2014-03-25 Richard Biener <rguenther@suse.de>
[official-gcc.git] / gcc / ipa-cp.c
blobb71048acc71a983c3ecc5ade17e4ea56f4bec8d9
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);
1642 #ifdef ENABLE_CHECKING
1643 if (target)
1644 gcc_assert (possible_polymorphic_call_target_p
1645 (ie, cgraph_get_node (target)));
1646 #endif
1648 return target;
1652 /* If an indirect edge IE can be turned into a direct one based on KNOWN_VALS
1653 (which can contain both constants and binfos), KNOWN_BINFOS (which can be
1654 NULL) or KNOWN_AGGS (which also can be NULL) return the destination. */
1656 tree
1657 ipa_get_indirect_edge_target (struct cgraph_edge *ie,
1658 vec<tree> known_vals,
1659 vec<tree> known_binfos,
1660 vec<ipa_agg_jump_function_p> known_aggs)
1662 return ipa_get_indirect_edge_target_1 (ie, known_vals, known_binfos,
1663 known_aggs, NULL);
1666 /* Calculate devirtualization time bonus for NODE, assuming we know KNOWN_CSTS
1667 and KNOWN_BINFOS. */
1669 static int
1670 devirtualization_time_bonus (struct cgraph_node *node,
1671 vec<tree> known_csts,
1672 vec<tree> known_binfos,
1673 vec<ipa_agg_jump_function_p> known_aggs)
1675 struct cgraph_edge *ie;
1676 int res = 0;
1678 for (ie = node->indirect_calls; ie; ie = ie->next_callee)
1680 struct cgraph_node *callee;
1681 struct inline_summary *isummary;
1682 tree target;
1684 target = ipa_get_indirect_edge_target (ie, known_csts, known_binfos,
1685 known_aggs);
1686 if (!target)
1687 continue;
1689 /* Only bare minimum benefit for clearly un-inlineable targets. */
1690 res += 1;
1691 callee = cgraph_get_node (target);
1692 if (!callee || !callee->definition)
1693 continue;
1694 isummary = inline_summary (callee);
1695 if (!isummary->inlinable)
1696 continue;
1698 /* FIXME: The values below need re-considering and perhaps also
1699 integrating into the cost metrics, at lest in some very basic way. */
1700 if (isummary->size <= MAX_INLINE_INSNS_AUTO / 4)
1701 res += 31;
1702 else if (isummary->size <= MAX_INLINE_INSNS_AUTO / 2)
1703 res += 15;
1704 else if (isummary->size <= MAX_INLINE_INSNS_AUTO
1705 || DECL_DECLARED_INLINE_P (callee->decl))
1706 res += 7;
1709 return res;
1712 /* Return time bonus incurred because of HINTS. */
1714 static int
1715 hint_time_bonus (inline_hints hints)
1717 int result = 0;
1718 if (hints & (INLINE_HINT_loop_iterations | INLINE_HINT_loop_stride))
1719 result += PARAM_VALUE (PARAM_IPA_CP_LOOP_HINT_BONUS);
1720 if (hints & INLINE_HINT_array_index)
1721 result += PARAM_VALUE (PARAM_IPA_CP_ARRAY_INDEX_HINT_BONUS);
1722 return result;
1725 /* Return true if cloning NODE is a good idea, given the estimated TIME_BENEFIT
1726 and SIZE_COST and with the sum of frequencies of incoming edges to the
1727 potential new clone in FREQUENCIES. */
1729 static bool
1730 good_cloning_opportunity_p (struct cgraph_node *node, int time_benefit,
1731 int freq_sum, gcov_type count_sum, int size_cost)
1733 if (time_benefit == 0
1734 || !flag_ipa_cp_clone
1735 || !optimize_function_for_speed_p (DECL_STRUCT_FUNCTION (node->decl)))
1736 return false;
1738 gcc_assert (size_cost > 0);
1740 if (max_count)
1742 int factor = (count_sum * 1000) / max_count;
1743 HOST_WIDEST_INT evaluation = (((HOST_WIDEST_INT) time_benefit * factor)
1744 / size_cost);
1746 if (dump_file && (dump_flags & TDF_DETAILS))
1747 fprintf (dump_file, " good_cloning_opportunity_p (time: %i, "
1748 "size: %i, count_sum: " HOST_WIDE_INT_PRINT_DEC
1749 ") -> evaluation: " HOST_WIDEST_INT_PRINT_DEC
1750 ", threshold: %i\n",
1751 time_benefit, size_cost, (HOST_WIDE_INT) count_sum,
1752 evaluation, PARAM_VALUE (PARAM_IPA_CP_EVAL_THRESHOLD));
1754 return evaluation >= PARAM_VALUE (PARAM_IPA_CP_EVAL_THRESHOLD);
1756 else
1758 HOST_WIDEST_INT evaluation = (((HOST_WIDEST_INT) time_benefit * freq_sum)
1759 / size_cost);
1761 if (dump_file && (dump_flags & TDF_DETAILS))
1762 fprintf (dump_file, " good_cloning_opportunity_p (time: %i, "
1763 "size: %i, freq_sum: %i) -> evaluation: "
1764 HOST_WIDEST_INT_PRINT_DEC ", threshold: %i\n",
1765 time_benefit, size_cost, freq_sum, evaluation,
1766 PARAM_VALUE (PARAM_IPA_CP_EVAL_THRESHOLD));
1768 return evaluation >= PARAM_VALUE (PARAM_IPA_CP_EVAL_THRESHOLD);
1772 /* Return all context independent values from aggregate lattices in PLATS in a
1773 vector. Return NULL if there are none. */
1775 static vec<ipa_agg_jf_item, va_gc> *
1776 context_independent_aggregate_values (struct ipcp_param_lattices *plats)
1778 vec<ipa_agg_jf_item, va_gc> *res = NULL;
1780 if (plats->aggs_bottom
1781 || plats->aggs_contain_variable
1782 || plats->aggs_count == 0)
1783 return NULL;
1785 for (struct ipcp_agg_lattice *aglat = plats->aggs;
1786 aglat;
1787 aglat = aglat->next)
1788 if (ipa_lat_is_single_const (aglat))
1790 struct ipa_agg_jf_item item;
1791 item.offset = aglat->offset;
1792 item.value = aglat->values->value;
1793 vec_safe_push (res, item);
1795 return res;
1798 /* Allocate KNOWN_CSTS, KNOWN_BINFOS and, if non-NULL, KNOWN_AGGS and populate
1799 them with values of parameters that are known independent of the context.
1800 INFO describes the function. If REMOVABLE_PARAMS_COST is non-NULL, the
1801 movement cost of all removable parameters will be stored in it. */
1803 static bool
1804 gather_context_independent_values (struct ipa_node_params *info,
1805 vec<tree> *known_csts,
1806 vec<tree> *known_binfos,
1807 vec<ipa_agg_jump_function> *known_aggs,
1808 int *removable_params_cost)
1810 int i, count = ipa_get_param_count (info);
1811 bool ret = false;
1813 known_csts->create (0);
1814 known_binfos->create (0);
1815 known_csts->safe_grow_cleared (count);
1816 known_binfos->safe_grow_cleared (count);
1817 if (known_aggs)
1819 known_aggs->create (0);
1820 known_aggs->safe_grow_cleared (count);
1823 if (removable_params_cost)
1824 *removable_params_cost = 0;
1826 for (i = 0; i < count ; i++)
1828 struct ipcp_param_lattices *plats = ipa_get_parm_lattices (info, i);
1829 struct ipcp_lattice *lat = &plats->itself;
1831 if (ipa_lat_is_single_const (lat))
1833 struct ipcp_value *val = lat->values;
1834 if (TREE_CODE (val->value) != TREE_BINFO)
1836 (*known_csts)[i] = val->value;
1837 if (removable_params_cost)
1838 *removable_params_cost
1839 += estimate_move_cost (TREE_TYPE (val->value));
1840 ret = true;
1842 else if (plats->virt_call)
1844 (*known_binfos)[i] = val->value;
1845 ret = true;
1847 else if (removable_params_cost
1848 && !ipa_is_param_used (info, i))
1849 *removable_params_cost += ipa_get_param_move_cost (info, i);
1851 else if (removable_params_cost
1852 && !ipa_is_param_used (info, i))
1853 *removable_params_cost
1854 += ipa_get_param_move_cost (info, i);
1856 if (known_aggs)
1858 vec<ipa_agg_jf_item, va_gc> *agg_items;
1859 struct ipa_agg_jump_function *ajf;
1861 agg_items = context_independent_aggregate_values (plats);
1862 ajf = &(*known_aggs)[i];
1863 ajf->items = agg_items;
1864 ajf->by_ref = plats->aggs_by_ref;
1865 ret |= agg_items != NULL;
1869 return ret;
1872 /* The current interface in ipa-inline-analysis requires a pointer vector.
1873 Create it.
1875 FIXME: That interface should be re-worked, this is slightly silly. Still,
1876 I'd like to discuss how to change it first and this demonstrates the
1877 issue. */
1879 static vec<ipa_agg_jump_function_p>
1880 agg_jmp_p_vec_for_t_vec (vec<ipa_agg_jump_function> known_aggs)
1882 vec<ipa_agg_jump_function_p> ret;
1883 struct ipa_agg_jump_function *ajf;
1884 int i;
1886 ret.create (known_aggs.length ());
1887 FOR_EACH_VEC_ELT (known_aggs, i, ajf)
1888 ret.quick_push (ajf);
1889 return ret;
1892 /* Iterate over known values of parameters of NODE and estimate the local
1893 effects in terms of time and size they have. */
1895 static void
1896 estimate_local_effects (struct cgraph_node *node)
1898 struct ipa_node_params *info = IPA_NODE_REF (node);
1899 int i, count = ipa_get_param_count (info);
1900 vec<tree> known_csts, known_binfos;
1901 vec<ipa_agg_jump_function> known_aggs;
1902 vec<ipa_agg_jump_function_p> known_aggs_ptrs;
1903 bool always_const;
1904 int base_time = inline_summary (node)->time;
1905 int removable_params_cost;
1907 if (!count || !ipcp_versionable_function_p (node))
1908 return;
1910 if (dump_file && (dump_flags & TDF_DETAILS))
1911 fprintf (dump_file, "\nEstimating effects for %s/%i, base_time: %i.\n",
1912 node->name (), node->order, base_time);
1914 always_const = gather_context_independent_values (info, &known_csts,
1915 &known_binfos, &known_aggs,
1916 &removable_params_cost);
1917 known_aggs_ptrs = agg_jmp_p_vec_for_t_vec (known_aggs);
1918 if (always_const)
1920 struct caller_statistics stats;
1921 inline_hints hints;
1922 int time, size;
1924 init_caller_stats (&stats);
1925 cgraph_for_node_and_aliases (node, gather_caller_stats, &stats, false);
1926 estimate_ipcp_clone_size_and_time (node, known_csts, known_binfos,
1927 known_aggs_ptrs, &size, &time, &hints);
1928 time -= devirtualization_time_bonus (node, known_csts, known_binfos,
1929 known_aggs_ptrs);
1930 time -= hint_time_bonus (hints);
1931 time -= removable_params_cost;
1932 size -= stats.n_calls * removable_params_cost;
1934 if (dump_file)
1935 fprintf (dump_file, " - context independent values, size: %i, "
1936 "time_benefit: %i\n", size, base_time - time);
1938 if (size <= 0
1939 || cgraph_will_be_removed_from_program_if_no_direct_calls (node))
1941 info->do_clone_for_all_contexts = true;
1942 base_time = time;
1944 if (dump_file)
1945 fprintf (dump_file, " Decided to specialize for all "
1946 "known contexts, code not going to grow.\n");
1948 else if (good_cloning_opportunity_p (node, base_time - time,
1949 stats.freq_sum, stats.count_sum,
1950 size))
1952 if (size + overall_size <= max_new_size)
1954 info->do_clone_for_all_contexts = true;
1955 base_time = time;
1956 overall_size += size;
1958 if (dump_file)
1959 fprintf (dump_file, " Decided to specialize for all "
1960 "known contexts, growth deemed beneficial.\n");
1962 else if (dump_file && (dump_flags & TDF_DETAILS))
1963 fprintf (dump_file, " Not cloning for all contexts because "
1964 "max_new_size would be reached with %li.\n",
1965 size + overall_size);
1969 for (i = 0; i < count ; i++)
1971 struct ipcp_param_lattices *plats = ipa_get_parm_lattices (info, i);
1972 struct ipcp_lattice *lat = &plats->itself;
1973 struct ipcp_value *val;
1974 int emc;
1976 if (lat->bottom
1977 || !lat->values
1978 || known_csts[i]
1979 || known_binfos[i])
1980 continue;
1982 for (val = lat->values; val; val = val->next)
1984 int time, size, time_benefit;
1985 inline_hints hints;
1987 if (TREE_CODE (val->value) != TREE_BINFO)
1989 known_csts[i] = val->value;
1990 known_binfos[i] = NULL_TREE;
1991 emc = estimate_move_cost (TREE_TYPE (val->value));
1993 else if (plats->virt_call)
1995 known_csts[i] = NULL_TREE;
1996 known_binfos[i] = val->value;
1997 emc = 0;
1999 else
2000 continue;
2002 estimate_ipcp_clone_size_and_time (node, known_csts, known_binfos,
2003 known_aggs_ptrs, &size, &time,
2004 &hints);
2005 time_benefit = base_time - time
2006 + devirtualization_time_bonus (node, known_csts, known_binfos,
2007 known_aggs_ptrs)
2008 + hint_time_bonus (hints)
2009 + removable_params_cost + emc;
2011 gcc_checking_assert (size >=0);
2012 /* The inliner-heuristics based estimates may think that in certain
2013 contexts some functions do not have any size at all but we want
2014 all specializations to have at least a tiny cost, not least not to
2015 divide by zero. */
2016 if (size == 0)
2017 size = 1;
2019 if (dump_file && (dump_flags & TDF_DETAILS))
2021 fprintf (dump_file, " - estimates for value ");
2022 print_ipcp_constant_value (dump_file, val->value);
2023 fprintf (dump_file, " for ");
2024 ipa_dump_param (dump_file, info, i);
2025 fprintf (dump_file, ": time_benefit: %i, size: %i\n",
2026 time_benefit, size);
2029 val->local_time_benefit = time_benefit;
2030 val->local_size_cost = size;
2032 known_binfos[i] = NULL_TREE;
2033 known_csts[i] = NULL_TREE;
2036 for (i = 0; i < count ; i++)
2038 struct ipcp_param_lattices *plats = ipa_get_parm_lattices (info, i);
2039 struct ipa_agg_jump_function *ajf;
2040 struct ipcp_agg_lattice *aglat;
2042 if (plats->aggs_bottom || !plats->aggs)
2043 continue;
2045 ajf = &known_aggs[i];
2046 for (aglat = plats->aggs; aglat; aglat = aglat->next)
2048 struct ipcp_value *val;
2049 if (aglat->bottom || !aglat->values
2050 /* If the following is true, the one value is in known_aggs. */
2051 || (!plats->aggs_contain_variable
2052 && ipa_lat_is_single_const (aglat)))
2053 continue;
2055 for (val = aglat->values; val; val = val->next)
2057 int time, size, time_benefit;
2058 struct ipa_agg_jf_item item;
2059 inline_hints hints;
2061 item.offset = aglat->offset;
2062 item.value = val->value;
2063 vec_safe_push (ajf->items, item);
2065 estimate_ipcp_clone_size_and_time (node, known_csts, known_binfos,
2066 known_aggs_ptrs, &size, &time,
2067 &hints);
2068 time_benefit = base_time - time
2069 + devirtualization_time_bonus (node, known_csts, known_binfos,
2070 known_aggs_ptrs)
2071 + hint_time_bonus (hints);
2072 gcc_checking_assert (size >=0);
2073 if (size == 0)
2074 size = 1;
2076 if (dump_file && (dump_flags & TDF_DETAILS))
2078 fprintf (dump_file, " - estimates for value ");
2079 print_ipcp_constant_value (dump_file, val->value);
2080 fprintf (dump_file, " for ");
2081 ipa_dump_param (dump_file, info, i);
2082 fprintf (dump_file, "[%soffset: " HOST_WIDE_INT_PRINT_DEC
2083 "]: time_benefit: %i, size: %i\n",
2084 plats->aggs_by_ref ? "ref " : "",
2085 aglat->offset, time_benefit, size);
2088 val->local_time_benefit = time_benefit;
2089 val->local_size_cost = size;
2090 ajf->items->pop ();
2095 for (i = 0; i < count ; i++)
2096 vec_free (known_aggs[i].items);
2098 known_csts.release ();
2099 known_binfos.release ();
2100 known_aggs.release ();
2101 known_aggs_ptrs.release ();
2105 /* Add value CUR_VAL and all yet-unsorted values it is dependent on to the
2106 topological sort of values. */
2108 static void
2109 add_val_to_toposort (struct ipcp_value *cur_val)
2111 static int dfs_counter = 0;
2112 static struct ipcp_value *stack;
2113 struct ipcp_value_source *src;
2115 if (cur_val->dfs)
2116 return;
2118 dfs_counter++;
2119 cur_val->dfs = dfs_counter;
2120 cur_val->low_link = dfs_counter;
2122 cur_val->topo_next = stack;
2123 stack = cur_val;
2124 cur_val->on_stack = true;
2126 for (src = cur_val->sources; src; src = src->next)
2127 if (src->val)
2129 if (src->val->dfs == 0)
2131 add_val_to_toposort (src->val);
2132 if (src->val->low_link < cur_val->low_link)
2133 cur_val->low_link = src->val->low_link;
2135 else if (src->val->on_stack
2136 && src->val->dfs < cur_val->low_link)
2137 cur_val->low_link = src->val->dfs;
2140 if (cur_val->dfs == cur_val->low_link)
2142 struct ipcp_value *v, *scc_list = NULL;
2146 v = stack;
2147 stack = v->topo_next;
2148 v->on_stack = false;
2150 v->scc_next = scc_list;
2151 scc_list = v;
2153 while (v != cur_val);
2155 cur_val->topo_next = values_topo;
2156 values_topo = cur_val;
2160 /* Add all values in lattices associated with NODE to the topological sort if
2161 they are not there yet. */
2163 static void
2164 add_all_node_vals_to_toposort (struct cgraph_node *node)
2166 struct ipa_node_params *info = IPA_NODE_REF (node);
2167 int i, count = ipa_get_param_count (info);
2169 for (i = 0; i < count ; i++)
2171 struct ipcp_param_lattices *plats = ipa_get_parm_lattices (info, i);
2172 struct ipcp_lattice *lat = &plats->itself;
2173 struct ipcp_agg_lattice *aglat;
2174 struct ipcp_value *val;
2176 if (!lat->bottom)
2177 for (val = lat->values; val; val = val->next)
2178 add_val_to_toposort (val);
2180 if (!plats->aggs_bottom)
2181 for (aglat = plats->aggs; aglat; aglat = aglat->next)
2182 if (!aglat->bottom)
2183 for (val = aglat->values; val; val = val->next)
2184 add_val_to_toposort (val);
2188 /* One pass of constants propagation along the call graph edges, from callers
2189 to callees (requires topological ordering in TOPO), iterate over strongly
2190 connected components. */
2192 static void
2193 propagate_constants_topo (struct topo_info *topo)
2195 int i;
2197 for (i = topo->nnodes - 1; i >= 0; i--)
2199 unsigned j;
2200 struct cgraph_node *v, *node = topo->order[i];
2201 vec<cgraph_node_ptr> cycle_nodes = ipa_get_nodes_in_cycle (node);
2203 /* First, iteratively propagate within the strongly connected component
2204 until all lattices stabilize. */
2205 FOR_EACH_VEC_ELT (cycle_nodes, j, v)
2206 if (cgraph_function_with_gimple_body_p (v))
2207 push_node_to_stack (topo, v);
2209 v = pop_node_from_stack (topo);
2210 while (v)
2212 struct cgraph_edge *cs;
2214 for (cs = v->callees; cs; cs = cs->next_callee)
2215 if (ipa_edge_within_scc (cs)
2216 && propagate_constants_accross_call (cs))
2217 push_node_to_stack (topo, cs->callee);
2218 v = pop_node_from_stack (topo);
2221 /* Afterwards, propagate along edges leading out of the SCC, calculates
2222 the local effects of the discovered constants and all valid values to
2223 their topological sort. */
2224 FOR_EACH_VEC_ELT (cycle_nodes, j, v)
2225 if (cgraph_function_with_gimple_body_p (v))
2227 struct cgraph_edge *cs;
2229 estimate_local_effects (v);
2230 add_all_node_vals_to_toposort (v);
2231 for (cs = v->callees; cs; cs = cs->next_callee)
2232 if (!ipa_edge_within_scc (cs))
2233 propagate_constants_accross_call (cs);
2235 cycle_nodes.release ();
2240 /* Return the sum of A and B if none of them is bigger than INT_MAX/2, return
2241 the bigger one if otherwise. */
2243 static int
2244 safe_add (int a, int b)
2246 if (a > INT_MAX/2 || b > INT_MAX/2)
2247 return a > b ? a : b;
2248 else
2249 return a + b;
2253 /* Propagate the estimated effects of individual values along the topological
2254 from the dependent values to those they depend on. */
2256 static void
2257 propagate_effects (void)
2259 struct ipcp_value *base;
2261 for (base = values_topo; base; base = base->topo_next)
2263 struct ipcp_value_source *src;
2264 struct ipcp_value *val;
2265 int time = 0, size = 0;
2267 for (val = base; val; val = val->scc_next)
2269 time = safe_add (time,
2270 val->local_time_benefit + val->prop_time_benefit);
2271 size = safe_add (size, val->local_size_cost + val->prop_size_cost);
2274 for (val = base; val; val = val->scc_next)
2275 for (src = val->sources; src; src = src->next)
2276 if (src->val
2277 && cgraph_maybe_hot_edge_p (src->cs))
2279 src->val->prop_time_benefit = safe_add (time,
2280 src->val->prop_time_benefit);
2281 src->val->prop_size_cost = safe_add (size,
2282 src->val->prop_size_cost);
2288 /* Propagate constants, binfos and their effects from the summaries
2289 interprocedurally. */
2291 static void
2292 ipcp_propagate_stage (struct topo_info *topo)
2294 struct cgraph_node *node;
2296 if (dump_file)
2297 fprintf (dump_file, "\n Propagating constants:\n\n");
2299 if (in_lto_p)
2300 ipa_update_after_lto_read ();
2303 FOR_EACH_DEFINED_FUNCTION (node)
2305 struct ipa_node_params *info = IPA_NODE_REF (node);
2307 determine_versionability (node);
2308 if (cgraph_function_with_gimple_body_p (node))
2310 info->lattices = XCNEWVEC (struct ipcp_param_lattices,
2311 ipa_get_param_count (info));
2312 initialize_node_lattices (node);
2314 if (node->definition && !node->alias)
2315 overall_size += inline_summary (node)->self_size;
2316 if (node->count > max_count)
2317 max_count = node->count;
2320 max_new_size = overall_size;
2321 if (max_new_size < PARAM_VALUE (PARAM_LARGE_UNIT_INSNS))
2322 max_new_size = PARAM_VALUE (PARAM_LARGE_UNIT_INSNS);
2323 max_new_size += max_new_size * PARAM_VALUE (PARAM_IPCP_UNIT_GROWTH) / 100 + 1;
2325 if (dump_file)
2326 fprintf (dump_file, "\noverall_size: %li, max_new_size: %li\n",
2327 overall_size, max_new_size);
2329 propagate_constants_topo (topo);
2330 #ifdef ENABLE_CHECKING
2331 ipcp_verify_propagated_values ();
2332 #endif
2333 propagate_effects ();
2335 if (dump_file)
2337 fprintf (dump_file, "\nIPA lattices after all propagation:\n");
2338 print_all_lattices (dump_file, (dump_flags & TDF_DETAILS), true);
2342 /* Discover newly direct outgoing edges from NODE which is a new clone with
2343 known KNOWN_VALS and make them direct. */
2345 static void
2346 ipcp_discover_new_direct_edges (struct cgraph_node *node,
2347 vec<tree> known_vals,
2348 struct ipa_agg_replacement_value *aggvals)
2350 struct cgraph_edge *ie, *next_ie;
2351 bool found = false;
2353 for (ie = node->indirect_calls; ie; ie = next_ie)
2355 tree target;
2357 next_ie = ie->next_callee;
2358 target = ipa_get_indirect_edge_target_1 (ie, known_vals, vNULL, vNULL,
2359 aggvals);
2360 if (target)
2362 bool agg_contents = ie->indirect_info->agg_contents;
2363 bool polymorphic = ie->indirect_info->polymorphic;
2364 int param_index = ie->indirect_info->param_index;
2365 struct cgraph_edge *cs = ipa_make_edge_direct_to_target (ie, target);
2366 found = true;
2368 if (cs && !agg_contents && !polymorphic)
2370 struct ipa_node_params *info = IPA_NODE_REF (node);
2371 int c = ipa_get_controlled_uses (info, param_index);
2372 if (c != IPA_UNDESCRIBED_USE)
2374 struct ipa_ref *to_del;
2376 c--;
2377 ipa_set_controlled_uses (info, param_index, c);
2378 if (dump_file && (dump_flags & TDF_DETAILS))
2379 fprintf (dump_file, " controlled uses count of param "
2380 "%i bumped down to %i\n", param_index, c);
2381 if (c == 0
2382 && (to_del = ipa_find_reference (node,
2383 cs->callee,
2384 NULL, 0)))
2386 if (dump_file && (dump_flags & TDF_DETAILS))
2387 fprintf (dump_file, " and even removing its "
2388 "cloning-created reference\n");
2389 ipa_remove_reference (to_del);
2395 /* Turning calls to direct calls will improve overall summary. */
2396 if (found)
2397 inline_update_overall_summary (node);
2400 /* Vector of pointers which for linked lists of clones of an original crgaph
2401 edge. */
2403 static vec<cgraph_edge_p> next_edge_clone;
2404 static vec<cgraph_edge_p> prev_edge_clone;
2406 static inline void
2407 grow_edge_clone_vectors (void)
2409 if (next_edge_clone.length ()
2410 <= (unsigned) cgraph_edge_max_uid)
2411 next_edge_clone.safe_grow_cleared (cgraph_edge_max_uid + 1);
2412 if (prev_edge_clone.length ()
2413 <= (unsigned) cgraph_edge_max_uid)
2414 prev_edge_clone.safe_grow_cleared (cgraph_edge_max_uid + 1);
2417 /* Edge duplication hook to grow the appropriate linked list in
2418 next_edge_clone. */
2420 static void
2421 ipcp_edge_duplication_hook (struct cgraph_edge *src, struct cgraph_edge *dst,
2422 void *)
2424 grow_edge_clone_vectors ();
2426 struct cgraph_edge *old_next = next_edge_clone[src->uid];
2427 if (old_next)
2428 prev_edge_clone[old_next->uid] = dst;
2429 prev_edge_clone[dst->uid] = src;
2431 next_edge_clone[dst->uid] = old_next;
2432 next_edge_clone[src->uid] = dst;
2435 /* Hook that is called by cgraph.c when an edge is removed. */
2437 static void
2438 ipcp_edge_removal_hook (struct cgraph_edge *cs, void *)
2440 grow_edge_clone_vectors ();
2442 struct cgraph_edge *prev = prev_edge_clone[cs->uid];
2443 struct cgraph_edge *next = next_edge_clone[cs->uid];
2444 if (prev)
2445 next_edge_clone[prev->uid] = next;
2446 if (next)
2447 prev_edge_clone[next->uid] = prev;
2450 /* See if NODE is a clone with a known aggregate value at a given OFFSET of a
2451 parameter with the given INDEX. */
2453 static tree
2454 get_clone_agg_value (struct cgraph_node *node, HOST_WIDEST_INT offset,
2455 int index)
2457 struct ipa_agg_replacement_value *aggval;
2459 aggval = ipa_get_agg_replacements_for_node (node);
2460 while (aggval)
2462 if (aggval->offset == offset
2463 && aggval->index == index)
2464 return aggval->value;
2465 aggval = aggval->next;
2467 return NULL_TREE;
2470 /* Return true if edge CS does bring about the value described by SRC. */
2472 static bool
2473 cgraph_edge_brings_value_p (struct cgraph_edge *cs,
2474 struct ipcp_value_source *src)
2476 struct ipa_node_params *caller_info = IPA_NODE_REF (cs->caller);
2477 struct ipa_node_params *dst_info = IPA_NODE_REF (cs->callee);
2479 if ((dst_info->ipcp_orig_node && !dst_info->is_all_contexts_clone)
2480 || caller_info->node_dead)
2481 return false;
2482 if (!src->val)
2483 return true;
2485 if (caller_info->ipcp_orig_node)
2487 tree t;
2488 if (src->offset == -1)
2489 t = caller_info->known_vals[src->index];
2490 else
2491 t = get_clone_agg_value (cs->caller, src->offset, src->index);
2492 return (t != NULL_TREE
2493 && values_equal_for_ipcp_p (src->val->value, t));
2495 else
2497 struct ipcp_agg_lattice *aglat;
2498 struct ipcp_param_lattices *plats = ipa_get_parm_lattices (caller_info,
2499 src->index);
2500 if (src->offset == -1)
2501 return (ipa_lat_is_single_const (&plats->itself)
2502 && values_equal_for_ipcp_p (src->val->value,
2503 plats->itself.values->value));
2504 else
2506 if (plats->aggs_bottom || plats->aggs_contain_variable)
2507 return false;
2508 for (aglat = plats->aggs; aglat; aglat = aglat->next)
2509 if (aglat->offset == src->offset)
2510 return (ipa_lat_is_single_const (aglat)
2511 && values_equal_for_ipcp_p (src->val->value,
2512 aglat->values->value));
2514 return false;
2518 /* Get the next clone in the linked list of clones of an edge. */
2520 static inline struct cgraph_edge *
2521 get_next_cgraph_edge_clone (struct cgraph_edge *cs)
2523 return next_edge_clone[cs->uid];
2526 /* Given VAL, iterate over all its sources and if they still hold, add their
2527 edge frequency and their number into *FREQUENCY and *CALLER_COUNT
2528 respectively. */
2530 static bool
2531 get_info_about_necessary_edges (struct ipcp_value *val, int *freq_sum,
2532 gcov_type *count_sum, int *caller_count)
2534 struct ipcp_value_source *src;
2535 int freq = 0, count = 0;
2536 gcov_type cnt = 0;
2537 bool hot = false;
2539 for (src = val->sources; src; src = src->next)
2541 struct cgraph_edge *cs = src->cs;
2542 while (cs)
2544 if (cgraph_edge_brings_value_p (cs, src))
2546 count++;
2547 freq += cs->frequency;
2548 cnt += cs->count;
2549 hot |= cgraph_maybe_hot_edge_p (cs);
2551 cs = get_next_cgraph_edge_clone (cs);
2555 *freq_sum = freq;
2556 *count_sum = cnt;
2557 *caller_count = count;
2558 return hot;
2561 /* Return a vector of incoming edges that do bring value VAL. It is assumed
2562 their number is known and equal to CALLER_COUNT. */
2564 static vec<cgraph_edge_p>
2565 gather_edges_for_value (struct ipcp_value *val, int caller_count)
2567 struct ipcp_value_source *src;
2568 vec<cgraph_edge_p> ret;
2570 ret.create (caller_count);
2571 for (src = val->sources; src; src = src->next)
2573 struct cgraph_edge *cs = src->cs;
2574 while (cs)
2576 if (cgraph_edge_brings_value_p (cs, src))
2577 ret.quick_push (cs);
2578 cs = get_next_cgraph_edge_clone (cs);
2582 return ret;
2585 /* Construct a replacement map for a know VALUE for a formal parameter PARAM.
2586 Return it or NULL if for some reason it cannot be created. */
2588 static struct ipa_replace_map *
2589 get_replacement_map (struct ipa_node_params *info, tree value, int parm_num)
2591 struct ipa_replace_map *replace_map;
2594 replace_map = ggc_alloc_ipa_replace_map ();
2595 if (dump_file)
2597 fprintf (dump_file, " replacing ");
2598 ipa_dump_param (dump_file, info, parm_num);
2600 fprintf (dump_file, " with const ");
2601 print_generic_expr (dump_file, value, 0);
2602 fprintf (dump_file, "\n");
2604 replace_map->old_tree = NULL;
2605 replace_map->parm_num = parm_num;
2606 replace_map->new_tree = value;
2607 replace_map->replace_p = true;
2608 replace_map->ref_p = false;
2610 return replace_map;
2613 /* Dump new profiling counts */
2615 static void
2616 dump_profile_updates (struct cgraph_node *orig_node,
2617 struct cgraph_node *new_node)
2619 struct cgraph_edge *cs;
2621 fprintf (dump_file, " setting count of the specialized node to "
2622 HOST_WIDE_INT_PRINT_DEC "\n", (HOST_WIDE_INT) new_node->count);
2623 for (cs = new_node->callees; cs ; cs = cs->next_callee)
2624 fprintf (dump_file, " edge to %s has count "
2625 HOST_WIDE_INT_PRINT_DEC "\n",
2626 cs->callee->name (), (HOST_WIDE_INT) cs->count);
2628 fprintf (dump_file, " setting count of the original node to "
2629 HOST_WIDE_INT_PRINT_DEC "\n", (HOST_WIDE_INT) orig_node->count);
2630 for (cs = orig_node->callees; cs ; cs = cs->next_callee)
2631 fprintf (dump_file, " edge to %s is left with "
2632 HOST_WIDE_INT_PRINT_DEC "\n",
2633 cs->callee->name (), (HOST_WIDE_INT) cs->count);
2636 /* After a specialized NEW_NODE version of ORIG_NODE has been created, update
2637 their profile information to reflect this. */
2639 static void
2640 update_profiling_info (struct cgraph_node *orig_node,
2641 struct cgraph_node *new_node)
2643 struct cgraph_edge *cs;
2644 struct caller_statistics stats;
2645 gcov_type new_sum, orig_sum;
2646 gcov_type remainder, orig_node_count = orig_node->count;
2648 if (orig_node_count == 0)
2649 return;
2651 init_caller_stats (&stats);
2652 cgraph_for_node_and_aliases (orig_node, gather_caller_stats, &stats, false);
2653 orig_sum = stats.count_sum;
2654 init_caller_stats (&stats);
2655 cgraph_for_node_and_aliases (new_node, gather_caller_stats, &stats, false);
2656 new_sum = stats.count_sum;
2658 if (orig_node_count < orig_sum + new_sum)
2660 if (dump_file)
2661 fprintf (dump_file, " Problem: node %s/%i has too low count "
2662 HOST_WIDE_INT_PRINT_DEC " while the sum of incoming "
2663 "counts is " HOST_WIDE_INT_PRINT_DEC "\n",
2664 orig_node->name (), orig_node->order,
2665 (HOST_WIDE_INT) orig_node_count,
2666 (HOST_WIDE_INT) (orig_sum + new_sum));
2668 orig_node_count = (orig_sum + new_sum) * 12 / 10;
2669 if (dump_file)
2670 fprintf (dump_file, " proceeding by pretending it was "
2671 HOST_WIDE_INT_PRINT_DEC "\n",
2672 (HOST_WIDE_INT) orig_node_count);
2675 new_node->count = new_sum;
2676 remainder = orig_node_count - new_sum;
2677 orig_node->count = remainder;
2679 for (cs = new_node->callees; cs ; cs = cs->next_callee)
2680 if (cs->frequency)
2681 cs->count = apply_probability (cs->count,
2682 GCOV_COMPUTE_SCALE (new_sum,
2683 orig_node_count));
2684 else
2685 cs->count = 0;
2687 for (cs = orig_node->callees; cs ; cs = cs->next_callee)
2688 cs->count = apply_probability (cs->count,
2689 GCOV_COMPUTE_SCALE (remainder,
2690 orig_node_count));
2692 if (dump_file)
2693 dump_profile_updates (orig_node, new_node);
2696 /* Update the respective profile of specialized NEW_NODE and the original
2697 ORIG_NODE after additional edges with cumulative count sum REDIRECTED_SUM
2698 have been redirected to the specialized version. */
2700 static void
2701 update_specialized_profile (struct cgraph_node *new_node,
2702 struct cgraph_node *orig_node,
2703 gcov_type redirected_sum)
2705 struct cgraph_edge *cs;
2706 gcov_type new_node_count, orig_node_count = orig_node->count;
2708 if (dump_file)
2709 fprintf (dump_file, " the sum of counts of redirected edges is "
2710 HOST_WIDE_INT_PRINT_DEC "\n", (HOST_WIDE_INT) redirected_sum);
2711 if (orig_node_count == 0)
2712 return;
2714 gcc_assert (orig_node_count >= redirected_sum);
2716 new_node_count = new_node->count;
2717 new_node->count += redirected_sum;
2718 orig_node->count -= redirected_sum;
2720 for (cs = new_node->callees; cs ; cs = cs->next_callee)
2721 if (cs->frequency)
2722 cs->count += apply_probability (cs->count,
2723 GCOV_COMPUTE_SCALE (redirected_sum,
2724 new_node_count));
2725 else
2726 cs->count = 0;
2728 for (cs = orig_node->callees; cs ; cs = cs->next_callee)
2730 gcov_type dec = apply_probability (cs->count,
2731 GCOV_COMPUTE_SCALE (redirected_sum,
2732 orig_node_count));
2733 if (dec < cs->count)
2734 cs->count -= dec;
2735 else
2736 cs->count = 0;
2739 if (dump_file)
2740 dump_profile_updates (orig_node, new_node);
2743 /* Create a specialized version of NODE with known constants and types of
2744 parameters in KNOWN_VALS and redirect all edges in CALLERS to it. */
2746 static struct cgraph_node *
2747 create_specialized_node (struct cgraph_node *node,
2748 vec<tree> known_vals,
2749 struct ipa_agg_replacement_value *aggvals,
2750 vec<cgraph_edge_p> callers)
2752 struct ipa_node_params *new_info, *info = IPA_NODE_REF (node);
2753 vec<ipa_replace_map_p, va_gc> *replace_trees = NULL;
2754 struct ipa_agg_replacement_value *av;
2755 struct cgraph_node *new_node;
2756 int i, count = ipa_get_param_count (info);
2757 bitmap args_to_skip;
2759 gcc_assert (!info->ipcp_orig_node);
2761 if (node->local.can_change_signature)
2763 args_to_skip = BITMAP_GGC_ALLOC ();
2764 for (i = 0; i < count; i++)
2766 tree t = known_vals[i];
2768 if ((t && TREE_CODE (t) != TREE_BINFO)
2769 || !ipa_is_param_used (info, i))
2770 bitmap_set_bit (args_to_skip, i);
2773 else
2775 args_to_skip = NULL;
2776 if (dump_file && (dump_flags & TDF_DETAILS))
2777 fprintf (dump_file, " cannot change function signature\n");
2780 for (i = 0; i < count ; i++)
2782 tree t = known_vals[i];
2783 if (t && TREE_CODE (t) != TREE_BINFO)
2785 struct ipa_replace_map *replace_map;
2787 replace_map = get_replacement_map (info, t, i);
2788 if (replace_map)
2789 vec_safe_push (replace_trees, replace_map);
2793 new_node = cgraph_create_virtual_clone (node, callers, replace_trees,
2794 args_to_skip, "constprop");
2795 ipa_set_node_agg_value_chain (new_node, aggvals);
2796 for (av = aggvals; av; av = av->next)
2797 ipa_maybe_record_reference (new_node, av->value,
2798 IPA_REF_ADDR, NULL);
2800 if (dump_file && (dump_flags & TDF_DETAILS))
2802 fprintf (dump_file, " the new node is %s/%i.\n",
2803 new_node->name (), new_node->order);
2804 if (aggvals)
2805 ipa_dump_agg_replacement_values (dump_file, aggvals);
2807 gcc_checking_assert (ipa_node_params_vector.exists ()
2808 && (ipa_node_params_vector.length ()
2809 > (unsigned) cgraph_max_uid));
2810 update_profiling_info (node, new_node);
2811 new_info = IPA_NODE_REF (new_node);
2812 new_info->ipcp_orig_node = node;
2813 new_info->known_vals = known_vals;
2815 ipcp_discover_new_direct_edges (new_node, known_vals, aggvals);
2817 callers.release ();
2818 return new_node;
2821 /* Given a NODE, and a subset of its CALLERS, try to populate blanks slots in
2822 KNOWN_VALS with constants and types that are also known for all of the
2823 CALLERS. */
2825 static void
2826 find_more_scalar_values_for_callers_subset (struct cgraph_node *node,
2827 vec<tree> known_vals,
2828 vec<cgraph_edge_p> callers)
2830 struct ipa_node_params *info = IPA_NODE_REF (node);
2831 int i, count = ipa_get_param_count (info);
2833 for (i = 0; i < count ; i++)
2835 struct cgraph_edge *cs;
2836 tree newval = NULL_TREE;
2837 int j;
2839 if (ipa_get_scalar_lat (info, i)->bottom || known_vals[i])
2840 continue;
2842 FOR_EACH_VEC_ELT (callers, j, cs)
2844 struct ipa_jump_func *jump_func;
2845 tree t;
2847 if (i >= ipa_get_cs_argument_count (IPA_EDGE_REF (cs)))
2849 newval = NULL_TREE;
2850 break;
2852 jump_func = ipa_get_ith_jump_func (IPA_EDGE_REF (cs), i);
2853 t = ipa_value_from_jfunc (IPA_NODE_REF (cs->caller), jump_func);
2854 if (!t
2855 || (newval
2856 && !values_equal_for_ipcp_p (t, newval)))
2858 newval = NULL_TREE;
2859 break;
2861 else
2862 newval = t;
2865 if (newval)
2867 if (dump_file && (dump_flags & TDF_DETAILS))
2869 fprintf (dump_file, " adding an extra known scalar value ");
2870 print_ipcp_constant_value (dump_file, newval);
2871 fprintf (dump_file, " for ");
2872 ipa_dump_param (dump_file, info, i);
2873 fprintf (dump_file, "\n");
2876 known_vals[i] = newval;
2881 /* Go through PLATS and create a vector of values consisting of values and
2882 offsets (minus OFFSET) of lattices that contain only a single value. */
2884 static vec<ipa_agg_jf_item>
2885 copy_plats_to_inter (struct ipcp_param_lattices *plats, HOST_WIDE_INT offset)
2887 vec<ipa_agg_jf_item> res = vNULL;
2889 if (!plats->aggs || plats->aggs_contain_variable || plats->aggs_bottom)
2890 return vNULL;
2892 for (struct ipcp_agg_lattice *aglat = plats->aggs; aglat; aglat = aglat->next)
2893 if (ipa_lat_is_single_const (aglat))
2895 struct ipa_agg_jf_item ti;
2896 ti.offset = aglat->offset - offset;
2897 ti.value = aglat->values->value;
2898 res.safe_push (ti);
2900 return res;
2903 /* Intersect all values in INTER with single value lattices in PLATS (while
2904 subtracting OFFSET). */
2906 static void
2907 intersect_with_plats (struct ipcp_param_lattices *plats,
2908 vec<ipa_agg_jf_item> *inter,
2909 HOST_WIDE_INT offset)
2911 struct ipcp_agg_lattice *aglat;
2912 struct ipa_agg_jf_item *item;
2913 int k;
2915 if (!plats->aggs || plats->aggs_contain_variable || plats->aggs_bottom)
2917 inter->release ();
2918 return;
2921 aglat = plats->aggs;
2922 FOR_EACH_VEC_ELT (*inter, k, item)
2924 bool found = false;
2925 if (!item->value)
2926 continue;
2927 while (aglat)
2929 if (aglat->offset - offset > item->offset)
2930 break;
2931 if (aglat->offset - offset == item->offset)
2933 gcc_checking_assert (item->value);
2934 if (values_equal_for_ipcp_p (item->value, aglat->values->value))
2935 found = true;
2936 break;
2938 aglat = aglat->next;
2940 if (!found)
2941 item->value = NULL_TREE;
2945 /* Copy agggregate replacement values of NODE (which is an IPA-CP clone) to the
2946 vector result while subtracting OFFSET from the individual value offsets. */
2948 static vec<ipa_agg_jf_item>
2949 agg_replacements_to_vector (struct cgraph_node *node, int index,
2950 HOST_WIDE_INT offset)
2952 struct ipa_agg_replacement_value *av;
2953 vec<ipa_agg_jf_item> res = vNULL;
2955 for (av = ipa_get_agg_replacements_for_node (node); av; av = av->next)
2956 if (av->index == index
2957 && (av->offset - offset) >= 0)
2959 struct ipa_agg_jf_item item;
2960 gcc_checking_assert (av->value);
2961 item.offset = av->offset - offset;
2962 item.value = av->value;
2963 res.safe_push (item);
2966 return res;
2969 /* Intersect all values in INTER with those that we have already scheduled to
2970 be replaced in parameter number INDEX of NODE, which is an IPA-CP clone
2971 (while subtracting OFFSET). */
2973 static void
2974 intersect_with_agg_replacements (struct cgraph_node *node, int index,
2975 vec<ipa_agg_jf_item> *inter,
2976 HOST_WIDE_INT offset)
2978 struct ipa_agg_replacement_value *srcvals;
2979 struct ipa_agg_jf_item *item;
2980 int i;
2982 srcvals = ipa_get_agg_replacements_for_node (node);
2983 if (!srcvals)
2985 inter->release ();
2986 return;
2989 FOR_EACH_VEC_ELT (*inter, i, item)
2991 struct ipa_agg_replacement_value *av;
2992 bool found = false;
2993 if (!item->value)
2994 continue;
2995 for (av = srcvals; av; av = av->next)
2997 gcc_checking_assert (av->value);
2998 if (av->index == index
2999 && av->offset - offset == item->offset)
3001 if (values_equal_for_ipcp_p (item->value, av->value))
3002 found = true;
3003 break;
3006 if (!found)
3007 item->value = NULL_TREE;
3011 /* Intersect values in INTER with aggregate values that come along edge CS to
3012 parameter number INDEX and return it. If INTER does not actually exist yet,
3013 copy all incoming values to it. If we determine we ended up with no values
3014 whatsoever, return a released vector. */
3016 static vec<ipa_agg_jf_item>
3017 intersect_aggregates_with_edge (struct cgraph_edge *cs, int index,
3018 vec<ipa_agg_jf_item> inter)
3020 struct ipa_jump_func *jfunc;
3021 jfunc = ipa_get_ith_jump_func (IPA_EDGE_REF (cs), index);
3022 if (jfunc->type == IPA_JF_PASS_THROUGH
3023 && ipa_get_jf_pass_through_operation (jfunc) == NOP_EXPR)
3025 struct ipa_node_params *caller_info = IPA_NODE_REF (cs->caller);
3026 int src_idx = ipa_get_jf_pass_through_formal_id (jfunc);
3028 if (caller_info->ipcp_orig_node)
3030 struct cgraph_node *orig_node = caller_info->ipcp_orig_node;
3031 struct ipcp_param_lattices *orig_plats;
3032 orig_plats = ipa_get_parm_lattices (IPA_NODE_REF (orig_node),
3033 src_idx);
3034 if (agg_pass_through_permissible_p (orig_plats, jfunc))
3036 if (!inter.exists ())
3037 inter = agg_replacements_to_vector (cs->caller, src_idx, 0);
3038 else
3039 intersect_with_agg_replacements (cs->caller, src_idx,
3040 &inter, 0);
3043 else
3045 struct ipcp_param_lattices *src_plats;
3046 src_plats = ipa_get_parm_lattices (caller_info, src_idx);
3047 if (agg_pass_through_permissible_p (src_plats, jfunc))
3049 /* Currently we do not produce clobber aggregate jump
3050 functions, adjust when we do. */
3051 gcc_checking_assert (!jfunc->agg.items);
3052 if (!inter.exists ())
3053 inter = copy_plats_to_inter (src_plats, 0);
3054 else
3055 intersect_with_plats (src_plats, &inter, 0);
3059 else if (jfunc->type == IPA_JF_ANCESTOR
3060 && ipa_get_jf_ancestor_agg_preserved (jfunc))
3062 struct ipa_node_params *caller_info = IPA_NODE_REF (cs->caller);
3063 int src_idx = ipa_get_jf_ancestor_formal_id (jfunc);
3064 struct ipcp_param_lattices *src_plats;
3065 HOST_WIDE_INT delta = ipa_get_jf_ancestor_offset (jfunc);
3067 if (caller_info->ipcp_orig_node)
3069 if (!inter.exists ())
3070 inter = agg_replacements_to_vector (cs->caller, src_idx, delta);
3071 else
3072 intersect_with_agg_replacements (cs->caller, src_idx, &inter,
3073 delta);
3075 else
3077 src_plats = ipa_get_parm_lattices (caller_info, src_idx);;
3078 /* Currently we do not produce clobber aggregate jump
3079 functions, adjust when we do. */
3080 gcc_checking_assert (!src_plats->aggs || !jfunc->agg.items);
3081 if (!inter.exists ())
3082 inter = copy_plats_to_inter (src_plats, delta);
3083 else
3084 intersect_with_plats (src_plats, &inter, delta);
3087 else if (jfunc->agg.items)
3089 struct ipa_agg_jf_item *item;
3090 int k;
3092 if (!inter.exists ())
3093 for (unsigned i = 0; i < jfunc->agg.items->length (); i++)
3094 inter.safe_push ((*jfunc->agg.items)[i]);
3095 else
3096 FOR_EACH_VEC_ELT (inter, k, item)
3098 int l = 0;
3099 bool found = false;;
3101 if (!item->value)
3102 continue;
3104 while ((unsigned) l < jfunc->agg.items->length ())
3106 struct ipa_agg_jf_item *ti;
3107 ti = &(*jfunc->agg.items)[l];
3108 if (ti->offset > item->offset)
3109 break;
3110 if (ti->offset == item->offset)
3112 gcc_checking_assert (ti->value);
3113 if (values_equal_for_ipcp_p (item->value,
3114 ti->value))
3115 found = true;
3116 break;
3118 l++;
3120 if (!found)
3121 item->value = NULL;
3124 else
3126 inter.release ();
3127 return vec<ipa_agg_jf_item>();
3129 return inter;
3132 /* Look at edges in CALLERS and collect all known aggregate values that arrive
3133 from all of them. */
3135 static struct ipa_agg_replacement_value *
3136 find_aggregate_values_for_callers_subset (struct cgraph_node *node,
3137 vec<cgraph_edge_p> callers)
3139 struct ipa_node_params *dest_info = IPA_NODE_REF (node);
3140 struct ipa_agg_replacement_value *res = NULL;
3141 struct cgraph_edge *cs;
3142 int i, j, count = ipa_get_param_count (dest_info);
3144 FOR_EACH_VEC_ELT (callers, j, cs)
3146 int c = ipa_get_cs_argument_count (IPA_EDGE_REF (cs));
3147 if (c < count)
3148 count = c;
3151 for (i = 0; i < count ; i++)
3153 struct cgraph_edge *cs;
3154 vec<ipa_agg_jf_item> inter = vNULL;
3155 struct ipa_agg_jf_item *item;
3156 struct ipcp_param_lattices *plats = ipa_get_parm_lattices (dest_info, i);
3157 int j;
3159 /* Among other things, the following check should deal with all by_ref
3160 mismatches. */
3161 if (plats->aggs_bottom)
3162 continue;
3164 FOR_EACH_VEC_ELT (callers, j, cs)
3166 inter = intersect_aggregates_with_edge (cs, i, inter);
3168 if (!inter.exists ())
3169 goto next_param;
3172 FOR_EACH_VEC_ELT (inter, j, item)
3174 struct ipa_agg_replacement_value *v;
3176 if (!item->value)
3177 continue;
3179 v = ggc_alloc_ipa_agg_replacement_value ();
3180 v->index = i;
3181 v->offset = item->offset;
3182 v->value = item->value;
3183 v->by_ref = plats->aggs_by_ref;
3184 v->next = res;
3185 res = v;
3188 next_param:
3189 if (inter.exists ())
3190 inter.release ();
3192 return res;
3195 /* Turn KNOWN_AGGS into a list of aggreate replacement values. */
3197 static struct ipa_agg_replacement_value *
3198 known_aggs_to_agg_replacement_list (vec<ipa_agg_jump_function> known_aggs)
3200 struct ipa_agg_replacement_value *res = NULL;
3201 struct ipa_agg_jump_function *aggjf;
3202 struct ipa_agg_jf_item *item;
3203 int i, j;
3205 FOR_EACH_VEC_ELT (known_aggs, i, aggjf)
3206 FOR_EACH_VEC_SAFE_ELT (aggjf->items, j, item)
3208 struct ipa_agg_replacement_value *v;
3209 v = ggc_alloc_ipa_agg_replacement_value ();
3210 v->index = i;
3211 v->offset = item->offset;
3212 v->value = item->value;
3213 v->by_ref = aggjf->by_ref;
3214 v->next = res;
3215 res = v;
3217 return res;
3220 /* Determine whether CS also brings all scalar values that the NODE is
3221 specialized for. */
3223 static bool
3224 cgraph_edge_brings_all_scalars_for_node (struct cgraph_edge *cs,
3225 struct cgraph_node *node)
3227 struct ipa_node_params *dest_info = IPA_NODE_REF (node);
3228 int count = ipa_get_param_count (dest_info);
3229 struct ipa_node_params *caller_info;
3230 struct ipa_edge_args *args;
3231 int i;
3233 caller_info = IPA_NODE_REF (cs->caller);
3234 args = IPA_EDGE_REF (cs);
3235 for (i = 0; i < count; i++)
3237 struct ipa_jump_func *jump_func;
3238 tree val, t;
3240 val = dest_info->known_vals[i];
3241 if (!val)
3242 continue;
3244 if (i >= ipa_get_cs_argument_count (args))
3245 return false;
3246 jump_func = ipa_get_ith_jump_func (args, i);
3247 t = ipa_value_from_jfunc (caller_info, jump_func);
3248 if (!t || !values_equal_for_ipcp_p (val, t))
3249 return false;
3251 return true;
3254 /* Determine whether CS also brings all aggregate values that NODE is
3255 specialized for. */
3256 static bool
3257 cgraph_edge_brings_all_agg_vals_for_node (struct cgraph_edge *cs,
3258 struct cgraph_node *node)
3260 struct ipa_node_params *orig_caller_info = IPA_NODE_REF (cs->caller);
3261 struct ipa_node_params *orig_node_info;
3262 struct ipa_agg_replacement_value *aggval;
3263 int i, ec, count;
3265 aggval = ipa_get_agg_replacements_for_node (node);
3266 if (!aggval)
3267 return true;
3269 count = ipa_get_param_count (IPA_NODE_REF (node));
3270 ec = ipa_get_cs_argument_count (IPA_EDGE_REF (cs));
3271 if (ec < count)
3272 for (struct ipa_agg_replacement_value *av = aggval; av; av = av->next)
3273 if (aggval->index >= ec)
3274 return false;
3276 orig_node_info = IPA_NODE_REF (IPA_NODE_REF (node)->ipcp_orig_node);
3277 if (orig_caller_info->ipcp_orig_node)
3278 orig_caller_info = IPA_NODE_REF (orig_caller_info->ipcp_orig_node);
3280 for (i = 0; i < count; i++)
3282 static vec<ipa_agg_jf_item> values = vec<ipa_agg_jf_item>();
3283 struct ipcp_param_lattices *plats;
3284 bool interesting = false;
3285 for (struct ipa_agg_replacement_value *av = aggval; av; av = av->next)
3286 if (aggval->index == i)
3288 interesting = true;
3289 break;
3291 if (!interesting)
3292 continue;
3294 plats = ipa_get_parm_lattices (orig_node_info, aggval->index);
3295 if (plats->aggs_bottom)
3296 return false;
3298 values = intersect_aggregates_with_edge (cs, i, values);
3299 if (!values.exists ())
3300 return false;
3302 for (struct ipa_agg_replacement_value *av = aggval; av; av = av->next)
3303 if (aggval->index == i)
3305 struct ipa_agg_jf_item *item;
3306 int j;
3307 bool found = false;
3308 FOR_EACH_VEC_ELT (values, j, item)
3309 if (item->value
3310 && item->offset == av->offset
3311 && values_equal_for_ipcp_p (item->value, av->value))
3313 found = true;
3314 break;
3316 if (!found)
3318 values.release ();
3319 return false;
3323 return true;
3326 /* Given an original NODE and a VAL for which we have already created a
3327 specialized clone, look whether there are incoming edges that still lead
3328 into the old node but now also bring the requested value and also conform to
3329 all other criteria such that they can be redirected the the special node.
3330 This function can therefore redirect the final edge in a SCC. */
3332 static void
3333 perhaps_add_new_callers (struct cgraph_node *node, struct ipcp_value *val)
3335 struct ipcp_value_source *src;
3336 gcov_type redirected_sum = 0;
3338 for (src = val->sources; src; src = src->next)
3340 struct cgraph_edge *cs = src->cs;
3341 while (cs)
3343 enum availability availability;
3344 struct cgraph_node *dst = cgraph_function_node (cs->callee,
3345 &availability);
3346 if ((dst == node || IPA_NODE_REF (dst)->is_all_contexts_clone)
3347 && availability > AVAIL_OVERWRITABLE
3348 && cgraph_edge_brings_value_p (cs, src))
3350 if (cgraph_edge_brings_all_scalars_for_node (cs, val->spec_node)
3351 && cgraph_edge_brings_all_agg_vals_for_node (cs,
3352 val->spec_node))
3354 if (dump_file)
3355 fprintf (dump_file, " - adding an extra caller %s/%i"
3356 " of %s/%i\n",
3357 xstrdup (cs->caller->name ()),
3358 cs->caller->order,
3359 xstrdup (val->spec_node->name ()),
3360 val->spec_node->order);
3362 cgraph_redirect_edge_callee (cs, val->spec_node);
3363 redirected_sum += cs->count;
3366 cs = get_next_cgraph_edge_clone (cs);
3370 if (redirected_sum)
3371 update_specialized_profile (val->spec_node, node, redirected_sum);
3375 /* Copy KNOWN_BINFOS to KNOWN_VALS. */
3377 static void
3378 move_binfos_to_values (vec<tree> known_vals,
3379 vec<tree> known_binfos)
3381 tree t;
3382 int i;
3384 for (i = 0; known_binfos.iterate (i, &t); i++)
3385 if (t)
3386 known_vals[i] = t;
3389 /* Return true if there is a replacement equivalent to VALUE, INDEX and OFFSET
3390 among those in the AGGVALS list. */
3392 DEBUG_FUNCTION bool
3393 ipcp_val_in_agg_replacements_p (struct ipa_agg_replacement_value *aggvals,
3394 int index, HOST_WIDE_INT offset, tree value)
3396 while (aggvals)
3398 if (aggvals->index == index
3399 && aggvals->offset == offset
3400 && values_equal_for_ipcp_p (aggvals->value, value))
3401 return true;
3402 aggvals = aggvals->next;
3404 return false;
3407 /* Decide wheter to create a special version of NODE for value VAL of parameter
3408 at the given INDEX. If OFFSET is -1, the value is for the parameter itself,
3409 otherwise it is stored at the given OFFSET of the parameter. KNOWN_CSTS,
3410 KNOWN_BINFOS and KNOWN_AGGS describe the other already known values. */
3412 static bool
3413 decide_about_value (struct cgraph_node *node, int index, HOST_WIDE_INT offset,
3414 struct ipcp_value *val, vec<tree> known_csts,
3415 vec<tree> known_binfos)
3417 struct ipa_agg_replacement_value *aggvals;
3418 int freq_sum, caller_count;
3419 gcov_type count_sum;
3420 vec<cgraph_edge_p> callers;
3421 vec<tree> kv;
3423 if (val->spec_node)
3425 perhaps_add_new_callers (node, val);
3426 return false;
3428 else if (val->local_size_cost + overall_size > max_new_size)
3430 if (dump_file && (dump_flags & TDF_DETAILS))
3431 fprintf (dump_file, " Ignoring candidate value because "
3432 "max_new_size would be reached with %li.\n",
3433 val->local_size_cost + overall_size);
3434 return false;
3436 else if (!get_info_about_necessary_edges (val, &freq_sum, &count_sum,
3437 &caller_count))
3438 return false;
3440 if (dump_file && (dump_flags & TDF_DETAILS))
3442 fprintf (dump_file, " - considering value ");
3443 print_ipcp_constant_value (dump_file, val->value);
3444 fprintf (dump_file, " for ");
3445 ipa_dump_param (dump_file, IPA_NODE_REF (node), index);
3446 if (offset != -1)
3447 fprintf (dump_file, ", offset: " HOST_WIDE_INT_PRINT_DEC, offset);
3448 fprintf (dump_file, " (caller_count: %i)\n", caller_count);
3451 if (!good_cloning_opportunity_p (node, val->local_time_benefit,
3452 freq_sum, count_sum,
3453 val->local_size_cost)
3454 && !good_cloning_opportunity_p (node,
3455 val->local_time_benefit
3456 + val->prop_time_benefit,
3457 freq_sum, count_sum,
3458 val->local_size_cost
3459 + val->prop_size_cost))
3460 return false;
3462 if (dump_file)
3463 fprintf (dump_file, " Creating a specialized node of %s/%i.\n",
3464 node->name (), node->order);
3466 callers = gather_edges_for_value (val, caller_count);
3467 kv = known_csts.copy ();
3468 move_binfos_to_values (kv, known_binfos);
3469 if (offset == -1)
3470 kv[index] = val->value;
3471 find_more_scalar_values_for_callers_subset (node, kv, callers);
3472 aggvals = find_aggregate_values_for_callers_subset (node, callers);
3473 gcc_checking_assert (offset == -1
3474 || ipcp_val_in_agg_replacements_p (aggvals, index,
3475 offset, val->value));
3476 val->spec_node = create_specialized_node (node, kv, aggvals, callers);
3477 overall_size += val->local_size_cost;
3479 /* TODO: If for some lattice there is only one other known value
3480 left, make a special node for it too. */
3482 return true;
3485 /* Decide whether and what specialized clones of NODE should be created. */
3487 static bool
3488 decide_whether_version_node (struct cgraph_node *node)
3490 struct ipa_node_params *info = IPA_NODE_REF (node);
3491 int i, count = ipa_get_param_count (info);
3492 vec<tree> known_csts, known_binfos;
3493 vec<ipa_agg_jump_function> known_aggs = vNULL;
3494 bool ret = false;
3496 if (count == 0)
3497 return false;
3499 if (dump_file && (dump_flags & TDF_DETAILS))
3500 fprintf (dump_file, "\nEvaluating opportunities for %s/%i.\n",
3501 node->name (), node->order);
3503 gather_context_independent_values (info, &known_csts, &known_binfos,
3504 info->do_clone_for_all_contexts ? &known_aggs
3505 : NULL, NULL);
3507 for (i = 0; i < count ;i++)
3509 struct ipcp_param_lattices *plats = ipa_get_parm_lattices (info, i);
3510 struct ipcp_lattice *lat = &plats->itself;
3511 struct ipcp_value *val;
3513 if (!lat->bottom
3514 && !known_csts[i]
3515 && !known_binfos[i])
3516 for (val = lat->values; val; val = val->next)
3517 ret |= decide_about_value (node, i, -1, val, known_csts,
3518 known_binfos);
3520 if (!plats->aggs_bottom)
3522 struct ipcp_agg_lattice *aglat;
3523 struct ipcp_value *val;
3524 for (aglat = plats->aggs; aglat; aglat = aglat->next)
3525 if (!aglat->bottom && aglat->values
3526 /* If the following is false, the one value is in
3527 known_aggs. */
3528 && (plats->aggs_contain_variable
3529 || !ipa_lat_is_single_const (aglat)))
3530 for (val = aglat->values; val; val = val->next)
3531 ret |= decide_about_value (node, i, aglat->offset, val,
3532 known_csts, known_binfos);
3534 info = IPA_NODE_REF (node);
3537 if (info->do_clone_for_all_contexts)
3539 struct cgraph_node *clone;
3540 vec<cgraph_edge_p> callers;
3542 if (dump_file)
3543 fprintf (dump_file, " - Creating a specialized node of %s/%i "
3544 "for all known contexts.\n", node->name (),
3545 node->order);
3547 callers = collect_callers_of_node (node);
3548 move_binfos_to_values (known_csts, known_binfos);
3549 clone = create_specialized_node (node, known_csts,
3550 known_aggs_to_agg_replacement_list (known_aggs),
3551 callers);
3552 info = IPA_NODE_REF (node);
3553 info->do_clone_for_all_contexts = false;
3554 IPA_NODE_REF (clone)->is_all_contexts_clone = true;
3555 for (i = 0; i < count ; i++)
3556 vec_free (known_aggs[i].items);
3557 known_aggs.release ();
3558 ret = true;
3560 else
3561 known_csts.release ();
3563 known_binfos.release ();
3564 return ret;
3567 /* Transitively mark all callees of NODE within the same SCC as not dead. */
3569 static void
3570 spread_undeadness (struct cgraph_node *node)
3572 struct cgraph_edge *cs;
3574 for (cs = node->callees; cs; cs = cs->next_callee)
3575 if (ipa_edge_within_scc (cs))
3577 struct cgraph_node *callee;
3578 struct ipa_node_params *info;
3580 callee = cgraph_function_node (cs->callee, NULL);
3581 info = IPA_NODE_REF (callee);
3583 if (info->node_dead)
3585 info->node_dead = 0;
3586 spread_undeadness (callee);
3591 /* Return true if NODE has a caller from outside of its SCC that is not
3592 dead. Worker callback for cgraph_for_node_and_aliases. */
3594 static bool
3595 has_undead_caller_from_outside_scc_p (struct cgraph_node *node,
3596 void *data ATTRIBUTE_UNUSED)
3598 struct cgraph_edge *cs;
3600 for (cs = node->callers; cs; cs = cs->next_caller)
3601 if (cs->caller->thunk.thunk_p
3602 && cgraph_for_node_and_aliases (cs->caller,
3603 has_undead_caller_from_outside_scc_p,
3604 NULL, true))
3605 return true;
3606 else if (!ipa_edge_within_scc (cs)
3607 && !IPA_NODE_REF (cs->caller)->node_dead)
3608 return true;
3609 return false;
3613 /* Identify nodes within the same SCC as NODE which are no longer needed
3614 because of new clones and will be removed as unreachable. */
3616 static void
3617 identify_dead_nodes (struct cgraph_node *node)
3619 struct cgraph_node *v;
3620 for (v = node; v ; v = ((struct ipa_dfs_info *) v->aux)->next_cycle)
3621 if (cgraph_will_be_removed_from_program_if_no_direct_calls (v)
3622 && !cgraph_for_node_and_aliases (v,
3623 has_undead_caller_from_outside_scc_p,
3624 NULL, true))
3625 IPA_NODE_REF (v)->node_dead = 1;
3627 for (v = node; v ; v = ((struct ipa_dfs_info *) v->aux)->next_cycle)
3628 if (!IPA_NODE_REF (v)->node_dead)
3629 spread_undeadness (v);
3631 if (dump_file && (dump_flags & TDF_DETAILS))
3633 for (v = node; v ; v = ((struct ipa_dfs_info *) v->aux)->next_cycle)
3634 if (IPA_NODE_REF (v)->node_dead)
3635 fprintf (dump_file, " Marking node as dead: %s/%i.\n",
3636 v->name (), v->order);
3640 /* The decision stage. Iterate over the topological order of call graph nodes
3641 TOPO and make specialized clones if deemed beneficial. */
3643 static void
3644 ipcp_decision_stage (struct topo_info *topo)
3646 int i;
3648 if (dump_file)
3649 fprintf (dump_file, "\nIPA decision stage:\n\n");
3651 for (i = topo->nnodes - 1; i >= 0; i--)
3653 struct cgraph_node *node = topo->order[i];
3654 bool change = false, iterate = true;
3656 while (iterate)
3658 struct cgraph_node *v;
3659 iterate = false;
3660 for (v = node; v ; v = ((struct ipa_dfs_info *) v->aux)->next_cycle)
3661 if (cgraph_function_with_gimple_body_p (v)
3662 && ipcp_versionable_function_p (v))
3663 iterate |= decide_whether_version_node (v);
3665 change |= iterate;
3667 if (change)
3668 identify_dead_nodes (node);
3672 /* The IPCP driver. */
3674 static unsigned int
3675 ipcp_driver (void)
3677 struct cgraph_2edge_hook_list *edge_duplication_hook_holder;
3678 struct cgraph_edge_hook_list *edge_removal_hook_holder;
3679 struct topo_info topo;
3681 ipa_check_create_node_params ();
3682 ipa_check_create_edge_args ();
3683 grow_edge_clone_vectors ();
3684 edge_duplication_hook_holder =
3685 cgraph_add_edge_duplication_hook (&ipcp_edge_duplication_hook, NULL);
3686 edge_removal_hook_holder =
3687 cgraph_add_edge_removal_hook (&ipcp_edge_removal_hook, NULL);
3689 ipcp_values_pool = create_alloc_pool ("IPA-CP values",
3690 sizeof (struct ipcp_value), 32);
3691 ipcp_sources_pool = create_alloc_pool ("IPA-CP value sources",
3692 sizeof (struct ipcp_value_source), 64);
3693 ipcp_agg_lattice_pool = create_alloc_pool ("IPA_CP aggregate lattices",
3694 sizeof (struct ipcp_agg_lattice),
3695 32);
3696 if (dump_file)
3698 fprintf (dump_file, "\nIPA structures before propagation:\n");
3699 if (dump_flags & TDF_DETAILS)
3700 ipa_print_all_params (dump_file);
3701 ipa_print_all_jump_functions (dump_file);
3704 /* Topological sort. */
3705 build_toporder_info (&topo);
3706 /* Do the interprocedural propagation. */
3707 ipcp_propagate_stage (&topo);
3708 /* Decide what constant propagation and cloning should be performed. */
3709 ipcp_decision_stage (&topo);
3711 /* Free all IPCP structures. */
3712 free_toporder_info (&topo);
3713 next_edge_clone.release ();
3714 cgraph_remove_edge_removal_hook (edge_removal_hook_holder);
3715 cgraph_remove_edge_duplication_hook (edge_duplication_hook_holder);
3716 ipa_free_all_structures_after_ipa_cp ();
3717 if (dump_file)
3718 fprintf (dump_file, "\nIPA constant propagation end\n");
3719 return 0;
3722 /* Initialization and computation of IPCP data structures. This is the initial
3723 intraprocedural analysis of functions, which gathers information to be
3724 propagated later on. */
3726 static void
3727 ipcp_generate_summary (void)
3729 struct cgraph_node *node;
3731 if (dump_file)
3732 fprintf (dump_file, "\nIPA constant propagation start:\n");
3733 ipa_register_cgraph_hooks ();
3735 FOR_EACH_FUNCTION_WITH_GIMPLE_BODY (node)
3737 node->local.versionable
3738 = tree_versionable_function_p (node->decl);
3739 ipa_analyze_node (node);
3743 /* Write ipcp summary for nodes in SET. */
3745 static void
3746 ipcp_write_summary (void)
3748 ipa_prop_write_jump_functions ();
3751 /* Read ipcp summary. */
3753 static void
3754 ipcp_read_summary (void)
3756 ipa_prop_read_jump_functions ();
3759 /* Gate for IPCP optimization. */
3761 static bool
3762 cgraph_gate_cp (void)
3764 /* FIXME: We should remove the optimize check after we ensure we never run
3765 IPA passes when not optimizing. */
3766 return flag_ipa_cp && optimize;
3769 namespace {
3771 const pass_data pass_data_ipa_cp =
3773 IPA_PASS, /* type */
3774 "cp", /* name */
3775 OPTGROUP_NONE, /* optinfo_flags */
3776 true, /* has_gate */
3777 true, /* has_execute */
3778 TV_IPA_CONSTANT_PROP, /* tv_id */
3779 0, /* properties_required */
3780 0, /* properties_provided */
3781 0, /* properties_destroyed */
3782 0, /* todo_flags_start */
3783 ( TODO_dump_symtab | TODO_remove_functions ), /* todo_flags_finish */
3786 class pass_ipa_cp : public ipa_opt_pass_d
3788 public:
3789 pass_ipa_cp (gcc::context *ctxt)
3790 : ipa_opt_pass_d (pass_data_ipa_cp, ctxt,
3791 ipcp_generate_summary, /* generate_summary */
3792 ipcp_write_summary, /* write_summary */
3793 ipcp_read_summary, /* read_summary */
3794 ipa_prop_write_all_agg_replacement, /*
3795 write_optimization_summary */
3796 ipa_prop_read_all_agg_replacement, /*
3797 read_optimization_summary */
3798 NULL, /* stmt_fixup */
3799 0, /* function_transform_todo_flags_start */
3800 ipcp_transform_function, /* function_transform */
3801 NULL) /* variable_transform */
3804 /* opt_pass methods: */
3805 bool gate () { return cgraph_gate_cp (); }
3806 unsigned int execute () { return ipcp_driver (); }
3808 }; // class pass_ipa_cp
3810 } // anon namespace
3812 ipa_opt_pass_d *
3813 make_pass_ipa_cp (gcc::context *ctxt)
3815 return new pass_ipa_cp (ctxt);