ipa-polymorphic-call.c (walk_ssa_copies): Recognize NULL pointer checks.
[gcc.git] / gcc / ipa-cp.c
1 /* Interprocedural constant propagation
2 Copyright (C) 2005-2014 Free Software Foundation, Inc.
3
4 Contributed by Razya Ladelsky <RAZYA@il.ibm.com> and Martin Jambor
5 <mjambor@suse.cz>
6
7 This file is part of GCC.
8
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.
13
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.
18
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/>. */
22
23 /* Interprocedural constant propagation (IPA-CP).
24
25 The goal of this transformation is to
26
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
30
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.
35
36 The algorithm also propagates types and attempts to perform type based
37 devirtualization. Types are propagated much like constants.
38
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.
47
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.
52
53
54 First stage - intraprocedural analysis
55 =======================================
56
57 This phase computes jump_function and modification flags.
58
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:
62
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.
67
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.
70
71 ipcp_generate_summary() is the main function of the first stage.
72
73 Second stage - interprocedural analysis
74 ========================================
75
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.
79
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.
86
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.
92
93 ipcp_propagate_stage() and ipcp_decision_stage() together constitute the
94 third stage.
95
96 Third phase - materialization of clones, call statement updates.
97 ============================================
98
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. */
102
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"
120
121 struct ipcp_value;
122
123 /* Describes a particular source for an IPA-CP value. */
124
125 struct ipcp_value_source
126 {
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;
142 };
143
144 /* Describes one particular value stored in struct ipcp_lattice. */
145
146 struct ipcp_value
147 {
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;
176 };
177
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. */
183
184 struct ipcp_lattice
185 {
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;
197 };
198
199 /* Lattice with an offset to describe a part of an aggregate. */
200
201 struct ipcp_agg_lattice : public ipcp_lattice
202 {
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;
210 };
211
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. */
215
216 struct ipcp_param_lattices
217 {
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;
233
234 /* There is a virtual call based on this parameter. */
235 bool virt_call;
236 };
237
238 /* Allocation pools for values and their sources in ipa-cp. */
239
240 alloc_pool ipcp_values_pool;
241 alloc_pool ipcp_sources_pool;
242 alloc_pool ipcp_agg_lattice_pool;
243
244 /* Maximal count found in program. */
245
246 static gcov_type max_count;
247
248 /* Original overall size of the program. */
249
250 static long overall_size, max_new_size;
251
252 /* Head of the linked list of topologically sorted values. */
253
254 static struct ipcp_value *values_topo;
255
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)
260 {
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]);
265 }
266
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)
271 {
272 struct ipcp_param_lattices *plats = ipa_get_parm_lattices (info, i);
273 return &plats->itself;
274 }
275
276 /* Return whether LAT is a lattice with a single constant and without an
277 undefined value. */
278
279 static inline bool
280 ipa_lat_is_single_const (struct ipcp_lattice *lat)
281 {
282 if (lat->bottom
283 || lat->contains_variable
284 || lat->values_count != 1)
285 return false;
286 else
287 return true;
288 }
289
290 /* Print V which is extracted from a value in a lattice to F. */
291
292 static void
293 print_ipcp_constant_value (FILE * f, tree v)
294 {
295 if (TREE_CODE (v) == TREE_BINFO)
296 {
297 fprintf (f, "BINFO ");
298 print_generic_expr (f, BINFO_TYPE (v), 0);
299 }
300 else if (TREE_CODE (v) == ADDR_EXPR
301 && TREE_CODE (TREE_OPERAND (v, 0)) == CONST_DECL)
302 {
303 fprintf (f, "& ");
304 print_generic_expr (f, DECL_INITIAL (TREE_OPERAND (v, 0)), 0);
305 }
306 else
307 print_generic_expr (f, v, 0);
308 }
309
310 /* Print a lattice LAT to F. */
311
312 static void
313 print_lattice (FILE * f, struct ipcp_lattice *lat,
314 bool dump_sources, bool dump_benefits)
315 {
316 struct ipcp_value *val;
317 bool prev = false;
318
319 if (lat->bottom)
320 {
321 fprintf (f, "BOTTOM\n");
322 return;
323 }
324
325 if (!lat->values_count && !lat->contains_variable)
326 {
327 fprintf (f, "TOP\n");
328 return;
329 }
330
331 if (lat->contains_variable)
332 {
333 fprintf (f, "VARIABLE");
334 prev = true;
335 if (dump_benefits)
336 fprintf (f, "\n");
337 }
338
339 for (val = lat->values; val; val = val->next)
340 {
341 if (dump_benefits && prev)
342 fprintf (f, " ");
343 else if (!dump_benefits && prev)
344 fprintf (f, ", ");
345 else
346 prev = true;
347
348 print_ipcp_constant_value (f, val->value);
349
350 if (dump_sources)
351 {
352 struct ipcp_value_source *s;
353
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, "]");
359 }
360
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);
366 }
367 if (!dump_benefits)
368 fprintf (f, "\n");
369 }
370
371 /* Print all ipcp_lattices of all functions to F. */
372
373 static void
374 print_all_lattices (FILE * f, bool dump_sources, bool dump_benefits)
375 {
376 struct cgraph_node *node;
377 int i, count;
378
379 fprintf (f, "\nLattices:\n");
380 FOR_EACH_FUNCTION_WITH_GIMPLE_BODY (node)
381 {
382 struct ipa_node_params *info;
383
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++)
389 {
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);
394
395 if (plats->virt_call)
396 fprintf (f, " virt_call flag set\n");
397
398 if (plats->aggs_bottom)
399 {
400 fprintf (f, " AGGS BOTTOM\n");
401 continue;
402 }
403 if (plats->aggs_contain_variable)
404 fprintf (f, " AGGS VARIABLE\n");
405 for (aglat = plats->aggs; aglat; aglat = aglat->next)
406 {
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);
410 }
411 }
412 }
413 }
414
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. */
418
419 static void
420 determine_versionability (struct cgraph_node *node)
421 {
422 const char *reason = NULL;
423
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 (node->get_availability () <= AVAIL_INTERPOSABLE)
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)))
437 {
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";
442 }
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 (node->comdat_local_p ())
446 reason = "comdat-local function";
447
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);
451
452 node->local.versionable = (reason == NULL);
453 }
454
455 /* Return true if it is at all technically possible to create clones of a
456 NODE. */
457
458 static bool
459 ipcp_versionable_function_p (struct cgraph_node *node)
460 {
461 return node->local.versionable;
462 }
463
464 /* Structure holding accumulated information about callers of a node. */
465
466 struct caller_statistics
467 {
468 gcov_type count_sum;
469 int n_calls, n_hot_calls, freq_sum;
470 };
471
472 /* Initialize fields of STAT to zeroes. */
473
474 static inline void
475 init_caller_stats (struct caller_statistics *stats)
476 {
477 stats->count_sum = 0;
478 stats->n_calls = 0;
479 stats->n_hot_calls = 0;
480 stats->freq_sum = 0;
481 }
482
483 /* Worker callback of cgraph_for_node_and_aliases accumulating statistics of
484 non-thunk incoming edges to NODE. */
485
486 static bool
487 gather_caller_stats (struct cgraph_node *node, void *data)
488 {
489 struct caller_statistics *stats = (struct caller_statistics *) data;
490 struct cgraph_edge *cs;
491
492 for (cs = node->callers; cs; cs = cs->next_caller)
493 if (cs->caller->thunk.thunk_p)
494 cs->caller->call_for_symbol_thunks_and_aliases (gather_caller_stats,
495 stats, false);
496 else
497 {
498 stats->count_sum += cs->count;
499 stats->freq_sum += cs->frequency;
500 stats->n_calls++;
501 if (cs->maybe_hot_p ())
502 stats->n_hot_calls ++;
503 }
504 return false;
505
506 }
507
508 /* Return true if this NODE is viable candidate for cloning. */
509
510 static bool
511 ipcp_cloning_candidate_p (struct cgraph_node *node)
512 {
513 struct caller_statistics stats;
514
515 gcc_checking_assert (node->has_gimple_body_p ());
516
517 if (!flag_ipa_cp_clone)
518 {
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;
524 }
525
526 if (!optimize_function_for_speed_p (DECL_STRUCT_FUNCTION (node->decl)))
527 {
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;
533 }
534
535 init_caller_stats (&stats);
536 node->call_for_symbol_thunks_and_aliases (gather_caller_stats, &stats, false);
537
538 if (inline_summary (node)->self_size < stats.n_calls)
539 {
540 if (dump_file)
541 fprintf (dump_file, "Considering %s for cloning; code might shrink.\n",
542 node->name ());
543 return true;
544 }
545
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)
550 {
551 if (stats.count_sum > node->count * 90 / 100)
552 {
553 if (dump_file)
554 fprintf (dump_file, "Considering %s for cloning; "
555 "usually called directly.\n",
556 node->name ());
557 return true;
558 }
559 }
560 if (!stats.n_hot_calls)
561 {
562 if (dump_file)
563 fprintf (dump_file, "Not considering %s for cloning; no hot calls.\n",
564 node->name ());
565 return false;
566 }
567 if (dump_file)
568 fprintf (dump_file, "Considering %s for cloning.\n",
569 node->name ());
570 return true;
571 }
572
573 /* Arrays representing a topological ordering of call graph nodes and a stack
574 of noes used during constant propagation. */
575
576 struct ipa_topo_info
577 {
578 struct cgraph_node **order;
579 struct cgraph_node **stack;
580 int nnodes, stack_top;
581 };
582
583 /* Allocate the arrays in TOPO and topologically sort the nodes into order. */
584
585 static void
586 build_toporder_info (struct ipa_topo_info *topo)
587 {
588 topo->order = XCNEWVEC (struct cgraph_node *, symtab->cgraph_count);
589 topo->stack = XCNEWVEC (struct cgraph_node *, symtab->cgraph_count);
590
591 topo->stack_top = 0;
592 topo->nnodes = ipa_reduced_postorder (topo->order, true, true, NULL);
593 }
594
595 /* Free information about strongly connected components and the arrays in
596 TOPO. */
597
598 static void
599 free_toporder_info (struct ipa_topo_info *topo)
600 {
601 ipa_free_postorder_info ();
602 free (topo->order);
603 free (topo->stack);
604 }
605
606 /* Add NODE to the stack in TOPO, unless it is already there. */
607
608 static inline void
609 push_node_to_stack (struct ipa_topo_info *topo, struct cgraph_node *node)
610 {
611 struct ipa_node_params *info = IPA_NODE_REF (node);
612 if (info->node_enqueued)
613 return;
614 info->node_enqueued = 1;
615 topo->stack[topo->stack_top++] = node;
616 }
617
618 /* Pop a node from the stack in TOPO and return it or return NULL if the stack
619 is empty. */
620
621 static struct cgraph_node *
622 pop_node_from_stack (struct ipa_topo_info *topo)
623 {
624 if (topo->stack_top)
625 {
626 struct cgraph_node *node;
627 topo->stack_top--;
628 node = topo->stack[topo->stack_top];
629 IPA_NODE_REF (node)->node_enqueued = 0;
630 return node;
631 }
632 else
633 return NULL;
634 }
635
636 /* Set lattice LAT to bottom and return true if it previously was not set as
637 such. */
638
639 static inline bool
640 set_lattice_to_bottom (struct ipcp_lattice *lat)
641 {
642 bool ret = !lat->bottom;
643 lat->bottom = true;
644 return ret;
645 }
646
647 /* Mark lattice as containing an unknown value and return true if it previously
648 was not marked as such. */
649
650 static inline bool
651 set_lattice_contains_variable (struct ipcp_lattice *lat)
652 {
653 bool ret = !lat->contains_variable;
654 lat->contains_variable = true;
655 return ret;
656 }
657
658 /* Set all aggegate lattices in PLATS to bottom and return true if they were
659 not previously set as such. */
660
661 static inline bool
662 set_agg_lats_to_bottom (struct ipcp_param_lattices *plats)
663 {
664 bool ret = !plats->aggs_bottom;
665 plats->aggs_bottom = true;
666 return ret;
667 }
668
669 /* Mark all aggegate lattices in PLATS as containing an unknown value and
670 return true if they were not previously marked as such. */
671
672 static inline bool
673 set_agg_lats_contain_variable (struct ipcp_param_lattices *plats)
674 {
675 bool ret = !plats->aggs_contain_variable;
676 plats->aggs_contain_variable = true;
677 return ret;
678 }
679
680 /* Mark bot aggregate and scalar lattices as containing an unknown variable,
681 return true is any of them has not been marked as such so far. */
682
683 static inline bool
684 set_all_contains_variable (struct ipcp_param_lattices *plats)
685 {
686 bool ret = !plats->itself.contains_variable || !plats->aggs_contain_variable;
687 plats->itself.contains_variable = true;
688 plats->aggs_contain_variable = true;
689 return ret;
690 }
691
692 /* Initialize ipcp_lattices. */
693
694 static void
695 initialize_node_lattices (struct cgraph_node *node)
696 {
697 struct ipa_node_params *info = IPA_NODE_REF (node);
698 struct cgraph_edge *ie;
699 bool disable = false, variable = false;
700 int i;
701
702 gcc_checking_assert (node->has_gimple_body_p ());
703 if (!node->local.local)
704 {
705 /* When cloning is allowed, we can assume that externally visible
706 functions are not called. We will compensate this by cloning
707 later. */
708 if (ipcp_versionable_function_p (node)
709 && ipcp_cloning_candidate_p (node))
710 variable = true;
711 else
712 disable = true;
713 }
714
715 if (disable || variable)
716 {
717 for (i = 0; i < ipa_get_param_count (info) ; i++)
718 {
719 struct ipcp_param_lattices *plats = ipa_get_parm_lattices (info, i);
720 if (disable)
721 {
722 set_lattice_to_bottom (&plats->itself);
723 set_agg_lats_to_bottom (plats);
724 }
725 else
726 set_all_contains_variable (plats);
727 }
728 if (dump_file && (dump_flags & TDF_DETAILS)
729 && !node->alias && !node->thunk.thunk_p)
730 fprintf (dump_file, "Marking all lattices of %s/%i as %s\n",
731 node->name (), node->order,
732 disable ? "BOTTOM" : "VARIABLE");
733 }
734
735 for (ie = node->indirect_calls; ie; ie = ie->next_callee)
736 if (ie->indirect_info->polymorphic
737 && ie->indirect_info->param_index >= 0)
738 {
739 gcc_checking_assert (ie->indirect_info->param_index >= 0);
740 ipa_get_parm_lattices (info,
741 ie->indirect_info->param_index)->virt_call = 1;
742 }
743 }
744
745 /* Return the result of a (possibly arithmetic) pass through jump function
746 JFUNC on the constant value INPUT. Return NULL_TREE if that cannot be
747 determined or be considered an interprocedural invariant. */
748
749 static tree
750 ipa_get_jf_pass_through_result (struct ipa_jump_func *jfunc, tree input)
751 {
752 tree restype, res;
753
754 if (TREE_CODE (input) == TREE_BINFO)
755 {
756 if (ipa_get_jf_pass_through_type_preserved (jfunc))
757 {
758 gcc_checking_assert (ipa_get_jf_pass_through_operation (jfunc)
759 == NOP_EXPR);
760 return input;
761 }
762 return NULL_TREE;
763 }
764
765 if (ipa_get_jf_pass_through_operation (jfunc) == NOP_EXPR)
766 return input;
767
768 gcc_checking_assert (is_gimple_ip_invariant (input));
769 if (TREE_CODE_CLASS (ipa_get_jf_pass_through_operation (jfunc))
770 == tcc_comparison)
771 restype = boolean_type_node;
772 else
773 restype = TREE_TYPE (input);
774 res = fold_binary (ipa_get_jf_pass_through_operation (jfunc), restype,
775 input, ipa_get_jf_pass_through_operand (jfunc));
776
777 if (res && !is_gimple_ip_invariant (res))
778 return NULL_TREE;
779
780 return res;
781 }
782
783 /* Return the result of an ancestor jump function JFUNC on the constant value
784 INPUT. Return NULL_TREE if that cannot be determined. */
785
786 static tree
787 ipa_get_jf_ancestor_result (struct ipa_jump_func *jfunc, tree input)
788 {
789 if (TREE_CODE (input) == TREE_BINFO)
790 {
791 if (!ipa_get_jf_ancestor_type_preserved (jfunc))
792 return NULL;
793 /* FIXME: At LTO we can't propagate to non-polymorphic type, because
794 we have no ODR equivalency on those. This should be fixed by
795 propagating on types rather than binfos that would make type
796 matching here unnecesary. */
797 if (in_lto_p
798 && (TREE_CODE (ipa_get_jf_ancestor_type (jfunc)) != RECORD_TYPE
799 || !TYPE_BINFO (ipa_get_jf_ancestor_type (jfunc))
800 || !BINFO_VTABLE (TYPE_BINFO (ipa_get_jf_ancestor_type (jfunc)))))
801 {
802 if (!ipa_get_jf_ancestor_offset (jfunc))
803 return input;
804 return NULL;
805 }
806 return get_binfo_at_offset (input,
807 ipa_get_jf_ancestor_offset (jfunc),
808 ipa_get_jf_ancestor_type (jfunc));
809 }
810 else if (TREE_CODE (input) == ADDR_EXPR)
811 {
812 tree t = TREE_OPERAND (input, 0);
813 t = build_ref_for_offset (EXPR_LOCATION (t), t,
814 ipa_get_jf_ancestor_offset (jfunc),
815 ipa_get_jf_ancestor_type (jfunc)
816 ? ipa_get_jf_ancestor_type (jfunc)
817 : ptr_type_node, NULL, false);
818 return build_fold_addr_expr (t);
819 }
820 else
821 return NULL_TREE;
822 }
823
824 /* Determine whether JFUNC evaluates to a known value (that is either a
825 constant or a binfo) and if so, return it. Otherwise return NULL. INFO
826 describes the caller node so that pass-through jump functions can be
827 evaluated. */
828
829 tree
830 ipa_value_from_jfunc (struct ipa_node_params *info, struct ipa_jump_func *jfunc)
831 {
832 if (jfunc->type == IPA_JF_CONST)
833 return ipa_get_jf_constant (jfunc);
834 else if (jfunc->type == IPA_JF_KNOWN_TYPE)
835 return ipa_binfo_from_known_type_jfunc (jfunc);
836 else if (jfunc->type == IPA_JF_PASS_THROUGH
837 || jfunc->type == IPA_JF_ANCESTOR)
838 {
839 tree input;
840 int idx;
841
842 if (jfunc->type == IPA_JF_PASS_THROUGH)
843 idx = ipa_get_jf_pass_through_formal_id (jfunc);
844 else
845 idx = ipa_get_jf_ancestor_formal_id (jfunc);
846
847 if (info->ipcp_orig_node)
848 input = info->known_vals[idx];
849 else
850 {
851 struct ipcp_lattice *lat;
852
853 if (!info->lattices)
854 {
855 gcc_checking_assert (!flag_ipa_cp);
856 return NULL_TREE;
857 }
858 lat = ipa_get_scalar_lat (info, idx);
859 if (!ipa_lat_is_single_const (lat))
860 return NULL_TREE;
861 input = lat->values->value;
862 }
863
864 if (!input)
865 return NULL_TREE;
866
867 if (jfunc->type == IPA_JF_PASS_THROUGH)
868 return ipa_get_jf_pass_through_result (jfunc, input);
869 else
870 return ipa_get_jf_ancestor_result (jfunc, input);
871 }
872 else
873 return NULL_TREE;
874 }
875
876
877 /* If checking is enabled, verify that no lattice is in the TOP state, i.e. not
878 bottom, not containing a variable component and without any known value at
879 the same time. */
880
881 DEBUG_FUNCTION void
882 ipcp_verify_propagated_values (void)
883 {
884 struct cgraph_node *node;
885
886 FOR_EACH_FUNCTION_WITH_GIMPLE_BODY (node)
887 {
888 struct ipa_node_params *info = IPA_NODE_REF (node);
889 int i, count = ipa_get_param_count (info);
890
891 for (i = 0; i < count; i++)
892 {
893 struct ipcp_lattice *lat = ipa_get_scalar_lat (info, i);
894
895 if (!lat->bottom
896 && !lat->contains_variable
897 && lat->values_count == 0)
898 {
899 if (dump_file)
900 {
901 symtab_node::dump_table (dump_file);
902 fprintf (dump_file, "\nIPA lattices after constant "
903 "propagation, before gcc_unreachable:\n");
904 print_all_lattices (dump_file, true, false);
905 }
906
907 gcc_unreachable ();
908 }
909 }
910 }
911 }
912
913 /* Return true iff X and Y should be considered equal values by IPA-CP. */
914
915 static bool
916 values_equal_for_ipcp_p (tree x, tree y)
917 {
918 gcc_checking_assert (x != NULL_TREE && y != NULL_TREE);
919
920 if (x == y)
921 return true;
922
923 if (TREE_CODE (x) == TREE_BINFO || TREE_CODE (y) == TREE_BINFO)
924 return false;
925
926 if (TREE_CODE (x) == ADDR_EXPR
927 && TREE_CODE (y) == ADDR_EXPR
928 && TREE_CODE (TREE_OPERAND (x, 0)) == CONST_DECL
929 && TREE_CODE (TREE_OPERAND (y, 0)) == CONST_DECL)
930 return operand_equal_p (DECL_INITIAL (TREE_OPERAND (x, 0)),
931 DECL_INITIAL (TREE_OPERAND (y, 0)), 0);
932 else
933 return operand_equal_p (x, y, 0);
934 }
935
936 /* Add a new value source to VAL, marking that a value comes from edge CS and
937 (if the underlying jump function is a pass-through or an ancestor one) from
938 a caller value SRC_VAL of a caller parameter described by SRC_INDEX. OFFSET
939 is negative if the source was the scalar value of the parameter itself or
940 the offset within an aggregate. */
941
942 static void
943 add_value_source (struct ipcp_value *val, struct cgraph_edge *cs,
944 struct ipcp_value *src_val, int src_idx, HOST_WIDE_INT offset)
945 {
946 struct ipcp_value_source *src;
947
948 src = (struct ipcp_value_source *) pool_alloc (ipcp_sources_pool);
949 src->offset = offset;
950 src->cs = cs;
951 src->val = src_val;
952 src->index = src_idx;
953
954 src->next = val->sources;
955 val->sources = src;
956 }
957
958 /* Try to add NEWVAL to LAT, potentially creating a new struct ipcp_value for
959 it. CS, SRC_VAL SRC_INDEX and OFFSET are meant for add_value_source and
960 have the same meaning. */
961
962 static bool
963 add_value_to_lattice (struct ipcp_lattice *lat, tree newval,
964 struct cgraph_edge *cs, struct ipcp_value *src_val,
965 int src_idx, HOST_WIDE_INT offset)
966 {
967 struct ipcp_value *val;
968
969 if (lat->bottom)
970 return false;
971
972 for (val = lat->values; val; val = val->next)
973 if (values_equal_for_ipcp_p (val->value, newval))
974 {
975 if (ipa_edge_within_scc (cs))
976 {
977 struct ipcp_value_source *s;
978 for (s = val->sources; s ; s = s->next)
979 if (s->cs == cs)
980 break;
981 if (s)
982 return false;
983 }
984
985 add_value_source (val, cs, src_val, src_idx, offset);
986 return false;
987 }
988
989 if (lat->values_count == PARAM_VALUE (PARAM_IPA_CP_VALUE_LIST_SIZE))
990 {
991 /* We can only free sources, not the values themselves, because sources
992 of other values in this this SCC might point to them. */
993 for (val = lat->values; val; val = val->next)
994 {
995 while (val->sources)
996 {
997 struct ipcp_value_source *src = val->sources;
998 val->sources = src->next;
999 pool_free (ipcp_sources_pool, src);
1000 }
1001 }
1002
1003 lat->values = NULL;
1004 return set_lattice_to_bottom (lat);
1005 }
1006
1007 lat->values_count++;
1008 val = (struct ipcp_value *) pool_alloc (ipcp_values_pool);
1009 memset (val, 0, sizeof (*val));
1010
1011 add_value_source (val, cs, src_val, src_idx, offset);
1012 val->value = newval;
1013 val->next = lat->values;
1014 lat->values = val;
1015 return true;
1016 }
1017
1018 /* Like above but passes a special value of offset to distinguish that the
1019 origin is the scalar value of the parameter rather than a part of an
1020 aggregate. */
1021
1022 static inline bool
1023 add_scalar_value_to_lattice (struct ipcp_lattice *lat, tree newval,
1024 struct cgraph_edge *cs,
1025 struct ipcp_value *src_val, int src_idx)
1026 {
1027 return add_value_to_lattice (lat, newval, cs, src_val, src_idx, -1);
1028 }
1029
1030 /* Propagate values through a pass-through jump function JFUNC associated with
1031 edge CS, taking values from SRC_LAT and putting them into DEST_LAT. SRC_IDX
1032 is the index of the source parameter. */
1033
1034 static bool
1035 propagate_vals_accross_pass_through (struct cgraph_edge *cs,
1036 struct ipa_jump_func *jfunc,
1037 struct ipcp_lattice *src_lat,
1038 struct ipcp_lattice *dest_lat,
1039 int src_idx)
1040 {
1041 struct ipcp_value *src_val;
1042 bool ret = false;
1043
1044 /* Do not create new values when propagating within an SCC because if there
1045 are arithmetic functions with circular dependencies, there is infinite
1046 number of them and we would just make lattices bottom. */
1047 if ((ipa_get_jf_pass_through_operation (jfunc) != NOP_EXPR)
1048 && ipa_edge_within_scc (cs))
1049 ret = set_lattice_contains_variable (dest_lat);
1050 else
1051 for (src_val = src_lat->values; src_val; src_val = src_val->next)
1052 {
1053 tree cstval = ipa_get_jf_pass_through_result (jfunc, src_val->value);
1054
1055 if (cstval)
1056 ret |= add_scalar_value_to_lattice (dest_lat, cstval, cs, src_val,
1057 src_idx);
1058 else
1059 ret |= set_lattice_contains_variable (dest_lat);
1060 }
1061
1062 return ret;
1063 }
1064
1065 /* Propagate values through an ancestor jump function JFUNC associated with
1066 edge CS, taking values from SRC_LAT and putting them into DEST_LAT. SRC_IDX
1067 is the index of the source parameter. */
1068
1069 static bool
1070 propagate_vals_accross_ancestor (struct cgraph_edge *cs,
1071 struct ipa_jump_func *jfunc,
1072 struct ipcp_lattice *src_lat,
1073 struct ipcp_lattice *dest_lat,
1074 int src_idx)
1075 {
1076 struct ipcp_value *src_val;
1077 bool ret = false;
1078
1079 if (ipa_edge_within_scc (cs))
1080 return set_lattice_contains_variable (dest_lat);
1081
1082 for (src_val = src_lat->values; src_val; src_val = src_val->next)
1083 {
1084 tree t = ipa_get_jf_ancestor_result (jfunc, src_val->value);
1085
1086 if (t)
1087 ret |= add_scalar_value_to_lattice (dest_lat, t, cs, src_val, src_idx);
1088 else
1089 ret |= set_lattice_contains_variable (dest_lat);
1090 }
1091
1092 return ret;
1093 }
1094
1095 /* Propagate scalar values across jump function JFUNC that is associated with
1096 edge CS and put the values into DEST_LAT. */
1097
1098 static bool
1099 propagate_scalar_accross_jump_function (struct cgraph_edge *cs,
1100 struct ipa_jump_func *jfunc,
1101 struct ipcp_lattice *dest_lat)
1102 {
1103 if (dest_lat->bottom)
1104 return false;
1105
1106 if (jfunc->type == IPA_JF_CONST
1107 || jfunc->type == IPA_JF_KNOWN_TYPE)
1108 {
1109 tree val;
1110
1111 if (jfunc->type == IPA_JF_KNOWN_TYPE)
1112 {
1113 val = ipa_binfo_from_known_type_jfunc (jfunc);
1114 if (!val)
1115 return set_lattice_contains_variable (dest_lat);
1116 }
1117 else
1118 val = ipa_get_jf_constant (jfunc);
1119 return add_scalar_value_to_lattice (dest_lat, val, cs, NULL, 0);
1120 }
1121 else if (jfunc->type == IPA_JF_PASS_THROUGH
1122 || jfunc->type == IPA_JF_ANCESTOR)
1123 {
1124 struct ipa_node_params *caller_info = IPA_NODE_REF (cs->caller);
1125 struct ipcp_lattice *src_lat;
1126 int src_idx;
1127 bool ret;
1128
1129 if (jfunc->type == IPA_JF_PASS_THROUGH)
1130 src_idx = ipa_get_jf_pass_through_formal_id (jfunc);
1131 else
1132 src_idx = ipa_get_jf_ancestor_formal_id (jfunc);
1133
1134 src_lat = ipa_get_scalar_lat (caller_info, src_idx);
1135 if (src_lat->bottom)
1136 return set_lattice_contains_variable (dest_lat);
1137
1138 /* If we would need to clone the caller and cannot, do not propagate. */
1139 if (!ipcp_versionable_function_p (cs->caller)
1140 && (src_lat->contains_variable
1141 || (src_lat->values_count > 1)))
1142 return set_lattice_contains_variable (dest_lat);
1143
1144 if (jfunc->type == IPA_JF_PASS_THROUGH)
1145 ret = propagate_vals_accross_pass_through (cs, jfunc, src_lat,
1146 dest_lat, src_idx);
1147 else
1148 ret = propagate_vals_accross_ancestor (cs, jfunc, src_lat, dest_lat,
1149 src_idx);
1150
1151 if (src_lat->contains_variable)
1152 ret |= set_lattice_contains_variable (dest_lat);
1153
1154 return ret;
1155 }
1156
1157 /* TODO: We currently do not handle member method pointers in IPA-CP (we only
1158 use it for indirect inlining), we should propagate them too. */
1159 return set_lattice_contains_variable (dest_lat);
1160 }
1161
1162 /* If DEST_PLATS already has aggregate items, check that aggs_by_ref matches
1163 NEW_AGGS_BY_REF and if not, mark all aggs as bottoms and return true (in all
1164 other cases, return false). If there are no aggregate items, set
1165 aggs_by_ref to NEW_AGGS_BY_REF. */
1166
1167 static bool
1168 set_check_aggs_by_ref (struct ipcp_param_lattices *dest_plats,
1169 bool new_aggs_by_ref)
1170 {
1171 if (dest_plats->aggs)
1172 {
1173 if (dest_plats->aggs_by_ref != new_aggs_by_ref)
1174 {
1175 set_agg_lats_to_bottom (dest_plats);
1176 return true;
1177 }
1178 }
1179 else
1180 dest_plats->aggs_by_ref = new_aggs_by_ref;
1181 return false;
1182 }
1183
1184 /* Walk aggregate lattices in DEST_PLATS from ***AGLAT on, until ***aglat is an
1185 already existing lattice for the given OFFSET and SIZE, marking all skipped
1186 lattices as containing variable and checking for overlaps. If there is no
1187 already existing lattice for the OFFSET and VAL_SIZE, create one, initialize
1188 it with offset, size and contains_variable to PRE_EXISTING, and return true,
1189 unless there are too many already. If there are two many, return false. If
1190 there are overlaps turn whole DEST_PLATS to bottom and return false. If any
1191 skipped lattices were newly marked as containing variable, set *CHANGE to
1192 true. */
1193
1194 static bool
1195 merge_agg_lats_step (struct ipcp_param_lattices *dest_plats,
1196 HOST_WIDE_INT offset, HOST_WIDE_INT val_size,
1197 struct ipcp_agg_lattice ***aglat,
1198 bool pre_existing, bool *change)
1199 {
1200 gcc_checking_assert (offset >= 0);
1201
1202 while (**aglat && (**aglat)->offset < offset)
1203 {
1204 if ((**aglat)->offset + (**aglat)->size > offset)
1205 {
1206 set_agg_lats_to_bottom (dest_plats);
1207 return false;
1208 }
1209 *change |= set_lattice_contains_variable (**aglat);
1210 *aglat = &(**aglat)->next;
1211 }
1212
1213 if (**aglat && (**aglat)->offset == offset)
1214 {
1215 if ((**aglat)->size != val_size
1216 || ((**aglat)->next
1217 && (**aglat)->next->offset < offset + val_size))
1218 {
1219 set_agg_lats_to_bottom (dest_plats);
1220 return false;
1221 }
1222 gcc_checking_assert (!(**aglat)->next
1223 || (**aglat)->next->offset >= offset + val_size);
1224 return true;
1225 }
1226 else
1227 {
1228 struct ipcp_agg_lattice *new_al;
1229
1230 if (**aglat && (**aglat)->offset < offset + val_size)
1231 {
1232 set_agg_lats_to_bottom (dest_plats);
1233 return false;
1234 }
1235 if (dest_plats->aggs_count == PARAM_VALUE (PARAM_IPA_MAX_AGG_ITEMS))
1236 return false;
1237 dest_plats->aggs_count++;
1238 new_al = (struct ipcp_agg_lattice *) pool_alloc (ipcp_agg_lattice_pool);
1239 memset (new_al, 0, sizeof (*new_al));
1240
1241 new_al->offset = offset;
1242 new_al->size = val_size;
1243 new_al->contains_variable = pre_existing;
1244
1245 new_al->next = **aglat;
1246 **aglat = new_al;
1247 return true;
1248 }
1249 }
1250
1251 /* Set all AGLAT and all other aggregate lattices reachable by next pointers as
1252 containing an unknown value. */
1253
1254 static bool
1255 set_chain_of_aglats_contains_variable (struct ipcp_agg_lattice *aglat)
1256 {
1257 bool ret = false;
1258 while (aglat)
1259 {
1260 ret |= set_lattice_contains_variable (aglat);
1261 aglat = aglat->next;
1262 }
1263 return ret;
1264 }
1265
1266 /* Merge existing aggregate lattices in SRC_PLATS to DEST_PLATS, subtracting
1267 DELTA_OFFSET. CS is the call graph edge and SRC_IDX the index of the source
1268 parameter used for lattice value sources. Return true if DEST_PLATS changed
1269 in any way. */
1270
1271 static bool
1272 merge_aggregate_lattices (struct cgraph_edge *cs,
1273 struct ipcp_param_lattices *dest_plats,
1274 struct ipcp_param_lattices *src_plats,
1275 int src_idx, HOST_WIDE_INT offset_delta)
1276 {
1277 bool pre_existing = dest_plats->aggs != NULL;
1278 struct ipcp_agg_lattice **dst_aglat;
1279 bool ret = false;
1280
1281 if (set_check_aggs_by_ref (dest_plats, src_plats->aggs_by_ref))
1282 return true;
1283 if (src_plats->aggs_bottom)
1284 return set_agg_lats_contain_variable (dest_plats);
1285 if (src_plats->aggs_contain_variable)
1286 ret |= set_agg_lats_contain_variable (dest_plats);
1287 dst_aglat = &dest_plats->aggs;
1288
1289 for (struct ipcp_agg_lattice *src_aglat = src_plats->aggs;
1290 src_aglat;
1291 src_aglat = src_aglat->next)
1292 {
1293 HOST_WIDE_INT new_offset = src_aglat->offset - offset_delta;
1294
1295 if (new_offset < 0)
1296 continue;
1297 if (merge_agg_lats_step (dest_plats, new_offset, src_aglat->size,
1298 &dst_aglat, pre_existing, &ret))
1299 {
1300 struct ipcp_agg_lattice *new_al = *dst_aglat;
1301
1302 dst_aglat = &(*dst_aglat)->next;
1303 if (src_aglat->bottom)
1304 {
1305 ret |= set_lattice_contains_variable (new_al);
1306 continue;
1307 }
1308 if (src_aglat->contains_variable)
1309 ret |= set_lattice_contains_variable (new_al);
1310 for (struct ipcp_value *val = src_aglat->values;
1311 val;
1312 val = val->next)
1313 ret |= add_value_to_lattice (new_al, val->value, cs, val, src_idx,
1314 src_aglat->offset);
1315 }
1316 else if (dest_plats->aggs_bottom)
1317 return true;
1318 }
1319 ret |= set_chain_of_aglats_contains_variable (*dst_aglat);
1320 return ret;
1321 }
1322
1323 /* Determine whether there is anything to propagate FROM SRC_PLATS through a
1324 pass-through JFUNC and if so, whether it has conform and conforms to the
1325 rules about propagating values passed by reference. */
1326
1327 static bool
1328 agg_pass_through_permissible_p (struct ipcp_param_lattices *src_plats,
1329 struct ipa_jump_func *jfunc)
1330 {
1331 return src_plats->aggs
1332 && (!src_plats->aggs_by_ref
1333 || ipa_get_jf_pass_through_agg_preserved (jfunc));
1334 }
1335
1336 /* Propagate scalar values across jump function JFUNC that is associated with
1337 edge CS and put the values into DEST_LAT. */
1338
1339 static bool
1340 propagate_aggs_accross_jump_function (struct cgraph_edge *cs,
1341 struct ipa_jump_func *jfunc,
1342 struct ipcp_param_lattices *dest_plats)
1343 {
1344 bool ret = false;
1345
1346 if (dest_plats->aggs_bottom)
1347 return false;
1348
1349 if (jfunc->type == IPA_JF_PASS_THROUGH
1350 && ipa_get_jf_pass_through_operation (jfunc) == NOP_EXPR)
1351 {
1352 struct ipa_node_params *caller_info = IPA_NODE_REF (cs->caller);
1353 int src_idx = ipa_get_jf_pass_through_formal_id (jfunc);
1354 struct ipcp_param_lattices *src_plats;
1355
1356 src_plats = ipa_get_parm_lattices (caller_info, src_idx);
1357 if (agg_pass_through_permissible_p (src_plats, jfunc))
1358 {
1359 /* Currently we do not produce clobber aggregate jump
1360 functions, replace with merging when we do. */
1361 gcc_assert (!jfunc->agg.items);
1362 ret |= merge_aggregate_lattices (cs, dest_plats, src_plats,
1363 src_idx, 0);
1364 }
1365 else
1366 ret |= set_agg_lats_contain_variable (dest_plats);
1367 }
1368 else if (jfunc->type == IPA_JF_ANCESTOR
1369 && ipa_get_jf_ancestor_agg_preserved (jfunc))
1370 {
1371 struct ipa_node_params *caller_info = IPA_NODE_REF (cs->caller);
1372 int src_idx = ipa_get_jf_ancestor_formal_id (jfunc);
1373 struct ipcp_param_lattices *src_plats;
1374
1375 src_plats = ipa_get_parm_lattices (caller_info, src_idx);
1376 if (src_plats->aggs && src_plats->aggs_by_ref)
1377 {
1378 /* Currently we do not produce clobber aggregate jump
1379 functions, replace with merging when we do. */
1380 gcc_assert (!jfunc->agg.items);
1381 ret |= merge_aggregate_lattices (cs, dest_plats, src_plats, src_idx,
1382 ipa_get_jf_ancestor_offset (jfunc));
1383 }
1384 else if (!src_plats->aggs_by_ref)
1385 ret |= set_agg_lats_to_bottom (dest_plats);
1386 else
1387 ret |= set_agg_lats_contain_variable (dest_plats);
1388 }
1389 else if (jfunc->agg.items)
1390 {
1391 bool pre_existing = dest_plats->aggs != NULL;
1392 struct ipcp_agg_lattice **aglat = &dest_plats->aggs;
1393 struct ipa_agg_jf_item *item;
1394 int i;
1395
1396 if (set_check_aggs_by_ref (dest_plats, jfunc->agg.by_ref))
1397 return true;
1398
1399 FOR_EACH_VEC_ELT (*jfunc->agg.items, i, item)
1400 {
1401 HOST_WIDE_INT val_size;
1402
1403 if (item->offset < 0)
1404 continue;
1405 gcc_checking_assert (is_gimple_ip_invariant (item->value));
1406 val_size = tree_to_uhwi (TYPE_SIZE (TREE_TYPE (item->value)));
1407
1408 if (merge_agg_lats_step (dest_plats, item->offset, val_size,
1409 &aglat, pre_existing, &ret))
1410 {
1411 ret |= add_value_to_lattice (*aglat, item->value, cs, NULL, 0, 0);
1412 aglat = &(*aglat)->next;
1413 }
1414 else if (dest_plats->aggs_bottom)
1415 return true;
1416 }
1417
1418 ret |= set_chain_of_aglats_contains_variable (*aglat);
1419 }
1420 else
1421 ret |= set_agg_lats_contain_variable (dest_plats);
1422
1423 return ret;
1424 }
1425
1426 /* Propagate constants from the caller to the callee of CS. INFO describes the
1427 caller. */
1428
1429 static bool
1430 propagate_constants_accross_call (struct cgraph_edge *cs)
1431 {
1432 struct ipa_node_params *callee_info;
1433 enum availability availability;
1434 struct cgraph_node *callee, *alias_or_thunk;
1435 struct ipa_edge_args *args;
1436 bool ret = false;
1437 int i, args_count, parms_count;
1438
1439 callee = cs->callee->function_symbol (&availability);
1440 if (!callee->definition)
1441 return false;
1442 gcc_checking_assert (callee->has_gimple_body_p ());
1443 callee_info = IPA_NODE_REF (callee);
1444
1445 args = IPA_EDGE_REF (cs);
1446 args_count = ipa_get_cs_argument_count (args);
1447 parms_count = ipa_get_param_count (callee_info);
1448 if (parms_count == 0)
1449 return false;
1450
1451 /* If this call goes through a thunk we must not propagate to the first (0th)
1452 parameter. However, we might need to uncover a thunk from below a series
1453 of aliases first. */
1454 alias_or_thunk = cs->callee;
1455 while (alias_or_thunk->alias)
1456 alias_or_thunk = alias_or_thunk->get_alias_target ();
1457 if (alias_or_thunk->thunk.thunk_p)
1458 {
1459 ret |= set_all_contains_variable (ipa_get_parm_lattices (callee_info,
1460 0));
1461 i = 1;
1462 }
1463 else
1464 i = 0;
1465
1466 for (; (i < args_count) && (i < parms_count); i++)
1467 {
1468 struct ipa_jump_func *jump_func = ipa_get_ith_jump_func (args, i);
1469 struct ipcp_param_lattices *dest_plats;
1470
1471 dest_plats = ipa_get_parm_lattices (callee_info, i);
1472 if (availability == AVAIL_INTERPOSABLE)
1473 ret |= set_all_contains_variable (dest_plats);
1474 else
1475 {
1476 ret |= propagate_scalar_accross_jump_function (cs, jump_func,
1477 &dest_plats->itself);
1478 ret |= propagate_aggs_accross_jump_function (cs, jump_func,
1479 dest_plats);
1480 }
1481 }
1482 for (; i < parms_count; i++)
1483 ret |= set_all_contains_variable (ipa_get_parm_lattices (callee_info, i));
1484
1485 return ret;
1486 }
1487
1488 /* If an indirect edge IE can be turned into a direct one based on KNOWN_VALS
1489 (which can contain both constants and binfos), KNOWN_BINFOS, KNOWN_AGGS or
1490 AGG_REPS return the destination. The latter three can be NULL. If AGG_REPS
1491 is not NULL, KNOWN_AGGS is ignored. */
1492
1493 static tree
1494 ipa_get_indirect_edge_target_1 (struct cgraph_edge *ie,
1495 vec<tree> known_vals,
1496 vec<tree> known_binfos,
1497 vec<ipa_agg_jump_function_p> known_aggs,
1498 struct ipa_agg_replacement_value *agg_reps)
1499 {
1500 int param_index = ie->indirect_info->param_index;
1501 HOST_WIDE_INT token, anc_offset;
1502 tree otr_type;
1503 tree t;
1504 tree target = NULL;
1505
1506 if (param_index == -1
1507 || known_vals.length () <= (unsigned int) param_index)
1508 return NULL_TREE;
1509
1510 if (!ie->indirect_info->polymorphic)
1511 {
1512 tree t;
1513
1514 if (ie->indirect_info->agg_contents)
1515 {
1516 if (agg_reps)
1517 {
1518 t = NULL;
1519 while (agg_reps)
1520 {
1521 if (agg_reps->index == param_index
1522 && agg_reps->offset == ie->indirect_info->offset
1523 && agg_reps->by_ref == ie->indirect_info->by_ref)
1524 {
1525 t = agg_reps->value;
1526 break;
1527 }
1528 agg_reps = agg_reps->next;
1529 }
1530 }
1531 else if (known_aggs.length () > (unsigned int) param_index)
1532 {
1533 struct ipa_agg_jump_function *agg;
1534 agg = known_aggs[param_index];
1535 t = ipa_find_agg_cst_for_param (agg, ie->indirect_info->offset,
1536 ie->indirect_info->by_ref);
1537 }
1538 else
1539 t = NULL;
1540 }
1541 else
1542 t = known_vals[param_index];
1543
1544 if (t &&
1545 TREE_CODE (t) == ADDR_EXPR
1546 && TREE_CODE (TREE_OPERAND (t, 0)) == FUNCTION_DECL)
1547 return TREE_OPERAND (t, 0);
1548 else
1549 return NULL_TREE;
1550 }
1551
1552 if (!flag_devirtualize)
1553 return NULL_TREE;
1554
1555 gcc_assert (!ie->indirect_info->agg_contents);
1556 token = ie->indirect_info->otr_token;
1557 anc_offset = ie->indirect_info->offset;
1558 otr_type = ie->indirect_info->otr_type;
1559
1560 t = NULL;
1561
1562 /* Try to work out value of virtual table pointer value in replacemnets. */
1563 if (!t && agg_reps && !ie->indirect_info->by_ref
1564 && !ie->indirect_info->vptr_changed)
1565 {
1566 while (agg_reps)
1567 {
1568 if (agg_reps->index == param_index
1569 && agg_reps->offset == ie->indirect_info->offset
1570 && agg_reps->by_ref)
1571 {
1572 t = agg_reps->value;
1573 break;
1574 }
1575 agg_reps = agg_reps->next;
1576 }
1577 }
1578
1579 /* Try to work out value of virtual table pointer value in known
1580 aggregate values. */
1581 if (!t && known_aggs.length () > (unsigned int) param_index
1582 && !ie->indirect_info->by_ref
1583 && !ie->indirect_info->vptr_changed)
1584 {
1585 struct ipa_agg_jump_function *agg;
1586 agg = known_aggs[param_index];
1587 t = ipa_find_agg_cst_for_param (agg, ie->indirect_info->offset,
1588 true);
1589 }
1590
1591 /* If we found the virtual table pointer, lookup the target. */
1592 if (t)
1593 {
1594 tree vtable;
1595 unsigned HOST_WIDE_INT offset;
1596 if (vtable_pointer_value_to_vtable (t, &vtable, &offset))
1597 {
1598 target = gimple_get_virt_method_for_vtable (ie->indirect_info->otr_token,
1599 vtable, offset);
1600 if (target)
1601 {
1602 if ((TREE_CODE (TREE_TYPE (target)) == FUNCTION_TYPE
1603 && DECL_FUNCTION_CODE (target) == BUILT_IN_UNREACHABLE)
1604 || !possible_polymorphic_call_target_p
1605 (ie, cgraph_node::get (target)))
1606 target = ipa_impossible_devirt_target (ie, target);
1607 return target;
1608 }
1609 }
1610 }
1611
1612 /* Did we work out BINFO via type propagation? */
1613 if (!t && known_binfos.length () > (unsigned int) param_index)
1614 t = known_binfos[param_index];
1615 /* Or do we know the constant value of pointer? */
1616 if (!t)
1617 t = known_vals[param_index];
1618 if (!t)
1619 return NULL_TREE;
1620
1621 if (TREE_CODE (t) != TREE_BINFO)
1622 {
1623 ipa_polymorphic_call_context context (t, ie->indirect_info->otr_type,
1624 anc_offset);
1625 vec <cgraph_node *>targets;
1626 bool final;
1627
1628 targets = possible_polymorphic_call_targets
1629 (ie->indirect_info->otr_type,
1630 ie->indirect_info->otr_token,
1631 context, &final);
1632 if (!final || targets.length () > 1)
1633 return NULL_TREE;
1634 if (targets.length () == 1)
1635 target = targets[0]->decl;
1636 else
1637 target = ipa_impossible_devirt_target (ie, NULL_TREE);
1638 }
1639 else
1640 {
1641 tree binfo;
1642
1643 binfo = get_binfo_at_offset (t, anc_offset, otr_type);
1644 if (!binfo)
1645 return NULL_TREE;
1646 target = gimple_get_virt_method_for_binfo (token, binfo);
1647 }
1648
1649 if (target && !possible_polymorphic_call_target_p (ie,
1650 cgraph_node::get (target)))
1651 target = ipa_impossible_devirt_target (ie, target);
1652
1653 return target;
1654 }
1655
1656
1657 /* If an indirect edge IE can be turned into a direct one based on KNOWN_VALS
1658 (which can contain both constants and binfos), KNOWN_BINFOS (which can be
1659 NULL) or KNOWN_AGGS (which also can be NULL) return the destination. */
1660
1661 tree
1662 ipa_get_indirect_edge_target (struct cgraph_edge *ie,
1663 vec<tree> known_vals,
1664 vec<tree> known_binfos,
1665 vec<ipa_agg_jump_function_p> known_aggs)
1666 {
1667 return ipa_get_indirect_edge_target_1 (ie, known_vals, known_binfos,
1668 known_aggs, NULL);
1669 }
1670
1671 /* Calculate devirtualization time bonus for NODE, assuming we know KNOWN_CSTS
1672 and KNOWN_BINFOS. */
1673
1674 static int
1675 devirtualization_time_bonus (struct cgraph_node *node,
1676 vec<tree> known_csts,
1677 vec<tree> known_binfos,
1678 vec<ipa_agg_jump_function_p> known_aggs)
1679 {
1680 struct cgraph_edge *ie;
1681 int res = 0;
1682
1683 for (ie = node->indirect_calls; ie; ie = ie->next_callee)
1684 {
1685 struct cgraph_node *callee;
1686 struct inline_summary *isummary;
1687 enum availability avail;
1688 tree target;
1689
1690 target = ipa_get_indirect_edge_target (ie, known_csts, known_binfos,
1691 known_aggs);
1692 if (!target)
1693 continue;
1694
1695 /* Only bare minimum benefit for clearly un-inlineable targets. */
1696 res += 1;
1697 callee = cgraph_node::get (target);
1698 if (!callee || !callee->definition)
1699 continue;
1700 callee = callee->function_symbol (&avail);
1701 if (avail < AVAIL_AVAILABLE)
1702 continue;
1703 isummary = inline_summary (callee);
1704 if (!isummary->inlinable)
1705 continue;
1706
1707 /* FIXME: The values below need re-considering and perhaps also
1708 integrating into the cost metrics, at lest in some very basic way. */
1709 if (isummary->size <= MAX_INLINE_INSNS_AUTO / 4)
1710 res += 31;
1711 else if (isummary->size <= MAX_INLINE_INSNS_AUTO / 2)
1712 res += 15;
1713 else if (isummary->size <= MAX_INLINE_INSNS_AUTO
1714 || DECL_DECLARED_INLINE_P (callee->decl))
1715 res += 7;
1716 }
1717
1718 return res;
1719 }
1720
1721 /* Return time bonus incurred because of HINTS. */
1722
1723 static int
1724 hint_time_bonus (inline_hints hints)
1725 {
1726 int result = 0;
1727 if (hints & (INLINE_HINT_loop_iterations | INLINE_HINT_loop_stride))
1728 result += PARAM_VALUE (PARAM_IPA_CP_LOOP_HINT_BONUS);
1729 if (hints & INLINE_HINT_array_index)
1730 result += PARAM_VALUE (PARAM_IPA_CP_ARRAY_INDEX_HINT_BONUS);
1731 return result;
1732 }
1733
1734 /* Return true if cloning NODE is a good idea, given the estimated TIME_BENEFIT
1735 and SIZE_COST and with the sum of frequencies of incoming edges to the
1736 potential new clone in FREQUENCIES. */
1737
1738 static bool
1739 good_cloning_opportunity_p (struct cgraph_node *node, int time_benefit,
1740 int freq_sum, gcov_type count_sum, int size_cost)
1741 {
1742 if (time_benefit == 0
1743 || !flag_ipa_cp_clone
1744 || !optimize_function_for_speed_p (DECL_STRUCT_FUNCTION (node->decl)))
1745 return false;
1746
1747 gcc_assert (size_cost > 0);
1748
1749 if (max_count)
1750 {
1751 int factor = (count_sum * 1000) / max_count;
1752 int64_t evaluation = (((int64_t) time_benefit * factor)
1753 / size_cost);
1754
1755 if (dump_file && (dump_flags & TDF_DETAILS))
1756 fprintf (dump_file, " good_cloning_opportunity_p (time: %i, "
1757 "size: %i, count_sum: " HOST_WIDE_INT_PRINT_DEC
1758 ") -> evaluation: " "%"PRId64
1759 ", threshold: %i\n",
1760 time_benefit, size_cost, (HOST_WIDE_INT) count_sum,
1761 evaluation, PARAM_VALUE (PARAM_IPA_CP_EVAL_THRESHOLD));
1762
1763 return evaluation >= PARAM_VALUE (PARAM_IPA_CP_EVAL_THRESHOLD);
1764 }
1765 else
1766 {
1767 int64_t evaluation = (((int64_t) time_benefit * freq_sum)
1768 / size_cost);
1769
1770 if (dump_file && (dump_flags & TDF_DETAILS))
1771 fprintf (dump_file, " good_cloning_opportunity_p (time: %i, "
1772 "size: %i, freq_sum: %i) -> evaluation: "
1773 "%"PRId64 ", threshold: %i\n",
1774 time_benefit, size_cost, freq_sum, evaluation,
1775 PARAM_VALUE (PARAM_IPA_CP_EVAL_THRESHOLD));
1776
1777 return evaluation >= PARAM_VALUE (PARAM_IPA_CP_EVAL_THRESHOLD);
1778 }
1779 }
1780
1781 /* Return all context independent values from aggregate lattices in PLATS in a
1782 vector. Return NULL if there are none. */
1783
1784 static vec<ipa_agg_jf_item, va_gc> *
1785 context_independent_aggregate_values (struct ipcp_param_lattices *plats)
1786 {
1787 vec<ipa_agg_jf_item, va_gc> *res = NULL;
1788
1789 if (plats->aggs_bottom
1790 || plats->aggs_contain_variable
1791 || plats->aggs_count == 0)
1792 return NULL;
1793
1794 for (struct ipcp_agg_lattice *aglat = plats->aggs;
1795 aglat;
1796 aglat = aglat->next)
1797 if (ipa_lat_is_single_const (aglat))
1798 {
1799 struct ipa_agg_jf_item item;
1800 item.offset = aglat->offset;
1801 item.value = aglat->values->value;
1802 vec_safe_push (res, item);
1803 }
1804 return res;
1805 }
1806
1807 /* Allocate KNOWN_CSTS, KNOWN_BINFOS and, if non-NULL, KNOWN_AGGS and populate
1808 them with values of parameters that are known independent of the context.
1809 INFO describes the function. If REMOVABLE_PARAMS_COST is non-NULL, the
1810 movement cost of all removable parameters will be stored in it. */
1811
1812 static bool
1813 gather_context_independent_values (struct ipa_node_params *info,
1814 vec<tree> *known_csts,
1815 vec<tree> *known_binfos,
1816 vec<ipa_agg_jump_function> *known_aggs,
1817 int *removable_params_cost)
1818 {
1819 int i, count = ipa_get_param_count (info);
1820 bool ret = false;
1821
1822 known_csts->create (0);
1823 known_binfos->create (0);
1824 known_csts->safe_grow_cleared (count);
1825 known_binfos->safe_grow_cleared (count);
1826 if (known_aggs)
1827 {
1828 known_aggs->create (0);
1829 known_aggs->safe_grow_cleared (count);
1830 }
1831
1832 if (removable_params_cost)
1833 *removable_params_cost = 0;
1834
1835 for (i = 0; i < count ; i++)
1836 {
1837 struct ipcp_param_lattices *plats = ipa_get_parm_lattices (info, i);
1838 struct ipcp_lattice *lat = &plats->itself;
1839
1840 if (ipa_lat_is_single_const (lat))
1841 {
1842 struct ipcp_value *val = lat->values;
1843 if (TREE_CODE (val->value) != TREE_BINFO)
1844 {
1845 (*known_csts)[i] = val->value;
1846 if (removable_params_cost)
1847 *removable_params_cost
1848 += estimate_move_cost (TREE_TYPE (val->value), false);
1849 ret = true;
1850 }
1851 else if (plats->virt_call)
1852 {
1853 (*known_binfos)[i] = val->value;
1854 ret = true;
1855 }
1856 else if (removable_params_cost
1857 && !ipa_is_param_used (info, i))
1858 *removable_params_cost += ipa_get_param_move_cost (info, i);
1859 }
1860 else if (removable_params_cost
1861 && !ipa_is_param_used (info, i))
1862 *removable_params_cost
1863 += ipa_get_param_move_cost (info, i);
1864
1865 if (known_aggs)
1866 {
1867 vec<ipa_agg_jf_item, va_gc> *agg_items;
1868 struct ipa_agg_jump_function *ajf;
1869
1870 agg_items = context_independent_aggregate_values (plats);
1871 ajf = &(*known_aggs)[i];
1872 ajf->items = agg_items;
1873 ajf->by_ref = plats->aggs_by_ref;
1874 ret |= agg_items != NULL;
1875 }
1876 }
1877
1878 return ret;
1879 }
1880
1881 /* The current interface in ipa-inline-analysis requires a pointer vector.
1882 Create it.
1883
1884 FIXME: That interface should be re-worked, this is slightly silly. Still,
1885 I'd like to discuss how to change it first and this demonstrates the
1886 issue. */
1887
1888 static vec<ipa_agg_jump_function_p>
1889 agg_jmp_p_vec_for_t_vec (vec<ipa_agg_jump_function> known_aggs)
1890 {
1891 vec<ipa_agg_jump_function_p> ret;
1892 struct ipa_agg_jump_function *ajf;
1893 int i;
1894
1895 ret.create (known_aggs.length ());
1896 FOR_EACH_VEC_ELT (known_aggs, i, ajf)
1897 ret.quick_push (ajf);
1898 return ret;
1899 }
1900
1901 /* Iterate over known values of parameters of NODE and estimate the local
1902 effects in terms of time and size they have. */
1903
1904 static void
1905 estimate_local_effects (struct cgraph_node *node)
1906 {
1907 struct ipa_node_params *info = IPA_NODE_REF (node);
1908 int i, count = ipa_get_param_count (info);
1909 vec<tree> known_csts, known_binfos;
1910 vec<ipa_agg_jump_function> known_aggs;
1911 vec<ipa_agg_jump_function_p> known_aggs_ptrs;
1912 bool always_const;
1913 int base_time = inline_summary (node)->time;
1914 int removable_params_cost;
1915
1916 if (!count || !ipcp_versionable_function_p (node))
1917 return;
1918
1919 if (dump_file && (dump_flags & TDF_DETAILS))
1920 fprintf (dump_file, "\nEstimating effects for %s/%i, base_time: %i.\n",
1921 node->name (), node->order, base_time);
1922
1923 always_const = gather_context_independent_values (info, &known_csts,
1924 &known_binfos, &known_aggs,
1925 &removable_params_cost);
1926 known_aggs_ptrs = agg_jmp_p_vec_for_t_vec (known_aggs);
1927 if (always_const)
1928 {
1929 struct caller_statistics stats;
1930 inline_hints hints;
1931 int time, size;
1932
1933 init_caller_stats (&stats);
1934 node->call_for_symbol_thunks_and_aliases (gather_caller_stats, &stats,
1935 false);
1936 estimate_ipcp_clone_size_and_time (node, known_csts, known_binfos,
1937 known_aggs_ptrs, &size, &time, &hints);
1938 time -= devirtualization_time_bonus (node, known_csts, known_binfos,
1939 known_aggs_ptrs);
1940 time -= hint_time_bonus (hints);
1941 time -= removable_params_cost;
1942 size -= stats.n_calls * removable_params_cost;
1943
1944 if (dump_file)
1945 fprintf (dump_file, " - context independent values, size: %i, "
1946 "time_benefit: %i\n", size, base_time - time);
1947
1948 if (size <= 0
1949 || node->will_be_removed_from_program_if_no_direct_calls_p ())
1950 {
1951 info->do_clone_for_all_contexts = true;
1952 base_time = time;
1953
1954 if (dump_file)
1955 fprintf (dump_file, " Decided to specialize for all "
1956 "known contexts, code not going to grow.\n");
1957 }
1958 else if (good_cloning_opportunity_p (node, base_time - time,
1959 stats.freq_sum, stats.count_sum,
1960 size))
1961 {
1962 if (size + overall_size <= max_new_size)
1963 {
1964 info->do_clone_for_all_contexts = true;
1965 base_time = time;
1966 overall_size += size;
1967
1968 if (dump_file)
1969 fprintf (dump_file, " Decided to specialize for all "
1970 "known contexts, growth deemed beneficial.\n");
1971 }
1972 else if (dump_file && (dump_flags & TDF_DETAILS))
1973 fprintf (dump_file, " Not cloning for all contexts because "
1974 "max_new_size would be reached with %li.\n",
1975 size + overall_size);
1976 }
1977 }
1978
1979 for (i = 0; i < count ; i++)
1980 {
1981 struct ipcp_param_lattices *plats = ipa_get_parm_lattices (info, i);
1982 struct ipcp_lattice *lat = &plats->itself;
1983 struct ipcp_value *val;
1984 int emc;
1985
1986 if (lat->bottom
1987 || !lat->values
1988 || known_csts[i]
1989 || known_binfos[i])
1990 continue;
1991
1992 for (val = lat->values; val; val = val->next)
1993 {
1994 int time, size, time_benefit;
1995 inline_hints hints;
1996
1997 if (TREE_CODE (val->value) != TREE_BINFO)
1998 {
1999 known_csts[i] = val->value;
2000 known_binfos[i] = NULL_TREE;
2001 emc = estimate_move_cost (TREE_TYPE (val->value), true);
2002 }
2003 else if (plats->virt_call)
2004 {
2005 known_csts[i] = NULL_TREE;
2006 known_binfos[i] = val->value;
2007 emc = 0;
2008 }
2009 else
2010 continue;
2011
2012 estimate_ipcp_clone_size_and_time (node, known_csts, known_binfos,
2013 known_aggs_ptrs, &size, &time,
2014 &hints);
2015 time_benefit = base_time - time
2016 + devirtualization_time_bonus (node, known_csts, known_binfos,
2017 known_aggs_ptrs)
2018 + hint_time_bonus (hints)
2019 + removable_params_cost + emc;
2020
2021 gcc_checking_assert (size >=0);
2022 /* The inliner-heuristics based estimates may think that in certain
2023 contexts some functions do not have any size at all but we want
2024 all specializations to have at least a tiny cost, not least not to
2025 divide by zero. */
2026 if (size == 0)
2027 size = 1;
2028
2029 if (dump_file && (dump_flags & TDF_DETAILS))
2030 {
2031 fprintf (dump_file, " - estimates for value ");
2032 print_ipcp_constant_value (dump_file, val->value);
2033 fprintf (dump_file, " for ");
2034 ipa_dump_param (dump_file, info, i);
2035 fprintf (dump_file, ": time_benefit: %i, size: %i\n",
2036 time_benefit, size);
2037 }
2038
2039 val->local_time_benefit = time_benefit;
2040 val->local_size_cost = size;
2041 }
2042 known_binfos[i] = NULL_TREE;
2043 known_csts[i] = NULL_TREE;
2044 }
2045
2046 for (i = 0; i < count ; i++)
2047 {
2048 struct ipcp_param_lattices *plats = ipa_get_parm_lattices (info, i);
2049 struct ipa_agg_jump_function *ajf;
2050 struct ipcp_agg_lattice *aglat;
2051
2052 if (plats->aggs_bottom || !plats->aggs)
2053 continue;
2054
2055 ajf = &known_aggs[i];
2056 for (aglat = plats->aggs; aglat; aglat = aglat->next)
2057 {
2058 struct ipcp_value *val;
2059 if (aglat->bottom || !aglat->values
2060 /* If the following is true, the one value is in known_aggs. */
2061 || (!plats->aggs_contain_variable
2062 && ipa_lat_is_single_const (aglat)))
2063 continue;
2064
2065 for (val = aglat->values; val; val = val->next)
2066 {
2067 int time, size, time_benefit;
2068 struct ipa_agg_jf_item item;
2069 inline_hints hints;
2070
2071 item.offset = aglat->offset;
2072 item.value = val->value;
2073 vec_safe_push (ajf->items, item);
2074
2075 estimate_ipcp_clone_size_and_time (node, known_csts, known_binfos,
2076 known_aggs_ptrs, &size, &time,
2077 &hints);
2078 time_benefit = base_time - time
2079 + devirtualization_time_bonus (node, known_csts, known_binfos,
2080 known_aggs_ptrs)
2081 + hint_time_bonus (hints);
2082 gcc_checking_assert (size >=0);
2083 if (size == 0)
2084 size = 1;
2085
2086 if (dump_file && (dump_flags & TDF_DETAILS))
2087 {
2088 fprintf (dump_file, " - estimates for value ");
2089 print_ipcp_constant_value (dump_file, val->value);
2090 fprintf (dump_file, " for ");
2091 ipa_dump_param (dump_file, info, i);
2092 fprintf (dump_file, "[%soffset: " HOST_WIDE_INT_PRINT_DEC
2093 "]: time_benefit: %i, size: %i\n",
2094 plats->aggs_by_ref ? "ref " : "",
2095 aglat->offset, time_benefit, size);
2096 }
2097
2098 val->local_time_benefit = time_benefit;
2099 val->local_size_cost = size;
2100 ajf->items->pop ();
2101 }
2102 }
2103 }
2104
2105 for (i = 0; i < count ; i++)
2106 vec_free (known_aggs[i].items);
2107
2108 known_csts.release ();
2109 known_binfos.release ();
2110 known_aggs.release ();
2111 known_aggs_ptrs.release ();
2112 }
2113
2114
2115 /* Add value CUR_VAL and all yet-unsorted values it is dependent on to the
2116 topological sort of values. */
2117
2118 static void
2119 add_val_to_toposort (struct ipcp_value *cur_val)
2120 {
2121 static int dfs_counter = 0;
2122 static struct ipcp_value *stack;
2123 struct ipcp_value_source *src;
2124
2125 if (cur_val->dfs)
2126 return;
2127
2128 dfs_counter++;
2129 cur_val->dfs = dfs_counter;
2130 cur_val->low_link = dfs_counter;
2131
2132 cur_val->topo_next = stack;
2133 stack = cur_val;
2134 cur_val->on_stack = true;
2135
2136 for (src = cur_val->sources; src; src = src->next)
2137 if (src->val)
2138 {
2139 if (src->val->dfs == 0)
2140 {
2141 add_val_to_toposort (src->val);
2142 if (src->val->low_link < cur_val->low_link)
2143 cur_val->low_link = src->val->low_link;
2144 }
2145 else if (src->val->on_stack
2146 && src->val->dfs < cur_val->low_link)
2147 cur_val->low_link = src->val->dfs;
2148 }
2149
2150 if (cur_val->dfs == cur_val->low_link)
2151 {
2152 struct ipcp_value *v, *scc_list = NULL;
2153
2154 do
2155 {
2156 v = stack;
2157 stack = v->topo_next;
2158 v->on_stack = false;
2159
2160 v->scc_next = scc_list;
2161 scc_list = v;
2162 }
2163 while (v != cur_val);
2164
2165 cur_val->topo_next = values_topo;
2166 values_topo = cur_val;
2167 }
2168 }
2169
2170 /* Add all values in lattices associated with NODE to the topological sort if
2171 they are not there yet. */
2172
2173 static void
2174 add_all_node_vals_to_toposort (struct cgraph_node *node)
2175 {
2176 struct ipa_node_params *info = IPA_NODE_REF (node);
2177 int i, count = ipa_get_param_count (info);
2178
2179 for (i = 0; i < count ; i++)
2180 {
2181 struct ipcp_param_lattices *plats = ipa_get_parm_lattices (info, i);
2182 struct ipcp_lattice *lat = &plats->itself;
2183 struct ipcp_agg_lattice *aglat;
2184 struct ipcp_value *val;
2185
2186 if (!lat->bottom)
2187 for (val = lat->values; val; val = val->next)
2188 add_val_to_toposort (val);
2189
2190 if (!plats->aggs_bottom)
2191 for (aglat = plats->aggs; aglat; aglat = aglat->next)
2192 if (!aglat->bottom)
2193 for (val = aglat->values; val; val = val->next)
2194 add_val_to_toposort (val);
2195 }
2196 }
2197
2198 /* One pass of constants propagation along the call graph edges, from callers
2199 to callees (requires topological ordering in TOPO), iterate over strongly
2200 connected components. */
2201
2202 static void
2203 propagate_constants_topo (struct ipa_topo_info *topo)
2204 {
2205 int i;
2206
2207 for (i = topo->nnodes - 1; i >= 0; i--)
2208 {
2209 unsigned j;
2210 struct cgraph_node *v, *node = topo->order[i];
2211 vec<cgraph_node *> cycle_nodes = ipa_get_nodes_in_cycle (node);
2212
2213 /* First, iteratively propagate within the strongly connected component
2214 until all lattices stabilize. */
2215 FOR_EACH_VEC_ELT (cycle_nodes, j, v)
2216 if (v->has_gimple_body_p ())
2217 push_node_to_stack (topo, v);
2218
2219 v = pop_node_from_stack (topo);
2220 while (v)
2221 {
2222 struct cgraph_edge *cs;
2223
2224 for (cs = v->callees; cs; cs = cs->next_callee)
2225 if (ipa_edge_within_scc (cs)
2226 && propagate_constants_accross_call (cs))
2227 push_node_to_stack (topo, cs->callee);
2228 v = pop_node_from_stack (topo);
2229 }
2230
2231 /* Afterwards, propagate along edges leading out of the SCC, calculates
2232 the local effects of the discovered constants and all valid values to
2233 their topological sort. */
2234 FOR_EACH_VEC_ELT (cycle_nodes, j, v)
2235 if (v->has_gimple_body_p ())
2236 {
2237 struct cgraph_edge *cs;
2238
2239 estimate_local_effects (v);
2240 add_all_node_vals_to_toposort (v);
2241 for (cs = v->callees; cs; cs = cs->next_callee)
2242 if (!ipa_edge_within_scc (cs))
2243 propagate_constants_accross_call (cs);
2244 }
2245 cycle_nodes.release ();
2246 }
2247 }
2248
2249
2250 /* Return the sum of A and B if none of them is bigger than INT_MAX/2, return
2251 the bigger one if otherwise. */
2252
2253 static int
2254 safe_add (int a, int b)
2255 {
2256 if (a > INT_MAX/2 || b > INT_MAX/2)
2257 return a > b ? a : b;
2258 else
2259 return a + b;
2260 }
2261
2262
2263 /* Propagate the estimated effects of individual values along the topological
2264 from the dependent values to those they depend on. */
2265
2266 static void
2267 propagate_effects (void)
2268 {
2269 struct ipcp_value *base;
2270
2271 for (base = values_topo; base; base = base->topo_next)
2272 {
2273 struct ipcp_value_source *src;
2274 struct ipcp_value *val;
2275 int time = 0, size = 0;
2276
2277 for (val = base; val; val = val->scc_next)
2278 {
2279 time = safe_add (time,
2280 val->local_time_benefit + val->prop_time_benefit);
2281 size = safe_add (size, val->local_size_cost + val->prop_size_cost);
2282 }
2283
2284 for (val = base; val; val = val->scc_next)
2285 for (src = val->sources; src; src = src->next)
2286 if (src->val
2287 && src->cs->maybe_hot_p ())
2288 {
2289 src->val->prop_time_benefit = safe_add (time,
2290 src->val->prop_time_benefit);
2291 src->val->prop_size_cost = safe_add (size,
2292 src->val->prop_size_cost);
2293 }
2294 }
2295 }
2296
2297
2298 /* Propagate constants, binfos and their effects from the summaries
2299 interprocedurally. */
2300
2301 static void
2302 ipcp_propagate_stage (struct ipa_topo_info *topo)
2303 {
2304 struct cgraph_node *node;
2305
2306 if (dump_file)
2307 fprintf (dump_file, "\n Propagating constants:\n\n");
2308
2309 if (in_lto_p)
2310 ipa_update_after_lto_read ();
2311
2312
2313 FOR_EACH_DEFINED_FUNCTION (node)
2314 {
2315 struct ipa_node_params *info = IPA_NODE_REF (node);
2316
2317 determine_versionability (node);
2318 if (node->has_gimple_body_p ())
2319 {
2320 info->lattices = XCNEWVEC (struct ipcp_param_lattices,
2321 ipa_get_param_count (info));
2322 initialize_node_lattices (node);
2323 }
2324 if (node->definition && !node->alias)
2325 overall_size += inline_summary (node)->self_size;
2326 if (node->count > max_count)
2327 max_count = node->count;
2328 }
2329
2330 max_new_size = overall_size;
2331 if (max_new_size < PARAM_VALUE (PARAM_LARGE_UNIT_INSNS))
2332 max_new_size = PARAM_VALUE (PARAM_LARGE_UNIT_INSNS);
2333 max_new_size += max_new_size * PARAM_VALUE (PARAM_IPCP_UNIT_GROWTH) / 100 + 1;
2334
2335 if (dump_file)
2336 fprintf (dump_file, "\noverall_size: %li, max_new_size: %li\n",
2337 overall_size, max_new_size);
2338
2339 propagate_constants_topo (topo);
2340 #ifdef ENABLE_CHECKING
2341 ipcp_verify_propagated_values ();
2342 #endif
2343 propagate_effects ();
2344
2345 if (dump_file)
2346 {
2347 fprintf (dump_file, "\nIPA lattices after all propagation:\n");
2348 print_all_lattices (dump_file, (dump_flags & TDF_DETAILS), true);
2349 }
2350 }
2351
2352 /* Discover newly direct outgoing edges from NODE which is a new clone with
2353 known KNOWN_VALS and make them direct. */
2354
2355 static void
2356 ipcp_discover_new_direct_edges (struct cgraph_node *node,
2357 vec<tree> known_vals,
2358 struct ipa_agg_replacement_value *aggvals)
2359 {
2360 struct cgraph_edge *ie, *next_ie;
2361 bool found = false;
2362
2363 for (ie = node->indirect_calls; ie; ie = next_ie)
2364 {
2365 tree target;
2366
2367 next_ie = ie->next_callee;
2368 target = ipa_get_indirect_edge_target_1 (ie, known_vals, vNULL, vNULL,
2369 aggvals);
2370 if (target)
2371 {
2372 bool agg_contents = ie->indirect_info->agg_contents;
2373 bool polymorphic = ie->indirect_info->polymorphic;
2374 int param_index = ie->indirect_info->param_index;
2375 struct cgraph_edge *cs = ipa_make_edge_direct_to_target (ie, target);
2376 found = true;
2377
2378 if (cs && !agg_contents && !polymorphic)
2379 {
2380 struct ipa_node_params *info = IPA_NODE_REF (node);
2381 int c = ipa_get_controlled_uses (info, param_index);
2382 if (c != IPA_UNDESCRIBED_USE)
2383 {
2384 struct ipa_ref *to_del;
2385
2386 c--;
2387 ipa_set_controlled_uses (info, param_index, c);
2388 if (dump_file && (dump_flags & TDF_DETAILS))
2389 fprintf (dump_file, " controlled uses count of param "
2390 "%i bumped down to %i\n", param_index, c);
2391 if (c == 0
2392 && (to_del = node->find_reference (cs->callee, NULL, 0)))
2393 {
2394 if (dump_file && (dump_flags & TDF_DETAILS))
2395 fprintf (dump_file, " and even removing its "
2396 "cloning-created reference\n");
2397 to_del->remove_reference ();
2398 }
2399 }
2400 }
2401 }
2402 }
2403 /* Turning calls to direct calls will improve overall summary. */
2404 if (found)
2405 inline_update_overall_summary (node);
2406 }
2407
2408 /* Vector of pointers which for linked lists of clones of an original crgaph
2409 edge. */
2410
2411 static vec<cgraph_edge *> next_edge_clone;
2412 static vec<cgraph_edge *> prev_edge_clone;
2413
2414 static inline void
2415 grow_edge_clone_vectors (void)
2416 {
2417 if (next_edge_clone.length ()
2418 <= (unsigned) symtab->edges_max_uid)
2419 next_edge_clone.safe_grow_cleared (symtab->edges_max_uid + 1);
2420 if (prev_edge_clone.length ()
2421 <= (unsigned) symtab->edges_max_uid)
2422 prev_edge_clone.safe_grow_cleared (symtab->edges_max_uid + 1);
2423 }
2424
2425 /* Edge duplication hook to grow the appropriate linked list in
2426 next_edge_clone. */
2427
2428 static void
2429 ipcp_edge_duplication_hook (struct cgraph_edge *src, struct cgraph_edge *dst,
2430 void *)
2431 {
2432 grow_edge_clone_vectors ();
2433
2434 struct cgraph_edge *old_next = next_edge_clone[src->uid];
2435 if (old_next)
2436 prev_edge_clone[old_next->uid] = dst;
2437 prev_edge_clone[dst->uid] = src;
2438
2439 next_edge_clone[dst->uid] = old_next;
2440 next_edge_clone[src->uid] = dst;
2441 }
2442
2443 /* Hook that is called by cgraph.c when an edge is removed. */
2444
2445 static void
2446 ipcp_edge_removal_hook (struct cgraph_edge *cs, void *)
2447 {
2448 grow_edge_clone_vectors ();
2449
2450 struct cgraph_edge *prev = prev_edge_clone[cs->uid];
2451 struct cgraph_edge *next = next_edge_clone[cs->uid];
2452 if (prev)
2453 next_edge_clone[prev->uid] = next;
2454 if (next)
2455 prev_edge_clone[next->uid] = prev;
2456 }
2457
2458 /* See if NODE is a clone with a known aggregate value at a given OFFSET of a
2459 parameter with the given INDEX. */
2460
2461 static tree
2462 get_clone_agg_value (struct cgraph_node *node, HOST_WIDE_INT offset,
2463 int index)
2464 {
2465 struct ipa_agg_replacement_value *aggval;
2466
2467 aggval = ipa_get_agg_replacements_for_node (node);
2468 while (aggval)
2469 {
2470 if (aggval->offset == offset
2471 && aggval->index == index)
2472 return aggval->value;
2473 aggval = aggval->next;
2474 }
2475 return NULL_TREE;
2476 }
2477
2478 /* Return true if edge CS does bring about the value described by SRC. */
2479
2480 static bool
2481 cgraph_edge_brings_value_p (struct cgraph_edge *cs,
2482 struct ipcp_value_source *src)
2483 {
2484 struct ipa_node_params *caller_info = IPA_NODE_REF (cs->caller);
2485 cgraph_node *real_dest = cs->callee->function_symbol ();
2486 struct ipa_node_params *dst_info = IPA_NODE_REF (real_dest);
2487
2488 if ((dst_info->ipcp_orig_node && !dst_info->is_all_contexts_clone)
2489 || caller_info->node_dead)
2490 return false;
2491 if (!src->val)
2492 return true;
2493
2494 if (caller_info->ipcp_orig_node)
2495 {
2496 tree t;
2497 if (src->offset == -1)
2498 t = caller_info->known_vals[src->index];
2499 else
2500 t = get_clone_agg_value (cs->caller, src->offset, src->index);
2501 return (t != NULL_TREE
2502 && values_equal_for_ipcp_p (src->val->value, t));
2503 }
2504 else
2505 {
2506 struct ipcp_agg_lattice *aglat;
2507 struct ipcp_param_lattices *plats = ipa_get_parm_lattices (caller_info,
2508 src->index);
2509 if (src->offset == -1)
2510 return (ipa_lat_is_single_const (&plats->itself)
2511 && values_equal_for_ipcp_p (src->val->value,
2512 plats->itself.values->value));
2513 else
2514 {
2515 if (plats->aggs_bottom || plats->aggs_contain_variable)
2516 return false;
2517 for (aglat = plats->aggs; aglat; aglat = aglat->next)
2518 if (aglat->offset == src->offset)
2519 return (ipa_lat_is_single_const (aglat)
2520 && values_equal_for_ipcp_p (src->val->value,
2521 aglat->values->value));
2522 }
2523 return false;
2524 }
2525 }
2526
2527 /* Get the next clone in the linked list of clones of an edge. */
2528
2529 static inline struct cgraph_edge *
2530 get_next_cgraph_edge_clone (struct cgraph_edge *cs)
2531 {
2532 return next_edge_clone[cs->uid];
2533 }
2534
2535 /* Given VAL, iterate over all its sources and if they still hold, add their
2536 edge frequency and their number into *FREQUENCY and *CALLER_COUNT
2537 respectively. */
2538
2539 static bool
2540 get_info_about_necessary_edges (struct ipcp_value *val, int *freq_sum,
2541 gcov_type *count_sum, int *caller_count)
2542 {
2543 struct ipcp_value_source *src;
2544 int freq = 0, count = 0;
2545 gcov_type cnt = 0;
2546 bool hot = false;
2547
2548 for (src = val->sources; src; src = src->next)
2549 {
2550 struct cgraph_edge *cs = src->cs;
2551 while (cs)
2552 {
2553 if (cgraph_edge_brings_value_p (cs, src))
2554 {
2555 count++;
2556 freq += cs->frequency;
2557 cnt += cs->count;
2558 hot |= cs->maybe_hot_p ();
2559 }
2560 cs = get_next_cgraph_edge_clone (cs);
2561 }
2562 }
2563
2564 *freq_sum = freq;
2565 *count_sum = cnt;
2566 *caller_count = count;
2567 return hot;
2568 }
2569
2570 /* Return a vector of incoming edges that do bring value VAL. It is assumed
2571 their number is known and equal to CALLER_COUNT. */
2572
2573 static vec<cgraph_edge *>
2574 gather_edges_for_value (struct ipcp_value *val, int caller_count)
2575 {
2576 struct ipcp_value_source *src;
2577 vec<cgraph_edge *> ret;
2578
2579 ret.create (caller_count);
2580 for (src = val->sources; src; src = src->next)
2581 {
2582 struct cgraph_edge *cs = src->cs;
2583 while (cs)
2584 {
2585 if (cgraph_edge_brings_value_p (cs, src))
2586 ret.quick_push (cs);
2587 cs = get_next_cgraph_edge_clone (cs);
2588 }
2589 }
2590
2591 return ret;
2592 }
2593
2594 /* Construct a replacement map for a know VALUE for a formal parameter PARAM.
2595 Return it or NULL if for some reason it cannot be created. */
2596
2597 static struct ipa_replace_map *
2598 get_replacement_map (struct ipa_node_params *info, tree value, int parm_num)
2599 {
2600 struct ipa_replace_map *replace_map;
2601
2602
2603 replace_map = ggc_alloc<ipa_replace_map> ();
2604 if (dump_file)
2605 {
2606 fprintf (dump_file, " replacing ");
2607 ipa_dump_param (dump_file, info, parm_num);
2608
2609 fprintf (dump_file, " with const ");
2610 print_generic_expr (dump_file, value, 0);
2611 fprintf (dump_file, "\n");
2612 }
2613 replace_map->old_tree = NULL;
2614 replace_map->parm_num = parm_num;
2615 replace_map->new_tree = value;
2616 replace_map->replace_p = true;
2617 replace_map->ref_p = false;
2618
2619 return replace_map;
2620 }
2621
2622 /* Dump new profiling counts */
2623
2624 static void
2625 dump_profile_updates (struct cgraph_node *orig_node,
2626 struct cgraph_node *new_node)
2627 {
2628 struct cgraph_edge *cs;
2629
2630 fprintf (dump_file, " setting count of the specialized node to "
2631 HOST_WIDE_INT_PRINT_DEC "\n", (HOST_WIDE_INT) new_node->count);
2632 for (cs = new_node->callees; cs ; cs = cs->next_callee)
2633 fprintf (dump_file, " edge to %s has count "
2634 HOST_WIDE_INT_PRINT_DEC "\n",
2635 cs->callee->name (), (HOST_WIDE_INT) cs->count);
2636
2637 fprintf (dump_file, " setting count of the original node to "
2638 HOST_WIDE_INT_PRINT_DEC "\n", (HOST_WIDE_INT) orig_node->count);
2639 for (cs = orig_node->callees; cs ; cs = cs->next_callee)
2640 fprintf (dump_file, " edge to %s is left with "
2641 HOST_WIDE_INT_PRINT_DEC "\n",
2642 cs->callee->name (), (HOST_WIDE_INT) cs->count);
2643 }
2644
2645 /* After a specialized NEW_NODE version of ORIG_NODE has been created, update
2646 their profile information to reflect this. */
2647
2648 static void
2649 update_profiling_info (struct cgraph_node *orig_node,
2650 struct cgraph_node *new_node)
2651 {
2652 struct cgraph_edge *cs;
2653 struct caller_statistics stats;
2654 gcov_type new_sum, orig_sum;
2655 gcov_type remainder, orig_node_count = orig_node->count;
2656
2657 if (orig_node_count == 0)
2658 return;
2659
2660 init_caller_stats (&stats);
2661 orig_node->call_for_symbol_thunks_and_aliases (gather_caller_stats, &stats,
2662 false);
2663 orig_sum = stats.count_sum;
2664 init_caller_stats (&stats);
2665 new_node->call_for_symbol_thunks_and_aliases (gather_caller_stats, &stats,
2666 false);
2667 new_sum = stats.count_sum;
2668
2669 if (orig_node_count < orig_sum + new_sum)
2670 {
2671 if (dump_file)
2672 fprintf (dump_file, " Problem: node %s/%i has too low count "
2673 HOST_WIDE_INT_PRINT_DEC " while the sum of incoming "
2674 "counts is " HOST_WIDE_INT_PRINT_DEC "\n",
2675 orig_node->name (), orig_node->order,
2676 (HOST_WIDE_INT) orig_node_count,
2677 (HOST_WIDE_INT) (orig_sum + new_sum));
2678
2679 orig_node_count = (orig_sum + new_sum) * 12 / 10;
2680 if (dump_file)
2681 fprintf (dump_file, " proceeding by pretending it was "
2682 HOST_WIDE_INT_PRINT_DEC "\n",
2683 (HOST_WIDE_INT) orig_node_count);
2684 }
2685
2686 new_node->count = new_sum;
2687 remainder = orig_node_count - new_sum;
2688 orig_node->count = remainder;
2689
2690 for (cs = new_node->callees; cs ; cs = cs->next_callee)
2691 if (cs->frequency)
2692 cs->count = apply_probability (cs->count,
2693 GCOV_COMPUTE_SCALE (new_sum,
2694 orig_node_count));
2695 else
2696 cs->count = 0;
2697
2698 for (cs = orig_node->callees; cs ; cs = cs->next_callee)
2699 cs->count = apply_probability (cs->count,
2700 GCOV_COMPUTE_SCALE (remainder,
2701 orig_node_count));
2702
2703 if (dump_file)
2704 dump_profile_updates (orig_node, new_node);
2705 }
2706
2707 /* Update the respective profile of specialized NEW_NODE and the original
2708 ORIG_NODE after additional edges with cumulative count sum REDIRECTED_SUM
2709 have been redirected to the specialized version. */
2710
2711 static void
2712 update_specialized_profile (struct cgraph_node *new_node,
2713 struct cgraph_node *orig_node,
2714 gcov_type redirected_sum)
2715 {
2716 struct cgraph_edge *cs;
2717 gcov_type new_node_count, orig_node_count = orig_node->count;
2718
2719 if (dump_file)
2720 fprintf (dump_file, " the sum of counts of redirected edges is "
2721 HOST_WIDE_INT_PRINT_DEC "\n", (HOST_WIDE_INT) redirected_sum);
2722 if (orig_node_count == 0)
2723 return;
2724
2725 gcc_assert (orig_node_count >= redirected_sum);
2726
2727 new_node_count = new_node->count;
2728 new_node->count += redirected_sum;
2729 orig_node->count -= redirected_sum;
2730
2731 for (cs = new_node->callees; cs ; cs = cs->next_callee)
2732 if (cs->frequency)
2733 cs->count += apply_probability (cs->count,
2734 GCOV_COMPUTE_SCALE (redirected_sum,
2735 new_node_count));
2736 else
2737 cs->count = 0;
2738
2739 for (cs = orig_node->callees; cs ; cs = cs->next_callee)
2740 {
2741 gcov_type dec = apply_probability (cs->count,
2742 GCOV_COMPUTE_SCALE (redirected_sum,
2743 orig_node_count));
2744 if (dec < cs->count)
2745 cs->count -= dec;
2746 else
2747 cs->count = 0;
2748 }
2749
2750 if (dump_file)
2751 dump_profile_updates (orig_node, new_node);
2752 }
2753
2754 /* Create a specialized version of NODE with known constants and types of
2755 parameters in KNOWN_VALS and redirect all edges in CALLERS to it. */
2756
2757 static struct cgraph_node *
2758 create_specialized_node (struct cgraph_node *node,
2759 vec<tree> known_vals,
2760 struct ipa_agg_replacement_value *aggvals,
2761 vec<cgraph_edge *> callers)
2762 {
2763 struct ipa_node_params *new_info, *info = IPA_NODE_REF (node);
2764 vec<ipa_replace_map *, va_gc> *replace_trees = NULL;
2765 struct ipa_agg_replacement_value *av;
2766 struct cgraph_node *new_node;
2767 int i, count = ipa_get_param_count (info);
2768 bitmap args_to_skip;
2769
2770 gcc_assert (!info->ipcp_orig_node);
2771
2772 if (node->local.can_change_signature)
2773 {
2774 args_to_skip = BITMAP_GGC_ALLOC ();
2775 for (i = 0; i < count; i++)
2776 {
2777 tree t = known_vals[i];
2778
2779 if ((t && TREE_CODE (t) != TREE_BINFO)
2780 || !ipa_is_param_used (info, i))
2781 bitmap_set_bit (args_to_skip, i);
2782 }
2783 }
2784 else
2785 {
2786 args_to_skip = NULL;
2787 if (dump_file && (dump_flags & TDF_DETAILS))
2788 fprintf (dump_file, " cannot change function signature\n");
2789 }
2790
2791 for (i = 0; i < count ; i++)
2792 {
2793 tree t = known_vals[i];
2794 if (t && TREE_CODE (t) != TREE_BINFO)
2795 {
2796 struct ipa_replace_map *replace_map;
2797
2798 replace_map = get_replacement_map (info, t, i);
2799 if (replace_map)
2800 vec_safe_push (replace_trees, replace_map);
2801 }
2802 }
2803
2804 new_node = node->create_virtual_clone (callers, replace_trees,
2805 args_to_skip, "constprop");
2806 ipa_set_node_agg_value_chain (new_node, aggvals);
2807 for (av = aggvals; av; av = av->next)
2808 new_node->maybe_create_reference (av->value, IPA_REF_ADDR, NULL);
2809
2810 if (dump_file && (dump_flags & TDF_DETAILS))
2811 {
2812 fprintf (dump_file, " the new node is %s/%i.\n",
2813 new_node->name (), new_node->order);
2814 if (aggvals)
2815 ipa_dump_agg_replacement_values (dump_file, aggvals);
2816 }
2817 ipa_check_create_node_params ();
2818 update_profiling_info (node, new_node);
2819 new_info = IPA_NODE_REF (new_node);
2820 new_info->ipcp_orig_node = node;
2821 new_info->known_vals = known_vals;
2822
2823 ipcp_discover_new_direct_edges (new_node, known_vals, aggvals);
2824
2825 callers.release ();
2826 return new_node;
2827 }
2828
2829 /* Given a NODE, and a subset of its CALLERS, try to populate blanks slots in
2830 KNOWN_VALS with constants and types that are also known for all of the
2831 CALLERS. */
2832
2833 static void
2834 find_more_scalar_values_for_callers_subset (struct cgraph_node *node,
2835 vec<tree> known_vals,
2836 vec<cgraph_edge *> callers)
2837 {
2838 struct ipa_node_params *info = IPA_NODE_REF (node);
2839 int i, count = ipa_get_param_count (info);
2840
2841 for (i = 0; i < count ; i++)
2842 {
2843 struct cgraph_edge *cs;
2844 tree newval = NULL_TREE;
2845 int j;
2846
2847 if (ipa_get_scalar_lat (info, i)->bottom || known_vals[i])
2848 continue;
2849
2850 FOR_EACH_VEC_ELT (callers, j, cs)
2851 {
2852 struct ipa_jump_func *jump_func;
2853 tree t;
2854
2855 if (i >= ipa_get_cs_argument_count (IPA_EDGE_REF (cs)))
2856 {
2857 newval = NULL_TREE;
2858 break;
2859 }
2860 jump_func = ipa_get_ith_jump_func (IPA_EDGE_REF (cs), i);
2861 t = ipa_value_from_jfunc (IPA_NODE_REF (cs->caller), jump_func);
2862 if (!t
2863 || (newval
2864 && !values_equal_for_ipcp_p (t, newval)))
2865 {
2866 newval = NULL_TREE;
2867 break;
2868 }
2869 else
2870 newval = t;
2871 }
2872
2873 if (newval)
2874 {
2875 if (dump_file && (dump_flags & TDF_DETAILS))
2876 {
2877 fprintf (dump_file, " adding an extra known scalar value ");
2878 print_ipcp_constant_value (dump_file, newval);
2879 fprintf (dump_file, " for ");
2880 ipa_dump_param (dump_file, info, i);
2881 fprintf (dump_file, "\n");
2882 }
2883
2884 known_vals[i] = newval;
2885 }
2886 }
2887 }
2888
2889 /* Go through PLATS and create a vector of values consisting of values and
2890 offsets (minus OFFSET) of lattices that contain only a single value. */
2891
2892 static vec<ipa_agg_jf_item>
2893 copy_plats_to_inter (struct ipcp_param_lattices *plats, HOST_WIDE_INT offset)
2894 {
2895 vec<ipa_agg_jf_item> res = vNULL;
2896
2897 if (!plats->aggs || plats->aggs_contain_variable || plats->aggs_bottom)
2898 return vNULL;
2899
2900 for (struct ipcp_agg_lattice *aglat = plats->aggs; aglat; aglat = aglat->next)
2901 if (ipa_lat_is_single_const (aglat))
2902 {
2903 struct ipa_agg_jf_item ti;
2904 ti.offset = aglat->offset - offset;
2905 ti.value = aglat->values->value;
2906 res.safe_push (ti);
2907 }
2908 return res;
2909 }
2910
2911 /* Intersect all values in INTER with single value lattices in PLATS (while
2912 subtracting OFFSET). */
2913
2914 static void
2915 intersect_with_plats (struct ipcp_param_lattices *plats,
2916 vec<ipa_agg_jf_item> *inter,
2917 HOST_WIDE_INT offset)
2918 {
2919 struct ipcp_agg_lattice *aglat;
2920 struct ipa_agg_jf_item *item;
2921 int k;
2922
2923 if (!plats->aggs || plats->aggs_contain_variable || plats->aggs_bottom)
2924 {
2925 inter->release ();
2926 return;
2927 }
2928
2929 aglat = plats->aggs;
2930 FOR_EACH_VEC_ELT (*inter, k, item)
2931 {
2932 bool found = false;
2933 if (!item->value)
2934 continue;
2935 while (aglat)
2936 {
2937 if (aglat->offset - offset > item->offset)
2938 break;
2939 if (aglat->offset - offset == item->offset)
2940 {
2941 gcc_checking_assert (item->value);
2942 if (values_equal_for_ipcp_p (item->value, aglat->values->value))
2943 found = true;
2944 break;
2945 }
2946 aglat = aglat->next;
2947 }
2948 if (!found)
2949 item->value = NULL_TREE;
2950 }
2951 }
2952
2953 /* Copy agggregate replacement values of NODE (which is an IPA-CP clone) to the
2954 vector result while subtracting OFFSET from the individual value offsets. */
2955
2956 static vec<ipa_agg_jf_item>
2957 agg_replacements_to_vector (struct cgraph_node *node, int index,
2958 HOST_WIDE_INT offset)
2959 {
2960 struct ipa_agg_replacement_value *av;
2961 vec<ipa_agg_jf_item> res = vNULL;
2962
2963 for (av = ipa_get_agg_replacements_for_node (node); av; av = av->next)
2964 if (av->index == index
2965 && (av->offset - offset) >= 0)
2966 {
2967 struct ipa_agg_jf_item item;
2968 gcc_checking_assert (av->value);
2969 item.offset = av->offset - offset;
2970 item.value = av->value;
2971 res.safe_push (item);
2972 }
2973
2974 return res;
2975 }
2976
2977 /* Intersect all values in INTER with those that we have already scheduled to
2978 be replaced in parameter number INDEX of NODE, which is an IPA-CP clone
2979 (while subtracting OFFSET). */
2980
2981 static void
2982 intersect_with_agg_replacements (struct cgraph_node *node, int index,
2983 vec<ipa_agg_jf_item> *inter,
2984 HOST_WIDE_INT offset)
2985 {
2986 struct ipa_agg_replacement_value *srcvals;
2987 struct ipa_agg_jf_item *item;
2988 int i;
2989
2990 srcvals = ipa_get_agg_replacements_for_node (node);
2991 if (!srcvals)
2992 {
2993 inter->release ();
2994 return;
2995 }
2996
2997 FOR_EACH_VEC_ELT (*inter, i, item)
2998 {
2999 struct ipa_agg_replacement_value *av;
3000 bool found = false;
3001 if (!item->value)
3002 continue;
3003 for (av = srcvals; av; av = av->next)
3004 {
3005 gcc_checking_assert (av->value);
3006 if (av->index == index
3007 && av->offset - offset == item->offset)
3008 {
3009 if (values_equal_for_ipcp_p (item->value, av->value))
3010 found = true;
3011 break;
3012 }
3013 }
3014 if (!found)
3015 item->value = NULL_TREE;
3016 }
3017 }
3018
3019 /* Intersect values in INTER with aggregate values that come along edge CS to
3020 parameter number INDEX and return it. If INTER does not actually exist yet,
3021 copy all incoming values to it. If we determine we ended up with no values
3022 whatsoever, return a released vector. */
3023
3024 static vec<ipa_agg_jf_item>
3025 intersect_aggregates_with_edge (struct cgraph_edge *cs, int index,
3026 vec<ipa_agg_jf_item> inter)
3027 {
3028 struct ipa_jump_func *jfunc;
3029 jfunc = ipa_get_ith_jump_func (IPA_EDGE_REF (cs), index);
3030 if (jfunc->type == IPA_JF_PASS_THROUGH
3031 && ipa_get_jf_pass_through_operation (jfunc) == NOP_EXPR)
3032 {
3033 struct ipa_node_params *caller_info = IPA_NODE_REF (cs->caller);
3034 int src_idx = ipa_get_jf_pass_through_formal_id (jfunc);
3035
3036 if (caller_info->ipcp_orig_node)
3037 {
3038 struct cgraph_node *orig_node = caller_info->ipcp_orig_node;
3039 struct ipcp_param_lattices *orig_plats;
3040 orig_plats = ipa_get_parm_lattices (IPA_NODE_REF (orig_node),
3041 src_idx);
3042 if (agg_pass_through_permissible_p (orig_plats, jfunc))
3043 {
3044 if (!inter.exists ())
3045 inter = agg_replacements_to_vector (cs->caller, src_idx, 0);
3046 else
3047 intersect_with_agg_replacements (cs->caller, src_idx,
3048 &inter, 0);
3049 }
3050 else
3051 {
3052 inter.release ();
3053 return vNULL;
3054 }
3055 }
3056 else
3057 {
3058 struct ipcp_param_lattices *src_plats;
3059 src_plats = ipa_get_parm_lattices (caller_info, src_idx);
3060 if (agg_pass_through_permissible_p (src_plats, jfunc))
3061 {
3062 /* Currently we do not produce clobber aggregate jump
3063 functions, adjust when we do. */
3064 gcc_checking_assert (!jfunc->agg.items);
3065 if (!inter.exists ())
3066 inter = copy_plats_to_inter (src_plats, 0);
3067 else
3068 intersect_with_plats (src_plats, &inter, 0);
3069 }
3070 else
3071 {
3072 inter.release ();
3073 return vNULL;
3074 }
3075 }
3076 }
3077 else if (jfunc->type == IPA_JF_ANCESTOR
3078 && ipa_get_jf_ancestor_agg_preserved (jfunc))
3079 {
3080 struct ipa_node_params *caller_info = IPA_NODE_REF (cs->caller);
3081 int src_idx = ipa_get_jf_ancestor_formal_id (jfunc);
3082 struct ipcp_param_lattices *src_plats;
3083 HOST_WIDE_INT delta = ipa_get_jf_ancestor_offset (jfunc);
3084
3085 if (caller_info->ipcp_orig_node)
3086 {
3087 if (!inter.exists ())
3088 inter = agg_replacements_to_vector (cs->caller, src_idx, delta);
3089 else
3090 intersect_with_agg_replacements (cs->caller, src_idx, &inter,
3091 delta);
3092 }
3093 else
3094 {
3095 src_plats = ipa_get_parm_lattices (caller_info, src_idx);;
3096 /* Currently we do not produce clobber aggregate jump
3097 functions, adjust when we do. */
3098 gcc_checking_assert (!src_plats->aggs || !jfunc->agg.items);
3099 if (!inter.exists ())
3100 inter = copy_plats_to_inter (src_plats, delta);
3101 else
3102 intersect_with_plats (src_plats, &inter, delta);
3103 }
3104 }
3105 else if (jfunc->agg.items)
3106 {
3107 struct ipa_agg_jf_item *item;
3108 int k;
3109
3110 if (!inter.exists ())
3111 for (unsigned i = 0; i < jfunc->agg.items->length (); i++)
3112 inter.safe_push ((*jfunc->agg.items)[i]);
3113 else
3114 FOR_EACH_VEC_ELT (inter, k, item)
3115 {
3116 int l = 0;
3117 bool found = false;;
3118
3119 if (!item->value)
3120 continue;
3121
3122 while ((unsigned) l < jfunc->agg.items->length ())
3123 {
3124 struct ipa_agg_jf_item *ti;
3125 ti = &(*jfunc->agg.items)[l];
3126 if (ti->offset > item->offset)
3127 break;
3128 if (ti->offset == item->offset)
3129 {
3130 gcc_checking_assert (ti->value);
3131 if (values_equal_for_ipcp_p (item->value,
3132 ti->value))
3133 found = true;
3134 break;
3135 }
3136 l++;
3137 }
3138 if (!found)
3139 item->value = NULL;
3140 }
3141 }
3142 else
3143 {
3144 inter.release ();
3145 return vec<ipa_agg_jf_item>();
3146 }
3147 return inter;
3148 }
3149
3150 /* Look at edges in CALLERS and collect all known aggregate values that arrive
3151 from all of them. */
3152
3153 static struct ipa_agg_replacement_value *
3154 find_aggregate_values_for_callers_subset (struct cgraph_node *node,
3155 vec<cgraph_edge *> callers)
3156 {
3157 struct ipa_node_params *dest_info = IPA_NODE_REF (node);
3158 struct ipa_agg_replacement_value *res;
3159 struct ipa_agg_replacement_value **tail = &res;
3160 struct cgraph_edge *cs;
3161 int i, j, count = ipa_get_param_count (dest_info);
3162
3163 FOR_EACH_VEC_ELT (callers, j, cs)
3164 {
3165 int c = ipa_get_cs_argument_count (IPA_EDGE_REF (cs));
3166 if (c < count)
3167 count = c;
3168 }
3169
3170 for (i = 0; i < count ; i++)
3171 {
3172 struct cgraph_edge *cs;
3173 vec<ipa_agg_jf_item> inter = vNULL;
3174 struct ipa_agg_jf_item *item;
3175 struct ipcp_param_lattices *plats = ipa_get_parm_lattices (dest_info, i);
3176 int j;
3177
3178 /* Among other things, the following check should deal with all by_ref
3179 mismatches. */
3180 if (plats->aggs_bottom)
3181 continue;
3182
3183 FOR_EACH_VEC_ELT (callers, j, cs)
3184 {
3185 inter = intersect_aggregates_with_edge (cs, i, inter);
3186
3187 if (!inter.exists ())
3188 goto next_param;
3189 }
3190
3191 FOR_EACH_VEC_ELT (inter, j, item)
3192 {
3193 struct ipa_agg_replacement_value *v;
3194
3195 if (!item->value)
3196 continue;
3197
3198 v = ggc_alloc<ipa_agg_replacement_value> ();
3199 v->index = i;
3200 v->offset = item->offset;
3201 v->value = item->value;
3202 v->by_ref = plats->aggs_by_ref;
3203 *tail = v;
3204 tail = &v->next;
3205 }
3206
3207 next_param:
3208 if (inter.exists ())
3209 inter.release ();
3210 }
3211 *tail = NULL;
3212 return res;
3213 }
3214
3215 /* Turn KNOWN_AGGS into a list of aggreate replacement values. */
3216
3217 static struct ipa_agg_replacement_value *
3218 known_aggs_to_agg_replacement_list (vec<ipa_agg_jump_function> known_aggs)
3219 {
3220 struct ipa_agg_replacement_value *res;
3221 struct ipa_agg_replacement_value **tail = &res;
3222 struct ipa_agg_jump_function *aggjf;
3223 struct ipa_agg_jf_item *item;
3224 int i, j;
3225
3226 FOR_EACH_VEC_ELT (known_aggs, i, aggjf)
3227 FOR_EACH_VEC_SAFE_ELT (aggjf->items, j, item)
3228 {
3229 struct ipa_agg_replacement_value *v;
3230 v = ggc_alloc<ipa_agg_replacement_value> ();
3231 v->index = i;
3232 v->offset = item->offset;
3233 v->value = item->value;
3234 v->by_ref = aggjf->by_ref;
3235 *tail = v;
3236 tail = &v->next;
3237 }
3238 *tail = NULL;
3239 return res;
3240 }
3241
3242 /* Determine whether CS also brings all scalar values that the NODE is
3243 specialized for. */
3244
3245 static bool
3246 cgraph_edge_brings_all_scalars_for_node (struct cgraph_edge *cs,
3247 struct cgraph_node *node)
3248 {
3249 struct ipa_node_params *dest_info = IPA_NODE_REF (node);
3250 int count = ipa_get_param_count (dest_info);
3251 struct ipa_node_params *caller_info;
3252 struct ipa_edge_args *args;
3253 int i;
3254
3255 caller_info = IPA_NODE_REF (cs->caller);
3256 args = IPA_EDGE_REF (cs);
3257 for (i = 0; i < count; i++)
3258 {
3259 struct ipa_jump_func *jump_func;
3260 tree val, t;
3261
3262 val = dest_info->known_vals[i];
3263 if (!val)
3264 continue;
3265
3266 if (i >= ipa_get_cs_argument_count (args))
3267 return false;
3268 jump_func = ipa_get_ith_jump_func (args, i);
3269 t = ipa_value_from_jfunc (caller_info, jump_func);
3270 if (!t || !values_equal_for_ipcp_p (val, t))
3271 return false;
3272 }
3273 return true;
3274 }
3275
3276 /* Determine whether CS also brings all aggregate values that NODE is
3277 specialized for. */
3278 static bool
3279 cgraph_edge_brings_all_agg_vals_for_node (struct cgraph_edge *cs,
3280 struct cgraph_node *node)
3281 {
3282 struct ipa_node_params *orig_caller_info = IPA_NODE_REF (cs->caller);
3283 struct ipa_node_params *orig_node_info;
3284 struct ipa_agg_replacement_value *aggval;
3285 int i, ec, count;
3286
3287 aggval = ipa_get_agg_replacements_for_node (node);
3288 if (!aggval)
3289 return true;
3290
3291 count = ipa_get_param_count (IPA_NODE_REF (node));
3292 ec = ipa_get_cs_argument_count (IPA_EDGE_REF (cs));
3293 if (ec < count)
3294 for (struct ipa_agg_replacement_value *av = aggval; av; av = av->next)
3295 if (aggval->index >= ec)
3296 return false;
3297
3298 orig_node_info = IPA_NODE_REF (IPA_NODE_REF (node)->ipcp_orig_node);
3299 if (orig_caller_info->ipcp_orig_node)
3300 orig_caller_info = IPA_NODE_REF (orig_caller_info->ipcp_orig_node);
3301
3302 for (i = 0; i < count; i++)
3303 {
3304 static vec<ipa_agg_jf_item> values = vec<ipa_agg_jf_item>();
3305 struct ipcp_param_lattices *plats;
3306 bool interesting = false;
3307 for (struct ipa_agg_replacement_value *av = aggval; av; av = av->next)
3308 if (aggval->index == i)
3309 {
3310 interesting = true;
3311 break;
3312 }
3313 if (!interesting)
3314 continue;
3315
3316 plats = ipa_get_parm_lattices (orig_node_info, aggval->index);
3317 if (plats->aggs_bottom)
3318 return false;
3319
3320 values = intersect_aggregates_with_edge (cs, i, values);
3321 if (!values.exists ())
3322 return false;
3323
3324 for (struct ipa_agg_replacement_value *av = aggval; av; av = av->next)
3325 if (aggval->index == i)
3326 {
3327 struct ipa_agg_jf_item *item;
3328 int j;
3329 bool found = false;
3330 FOR_EACH_VEC_ELT (values, j, item)
3331 if (item->value
3332 && item->offset == av->offset
3333 && values_equal_for_ipcp_p (item->value, av->value))
3334 {
3335 found = true;
3336 break;
3337 }
3338 if (!found)
3339 {
3340 values.release ();
3341 return false;
3342 }
3343 }
3344 }
3345 return true;
3346 }
3347
3348 /* Given an original NODE and a VAL for which we have already created a
3349 specialized clone, look whether there are incoming edges that still lead
3350 into the old node but now also bring the requested value and also conform to
3351 all other criteria such that they can be redirected the the special node.
3352 This function can therefore redirect the final edge in a SCC. */
3353
3354 static void
3355 perhaps_add_new_callers (struct cgraph_node *node, struct ipcp_value *val)
3356 {
3357 struct ipcp_value_source *src;
3358 gcov_type redirected_sum = 0;
3359
3360 for (src = val->sources; src; src = src->next)
3361 {
3362 struct cgraph_edge *cs = src->cs;
3363 while (cs)
3364 {
3365 enum availability availability;
3366 struct cgraph_node *dst = cs->callee->function_symbol (&availability);
3367 if ((dst == node || IPA_NODE_REF (dst)->is_all_contexts_clone)
3368 && availability > AVAIL_INTERPOSABLE
3369 && cgraph_edge_brings_value_p (cs, src))
3370 {
3371 if (cgraph_edge_brings_all_scalars_for_node (cs, val->spec_node)
3372 && cgraph_edge_brings_all_agg_vals_for_node (cs,
3373 val->spec_node))
3374 {
3375 if (dump_file)
3376 fprintf (dump_file, " - adding an extra caller %s/%i"
3377 " of %s/%i\n",
3378 xstrdup (cs->caller->name ()),
3379 cs->caller->order,
3380 xstrdup (val->spec_node->name ()),
3381 val->spec_node->order);
3382
3383 cs->redirect_callee (val->spec_node);
3384 redirected_sum += cs->count;
3385 }
3386 }
3387 cs = get_next_cgraph_edge_clone (cs);
3388 }
3389 }
3390
3391 if (redirected_sum)
3392 update_specialized_profile (val->spec_node, node, redirected_sum);
3393 }
3394
3395
3396 /* Copy KNOWN_BINFOS to KNOWN_VALS. */
3397
3398 static void
3399 move_binfos_to_values (vec<tree> known_vals,
3400 vec<tree> known_binfos)
3401 {
3402 tree t;
3403 int i;
3404
3405 for (i = 0; known_binfos.iterate (i, &t); i++)
3406 if (t)
3407 known_vals[i] = t;
3408 }
3409
3410 /* Return true if there is a replacement equivalent to VALUE, INDEX and OFFSET
3411 among those in the AGGVALS list. */
3412
3413 DEBUG_FUNCTION bool
3414 ipcp_val_in_agg_replacements_p (struct ipa_agg_replacement_value *aggvals,
3415 int index, HOST_WIDE_INT offset, tree value)
3416 {
3417 while (aggvals)
3418 {
3419 if (aggvals->index == index
3420 && aggvals->offset == offset
3421 && values_equal_for_ipcp_p (aggvals->value, value))
3422 return true;
3423 aggvals = aggvals->next;
3424 }
3425 return false;
3426 }
3427
3428 /* Decide wheter to create a special version of NODE for value VAL of parameter
3429 at the given INDEX. If OFFSET is -1, the value is for the parameter itself,
3430 otherwise it is stored at the given OFFSET of the parameter. KNOWN_CSTS,
3431 KNOWN_BINFOS and KNOWN_AGGS describe the other already known values. */
3432
3433 static bool
3434 decide_about_value (struct cgraph_node *node, int index, HOST_WIDE_INT offset,
3435 struct ipcp_value *val, vec<tree> known_csts,
3436 vec<tree> known_binfos)
3437 {
3438 struct ipa_agg_replacement_value *aggvals;
3439 int freq_sum, caller_count;
3440 gcov_type count_sum;
3441 vec<cgraph_edge *> callers;
3442 vec<tree> kv;
3443
3444 if (val->spec_node)
3445 {
3446 perhaps_add_new_callers (node, val);
3447 return false;
3448 }
3449 else if (val->local_size_cost + overall_size > max_new_size)
3450 {
3451 if (dump_file && (dump_flags & TDF_DETAILS))
3452 fprintf (dump_file, " Ignoring candidate value because "
3453 "max_new_size would be reached with %li.\n",
3454 val->local_size_cost + overall_size);
3455 return false;
3456 }
3457 else if (!get_info_about_necessary_edges (val, &freq_sum, &count_sum,
3458 &caller_count))
3459 return false;
3460
3461 if (dump_file && (dump_flags & TDF_DETAILS))
3462 {
3463 fprintf (dump_file, " - considering value ");
3464 print_ipcp_constant_value (dump_file, val->value);
3465 fprintf (dump_file, " for ");
3466 ipa_dump_param (dump_file, IPA_NODE_REF (node), index);
3467 if (offset != -1)
3468 fprintf (dump_file, ", offset: " HOST_WIDE_INT_PRINT_DEC, offset);
3469 fprintf (dump_file, " (caller_count: %i)\n", caller_count);
3470 }
3471
3472 if (!good_cloning_opportunity_p (node, val->local_time_benefit,
3473 freq_sum, count_sum,
3474 val->local_size_cost)
3475 && !good_cloning_opportunity_p (node,
3476 val->local_time_benefit
3477 + val->prop_time_benefit,
3478 freq_sum, count_sum,
3479 val->local_size_cost
3480 + val->prop_size_cost))
3481 return false;
3482
3483 if (dump_file)
3484 fprintf (dump_file, " Creating a specialized node of %s/%i.\n",
3485 node->name (), node->order);
3486
3487 callers = gather_edges_for_value (val, caller_count);
3488 kv = known_csts.copy ();
3489 move_binfos_to_values (kv, known_binfos);
3490 if (offset == -1)
3491 kv[index] = val->value;
3492 find_more_scalar_values_for_callers_subset (node, kv, callers);
3493 aggvals = find_aggregate_values_for_callers_subset (node, callers);
3494 gcc_checking_assert (offset == -1
3495 || ipcp_val_in_agg_replacements_p (aggvals, index,
3496 offset, val->value));
3497 val->spec_node = create_specialized_node (node, kv, aggvals, callers);
3498 overall_size += val->local_size_cost;
3499
3500 /* TODO: If for some lattice there is only one other known value
3501 left, make a special node for it too. */
3502
3503 return true;
3504 }
3505
3506 /* Decide whether and what specialized clones of NODE should be created. */
3507
3508 static bool
3509 decide_whether_version_node (struct cgraph_node *node)
3510 {
3511 struct ipa_node_params *info = IPA_NODE_REF (node);
3512 int i, count = ipa_get_param_count (info);
3513 vec<tree> known_csts, known_binfos;
3514 vec<ipa_agg_jump_function> known_aggs = vNULL;
3515 bool ret = false;
3516
3517 if (count == 0)
3518 return false;
3519
3520 if (dump_file && (dump_flags & TDF_DETAILS))
3521 fprintf (dump_file, "\nEvaluating opportunities for %s/%i.\n",
3522 node->name (), node->order);
3523
3524 gather_context_independent_values (info, &known_csts, &known_binfos,
3525 info->do_clone_for_all_contexts ? &known_aggs
3526 : NULL, NULL);
3527
3528 for (i = 0; i < count ;i++)
3529 {
3530 struct ipcp_param_lattices *plats = ipa_get_parm_lattices (info, i);
3531 struct ipcp_lattice *lat = &plats->itself;
3532 struct ipcp_value *val;
3533
3534 if (!lat->bottom
3535 && !known_csts[i]
3536 && !known_binfos[i])
3537 for (val = lat->values; val; val = val->next)
3538 ret |= decide_about_value (node, i, -1, val, known_csts,
3539 known_binfos);
3540
3541 if (!plats->aggs_bottom)
3542 {
3543 struct ipcp_agg_lattice *aglat;
3544 struct ipcp_value *val;
3545 for (aglat = plats->aggs; aglat; aglat = aglat->next)
3546 if (!aglat->bottom && aglat->values
3547 /* If the following is false, the one value is in
3548 known_aggs. */
3549 && (plats->aggs_contain_variable
3550 || !ipa_lat_is_single_const (aglat)))
3551 for (val = aglat->values; val; val = val->next)
3552 ret |= decide_about_value (node, i, aglat->offset, val,
3553 known_csts, known_binfos);
3554 }
3555 info = IPA_NODE_REF (node);
3556 }
3557
3558 if (info->do_clone_for_all_contexts)
3559 {
3560 struct cgraph_node *clone;
3561 vec<cgraph_edge *> callers;
3562
3563 if (dump_file)
3564 fprintf (dump_file, " - Creating a specialized node of %s/%i "
3565 "for all known contexts.\n", node->name (),
3566 node->order);
3567
3568 callers = node->collect_callers ();
3569 move_binfos_to_values (known_csts, known_binfos);
3570 clone = create_specialized_node (node, known_csts,
3571 known_aggs_to_agg_replacement_list (known_aggs),
3572 callers);
3573 info = IPA_NODE_REF (node);
3574 info->do_clone_for_all_contexts = false;
3575 IPA_NODE_REF (clone)->is_all_contexts_clone = true;
3576 for (i = 0; i < count ; i++)
3577 vec_free (known_aggs[i].items);
3578 known_aggs.release ();
3579 ret = true;
3580 }
3581 else
3582 known_csts.release ();
3583
3584 known_binfos.release ();
3585 return ret;
3586 }
3587
3588 /* Transitively mark all callees of NODE within the same SCC as not dead. */
3589
3590 static void
3591 spread_undeadness (struct cgraph_node *node)
3592 {
3593 struct cgraph_edge *cs;
3594
3595 for (cs = node->callees; cs; cs = cs->next_callee)
3596 if (ipa_edge_within_scc (cs))
3597 {
3598 struct cgraph_node *callee;
3599 struct ipa_node_params *info;
3600
3601 callee = cs->callee->function_symbol (NULL);
3602 info = IPA_NODE_REF (callee);
3603
3604 if (info->node_dead)
3605 {
3606 info->node_dead = 0;
3607 spread_undeadness (callee);
3608 }
3609 }
3610 }
3611
3612 /* Return true if NODE has a caller from outside of its SCC that is not
3613 dead. Worker callback for cgraph_for_node_and_aliases. */
3614
3615 static bool
3616 has_undead_caller_from_outside_scc_p (struct cgraph_node *node,
3617 void *data ATTRIBUTE_UNUSED)
3618 {
3619 struct cgraph_edge *cs;
3620
3621 for (cs = node->callers; cs; cs = cs->next_caller)
3622 if (cs->caller->thunk.thunk_p
3623 && cs->caller->call_for_symbol_thunks_and_aliases
3624 (has_undead_caller_from_outside_scc_p, NULL, true))
3625 return true;
3626 else if (!ipa_edge_within_scc (cs)
3627 && !IPA_NODE_REF (cs->caller)->node_dead)
3628 return true;
3629 return false;
3630 }
3631
3632
3633 /* Identify nodes within the same SCC as NODE which are no longer needed
3634 because of new clones and will be removed as unreachable. */
3635
3636 static void
3637 identify_dead_nodes (struct cgraph_node *node)
3638 {
3639 struct cgraph_node *v;
3640 for (v = node; v ; v = ((struct ipa_dfs_info *) v->aux)->next_cycle)
3641 if (v->will_be_removed_from_program_if_no_direct_calls_p ()
3642 && !v->call_for_symbol_thunks_and_aliases
3643 (has_undead_caller_from_outside_scc_p, NULL, true))
3644 IPA_NODE_REF (v)->node_dead = 1;
3645
3646 for (v = node; v ; v = ((struct ipa_dfs_info *) v->aux)->next_cycle)
3647 if (!IPA_NODE_REF (v)->node_dead)
3648 spread_undeadness (v);
3649
3650 if (dump_file && (dump_flags & TDF_DETAILS))
3651 {
3652 for (v = node; v ; v = ((struct ipa_dfs_info *) v->aux)->next_cycle)
3653 if (IPA_NODE_REF (v)->node_dead)
3654 fprintf (dump_file, " Marking node as dead: %s/%i.\n",
3655 v->name (), v->order);
3656 }
3657 }
3658
3659 /* The decision stage. Iterate over the topological order of call graph nodes
3660 TOPO and make specialized clones if deemed beneficial. */
3661
3662 static void
3663 ipcp_decision_stage (struct ipa_topo_info *topo)
3664 {
3665 int i;
3666
3667 if (dump_file)
3668 fprintf (dump_file, "\nIPA decision stage:\n\n");
3669
3670 for (i = topo->nnodes - 1; i >= 0; i--)
3671 {
3672 struct cgraph_node *node = topo->order[i];
3673 bool change = false, iterate = true;
3674
3675 while (iterate)
3676 {
3677 struct cgraph_node *v;
3678 iterate = false;
3679 for (v = node; v ; v = ((struct ipa_dfs_info *) v->aux)->next_cycle)
3680 if (v->has_gimple_body_p ()
3681 && ipcp_versionable_function_p (v))
3682 iterate |= decide_whether_version_node (v);
3683
3684 change |= iterate;
3685 }
3686 if (change)
3687 identify_dead_nodes (node);
3688 }
3689 }
3690
3691 /* The IPCP driver. */
3692
3693 static unsigned int
3694 ipcp_driver (void)
3695 {
3696 struct cgraph_2edge_hook_list *edge_duplication_hook_holder;
3697 struct cgraph_edge_hook_list *edge_removal_hook_holder;
3698 struct ipa_topo_info topo;
3699
3700 ipa_check_create_node_params ();
3701 ipa_check_create_edge_args ();
3702 grow_edge_clone_vectors ();
3703 edge_duplication_hook_holder =
3704 symtab->add_edge_duplication_hook (&ipcp_edge_duplication_hook, NULL);
3705 edge_removal_hook_holder =
3706 symtab->add_edge_removal_hook (&ipcp_edge_removal_hook, NULL);
3707
3708 ipcp_values_pool = create_alloc_pool ("IPA-CP values",
3709 sizeof (struct ipcp_value), 32);
3710 ipcp_sources_pool = create_alloc_pool ("IPA-CP value sources",
3711 sizeof (struct ipcp_value_source), 64);
3712 ipcp_agg_lattice_pool = create_alloc_pool ("IPA_CP aggregate lattices",
3713 sizeof (struct ipcp_agg_lattice),
3714 32);
3715 if (dump_file)
3716 {
3717 fprintf (dump_file, "\nIPA structures before propagation:\n");
3718 if (dump_flags & TDF_DETAILS)
3719 ipa_print_all_params (dump_file);
3720 ipa_print_all_jump_functions (dump_file);
3721 }
3722
3723 /* Topological sort. */
3724 build_toporder_info (&topo);
3725 /* Do the interprocedural propagation. */
3726 ipcp_propagate_stage (&topo);
3727 /* Decide what constant propagation and cloning should be performed. */
3728 ipcp_decision_stage (&topo);
3729
3730 /* Free all IPCP structures. */
3731 free_toporder_info (&topo);
3732 next_edge_clone.release ();
3733 symtab->remove_edge_removal_hook (edge_removal_hook_holder);
3734 symtab->remove_edge_duplication_hook (edge_duplication_hook_holder);
3735 ipa_free_all_structures_after_ipa_cp ();
3736 if (dump_file)
3737 fprintf (dump_file, "\nIPA constant propagation end\n");
3738 return 0;
3739 }
3740
3741 /* Initialization and computation of IPCP data structures. This is the initial
3742 intraprocedural analysis of functions, which gathers information to be
3743 propagated later on. */
3744
3745 static void
3746 ipcp_generate_summary (void)
3747 {
3748 struct cgraph_node *node;
3749
3750 if (dump_file)
3751 fprintf (dump_file, "\nIPA constant propagation start:\n");
3752 ipa_register_cgraph_hooks ();
3753
3754 FOR_EACH_FUNCTION_WITH_GIMPLE_BODY (node)
3755 {
3756 node->local.versionable
3757 = tree_versionable_function_p (node->decl);
3758 ipa_analyze_node (node);
3759 }
3760 }
3761
3762 /* Write ipcp summary for nodes in SET. */
3763
3764 static void
3765 ipcp_write_summary (void)
3766 {
3767 ipa_prop_write_jump_functions ();
3768 }
3769
3770 /* Read ipcp summary. */
3771
3772 static void
3773 ipcp_read_summary (void)
3774 {
3775 ipa_prop_read_jump_functions ();
3776 }
3777
3778 namespace {
3779
3780 const pass_data pass_data_ipa_cp =
3781 {
3782 IPA_PASS, /* type */
3783 "cp", /* name */
3784 OPTGROUP_NONE, /* optinfo_flags */
3785 TV_IPA_CONSTANT_PROP, /* tv_id */
3786 0, /* properties_required */
3787 0, /* properties_provided */
3788 0, /* properties_destroyed */
3789 0, /* todo_flags_start */
3790 ( TODO_dump_symtab | TODO_remove_functions ), /* todo_flags_finish */
3791 };
3792
3793 class pass_ipa_cp : public ipa_opt_pass_d
3794 {
3795 public:
3796 pass_ipa_cp (gcc::context *ctxt)
3797 : ipa_opt_pass_d (pass_data_ipa_cp, ctxt,
3798 ipcp_generate_summary, /* generate_summary */
3799 ipcp_write_summary, /* write_summary */
3800 ipcp_read_summary, /* read_summary */
3801 ipa_prop_write_all_agg_replacement, /*
3802 write_optimization_summary */
3803 ipa_prop_read_all_agg_replacement, /*
3804 read_optimization_summary */
3805 NULL, /* stmt_fixup */
3806 0, /* function_transform_todo_flags_start */
3807 ipcp_transform_function, /* function_transform */
3808 NULL) /* variable_transform */
3809 {}
3810
3811 /* opt_pass methods: */
3812 virtual bool gate (function *)
3813 {
3814 /* FIXME: We should remove the optimize check after we ensure we never run
3815 IPA passes when not optimizing. */
3816 return flag_ipa_cp && optimize;
3817 }
3818
3819 virtual unsigned int execute (function *) { return ipcp_driver (); }
3820
3821 }; // class pass_ipa_cp
3822
3823 } // anon namespace
3824
3825 ipa_opt_pass_d *
3826 make_pass_ipa_cp (gcc::context *ctxt)
3827 {
3828 return new pass_ipa_cp (ctxt);
3829 }