929 time_benefit = base_time - time
1930 + devirtualization_time_bonus (node, known_csts, known_binfos,
1931 known_aggs_ptrs)
1932 + hint_time_bonus (hints)
1933 + removable_params_cost + emc;
1935 gcc_checking_assert (size >=0);
1936 /* The inliner-heuristics based estimates may think that in certain
1937 contexts some functions do not have any size at all but we want
1938 all specializations to have at least a tiny cost, not least not to
1939 divide by zero. */
1940 if (size == 0)
1941 size = 1;
1943 if (dump_file && (dump_flags & TDF_DETAILS))
1945 fprintf (dump_file, " - estimates for value ");
1946 print_ipcp_constant_value (dump_file, val->value);
1947 fprintf (dump_file, " for parameter ");
1948 print_generic_expr (dump_file, ipa_get_param (info, i), 0);
1949 fprintf (dump_file, ": time_benefit: %i, size: %i\n",
1950 time_benefit, size);
1953 val->local_time_benefit = time_benefit;
1954 val->local_size_cost = size;
1956 known_binfos[i] = NULL_TREE;
1957 known_csts[i] = NULL_TREE;
1960 for (i = 0; i < count ; i++)
1962 struct ipcp_param_lattices *plats = ipa_get_parm_lattices (info, i);
1963 struct ipa_agg_jump_function *ajf;
1964 struct ipcp_agg_lattice *aglat;
1966 if (plats->aggs_bottom || !plats->aggs)
1967 continue;
1969 ajf = &known_aggs[i];
1970 for (aglat = plats->aggs; aglat; aglat = aglat->next)
1972 struct ipcp_value *val;
1973 if (aglat->bottom || !aglat->values
1974 /* If the following is true, the one value is in known_aggs. */
1975 || (!plats->aggs_contain_variable
1976 && ipa_lat_is_single_const (aglat)))
1977 continue;
1979 for (val = aglat->values; val; val = val->next)
1981 int time, size, time_benefit;
1982 struct ipa_agg_jf_item item;
1983 inline_hints hints;
1985 item.offset = aglat->offset;
1986 item.value = val->value;
1987 vec_safe_push (ajf->items, item);
1989 estimate_ipcp_clone_size_and_time (node, known_csts, known_binfos,
1990 known_aggs_ptrs, &size, &time,
1991 &hints);
1992 time_benefit = base_time - time
1993 + devirtualization_time_bonus (node, known_csts, known_binfos,
1994 known_aggs_ptrs)
1995 + hint_time_bonus (hints);
1996 gcc_checking_assert (size >=0);
1997 if (size == 0)
1998 size = 1;
2000 if (dump_file && (dump_flags & TDF_DETAILS))
2002 fprintf (dump_file, " - estimates for value ");
2003 print_ipcp_constant_value (dump_file, val->value);
2004 fprintf (dump_file, " for parameter ");
2005 print_generic_expr (dump_file, ipa_get_param (info, i), 0);
2006 fprintf (dump_file, "[%soffset: " HOST_WIDE_INT_PRINT_DEC
2007 "]: time_benefit: %i, size: %i\n",
2008 plats->aggs_by_ref ? "ref " : "",
2009 aglat->offset, time_benefit, size);
2012 val->local_time_benefit = time_benefit;
2013 val->local_size_cost = size;
2014 ajf->items->pop ();
2019 for (i = 0; i < count ; i++)
2020 vec_free (known_aggs[i].items);
2022 known_csts.release ();
2023 known_binfos.release ();
2024 known_aggs.release ();
2025 known_aggs_ptrs.release ();
2029 /* Add value CUR_VAL and all yet-unsorted values it is dependent on to the
2030 topological sort of values. */
2032 static void
2033 add_val_to_toposort (struct ipcp_value *cur_val)
2035 static int dfs_counter = 0;
2036 static struct ipcp_value *stack;
2037 struct ipcp_value_source *src;
2039 if (cur_val->dfs)
2040 return;
2042 dfs_counter++;
2043 cur_val->dfs = dfs_counter;
2044 cur_val->low_link = dfs_counter;
2046 cur_val->topo_next = stack;
2047 stack = cur_val;
2048 cur_val->on_stack = true;
2050 for (src = cur_val->sources; src; src = src->next)
2051 if (src->val)
2053 if (src->val->dfs == 0)
2055 add_val_to_toposort (src->val);
2056 if (src->val->low_link < cur_val->low_link)
2057 cur_val->low_link = src->val->low_link;
2059 else if (src->val->on_stack
2060 && src->val->dfs < cur_val->low_link)
2061 cur_val->low_link = src->val->dfs;
2064 if (cur_val->dfs == cur_val->low_link)
2066 struct ipcp_value *v, *scc_list = NULL;
2070 v = stack;
2071 stack = v->topo_next;
2072 v->on_stack = false;
2074 v->scc_next = scc_list;
2075 scc_list = v;
2077 while (v != cur_val);
2079 cur_val->topo_next = values_topo;
2080 values_topo = cur_val;
2084 /* Add all values in lattices associated with NODE to the topological sort if
2085 they are not there yet. */
2087 static void
2088 add_all_node_vals_to_toposort (struct cgraph_node *node)
2090 struct ipa_node_params *info = IPA_NODE_REF (node);
2091 int i, count = ipa_get_param_count (info);
2093 for (i = 0; i < count ; i++)
2095 struct ipcp_param_lattices *plats = ipa_get_parm_lattices (info, i);
2096 struct ipcp_lattice *lat = &plats->itself;
2097 struct ipcp_agg_lattice *aglat;
2098 struct ipcp_value *val;
2100 if (!lat->bottom)
2101 for (val = lat->values; val; val = val->next)
2102 add_val_to_toposort (val);
2104 if (!plats->aggs_bottom)
2105 for (aglat = plats->aggs; aglat; aglat = aglat->next)
2106 if (!aglat->bottom)
2107 for (val = aglat->values; val; val = val->next)
2108 add_val_to_toposort (val);
2112 /* One pass of constants propagation along the call graph edges, from callers
2113 to callees (requires topological ordering in TOPO), iterate over strongly
2114 connected components. */
2116 static void
2117 propagate_constants_topo (struct topo_info *topo)
2119 int i;
2121 for (i = topo->nnodes - 1; i >= 0; i--)
2123 struct cgraph_node *v, *node = topo->order[i];
2124 struct ipa_dfs_info *node_dfs_info;
2126 if (!cgraph_function_with_gimple_body_p (node))
2127 continue;
2129 node_dfs_info = (struct ipa_dfs_info *) node->symbol.aux;
2130 /* First, iteratively propagate within the strongly connected component
2131 until all lattices stabilize. */
2132 v = node_dfs_info->next_cycle;
2133 while (v)
2135 push_node_to_stack (topo, v);
2136 v = ((struct ipa_dfs_info *) v->symbol.aux)->next_cycle;
2139 v = node;
2140 while (v)
2142 struct cgraph_edge *cs;
2144 for (cs = v->callees; cs; cs = cs->next_callee)
2145 if (edge_within_scc (cs)
2146 && propagate_constants_accross_call (cs))
2147 push_node_to_stack (topo, cs->callee);
2148 v = pop_node_from_stack (topo);
2151 /* Afterwards, propagate along edges leading out of the SCC, calculates
2152 the local effects of the discovered constants and all valid values to
2153 their topological sort. */
2154 v = node;
2155 while (v)
2157 struct cgraph_edge *cs;
2159 estimate_local_effects (v);
2160 add_all_node_vals_to_toposort (v);
2161 for (cs = v->callees; cs; cs = cs->next_callee)
2162 if (!edge_within_scc (cs))
2163 propagate_constants_accross_call (cs);
2165 v = ((struct ipa_dfs_info *) v->symbol.aux)->next_cycle;
2171 /* Return the sum of A and B if none of them is bigger than INT_MAX/2, return
2172 the bigger one if otherwise. */
2174 static int
2175 safe_add (int a, int b)
2177 if (a > INT_MAX/2 || b > INT_MAX/2)
2178 return a > b ? a : b;
2179 else
2180 return a + b;
2184 /* Propagate the estimated effects of individual values along the topological
2185 from the dependent values to those they depend on. */
2187 static void
2188 propagate_effects (void)
2190 struct ipcp_value *base;
2192 for (base = values_topo; base; base = base->topo_next)
2194 struct ipcp_value_source *src;
2195 struct ipcp_value *val;
2196 int time = 0, size = 0;
2198 for (val = base; val; val = val->scc_next)
2200 time = safe_add (time,
2201 val->local_time_benefit + val->prop_time_benefit);
2202 size = safe_add (size, val->local_size_cost + val->prop_size_cost);
2205 for (val = base; val; val = val->scc_next)
2206 for (src = val->sources; src; src = src->next)
2207 if (src->val
2208 && cgraph_maybe_hot_edge_p (src->cs))
2210 src->val->prop_time_benefit = safe_add (time,
2211 src->val->prop_time_benefit);
2212 src->val->prop_size_cost = safe_add (size,
2213 src->val->prop_size_cost);
2219 /* Propagate constants, binfos and their effects from the summaries
2220 interprocedurally. */
2222 static void
2223 ipcp_propagate_stage (struct topo_info *topo)
2225 struct cgraph_node *node;
2227 if (dump_file)
2228 fprintf (dump_file, "\n Propagating constants:\n\n");
2230 if (in_lto_p)
2231 ipa_update_after_lto_read ();
2234 FOR_EACH_DEFINED_FUNCTION (node)
2236 struct ipa_node_params *info = IPA_NODE_REF (node);
2238 determine_versionability (node);
2239 if (cgraph_function_with_gimple_body_p (node))
2241 info->lattices = XCNEWVEC (struct ipcp_param_lattices,
2242 ipa_get_param_count (info));
2243 initialize_node_lattices (node);
2245 if (node->count > max_count)
2246 max_count = node->count;
2247 overall_size += inline_summary (node)->self_size;
2250 max_new_size = overall_size;
2251 if (max_new_size < PARAM_VALUE (PARAM_LARGE_UNIT_INSNS))
2252 max_new_size = PARAM_VALUE (PARAM_LARGE_UNIT_INSNS);
2253 max_new_size += max_new_size * PARAM_VALUE (PARAM_IPCP_UNIT_GROWTH) / 100 + 1;
2255 if (dump_file)
2256 fprintf (dump_file, "\noverall_size: %li, max_new_size: %li\n",
2257 overall_size, max_new_size);
2259 propagate_constants_topo (topo);
2260 #ifdef ENABLE_CHECKING
2261 ipcp_verify_propagated_values ();
2262 #endif
2263 propagate_effects ();
2265 if (dump_file)
2267 fprintf (dump_file, "\nIPA lattices after all propagation:\n");
2268 print_all_lattices (dump_file, (dump_flags & TDF_DETAILS), true);
2272 /* Discover newly direct outgoing edges from NODE which is a new clone with
2273 known KNOWN_VALS and make them direct. */
2275 static void
2276 ipcp_discover_new_direct_edges (struct cgraph_node *node,
2277 vec<tree> known_vals,
2278 struct ipa_agg_replacement_value *aggvals)
2280 struct cgraph_edge *ie, *next_ie;
2281 bool found = false;
2283 for (ie = node->indirect_calls; ie; ie = next_ie)
2285 tree target;
2287 next_ie = ie->next_callee;
2288 target = ipa_get_indirect_edge_target_1 (ie, known_vals, vNULL, vNULL,
2289 aggvals);
2290 if (target)
2292 ipa_make_edge_direct_to_target (ie, target);
2293 found = true;
2296 /* Turning calls to direct calls will improve overall summary. */
2297 if (found)
2298 inline_update_overall_summary (node);
2301 /* Vector of pointers which for linked lists of clones of an original crgaph
2302 edge. */
2304 static vec<cgraph_edge_p> next_edge_clone;
2306 static inline void
2307 grow_next_edge_clone_vector (void)
2309 if (next_edge_clone.length ()
2310 <= (unsigned) cgraph_edge_max_uid)
2311 next_edge_clone.safe_grow_cleared (cgraph_edge_max_uid + 1);
2314 /* Edge duplication hook to grow the appropriate linked list in
2315 next_edge_clone. */
2317 static void
2318 ipcp_edge_duplication_hook (struct cgraph_edge *src, struct cgraph_edge *dst,
2319 __attribute__((unused)) void *data)
2321 grow_next_edge_clone_vector ();
2322 next_edge_clone[dst->uid] = next_edge_clone[src->uid];
2323 next_edge_clone[src->uid] = dst;
2326 /* See if NODE is a clone with a known aggregate value at a given OFFSET of a
2327 parameter with the given INDEX. */
2329 static tree
2330 get_clone_agg_value (struct cgraph_node *node, HOST_WIDEST_INT offset,
2331 int index)
2333 struct ipa_agg_replacement_value *aggval;
2335 aggval = ipa_get_agg_replacements_for_node (node);
2336 while (aggval)
2338 if (aggval->offset == offset
2339 && aggval->index == index)
2340 return aggval->value;
2341 aggval = aggval->next;
2343 return NULL_TREE;
2346 /* Return true if edge CS does bring about the value described by SRC. */
2348 static bool
2349 cgraph_edge_brings_value_p (struct cgraph_edge *cs,
2350 struct ipcp_value_source *src)
2352 struct ipa_node_params *caller_info = IPA_NODE_REF (cs->caller);
2353 struct ipa_node_params *dst_info = IPA_NODE_REF (cs->callee);
2355 if ((dst_info->ipcp_orig_node && !dst_info->is_all_contexts_clone)
2356 || caller_info->node_dead)
2357 return false;
2358 if (!src->val)
2359 return true;
2361 if (caller_info->ipcp_orig_node)
2363 tree t;
2364 if (src->offset == -1)
2365 t = caller_info->known_vals[src->index];
2366 else
2367 t = get_clone_agg_value (cs->caller, src->offset, src->index);
2368 return (t != NULL_TREE
2369 && values_equal_for_ipcp_p (src->val->value, t));
2371 else
2373 struct ipcp_agg_lattice *aglat;
2374 struct ipcp_param_lattices *plats = ipa_get_parm_lattices (caller_info,
2375 src->index);
2376 if (src->offset == -1)
2377 return (ipa_lat_is_single_const (&plats->itself)
2378 && values_equal_for_ipcp_p (src->val->value,
2379 plats->itself.values->value));
2380 else
2382 if (plats->aggs_bottom || plats->aggs_contain_variable)
2383 return false;
2384 for (aglat = plats->aggs; aglat; aglat = aglat->next)
2385 if (aglat->offset == src->offset)
2386 return (ipa_lat_is_single_const (aglat)
2387 && values_equal_for_ipcp_p (src->val->value,
2388 aglat->values->value));
2390 return false;
2394 /* Get the next clone in the linked list of clones of an edge. */
2396 static inline struct cgraph_edge *
2397 get_next_cgraph_edge_clone (struct cgraph_edge *cs)
2399 return next_edge_clone[cs->uid];
2402 /* Given VAL, iterate over all its sources and if they still hold, add their
2403 edge frequency and their number into *FREQUENCY and *CALLER_COUNT
2404 respectively. */
2406 static bool
2407 get_info_about_necessary_edges (struct ipcp_value *val, int *freq_sum,
2408 gcov_type *count_sum, int *caller_count)
2410 struct ipcp_value_source *src;
2411 int freq = 0, count = 0;
2412 gcov_type cnt = 0;
2413 bool hot = false;
2415 for (src = val->sources; src; src = src->next)
2417 struct cgraph_edge *cs = src->cs;
2418 while (cs)
2420 if (cgraph_edge_brings_value_p (cs, src))
2422 count++;
2423 freq += cs->frequency;
2424 cnt += cs->count;
2425 hot |= cgraph_maybe_hot_edge_p (cs);
2427 cs = get_next_cgraph_edge_clone (cs);
2431 *freq_sum = freq;
2432 *count_sum = cnt;
2433 *caller_count = count;
2434 return hot;
2437 /* Return a vector of incoming edges that do bring value VAL. It is assumed
2438 their number is known and equal to CALLER_COUNT. */
2440 static vec<cgraph_edge_p>
2441 gather_edges_for_value (struct ipcp_value *val, int caller_count)
2443 struct ipcp_value_source *src;
2444 vec<cgraph_edge_p> ret;
2446 ret.create (caller_count);
2447 for (src = val->sources; src; src = src->next)
2449 struct cgraph_edge *cs = src->cs;
2450 while (cs)
2452 if (cgraph_edge_brings_value_p (cs, src))
2453 ret.quick_push (cs);
2454 cs = get_next_cgraph_edge_clone (cs);
2458 return ret;
2461 /* Construct a replacement map for a know VALUE for a formal parameter PARAM.
2462 Return it or NULL if for some reason it cannot be created. */
2464 static struct ipa_replace_map *
2465 get_replacement_map (tree value, tree parm)
2467 tree req_type = TREE_TYPE (parm);
2468 struct ipa_replace_map *replace_map;
2470 if (!useless_type_conversion_p (req_type, TREE_TYPE (value)))
2472 if (fold_convertible_p (req_type, value))
2473 value = fold_build1 (NOP_EXPR, req_type, value);
2474 else if (TYPE_SIZE (req_type) == TYPE_SIZE (TREE_TYPE (value)))
2475 value = fold_build1 (VIEW_CONVERT_EXPR, req_type, value);
2476 else
2478 if (dump_file)
2480 fprintf (dump_file, " const ");
2481 print_generic_expr (dump_file, value, 0);
2482 fprintf (dump_file, " can't be converted to param ");
2483 print_generic_expr (dump_file, parm, 0);
2484 fprintf (dump_file, "\n");
2486 return NULL;
2490 replace_map = ggc_alloc_ipa_replace_map ();
2491 if (dump_file)
2493 fprintf (dump_file, " replacing param ");
2494 print_generic_expr (dump_file, parm, 0);
2495 fprintf (dump_file, " with const ");
2496 print_generic_expr (dump_file, value, 0);
2497 fprintf (dump_file, "\n");
2499 replace_map->old_tree = parm;
2500 replace_map->new_tree = value;
2501 replace_map->replace_p = true;
2502 replace_map->ref_p = false;
2504 return replace_map;
2507 /* Dump new profiling counts */
2509 static void
2510 dump_profile_updates (struct cgraph_node *orig_node,
2511 struct cgraph_node *new_node)
2513 struct cgraph_edge *cs;
2515 fprintf (dump_file, " setting count of the specialized node to "
2516 HOST_WIDE_INT_PRINT_DEC "\n", (HOST_WIDE_INT) new_node->count);
2517 for (cs = new_node->callees; cs ; cs = cs->next_callee)
2518 fprintf (dump_file, " edge to %s has count "
2519 HOST_WIDE_INT_PRINT_DEC "\n",
2520 cgraph_node_name (cs->callee), (HOST_WIDE_INT) cs->count);
2522 fprintf (dump_file, " setting count of the original node to "
2523 HOST_WIDE_INT_PRINT_DEC "\n", (HOST_WIDE_INT) orig_node->count);
2524 for (cs = orig_node->callees; cs ; cs = cs->next_callee)
2525 fprintf (dump_file, " edge to %s is left with "
2526 HOST_WIDE_INT_PRINT_DEC "\n",
2527 cgraph_node_name (cs->callee), (HOST_WIDE_INT) cs->count);
2530 /* After a specialized NEW_NODE version of ORIG_NODE has been created, update
2531 their profile information to reflect this. */
2533 static void
2534 update_profiling_info (struct cgraph_node *orig_node,
2535 struct cgraph_node *new_node)
2537 struct cgraph_edge *cs;
2538 struct caller_statistics stats;
2539 gcov_type new_sum, orig_sum;
2540 gcov_type remainder, orig_node_count = orig_node->count;
2542 if (orig_node_count == 0)
2543 return;
2545 init_caller_stats (&stats);
2546 cgraph_for_node_and_aliases (orig_node, gather_caller_stats, &stats, false);
2547 orig_sum = stats.count_sum;
2548 init_caller_stats (&stats);
2549 cgraph_for_node_and_aliases (new_node, gather_caller_stats, &stats, false);
2550 new_sum = stats.count_sum;
2552 if (orig_node_count < orig_sum + new_sum)
2554 if (dump_file)
2555 fprintf (dump_file, " Problem: node %s/%i has too low count "
2556 HOST_WIDE_INT_PRINT_DEC " while the sum of incoming "
2557 "counts is " HOST_WIDE_INT_PRINT_DEC "\n",
2558 cgraph_node_name (orig_node), orig_node->uid,
2559 (HOST_WIDE_INT) orig_node_count,
2560 (HOST_WIDE_INT) (orig_sum + new_sum));
2562 orig_node_count = (orig_sum + new_sum) * 12 / 10;
2563 if (dump_file)
2564 fprintf (dump_file, " proceeding by pretending it was "
2565 HOST_WIDE_INT_PRINT_DEC "\n",
2566 (HOST_WIDE_INT) orig_node_count);
2569 new_node->count = new_sum;
2570 remainder = orig_node_count - new_sum;
2571 orig_node->count = remainder;
2573 for (cs = new_node->callees; cs ; cs = cs->next_callee)
2574 if (cs->frequency)
2575 cs->count = cs->count * (new_sum * REG_BR_PROB_BASE
2576 / orig_node_count) / REG_BR_PROB_BASE;
2577 else
2578 cs->count = 0;
2580 for (cs = orig_node->callees; cs ; cs = cs->next_callee)
2581 cs->count = cs->count * (remainder * REG_BR_PROB_BASE
2582 / orig_node_count) / REG_BR_PROB_BASE;
2584 if (dump_file)
2585 dump_profile_updates (orig_node, new_node);
2588 /* Update the respective profile of specialized NEW_NODE and the original
2589 ORIG_NODE after additional edges with cumulative count sum REDIRECTED_SUM
2590 have been redirected to the specialized version. */
2592 static void
2593 update_specialized_profile (struct cgraph_node *new_node,
2594 struct cgraph_node *orig_node,
2595 gcov_type redirected_sum)
2597 struct cgraph_edge *cs;
2598 gcov_type new_node_count, orig_node_count = orig_node->count;
2600 if (dump_file)
2601 fprintf (dump_file, " the sum of counts of redirected edges is "
2602 HOST_WIDE_INT_PRINT_DEC "\n", (HOST_WIDE_INT) redirected_sum);
2603 if (orig_node_count == 0)
2604 return;
2606 gcc_assert (orig_node_count >= redirected_sum);
2608 new_node_count = new_node->count;
2609 new_node->count += redirected_sum;
2610 orig_node->count -= redirected_sum;
2612 for (cs = new_node->callees; cs ; cs = cs->next_callee)
2613 if (cs->frequency)
2614 cs->count += cs->count * redirected_sum / new_node_count;
2615 else
2616 cs->count = 0;
2618 for (cs = orig_node->callees; cs ; cs = cs->next_callee)
2620 gcov_type dec = cs->count * (redirected_sum * REG_BR_PROB_BASE
2621 / orig_node_count) / REG_BR_PROB_BASE;
2622 if (dec < cs->count)
2623 cs->count -= dec;
2624 else
2625 cs->count = 0;
2628 if (dump_file)
2629 dump_profile_updates (orig_node, new_node);
2632 /* Create a specialized version of NODE with known constants and types of
2633 parameters in KNOWN_VALS and redirect all edges in CALLERS to it. */
2635 static struct cgraph_node *
2636 create_specialized_node (struct cgraph_node *node,
2637 vec<tree> known_vals,
2638 struct ipa_agg_replacement_value *aggvals,
2639 vec<cgraph_edge_p> callers)
2641 struct ipa_node_params *new_info, *info = IPA_NODE_REF (node);
2642 vec<ipa_replace_map_p, va_gc> *replace_trees = NULL;
2643 struct cgraph_node *new_node;
2644 int i, count = ipa_get_param_count (info);
2645 bitmap args_to_skip;
2647 gcc_assert (!info->ipcp_orig_node);
2649 if (node->local.can_change_signature)
2651 args_to_skip = BITMAP_GGC_ALLOC ();
2652 for (i = 0; i < count; i++)
2654 tree t = known_vals[i];
2656 if ((t && TREE_CODE (t) != TREE_BINFO)
2657 || !ipa_is_param_used (info, i))
2658 bitmap_set_bit (args_to_skip, i);
2661 else
2663 args_to_skip = NULL;
2664 if (dump_file && (dump_flags & TDF_DETAILS))
2665 fprintf (dump_file, " cannot change function signature\n");
2668 for (i = 0; i < count ; i++)
2670 tree t = known_vals[i];
2671 if (t && TREE_CODE (t) != TREE_BINFO)
2673 struct ipa_replace_map *replace_map;
2675 replace_map = get_replacement_map (t, ipa_get_param (info, i));
2676 if (replace_map)
2677 vec_safe_push (replace_trees, replace_map);
2681 new_node = cgraph_create_virtual_clone (node, callers, replace_trees,
2682 args_to_skip, "constprop");
2683 ipa_set_node_agg_value_chain (new_node, aggvals);
2684 if (dump_file && (dump_flags & TDF_DETAILS))
2686 fprintf (dump_file, " the new node is %s/%i.\n",
2687 cgraph_node_name (new_node), new_node->uid);
2688 if (aggvals)
2689 ipa_dump_agg_replacement_values (dump_file, aggvals);
2691 gcc_checking_assert (ipa_node_params_vector.exists ()
2692 && (ipa_node_params_vector.length ()
2693 > (unsigned) cgraph_max_uid));
2694 update_profiling_info (node, new_node);
2695 new_info = IPA_NODE_REF (new_node);
2696 new_info->ipcp_orig_node = node;
2697 new_info->known_vals = known_vals;
2699 ipcp_discover_new_direct_edges (new_node, known_vals, aggvals);
2701 callers.release ();
2702 return new_node;
2705 /* Given a NODE, and a subset of its CALLERS, try to populate blanks slots in
2706 KNOWN_VALS with constants and types that are also known for all of the
2707 CALLERS. */
2709 static void
2710 find_more_scalar_values_for_callers_subset (struct cgraph_node *node,
2711 vec<tree> known_vals,
2712 vec<cgraph_edge_p> callers)
2714 struct ipa_node_params *info = IPA_NODE_REF (node);
2715 int i, count = ipa_get_param_count (info);
2717 for (i = 0; i < count ; i++)
2719 struct cgraph_edge *cs;
2720 tree newval = NULL_TREE;
2721 int j;
2723 if (ipa_get_scalar_lat (info, i)->bottom || known_vals[i])
2724 continue;
2726 FOR_EACH_VEC_ELT (callers, j, cs)
2728 struct ipa_jump_func *jump_func;
2729 tree t;
2731 if (i >= ipa_get_cs_argument_count (IPA_EDGE_REF (cs)))
2733 newval = NULL_TREE;
2734 break;
2736 jump_func = ipa_get_ith_jump_func (IPA_EDGE_REF (cs), i);
2737 t = ipa_value_from_jfunc (IPA_NODE_REF (cs->caller), jump_func);
2738 if (!t
2739 || (newval
2740 && !values_equal_for_ipcp_p (t, newval)))
2742 newval = NULL_TREE;
2743 break;
2745 else
2746 newval = t;
2749 if (newval)
2751 if (dump_file && (dump_flags & TDF_DETAILS))
2753 fprintf (dump_file, " adding an extra known scalar value ");
2754 print_ipcp_constant_value (dump_file, newval);
2755 fprintf (dump_file, " for parameter ");
2756 print_generic_expr (dump_file, ipa_get_param (info, i), 0);
2757 fprintf (dump_file, "\n");
2760 known_vals[i] = newval;
2765 /* Go through PLATS and create a vector of values consisting of values and
2766 offsets (minus OFFSET) of lattices that contain only a single value. */
2768 static vec<ipa_agg_jf_item_t>
2769 copy_plats_to_inter (struct ipcp_param_lattices *plats, HOST_WIDE_INT offset)
2771 vec<ipa_agg_jf_item_t> res = vNULL;
2773 if (!plats->aggs || plats->aggs_contain_variable || plats->aggs_bottom)
2774 return vNULL;
2776 for (struct ipcp_agg_lattice *aglat = plats->aggs; aglat; aglat = aglat->next)
2777 if (ipa_lat_is_single_const (aglat))
2779 struct ipa_agg_jf_item ti;
2780 ti.offset = aglat->offset - offset;
2781 ti.value = aglat->values->value;
2782 res.safe_push (ti);
2784 return res;
2787 /* Intersect all values in INTER with single value lattices in PLATS (while
2788 subtracting OFFSET). */
2790 static void
2791 intersect_with_plats (struct ipcp_param_lattices *plats,
2792 vec<ipa_agg_jf_item_t> *inter,
2793 HOST_WIDE_INT offset)
2795 struct ipcp_agg_lattice *aglat;
2796 struct ipa_agg_jf_item *item;
2797 int k;
2799 if (!plats->aggs || plats->aggs_contain_variable || plats->aggs_bottom)
2801 inter->release ();
2802 return;
2805 aglat = plats->aggs;
2806 FOR_EACH_VEC_ELT (*inter, k, item)
2808 bool found = false;
2809 if (!item->value)
2810 continue;
2811 while (aglat)
2813 if (aglat->offset - offset > item->offset)
2814 break;
2815 if (aglat->offset - offset == item->offset)
2817 gcc_checking_assert (item->value);
2818 if (values_equal_for_ipcp_p (item->value, aglat->values->value))
2819 found = true;
2820 break;
2822 aglat = aglat->next;
2824 if (!found)
2825 item->value = NULL_TREE;
2829 /* Copy agggregate replacement values of NODE (which is an IPA-CP clone) to the
2830 vector result while subtracting OFFSET from the individual value offsets. */
2832 static vec<ipa_agg_jf_item_t>
2833 agg_replacements_to_vector (struct cgraph_node *node, int index,
2834 HOST_WIDE_INT offset)
2836 struct ipa_agg_replacement_value *av;
2837 vec<ipa_agg_jf_item_t> res = vNULL;
2839 for (av = ipa_get_agg_replacements_for_node (node); av; av = av->next)
2840 if (av->index == index
2841 && (av->offset - offset) >= 0)
2843 struct ipa_agg_jf_item item;
2844 gcc_checking_assert (av->value);
2845 item.offset = av->offset - offset;
2846 item.value = av->value;
2847 res.safe_push (item);
2850 return res;
2853 /* Intersect all values in INTER with those that we have already scheduled to
2854 be replaced in parameter number INDEX of NODE, which is an IPA-CP clone
2855 (while subtracting OFFSET). */
2857 static void
2858 intersect_with_agg_replacements (struct cgraph_node *node, int index,
2859 vec<ipa_agg_jf_item_t> *inter,
2860 HOST_WIDE_INT offset)
2862 struct ipa_agg_replacement_value *srcvals;
2863 struct ipa_agg_jf_item *item;
2864 int i;
2866 srcvals = ipa_get_agg_replacements_for_node (node);
2867 if (!srcvals)
2869 inter->release ();
2870 return;
2873 FOR_EACH_VEC_ELT (*inter, i, item)
2875 struct ipa_agg_replacement_value *av;
2876 bool found = false;
2877 if (!item->value)
2878 continue;
2879 for (av = srcvals; av; av = av->next)
2881 gcc_checking_assert (av->value);
2882 if (av->index == index
2883 && av->offset - offset == item->offset)
2885 if (values_equal_for_ipcp_p (item->value, av->value))
2886 found = true;
2887 break;
2890 if (!found)
2891 item->value = NULL_TREE;
2895 /* Intersect values in INTER with aggregate values that come along edge CS to
2896 parameter number INDEX and return it. If INTER does not actually exist yet,
2897 copy all incoming values to it. If we determine we ended up with no values
2898 whatsoever, return a released vector. */
2900 static vec<ipa_agg_jf_item_t>
2901 intersect_aggregates_with_edge (struct cgraph_edge *cs, int index,
2902 vec<ipa_agg_jf_item_t> inter)
2904 struct ipa_jump_func *jfunc;
2905 jfunc = ipa_get_ith_jump_func (IPA_EDGE_REF (cs), index);
2906 if (jfunc->type == IPA_JF_PASS_THROUGH
2907 && ipa_get_jf_pass_through_operation (jfunc) == NOP_EXPR)
2909 struct ipa_node_params *caller_info = IPA_NODE_REF (cs->caller);
2910 int src_idx = ipa_get_jf_pass_through_formal_id (jfunc);
2912 if (caller_info->ipcp_orig_node)
2914 struct cgraph_node *orig_node = caller_info->ipcp_orig_node;
2915 struct ipcp_param_lattices *orig_plats;
2916 orig_plats = ipa_get_parm_lattices (IPA_NODE_REF (orig_node),
2917 src_idx);
2918 if (agg_pass_through_permissible_p (orig_plats, jfunc))
2920 if (!inter.exists ())
2921 inter = agg_replacements_to_vector (cs->caller, src_idx, 0);
2922 else
2923 intersect_with_agg_replacements (cs->caller, src_idx,
2924 &inter, 0);
2927 else
2929 struct ipcp_param_lattices *src_plats;
2930 src_plats = ipa_get_parm_lattices (caller_info, src_idx);
2931 if (agg_pass_through_permissible_p (src_plats, jfunc))
2933 /* Currently we do not produce clobber aggregate jump
2934 functions, adjust when we do. */
2935 gcc_checking_assert (!jfunc->agg.items);
2936 if (!inter.exists ())
2937 inter = copy_plats_to_inter (src_plats, 0);
2938 else
2939 intersect_with_plats (src_plats, &inter, 0);
2943 else if (jfunc->type == IPA_JF_ANCESTOR
2944 && ipa_get_jf_ancestor_agg_preserved (jfunc))
2946 struct ipa_node_params *caller_info = IPA_NODE_REF (cs->caller);
2947 int src_idx = ipa_get_jf_ancestor_formal_id (jfunc);
2948 struct ipcp_param_lattices *src_plats;
2949 HOST_WIDE_INT delta = ipa_get_jf_ancestor_offset (jfunc);
2951 if (caller_info->ipcp_orig_node)
2953 if (!inter.exists ())
2954 inter = agg_replacements_to_vector (cs->caller, src_idx, delta);
2955 else
2956 intersect_with_agg_replacements (cs->caller, src_idx, &inter,
2957 delta);
2959 else
2961 src_plats = ipa_get_parm_lattices (caller_info, src_idx);;
2962 /* Currently we do not produce clobber aggregate jump
2963 functions, adjust when we do. */
2964 gcc_checking_assert (!src_plats->aggs || !jfunc->agg.items);
2965 if (!inter.exists ())
2966 inter = copy_plats_to_inter (src_plats, delta);
2967 else
2968 intersect_with_plats (src_plats, &inter, delta);
2971 else if (jfunc->agg.items)
2973 struct ipa_agg_jf_item *item;
2974 int k;
2976 if (!inter.exists ())
2977 for (unsigned i = 0; i < jfunc->agg.items->length (); i++)
2978 inter.safe_push ((*jfunc->agg.items)[i]);
2979 else
2980 FOR_EACH_VEC_ELT (inter, k, item)
2982 int l = 0;
2983 bool found = false;;
2985 if (!item->value)
2986 continue;
2988 while ((unsigned) l < jfunc->agg.items->length ())
2990 struct ipa_agg_jf_item *ti;
2991 ti = &(*jfunc->agg.items)[l];
2992 if (ti->offset > item->offset)
2993 break;
2994 if (ti->offset == item->offset)
2996 gcc_checking_assert (ti->value);
2997 if (values_equal_for_ipcp_p (item->value,
2998 ti->value))
2999 found = true;
3000 break;
3002 l++;
3004 if (!found)
3005 item->value = NULL;
3008 else
3010 inter.release();
3011 return vec<ipa_agg_jf_item_t>();
3013 return inter;
3016 /* Look at edges in CALLERS and collect all known aggregate values that arrive
3017 from all of them. */
3019 static struct ipa_agg_replacement_value *
3020 find_aggregate_values_for_callers_subset (struct cgraph_node *node,
3021 vec<cgraph_edge_p> callers)
3023 struct ipa_node_params *dest_info = IPA_NODE_REF (node);
3024 struct ipa_agg_replacement_value *res = NULL;
3025 struct cgraph_edge *cs;
3026 int i, j, count = ipa_get_param_count (dest_info);
3028 FOR_EACH_VEC_ELT (callers, j, cs)
3030 int c = ipa_get_cs_argument_count (IPA_EDGE_REF (cs));
3031 if (c < count)
3032 count = c;
3035 for (i = 0; i < count ; i++)
3037 struct cgraph_edge *cs;
3038 vec<ipa_agg_jf_item_t> inter = vNULL;
3039 struct ipa_agg_jf_item *item;
3040 int j;
3042 /* Among other things, the following check should deal with all by_ref
3043 mismatches. */
3044 if (ipa_get_parm_lattices (dest_info, i)->aggs_bottom)
3045 continue;
3047 FOR_EACH_VEC_ELT (callers, j, cs)
3049 inter = intersect_aggregates_with_edge (cs, i, inter);
3051 if (!inter.exists ())
3052 goto next_param;
3055 FOR_EACH_VEC_ELT (inter, j, item)
3057 struct ipa_agg_replacement_value *v;
3059 if (!item->value)
3060 continue;
3062 v = ggc_alloc_ipa_agg_replacement_value ();
3063 v->index = i;
3064 v->offset = item->offset;
3065 v->value = item->value;
3066 v->next = res;
3067 res = v;
3070 next_param:
3071 if (inter.exists ())
3072 inter.release ();
3074 return res;
3077 /* Turn KNOWN_AGGS into a list of aggreate replacement values. */
3079 static struct ipa_agg_replacement_value *
3080 known_aggs_to_agg_replacement_list (vec<ipa_agg_jump_function_t> known_aggs)
3082 struct ipa_agg_replacement_value *res = NULL;
3083 struct ipa_agg_jump_function *aggjf;
3084 struct ipa_agg_jf_item *item;
3085 int i, j;
3087 FOR_EACH_VEC_ELT (known_aggs, i, aggjf)
3088 FOR_EACH_VEC_SAFE_ELT (aggjf->items, j, item)
3090 struct ipa_agg_replacement_value *v;
3091 v = ggc_alloc_ipa_agg_replacement_value ();
3092 v->index = i;
3093 v->offset = item->offset;
3094 v->value = item->value;
3095 v->next = res;
3096 res = v;
3098 return res;
3101 /* Determine whether CS also brings all scalar values that the NODE is
3102 specialized for. */
3104 static bool
3105 cgraph_edge_brings_all_scalars_for_node (struct cgraph_edge *cs,
3106 struct cgraph_node *node)
3108 struct ipa_node_params *dest_info = IPA_NODE_REF (node);
3109 int count = ipa_get_param_count (dest_info);
3110 struct ipa_node_params *caller_info;
3111 struct ipa_edge_args *args;
3112 int i;
3114 caller_info = IPA_NODE_REF (cs->caller);
3115 args = IPA_EDGE_REF (cs);
3116 for (i = 0; i < count; i++)
3118 struct ipa_jump_func *jump_func;
3119 tree val, t;
3121 val = dest_info->known_vals[i];
3122 if (!val)
3123 continue;
3125 if (i >= ipa_get_cs_argument_count (args))
3126 return false;
3127 jump_func = ipa_get_ith_jump_func (args, i);
3128 t = ipa_value_from_jfunc (caller_info, jump_func);
3129 if (!t || !values_equal_for_ipcp_p (val, t))
3130 return false;
3132 return true;
3135 /* Determine whether CS also brings all aggregate values that NODE is
3136 specialized for. */
3137 static bool
3138 cgraph_edge_brings_all_agg_vals_for_node (struct cgraph_edge *cs,
3139 struct cgraph_node *node)
3141 struct ipa_node_params *orig_caller_info = IPA_NODE_REF (cs->caller);
3142 struct ipa_agg_replacement_value *aggval;
3143 int i, ec, count;
3145 aggval = ipa_get_agg_replacements_for_node (node);
3146 if (!aggval)
3147 return true;
3149 count = ipa_get_param_count (IPA_NODE_REF (node));
3150 ec = ipa_get_cs_argument_count (IPA_EDGE_REF (cs));
3151 if (ec < count)
3152 for (struct ipa_agg_replacement_value *av = aggval; av; av = av->next)
3153 if (aggval->index >= ec)
3154 return false;
3156 if (orig_caller_info->ipcp_orig_node)
3157 orig_caller_info = IPA_NODE_REF (orig_caller_info->ipcp_orig_node);
3159 for (i = 0; i < count; i++)
3161 static vec<ipa_agg_jf_item_t> values = vec<ipa_agg_jf_item_t>();
3162 struct ipcp_param_lattices *plats;
3163 bool interesting = false;
3164 for (struct ipa_agg_replacement_value *av = aggval; av; av = av->next)
3165 if (aggval->index == i)
3167 interesting = true;
3168 break;
3170 if (!interesting)
3171 continue;
3173 plats = ipa_get_parm_lattices (orig_caller_info, aggval->index);
3174 if (plats->aggs_bottom)
3175 return false;
3177 values = intersect_aggregates_with_edge (cs, i, values);
3178 if (!values.exists())
3179 return false;
3181 for (struct ipa_agg_replacement_value *av = aggval; av; av = av->next)
3182 if (aggval->index == i)
3184 struct ipa_agg_jf_item *item;
3185 int j;
3186 bool found = false;
3187 FOR_EACH_VEC_ELT (values, j, item)
3188 if (item->value
3189 && item->offset == av->offset
3190 && values_equal_for_ipcp_p (item->value, av->value))
3191 found = true;
3192 if (!found)
3194 values.release();
3195 return false;
3199 return true;
3202 /* Given an original NODE and a VAL for which we have already created a
3203 specialized clone, look whether there are incoming edges that still lead
3204 into the old node but now also bring the requested value and also conform to
3205 all other criteria such that they can be redirected the the special node.
3206 This function can therefore redirect the final edge in a SCC. */
3208 static void
3209 perhaps_add_new_callers (struct cgraph_node *node, struct ipcp_value *val)
3211 struct ipcp_value_source *src;
3212 gcov_type redirected_sum = 0;
3214 for (src = val->sources; src; src = src->next)
3216 struct cgraph_edge *cs = src->cs;
3217 while (cs)
3219 enum availability availability;
3220 struct cgraph_node *dst = cgraph_function_node (cs->callee,
3221 &availability);
3222 if ((dst == node || IPA_NODE_REF (dst)->is_all_contexts_clone)
3223 && availability > AVAIL_OVERWRITABLE
3224 && cgraph_edge_brings_value_p (cs, src))
3226 if (cgraph_edge_brings_all_scalars_for_node (cs, val->spec_node)
3227 && cgraph_edge_brings_all_agg_vals_for_node (cs,
3228 val->spec_node))
3230 if (dump_file)
3231 fprintf (dump_file, " - adding an extra caller %s/%i"
3232 " of %s/%i\n",
3233 xstrdup (cgraph_node_name (cs->caller)),
3234 cs->caller->uid,
3235 xstrdup (cgraph_node_name (val->spec_node)),
3236 val->spec_node->uid);
3238 cgraph_redirect_edge_callee (cs, val->spec_node);
3239 redirected_sum += cs->count;
3242 cs = get_next_cgraph_edge_clone (cs);
3246 if (redirected_sum)
3247 update_specialized_profile (val->spec_node, node, redirected_sum);
3251 /* Copy KNOWN_BINFOS to KNOWN_VALS. */
3253 static void
3254 move_binfos_to_values (vec<tree> known_vals,
3255 vec<tree> known_binfos)
3257 tree t;
3258 int i;
3260 for (i = 0; known_binfos.iterate (i, &t); i++)
3261 if (t)
3262 known_vals[i] = t;
3265 /* Return true if there is a replacement equivalent to VALUE, INDEX and OFFSET
3266 among those in the AGGVALS list. */
3268 DEBUG_FUNCTION bool
3269 ipcp_val_in_agg_replacements_p (struct ipa_agg_replacement_value *aggvals,
3270 int index, HOST_WIDE_INT offset, tree value)
3272 while (aggvals)
3274 if (aggvals->index == index
3275 && aggvals->offset == offset
3276 && values_equal_for_ipcp_p (aggvals->value, value))
3277 return true;
3278 aggvals = aggvals->next;
3280 return false;
3283 /* Decide wheter to create a special version of NODE for value VAL of parameter
3284 at the given INDEX. If OFFSET is -1, the value is for the parameter itself,
3285 otherwise it is stored at the given OFFSET of the parameter. KNOWN_CSTS,
3286 KNOWN_BINFOS and KNOWN_AGGS describe the other already known values. */
3288 static bool
3289 decide_about_value (struct cgraph_node *node, int index, HOST_WIDE_INT offset,
3290 struct ipcp_value *val, vec<tree> known_csts,
3291 vec<tree> known_binfos)
3293 struct ipa_agg_replacement_value *aggvals;
3294 int freq_sum, caller_count;
3295 gcov_type count_sum;
3296 vec<cgraph_edge_p> callers;
3297 vec<tree> kv;
3299 if (val->spec_node)
3301 perhaps_add_new_callers (node, val);
3302 return false;
3304 else if (val->local_size_cost + overall_size > max_new_size)
3306 if (dump_file && (dump_flags & TDF_DETAILS))
3307 fprintf (dump_file, " Ignoring candidate value because "
3308 "max_new_size would be reached with %li.\n",
3309 val->local_size_cost + overall_size);
3310 return false;
3312 else if (!get_info_about_necessary_edges (val, &freq_sum, &count_sum,
3313 &caller_count))
3314 return false;
3316 if (dump_file && (dump_flags & TDF_DETAILS))
3318 fprintf (dump_file, " - considering value ");
3319 print_ipcp_constant_value (dump_file, val->value);
3320 fprintf (dump_file, " for parameter ");
3321 print_generic_expr (dump_file, ipa_get_param (IPA_NODE_REF (node),
3322 index), 0);
3323 if (offset != -1)
3324 fprintf (dump_file, ", offset: " HOST_WIDE_INT_PRINT_DEC, offset);
3325 fprintf (dump_file, " (caller_count: %i)\n", caller_count);
3328 if (!good_cloning_opportunity_p (node, val->local_time_benefit,
3329 freq_sum, count_sum,
3330 val->local_size_cost)
3331 && !good_cloning_opportunity_p (node,
3332 val->local_time_benefit
3333 + val->prop_time_benefit,
3334 freq_sum, count_sum,
3335 val->local_size_cost
3336 + val->prop_size_cost))
3337 return false;
3339 if (dump_file)
3340 fprintf (dump_file, " Creating a specialized node of %s/%i.\n",
3341 cgraph_node_name (node), node->uid);
3343 callers = gather_edges_for_value (val, caller_count);
3344 kv = known_csts.copy ();
3345 move_binfos_to_values (kv, known_binfos);
3346 if (offset == -1)
3347 kv[index] = val->value;
3348 find_more_scalar_values_for_callers_subset (node, kv, callers);
3349 aggvals = find_aggregate_values_for_callers_subset (node, callers);
3350 gcc_checking_assert (offset == -1
3351 || ipcp_val_in_agg_replacements_p (aggvals, index,
3352 offset, val->value));
3353 val->spec_node = create_specialized_node (node, kv, aggvals, callers);
3354 overall_size += val->local_size_cost;
3356 /* TODO: If for some lattice there is only one other known value
3357 left, make a special node for it too. */
3359 return true;
3362 /* Decide whether and what specialized clones of NODE should be created. */
3364 static bool
3365 decide_whether_version_node (struct cgraph_node *node)
3367 struct ipa_node_params *info = IPA_NODE_REF (node);
3368 int i, count = ipa_get_param_count (info);
3369 vec<tree> known_csts, known_binfos;
3370 vec<ipa_agg_jump_function_t> known_aggs = vNULL;
3371 bool ret = false;
3373 if (count == 0)
3374 return false;
3376 if (dump_file && (dump_flags & TDF_DETAILS))
3377 fprintf (dump_file, "\nEvaluating opportunities for %s/%i.\n",
3378 cgraph_node_name (node), node->uid);
3380 gather_context_independent_values (info, &known_csts, &known_binfos,
3381 info->do_clone_for_all_contexts ? &known_aggs
3382 : NULL, NULL);
3384 for (i = 0; i < count ;i++)
3386 struct ipcp_param_lattices *plats = ipa_get_parm_lattices (info, i);
3387 struct ipcp_lattice *lat = &plats->itself;
3388 struct ipcp_value *val;
3390 if (!lat->bottom
3391 && !known_csts[i]
3392 && !known_binfos[i])
3393 for (val = lat->values; val; val = val->next)
3394 ret |= decide_about_value (node, i, -1, val, known_csts,
3395 known_binfos);
3397 if (!plats->aggs_bottom)
3399 struct ipcp_agg_lattice *aglat;
3400 struct ipcp_value *val;
3401 for (aglat = plats->aggs; aglat; aglat = aglat->next)
3402 if (!aglat->bottom && aglat->values
3403 /* If the following is false, the one value is in
3404 known_aggs. */
3405 && (plats->aggs_contain_variable
3406 || !ipa_lat_is_single_const (aglat)))
3407 for (val = aglat->values; val; val = val->next)
3408 ret |= decide_about_value (node, i, aglat->offset, val,
3409 known_csts, known_binfos);
3411 info = IPA_NODE_REF (node);
3414 if (info->do_clone_for_all_contexts)
3416 struct cgraph_node *clone;
3417 vec<cgraph_edge_p> callers;
3419 if (dump_file)
3420 fprintf (dump_file, " - Creating a specialized node of %s/%i "
3421 "for all known contexts.\n", cgraph_node_name (node),
3422 node->uid);
3424 callers = collect_callers_of_node (node);
3425 move_binfos_to_values (known_csts, known_binfos);
3426 clone = create_specialized_node (node, known_csts,
3427 known_aggs_to_agg_replacement_list (known_aggs),
3428 callers);
3429 info = IPA_NODE_REF (node);
3430 info->do_clone_for_all_contexts = false;
3431 IPA_NODE_REF (clone)->is_all_contexts_clone = true;
3432 for (i = 0; i < count ; i++)
3433 vec_free (known_aggs[i].items);
3434 known_aggs.release ();
3435 ret = true;
3437 else
3438 known_csts.release ();
3440 known_binfos.release ();
3441 return ret;
3444 /* Transitively mark all callees of NODE within the same SCC as not dead. */
3446 static void
3447 spread_undeadness (struct cgraph_node *node)
3449 struct cgraph_edge *cs;
3451 for (cs = node->callees; cs; cs = cs->next_callee)
3452 if (edge_within_scc (cs))
3454 struct cgraph_node *callee;
3455 struct ipa_node_params *info;
3457 callee = cgraph_function_node (cs->callee, NULL);
3458 info = IPA_NODE_REF (callee);
3460 if (info->node_dead)
3462 info->node_dead = 0;
3463 spread_undeadness (callee);
3468 /* Return true if NODE has a caller from outside of its SCC that is not
3469 dead. Worker callback for cgraph_for_node_and_aliases. */
3471 static bool
3472 has_undead_caller_from_outside_scc_p (struct cgraph_node *node,
3473 void *data ATTRIBUTE_UNUSED)
3475 struct cgraph_edge *cs;
3477 for (cs = node->callers; cs; cs = cs->next_caller)
3478 if (cs->caller->thunk.thunk_p
3479 && cgraph_for_node_and_aliases (cs->caller,
3480 has_undead_caller_from_outside_scc_p,
3481 NULL, true))
3482 return true;
3483 else if (!edge_within_scc (cs)
3484 && !IPA_NODE_REF (cs->caller)->node_dead)
3485 return true;
3486 return false;
3490 /* Identify nodes within the same SCC as NODE which are no longer needed
3491 because of new clones and will be removed as unreachable. */
3493 static void
3494 identify_dead_nodes (struct cgraph_node *node)
3496 struct cgraph_node *v;
3497 for (v = node; v ; v = ((struct ipa_dfs_info *) v->symbol.aux)->next_cycle)
3498 if (cgraph_will_be_removed_from_program_if_no_direct_calls (v)
3499 && !cgraph_for_node_and_aliases (v,
3500 has_undead_caller_from_outside_scc_p,
3501 NULL, true))
3502 IPA_NODE_REF (v)->node_dead = 1;
3504 for (v = node; v ; v = ((struct ipa_dfs_info *) v->symbol.aux)->next_cycle)
3505 if (!IPA_NODE_REF (v)->node_dead)
3506 spread_undeadness (v);
3508 if (dump_file && (dump_flags & TDF_DETAILS))
3510 for (v = node; v ; v = ((struct ipa_dfs_info *) v->symbol.aux)->next_cycle)
3511 if (IPA_NODE_REF (v)->node_dead)
3512 fprintf (dump_file, " Marking node as dead: %s/%i.\n",
3513 cgraph_node_name (v), v->uid);
3517 /* The decision stage. Iterate over the topological order of call graph nodes
3518 TOPO and make specialized clones if deemed beneficial. */
3520 static void
3521 ipcp_decision_stage (struct topo_info *topo)
3523 int i;
3525 if (dump_file)
3526 fprintf (dump_file, "\nIPA decision stage:\n\n");
3528 for (i = topo->nnodes - 1; i >= 0; i--)
3530 struct cgraph_node *node = topo->order[i];
3531 bool change = false, iterate = true;
3533 while (iterate)
3535 struct cgraph_node *v;
3536 iterate = false;
3537 for (v = node; v ; v = ((struct ipa_dfs_info *) v->symbol.aux)->next_cycle)
3538 if (cgraph_function_with_gimple_body_p (v)
3539 && ipcp_versionable_function_p (v))
3540 iterate |= decide_whether_version_node (v);
3542 change |= iterate;
3544 if (change)
3545 identify_dead_nodes (node);
3549 /* The IPCP driver. */
3551 static unsigned int
3552 ipcp_driver (void)
3554 struct cgraph_2edge_hook_list *edge_duplication_hook_holder;
3555 struct topo_info topo;
3557 ipa_check_create_node_params ();
3558 ipa_check_create_edge_args ();
3559 grow_next_edge_clone_vector ();
3560 edge_duplication_hook_holder =
3561 cgraph_add_edge_duplication_hook (&ipcp_edge_duplication_hook, NULL);
3562 ipcp_values_pool = create_alloc_pool ("IPA-CP values",
3563 sizeof (struct ipcp_value), 32);
3564 ipcp_sources_pool = create_alloc_pool ("IPA-CP value sources",
3565 sizeof (struct ipcp_value_source), 64);
3566 ipcp_agg_lattice_pool = create_alloc_pool ("IPA_CP aggregate lattices",
3567 sizeof (struct ipcp_agg_lattice),
3568 32);
3569 if (dump_file)
3571 fprintf (dump_file, "\nIPA structures before propagation:\n");
3572 if (dump_flags & TDF_DETAILS)
3573 ipa_print_all_params (dump_file);
3574 ipa_print_all_jump_functions (dump_file);
3577 /* Topological sort. */
3578 build_toporder_info (&topo);
3579 /* Do the interprocedural propagation. */
3580 ipcp_propagate_stage (&topo);
3581 /* Decide what constant propagation and cloning should be performed. */
3582 ipcp_decision_stage (&topo);
3584 /* Free all IPCP structures. */
3585 free_toporder_info (&topo);
3586 next_edge_clone.release ();
3587 cgraph_remove_edge_duplication_hook (edge_duplication_hook_holder);
3588 ipa_free_all_structures_after_ipa_cp ();
3589 if (dump_file)
3590 fprintf (dump_file, "\nIPA constant propagation end\n");
3591 return 0;
3594 /* Initialization and computation of IPCP data structures. This is the initial
3595 intraprocedural analysis of functions, which gathers information to be
3596 propagated later on. */
3598 static void
3599 ipcp_generate_summary (void)
3601 struct cgraph_node *node;
3603 if (dump_file)
3604 fprintf (dump_file, "\nIPA constant propagation start:\n");
3605 ipa_register_cgraph_hooks ();
3607 FOR_EACH_FUNCTION_WITH_GIMPLE_BODY (node)
3609 node->local.versionable
3610 = tree_versionable_function_p (node->symbol.decl);
3611 ipa_analyze_node (node);
3615 /* Write ipcp summary for nodes in SET. */
3617 static void
3618 ipcp_write_summary (void)
3620 ipa_prop_write_jump_functions ();
3623 /* Read ipcp summary. */
3625 static void
3626 ipcp_read_summary (void)
3628 ipa_prop_read_jump_functions ();
3631 /* Gate for IPCP optimization. */
3633 static bool
3634 cgraph_gate_cp (void)
3636 /* FIXME: We should remove the optimize check after we ensure we never run
3637 IPA passes when not optimizing. */
3638 return flag_ipa_cp && optimize;
3641 struct ipa_opt_pass_d pass_ipa_cp =
3644 IPA_PASS,
3645 "cp", /* name */
3646 OPTGROUP_NONE, /* optinfo_flags */
3647 cgraph_gate_cp, /* gate */
3648 ipcp_driver, /* execute */
3649 NULL, /* sub */
3650 NULL, /* next */
3651 0, /* static_pass_number */
3652 TV_IPA_CONSTANT_PROP, /* tv_id */
3653 0, /* properties_required */
3654 0, /* properties_provided */
3655 0, /* properties_destroyed */
3656 0, /* todo_flags_start */
3657 TODO_dump_symtab |
3658 TODO_remove_functions | TODO_ggc_collect /* todo_flags_finish */
3660 ipcp_generate_summary, /* generate_summary */
3661 ipcp_write_summary, /* write_summary */
3662 ipcp_read_summary, /* read_summary */
3663 ipa_prop_write_all_agg_replacement, /* write_optimization_summary */
3664 ipa_prop_read_all_agg_replacement, /* read_optimization_summary */
3665 NULL, /* stmt_fixup */
3666 0, /* TODOs */
3667 ipcp_transform_function, /* function_transform */
3668 NULL, /* variable_transform */