[RTL-ifcvt] PR rtl-optimization/68506: Fix emitting order of insns in IF-THEN-JOIN...
[official-gcc.git] / gcc / ipa-cp.c
blob6ba2f1493f8610d03d2973cb5aaacadec21e2379
1 /* Interprocedural constant propagation
2 Copyright (C) 2005-2015 Free Software Foundation, Inc.
4 Contributed by Razya Ladelsky <RAZYA@il.ibm.com> and Martin Jambor
5 <mjambor@suse.cz>
7 This file is part of GCC.
9 GCC is free software; you can redistribute it and/or modify it under
10 the terms of the GNU General Public License as published by the Free
11 Software Foundation; either version 3, or (at your option) any later
12 version.
14 GCC is distributed in the hope that it will be useful, but WITHOUT ANY
15 WARRANTY; without even the implied warranty of MERCHANTABILITY or
16 FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
17 for more details.
19 You should have received a copy of the GNU General Public License
20 along with GCC; see the file COPYING3. If not see
21 <http://www.gnu.org/licenses/>. */
23 /* Interprocedural constant propagation (IPA-CP).
25 The goal of this transformation is to
27 1) discover functions which are always invoked with some arguments with the
28 same known constant values and modify the functions so that the
29 subsequent optimizations can take advantage of the knowledge, and
31 2) partial specialization - create specialized versions of functions
32 transformed in this way if some parameters are known constants only in
33 certain contexts but the estimated tradeoff between speedup and cost size
34 is deemed good.
36 The algorithm also propagates types and attempts to perform type based
37 devirtualization. Types are propagated much like constants.
39 The algorithm basically consists of three stages. In the first, functions
40 are analyzed one at a time and jump functions are constructed for all known
41 call-sites. In the second phase, the pass propagates information from the
42 jump functions across the call to reveal what values are available at what
43 call sites, performs estimations of effects of known values on functions and
44 their callees, and finally decides what specialized extra versions should be
45 created. In the third, the special versions materialize and appropriate
46 calls are redirected.
48 The algorithm used is to a certain extent based on "Interprocedural Constant
49 Propagation", by David Callahan, Keith D Cooper, Ken Kennedy, Linda Torczon,
50 Comp86, pg 152-161 and "A Methodology for Procedure Cloning" by Keith D
51 Cooper, Mary W. Hall, and Ken Kennedy.
54 First stage - intraprocedural analysis
55 =======================================
57 This phase computes jump_function and modification flags.
59 A jump function for a call-site represents the values passed as an actual
60 arguments of a given call-site. In principle, there are three types of
61 values:
63 Pass through - the caller's formal parameter is passed as an actual
64 argument, plus an operation on it can be performed.
65 Constant - a constant is passed as an actual argument.
66 Unknown - neither of the above.
68 All jump function types are described in detail in ipa-prop.h, together with
69 the data structures that represent them and methods of accessing them.
71 ipcp_generate_summary() is the main function of the first stage.
73 Second stage - interprocedural analysis
74 ========================================
76 This stage is itself divided into two phases. In the first, we propagate
77 known values over the call graph, in the second, we make cloning decisions.
78 It uses a different algorithm than the original Callahan's paper.
80 First, we traverse the functions topologically from callers to callees and,
81 for each strongly connected component (SCC), we propagate constants
82 according to previously computed jump functions. We also record what known
83 values depend on other known values and estimate local effects. Finally, we
84 propagate cumulative information about these effects from dependent values
85 to those on which they depend.
87 Second, we again traverse the call graph in the same topological order and
88 make clones for functions which we know are called with the same values in
89 all contexts and decide about extra specialized clones of functions just for
90 some contexts - these decisions are based on both local estimates and
91 cumulative estimates propagated from callees.
93 ipcp_propagate_stage() and ipcp_decision_stage() together constitute the
94 third stage.
96 Third phase - materialization of clones, call statement updates.
97 ============================================
99 This stage is currently performed by call graph code (mainly in cgraphunit.c
100 and tree-inline.c) according to instructions inserted to the call graph by
101 the second stage. */
103 #include "config.h"
104 #include "system.h"
105 #include "coretypes.h"
106 #include "backend.h"
107 #include "tree.h"
108 #include "gimple-expr.h"
109 #include "predict.h"
110 #include "alloc-pool.h"
111 #include "tree-pass.h"
112 #include "cgraph.h"
113 #include "diagnostic.h"
114 #include "fold-const.h"
115 #include "gimple-fold.h"
116 #include "symbol-summary.h"
117 #include "ipa-prop.h"
118 #include "tree-pretty-print.h"
119 #include "tree-inline.h"
120 #include "params.h"
121 #include "ipa-inline.h"
122 #include "ipa-utils.h"
124 template <typename valtype> class ipcp_value;
126 /* Describes a particular source for an IPA-CP value. */
128 template <typename valtype>
129 class ipcp_value_source
131 public:
132 /* Aggregate offset of the source, negative if the source is scalar value of
133 the argument itself. */
134 HOST_WIDE_INT offset;
135 /* The incoming edge that brought the value. */
136 cgraph_edge *cs;
137 /* If the jump function that resulted into his value was a pass-through or an
138 ancestor, this is the ipcp_value of the caller from which the described
139 value has been derived. Otherwise it is NULL. */
140 ipcp_value<valtype> *val;
141 /* Next pointer in a linked list of sources of a value. */
142 ipcp_value_source *next;
143 /* If the jump function that resulted into his value was a pass-through or an
144 ancestor, this is the index of the parameter of the caller the jump
145 function references. */
146 int index;
149 /* Common ancestor for all ipcp_value instantiations. */
151 class ipcp_value_base
153 public:
154 /* Time benefit and size cost that specializing the function for this value
155 would bring about in this function alone. */
156 int local_time_benefit, local_size_cost;
157 /* Time benefit and size cost that specializing the function for this value
158 can bring about in it's callees (transitively). */
159 int prop_time_benefit, prop_size_cost;
162 /* Describes one particular value stored in struct ipcp_lattice. */
164 template <typename valtype>
165 class ipcp_value : public ipcp_value_base
167 public:
168 /* The actual value for the given parameter. */
169 valtype value;
170 /* The list of sources from which this value originates. */
171 ipcp_value_source <valtype> *sources;
172 /* Next pointers in a linked list of all values in a lattice. */
173 ipcp_value *next;
174 /* Next pointers in a linked list of values in a strongly connected component
175 of values. */
176 ipcp_value *scc_next;
177 /* Next pointers in a linked list of SCCs of values sorted topologically
178 according their sources. */
179 ipcp_value *topo_next;
180 /* A specialized node created for this value, NULL if none has been (so far)
181 created. */
182 cgraph_node *spec_node;
183 /* Depth first search number and low link for topological sorting of
184 values. */
185 int dfs, low_link;
186 /* True if this valye is currently on the topo-sort stack. */
187 bool on_stack;
189 void add_source (cgraph_edge *cs, ipcp_value *src_val, int src_idx,
190 HOST_WIDE_INT offset);
193 /* Lattice describing potential values of a formal parameter of a function, or
194 a part of an aggreagate. TOP is represented by a lattice with zero values
195 and with contains_variable and bottom flags cleared. BOTTOM is represented
196 by a lattice with the bottom flag set. In that case, values and
197 contains_variable flag should be disregarded. */
199 template <typename valtype>
200 class ipcp_lattice
202 public:
203 /* The list of known values and types in this lattice. Note that values are
204 not deallocated if a lattice is set to bottom because there may be value
205 sources referencing them. */
206 ipcp_value<valtype> *values;
207 /* Number of known values and types in this lattice. */
208 int values_count;
209 /* The lattice contains a variable component (in addition to values). */
210 bool contains_variable;
211 /* The value of the lattice is bottom (i.e. variable and unusable for any
212 propagation). */
213 bool bottom;
215 inline bool is_single_const ();
216 inline bool set_to_bottom ();
217 inline bool set_contains_variable ();
218 bool add_value (valtype newval, cgraph_edge *cs,
219 ipcp_value<valtype> *src_val = NULL,
220 int src_idx = 0, HOST_WIDE_INT offset = -1);
221 void print (FILE * f, bool dump_sources, bool dump_benefits);
224 /* Lattice of tree values with an offset to describe a part of an
225 aggregate. */
227 class ipcp_agg_lattice : public ipcp_lattice<tree>
229 public:
230 /* Offset that is being described by this lattice. */
231 HOST_WIDE_INT offset;
232 /* Size so that we don't have to re-compute it every time we traverse the
233 list. Must correspond to TYPE_SIZE of all lat values. */
234 HOST_WIDE_INT size;
235 /* Next element of the linked list. */
236 struct ipcp_agg_lattice *next;
239 /* Lattice of pointer alignment. Unlike the previous types of lattices, this
240 one is only capable of holding one value. */
242 class ipcp_alignment_lattice
244 public:
245 /* If bottom and top are both false, these two fields hold values as given by
246 ptr_info_def and get_pointer_alignment_1. */
247 unsigned align;
248 unsigned misalign;
250 inline bool bottom_p () const;
251 inline bool top_p () const;
252 inline bool set_to_bottom ();
253 bool meet_with (unsigned new_align, unsigned new_misalign);
254 bool meet_with (const ipcp_alignment_lattice &other, HOST_WIDE_INT offset);
255 void print (FILE * f);
256 private:
257 /* If set, this lattice is bottom and all other fields should be
258 disregarded. */
259 bool bottom;
260 /* If bottom and not_top are false, the lattice is TOP. If not_top is true,
261 the known alignment is stored in the fields align and misalign. The field
262 is negated so that memset to zero initializes the lattice to TOP
263 state. */
264 bool not_top;
266 bool meet_with_1 (unsigned new_align, unsigned new_misalign);
269 /* Structure containing lattices for a parameter itself and for pieces of
270 aggregates that are passed in the parameter or by a reference in a parameter
271 plus some other useful flags. */
273 class ipcp_param_lattices
275 public:
276 /* Lattice describing the value of the parameter itself. */
277 ipcp_lattice<tree> itself;
278 /* Lattice describing the polymorphic contexts of a parameter. */
279 ipcp_lattice<ipa_polymorphic_call_context> ctxlat;
280 /* Lattices describing aggregate parts. */
281 ipcp_agg_lattice *aggs;
282 /* Lattice describing known alignment. */
283 ipcp_alignment_lattice alignment;
284 /* Number of aggregate lattices */
285 int aggs_count;
286 /* True if aggregate data were passed by reference (as opposed to by
287 value). */
288 bool aggs_by_ref;
289 /* All aggregate lattices contain a variable component (in addition to
290 values). */
291 bool aggs_contain_variable;
292 /* The value of all aggregate lattices is bottom (i.e. variable and unusable
293 for any propagation). */
294 bool aggs_bottom;
296 /* There is a virtual call based on this parameter. */
297 bool virt_call;
300 /* Allocation pools for values and their sources in ipa-cp. */
302 object_allocator<ipcp_value<tree> > ipcp_cst_values_pool
303 ("IPA-CP constant values");
305 object_allocator<ipcp_value<ipa_polymorphic_call_context> >
306 ipcp_poly_ctx_values_pool ("IPA-CP polymorphic contexts");
308 object_allocator<ipcp_value_source<tree> > ipcp_sources_pool
309 ("IPA-CP value sources");
311 object_allocator<ipcp_agg_lattice> ipcp_agg_lattice_pool
312 ("IPA_CP aggregate lattices");
314 /* Maximal count found in program. */
316 static gcov_type max_count;
318 /* Original overall size of the program. */
320 static long overall_size, max_new_size;
322 /* Return the param lattices structure corresponding to the Ith formal
323 parameter of the function described by INFO. */
324 static inline struct ipcp_param_lattices *
325 ipa_get_parm_lattices (struct ipa_node_params *info, int i)
327 gcc_assert (i >= 0 && i < ipa_get_param_count (info));
328 gcc_checking_assert (!info->ipcp_orig_node);
329 gcc_checking_assert (info->lattices);
330 return &(info->lattices[i]);
333 /* Return the lattice corresponding to the scalar value of the Ith formal
334 parameter of the function described by INFO. */
335 static inline ipcp_lattice<tree> *
336 ipa_get_scalar_lat (struct ipa_node_params *info, int i)
338 struct ipcp_param_lattices *plats = ipa_get_parm_lattices (info, i);
339 return &plats->itself;
342 /* Return the lattice corresponding to the scalar value of the Ith formal
343 parameter of the function described by INFO. */
344 static inline ipcp_lattice<ipa_polymorphic_call_context> *
345 ipa_get_poly_ctx_lat (struct ipa_node_params *info, int i)
347 struct ipcp_param_lattices *plats = ipa_get_parm_lattices (info, i);
348 return &plats->ctxlat;
351 /* Return whether LAT is a lattice with a single constant and without an
352 undefined value. */
354 template <typename valtype>
355 inline bool
356 ipcp_lattice<valtype>::is_single_const ()
358 if (bottom || contains_variable || values_count != 1)
359 return false;
360 else
361 return true;
364 /* Print V which is extracted from a value in a lattice to F. */
366 static void
367 print_ipcp_constant_value (FILE * f, tree v)
369 if (TREE_CODE (v) == ADDR_EXPR
370 && TREE_CODE (TREE_OPERAND (v, 0)) == CONST_DECL)
372 fprintf (f, "& ");
373 print_generic_expr (f, DECL_INITIAL (TREE_OPERAND (v, 0)), 0);
375 else
376 print_generic_expr (f, v, 0);
379 /* Print V which is extracted from a value in a lattice to F. */
381 static void
382 print_ipcp_constant_value (FILE * f, ipa_polymorphic_call_context v)
384 v.dump(f, false);
387 /* Print a lattice LAT to F. */
389 template <typename valtype>
390 void
391 ipcp_lattice<valtype>::print (FILE * f, bool dump_sources, bool dump_benefits)
393 ipcp_value<valtype> *val;
394 bool prev = false;
396 if (bottom)
398 fprintf (f, "BOTTOM\n");
399 return;
402 if (!values_count && !contains_variable)
404 fprintf (f, "TOP\n");
405 return;
408 if (contains_variable)
410 fprintf (f, "VARIABLE");
411 prev = true;
412 if (dump_benefits)
413 fprintf (f, "\n");
416 for (val = values; val; val = val->next)
418 if (dump_benefits && prev)
419 fprintf (f, " ");
420 else if (!dump_benefits && prev)
421 fprintf (f, ", ");
422 else
423 prev = true;
425 print_ipcp_constant_value (f, val->value);
427 if (dump_sources)
429 ipcp_value_source<valtype> *s;
431 fprintf (f, " [from:");
432 for (s = val->sources; s; s = s->next)
433 fprintf (f, " %i(%i)", s->cs->caller->order,
434 s->cs->frequency);
435 fprintf (f, "]");
438 if (dump_benefits)
439 fprintf (f, " [loc_time: %i, loc_size: %i, "
440 "prop_time: %i, prop_size: %i]\n",
441 val->local_time_benefit, val->local_size_cost,
442 val->prop_time_benefit, val->prop_size_cost);
444 if (!dump_benefits)
445 fprintf (f, "\n");
448 /* Print alignment lattice to F. */
450 void
451 ipcp_alignment_lattice::print (FILE * f)
453 if (top_p ())
454 fprintf (f, " Alignment unknown (TOP)\n");
455 else if (bottom_p ())
456 fprintf (f, " Alignment unusable (BOTTOM)\n");
457 else
458 fprintf (f, " Alignment %u, misalignment %u\n", align, misalign);
461 /* Print all ipcp_lattices of all functions to F. */
463 static void
464 print_all_lattices (FILE * f, bool dump_sources, bool dump_benefits)
466 struct cgraph_node *node;
467 int i, count;
469 fprintf (f, "\nLattices:\n");
470 FOR_EACH_FUNCTION_WITH_GIMPLE_BODY (node)
472 struct ipa_node_params *info;
474 info = IPA_NODE_REF (node);
475 fprintf (f, " Node: %s/%i:\n", node->name (),
476 node->order);
477 count = ipa_get_param_count (info);
478 for (i = 0; i < count; i++)
480 struct ipcp_agg_lattice *aglat;
481 struct ipcp_param_lattices *plats = ipa_get_parm_lattices (info, i);
482 fprintf (f, " param [%d]: ", i);
483 plats->itself.print (f, dump_sources, dump_benefits);
484 fprintf (f, " ctxs: ");
485 plats->ctxlat.print (f, dump_sources, dump_benefits);
486 plats->alignment.print (f);
487 if (plats->virt_call)
488 fprintf (f, " virt_call flag set\n");
490 if (plats->aggs_bottom)
492 fprintf (f, " AGGS BOTTOM\n");
493 continue;
495 if (plats->aggs_contain_variable)
496 fprintf (f, " AGGS VARIABLE\n");
497 for (aglat = plats->aggs; aglat; aglat = aglat->next)
499 fprintf (f, " %soffset " HOST_WIDE_INT_PRINT_DEC ": ",
500 plats->aggs_by_ref ? "ref " : "", aglat->offset);
501 aglat->print (f, dump_sources, dump_benefits);
507 /* Determine whether it is at all technically possible to create clones of NODE
508 and store this information in the ipa_node_params structure associated
509 with NODE. */
511 static void
512 determine_versionability (struct cgraph_node *node,
513 struct ipa_node_params *info)
515 const char *reason = NULL;
517 /* There are a number of generic reasons functions cannot be versioned. We
518 also cannot remove parameters if there are type attributes such as fnspec
519 present. */
520 if (node->alias || node->thunk.thunk_p)
521 reason = "alias or thunk";
522 else if (!node->local.versionable)
523 reason = "not a tree_versionable_function";
524 else if (node->get_availability () <= AVAIL_INTERPOSABLE)
525 reason = "insufficient body availability";
526 else if (!opt_for_fn (node->decl, optimize)
527 || !opt_for_fn (node->decl, flag_ipa_cp))
528 reason = "non-optimized function";
529 else if (lookup_attribute ("omp declare simd", DECL_ATTRIBUTES (node->decl)))
531 /* Ideally we should clone the SIMD clones themselves and create
532 vector copies of them, so IPA-cp and SIMD clones can happily
533 coexist, but that may not be worth the effort. */
534 reason = "function has SIMD clones";
536 /* Don't clone decls local to a comdat group; it breaks and for C++
537 decloned constructors, inlining is always better anyway. */
538 else if (node->comdat_local_p ())
539 reason = "comdat-local function";
541 if (reason && dump_file && !node->alias && !node->thunk.thunk_p)
542 fprintf (dump_file, "Function %s/%i is not versionable, reason: %s.\n",
543 node->name (), node->order, reason);
545 info->versionable = (reason == NULL);
548 /* Return true if it is at all technically possible to create clones of a
549 NODE. */
551 static bool
552 ipcp_versionable_function_p (struct cgraph_node *node)
554 return IPA_NODE_REF (node)->versionable;
557 /* Structure holding accumulated information about callers of a node. */
559 struct caller_statistics
561 gcov_type count_sum;
562 int n_calls, n_hot_calls, freq_sum;
565 /* Initialize fields of STAT to zeroes. */
567 static inline void
568 init_caller_stats (struct caller_statistics *stats)
570 stats->count_sum = 0;
571 stats->n_calls = 0;
572 stats->n_hot_calls = 0;
573 stats->freq_sum = 0;
576 /* Worker callback of cgraph_for_node_and_aliases accumulating statistics of
577 non-thunk incoming edges to NODE. */
579 static bool
580 gather_caller_stats (struct cgraph_node *node, void *data)
582 struct caller_statistics *stats = (struct caller_statistics *) data;
583 struct cgraph_edge *cs;
585 for (cs = node->callers; cs; cs = cs->next_caller)
586 if (!cs->caller->thunk.thunk_p)
588 stats->count_sum += cs->count;
589 stats->freq_sum += cs->frequency;
590 stats->n_calls++;
591 if (cs->maybe_hot_p ())
592 stats->n_hot_calls ++;
594 return false;
598 /* Return true if this NODE is viable candidate for cloning. */
600 static bool
601 ipcp_cloning_candidate_p (struct cgraph_node *node)
603 struct caller_statistics stats;
605 gcc_checking_assert (node->has_gimple_body_p ());
607 if (!opt_for_fn (node->decl, flag_ipa_cp_clone))
609 if (dump_file)
610 fprintf (dump_file, "Not considering %s for cloning; "
611 "-fipa-cp-clone disabled.\n",
612 node->name ());
613 return false;
616 if (!optimize_function_for_speed_p (DECL_STRUCT_FUNCTION (node->decl)))
618 if (dump_file)
619 fprintf (dump_file, "Not considering %s for cloning; "
620 "optimizing it for size.\n",
621 node->name ());
622 return false;
625 init_caller_stats (&stats);
626 node->call_for_symbol_thunks_and_aliases (gather_caller_stats, &stats, false);
628 if (inline_summaries->get (node)->self_size < stats.n_calls)
630 if (dump_file)
631 fprintf (dump_file, "Considering %s for cloning; code might shrink.\n",
632 node->name ());
633 return true;
636 /* When profile is available and function is hot, propagate into it even if
637 calls seems cold; constant propagation can improve function's speed
638 significantly. */
639 if (max_count)
641 if (stats.count_sum > node->count * 90 / 100)
643 if (dump_file)
644 fprintf (dump_file, "Considering %s for cloning; "
645 "usually called directly.\n",
646 node->name ());
647 return true;
650 if (!stats.n_hot_calls)
652 if (dump_file)
653 fprintf (dump_file, "Not considering %s for cloning; no hot calls.\n",
654 node->name ());
655 return false;
657 if (dump_file)
658 fprintf (dump_file, "Considering %s for cloning.\n",
659 node->name ());
660 return true;
663 template <typename valtype>
664 class value_topo_info
666 public:
667 /* Head of the linked list of topologically sorted values. */
668 ipcp_value<valtype> *values_topo;
669 /* Stack for creating SCCs, represented by a linked list too. */
670 ipcp_value<valtype> *stack;
671 /* Counter driving the algorithm in add_val_to_toposort. */
672 int dfs_counter;
674 value_topo_info () : values_topo (NULL), stack (NULL), dfs_counter (0)
676 void add_val (ipcp_value<valtype> *cur_val);
677 void propagate_effects ();
680 /* Arrays representing a topological ordering of call graph nodes and a stack
681 of nodes used during constant propagation and also data required to perform
682 topological sort of values and propagation of benefits in the determined
683 order. */
685 class ipa_topo_info
687 public:
688 /* Array with obtained topological order of cgraph nodes. */
689 struct cgraph_node **order;
690 /* Stack of cgraph nodes used during propagation within SCC until all values
691 in the SCC stabilize. */
692 struct cgraph_node **stack;
693 int nnodes, stack_top;
695 value_topo_info<tree> constants;
696 value_topo_info<ipa_polymorphic_call_context> contexts;
698 ipa_topo_info () : order(NULL), stack(NULL), nnodes(0), stack_top(0),
699 constants ()
703 /* Allocate the arrays in TOPO and topologically sort the nodes into order. */
705 static void
706 build_toporder_info (struct ipa_topo_info *topo)
708 topo->order = XCNEWVEC (struct cgraph_node *, symtab->cgraph_count);
709 topo->stack = XCNEWVEC (struct cgraph_node *, symtab->cgraph_count);
711 gcc_checking_assert (topo->stack_top == 0);
712 topo->nnodes = ipa_reduced_postorder (topo->order, true, true, NULL);
715 /* Free information about strongly connected components and the arrays in
716 TOPO. */
718 static void
719 free_toporder_info (struct ipa_topo_info *topo)
721 ipa_free_postorder_info ();
722 free (topo->order);
723 free (topo->stack);
726 /* Add NODE to the stack in TOPO, unless it is already there. */
728 static inline void
729 push_node_to_stack (struct ipa_topo_info *topo, struct cgraph_node *node)
731 struct ipa_node_params *info = IPA_NODE_REF (node);
732 if (info->node_enqueued)
733 return;
734 info->node_enqueued = 1;
735 topo->stack[topo->stack_top++] = node;
738 /* Pop a node from the stack in TOPO and return it or return NULL if the stack
739 is empty. */
741 static struct cgraph_node *
742 pop_node_from_stack (struct ipa_topo_info *topo)
744 if (topo->stack_top)
746 struct cgraph_node *node;
747 topo->stack_top--;
748 node = topo->stack[topo->stack_top];
749 IPA_NODE_REF (node)->node_enqueued = 0;
750 return node;
752 else
753 return NULL;
756 /* Set lattice LAT to bottom and return true if it previously was not set as
757 such. */
759 template <typename valtype>
760 inline bool
761 ipcp_lattice<valtype>::set_to_bottom ()
763 bool ret = !bottom;
764 bottom = true;
765 return ret;
768 /* Mark lattice as containing an unknown value and return true if it previously
769 was not marked as such. */
771 template <typename valtype>
772 inline bool
773 ipcp_lattice<valtype>::set_contains_variable ()
775 bool ret = !contains_variable;
776 contains_variable = true;
777 return ret;
780 /* Set all aggegate lattices in PLATS to bottom and return true if they were
781 not previously set as such. */
783 static inline bool
784 set_agg_lats_to_bottom (struct ipcp_param_lattices *plats)
786 bool ret = !plats->aggs_bottom;
787 plats->aggs_bottom = true;
788 return ret;
791 /* Mark all aggegate lattices in PLATS as containing an unknown value and
792 return true if they were not previously marked as such. */
794 static inline bool
795 set_agg_lats_contain_variable (struct ipcp_param_lattices *plats)
797 bool ret = !plats->aggs_contain_variable;
798 plats->aggs_contain_variable = true;
799 return ret;
802 /* Return true if alignment information in the lattice is yet unknown. */
804 bool
805 ipcp_alignment_lattice::top_p () const
807 return !bottom && !not_top;
810 /* Return true if alignment information in the lattice is known to be
811 unusable. */
813 bool
814 ipcp_alignment_lattice::bottom_p () const
816 return bottom;
819 /* Set alignment information in the lattice to bottom. Return true if it
820 previously was in a different state. */
822 bool
823 ipcp_alignment_lattice::set_to_bottom ()
825 if (bottom_p ())
826 return false;
827 bottom = true;
828 return true;
831 /* Meet the current value of the lattice with alignment described by NEW_ALIGN
832 and NEW_MISALIGN, assuming that we know the current value is neither TOP nor
833 BOTTOM. Return true if the value of lattice has changed. */
835 bool
836 ipcp_alignment_lattice::meet_with_1 (unsigned new_align, unsigned new_misalign)
838 gcc_checking_assert (new_align != 0);
839 if (align == new_align && misalign == new_misalign)
840 return false;
842 bool changed = false;
843 if (align > new_align)
845 align = new_align;
846 misalign = misalign % new_align;
847 changed = true;
849 if (misalign != (new_misalign % align))
851 int diff = abs ((int) misalign - (int) (new_misalign % align));
852 align = (unsigned) diff & -diff;
853 if (align)
854 misalign = misalign % align;
855 else
856 set_to_bottom ();
857 changed = true;
859 gcc_checking_assert (bottom_p () || align != 0);
860 return changed;
863 /* Meet the current value of the lattice with alignment described by NEW_ALIGN
864 and NEW_MISALIGN. Return true if the value of lattice has changed. */
866 bool
867 ipcp_alignment_lattice::meet_with (unsigned new_align, unsigned new_misalign)
869 gcc_assert (new_align != 0);
870 if (bottom_p ())
871 return false;
872 if (top_p ())
874 not_top = true;
875 align = new_align;
876 misalign = new_misalign;
877 return true;
879 return meet_with_1 (new_align, new_misalign);
882 /* Meet the current value of the lattice with OTHER, taking into account that
883 OFFSET has been added to the pointer value. Return true if the value of
884 lattice has changed. */
886 bool
887 ipcp_alignment_lattice::meet_with (const ipcp_alignment_lattice &other,
888 HOST_WIDE_INT offset)
890 if (other.bottom_p ())
891 return set_to_bottom ();
892 if (bottom_p () || other.top_p ())
893 return false;
895 unsigned adjusted_misalign = (other.misalign + offset) % other.align;
896 if (top_p ())
898 not_top = true;
899 align = other.align;
900 misalign = adjusted_misalign;
901 return true;
904 return meet_with_1 (other.align, adjusted_misalign);
907 /* Mark bot aggregate and scalar lattices as containing an unknown variable,
908 return true is any of them has not been marked as such so far. */
910 static inline bool
911 set_all_contains_variable (struct ipcp_param_lattices *plats)
913 bool ret;
914 ret = plats->itself.set_contains_variable ();
915 ret |= plats->ctxlat.set_contains_variable ();
916 ret |= set_agg_lats_contain_variable (plats);
917 ret |= plats->alignment.set_to_bottom ();
918 return ret;
921 /* Worker of call_for_symbol_thunks_and_aliases, increment the integer DATA
922 points to by the number of callers to NODE. */
924 static bool
925 count_callers (cgraph_node *node, void *data)
927 int *caller_count = (int *) data;
929 for (cgraph_edge *cs = node->callers; cs; cs = cs->next_caller)
930 /* Local thunks can be handled transparently, but if the thunk can not
931 be optimized out, count it as a real use. */
932 if (!cs->caller->thunk.thunk_p || !cs->caller->local.local)
933 ++*caller_count;
934 return false;
937 /* Worker of call_for_symbol_thunks_and_aliases, it is supposed to be called on
938 the one caller of some other node. Set the caller's corresponding flag. */
940 static bool
941 set_single_call_flag (cgraph_node *node, void *)
943 cgraph_edge *cs = node->callers;
944 /* Local thunks can be handled transparently, skip them. */
945 while (cs && cs->caller->thunk.thunk_p && cs->caller->local.local)
946 cs = cs->next_caller;
947 if (cs)
949 IPA_NODE_REF (cs->caller)->node_calling_single_call = true;
950 return true;
952 return false;
955 /* Initialize ipcp_lattices. */
957 static void
958 initialize_node_lattices (struct cgraph_node *node)
960 struct ipa_node_params *info = IPA_NODE_REF (node);
961 struct cgraph_edge *ie;
962 bool disable = false, variable = false;
963 int i;
965 gcc_checking_assert (node->has_gimple_body_p ());
966 if (cgraph_local_p (node))
968 int caller_count = 0;
969 node->call_for_symbol_thunks_and_aliases (count_callers, &caller_count,
970 true);
971 gcc_checking_assert (caller_count > 0);
972 if (caller_count == 1)
973 node->call_for_symbol_thunks_and_aliases (set_single_call_flag,
974 NULL, true);
976 else
978 /* When cloning is allowed, we can assume that externally visible
979 functions are not called. We will compensate this by cloning
980 later. */
981 if (ipcp_versionable_function_p (node)
982 && ipcp_cloning_candidate_p (node))
983 variable = true;
984 else
985 disable = true;
988 if (disable || variable)
990 for (i = 0; i < ipa_get_param_count (info) ; i++)
992 struct ipcp_param_lattices *plats = ipa_get_parm_lattices (info, i);
993 if (disable)
995 plats->itself.set_to_bottom ();
996 plats->ctxlat.set_to_bottom ();
997 set_agg_lats_to_bottom (plats);
998 plats->alignment.set_to_bottom ();
1000 else
1001 set_all_contains_variable (plats);
1003 if (dump_file && (dump_flags & TDF_DETAILS)
1004 && !node->alias && !node->thunk.thunk_p)
1005 fprintf (dump_file, "Marking all lattices of %s/%i as %s\n",
1006 node->name (), node->order,
1007 disable ? "BOTTOM" : "VARIABLE");
1010 for (ie = node->indirect_calls; ie; ie = ie->next_callee)
1011 if (ie->indirect_info->polymorphic
1012 && ie->indirect_info->param_index >= 0)
1014 gcc_checking_assert (ie->indirect_info->param_index >= 0);
1015 ipa_get_parm_lattices (info,
1016 ie->indirect_info->param_index)->virt_call = 1;
1020 /* Return the result of a (possibly arithmetic) pass through jump function
1021 JFUNC on the constant value INPUT. Return NULL_TREE if that cannot be
1022 determined or be considered an interprocedural invariant. */
1024 static tree
1025 ipa_get_jf_pass_through_result (struct ipa_jump_func *jfunc, tree input)
1027 tree restype, res;
1029 gcc_checking_assert (is_gimple_ip_invariant (input));
1030 if (ipa_get_jf_pass_through_operation (jfunc) == NOP_EXPR)
1031 return input;
1033 if (TREE_CODE_CLASS (ipa_get_jf_pass_through_operation (jfunc))
1034 == tcc_comparison)
1035 restype = boolean_type_node;
1036 else
1037 restype = TREE_TYPE (input);
1038 res = fold_binary (ipa_get_jf_pass_through_operation (jfunc), restype,
1039 input, ipa_get_jf_pass_through_operand (jfunc));
1041 if (res && !is_gimple_ip_invariant (res))
1042 return NULL_TREE;
1044 return res;
1047 /* Return the result of an ancestor jump function JFUNC on the constant value
1048 INPUT. Return NULL_TREE if that cannot be determined. */
1050 static tree
1051 ipa_get_jf_ancestor_result (struct ipa_jump_func *jfunc, tree input)
1053 gcc_checking_assert (TREE_CODE (input) != TREE_BINFO);
1054 if (TREE_CODE (input) == ADDR_EXPR)
1056 tree t = TREE_OPERAND (input, 0);
1057 t = build_ref_for_offset (EXPR_LOCATION (t), t,
1058 ipa_get_jf_ancestor_offset (jfunc), false,
1059 ptr_type_node, NULL, false);
1060 return build_fold_addr_expr (t);
1062 else
1063 return NULL_TREE;
1066 /* Determine whether JFUNC evaluates to a single known constant value and if
1067 so, return it. Otherwise return NULL. INFO describes the caller node or
1068 the one it is inlined to, so that pass-through jump functions can be
1069 evaluated. */
1071 tree
1072 ipa_value_from_jfunc (struct ipa_node_params *info, struct ipa_jump_func *jfunc)
1074 if (jfunc->type == IPA_JF_CONST)
1075 return ipa_get_jf_constant (jfunc);
1076 else if (jfunc->type == IPA_JF_PASS_THROUGH
1077 || jfunc->type == IPA_JF_ANCESTOR)
1079 tree input;
1080 int idx;
1082 if (jfunc->type == IPA_JF_PASS_THROUGH)
1083 idx = ipa_get_jf_pass_through_formal_id (jfunc);
1084 else
1085 idx = ipa_get_jf_ancestor_formal_id (jfunc);
1087 if (info->ipcp_orig_node)
1088 input = info->known_csts[idx];
1089 else
1091 ipcp_lattice<tree> *lat;
1093 if (!info->lattices
1094 || idx >= ipa_get_param_count (info))
1095 return NULL_TREE;
1096 lat = ipa_get_scalar_lat (info, idx);
1097 if (!lat->is_single_const ())
1098 return NULL_TREE;
1099 input = lat->values->value;
1102 if (!input)
1103 return NULL_TREE;
1105 if (jfunc->type == IPA_JF_PASS_THROUGH)
1106 return ipa_get_jf_pass_through_result (jfunc, input);
1107 else
1108 return ipa_get_jf_ancestor_result (jfunc, input);
1110 else
1111 return NULL_TREE;
1114 /* Determie whether JFUNC evaluates to single known polymorphic context, given
1115 that INFO describes the caller node or the one it is inlined to, CS is the
1116 call graph edge corresponding to JFUNC and CSIDX index of the described
1117 parameter. */
1119 ipa_polymorphic_call_context
1120 ipa_context_from_jfunc (ipa_node_params *info, cgraph_edge *cs, int csidx,
1121 ipa_jump_func *jfunc)
1123 ipa_edge_args *args = IPA_EDGE_REF (cs);
1124 ipa_polymorphic_call_context ctx;
1125 ipa_polymorphic_call_context *edge_ctx
1126 = cs ? ipa_get_ith_polymorhic_call_context (args, csidx) : NULL;
1128 if (edge_ctx && !edge_ctx->useless_p ())
1129 ctx = *edge_ctx;
1131 if (jfunc->type == IPA_JF_PASS_THROUGH
1132 || jfunc->type == IPA_JF_ANCESTOR)
1134 ipa_polymorphic_call_context srcctx;
1135 int srcidx;
1136 bool type_preserved = true;
1137 if (jfunc->type == IPA_JF_PASS_THROUGH)
1139 if (ipa_get_jf_pass_through_operation (jfunc) != NOP_EXPR)
1140 return ctx;
1141 type_preserved = ipa_get_jf_pass_through_type_preserved (jfunc);
1142 srcidx = ipa_get_jf_pass_through_formal_id (jfunc);
1144 else
1146 type_preserved = ipa_get_jf_ancestor_type_preserved (jfunc);
1147 srcidx = ipa_get_jf_ancestor_formal_id (jfunc);
1149 if (info->ipcp_orig_node)
1151 if (info->known_contexts.exists ())
1152 srcctx = info->known_contexts[srcidx];
1154 else
1156 if (!info->lattices
1157 || srcidx >= ipa_get_param_count (info))
1158 return ctx;
1159 ipcp_lattice<ipa_polymorphic_call_context> *lat;
1160 lat = ipa_get_poly_ctx_lat (info, srcidx);
1161 if (!lat->is_single_const ())
1162 return ctx;
1163 srcctx = lat->values->value;
1165 if (srcctx.useless_p ())
1166 return ctx;
1167 if (jfunc->type == IPA_JF_ANCESTOR)
1168 srcctx.offset_by (ipa_get_jf_ancestor_offset (jfunc));
1169 if (!type_preserved)
1170 srcctx.possible_dynamic_type_change (cs->in_polymorphic_cdtor);
1171 srcctx.combine_with (ctx);
1172 return srcctx;
1175 return ctx;
1178 /* If checking is enabled, verify that no lattice is in the TOP state, i.e. not
1179 bottom, not containing a variable component and without any known value at
1180 the same time. */
1182 DEBUG_FUNCTION void
1183 ipcp_verify_propagated_values (void)
1185 struct cgraph_node *node;
1187 FOR_EACH_FUNCTION_WITH_GIMPLE_BODY (node)
1189 struct ipa_node_params *info = IPA_NODE_REF (node);
1190 int i, count = ipa_get_param_count (info);
1192 for (i = 0; i < count; i++)
1194 ipcp_lattice<tree> *lat = ipa_get_scalar_lat (info, i);
1196 if (!lat->bottom
1197 && !lat->contains_variable
1198 && lat->values_count == 0)
1200 if (dump_file)
1202 symtab_node::dump_table (dump_file);
1203 fprintf (dump_file, "\nIPA lattices after constant "
1204 "propagation, before gcc_unreachable:\n");
1205 print_all_lattices (dump_file, true, false);
1208 gcc_unreachable ();
1214 /* Return true iff X and Y should be considered equal values by IPA-CP. */
1216 static bool
1217 values_equal_for_ipcp_p (tree x, tree y)
1219 gcc_checking_assert (x != NULL_TREE && y != NULL_TREE);
1221 if (x == y)
1222 return true;
1224 if (TREE_CODE (x) == ADDR_EXPR
1225 && TREE_CODE (y) == ADDR_EXPR
1226 && TREE_CODE (TREE_OPERAND (x, 0)) == CONST_DECL
1227 && TREE_CODE (TREE_OPERAND (y, 0)) == CONST_DECL)
1228 return operand_equal_p (DECL_INITIAL (TREE_OPERAND (x, 0)),
1229 DECL_INITIAL (TREE_OPERAND (y, 0)), 0);
1230 else
1231 return operand_equal_p (x, y, 0);
1234 /* Return true iff X and Y should be considered equal contexts by IPA-CP. */
1236 static bool
1237 values_equal_for_ipcp_p (ipa_polymorphic_call_context x,
1238 ipa_polymorphic_call_context y)
1240 return x.equal_to (y);
1244 /* Add a new value source to the value represented by THIS, marking that a
1245 value comes from edge CS and (if the underlying jump function is a
1246 pass-through or an ancestor one) from a caller value SRC_VAL of a caller
1247 parameter described by SRC_INDEX. OFFSET is negative if the source was the
1248 scalar value of the parameter itself or the offset within an aggregate. */
1250 template <typename valtype>
1251 void
1252 ipcp_value<valtype>::add_source (cgraph_edge *cs, ipcp_value *src_val,
1253 int src_idx, HOST_WIDE_INT offset)
1255 ipcp_value_source<valtype> *src;
1257 src = new (ipcp_sources_pool.allocate ()) ipcp_value_source<valtype>;
1258 src->offset = offset;
1259 src->cs = cs;
1260 src->val = src_val;
1261 src->index = src_idx;
1263 src->next = sources;
1264 sources = src;
1267 /* Allocate a new ipcp_value holding a tree constant, initialize its value to
1268 SOURCE and clear all other fields. */
1270 static ipcp_value<tree> *
1271 allocate_and_init_ipcp_value (tree source)
1273 ipcp_value<tree> *val;
1275 val = ipcp_cst_values_pool.allocate ();
1276 memset (val, 0, sizeof (*val));
1277 val->value = source;
1278 return val;
1281 /* Allocate a new ipcp_value holding a polymorphic context, initialize its
1282 value to SOURCE and clear all other fields. */
1284 static ipcp_value<ipa_polymorphic_call_context> *
1285 allocate_and_init_ipcp_value (ipa_polymorphic_call_context source)
1287 ipcp_value<ipa_polymorphic_call_context> *val;
1289 // TODO
1290 val = ipcp_poly_ctx_values_pool.allocate ();
1291 memset (val, 0, sizeof (*val));
1292 val->value = source;
1293 return val;
1296 /* Try to add NEWVAL to LAT, potentially creating a new ipcp_value for it. CS,
1297 SRC_VAL SRC_INDEX and OFFSET are meant for add_source and have the same
1298 meaning. OFFSET -1 means the source is scalar and not a part of an
1299 aggregate. */
1301 template <typename valtype>
1302 bool
1303 ipcp_lattice<valtype>::add_value (valtype newval, cgraph_edge *cs,
1304 ipcp_value<valtype> *src_val,
1305 int src_idx, HOST_WIDE_INT offset)
1307 ipcp_value<valtype> *val;
1309 if (bottom)
1310 return false;
1312 for (val = values; val; val = val->next)
1313 if (values_equal_for_ipcp_p (val->value, newval))
1315 if (ipa_edge_within_scc (cs))
1317 ipcp_value_source<valtype> *s;
1318 for (s = val->sources; s ; s = s->next)
1319 if (s->cs == cs)
1320 break;
1321 if (s)
1322 return false;
1325 val->add_source (cs, src_val, src_idx, offset);
1326 return false;
1329 if (values_count == PARAM_VALUE (PARAM_IPA_CP_VALUE_LIST_SIZE))
1331 /* We can only free sources, not the values themselves, because sources
1332 of other values in this SCC might point to them. */
1333 for (val = values; val; val = val->next)
1335 while (val->sources)
1337 ipcp_value_source<valtype> *src = val->sources;
1338 val->sources = src->next;
1339 ipcp_sources_pool.remove ((ipcp_value_source<tree>*)src);
1343 values = NULL;
1344 return set_to_bottom ();
1347 values_count++;
1348 val = allocate_and_init_ipcp_value (newval);
1349 val->add_source (cs, src_val, src_idx, offset);
1350 val->next = values;
1351 values = val;
1352 return true;
1355 /* Propagate values through a pass-through jump function JFUNC associated with
1356 edge CS, taking values from SRC_LAT and putting them into DEST_LAT. SRC_IDX
1357 is the index of the source parameter. */
1359 static bool
1360 propagate_vals_accross_pass_through (cgraph_edge *cs,
1361 ipa_jump_func *jfunc,
1362 ipcp_lattice<tree> *src_lat,
1363 ipcp_lattice<tree> *dest_lat,
1364 int src_idx)
1366 ipcp_value<tree> *src_val;
1367 bool ret = false;
1369 /* Do not create new values when propagating within an SCC because if there
1370 are arithmetic functions with circular dependencies, there is infinite
1371 number of them and we would just make lattices bottom. */
1372 if ((ipa_get_jf_pass_through_operation (jfunc) != NOP_EXPR)
1373 && ipa_edge_within_scc (cs))
1374 ret = dest_lat->set_contains_variable ();
1375 else
1376 for (src_val = src_lat->values; src_val; src_val = src_val->next)
1378 tree cstval = ipa_get_jf_pass_through_result (jfunc, src_val->value);
1380 if (cstval)
1381 ret |= dest_lat->add_value (cstval, cs, src_val, src_idx);
1382 else
1383 ret |= dest_lat->set_contains_variable ();
1386 return ret;
1389 /* Propagate values through an ancestor jump function JFUNC associated with
1390 edge CS, taking values from SRC_LAT and putting them into DEST_LAT. SRC_IDX
1391 is the index of the source parameter. */
1393 static bool
1394 propagate_vals_accross_ancestor (struct cgraph_edge *cs,
1395 struct ipa_jump_func *jfunc,
1396 ipcp_lattice<tree> *src_lat,
1397 ipcp_lattice<tree> *dest_lat,
1398 int src_idx)
1400 ipcp_value<tree> *src_val;
1401 bool ret = false;
1403 if (ipa_edge_within_scc (cs))
1404 return dest_lat->set_contains_variable ();
1406 for (src_val = src_lat->values; src_val; src_val = src_val->next)
1408 tree t = ipa_get_jf_ancestor_result (jfunc, src_val->value);
1410 if (t)
1411 ret |= dest_lat->add_value (t, cs, src_val, src_idx);
1412 else
1413 ret |= dest_lat->set_contains_variable ();
1416 return ret;
1419 /* Propagate scalar values across jump function JFUNC that is associated with
1420 edge CS and put the values into DEST_LAT. */
1422 static bool
1423 propagate_scalar_accross_jump_function (struct cgraph_edge *cs,
1424 struct ipa_jump_func *jfunc,
1425 ipcp_lattice<tree> *dest_lat)
1427 if (dest_lat->bottom)
1428 return false;
1430 if (jfunc->type == IPA_JF_CONST)
1432 tree val = ipa_get_jf_constant (jfunc);
1433 return dest_lat->add_value (val, cs, NULL, 0);
1435 else if (jfunc->type == IPA_JF_PASS_THROUGH
1436 || jfunc->type == IPA_JF_ANCESTOR)
1438 struct ipa_node_params *caller_info = IPA_NODE_REF (cs->caller);
1439 ipcp_lattice<tree> *src_lat;
1440 int src_idx;
1441 bool ret;
1443 if (jfunc->type == IPA_JF_PASS_THROUGH)
1444 src_idx = ipa_get_jf_pass_through_formal_id (jfunc);
1445 else
1446 src_idx = ipa_get_jf_ancestor_formal_id (jfunc);
1448 src_lat = ipa_get_scalar_lat (caller_info, src_idx);
1449 if (src_lat->bottom)
1450 return dest_lat->set_contains_variable ();
1452 /* If we would need to clone the caller and cannot, do not propagate. */
1453 if (!ipcp_versionable_function_p (cs->caller)
1454 && (src_lat->contains_variable
1455 || (src_lat->values_count > 1)))
1456 return dest_lat->set_contains_variable ();
1458 if (jfunc->type == IPA_JF_PASS_THROUGH)
1459 ret = propagate_vals_accross_pass_through (cs, jfunc, src_lat,
1460 dest_lat, src_idx);
1461 else
1462 ret = propagate_vals_accross_ancestor (cs, jfunc, src_lat, dest_lat,
1463 src_idx);
1465 if (src_lat->contains_variable)
1466 ret |= dest_lat->set_contains_variable ();
1468 return ret;
1471 /* TODO: We currently do not handle member method pointers in IPA-CP (we only
1472 use it for indirect inlining), we should propagate them too. */
1473 return dest_lat->set_contains_variable ();
1476 /* Propagate scalar values across jump function JFUNC that is associated with
1477 edge CS and describes argument IDX and put the values into DEST_LAT. */
1479 static bool
1480 propagate_context_accross_jump_function (cgraph_edge *cs,
1481 ipa_jump_func *jfunc, int idx,
1482 ipcp_lattice<ipa_polymorphic_call_context> *dest_lat)
1484 ipa_edge_args *args = IPA_EDGE_REF (cs);
1485 if (dest_lat->bottom)
1486 return false;
1487 bool ret = false;
1488 bool added_sth = false;
1489 bool type_preserved = true;
1491 ipa_polymorphic_call_context edge_ctx, *edge_ctx_ptr
1492 = ipa_get_ith_polymorhic_call_context (args, idx);
1494 if (edge_ctx_ptr)
1495 edge_ctx = *edge_ctx_ptr;
1497 if (jfunc->type == IPA_JF_PASS_THROUGH
1498 || jfunc->type == IPA_JF_ANCESTOR)
1500 struct ipa_node_params *caller_info = IPA_NODE_REF (cs->caller);
1501 int src_idx;
1502 ipcp_lattice<ipa_polymorphic_call_context> *src_lat;
1504 /* TODO: Once we figure out how to propagate speculations, it will
1505 probably be a good idea to switch to speculation if type_preserved is
1506 not set instead of punting. */
1507 if (jfunc->type == IPA_JF_PASS_THROUGH)
1509 if (ipa_get_jf_pass_through_operation (jfunc) != NOP_EXPR)
1510 goto prop_fail;
1511 type_preserved = ipa_get_jf_pass_through_type_preserved (jfunc);
1512 src_idx = ipa_get_jf_pass_through_formal_id (jfunc);
1514 else
1516 type_preserved = ipa_get_jf_ancestor_type_preserved (jfunc);
1517 src_idx = ipa_get_jf_ancestor_formal_id (jfunc);
1520 src_lat = ipa_get_poly_ctx_lat (caller_info, src_idx);
1521 /* If we would need to clone the caller and cannot, do not propagate. */
1522 if (!ipcp_versionable_function_p (cs->caller)
1523 && (src_lat->contains_variable
1524 || (src_lat->values_count > 1)))
1525 goto prop_fail;
1527 ipcp_value<ipa_polymorphic_call_context> *src_val;
1528 for (src_val = src_lat->values; src_val; src_val = src_val->next)
1530 ipa_polymorphic_call_context cur = src_val->value;
1532 if (!type_preserved)
1533 cur.possible_dynamic_type_change (cs->in_polymorphic_cdtor);
1534 if (jfunc->type == IPA_JF_ANCESTOR)
1535 cur.offset_by (ipa_get_jf_ancestor_offset (jfunc));
1536 /* TODO: In cases we know how the context is going to be used,
1537 we can improve the result by passing proper OTR_TYPE. */
1538 cur.combine_with (edge_ctx);
1539 if (!cur.useless_p ())
1541 if (src_lat->contains_variable
1542 && !edge_ctx.equal_to (cur))
1543 ret |= dest_lat->set_contains_variable ();
1544 ret |= dest_lat->add_value (cur, cs, src_val, src_idx);
1545 added_sth = true;
1551 prop_fail:
1552 if (!added_sth)
1554 if (!edge_ctx.useless_p ())
1555 ret |= dest_lat->add_value (edge_ctx, cs);
1556 else
1557 ret |= dest_lat->set_contains_variable ();
1560 return ret;
1563 /* Propagate alignments across jump function JFUNC that is associated with
1564 edge CS and update DEST_LAT accordingly. */
1566 static bool
1567 propagate_alignment_accross_jump_function (cgraph_edge *cs,
1568 ipa_jump_func *jfunc,
1569 ipcp_alignment_lattice *dest_lat)
1571 if (dest_lat->bottom_p ())
1572 return false;
1574 if (jfunc->type == IPA_JF_PASS_THROUGH
1575 || jfunc->type == IPA_JF_ANCESTOR)
1577 struct ipa_node_params *caller_info = IPA_NODE_REF (cs->caller);
1578 HOST_WIDE_INT offset = 0;
1579 int src_idx;
1581 if (jfunc->type == IPA_JF_PASS_THROUGH)
1583 enum tree_code op = ipa_get_jf_pass_through_operation (jfunc);
1584 if (op != NOP_EXPR)
1586 if (op != POINTER_PLUS_EXPR
1587 && op != PLUS_EXPR)
1588 return dest_lat->set_to_bottom ();
1589 tree operand = ipa_get_jf_pass_through_operand (jfunc);
1590 if (!tree_fits_shwi_p (operand))
1591 return dest_lat->set_to_bottom ();
1592 offset = tree_to_shwi (operand);
1594 src_idx = ipa_get_jf_pass_through_formal_id (jfunc);
1596 else
1598 src_idx = ipa_get_jf_ancestor_formal_id (jfunc);
1599 offset = ipa_get_jf_ancestor_offset (jfunc) / BITS_PER_UNIT;
1602 struct ipcp_param_lattices *src_lats;
1603 src_lats = ipa_get_parm_lattices (caller_info, src_idx);
1604 return dest_lat->meet_with (src_lats->alignment, offset);
1606 else
1608 if (jfunc->alignment.known)
1609 return dest_lat->meet_with (jfunc->alignment.align,
1610 jfunc->alignment.misalign);
1611 else
1612 return dest_lat->set_to_bottom ();
1616 /* If DEST_PLATS already has aggregate items, check that aggs_by_ref matches
1617 NEW_AGGS_BY_REF and if not, mark all aggs as bottoms and return true (in all
1618 other cases, return false). If there are no aggregate items, set
1619 aggs_by_ref to NEW_AGGS_BY_REF. */
1621 static bool
1622 set_check_aggs_by_ref (struct ipcp_param_lattices *dest_plats,
1623 bool new_aggs_by_ref)
1625 if (dest_plats->aggs)
1627 if (dest_plats->aggs_by_ref != new_aggs_by_ref)
1629 set_agg_lats_to_bottom (dest_plats);
1630 return true;
1633 else
1634 dest_plats->aggs_by_ref = new_aggs_by_ref;
1635 return false;
1638 /* Walk aggregate lattices in DEST_PLATS from ***AGLAT on, until ***aglat is an
1639 already existing lattice for the given OFFSET and SIZE, marking all skipped
1640 lattices as containing variable and checking for overlaps. If there is no
1641 already existing lattice for the OFFSET and VAL_SIZE, create one, initialize
1642 it with offset, size and contains_variable to PRE_EXISTING, and return true,
1643 unless there are too many already. If there are two many, return false. If
1644 there are overlaps turn whole DEST_PLATS to bottom and return false. If any
1645 skipped lattices were newly marked as containing variable, set *CHANGE to
1646 true. */
1648 static bool
1649 merge_agg_lats_step (struct ipcp_param_lattices *dest_plats,
1650 HOST_WIDE_INT offset, HOST_WIDE_INT val_size,
1651 struct ipcp_agg_lattice ***aglat,
1652 bool pre_existing, bool *change)
1654 gcc_checking_assert (offset >= 0);
1656 while (**aglat && (**aglat)->offset < offset)
1658 if ((**aglat)->offset + (**aglat)->size > offset)
1660 set_agg_lats_to_bottom (dest_plats);
1661 return false;
1663 *change |= (**aglat)->set_contains_variable ();
1664 *aglat = &(**aglat)->next;
1667 if (**aglat && (**aglat)->offset == offset)
1669 if ((**aglat)->size != val_size
1670 || ((**aglat)->next
1671 && (**aglat)->next->offset < offset + val_size))
1673 set_agg_lats_to_bottom (dest_plats);
1674 return false;
1676 gcc_checking_assert (!(**aglat)->next
1677 || (**aglat)->next->offset >= offset + val_size);
1678 return true;
1680 else
1682 struct ipcp_agg_lattice *new_al;
1684 if (**aglat && (**aglat)->offset < offset + val_size)
1686 set_agg_lats_to_bottom (dest_plats);
1687 return false;
1689 if (dest_plats->aggs_count == PARAM_VALUE (PARAM_IPA_MAX_AGG_ITEMS))
1690 return false;
1691 dest_plats->aggs_count++;
1692 new_al = ipcp_agg_lattice_pool.allocate ();
1693 memset (new_al, 0, sizeof (*new_al));
1695 new_al->offset = offset;
1696 new_al->size = val_size;
1697 new_al->contains_variable = pre_existing;
1699 new_al->next = **aglat;
1700 **aglat = new_al;
1701 return true;
1705 /* Set all AGLAT and all other aggregate lattices reachable by next pointers as
1706 containing an unknown value. */
1708 static bool
1709 set_chain_of_aglats_contains_variable (struct ipcp_agg_lattice *aglat)
1711 bool ret = false;
1712 while (aglat)
1714 ret |= aglat->set_contains_variable ();
1715 aglat = aglat->next;
1717 return ret;
1720 /* Merge existing aggregate lattices in SRC_PLATS to DEST_PLATS, subtracting
1721 DELTA_OFFSET. CS is the call graph edge and SRC_IDX the index of the source
1722 parameter used for lattice value sources. Return true if DEST_PLATS changed
1723 in any way. */
1725 static bool
1726 merge_aggregate_lattices (struct cgraph_edge *cs,
1727 struct ipcp_param_lattices *dest_plats,
1728 struct ipcp_param_lattices *src_plats,
1729 int src_idx, HOST_WIDE_INT offset_delta)
1731 bool pre_existing = dest_plats->aggs != NULL;
1732 struct ipcp_agg_lattice **dst_aglat;
1733 bool ret = false;
1735 if (set_check_aggs_by_ref (dest_plats, src_plats->aggs_by_ref))
1736 return true;
1737 if (src_plats->aggs_bottom)
1738 return set_agg_lats_contain_variable (dest_plats);
1739 if (src_plats->aggs_contain_variable)
1740 ret |= set_agg_lats_contain_variable (dest_plats);
1741 dst_aglat = &dest_plats->aggs;
1743 for (struct ipcp_agg_lattice *src_aglat = src_plats->aggs;
1744 src_aglat;
1745 src_aglat = src_aglat->next)
1747 HOST_WIDE_INT new_offset = src_aglat->offset - offset_delta;
1749 if (new_offset < 0)
1750 continue;
1751 if (merge_agg_lats_step (dest_plats, new_offset, src_aglat->size,
1752 &dst_aglat, pre_existing, &ret))
1754 struct ipcp_agg_lattice *new_al = *dst_aglat;
1756 dst_aglat = &(*dst_aglat)->next;
1757 if (src_aglat->bottom)
1759 ret |= new_al->set_contains_variable ();
1760 continue;
1762 if (src_aglat->contains_variable)
1763 ret |= new_al->set_contains_variable ();
1764 for (ipcp_value<tree> *val = src_aglat->values;
1765 val;
1766 val = val->next)
1767 ret |= new_al->add_value (val->value, cs, val, src_idx,
1768 src_aglat->offset);
1770 else if (dest_plats->aggs_bottom)
1771 return true;
1773 ret |= set_chain_of_aglats_contains_variable (*dst_aglat);
1774 return ret;
1777 /* Determine whether there is anything to propagate FROM SRC_PLATS through a
1778 pass-through JFUNC and if so, whether it has conform and conforms to the
1779 rules about propagating values passed by reference. */
1781 static bool
1782 agg_pass_through_permissible_p (struct ipcp_param_lattices *src_plats,
1783 struct ipa_jump_func *jfunc)
1785 return src_plats->aggs
1786 && (!src_plats->aggs_by_ref
1787 || ipa_get_jf_pass_through_agg_preserved (jfunc));
1790 /* Propagate scalar values across jump function JFUNC that is associated with
1791 edge CS and put the values into DEST_LAT. */
1793 static bool
1794 propagate_aggs_accross_jump_function (struct cgraph_edge *cs,
1795 struct ipa_jump_func *jfunc,
1796 struct ipcp_param_lattices *dest_plats)
1798 bool ret = false;
1800 if (dest_plats->aggs_bottom)
1801 return false;
1803 if (jfunc->type == IPA_JF_PASS_THROUGH
1804 && ipa_get_jf_pass_through_operation (jfunc) == NOP_EXPR)
1806 struct ipa_node_params *caller_info = IPA_NODE_REF (cs->caller);
1807 int src_idx = ipa_get_jf_pass_through_formal_id (jfunc);
1808 struct ipcp_param_lattices *src_plats;
1810 src_plats = ipa_get_parm_lattices (caller_info, src_idx);
1811 if (agg_pass_through_permissible_p (src_plats, jfunc))
1813 /* Currently we do not produce clobber aggregate jump
1814 functions, replace with merging when we do. */
1815 gcc_assert (!jfunc->agg.items);
1816 ret |= merge_aggregate_lattices (cs, dest_plats, src_plats,
1817 src_idx, 0);
1819 else
1820 ret |= set_agg_lats_contain_variable (dest_plats);
1822 else if (jfunc->type == IPA_JF_ANCESTOR
1823 && ipa_get_jf_ancestor_agg_preserved (jfunc))
1825 struct ipa_node_params *caller_info = IPA_NODE_REF (cs->caller);
1826 int src_idx = ipa_get_jf_ancestor_formal_id (jfunc);
1827 struct ipcp_param_lattices *src_plats;
1829 src_plats = ipa_get_parm_lattices (caller_info, src_idx);
1830 if (src_plats->aggs && src_plats->aggs_by_ref)
1832 /* Currently we do not produce clobber aggregate jump
1833 functions, replace with merging when we do. */
1834 gcc_assert (!jfunc->agg.items);
1835 ret |= merge_aggregate_lattices (cs, dest_plats, src_plats, src_idx,
1836 ipa_get_jf_ancestor_offset (jfunc));
1838 else if (!src_plats->aggs_by_ref)
1839 ret |= set_agg_lats_to_bottom (dest_plats);
1840 else
1841 ret |= set_agg_lats_contain_variable (dest_plats);
1843 else if (jfunc->agg.items)
1845 bool pre_existing = dest_plats->aggs != NULL;
1846 struct ipcp_agg_lattice **aglat = &dest_plats->aggs;
1847 struct ipa_agg_jf_item *item;
1848 int i;
1850 if (set_check_aggs_by_ref (dest_plats, jfunc->agg.by_ref))
1851 return true;
1853 FOR_EACH_VEC_ELT (*jfunc->agg.items, i, item)
1855 HOST_WIDE_INT val_size;
1857 if (item->offset < 0)
1858 continue;
1859 gcc_checking_assert (is_gimple_ip_invariant (item->value));
1860 val_size = tree_to_uhwi (TYPE_SIZE (TREE_TYPE (item->value)));
1862 if (merge_agg_lats_step (dest_plats, item->offset, val_size,
1863 &aglat, pre_existing, &ret))
1865 ret |= (*aglat)->add_value (item->value, cs, NULL, 0, 0);
1866 aglat = &(*aglat)->next;
1868 else if (dest_plats->aggs_bottom)
1869 return true;
1872 ret |= set_chain_of_aglats_contains_variable (*aglat);
1874 else
1875 ret |= set_agg_lats_contain_variable (dest_plats);
1877 return ret;
1880 /* Propagate constants from the caller to the callee of CS. INFO describes the
1881 caller. */
1883 static bool
1884 propagate_constants_accross_call (struct cgraph_edge *cs)
1886 struct ipa_node_params *callee_info;
1887 enum availability availability;
1888 struct cgraph_node *callee, *alias_or_thunk;
1889 struct ipa_edge_args *args;
1890 bool ret = false;
1891 int i, args_count, parms_count;
1893 callee = cs->callee->function_symbol (&availability);
1894 if (!callee->definition)
1895 return false;
1896 gcc_checking_assert (callee->has_gimple_body_p ());
1897 callee_info = IPA_NODE_REF (callee);
1899 args = IPA_EDGE_REF (cs);
1900 args_count = ipa_get_cs_argument_count (args);
1901 parms_count = ipa_get_param_count (callee_info);
1902 if (parms_count == 0)
1903 return false;
1905 /* No propagation through instrumentation thunks is available yet.
1906 It should be possible with proper mapping of call args and
1907 instrumented callee params in the propagation loop below. But
1908 this case mostly occurs when legacy code calls instrumented code
1909 and it is not a primary target for optimizations.
1910 We detect instrumentation thunks in aliases and thunks chain by
1911 checking instrumentation_clone flag for chain source and target.
1912 Going through instrumentation thunks we always have it changed
1913 from 0 to 1 and all other nodes do not change it. */
1914 if (!cs->callee->instrumentation_clone
1915 && callee->instrumentation_clone)
1917 for (i = 0; i < parms_count; i++)
1918 ret |= set_all_contains_variable (ipa_get_parm_lattices (callee_info,
1919 i));
1920 return ret;
1923 /* If this call goes through a thunk we must not propagate to the first (0th)
1924 parameter. However, we might need to uncover a thunk from below a series
1925 of aliases first. */
1926 alias_or_thunk = cs->callee;
1927 while (alias_or_thunk->alias)
1928 alias_or_thunk = alias_or_thunk->get_alias_target ();
1929 if (alias_or_thunk->thunk.thunk_p)
1931 ret |= set_all_contains_variable (ipa_get_parm_lattices (callee_info,
1932 0));
1933 i = 1;
1935 else
1936 i = 0;
1938 for (; (i < args_count) && (i < parms_count); i++)
1940 struct ipa_jump_func *jump_func = ipa_get_ith_jump_func (args, i);
1941 struct ipcp_param_lattices *dest_plats;
1943 dest_plats = ipa_get_parm_lattices (callee_info, i);
1944 if (availability == AVAIL_INTERPOSABLE)
1945 ret |= set_all_contains_variable (dest_plats);
1946 else
1948 ret |= propagate_scalar_accross_jump_function (cs, jump_func,
1949 &dest_plats->itself);
1950 ret |= propagate_context_accross_jump_function (cs, jump_func, i,
1951 &dest_plats->ctxlat);
1952 ret |= propagate_alignment_accross_jump_function (cs, jump_func,
1953 &dest_plats->alignment);
1954 ret |= propagate_aggs_accross_jump_function (cs, jump_func,
1955 dest_plats);
1958 for (; i < parms_count; i++)
1959 ret |= set_all_contains_variable (ipa_get_parm_lattices (callee_info, i));
1961 return ret;
1964 /* If an indirect edge IE can be turned into a direct one based on KNOWN_VALS
1965 KNOWN_CONTEXTS, KNOWN_AGGS or AGG_REPS return the destination. The latter
1966 three can be NULL. If AGG_REPS is not NULL, KNOWN_AGGS is ignored. */
1968 static tree
1969 ipa_get_indirect_edge_target_1 (struct cgraph_edge *ie,
1970 vec<tree> known_csts,
1971 vec<ipa_polymorphic_call_context> known_contexts,
1972 vec<ipa_agg_jump_function_p> known_aggs,
1973 struct ipa_agg_replacement_value *agg_reps,
1974 bool *speculative)
1976 int param_index = ie->indirect_info->param_index;
1977 HOST_WIDE_INT anc_offset;
1978 tree t;
1979 tree target = NULL;
1981 *speculative = false;
1983 if (param_index == -1
1984 || known_csts.length () <= (unsigned int) param_index)
1985 return NULL_TREE;
1987 if (!ie->indirect_info->polymorphic)
1989 tree t;
1991 if (ie->indirect_info->agg_contents)
1993 if (agg_reps)
1995 t = NULL;
1996 while (agg_reps)
1998 if (agg_reps->index == param_index
1999 && agg_reps->offset == ie->indirect_info->offset
2000 && agg_reps->by_ref == ie->indirect_info->by_ref)
2002 t = agg_reps->value;
2003 break;
2005 agg_reps = agg_reps->next;
2008 else if (known_aggs.length () > (unsigned int) param_index)
2010 struct ipa_agg_jump_function *agg;
2011 agg = known_aggs[param_index];
2012 t = ipa_find_agg_cst_for_param (agg, ie->indirect_info->offset,
2013 ie->indirect_info->by_ref);
2015 else
2016 t = NULL;
2018 else
2019 t = known_csts[param_index];
2021 if (t &&
2022 TREE_CODE (t) == ADDR_EXPR
2023 && TREE_CODE (TREE_OPERAND (t, 0)) == FUNCTION_DECL)
2024 return TREE_OPERAND (t, 0);
2025 else
2026 return NULL_TREE;
2029 if (!opt_for_fn (ie->caller->decl, flag_devirtualize))
2030 return NULL_TREE;
2032 gcc_assert (!ie->indirect_info->agg_contents);
2033 anc_offset = ie->indirect_info->offset;
2035 t = NULL;
2037 /* Try to work out value of virtual table pointer value in replacemnets. */
2038 if (!t && agg_reps && !ie->indirect_info->by_ref)
2040 while (agg_reps)
2042 if (agg_reps->index == param_index
2043 && agg_reps->offset == ie->indirect_info->offset
2044 && agg_reps->by_ref)
2046 t = agg_reps->value;
2047 break;
2049 agg_reps = agg_reps->next;
2053 /* Try to work out value of virtual table pointer value in known
2054 aggregate values. */
2055 if (!t && known_aggs.length () > (unsigned int) param_index
2056 && !ie->indirect_info->by_ref)
2058 struct ipa_agg_jump_function *agg;
2059 agg = known_aggs[param_index];
2060 t = ipa_find_agg_cst_for_param (agg, ie->indirect_info->offset,
2061 true);
2064 /* If we found the virtual table pointer, lookup the target. */
2065 if (t)
2067 tree vtable;
2068 unsigned HOST_WIDE_INT offset;
2069 if (vtable_pointer_value_to_vtable (t, &vtable, &offset))
2071 target = gimple_get_virt_method_for_vtable (ie->indirect_info->otr_token,
2072 vtable, offset);
2073 if (target)
2075 if ((TREE_CODE (TREE_TYPE (target)) == FUNCTION_TYPE
2076 && DECL_FUNCTION_CODE (target) == BUILT_IN_UNREACHABLE)
2077 || !possible_polymorphic_call_target_p
2078 (ie, cgraph_node::get (target)))
2079 target = ipa_impossible_devirt_target (ie, target);
2080 *speculative = ie->indirect_info->vptr_changed;
2081 if (!*speculative)
2082 return target;
2087 /* Do we know the constant value of pointer? */
2088 if (!t)
2089 t = known_csts[param_index];
2091 gcc_checking_assert (!t || TREE_CODE (t) != TREE_BINFO);
2093 ipa_polymorphic_call_context context;
2094 if (known_contexts.length () > (unsigned int) param_index)
2096 context = known_contexts[param_index];
2097 context.offset_by (anc_offset);
2098 if (ie->indirect_info->vptr_changed)
2099 context.possible_dynamic_type_change (ie->in_polymorphic_cdtor,
2100 ie->indirect_info->otr_type);
2101 if (t)
2103 ipa_polymorphic_call_context ctx2 = ipa_polymorphic_call_context
2104 (t, ie->indirect_info->otr_type, anc_offset);
2105 if (!ctx2.useless_p ())
2106 context.combine_with (ctx2, ie->indirect_info->otr_type);
2109 else if (t)
2111 context = ipa_polymorphic_call_context (t, ie->indirect_info->otr_type,
2112 anc_offset);
2113 if (ie->indirect_info->vptr_changed)
2114 context.possible_dynamic_type_change (ie->in_polymorphic_cdtor,
2115 ie->indirect_info->otr_type);
2117 else
2118 return NULL_TREE;
2120 vec <cgraph_node *>targets;
2121 bool final;
2123 targets = possible_polymorphic_call_targets
2124 (ie->indirect_info->otr_type,
2125 ie->indirect_info->otr_token,
2126 context, &final);
2127 if (!final || targets.length () > 1)
2129 struct cgraph_node *node;
2130 if (*speculative)
2131 return target;
2132 if (!opt_for_fn (ie->caller->decl, flag_devirtualize_speculatively)
2133 || ie->speculative || !ie->maybe_hot_p ())
2134 return NULL;
2135 node = try_speculative_devirtualization (ie->indirect_info->otr_type,
2136 ie->indirect_info->otr_token,
2137 context);
2138 if (node)
2140 *speculative = true;
2141 target = node->decl;
2143 else
2144 return NULL;
2146 else
2148 *speculative = false;
2149 if (targets.length () == 1)
2150 target = targets[0]->decl;
2151 else
2152 target = ipa_impossible_devirt_target (ie, NULL_TREE);
2155 if (target && !possible_polymorphic_call_target_p (ie,
2156 cgraph_node::get (target)))
2157 target = ipa_impossible_devirt_target (ie, target);
2159 return target;
2163 /* If an indirect edge IE can be turned into a direct one based on KNOWN_CSTS,
2164 KNOWN_CONTEXTS (which can be vNULL) or KNOWN_AGGS (which also can be vNULL)
2165 return the destination. */
2167 tree
2168 ipa_get_indirect_edge_target (struct cgraph_edge *ie,
2169 vec<tree> known_csts,
2170 vec<ipa_polymorphic_call_context> known_contexts,
2171 vec<ipa_agg_jump_function_p> known_aggs,
2172 bool *speculative)
2174 return ipa_get_indirect_edge_target_1 (ie, known_csts, known_contexts,
2175 known_aggs, NULL, speculative);
2178 /* Calculate devirtualization time bonus for NODE, assuming we know KNOWN_CSTS
2179 and KNOWN_CONTEXTS. */
2181 static int
2182 devirtualization_time_bonus (struct cgraph_node *node,
2183 vec<tree> known_csts,
2184 vec<ipa_polymorphic_call_context> known_contexts,
2185 vec<ipa_agg_jump_function_p> known_aggs)
2187 struct cgraph_edge *ie;
2188 int res = 0;
2190 for (ie = node->indirect_calls; ie; ie = ie->next_callee)
2192 struct cgraph_node *callee;
2193 struct inline_summary *isummary;
2194 enum availability avail;
2195 tree target;
2196 bool speculative;
2198 target = ipa_get_indirect_edge_target (ie, known_csts, known_contexts,
2199 known_aggs, &speculative);
2200 if (!target)
2201 continue;
2203 /* Only bare minimum benefit for clearly un-inlineable targets. */
2204 res += 1;
2205 callee = cgraph_node::get (target);
2206 if (!callee || !callee->definition)
2207 continue;
2208 callee = callee->function_symbol (&avail);
2209 if (avail < AVAIL_AVAILABLE)
2210 continue;
2211 isummary = inline_summaries->get (callee);
2212 if (!isummary->inlinable)
2213 continue;
2215 /* FIXME: The values below need re-considering and perhaps also
2216 integrating into the cost metrics, at lest in some very basic way. */
2217 if (isummary->size <= MAX_INLINE_INSNS_AUTO / 4)
2218 res += 31 / ((int)speculative + 1);
2219 else if (isummary->size <= MAX_INLINE_INSNS_AUTO / 2)
2220 res += 15 / ((int)speculative + 1);
2221 else if (isummary->size <= MAX_INLINE_INSNS_AUTO
2222 || DECL_DECLARED_INLINE_P (callee->decl))
2223 res += 7 / ((int)speculative + 1);
2226 return res;
2229 /* Return time bonus incurred because of HINTS. */
2231 static int
2232 hint_time_bonus (inline_hints hints)
2234 int result = 0;
2235 if (hints & (INLINE_HINT_loop_iterations | INLINE_HINT_loop_stride))
2236 result += PARAM_VALUE (PARAM_IPA_CP_LOOP_HINT_BONUS);
2237 if (hints & INLINE_HINT_array_index)
2238 result += PARAM_VALUE (PARAM_IPA_CP_ARRAY_INDEX_HINT_BONUS);
2239 return result;
2242 /* If there is a reason to penalize the function described by INFO in the
2243 cloning goodness evaluation, do so. */
2245 static inline int64_t
2246 incorporate_penalties (ipa_node_params *info, int64_t evaluation)
2248 if (info->node_within_scc)
2249 evaluation = (evaluation
2250 * (100 - PARAM_VALUE (PARAM_IPA_CP_RECURSION_PENALTY))) / 100;
2252 if (info->node_calling_single_call)
2253 evaluation = (evaluation
2254 * (100 - PARAM_VALUE (PARAM_IPA_CP_SINGLE_CALL_PENALTY)))
2255 / 100;
2257 return evaluation;
2260 /* Return true if cloning NODE is a good idea, given the estimated TIME_BENEFIT
2261 and SIZE_COST and with the sum of frequencies of incoming edges to the
2262 potential new clone in FREQUENCIES. */
2264 static bool
2265 good_cloning_opportunity_p (struct cgraph_node *node, int time_benefit,
2266 int freq_sum, gcov_type count_sum, int size_cost)
2268 if (time_benefit == 0
2269 || !opt_for_fn (node->decl, flag_ipa_cp_clone)
2270 || !optimize_function_for_speed_p (DECL_STRUCT_FUNCTION (node->decl)))
2271 return false;
2273 gcc_assert (size_cost > 0);
2275 struct ipa_node_params *info = IPA_NODE_REF (node);
2276 if (max_count)
2278 int factor = (count_sum * 1000) / max_count;
2279 int64_t evaluation = (((int64_t) time_benefit * factor)
2280 / size_cost);
2281 evaluation = incorporate_penalties (info, evaluation);
2283 if (dump_file && (dump_flags & TDF_DETAILS))
2284 fprintf (dump_file, " good_cloning_opportunity_p (time: %i, "
2285 "size: %i, count_sum: " HOST_WIDE_INT_PRINT_DEC
2286 "%s%s) -> evaluation: " "%" PRId64
2287 ", threshold: %i\n",
2288 time_benefit, size_cost, (HOST_WIDE_INT) count_sum,
2289 info->node_within_scc ? ", scc" : "",
2290 info->node_calling_single_call ? ", single_call" : "",
2291 evaluation, PARAM_VALUE (PARAM_IPA_CP_EVAL_THRESHOLD));
2293 return evaluation >= PARAM_VALUE (PARAM_IPA_CP_EVAL_THRESHOLD);
2295 else
2297 int64_t evaluation = (((int64_t) time_benefit * freq_sum)
2298 / size_cost);
2299 evaluation = incorporate_penalties (info, evaluation);
2301 if (dump_file && (dump_flags & TDF_DETAILS))
2302 fprintf (dump_file, " good_cloning_opportunity_p (time: %i, "
2303 "size: %i, freq_sum: %i%s%s) -> evaluation: "
2304 "%" PRId64 ", threshold: %i\n",
2305 time_benefit, size_cost, freq_sum,
2306 info->node_within_scc ? ", scc" : "",
2307 info->node_calling_single_call ? ", single_call" : "",
2308 evaluation, PARAM_VALUE (PARAM_IPA_CP_EVAL_THRESHOLD));
2310 return evaluation >= PARAM_VALUE (PARAM_IPA_CP_EVAL_THRESHOLD);
2314 /* Return all context independent values from aggregate lattices in PLATS in a
2315 vector. Return NULL if there are none. */
2317 static vec<ipa_agg_jf_item, va_gc> *
2318 context_independent_aggregate_values (struct ipcp_param_lattices *plats)
2320 vec<ipa_agg_jf_item, va_gc> *res = NULL;
2322 if (plats->aggs_bottom
2323 || plats->aggs_contain_variable
2324 || plats->aggs_count == 0)
2325 return NULL;
2327 for (struct ipcp_agg_lattice *aglat = plats->aggs;
2328 aglat;
2329 aglat = aglat->next)
2330 if (aglat->is_single_const ())
2332 struct ipa_agg_jf_item item;
2333 item.offset = aglat->offset;
2334 item.value = aglat->values->value;
2335 vec_safe_push (res, item);
2337 return res;
2340 /* Allocate KNOWN_CSTS, KNOWN_CONTEXTS and, if non-NULL, KNOWN_AGGS and
2341 populate them with values of parameters that are known independent of the
2342 context. INFO describes the function. If REMOVABLE_PARAMS_COST is
2343 non-NULL, the movement cost of all removable parameters will be stored in
2344 it. */
2346 static bool
2347 gather_context_independent_values (struct ipa_node_params *info,
2348 vec<tree> *known_csts,
2349 vec<ipa_polymorphic_call_context>
2350 *known_contexts,
2351 vec<ipa_agg_jump_function> *known_aggs,
2352 int *removable_params_cost)
2354 int i, count = ipa_get_param_count (info);
2355 bool ret = false;
2357 known_csts->create (0);
2358 known_contexts->create (0);
2359 known_csts->safe_grow_cleared (count);
2360 known_contexts->safe_grow_cleared (count);
2361 if (known_aggs)
2363 known_aggs->create (0);
2364 known_aggs->safe_grow_cleared (count);
2367 if (removable_params_cost)
2368 *removable_params_cost = 0;
2370 for (i = 0; i < count ; i++)
2372 struct ipcp_param_lattices *plats = ipa_get_parm_lattices (info, i);
2373 ipcp_lattice<tree> *lat = &plats->itself;
2375 if (lat->is_single_const ())
2377 ipcp_value<tree> *val = lat->values;
2378 gcc_checking_assert (TREE_CODE (val->value) != TREE_BINFO);
2379 (*known_csts)[i] = val->value;
2380 if (removable_params_cost)
2381 *removable_params_cost
2382 += estimate_move_cost (TREE_TYPE (val->value), false);
2383 ret = true;
2385 else if (removable_params_cost
2386 && !ipa_is_param_used (info, i))
2387 *removable_params_cost
2388 += ipa_get_param_move_cost (info, i);
2390 ipcp_lattice<ipa_polymorphic_call_context> *ctxlat = &plats->ctxlat;
2391 if (ctxlat->is_single_const ())
2393 (*known_contexts)[i] = ctxlat->values->value;
2394 ret = true;
2397 if (known_aggs)
2399 vec<ipa_agg_jf_item, va_gc> *agg_items;
2400 struct ipa_agg_jump_function *ajf;
2402 agg_items = context_independent_aggregate_values (plats);
2403 ajf = &(*known_aggs)[i];
2404 ajf->items = agg_items;
2405 ajf->by_ref = plats->aggs_by_ref;
2406 ret |= agg_items != NULL;
2410 return ret;
2413 /* The current interface in ipa-inline-analysis requires a pointer vector.
2414 Create it.
2416 FIXME: That interface should be re-worked, this is slightly silly. Still,
2417 I'd like to discuss how to change it first and this demonstrates the
2418 issue. */
2420 static vec<ipa_agg_jump_function_p>
2421 agg_jmp_p_vec_for_t_vec (vec<ipa_agg_jump_function> known_aggs)
2423 vec<ipa_agg_jump_function_p> ret;
2424 struct ipa_agg_jump_function *ajf;
2425 int i;
2427 ret.create (known_aggs.length ());
2428 FOR_EACH_VEC_ELT (known_aggs, i, ajf)
2429 ret.quick_push (ajf);
2430 return ret;
2433 /* Perform time and size measurement of NODE with the context given in
2434 KNOWN_CSTS, KNOWN_CONTEXTS and KNOWN_AGGS, calculate the benefit and cost
2435 given BASE_TIME of the node without specialization, REMOVABLE_PARAMS_COST of
2436 all context-independent removable parameters and EST_MOVE_COST of estimated
2437 movement of the considered parameter and store it into VAL. */
2439 static void
2440 perform_estimation_of_a_value (cgraph_node *node, vec<tree> known_csts,
2441 vec<ipa_polymorphic_call_context> known_contexts,
2442 vec<ipa_agg_jump_function_p> known_aggs_ptrs,
2443 int base_time, int removable_params_cost,
2444 int est_move_cost, ipcp_value_base *val)
2446 int time, size, time_benefit;
2447 inline_hints hints;
2449 estimate_ipcp_clone_size_and_time (node, known_csts, known_contexts,
2450 known_aggs_ptrs, &size, &time,
2451 &hints);
2452 time_benefit = base_time - time
2453 + devirtualization_time_bonus (node, known_csts, known_contexts,
2454 known_aggs_ptrs)
2455 + hint_time_bonus (hints)
2456 + removable_params_cost + est_move_cost;
2458 gcc_checking_assert (size >=0);
2459 /* The inliner-heuristics based estimates may think that in certain
2460 contexts some functions do not have any size at all but we want
2461 all specializations to have at least a tiny cost, not least not to
2462 divide by zero. */
2463 if (size == 0)
2464 size = 1;
2466 val->local_time_benefit = time_benefit;
2467 val->local_size_cost = size;
2470 /* Iterate over known values of parameters of NODE and estimate the local
2471 effects in terms of time and size they have. */
2473 static void
2474 estimate_local_effects (struct cgraph_node *node)
2476 struct ipa_node_params *info = IPA_NODE_REF (node);
2477 int i, count = ipa_get_param_count (info);
2478 vec<tree> known_csts;
2479 vec<ipa_polymorphic_call_context> known_contexts;
2480 vec<ipa_agg_jump_function> known_aggs;
2481 vec<ipa_agg_jump_function_p> known_aggs_ptrs;
2482 bool always_const;
2483 int base_time = inline_summaries->get (node)->time;
2484 int removable_params_cost;
2486 if (!count || !ipcp_versionable_function_p (node))
2487 return;
2489 if (dump_file && (dump_flags & TDF_DETAILS))
2490 fprintf (dump_file, "\nEstimating effects for %s/%i, base_time: %i.\n",
2491 node->name (), node->order, base_time);
2493 always_const = gather_context_independent_values (info, &known_csts,
2494 &known_contexts, &known_aggs,
2495 &removable_params_cost);
2496 known_aggs_ptrs = agg_jmp_p_vec_for_t_vec (known_aggs);
2497 if (always_const)
2499 struct caller_statistics stats;
2500 inline_hints hints;
2501 int time, size;
2503 init_caller_stats (&stats);
2504 node->call_for_symbol_thunks_and_aliases (gather_caller_stats, &stats,
2505 false);
2506 estimate_ipcp_clone_size_and_time (node, known_csts, known_contexts,
2507 known_aggs_ptrs, &size, &time, &hints);
2508 time -= devirtualization_time_bonus (node, known_csts, known_contexts,
2509 known_aggs_ptrs);
2510 time -= hint_time_bonus (hints);
2511 time -= removable_params_cost;
2512 size -= stats.n_calls * removable_params_cost;
2514 if (dump_file)
2515 fprintf (dump_file, " - context independent values, size: %i, "
2516 "time_benefit: %i\n", size, base_time - time);
2518 if (size <= 0
2519 || node->will_be_removed_from_program_if_no_direct_calls_p ())
2521 info->do_clone_for_all_contexts = true;
2522 base_time = time;
2524 if (dump_file)
2525 fprintf (dump_file, " Decided to specialize for all "
2526 "known contexts, code not going to grow.\n");
2528 else if (good_cloning_opportunity_p (node, base_time - time,
2529 stats.freq_sum, stats.count_sum,
2530 size))
2532 if (size + overall_size <= max_new_size)
2534 info->do_clone_for_all_contexts = true;
2535 base_time = time;
2536 overall_size += size;
2538 if (dump_file)
2539 fprintf (dump_file, " Decided to specialize for all "
2540 "known contexts, growth deemed beneficial.\n");
2542 else if (dump_file && (dump_flags & TDF_DETAILS))
2543 fprintf (dump_file, " Not cloning for all contexts because "
2544 "max_new_size would be reached with %li.\n",
2545 size + overall_size);
2549 for (i = 0; i < count ; i++)
2551 struct ipcp_param_lattices *plats = ipa_get_parm_lattices (info, i);
2552 ipcp_lattice<tree> *lat = &plats->itself;
2553 ipcp_value<tree> *val;
2555 if (lat->bottom
2556 || !lat->values
2557 || known_csts[i])
2558 continue;
2560 for (val = lat->values; val; val = val->next)
2562 gcc_checking_assert (TREE_CODE (val->value) != TREE_BINFO);
2563 known_csts[i] = val->value;
2565 int emc = estimate_move_cost (TREE_TYPE (val->value), true);
2566 perform_estimation_of_a_value (node, known_csts, known_contexts,
2567 known_aggs_ptrs, base_time,
2568 removable_params_cost, emc, val);
2570 if (dump_file && (dump_flags & TDF_DETAILS))
2572 fprintf (dump_file, " - estimates for value ");
2573 print_ipcp_constant_value (dump_file, val->value);
2574 fprintf (dump_file, " for ");
2575 ipa_dump_param (dump_file, info, i);
2576 fprintf (dump_file, ": time_benefit: %i, size: %i\n",
2577 val->local_time_benefit, val->local_size_cost);
2580 known_csts[i] = NULL_TREE;
2583 for (i = 0; i < count; i++)
2585 struct ipcp_param_lattices *plats = ipa_get_parm_lattices (info, i);
2587 if (!plats->virt_call)
2588 continue;
2590 ipcp_lattice<ipa_polymorphic_call_context> *ctxlat = &plats->ctxlat;
2591 ipcp_value<ipa_polymorphic_call_context> *val;
2593 if (ctxlat->bottom
2594 || !ctxlat->values
2595 || !known_contexts[i].useless_p ())
2596 continue;
2598 for (val = ctxlat->values; val; val = val->next)
2600 known_contexts[i] = val->value;
2601 perform_estimation_of_a_value (node, known_csts, known_contexts,
2602 known_aggs_ptrs, base_time,
2603 removable_params_cost, 0, val);
2605 if (dump_file && (dump_flags & TDF_DETAILS))
2607 fprintf (dump_file, " - estimates for polymorphic context ");
2608 print_ipcp_constant_value (dump_file, val->value);
2609 fprintf (dump_file, " for ");
2610 ipa_dump_param (dump_file, info, i);
2611 fprintf (dump_file, ": time_benefit: %i, size: %i\n",
2612 val->local_time_benefit, val->local_size_cost);
2615 known_contexts[i] = ipa_polymorphic_call_context ();
2618 for (i = 0; i < count ; i++)
2620 struct ipcp_param_lattices *plats = ipa_get_parm_lattices (info, i);
2621 struct ipa_agg_jump_function *ajf;
2622 struct ipcp_agg_lattice *aglat;
2624 if (plats->aggs_bottom || !plats->aggs)
2625 continue;
2627 ajf = &known_aggs[i];
2628 for (aglat = plats->aggs; aglat; aglat = aglat->next)
2630 ipcp_value<tree> *val;
2631 if (aglat->bottom || !aglat->values
2632 /* If the following is true, the one value is in known_aggs. */
2633 || (!plats->aggs_contain_variable
2634 && aglat->is_single_const ()))
2635 continue;
2637 for (val = aglat->values; val; val = val->next)
2639 struct ipa_agg_jf_item item;
2641 item.offset = aglat->offset;
2642 item.value = val->value;
2643 vec_safe_push (ajf->items, item);
2645 perform_estimation_of_a_value (node, known_csts, known_contexts,
2646 known_aggs_ptrs, base_time,
2647 removable_params_cost, 0, val);
2649 if (dump_file && (dump_flags & TDF_DETAILS))
2651 fprintf (dump_file, " - estimates for value ");
2652 print_ipcp_constant_value (dump_file, val->value);
2653 fprintf (dump_file, " for ");
2654 ipa_dump_param (dump_file, info, i);
2655 fprintf (dump_file, "[%soffset: " HOST_WIDE_INT_PRINT_DEC
2656 "]: time_benefit: %i, size: %i\n",
2657 plats->aggs_by_ref ? "ref " : "",
2658 aglat->offset,
2659 val->local_time_benefit, val->local_size_cost);
2662 ajf->items->pop ();
2667 for (i = 0; i < count ; i++)
2668 vec_free (known_aggs[i].items);
2670 known_csts.release ();
2671 known_contexts.release ();
2672 known_aggs.release ();
2673 known_aggs_ptrs.release ();
2677 /* Add value CUR_VAL and all yet-unsorted values it is dependent on to the
2678 topological sort of values. */
2680 template <typename valtype>
2681 void
2682 value_topo_info<valtype>::add_val (ipcp_value<valtype> *cur_val)
2684 ipcp_value_source<valtype> *src;
2686 if (cur_val->dfs)
2687 return;
2689 dfs_counter++;
2690 cur_val->dfs = dfs_counter;
2691 cur_val->low_link = dfs_counter;
2693 cur_val->topo_next = stack;
2694 stack = cur_val;
2695 cur_val->on_stack = true;
2697 for (src = cur_val->sources; src; src = src->next)
2698 if (src->val)
2700 if (src->val->dfs == 0)
2702 add_val (src->val);
2703 if (src->val->low_link < cur_val->low_link)
2704 cur_val->low_link = src->val->low_link;
2706 else if (src->val->on_stack
2707 && src->val->dfs < cur_val->low_link)
2708 cur_val->low_link = src->val->dfs;
2711 if (cur_val->dfs == cur_val->low_link)
2713 ipcp_value<valtype> *v, *scc_list = NULL;
2717 v = stack;
2718 stack = v->topo_next;
2719 v->on_stack = false;
2721 v->scc_next = scc_list;
2722 scc_list = v;
2724 while (v != cur_val);
2726 cur_val->topo_next = values_topo;
2727 values_topo = cur_val;
2731 /* Add all values in lattices associated with NODE to the topological sort if
2732 they are not there yet. */
2734 static void
2735 add_all_node_vals_to_toposort (cgraph_node *node, ipa_topo_info *topo)
2737 struct ipa_node_params *info = IPA_NODE_REF (node);
2738 int i, count = ipa_get_param_count (info);
2740 for (i = 0; i < count ; i++)
2742 struct ipcp_param_lattices *plats = ipa_get_parm_lattices (info, i);
2743 ipcp_lattice<tree> *lat = &plats->itself;
2744 struct ipcp_agg_lattice *aglat;
2746 if (!lat->bottom)
2748 ipcp_value<tree> *val;
2749 for (val = lat->values; val; val = val->next)
2750 topo->constants.add_val (val);
2753 if (!plats->aggs_bottom)
2754 for (aglat = plats->aggs; aglat; aglat = aglat->next)
2755 if (!aglat->bottom)
2757 ipcp_value<tree> *val;
2758 for (val = aglat->values; val; val = val->next)
2759 topo->constants.add_val (val);
2762 ipcp_lattice<ipa_polymorphic_call_context> *ctxlat = &plats->ctxlat;
2763 if (!ctxlat->bottom)
2765 ipcp_value<ipa_polymorphic_call_context> *ctxval;
2766 for (ctxval = ctxlat->values; ctxval; ctxval = ctxval->next)
2767 topo->contexts.add_val (ctxval);
2772 /* One pass of constants propagation along the call graph edges, from callers
2773 to callees (requires topological ordering in TOPO), iterate over strongly
2774 connected components. */
2776 static void
2777 propagate_constants_topo (struct ipa_topo_info *topo)
2779 int i;
2781 for (i = topo->nnodes - 1; i >= 0; i--)
2783 unsigned j;
2784 struct cgraph_node *v, *node = topo->order[i];
2785 vec<cgraph_node *> cycle_nodes = ipa_get_nodes_in_cycle (node);
2787 /* First, iteratively propagate within the strongly connected component
2788 until all lattices stabilize. */
2789 FOR_EACH_VEC_ELT (cycle_nodes, j, v)
2790 if (v->has_gimple_body_p ())
2791 push_node_to_stack (topo, v);
2793 v = pop_node_from_stack (topo);
2794 while (v)
2796 struct cgraph_edge *cs;
2798 for (cs = v->callees; cs; cs = cs->next_callee)
2799 if (ipa_edge_within_scc (cs))
2801 IPA_NODE_REF (v)->node_within_scc = true;
2802 if (propagate_constants_accross_call (cs))
2803 push_node_to_stack (topo, cs->callee->function_symbol ());
2805 v = pop_node_from_stack (topo);
2808 /* Afterwards, propagate along edges leading out of the SCC, calculates
2809 the local effects of the discovered constants and all valid values to
2810 their topological sort. */
2811 FOR_EACH_VEC_ELT (cycle_nodes, j, v)
2812 if (v->has_gimple_body_p ())
2814 struct cgraph_edge *cs;
2816 estimate_local_effects (v);
2817 add_all_node_vals_to_toposort (v, topo);
2818 for (cs = v->callees; cs; cs = cs->next_callee)
2819 if (!ipa_edge_within_scc (cs))
2820 propagate_constants_accross_call (cs);
2822 cycle_nodes.release ();
2827 /* Return the sum of A and B if none of them is bigger than INT_MAX/2, return
2828 the bigger one if otherwise. */
2830 static int
2831 safe_add (int a, int b)
2833 if (a > INT_MAX/2 || b > INT_MAX/2)
2834 return a > b ? a : b;
2835 else
2836 return a + b;
2840 /* Propagate the estimated effects of individual values along the topological
2841 from the dependent values to those they depend on. */
2843 template <typename valtype>
2844 void
2845 value_topo_info<valtype>::propagate_effects ()
2847 ipcp_value<valtype> *base;
2849 for (base = values_topo; base; base = base->topo_next)
2851 ipcp_value_source<valtype> *src;
2852 ipcp_value<valtype> *val;
2853 int time = 0, size = 0;
2855 for (val = base; val; val = val->scc_next)
2857 time = safe_add (time,
2858 val->local_time_benefit + val->prop_time_benefit);
2859 size = safe_add (size, val->local_size_cost + val->prop_size_cost);
2862 for (val = base; val; val = val->scc_next)
2863 for (src = val->sources; src; src = src->next)
2864 if (src->val
2865 && src->cs->maybe_hot_p ())
2867 src->val->prop_time_benefit = safe_add (time,
2868 src->val->prop_time_benefit);
2869 src->val->prop_size_cost = safe_add (size,
2870 src->val->prop_size_cost);
2876 /* Propagate constants, polymorphic contexts and their effects from the
2877 summaries interprocedurally. */
2879 static void
2880 ipcp_propagate_stage (struct ipa_topo_info *topo)
2882 struct cgraph_node *node;
2884 if (dump_file)
2885 fprintf (dump_file, "\n Propagating constants:\n\n");
2887 if (in_lto_p)
2888 ipa_update_after_lto_read ();
2891 FOR_EACH_DEFINED_FUNCTION (node)
2893 struct ipa_node_params *info = IPA_NODE_REF (node);
2895 determine_versionability (node, info);
2896 if (node->has_gimple_body_p ())
2898 info->lattices = XCNEWVEC (struct ipcp_param_lattices,
2899 ipa_get_param_count (info));
2900 initialize_node_lattices (node);
2902 if (node->definition && !node->alias)
2903 overall_size += inline_summaries->get (node)->self_size;
2904 if (node->count > max_count)
2905 max_count = node->count;
2908 max_new_size = overall_size;
2909 if (max_new_size < PARAM_VALUE (PARAM_LARGE_UNIT_INSNS))
2910 max_new_size = PARAM_VALUE (PARAM_LARGE_UNIT_INSNS);
2911 max_new_size += max_new_size * PARAM_VALUE (PARAM_IPCP_UNIT_GROWTH) / 100 + 1;
2913 if (dump_file)
2914 fprintf (dump_file, "\noverall_size: %li, max_new_size: %li\n",
2915 overall_size, max_new_size);
2917 propagate_constants_topo (topo);
2918 if (flag_checking)
2919 ipcp_verify_propagated_values ();
2920 topo->constants.propagate_effects ();
2921 topo->contexts.propagate_effects ();
2923 if (dump_file)
2925 fprintf (dump_file, "\nIPA lattices after all propagation:\n");
2926 print_all_lattices (dump_file, (dump_flags & TDF_DETAILS), true);
2930 /* Discover newly direct outgoing edges from NODE which is a new clone with
2931 known KNOWN_CSTS and make them direct. */
2933 static void
2934 ipcp_discover_new_direct_edges (struct cgraph_node *node,
2935 vec<tree> known_csts,
2936 vec<ipa_polymorphic_call_context>
2937 known_contexts,
2938 struct ipa_agg_replacement_value *aggvals)
2940 struct cgraph_edge *ie, *next_ie;
2941 bool found = false;
2943 for (ie = node->indirect_calls; ie; ie = next_ie)
2945 tree target;
2946 bool speculative;
2948 next_ie = ie->next_callee;
2949 target = ipa_get_indirect_edge_target_1 (ie, known_csts, known_contexts,
2950 vNULL, aggvals, &speculative);
2951 if (target)
2953 bool agg_contents = ie->indirect_info->agg_contents;
2954 bool polymorphic = ie->indirect_info->polymorphic;
2955 int param_index = ie->indirect_info->param_index;
2956 struct cgraph_edge *cs = ipa_make_edge_direct_to_target (ie, target,
2957 speculative);
2958 found = true;
2960 if (cs && !agg_contents && !polymorphic)
2962 struct ipa_node_params *info = IPA_NODE_REF (node);
2963 int c = ipa_get_controlled_uses (info, param_index);
2964 if (c != IPA_UNDESCRIBED_USE)
2966 struct ipa_ref *to_del;
2968 c--;
2969 ipa_set_controlled_uses (info, param_index, c);
2970 if (dump_file && (dump_flags & TDF_DETAILS))
2971 fprintf (dump_file, " controlled uses count of param "
2972 "%i bumped down to %i\n", param_index, c);
2973 if (c == 0
2974 && (to_del = node->find_reference (cs->callee, NULL, 0)))
2976 if (dump_file && (dump_flags & TDF_DETAILS))
2977 fprintf (dump_file, " and even removing its "
2978 "cloning-created reference\n");
2979 to_del->remove_reference ();
2985 /* Turning calls to direct calls will improve overall summary. */
2986 if (found)
2987 inline_update_overall_summary (node);
2990 /* Vector of pointers which for linked lists of clones of an original crgaph
2991 edge. */
2993 static vec<cgraph_edge *> next_edge_clone;
2994 static vec<cgraph_edge *> prev_edge_clone;
2996 static inline void
2997 grow_edge_clone_vectors (void)
2999 if (next_edge_clone.length ()
3000 <= (unsigned) symtab->edges_max_uid)
3001 next_edge_clone.safe_grow_cleared (symtab->edges_max_uid + 1);
3002 if (prev_edge_clone.length ()
3003 <= (unsigned) symtab->edges_max_uid)
3004 prev_edge_clone.safe_grow_cleared (symtab->edges_max_uid + 1);
3007 /* Edge duplication hook to grow the appropriate linked list in
3008 next_edge_clone. */
3010 static void
3011 ipcp_edge_duplication_hook (struct cgraph_edge *src, struct cgraph_edge *dst,
3012 void *)
3014 grow_edge_clone_vectors ();
3016 struct cgraph_edge *old_next = next_edge_clone[src->uid];
3017 if (old_next)
3018 prev_edge_clone[old_next->uid] = dst;
3019 prev_edge_clone[dst->uid] = src;
3021 next_edge_clone[dst->uid] = old_next;
3022 next_edge_clone[src->uid] = dst;
3025 /* Hook that is called by cgraph.c when an edge is removed. */
3027 static void
3028 ipcp_edge_removal_hook (struct cgraph_edge *cs, void *)
3030 grow_edge_clone_vectors ();
3032 struct cgraph_edge *prev = prev_edge_clone[cs->uid];
3033 struct cgraph_edge *next = next_edge_clone[cs->uid];
3034 if (prev)
3035 next_edge_clone[prev->uid] = next;
3036 if (next)
3037 prev_edge_clone[next->uid] = prev;
3040 /* See if NODE is a clone with a known aggregate value at a given OFFSET of a
3041 parameter with the given INDEX. */
3043 static tree
3044 get_clone_agg_value (struct cgraph_node *node, HOST_WIDE_INT offset,
3045 int index)
3047 struct ipa_agg_replacement_value *aggval;
3049 aggval = ipa_get_agg_replacements_for_node (node);
3050 while (aggval)
3052 if (aggval->offset == offset
3053 && aggval->index == index)
3054 return aggval->value;
3055 aggval = aggval->next;
3057 return NULL_TREE;
3060 /* Return true is NODE is DEST or its clone for all contexts. */
3062 static bool
3063 same_node_or_its_all_contexts_clone_p (cgraph_node *node, cgraph_node *dest)
3065 if (node == dest)
3066 return true;
3068 struct ipa_node_params *info = IPA_NODE_REF (node);
3069 return info->is_all_contexts_clone && info->ipcp_orig_node == dest;
3072 /* Return true if edge CS does bring about the value described by SRC to node
3073 DEST or its clone for all contexts. */
3075 static bool
3076 cgraph_edge_brings_value_p (cgraph_edge *cs, ipcp_value_source<tree> *src,
3077 cgraph_node *dest)
3079 struct ipa_node_params *caller_info = IPA_NODE_REF (cs->caller);
3080 enum availability availability;
3081 cgraph_node *real_dest = cs->callee->function_symbol (&availability);
3083 if (!same_node_or_its_all_contexts_clone_p (real_dest, dest)
3084 || availability <= AVAIL_INTERPOSABLE
3085 || caller_info->node_dead)
3086 return false;
3087 if (!src->val)
3088 return true;
3090 if (caller_info->ipcp_orig_node)
3092 tree t;
3093 if (src->offset == -1)
3094 t = caller_info->known_csts[src->index];
3095 else
3096 t = get_clone_agg_value (cs->caller, src->offset, src->index);
3097 return (t != NULL_TREE
3098 && values_equal_for_ipcp_p (src->val->value, t));
3100 else
3102 struct ipcp_agg_lattice *aglat;
3103 struct ipcp_param_lattices *plats = ipa_get_parm_lattices (caller_info,
3104 src->index);
3105 if (src->offset == -1)
3106 return (plats->itself.is_single_const ()
3107 && values_equal_for_ipcp_p (src->val->value,
3108 plats->itself.values->value));
3109 else
3111 if (plats->aggs_bottom || plats->aggs_contain_variable)
3112 return false;
3113 for (aglat = plats->aggs; aglat; aglat = aglat->next)
3114 if (aglat->offset == src->offset)
3115 return (aglat->is_single_const ()
3116 && values_equal_for_ipcp_p (src->val->value,
3117 aglat->values->value));
3119 return false;
3123 /* Return true if edge CS does bring about the value described by SRC to node
3124 DEST or its clone for all contexts. */
3126 static bool
3127 cgraph_edge_brings_value_p (cgraph_edge *cs,
3128 ipcp_value_source<ipa_polymorphic_call_context> *src,
3129 cgraph_node *dest)
3131 struct ipa_node_params *caller_info = IPA_NODE_REF (cs->caller);
3132 cgraph_node *real_dest = cs->callee->function_symbol ();
3134 if (!same_node_or_its_all_contexts_clone_p (real_dest, dest)
3135 || caller_info->node_dead)
3136 return false;
3137 if (!src->val)
3138 return true;
3140 if (caller_info->ipcp_orig_node)
3141 return (caller_info->known_contexts.length () > (unsigned) src->index)
3142 && values_equal_for_ipcp_p (src->val->value,
3143 caller_info->known_contexts[src->index]);
3145 struct ipcp_param_lattices *plats = ipa_get_parm_lattices (caller_info,
3146 src->index);
3147 return plats->ctxlat.is_single_const ()
3148 && values_equal_for_ipcp_p (src->val->value,
3149 plats->ctxlat.values->value);
3152 /* Get the next clone in the linked list of clones of an edge. */
3154 static inline struct cgraph_edge *
3155 get_next_cgraph_edge_clone (struct cgraph_edge *cs)
3157 return next_edge_clone[cs->uid];
3160 /* Given VAL that is intended for DEST, iterate over all its sources and if
3161 they still hold, add their edge frequency and their number into *FREQUENCY
3162 and *CALLER_COUNT respectively. */
3164 template <typename valtype>
3165 static bool
3166 get_info_about_necessary_edges (ipcp_value<valtype> *val, cgraph_node *dest,
3167 int *freq_sum,
3168 gcov_type *count_sum, int *caller_count)
3170 ipcp_value_source<valtype> *src;
3171 int freq = 0, count = 0;
3172 gcov_type cnt = 0;
3173 bool hot = false;
3175 for (src = val->sources; src; src = src->next)
3177 struct cgraph_edge *cs = src->cs;
3178 while (cs)
3180 if (cgraph_edge_brings_value_p (cs, src, dest))
3182 count++;
3183 freq += cs->frequency;
3184 cnt += cs->count;
3185 hot |= cs->maybe_hot_p ();
3187 cs = get_next_cgraph_edge_clone (cs);
3191 *freq_sum = freq;
3192 *count_sum = cnt;
3193 *caller_count = count;
3194 return hot;
3197 /* Return a vector of incoming edges that do bring value VAL to node DEST. It
3198 is assumed their number is known and equal to CALLER_COUNT. */
3200 template <typename valtype>
3201 static vec<cgraph_edge *>
3202 gather_edges_for_value (ipcp_value<valtype> *val, cgraph_node *dest,
3203 int caller_count)
3205 ipcp_value_source<valtype> *src;
3206 vec<cgraph_edge *> ret;
3208 ret.create (caller_count);
3209 for (src = val->sources; src; src = src->next)
3211 struct cgraph_edge *cs = src->cs;
3212 while (cs)
3214 if (cgraph_edge_brings_value_p (cs, src, dest))
3215 ret.quick_push (cs);
3216 cs = get_next_cgraph_edge_clone (cs);
3220 return ret;
3223 /* Construct a replacement map for a know VALUE for a formal parameter PARAM.
3224 Return it or NULL if for some reason it cannot be created. */
3226 static struct ipa_replace_map *
3227 get_replacement_map (struct ipa_node_params *info, tree value, int parm_num)
3229 struct ipa_replace_map *replace_map;
3232 replace_map = ggc_alloc<ipa_replace_map> ();
3233 if (dump_file)
3235 fprintf (dump_file, " replacing ");
3236 ipa_dump_param (dump_file, info, parm_num);
3238 fprintf (dump_file, " with const ");
3239 print_generic_expr (dump_file, value, 0);
3240 fprintf (dump_file, "\n");
3242 replace_map->old_tree = NULL;
3243 replace_map->parm_num = parm_num;
3244 replace_map->new_tree = value;
3245 replace_map->replace_p = true;
3246 replace_map->ref_p = false;
3248 return replace_map;
3251 /* Dump new profiling counts */
3253 static void
3254 dump_profile_updates (struct cgraph_node *orig_node,
3255 struct cgraph_node *new_node)
3257 struct cgraph_edge *cs;
3259 fprintf (dump_file, " setting count of the specialized node to "
3260 HOST_WIDE_INT_PRINT_DEC "\n", (HOST_WIDE_INT) new_node->count);
3261 for (cs = new_node->callees; cs ; cs = cs->next_callee)
3262 fprintf (dump_file, " edge to %s has count "
3263 HOST_WIDE_INT_PRINT_DEC "\n",
3264 cs->callee->name (), (HOST_WIDE_INT) cs->count);
3266 fprintf (dump_file, " setting count of the original node to "
3267 HOST_WIDE_INT_PRINT_DEC "\n", (HOST_WIDE_INT) orig_node->count);
3268 for (cs = orig_node->callees; cs ; cs = cs->next_callee)
3269 fprintf (dump_file, " edge to %s is left with "
3270 HOST_WIDE_INT_PRINT_DEC "\n",
3271 cs->callee->name (), (HOST_WIDE_INT) cs->count);
3274 /* After a specialized NEW_NODE version of ORIG_NODE has been created, update
3275 their profile information to reflect this. */
3277 static void
3278 update_profiling_info (struct cgraph_node *orig_node,
3279 struct cgraph_node *new_node)
3281 struct cgraph_edge *cs;
3282 struct caller_statistics stats;
3283 gcov_type new_sum, orig_sum;
3284 gcov_type remainder, orig_node_count = orig_node->count;
3286 if (orig_node_count == 0)
3287 return;
3289 init_caller_stats (&stats);
3290 orig_node->call_for_symbol_thunks_and_aliases (gather_caller_stats, &stats,
3291 false);
3292 orig_sum = stats.count_sum;
3293 init_caller_stats (&stats);
3294 new_node->call_for_symbol_thunks_and_aliases (gather_caller_stats, &stats,
3295 false);
3296 new_sum = stats.count_sum;
3298 if (orig_node_count < orig_sum + new_sum)
3300 if (dump_file)
3301 fprintf (dump_file, " Problem: node %s/%i has too low count "
3302 HOST_WIDE_INT_PRINT_DEC " while the sum of incoming "
3303 "counts is " HOST_WIDE_INT_PRINT_DEC "\n",
3304 orig_node->name (), orig_node->order,
3305 (HOST_WIDE_INT) orig_node_count,
3306 (HOST_WIDE_INT) (orig_sum + new_sum));
3308 orig_node_count = (orig_sum + new_sum) * 12 / 10;
3309 if (dump_file)
3310 fprintf (dump_file, " proceeding by pretending it was "
3311 HOST_WIDE_INT_PRINT_DEC "\n",
3312 (HOST_WIDE_INT) orig_node_count);
3315 new_node->count = new_sum;
3316 remainder = orig_node_count - new_sum;
3317 orig_node->count = remainder;
3319 for (cs = new_node->callees; cs ; cs = cs->next_callee)
3320 if (cs->frequency)
3321 cs->count = apply_probability (cs->count,
3322 GCOV_COMPUTE_SCALE (new_sum,
3323 orig_node_count));
3324 else
3325 cs->count = 0;
3327 for (cs = orig_node->callees; cs ; cs = cs->next_callee)
3328 cs->count = apply_probability (cs->count,
3329 GCOV_COMPUTE_SCALE (remainder,
3330 orig_node_count));
3332 if (dump_file)
3333 dump_profile_updates (orig_node, new_node);
3336 /* Update the respective profile of specialized NEW_NODE and the original
3337 ORIG_NODE after additional edges with cumulative count sum REDIRECTED_SUM
3338 have been redirected to the specialized version. */
3340 static void
3341 update_specialized_profile (struct cgraph_node *new_node,
3342 struct cgraph_node *orig_node,
3343 gcov_type redirected_sum)
3345 struct cgraph_edge *cs;
3346 gcov_type new_node_count, orig_node_count = orig_node->count;
3348 if (dump_file)
3349 fprintf (dump_file, " the sum of counts of redirected edges is "
3350 HOST_WIDE_INT_PRINT_DEC "\n", (HOST_WIDE_INT) redirected_sum);
3351 if (orig_node_count == 0)
3352 return;
3354 gcc_assert (orig_node_count >= redirected_sum);
3356 new_node_count = new_node->count;
3357 new_node->count += redirected_sum;
3358 orig_node->count -= redirected_sum;
3360 for (cs = new_node->callees; cs ; cs = cs->next_callee)
3361 if (cs->frequency)
3362 cs->count += apply_probability (cs->count,
3363 GCOV_COMPUTE_SCALE (redirected_sum,
3364 new_node_count));
3365 else
3366 cs->count = 0;
3368 for (cs = orig_node->callees; cs ; cs = cs->next_callee)
3370 gcov_type dec = apply_probability (cs->count,
3371 GCOV_COMPUTE_SCALE (redirected_sum,
3372 orig_node_count));
3373 if (dec < cs->count)
3374 cs->count -= dec;
3375 else
3376 cs->count = 0;
3379 if (dump_file)
3380 dump_profile_updates (orig_node, new_node);
3383 /* Create a specialized version of NODE with known constants in KNOWN_CSTS,
3384 known contexts in KNOWN_CONTEXTS and known aggregate values in AGGVALS and
3385 redirect all edges in CALLERS to it. */
3387 static struct cgraph_node *
3388 create_specialized_node (struct cgraph_node *node,
3389 vec<tree> known_csts,
3390 vec<ipa_polymorphic_call_context> known_contexts,
3391 struct ipa_agg_replacement_value *aggvals,
3392 vec<cgraph_edge *> callers)
3394 struct ipa_node_params *new_info, *info = IPA_NODE_REF (node);
3395 vec<ipa_replace_map *, va_gc> *replace_trees = NULL;
3396 struct ipa_agg_replacement_value *av;
3397 struct cgraph_node *new_node;
3398 int i, count = ipa_get_param_count (info);
3399 bitmap args_to_skip;
3401 gcc_assert (!info->ipcp_orig_node);
3403 if (node->local.can_change_signature)
3405 args_to_skip = BITMAP_GGC_ALLOC ();
3406 for (i = 0; i < count; i++)
3408 tree t = known_csts[i];
3410 if (t || !ipa_is_param_used (info, i))
3411 bitmap_set_bit (args_to_skip, i);
3414 else
3416 args_to_skip = NULL;
3417 if (dump_file && (dump_flags & TDF_DETAILS))
3418 fprintf (dump_file, " cannot change function signature\n");
3421 for (i = 0; i < count ; i++)
3423 tree t = known_csts[i];
3424 if (t)
3426 struct ipa_replace_map *replace_map;
3428 gcc_checking_assert (TREE_CODE (t) != TREE_BINFO);
3429 replace_map = get_replacement_map (info, t, i);
3430 if (replace_map)
3431 vec_safe_push (replace_trees, replace_map);
3435 new_node = node->create_virtual_clone (callers, replace_trees,
3436 args_to_skip, "constprop");
3437 ipa_set_node_agg_value_chain (new_node, aggvals);
3438 for (av = aggvals; av; av = av->next)
3439 new_node->maybe_create_reference (av->value, IPA_REF_ADDR, NULL);
3441 if (dump_file && (dump_flags & TDF_DETAILS))
3443 fprintf (dump_file, " the new node is %s/%i.\n",
3444 new_node->name (), new_node->order);
3445 if (known_contexts.exists ())
3447 for (i = 0; i < count ; i++)
3448 if (!known_contexts[i].useless_p ())
3450 fprintf (dump_file, " known ctx %i is ", i);
3451 known_contexts[i].dump (dump_file);
3454 if (aggvals)
3455 ipa_dump_agg_replacement_values (dump_file, aggvals);
3457 ipa_check_create_node_params ();
3458 update_profiling_info (node, new_node);
3459 new_info = IPA_NODE_REF (new_node);
3460 new_info->ipcp_orig_node = node;
3461 new_info->known_csts = known_csts;
3462 new_info->known_contexts = known_contexts;
3464 ipcp_discover_new_direct_edges (new_node, known_csts, known_contexts, aggvals);
3466 callers.release ();
3467 return new_node;
3470 /* Given a NODE, and a subset of its CALLERS, try to populate blanks slots in
3471 KNOWN_CSTS with constants that are also known for all of the CALLERS. */
3473 static void
3474 find_more_scalar_values_for_callers_subset (struct cgraph_node *node,
3475 vec<tree> known_csts,
3476 vec<cgraph_edge *> callers)
3478 struct ipa_node_params *info = IPA_NODE_REF (node);
3479 int i, count = ipa_get_param_count (info);
3481 for (i = 0; i < count ; i++)
3483 struct cgraph_edge *cs;
3484 tree newval = NULL_TREE;
3485 int j;
3486 bool first = true;
3488 if (ipa_get_scalar_lat (info, i)->bottom || known_csts[i])
3489 continue;
3491 FOR_EACH_VEC_ELT (callers, j, cs)
3493 struct ipa_jump_func *jump_func;
3494 tree t;
3496 if (i >= ipa_get_cs_argument_count (IPA_EDGE_REF (cs)))
3498 newval = NULL_TREE;
3499 break;
3501 jump_func = ipa_get_ith_jump_func (IPA_EDGE_REF (cs), i);
3502 t = ipa_value_from_jfunc (IPA_NODE_REF (cs->caller), jump_func);
3503 if (!t
3504 || (newval
3505 && !values_equal_for_ipcp_p (t, newval))
3506 || (!first && !newval))
3508 newval = NULL_TREE;
3509 break;
3511 else
3512 newval = t;
3513 first = false;
3516 if (newval)
3518 if (dump_file && (dump_flags & TDF_DETAILS))
3520 fprintf (dump_file, " adding an extra known scalar value ");
3521 print_ipcp_constant_value (dump_file, newval);
3522 fprintf (dump_file, " for ");
3523 ipa_dump_param (dump_file, info, i);
3524 fprintf (dump_file, "\n");
3527 known_csts[i] = newval;
3532 /* Given a NODE and a subset of its CALLERS, try to populate plank slots in
3533 KNOWN_CONTEXTS with polymorphic contexts that are also known for all of the
3534 CALLERS. */
3536 static void
3537 find_more_contexts_for_caller_subset (cgraph_node *node,
3538 vec<ipa_polymorphic_call_context>
3539 *known_contexts,
3540 vec<cgraph_edge *> callers)
3542 ipa_node_params *info = IPA_NODE_REF (node);
3543 int i, count = ipa_get_param_count (info);
3545 for (i = 0; i < count ; i++)
3547 cgraph_edge *cs;
3549 if (ipa_get_poly_ctx_lat (info, i)->bottom
3550 || (known_contexts->exists ()
3551 && !(*known_contexts)[i].useless_p ()))
3552 continue;
3554 ipa_polymorphic_call_context newval;
3555 bool first = true;
3556 int j;
3558 FOR_EACH_VEC_ELT (callers, j, cs)
3560 if (i >= ipa_get_cs_argument_count (IPA_EDGE_REF (cs)))
3561 return;
3562 ipa_jump_func *jfunc = ipa_get_ith_jump_func (IPA_EDGE_REF (cs),
3564 ipa_polymorphic_call_context ctx;
3565 ctx = ipa_context_from_jfunc (IPA_NODE_REF (cs->caller), cs, i,
3566 jfunc);
3567 if (first)
3569 newval = ctx;
3570 first = false;
3572 else
3573 newval.meet_with (ctx);
3574 if (newval.useless_p ())
3575 break;
3578 if (!newval.useless_p ())
3580 if (dump_file && (dump_flags & TDF_DETAILS))
3582 fprintf (dump_file, " adding an extra known polymorphic "
3583 "context ");
3584 print_ipcp_constant_value (dump_file, newval);
3585 fprintf (dump_file, " for ");
3586 ipa_dump_param (dump_file, info, i);
3587 fprintf (dump_file, "\n");
3590 if (!known_contexts->exists ())
3591 known_contexts->safe_grow_cleared (ipa_get_param_count (info));
3592 (*known_contexts)[i] = newval;
3598 /* Go through PLATS and create a vector of values consisting of values and
3599 offsets (minus OFFSET) of lattices that contain only a single value. */
3601 static vec<ipa_agg_jf_item>
3602 copy_plats_to_inter (struct ipcp_param_lattices *plats, HOST_WIDE_INT offset)
3604 vec<ipa_agg_jf_item> res = vNULL;
3606 if (!plats->aggs || plats->aggs_contain_variable || plats->aggs_bottom)
3607 return vNULL;
3609 for (struct ipcp_agg_lattice *aglat = plats->aggs; aglat; aglat = aglat->next)
3610 if (aglat->is_single_const ())
3612 struct ipa_agg_jf_item ti;
3613 ti.offset = aglat->offset - offset;
3614 ti.value = aglat->values->value;
3615 res.safe_push (ti);
3617 return res;
3620 /* Intersect all values in INTER with single value lattices in PLATS (while
3621 subtracting OFFSET). */
3623 static void
3624 intersect_with_plats (struct ipcp_param_lattices *plats,
3625 vec<ipa_agg_jf_item> *inter,
3626 HOST_WIDE_INT offset)
3628 struct ipcp_agg_lattice *aglat;
3629 struct ipa_agg_jf_item *item;
3630 int k;
3632 if (!plats->aggs || plats->aggs_contain_variable || plats->aggs_bottom)
3634 inter->release ();
3635 return;
3638 aglat = plats->aggs;
3639 FOR_EACH_VEC_ELT (*inter, k, item)
3641 bool found = false;
3642 if (!item->value)
3643 continue;
3644 while (aglat)
3646 if (aglat->offset - offset > item->offset)
3647 break;
3648 if (aglat->offset - offset == item->offset)
3650 gcc_checking_assert (item->value);
3651 if (values_equal_for_ipcp_p (item->value, aglat->values->value))
3652 found = true;
3653 break;
3655 aglat = aglat->next;
3657 if (!found)
3658 item->value = NULL_TREE;
3662 /* Copy agggregate replacement values of NODE (which is an IPA-CP clone) to the
3663 vector result while subtracting OFFSET from the individual value offsets. */
3665 static vec<ipa_agg_jf_item>
3666 agg_replacements_to_vector (struct cgraph_node *node, int index,
3667 HOST_WIDE_INT offset)
3669 struct ipa_agg_replacement_value *av;
3670 vec<ipa_agg_jf_item> res = vNULL;
3672 for (av = ipa_get_agg_replacements_for_node (node); av; av = av->next)
3673 if (av->index == index
3674 && (av->offset - offset) >= 0)
3676 struct ipa_agg_jf_item item;
3677 gcc_checking_assert (av->value);
3678 item.offset = av->offset - offset;
3679 item.value = av->value;
3680 res.safe_push (item);
3683 return res;
3686 /* Intersect all values in INTER with those that we have already scheduled to
3687 be replaced in parameter number INDEX of NODE, which is an IPA-CP clone
3688 (while subtracting OFFSET). */
3690 static void
3691 intersect_with_agg_replacements (struct cgraph_node *node, int index,
3692 vec<ipa_agg_jf_item> *inter,
3693 HOST_WIDE_INT offset)
3695 struct ipa_agg_replacement_value *srcvals;
3696 struct ipa_agg_jf_item *item;
3697 int i;
3699 srcvals = ipa_get_agg_replacements_for_node (node);
3700 if (!srcvals)
3702 inter->release ();
3703 return;
3706 FOR_EACH_VEC_ELT (*inter, i, item)
3708 struct ipa_agg_replacement_value *av;
3709 bool found = false;
3710 if (!item->value)
3711 continue;
3712 for (av = srcvals; av; av = av->next)
3714 gcc_checking_assert (av->value);
3715 if (av->index == index
3716 && av->offset - offset == item->offset)
3718 if (values_equal_for_ipcp_p (item->value, av->value))
3719 found = true;
3720 break;
3723 if (!found)
3724 item->value = NULL_TREE;
3728 /* Intersect values in INTER with aggregate values that come along edge CS to
3729 parameter number INDEX and return it. If INTER does not actually exist yet,
3730 copy all incoming values to it. If we determine we ended up with no values
3731 whatsoever, return a released vector. */
3733 static vec<ipa_agg_jf_item>
3734 intersect_aggregates_with_edge (struct cgraph_edge *cs, int index,
3735 vec<ipa_agg_jf_item> inter)
3737 struct ipa_jump_func *jfunc;
3738 jfunc = ipa_get_ith_jump_func (IPA_EDGE_REF (cs), index);
3739 if (jfunc->type == IPA_JF_PASS_THROUGH
3740 && ipa_get_jf_pass_through_operation (jfunc) == NOP_EXPR)
3742 struct ipa_node_params *caller_info = IPA_NODE_REF (cs->caller);
3743 int src_idx = ipa_get_jf_pass_through_formal_id (jfunc);
3745 if (caller_info->ipcp_orig_node)
3747 struct cgraph_node *orig_node = caller_info->ipcp_orig_node;
3748 struct ipcp_param_lattices *orig_plats;
3749 orig_plats = ipa_get_parm_lattices (IPA_NODE_REF (orig_node),
3750 src_idx);
3751 if (agg_pass_through_permissible_p (orig_plats, jfunc))
3753 if (!inter.exists ())
3754 inter = agg_replacements_to_vector (cs->caller, src_idx, 0);
3755 else
3756 intersect_with_agg_replacements (cs->caller, src_idx,
3757 &inter, 0);
3759 else
3761 inter.release ();
3762 return vNULL;
3765 else
3767 struct ipcp_param_lattices *src_plats;
3768 src_plats = ipa_get_parm_lattices (caller_info, src_idx);
3769 if (agg_pass_through_permissible_p (src_plats, jfunc))
3771 /* Currently we do not produce clobber aggregate jump
3772 functions, adjust when we do. */
3773 gcc_checking_assert (!jfunc->agg.items);
3774 if (!inter.exists ())
3775 inter = copy_plats_to_inter (src_plats, 0);
3776 else
3777 intersect_with_plats (src_plats, &inter, 0);
3779 else
3781 inter.release ();
3782 return vNULL;
3786 else if (jfunc->type == IPA_JF_ANCESTOR
3787 && ipa_get_jf_ancestor_agg_preserved (jfunc))
3789 struct ipa_node_params *caller_info = IPA_NODE_REF (cs->caller);
3790 int src_idx = ipa_get_jf_ancestor_formal_id (jfunc);
3791 struct ipcp_param_lattices *src_plats;
3792 HOST_WIDE_INT delta = ipa_get_jf_ancestor_offset (jfunc);
3794 if (caller_info->ipcp_orig_node)
3796 if (!inter.exists ())
3797 inter = agg_replacements_to_vector (cs->caller, src_idx, delta);
3798 else
3799 intersect_with_agg_replacements (cs->caller, src_idx, &inter,
3800 delta);
3802 else
3804 src_plats = ipa_get_parm_lattices (caller_info, src_idx);;
3805 /* Currently we do not produce clobber aggregate jump
3806 functions, adjust when we do. */
3807 gcc_checking_assert (!src_plats->aggs || !jfunc->agg.items);
3808 if (!inter.exists ())
3809 inter = copy_plats_to_inter (src_plats, delta);
3810 else
3811 intersect_with_plats (src_plats, &inter, delta);
3814 else if (jfunc->agg.items)
3816 struct ipa_agg_jf_item *item;
3817 int k;
3819 if (!inter.exists ())
3820 for (unsigned i = 0; i < jfunc->agg.items->length (); i++)
3821 inter.safe_push ((*jfunc->agg.items)[i]);
3822 else
3823 FOR_EACH_VEC_ELT (inter, k, item)
3825 int l = 0;
3826 bool found = false;;
3828 if (!item->value)
3829 continue;
3831 while ((unsigned) l < jfunc->agg.items->length ())
3833 struct ipa_agg_jf_item *ti;
3834 ti = &(*jfunc->agg.items)[l];
3835 if (ti->offset > item->offset)
3836 break;
3837 if (ti->offset == item->offset)
3839 gcc_checking_assert (ti->value);
3840 if (values_equal_for_ipcp_p (item->value,
3841 ti->value))
3842 found = true;
3843 break;
3845 l++;
3847 if (!found)
3848 item->value = NULL;
3851 else
3853 inter.release ();
3854 return vec<ipa_agg_jf_item>();
3856 return inter;
3859 /* Look at edges in CALLERS and collect all known aggregate values that arrive
3860 from all of them. */
3862 static struct ipa_agg_replacement_value *
3863 find_aggregate_values_for_callers_subset (struct cgraph_node *node,
3864 vec<cgraph_edge *> callers)
3866 struct ipa_node_params *dest_info = IPA_NODE_REF (node);
3867 struct ipa_agg_replacement_value *res;
3868 struct ipa_agg_replacement_value **tail = &res;
3869 struct cgraph_edge *cs;
3870 int i, j, count = ipa_get_param_count (dest_info);
3872 FOR_EACH_VEC_ELT (callers, j, cs)
3874 int c = ipa_get_cs_argument_count (IPA_EDGE_REF (cs));
3875 if (c < count)
3876 count = c;
3879 for (i = 0; i < count ; i++)
3881 struct cgraph_edge *cs;
3882 vec<ipa_agg_jf_item> inter = vNULL;
3883 struct ipa_agg_jf_item *item;
3884 struct ipcp_param_lattices *plats = ipa_get_parm_lattices (dest_info, i);
3885 int j;
3887 /* Among other things, the following check should deal with all by_ref
3888 mismatches. */
3889 if (plats->aggs_bottom)
3890 continue;
3892 FOR_EACH_VEC_ELT (callers, j, cs)
3894 inter = intersect_aggregates_with_edge (cs, i, inter);
3896 if (!inter.exists ())
3897 goto next_param;
3900 FOR_EACH_VEC_ELT (inter, j, item)
3902 struct ipa_agg_replacement_value *v;
3904 if (!item->value)
3905 continue;
3907 v = ggc_alloc<ipa_agg_replacement_value> ();
3908 v->index = i;
3909 v->offset = item->offset;
3910 v->value = item->value;
3911 v->by_ref = plats->aggs_by_ref;
3912 *tail = v;
3913 tail = &v->next;
3916 next_param:
3917 if (inter.exists ())
3918 inter.release ();
3920 *tail = NULL;
3921 return res;
3924 /* Turn KNOWN_AGGS into a list of aggreate replacement values. */
3926 static struct ipa_agg_replacement_value *
3927 known_aggs_to_agg_replacement_list (vec<ipa_agg_jump_function> known_aggs)
3929 struct ipa_agg_replacement_value *res;
3930 struct ipa_agg_replacement_value **tail = &res;
3931 struct ipa_agg_jump_function *aggjf;
3932 struct ipa_agg_jf_item *item;
3933 int i, j;
3935 FOR_EACH_VEC_ELT (known_aggs, i, aggjf)
3936 FOR_EACH_VEC_SAFE_ELT (aggjf->items, j, item)
3938 struct ipa_agg_replacement_value *v;
3939 v = ggc_alloc<ipa_agg_replacement_value> ();
3940 v->index = i;
3941 v->offset = item->offset;
3942 v->value = item->value;
3943 v->by_ref = aggjf->by_ref;
3944 *tail = v;
3945 tail = &v->next;
3947 *tail = NULL;
3948 return res;
3951 /* Determine whether CS also brings all scalar values that the NODE is
3952 specialized for. */
3954 static bool
3955 cgraph_edge_brings_all_scalars_for_node (struct cgraph_edge *cs,
3956 struct cgraph_node *node)
3958 struct ipa_node_params *dest_info = IPA_NODE_REF (node);
3959 int count = ipa_get_param_count (dest_info);
3960 struct ipa_node_params *caller_info;
3961 struct ipa_edge_args *args;
3962 int i;
3964 caller_info = IPA_NODE_REF (cs->caller);
3965 args = IPA_EDGE_REF (cs);
3966 for (i = 0; i < count; i++)
3968 struct ipa_jump_func *jump_func;
3969 tree val, t;
3971 val = dest_info->known_csts[i];
3972 if (!val)
3973 continue;
3975 if (i >= ipa_get_cs_argument_count (args))
3976 return false;
3977 jump_func = ipa_get_ith_jump_func (args, i);
3978 t = ipa_value_from_jfunc (caller_info, jump_func);
3979 if (!t || !values_equal_for_ipcp_p (val, t))
3980 return false;
3982 return true;
3985 /* Determine whether CS also brings all aggregate values that NODE is
3986 specialized for. */
3987 static bool
3988 cgraph_edge_brings_all_agg_vals_for_node (struct cgraph_edge *cs,
3989 struct cgraph_node *node)
3991 struct ipa_node_params *orig_caller_info = IPA_NODE_REF (cs->caller);
3992 struct ipa_node_params *orig_node_info;
3993 struct ipa_agg_replacement_value *aggval;
3994 int i, ec, count;
3996 aggval = ipa_get_agg_replacements_for_node (node);
3997 if (!aggval)
3998 return true;
4000 count = ipa_get_param_count (IPA_NODE_REF (node));
4001 ec = ipa_get_cs_argument_count (IPA_EDGE_REF (cs));
4002 if (ec < count)
4003 for (struct ipa_agg_replacement_value *av = aggval; av; av = av->next)
4004 if (aggval->index >= ec)
4005 return false;
4007 orig_node_info = IPA_NODE_REF (IPA_NODE_REF (node)->ipcp_orig_node);
4008 if (orig_caller_info->ipcp_orig_node)
4009 orig_caller_info = IPA_NODE_REF (orig_caller_info->ipcp_orig_node);
4011 for (i = 0; i < count; i++)
4013 static vec<ipa_agg_jf_item> values = vec<ipa_agg_jf_item>();
4014 struct ipcp_param_lattices *plats;
4015 bool interesting = false;
4016 for (struct ipa_agg_replacement_value *av = aggval; av; av = av->next)
4017 if (aggval->index == i)
4019 interesting = true;
4020 break;
4022 if (!interesting)
4023 continue;
4025 plats = ipa_get_parm_lattices (orig_node_info, aggval->index);
4026 if (plats->aggs_bottom)
4027 return false;
4029 values = intersect_aggregates_with_edge (cs, i, values);
4030 if (!values.exists ())
4031 return false;
4033 for (struct ipa_agg_replacement_value *av = aggval; av; av = av->next)
4034 if (aggval->index == i)
4036 struct ipa_agg_jf_item *item;
4037 int j;
4038 bool found = false;
4039 FOR_EACH_VEC_ELT (values, j, item)
4040 if (item->value
4041 && item->offset == av->offset
4042 && values_equal_for_ipcp_p (item->value, av->value))
4044 found = true;
4045 break;
4047 if (!found)
4049 values.release ();
4050 return false;
4054 return true;
4057 /* Given an original NODE and a VAL for which we have already created a
4058 specialized clone, look whether there are incoming edges that still lead
4059 into the old node but now also bring the requested value and also conform to
4060 all other criteria such that they can be redirected the special node.
4061 This function can therefore redirect the final edge in a SCC. */
4063 template <typename valtype>
4064 static void
4065 perhaps_add_new_callers (cgraph_node *node, ipcp_value<valtype> *val)
4067 ipcp_value_source<valtype> *src;
4068 gcov_type redirected_sum = 0;
4070 for (src = val->sources; src; src = src->next)
4072 struct cgraph_edge *cs = src->cs;
4073 while (cs)
4075 if (cgraph_edge_brings_value_p (cs, src, node)
4076 && cgraph_edge_brings_all_scalars_for_node (cs, val->spec_node)
4077 && cgraph_edge_brings_all_agg_vals_for_node (cs, val->spec_node))
4079 if (dump_file)
4080 fprintf (dump_file, " - adding an extra caller %s/%i"
4081 " of %s/%i\n",
4082 xstrdup_for_dump (cs->caller->name ()),
4083 cs->caller->order,
4084 xstrdup_for_dump (val->spec_node->name ()),
4085 val->spec_node->order);
4087 cs->redirect_callee_duplicating_thunks (val->spec_node);
4088 val->spec_node->expand_all_artificial_thunks ();
4089 redirected_sum += cs->count;
4091 cs = get_next_cgraph_edge_clone (cs);
4095 if (redirected_sum)
4096 update_specialized_profile (val->spec_node, node, redirected_sum);
4099 /* Return true if KNOWN_CONTEXTS contain at least one useful context. */
4101 static bool
4102 known_contexts_useful_p (vec<ipa_polymorphic_call_context> known_contexts)
4104 ipa_polymorphic_call_context *ctx;
4105 int i;
4107 FOR_EACH_VEC_ELT (known_contexts, i, ctx)
4108 if (!ctx->useless_p ())
4109 return true;
4110 return false;
4113 /* Return a copy of KNOWN_CSTS if it is not empty, otherwise return vNULL. */
4115 static vec<ipa_polymorphic_call_context>
4116 copy_useful_known_contexts (vec<ipa_polymorphic_call_context> known_contexts)
4118 if (known_contexts_useful_p (known_contexts))
4119 return known_contexts.copy ();
4120 else
4121 return vNULL;
4124 /* Copy KNOWN_CSTS and modify the copy according to VAL and INDEX. If
4125 non-empty, replace KNOWN_CONTEXTS with its copy too. */
4127 static void
4128 modify_known_vectors_with_val (vec<tree> *known_csts,
4129 vec<ipa_polymorphic_call_context> *known_contexts,
4130 ipcp_value<tree> *val,
4131 int index)
4133 *known_csts = known_csts->copy ();
4134 *known_contexts = copy_useful_known_contexts (*known_contexts);
4135 (*known_csts)[index] = val->value;
4138 /* Replace KNOWN_CSTS with its copy. Also copy KNOWN_CONTEXTS and modify the
4139 copy according to VAL and INDEX. */
4141 static void
4142 modify_known_vectors_with_val (vec<tree> *known_csts,
4143 vec<ipa_polymorphic_call_context> *known_contexts,
4144 ipcp_value<ipa_polymorphic_call_context> *val,
4145 int index)
4147 *known_csts = known_csts->copy ();
4148 *known_contexts = known_contexts->copy ();
4149 (*known_contexts)[index] = val->value;
4152 /* Return true if OFFSET indicates this was not an aggregate value or there is
4153 a replacement equivalent to VALUE, INDEX and OFFSET among those in the
4154 AGGVALS list. */
4156 DEBUG_FUNCTION bool
4157 ipcp_val_agg_replacement_ok_p (ipa_agg_replacement_value *aggvals,
4158 int index, HOST_WIDE_INT offset, tree value)
4160 if (offset == -1)
4161 return true;
4163 while (aggvals)
4165 if (aggvals->index == index
4166 && aggvals->offset == offset
4167 && values_equal_for_ipcp_p (aggvals->value, value))
4168 return true;
4169 aggvals = aggvals->next;
4171 return false;
4174 /* Return true if offset is minus one because source of a polymorphic contect
4175 cannot be an aggregate value. */
4177 DEBUG_FUNCTION bool
4178 ipcp_val_agg_replacement_ok_p (ipa_agg_replacement_value *,
4179 int , HOST_WIDE_INT offset,
4180 ipa_polymorphic_call_context)
4182 return offset == -1;
4185 /* Decide wheter to create a special version of NODE for value VAL of parameter
4186 at the given INDEX. If OFFSET is -1, the value is for the parameter itself,
4187 otherwise it is stored at the given OFFSET of the parameter. KNOWN_CSTS,
4188 KNOWN_CONTEXTS and KNOWN_AGGS describe the other already known values. */
4190 template <typename valtype>
4191 static bool
4192 decide_about_value (struct cgraph_node *node, int index, HOST_WIDE_INT offset,
4193 ipcp_value<valtype> *val, vec<tree> known_csts,
4194 vec<ipa_polymorphic_call_context> known_contexts)
4196 struct ipa_agg_replacement_value *aggvals;
4197 int freq_sum, caller_count;
4198 gcov_type count_sum;
4199 vec<cgraph_edge *> callers;
4201 if (val->spec_node)
4203 perhaps_add_new_callers (node, val);
4204 return false;
4206 else if (val->local_size_cost + overall_size > max_new_size)
4208 if (dump_file && (dump_flags & TDF_DETAILS))
4209 fprintf (dump_file, " Ignoring candidate value because "
4210 "max_new_size would be reached with %li.\n",
4211 val->local_size_cost + overall_size);
4212 return false;
4214 else if (!get_info_about_necessary_edges (val, node, &freq_sum, &count_sum,
4215 &caller_count))
4216 return false;
4218 if (dump_file && (dump_flags & TDF_DETAILS))
4220 fprintf (dump_file, " - considering value ");
4221 print_ipcp_constant_value (dump_file, val->value);
4222 fprintf (dump_file, " for ");
4223 ipa_dump_param (dump_file, IPA_NODE_REF (node), index);
4224 if (offset != -1)
4225 fprintf (dump_file, ", offset: " HOST_WIDE_INT_PRINT_DEC, offset);
4226 fprintf (dump_file, " (caller_count: %i)\n", caller_count);
4229 if (!good_cloning_opportunity_p (node, val->local_time_benefit,
4230 freq_sum, count_sum,
4231 val->local_size_cost)
4232 && !good_cloning_opportunity_p (node,
4233 val->local_time_benefit
4234 + val->prop_time_benefit,
4235 freq_sum, count_sum,
4236 val->local_size_cost
4237 + val->prop_size_cost))
4238 return false;
4240 if (dump_file)
4241 fprintf (dump_file, " Creating a specialized node of %s/%i.\n",
4242 node->name (), node->order);
4244 callers = gather_edges_for_value (val, node, caller_count);
4245 if (offset == -1)
4246 modify_known_vectors_with_val (&known_csts, &known_contexts, val, index);
4247 else
4249 known_csts = known_csts.copy ();
4250 known_contexts = copy_useful_known_contexts (known_contexts);
4252 find_more_scalar_values_for_callers_subset (node, known_csts, callers);
4253 find_more_contexts_for_caller_subset (node, &known_contexts, callers);
4254 aggvals = find_aggregate_values_for_callers_subset (node, callers);
4255 gcc_checking_assert (ipcp_val_agg_replacement_ok_p (aggvals, index,
4256 offset, val->value));
4257 val->spec_node = create_specialized_node (node, known_csts, known_contexts,
4258 aggvals, callers);
4259 overall_size += val->local_size_cost;
4261 /* TODO: If for some lattice there is only one other known value
4262 left, make a special node for it too. */
4264 return true;
4267 /* Decide whether and what specialized clones of NODE should be created. */
4269 static bool
4270 decide_whether_version_node (struct cgraph_node *node)
4272 struct ipa_node_params *info = IPA_NODE_REF (node);
4273 int i, count = ipa_get_param_count (info);
4274 vec<tree> known_csts;
4275 vec<ipa_polymorphic_call_context> known_contexts;
4276 vec<ipa_agg_jump_function> known_aggs = vNULL;
4277 bool ret = false;
4279 if (count == 0)
4280 return false;
4282 if (dump_file && (dump_flags & TDF_DETAILS))
4283 fprintf (dump_file, "\nEvaluating opportunities for %s/%i.\n",
4284 node->name (), node->order);
4286 gather_context_independent_values (info, &known_csts, &known_contexts,
4287 info->do_clone_for_all_contexts ? &known_aggs
4288 : NULL, NULL);
4290 for (i = 0; i < count ;i++)
4292 struct ipcp_param_lattices *plats = ipa_get_parm_lattices (info, i);
4293 ipcp_lattice<tree> *lat = &plats->itself;
4294 ipcp_lattice<ipa_polymorphic_call_context> *ctxlat = &plats->ctxlat;
4296 if (!lat->bottom
4297 && !known_csts[i])
4299 ipcp_value<tree> *val;
4300 for (val = lat->values; val; val = val->next)
4301 ret |= decide_about_value (node, i, -1, val, known_csts,
4302 known_contexts);
4305 if (!plats->aggs_bottom)
4307 struct ipcp_agg_lattice *aglat;
4308 ipcp_value<tree> *val;
4309 for (aglat = plats->aggs; aglat; aglat = aglat->next)
4310 if (!aglat->bottom && aglat->values
4311 /* If the following is false, the one value is in
4312 known_aggs. */
4313 && (plats->aggs_contain_variable
4314 || !aglat->is_single_const ()))
4315 for (val = aglat->values; val; val = val->next)
4316 ret |= decide_about_value (node, i, aglat->offset, val,
4317 known_csts, known_contexts);
4320 if (!ctxlat->bottom
4321 && known_contexts[i].useless_p ())
4323 ipcp_value<ipa_polymorphic_call_context> *val;
4324 for (val = ctxlat->values; val; val = val->next)
4325 ret |= decide_about_value (node, i, -1, val, known_csts,
4326 known_contexts);
4329 info = IPA_NODE_REF (node);
4332 if (info->do_clone_for_all_contexts)
4334 struct cgraph_node *clone;
4335 vec<cgraph_edge *> callers;
4337 if (dump_file)
4338 fprintf (dump_file, " - Creating a specialized node of %s/%i "
4339 "for all known contexts.\n", node->name (),
4340 node->order);
4342 callers = node->collect_callers ();
4344 if (!known_contexts_useful_p (known_contexts))
4346 known_contexts.release ();
4347 known_contexts = vNULL;
4349 clone = create_specialized_node (node, known_csts, known_contexts,
4350 known_aggs_to_agg_replacement_list (known_aggs),
4351 callers);
4352 info = IPA_NODE_REF (node);
4353 info->do_clone_for_all_contexts = false;
4354 IPA_NODE_REF (clone)->is_all_contexts_clone = true;
4355 for (i = 0; i < count ; i++)
4356 vec_free (known_aggs[i].items);
4357 known_aggs.release ();
4358 ret = true;
4360 else
4362 known_csts.release ();
4363 known_contexts.release ();
4366 return ret;
4369 /* Transitively mark all callees of NODE within the same SCC as not dead. */
4371 static void
4372 spread_undeadness (struct cgraph_node *node)
4374 struct cgraph_edge *cs;
4376 for (cs = node->callees; cs; cs = cs->next_callee)
4377 if (ipa_edge_within_scc (cs))
4379 struct cgraph_node *callee;
4380 struct ipa_node_params *info;
4382 callee = cs->callee->function_symbol (NULL);
4383 info = IPA_NODE_REF (callee);
4385 if (info->node_dead)
4387 info->node_dead = 0;
4388 spread_undeadness (callee);
4393 /* Return true if NODE has a caller from outside of its SCC that is not
4394 dead. Worker callback for cgraph_for_node_and_aliases. */
4396 static bool
4397 has_undead_caller_from_outside_scc_p (struct cgraph_node *node,
4398 void *data ATTRIBUTE_UNUSED)
4400 struct cgraph_edge *cs;
4402 for (cs = node->callers; cs; cs = cs->next_caller)
4403 if (cs->caller->thunk.thunk_p
4404 && cs->caller->call_for_symbol_thunks_and_aliases
4405 (has_undead_caller_from_outside_scc_p, NULL, true))
4406 return true;
4407 else if (!ipa_edge_within_scc (cs)
4408 && !IPA_NODE_REF (cs->caller)->node_dead)
4409 return true;
4410 return false;
4414 /* Identify nodes within the same SCC as NODE which are no longer needed
4415 because of new clones and will be removed as unreachable. */
4417 static void
4418 identify_dead_nodes (struct cgraph_node *node)
4420 struct cgraph_node *v;
4421 for (v = node; v ; v = ((struct ipa_dfs_info *) v->aux)->next_cycle)
4422 if (v->will_be_removed_from_program_if_no_direct_calls_p ()
4423 && !v->call_for_symbol_thunks_and_aliases
4424 (has_undead_caller_from_outside_scc_p, NULL, true))
4425 IPA_NODE_REF (v)->node_dead = 1;
4427 for (v = node; v ; v = ((struct ipa_dfs_info *) v->aux)->next_cycle)
4428 if (!IPA_NODE_REF (v)->node_dead)
4429 spread_undeadness (v);
4431 if (dump_file && (dump_flags & TDF_DETAILS))
4433 for (v = node; v ; v = ((struct ipa_dfs_info *) v->aux)->next_cycle)
4434 if (IPA_NODE_REF (v)->node_dead)
4435 fprintf (dump_file, " Marking node as dead: %s/%i.\n",
4436 v->name (), v->order);
4440 /* The decision stage. Iterate over the topological order of call graph nodes
4441 TOPO and make specialized clones if deemed beneficial. */
4443 static void
4444 ipcp_decision_stage (struct ipa_topo_info *topo)
4446 int i;
4448 if (dump_file)
4449 fprintf (dump_file, "\nIPA decision stage:\n\n");
4451 for (i = topo->nnodes - 1; i >= 0; i--)
4453 struct cgraph_node *node = topo->order[i];
4454 bool change = false, iterate = true;
4456 while (iterate)
4458 struct cgraph_node *v;
4459 iterate = false;
4460 for (v = node; v ; v = ((struct ipa_dfs_info *) v->aux)->next_cycle)
4461 if (v->has_gimple_body_p ()
4462 && ipcp_versionable_function_p (v))
4463 iterate |= decide_whether_version_node (v);
4465 change |= iterate;
4467 if (change)
4468 identify_dead_nodes (node);
4472 /* Look up all alignment information that we have discovered and copy it over
4473 to the transformation summary. */
4475 static void
4476 ipcp_store_alignment_results (void)
4478 cgraph_node *node;
4480 FOR_EACH_FUNCTION_WITH_GIMPLE_BODY (node)
4482 ipa_node_params *info = IPA_NODE_REF (node);
4483 bool dumped_sth = false;
4484 bool found_useful_result = false;
4486 if (!opt_for_fn (node->decl, flag_ipa_cp_alignment))
4488 if (dump_file)
4489 fprintf (dump_file, "Not considering %s for alignment discovery "
4490 "and propagate; -fipa-cp-alignment: disabled.\n",
4491 node->name ());
4492 continue;
4495 if (info->ipcp_orig_node)
4496 info = IPA_NODE_REF (info->ipcp_orig_node);
4498 unsigned count = ipa_get_param_count (info);
4499 for (unsigned i = 0; i < count ; i++)
4501 ipcp_param_lattices *plats = ipa_get_parm_lattices (info, i);
4502 if (!plats->alignment.bottom_p ()
4503 && !plats->alignment.top_p ())
4505 gcc_checking_assert (plats->alignment.align > 0);
4506 found_useful_result = true;
4507 break;
4510 if (!found_useful_result)
4511 continue;
4513 ipcp_grow_transformations_if_necessary ();
4514 ipcp_transformation_summary *ts = ipcp_get_transformation_summary (node);
4515 vec_safe_reserve_exact (ts->alignments, count);
4517 for (unsigned i = 0; i < count ; i++)
4519 ipcp_param_lattices *plats = ipa_get_parm_lattices (info, i);
4520 ipa_alignment al;
4522 if (!plats->alignment.bottom_p ()
4523 && !plats->alignment.top_p ())
4525 al.known = true;
4526 al.align = plats->alignment.align;
4527 al.misalign = plats->alignment.misalign;
4529 else
4530 al.known = false;
4532 ts->alignments->quick_push (al);
4533 if (!dump_file || !al.known)
4534 continue;
4535 if (!dumped_sth)
4537 fprintf (dump_file, "Propagated alignment info for function %s/%i:\n",
4538 node->name (), node->order);
4539 dumped_sth = true;
4541 fprintf (dump_file, " param %i: align: %u, misalign: %u\n",
4542 i, al.align, al.misalign);
4547 /* The IPCP driver. */
4549 static unsigned int
4550 ipcp_driver (void)
4552 struct cgraph_2edge_hook_list *edge_duplication_hook_holder;
4553 struct cgraph_edge_hook_list *edge_removal_hook_holder;
4554 struct ipa_topo_info topo;
4556 ipa_check_create_node_params ();
4557 ipa_check_create_edge_args ();
4558 grow_edge_clone_vectors ();
4559 edge_duplication_hook_holder =
4560 symtab->add_edge_duplication_hook (&ipcp_edge_duplication_hook, NULL);
4561 edge_removal_hook_holder =
4562 symtab->add_edge_removal_hook (&ipcp_edge_removal_hook, NULL);
4564 if (dump_file)
4566 fprintf (dump_file, "\nIPA structures before propagation:\n");
4567 if (dump_flags & TDF_DETAILS)
4568 ipa_print_all_params (dump_file);
4569 ipa_print_all_jump_functions (dump_file);
4572 /* Topological sort. */
4573 build_toporder_info (&topo);
4574 /* Do the interprocedural propagation. */
4575 ipcp_propagate_stage (&topo);
4576 /* Decide what constant propagation and cloning should be performed. */
4577 ipcp_decision_stage (&topo);
4578 /* Store results of alignment propagation. */
4579 ipcp_store_alignment_results ();
4581 /* Free all IPCP structures. */
4582 free_toporder_info (&topo);
4583 next_edge_clone.release ();
4584 prev_edge_clone.release ();
4585 symtab->remove_edge_removal_hook (edge_removal_hook_holder);
4586 symtab->remove_edge_duplication_hook (edge_duplication_hook_holder);
4587 ipa_free_all_structures_after_ipa_cp ();
4588 if (dump_file)
4589 fprintf (dump_file, "\nIPA constant propagation end\n");
4590 return 0;
4593 /* Initialization and computation of IPCP data structures. This is the initial
4594 intraprocedural analysis of functions, which gathers information to be
4595 propagated later on. */
4597 static void
4598 ipcp_generate_summary (void)
4600 struct cgraph_node *node;
4602 if (dump_file)
4603 fprintf (dump_file, "\nIPA constant propagation start:\n");
4604 ipa_register_cgraph_hooks ();
4606 FOR_EACH_FUNCTION_WITH_GIMPLE_BODY (node)
4607 ipa_analyze_node (node);
4610 /* Write ipcp summary for nodes in SET. */
4612 static void
4613 ipcp_write_summary (void)
4615 ipa_prop_write_jump_functions ();
4618 /* Read ipcp summary. */
4620 static void
4621 ipcp_read_summary (void)
4623 ipa_prop_read_jump_functions ();
4626 namespace {
4628 const pass_data pass_data_ipa_cp =
4630 IPA_PASS, /* type */
4631 "cp", /* name */
4632 OPTGROUP_NONE, /* optinfo_flags */
4633 TV_IPA_CONSTANT_PROP, /* tv_id */
4634 0, /* properties_required */
4635 0, /* properties_provided */
4636 0, /* properties_destroyed */
4637 0, /* todo_flags_start */
4638 ( TODO_dump_symtab | TODO_remove_functions ), /* todo_flags_finish */
4641 class pass_ipa_cp : public ipa_opt_pass_d
4643 public:
4644 pass_ipa_cp (gcc::context *ctxt)
4645 : ipa_opt_pass_d (pass_data_ipa_cp, ctxt,
4646 ipcp_generate_summary, /* generate_summary */
4647 ipcp_write_summary, /* write_summary */
4648 ipcp_read_summary, /* read_summary */
4649 ipcp_write_transformation_summaries, /*
4650 write_optimization_summary */
4651 ipcp_read_transformation_summaries, /*
4652 read_optimization_summary */
4653 NULL, /* stmt_fixup */
4654 0, /* function_transform_todo_flags_start */
4655 ipcp_transform_function, /* function_transform */
4656 NULL) /* variable_transform */
4659 /* opt_pass methods: */
4660 virtual bool gate (function *)
4662 /* FIXME: We should remove the optimize check after we ensure we never run
4663 IPA passes when not optimizing. */
4664 return (flag_ipa_cp && optimize) || in_lto_p;
4667 virtual unsigned int execute (function *) { return ipcp_driver (); }
4669 }; // class pass_ipa_cp
4671 } // anon namespace
4673 ipa_opt_pass_d *
4674 make_pass_ipa_cp (gcc::context *ctxt)
4676 return new pass_ipa_cp (ctxt);
4679 /* Reset all state within ipa-cp.c so that we can rerun the compiler
4680 within the same process. For use by toplev::finalize. */
4682 void
4683 ipa_cp_c_finalize (void)
4685 max_count = 0;
4686 overall_size = 0;
4687 max_new_size = 0;