* libgfortran.h (support_fpu_underflow_control,
[official-gcc.git] / gcc / ipa-cp.c
blobb6d66d90eafa37293f2f466cba61925f12ddca54
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 dump_symtab (dump_file);
888 fprintf (dump_file, "\nIPA lattices after constant "
889 "propagation, before gcc_unreachable:\n");
890 print_all_lattices (dump_file, true, false);
893 gcc_unreachable ();
899 /* Return true iff X and Y should be considered equal values by IPA-CP. */
901 static bool
902 values_equal_for_ipcp_p (tree x, tree y)
904 gcc_checking_assert (x != NULL_TREE && y != NULL_TREE);
906 if (x == y)
907 return true;
909 if (TREE_CODE (x) == TREE_BINFO || TREE_CODE (y) == TREE_BINFO)
910 return false;
912 if (TREE_CODE (x) == ADDR_EXPR
913 && TREE_CODE (y) == ADDR_EXPR
914 && TREE_CODE (TREE_OPERAND (x, 0)) == CONST_DECL
915 && TREE_CODE (TREE_OPERAND (y, 0)) == CONST_DECL)
916 return operand_equal_p (DECL_INITIAL (TREE_OPERAND (x, 0)),
917 DECL_INITIAL (TREE_OPERAND (y, 0)), 0);
918 else
919 return operand_equal_p (x, y, 0);
922 /* Add a new value source to VAL, marking that a value comes from edge CS and
923 (if the underlying jump function is a pass-through or an ancestor one) from
924 a caller value SRC_VAL of a caller parameter described by SRC_INDEX. OFFSET
925 is negative if the source was the scalar value of the parameter itself or
926 the offset within an aggregate. */
928 static void
929 add_value_source (struct ipcp_value *val, struct cgraph_edge *cs,
930 struct ipcp_value *src_val, int src_idx, HOST_WIDE_INT offset)
932 struct ipcp_value_source *src;
934 src = (struct ipcp_value_source *) pool_alloc (ipcp_sources_pool);
935 src->offset = offset;
936 src->cs = cs;
937 src->val = src_val;
938 src->index = src_idx;
940 src->next = val->sources;
941 val->sources = src;
944 /* Try to add NEWVAL to LAT, potentially creating a new struct ipcp_value for
945 it. CS, SRC_VAL SRC_INDEX and OFFSET are meant for add_value_source and
946 have the same meaning. */
948 static bool
949 add_value_to_lattice (struct ipcp_lattice *lat, tree newval,
950 struct cgraph_edge *cs, struct ipcp_value *src_val,
951 int src_idx, HOST_WIDE_INT offset)
953 struct ipcp_value *val;
955 if (lat->bottom)
956 return false;
958 for (val = lat->values; val; val = val->next)
959 if (values_equal_for_ipcp_p (val->value, newval))
961 if (ipa_edge_within_scc (cs))
963 struct ipcp_value_source *s;
964 for (s = val->sources; s ; s = s->next)
965 if (s->cs == cs)
966 break;
967 if (s)
968 return false;
971 add_value_source (val, cs, src_val, src_idx, offset);
972 return false;
975 if (lat->values_count == PARAM_VALUE (PARAM_IPA_CP_VALUE_LIST_SIZE))
977 /* We can only free sources, not the values themselves, because sources
978 of other values in this this SCC might point to them. */
979 for (val = lat->values; val; val = val->next)
981 while (val->sources)
983 struct ipcp_value_source *src = val->sources;
984 val->sources = src->next;
985 pool_free (ipcp_sources_pool, src);
989 lat->values = NULL;
990 return set_lattice_to_bottom (lat);
993 lat->values_count++;
994 val = (struct ipcp_value *) pool_alloc (ipcp_values_pool);
995 memset (val, 0, sizeof (*val));
997 add_value_source (val, cs, src_val, src_idx, offset);
998 val->value = newval;
999 val->next = lat->values;
1000 lat->values = val;
1001 return true;
1004 /* Like above but passes a special value of offset to distinguish that the
1005 origin is the scalar value of the parameter rather than a part of an
1006 aggregate. */
1008 static inline bool
1009 add_scalar_value_to_lattice (struct ipcp_lattice *lat, tree newval,
1010 struct cgraph_edge *cs,
1011 struct ipcp_value *src_val, int src_idx)
1013 return add_value_to_lattice (lat, newval, cs, src_val, src_idx, -1);
1016 /* Propagate values through a pass-through jump function JFUNC associated with
1017 edge CS, taking values from SRC_LAT and putting them into DEST_LAT. SRC_IDX
1018 is the index of the source parameter. */
1020 static bool
1021 propagate_vals_accross_pass_through (struct cgraph_edge *cs,
1022 struct ipa_jump_func *jfunc,
1023 struct ipcp_lattice *src_lat,
1024 struct ipcp_lattice *dest_lat,
1025 int src_idx)
1027 struct ipcp_value *src_val;
1028 bool ret = false;
1030 /* Do not create new values when propagating within an SCC because if there
1031 are arithmetic functions with circular dependencies, there is infinite
1032 number of them and we would just make lattices bottom. */
1033 if ((ipa_get_jf_pass_through_operation (jfunc) != NOP_EXPR)
1034 && ipa_edge_within_scc (cs))
1035 ret = set_lattice_contains_variable (dest_lat);
1036 else
1037 for (src_val = src_lat->values; src_val; src_val = src_val->next)
1039 tree cstval = ipa_get_jf_pass_through_result (jfunc, src_val->value);
1041 if (cstval)
1042 ret |= add_scalar_value_to_lattice (dest_lat, cstval, cs, src_val,
1043 src_idx);
1044 else
1045 ret |= set_lattice_contains_variable (dest_lat);
1048 return ret;
1051 /* Propagate values through an ancestor jump function JFUNC associated with
1052 edge CS, taking values from SRC_LAT and putting them into DEST_LAT. SRC_IDX
1053 is the index of the source parameter. */
1055 static bool
1056 propagate_vals_accross_ancestor (struct cgraph_edge *cs,
1057 struct ipa_jump_func *jfunc,
1058 struct ipcp_lattice *src_lat,
1059 struct ipcp_lattice *dest_lat,
1060 int src_idx)
1062 struct ipcp_value *src_val;
1063 bool ret = false;
1065 if (ipa_edge_within_scc (cs))
1066 return set_lattice_contains_variable (dest_lat);
1068 for (src_val = src_lat->values; src_val; src_val = src_val->next)
1070 tree t = ipa_get_jf_ancestor_result (jfunc, src_val->value);
1072 if (t)
1073 ret |= add_scalar_value_to_lattice (dest_lat, t, cs, src_val, src_idx);
1074 else
1075 ret |= set_lattice_contains_variable (dest_lat);
1078 return ret;
1081 /* Propagate scalar values across jump function JFUNC that is associated with
1082 edge CS and put the values into DEST_LAT. */
1084 static bool
1085 propagate_scalar_accross_jump_function (struct cgraph_edge *cs,
1086 struct ipa_jump_func *jfunc,
1087 struct ipcp_lattice *dest_lat)
1089 if (dest_lat->bottom)
1090 return false;
1092 if (jfunc->type == IPA_JF_CONST
1093 || jfunc->type == IPA_JF_KNOWN_TYPE)
1095 tree val;
1097 if (jfunc->type == IPA_JF_KNOWN_TYPE)
1099 val = ipa_binfo_from_known_type_jfunc (jfunc);
1100 if (!val)
1101 return set_lattice_contains_variable (dest_lat);
1103 else
1104 val = ipa_get_jf_constant (jfunc);
1105 return add_scalar_value_to_lattice (dest_lat, val, cs, NULL, 0);
1107 else if (jfunc->type == IPA_JF_PASS_THROUGH
1108 || jfunc->type == IPA_JF_ANCESTOR)
1110 struct ipa_node_params *caller_info = IPA_NODE_REF (cs->caller);
1111 struct ipcp_lattice *src_lat;
1112 int src_idx;
1113 bool ret;
1115 if (jfunc->type == IPA_JF_PASS_THROUGH)
1116 src_idx = ipa_get_jf_pass_through_formal_id (jfunc);
1117 else
1118 src_idx = ipa_get_jf_ancestor_formal_id (jfunc);
1120 src_lat = ipa_get_scalar_lat (caller_info, src_idx);
1121 if (src_lat->bottom)
1122 return set_lattice_contains_variable (dest_lat);
1124 /* If we would need to clone the caller and cannot, do not propagate. */
1125 if (!ipcp_versionable_function_p (cs->caller)
1126 && (src_lat->contains_variable
1127 || (src_lat->values_count > 1)))
1128 return set_lattice_contains_variable (dest_lat);
1130 if (jfunc->type == IPA_JF_PASS_THROUGH)
1131 ret = propagate_vals_accross_pass_through (cs, jfunc, src_lat,
1132 dest_lat, src_idx);
1133 else
1134 ret = propagate_vals_accross_ancestor (cs, jfunc, src_lat, dest_lat,
1135 src_idx);
1137 if (src_lat->contains_variable)
1138 ret |= set_lattice_contains_variable (dest_lat);
1140 return ret;
1143 /* TODO: We currently do not handle member method pointers in IPA-CP (we only
1144 use it for indirect inlining), we should propagate them too. */
1145 return set_lattice_contains_variable (dest_lat);
1148 /* If DEST_PLATS already has aggregate items, check that aggs_by_ref matches
1149 NEW_AGGS_BY_REF and if not, mark all aggs as bottoms and return true (in all
1150 other cases, return false). If there are no aggregate items, set
1151 aggs_by_ref to NEW_AGGS_BY_REF. */
1153 static bool
1154 set_check_aggs_by_ref (struct ipcp_param_lattices *dest_plats,
1155 bool new_aggs_by_ref)
1157 if (dest_plats->aggs)
1159 if (dest_plats->aggs_by_ref != new_aggs_by_ref)
1161 set_agg_lats_to_bottom (dest_plats);
1162 return true;
1165 else
1166 dest_plats->aggs_by_ref = new_aggs_by_ref;
1167 return false;
1170 /* Walk aggregate lattices in DEST_PLATS from ***AGLAT on, until ***aglat is an
1171 already existing lattice for the given OFFSET and SIZE, marking all skipped
1172 lattices as containing variable and checking for overlaps. If there is no
1173 already existing lattice for the OFFSET and VAL_SIZE, create one, initialize
1174 it with offset, size and contains_variable to PRE_EXISTING, and return true,
1175 unless there are too many already. If there are two many, return false. If
1176 there are overlaps turn whole DEST_PLATS to bottom and return false. If any
1177 skipped lattices were newly marked as containing variable, set *CHANGE to
1178 true. */
1180 static bool
1181 merge_agg_lats_step (struct ipcp_param_lattices *dest_plats,
1182 HOST_WIDE_INT offset, HOST_WIDE_INT val_size,
1183 struct ipcp_agg_lattice ***aglat,
1184 bool pre_existing, bool *change)
1186 gcc_checking_assert (offset >= 0);
1188 while (**aglat && (**aglat)->offset < offset)
1190 if ((**aglat)->offset + (**aglat)->size > offset)
1192 set_agg_lats_to_bottom (dest_plats);
1193 return false;
1195 *change |= set_lattice_contains_variable (**aglat);
1196 *aglat = &(**aglat)->next;
1199 if (**aglat && (**aglat)->offset == offset)
1201 if ((**aglat)->size != val_size
1202 || ((**aglat)->next
1203 && (**aglat)->next->offset < offset + val_size))
1205 set_agg_lats_to_bottom (dest_plats);
1206 return false;
1208 gcc_checking_assert (!(**aglat)->next
1209 || (**aglat)->next->offset >= offset + val_size);
1210 return true;
1212 else
1214 struct ipcp_agg_lattice *new_al;
1216 if (**aglat && (**aglat)->offset < offset + val_size)
1218 set_agg_lats_to_bottom (dest_plats);
1219 return false;
1221 if (dest_plats->aggs_count == PARAM_VALUE (PARAM_IPA_MAX_AGG_ITEMS))
1222 return false;
1223 dest_plats->aggs_count++;
1224 new_al = (struct ipcp_agg_lattice *) pool_alloc (ipcp_agg_lattice_pool);
1225 memset (new_al, 0, sizeof (*new_al));
1227 new_al->offset = offset;
1228 new_al->size = val_size;
1229 new_al->contains_variable = pre_existing;
1231 new_al->next = **aglat;
1232 **aglat = new_al;
1233 return true;
1237 /* Set all AGLAT and all other aggregate lattices reachable by next pointers as
1238 containing an unknown value. */
1240 static bool
1241 set_chain_of_aglats_contains_variable (struct ipcp_agg_lattice *aglat)
1243 bool ret = false;
1244 while (aglat)
1246 ret |= set_lattice_contains_variable (aglat);
1247 aglat = aglat->next;
1249 return ret;
1252 /* Merge existing aggregate lattices in SRC_PLATS to DEST_PLATS, subtracting
1253 DELTA_OFFSET. CS is the call graph edge and SRC_IDX the index of the source
1254 parameter used for lattice value sources. Return true if DEST_PLATS changed
1255 in any way. */
1257 static bool
1258 merge_aggregate_lattices (struct cgraph_edge *cs,
1259 struct ipcp_param_lattices *dest_plats,
1260 struct ipcp_param_lattices *src_plats,
1261 int src_idx, HOST_WIDE_INT offset_delta)
1263 bool pre_existing = dest_plats->aggs != NULL;
1264 struct ipcp_agg_lattice **dst_aglat;
1265 bool ret = false;
1267 if (set_check_aggs_by_ref (dest_plats, src_plats->aggs_by_ref))
1268 return true;
1269 if (src_plats->aggs_bottom)
1270 return set_agg_lats_contain_variable (dest_plats);
1271 if (src_plats->aggs_contain_variable)
1272 ret |= set_agg_lats_contain_variable (dest_plats);
1273 dst_aglat = &dest_plats->aggs;
1275 for (struct ipcp_agg_lattice *src_aglat = src_plats->aggs;
1276 src_aglat;
1277 src_aglat = src_aglat->next)
1279 HOST_WIDE_INT new_offset = src_aglat->offset - offset_delta;
1281 if (new_offset < 0)
1282 continue;
1283 if (merge_agg_lats_step (dest_plats, new_offset, src_aglat->size,
1284 &dst_aglat, pre_existing, &ret))
1286 struct ipcp_agg_lattice *new_al = *dst_aglat;
1288 dst_aglat = &(*dst_aglat)->next;
1289 if (src_aglat->bottom)
1291 ret |= set_lattice_contains_variable (new_al);
1292 continue;
1294 if (src_aglat->contains_variable)
1295 ret |= set_lattice_contains_variable (new_al);
1296 for (struct ipcp_value *val = src_aglat->values;
1297 val;
1298 val = val->next)
1299 ret |= add_value_to_lattice (new_al, val->value, cs, val, src_idx,
1300 src_aglat->offset);
1302 else if (dest_plats->aggs_bottom)
1303 return true;
1305 ret |= set_chain_of_aglats_contains_variable (*dst_aglat);
1306 return ret;
1309 /* Determine whether there is anything to propagate FROM SRC_PLATS through a
1310 pass-through JFUNC and if so, whether it has conform and conforms to the
1311 rules about propagating values passed by reference. */
1313 static bool
1314 agg_pass_through_permissible_p (struct ipcp_param_lattices *src_plats,
1315 struct ipa_jump_func *jfunc)
1317 return src_plats->aggs
1318 && (!src_plats->aggs_by_ref
1319 || ipa_get_jf_pass_through_agg_preserved (jfunc));
1322 /* Propagate scalar values across jump function JFUNC that is associated with
1323 edge CS and put the values into DEST_LAT. */
1325 static bool
1326 propagate_aggs_accross_jump_function (struct cgraph_edge *cs,
1327 struct ipa_jump_func *jfunc,
1328 struct ipcp_param_lattices *dest_plats)
1330 bool ret = false;
1332 if (dest_plats->aggs_bottom)
1333 return false;
1335 if (jfunc->type == IPA_JF_PASS_THROUGH
1336 && ipa_get_jf_pass_through_operation (jfunc) == NOP_EXPR)
1338 struct ipa_node_params *caller_info = IPA_NODE_REF (cs->caller);
1339 int src_idx = ipa_get_jf_pass_through_formal_id (jfunc);
1340 struct ipcp_param_lattices *src_plats;
1342 src_plats = ipa_get_parm_lattices (caller_info, src_idx);
1343 if (agg_pass_through_permissible_p (src_plats, jfunc))
1345 /* Currently we do not produce clobber aggregate jump
1346 functions, replace with merging when we do. */
1347 gcc_assert (!jfunc->agg.items);
1348 ret |= merge_aggregate_lattices (cs, dest_plats, src_plats,
1349 src_idx, 0);
1351 else
1352 ret |= set_agg_lats_contain_variable (dest_plats);
1354 else if (jfunc->type == IPA_JF_ANCESTOR
1355 && ipa_get_jf_ancestor_agg_preserved (jfunc))
1357 struct ipa_node_params *caller_info = IPA_NODE_REF (cs->caller);
1358 int src_idx = ipa_get_jf_ancestor_formal_id (jfunc);
1359 struct ipcp_param_lattices *src_plats;
1361 src_plats = ipa_get_parm_lattices (caller_info, src_idx);
1362 if (src_plats->aggs && src_plats->aggs_by_ref)
1364 /* Currently we do not produce clobber aggregate jump
1365 functions, replace with merging when we do. */
1366 gcc_assert (!jfunc->agg.items);
1367 ret |= merge_aggregate_lattices (cs, dest_plats, src_plats, src_idx,
1368 ipa_get_jf_ancestor_offset (jfunc));
1370 else if (!src_plats->aggs_by_ref)
1371 ret |= set_agg_lats_to_bottom (dest_plats);
1372 else
1373 ret |= set_agg_lats_contain_variable (dest_plats);
1375 else if (jfunc->agg.items)
1377 bool pre_existing = dest_plats->aggs != NULL;
1378 struct ipcp_agg_lattice **aglat = &dest_plats->aggs;
1379 struct ipa_agg_jf_item *item;
1380 int i;
1382 if (set_check_aggs_by_ref (dest_plats, jfunc->agg.by_ref))
1383 return true;
1385 FOR_EACH_VEC_ELT (*jfunc->agg.items, i, item)
1387 HOST_WIDE_INT val_size;
1389 if (item->offset < 0)
1390 continue;
1391 gcc_checking_assert (is_gimple_ip_invariant (item->value));
1392 val_size = tree_to_uhwi (TYPE_SIZE (TREE_TYPE (item->value)));
1394 if (merge_agg_lats_step (dest_plats, item->offset, val_size,
1395 &aglat, pre_existing, &ret))
1397 ret |= add_value_to_lattice (*aglat, item->value, cs, NULL, 0, 0);
1398 aglat = &(*aglat)->next;
1400 else if (dest_plats->aggs_bottom)
1401 return true;
1404 ret |= set_chain_of_aglats_contains_variable (*aglat);
1406 else
1407 ret |= set_agg_lats_contain_variable (dest_plats);
1409 return ret;
1412 /* Propagate constants from the caller to the callee of CS. INFO describes the
1413 caller. */
1415 static bool
1416 propagate_constants_accross_call (struct cgraph_edge *cs)
1418 struct ipa_node_params *callee_info;
1419 enum availability availability;
1420 struct cgraph_node *callee, *alias_or_thunk;
1421 struct ipa_edge_args *args;
1422 bool ret = false;
1423 int i, args_count, parms_count;
1425 callee = cgraph_function_node (cs->callee, &availability);
1426 if (!callee->definition)
1427 return false;
1428 gcc_checking_assert (cgraph_function_with_gimple_body_p (callee));
1429 callee_info = IPA_NODE_REF (callee);
1431 args = IPA_EDGE_REF (cs);
1432 args_count = ipa_get_cs_argument_count (args);
1433 parms_count = ipa_get_param_count (callee_info);
1434 if (parms_count == 0)
1435 return false;
1437 /* If this call goes through a thunk we must not propagate to the first (0th)
1438 parameter. However, we might need to uncover a thunk from below a series
1439 of aliases first. */
1440 alias_or_thunk = cs->callee;
1441 while (alias_or_thunk->alias)
1442 alias_or_thunk = cgraph_alias_target (alias_or_thunk);
1443 if (alias_or_thunk->thunk.thunk_p)
1445 ret |= set_all_contains_variable (ipa_get_parm_lattices (callee_info,
1446 0));
1447 i = 1;
1449 else
1450 i = 0;
1452 for (; (i < args_count) && (i < parms_count); i++)
1454 struct ipa_jump_func *jump_func = ipa_get_ith_jump_func (args, i);
1455 struct ipcp_param_lattices *dest_plats;
1457 dest_plats = ipa_get_parm_lattices (callee_info, i);
1458 if (availability == AVAIL_OVERWRITABLE)
1459 ret |= set_all_contains_variable (dest_plats);
1460 else
1462 ret |= propagate_scalar_accross_jump_function (cs, jump_func,
1463 &dest_plats->itself);
1464 ret |= propagate_aggs_accross_jump_function (cs, jump_func,
1465 dest_plats);
1468 for (; i < parms_count; i++)
1469 ret |= set_all_contains_variable (ipa_get_parm_lattices (callee_info, i));
1471 return ret;
1474 /* If an indirect edge IE can be turned into a direct one based on KNOWN_VALS
1475 (which can contain both constants and binfos), KNOWN_BINFOS, KNOWN_AGGS or
1476 AGG_REPS return the destination. The latter three can be NULL. If AGG_REPS
1477 is not NULL, KNOWN_AGGS is ignored. */
1479 static tree
1480 ipa_get_indirect_edge_target_1 (struct cgraph_edge *ie,
1481 vec<tree> known_vals,
1482 vec<tree> known_binfos,
1483 vec<ipa_agg_jump_function_p> known_aggs,
1484 struct ipa_agg_replacement_value *agg_reps)
1486 int param_index = ie->indirect_info->param_index;
1487 HOST_WIDE_INT token, anc_offset;
1488 tree otr_type;
1489 tree t;
1490 tree target = NULL;
1492 if (param_index == -1
1493 || known_vals.length () <= (unsigned int) param_index)
1494 return NULL_TREE;
1496 if (!ie->indirect_info->polymorphic)
1498 tree t;
1500 if (ie->indirect_info->agg_contents)
1502 if (agg_reps)
1504 t = NULL;
1505 while (agg_reps)
1507 if (agg_reps->index == param_index
1508 && agg_reps->offset == ie->indirect_info->offset
1509 && agg_reps->by_ref == ie->indirect_info->by_ref)
1511 t = agg_reps->value;
1512 break;
1514 agg_reps = agg_reps->next;
1517 else if (known_aggs.length () > (unsigned int) param_index)
1519 struct ipa_agg_jump_function *agg;
1520 agg = known_aggs[param_index];
1521 t = ipa_find_agg_cst_for_param (agg, ie->indirect_info->offset,
1522 ie->indirect_info->by_ref);
1524 else
1525 t = NULL;
1527 else
1528 t = known_vals[param_index];
1530 if (t &&
1531 TREE_CODE (t) == ADDR_EXPR
1532 && TREE_CODE (TREE_OPERAND (t, 0)) == FUNCTION_DECL)
1533 return TREE_OPERAND (t, 0);
1534 else
1535 return NULL_TREE;
1538 if (!flag_devirtualize)
1539 return NULL_TREE;
1541 gcc_assert (!ie->indirect_info->agg_contents);
1542 token = ie->indirect_info->otr_token;
1543 anc_offset = ie->indirect_info->offset;
1544 otr_type = ie->indirect_info->otr_type;
1546 t = NULL;
1548 /* Try to work out value of virtual table pointer value in replacemnets. */
1549 if (!t && agg_reps && !ie->indirect_info->by_ref)
1551 while (agg_reps)
1553 if (agg_reps->index == param_index
1554 && agg_reps->offset == ie->indirect_info->offset
1555 && agg_reps->by_ref)
1557 t = agg_reps->value;
1558 break;
1560 agg_reps = agg_reps->next;
1564 /* Try to work out value of virtual table pointer value in known
1565 aggregate values. */
1566 if (!t && known_aggs.length () > (unsigned int) param_index
1567 && !ie->indirect_info->by_ref)
1569 struct ipa_agg_jump_function *agg;
1570 agg = known_aggs[param_index];
1571 t = ipa_find_agg_cst_for_param (agg, ie->indirect_info->offset,
1572 true);
1575 /* If we found the virtual table pointer, lookup the target. */
1576 if (t)
1578 tree vtable;
1579 unsigned HOST_WIDE_INT offset;
1580 if (vtable_pointer_value_to_vtable (t, &vtable, &offset))
1582 target = gimple_get_virt_method_for_vtable (ie->indirect_info->otr_token,
1583 vtable, offset);
1584 if (target)
1586 if ((TREE_CODE (TREE_TYPE (target)) == FUNCTION_TYPE
1587 && DECL_FUNCTION_CODE (target) == BUILT_IN_UNREACHABLE)
1588 || !possible_polymorphic_call_target_p
1589 (ie, cgraph_get_node (target)))
1590 target = ipa_impossible_devirt_target (ie, target);
1591 return target;
1596 /* Did we work out BINFO via type propagation? */
1597 if (!t && known_binfos.length () > (unsigned int) param_index)
1598 t = known_binfos[param_index];
1599 /* Or do we know the constant value of pointer? */
1600 if (!t)
1601 t = known_vals[param_index];
1602 if (!t)
1603 return NULL_TREE;
1605 if (TREE_CODE (t) != TREE_BINFO)
1607 ipa_polymorphic_call_context context;
1608 vec <cgraph_node *>targets;
1609 bool final;
1611 if (!get_polymorphic_call_info_from_invariant
1612 (&context, t, ie->indirect_info->otr_type,
1613 anc_offset))
1614 return NULL_TREE;
1615 targets = possible_polymorphic_call_targets
1616 (ie->indirect_info->otr_type,
1617 ie->indirect_info->otr_token,
1618 context, &final);
1619 if (!final || targets.length () > 1)
1620 return NULL_TREE;
1621 if (targets.length () == 1)
1622 target = targets[0]->decl;
1623 else
1624 target = ipa_impossible_devirt_target (ie, NULL_TREE);
1626 else
1628 tree binfo;
1630 binfo = get_binfo_at_offset (t, anc_offset, otr_type);
1631 if (!binfo)
1632 return NULL_TREE;
1633 target = gimple_get_virt_method_for_binfo (token, binfo);
1636 if (target && !possible_polymorphic_call_target_p (ie,
1637 cgraph_get_node (target)))
1638 target = ipa_impossible_devirt_target (ie, target);
1640 return target;
1644 /* If an indirect edge IE can be turned into a direct one based on KNOWN_VALS
1645 (which can contain both constants and binfos), KNOWN_BINFOS (which can be
1646 NULL) or KNOWN_AGGS (which also can be NULL) return the destination. */
1648 tree
1649 ipa_get_indirect_edge_target (struct cgraph_edge *ie,
1650 vec<tree> known_vals,
1651 vec<tree> known_binfos,
1652 vec<ipa_agg_jump_function_p> known_aggs)
1654 return ipa_get_indirect_edge_target_1 (ie, known_vals, known_binfos,
1655 known_aggs, NULL);
1658 /* Calculate devirtualization time bonus for NODE, assuming we know KNOWN_CSTS
1659 and KNOWN_BINFOS. */
1661 static int
1662 devirtualization_time_bonus (struct cgraph_node *node,
1663 vec<tree> known_csts,
1664 vec<tree> known_binfos,
1665 vec<ipa_agg_jump_function_p> known_aggs)
1667 struct cgraph_edge *ie;
1668 int res = 0;
1670 for (ie = node->indirect_calls; ie; ie = ie->next_callee)
1672 struct cgraph_node *callee;
1673 struct inline_summary *isummary;
1674 enum availability avail;
1675 tree target;
1677 target = ipa_get_indirect_edge_target (ie, known_csts, known_binfos,
1678 known_aggs);
1679 if (!target)
1680 continue;
1682 /* Only bare minimum benefit for clearly un-inlineable targets. */
1683 res += 1;
1684 callee = cgraph_get_node (target);
1685 if (!callee || !callee->definition)
1686 continue;
1687 callee = cgraph_function_node (callee, &avail);
1688 if (avail < AVAIL_AVAILABLE)
1689 continue;
1690 isummary = inline_summary (callee);
1691 if (!isummary->inlinable)
1692 continue;
1694 /* FIXME: The values below need re-considering and perhaps also
1695 integrating into the cost metrics, at lest in some very basic way. */
1696 if (isummary->size <= MAX_INLINE_INSNS_AUTO / 4)
1697 res += 31;
1698 else if (isummary->size <= MAX_INLINE_INSNS_AUTO / 2)
1699 res += 15;
1700 else if (isummary->size <= MAX_INLINE_INSNS_AUTO
1701 || DECL_DECLARED_INLINE_P (callee->decl))
1702 res += 7;
1705 return res;
1708 /* Return time bonus incurred because of HINTS. */
1710 static int
1711 hint_time_bonus (inline_hints hints)
1713 int result = 0;
1714 if (hints & (INLINE_HINT_loop_iterations | INLINE_HINT_loop_stride))
1715 result += PARAM_VALUE (PARAM_IPA_CP_LOOP_HINT_BONUS);
1716 if (hints & INLINE_HINT_array_index)
1717 result += PARAM_VALUE (PARAM_IPA_CP_ARRAY_INDEX_HINT_BONUS);
1718 return result;
1721 /* Return true if cloning NODE is a good idea, given the estimated TIME_BENEFIT
1722 and SIZE_COST and with the sum of frequencies of incoming edges to the
1723 potential new clone in FREQUENCIES. */
1725 static bool
1726 good_cloning_opportunity_p (struct cgraph_node *node, int time_benefit,
1727 int freq_sum, gcov_type count_sum, int size_cost)
1729 if (time_benefit == 0
1730 || !flag_ipa_cp_clone
1731 || !optimize_function_for_speed_p (DECL_STRUCT_FUNCTION (node->decl)))
1732 return false;
1734 gcc_assert (size_cost > 0);
1736 if (max_count)
1738 int factor = (count_sum * 1000) / max_count;
1739 int64_t evaluation = (((int64_t) time_benefit * factor)
1740 / size_cost);
1742 if (dump_file && (dump_flags & TDF_DETAILS))
1743 fprintf (dump_file, " good_cloning_opportunity_p (time: %i, "
1744 "size: %i, count_sum: " HOST_WIDE_INT_PRINT_DEC
1745 ") -> evaluation: " "%"PRId64
1746 ", threshold: %i\n",
1747 time_benefit, size_cost, (HOST_WIDE_INT) count_sum,
1748 evaluation, PARAM_VALUE (PARAM_IPA_CP_EVAL_THRESHOLD));
1750 return evaluation >= PARAM_VALUE (PARAM_IPA_CP_EVAL_THRESHOLD);
1752 else
1754 int64_t evaluation = (((int64_t) time_benefit * freq_sum)
1755 / size_cost);
1757 if (dump_file && (dump_flags & TDF_DETAILS))
1758 fprintf (dump_file, " good_cloning_opportunity_p (time: %i, "
1759 "size: %i, freq_sum: %i) -> evaluation: "
1760 "%"PRId64 ", threshold: %i\n",
1761 time_benefit, size_cost, freq_sum, evaluation,
1762 PARAM_VALUE (PARAM_IPA_CP_EVAL_THRESHOLD));
1764 return evaluation >= PARAM_VALUE (PARAM_IPA_CP_EVAL_THRESHOLD);
1768 /* Return all context independent values from aggregate lattices in PLATS in a
1769 vector. Return NULL if there are none. */
1771 static vec<ipa_agg_jf_item, va_gc> *
1772 context_independent_aggregate_values (struct ipcp_param_lattices *plats)
1774 vec<ipa_agg_jf_item, va_gc> *res = NULL;
1776 if (plats->aggs_bottom
1777 || plats->aggs_contain_variable
1778 || plats->aggs_count == 0)
1779 return NULL;
1781 for (struct ipcp_agg_lattice *aglat = plats->aggs;
1782 aglat;
1783 aglat = aglat->next)
1784 if (ipa_lat_is_single_const (aglat))
1786 struct ipa_agg_jf_item item;
1787 item.offset = aglat->offset;
1788 item.value = aglat->values->value;
1789 vec_safe_push (res, item);
1791 return res;
1794 /* Allocate KNOWN_CSTS, KNOWN_BINFOS and, if non-NULL, KNOWN_AGGS and populate
1795 them with values of parameters that are known independent of the context.
1796 INFO describes the function. If REMOVABLE_PARAMS_COST is non-NULL, the
1797 movement cost of all removable parameters will be stored in it. */
1799 static bool
1800 gather_context_independent_values (struct ipa_node_params *info,
1801 vec<tree> *known_csts,
1802 vec<tree> *known_binfos,
1803 vec<ipa_agg_jump_function> *known_aggs,
1804 int *removable_params_cost)
1806 int i, count = ipa_get_param_count (info);
1807 bool ret = false;
1809 known_csts->create (0);
1810 known_binfos->create (0);
1811 known_csts->safe_grow_cleared (count);
1812 known_binfos->safe_grow_cleared (count);
1813 if (known_aggs)
1815 known_aggs->create (0);
1816 known_aggs->safe_grow_cleared (count);
1819 if (removable_params_cost)
1820 *removable_params_cost = 0;
1822 for (i = 0; i < count ; i++)
1824 struct ipcp_param_lattices *plats = ipa_get_parm_lattices (info, i);
1825 struct ipcp_lattice *lat = &plats->itself;
1827 if (ipa_lat_is_single_const (lat))
1829 struct ipcp_value *val = lat->values;
1830 if (TREE_CODE (val->value) != TREE_BINFO)
1832 (*known_csts)[i] = val->value;
1833 if (removable_params_cost)
1834 *removable_params_cost
1835 += estimate_move_cost (TREE_TYPE (val->value));
1836 ret = true;
1838 else if (plats->virt_call)
1840 (*known_binfos)[i] = val->value;
1841 ret = true;
1843 else if (removable_params_cost
1844 && !ipa_is_param_used (info, i))
1845 *removable_params_cost += ipa_get_param_move_cost (info, i);
1847 else if (removable_params_cost
1848 && !ipa_is_param_used (info, i))
1849 *removable_params_cost
1850 += ipa_get_param_move_cost (info, i);
1852 if (known_aggs)
1854 vec<ipa_agg_jf_item, va_gc> *agg_items;
1855 struct ipa_agg_jump_function *ajf;
1857 agg_items = context_independent_aggregate_values (plats);
1858 ajf = &(*known_aggs)[i];
1859 ajf->items = agg_items;
1860 ajf->by_ref = plats->aggs_by_ref;
1861 ret |= agg_items != NULL;
1865 return ret;
1868 /* The current interface in ipa-inline-analysis requires a pointer vector.
1869 Create it.
1871 FIXME: That interface should be re-worked, this is slightly silly. Still,
1872 I'd like to discuss how to change it first and this demonstrates the
1873 issue. */
1875 static vec<ipa_agg_jump_function_p>
1876 agg_jmp_p_vec_for_t_vec (vec<ipa_agg_jump_function> known_aggs)
1878 vec<ipa_agg_jump_function_p> ret;
1879 struct ipa_agg_jump_function *ajf;
1880 int i;
1882 ret.create (known_aggs.length ());
1883 FOR_EACH_VEC_ELT (known_aggs, i, ajf)
1884 ret.quick_push (ajf);
1885 return ret;
1888 /* Iterate over known values of parameters of NODE and estimate the local
1889 effects in terms of time and size they have. */
1891 static void
1892 estimate_local_effects (struct cgraph_node *node)
1894 struct ipa_node_params *info = IPA_NODE_REF (node);
1895 int i, count = ipa_get_param_count (info);
1896 vec<tree> known_csts, known_binfos;
1897 vec<ipa_agg_jump_function> known_aggs;
1898 vec<ipa_agg_jump_function_p> known_aggs_ptrs;
1899 bool always_const;
1900 int base_time = inline_summary (node)->time;
1901 int removable_params_cost;
1903 if (!count || !ipcp_versionable_function_p (node))
1904 return;
1906 if (dump_file && (dump_flags & TDF_DETAILS))
1907 fprintf (dump_file, "\nEstimating effects for %s/%i, base_time: %i.\n",
1908 node->name (), node->order, base_time);
1910 always_const = gather_context_independent_values (info, &known_csts,
1911 &known_binfos, &known_aggs,
1912 &removable_params_cost);
1913 known_aggs_ptrs = agg_jmp_p_vec_for_t_vec (known_aggs);
1914 if (always_const)
1916 struct caller_statistics stats;
1917 inline_hints hints;
1918 int time, size;
1920 init_caller_stats (&stats);
1921 cgraph_for_node_and_aliases (node, gather_caller_stats, &stats, false);
1922 estimate_ipcp_clone_size_and_time (node, known_csts, known_binfos,
1923 known_aggs_ptrs, &size, &time, &hints);
1924 time -= devirtualization_time_bonus (node, known_csts, known_binfos,
1925 known_aggs_ptrs);
1926 time -= hint_time_bonus (hints);
1927 time -= removable_params_cost;
1928 size -= stats.n_calls * removable_params_cost;
1930 if (dump_file)
1931 fprintf (dump_file, " - context independent values, size: %i, "
1932 "time_benefit: %i\n", size, base_time - time);
1934 if (size <= 0
1935 || cgraph_will_be_removed_from_program_if_no_direct_calls (node))
1937 info->do_clone_for_all_contexts = true;
1938 base_time = time;
1940 if (dump_file)
1941 fprintf (dump_file, " Decided to specialize for all "
1942 "known contexts, code not going to grow.\n");
1944 else if (good_cloning_opportunity_p (node, base_time - time,
1945 stats.freq_sum, stats.count_sum,
1946 size))
1948 if (size + overall_size <= max_new_size)
1950 info->do_clone_for_all_contexts = true;
1951 base_time = time;
1952 overall_size += size;
1954 if (dump_file)
1955 fprintf (dump_file, " Decided to specialize for all "
1956 "known contexts, growth deemed beneficial.\n");
1958 else if (dump_file && (dump_flags & TDF_DETAILS))
1959 fprintf (dump_file, " Not cloning for all contexts because "
1960 "max_new_size would be reached with %li.\n",
1961 size + overall_size);
1965 for (i = 0; i < count ; i++)
1967 struct ipcp_param_lattices *plats = ipa_get_parm_lattices (info, i);
1968 struct ipcp_lattice *lat = &plats->itself;
1969 struct ipcp_value *val;
1970 int emc;
1972 if (lat->bottom
1973 || !lat->values
1974 || known_csts[i]
1975 || known_binfos[i])
1976 continue;
1978 for (val = lat->values; val; val = val->next)
1980 int time, size, time_benefit;
1981 inline_hints hints;
1983 if (TREE_CODE (val->value) != TREE_BINFO)
1985 known_csts[i] = val->value;
1986 known_binfos[i] = NULL_TREE;
1987 emc = estimate_move_cost (TREE_TYPE (val->value));
1989 else if (plats->virt_call)
1991 known_csts[i] = NULL_TREE;
1992 known_binfos[i] = val->value;
1993 emc = 0;
1995 else
1996 continue;
1998 estimate_ipcp_clone_size_and_time (node, known_csts, known_binfos,
1999 known_aggs_ptrs, &size, &time,
2000 &hints);
2001 time_benefit = base_time - time
2002 + devirtualization_time_bonus (node, known_csts, known_binfos,
2003 known_aggs_ptrs)
2004 + hint_time_bonus (hints)
2005 + removable_params_cost + emc;
2007 gcc_checking_assert (size >=0);
2008 /* The inliner-heuristics based estimates may think that in certain
2009 contexts some functions do not have any size at all but we want
2010 all specializations to have at least a tiny cost, not least not to
2011 divide by zero. */
2012 if (size == 0)
2013 size = 1;
2015 if (dump_file && (dump_flags & TDF_DETAILS))
2017 fprintf (dump_file, " - estimates for value ");
2018 print_ipcp_constant_value (dump_file, val->value);
2019 fprintf (dump_file, " for ");
2020 ipa_dump_param (dump_file, info, i);
2021 fprintf (dump_file, ": time_benefit: %i, size: %i\n",
2022 time_benefit, size);
2025 val->local_time_benefit = time_benefit;
2026 val->local_size_cost = size;
2028 known_binfos[i] = NULL_TREE;
2029 known_csts[i] = NULL_TREE;
2032 for (i = 0; i < count ; i++)
2034 struct ipcp_param_lattices *plats = ipa_get_parm_lattices (info, i);
2035 struct ipa_agg_jump_function *ajf;
2036 struct ipcp_agg_lattice *aglat;
2038 if (plats->aggs_bottom || !plats->aggs)
2039 continue;
2041 ajf = &known_aggs[i];
2042 for (aglat = plats->aggs; aglat; aglat = aglat->next)
2044 struct ipcp_value *val;
2045 if (aglat->bottom || !aglat->values
2046 /* If the following is true, the one value is in known_aggs. */
2047 || (!plats->aggs_contain_variable
2048 && ipa_lat_is_single_const (aglat)))
2049 continue;
2051 for (val = aglat->values; val; val = val->next)
2053 int time, size, time_benefit;
2054 struct ipa_agg_jf_item item;
2055 inline_hints hints;
2057 item.offset = aglat->offset;
2058 item.value = val->value;
2059 vec_safe_push (ajf->items, item);
2061 estimate_ipcp_clone_size_and_time (node, known_csts, known_binfos,
2062 known_aggs_ptrs, &size, &time,
2063 &hints);
2064 time_benefit = base_time - time
2065 + devirtualization_time_bonus (node, known_csts, known_binfos,
2066 known_aggs_ptrs)
2067 + hint_time_bonus (hints);
2068 gcc_checking_assert (size >=0);
2069 if (size == 0)
2070 size = 1;
2072 if (dump_file && (dump_flags & TDF_DETAILS))
2074 fprintf (dump_file, " - estimates for value ");
2075 print_ipcp_constant_value (dump_file, val->value);
2076 fprintf (dump_file, " for ");
2077 ipa_dump_param (dump_file, info, i);
2078 fprintf (dump_file, "[%soffset: " HOST_WIDE_INT_PRINT_DEC
2079 "]: time_benefit: %i, size: %i\n",
2080 plats->aggs_by_ref ? "ref " : "",
2081 aglat->offset, time_benefit, size);
2084 val->local_time_benefit = time_benefit;
2085 val->local_size_cost = size;
2086 ajf->items->pop ();
2091 for (i = 0; i < count ; i++)
2092 vec_free (known_aggs[i].items);
2094 known_csts.release ();
2095 known_binfos.release ();
2096 known_aggs.release ();
2097 known_aggs_ptrs.release ();
2101 /* Add value CUR_VAL and all yet-unsorted values it is dependent on to the
2102 topological sort of values. */
2104 static void
2105 add_val_to_toposort (struct ipcp_value *cur_val)
2107 static int dfs_counter = 0;
2108 static struct ipcp_value *stack;
2109 struct ipcp_value_source *src;
2111 if (cur_val->dfs)
2112 return;
2114 dfs_counter++;
2115 cur_val->dfs = dfs_counter;
2116 cur_val->low_link = dfs_counter;
2118 cur_val->topo_next = stack;
2119 stack = cur_val;
2120 cur_val->on_stack = true;
2122 for (src = cur_val->sources; src; src = src->next)
2123 if (src->val)
2125 if (src->val->dfs == 0)
2127 add_val_to_toposort (src->val);
2128 if (src->val->low_link < cur_val->low_link)
2129 cur_val->low_link = src->val->low_link;
2131 else if (src->val->on_stack
2132 && src->val->dfs < cur_val->low_link)
2133 cur_val->low_link = src->val->dfs;
2136 if (cur_val->dfs == cur_val->low_link)
2138 struct ipcp_value *v, *scc_list = NULL;
2142 v = stack;
2143 stack = v->topo_next;
2144 v->on_stack = false;
2146 v->scc_next = scc_list;
2147 scc_list = v;
2149 while (v != cur_val);
2151 cur_val->topo_next = values_topo;
2152 values_topo = cur_val;
2156 /* Add all values in lattices associated with NODE to the topological sort if
2157 they are not there yet. */
2159 static void
2160 add_all_node_vals_to_toposort (struct cgraph_node *node)
2162 struct ipa_node_params *info = IPA_NODE_REF (node);
2163 int i, count = ipa_get_param_count (info);
2165 for (i = 0; i < count ; i++)
2167 struct ipcp_param_lattices *plats = ipa_get_parm_lattices (info, i);
2168 struct ipcp_lattice *lat = &plats->itself;
2169 struct ipcp_agg_lattice *aglat;
2170 struct ipcp_value *val;
2172 if (!lat->bottom)
2173 for (val = lat->values; val; val = val->next)
2174 add_val_to_toposort (val);
2176 if (!plats->aggs_bottom)
2177 for (aglat = plats->aggs; aglat; aglat = aglat->next)
2178 if (!aglat->bottom)
2179 for (val = aglat->values; val; val = val->next)
2180 add_val_to_toposort (val);
2184 /* One pass of constants propagation along the call graph edges, from callers
2185 to callees (requires topological ordering in TOPO), iterate over strongly
2186 connected components. */
2188 static void
2189 propagate_constants_topo (struct topo_info *topo)
2191 int i;
2193 for (i = topo->nnodes - 1; i >= 0; i--)
2195 unsigned j;
2196 struct cgraph_node *v, *node = topo->order[i];
2197 vec<cgraph_node_ptr> cycle_nodes = ipa_get_nodes_in_cycle (node);
2199 /* First, iteratively propagate within the strongly connected component
2200 until all lattices stabilize. */
2201 FOR_EACH_VEC_ELT (cycle_nodes, j, v)
2202 if (cgraph_function_with_gimple_body_p (v))
2203 push_node_to_stack (topo, v);
2205 v = pop_node_from_stack (topo);
2206 while (v)
2208 struct cgraph_edge *cs;
2210 for (cs = v->callees; cs; cs = cs->next_callee)
2211 if (ipa_edge_within_scc (cs)
2212 && propagate_constants_accross_call (cs))
2213 push_node_to_stack (topo, cs->callee);
2214 v = pop_node_from_stack (topo);
2217 /* Afterwards, propagate along edges leading out of the SCC, calculates
2218 the local effects of the discovered constants and all valid values to
2219 their topological sort. */
2220 FOR_EACH_VEC_ELT (cycle_nodes, j, v)
2221 if (cgraph_function_with_gimple_body_p (v))
2223 struct cgraph_edge *cs;
2225 estimate_local_effects (v);
2226 add_all_node_vals_to_toposort (v);
2227 for (cs = v->callees; cs; cs = cs->next_callee)
2228 if (!ipa_edge_within_scc (cs))
2229 propagate_constants_accross_call (cs);
2231 cycle_nodes.release ();
2236 /* Return the sum of A and B if none of them is bigger than INT_MAX/2, return
2237 the bigger one if otherwise. */
2239 static int
2240 safe_add (int a, int b)
2242 if (a > INT_MAX/2 || b > INT_MAX/2)
2243 return a > b ? a : b;
2244 else
2245 return a + b;
2249 /* Propagate the estimated effects of individual values along the topological
2250 from the dependent values to those they depend on. */
2252 static void
2253 propagate_effects (void)
2255 struct ipcp_value *base;
2257 for (base = values_topo; base; base = base->topo_next)
2259 struct ipcp_value_source *src;
2260 struct ipcp_value *val;
2261 int time = 0, size = 0;
2263 for (val = base; val; val = val->scc_next)
2265 time = safe_add (time,
2266 val->local_time_benefit + val->prop_time_benefit);
2267 size = safe_add (size, val->local_size_cost + val->prop_size_cost);
2270 for (val = base; val; val = val->scc_next)
2271 for (src = val->sources; src; src = src->next)
2272 if (src->val
2273 && cgraph_maybe_hot_edge_p (src->cs))
2275 src->val->prop_time_benefit = safe_add (time,
2276 src->val->prop_time_benefit);
2277 src->val->prop_size_cost = safe_add (size,
2278 src->val->prop_size_cost);
2284 /* Propagate constants, binfos and their effects from the summaries
2285 interprocedurally. */
2287 static void
2288 ipcp_propagate_stage (struct topo_info *topo)
2290 struct cgraph_node *node;
2292 if (dump_file)
2293 fprintf (dump_file, "\n Propagating constants:\n\n");
2295 if (in_lto_p)
2296 ipa_update_after_lto_read ();
2299 FOR_EACH_DEFINED_FUNCTION (node)
2301 struct ipa_node_params *info = IPA_NODE_REF (node);
2303 determine_versionability (node);
2304 if (cgraph_function_with_gimple_body_p (node))
2306 info->lattices = XCNEWVEC (struct ipcp_param_lattices,
2307 ipa_get_param_count (info));
2308 initialize_node_lattices (node);
2310 if (node->definition && !node->alias)
2311 overall_size += inline_summary (node)->self_size;
2312 if (node->count > max_count)
2313 max_count = node->count;
2316 max_new_size = overall_size;
2317 if (max_new_size < PARAM_VALUE (PARAM_LARGE_UNIT_INSNS))
2318 max_new_size = PARAM_VALUE (PARAM_LARGE_UNIT_INSNS);
2319 max_new_size += max_new_size * PARAM_VALUE (PARAM_IPCP_UNIT_GROWTH) / 100 + 1;
2321 if (dump_file)
2322 fprintf (dump_file, "\noverall_size: %li, max_new_size: %li\n",
2323 overall_size, max_new_size);
2325 propagate_constants_topo (topo);
2326 #ifdef ENABLE_CHECKING
2327 ipcp_verify_propagated_values ();
2328 #endif
2329 propagate_effects ();
2331 if (dump_file)
2333 fprintf (dump_file, "\nIPA lattices after all propagation:\n");
2334 print_all_lattices (dump_file, (dump_flags & TDF_DETAILS), true);
2338 /* Discover newly direct outgoing edges from NODE which is a new clone with
2339 known KNOWN_VALS and make them direct. */
2341 static void
2342 ipcp_discover_new_direct_edges (struct cgraph_node *node,
2343 vec<tree> known_vals,
2344 struct ipa_agg_replacement_value *aggvals)
2346 struct cgraph_edge *ie, *next_ie;
2347 bool found = false;
2349 for (ie = node->indirect_calls; ie; ie = next_ie)
2351 tree target;
2353 next_ie = ie->next_callee;
2354 target = ipa_get_indirect_edge_target_1 (ie, known_vals, vNULL, vNULL,
2355 aggvals);
2356 if (target)
2358 bool agg_contents = ie->indirect_info->agg_contents;
2359 bool polymorphic = ie->indirect_info->polymorphic;
2360 int param_index = ie->indirect_info->param_index;
2361 struct cgraph_edge *cs = ipa_make_edge_direct_to_target (ie, target);
2362 found = true;
2364 if (cs && !agg_contents && !polymorphic)
2366 struct ipa_node_params *info = IPA_NODE_REF (node);
2367 int c = ipa_get_controlled_uses (info, param_index);
2368 if (c != IPA_UNDESCRIBED_USE)
2370 struct ipa_ref *to_del;
2372 c--;
2373 ipa_set_controlled_uses (info, param_index, c);
2374 if (dump_file && (dump_flags & TDF_DETAILS))
2375 fprintf (dump_file, " controlled uses count of param "
2376 "%i bumped down to %i\n", param_index, c);
2377 if (c == 0
2378 && (to_del = node->find_reference (cs->callee, NULL, 0)))
2380 if (dump_file && (dump_flags & TDF_DETAILS))
2381 fprintf (dump_file, " and even removing its "
2382 "cloning-created reference\n");
2383 to_del->remove_reference ();
2389 /* Turning calls to direct calls will improve overall summary. */
2390 if (found)
2391 inline_update_overall_summary (node);
2394 /* Vector of pointers which for linked lists of clones of an original crgaph
2395 edge. */
2397 static vec<cgraph_edge_p> next_edge_clone;
2398 static vec<cgraph_edge_p> prev_edge_clone;
2400 static inline void
2401 grow_edge_clone_vectors (void)
2403 if (next_edge_clone.length ()
2404 <= (unsigned) cgraph_edge_max_uid)
2405 next_edge_clone.safe_grow_cleared (cgraph_edge_max_uid + 1);
2406 if (prev_edge_clone.length ()
2407 <= (unsigned) cgraph_edge_max_uid)
2408 prev_edge_clone.safe_grow_cleared (cgraph_edge_max_uid + 1);
2411 /* Edge duplication hook to grow the appropriate linked list in
2412 next_edge_clone. */
2414 static void
2415 ipcp_edge_duplication_hook (struct cgraph_edge *src, struct cgraph_edge *dst,
2416 void *)
2418 grow_edge_clone_vectors ();
2420 struct cgraph_edge *old_next = next_edge_clone[src->uid];
2421 if (old_next)
2422 prev_edge_clone[old_next->uid] = dst;
2423 prev_edge_clone[dst->uid] = src;
2425 next_edge_clone[dst->uid] = old_next;
2426 next_edge_clone[src->uid] = dst;
2429 /* Hook that is called by cgraph.c when an edge is removed. */
2431 static void
2432 ipcp_edge_removal_hook (struct cgraph_edge *cs, void *)
2434 grow_edge_clone_vectors ();
2436 struct cgraph_edge *prev = prev_edge_clone[cs->uid];
2437 struct cgraph_edge *next = next_edge_clone[cs->uid];
2438 if (prev)
2439 next_edge_clone[prev->uid] = next;
2440 if (next)
2441 prev_edge_clone[next->uid] = prev;
2444 /* See if NODE is a clone with a known aggregate value at a given OFFSET of a
2445 parameter with the given INDEX. */
2447 static tree
2448 get_clone_agg_value (struct cgraph_node *node, HOST_WIDE_INT offset,
2449 int index)
2451 struct ipa_agg_replacement_value *aggval;
2453 aggval = ipa_get_agg_replacements_for_node (node);
2454 while (aggval)
2456 if (aggval->offset == offset
2457 && aggval->index == index)
2458 return aggval->value;
2459 aggval = aggval->next;
2461 return NULL_TREE;
2464 /* Return true if edge CS does bring about the value described by SRC. */
2466 static bool
2467 cgraph_edge_brings_value_p (struct cgraph_edge *cs,
2468 struct ipcp_value_source *src)
2470 struct ipa_node_params *caller_info = IPA_NODE_REF (cs->caller);
2471 cgraph_node *real_dest = cgraph_function_node (cs->callee);
2472 struct ipa_node_params *dst_info = IPA_NODE_REF (real_dest);
2474 if ((dst_info->ipcp_orig_node && !dst_info->is_all_contexts_clone)
2475 || caller_info->node_dead)
2476 return false;
2477 if (!src->val)
2478 return true;
2480 if (caller_info->ipcp_orig_node)
2482 tree t;
2483 if (src->offset == -1)
2484 t = caller_info->known_vals[src->index];
2485 else
2486 t = get_clone_agg_value (cs->caller, src->offset, src->index);
2487 return (t != NULL_TREE
2488 && values_equal_for_ipcp_p (src->val->value, t));
2490 else
2492 struct ipcp_agg_lattice *aglat;
2493 struct ipcp_param_lattices *plats = ipa_get_parm_lattices (caller_info,
2494 src->index);
2495 if (src->offset == -1)
2496 return (ipa_lat_is_single_const (&plats->itself)
2497 && values_equal_for_ipcp_p (src->val->value,
2498 plats->itself.values->value));
2499 else
2501 if (plats->aggs_bottom || plats->aggs_contain_variable)
2502 return false;
2503 for (aglat = plats->aggs; aglat; aglat = aglat->next)
2504 if (aglat->offset == src->offset)
2505 return (ipa_lat_is_single_const (aglat)
2506 && values_equal_for_ipcp_p (src->val->value,
2507 aglat->values->value));
2509 return false;
2513 /* Get the next clone in the linked list of clones of an edge. */
2515 static inline struct cgraph_edge *
2516 get_next_cgraph_edge_clone (struct cgraph_edge *cs)
2518 return next_edge_clone[cs->uid];
2521 /* Given VAL, iterate over all its sources and if they still hold, add their
2522 edge frequency and their number into *FREQUENCY and *CALLER_COUNT
2523 respectively. */
2525 static bool
2526 get_info_about_necessary_edges (struct ipcp_value *val, int *freq_sum,
2527 gcov_type *count_sum, int *caller_count)
2529 struct ipcp_value_source *src;
2530 int freq = 0, count = 0;
2531 gcov_type cnt = 0;
2532 bool hot = false;
2534 for (src = val->sources; src; src = src->next)
2536 struct cgraph_edge *cs = src->cs;
2537 while (cs)
2539 if (cgraph_edge_brings_value_p (cs, src))
2541 count++;
2542 freq += cs->frequency;
2543 cnt += cs->count;
2544 hot |= cgraph_maybe_hot_edge_p (cs);
2546 cs = get_next_cgraph_edge_clone (cs);
2550 *freq_sum = freq;
2551 *count_sum = cnt;
2552 *caller_count = count;
2553 return hot;
2556 /* Return a vector of incoming edges that do bring value VAL. It is assumed
2557 their number is known and equal to CALLER_COUNT. */
2559 static vec<cgraph_edge_p>
2560 gather_edges_for_value (struct ipcp_value *val, int caller_count)
2562 struct ipcp_value_source *src;
2563 vec<cgraph_edge_p> ret;
2565 ret.create (caller_count);
2566 for (src = val->sources; src; src = src->next)
2568 struct cgraph_edge *cs = src->cs;
2569 while (cs)
2571 if (cgraph_edge_brings_value_p (cs, src))
2572 ret.quick_push (cs);
2573 cs = get_next_cgraph_edge_clone (cs);
2577 return ret;
2580 /* Construct a replacement map for a know VALUE for a formal parameter PARAM.
2581 Return it or NULL if for some reason it cannot be created. */
2583 static struct ipa_replace_map *
2584 get_replacement_map (struct ipa_node_params *info, tree value, int parm_num)
2586 struct ipa_replace_map *replace_map;
2589 replace_map = ggc_alloc<ipa_replace_map> ();
2590 if (dump_file)
2592 fprintf (dump_file, " replacing ");
2593 ipa_dump_param (dump_file, info, parm_num);
2595 fprintf (dump_file, " with const ");
2596 print_generic_expr (dump_file, value, 0);
2597 fprintf (dump_file, "\n");
2599 replace_map->old_tree = NULL;
2600 replace_map->parm_num = parm_num;
2601 replace_map->new_tree = value;
2602 replace_map->replace_p = true;
2603 replace_map->ref_p = false;
2605 return replace_map;
2608 /* Dump new profiling counts */
2610 static void
2611 dump_profile_updates (struct cgraph_node *orig_node,
2612 struct cgraph_node *new_node)
2614 struct cgraph_edge *cs;
2616 fprintf (dump_file, " setting count of the specialized node to "
2617 HOST_WIDE_INT_PRINT_DEC "\n", (HOST_WIDE_INT) new_node->count);
2618 for (cs = new_node->callees; cs ; cs = cs->next_callee)
2619 fprintf (dump_file, " edge to %s has count "
2620 HOST_WIDE_INT_PRINT_DEC "\n",
2621 cs->callee->name (), (HOST_WIDE_INT) cs->count);
2623 fprintf (dump_file, " setting count of the original node to "
2624 HOST_WIDE_INT_PRINT_DEC "\n", (HOST_WIDE_INT) orig_node->count);
2625 for (cs = orig_node->callees; cs ; cs = cs->next_callee)
2626 fprintf (dump_file, " edge to %s is left with "
2627 HOST_WIDE_INT_PRINT_DEC "\n",
2628 cs->callee->name (), (HOST_WIDE_INT) cs->count);
2631 /* After a specialized NEW_NODE version of ORIG_NODE has been created, update
2632 their profile information to reflect this. */
2634 static void
2635 update_profiling_info (struct cgraph_node *orig_node,
2636 struct cgraph_node *new_node)
2638 struct cgraph_edge *cs;
2639 struct caller_statistics stats;
2640 gcov_type new_sum, orig_sum;
2641 gcov_type remainder, orig_node_count = orig_node->count;
2643 if (orig_node_count == 0)
2644 return;
2646 init_caller_stats (&stats);
2647 cgraph_for_node_and_aliases (orig_node, gather_caller_stats, &stats, false);
2648 orig_sum = stats.count_sum;
2649 init_caller_stats (&stats);
2650 cgraph_for_node_and_aliases (new_node, gather_caller_stats, &stats, false);
2651 new_sum = stats.count_sum;
2653 if (orig_node_count < orig_sum + new_sum)
2655 if (dump_file)
2656 fprintf (dump_file, " Problem: node %s/%i has too low count "
2657 HOST_WIDE_INT_PRINT_DEC " while the sum of incoming "
2658 "counts is " HOST_WIDE_INT_PRINT_DEC "\n",
2659 orig_node->name (), orig_node->order,
2660 (HOST_WIDE_INT) orig_node_count,
2661 (HOST_WIDE_INT) (orig_sum + new_sum));
2663 orig_node_count = (orig_sum + new_sum) * 12 / 10;
2664 if (dump_file)
2665 fprintf (dump_file, " proceeding by pretending it was "
2666 HOST_WIDE_INT_PRINT_DEC "\n",
2667 (HOST_WIDE_INT) orig_node_count);
2670 new_node->count = new_sum;
2671 remainder = orig_node_count - new_sum;
2672 orig_node->count = remainder;
2674 for (cs = new_node->callees; cs ; cs = cs->next_callee)
2675 if (cs->frequency)
2676 cs->count = apply_probability (cs->count,
2677 GCOV_COMPUTE_SCALE (new_sum,
2678 orig_node_count));
2679 else
2680 cs->count = 0;
2682 for (cs = orig_node->callees; cs ; cs = cs->next_callee)
2683 cs->count = apply_probability (cs->count,
2684 GCOV_COMPUTE_SCALE (remainder,
2685 orig_node_count));
2687 if (dump_file)
2688 dump_profile_updates (orig_node, new_node);
2691 /* Update the respective profile of specialized NEW_NODE and the original
2692 ORIG_NODE after additional edges with cumulative count sum REDIRECTED_SUM
2693 have been redirected to the specialized version. */
2695 static void
2696 update_specialized_profile (struct cgraph_node *new_node,
2697 struct cgraph_node *orig_node,
2698 gcov_type redirected_sum)
2700 struct cgraph_edge *cs;
2701 gcov_type new_node_count, orig_node_count = orig_node->count;
2703 if (dump_file)
2704 fprintf (dump_file, " the sum of counts of redirected edges is "
2705 HOST_WIDE_INT_PRINT_DEC "\n", (HOST_WIDE_INT) redirected_sum);
2706 if (orig_node_count == 0)
2707 return;
2709 gcc_assert (orig_node_count >= redirected_sum);
2711 new_node_count = new_node->count;
2712 new_node->count += redirected_sum;
2713 orig_node->count -= redirected_sum;
2715 for (cs = new_node->callees; cs ; cs = cs->next_callee)
2716 if (cs->frequency)
2717 cs->count += apply_probability (cs->count,
2718 GCOV_COMPUTE_SCALE (redirected_sum,
2719 new_node_count));
2720 else
2721 cs->count = 0;
2723 for (cs = orig_node->callees; cs ; cs = cs->next_callee)
2725 gcov_type dec = apply_probability (cs->count,
2726 GCOV_COMPUTE_SCALE (redirected_sum,
2727 orig_node_count));
2728 if (dec < cs->count)
2729 cs->count -= dec;
2730 else
2731 cs->count = 0;
2734 if (dump_file)
2735 dump_profile_updates (orig_node, new_node);
2738 /* Create a specialized version of NODE with known constants and types of
2739 parameters in KNOWN_VALS and redirect all edges in CALLERS to it. */
2741 static struct cgraph_node *
2742 create_specialized_node (struct cgraph_node *node,
2743 vec<tree> known_vals,
2744 struct ipa_agg_replacement_value *aggvals,
2745 vec<cgraph_edge_p> callers)
2747 struct ipa_node_params *new_info, *info = IPA_NODE_REF (node);
2748 vec<ipa_replace_map_p, va_gc> *replace_trees = NULL;
2749 struct ipa_agg_replacement_value *av;
2750 struct cgraph_node *new_node;
2751 int i, count = ipa_get_param_count (info);
2752 bitmap args_to_skip;
2754 gcc_assert (!info->ipcp_orig_node);
2756 if (node->local.can_change_signature)
2758 args_to_skip = BITMAP_GGC_ALLOC ();
2759 for (i = 0; i < count; i++)
2761 tree t = known_vals[i];
2763 if ((t && TREE_CODE (t) != TREE_BINFO)
2764 || !ipa_is_param_used (info, i))
2765 bitmap_set_bit (args_to_skip, i);
2768 else
2770 args_to_skip = NULL;
2771 if (dump_file && (dump_flags & TDF_DETAILS))
2772 fprintf (dump_file, " cannot change function signature\n");
2775 for (i = 0; i < count ; i++)
2777 tree t = known_vals[i];
2778 if (t && TREE_CODE (t) != TREE_BINFO)
2780 struct ipa_replace_map *replace_map;
2782 replace_map = get_replacement_map (info, t, i);
2783 if (replace_map)
2784 vec_safe_push (replace_trees, replace_map);
2788 new_node = cgraph_create_virtual_clone (node, callers, replace_trees,
2789 args_to_skip, "constprop");
2790 ipa_set_node_agg_value_chain (new_node, aggvals);
2791 for (av = aggvals; av; av = av->next)
2792 new_node->maybe_add_reference (av->value, IPA_REF_ADDR, NULL);
2794 if (dump_file && (dump_flags & TDF_DETAILS))
2796 fprintf (dump_file, " the new node is %s/%i.\n",
2797 new_node->name (), new_node->order);
2798 if (aggvals)
2799 ipa_dump_agg_replacement_values (dump_file, aggvals);
2801 ipa_check_create_node_params ();
2802 update_profiling_info (node, new_node);
2803 new_info = IPA_NODE_REF (new_node);
2804 new_info->ipcp_orig_node = node;
2805 new_info->known_vals = known_vals;
2807 ipcp_discover_new_direct_edges (new_node, known_vals, aggvals);
2809 callers.release ();
2810 return new_node;
2813 /* Given a NODE, and a subset of its CALLERS, try to populate blanks slots in
2814 KNOWN_VALS with constants and types that are also known for all of the
2815 CALLERS. */
2817 static void
2818 find_more_scalar_values_for_callers_subset (struct cgraph_node *node,
2819 vec<tree> known_vals,
2820 vec<cgraph_edge_p> callers)
2822 struct ipa_node_params *info = IPA_NODE_REF (node);
2823 int i, count = ipa_get_param_count (info);
2825 for (i = 0; i < count ; i++)
2827 struct cgraph_edge *cs;
2828 tree newval = NULL_TREE;
2829 int j;
2831 if (ipa_get_scalar_lat (info, i)->bottom || known_vals[i])
2832 continue;
2834 FOR_EACH_VEC_ELT (callers, j, cs)
2836 struct ipa_jump_func *jump_func;
2837 tree t;
2839 if (i >= ipa_get_cs_argument_count (IPA_EDGE_REF (cs)))
2841 newval = NULL_TREE;
2842 break;
2844 jump_func = ipa_get_ith_jump_func (IPA_EDGE_REF (cs), i);
2845 t = ipa_value_from_jfunc (IPA_NODE_REF (cs->caller), jump_func);
2846 if (!t
2847 || (newval
2848 && !values_equal_for_ipcp_p (t, newval)))
2850 newval = NULL_TREE;
2851 break;
2853 else
2854 newval = t;
2857 if (newval)
2859 if (dump_file && (dump_flags & TDF_DETAILS))
2861 fprintf (dump_file, " adding an extra known scalar value ");
2862 print_ipcp_constant_value (dump_file, newval);
2863 fprintf (dump_file, " for ");
2864 ipa_dump_param (dump_file, info, i);
2865 fprintf (dump_file, "\n");
2868 known_vals[i] = newval;
2873 /* Go through PLATS and create a vector of values consisting of values and
2874 offsets (minus OFFSET) of lattices that contain only a single value. */
2876 static vec<ipa_agg_jf_item>
2877 copy_plats_to_inter (struct ipcp_param_lattices *plats, HOST_WIDE_INT offset)
2879 vec<ipa_agg_jf_item> res = vNULL;
2881 if (!plats->aggs || plats->aggs_contain_variable || plats->aggs_bottom)
2882 return vNULL;
2884 for (struct ipcp_agg_lattice *aglat = plats->aggs; aglat; aglat = aglat->next)
2885 if (ipa_lat_is_single_const (aglat))
2887 struct ipa_agg_jf_item ti;
2888 ti.offset = aglat->offset - offset;
2889 ti.value = aglat->values->value;
2890 res.safe_push (ti);
2892 return res;
2895 /* Intersect all values in INTER with single value lattices in PLATS (while
2896 subtracting OFFSET). */
2898 static void
2899 intersect_with_plats (struct ipcp_param_lattices *plats,
2900 vec<ipa_agg_jf_item> *inter,
2901 HOST_WIDE_INT offset)
2903 struct ipcp_agg_lattice *aglat;
2904 struct ipa_agg_jf_item *item;
2905 int k;
2907 if (!plats->aggs || plats->aggs_contain_variable || plats->aggs_bottom)
2909 inter->release ();
2910 return;
2913 aglat = plats->aggs;
2914 FOR_EACH_VEC_ELT (*inter, k, item)
2916 bool found = false;
2917 if (!item->value)
2918 continue;
2919 while (aglat)
2921 if (aglat->offset - offset > item->offset)
2922 break;
2923 if (aglat->offset - offset == item->offset)
2925 gcc_checking_assert (item->value);
2926 if (values_equal_for_ipcp_p (item->value, aglat->values->value))
2927 found = true;
2928 break;
2930 aglat = aglat->next;
2932 if (!found)
2933 item->value = NULL_TREE;
2937 /* Copy agggregate replacement values of NODE (which is an IPA-CP clone) to the
2938 vector result while subtracting OFFSET from the individual value offsets. */
2940 static vec<ipa_agg_jf_item>
2941 agg_replacements_to_vector (struct cgraph_node *node, int index,
2942 HOST_WIDE_INT offset)
2944 struct ipa_agg_replacement_value *av;
2945 vec<ipa_agg_jf_item> res = vNULL;
2947 for (av = ipa_get_agg_replacements_for_node (node); av; av = av->next)
2948 if (av->index == index
2949 && (av->offset - offset) >= 0)
2951 struct ipa_agg_jf_item item;
2952 gcc_checking_assert (av->value);
2953 item.offset = av->offset - offset;
2954 item.value = av->value;
2955 res.safe_push (item);
2958 return res;
2961 /* Intersect all values in INTER with those that we have already scheduled to
2962 be replaced in parameter number INDEX of NODE, which is an IPA-CP clone
2963 (while subtracting OFFSET). */
2965 static void
2966 intersect_with_agg_replacements (struct cgraph_node *node, int index,
2967 vec<ipa_agg_jf_item> *inter,
2968 HOST_WIDE_INT offset)
2970 struct ipa_agg_replacement_value *srcvals;
2971 struct ipa_agg_jf_item *item;
2972 int i;
2974 srcvals = ipa_get_agg_replacements_for_node (node);
2975 if (!srcvals)
2977 inter->release ();
2978 return;
2981 FOR_EACH_VEC_ELT (*inter, i, item)
2983 struct ipa_agg_replacement_value *av;
2984 bool found = false;
2985 if (!item->value)
2986 continue;
2987 for (av = srcvals; av; av = av->next)
2989 gcc_checking_assert (av->value);
2990 if (av->index == index
2991 && av->offset - offset == item->offset)
2993 if (values_equal_for_ipcp_p (item->value, av->value))
2994 found = true;
2995 break;
2998 if (!found)
2999 item->value = NULL_TREE;
3003 /* Intersect values in INTER with aggregate values that come along edge CS to
3004 parameter number INDEX and return it. If INTER does not actually exist yet,
3005 copy all incoming values to it. If we determine we ended up with no values
3006 whatsoever, return a released vector. */
3008 static vec<ipa_agg_jf_item>
3009 intersect_aggregates_with_edge (struct cgraph_edge *cs, int index,
3010 vec<ipa_agg_jf_item> inter)
3012 struct ipa_jump_func *jfunc;
3013 jfunc = ipa_get_ith_jump_func (IPA_EDGE_REF (cs), index);
3014 if (jfunc->type == IPA_JF_PASS_THROUGH
3015 && ipa_get_jf_pass_through_operation (jfunc) == NOP_EXPR)
3017 struct ipa_node_params *caller_info = IPA_NODE_REF (cs->caller);
3018 int src_idx = ipa_get_jf_pass_through_formal_id (jfunc);
3020 if (caller_info->ipcp_orig_node)
3022 struct cgraph_node *orig_node = caller_info->ipcp_orig_node;
3023 struct ipcp_param_lattices *orig_plats;
3024 orig_plats = ipa_get_parm_lattices (IPA_NODE_REF (orig_node),
3025 src_idx);
3026 if (agg_pass_through_permissible_p (orig_plats, jfunc))
3028 if (!inter.exists ())
3029 inter = agg_replacements_to_vector (cs->caller, src_idx, 0);
3030 else
3031 intersect_with_agg_replacements (cs->caller, src_idx,
3032 &inter, 0);
3035 else
3037 struct ipcp_param_lattices *src_plats;
3038 src_plats = ipa_get_parm_lattices (caller_info, src_idx);
3039 if (agg_pass_through_permissible_p (src_plats, jfunc))
3041 /* Currently we do not produce clobber aggregate jump
3042 functions, adjust when we do. */
3043 gcc_checking_assert (!jfunc->agg.items);
3044 if (!inter.exists ())
3045 inter = copy_plats_to_inter (src_plats, 0);
3046 else
3047 intersect_with_plats (src_plats, &inter, 0);
3051 else if (jfunc->type == IPA_JF_ANCESTOR
3052 && ipa_get_jf_ancestor_agg_preserved (jfunc))
3054 struct ipa_node_params *caller_info = IPA_NODE_REF (cs->caller);
3055 int src_idx = ipa_get_jf_ancestor_formal_id (jfunc);
3056 struct ipcp_param_lattices *src_plats;
3057 HOST_WIDE_INT delta = ipa_get_jf_ancestor_offset (jfunc);
3059 if (caller_info->ipcp_orig_node)
3061 if (!inter.exists ())
3062 inter = agg_replacements_to_vector (cs->caller, src_idx, delta);
3063 else
3064 intersect_with_agg_replacements (cs->caller, src_idx, &inter,
3065 delta);
3067 else
3069 src_plats = ipa_get_parm_lattices (caller_info, src_idx);;
3070 /* Currently we do not produce clobber aggregate jump
3071 functions, adjust when we do. */
3072 gcc_checking_assert (!src_plats->aggs || !jfunc->agg.items);
3073 if (!inter.exists ())
3074 inter = copy_plats_to_inter (src_plats, delta);
3075 else
3076 intersect_with_plats (src_plats, &inter, delta);
3079 else if (jfunc->agg.items)
3081 struct ipa_agg_jf_item *item;
3082 int k;
3084 if (!inter.exists ())
3085 for (unsigned i = 0; i < jfunc->agg.items->length (); i++)
3086 inter.safe_push ((*jfunc->agg.items)[i]);
3087 else
3088 FOR_EACH_VEC_ELT (inter, k, item)
3090 int l = 0;
3091 bool found = false;;
3093 if (!item->value)
3094 continue;
3096 while ((unsigned) l < jfunc->agg.items->length ())
3098 struct ipa_agg_jf_item *ti;
3099 ti = &(*jfunc->agg.items)[l];
3100 if (ti->offset > item->offset)
3101 break;
3102 if (ti->offset == item->offset)
3104 gcc_checking_assert (ti->value);
3105 if (values_equal_for_ipcp_p (item->value,
3106 ti->value))
3107 found = true;
3108 break;
3110 l++;
3112 if (!found)
3113 item->value = NULL;
3116 else
3118 inter.release ();
3119 return vec<ipa_agg_jf_item>();
3121 return inter;
3124 /* Look at edges in CALLERS and collect all known aggregate values that arrive
3125 from all of them. */
3127 static struct ipa_agg_replacement_value *
3128 find_aggregate_values_for_callers_subset (struct cgraph_node *node,
3129 vec<cgraph_edge_p> callers)
3131 struct ipa_node_params *dest_info = IPA_NODE_REF (node);
3132 struct ipa_agg_replacement_value *res = NULL;
3133 struct cgraph_edge *cs;
3134 int i, j, count = ipa_get_param_count (dest_info);
3136 FOR_EACH_VEC_ELT (callers, j, cs)
3138 int c = ipa_get_cs_argument_count (IPA_EDGE_REF (cs));
3139 if (c < count)
3140 count = c;
3143 for (i = 0; i < count ; i++)
3145 struct cgraph_edge *cs;
3146 vec<ipa_agg_jf_item> inter = vNULL;
3147 struct ipa_agg_jf_item *item;
3148 struct ipcp_param_lattices *plats = ipa_get_parm_lattices (dest_info, i);
3149 int j;
3151 /* Among other things, the following check should deal with all by_ref
3152 mismatches. */
3153 if (plats->aggs_bottom)
3154 continue;
3156 FOR_EACH_VEC_ELT (callers, j, cs)
3158 inter = intersect_aggregates_with_edge (cs, i, inter);
3160 if (!inter.exists ())
3161 goto next_param;
3164 FOR_EACH_VEC_ELT (inter, j, item)
3166 struct ipa_agg_replacement_value *v;
3168 if (!item->value)
3169 continue;
3171 v = ggc_alloc<ipa_agg_replacement_value> ();
3172 v->index = i;
3173 v->offset = item->offset;
3174 v->value = item->value;
3175 v->by_ref = plats->aggs_by_ref;
3176 v->next = res;
3177 res = v;
3180 next_param:
3181 if (inter.exists ())
3182 inter.release ();
3184 return res;
3187 /* Turn KNOWN_AGGS into a list of aggreate replacement values. */
3189 static struct ipa_agg_replacement_value *
3190 known_aggs_to_agg_replacement_list (vec<ipa_agg_jump_function> known_aggs)
3192 struct ipa_agg_replacement_value *res = NULL;
3193 struct ipa_agg_jump_function *aggjf;
3194 struct ipa_agg_jf_item *item;
3195 int i, j;
3197 FOR_EACH_VEC_ELT (known_aggs, i, aggjf)
3198 FOR_EACH_VEC_SAFE_ELT (aggjf->items, j, item)
3200 struct ipa_agg_replacement_value *v;
3201 v = ggc_alloc<ipa_agg_replacement_value> ();
3202 v->index = i;
3203 v->offset = item->offset;
3204 v->value = item->value;
3205 v->by_ref = aggjf->by_ref;
3206 v->next = res;
3207 res = v;
3209 return res;
3212 /* Determine whether CS also brings all scalar values that the NODE is
3213 specialized for. */
3215 static bool
3216 cgraph_edge_brings_all_scalars_for_node (struct cgraph_edge *cs,
3217 struct cgraph_node *node)
3219 struct ipa_node_params *dest_info = IPA_NODE_REF (node);
3220 int count = ipa_get_param_count (dest_info);
3221 struct ipa_node_params *caller_info;
3222 struct ipa_edge_args *args;
3223 int i;
3225 caller_info = IPA_NODE_REF (cs->caller);
3226 args = IPA_EDGE_REF (cs);
3227 for (i = 0; i < count; i++)
3229 struct ipa_jump_func *jump_func;
3230 tree val, t;
3232 val = dest_info->known_vals[i];
3233 if (!val)
3234 continue;
3236 if (i >= ipa_get_cs_argument_count (args))
3237 return false;
3238 jump_func = ipa_get_ith_jump_func (args, i);
3239 t = ipa_value_from_jfunc (caller_info, jump_func);
3240 if (!t || !values_equal_for_ipcp_p (val, t))
3241 return false;
3243 return true;
3246 /* Determine whether CS also brings all aggregate values that NODE is
3247 specialized for. */
3248 static bool
3249 cgraph_edge_brings_all_agg_vals_for_node (struct cgraph_edge *cs,
3250 struct cgraph_node *node)
3252 struct ipa_node_params *orig_caller_info = IPA_NODE_REF (cs->caller);
3253 struct ipa_node_params *orig_node_info;
3254 struct ipa_agg_replacement_value *aggval;
3255 int i, ec, count;
3257 aggval = ipa_get_agg_replacements_for_node (node);
3258 if (!aggval)
3259 return true;
3261 count = ipa_get_param_count (IPA_NODE_REF (node));
3262 ec = ipa_get_cs_argument_count (IPA_EDGE_REF (cs));
3263 if (ec < count)
3264 for (struct ipa_agg_replacement_value *av = aggval; av; av = av->next)
3265 if (aggval->index >= ec)
3266 return false;
3268 orig_node_info = IPA_NODE_REF (IPA_NODE_REF (node)->ipcp_orig_node);
3269 if (orig_caller_info->ipcp_orig_node)
3270 orig_caller_info = IPA_NODE_REF (orig_caller_info->ipcp_orig_node);
3272 for (i = 0; i < count; i++)
3274 static vec<ipa_agg_jf_item> values = vec<ipa_agg_jf_item>();
3275 struct ipcp_param_lattices *plats;
3276 bool interesting = false;
3277 for (struct ipa_agg_replacement_value *av = aggval; av; av = av->next)
3278 if (aggval->index == i)
3280 interesting = true;
3281 break;
3283 if (!interesting)
3284 continue;
3286 plats = ipa_get_parm_lattices (orig_node_info, aggval->index);
3287 if (plats->aggs_bottom)
3288 return false;
3290 values = intersect_aggregates_with_edge (cs, i, values);
3291 if (!values.exists ())
3292 return false;
3294 for (struct ipa_agg_replacement_value *av = aggval; av; av = av->next)
3295 if (aggval->index == i)
3297 struct ipa_agg_jf_item *item;
3298 int j;
3299 bool found = false;
3300 FOR_EACH_VEC_ELT (values, j, item)
3301 if (item->value
3302 && item->offset == av->offset
3303 && values_equal_for_ipcp_p (item->value, av->value))
3305 found = true;
3306 break;
3308 if (!found)
3310 values.release ();
3311 return false;
3315 return true;
3318 /* Given an original NODE and a VAL for which we have already created a
3319 specialized clone, look whether there are incoming edges that still lead
3320 into the old node but now also bring the requested value and also conform to
3321 all other criteria such that they can be redirected the the special node.
3322 This function can therefore redirect the final edge in a SCC. */
3324 static void
3325 perhaps_add_new_callers (struct cgraph_node *node, struct ipcp_value *val)
3327 struct ipcp_value_source *src;
3328 gcov_type redirected_sum = 0;
3330 for (src = val->sources; src; src = src->next)
3332 struct cgraph_edge *cs = src->cs;
3333 while (cs)
3335 enum availability availability;
3336 struct cgraph_node *dst = cgraph_function_node (cs->callee,
3337 &availability);
3338 if ((dst == node || IPA_NODE_REF (dst)->is_all_contexts_clone)
3339 && availability > AVAIL_OVERWRITABLE
3340 && cgraph_edge_brings_value_p (cs, src))
3342 if (cgraph_edge_brings_all_scalars_for_node (cs, val->spec_node)
3343 && cgraph_edge_brings_all_agg_vals_for_node (cs,
3344 val->spec_node))
3346 if (dump_file)
3347 fprintf (dump_file, " - adding an extra caller %s/%i"
3348 " of %s/%i\n",
3349 xstrdup (cs->caller->name ()),
3350 cs->caller->order,
3351 xstrdup (val->spec_node->name ()),
3352 val->spec_node->order);
3354 cgraph_redirect_edge_callee (cs, val->spec_node);
3355 redirected_sum += cs->count;
3358 cs = get_next_cgraph_edge_clone (cs);
3362 if (redirected_sum)
3363 update_specialized_profile (val->spec_node, node, redirected_sum);
3367 /* Copy KNOWN_BINFOS to KNOWN_VALS. */
3369 static void
3370 move_binfos_to_values (vec<tree> known_vals,
3371 vec<tree> known_binfos)
3373 tree t;
3374 int i;
3376 for (i = 0; known_binfos.iterate (i, &t); i++)
3377 if (t)
3378 known_vals[i] = t;
3381 /* Return true if there is a replacement equivalent to VALUE, INDEX and OFFSET
3382 among those in the AGGVALS list. */
3384 DEBUG_FUNCTION bool
3385 ipcp_val_in_agg_replacements_p (struct ipa_agg_replacement_value *aggvals,
3386 int index, HOST_WIDE_INT offset, tree value)
3388 while (aggvals)
3390 if (aggvals->index == index
3391 && aggvals->offset == offset
3392 && values_equal_for_ipcp_p (aggvals->value, value))
3393 return true;
3394 aggvals = aggvals->next;
3396 return false;
3399 /* Decide wheter to create a special version of NODE for value VAL of parameter
3400 at the given INDEX. If OFFSET is -1, the value is for the parameter itself,
3401 otherwise it is stored at the given OFFSET of the parameter. KNOWN_CSTS,
3402 KNOWN_BINFOS and KNOWN_AGGS describe the other already known values. */
3404 static bool
3405 decide_about_value (struct cgraph_node *node, int index, HOST_WIDE_INT offset,
3406 struct ipcp_value *val, vec<tree> known_csts,
3407 vec<tree> known_binfos)
3409 struct ipa_agg_replacement_value *aggvals;
3410 int freq_sum, caller_count;
3411 gcov_type count_sum;
3412 vec<cgraph_edge_p> callers;
3413 vec<tree> kv;
3415 if (val->spec_node)
3417 perhaps_add_new_callers (node, val);
3418 return false;
3420 else if (val->local_size_cost + overall_size > max_new_size)
3422 if (dump_file && (dump_flags & TDF_DETAILS))
3423 fprintf (dump_file, " Ignoring candidate value because "
3424 "max_new_size would be reached with %li.\n",
3425 val->local_size_cost + overall_size);
3426 return false;
3428 else if (!get_info_about_necessary_edges (val, &freq_sum, &count_sum,
3429 &caller_count))
3430 return false;
3432 if (dump_file && (dump_flags & TDF_DETAILS))
3434 fprintf (dump_file, " - considering value ");
3435 print_ipcp_constant_value (dump_file, val->value);
3436 fprintf (dump_file, " for ");
3437 ipa_dump_param (dump_file, IPA_NODE_REF (node), index);
3438 if (offset != -1)
3439 fprintf (dump_file, ", offset: " HOST_WIDE_INT_PRINT_DEC, offset);
3440 fprintf (dump_file, " (caller_count: %i)\n", caller_count);
3443 if (!good_cloning_opportunity_p (node, val->local_time_benefit,
3444 freq_sum, count_sum,
3445 val->local_size_cost)
3446 && !good_cloning_opportunity_p (node,
3447 val->local_time_benefit
3448 + val->prop_time_benefit,
3449 freq_sum, count_sum,
3450 val->local_size_cost
3451 + val->prop_size_cost))
3452 return false;
3454 if (dump_file)
3455 fprintf (dump_file, " Creating a specialized node of %s/%i.\n",
3456 node->name (), node->order);
3458 callers = gather_edges_for_value (val, caller_count);
3459 kv = known_csts.copy ();
3460 move_binfos_to_values (kv, known_binfos);
3461 if (offset == -1)
3462 kv[index] = val->value;
3463 find_more_scalar_values_for_callers_subset (node, kv, callers);
3464 aggvals = find_aggregate_values_for_callers_subset (node, callers);
3465 gcc_checking_assert (offset == -1
3466 || ipcp_val_in_agg_replacements_p (aggvals, index,
3467 offset, val->value));
3468 val->spec_node = create_specialized_node (node, kv, aggvals, callers);
3469 overall_size += val->local_size_cost;
3471 /* TODO: If for some lattice there is only one other known value
3472 left, make a special node for it too. */
3474 return true;
3477 /* Decide whether and what specialized clones of NODE should be created. */
3479 static bool
3480 decide_whether_version_node (struct cgraph_node *node)
3482 struct ipa_node_params *info = IPA_NODE_REF (node);
3483 int i, count = ipa_get_param_count (info);
3484 vec<tree> known_csts, known_binfos;
3485 vec<ipa_agg_jump_function> known_aggs = vNULL;
3486 bool ret = false;
3488 if (count == 0)
3489 return false;
3491 if (dump_file && (dump_flags & TDF_DETAILS))
3492 fprintf (dump_file, "\nEvaluating opportunities for %s/%i.\n",
3493 node->name (), node->order);
3495 gather_context_independent_values (info, &known_csts, &known_binfos,
3496 info->do_clone_for_all_contexts ? &known_aggs
3497 : NULL, NULL);
3499 for (i = 0; i < count ;i++)
3501 struct ipcp_param_lattices *plats = ipa_get_parm_lattices (info, i);
3502 struct ipcp_lattice *lat = &plats->itself;
3503 struct ipcp_value *val;
3505 if (!lat->bottom
3506 && !known_csts[i]
3507 && !known_binfos[i])
3508 for (val = lat->values; val; val = val->next)
3509 ret |= decide_about_value (node, i, -1, val, known_csts,
3510 known_binfos);
3512 if (!plats->aggs_bottom)
3514 struct ipcp_agg_lattice *aglat;
3515 struct ipcp_value *val;
3516 for (aglat = plats->aggs; aglat; aglat = aglat->next)
3517 if (!aglat->bottom && aglat->values
3518 /* If the following is false, the one value is in
3519 known_aggs. */
3520 && (plats->aggs_contain_variable
3521 || !ipa_lat_is_single_const (aglat)))
3522 for (val = aglat->values; val; val = val->next)
3523 ret |= decide_about_value (node, i, aglat->offset, val,
3524 known_csts, known_binfos);
3526 info = IPA_NODE_REF (node);
3529 if (info->do_clone_for_all_contexts)
3531 struct cgraph_node *clone;
3532 vec<cgraph_edge_p> callers;
3534 if (dump_file)
3535 fprintf (dump_file, " - Creating a specialized node of %s/%i "
3536 "for all known contexts.\n", node->name (),
3537 node->order);
3539 callers = collect_callers_of_node (node);
3540 move_binfos_to_values (known_csts, known_binfos);
3541 clone = create_specialized_node (node, known_csts,
3542 known_aggs_to_agg_replacement_list (known_aggs),
3543 callers);
3544 info = IPA_NODE_REF (node);
3545 info->do_clone_for_all_contexts = false;
3546 IPA_NODE_REF (clone)->is_all_contexts_clone = true;
3547 for (i = 0; i < count ; i++)
3548 vec_free (known_aggs[i].items);
3549 known_aggs.release ();
3550 ret = true;
3552 else
3553 known_csts.release ();
3555 known_binfos.release ();
3556 return ret;
3559 /* Transitively mark all callees of NODE within the same SCC as not dead. */
3561 static void
3562 spread_undeadness (struct cgraph_node *node)
3564 struct cgraph_edge *cs;
3566 for (cs = node->callees; cs; cs = cs->next_callee)
3567 if (ipa_edge_within_scc (cs))
3569 struct cgraph_node *callee;
3570 struct ipa_node_params *info;
3572 callee = cgraph_function_node (cs->callee, NULL);
3573 info = IPA_NODE_REF (callee);
3575 if (info->node_dead)
3577 info->node_dead = 0;
3578 spread_undeadness (callee);
3583 /* Return true if NODE has a caller from outside of its SCC that is not
3584 dead. Worker callback for cgraph_for_node_and_aliases. */
3586 static bool
3587 has_undead_caller_from_outside_scc_p (struct cgraph_node *node,
3588 void *data ATTRIBUTE_UNUSED)
3590 struct cgraph_edge *cs;
3592 for (cs = node->callers; cs; cs = cs->next_caller)
3593 if (cs->caller->thunk.thunk_p
3594 && cgraph_for_node_and_aliases (cs->caller,
3595 has_undead_caller_from_outside_scc_p,
3596 NULL, true))
3597 return true;
3598 else if (!ipa_edge_within_scc (cs)
3599 && !IPA_NODE_REF (cs->caller)->node_dead)
3600 return true;
3601 return false;
3605 /* Identify nodes within the same SCC as NODE which are no longer needed
3606 because of new clones and will be removed as unreachable. */
3608 static void
3609 identify_dead_nodes (struct cgraph_node *node)
3611 struct cgraph_node *v;
3612 for (v = node; v ; v = ((struct ipa_dfs_info *) v->aux)->next_cycle)
3613 if (cgraph_will_be_removed_from_program_if_no_direct_calls (v)
3614 && !cgraph_for_node_and_aliases (v,
3615 has_undead_caller_from_outside_scc_p,
3616 NULL, true))
3617 IPA_NODE_REF (v)->node_dead = 1;
3619 for (v = node; v ; v = ((struct ipa_dfs_info *) v->aux)->next_cycle)
3620 if (!IPA_NODE_REF (v)->node_dead)
3621 spread_undeadness (v);
3623 if (dump_file && (dump_flags & TDF_DETAILS))
3625 for (v = node; v ; v = ((struct ipa_dfs_info *) v->aux)->next_cycle)
3626 if (IPA_NODE_REF (v)->node_dead)
3627 fprintf (dump_file, " Marking node as dead: %s/%i.\n",
3628 v->name (), v->order);
3632 /* The decision stage. Iterate over the topological order of call graph nodes
3633 TOPO and make specialized clones if deemed beneficial. */
3635 static void
3636 ipcp_decision_stage (struct topo_info *topo)
3638 int i;
3640 if (dump_file)
3641 fprintf (dump_file, "\nIPA decision stage:\n\n");
3643 for (i = topo->nnodes - 1; i >= 0; i--)
3645 struct cgraph_node *node = topo->order[i];
3646 bool change = false, iterate = true;
3648 while (iterate)
3650 struct cgraph_node *v;
3651 iterate = false;
3652 for (v = node; v ; v = ((struct ipa_dfs_info *) v->aux)->next_cycle)
3653 if (cgraph_function_with_gimple_body_p (v)
3654 && ipcp_versionable_function_p (v))
3655 iterate |= decide_whether_version_node (v);
3657 change |= iterate;
3659 if (change)
3660 identify_dead_nodes (node);
3664 /* The IPCP driver. */
3666 static unsigned int
3667 ipcp_driver (void)
3669 struct cgraph_2edge_hook_list *edge_duplication_hook_holder;
3670 struct cgraph_edge_hook_list *edge_removal_hook_holder;
3671 struct topo_info topo;
3673 ipa_check_create_node_params ();
3674 ipa_check_create_edge_args ();
3675 grow_edge_clone_vectors ();
3676 edge_duplication_hook_holder =
3677 cgraph_add_edge_duplication_hook (&ipcp_edge_duplication_hook, NULL);
3678 edge_removal_hook_holder =
3679 cgraph_add_edge_removal_hook (&ipcp_edge_removal_hook, NULL);
3681 ipcp_values_pool = create_alloc_pool ("IPA-CP values",
3682 sizeof (struct ipcp_value), 32);
3683 ipcp_sources_pool = create_alloc_pool ("IPA-CP value sources",
3684 sizeof (struct ipcp_value_source), 64);
3685 ipcp_agg_lattice_pool = create_alloc_pool ("IPA_CP aggregate lattices",
3686 sizeof (struct ipcp_agg_lattice),
3687 32);
3688 if (dump_file)
3690 fprintf (dump_file, "\nIPA structures before propagation:\n");
3691 if (dump_flags & TDF_DETAILS)
3692 ipa_print_all_params (dump_file);
3693 ipa_print_all_jump_functions (dump_file);
3696 /* Topological sort. */
3697 build_toporder_info (&topo);
3698 /* Do the interprocedural propagation. */
3699 ipcp_propagate_stage (&topo);
3700 /* Decide what constant propagation and cloning should be performed. */
3701 ipcp_decision_stage (&topo);
3703 /* Free all IPCP structures. */
3704 free_toporder_info (&topo);
3705 next_edge_clone.release ();
3706 cgraph_remove_edge_removal_hook (edge_removal_hook_holder);
3707 cgraph_remove_edge_duplication_hook (edge_duplication_hook_holder);
3708 ipa_free_all_structures_after_ipa_cp ();
3709 if (dump_file)
3710 fprintf (dump_file, "\nIPA constant propagation end\n");
3711 return 0;
3714 /* Initialization and computation of IPCP data structures. This is the initial
3715 intraprocedural analysis of functions, which gathers information to be
3716 propagated later on. */
3718 static void
3719 ipcp_generate_summary (void)
3721 struct cgraph_node *node;
3723 if (dump_file)
3724 fprintf (dump_file, "\nIPA constant propagation start:\n");
3725 ipa_register_cgraph_hooks ();
3727 FOR_EACH_FUNCTION_WITH_GIMPLE_BODY (node)
3729 node->local.versionable
3730 = tree_versionable_function_p (node->decl);
3731 ipa_analyze_node (node);
3735 /* Write ipcp summary for nodes in SET. */
3737 static void
3738 ipcp_write_summary (void)
3740 ipa_prop_write_jump_functions ();
3743 /* Read ipcp summary. */
3745 static void
3746 ipcp_read_summary (void)
3748 ipa_prop_read_jump_functions ();
3751 namespace {
3753 const pass_data pass_data_ipa_cp =
3755 IPA_PASS, /* type */
3756 "cp", /* name */
3757 OPTGROUP_NONE, /* optinfo_flags */
3758 TV_IPA_CONSTANT_PROP, /* tv_id */
3759 0, /* properties_required */
3760 0, /* properties_provided */
3761 0, /* properties_destroyed */
3762 0, /* todo_flags_start */
3763 ( TODO_dump_symtab | TODO_remove_functions ), /* todo_flags_finish */
3766 class pass_ipa_cp : public ipa_opt_pass_d
3768 public:
3769 pass_ipa_cp (gcc::context *ctxt)
3770 : ipa_opt_pass_d (pass_data_ipa_cp, ctxt,
3771 ipcp_generate_summary, /* generate_summary */
3772 ipcp_write_summary, /* write_summary */
3773 ipcp_read_summary, /* read_summary */
3774 ipa_prop_write_all_agg_replacement, /*
3775 write_optimization_summary */
3776 ipa_prop_read_all_agg_replacement, /*
3777 read_optimization_summary */
3778 NULL, /* stmt_fixup */
3779 0, /* function_transform_todo_flags_start */
3780 ipcp_transform_function, /* function_transform */
3781 NULL) /* variable_transform */
3784 /* opt_pass methods: */
3785 virtual bool gate (function *)
3787 /* FIXME: We should remove the optimize check after we ensure we never run
3788 IPA passes when not optimizing. */
3789 return flag_ipa_cp && optimize;
3792 virtual unsigned int execute (function *) { return ipcp_driver (); }
3794 }; // class pass_ipa_cp
3796 } // anon namespace
3798 ipa_opt_pass_d *
3799 make_pass_ipa_cp (gcc::context *ctxt)
3801 return new pass_ipa_cp (ctxt);