Merge trunk version 208955 into gupc branch.
[official-gcc.git] / gcc / ipa-cp.c
blob05de8572492c066dd8b708f5b72b3c04525c4784
1 /* Interprocedural constant propagation
2 Copyright (C) 2005-2014 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 "gimple-fold.h"
108 #include "gimple-expr.h"
109 #include "target.h"
110 #include "ipa-prop.h"
111 #include "bitmap.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 /* Print V which is extracted from a value in a lattice to F. */
292 static void
293 print_ipcp_constant_value (FILE * f, tree v)
295 if (TREE_CODE (v) == TREE_BINFO)
297 fprintf (f, "BINFO ");
298 print_generic_expr (f, BINFO_TYPE (v), 0);
300 else if (TREE_CODE (v) == ADDR_EXPR
301 && TREE_CODE (TREE_OPERAND (v, 0)) == CONST_DECL)
303 fprintf (f, "& ");
304 print_generic_expr (f, DECL_INITIAL (TREE_OPERAND (v, 0)), 0);
306 else
307 print_generic_expr (f, v, 0);
310 /* Print a lattice LAT to F. */
312 static void
313 print_lattice (FILE * f, struct ipcp_lattice *lat,
314 bool dump_sources, bool dump_benefits)
316 struct ipcp_value *val;
317 bool prev = false;
319 if (lat->bottom)
321 fprintf (f, "BOTTOM\n");
322 return;
325 if (!lat->values_count && !lat->contains_variable)
327 fprintf (f, "TOP\n");
328 return;
331 if (lat->contains_variable)
333 fprintf (f, "VARIABLE");
334 prev = true;
335 if (dump_benefits)
336 fprintf (f, "\n");
339 for (val = lat->values; val; val = val->next)
341 if (dump_benefits && prev)
342 fprintf (f, " ");
343 else if (!dump_benefits && prev)
344 fprintf (f, ", ");
345 else
346 prev = true;
348 print_ipcp_constant_value (f, val->value);
350 if (dump_sources)
352 struct ipcp_value_source *s;
354 fprintf (f, " [from:");
355 for (s = val->sources; s; s = s->next)
356 fprintf (f, " %i(%i)", s->cs->caller->order,
357 s->cs->frequency);
358 fprintf (f, "]");
361 if (dump_benefits)
362 fprintf (f, " [loc_time: %i, loc_size: %i, "
363 "prop_time: %i, prop_size: %i]\n",
364 val->local_time_benefit, val->local_size_cost,
365 val->prop_time_benefit, val->prop_size_cost);
367 if (!dump_benefits)
368 fprintf (f, "\n");
371 /* Print all ipcp_lattices of all functions to F. */
373 static void
374 print_all_lattices (FILE * f, bool dump_sources, bool dump_benefits)
376 struct cgraph_node *node;
377 int i, count;
379 fprintf (f, "\nLattices:\n");
380 FOR_EACH_FUNCTION_WITH_GIMPLE_BODY (node)
382 struct ipa_node_params *info;
384 info = IPA_NODE_REF (node);
385 fprintf (f, " Node: %s/%i:\n", node->name (),
386 node->order);
387 count = ipa_get_param_count (info);
388 for (i = 0; i < count; i++)
390 struct ipcp_agg_lattice *aglat;
391 struct ipcp_param_lattices *plats = ipa_get_parm_lattices (info, i);
392 fprintf (f, " param [%d]: ", i);
393 print_lattice (f, &plats->itself, dump_sources, dump_benefits);
395 if (plats->virt_call)
396 fprintf (f, " virt_call flag set\n");
398 if (plats->aggs_bottom)
400 fprintf (f, " AGGS BOTTOM\n");
401 continue;
403 if (plats->aggs_contain_variable)
404 fprintf (f, " AGGS VARIABLE\n");
405 for (aglat = plats->aggs; aglat; aglat = aglat->next)
407 fprintf (f, " %soffset " HOST_WIDE_INT_PRINT_DEC ": ",
408 plats->aggs_by_ref ? "ref " : "", aglat->offset);
409 print_lattice (f, aglat, dump_sources, dump_benefits);
415 /* Determine whether it is at all technically possible to create clones of NODE
416 and store this information in the ipa_node_params structure associated
417 with NODE. */
419 static void
420 determine_versionability (struct cgraph_node *node)
422 const char *reason = NULL;
424 /* There are a number of generic reasons functions cannot be versioned. We
425 also cannot remove parameters if there are type attributes such as fnspec
426 present. */
427 if (node->alias || node->thunk.thunk_p)
428 reason = "alias or thunk";
429 else if (!node->local.versionable)
430 reason = "not a tree_versionable_function";
431 else if (cgraph_function_body_availability (node) <= AVAIL_OVERWRITABLE)
432 reason = "insufficient body availability";
433 else if (!opt_for_fn (node->decl, optimize)
434 || !opt_for_fn (node->decl, flag_ipa_cp))
435 reason = "non-optimized function";
436 else if (lookup_attribute ("omp declare simd", DECL_ATTRIBUTES (node->decl)))
438 /* Ideally we should clone the SIMD clones themselves and create
439 vector copies of them, so IPA-cp and SIMD clones can happily
440 coexist, but that may not be worth the effort. */
441 reason = "function has SIMD clones";
443 /* Don't clone decls local to a comdat group; it breaks and for C++
444 decloned constructors, inlining is always better anyway. */
445 else if (symtab_comdat_local_p (node))
446 reason = "comdat-local function";
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 node->name (), node->order, 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 node->name ());
523 return false;
526 if (!optimize_function_for_speed_p (DECL_STRUCT_FUNCTION (node->decl)))
528 if (dump_file)
529 fprintf (dump_file, "Not considering %s for cloning; "
530 "optimizing it for size.\n",
531 node->name ());
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 node->name ());
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 node->name ());
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 node->name ());
565 return false;
567 if (dump_file)
568 fprintf (dump_file, "Considering %s for cloning.\n",
569 node->name ());
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 node->name (), node->order,
731 disable ? "BOTTOM" : "VARIABLE");
734 for (ie = node->indirect_calls; ie; ie = ie->next_callee)
735 if (ie->indirect_info->polymorphic
736 && ie->indirect_info->param_index >= 0)
738 gcc_checking_assert (ie->indirect_info->param_index >= 0);
739 ipa_get_parm_lattices (info,
740 ie->indirect_info->param_index)->virt_call = 1;
744 /* Return the result of a (possibly arithmetic) pass through jump function
745 JFUNC on the constant value INPUT. Return NULL_TREE if that cannot be
746 determined or be considered an interprocedural invariant. */
748 static tree
749 ipa_get_jf_pass_through_result (struct ipa_jump_func *jfunc, tree input)
751 tree restype, res;
753 if (TREE_CODE (input) == TREE_BINFO)
755 if (ipa_get_jf_pass_through_type_preserved (jfunc))
757 gcc_checking_assert (ipa_get_jf_pass_through_operation (jfunc)
758 == NOP_EXPR);
759 return input;
761 return NULL_TREE;
764 if (ipa_get_jf_pass_through_operation (jfunc) == NOP_EXPR)
765 return input;
767 gcc_checking_assert (is_gimple_ip_invariant (input));
768 if (TREE_CODE_CLASS (ipa_get_jf_pass_through_operation (jfunc))
769 == tcc_comparison)
770 restype = boolean_type_node;
771 else
772 restype = TREE_TYPE (input);
773 res = fold_binary (ipa_get_jf_pass_through_operation (jfunc), restype,
774 input, ipa_get_jf_pass_through_operand (jfunc));
776 if (res && !is_gimple_ip_invariant (res))
777 return NULL_TREE;
779 return res;
782 /* Return the result of an ancestor jump function JFUNC on the constant value
783 INPUT. Return NULL_TREE if that cannot be determined. */
785 static tree
786 ipa_get_jf_ancestor_result (struct ipa_jump_func *jfunc, tree input)
788 if (TREE_CODE (input) == TREE_BINFO)
790 if (!ipa_get_jf_ancestor_type_preserved (jfunc))
791 return NULL;
792 return get_binfo_at_offset (input,
793 ipa_get_jf_ancestor_offset (jfunc),
794 ipa_get_jf_ancestor_type (jfunc));
796 else if (TREE_CODE (input) == ADDR_EXPR)
798 tree t = TREE_OPERAND (input, 0);
799 t = build_ref_for_offset (EXPR_LOCATION (t), t,
800 ipa_get_jf_ancestor_offset (jfunc),
801 ipa_get_jf_ancestor_type (jfunc)
802 ? ipa_get_jf_ancestor_type (jfunc)
803 : ptr_type_node, NULL, false);
804 return build_fold_addr_expr (t);
806 else
807 return NULL_TREE;
810 /* Determine whether JFUNC evaluates to a known value (that is either a
811 constant or a binfo) and if so, return it. Otherwise return NULL. INFO
812 describes the caller node so that pass-through jump functions can be
813 evaluated. */
815 tree
816 ipa_value_from_jfunc (struct ipa_node_params *info, struct ipa_jump_func *jfunc)
818 if (jfunc->type == IPA_JF_CONST)
819 return ipa_get_jf_constant (jfunc);
820 else if (jfunc->type == IPA_JF_KNOWN_TYPE)
821 return ipa_binfo_from_known_type_jfunc (jfunc);
822 else if (jfunc->type == IPA_JF_PASS_THROUGH
823 || jfunc->type == IPA_JF_ANCESTOR)
825 tree input;
826 int idx;
828 if (jfunc->type == IPA_JF_PASS_THROUGH)
829 idx = ipa_get_jf_pass_through_formal_id (jfunc);
830 else
831 idx = ipa_get_jf_ancestor_formal_id (jfunc);
833 if (info->ipcp_orig_node)
834 input = info->known_vals[idx];
835 else
837 struct ipcp_lattice *lat;
839 if (!info->lattices)
841 gcc_checking_assert (!flag_ipa_cp);
842 return NULL_TREE;
844 lat = ipa_get_scalar_lat (info, idx);
845 if (!ipa_lat_is_single_const (lat))
846 return NULL_TREE;
847 input = lat->values->value;
850 if (!input)
851 return NULL_TREE;
853 if (jfunc->type == IPA_JF_PASS_THROUGH)
854 return ipa_get_jf_pass_through_result (jfunc, input);
855 else
856 return ipa_get_jf_ancestor_result (jfunc, input);
858 else
859 return NULL_TREE;
863 /* If checking is enabled, verify that no lattice is in the TOP state, i.e. not
864 bottom, not containing a variable component and without any known value at
865 the same time. */
867 DEBUG_FUNCTION void
868 ipcp_verify_propagated_values (void)
870 struct cgraph_node *node;
872 FOR_EACH_FUNCTION_WITH_GIMPLE_BODY (node)
874 struct ipa_node_params *info = IPA_NODE_REF (node);
875 int i, count = ipa_get_param_count (info);
877 for (i = 0; i < count; i++)
879 struct ipcp_lattice *lat = ipa_get_scalar_lat (info, i);
881 if (!lat->bottom
882 && !lat->contains_variable
883 && lat->values_count == 0)
885 if (dump_file)
887 fprintf (dump_file, "\nIPA lattices after constant "
888 "propagation:\n");
889 print_all_lattices (dump_file, true, false);
892 gcc_unreachable ();
898 /* Return true iff X and Y should be considered equal values by IPA-CP. */
900 static bool
901 values_equal_for_ipcp_p (tree x, tree y)
903 gcc_checking_assert (x != NULL_TREE && y != NULL_TREE);
905 if (x == y)
906 return true;
908 if (TREE_CODE (x) == TREE_BINFO || TREE_CODE (y) == TREE_BINFO)
909 return false;
911 if (TREE_CODE (x) == ADDR_EXPR
912 && TREE_CODE (y) == ADDR_EXPR
913 && TREE_CODE (TREE_OPERAND (x, 0)) == CONST_DECL
914 && TREE_CODE (TREE_OPERAND (y, 0)) == CONST_DECL)
915 return operand_equal_p (DECL_INITIAL (TREE_OPERAND (x, 0)),
916 DECL_INITIAL (TREE_OPERAND (y, 0)), 0);
917 else
918 return operand_equal_p (x, y, 0);
921 /* Add a new value source to VAL, marking that a value comes from edge CS and
922 (if the underlying jump function is a pass-through or an ancestor one) from
923 a caller value SRC_VAL of a caller parameter described by SRC_INDEX. OFFSET
924 is negative if the source was the scalar value of the parameter itself or
925 the offset within an aggregate. */
927 static void
928 add_value_source (struct ipcp_value *val, struct cgraph_edge *cs,
929 struct ipcp_value *src_val, int src_idx, HOST_WIDE_INT offset)
931 struct ipcp_value_source *src;
933 src = (struct ipcp_value_source *) pool_alloc (ipcp_sources_pool);
934 src->offset = offset;
935 src->cs = cs;
936 src->val = src_val;
937 src->index = src_idx;
939 src->next = val->sources;
940 val->sources = src;
943 /* Try to add NEWVAL to LAT, potentially creating a new struct ipcp_value for
944 it. CS, SRC_VAL SRC_INDEX and OFFSET are meant for add_value_source and
945 have the same meaning. */
947 static bool
948 add_value_to_lattice (struct ipcp_lattice *lat, tree newval,
949 struct cgraph_edge *cs, struct ipcp_value *src_val,
950 int src_idx, HOST_WIDE_INT offset)
952 struct ipcp_value *val;
954 if (lat->bottom)
955 return false;
957 for (val = lat->values; val; val = val->next)
958 if (values_equal_for_ipcp_p (val->value, newval))
960 if (ipa_edge_within_scc (cs))
962 struct ipcp_value_source *s;
963 for (s = val->sources; s ; s = s->next)
964 if (s->cs == cs)
965 break;
966 if (s)
967 return false;
970 add_value_source (val, cs, src_val, src_idx, offset);
971 return false;
974 if (lat->values_count == PARAM_VALUE (PARAM_IPA_CP_VALUE_LIST_SIZE))
976 /* We can only free sources, not the values themselves, because sources
977 of other values in this this SCC might point to them. */
978 for (val = lat->values; val; val = val->next)
980 while (val->sources)
982 struct ipcp_value_source *src = val->sources;
983 val->sources = src->next;
984 pool_free (ipcp_sources_pool, src);
988 lat->values = NULL;
989 return set_lattice_to_bottom (lat);
992 lat->values_count++;
993 val = (struct ipcp_value *) pool_alloc (ipcp_values_pool);
994 memset (val, 0, sizeof (*val));
996 add_value_source (val, cs, src_val, src_idx, offset);
997 val->value = newval;
998 val->next = lat->values;
999 lat->values = val;
1000 return true;
1003 /* Like above but passes a special value of offset to distinguish that the
1004 origin is the scalar value of the parameter rather than a part of an
1005 aggregate. */
1007 static inline bool
1008 add_scalar_value_to_lattice (struct ipcp_lattice *lat, tree newval,
1009 struct cgraph_edge *cs,
1010 struct ipcp_value *src_val, int src_idx)
1012 return add_value_to_lattice (lat, newval, cs, src_val, src_idx, -1);
1015 /* Propagate values through a pass-through jump function JFUNC associated with
1016 edge CS, taking values from SRC_LAT and putting them into DEST_LAT. SRC_IDX
1017 is the index of the source parameter. */
1019 static bool
1020 propagate_vals_accross_pass_through (struct cgraph_edge *cs,
1021 struct ipa_jump_func *jfunc,
1022 struct ipcp_lattice *src_lat,
1023 struct ipcp_lattice *dest_lat,
1024 int src_idx)
1026 struct ipcp_value *src_val;
1027 bool ret = false;
1029 /* Do not create new values when propagating within an SCC because if there
1030 are arithmetic functions with circular dependencies, there is infinite
1031 number of them and we would just make lattices bottom. */
1032 if ((ipa_get_jf_pass_through_operation (jfunc) != NOP_EXPR)
1033 && ipa_edge_within_scc (cs))
1034 ret = set_lattice_contains_variable (dest_lat);
1035 else
1036 for (src_val = src_lat->values; src_val; src_val = src_val->next)
1038 tree cstval = ipa_get_jf_pass_through_result (jfunc, src_val->value);
1040 if (cstval)
1041 ret |= add_scalar_value_to_lattice (dest_lat, cstval, cs, src_val,
1042 src_idx);
1043 else
1044 ret |= set_lattice_contains_variable (dest_lat);
1047 return ret;
1050 /* Propagate values through an ancestor jump function JFUNC associated with
1051 edge CS, taking values from SRC_LAT and putting them into DEST_LAT. SRC_IDX
1052 is the index of the source parameter. */
1054 static bool
1055 propagate_vals_accross_ancestor (struct cgraph_edge *cs,
1056 struct ipa_jump_func *jfunc,
1057 struct ipcp_lattice *src_lat,
1058 struct ipcp_lattice *dest_lat,
1059 int src_idx)
1061 struct ipcp_value *src_val;
1062 bool ret = false;
1064 if (ipa_edge_within_scc (cs))
1065 return set_lattice_contains_variable (dest_lat);
1067 for (src_val = src_lat->values; src_val; src_val = src_val->next)
1069 tree t = ipa_get_jf_ancestor_result (jfunc, src_val->value);
1071 if (t)
1072 ret |= add_scalar_value_to_lattice (dest_lat, t, cs, src_val, src_idx);
1073 else
1074 ret |= set_lattice_contains_variable (dest_lat);
1077 return ret;
1080 /* Propagate scalar values across jump function JFUNC that is associated with
1081 edge CS and put the values into DEST_LAT. */
1083 static bool
1084 propagate_scalar_accross_jump_function (struct cgraph_edge *cs,
1085 struct ipa_jump_func *jfunc,
1086 struct ipcp_lattice *dest_lat)
1088 if (dest_lat->bottom)
1089 return false;
1091 if (jfunc->type == IPA_JF_CONST
1092 || jfunc->type == IPA_JF_KNOWN_TYPE)
1094 tree val;
1096 if (jfunc->type == IPA_JF_KNOWN_TYPE)
1098 val = ipa_binfo_from_known_type_jfunc (jfunc);
1099 if (!val)
1100 return set_lattice_contains_variable (dest_lat);
1102 else
1103 val = ipa_get_jf_constant (jfunc);
1104 return add_scalar_value_to_lattice (dest_lat, val, cs, NULL, 0);
1106 else if (jfunc->type == IPA_JF_PASS_THROUGH
1107 || jfunc->type == IPA_JF_ANCESTOR)
1109 struct ipa_node_params *caller_info = IPA_NODE_REF (cs->caller);
1110 struct ipcp_lattice *src_lat;
1111 int src_idx;
1112 bool ret;
1114 if (jfunc->type == IPA_JF_PASS_THROUGH)
1115 src_idx = ipa_get_jf_pass_through_formal_id (jfunc);
1116 else
1117 src_idx = ipa_get_jf_ancestor_formal_id (jfunc);
1119 src_lat = ipa_get_scalar_lat (caller_info, src_idx);
1120 if (src_lat->bottom)
1121 return set_lattice_contains_variable (dest_lat);
1123 /* If we would need to clone the caller and cannot, do not propagate. */
1124 if (!ipcp_versionable_function_p (cs->caller)
1125 && (src_lat->contains_variable
1126 || (src_lat->values_count > 1)))
1127 return set_lattice_contains_variable (dest_lat);
1129 if (jfunc->type == IPA_JF_PASS_THROUGH)
1130 ret = propagate_vals_accross_pass_through (cs, jfunc, src_lat,
1131 dest_lat, src_idx);
1132 else
1133 ret = propagate_vals_accross_ancestor (cs, jfunc, src_lat, dest_lat,
1134 src_idx);
1136 if (src_lat->contains_variable)
1137 ret |= set_lattice_contains_variable (dest_lat);
1139 return ret;
1142 /* TODO: We currently do not handle member method pointers in IPA-CP (we only
1143 use it for indirect inlining), we should propagate them too. */
1144 return set_lattice_contains_variable (dest_lat);
1147 /* If DEST_PLATS already has aggregate items, check that aggs_by_ref matches
1148 NEW_AGGS_BY_REF and if not, mark all aggs as bottoms and return true (in all
1149 other cases, return false). If there are no aggregate items, set
1150 aggs_by_ref to NEW_AGGS_BY_REF. */
1152 static bool
1153 set_check_aggs_by_ref (struct ipcp_param_lattices *dest_plats,
1154 bool new_aggs_by_ref)
1156 if (dest_plats->aggs)
1158 if (dest_plats->aggs_by_ref != new_aggs_by_ref)
1160 set_agg_lats_to_bottom (dest_plats);
1161 return true;
1164 else
1165 dest_plats->aggs_by_ref = new_aggs_by_ref;
1166 return false;
1169 /* Walk aggregate lattices in DEST_PLATS from ***AGLAT on, until ***aglat is an
1170 already existing lattice for the given OFFSET and SIZE, marking all skipped
1171 lattices as containing variable and checking for overlaps. If there is no
1172 already existing lattice for the OFFSET and VAL_SIZE, create one, initialize
1173 it with offset, size and contains_variable to PRE_EXISTING, and return true,
1174 unless there are too many already. If there are two many, return false. If
1175 there are overlaps turn whole DEST_PLATS to bottom and return false. If any
1176 skipped lattices were newly marked as containing variable, set *CHANGE to
1177 true. */
1179 static bool
1180 merge_agg_lats_step (struct ipcp_param_lattices *dest_plats,
1181 HOST_WIDE_INT offset, HOST_WIDE_INT val_size,
1182 struct ipcp_agg_lattice ***aglat,
1183 bool pre_existing, bool *change)
1185 gcc_checking_assert (offset >= 0);
1187 while (**aglat && (**aglat)->offset < offset)
1189 if ((**aglat)->offset + (**aglat)->size > offset)
1191 set_agg_lats_to_bottom (dest_plats);
1192 return false;
1194 *change |= set_lattice_contains_variable (**aglat);
1195 *aglat = &(**aglat)->next;
1198 if (**aglat && (**aglat)->offset == offset)
1200 if ((**aglat)->size != val_size
1201 || ((**aglat)->next
1202 && (**aglat)->next->offset < offset + val_size))
1204 set_agg_lats_to_bottom (dest_plats);
1205 return false;
1207 gcc_checking_assert (!(**aglat)->next
1208 || (**aglat)->next->offset >= offset + val_size);
1209 return true;
1211 else
1213 struct ipcp_agg_lattice *new_al;
1215 if (**aglat && (**aglat)->offset < offset + val_size)
1217 set_agg_lats_to_bottom (dest_plats);
1218 return false;
1220 if (dest_plats->aggs_count == PARAM_VALUE (PARAM_IPA_MAX_AGG_ITEMS))
1221 return false;
1222 dest_plats->aggs_count++;
1223 new_al = (struct ipcp_agg_lattice *) pool_alloc (ipcp_agg_lattice_pool);
1224 memset (new_al, 0, sizeof (*new_al));
1226 new_al->offset = offset;
1227 new_al->size = val_size;
1228 new_al->contains_variable = pre_existing;
1230 new_al->next = **aglat;
1231 **aglat = new_al;
1232 return true;
1236 /* Set all AGLAT and all other aggregate lattices reachable by next pointers as
1237 containing an unknown value. */
1239 static bool
1240 set_chain_of_aglats_contains_variable (struct ipcp_agg_lattice *aglat)
1242 bool ret = false;
1243 while (aglat)
1245 ret |= set_lattice_contains_variable (aglat);
1246 aglat = aglat->next;
1248 return ret;
1251 /* Merge existing aggregate lattices in SRC_PLATS to DEST_PLATS, subtracting
1252 DELTA_OFFSET. CS is the call graph edge and SRC_IDX the index of the source
1253 parameter used for lattice value sources. Return true if DEST_PLATS changed
1254 in any way. */
1256 static bool
1257 merge_aggregate_lattices (struct cgraph_edge *cs,
1258 struct ipcp_param_lattices *dest_plats,
1259 struct ipcp_param_lattices *src_plats,
1260 int src_idx, HOST_WIDE_INT offset_delta)
1262 bool pre_existing = dest_plats->aggs != NULL;
1263 struct ipcp_agg_lattice **dst_aglat;
1264 bool ret = false;
1266 if (set_check_aggs_by_ref (dest_plats, src_plats->aggs_by_ref))
1267 return true;
1268 if (src_plats->aggs_bottom)
1269 return set_agg_lats_contain_variable (dest_plats);
1270 if (src_plats->aggs_contain_variable)
1271 ret |= set_agg_lats_contain_variable (dest_plats);
1272 dst_aglat = &dest_plats->aggs;
1274 for (struct ipcp_agg_lattice *src_aglat = src_plats->aggs;
1275 src_aglat;
1276 src_aglat = src_aglat->next)
1278 HOST_WIDE_INT new_offset = src_aglat->offset - offset_delta;
1280 if (new_offset < 0)
1281 continue;
1282 if (merge_agg_lats_step (dest_plats, new_offset, src_aglat->size,
1283 &dst_aglat, pre_existing, &ret))
1285 struct ipcp_agg_lattice *new_al = *dst_aglat;
1287 dst_aglat = &(*dst_aglat)->next;
1288 if (src_aglat->bottom)
1290 ret |= set_lattice_contains_variable (new_al);
1291 continue;
1293 if (src_aglat->contains_variable)
1294 ret |= set_lattice_contains_variable (new_al);
1295 for (struct ipcp_value *val = src_aglat->values;
1296 val;
1297 val = val->next)
1298 ret |= add_value_to_lattice (new_al, val->value, cs, val, src_idx,
1299 src_aglat->offset);
1301 else if (dest_plats->aggs_bottom)
1302 return true;
1304 ret |= set_chain_of_aglats_contains_variable (*dst_aglat);
1305 return ret;
1308 /* Determine whether there is anything to propagate FROM SRC_PLATS through a
1309 pass-through JFUNC and if so, whether it has conform and conforms to the
1310 rules about propagating values passed by reference. */
1312 static bool
1313 agg_pass_through_permissible_p (struct ipcp_param_lattices *src_plats,
1314 struct ipa_jump_func *jfunc)
1316 return src_plats->aggs
1317 && (!src_plats->aggs_by_ref
1318 || ipa_get_jf_pass_through_agg_preserved (jfunc));
1321 /* Propagate scalar values across jump function JFUNC that is associated with
1322 edge CS and put the values into DEST_LAT. */
1324 static bool
1325 propagate_aggs_accross_jump_function (struct cgraph_edge *cs,
1326 struct ipa_jump_func *jfunc,
1327 struct ipcp_param_lattices *dest_plats)
1329 bool ret = false;
1331 if (dest_plats->aggs_bottom)
1332 return false;
1334 if (jfunc->type == IPA_JF_PASS_THROUGH
1335 && ipa_get_jf_pass_through_operation (jfunc) == NOP_EXPR)
1337 struct ipa_node_params *caller_info = IPA_NODE_REF (cs->caller);
1338 int src_idx = ipa_get_jf_pass_through_formal_id (jfunc);
1339 struct ipcp_param_lattices *src_plats;
1341 src_plats = ipa_get_parm_lattices (caller_info, src_idx);
1342 if (agg_pass_through_permissible_p (src_plats, jfunc))
1344 /* Currently we do not produce clobber aggregate jump
1345 functions, replace with merging when we do. */
1346 gcc_assert (!jfunc->agg.items);
1347 ret |= merge_aggregate_lattices (cs, dest_plats, src_plats,
1348 src_idx, 0);
1350 else
1351 ret |= set_agg_lats_contain_variable (dest_plats);
1353 else if (jfunc->type == IPA_JF_ANCESTOR
1354 && ipa_get_jf_ancestor_agg_preserved (jfunc))
1356 struct ipa_node_params *caller_info = IPA_NODE_REF (cs->caller);
1357 int src_idx = ipa_get_jf_ancestor_formal_id (jfunc);
1358 struct ipcp_param_lattices *src_plats;
1360 src_plats = ipa_get_parm_lattices (caller_info, src_idx);
1361 if (src_plats->aggs && src_plats->aggs_by_ref)
1363 /* Currently we do not produce clobber aggregate jump
1364 functions, replace with merging when we do. */
1365 gcc_assert (!jfunc->agg.items);
1366 ret |= merge_aggregate_lattices (cs, dest_plats, src_plats, src_idx,
1367 ipa_get_jf_ancestor_offset (jfunc));
1369 else if (!src_plats->aggs_by_ref)
1370 ret |= set_agg_lats_to_bottom (dest_plats);
1371 else
1372 ret |= set_agg_lats_contain_variable (dest_plats);
1374 else if (jfunc->agg.items)
1376 bool pre_existing = dest_plats->aggs != NULL;
1377 struct ipcp_agg_lattice **aglat = &dest_plats->aggs;
1378 struct ipa_agg_jf_item *item;
1379 int i;
1381 if (set_check_aggs_by_ref (dest_plats, jfunc->agg.by_ref))
1382 return true;
1384 FOR_EACH_VEC_ELT (*jfunc->agg.items, i, item)
1386 HOST_WIDE_INT val_size;
1388 if (item->offset < 0)
1389 continue;
1390 gcc_checking_assert (is_gimple_ip_invariant (item->value));
1391 val_size = tree_to_uhwi (TYPE_SIZE (TREE_TYPE (item->value)));
1393 if (merge_agg_lats_step (dest_plats, item->offset, val_size,
1394 &aglat, pre_existing, &ret))
1396 ret |= add_value_to_lattice (*aglat, item->value, cs, NULL, 0, 0);
1397 aglat = &(*aglat)->next;
1399 else if (dest_plats->aggs_bottom)
1400 return true;
1403 ret |= set_chain_of_aglats_contains_variable (*aglat);
1405 else
1406 ret |= set_agg_lats_contain_variable (dest_plats);
1408 return ret;
1411 /* Propagate constants from the caller to the callee of CS. INFO describes the
1412 caller. */
1414 static bool
1415 propagate_constants_accross_call (struct cgraph_edge *cs)
1417 struct ipa_node_params *callee_info;
1418 enum availability availability;
1419 struct cgraph_node *callee, *alias_or_thunk;
1420 struct ipa_edge_args *args;
1421 bool ret = false;
1422 int i, args_count, parms_count;
1424 callee = cgraph_function_node (cs->callee, &availability);
1425 if (!callee->definition)
1426 return false;
1427 gcc_checking_assert (cgraph_function_with_gimple_body_p (callee));
1428 callee_info = IPA_NODE_REF (callee);
1430 args = IPA_EDGE_REF (cs);
1431 args_count = ipa_get_cs_argument_count (args);
1432 parms_count = ipa_get_param_count (callee_info);
1433 if (parms_count == 0)
1434 return false;
1436 /* If this call goes through a thunk we must not propagate to the first (0th)
1437 parameter. However, we might need to uncover a thunk from below a series
1438 of aliases first. */
1439 alias_or_thunk = cs->callee;
1440 while (alias_or_thunk->alias)
1441 alias_or_thunk = cgraph_alias_target (alias_or_thunk);
1442 if (alias_or_thunk->thunk.thunk_p)
1444 ret |= set_all_contains_variable (ipa_get_parm_lattices (callee_info,
1445 0));
1446 i = 1;
1448 else
1449 i = 0;
1451 for (; (i < args_count) && (i < parms_count); i++)
1453 struct ipa_jump_func *jump_func = ipa_get_ith_jump_func (args, i);
1454 struct ipcp_param_lattices *dest_plats;
1456 dest_plats = ipa_get_parm_lattices (callee_info, i);
1457 if (availability == AVAIL_OVERWRITABLE)
1458 ret |= set_all_contains_variable (dest_plats);
1459 else
1461 ret |= propagate_scalar_accross_jump_function (cs, jump_func,
1462 &dest_plats->itself);
1463 ret |= propagate_aggs_accross_jump_function (cs, jump_func,
1464 dest_plats);
1467 for (; i < parms_count; i++)
1468 ret |= set_all_contains_variable (ipa_get_parm_lattices (callee_info, i));
1470 return ret;
1473 /* If an indirect edge IE can be turned into a direct one based on KNOWN_VALS
1474 (which can contain both constants and binfos), KNOWN_BINFOS, KNOWN_AGGS or
1475 AGG_REPS return the destination. The latter three can be NULL. If AGG_REPS
1476 is not NULL, KNOWN_AGGS is ignored. */
1478 static tree
1479 ipa_get_indirect_edge_target_1 (struct cgraph_edge *ie,
1480 vec<tree> known_vals,
1481 vec<tree> known_binfos,
1482 vec<ipa_agg_jump_function_p> known_aggs,
1483 struct ipa_agg_replacement_value *agg_reps)
1485 int param_index = ie->indirect_info->param_index;
1486 HOST_WIDE_INT token, anc_offset;
1487 tree otr_type;
1488 tree t;
1489 tree target = NULL;
1491 if (param_index == -1
1492 || known_vals.length () <= (unsigned int) param_index)
1493 return NULL_TREE;
1495 if (!ie->indirect_info->polymorphic)
1497 tree t;
1499 if (ie->indirect_info->agg_contents)
1501 if (agg_reps)
1503 t = NULL;
1504 while (agg_reps)
1506 if (agg_reps->index == param_index
1507 && agg_reps->offset == ie->indirect_info->offset
1508 && agg_reps->by_ref == ie->indirect_info->by_ref)
1510 t = agg_reps->value;
1511 break;
1513 agg_reps = agg_reps->next;
1516 else if (known_aggs.length () > (unsigned int) param_index)
1518 struct ipa_agg_jump_function *agg;
1519 agg = known_aggs[param_index];
1520 t = ipa_find_agg_cst_for_param (agg, ie->indirect_info->offset,
1521 ie->indirect_info->by_ref);
1523 else
1524 t = NULL;
1526 else
1527 t = known_vals[param_index];
1529 if (t &&
1530 TREE_CODE (t) == ADDR_EXPR
1531 && TREE_CODE (TREE_OPERAND (t, 0)) == FUNCTION_DECL)
1532 return TREE_OPERAND (t, 0);
1533 else
1534 return NULL_TREE;
1537 if (!flag_devirtualize)
1538 return NULL_TREE;
1540 gcc_assert (!ie->indirect_info->agg_contents);
1541 token = ie->indirect_info->otr_token;
1542 anc_offset = ie->indirect_info->offset;
1543 otr_type = ie->indirect_info->otr_type;
1545 t = NULL;
1547 /* Try to work out value of virtual table pointer value in replacemnets. */
1548 if (!t && agg_reps && !ie->indirect_info->by_ref)
1550 while (agg_reps)
1552 if (agg_reps->index == param_index
1553 && agg_reps->offset == ie->indirect_info->offset
1554 && agg_reps->by_ref)
1556 t = agg_reps->value;
1557 break;
1559 agg_reps = agg_reps->next;
1563 /* Try to work out value of virtual table pointer value in known
1564 aggregate values. */
1565 if (!t && known_aggs.length () > (unsigned int) param_index
1566 && !ie->indirect_info->by_ref)
1568 struct ipa_agg_jump_function *agg;
1569 agg = known_aggs[param_index];
1570 t = ipa_find_agg_cst_for_param (agg, ie->indirect_info->offset,
1571 true);
1574 /* If we found the virtual table pointer, lookup the target. */
1575 if (t)
1577 tree vtable;
1578 unsigned HOST_WIDE_INT offset;
1579 if (vtable_pointer_value_to_vtable (t, &vtable, &offset))
1581 target = gimple_get_virt_method_for_vtable (ie->indirect_info->otr_token,
1582 vtable, offset);
1583 if (target)
1585 if ((TREE_CODE (TREE_TYPE (target)) == FUNCTION_TYPE
1586 && DECL_FUNCTION_CODE (target) == BUILT_IN_UNREACHABLE)
1587 || !possible_polymorphic_call_target_p
1588 (ie, cgraph_get_node (target)))
1590 if (dump_file)
1591 fprintf (dump_file,
1592 "Type inconsident devirtualization: %s/%i->%s\n",
1593 ie->caller->name (), ie->caller->order,
1594 IDENTIFIER_POINTER (DECL_ASSEMBLER_NAME (target)));
1595 target = builtin_decl_implicit (BUILT_IN_UNREACHABLE);
1596 cgraph_get_create_node (target);
1598 return target;
1603 /* Did we work out BINFO via type propagation? */
1604 if (!t && known_binfos.length () > (unsigned int) param_index)
1605 t = known_binfos[param_index];
1606 /* Or do we know the constant value of pointer? */
1607 if (!t)
1608 t = known_vals[param_index];
1609 if (!t)
1610 return NULL_TREE;
1612 if (TREE_CODE (t) != TREE_BINFO)
1614 ipa_polymorphic_call_context context;
1615 vec <cgraph_node *>targets;
1616 bool final;
1618 if (!get_polymorphic_call_info_from_invariant
1619 (&context, t, ie->indirect_info->otr_type,
1620 anc_offset))
1621 return NULL_TREE;
1622 targets = possible_polymorphic_call_targets
1623 (ie->indirect_info->otr_type,
1624 ie->indirect_info->otr_token,
1625 context, &final);
1626 if (!final || targets.length () > 1)
1627 return NULL_TREE;
1628 if (targets.length () == 1)
1629 target = targets[0]->decl;
1630 else
1631 target = builtin_decl_implicit (BUILT_IN_UNREACHABLE);
1633 else
1635 tree binfo;
1637 binfo = get_binfo_at_offset (t, anc_offset, otr_type);
1638 if (!binfo)
1639 return NULL_TREE;
1640 target = gimple_get_virt_method_for_binfo (token, binfo);
1643 if (target && !possible_polymorphic_call_target_p (ie,
1644 cgraph_get_node (target)))
1646 if (dump_file)
1647 fprintf (dump_file,
1648 "Type inconsident devirtualization: %s/%i->%s\n",
1649 ie->caller->name (), ie->caller->order,
1650 IDENTIFIER_POINTER (DECL_ASSEMBLER_NAME (target)));
1651 target = builtin_decl_implicit (BUILT_IN_UNREACHABLE);
1652 cgraph_get_create_node (target);
1655 return target;
1659 /* If an indirect edge IE can be turned into a direct one based on KNOWN_VALS
1660 (which can contain both constants and binfos), KNOWN_BINFOS (which can be
1661 NULL) or KNOWN_AGGS (which also can be NULL) return the destination. */
1663 tree
1664 ipa_get_indirect_edge_target (struct cgraph_edge *ie,
1665 vec<tree> known_vals,
1666 vec<tree> known_binfos,
1667 vec<ipa_agg_jump_function_p> known_aggs)
1669 return ipa_get_indirect_edge_target_1 (ie, known_vals, known_binfos,
1670 known_aggs, NULL);
1673 /* Calculate devirtualization time bonus for NODE, assuming we know KNOWN_CSTS
1674 and KNOWN_BINFOS. */
1676 static int
1677 devirtualization_time_bonus (struct cgraph_node *node,
1678 vec<tree> known_csts,
1679 vec<tree> known_binfos,
1680 vec<ipa_agg_jump_function_p> known_aggs)
1682 struct cgraph_edge *ie;
1683 int res = 0;
1685 for (ie = node->indirect_calls; ie; ie = ie->next_callee)
1687 struct cgraph_node *callee;
1688 struct inline_summary *isummary;
1689 tree target;
1691 target = ipa_get_indirect_edge_target (ie, known_csts, known_binfos,
1692 known_aggs);
1693 if (!target)
1694 continue;
1696 /* Only bare minimum benefit for clearly un-inlineable targets. */
1697 res += 1;
1698 callee = cgraph_get_node (target);
1699 if (!callee || !callee->definition)
1700 continue;
1701 isummary = inline_summary (callee);
1702 if (!isummary->inlinable)
1703 continue;
1705 /* FIXME: The values below need re-considering and perhaps also
1706 integrating into the cost metrics, at lest in some very basic way. */
1707 if (isummary->size <= MAX_INLINE_INSNS_AUTO / 4)
1708 res += 31;
1709 else if (isummary->size <= MAX_INLINE_INSNS_AUTO / 2)
1710 res += 15;
1711 else if (isummary->size <= MAX_INLINE_INSNS_AUTO
1712 || DECL_DECLARED_INLINE_P (callee->decl))
1713 res += 7;
1716 return res;
1719 /* Return time bonus incurred because of HINTS. */
1721 static int
1722 hint_time_bonus (inline_hints hints)
1724 int result = 0;
1725 if (hints & (INLINE_HINT_loop_iterations | INLINE_HINT_loop_stride))
1726 result += PARAM_VALUE (PARAM_IPA_CP_LOOP_HINT_BONUS);
1727 if (hints & INLINE_HINT_array_index)
1728 result += PARAM_VALUE (PARAM_IPA_CP_ARRAY_INDEX_HINT_BONUS);
1729 return result;
1732 /* Return true if cloning NODE is a good idea, given the estimated TIME_BENEFIT
1733 and SIZE_COST and with the sum of frequencies of incoming edges to the
1734 potential new clone in FREQUENCIES. */
1736 static bool
1737 good_cloning_opportunity_p (struct cgraph_node *node, int time_benefit,
1738 int freq_sum, gcov_type count_sum, int size_cost)
1740 if (time_benefit == 0
1741 || !flag_ipa_cp_clone
1742 || !optimize_function_for_speed_p (DECL_STRUCT_FUNCTION (node->decl)))
1743 return false;
1745 gcc_assert (size_cost > 0);
1747 if (max_count)
1749 int factor = (count_sum * 1000) / max_count;
1750 HOST_WIDEST_INT evaluation = (((HOST_WIDEST_INT) time_benefit * factor)
1751 / size_cost);
1753 if (dump_file && (dump_flags & TDF_DETAILS))
1754 fprintf (dump_file, " good_cloning_opportunity_p (time: %i, "
1755 "size: %i, count_sum: " HOST_WIDE_INT_PRINT_DEC
1756 ") -> evaluation: " HOST_WIDEST_INT_PRINT_DEC
1757 ", threshold: %i\n",
1758 time_benefit, size_cost, (HOST_WIDE_INT) count_sum,
1759 evaluation, PARAM_VALUE (PARAM_IPA_CP_EVAL_THRESHOLD));
1761 return evaluation >= PARAM_VALUE (PARAM_IPA_CP_EVAL_THRESHOLD);
1763 else
1765 HOST_WIDEST_INT evaluation = (((HOST_WIDEST_INT) time_benefit * freq_sum)
1766 / size_cost);
1768 if (dump_file && (dump_flags & TDF_DETAILS))
1769 fprintf (dump_file, " good_cloning_opportunity_p (time: %i, "
1770 "size: %i, freq_sum: %i) -> evaluation: "
1771 HOST_WIDEST_INT_PRINT_DEC ", threshold: %i\n",
1772 time_benefit, size_cost, freq_sum, evaluation,
1773 PARAM_VALUE (PARAM_IPA_CP_EVAL_THRESHOLD));
1775 return evaluation >= PARAM_VALUE (PARAM_IPA_CP_EVAL_THRESHOLD);
1779 /* Return all context independent values from aggregate lattices in PLATS in a
1780 vector. Return NULL if there are none. */
1782 static vec<ipa_agg_jf_item, va_gc> *
1783 context_independent_aggregate_values (struct ipcp_param_lattices *plats)
1785 vec<ipa_agg_jf_item, va_gc> *res = NULL;
1787 if (plats->aggs_bottom
1788 || plats->aggs_contain_variable
1789 || plats->aggs_count == 0)
1790 return NULL;
1792 for (struct ipcp_agg_lattice *aglat = plats->aggs;
1793 aglat;
1794 aglat = aglat->next)
1795 if (ipa_lat_is_single_const (aglat))
1797 struct ipa_agg_jf_item item;
1798 item.offset = aglat->offset;
1799 item.value = aglat->values->value;
1800 vec_safe_push (res, item);
1802 return res;
1805 /* Allocate KNOWN_CSTS, KNOWN_BINFOS and, if non-NULL, KNOWN_AGGS and populate
1806 them with values of parameters that are known independent of the context.
1807 INFO describes the function. If REMOVABLE_PARAMS_COST is non-NULL, the
1808 movement cost of all removable parameters will be stored in it. */
1810 static bool
1811 gather_context_independent_values (struct ipa_node_params *info,
1812 vec<tree> *known_csts,
1813 vec<tree> *known_binfos,
1814 vec<ipa_agg_jump_function> *known_aggs,
1815 int *removable_params_cost)
1817 int i, count = ipa_get_param_count (info);
1818 bool ret = false;
1820 known_csts->create (0);
1821 known_binfos->create (0);
1822 known_csts->safe_grow_cleared (count);
1823 known_binfos->safe_grow_cleared (count);
1824 if (known_aggs)
1826 known_aggs->create (0);
1827 known_aggs->safe_grow_cleared (count);
1830 if (removable_params_cost)
1831 *removable_params_cost = 0;
1833 for (i = 0; i < count ; i++)
1835 struct ipcp_param_lattices *plats = ipa_get_parm_lattices (info, i);
1836 struct ipcp_lattice *lat = &plats->itself;
1838 if (ipa_lat_is_single_const (lat))
1840 struct ipcp_value *val = lat->values;
1841 if (TREE_CODE (val->value) != TREE_BINFO)
1843 (*known_csts)[i] = val->value;
1844 if (removable_params_cost)
1845 *removable_params_cost
1846 += estimate_move_cost (TREE_TYPE (val->value));
1847 ret = true;
1849 else if (plats->virt_call)
1851 (*known_binfos)[i] = val->value;
1852 ret = true;
1854 else if (removable_params_cost
1855 && !ipa_is_param_used (info, i))
1856 *removable_params_cost += ipa_get_param_move_cost (info, i);
1858 else if (removable_params_cost
1859 && !ipa_is_param_used (info, i))
1860 *removable_params_cost
1861 += ipa_get_param_move_cost (info, i);
1863 if (known_aggs)
1865 vec<ipa_agg_jf_item, va_gc> *agg_items;
1866 struct ipa_agg_jump_function *ajf;
1868 agg_items = context_independent_aggregate_values (plats);
1869 ajf = &(*known_aggs)[i];
1870 ajf->items = agg_items;
1871 ajf->by_ref = plats->aggs_by_ref;
1872 ret |= agg_items != NULL;
1876 return ret;
1879 /* The current interface in ipa-inline-analysis requires a pointer vector.
1880 Create it.
1882 FIXME: That interface should be re-worked, this is slightly silly. Still,
1883 I'd like to discuss how to change it first and this demonstrates the
1884 issue. */
1886 static vec<ipa_agg_jump_function_p>
1887 agg_jmp_p_vec_for_t_vec (vec<ipa_agg_jump_function> known_aggs)
1889 vec<ipa_agg_jump_function_p> ret;
1890 struct ipa_agg_jump_function *ajf;
1891 int i;
1893 ret.create (known_aggs.length ());
1894 FOR_EACH_VEC_ELT (known_aggs, i, ajf)
1895 ret.quick_push (ajf);
1896 return ret;
1899 /* Iterate over known values of parameters of NODE and estimate the local
1900 effects in terms of time and size they have. */
1902 static void
1903 estimate_local_effects (struct cgraph_node *node)
1905 struct ipa_node_params *info = IPA_NODE_REF (node);
1906 int i, count = ipa_get_param_count (info);
1907 vec<tree> known_csts, known_binfos;
1908 vec<ipa_agg_jump_function> known_aggs;
1909 vec<ipa_agg_jump_function_p> known_aggs_ptrs;
1910 bool always_const;
1911 int base_time = inline_summary (node)->time;
1912 int removable_params_cost;
1914 if (!count || !ipcp_versionable_function_p (node))
1915 return;
1917 if (dump_file && (dump_flags & TDF_DETAILS))
1918 fprintf (dump_file, "\nEstimating effects for %s/%i, base_time: %i.\n",
1919 node->name (), node->order, base_time);
1921 always_const = gather_context_independent_values (info, &known_csts,
1922 &known_binfos, &known_aggs,
1923 &removable_params_cost);
1924 known_aggs_ptrs = agg_jmp_p_vec_for_t_vec (known_aggs);
1925 if (always_const)
1927 struct caller_statistics stats;
1928 inline_hints hints;
1929 int time, size;
1931 init_caller_stats (&stats);
1932 cgraph_for_node_and_aliases (node, gather_caller_stats, &stats, false);
1933 estimate_ipcp_clone_size_and_time (node, known_csts, known_binfos,
1934 known_aggs_ptrs, &size, &time, &hints);
1935 time -= devirtualization_time_bonus (node, known_csts, known_binfos,
1936 known_aggs_ptrs);
1937 time -= hint_time_bonus (hints);
1938 time -= removable_params_cost;
1939 size -= stats.n_calls * removable_params_cost;
1941 if (dump_file)
1942 fprintf (dump_file, " - context independent values, size: %i, "
1943 "time_benefit: %i\n", size, base_time - time);
1945 if (size <= 0
1946 || cgraph_will_be_removed_from_program_if_no_direct_calls (node))
1948 info->do_clone_for_all_contexts = true;
1949 base_time = time;
1951 if (dump_file)
1952 fprintf (dump_file, " Decided to specialize for all "
1953 "known contexts, code not going to grow.\n");
1955 else if (good_cloning_opportunity_p (node, base_time - time,
1956 stats.freq_sum, stats.count_sum,
1957 size))
1959 if (size + overall_size <= max_new_size)
1961 info->do_clone_for_all_contexts = true;
1962 base_time = time;
1963 overall_size += size;
1965 if (dump_file)
1966 fprintf (dump_file, " Decided to specialize for all "
1967 "known contexts, growth deemed beneficial.\n");
1969 else if (dump_file && (dump_flags & TDF_DETAILS))
1970 fprintf (dump_file, " Not cloning for all contexts because "
1971 "max_new_size would be reached with %li.\n",
1972 size + overall_size);
1976 for (i = 0; i < count ; i++)
1978 struct ipcp_param_lattices *plats = ipa_get_parm_lattices (info, i);
1979 struct ipcp_lattice *lat = &plats->itself;
1980 struct ipcp_value *val;
1981 int emc;
1983 if (lat->bottom
1984 || !lat->values
1985 || known_csts[i]
1986 || known_binfos[i])
1987 continue;
1989 for (val = lat->values; val; val = val->next)
1991 int time, size, time_benefit;
1992 inline_hints hints;
1994 if (TREE_CODE (val->value) != TREE_BINFO)
1996 known_csts[i] = val->value;
1997 known_binfos[i] = NULL_TREE;
1998 emc = estimate_move_cost (TREE_TYPE (val->value));
2000 else if (plats->virt_call)
2002 known_csts[i] = NULL_TREE;
2003 known_binfos[i] = val->value;
2004 emc = 0;
2006 else
2007 continue;
2009 estimate_ipcp_clone_size_and_time (node, known_csts, known_binfos,
2010 known_aggs_ptrs, &size, &time,
2011 &hints);
2012 time_benefit = base_time - time
2013 + devirtualization_time_bonus (node, known_csts, known_binfos,
2014 known_aggs_ptrs)
2015 + hint_time_bonus (hints)
2016 + removable_params_cost + emc;
2018 gcc_checking_assert (size >=0);
2019 /* The inliner-heuristics based estimates may think that in certain
2020 contexts some functions do not have any size at all but we want
2021 all specializations to have at least a tiny cost, not least not to
2022 divide by zero. */
2023 if (size == 0)
2024 size = 1;
2026 if (dump_file && (dump_flags & TDF_DETAILS))
2028 fprintf (dump_file, " - estimates for value ");
2029 print_ipcp_constant_value (dump_file, val->value);
2030 fprintf (dump_file, " for ");
2031 ipa_dump_param (dump_file, info, i);
2032 fprintf (dump_file, ": time_benefit: %i, size: %i\n",
2033 time_benefit, size);
2036 val->local_time_benefit = time_benefit;
2037 val->local_size_cost = size;
2039 known_binfos[i] = NULL_TREE;
2040 known_csts[i] = NULL_TREE;
2043 for (i = 0; i < count ; i++)
2045 struct ipcp_param_lattices *plats = ipa_get_parm_lattices (info, i);
2046 struct ipa_agg_jump_function *ajf;
2047 struct ipcp_agg_lattice *aglat;
2049 if (plats->aggs_bottom || !plats->aggs)
2050 continue;
2052 ajf = &known_aggs[i];
2053 for (aglat = plats->aggs; aglat; aglat = aglat->next)
2055 struct ipcp_value *val;
2056 if (aglat->bottom || !aglat->values
2057 /* If the following is true, the one value is in known_aggs. */
2058 || (!plats->aggs_contain_variable
2059 && ipa_lat_is_single_const (aglat)))
2060 continue;
2062 for (val = aglat->values; val; val = val->next)
2064 int time, size, time_benefit;
2065 struct ipa_agg_jf_item item;
2066 inline_hints hints;
2068 item.offset = aglat->offset;
2069 item.value = val->value;
2070 vec_safe_push (ajf->items, item);
2072 estimate_ipcp_clone_size_and_time (node, known_csts, known_binfos,
2073 known_aggs_ptrs, &size, &time,
2074 &hints);
2075 time_benefit = base_time - time
2076 + devirtualization_time_bonus (node, known_csts, known_binfos,
2077 known_aggs_ptrs)
2078 + hint_time_bonus (hints);
2079 gcc_checking_assert (size >=0);
2080 if (size == 0)
2081 size = 1;
2083 if (dump_file && (dump_flags & TDF_DETAILS))
2085 fprintf (dump_file, " - estimates for value ");
2086 print_ipcp_constant_value (dump_file, val->value);
2087 fprintf (dump_file, " for ");
2088 ipa_dump_param (dump_file, info, i);
2089 fprintf (dump_file, "[%soffset: " HOST_WIDE_INT_PRINT_DEC
2090 "]: time_benefit: %i, size: %i\n",
2091 plats->aggs_by_ref ? "ref " : "",
2092 aglat->offset, time_benefit, size);
2095 val->local_time_benefit = time_benefit;
2096 val->local_size_cost = size;
2097 ajf->items->pop ();
2102 for (i = 0; i < count ; i++)
2103 vec_free (known_aggs[i].items);
2105 known_csts.release ();
2106 known_binfos.release ();
2107 known_aggs.release ();
2108 known_aggs_ptrs.release ();
2112 /* Add value CUR_VAL and all yet-unsorted values it is dependent on to the
2113 topological sort of values. */
2115 static void
2116 add_val_to_toposort (struct ipcp_value *cur_val)
2118 static int dfs_counter = 0;
2119 static struct ipcp_value *stack;
2120 struct ipcp_value_source *src;
2122 if (cur_val->dfs)
2123 return;
2125 dfs_counter++;
2126 cur_val->dfs = dfs_counter;
2127 cur_val->low_link = dfs_counter;
2129 cur_val->topo_next = stack;
2130 stack = cur_val;
2131 cur_val->on_stack = true;
2133 for (src = cur_val->sources; src; src = src->next)
2134 if (src->val)
2136 if (src->val->dfs == 0)
2138 add_val_to_toposort (src->val);
2139 if (src->val->low_link < cur_val->low_link)
2140 cur_val->low_link = src->val->low_link;
2142 else if (src->val->on_stack
2143 && src->val->dfs < cur_val->low_link)
2144 cur_val->low_link = src->val->dfs;
2147 if (cur_val->dfs == cur_val->low_link)
2149 struct ipcp_value *v, *scc_list = NULL;
2153 v = stack;
2154 stack = v->topo_next;
2155 v->on_stack = false;
2157 v->scc_next = scc_list;
2158 scc_list = v;
2160 while (v != cur_val);
2162 cur_val->topo_next = values_topo;
2163 values_topo = cur_val;
2167 /* Add all values in lattices associated with NODE to the topological sort if
2168 they are not there yet. */
2170 static void
2171 add_all_node_vals_to_toposort (struct cgraph_node *node)
2173 struct ipa_node_params *info = IPA_NODE_REF (node);
2174 int i, count = ipa_get_param_count (info);
2176 for (i = 0; i < count ; i++)
2178 struct ipcp_param_lattices *plats = ipa_get_parm_lattices (info, i);
2179 struct ipcp_lattice *lat = &plats->itself;
2180 struct ipcp_agg_lattice *aglat;
2181 struct ipcp_value *val;
2183 if (!lat->bottom)
2184 for (val = lat->values; val; val = val->next)
2185 add_val_to_toposort (val);
2187 if (!plats->aggs_bottom)
2188 for (aglat = plats->aggs; aglat; aglat = aglat->next)
2189 if (!aglat->bottom)
2190 for (val = aglat->values; val; val = val->next)
2191 add_val_to_toposort (val);
2195 /* One pass of constants propagation along the call graph edges, from callers
2196 to callees (requires topological ordering in TOPO), iterate over strongly
2197 connected components. */
2199 static void
2200 propagate_constants_topo (struct topo_info *topo)
2202 int i;
2204 for (i = topo->nnodes - 1; i >= 0; i--)
2206 unsigned j;
2207 struct cgraph_node *v, *node = topo->order[i];
2208 vec<cgraph_node_ptr> cycle_nodes = ipa_get_nodes_in_cycle (node);
2210 /* First, iteratively propagate within the strongly connected component
2211 until all lattices stabilize. */
2212 FOR_EACH_VEC_ELT (cycle_nodes, j, v)
2213 if (cgraph_function_with_gimple_body_p (v))
2214 push_node_to_stack (topo, v);
2216 v = pop_node_from_stack (topo);
2217 while (v)
2219 struct cgraph_edge *cs;
2221 for (cs = v->callees; cs; cs = cs->next_callee)
2222 if (ipa_edge_within_scc (cs)
2223 && propagate_constants_accross_call (cs))
2224 push_node_to_stack (topo, cs->callee);
2225 v = pop_node_from_stack (topo);
2228 /* Afterwards, propagate along edges leading out of the SCC, calculates
2229 the local effects of the discovered constants and all valid values to
2230 their topological sort. */
2231 FOR_EACH_VEC_ELT (cycle_nodes, j, v)
2232 if (cgraph_function_with_gimple_body_p (v))
2234 struct cgraph_edge *cs;
2236 estimate_local_effects (v);
2237 add_all_node_vals_to_toposort (v);
2238 for (cs = v->callees; cs; cs = cs->next_callee)
2239 if (!ipa_edge_within_scc (cs))
2240 propagate_constants_accross_call (cs);
2242 cycle_nodes.release ();
2247 /* Return the sum of A and B if none of them is bigger than INT_MAX/2, return
2248 the bigger one if otherwise. */
2250 static int
2251 safe_add (int a, int b)
2253 if (a > INT_MAX/2 || b > INT_MAX/2)
2254 return a > b ? a : b;
2255 else
2256 return a + b;
2260 /* Propagate the estimated effects of individual values along the topological
2261 from the dependent values to those they depend on. */
2263 static void
2264 propagate_effects (void)
2266 struct ipcp_value *base;
2268 for (base = values_topo; base; base = base->topo_next)
2270 struct ipcp_value_source *src;
2271 struct ipcp_value *val;
2272 int time = 0, size = 0;
2274 for (val = base; val; val = val->scc_next)
2276 time = safe_add (time,
2277 val->local_time_benefit + val->prop_time_benefit);
2278 size = safe_add (size, val->local_size_cost + val->prop_size_cost);
2281 for (val = base; val; val = val->scc_next)
2282 for (src = val->sources; src; src = src->next)
2283 if (src->val
2284 && cgraph_maybe_hot_edge_p (src->cs))
2286 src->val->prop_time_benefit = safe_add (time,
2287 src->val->prop_time_benefit);
2288 src->val->prop_size_cost = safe_add (size,
2289 src->val->prop_size_cost);
2295 /* Propagate constants, binfos and their effects from the summaries
2296 interprocedurally. */
2298 static void
2299 ipcp_propagate_stage (struct topo_info *topo)
2301 struct cgraph_node *node;
2303 if (dump_file)
2304 fprintf (dump_file, "\n Propagating constants:\n\n");
2306 if (in_lto_p)
2307 ipa_update_after_lto_read ();
2310 FOR_EACH_DEFINED_FUNCTION (node)
2312 struct ipa_node_params *info = IPA_NODE_REF (node);
2314 determine_versionability (node);
2315 if (cgraph_function_with_gimple_body_p (node))
2317 info->lattices = XCNEWVEC (struct ipcp_param_lattices,
2318 ipa_get_param_count (info));
2319 initialize_node_lattices (node);
2321 if (node->definition && !node->alias)
2322 overall_size += inline_summary (node)->self_size;
2323 if (node->count > max_count)
2324 max_count = node->count;
2327 max_new_size = overall_size;
2328 if (max_new_size < PARAM_VALUE (PARAM_LARGE_UNIT_INSNS))
2329 max_new_size = PARAM_VALUE (PARAM_LARGE_UNIT_INSNS);
2330 max_new_size += max_new_size * PARAM_VALUE (PARAM_IPCP_UNIT_GROWTH) / 100 + 1;
2332 if (dump_file)
2333 fprintf (dump_file, "\noverall_size: %li, max_new_size: %li\n",
2334 overall_size, max_new_size);
2336 propagate_constants_topo (topo);
2337 #ifdef ENABLE_CHECKING
2338 ipcp_verify_propagated_values ();
2339 #endif
2340 propagate_effects ();
2342 if (dump_file)
2344 fprintf (dump_file, "\nIPA lattices after all propagation:\n");
2345 print_all_lattices (dump_file, (dump_flags & TDF_DETAILS), true);
2349 /* Discover newly direct outgoing edges from NODE which is a new clone with
2350 known KNOWN_VALS and make them direct. */
2352 static void
2353 ipcp_discover_new_direct_edges (struct cgraph_node *node,
2354 vec<tree> known_vals,
2355 struct ipa_agg_replacement_value *aggvals)
2357 struct cgraph_edge *ie, *next_ie;
2358 bool found = false;
2360 for (ie = node->indirect_calls; ie; ie = next_ie)
2362 tree target;
2364 next_ie = ie->next_callee;
2365 target = ipa_get_indirect_edge_target_1 (ie, known_vals, vNULL, vNULL,
2366 aggvals);
2367 if (target)
2369 bool agg_contents = ie->indirect_info->agg_contents;
2370 bool polymorphic = ie->indirect_info->polymorphic;
2371 int param_index = ie->indirect_info->param_index;
2372 struct cgraph_edge *cs = ipa_make_edge_direct_to_target (ie, target);
2373 found = true;
2375 if (cs && !agg_contents && !polymorphic)
2377 struct ipa_node_params *info = IPA_NODE_REF (node);
2378 int c = ipa_get_controlled_uses (info, param_index);
2379 if (c != IPA_UNDESCRIBED_USE)
2381 struct ipa_ref *to_del;
2383 c--;
2384 ipa_set_controlled_uses (info, param_index, c);
2385 if (dump_file && (dump_flags & TDF_DETAILS))
2386 fprintf (dump_file, " controlled uses count of param "
2387 "%i bumped down to %i\n", param_index, c);
2388 if (c == 0
2389 && (to_del = ipa_find_reference (node,
2390 cs->callee,
2391 NULL, 0)))
2393 if (dump_file && (dump_flags & TDF_DETAILS))
2394 fprintf (dump_file, " and even removing its "
2395 "cloning-created reference\n");
2396 ipa_remove_reference (to_del);
2402 /* Turning calls to direct calls will improve overall summary. */
2403 if (found)
2404 inline_update_overall_summary (node);
2407 /* Vector of pointers which for linked lists of clones of an original crgaph
2408 edge. */
2410 static vec<cgraph_edge_p> next_edge_clone;
2411 static vec<cgraph_edge_p> prev_edge_clone;
2413 static inline void
2414 grow_edge_clone_vectors (void)
2416 if (next_edge_clone.length ()
2417 <= (unsigned) cgraph_edge_max_uid)
2418 next_edge_clone.safe_grow_cleared (cgraph_edge_max_uid + 1);
2419 if (prev_edge_clone.length ()
2420 <= (unsigned) cgraph_edge_max_uid)
2421 prev_edge_clone.safe_grow_cleared (cgraph_edge_max_uid + 1);
2424 /* Edge duplication hook to grow the appropriate linked list in
2425 next_edge_clone. */
2427 static void
2428 ipcp_edge_duplication_hook (struct cgraph_edge *src, struct cgraph_edge *dst,
2429 void *)
2431 grow_edge_clone_vectors ();
2433 struct cgraph_edge *old_next = next_edge_clone[src->uid];
2434 if (old_next)
2435 prev_edge_clone[old_next->uid] = dst;
2436 prev_edge_clone[dst->uid] = src;
2438 next_edge_clone[dst->uid] = old_next;
2439 next_edge_clone[src->uid] = dst;
2442 /* Hook that is called by cgraph.c when an edge is removed. */
2444 static void
2445 ipcp_edge_removal_hook (struct cgraph_edge *cs, void *)
2447 grow_edge_clone_vectors ();
2449 struct cgraph_edge *prev = prev_edge_clone[cs->uid];
2450 struct cgraph_edge *next = next_edge_clone[cs->uid];
2451 if (prev)
2452 next_edge_clone[prev->uid] = next;
2453 if (next)
2454 prev_edge_clone[next->uid] = prev;
2457 /* See if NODE is a clone with a known aggregate value at a given OFFSET of a
2458 parameter with the given INDEX. */
2460 static tree
2461 get_clone_agg_value (struct cgraph_node *node, HOST_WIDEST_INT offset,
2462 int index)
2464 struct ipa_agg_replacement_value *aggval;
2466 aggval = ipa_get_agg_replacements_for_node (node);
2467 while (aggval)
2469 if (aggval->offset == offset
2470 && aggval->index == index)
2471 return aggval->value;
2472 aggval = aggval->next;
2474 return NULL_TREE;
2477 /* Return true if edge CS does bring about the value described by SRC. */
2479 static bool
2480 cgraph_edge_brings_value_p (struct cgraph_edge *cs,
2481 struct ipcp_value_source *src)
2483 struct ipa_node_params *caller_info = IPA_NODE_REF (cs->caller);
2484 struct ipa_node_params *dst_info = IPA_NODE_REF (cs->callee);
2486 if ((dst_info->ipcp_orig_node && !dst_info->is_all_contexts_clone)
2487 || caller_info->node_dead)
2488 return false;
2489 if (!src->val)
2490 return true;
2492 if (caller_info->ipcp_orig_node)
2494 tree t;
2495 if (src->offset == -1)
2496 t = caller_info->known_vals[src->index];
2497 else
2498 t = get_clone_agg_value (cs->caller, src->offset, src->index);
2499 return (t != NULL_TREE
2500 && values_equal_for_ipcp_p (src->val->value, t));
2502 else
2504 struct ipcp_agg_lattice *aglat;
2505 struct ipcp_param_lattices *plats = ipa_get_parm_lattices (caller_info,
2506 src->index);
2507 if (src->offset == -1)
2508 return (ipa_lat_is_single_const (&plats->itself)
2509 && values_equal_for_ipcp_p (src->val->value,
2510 plats->itself.values->value));
2511 else
2513 if (plats->aggs_bottom || plats->aggs_contain_variable)
2514 return false;
2515 for (aglat = plats->aggs; aglat; aglat = aglat->next)
2516 if (aglat->offset == src->offset)
2517 return (ipa_lat_is_single_const (aglat)
2518 && values_equal_for_ipcp_p (src->val->value,
2519 aglat->values->value));
2521 return false;
2525 /* Get the next clone in the linked list of clones of an edge. */
2527 static inline struct cgraph_edge *
2528 get_next_cgraph_edge_clone (struct cgraph_edge *cs)
2530 return next_edge_clone[cs->uid];
2533 /* Given VAL, iterate over all its sources and if they still hold, add their
2534 edge frequency and their number into *FREQUENCY and *CALLER_COUNT
2535 respectively. */
2537 static bool
2538 get_info_about_necessary_edges (struct ipcp_value *val, int *freq_sum,
2539 gcov_type *count_sum, int *caller_count)
2541 struct ipcp_value_source *src;
2542 int freq = 0, count = 0;
2543 gcov_type cnt = 0;
2544 bool hot = false;
2546 for (src = val->sources; src; src = src->next)
2548 struct cgraph_edge *cs = src->cs;
2549 while (cs)
2551 if (cgraph_edge_brings_value_p (cs, src))
2553 count++;
2554 freq += cs->frequency;
2555 cnt += cs->count;
2556 hot |= cgraph_maybe_hot_edge_p (cs);
2558 cs = get_next_cgraph_edge_clone (cs);
2562 *freq_sum = freq;
2563 *count_sum = cnt;
2564 *caller_count = count;
2565 return hot;
2568 /* Return a vector of incoming edges that do bring value VAL. It is assumed
2569 their number is known and equal to CALLER_COUNT. */
2571 static vec<cgraph_edge_p>
2572 gather_edges_for_value (struct ipcp_value *val, int caller_count)
2574 struct ipcp_value_source *src;
2575 vec<cgraph_edge_p> ret;
2577 ret.create (caller_count);
2578 for (src = val->sources; src; src = src->next)
2580 struct cgraph_edge *cs = src->cs;
2581 while (cs)
2583 if (cgraph_edge_brings_value_p (cs, src))
2584 ret.quick_push (cs);
2585 cs = get_next_cgraph_edge_clone (cs);
2589 return ret;
2592 /* Construct a replacement map for a know VALUE for a formal parameter PARAM.
2593 Return it or NULL if for some reason it cannot be created. */
2595 static struct ipa_replace_map *
2596 get_replacement_map (struct ipa_node_params *info, tree value, int parm_num)
2598 struct ipa_replace_map *replace_map;
2601 replace_map = ggc_alloc_ipa_replace_map ();
2602 if (dump_file)
2604 fprintf (dump_file, " replacing ");
2605 ipa_dump_param (dump_file, info, parm_num);
2607 fprintf (dump_file, " with const ");
2608 print_generic_expr (dump_file, value, 0);
2609 fprintf (dump_file, "\n");
2611 replace_map->old_tree = NULL;
2612 replace_map->parm_num = parm_num;
2613 replace_map->new_tree = value;
2614 replace_map->replace_p = true;
2615 replace_map->ref_p = false;
2617 return replace_map;
2620 /* Dump new profiling counts */
2622 static void
2623 dump_profile_updates (struct cgraph_node *orig_node,
2624 struct cgraph_node *new_node)
2626 struct cgraph_edge *cs;
2628 fprintf (dump_file, " setting count of the specialized node to "
2629 HOST_WIDE_INT_PRINT_DEC "\n", (HOST_WIDE_INT) new_node->count);
2630 for (cs = new_node->callees; cs ; cs = cs->next_callee)
2631 fprintf (dump_file, " edge to %s has count "
2632 HOST_WIDE_INT_PRINT_DEC "\n",
2633 cs->callee->name (), (HOST_WIDE_INT) cs->count);
2635 fprintf (dump_file, " setting count of the original node to "
2636 HOST_WIDE_INT_PRINT_DEC "\n", (HOST_WIDE_INT) orig_node->count);
2637 for (cs = orig_node->callees; cs ; cs = cs->next_callee)
2638 fprintf (dump_file, " edge to %s is left with "
2639 HOST_WIDE_INT_PRINT_DEC "\n",
2640 cs->callee->name (), (HOST_WIDE_INT) cs->count);
2643 /* After a specialized NEW_NODE version of ORIG_NODE has been created, update
2644 their profile information to reflect this. */
2646 static void
2647 update_profiling_info (struct cgraph_node *orig_node,
2648 struct cgraph_node *new_node)
2650 struct cgraph_edge *cs;
2651 struct caller_statistics stats;
2652 gcov_type new_sum, orig_sum;
2653 gcov_type remainder, orig_node_count = orig_node->count;
2655 if (orig_node_count == 0)
2656 return;
2658 init_caller_stats (&stats);
2659 cgraph_for_node_and_aliases (orig_node, gather_caller_stats, &stats, false);
2660 orig_sum = stats.count_sum;
2661 init_caller_stats (&stats);
2662 cgraph_for_node_and_aliases (new_node, gather_caller_stats, &stats, false);
2663 new_sum = stats.count_sum;
2665 if (orig_node_count < orig_sum + new_sum)
2667 if (dump_file)
2668 fprintf (dump_file, " Problem: node %s/%i has too low count "
2669 HOST_WIDE_INT_PRINT_DEC " while the sum of incoming "
2670 "counts is " HOST_WIDE_INT_PRINT_DEC "\n",
2671 orig_node->name (), orig_node->order,
2672 (HOST_WIDE_INT) orig_node_count,
2673 (HOST_WIDE_INT) (orig_sum + new_sum));
2675 orig_node_count = (orig_sum + new_sum) * 12 / 10;
2676 if (dump_file)
2677 fprintf (dump_file, " proceeding by pretending it was "
2678 HOST_WIDE_INT_PRINT_DEC "\n",
2679 (HOST_WIDE_INT) orig_node_count);
2682 new_node->count = new_sum;
2683 remainder = orig_node_count - new_sum;
2684 orig_node->count = remainder;
2686 for (cs = new_node->callees; cs ; cs = cs->next_callee)
2687 if (cs->frequency)
2688 cs->count = apply_probability (cs->count,
2689 GCOV_COMPUTE_SCALE (new_sum,
2690 orig_node_count));
2691 else
2692 cs->count = 0;
2694 for (cs = orig_node->callees; cs ; cs = cs->next_callee)
2695 cs->count = apply_probability (cs->count,
2696 GCOV_COMPUTE_SCALE (remainder,
2697 orig_node_count));
2699 if (dump_file)
2700 dump_profile_updates (orig_node, new_node);
2703 /* Update the respective profile of specialized NEW_NODE and the original
2704 ORIG_NODE after additional edges with cumulative count sum REDIRECTED_SUM
2705 have been redirected to the specialized version. */
2707 static void
2708 update_specialized_profile (struct cgraph_node *new_node,
2709 struct cgraph_node *orig_node,
2710 gcov_type redirected_sum)
2712 struct cgraph_edge *cs;
2713 gcov_type new_node_count, orig_node_count = orig_node->count;
2715 if (dump_file)
2716 fprintf (dump_file, " the sum of counts of redirected edges is "
2717 HOST_WIDE_INT_PRINT_DEC "\n", (HOST_WIDE_INT) redirected_sum);
2718 if (orig_node_count == 0)
2719 return;
2721 gcc_assert (orig_node_count >= redirected_sum);
2723 new_node_count = new_node->count;
2724 new_node->count += redirected_sum;
2725 orig_node->count -= redirected_sum;
2727 for (cs = new_node->callees; cs ; cs = cs->next_callee)
2728 if (cs->frequency)
2729 cs->count += apply_probability (cs->count,
2730 GCOV_COMPUTE_SCALE (redirected_sum,
2731 new_node_count));
2732 else
2733 cs->count = 0;
2735 for (cs = orig_node->callees; cs ; cs = cs->next_callee)
2737 gcov_type dec = apply_probability (cs->count,
2738 GCOV_COMPUTE_SCALE (redirected_sum,
2739 orig_node_count));
2740 if (dec < cs->count)
2741 cs->count -= dec;
2742 else
2743 cs->count = 0;
2746 if (dump_file)
2747 dump_profile_updates (orig_node, new_node);
2750 /* Create a specialized version of NODE with known constants and types of
2751 parameters in KNOWN_VALS and redirect all edges in CALLERS to it. */
2753 static struct cgraph_node *
2754 create_specialized_node (struct cgraph_node *node,
2755 vec<tree> known_vals,
2756 struct ipa_agg_replacement_value *aggvals,
2757 vec<cgraph_edge_p> callers)
2759 struct ipa_node_params *new_info, *info = IPA_NODE_REF (node);
2760 vec<ipa_replace_map_p, va_gc> *replace_trees = NULL;
2761 struct ipa_agg_replacement_value *av;
2762 struct cgraph_node *new_node;
2763 int i, count = ipa_get_param_count (info);
2764 bitmap args_to_skip;
2766 gcc_assert (!info->ipcp_orig_node);
2768 if (node->local.can_change_signature)
2770 args_to_skip = BITMAP_GGC_ALLOC ();
2771 for (i = 0; i < count; i++)
2773 tree t = known_vals[i];
2775 if ((t && TREE_CODE (t) != TREE_BINFO)
2776 || !ipa_is_param_used (info, i))
2777 bitmap_set_bit (args_to_skip, i);
2780 else
2782 args_to_skip = NULL;
2783 if (dump_file && (dump_flags & TDF_DETAILS))
2784 fprintf (dump_file, " cannot change function signature\n");
2787 for (i = 0; i < count ; i++)
2789 tree t = known_vals[i];
2790 if (t && TREE_CODE (t) != TREE_BINFO)
2792 struct ipa_replace_map *replace_map;
2794 replace_map = get_replacement_map (info, t, i);
2795 if (replace_map)
2796 vec_safe_push (replace_trees, replace_map);
2800 new_node = cgraph_create_virtual_clone (node, callers, replace_trees,
2801 args_to_skip, "constprop");
2802 ipa_set_node_agg_value_chain (new_node, aggvals);
2803 for (av = aggvals; av; av = av->next)
2804 ipa_maybe_record_reference (new_node, av->value,
2805 IPA_REF_ADDR, NULL);
2807 if (dump_file && (dump_flags & TDF_DETAILS))
2809 fprintf (dump_file, " the new node is %s/%i.\n",
2810 new_node->name (), new_node->order);
2811 if (aggvals)
2812 ipa_dump_agg_replacement_values (dump_file, aggvals);
2814 ipa_check_create_node_params ();
2815 update_profiling_info (node, new_node);
2816 new_info = IPA_NODE_REF (new_node);
2817 new_info->ipcp_orig_node = node;
2818 new_info->known_vals = known_vals;
2820 ipcp_discover_new_direct_edges (new_node, known_vals, aggvals);
2822 callers.release ();
2823 return new_node;
2826 /* Given a NODE, and a subset of its CALLERS, try to populate blanks slots in
2827 KNOWN_VALS with constants and types that are also known for all of the
2828 CALLERS. */
2830 static void
2831 find_more_scalar_values_for_callers_subset (struct cgraph_node *node,
2832 vec<tree> known_vals,
2833 vec<cgraph_edge_p> callers)
2835 struct ipa_node_params *info = IPA_NODE_REF (node);
2836 int i, count = ipa_get_param_count (info);
2838 for (i = 0; i < count ; i++)
2840 struct cgraph_edge *cs;
2841 tree newval = NULL_TREE;
2842 int j;
2844 if (ipa_get_scalar_lat (info, i)->bottom || known_vals[i])
2845 continue;
2847 FOR_EACH_VEC_ELT (callers, j, cs)
2849 struct ipa_jump_func *jump_func;
2850 tree t;
2852 if (i >= ipa_get_cs_argument_count (IPA_EDGE_REF (cs)))
2854 newval = NULL_TREE;
2855 break;
2857 jump_func = ipa_get_ith_jump_func (IPA_EDGE_REF (cs), i);
2858 t = ipa_value_from_jfunc (IPA_NODE_REF (cs->caller), jump_func);
2859 if (!t
2860 || (newval
2861 && !values_equal_for_ipcp_p (t, newval)))
2863 newval = NULL_TREE;
2864 break;
2866 else
2867 newval = t;
2870 if (newval)
2872 if (dump_file && (dump_flags & TDF_DETAILS))
2874 fprintf (dump_file, " adding an extra known scalar value ");
2875 print_ipcp_constant_value (dump_file, newval);
2876 fprintf (dump_file, " for ");
2877 ipa_dump_param (dump_file, info, i);
2878 fprintf (dump_file, "\n");
2881 known_vals[i] = newval;
2886 /* Go through PLATS and create a vector of values consisting of values and
2887 offsets (minus OFFSET) of lattices that contain only a single value. */
2889 static vec<ipa_agg_jf_item>
2890 copy_plats_to_inter (struct ipcp_param_lattices *plats, HOST_WIDE_INT offset)
2892 vec<ipa_agg_jf_item> res = vNULL;
2894 if (!plats->aggs || plats->aggs_contain_variable || plats->aggs_bottom)
2895 return vNULL;
2897 for (struct ipcp_agg_lattice *aglat = plats->aggs; aglat; aglat = aglat->next)
2898 if (ipa_lat_is_single_const (aglat))
2900 struct ipa_agg_jf_item ti;
2901 ti.offset = aglat->offset - offset;
2902 ti.value = aglat->values->value;
2903 res.safe_push (ti);
2905 return res;
2908 /* Intersect all values in INTER with single value lattices in PLATS (while
2909 subtracting OFFSET). */
2911 static void
2912 intersect_with_plats (struct ipcp_param_lattices *plats,
2913 vec<ipa_agg_jf_item> *inter,
2914 HOST_WIDE_INT offset)
2916 struct ipcp_agg_lattice *aglat;
2917 struct ipa_agg_jf_item *item;
2918 int k;
2920 if (!plats->aggs || plats->aggs_contain_variable || plats->aggs_bottom)
2922 inter->release ();
2923 return;
2926 aglat = plats->aggs;
2927 FOR_EACH_VEC_ELT (*inter, k, item)
2929 bool found = false;
2930 if (!item->value)
2931 continue;
2932 while (aglat)
2934 if (aglat->offset - offset > item->offset)
2935 break;
2936 if (aglat->offset - offset == item->offset)
2938 gcc_checking_assert (item->value);
2939 if (values_equal_for_ipcp_p (item->value, aglat->values->value))
2940 found = true;
2941 break;
2943 aglat = aglat->next;
2945 if (!found)
2946 item->value = NULL_TREE;
2950 /* Copy agggregate replacement values of NODE (which is an IPA-CP clone) to the
2951 vector result while subtracting OFFSET from the individual value offsets. */
2953 static vec<ipa_agg_jf_item>
2954 agg_replacements_to_vector (struct cgraph_node *node, int index,
2955 HOST_WIDE_INT offset)
2957 struct ipa_agg_replacement_value *av;
2958 vec<ipa_agg_jf_item> res = vNULL;
2960 for (av = ipa_get_agg_replacements_for_node (node); av; av = av->next)
2961 if (av->index == index
2962 && (av->offset - offset) >= 0)
2964 struct ipa_agg_jf_item item;
2965 gcc_checking_assert (av->value);
2966 item.offset = av->offset - offset;
2967 item.value = av->value;
2968 res.safe_push (item);
2971 return res;
2974 /* Intersect all values in INTER with those that we have already scheduled to
2975 be replaced in parameter number INDEX of NODE, which is an IPA-CP clone
2976 (while subtracting OFFSET). */
2978 static void
2979 intersect_with_agg_replacements (struct cgraph_node *node, int index,
2980 vec<ipa_agg_jf_item> *inter,
2981 HOST_WIDE_INT offset)
2983 struct ipa_agg_replacement_value *srcvals;
2984 struct ipa_agg_jf_item *item;
2985 int i;
2987 srcvals = ipa_get_agg_replacements_for_node (node);
2988 if (!srcvals)
2990 inter->release ();
2991 return;
2994 FOR_EACH_VEC_ELT (*inter, i, item)
2996 struct ipa_agg_replacement_value *av;
2997 bool found = false;
2998 if (!item->value)
2999 continue;
3000 for (av = srcvals; av; av = av->next)
3002 gcc_checking_assert (av->value);
3003 if (av->index == index
3004 && av->offset - offset == item->offset)
3006 if (values_equal_for_ipcp_p (item->value, av->value))
3007 found = true;
3008 break;
3011 if (!found)
3012 item->value = NULL_TREE;
3016 /* Intersect values in INTER with aggregate values that come along edge CS to
3017 parameter number INDEX and return it. If INTER does not actually exist yet,
3018 copy all incoming values to it. If we determine we ended up with no values
3019 whatsoever, return a released vector. */
3021 static vec<ipa_agg_jf_item>
3022 intersect_aggregates_with_edge (struct cgraph_edge *cs, int index,
3023 vec<ipa_agg_jf_item> inter)
3025 struct ipa_jump_func *jfunc;
3026 jfunc = ipa_get_ith_jump_func (IPA_EDGE_REF (cs), index);
3027 if (jfunc->type == IPA_JF_PASS_THROUGH
3028 && ipa_get_jf_pass_through_operation (jfunc) == NOP_EXPR)
3030 struct ipa_node_params *caller_info = IPA_NODE_REF (cs->caller);
3031 int src_idx = ipa_get_jf_pass_through_formal_id (jfunc);
3033 if (caller_info->ipcp_orig_node)
3035 struct cgraph_node *orig_node = caller_info->ipcp_orig_node;
3036 struct ipcp_param_lattices *orig_plats;
3037 orig_plats = ipa_get_parm_lattices (IPA_NODE_REF (orig_node),
3038 src_idx);
3039 if (agg_pass_through_permissible_p (orig_plats, jfunc))
3041 if (!inter.exists ())
3042 inter = agg_replacements_to_vector (cs->caller, src_idx, 0);
3043 else
3044 intersect_with_agg_replacements (cs->caller, src_idx,
3045 &inter, 0);
3048 else
3050 struct ipcp_param_lattices *src_plats;
3051 src_plats = ipa_get_parm_lattices (caller_info, src_idx);
3052 if (agg_pass_through_permissible_p (src_plats, jfunc))
3054 /* Currently we do not produce clobber aggregate jump
3055 functions, adjust when we do. */
3056 gcc_checking_assert (!jfunc->agg.items);
3057 if (!inter.exists ())
3058 inter = copy_plats_to_inter (src_plats, 0);
3059 else
3060 intersect_with_plats (src_plats, &inter, 0);
3064 else if (jfunc->type == IPA_JF_ANCESTOR
3065 && ipa_get_jf_ancestor_agg_preserved (jfunc))
3067 struct ipa_node_params *caller_info = IPA_NODE_REF (cs->caller);
3068 int src_idx = ipa_get_jf_ancestor_formal_id (jfunc);
3069 struct ipcp_param_lattices *src_plats;
3070 HOST_WIDE_INT delta = ipa_get_jf_ancestor_offset (jfunc);
3072 if (caller_info->ipcp_orig_node)
3074 if (!inter.exists ())
3075 inter = agg_replacements_to_vector (cs->caller, src_idx, delta);
3076 else
3077 intersect_with_agg_replacements (cs->caller, src_idx, &inter,
3078 delta);
3080 else
3082 src_plats = ipa_get_parm_lattices (caller_info, src_idx);;
3083 /* Currently we do not produce clobber aggregate jump
3084 functions, adjust when we do. */
3085 gcc_checking_assert (!src_plats->aggs || !jfunc->agg.items);
3086 if (!inter.exists ())
3087 inter = copy_plats_to_inter (src_plats, delta);
3088 else
3089 intersect_with_plats (src_plats, &inter, delta);
3092 else if (jfunc->agg.items)
3094 struct ipa_agg_jf_item *item;
3095 int k;
3097 if (!inter.exists ())
3098 for (unsigned i = 0; i < jfunc->agg.items->length (); i++)
3099 inter.safe_push ((*jfunc->agg.items)[i]);
3100 else
3101 FOR_EACH_VEC_ELT (inter, k, item)
3103 int l = 0;
3104 bool found = false;;
3106 if (!item->value)
3107 continue;
3109 while ((unsigned) l < jfunc->agg.items->length ())
3111 struct ipa_agg_jf_item *ti;
3112 ti = &(*jfunc->agg.items)[l];
3113 if (ti->offset > item->offset)
3114 break;
3115 if (ti->offset == item->offset)
3117 gcc_checking_assert (ti->value);
3118 if (values_equal_for_ipcp_p (item->value,
3119 ti->value))
3120 found = true;
3121 break;
3123 l++;
3125 if (!found)
3126 item->value = NULL;
3129 else
3131 inter.release ();
3132 return vec<ipa_agg_jf_item>();
3134 return inter;
3137 /* Look at edges in CALLERS and collect all known aggregate values that arrive
3138 from all of them. */
3140 static struct ipa_agg_replacement_value *
3141 find_aggregate_values_for_callers_subset (struct cgraph_node *node,
3142 vec<cgraph_edge_p> callers)
3144 struct ipa_node_params *dest_info = IPA_NODE_REF (node);
3145 struct ipa_agg_replacement_value *res = NULL;
3146 struct cgraph_edge *cs;
3147 int i, j, count = ipa_get_param_count (dest_info);
3149 FOR_EACH_VEC_ELT (callers, j, cs)
3151 int c = ipa_get_cs_argument_count (IPA_EDGE_REF (cs));
3152 if (c < count)
3153 count = c;
3156 for (i = 0; i < count ; i++)
3158 struct cgraph_edge *cs;
3159 vec<ipa_agg_jf_item> inter = vNULL;
3160 struct ipa_agg_jf_item *item;
3161 struct ipcp_param_lattices *plats = ipa_get_parm_lattices (dest_info, i);
3162 int j;
3164 /* Among other things, the following check should deal with all by_ref
3165 mismatches. */
3166 if (plats->aggs_bottom)
3167 continue;
3169 FOR_EACH_VEC_ELT (callers, j, cs)
3171 inter = intersect_aggregates_with_edge (cs, i, inter);
3173 if (!inter.exists ())
3174 goto next_param;
3177 FOR_EACH_VEC_ELT (inter, j, item)
3179 struct ipa_agg_replacement_value *v;
3181 if (!item->value)
3182 continue;
3184 v = ggc_alloc_ipa_agg_replacement_value ();
3185 v->index = i;
3186 v->offset = item->offset;
3187 v->value = item->value;
3188 v->by_ref = plats->aggs_by_ref;
3189 v->next = res;
3190 res = v;
3193 next_param:
3194 if (inter.exists ())
3195 inter.release ();
3197 return res;
3200 /* Turn KNOWN_AGGS into a list of aggreate replacement values. */
3202 static struct ipa_agg_replacement_value *
3203 known_aggs_to_agg_replacement_list (vec<ipa_agg_jump_function> known_aggs)
3205 struct ipa_agg_replacement_value *res = NULL;
3206 struct ipa_agg_jump_function *aggjf;
3207 struct ipa_agg_jf_item *item;
3208 int i, j;
3210 FOR_EACH_VEC_ELT (known_aggs, i, aggjf)
3211 FOR_EACH_VEC_SAFE_ELT (aggjf->items, j, item)
3213 struct ipa_agg_replacement_value *v;
3214 v = ggc_alloc_ipa_agg_replacement_value ();
3215 v->index = i;
3216 v->offset = item->offset;
3217 v->value = item->value;
3218 v->by_ref = aggjf->by_ref;
3219 v->next = res;
3220 res = v;
3222 return res;
3225 /* Determine whether CS also brings all scalar values that the NODE is
3226 specialized for. */
3228 static bool
3229 cgraph_edge_brings_all_scalars_for_node (struct cgraph_edge *cs,
3230 struct cgraph_node *node)
3232 struct ipa_node_params *dest_info = IPA_NODE_REF (node);
3233 int count = ipa_get_param_count (dest_info);
3234 struct ipa_node_params *caller_info;
3235 struct ipa_edge_args *args;
3236 int i;
3238 caller_info = IPA_NODE_REF (cs->caller);
3239 args = IPA_EDGE_REF (cs);
3240 for (i = 0; i < count; i++)
3242 struct ipa_jump_func *jump_func;
3243 tree val, t;
3245 val = dest_info->known_vals[i];
3246 if (!val)
3247 continue;
3249 if (i >= ipa_get_cs_argument_count (args))
3250 return false;
3251 jump_func = ipa_get_ith_jump_func (args, i);
3252 t = ipa_value_from_jfunc (caller_info, jump_func);
3253 if (!t || !values_equal_for_ipcp_p (val, t))
3254 return false;
3256 return true;
3259 /* Determine whether CS also brings all aggregate values that NODE is
3260 specialized for. */
3261 static bool
3262 cgraph_edge_brings_all_agg_vals_for_node (struct cgraph_edge *cs,
3263 struct cgraph_node *node)
3265 struct ipa_node_params *orig_caller_info = IPA_NODE_REF (cs->caller);
3266 struct ipa_node_params *orig_node_info;
3267 struct ipa_agg_replacement_value *aggval;
3268 int i, ec, count;
3270 aggval = ipa_get_agg_replacements_for_node (node);
3271 if (!aggval)
3272 return true;
3274 count = ipa_get_param_count (IPA_NODE_REF (node));
3275 ec = ipa_get_cs_argument_count (IPA_EDGE_REF (cs));
3276 if (ec < count)
3277 for (struct ipa_agg_replacement_value *av = aggval; av; av = av->next)
3278 if (aggval->index >= ec)
3279 return false;
3281 orig_node_info = IPA_NODE_REF (IPA_NODE_REF (node)->ipcp_orig_node);
3282 if (orig_caller_info->ipcp_orig_node)
3283 orig_caller_info = IPA_NODE_REF (orig_caller_info->ipcp_orig_node);
3285 for (i = 0; i < count; i++)
3287 static vec<ipa_agg_jf_item> values = vec<ipa_agg_jf_item>();
3288 struct ipcp_param_lattices *plats;
3289 bool interesting = false;
3290 for (struct ipa_agg_replacement_value *av = aggval; av; av = av->next)
3291 if (aggval->index == i)
3293 interesting = true;
3294 break;
3296 if (!interesting)
3297 continue;
3299 plats = ipa_get_parm_lattices (orig_node_info, aggval->index);
3300 if (plats->aggs_bottom)
3301 return false;
3303 values = intersect_aggregates_with_edge (cs, i, values);
3304 if (!values.exists ())
3305 return false;
3307 for (struct ipa_agg_replacement_value *av = aggval; av; av = av->next)
3308 if (aggval->index == i)
3310 struct ipa_agg_jf_item *item;
3311 int j;
3312 bool found = false;
3313 FOR_EACH_VEC_ELT (values, j, item)
3314 if (item->value
3315 && item->offset == av->offset
3316 && values_equal_for_ipcp_p (item->value, av->value))
3318 found = true;
3319 break;
3321 if (!found)
3323 values.release ();
3324 return false;
3328 return true;
3331 /* Given an original NODE and a VAL for which we have already created a
3332 specialized clone, look whether there are incoming edges that still lead
3333 into the old node but now also bring the requested value and also conform to
3334 all other criteria such that they can be redirected the the special node.
3335 This function can therefore redirect the final edge in a SCC. */
3337 static void
3338 perhaps_add_new_callers (struct cgraph_node *node, struct ipcp_value *val)
3340 struct ipcp_value_source *src;
3341 gcov_type redirected_sum = 0;
3343 for (src = val->sources; src; src = src->next)
3345 struct cgraph_edge *cs = src->cs;
3346 while (cs)
3348 enum availability availability;
3349 struct cgraph_node *dst = cgraph_function_node (cs->callee,
3350 &availability);
3351 if ((dst == node || IPA_NODE_REF (dst)->is_all_contexts_clone)
3352 && availability > AVAIL_OVERWRITABLE
3353 && cgraph_edge_brings_value_p (cs, src))
3355 if (cgraph_edge_brings_all_scalars_for_node (cs, val->spec_node)
3356 && cgraph_edge_brings_all_agg_vals_for_node (cs,
3357 val->spec_node))
3359 if (dump_file)
3360 fprintf (dump_file, " - adding an extra caller %s/%i"
3361 " of %s/%i\n",
3362 xstrdup (cs->caller->name ()),
3363 cs->caller->order,
3364 xstrdup (val->spec_node->name ()),
3365 val->spec_node->order);
3367 cgraph_redirect_edge_callee (cs, val->spec_node);
3368 redirected_sum += cs->count;
3371 cs = get_next_cgraph_edge_clone (cs);
3375 if (redirected_sum)
3376 update_specialized_profile (val->spec_node, node, redirected_sum);
3380 /* Copy KNOWN_BINFOS to KNOWN_VALS. */
3382 static void
3383 move_binfos_to_values (vec<tree> known_vals,
3384 vec<tree> known_binfos)
3386 tree t;
3387 int i;
3389 for (i = 0; known_binfos.iterate (i, &t); i++)
3390 if (t)
3391 known_vals[i] = t;
3394 /* Return true if there is a replacement equivalent to VALUE, INDEX and OFFSET
3395 among those in the AGGVALS list. */
3397 DEBUG_FUNCTION bool
3398 ipcp_val_in_agg_replacements_p (struct ipa_agg_replacement_value *aggvals,
3399 int index, HOST_WIDE_INT offset, tree value)
3401 while (aggvals)
3403 if (aggvals->index == index
3404 && aggvals->offset == offset
3405 && values_equal_for_ipcp_p (aggvals->value, value))
3406 return true;
3407 aggvals = aggvals->next;
3409 return false;
3412 /* Decide wheter to create a special version of NODE for value VAL of parameter
3413 at the given INDEX. If OFFSET is -1, the value is for the parameter itself,
3414 otherwise it is stored at the given OFFSET of the parameter. KNOWN_CSTS,
3415 KNOWN_BINFOS and KNOWN_AGGS describe the other already known values. */
3417 static bool
3418 decide_about_value (struct cgraph_node *node, int index, HOST_WIDE_INT offset,
3419 struct ipcp_value *val, vec<tree> known_csts,
3420 vec<tree> known_binfos)
3422 struct ipa_agg_replacement_value *aggvals;
3423 int freq_sum, caller_count;
3424 gcov_type count_sum;
3425 vec<cgraph_edge_p> callers;
3426 vec<tree> kv;
3428 if (val->spec_node)
3430 perhaps_add_new_callers (node, val);
3431 return false;
3433 else if (val->local_size_cost + overall_size > max_new_size)
3435 if (dump_file && (dump_flags & TDF_DETAILS))
3436 fprintf (dump_file, " Ignoring candidate value because "
3437 "max_new_size would be reached with %li.\n",
3438 val->local_size_cost + overall_size);
3439 return false;
3441 else if (!get_info_about_necessary_edges (val, &freq_sum, &count_sum,
3442 &caller_count))
3443 return false;
3445 if (dump_file && (dump_flags & TDF_DETAILS))
3447 fprintf (dump_file, " - considering value ");
3448 print_ipcp_constant_value (dump_file, val->value);
3449 fprintf (dump_file, " for ");
3450 ipa_dump_param (dump_file, IPA_NODE_REF (node), index);
3451 if (offset != -1)
3452 fprintf (dump_file, ", offset: " HOST_WIDE_INT_PRINT_DEC, offset);
3453 fprintf (dump_file, " (caller_count: %i)\n", caller_count);
3456 if (!good_cloning_opportunity_p (node, val->local_time_benefit,
3457 freq_sum, count_sum,
3458 val->local_size_cost)
3459 && !good_cloning_opportunity_p (node,
3460 val->local_time_benefit
3461 + val->prop_time_benefit,
3462 freq_sum, count_sum,
3463 val->local_size_cost
3464 + val->prop_size_cost))
3465 return false;
3467 if (dump_file)
3468 fprintf (dump_file, " Creating a specialized node of %s/%i.\n",
3469 node->name (), node->order);
3471 callers = gather_edges_for_value (val, caller_count);
3472 kv = known_csts.copy ();
3473 move_binfos_to_values (kv, known_binfos);
3474 if (offset == -1)
3475 kv[index] = val->value;
3476 find_more_scalar_values_for_callers_subset (node, kv, callers);
3477 aggvals = find_aggregate_values_for_callers_subset (node, callers);
3478 gcc_checking_assert (offset == -1
3479 || ipcp_val_in_agg_replacements_p (aggvals, index,
3480 offset, val->value));
3481 val->spec_node = create_specialized_node (node, kv, aggvals, callers);
3482 overall_size += val->local_size_cost;
3484 /* TODO: If for some lattice there is only one other known value
3485 left, make a special node for it too. */
3487 return true;
3490 /* Decide whether and what specialized clones of NODE should be created. */
3492 static bool
3493 decide_whether_version_node (struct cgraph_node *node)
3495 struct ipa_node_params *info = IPA_NODE_REF (node);
3496 int i, count = ipa_get_param_count (info);
3497 vec<tree> known_csts, known_binfos;
3498 vec<ipa_agg_jump_function> known_aggs = vNULL;
3499 bool ret = false;
3501 if (count == 0)
3502 return false;
3504 if (dump_file && (dump_flags & TDF_DETAILS))
3505 fprintf (dump_file, "\nEvaluating opportunities for %s/%i.\n",
3506 node->name (), node->order);
3508 gather_context_independent_values (info, &known_csts, &known_binfos,
3509 info->do_clone_for_all_contexts ? &known_aggs
3510 : NULL, NULL);
3512 for (i = 0; i < count ;i++)
3514 struct ipcp_param_lattices *plats = ipa_get_parm_lattices (info, i);
3515 struct ipcp_lattice *lat = &plats->itself;
3516 struct ipcp_value *val;
3518 if (!lat->bottom
3519 && !known_csts[i]
3520 && !known_binfos[i])
3521 for (val = lat->values; val; val = val->next)
3522 ret |= decide_about_value (node, i, -1, val, known_csts,
3523 known_binfos);
3525 if (!plats->aggs_bottom)
3527 struct ipcp_agg_lattice *aglat;
3528 struct ipcp_value *val;
3529 for (aglat = plats->aggs; aglat; aglat = aglat->next)
3530 if (!aglat->bottom && aglat->values
3531 /* If the following is false, the one value is in
3532 known_aggs. */
3533 && (plats->aggs_contain_variable
3534 || !ipa_lat_is_single_const (aglat)))
3535 for (val = aglat->values; val; val = val->next)
3536 ret |= decide_about_value (node, i, aglat->offset, val,
3537 known_csts, known_binfos);
3539 info = IPA_NODE_REF (node);
3542 if (info->do_clone_for_all_contexts)
3544 struct cgraph_node *clone;
3545 vec<cgraph_edge_p> callers;
3547 if (dump_file)
3548 fprintf (dump_file, " - Creating a specialized node of %s/%i "
3549 "for all known contexts.\n", node->name (),
3550 node->order);
3552 callers = collect_callers_of_node (node);
3553 move_binfos_to_values (known_csts, known_binfos);
3554 clone = create_specialized_node (node, known_csts,
3555 known_aggs_to_agg_replacement_list (known_aggs),
3556 callers);
3557 info = IPA_NODE_REF (node);
3558 info->do_clone_for_all_contexts = false;
3559 IPA_NODE_REF (clone)->is_all_contexts_clone = true;
3560 for (i = 0; i < count ; i++)
3561 vec_free (known_aggs[i].items);
3562 known_aggs.release ();
3563 ret = true;
3565 else
3566 known_csts.release ();
3568 known_binfos.release ();
3569 return ret;
3572 /* Transitively mark all callees of NODE within the same SCC as not dead. */
3574 static void
3575 spread_undeadness (struct cgraph_node *node)
3577 struct cgraph_edge *cs;
3579 for (cs = node->callees; cs; cs = cs->next_callee)
3580 if (ipa_edge_within_scc (cs))
3582 struct cgraph_node *callee;
3583 struct ipa_node_params *info;
3585 callee = cgraph_function_node (cs->callee, NULL);
3586 info = IPA_NODE_REF (callee);
3588 if (info->node_dead)
3590 info->node_dead = 0;
3591 spread_undeadness (callee);
3596 /* Return true if NODE has a caller from outside of its SCC that is not
3597 dead. Worker callback for cgraph_for_node_and_aliases. */
3599 static bool
3600 has_undead_caller_from_outside_scc_p (struct cgraph_node *node,
3601 void *data ATTRIBUTE_UNUSED)
3603 struct cgraph_edge *cs;
3605 for (cs = node->callers; cs; cs = cs->next_caller)
3606 if (cs->caller->thunk.thunk_p
3607 && cgraph_for_node_and_aliases (cs->caller,
3608 has_undead_caller_from_outside_scc_p,
3609 NULL, true))
3610 return true;
3611 else if (!ipa_edge_within_scc (cs)
3612 && !IPA_NODE_REF (cs->caller)->node_dead)
3613 return true;
3614 return false;
3618 /* Identify nodes within the same SCC as NODE which are no longer needed
3619 because of new clones and will be removed as unreachable. */
3621 static void
3622 identify_dead_nodes (struct cgraph_node *node)
3624 struct cgraph_node *v;
3625 for (v = node; v ; v = ((struct ipa_dfs_info *) v->aux)->next_cycle)
3626 if (cgraph_will_be_removed_from_program_if_no_direct_calls (v)
3627 && !cgraph_for_node_and_aliases (v,
3628 has_undead_caller_from_outside_scc_p,
3629 NULL, true))
3630 IPA_NODE_REF (v)->node_dead = 1;
3632 for (v = node; v ; v = ((struct ipa_dfs_info *) v->aux)->next_cycle)
3633 if (!IPA_NODE_REF (v)->node_dead)
3634 spread_undeadness (v);
3636 if (dump_file && (dump_flags & TDF_DETAILS))
3638 for (v = node; v ; v = ((struct ipa_dfs_info *) v->aux)->next_cycle)
3639 if (IPA_NODE_REF (v)->node_dead)
3640 fprintf (dump_file, " Marking node as dead: %s/%i.\n",
3641 v->name (), v->order);
3645 /* The decision stage. Iterate over the topological order of call graph nodes
3646 TOPO and make specialized clones if deemed beneficial. */
3648 static void
3649 ipcp_decision_stage (struct topo_info *topo)
3651 int i;
3653 if (dump_file)
3654 fprintf (dump_file, "\nIPA decision stage:\n\n");
3656 for (i = topo->nnodes - 1; i >= 0; i--)
3658 struct cgraph_node *node = topo->order[i];
3659 bool change = false, iterate = true;
3661 while (iterate)
3663 struct cgraph_node *v;
3664 iterate = false;
3665 for (v = node; v ; v = ((struct ipa_dfs_info *) v->aux)->next_cycle)
3666 if (cgraph_function_with_gimple_body_p (v)
3667 && ipcp_versionable_function_p (v))
3668 iterate |= decide_whether_version_node (v);
3670 change |= iterate;
3672 if (change)
3673 identify_dead_nodes (node);
3677 /* The IPCP driver. */
3679 static unsigned int
3680 ipcp_driver (void)
3682 struct cgraph_2edge_hook_list *edge_duplication_hook_holder;
3683 struct cgraph_edge_hook_list *edge_removal_hook_holder;
3684 struct topo_info topo;
3686 ipa_check_create_node_params ();
3687 ipa_check_create_edge_args ();
3688 grow_edge_clone_vectors ();
3689 edge_duplication_hook_holder =
3690 cgraph_add_edge_duplication_hook (&ipcp_edge_duplication_hook, NULL);
3691 edge_removal_hook_holder =
3692 cgraph_add_edge_removal_hook (&ipcp_edge_removal_hook, NULL);
3694 ipcp_values_pool = create_alloc_pool ("IPA-CP values",
3695 sizeof (struct ipcp_value), 32);
3696 ipcp_sources_pool = create_alloc_pool ("IPA-CP value sources",
3697 sizeof (struct ipcp_value_source), 64);
3698 ipcp_agg_lattice_pool = create_alloc_pool ("IPA_CP aggregate lattices",
3699 sizeof (struct ipcp_agg_lattice),
3700 32);
3701 if (dump_file)
3703 fprintf (dump_file, "\nIPA structures before propagation:\n");
3704 if (dump_flags & TDF_DETAILS)
3705 ipa_print_all_params (dump_file);
3706 ipa_print_all_jump_functions (dump_file);
3709 /* Topological sort. */
3710 build_toporder_info (&topo);
3711 /* Do the interprocedural propagation. */
3712 ipcp_propagate_stage (&topo);
3713 /* Decide what constant propagation and cloning should be performed. */
3714 ipcp_decision_stage (&topo);
3716 /* Free all IPCP structures. */
3717 free_toporder_info (&topo);
3718 next_edge_clone.release ();
3719 cgraph_remove_edge_removal_hook (edge_removal_hook_holder);
3720 cgraph_remove_edge_duplication_hook (edge_duplication_hook_holder);
3721 ipa_free_all_structures_after_ipa_cp ();
3722 if (dump_file)
3723 fprintf (dump_file, "\nIPA constant propagation end\n");
3724 return 0;
3727 /* Initialization and computation of IPCP data structures. This is the initial
3728 intraprocedural analysis of functions, which gathers information to be
3729 propagated later on. */
3731 static void
3732 ipcp_generate_summary (void)
3734 struct cgraph_node *node;
3736 if (dump_file)
3737 fprintf (dump_file, "\nIPA constant propagation start:\n");
3738 ipa_register_cgraph_hooks ();
3740 FOR_EACH_FUNCTION_WITH_GIMPLE_BODY (node)
3742 node->local.versionable
3743 = tree_versionable_function_p (node->decl);
3744 ipa_analyze_node (node);
3748 /* Write ipcp summary for nodes in SET. */
3750 static void
3751 ipcp_write_summary (void)
3753 ipa_prop_write_jump_functions ();
3756 /* Read ipcp summary. */
3758 static void
3759 ipcp_read_summary (void)
3761 ipa_prop_read_jump_functions ();
3764 /* Gate for IPCP optimization. */
3766 static bool
3767 cgraph_gate_cp (void)
3769 /* FIXME: We should remove the optimize check after we ensure we never run
3770 IPA passes when not optimizing. */
3771 return flag_ipa_cp && optimize;
3774 namespace {
3776 const pass_data pass_data_ipa_cp =
3778 IPA_PASS, /* type */
3779 "cp", /* name */
3780 OPTGROUP_NONE, /* optinfo_flags */
3781 true, /* has_gate */
3782 true, /* has_execute */
3783 TV_IPA_CONSTANT_PROP, /* tv_id */
3784 0, /* properties_required */
3785 0, /* properties_provided */
3786 0, /* properties_destroyed */
3787 0, /* todo_flags_start */
3788 ( TODO_dump_symtab | TODO_remove_functions ), /* todo_flags_finish */
3791 class pass_ipa_cp : public ipa_opt_pass_d
3793 public:
3794 pass_ipa_cp (gcc::context *ctxt)
3795 : ipa_opt_pass_d (pass_data_ipa_cp, ctxt,
3796 ipcp_generate_summary, /* generate_summary */
3797 ipcp_write_summary, /* write_summary */
3798 ipcp_read_summary, /* read_summary */
3799 ipa_prop_write_all_agg_replacement, /*
3800 write_optimization_summary */
3801 ipa_prop_read_all_agg_replacement, /*
3802 read_optimization_summary */
3803 NULL, /* stmt_fixup */
3804 0, /* function_transform_todo_flags_start */
3805 ipcp_transform_function, /* function_transform */
3806 NULL) /* variable_transform */
3809 /* opt_pass methods: */
3810 bool gate () { return cgraph_gate_cp (); }
3811 unsigned int execute () { return ipcp_driver (); }
3813 }; // class pass_ipa_cp
3815 } // anon namespace
3817 ipa_opt_pass_d *
3818 make_pass_ipa_cp (gcc::context *ctxt)
3820 return new pass_ipa_cp (ctxt);