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