/cp
[official-gcc.git] / gcc / ipa-cp.c
blob5e1d94c67c731b6f95e9bb250533f7c03cee50e1
1 /* Interprocedural constant propagation
2 Copyright (C) 2005-2017 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 "tree-vrp.h"
118 #include "ipa-prop.h"
119 #include "tree-pretty-print.h"
120 #include "tree-inline.h"
121 #include "params.h"
122 #include "ipa-inline.h"
123 #include "ipa-utils.h"
124 #include "tree-ssa-ccp.h"
126 template <typename valtype> class ipcp_value;
128 /* Describes a particular source for an IPA-CP value. */
130 template <typename valtype>
131 class ipcp_value_source
133 public:
134 /* Aggregate offset of the source, negative if the source is scalar value of
135 the argument itself. */
136 HOST_WIDE_INT offset;
137 /* The incoming edge that brought the value. */
138 cgraph_edge *cs;
139 /* If the jump function that resulted into his value was a pass-through or an
140 ancestor, this is the ipcp_value of the caller from which the described
141 value has been derived. Otherwise it is NULL. */
142 ipcp_value<valtype> *val;
143 /* Next pointer in a linked list of sources of a value. */
144 ipcp_value_source *next;
145 /* If the jump function that resulted into his value was a pass-through or an
146 ancestor, this is the index of the parameter of the caller the jump
147 function references. */
148 int index;
151 /* Common ancestor for all ipcp_value instantiations. */
153 class ipcp_value_base
155 public:
156 /* Time benefit and size cost that specializing the function for this value
157 would bring about in this function alone. */
158 int local_time_benefit, local_size_cost;
159 /* Time benefit and size cost that specializing the function for this value
160 can bring about in it's callees (transitively). */
161 int prop_time_benefit, prop_size_cost;
164 /* Describes one particular value stored in struct ipcp_lattice. */
166 template <typename valtype>
167 class ipcp_value : public ipcp_value_base
169 public:
170 /* The actual value for the given parameter. */
171 valtype value;
172 /* The list of sources from which this value originates. */
173 ipcp_value_source <valtype> *sources;
174 /* Next pointers in a linked list of all values in a lattice. */
175 ipcp_value *next;
176 /* Next pointers in a linked list of values in a strongly connected component
177 of values. */
178 ipcp_value *scc_next;
179 /* Next pointers in a linked list of SCCs of values sorted topologically
180 according their sources. */
181 ipcp_value *topo_next;
182 /* A specialized node created for this value, NULL if none has been (so far)
183 created. */
184 cgraph_node *spec_node;
185 /* Depth first search number and low link for topological sorting of
186 values. */
187 int dfs, low_link;
188 /* True if this valye is currently on the topo-sort stack. */
189 bool on_stack;
191 void add_source (cgraph_edge *cs, ipcp_value *src_val, int src_idx,
192 HOST_WIDE_INT offset);
195 /* Lattice describing potential values of a formal parameter of a function, or
196 a part of an aggregate. TOP is represented by a lattice with zero values
197 and with contains_variable and bottom flags cleared. BOTTOM is represented
198 by a lattice with the bottom flag set. In that case, values and
199 contains_variable flag should be disregarded. */
201 template <typename valtype>
202 class ipcp_lattice
204 public:
205 /* The list of known values and types in this lattice. Note that values are
206 not deallocated if a lattice is set to bottom because there may be value
207 sources referencing them. */
208 ipcp_value<valtype> *values;
209 /* Number of known values and types in this lattice. */
210 int values_count;
211 /* The lattice contains a variable component (in addition to values). */
212 bool contains_variable;
213 /* The value of the lattice is bottom (i.e. variable and unusable for any
214 propagation). */
215 bool bottom;
217 inline bool is_single_const ();
218 inline bool set_to_bottom ();
219 inline bool set_contains_variable ();
220 bool add_value (valtype newval, cgraph_edge *cs,
221 ipcp_value<valtype> *src_val = NULL,
222 int src_idx = 0, HOST_WIDE_INT offset = -1);
223 void print (FILE * f, bool dump_sources, bool dump_benefits);
226 /* Lattice of tree values with an offset to describe a part of an
227 aggregate. */
229 class ipcp_agg_lattice : public ipcp_lattice<tree>
231 public:
232 /* Offset that is being described by this lattice. */
233 HOST_WIDE_INT offset;
234 /* Size so that we don't have to re-compute it every time we traverse the
235 list. Must correspond to TYPE_SIZE of all lat values. */
236 HOST_WIDE_INT size;
237 /* Next element of the linked list. */
238 struct ipcp_agg_lattice *next;
241 /* Lattice of known bits, only capable of holding one value.
242 Bitwise constant propagation propagates which bits of a
243 value are constant.
244 For eg:
245 int f(int x)
247 return some_op (x);
250 int f1(int y)
252 if (cond)
253 return f (y & 0xff);
254 else
255 return f (y & 0xf);
258 In the above case, the param 'x' will always have all
259 the bits (except the bits in lsb) set to 0.
260 Hence the mask of 'x' would be 0xff. The mask
261 reflects that the bits in lsb are unknown.
262 The actual propagated value is given by m_value & ~m_mask. */
264 class ipcp_bits_lattice
266 public:
267 bool bottom_p () { return m_lattice_val == IPA_BITS_VARYING; }
268 bool top_p () { return m_lattice_val == IPA_BITS_UNDEFINED; }
269 bool constant_p () { return m_lattice_val == IPA_BITS_CONSTANT; }
270 bool set_to_bottom ();
271 bool set_to_constant (widest_int, widest_int);
273 widest_int get_value () { return m_value; }
274 widest_int get_mask () { return m_mask; }
276 bool meet_with (ipcp_bits_lattice& other, unsigned, signop,
277 enum tree_code, tree);
279 bool meet_with (widest_int, widest_int, unsigned);
281 void print (FILE *);
283 private:
284 enum { IPA_BITS_UNDEFINED, IPA_BITS_CONSTANT, IPA_BITS_VARYING } m_lattice_val;
286 /* Similar to ccp_lattice_t, mask represents which bits of value are constant.
287 If a bit in mask is set to 0, then the corresponding bit in
288 value is known to be constant. */
289 widest_int m_value, m_mask;
291 bool meet_with_1 (widest_int, widest_int, unsigned);
292 void get_value_and_mask (tree, widest_int *, widest_int *);
295 /* Lattice of value ranges. */
297 class ipcp_vr_lattice
299 public:
300 value_range m_vr;
302 inline bool bottom_p () const;
303 inline bool top_p () const;
304 inline bool set_to_bottom ();
305 bool meet_with (const value_range *p_vr);
306 bool meet_with (const ipcp_vr_lattice &other);
307 void init () { m_vr.type = VR_UNDEFINED; }
308 void print (FILE * f);
310 private:
311 bool meet_with_1 (const value_range *other_vr);
314 /* Structure containing lattices for a parameter itself and for pieces of
315 aggregates that are passed in the parameter or by a reference in a parameter
316 plus some other useful flags. */
318 class ipcp_param_lattices
320 public:
321 /* Lattice describing the value of the parameter itself. */
322 ipcp_lattice<tree> itself;
323 /* Lattice describing the polymorphic contexts of a parameter. */
324 ipcp_lattice<ipa_polymorphic_call_context> ctxlat;
325 /* Lattices describing aggregate parts. */
326 ipcp_agg_lattice *aggs;
327 /* Lattice describing known bits. */
328 ipcp_bits_lattice bits_lattice;
329 /* Lattice describing value range. */
330 ipcp_vr_lattice m_value_range;
331 /* Number of aggregate lattices */
332 int aggs_count;
333 /* True if aggregate data were passed by reference (as opposed to by
334 value). */
335 bool aggs_by_ref;
336 /* All aggregate lattices contain a variable component (in addition to
337 values). */
338 bool aggs_contain_variable;
339 /* The value of all aggregate lattices is bottom (i.e. variable and unusable
340 for any propagation). */
341 bool aggs_bottom;
343 /* There is a virtual call based on this parameter. */
344 bool virt_call;
347 /* Allocation pools for values and their sources in ipa-cp. */
349 object_allocator<ipcp_value<tree> > ipcp_cst_values_pool
350 ("IPA-CP constant values");
352 object_allocator<ipcp_value<ipa_polymorphic_call_context> >
353 ipcp_poly_ctx_values_pool ("IPA-CP polymorphic contexts");
355 object_allocator<ipcp_value_source<tree> > ipcp_sources_pool
356 ("IPA-CP value sources");
358 object_allocator<ipcp_agg_lattice> ipcp_agg_lattice_pool
359 ("IPA_CP aggregate lattices");
361 /* Maximal count found in program. */
363 static gcov_type max_count;
365 /* Original overall size of the program. */
367 static long overall_size, max_new_size;
369 /* Return the param lattices structure corresponding to the Ith formal
370 parameter of the function described by INFO. */
371 static inline struct ipcp_param_lattices *
372 ipa_get_parm_lattices (struct ipa_node_params *info, int i)
374 gcc_assert (i >= 0 && i < ipa_get_param_count (info));
375 gcc_checking_assert (!info->ipcp_orig_node);
376 gcc_checking_assert (info->lattices);
377 return &(info->lattices[i]);
380 /* Return the lattice corresponding to the scalar value of the Ith formal
381 parameter of the function described by INFO. */
382 static inline ipcp_lattice<tree> *
383 ipa_get_scalar_lat (struct ipa_node_params *info, int i)
385 struct ipcp_param_lattices *plats = ipa_get_parm_lattices (info, i);
386 return &plats->itself;
389 /* Return the lattice corresponding to the scalar value of the Ith formal
390 parameter of the function described by INFO. */
391 static inline ipcp_lattice<ipa_polymorphic_call_context> *
392 ipa_get_poly_ctx_lat (struct ipa_node_params *info, int i)
394 struct ipcp_param_lattices *plats = ipa_get_parm_lattices (info, i);
395 return &plats->ctxlat;
398 /* Return the lattice corresponding to the value range of the Ith formal
399 parameter of the function described by INFO. */
401 static inline ipcp_vr_lattice *
402 ipa_get_vr_lat (struct ipa_node_params *info, int i)
404 struct ipcp_param_lattices *plats = ipa_get_parm_lattices (info, i);
405 return &plats->m_value_range;
408 /* Return whether LAT is a lattice with a single constant and without an
409 undefined value. */
411 template <typename valtype>
412 inline bool
413 ipcp_lattice<valtype>::is_single_const ()
415 if (bottom || contains_variable || values_count != 1)
416 return false;
417 else
418 return true;
421 /* Print V which is extracted from a value in a lattice to F. */
423 static void
424 print_ipcp_constant_value (FILE * f, tree v)
426 if (TREE_CODE (v) == ADDR_EXPR
427 && TREE_CODE (TREE_OPERAND (v, 0)) == CONST_DECL)
429 fprintf (f, "& ");
430 print_generic_expr (f, DECL_INITIAL (TREE_OPERAND (v, 0)), 0);
432 else
433 print_generic_expr (f, v, 0);
436 /* Print V which is extracted from a value in a lattice to F. */
438 static void
439 print_ipcp_constant_value (FILE * f, ipa_polymorphic_call_context v)
441 v.dump(f, false);
444 /* Print a lattice LAT to F. */
446 template <typename valtype>
447 void
448 ipcp_lattice<valtype>::print (FILE * f, bool dump_sources, bool dump_benefits)
450 ipcp_value<valtype> *val;
451 bool prev = false;
453 if (bottom)
455 fprintf (f, "BOTTOM\n");
456 return;
459 if (!values_count && !contains_variable)
461 fprintf (f, "TOP\n");
462 return;
465 if (contains_variable)
467 fprintf (f, "VARIABLE");
468 prev = true;
469 if (dump_benefits)
470 fprintf (f, "\n");
473 for (val = values; val; val = val->next)
475 if (dump_benefits && prev)
476 fprintf (f, " ");
477 else if (!dump_benefits && prev)
478 fprintf (f, ", ");
479 else
480 prev = true;
482 print_ipcp_constant_value (f, val->value);
484 if (dump_sources)
486 ipcp_value_source<valtype> *s;
488 fprintf (f, " [from:");
489 for (s = val->sources; s; s = s->next)
490 fprintf (f, " %i(%i)", s->cs->caller->order,
491 s->cs->frequency);
492 fprintf (f, "]");
495 if (dump_benefits)
496 fprintf (f, " [loc_time: %i, loc_size: %i, "
497 "prop_time: %i, prop_size: %i]\n",
498 val->local_time_benefit, val->local_size_cost,
499 val->prop_time_benefit, val->prop_size_cost);
501 if (!dump_benefits)
502 fprintf (f, "\n");
505 void
506 ipcp_bits_lattice::print (FILE *f)
508 if (top_p ())
509 fprintf (f, " Bits unknown (TOP)\n");
510 else if (bottom_p ())
511 fprintf (f, " Bits unusable (BOTTOM)\n");
512 else
514 fprintf (f, " Bits: value = "); print_hex (get_value (), f);
515 fprintf (f, ", mask = "); print_hex (get_mask (), f);
516 fprintf (f, "\n");
520 /* Print value range lattice to F. */
522 void
523 ipcp_vr_lattice::print (FILE * f)
525 dump_value_range (f, &m_vr);
528 /* Print all ipcp_lattices of all functions to F. */
530 static void
531 print_all_lattices (FILE * f, bool dump_sources, bool dump_benefits)
533 struct cgraph_node *node;
534 int i, count;
536 fprintf (f, "\nLattices:\n");
537 FOR_EACH_FUNCTION_WITH_GIMPLE_BODY (node)
539 struct ipa_node_params *info;
541 info = IPA_NODE_REF (node);
542 fprintf (f, " Node: %s/%i:\n", node->name (),
543 node->order);
544 count = ipa_get_param_count (info);
545 for (i = 0; i < count; i++)
547 struct ipcp_agg_lattice *aglat;
548 struct ipcp_param_lattices *plats = ipa_get_parm_lattices (info, i);
549 fprintf (f, " param [%d]: ", i);
550 plats->itself.print (f, dump_sources, dump_benefits);
551 fprintf (f, " ctxs: ");
552 plats->ctxlat.print (f, dump_sources, dump_benefits);
553 plats->bits_lattice.print (f);
554 fprintf (f, " ");
555 plats->m_value_range.print (f);
556 fprintf (f, "\n");
557 if (plats->virt_call)
558 fprintf (f, " virt_call flag set\n");
560 if (plats->aggs_bottom)
562 fprintf (f, " AGGS BOTTOM\n");
563 continue;
565 if (plats->aggs_contain_variable)
566 fprintf (f, " AGGS VARIABLE\n");
567 for (aglat = plats->aggs; aglat; aglat = aglat->next)
569 fprintf (f, " %soffset " HOST_WIDE_INT_PRINT_DEC ": ",
570 plats->aggs_by_ref ? "ref " : "", aglat->offset);
571 aglat->print (f, dump_sources, dump_benefits);
577 /* Determine whether it is at all technically possible to create clones of NODE
578 and store this information in the ipa_node_params structure associated
579 with NODE. */
581 static void
582 determine_versionability (struct cgraph_node *node,
583 struct ipa_node_params *info)
585 const char *reason = NULL;
587 /* There are a number of generic reasons functions cannot be versioned. We
588 also cannot remove parameters if there are type attributes such as fnspec
589 present. */
590 if (node->alias || node->thunk.thunk_p)
591 reason = "alias or thunk";
592 else if (!node->local.versionable)
593 reason = "not a tree_versionable_function";
594 else if (node->get_availability () <= AVAIL_INTERPOSABLE)
595 reason = "insufficient body availability";
596 else if (!opt_for_fn (node->decl, optimize)
597 || !opt_for_fn (node->decl, flag_ipa_cp))
598 reason = "non-optimized function";
599 else if (lookup_attribute ("omp declare simd", DECL_ATTRIBUTES (node->decl)))
601 /* Ideally we should clone the SIMD clones themselves and create
602 vector copies of them, so IPA-cp and SIMD clones can happily
603 coexist, but that may not be worth the effort. */
604 reason = "function has SIMD clones";
606 else if (lookup_attribute ("target_clones", DECL_ATTRIBUTES (node->decl)))
608 /* Ideally we should clone the target clones themselves and create
609 copies of them, so IPA-cp and target clones can happily
610 coexist, but that may not be worth the effort. */
611 reason = "function target_clones attribute";
613 /* Don't clone decls local to a comdat group; it breaks and for C++
614 decloned constructors, inlining is always better anyway. */
615 else if (node->comdat_local_p ())
616 reason = "comdat-local function";
617 else if (node->calls_comdat_local)
619 /* TODO: call is versionable if we make sure that all
620 callers are inside of a comdat group. */
621 reason = "calls comdat-local function";
624 if (reason && dump_file && !node->alias && !node->thunk.thunk_p)
625 fprintf (dump_file, "Function %s/%i is not versionable, reason: %s.\n",
626 node->name (), node->order, reason);
628 info->versionable = (reason == NULL);
631 /* Return true if it is at all technically possible to create clones of a
632 NODE. */
634 static bool
635 ipcp_versionable_function_p (struct cgraph_node *node)
637 return IPA_NODE_REF (node)->versionable;
640 /* Structure holding accumulated information about callers of a node. */
642 struct caller_statistics
644 gcov_type count_sum;
645 int n_calls, n_hot_calls, freq_sum;
648 /* Initialize fields of STAT to zeroes. */
650 static inline void
651 init_caller_stats (struct caller_statistics *stats)
653 stats->count_sum = 0;
654 stats->n_calls = 0;
655 stats->n_hot_calls = 0;
656 stats->freq_sum = 0;
659 /* Worker callback of cgraph_for_node_and_aliases accumulating statistics of
660 non-thunk incoming edges to NODE. */
662 static bool
663 gather_caller_stats (struct cgraph_node *node, void *data)
665 struct caller_statistics *stats = (struct caller_statistics *) data;
666 struct cgraph_edge *cs;
668 for (cs = node->callers; cs; cs = cs->next_caller)
669 if (!cs->caller->thunk.thunk_p)
671 stats->count_sum += cs->count;
672 stats->freq_sum += cs->frequency;
673 stats->n_calls++;
674 if (cs->maybe_hot_p ())
675 stats->n_hot_calls ++;
677 return false;
681 /* Return true if this NODE is viable candidate for cloning. */
683 static bool
684 ipcp_cloning_candidate_p (struct cgraph_node *node)
686 struct caller_statistics stats;
688 gcc_checking_assert (node->has_gimple_body_p ());
690 if (!opt_for_fn (node->decl, flag_ipa_cp_clone))
692 if (dump_file)
693 fprintf (dump_file, "Not considering %s for cloning; "
694 "-fipa-cp-clone disabled.\n",
695 node->name ());
696 return false;
699 if (node->optimize_for_size_p ())
701 if (dump_file)
702 fprintf (dump_file, "Not considering %s for cloning; "
703 "optimizing it for size.\n",
704 node->name ());
705 return false;
708 init_caller_stats (&stats);
709 node->call_for_symbol_thunks_and_aliases (gather_caller_stats, &stats, false);
711 if (inline_summaries->get (node)->self_size < stats.n_calls)
713 if (dump_file)
714 fprintf (dump_file, "Considering %s for cloning; code might shrink.\n",
715 node->name ());
716 return true;
719 /* When profile is available and function is hot, propagate into it even if
720 calls seems cold; constant propagation can improve function's speed
721 significantly. */
722 if (max_count)
724 if (stats.count_sum > node->count * 90 / 100)
726 if (dump_file)
727 fprintf (dump_file, "Considering %s for cloning; "
728 "usually called directly.\n",
729 node->name ());
730 return true;
733 if (!stats.n_hot_calls)
735 if (dump_file)
736 fprintf (dump_file, "Not considering %s for cloning; no hot calls.\n",
737 node->name ());
738 return false;
740 if (dump_file)
741 fprintf (dump_file, "Considering %s for cloning.\n",
742 node->name ());
743 return true;
746 template <typename valtype>
747 class value_topo_info
749 public:
750 /* Head of the linked list of topologically sorted values. */
751 ipcp_value<valtype> *values_topo;
752 /* Stack for creating SCCs, represented by a linked list too. */
753 ipcp_value<valtype> *stack;
754 /* Counter driving the algorithm in add_val_to_toposort. */
755 int dfs_counter;
757 value_topo_info () : values_topo (NULL), stack (NULL), dfs_counter (0)
759 void add_val (ipcp_value<valtype> *cur_val);
760 void propagate_effects ();
763 /* Arrays representing a topological ordering of call graph nodes and a stack
764 of nodes used during constant propagation and also data required to perform
765 topological sort of values and propagation of benefits in the determined
766 order. */
768 class ipa_topo_info
770 public:
771 /* Array with obtained topological order of cgraph nodes. */
772 struct cgraph_node **order;
773 /* Stack of cgraph nodes used during propagation within SCC until all values
774 in the SCC stabilize. */
775 struct cgraph_node **stack;
776 int nnodes, stack_top;
778 value_topo_info<tree> constants;
779 value_topo_info<ipa_polymorphic_call_context> contexts;
781 ipa_topo_info () : order(NULL), stack(NULL), nnodes(0), stack_top(0),
782 constants ()
786 /* Allocate the arrays in TOPO and topologically sort the nodes into order. */
788 static void
789 build_toporder_info (struct ipa_topo_info *topo)
791 topo->order = XCNEWVEC (struct cgraph_node *, symtab->cgraph_count);
792 topo->stack = XCNEWVEC (struct cgraph_node *, symtab->cgraph_count);
794 gcc_checking_assert (topo->stack_top == 0);
795 topo->nnodes = ipa_reduced_postorder (topo->order, true, true, NULL);
798 /* Free information about strongly connected components and the arrays in
799 TOPO. */
801 static void
802 free_toporder_info (struct ipa_topo_info *topo)
804 ipa_free_postorder_info ();
805 free (topo->order);
806 free (topo->stack);
809 /* Add NODE to the stack in TOPO, unless it is already there. */
811 static inline void
812 push_node_to_stack (struct ipa_topo_info *topo, struct cgraph_node *node)
814 struct ipa_node_params *info = IPA_NODE_REF (node);
815 if (info->node_enqueued)
816 return;
817 info->node_enqueued = 1;
818 topo->stack[topo->stack_top++] = node;
821 /* Pop a node from the stack in TOPO and return it or return NULL if the stack
822 is empty. */
824 static struct cgraph_node *
825 pop_node_from_stack (struct ipa_topo_info *topo)
827 if (topo->stack_top)
829 struct cgraph_node *node;
830 topo->stack_top--;
831 node = topo->stack[topo->stack_top];
832 IPA_NODE_REF (node)->node_enqueued = 0;
833 return node;
835 else
836 return NULL;
839 /* Set lattice LAT to bottom and return true if it previously was not set as
840 such. */
842 template <typename valtype>
843 inline bool
844 ipcp_lattice<valtype>::set_to_bottom ()
846 bool ret = !bottom;
847 bottom = true;
848 return ret;
851 /* Mark lattice as containing an unknown value and return true if it previously
852 was not marked as such. */
854 template <typename valtype>
855 inline bool
856 ipcp_lattice<valtype>::set_contains_variable ()
858 bool ret = !contains_variable;
859 contains_variable = true;
860 return ret;
863 /* Set all aggegate lattices in PLATS to bottom and return true if they were
864 not previously set as such. */
866 static inline bool
867 set_agg_lats_to_bottom (struct ipcp_param_lattices *plats)
869 bool ret = !plats->aggs_bottom;
870 plats->aggs_bottom = true;
871 return ret;
874 /* Mark all aggegate lattices in PLATS as containing an unknown value and
875 return true if they were not previously marked as such. */
877 static inline bool
878 set_agg_lats_contain_variable (struct ipcp_param_lattices *plats)
880 bool ret = !plats->aggs_contain_variable;
881 plats->aggs_contain_variable = true;
882 return ret;
885 bool
886 ipcp_vr_lattice::meet_with (const ipcp_vr_lattice &other)
888 return meet_with_1 (&other.m_vr);
891 /* Meet the current value of the lattice with value ranfge described by VR
892 lattice. */
894 bool
895 ipcp_vr_lattice::meet_with (const value_range *p_vr)
897 return meet_with_1 (p_vr);
900 /* Meet the current value of the lattice with value ranfge described by
901 OTHER_VR lattice. */
903 bool
904 ipcp_vr_lattice::meet_with_1 (const value_range *other_vr)
906 tree min = m_vr.min, max = m_vr.max;
907 value_range_type type = m_vr.type;
909 if (bottom_p ())
910 return false;
912 if (other_vr->type == VR_VARYING)
913 return set_to_bottom ();
915 vrp_meet (&m_vr, other_vr);
916 if (type != m_vr.type
917 || min != m_vr.min
918 || max != m_vr.max)
919 return true;
920 else
921 return false;
924 /* Return true if value range information in the lattice is yet unknown. */
926 bool
927 ipcp_vr_lattice::top_p () const
929 return m_vr.type == VR_UNDEFINED;
932 /* Return true if value range information in the lattice is known to be
933 unusable. */
935 bool
936 ipcp_vr_lattice::bottom_p () const
938 return m_vr.type == VR_VARYING;
941 /* Set value range information in the lattice to bottom. Return true if it
942 previously was in a different state. */
944 bool
945 ipcp_vr_lattice::set_to_bottom ()
947 if (m_vr.type == VR_VARYING)
948 return false;
949 m_vr.type = VR_VARYING;
950 return true;
953 /* Set lattice value to bottom, if it already isn't the case. */
955 bool
956 ipcp_bits_lattice::set_to_bottom ()
958 if (bottom_p ())
959 return false;
960 m_lattice_val = IPA_BITS_VARYING;
961 m_value = 0;
962 m_mask = -1;
963 return true;
966 /* Set to constant if it isn't already. Only meant to be called
967 when switching state from TOP. */
969 bool
970 ipcp_bits_lattice::set_to_constant (widest_int value, widest_int mask)
972 gcc_assert (top_p ());
973 m_lattice_val = IPA_BITS_CONSTANT;
974 m_value = value;
975 m_mask = mask;
976 return true;
979 /* Convert operand to value, mask form. */
981 void
982 ipcp_bits_lattice::get_value_and_mask (tree operand, widest_int *valuep, widest_int *maskp)
984 wide_int get_nonzero_bits (const_tree);
986 if (TREE_CODE (operand) == INTEGER_CST)
988 *valuep = wi::to_widest (operand);
989 *maskp = 0;
991 else
993 *valuep = 0;
994 *maskp = -1;
998 /* Meet operation, similar to ccp_lattice_meet, we xor values
999 if this->value, value have different values at same bit positions, we want
1000 to drop that bit to varying. Return true if mask is changed.
1001 This function assumes that the lattice value is in CONSTANT state */
1003 bool
1004 ipcp_bits_lattice::meet_with_1 (widest_int value, widest_int mask,
1005 unsigned precision)
1007 gcc_assert (constant_p ());
1009 widest_int old_mask = m_mask;
1010 m_mask = (m_mask | mask) | (m_value ^ value);
1012 if (wi::sext (m_mask, precision) == -1)
1013 return set_to_bottom ();
1015 return m_mask != old_mask;
1018 /* Meet the bits lattice with operand
1019 described by <value, mask, sgn, precision. */
1021 bool
1022 ipcp_bits_lattice::meet_with (widest_int value, widest_int mask,
1023 unsigned precision)
1025 if (bottom_p ())
1026 return false;
1028 if (top_p ())
1030 if (wi::sext (mask, precision) == -1)
1031 return set_to_bottom ();
1032 return set_to_constant (value, mask);
1035 return meet_with_1 (value, mask, precision);
1038 /* Meet bits lattice with the result of bit_value_binop (other, operand)
1039 if code is binary operation or bit_value_unop (other) if code is unary op.
1040 In the case when code is nop_expr, no adjustment is required. */
1042 bool
1043 ipcp_bits_lattice::meet_with (ipcp_bits_lattice& other, unsigned precision,
1044 signop sgn, enum tree_code code, tree operand)
1046 if (other.bottom_p ())
1047 return set_to_bottom ();
1049 if (bottom_p () || other.top_p ())
1050 return false;
1052 widest_int adjusted_value, adjusted_mask;
1054 if (TREE_CODE_CLASS (code) == tcc_binary)
1056 tree type = TREE_TYPE (operand);
1057 gcc_assert (INTEGRAL_TYPE_P (type));
1058 widest_int o_value, o_mask;
1059 get_value_and_mask (operand, &o_value, &o_mask);
1061 bit_value_binop (code, sgn, precision, &adjusted_value, &adjusted_mask,
1062 sgn, precision, other.get_value (), other.get_mask (),
1063 TYPE_SIGN (type), TYPE_PRECISION (type), o_value, o_mask);
1065 if (wi::sext (adjusted_mask, precision) == -1)
1066 return set_to_bottom ();
1069 else if (TREE_CODE_CLASS (code) == tcc_unary)
1071 bit_value_unop (code, sgn, precision, &adjusted_value,
1072 &adjusted_mask, sgn, precision, other.get_value (),
1073 other.get_mask ());
1075 if (wi::sext (adjusted_mask, precision) == -1)
1076 return set_to_bottom ();
1079 else
1080 return set_to_bottom ();
1082 if (top_p ())
1084 if (wi::sext (adjusted_mask, precision) == -1)
1085 return set_to_bottom ();
1086 return set_to_constant (adjusted_value, adjusted_mask);
1088 else
1089 return meet_with_1 (adjusted_value, adjusted_mask, precision);
1092 /* Mark bot aggregate and scalar lattices as containing an unknown variable,
1093 return true is any of them has not been marked as such so far. */
1095 static inline bool
1096 set_all_contains_variable (struct ipcp_param_lattices *plats)
1098 bool ret;
1099 ret = plats->itself.set_contains_variable ();
1100 ret |= plats->ctxlat.set_contains_variable ();
1101 ret |= set_agg_lats_contain_variable (plats);
1102 ret |= plats->bits_lattice.set_to_bottom ();
1103 ret |= plats->m_value_range.set_to_bottom ();
1104 return ret;
1107 /* Worker of call_for_symbol_thunks_and_aliases, increment the integer DATA
1108 points to by the number of callers to NODE. */
1110 static bool
1111 count_callers (cgraph_node *node, void *data)
1113 int *caller_count = (int *) data;
1115 for (cgraph_edge *cs = node->callers; cs; cs = cs->next_caller)
1116 /* Local thunks can be handled transparently, but if the thunk can not
1117 be optimized out, count it as a real use. */
1118 if (!cs->caller->thunk.thunk_p || !cs->caller->local.local)
1119 ++*caller_count;
1120 return false;
1123 /* Worker of call_for_symbol_thunks_and_aliases, it is supposed to be called on
1124 the one caller of some other node. Set the caller's corresponding flag. */
1126 static bool
1127 set_single_call_flag (cgraph_node *node, void *)
1129 cgraph_edge *cs = node->callers;
1130 /* Local thunks can be handled transparently, skip them. */
1131 while (cs && cs->caller->thunk.thunk_p && cs->caller->local.local)
1132 cs = cs->next_caller;
1133 if (cs)
1135 IPA_NODE_REF (cs->caller)->node_calling_single_call = true;
1136 return true;
1138 return false;
1141 /* Initialize ipcp_lattices. */
1143 static void
1144 initialize_node_lattices (struct cgraph_node *node)
1146 struct ipa_node_params *info = IPA_NODE_REF (node);
1147 struct cgraph_edge *ie;
1148 bool disable = false, variable = false;
1149 int i;
1151 gcc_checking_assert (node->has_gimple_body_p ());
1152 if (cgraph_local_p (node))
1154 int caller_count = 0;
1155 node->call_for_symbol_thunks_and_aliases (count_callers, &caller_count,
1156 true);
1157 gcc_checking_assert (caller_count > 0);
1158 if (caller_count == 1)
1159 node->call_for_symbol_thunks_and_aliases (set_single_call_flag,
1160 NULL, true);
1162 else
1164 /* When cloning is allowed, we can assume that externally visible
1165 functions are not called. We will compensate this by cloning
1166 later. */
1167 if (ipcp_versionable_function_p (node)
1168 && ipcp_cloning_candidate_p (node))
1169 variable = true;
1170 else
1171 disable = true;
1174 for (i = 0; i < ipa_get_param_count (info); i++)
1176 struct ipcp_param_lattices *plats = ipa_get_parm_lattices (info, i);
1177 plats->m_value_range.init ();
1180 if (disable || variable)
1182 for (i = 0; i < ipa_get_param_count (info); i++)
1184 struct ipcp_param_lattices *plats = ipa_get_parm_lattices (info, i);
1185 if (disable)
1187 plats->itself.set_to_bottom ();
1188 plats->ctxlat.set_to_bottom ();
1189 set_agg_lats_to_bottom (plats);
1190 plats->bits_lattice.set_to_bottom ();
1191 plats->m_value_range.set_to_bottom ();
1193 else
1194 set_all_contains_variable (plats);
1196 if (dump_file && (dump_flags & TDF_DETAILS)
1197 && !node->alias && !node->thunk.thunk_p)
1198 fprintf (dump_file, "Marking all lattices of %s/%i as %s\n",
1199 node->name (), node->order,
1200 disable ? "BOTTOM" : "VARIABLE");
1203 for (ie = node->indirect_calls; ie; ie = ie->next_callee)
1204 if (ie->indirect_info->polymorphic
1205 && ie->indirect_info->param_index >= 0)
1207 gcc_checking_assert (ie->indirect_info->param_index >= 0);
1208 ipa_get_parm_lattices (info,
1209 ie->indirect_info->param_index)->virt_call = 1;
1213 /* Return the result of a (possibly arithmetic) pass through jump function
1214 JFUNC on the constant value INPUT. Return NULL_TREE if that cannot be
1215 determined or be considered an interprocedural invariant. */
1217 static tree
1218 ipa_get_jf_pass_through_result (struct ipa_jump_func *jfunc, tree input)
1220 tree restype, res;
1222 if (ipa_get_jf_pass_through_operation (jfunc) == NOP_EXPR)
1223 return input;
1224 if (!is_gimple_ip_invariant (input))
1225 return NULL_TREE;
1227 if (TREE_CODE_CLASS (ipa_get_jf_pass_through_operation (jfunc))
1228 == tcc_unary)
1229 res = fold_unary (ipa_get_jf_pass_through_operation (jfunc),
1230 TREE_TYPE (input), input);
1231 else
1233 if (TREE_CODE_CLASS (ipa_get_jf_pass_through_operation (jfunc))
1234 == tcc_comparison)
1235 restype = boolean_type_node;
1236 else
1237 restype = TREE_TYPE (input);
1238 res = fold_binary (ipa_get_jf_pass_through_operation (jfunc), restype,
1239 input, ipa_get_jf_pass_through_operand (jfunc));
1241 if (res && !is_gimple_ip_invariant (res))
1242 return NULL_TREE;
1244 return res;
1247 /* Return the result of an ancestor jump function JFUNC on the constant value
1248 INPUT. Return NULL_TREE if that cannot be determined. */
1250 static tree
1251 ipa_get_jf_ancestor_result (struct ipa_jump_func *jfunc, tree input)
1253 gcc_checking_assert (TREE_CODE (input) != TREE_BINFO);
1254 if (TREE_CODE (input) == ADDR_EXPR)
1256 tree t = TREE_OPERAND (input, 0);
1257 t = build_ref_for_offset (EXPR_LOCATION (t), t,
1258 ipa_get_jf_ancestor_offset (jfunc), false,
1259 ptr_type_node, NULL, false);
1260 return build_fold_addr_expr (t);
1262 else
1263 return NULL_TREE;
1266 /* Determine whether JFUNC evaluates to a single known constant value and if
1267 so, return it. Otherwise return NULL. INFO describes the caller node or
1268 the one it is inlined to, so that pass-through jump functions can be
1269 evaluated. */
1271 tree
1272 ipa_value_from_jfunc (struct ipa_node_params *info, struct ipa_jump_func *jfunc)
1274 if (jfunc->type == IPA_JF_CONST)
1275 return ipa_get_jf_constant (jfunc);
1276 else if (jfunc->type == IPA_JF_PASS_THROUGH
1277 || jfunc->type == IPA_JF_ANCESTOR)
1279 tree input;
1280 int idx;
1282 if (jfunc->type == IPA_JF_PASS_THROUGH)
1283 idx = ipa_get_jf_pass_through_formal_id (jfunc);
1284 else
1285 idx = ipa_get_jf_ancestor_formal_id (jfunc);
1287 if (info->ipcp_orig_node)
1288 input = info->known_csts[idx];
1289 else
1291 ipcp_lattice<tree> *lat;
1293 if (!info->lattices
1294 || idx >= ipa_get_param_count (info))
1295 return NULL_TREE;
1296 lat = ipa_get_scalar_lat (info, idx);
1297 if (!lat->is_single_const ())
1298 return NULL_TREE;
1299 input = lat->values->value;
1302 if (!input)
1303 return NULL_TREE;
1305 if (jfunc->type == IPA_JF_PASS_THROUGH)
1306 return ipa_get_jf_pass_through_result (jfunc, input);
1307 else
1308 return ipa_get_jf_ancestor_result (jfunc, input);
1310 else
1311 return NULL_TREE;
1314 /* Determie whether JFUNC evaluates to single known polymorphic context, given
1315 that INFO describes the caller node or the one it is inlined to, CS is the
1316 call graph edge corresponding to JFUNC and CSIDX index of the described
1317 parameter. */
1319 ipa_polymorphic_call_context
1320 ipa_context_from_jfunc (ipa_node_params *info, cgraph_edge *cs, int csidx,
1321 ipa_jump_func *jfunc)
1323 ipa_edge_args *args = IPA_EDGE_REF (cs);
1324 ipa_polymorphic_call_context ctx;
1325 ipa_polymorphic_call_context *edge_ctx
1326 = cs ? ipa_get_ith_polymorhic_call_context (args, csidx) : NULL;
1328 if (edge_ctx && !edge_ctx->useless_p ())
1329 ctx = *edge_ctx;
1331 if (jfunc->type == IPA_JF_PASS_THROUGH
1332 || jfunc->type == IPA_JF_ANCESTOR)
1334 ipa_polymorphic_call_context srcctx;
1335 int srcidx;
1336 bool type_preserved = true;
1337 if (jfunc->type == IPA_JF_PASS_THROUGH)
1339 if (ipa_get_jf_pass_through_operation (jfunc) != NOP_EXPR)
1340 return ctx;
1341 type_preserved = ipa_get_jf_pass_through_type_preserved (jfunc);
1342 srcidx = ipa_get_jf_pass_through_formal_id (jfunc);
1344 else
1346 type_preserved = ipa_get_jf_ancestor_type_preserved (jfunc);
1347 srcidx = ipa_get_jf_ancestor_formal_id (jfunc);
1349 if (info->ipcp_orig_node)
1351 if (info->known_contexts.exists ())
1352 srcctx = info->known_contexts[srcidx];
1354 else
1356 if (!info->lattices
1357 || srcidx >= ipa_get_param_count (info))
1358 return ctx;
1359 ipcp_lattice<ipa_polymorphic_call_context> *lat;
1360 lat = ipa_get_poly_ctx_lat (info, srcidx);
1361 if (!lat->is_single_const ())
1362 return ctx;
1363 srcctx = lat->values->value;
1365 if (srcctx.useless_p ())
1366 return ctx;
1367 if (jfunc->type == IPA_JF_ANCESTOR)
1368 srcctx.offset_by (ipa_get_jf_ancestor_offset (jfunc));
1369 if (!type_preserved)
1370 srcctx.possible_dynamic_type_change (cs->in_polymorphic_cdtor);
1371 srcctx.combine_with (ctx);
1372 return srcctx;
1375 return ctx;
1378 /* If checking is enabled, verify that no lattice is in the TOP state, i.e. not
1379 bottom, not containing a variable component and without any known value at
1380 the same time. */
1382 DEBUG_FUNCTION void
1383 ipcp_verify_propagated_values (void)
1385 struct cgraph_node *node;
1387 FOR_EACH_FUNCTION_WITH_GIMPLE_BODY (node)
1389 struct ipa_node_params *info = IPA_NODE_REF (node);
1390 int i, count = ipa_get_param_count (info);
1392 for (i = 0; i < count; i++)
1394 ipcp_lattice<tree> *lat = ipa_get_scalar_lat (info, i);
1396 if (!lat->bottom
1397 && !lat->contains_variable
1398 && lat->values_count == 0)
1400 if (dump_file)
1402 symtab_node::dump_table (dump_file);
1403 fprintf (dump_file, "\nIPA lattices after constant "
1404 "propagation, before gcc_unreachable:\n");
1405 print_all_lattices (dump_file, true, false);
1408 gcc_unreachable ();
1414 /* Return true iff X and Y should be considered equal values by IPA-CP. */
1416 static bool
1417 values_equal_for_ipcp_p (tree x, tree y)
1419 gcc_checking_assert (x != NULL_TREE && y != NULL_TREE);
1421 if (x == y)
1422 return true;
1424 if (TREE_CODE (x) == ADDR_EXPR
1425 && TREE_CODE (y) == ADDR_EXPR
1426 && TREE_CODE (TREE_OPERAND (x, 0)) == CONST_DECL
1427 && TREE_CODE (TREE_OPERAND (y, 0)) == CONST_DECL)
1428 return operand_equal_p (DECL_INITIAL (TREE_OPERAND (x, 0)),
1429 DECL_INITIAL (TREE_OPERAND (y, 0)), 0);
1430 else
1431 return operand_equal_p (x, y, 0);
1434 /* Return true iff X and Y should be considered equal contexts by IPA-CP. */
1436 static bool
1437 values_equal_for_ipcp_p (ipa_polymorphic_call_context x,
1438 ipa_polymorphic_call_context y)
1440 return x.equal_to (y);
1444 /* Add a new value source to the value represented by THIS, marking that a
1445 value comes from edge CS and (if the underlying jump function is a
1446 pass-through or an ancestor one) from a caller value SRC_VAL of a caller
1447 parameter described by SRC_INDEX. OFFSET is negative if the source was the
1448 scalar value of the parameter itself or the offset within an aggregate. */
1450 template <typename valtype>
1451 void
1452 ipcp_value<valtype>::add_source (cgraph_edge *cs, ipcp_value *src_val,
1453 int src_idx, HOST_WIDE_INT offset)
1455 ipcp_value_source<valtype> *src;
1457 src = new (ipcp_sources_pool.allocate ()) ipcp_value_source<valtype>;
1458 src->offset = offset;
1459 src->cs = cs;
1460 src->val = src_val;
1461 src->index = src_idx;
1463 src->next = sources;
1464 sources = src;
1467 /* Allocate a new ipcp_value holding a tree constant, initialize its value to
1468 SOURCE and clear all other fields. */
1470 static ipcp_value<tree> *
1471 allocate_and_init_ipcp_value (tree source)
1473 ipcp_value<tree> *val;
1475 val = ipcp_cst_values_pool.allocate ();
1476 memset (val, 0, sizeof (*val));
1477 val->value = source;
1478 return val;
1481 /* Allocate a new ipcp_value holding a polymorphic context, initialize its
1482 value to SOURCE and clear all other fields. */
1484 static ipcp_value<ipa_polymorphic_call_context> *
1485 allocate_and_init_ipcp_value (ipa_polymorphic_call_context source)
1487 ipcp_value<ipa_polymorphic_call_context> *val;
1489 // TODO
1490 val = ipcp_poly_ctx_values_pool.allocate ();
1491 memset (val, 0, sizeof (*val));
1492 val->value = source;
1493 return val;
1496 /* Try to add NEWVAL to LAT, potentially creating a new ipcp_value for it. CS,
1497 SRC_VAL SRC_INDEX and OFFSET are meant for add_source and have the same
1498 meaning. OFFSET -1 means the source is scalar and not a part of an
1499 aggregate. */
1501 template <typename valtype>
1502 bool
1503 ipcp_lattice<valtype>::add_value (valtype newval, cgraph_edge *cs,
1504 ipcp_value<valtype> *src_val,
1505 int src_idx, HOST_WIDE_INT offset)
1507 ipcp_value<valtype> *val;
1509 if (bottom)
1510 return false;
1512 for (val = values; val; val = val->next)
1513 if (values_equal_for_ipcp_p (val->value, newval))
1515 if (ipa_edge_within_scc (cs))
1517 ipcp_value_source<valtype> *s;
1518 for (s = val->sources; s; s = s->next)
1519 if (s->cs == cs)
1520 break;
1521 if (s)
1522 return false;
1525 val->add_source (cs, src_val, src_idx, offset);
1526 return false;
1529 if (values_count == PARAM_VALUE (PARAM_IPA_CP_VALUE_LIST_SIZE))
1531 /* We can only free sources, not the values themselves, because sources
1532 of other values in this SCC might point to them. */
1533 for (val = values; val; val = val->next)
1535 while (val->sources)
1537 ipcp_value_source<valtype> *src = val->sources;
1538 val->sources = src->next;
1539 ipcp_sources_pool.remove ((ipcp_value_source<tree>*)src);
1543 values = NULL;
1544 return set_to_bottom ();
1547 values_count++;
1548 val = allocate_and_init_ipcp_value (newval);
1549 val->add_source (cs, src_val, src_idx, offset);
1550 val->next = values;
1551 values = val;
1552 return true;
1555 /* Propagate values through a pass-through jump function JFUNC associated with
1556 edge CS, taking values from SRC_LAT and putting them into DEST_LAT. SRC_IDX
1557 is the index of the source parameter. */
1559 static bool
1560 propagate_vals_across_pass_through (cgraph_edge *cs, ipa_jump_func *jfunc,
1561 ipcp_lattice<tree> *src_lat,
1562 ipcp_lattice<tree> *dest_lat, int src_idx)
1564 ipcp_value<tree> *src_val;
1565 bool ret = false;
1567 /* Do not create new values when propagating within an SCC because if there
1568 are arithmetic functions with circular dependencies, there is infinite
1569 number of them and we would just make lattices bottom. */
1570 if ((ipa_get_jf_pass_through_operation (jfunc) != NOP_EXPR)
1571 && ipa_edge_within_scc (cs))
1572 ret = dest_lat->set_contains_variable ();
1573 else
1574 for (src_val = src_lat->values; src_val; src_val = src_val->next)
1576 tree cstval = ipa_get_jf_pass_through_result (jfunc, src_val->value);
1578 if (cstval)
1579 ret |= dest_lat->add_value (cstval, cs, src_val, src_idx);
1580 else
1581 ret |= dest_lat->set_contains_variable ();
1584 return ret;
1587 /* Propagate values through an ancestor jump function JFUNC associated with
1588 edge CS, taking values from SRC_LAT and putting them into DEST_LAT. SRC_IDX
1589 is the index of the source parameter. */
1591 static bool
1592 propagate_vals_across_ancestor (struct cgraph_edge *cs,
1593 struct ipa_jump_func *jfunc,
1594 ipcp_lattice<tree> *src_lat,
1595 ipcp_lattice<tree> *dest_lat, int src_idx)
1597 ipcp_value<tree> *src_val;
1598 bool ret = false;
1600 if (ipa_edge_within_scc (cs))
1601 return dest_lat->set_contains_variable ();
1603 for (src_val = src_lat->values; src_val; src_val = src_val->next)
1605 tree t = ipa_get_jf_ancestor_result (jfunc, src_val->value);
1607 if (t)
1608 ret |= dest_lat->add_value (t, cs, src_val, src_idx);
1609 else
1610 ret |= dest_lat->set_contains_variable ();
1613 return ret;
1616 /* Propagate scalar values across jump function JFUNC that is associated with
1617 edge CS and put the values into DEST_LAT. */
1619 static bool
1620 propagate_scalar_across_jump_function (struct cgraph_edge *cs,
1621 struct ipa_jump_func *jfunc,
1622 ipcp_lattice<tree> *dest_lat)
1624 if (dest_lat->bottom)
1625 return false;
1627 if (jfunc->type == IPA_JF_CONST)
1629 tree val = ipa_get_jf_constant (jfunc);
1630 return dest_lat->add_value (val, cs, NULL, 0);
1632 else if (jfunc->type == IPA_JF_PASS_THROUGH
1633 || jfunc->type == IPA_JF_ANCESTOR)
1635 struct ipa_node_params *caller_info = IPA_NODE_REF (cs->caller);
1636 ipcp_lattice<tree> *src_lat;
1637 int src_idx;
1638 bool ret;
1640 if (jfunc->type == IPA_JF_PASS_THROUGH)
1641 src_idx = ipa_get_jf_pass_through_formal_id (jfunc);
1642 else
1643 src_idx = ipa_get_jf_ancestor_formal_id (jfunc);
1645 src_lat = ipa_get_scalar_lat (caller_info, src_idx);
1646 if (src_lat->bottom)
1647 return dest_lat->set_contains_variable ();
1649 /* If we would need to clone the caller and cannot, do not propagate. */
1650 if (!ipcp_versionable_function_p (cs->caller)
1651 && (src_lat->contains_variable
1652 || (src_lat->values_count > 1)))
1653 return dest_lat->set_contains_variable ();
1655 if (jfunc->type == IPA_JF_PASS_THROUGH)
1656 ret = propagate_vals_across_pass_through (cs, jfunc, src_lat,
1657 dest_lat, src_idx);
1658 else
1659 ret = propagate_vals_across_ancestor (cs, jfunc, src_lat, dest_lat,
1660 src_idx);
1662 if (src_lat->contains_variable)
1663 ret |= dest_lat->set_contains_variable ();
1665 return ret;
1668 /* TODO: We currently do not handle member method pointers in IPA-CP (we only
1669 use it for indirect inlining), we should propagate them too. */
1670 return dest_lat->set_contains_variable ();
1673 /* Propagate scalar values across jump function JFUNC that is associated with
1674 edge CS and describes argument IDX and put the values into DEST_LAT. */
1676 static bool
1677 propagate_context_across_jump_function (cgraph_edge *cs,
1678 ipa_jump_func *jfunc, int idx,
1679 ipcp_lattice<ipa_polymorphic_call_context> *dest_lat)
1681 ipa_edge_args *args = IPA_EDGE_REF (cs);
1682 if (dest_lat->bottom)
1683 return false;
1684 bool ret = false;
1685 bool added_sth = false;
1686 bool type_preserved = true;
1688 ipa_polymorphic_call_context edge_ctx, *edge_ctx_ptr
1689 = ipa_get_ith_polymorhic_call_context (args, idx);
1691 if (edge_ctx_ptr)
1692 edge_ctx = *edge_ctx_ptr;
1694 if (jfunc->type == IPA_JF_PASS_THROUGH
1695 || jfunc->type == IPA_JF_ANCESTOR)
1697 struct ipa_node_params *caller_info = IPA_NODE_REF (cs->caller);
1698 int src_idx;
1699 ipcp_lattice<ipa_polymorphic_call_context> *src_lat;
1701 /* TODO: Once we figure out how to propagate speculations, it will
1702 probably be a good idea to switch to speculation if type_preserved is
1703 not set instead of punting. */
1704 if (jfunc->type == IPA_JF_PASS_THROUGH)
1706 if (ipa_get_jf_pass_through_operation (jfunc) != NOP_EXPR)
1707 goto prop_fail;
1708 type_preserved = ipa_get_jf_pass_through_type_preserved (jfunc);
1709 src_idx = ipa_get_jf_pass_through_formal_id (jfunc);
1711 else
1713 type_preserved = ipa_get_jf_ancestor_type_preserved (jfunc);
1714 src_idx = ipa_get_jf_ancestor_formal_id (jfunc);
1717 src_lat = ipa_get_poly_ctx_lat (caller_info, src_idx);
1718 /* If we would need to clone the caller and cannot, do not propagate. */
1719 if (!ipcp_versionable_function_p (cs->caller)
1720 && (src_lat->contains_variable
1721 || (src_lat->values_count > 1)))
1722 goto prop_fail;
1724 ipcp_value<ipa_polymorphic_call_context> *src_val;
1725 for (src_val = src_lat->values; src_val; src_val = src_val->next)
1727 ipa_polymorphic_call_context cur = src_val->value;
1729 if (!type_preserved)
1730 cur.possible_dynamic_type_change (cs->in_polymorphic_cdtor);
1731 if (jfunc->type == IPA_JF_ANCESTOR)
1732 cur.offset_by (ipa_get_jf_ancestor_offset (jfunc));
1733 /* TODO: In cases we know how the context is going to be used,
1734 we can improve the result by passing proper OTR_TYPE. */
1735 cur.combine_with (edge_ctx);
1736 if (!cur.useless_p ())
1738 if (src_lat->contains_variable
1739 && !edge_ctx.equal_to (cur))
1740 ret |= dest_lat->set_contains_variable ();
1741 ret |= dest_lat->add_value (cur, cs, src_val, src_idx);
1742 added_sth = true;
1748 prop_fail:
1749 if (!added_sth)
1751 if (!edge_ctx.useless_p ())
1752 ret |= dest_lat->add_value (edge_ctx, cs);
1753 else
1754 ret |= dest_lat->set_contains_variable ();
1757 return ret;
1760 /* Propagate bits across jfunc that is associated with
1761 edge cs and update dest_lattice accordingly. */
1763 bool
1764 propagate_bits_across_jump_function (cgraph_edge *cs, int idx,
1765 ipa_jump_func *jfunc,
1766 ipcp_bits_lattice *dest_lattice)
1768 if (dest_lattice->bottom_p ())
1769 return false;
1771 enum availability availability;
1772 cgraph_node *callee = cs->callee->function_symbol (&availability);
1773 struct ipa_node_params *callee_info = IPA_NODE_REF (callee);
1774 tree parm_type = ipa_get_type (callee_info, idx);
1776 /* For K&R C programs, ipa_get_type() could return NULL_TREE.
1777 Avoid the transform for these cases. */
1778 if (!parm_type)
1780 if (dump_file && (dump_flags & TDF_DETAILS))
1781 fprintf (dump_file, "Setting dest_lattice to bottom, because"
1782 " param %i type is NULL for %s\n", idx,
1783 cs->callee->name ());
1785 return dest_lattice->set_to_bottom ();
1788 unsigned precision = TYPE_PRECISION (parm_type);
1789 signop sgn = TYPE_SIGN (parm_type);
1791 if (jfunc->type == IPA_JF_PASS_THROUGH
1792 || jfunc->type == IPA_JF_ANCESTOR)
1794 struct ipa_node_params *caller_info = IPA_NODE_REF (cs->caller);
1795 tree operand = NULL_TREE;
1796 enum tree_code code;
1797 unsigned src_idx;
1799 if (jfunc->type == IPA_JF_PASS_THROUGH)
1801 code = ipa_get_jf_pass_through_operation (jfunc);
1802 src_idx = ipa_get_jf_pass_through_formal_id (jfunc);
1803 if (code != NOP_EXPR)
1804 operand = ipa_get_jf_pass_through_operand (jfunc);
1806 else
1808 code = POINTER_PLUS_EXPR;
1809 src_idx = ipa_get_jf_ancestor_formal_id (jfunc);
1810 unsigned HOST_WIDE_INT offset = ipa_get_jf_ancestor_offset (jfunc) / BITS_PER_UNIT;
1811 operand = build_int_cstu (size_type_node, offset);
1814 struct ipcp_param_lattices *src_lats
1815 = ipa_get_parm_lattices (caller_info, src_idx);
1817 /* Try to propagate bits if src_lattice is bottom, but jfunc is known.
1818 for eg consider:
1819 int f(int x)
1821 g (x & 0xff);
1823 Assume lattice for x is bottom, however we can still propagate
1824 result of x & 0xff == 0xff, which gets computed during ccp1 pass
1825 and we store it in jump function during analysis stage. */
1827 if (src_lats->bits_lattice.bottom_p ()
1828 && jfunc->bits)
1829 return dest_lattice->meet_with (jfunc->bits->value, jfunc->bits->mask,
1830 precision);
1831 else
1832 return dest_lattice->meet_with (src_lats->bits_lattice, precision, sgn,
1833 code, operand);
1836 else if (jfunc->type == IPA_JF_ANCESTOR)
1837 return dest_lattice->set_to_bottom ();
1838 else if (jfunc->bits)
1839 return dest_lattice->meet_with (jfunc->bits->value, jfunc->bits->mask,
1840 precision);
1841 else
1842 return dest_lattice->set_to_bottom ();
1845 /* Emulate effects of unary OPERATION and/or conversion from SRC_TYPE to
1846 DST_TYPE on value range in SRC_VR and store it to DST_VR. Return true if
1847 the result is a range or an anti-range. */
1849 static bool
1850 ipa_vr_operation_and_type_effects (value_range *dst_vr, value_range *src_vr,
1851 enum tree_code operation,
1852 tree dst_type, tree src_type)
1854 memset (dst_vr, 0, sizeof (*dst_vr));
1855 extract_range_from_unary_expr (dst_vr, operation, dst_type, src_vr, src_type);
1856 if (dst_vr->type == VR_RANGE || dst_vr->type == VR_ANTI_RANGE)
1857 return true;
1858 else
1859 return false;
1862 /* Propagate value range across jump function JFUNC that is associated with
1863 edge CS with param of callee of PARAM_TYPE and update DEST_PLATS
1864 accordingly. */
1866 static bool
1867 propagate_vr_across_jump_function (cgraph_edge *cs, ipa_jump_func *jfunc,
1868 struct ipcp_param_lattices *dest_plats,
1869 tree param_type)
1871 ipcp_vr_lattice *dest_lat = &dest_plats->m_value_range;
1873 if (dest_lat->bottom_p ())
1874 return false;
1876 if (!param_type
1877 || (!INTEGRAL_TYPE_P (param_type)
1878 && !POINTER_TYPE_P (param_type)))
1879 return dest_lat->set_to_bottom ();
1881 if (jfunc->type == IPA_JF_PASS_THROUGH)
1883 enum tree_code operation = ipa_get_jf_pass_through_operation (jfunc);
1885 if (TREE_CODE_CLASS (operation) == tcc_unary)
1887 struct ipa_node_params *caller_info = IPA_NODE_REF (cs->caller);
1888 int src_idx = ipa_get_jf_pass_through_formal_id (jfunc);
1889 tree operand_type = ipa_get_type (caller_info, src_idx);
1890 struct ipcp_param_lattices *src_lats
1891 = ipa_get_parm_lattices (caller_info, src_idx);
1893 if (src_lats->m_value_range.bottom_p ())
1894 return dest_lat->set_to_bottom ();
1895 value_range vr;
1896 if (ipa_vr_operation_and_type_effects (&vr,
1897 &src_lats->m_value_range.m_vr,
1898 operation, param_type,
1899 operand_type))
1900 return dest_lat->meet_with (&vr);
1903 else if (jfunc->type == IPA_JF_CONST)
1905 tree val = ipa_get_jf_constant (jfunc);
1906 if (TREE_CODE (val) == INTEGER_CST)
1908 val = fold_convert (param_type, val);
1909 if (TREE_OVERFLOW_P (val))
1910 val = drop_tree_overflow (val);
1912 value_range tmpvr;
1913 memset (&tmpvr, 0, sizeof (tmpvr));
1914 tmpvr.type = VR_RANGE;
1915 tmpvr.min = val;
1916 tmpvr.max = val;
1917 return dest_lat->meet_with (&tmpvr);
1921 value_range vr;
1922 if (jfunc->m_vr
1923 && ipa_vr_operation_and_type_effects (&vr, jfunc->m_vr, NOP_EXPR,
1924 param_type,
1925 TREE_TYPE (jfunc->m_vr->min)))
1926 return dest_lat->meet_with (&vr);
1927 else
1928 return dest_lat->set_to_bottom ();
1931 /* If DEST_PLATS already has aggregate items, check that aggs_by_ref matches
1932 NEW_AGGS_BY_REF and if not, mark all aggs as bottoms and return true (in all
1933 other cases, return false). If there are no aggregate items, set
1934 aggs_by_ref to NEW_AGGS_BY_REF. */
1936 static bool
1937 set_check_aggs_by_ref (struct ipcp_param_lattices *dest_plats,
1938 bool new_aggs_by_ref)
1940 if (dest_plats->aggs)
1942 if (dest_plats->aggs_by_ref != new_aggs_by_ref)
1944 set_agg_lats_to_bottom (dest_plats);
1945 return true;
1948 else
1949 dest_plats->aggs_by_ref = new_aggs_by_ref;
1950 return false;
1953 /* Walk aggregate lattices in DEST_PLATS from ***AGLAT on, until ***aglat is an
1954 already existing lattice for the given OFFSET and SIZE, marking all skipped
1955 lattices as containing variable and checking for overlaps. If there is no
1956 already existing lattice for the OFFSET and VAL_SIZE, create one, initialize
1957 it with offset, size and contains_variable to PRE_EXISTING, and return true,
1958 unless there are too many already. If there are two many, return false. If
1959 there are overlaps turn whole DEST_PLATS to bottom and return false. If any
1960 skipped lattices were newly marked as containing variable, set *CHANGE to
1961 true. */
1963 static bool
1964 merge_agg_lats_step (struct ipcp_param_lattices *dest_plats,
1965 HOST_WIDE_INT offset, HOST_WIDE_INT val_size,
1966 struct ipcp_agg_lattice ***aglat,
1967 bool pre_existing, bool *change)
1969 gcc_checking_assert (offset >= 0);
1971 while (**aglat && (**aglat)->offset < offset)
1973 if ((**aglat)->offset + (**aglat)->size > offset)
1975 set_agg_lats_to_bottom (dest_plats);
1976 return false;
1978 *change |= (**aglat)->set_contains_variable ();
1979 *aglat = &(**aglat)->next;
1982 if (**aglat && (**aglat)->offset == offset)
1984 if ((**aglat)->size != val_size
1985 || ((**aglat)->next
1986 && (**aglat)->next->offset < offset + val_size))
1988 set_agg_lats_to_bottom (dest_plats);
1989 return false;
1991 gcc_checking_assert (!(**aglat)->next
1992 || (**aglat)->next->offset >= offset + val_size);
1993 return true;
1995 else
1997 struct ipcp_agg_lattice *new_al;
1999 if (**aglat && (**aglat)->offset < offset + val_size)
2001 set_agg_lats_to_bottom (dest_plats);
2002 return false;
2004 if (dest_plats->aggs_count == PARAM_VALUE (PARAM_IPA_MAX_AGG_ITEMS))
2005 return false;
2006 dest_plats->aggs_count++;
2007 new_al = ipcp_agg_lattice_pool.allocate ();
2008 memset (new_al, 0, sizeof (*new_al));
2010 new_al->offset = offset;
2011 new_al->size = val_size;
2012 new_al->contains_variable = pre_existing;
2014 new_al->next = **aglat;
2015 **aglat = new_al;
2016 return true;
2020 /* Set all AGLAT and all other aggregate lattices reachable by next pointers as
2021 containing an unknown value. */
2023 static bool
2024 set_chain_of_aglats_contains_variable (struct ipcp_agg_lattice *aglat)
2026 bool ret = false;
2027 while (aglat)
2029 ret |= aglat->set_contains_variable ();
2030 aglat = aglat->next;
2032 return ret;
2035 /* Merge existing aggregate lattices in SRC_PLATS to DEST_PLATS, subtracting
2036 DELTA_OFFSET. CS is the call graph edge and SRC_IDX the index of the source
2037 parameter used for lattice value sources. Return true if DEST_PLATS changed
2038 in any way. */
2040 static bool
2041 merge_aggregate_lattices (struct cgraph_edge *cs,
2042 struct ipcp_param_lattices *dest_plats,
2043 struct ipcp_param_lattices *src_plats,
2044 int src_idx, HOST_WIDE_INT offset_delta)
2046 bool pre_existing = dest_plats->aggs != NULL;
2047 struct ipcp_agg_lattice **dst_aglat;
2048 bool ret = false;
2050 if (set_check_aggs_by_ref (dest_plats, src_plats->aggs_by_ref))
2051 return true;
2052 if (src_plats->aggs_bottom)
2053 return set_agg_lats_contain_variable (dest_plats);
2054 if (src_plats->aggs_contain_variable)
2055 ret |= set_agg_lats_contain_variable (dest_plats);
2056 dst_aglat = &dest_plats->aggs;
2058 for (struct ipcp_agg_lattice *src_aglat = src_plats->aggs;
2059 src_aglat;
2060 src_aglat = src_aglat->next)
2062 HOST_WIDE_INT new_offset = src_aglat->offset - offset_delta;
2064 if (new_offset < 0)
2065 continue;
2066 if (merge_agg_lats_step (dest_plats, new_offset, src_aglat->size,
2067 &dst_aglat, pre_existing, &ret))
2069 struct ipcp_agg_lattice *new_al = *dst_aglat;
2071 dst_aglat = &(*dst_aglat)->next;
2072 if (src_aglat->bottom)
2074 ret |= new_al->set_contains_variable ();
2075 continue;
2077 if (src_aglat->contains_variable)
2078 ret |= new_al->set_contains_variable ();
2079 for (ipcp_value<tree> *val = src_aglat->values;
2080 val;
2081 val = val->next)
2082 ret |= new_al->add_value (val->value, cs, val, src_idx,
2083 src_aglat->offset);
2085 else if (dest_plats->aggs_bottom)
2086 return true;
2088 ret |= set_chain_of_aglats_contains_variable (*dst_aglat);
2089 return ret;
2092 /* Determine whether there is anything to propagate FROM SRC_PLATS through a
2093 pass-through JFUNC and if so, whether it has conform and conforms to the
2094 rules about propagating values passed by reference. */
2096 static bool
2097 agg_pass_through_permissible_p (struct ipcp_param_lattices *src_plats,
2098 struct ipa_jump_func *jfunc)
2100 return src_plats->aggs
2101 && (!src_plats->aggs_by_ref
2102 || ipa_get_jf_pass_through_agg_preserved (jfunc));
2105 /* Propagate scalar values across jump function JFUNC that is associated with
2106 edge CS and put the values into DEST_LAT. */
2108 static bool
2109 propagate_aggs_across_jump_function (struct cgraph_edge *cs,
2110 struct ipa_jump_func *jfunc,
2111 struct ipcp_param_lattices *dest_plats)
2113 bool ret = false;
2115 if (dest_plats->aggs_bottom)
2116 return false;
2118 if (jfunc->type == IPA_JF_PASS_THROUGH
2119 && ipa_get_jf_pass_through_operation (jfunc) == NOP_EXPR)
2121 struct ipa_node_params *caller_info = IPA_NODE_REF (cs->caller);
2122 int src_idx = ipa_get_jf_pass_through_formal_id (jfunc);
2123 struct ipcp_param_lattices *src_plats;
2125 src_plats = ipa_get_parm_lattices (caller_info, src_idx);
2126 if (agg_pass_through_permissible_p (src_plats, jfunc))
2128 /* Currently we do not produce clobber aggregate jump
2129 functions, replace with merging when we do. */
2130 gcc_assert (!jfunc->agg.items);
2131 ret |= merge_aggregate_lattices (cs, dest_plats, src_plats,
2132 src_idx, 0);
2134 else
2135 ret |= set_agg_lats_contain_variable (dest_plats);
2137 else if (jfunc->type == IPA_JF_ANCESTOR
2138 && ipa_get_jf_ancestor_agg_preserved (jfunc))
2140 struct ipa_node_params *caller_info = IPA_NODE_REF (cs->caller);
2141 int src_idx = ipa_get_jf_ancestor_formal_id (jfunc);
2142 struct ipcp_param_lattices *src_plats;
2144 src_plats = ipa_get_parm_lattices (caller_info, src_idx);
2145 if (src_plats->aggs && src_plats->aggs_by_ref)
2147 /* Currently we do not produce clobber aggregate jump
2148 functions, replace with merging when we do. */
2149 gcc_assert (!jfunc->agg.items);
2150 ret |= merge_aggregate_lattices (cs, dest_plats, src_plats, src_idx,
2151 ipa_get_jf_ancestor_offset (jfunc));
2153 else if (!src_plats->aggs_by_ref)
2154 ret |= set_agg_lats_to_bottom (dest_plats);
2155 else
2156 ret |= set_agg_lats_contain_variable (dest_plats);
2158 else if (jfunc->agg.items)
2160 bool pre_existing = dest_plats->aggs != NULL;
2161 struct ipcp_agg_lattice **aglat = &dest_plats->aggs;
2162 struct ipa_agg_jf_item *item;
2163 int i;
2165 if (set_check_aggs_by_ref (dest_plats, jfunc->agg.by_ref))
2166 return true;
2168 FOR_EACH_VEC_ELT (*jfunc->agg.items, i, item)
2170 HOST_WIDE_INT val_size;
2172 if (item->offset < 0)
2173 continue;
2174 gcc_checking_assert (is_gimple_ip_invariant (item->value));
2175 val_size = tree_to_uhwi (TYPE_SIZE (TREE_TYPE (item->value)));
2177 if (merge_agg_lats_step (dest_plats, item->offset, val_size,
2178 &aglat, pre_existing, &ret))
2180 ret |= (*aglat)->add_value (item->value, cs, NULL, 0, 0);
2181 aglat = &(*aglat)->next;
2183 else if (dest_plats->aggs_bottom)
2184 return true;
2187 ret |= set_chain_of_aglats_contains_variable (*aglat);
2189 else
2190 ret |= set_agg_lats_contain_variable (dest_plats);
2192 return ret;
2195 /* Return true if on the way cfrom CS->caller to the final (non-alias and
2196 non-thunk) destination, the call passes through a thunk. */
2198 static bool
2199 call_passes_through_thunk_p (cgraph_edge *cs)
2201 cgraph_node *alias_or_thunk = cs->callee;
2202 while (alias_or_thunk->alias)
2203 alias_or_thunk = alias_or_thunk->get_alias_target ();
2204 return alias_or_thunk->thunk.thunk_p;
2207 /* Propagate constants from the caller to the callee of CS. INFO describes the
2208 caller. */
2210 static bool
2211 propagate_constants_across_call (struct cgraph_edge *cs)
2213 struct ipa_node_params *callee_info;
2214 enum availability availability;
2215 cgraph_node *callee;
2216 struct ipa_edge_args *args;
2217 bool ret = false;
2218 int i, args_count, parms_count;
2220 callee = cs->callee->function_symbol (&availability);
2221 if (!callee->definition)
2222 return false;
2223 gcc_checking_assert (callee->has_gimple_body_p ());
2224 callee_info = IPA_NODE_REF (callee);
2226 args = IPA_EDGE_REF (cs);
2227 args_count = ipa_get_cs_argument_count (args);
2228 parms_count = ipa_get_param_count (callee_info);
2229 if (parms_count == 0)
2230 return false;
2232 /* No propagation through instrumentation thunks is available yet.
2233 It should be possible with proper mapping of call args and
2234 instrumented callee params in the propagation loop below. But
2235 this case mostly occurs when legacy code calls instrumented code
2236 and it is not a primary target for optimizations.
2237 We detect instrumentation thunks in aliases and thunks chain by
2238 checking instrumentation_clone flag for chain source and target.
2239 Going through instrumentation thunks we always have it changed
2240 from 0 to 1 and all other nodes do not change it. */
2241 if (!cs->callee->instrumentation_clone
2242 && callee->instrumentation_clone)
2244 for (i = 0; i < parms_count; i++)
2245 ret |= set_all_contains_variable (ipa_get_parm_lattices (callee_info,
2246 i));
2247 return ret;
2250 /* If this call goes through a thunk we must not propagate to the first (0th)
2251 parameter. However, we might need to uncover a thunk from below a series
2252 of aliases first. */
2253 if (call_passes_through_thunk_p (cs))
2255 ret |= set_all_contains_variable (ipa_get_parm_lattices (callee_info,
2256 0));
2257 i = 1;
2259 else
2260 i = 0;
2262 for (; (i < args_count) && (i < parms_count); i++)
2264 struct ipa_jump_func *jump_func = ipa_get_ith_jump_func (args, i);
2265 struct ipcp_param_lattices *dest_plats;
2266 tree param_type = ipa_get_type (callee_info, i);
2268 dest_plats = ipa_get_parm_lattices (callee_info, i);
2269 if (availability == AVAIL_INTERPOSABLE)
2270 ret |= set_all_contains_variable (dest_plats);
2271 else
2273 ret |= propagate_scalar_across_jump_function (cs, jump_func,
2274 &dest_plats->itself);
2275 ret |= propagate_context_across_jump_function (cs, jump_func, i,
2276 &dest_plats->ctxlat);
2278 |= propagate_bits_across_jump_function (cs, i, jump_func,
2279 &dest_plats->bits_lattice);
2280 ret |= propagate_aggs_across_jump_function (cs, jump_func,
2281 dest_plats);
2282 if (opt_for_fn (callee->decl, flag_ipa_vrp))
2283 ret |= propagate_vr_across_jump_function (cs, jump_func,
2284 dest_plats, param_type);
2285 else
2286 ret |= dest_plats->m_value_range.set_to_bottom ();
2289 for (; i < parms_count; i++)
2290 ret |= set_all_contains_variable (ipa_get_parm_lattices (callee_info, i));
2292 return ret;
2295 /* If an indirect edge IE can be turned into a direct one based on KNOWN_VALS
2296 KNOWN_CONTEXTS, KNOWN_AGGS or AGG_REPS return the destination. The latter
2297 three can be NULL. If AGG_REPS is not NULL, KNOWN_AGGS is ignored. */
2299 static tree
2300 ipa_get_indirect_edge_target_1 (struct cgraph_edge *ie,
2301 vec<tree> known_csts,
2302 vec<ipa_polymorphic_call_context> known_contexts,
2303 vec<ipa_agg_jump_function_p> known_aggs,
2304 struct ipa_agg_replacement_value *agg_reps,
2305 bool *speculative)
2307 int param_index = ie->indirect_info->param_index;
2308 HOST_WIDE_INT anc_offset;
2309 tree t;
2310 tree target = NULL;
2312 *speculative = false;
2314 if (param_index == -1
2315 || known_csts.length () <= (unsigned int) param_index)
2316 return NULL_TREE;
2318 if (!ie->indirect_info->polymorphic)
2320 tree t;
2322 if (ie->indirect_info->agg_contents)
2324 t = NULL;
2325 if (agg_reps && ie->indirect_info->guaranteed_unmodified)
2327 while (agg_reps)
2329 if (agg_reps->index == param_index
2330 && agg_reps->offset == ie->indirect_info->offset
2331 && agg_reps->by_ref == ie->indirect_info->by_ref)
2333 t = agg_reps->value;
2334 break;
2336 agg_reps = agg_reps->next;
2339 if (!t)
2341 struct ipa_agg_jump_function *agg;
2342 if (known_aggs.length () > (unsigned int) param_index)
2343 agg = known_aggs[param_index];
2344 else
2345 agg = NULL;
2346 bool from_global_constant;
2347 t = ipa_find_agg_cst_for_param (agg, known_csts[param_index],
2348 ie->indirect_info->offset,
2349 ie->indirect_info->by_ref,
2350 &from_global_constant);
2351 if (t
2352 && !from_global_constant
2353 && !ie->indirect_info->guaranteed_unmodified)
2354 t = NULL_TREE;
2357 else
2358 t = known_csts[param_index];
2360 if (t
2361 && TREE_CODE (t) == ADDR_EXPR
2362 && TREE_CODE (TREE_OPERAND (t, 0)) == FUNCTION_DECL)
2363 return TREE_OPERAND (t, 0);
2364 else
2365 return NULL_TREE;
2368 if (!opt_for_fn (ie->caller->decl, flag_devirtualize))
2369 return NULL_TREE;
2371 gcc_assert (!ie->indirect_info->agg_contents);
2372 anc_offset = ie->indirect_info->offset;
2374 t = NULL;
2376 /* Try to work out value of virtual table pointer value in replacemnets. */
2377 if (!t && agg_reps && !ie->indirect_info->by_ref)
2379 while (agg_reps)
2381 if (agg_reps->index == param_index
2382 && agg_reps->offset == ie->indirect_info->offset
2383 && agg_reps->by_ref)
2385 t = agg_reps->value;
2386 break;
2388 agg_reps = agg_reps->next;
2392 /* Try to work out value of virtual table pointer value in known
2393 aggregate values. */
2394 if (!t && known_aggs.length () > (unsigned int) param_index
2395 && !ie->indirect_info->by_ref)
2397 struct ipa_agg_jump_function *agg;
2398 agg = known_aggs[param_index];
2399 t = ipa_find_agg_cst_for_param (agg, known_csts[param_index],
2400 ie->indirect_info->offset, true);
2403 /* If we found the virtual table pointer, lookup the target. */
2404 if (t)
2406 tree vtable;
2407 unsigned HOST_WIDE_INT offset;
2408 if (vtable_pointer_value_to_vtable (t, &vtable, &offset))
2410 bool can_refer;
2411 target = gimple_get_virt_method_for_vtable (ie->indirect_info->otr_token,
2412 vtable, offset, &can_refer);
2413 if (can_refer)
2415 if (!target
2416 || (TREE_CODE (TREE_TYPE (target)) == FUNCTION_TYPE
2417 && DECL_FUNCTION_CODE (target) == BUILT_IN_UNREACHABLE)
2418 || !possible_polymorphic_call_target_p
2419 (ie, cgraph_node::get (target)))
2421 /* Do not speculate builtin_unreachable, it is stupid! */
2422 if (ie->indirect_info->vptr_changed)
2423 return NULL;
2424 target = ipa_impossible_devirt_target (ie, target);
2426 *speculative = ie->indirect_info->vptr_changed;
2427 if (!*speculative)
2428 return target;
2433 /* Do we know the constant value of pointer? */
2434 if (!t)
2435 t = known_csts[param_index];
2437 gcc_checking_assert (!t || TREE_CODE (t) != TREE_BINFO);
2439 ipa_polymorphic_call_context context;
2440 if (known_contexts.length () > (unsigned int) param_index)
2442 context = known_contexts[param_index];
2443 context.offset_by (anc_offset);
2444 if (ie->indirect_info->vptr_changed)
2445 context.possible_dynamic_type_change (ie->in_polymorphic_cdtor,
2446 ie->indirect_info->otr_type);
2447 if (t)
2449 ipa_polymorphic_call_context ctx2 = ipa_polymorphic_call_context
2450 (t, ie->indirect_info->otr_type, anc_offset);
2451 if (!ctx2.useless_p ())
2452 context.combine_with (ctx2, ie->indirect_info->otr_type);
2455 else if (t)
2457 context = ipa_polymorphic_call_context (t, ie->indirect_info->otr_type,
2458 anc_offset);
2459 if (ie->indirect_info->vptr_changed)
2460 context.possible_dynamic_type_change (ie->in_polymorphic_cdtor,
2461 ie->indirect_info->otr_type);
2463 else
2464 return NULL_TREE;
2466 vec <cgraph_node *>targets;
2467 bool final;
2469 targets = possible_polymorphic_call_targets
2470 (ie->indirect_info->otr_type,
2471 ie->indirect_info->otr_token,
2472 context, &final);
2473 if (!final || targets.length () > 1)
2475 struct cgraph_node *node;
2476 if (*speculative)
2477 return target;
2478 if (!opt_for_fn (ie->caller->decl, flag_devirtualize_speculatively)
2479 || ie->speculative || !ie->maybe_hot_p ())
2480 return NULL;
2481 node = try_speculative_devirtualization (ie->indirect_info->otr_type,
2482 ie->indirect_info->otr_token,
2483 context);
2484 if (node)
2486 *speculative = true;
2487 target = node->decl;
2489 else
2490 return NULL;
2492 else
2494 *speculative = false;
2495 if (targets.length () == 1)
2496 target = targets[0]->decl;
2497 else
2498 target = ipa_impossible_devirt_target (ie, NULL_TREE);
2501 if (target && !possible_polymorphic_call_target_p (ie,
2502 cgraph_node::get (target)))
2504 if (*speculative)
2505 return NULL;
2506 target = ipa_impossible_devirt_target (ie, target);
2509 return target;
2513 /* If an indirect edge IE can be turned into a direct one based on KNOWN_CSTS,
2514 KNOWN_CONTEXTS (which can be vNULL) or KNOWN_AGGS (which also can be vNULL)
2515 return the destination. */
2517 tree
2518 ipa_get_indirect_edge_target (struct cgraph_edge *ie,
2519 vec<tree> known_csts,
2520 vec<ipa_polymorphic_call_context> known_contexts,
2521 vec<ipa_agg_jump_function_p> known_aggs,
2522 bool *speculative)
2524 return ipa_get_indirect_edge_target_1 (ie, known_csts, known_contexts,
2525 known_aggs, NULL, speculative);
2528 /* Calculate devirtualization time bonus for NODE, assuming we know KNOWN_CSTS
2529 and KNOWN_CONTEXTS. */
2531 static int
2532 devirtualization_time_bonus (struct cgraph_node *node,
2533 vec<tree> known_csts,
2534 vec<ipa_polymorphic_call_context> known_contexts,
2535 vec<ipa_agg_jump_function_p> known_aggs)
2537 struct cgraph_edge *ie;
2538 int res = 0;
2540 for (ie = node->indirect_calls; ie; ie = ie->next_callee)
2542 struct cgraph_node *callee;
2543 struct inline_summary *isummary;
2544 enum availability avail;
2545 tree target;
2546 bool speculative;
2548 target = ipa_get_indirect_edge_target (ie, known_csts, known_contexts,
2549 known_aggs, &speculative);
2550 if (!target)
2551 continue;
2553 /* Only bare minimum benefit for clearly un-inlineable targets. */
2554 res += 1;
2555 callee = cgraph_node::get (target);
2556 if (!callee || !callee->definition)
2557 continue;
2558 callee = callee->function_symbol (&avail);
2559 if (avail < AVAIL_AVAILABLE)
2560 continue;
2561 isummary = inline_summaries->get (callee);
2562 if (!isummary->inlinable)
2563 continue;
2565 /* FIXME: The values below need re-considering and perhaps also
2566 integrating into the cost metrics, at lest in some very basic way. */
2567 if (isummary->size <= MAX_INLINE_INSNS_AUTO / 4)
2568 res += 31 / ((int)speculative + 1);
2569 else if (isummary->size <= MAX_INLINE_INSNS_AUTO / 2)
2570 res += 15 / ((int)speculative + 1);
2571 else if (isummary->size <= MAX_INLINE_INSNS_AUTO
2572 || DECL_DECLARED_INLINE_P (callee->decl))
2573 res += 7 / ((int)speculative + 1);
2576 return res;
2579 /* Return time bonus incurred because of HINTS. */
2581 static int
2582 hint_time_bonus (inline_hints hints)
2584 int result = 0;
2585 if (hints & (INLINE_HINT_loop_iterations | INLINE_HINT_loop_stride))
2586 result += PARAM_VALUE (PARAM_IPA_CP_LOOP_HINT_BONUS);
2587 if (hints & INLINE_HINT_array_index)
2588 result += PARAM_VALUE (PARAM_IPA_CP_ARRAY_INDEX_HINT_BONUS);
2589 return result;
2592 /* If there is a reason to penalize the function described by INFO in the
2593 cloning goodness evaluation, do so. */
2595 static inline int64_t
2596 incorporate_penalties (ipa_node_params *info, int64_t evaluation)
2598 if (info->node_within_scc)
2599 evaluation = (evaluation
2600 * (100 - PARAM_VALUE (PARAM_IPA_CP_RECURSION_PENALTY))) / 100;
2602 if (info->node_calling_single_call)
2603 evaluation = (evaluation
2604 * (100 - PARAM_VALUE (PARAM_IPA_CP_SINGLE_CALL_PENALTY)))
2605 / 100;
2607 return evaluation;
2610 /* Return true if cloning NODE is a good idea, given the estimated TIME_BENEFIT
2611 and SIZE_COST and with the sum of frequencies of incoming edges to the
2612 potential new clone in FREQUENCIES. */
2614 static bool
2615 good_cloning_opportunity_p (struct cgraph_node *node, int time_benefit,
2616 int freq_sum, gcov_type count_sum, int size_cost)
2618 if (time_benefit == 0
2619 || !opt_for_fn (node->decl, flag_ipa_cp_clone)
2620 || node->optimize_for_size_p ())
2621 return false;
2623 gcc_assert (size_cost > 0);
2625 struct ipa_node_params *info = IPA_NODE_REF (node);
2626 if (max_count)
2628 int factor = (count_sum * 1000) / max_count;
2629 int64_t evaluation = (((int64_t) time_benefit * factor)
2630 / size_cost);
2631 evaluation = incorporate_penalties (info, evaluation);
2633 if (dump_file && (dump_flags & TDF_DETAILS))
2634 fprintf (dump_file, " good_cloning_opportunity_p (time: %i, "
2635 "size: %i, count_sum: " HOST_WIDE_INT_PRINT_DEC
2636 "%s%s) -> evaluation: " "%" PRId64
2637 ", threshold: %i\n",
2638 time_benefit, size_cost, (HOST_WIDE_INT) count_sum,
2639 info->node_within_scc ? ", scc" : "",
2640 info->node_calling_single_call ? ", single_call" : "",
2641 evaluation, PARAM_VALUE (PARAM_IPA_CP_EVAL_THRESHOLD));
2643 return evaluation >= PARAM_VALUE (PARAM_IPA_CP_EVAL_THRESHOLD);
2645 else
2647 int64_t evaluation = (((int64_t) time_benefit * freq_sum)
2648 / size_cost);
2649 evaluation = incorporate_penalties (info, evaluation);
2651 if (dump_file && (dump_flags & TDF_DETAILS))
2652 fprintf (dump_file, " good_cloning_opportunity_p (time: %i, "
2653 "size: %i, freq_sum: %i%s%s) -> evaluation: "
2654 "%" PRId64 ", threshold: %i\n",
2655 time_benefit, size_cost, freq_sum,
2656 info->node_within_scc ? ", scc" : "",
2657 info->node_calling_single_call ? ", single_call" : "",
2658 evaluation, PARAM_VALUE (PARAM_IPA_CP_EVAL_THRESHOLD));
2660 return evaluation >= PARAM_VALUE (PARAM_IPA_CP_EVAL_THRESHOLD);
2664 /* Return all context independent values from aggregate lattices in PLATS in a
2665 vector. Return NULL if there are none. */
2667 static vec<ipa_agg_jf_item, va_gc> *
2668 context_independent_aggregate_values (struct ipcp_param_lattices *plats)
2670 vec<ipa_agg_jf_item, va_gc> *res = NULL;
2672 if (plats->aggs_bottom
2673 || plats->aggs_contain_variable
2674 || plats->aggs_count == 0)
2675 return NULL;
2677 for (struct ipcp_agg_lattice *aglat = plats->aggs;
2678 aglat;
2679 aglat = aglat->next)
2680 if (aglat->is_single_const ())
2682 struct ipa_agg_jf_item item;
2683 item.offset = aglat->offset;
2684 item.value = aglat->values->value;
2685 vec_safe_push (res, item);
2687 return res;
2690 /* Allocate KNOWN_CSTS, KNOWN_CONTEXTS and, if non-NULL, KNOWN_AGGS and
2691 populate them with values of parameters that are known independent of the
2692 context. INFO describes the function. If REMOVABLE_PARAMS_COST is
2693 non-NULL, the movement cost of all removable parameters will be stored in
2694 it. */
2696 static bool
2697 gather_context_independent_values (struct ipa_node_params *info,
2698 vec<tree> *known_csts,
2699 vec<ipa_polymorphic_call_context>
2700 *known_contexts,
2701 vec<ipa_agg_jump_function> *known_aggs,
2702 int *removable_params_cost)
2704 int i, count = ipa_get_param_count (info);
2705 bool ret = false;
2707 known_csts->create (0);
2708 known_contexts->create (0);
2709 known_csts->safe_grow_cleared (count);
2710 known_contexts->safe_grow_cleared (count);
2711 if (known_aggs)
2713 known_aggs->create (0);
2714 known_aggs->safe_grow_cleared (count);
2717 if (removable_params_cost)
2718 *removable_params_cost = 0;
2720 for (i = 0; i < count; i++)
2722 struct ipcp_param_lattices *plats = ipa_get_parm_lattices (info, i);
2723 ipcp_lattice<tree> *lat = &plats->itself;
2725 if (lat->is_single_const ())
2727 ipcp_value<tree> *val = lat->values;
2728 gcc_checking_assert (TREE_CODE (val->value) != TREE_BINFO);
2729 (*known_csts)[i] = val->value;
2730 if (removable_params_cost)
2731 *removable_params_cost
2732 += estimate_move_cost (TREE_TYPE (val->value), false);
2733 ret = true;
2735 else if (removable_params_cost
2736 && !ipa_is_param_used (info, i))
2737 *removable_params_cost
2738 += ipa_get_param_move_cost (info, i);
2740 if (!ipa_is_param_used (info, i))
2741 continue;
2743 ipcp_lattice<ipa_polymorphic_call_context> *ctxlat = &plats->ctxlat;
2744 /* Do not account known context as reason for cloning. We can see
2745 if it permits devirtualization. */
2746 if (ctxlat->is_single_const ())
2747 (*known_contexts)[i] = ctxlat->values->value;
2749 if (known_aggs)
2751 vec<ipa_agg_jf_item, va_gc> *agg_items;
2752 struct ipa_agg_jump_function *ajf;
2754 agg_items = context_independent_aggregate_values (plats);
2755 ajf = &(*known_aggs)[i];
2756 ajf->items = agg_items;
2757 ajf->by_ref = plats->aggs_by_ref;
2758 ret |= agg_items != NULL;
2762 return ret;
2765 /* The current interface in ipa-inline-analysis requires a pointer vector.
2766 Create it.
2768 FIXME: That interface should be re-worked, this is slightly silly. Still,
2769 I'd like to discuss how to change it first and this demonstrates the
2770 issue. */
2772 static vec<ipa_agg_jump_function_p>
2773 agg_jmp_p_vec_for_t_vec (vec<ipa_agg_jump_function> known_aggs)
2775 vec<ipa_agg_jump_function_p> ret;
2776 struct ipa_agg_jump_function *ajf;
2777 int i;
2779 ret.create (known_aggs.length ());
2780 FOR_EACH_VEC_ELT (known_aggs, i, ajf)
2781 ret.quick_push (ajf);
2782 return ret;
2785 /* Perform time and size measurement of NODE with the context given in
2786 KNOWN_CSTS, KNOWN_CONTEXTS and KNOWN_AGGS, calculate the benefit and cost
2787 given BASE_TIME of the node without specialization, REMOVABLE_PARAMS_COST of
2788 all context-independent removable parameters and EST_MOVE_COST of estimated
2789 movement of the considered parameter and store it into VAL. */
2791 static void
2792 perform_estimation_of_a_value (cgraph_node *node, vec<tree> known_csts,
2793 vec<ipa_polymorphic_call_context> known_contexts,
2794 vec<ipa_agg_jump_function_p> known_aggs_ptrs,
2795 int removable_params_cost,
2796 int est_move_cost, ipcp_value_base *val)
2798 int size, time_benefit;
2799 sreal time, base_time;
2800 inline_hints hints;
2802 estimate_ipcp_clone_size_and_time (node, known_csts, known_contexts,
2803 known_aggs_ptrs, &size, &time,
2804 &base_time, &hints);
2805 base_time -= time;
2806 if (base_time > 65535)
2807 base_time = 65535;
2808 time_benefit = base_time.to_int ()
2809 + devirtualization_time_bonus (node, known_csts, known_contexts,
2810 known_aggs_ptrs)
2811 + hint_time_bonus (hints)
2812 + removable_params_cost + est_move_cost;
2814 gcc_checking_assert (size >=0);
2815 /* The inliner-heuristics based estimates may think that in certain
2816 contexts some functions do not have any size at all but we want
2817 all specializations to have at least a tiny cost, not least not to
2818 divide by zero. */
2819 if (size == 0)
2820 size = 1;
2822 val->local_time_benefit = time_benefit;
2823 val->local_size_cost = size;
2826 /* Iterate over known values of parameters of NODE and estimate the local
2827 effects in terms of time and size they have. */
2829 static void
2830 estimate_local_effects (struct cgraph_node *node)
2832 struct ipa_node_params *info = IPA_NODE_REF (node);
2833 int i, count = ipa_get_param_count (info);
2834 vec<tree> known_csts;
2835 vec<ipa_polymorphic_call_context> known_contexts;
2836 vec<ipa_agg_jump_function> known_aggs;
2837 vec<ipa_agg_jump_function_p> known_aggs_ptrs;
2838 bool always_const;
2839 int removable_params_cost;
2841 if (!count || !ipcp_versionable_function_p (node))
2842 return;
2844 if (dump_file && (dump_flags & TDF_DETAILS))
2845 fprintf (dump_file, "\nEstimating effects for %s/%i.\n",
2846 node->name (), node->order);
2848 always_const = gather_context_independent_values (info, &known_csts,
2849 &known_contexts, &known_aggs,
2850 &removable_params_cost);
2851 known_aggs_ptrs = agg_jmp_p_vec_for_t_vec (known_aggs);
2852 int devirt_bonus = devirtualization_time_bonus (node, known_csts,
2853 known_contexts, known_aggs_ptrs);
2854 if (always_const || devirt_bonus
2855 || (removable_params_cost && node->local.can_change_signature))
2857 struct caller_statistics stats;
2858 inline_hints hints;
2859 sreal time, base_time;
2860 int size;
2862 init_caller_stats (&stats);
2863 node->call_for_symbol_thunks_and_aliases (gather_caller_stats, &stats,
2864 false);
2865 estimate_ipcp_clone_size_and_time (node, known_csts, known_contexts,
2866 known_aggs_ptrs, &size, &time,
2867 &base_time, &hints);
2868 time -= devirt_bonus;
2869 time -= hint_time_bonus (hints);
2870 time -= removable_params_cost;
2871 size -= stats.n_calls * removable_params_cost;
2873 if (dump_file)
2874 fprintf (dump_file, " - context independent values, size: %i, "
2875 "time_benefit: %f\n", size, (base_time - time).to_double ());
2877 if (size <= 0 || node->local.local)
2879 info->do_clone_for_all_contexts = true;
2881 if (dump_file)
2882 fprintf (dump_file, " Decided to specialize for all "
2883 "known contexts, code not going to grow.\n");
2885 else if (good_cloning_opportunity_p (node,
2886 MAX ((base_time - time).to_int (),
2887 65536),
2888 stats.freq_sum, stats.count_sum,
2889 size))
2891 if (size + overall_size <= max_new_size)
2893 info->do_clone_for_all_contexts = true;
2894 overall_size += size;
2896 if (dump_file)
2897 fprintf (dump_file, " Decided to specialize for all "
2898 "known contexts, growth deemed beneficial.\n");
2900 else if (dump_file && (dump_flags & TDF_DETAILS))
2901 fprintf (dump_file, " Not cloning for all contexts because "
2902 "max_new_size would be reached with %li.\n",
2903 size + overall_size);
2905 else if (dump_file && (dump_flags & TDF_DETAILS))
2906 fprintf (dump_file, " Not cloning for all contexts because "
2907 "!good_cloning_opportunity_p.\n");
2911 for (i = 0; i < count; i++)
2913 struct ipcp_param_lattices *plats = ipa_get_parm_lattices (info, i);
2914 ipcp_lattice<tree> *lat = &plats->itself;
2915 ipcp_value<tree> *val;
2917 if (lat->bottom
2918 || !lat->values
2919 || known_csts[i])
2920 continue;
2922 for (val = lat->values; val; val = val->next)
2924 gcc_checking_assert (TREE_CODE (val->value) != TREE_BINFO);
2925 known_csts[i] = val->value;
2927 int emc = estimate_move_cost (TREE_TYPE (val->value), true);
2928 perform_estimation_of_a_value (node, known_csts, known_contexts,
2929 known_aggs_ptrs,
2930 removable_params_cost, emc, val);
2932 if (dump_file && (dump_flags & TDF_DETAILS))
2934 fprintf (dump_file, " - estimates for value ");
2935 print_ipcp_constant_value (dump_file, val->value);
2936 fprintf (dump_file, " for ");
2937 ipa_dump_param (dump_file, info, i);
2938 fprintf (dump_file, ": time_benefit: %i, size: %i\n",
2939 val->local_time_benefit, val->local_size_cost);
2942 known_csts[i] = NULL_TREE;
2945 for (i = 0; i < count; i++)
2947 struct ipcp_param_lattices *plats = ipa_get_parm_lattices (info, i);
2949 if (!plats->virt_call)
2950 continue;
2952 ipcp_lattice<ipa_polymorphic_call_context> *ctxlat = &plats->ctxlat;
2953 ipcp_value<ipa_polymorphic_call_context> *val;
2955 if (ctxlat->bottom
2956 || !ctxlat->values
2957 || !known_contexts[i].useless_p ())
2958 continue;
2960 for (val = ctxlat->values; val; val = val->next)
2962 known_contexts[i] = val->value;
2963 perform_estimation_of_a_value (node, known_csts, known_contexts,
2964 known_aggs_ptrs,
2965 removable_params_cost, 0, val);
2967 if (dump_file && (dump_flags & TDF_DETAILS))
2969 fprintf (dump_file, " - estimates for polymorphic context ");
2970 print_ipcp_constant_value (dump_file, val->value);
2971 fprintf (dump_file, " for ");
2972 ipa_dump_param (dump_file, info, i);
2973 fprintf (dump_file, ": time_benefit: %i, size: %i\n",
2974 val->local_time_benefit, val->local_size_cost);
2977 known_contexts[i] = ipa_polymorphic_call_context ();
2980 for (i = 0; i < count; i++)
2982 struct ipcp_param_lattices *plats = ipa_get_parm_lattices (info, i);
2983 struct ipa_agg_jump_function *ajf;
2984 struct ipcp_agg_lattice *aglat;
2986 if (plats->aggs_bottom || !plats->aggs)
2987 continue;
2989 ajf = &known_aggs[i];
2990 for (aglat = plats->aggs; aglat; aglat = aglat->next)
2992 ipcp_value<tree> *val;
2993 if (aglat->bottom || !aglat->values
2994 /* If the following is true, the one value is in known_aggs. */
2995 || (!plats->aggs_contain_variable
2996 && aglat->is_single_const ()))
2997 continue;
2999 for (val = aglat->values; val; val = val->next)
3001 struct ipa_agg_jf_item item;
3003 item.offset = aglat->offset;
3004 item.value = val->value;
3005 vec_safe_push (ajf->items, item);
3007 perform_estimation_of_a_value (node, known_csts, known_contexts,
3008 known_aggs_ptrs,
3009 removable_params_cost, 0, val);
3011 if (dump_file && (dump_flags & TDF_DETAILS))
3013 fprintf (dump_file, " - estimates for value ");
3014 print_ipcp_constant_value (dump_file, val->value);
3015 fprintf (dump_file, " for ");
3016 ipa_dump_param (dump_file, info, i);
3017 fprintf (dump_file, "[%soffset: " HOST_WIDE_INT_PRINT_DEC
3018 "]: time_benefit: %i, size: %i\n",
3019 plats->aggs_by_ref ? "ref " : "",
3020 aglat->offset,
3021 val->local_time_benefit, val->local_size_cost);
3024 ajf->items->pop ();
3029 for (i = 0; i < count; i++)
3030 vec_free (known_aggs[i].items);
3032 known_csts.release ();
3033 known_contexts.release ();
3034 known_aggs.release ();
3035 known_aggs_ptrs.release ();
3039 /* Add value CUR_VAL and all yet-unsorted values it is dependent on to the
3040 topological sort of values. */
3042 template <typename valtype>
3043 void
3044 value_topo_info<valtype>::add_val (ipcp_value<valtype> *cur_val)
3046 ipcp_value_source<valtype> *src;
3048 if (cur_val->dfs)
3049 return;
3051 dfs_counter++;
3052 cur_val->dfs = dfs_counter;
3053 cur_val->low_link = dfs_counter;
3055 cur_val->topo_next = stack;
3056 stack = cur_val;
3057 cur_val->on_stack = true;
3059 for (src = cur_val->sources; src; src = src->next)
3060 if (src->val)
3062 if (src->val->dfs == 0)
3064 add_val (src->val);
3065 if (src->val->low_link < cur_val->low_link)
3066 cur_val->low_link = src->val->low_link;
3068 else if (src->val->on_stack
3069 && src->val->dfs < cur_val->low_link)
3070 cur_val->low_link = src->val->dfs;
3073 if (cur_val->dfs == cur_val->low_link)
3075 ipcp_value<valtype> *v, *scc_list = NULL;
3079 v = stack;
3080 stack = v->topo_next;
3081 v->on_stack = false;
3083 v->scc_next = scc_list;
3084 scc_list = v;
3086 while (v != cur_val);
3088 cur_val->topo_next = values_topo;
3089 values_topo = cur_val;
3093 /* Add all values in lattices associated with NODE to the topological sort if
3094 they are not there yet. */
3096 static void
3097 add_all_node_vals_to_toposort (cgraph_node *node, ipa_topo_info *topo)
3099 struct ipa_node_params *info = IPA_NODE_REF (node);
3100 int i, count = ipa_get_param_count (info);
3102 for (i = 0; i < count; i++)
3104 struct ipcp_param_lattices *plats = ipa_get_parm_lattices (info, i);
3105 ipcp_lattice<tree> *lat = &plats->itself;
3106 struct ipcp_agg_lattice *aglat;
3108 if (!lat->bottom)
3110 ipcp_value<tree> *val;
3111 for (val = lat->values; val; val = val->next)
3112 topo->constants.add_val (val);
3115 if (!plats->aggs_bottom)
3116 for (aglat = plats->aggs; aglat; aglat = aglat->next)
3117 if (!aglat->bottom)
3119 ipcp_value<tree> *val;
3120 for (val = aglat->values; val; val = val->next)
3121 topo->constants.add_val (val);
3124 ipcp_lattice<ipa_polymorphic_call_context> *ctxlat = &plats->ctxlat;
3125 if (!ctxlat->bottom)
3127 ipcp_value<ipa_polymorphic_call_context> *ctxval;
3128 for (ctxval = ctxlat->values; ctxval; ctxval = ctxval->next)
3129 topo->contexts.add_val (ctxval);
3134 /* One pass of constants propagation along the call graph edges, from callers
3135 to callees (requires topological ordering in TOPO), iterate over strongly
3136 connected components. */
3138 static void
3139 propagate_constants_topo (struct ipa_topo_info *topo)
3141 int i;
3143 for (i = topo->nnodes - 1; i >= 0; i--)
3145 unsigned j;
3146 struct cgraph_node *v, *node = topo->order[i];
3147 vec<cgraph_node *> cycle_nodes = ipa_get_nodes_in_cycle (node);
3149 /* First, iteratively propagate within the strongly connected component
3150 until all lattices stabilize. */
3151 FOR_EACH_VEC_ELT (cycle_nodes, j, v)
3152 if (v->has_gimple_body_p ())
3153 push_node_to_stack (topo, v);
3155 v = pop_node_from_stack (topo);
3156 while (v)
3158 struct cgraph_edge *cs;
3160 for (cs = v->callees; cs; cs = cs->next_callee)
3161 if (ipa_edge_within_scc (cs))
3163 IPA_NODE_REF (v)->node_within_scc = true;
3164 if (propagate_constants_across_call (cs))
3165 push_node_to_stack (topo, cs->callee->function_symbol ());
3167 v = pop_node_from_stack (topo);
3170 /* Afterwards, propagate along edges leading out of the SCC, calculates
3171 the local effects of the discovered constants and all valid values to
3172 their topological sort. */
3173 FOR_EACH_VEC_ELT (cycle_nodes, j, v)
3174 if (v->has_gimple_body_p ())
3176 struct cgraph_edge *cs;
3178 estimate_local_effects (v);
3179 add_all_node_vals_to_toposort (v, topo);
3180 for (cs = v->callees; cs; cs = cs->next_callee)
3181 if (!ipa_edge_within_scc (cs))
3182 propagate_constants_across_call (cs);
3184 cycle_nodes.release ();
3189 /* Return the sum of A and B if none of them is bigger than INT_MAX/2, return
3190 the bigger one if otherwise. */
3192 static int
3193 safe_add (int a, int b)
3195 if (a > INT_MAX/2 || b > INT_MAX/2)
3196 return a > b ? a : b;
3197 else
3198 return a + b;
3202 /* Propagate the estimated effects of individual values along the topological
3203 from the dependent values to those they depend on. */
3205 template <typename valtype>
3206 void
3207 value_topo_info<valtype>::propagate_effects ()
3209 ipcp_value<valtype> *base;
3211 for (base = values_topo; base; base = base->topo_next)
3213 ipcp_value_source<valtype> *src;
3214 ipcp_value<valtype> *val;
3215 int time = 0, size = 0;
3217 for (val = base; val; val = val->scc_next)
3219 time = safe_add (time,
3220 val->local_time_benefit + val->prop_time_benefit);
3221 size = safe_add (size, val->local_size_cost + val->prop_size_cost);
3224 for (val = base; val; val = val->scc_next)
3225 for (src = val->sources; src; src = src->next)
3226 if (src->val
3227 && src->cs->maybe_hot_p ())
3229 src->val->prop_time_benefit = safe_add (time,
3230 src->val->prop_time_benefit);
3231 src->val->prop_size_cost = safe_add (size,
3232 src->val->prop_size_cost);
3238 /* Propagate constants, polymorphic contexts and their effects from the
3239 summaries interprocedurally. */
3241 static void
3242 ipcp_propagate_stage (struct ipa_topo_info *topo)
3244 struct cgraph_node *node;
3246 if (dump_file)
3247 fprintf (dump_file, "\n Propagating constants:\n\n");
3249 FOR_EACH_DEFINED_FUNCTION (node)
3251 struct ipa_node_params *info = IPA_NODE_REF (node);
3253 determine_versionability (node, info);
3254 if (node->has_gimple_body_p ())
3256 info->lattices = XCNEWVEC (struct ipcp_param_lattices,
3257 ipa_get_param_count (info));
3258 initialize_node_lattices (node);
3260 if (node->definition && !node->alias)
3261 overall_size += inline_summaries->get (node)->self_size;
3262 if (node->count > max_count)
3263 max_count = node->count;
3266 max_new_size = overall_size;
3267 if (max_new_size < PARAM_VALUE (PARAM_LARGE_UNIT_INSNS))
3268 max_new_size = PARAM_VALUE (PARAM_LARGE_UNIT_INSNS);
3269 max_new_size += max_new_size * PARAM_VALUE (PARAM_IPCP_UNIT_GROWTH) / 100 + 1;
3271 if (dump_file)
3272 fprintf (dump_file, "\noverall_size: %li, max_new_size: %li\n",
3273 overall_size, max_new_size);
3275 propagate_constants_topo (topo);
3276 if (flag_checking)
3277 ipcp_verify_propagated_values ();
3278 topo->constants.propagate_effects ();
3279 topo->contexts.propagate_effects ();
3281 if (dump_file)
3283 fprintf (dump_file, "\nIPA lattices after all propagation:\n");
3284 print_all_lattices (dump_file, (dump_flags & TDF_DETAILS), true);
3288 /* Discover newly direct outgoing edges from NODE which is a new clone with
3289 known KNOWN_CSTS and make them direct. */
3291 static void
3292 ipcp_discover_new_direct_edges (struct cgraph_node *node,
3293 vec<tree> known_csts,
3294 vec<ipa_polymorphic_call_context>
3295 known_contexts,
3296 struct ipa_agg_replacement_value *aggvals)
3298 struct cgraph_edge *ie, *next_ie;
3299 bool found = false;
3301 for (ie = node->indirect_calls; ie; ie = next_ie)
3303 tree target;
3304 bool speculative;
3306 next_ie = ie->next_callee;
3307 target = ipa_get_indirect_edge_target_1 (ie, known_csts, known_contexts,
3308 vNULL, aggvals, &speculative);
3309 if (target)
3311 bool agg_contents = ie->indirect_info->agg_contents;
3312 bool polymorphic = ie->indirect_info->polymorphic;
3313 int param_index = ie->indirect_info->param_index;
3314 struct cgraph_edge *cs = ipa_make_edge_direct_to_target (ie, target,
3315 speculative);
3316 found = true;
3318 if (cs && !agg_contents && !polymorphic)
3320 struct ipa_node_params *info = IPA_NODE_REF (node);
3321 int c = ipa_get_controlled_uses (info, param_index);
3322 if (c != IPA_UNDESCRIBED_USE)
3324 struct ipa_ref *to_del;
3326 c--;
3327 ipa_set_controlled_uses (info, param_index, c);
3328 if (dump_file && (dump_flags & TDF_DETAILS))
3329 fprintf (dump_file, " controlled uses count of param "
3330 "%i bumped down to %i\n", param_index, c);
3331 if (c == 0
3332 && (to_del = node->find_reference (cs->callee, NULL, 0)))
3334 if (dump_file && (dump_flags & TDF_DETAILS))
3335 fprintf (dump_file, " and even removing its "
3336 "cloning-created reference\n");
3337 to_del->remove_reference ();
3343 /* Turning calls to direct calls will improve overall summary. */
3344 if (found)
3345 inline_update_overall_summary (node);
3348 /* Vector of pointers which for linked lists of clones of an original crgaph
3349 edge. */
3351 static vec<cgraph_edge *> next_edge_clone;
3352 static vec<cgraph_edge *> prev_edge_clone;
3354 static inline void
3355 grow_edge_clone_vectors (void)
3357 if (next_edge_clone.length ()
3358 <= (unsigned) symtab->edges_max_uid)
3359 next_edge_clone.safe_grow_cleared (symtab->edges_max_uid + 1);
3360 if (prev_edge_clone.length ()
3361 <= (unsigned) symtab->edges_max_uid)
3362 prev_edge_clone.safe_grow_cleared (symtab->edges_max_uid + 1);
3365 /* Edge duplication hook to grow the appropriate linked list in
3366 next_edge_clone. */
3368 static void
3369 ipcp_edge_duplication_hook (struct cgraph_edge *src, struct cgraph_edge *dst,
3370 void *)
3372 grow_edge_clone_vectors ();
3374 struct cgraph_edge *old_next = next_edge_clone[src->uid];
3375 if (old_next)
3376 prev_edge_clone[old_next->uid] = dst;
3377 prev_edge_clone[dst->uid] = src;
3379 next_edge_clone[dst->uid] = old_next;
3380 next_edge_clone[src->uid] = dst;
3383 /* Hook that is called by cgraph.c when an edge is removed. */
3385 static void
3386 ipcp_edge_removal_hook (struct cgraph_edge *cs, void *)
3388 grow_edge_clone_vectors ();
3390 struct cgraph_edge *prev = prev_edge_clone[cs->uid];
3391 struct cgraph_edge *next = next_edge_clone[cs->uid];
3392 if (prev)
3393 next_edge_clone[prev->uid] = next;
3394 if (next)
3395 prev_edge_clone[next->uid] = prev;
3398 /* See if NODE is a clone with a known aggregate value at a given OFFSET of a
3399 parameter with the given INDEX. */
3401 static tree
3402 get_clone_agg_value (struct cgraph_node *node, HOST_WIDE_INT offset,
3403 int index)
3405 struct ipa_agg_replacement_value *aggval;
3407 aggval = ipa_get_agg_replacements_for_node (node);
3408 while (aggval)
3410 if (aggval->offset == offset
3411 && aggval->index == index)
3412 return aggval->value;
3413 aggval = aggval->next;
3415 return NULL_TREE;
3418 /* Return true is NODE is DEST or its clone for all contexts. */
3420 static bool
3421 same_node_or_its_all_contexts_clone_p (cgraph_node *node, cgraph_node *dest)
3423 if (node == dest)
3424 return true;
3426 struct ipa_node_params *info = IPA_NODE_REF (node);
3427 return info->is_all_contexts_clone && info->ipcp_orig_node == dest;
3430 /* Return true if edge CS does bring about the value described by SRC to node
3431 DEST or its clone for all contexts. */
3433 static bool
3434 cgraph_edge_brings_value_p (cgraph_edge *cs, ipcp_value_source<tree> *src,
3435 cgraph_node *dest)
3437 struct ipa_node_params *caller_info = IPA_NODE_REF (cs->caller);
3438 enum availability availability;
3439 cgraph_node *real_dest = cs->callee->function_symbol (&availability);
3441 if (!same_node_or_its_all_contexts_clone_p (real_dest, dest)
3442 || availability <= AVAIL_INTERPOSABLE
3443 || caller_info->node_dead)
3444 return false;
3445 if (!src->val)
3446 return true;
3448 if (caller_info->ipcp_orig_node)
3450 tree t;
3451 if (src->offset == -1)
3452 t = caller_info->known_csts[src->index];
3453 else
3454 t = get_clone_agg_value (cs->caller, src->offset, src->index);
3455 return (t != NULL_TREE
3456 && values_equal_for_ipcp_p (src->val->value, t));
3458 else
3460 struct ipcp_agg_lattice *aglat;
3461 struct ipcp_param_lattices *plats = ipa_get_parm_lattices (caller_info,
3462 src->index);
3463 if (src->offset == -1)
3464 return (plats->itself.is_single_const ()
3465 && values_equal_for_ipcp_p (src->val->value,
3466 plats->itself.values->value));
3467 else
3469 if (plats->aggs_bottom || plats->aggs_contain_variable)
3470 return false;
3471 for (aglat = plats->aggs; aglat; aglat = aglat->next)
3472 if (aglat->offset == src->offset)
3473 return (aglat->is_single_const ()
3474 && values_equal_for_ipcp_p (src->val->value,
3475 aglat->values->value));
3477 return false;
3481 /* Return true if edge CS does bring about the value described by SRC to node
3482 DEST or its clone for all contexts. */
3484 static bool
3485 cgraph_edge_brings_value_p (cgraph_edge *cs,
3486 ipcp_value_source<ipa_polymorphic_call_context> *src,
3487 cgraph_node *dest)
3489 struct ipa_node_params *caller_info = IPA_NODE_REF (cs->caller);
3490 cgraph_node *real_dest = cs->callee->function_symbol ();
3492 if (!same_node_or_its_all_contexts_clone_p (real_dest, dest)
3493 || caller_info->node_dead)
3494 return false;
3495 if (!src->val)
3496 return true;
3498 if (caller_info->ipcp_orig_node)
3499 return (caller_info->known_contexts.length () > (unsigned) src->index)
3500 && values_equal_for_ipcp_p (src->val->value,
3501 caller_info->known_contexts[src->index]);
3503 struct ipcp_param_lattices *plats = ipa_get_parm_lattices (caller_info,
3504 src->index);
3505 return plats->ctxlat.is_single_const ()
3506 && values_equal_for_ipcp_p (src->val->value,
3507 plats->ctxlat.values->value);
3510 /* Get the next clone in the linked list of clones of an edge. */
3512 static inline struct cgraph_edge *
3513 get_next_cgraph_edge_clone (struct cgraph_edge *cs)
3515 return next_edge_clone[cs->uid];
3518 /* Given VAL that is intended for DEST, iterate over all its sources and if
3519 they still hold, add their edge frequency and their number into *FREQUENCY
3520 and *CALLER_COUNT respectively. */
3522 template <typename valtype>
3523 static bool
3524 get_info_about_necessary_edges (ipcp_value<valtype> *val, cgraph_node *dest,
3525 int *freq_sum,
3526 gcov_type *count_sum, int *caller_count)
3528 ipcp_value_source<valtype> *src;
3529 int freq = 0, count = 0;
3530 gcov_type cnt = 0;
3531 bool hot = false;
3533 for (src = val->sources; src; src = src->next)
3535 struct cgraph_edge *cs = src->cs;
3536 while (cs)
3538 if (cgraph_edge_brings_value_p (cs, src, dest))
3540 count++;
3541 freq += cs->frequency;
3542 cnt += cs->count;
3543 hot |= cs->maybe_hot_p ();
3545 cs = get_next_cgraph_edge_clone (cs);
3549 *freq_sum = freq;
3550 *count_sum = cnt;
3551 *caller_count = count;
3552 return hot;
3555 /* Return a vector of incoming edges that do bring value VAL to node DEST. It
3556 is assumed their number is known and equal to CALLER_COUNT. */
3558 template <typename valtype>
3559 static vec<cgraph_edge *>
3560 gather_edges_for_value (ipcp_value<valtype> *val, cgraph_node *dest,
3561 int caller_count)
3563 ipcp_value_source<valtype> *src;
3564 vec<cgraph_edge *> ret;
3566 ret.create (caller_count);
3567 for (src = val->sources; src; src = src->next)
3569 struct cgraph_edge *cs = src->cs;
3570 while (cs)
3572 if (cgraph_edge_brings_value_p (cs, src, dest))
3573 ret.quick_push (cs);
3574 cs = get_next_cgraph_edge_clone (cs);
3578 return ret;
3581 /* Construct a replacement map for a know VALUE for a formal parameter PARAM.
3582 Return it or NULL if for some reason it cannot be created. */
3584 static struct ipa_replace_map *
3585 get_replacement_map (struct ipa_node_params *info, tree value, int parm_num)
3587 struct ipa_replace_map *replace_map;
3590 replace_map = ggc_alloc<ipa_replace_map> ();
3591 if (dump_file)
3593 fprintf (dump_file, " replacing ");
3594 ipa_dump_param (dump_file, info, parm_num);
3596 fprintf (dump_file, " with const ");
3597 print_generic_expr (dump_file, value, 0);
3598 fprintf (dump_file, "\n");
3600 replace_map->old_tree = NULL;
3601 replace_map->parm_num = parm_num;
3602 replace_map->new_tree = value;
3603 replace_map->replace_p = true;
3604 replace_map->ref_p = false;
3606 return replace_map;
3609 /* Dump new profiling counts */
3611 static void
3612 dump_profile_updates (struct cgraph_node *orig_node,
3613 struct cgraph_node *new_node)
3615 struct cgraph_edge *cs;
3617 fprintf (dump_file, " setting count of the specialized node to "
3618 HOST_WIDE_INT_PRINT_DEC "\n", (HOST_WIDE_INT) new_node->count);
3619 for (cs = new_node->callees; cs; cs = cs->next_callee)
3620 fprintf (dump_file, " edge to %s has count "
3621 HOST_WIDE_INT_PRINT_DEC "\n",
3622 cs->callee->name (), (HOST_WIDE_INT) cs->count);
3624 fprintf (dump_file, " setting count of the original node to "
3625 HOST_WIDE_INT_PRINT_DEC "\n", (HOST_WIDE_INT) orig_node->count);
3626 for (cs = orig_node->callees; cs; cs = cs->next_callee)
3627 fprintf (dump_file, " edge to %s is left with "
3628 HOST_WIDE_INT_PRINT_DEC "\n",
3629 cs->callee->name (), (HOST_WIDE_INT) cs->count);
3632 /* After a specialized NEW_NODE version of ORIG_NODE has been created, update
3633 their profile information to reflect this. */
3635 static void
3636 update_profiling_info (struct cgraph_node *orig_node,
3637 struct cgraph_node *new_node)
3639 struct cgraph_edge *cs;
3640 struct caller_statistics stats;
3641 gcov_type new_sum, orig_sum;
3642 gcov_type remainder, orig_node_count = orig_node->count;
3644 if (orig_node_count == 0)
3645 return;
3647 init_caller_stats (&stats);
3648 orig_node->call_for_symbol_thunks_and_aliases (gather_caller_stats, &stats,
3649 false);
3650 orig_sum = stats.count_sum;
3651 init_caller_stats (&stats);
3652 new_node->call_for_symbol_thunks_and_aliases (gather_caller_stats, &stats,
3653 false);
3654 new_sum = stats.count_sum;
3656 if (orig_node_count < orig_sum + new_sum)
3658 if (dump_file)
3659 fprintf (dump_file, " Problem: node %s/%i has too low count "
3660 HOST_WIDE_INT_PRINT_DEC " while the sum of incoming "
3661 "counts is " HOST_WIDE_INT_PRINT_DEC "\n",
3662 orig_node->name (), orig_node->order,
3663 (HOST_WIDE_INT) orig_node_count,
3664 (HOST_WIDE_INT) (orig_sum + new_sum));
3666 orig_node_count = (orig_sum + new_sum) * 12 / 10;
3667 if (dump_file)
3668 fprintf (dump_file, " proceeding by pretending it was "
3669 HOST_WIDE_INT_PRINT_DEC "\n",
3670 (HOST_WIDE_INT) orig_node_count);
3673 new_node->count = new_sum;
3674 remainder = orig_node_count - new_sum;
3675 orig_node->count = remainder;
3677 for (cs = new_node->callees; cs; cs = cs->next_callee)
3678 if (cs->frequency)
3679 cs->count = apply_probability (cs->count,
3680 GCOV_COMPUTE_SCALE (new_sum,
3681 orig_node_count));
3682 else
3683 cs->count = 0;
3685 for (cs = orig_node->callees; cs; cs = cs->next_callee)
3686 cs->count = apply_probability (cs->count,
3687 GCOV_COMPUTE_SCALE (remainder,
3688 orig_node_count));
3690 if (dump_file)
3691 dump_profile_updates (orig_node, new_node);
3694 /* Update the respective profile of specialized NEW_NODE and the original
3695 ORIG_NODE after additional edges with cumulative count sum REDIRECTED_SUM
3696 have been redirected to the specialized version. */
3698 static void
3699 update_specialized_profile (struct cgraph_node *new_node,
3700 struct cgraph_node *orig_node,
3701 gcov_type redirected_sum)
3703 struct cgraph_edge *cs;
3704 gcov_type new_node_count, orig_node_count = orig_node->count;
3706 if (dump_file)
3707 fprintf (dump_file, " the sum of counts of redirected edges is "
3708 HOST_WIDE_INT_PRINT_DEC "\n", (HOST_WIDE_INT) redirected_sum);
3709 if (orig_node_count == 0)
3710 return;
3712 gcc_assert (orig_node_count >= redirected_sum);
3714 new_node_count = new_node->count;
3715 new_node->count += redirected_sum;
3716 orig_node->count -= redirected_sum;
3718 for (cs = new_node->callees; cs; cs = cs->next_callee)
3719 if (cs->frequency)
3720 cs->count += apply_probability (cs->count,
3721 GCOV_COMPUTE_SCALE (redirected_sum,
3722 new_node_count));
3723 else
3724 cs->count = 0;
3726 for (cs = orig_node->callees; cs; cs = cs->next_callee)
3728 gcov_type dec = apply_probability (cs->count,
3729 GCOV_COMPUTE_SCALE (redirected_sum,
3730 orig_node_count));
3731 if (dec < cs->count)
3732 cs->count -= dec;
3733 else
3734 cs->count = 0;
3737 if (dump_file)
3738 dump_profile_updates (orig_node, new_node);
3741 /* Create a specialized version of NODE with known constants in KNOWN_CSTS,
3742 known contexts in KNOWN_CONTEXTS and known aggregate values in AGGVALS and
3743 redirect all edges in CALLERS to it. */
3745 static struct cgraph_node *
3746 create_specialized_node (struct cgraph_node *node,
3747 vec<tree> known_csts,
3748 vec<ipa_polymorphic_call_context> known_contexts,
3749 struct ipa_agg_replacement_value *aggvals,
3750 vec<cgraph_edge *> callers)
3752 struct ipa_node_params *new_info, *info = IPA_NODE_REF (node);
3753 vec<ipa_replace_map *, va_gc> *replace_trees = NULL;
3754 struct ipa_agg_replacement_value *av;
3755 struct cgraph_node *new_node;
3756 int i, count = ipa_get_param_count (info);
3757 bitmap args_to_skip;
3759 gcc_assert (!info->ipcp_orig_node);
3761 if (node->local.can_change_signature)
3763 args_to_skip = BITMAP_GGC_ALLOC ();
3764 for (i = 0; i < count; i++)
3766 tree t = known_csts[i];
3768 if (t || !ipa_is_param_used (info, i))
3769 bitmap_set_bit (args_to_skip, i);
3772 else
3774 args_to_skip = NULL;
3775 if (dump_file && (dump_flags & TDF_DETAILS))
3776 fprintf (dump_file, " cannot change function signature\n");
3779 for (i = 0; i < count; i++)
3781 tree t = known_csts[i];
3782 if (t)
3784 struct ipa_replace_map *replace_map;
3786 gcc_checking_assert (TREE_CODE (t) != TREE_BINFO);
3787 replace_map = get_replacement_map (info, t, i);
3788 if (replace_map)
3789 vec_safe_push (replace_trees, replace_map);
3793 new_node = node->create_virtual_clone (callers, replace_trees,
3794 args_to_skip, "constprop");
3795 ipa_set_node_agg_value_chain (new_node, aggvals);
3796 for (av = aggvals; av; av = av->next)
3797 new_node->maybe_create_reference (av->value, NULL);
3799 if (dump_file && (dump_flags & TDF_DETAILS))
3801 fprintf (dump_file, " the new node is %s/%i.\n",
3802 new_node->name (), new_node->order);
3803 if (known_contexts.exists ())
3805 for (i = 0; i < count; i++)
3806 if (!known_contexts[i].useless_p ())
3808 fprintf (dump_file, " known ctx %i is ", i);
3809 known_contexts[i].dump (dump_file);
3812 if (aggvals)
3813 ipa_dump_agg_replacement_values (dump_file, aggvals);
3815 ipa_check_create_node_params ();
3816 update_profiling_info (node, new_node);
3817 new_info = IPA_NODE_REF (new_node);
3818 new_info->ipcp_orig_node = node;
3819 new_info->known_csts = known_csts;
3820 new_info->known_contexts = known_contexts;
3822 ipcp_discover_new_direct_edges (new_node, known_csts, known_contexts, aggvals);
3824 callers.release ();
3825 return new_node;
3828 /* Given a NODE, and a subset of its CALLERS, try to populate blanks slots in
3829 KNOWN_CSTS with constants that are also known for all of the CALLERS. */
3831 static void
3832 find_more_scalar_values_for_callers_subset (struct cgraph_node *node,
3833 vec<tree> known_csts,
3834 vec<cgraph_edge *> callers)
3836 struct ipa_node_params *info = IPA_NODE_REF (node);
3837 int i, count = ipa_get_param_count (info);
3839 for (i = 0; i < count; i++)
3841 struct cgraph_edge *cs;
3842 tree newval = NULL_TREE;
3843 int j;
3844 bool first = true;
3846 if (ipa_get_scalar_lat (info, i)->bottom || known_csts[i])
3847 continue;
3849 FOR_EACH_VEC_ELT (callers, j, cs)
3851 struct ipa_jump_func *jump_func;
3852 tree t;
3854 if (i >= ipa_get_cs_argument_count (IPA_EDGE_REF (cs))
3855 || (i == 0
3856 && call_passes_through_thunk_p (cs))
3857 || (!cs->callee->instrumentation_clone
3858 && cs->callee->function_symbol ()->instrumentation_clone))
3860 newval = NULL_TREE;
3861 break;
3863 jump_func = ipa_get_ith_jump_func (IPA_EDGE_REF (cs), i);
3864 t = ipa_value_from_jfunc (IPA_NODE_REF (cs->caller), jump_func);
3865 if (!t
3866 || (newval
3867 && !values_equal_for_ipcp_p (t, newval))
3868 || (!first && !newval))
3870 newval = NULL_TREE;
3871 break;
3873 else
3874 newval = t;
3875 first = false;
3878 if (newval)
3880 if (dump_file && (dump_flags & TDF_DETAILS))
3882 fprintf (dump_file, " adding an extra known scalar value ");
3883 print_ipcp_constant_value (dump_file, newval);
3884 fprintf (dump_file, " for ");
3885 ipa_dump_param (dump_file, info, i);
3886 fprintf (dump_file, "\n");
3889 known_csts[i] = newval;
3894 /* Given a NODE and a subset of its CALLERS, try to populate plank slots in
3895 KNOWN_CONTEXTS with polymorphic contexts that are also known for all of the
3896 CALLERS. */
3898 static void
3899 find_more_contexts_for_caller_subset (cgraph_node *node,
3900 vec<ipa_polymorphic_call_context>
3901 *known_contexts,
3902 vec<cgraph_edge *> callers)
3904 ipa_node_params *info = IPA_NODE_REF (node);
3905 int i, count = ipa_get_param_count (info);
3907 for (i = 0; i < count; i++)
3909 cgraph_edge *cs;
3911 if (ipa_get_poly_ctx_lat (info, i)->bottom
3912 || (known_contexts->exists ()
3913 && !(*known_contexts)[i].useless_p ()))
3914 continue;
3916 ipa_polymorphic_call_context newval;
3917 bool first = true;
3918 int j;
3920 FOR_EACH_VEC_ELT (callers, j, cs)
3922 if (i >= ipa_get_cs_argument_count (IPA_EDGE_REF (cs)))
3923 return;
3924 ipa_jump_func *jfunc = ipa_get_ith_jump_func (IPA_EDGE_REF (cs),
3926 ipa_polymorphic_call_context ctx;
3927 ctx = ipa_context_from_jfunc (IPA_NODE_REF (cs->caller), cs, i,
3928 jfunc);
3929 if (first)
3931 newval = ctx;
3932 first = false;
3934 else
3935 newval.meet_with (ctx);
3936 if (newval.useless_p ())
3937 break;
3940 if (!newval.useless_p ())
3942 if (dump_file && (dump_flags & TDF_DETAILS))
3944 fprintf (dump_file, " adding an extra known polymorphic "
3945 "context ");
3946 print_ipcp_constant_value (dump_file, newval);
3947 fprintf (dump_file, " for ");
3948 ipa_dump_param (dump_file, info, i);
3949 fprintf (dump_file, "\n");
3952 if (!known_contexts->exists ())
3953 known_contexts->safe_grow_cleared (ipa_get_param_count (info));
3954 (*known_contexts)[i] = newval;
3960 /* Go through PLATS and create a vector of values consisting of values and
3961 offsets (minus OFFSET) of lattices that contain only a single value. */
3963 static vec<ipa_agg_jf_item>
3964 copy_plats_to_inter (struct ipcp_param_lattices *plats, HOST_WIDE_INT offset)
3966 vec<ipa_agg_jf_item> res = vNULL;
3968 if (!plats->aggs || plats->aggs_contain_variable || plats->aggs_bottom)
3969 return vNULL;
3971 for (struct ipcp_agg_lattice *aglat = plats->aggs; aglat; aglat = aglat->next)
3972 if (aglat->is_single_const ())
3974 struct ipa_agg_jf_item ti;
3975 ti.offset = aglat->offset - offset;
3976 ti.value = aglat->values->value;
3977 res.safe_push (ti);
3979 return res;
3982 /* Intersect all values in INTER with single value lattices in PLATS (while
3983 subtracting OFFSET). */
3985 static void
3986 intersect_with_plats (struct ipcp_param_lattices *plats,
3987 vec<ipa_agg_jf_item> *inter,
3988 HOST_WIDE_INT offset)
3990 struct ipcp_agg_lattice *aglat;
3991 struct ipa_agg_jf_item *item;
3992 int k;
3994 if (!plats->aggs || plats->aggs_contain_variable || plats->aggs_bottom)
3996 inter->release ();
3997 return;
4000 aglat = plats->aggs;
4001 FOR_EACH_VEC_ELT (*inter, k, item)
4003 bool found = false;
4004 if (!item->value)
4005 continue;
4006 while (aglat)
4008 if (aglat->offset - offset > item->offset)
4009 break;
4010 if (aglat->offset - offset == item->offset)
4012 gcc_checking_assert (item->value);
4013 if (values_equal_for_ipcp_p (item->value, aglat->values->value))
4014 found = true;
4015 break;
4017 aglat = aglat->next;
4019 if (!found)
4020 item->value = NULL_TREE;
4024 /* Copy aggregate replacement values of NODE (which is an IPA-CP clone) to the
4025 vector result while subtracting OFFSET from the individual value offsets. */
4027 static vec<ipa_agg_jf_item>
4028 agg_replacements_to_vector (struct cgraph_node *node, int index,
4029 HOST_WIDE_INT offset)
4031 struct ipa_agg_replacement_value *av;
4032 vec<ipa_agg_jf_item> res = vNULL;
4034 for (av = ipa_get_agg_replacements_for_node (node); av; av = av->next)
4035 if (av->index == index
4036 && (av->offset - offset) >= 0)
4038 struct ipa_agg_jf_item item;
4039 gcc_checking_assert (av->value);
4040 item.offset = av->offset - offset;
4041 item.value = av->value;
4042 res.safe_push (item);
4045 return res;
4048 /* Intersect all values in INTER with those that we have already scheduled to
4049 be replaced in parameter number INDEX of NODE, which is an IPA-CP clone
4050 (while subtracting OFFSET). */
4052 static void
4053 intersect_with_agg_replacements (struct cgraph_node *node, int index,
4054 vec<ipa_agg_jf_item> *inter,
4055 HOST_WIDE_INT offset)
4057 struct ipa_agg_replacement_value *srcvals;
4058 struct ipa_agg_jf_item *item;
4059 int i;
4061 srcvals = ipa_get_agg_replacements_for_node (node);
4062 if (!srcvals)
4064 inter->release ();
4065 return;
4068 FOR_EACH_VEC_ELT (*inter, i, item)
4070 struct ipa_agg_replacement_value *av;
4071 bool found = false;
4072 if (!item->value)
4073 continue;
4074 for (av = srcvals; av; av = av->next)
4076 gcc_checking_assert (av->value);
4077 if (av->index == index
4078 && av->offset - offset == item->offset)
4080 if (values_equal_for_ipcp_p (item->value, av->value))
4081 found = true;
4082 break;
4085 if (!found)
4086 item->value = NULL_TREE;
4090 /* Intersect values in INTER with aggregate values that come along edge CS to
4091 parameter number INDEX and return it. If INTER does not actually exist yet,
4092 copy all incoming values to it. If we determine we ended up with no values
4093 whatsoever, return a released vector. */
4095 static vec<ipa_agg_jf_item>
4096 intersect_aggregates_with_edge (struct cgraph_edge *cs, int index,
4097 vec<ipa_agg_jf_item> inter)
4099 struct ipa_jump_func *jfunc;
4100 jfunc = ipa_get_ith_jump_func (IPA_EDGE_REF (cs), index);
4101 if (jfunc->type == IPA_JF_PASS_THROUGH
4102 && ipa_get_jf_pass_through_operation (jfunc) == NOP_EXPR)
4104 struct ipa_node_params *caller_info = IPA_NODE_REF (cs->caller);
4105 int src_idx = ipa_get_jf_pass_through_formal_id (jfunc);
4107 if (caller_info->ipcp_orig_node)
4109 struct cgraph_node *orig_node = caller_info->ipcp_orig_node;
4110 struct ipcp_param_lattices *orig_plats;
4111 orig_plats = ipa_get_parm_lattices (IPA_NODE_REF (orig_node),
4112 src_idx);
4113 if (agg_pass_through_permissible_p (orig_plats, jfunc))
4115 if (!inter.exists ())
4116 inter = agg_replacements_to_vector (cs->caller, src_idx, 0);
4117 else
4118 intersect_with_agg_replacements (cs->caller, src_idx,
4119 &inter, 0);
4121 else
4123 inter.release ();
4124 return vNULL;
4127 else
4129 struct ipcp_param_lattices *src_plats;
4130 src_plats = ipa_get_parm_lattices (caller_info, src_idx);
4131 if (agg_pass_through_permissible_p (src_plats, jfunc))
4133 /* Currently we do not produce clobber aggregate jump
4134 functions, adjust when we do. */
4135 gcc_checking_assert (!jfunc->agg.items);
4136 if (!inter.exists ())
4137 inter = copy_plats_to_inter (src_plats, 0);
4138 else
4139 intersect_with_plats (src_plats, &inter, 0);
4141 else
4143 inter.release ();
4144 return vNULL;
4148 else if (jfunc->type == IPA_JF_ANCESTOR
4149 && ipa_get_jf_ancestor_agg_preserved (jfunc))
4151 struct ipa_node_params *caller_info = IPA_NODE_REF (cs->caller);
4152 int src_idx = ipa_get_jf_ancestor_formal_id (jfunc);
4153 struct ipcp_param_lattices *src_plats;
4154 HOST_WIDE_INT delta = ipa_get_jf_ancestor_offset (jfunc);
4156 if (caller_info->ipcp_orig_node)
4158 if (!inter.exists ())
4159 inter = agg_replacements_to_vector (cs->caller, src_idx, delta);
4160 else
4161 intersect_with_agg_replacements (cs->caller, src_idx, &inter,
4162 delta);
4164 else
4166 src_plats = ipa_get_parm_lattices (caller_info, src_idx);;
4167 /* Currently we do not produce clobber aggregate jump
4168 functions, adjust when we do. */
4169 gcc_checking_assert (!src_plats->aggs || !jfunc->agg.items);
4170 if (!inter.exists ())
4171 inter = copy_plats_to_inter (src_plats, delta);
4172 else
4173 intersect_with_plats (src_plats, &inter, delta);
4176 else if (jfunc->agg.items)
4178 struct ipa_agg_jf_item *item;
4179 int k;
4181 if (!inter.exists ())
4182 for (unsigned i = 0; i < jfunc->agg.items->length (); i++)
4183 inter.safe_push ((*jfunc->agg.items)[i]);
4184 else
4185 FOR_EACH_VEC_ELT (inter, k, item)
4187 int l = 0;
4188 bool found = false;;
4190 if (!item->value)
4191 continue;
4193 while ((unsigned) l < jfunc->agg.items->length ())
4195 struct ipa_agg_jf_item *ti;
4196 ti = &(*jfunc->agg.items)[l];
4197 if (ti->offset > item->offset)
4198 break;
4199 if (ti->offset == item->offset)
4201 gcc_checking_assert (ti->value);
4202 if (values_equal_for_ipcp_p (item->value,
4203 ti->value))
4204 found = true;
4205 break;
4207 l++;
4209 if (!found)
4210 item->value = NULL;
4213 else
4215 inter.release ();
4216 return vec<ipa_agg_jf_item>();
4218 return inter;
4221 /* Look at edges in CALLERS and collect all known aggregate values that arrive
4222 from all of them. */
4224 static struct ipa_agg_replacement_value *
4225 find_aggregate_values_for_callers_subset (struct cgraph_node *node,
4226 vec<cgraph_edge *> callers)
4228 struct ipa_node_params *dest_info = IPA_NODE_REF (node);
4229 struct ipa_agg_replacement_value *res;
4230 struct ipa_agg_replacement_value **tail = &res;
4231 struct cgraph_edge *cs;
4232 int i, j, count = ipa_get_param_count (dest_info);
4234 FOR_EACH_VEC_ELT (callers, j, cs)
4236 int c = ipa_get_cs_argument_count (IPA_EDGE_REF (cs));
4237 if (c < count)
4238 count = c;
4241 for (i = 0; i < count; i++)
4243 struct cgraph_edge *cs;
4244 vec<ipa_agg_jf_item> inter = vNULL;
4245 struct ipa_agg_jf_item *item;
4246 struct ipcp_param_lattices *plats = ipa_get_parm_lattices (dest_info, i);
4247 int j;
4249 /* Among other things, the following check should deal with all by_ref
4250 mismatches. */
4251 if (plats->aggs_bottom)
4252 continue;
4254 FOR_EACH_VEC_ELT (callers, j, cs)
4256 inter = intersect_aggregates_with_edge (cs, i, inter);
4258 if (!inter.exists ())
4259 goto next_param;
4262 FOR_EACH_VEC_ELT (inter, j, item)
4264 struct ipa_agg_replacement_value *v;
4266 if (!item->value)
4267 continue;
4269 v = ggc_alloc<ipa_agg_replacement_value> ();
4270 v->index = i;
4271 v->offset = item->offset;
4272 v->value = item->value;
4273 v->by_ref = plats->aggs_by_ref;
4274 *tail = v;
4275 tail = &v->next;
4278 next_param:
4279 if (inter.exists ())
4280 inter.release ();
4282 *tail = NULL;
4283 return res;
4286 /* Turn KNOWN_AGGS into a list of aggregate replacement values. */
4288 static struct ipa_agg_replacement_value *
4289 known_aggs_to_agg_replacement_list (vec<ipa_agg_jump_function> known_aggs)
4291 struct ipa_agg_replacement_value *res;
4292 struct ipa_agg_replacement_value **tail = &res;
4293 struct ipa_agg_jump_function *aggjf;
4294 struct ipa_agg_jf_item *item;
4295 int i, j;
4297 FOR_EACH_VEC_ELT (known_aggs, i, aggjf)
4298 FOR_EACH_VEC_SAFE_ELT (aggjf->items, j, item)
4300 struct ipa_agg_replacement_value *v;
4301 v = ggc_alloc<ipa_agg_replacement_value> ();
4302 v->index = i;
4303 v->offset = item->offset;
4304 v->value = item->value;
4305 v->by_ref = aggjf->by_ref;
4306 *tail = v;
4307 tail = &v->next;
4309 *tail = NULL;
4310 return res;
4313 /* Determine whether CS also brings all scalar values that the NODE is
4314 specialized for. */
4316 static bool
4317 cgraph_edge_brings_all_scalars_for_node (struct cgraph_edge *cs,
4318 struct cgraph_node *node)
4320 struct ipa_node_params *dest_info = IPA_NODE_REF (node);
4321 int count = ipa_get_param_count (dest_info);
4322 struct ipa_node_params *caller_info;
4323 struct ipa_edge_args *args;
4324 int i;
4326 caller_info = IPA_NODE_REF (cs->caller);
4327 args = IPA_EDGE_REF (cs);
4328 for (i = 0; i < count; i++)
4330 struct ipa_jump_func *jump_func;
4331 tree val, t;
4333 val = dest_info->known_csts[i];
4334 if (!val)
4335 continue;
4337 if (i >= ipa_get_cs_argument_count (args))
4338 return false;
4339 jump_func = ipa_get_ith_jump_func (args, i);
4340 t = ipa_value_from_jfunc (caller_info, jump_func);
4341 if (!t || !values_equal_for_ipcp_p (val, t))
4342 return false;
4344 return true;
4347 /* Determine whether CS also brings all aggregate values that NODE is
4348 specialized for. */
4349 static bool
4350 cgraph_edge_brings_all_agg_vals_for_node (struct cgraph_edge *cs,
4351 struct cgraph_node *node)
4353 struct ipa_node_params *orig_caller_info = IPA_NODE_REF (cs->caller);
4354 struct ipa_node_params *orig_node_info;
4355 struct ipa_agg_replacement_value *aggval;
4356 int i, ec, count;
4358 aggval = ipa_get_agg_replacements_for_node (node);
4359 if (!aggval)
4360 return true;
4362 count = ipa_get_param_count (IPA_NODE_REF (node));
4363 ec = ipa_get_cs_argument_count (IPA_EDGE_REF (cs));
4364 if (ec < count)
4365 for (struct ipa_agg_replacement_value *av = aggval; av; av = av->next)
4366 if (aggval->index >= ec)
4367 return false;
4369 orig_node_info = IPA_NODE_REF (IPA_NODE_REF (node)->ipcp_orig_node);
4370 if (orig_caller_info->ipcp_orig_node)
4371 orig_caller_info = IPA_NODE_REF (orig_caller_info->ipcp_orig_node);
4373 for (i = 0; i < count; i++)
4375 static vec<ipa_agg_jf_item> values = vec<ipa_agg_jf_item>();
4376 struct ipcp_param_lattices *plats;
4377 bool interesting = false;
4378 for (struct ipa_agg_replacement_value *av = aggval; av; av = av->next)
4379 if (aggval->index == i)
4381 interesting = true;
4382 break;
4384 if (!interesting)
4385 continue;
4387 plats = ipa_get_parm_lattices (orig_node_info, aggval->index);
4388 if (plats->aggs_bottom)
4389 return false;
4391 values = intersect_aggregates_with_edge (cs, i, values);
4392 if (!values.exists ())
4393 return false;
4395 for (struct ipa_agg_replacement_value *av = aggval; av; av = av->next)
4396 if (aggval->index == i)
4398 struct ipa_agg_jf_item *item;
4399 int j;
4400 bool found = false;
4401 FOR_EACH_VEC_ELT (values, j, item)
4402 if (item->value
4403 && item->offset == av->offset
4404 && values_equal_for_ipcp_p (item->value, av->value))
4406 found = true;
4407 break;
4409 if (!found)
4411 values.release ();
4412 return false;
4416 return true;
4419 /* Given an original NODE and a VAL for which we have already created a
4420 specialized clone, look whether there are incoming edges that still lead
4421 into the old node but now also bring the requested value and also conform to
4422 all other criteria such that they can be redirected the special node.
4423 This function can therefore redirect the final edge in a SCC. */
4425 template <typename valtype>
4426 static void
4427 perhaps_add_new_callers (cgraph_node *node, ipcp_value<valtype> *val)
4429 ipcp_value_source<valtype> *src;
4430 gcov_type redirected_sum = 0;
4432 for (src = val->sources; src; src = src->next)
4434 struct cgraph_edge *cs = src->cs;
4435 while (cs)
4437 if (cgraph_edge_brings_value_p (cs, src, node)
4438 && cgraph_edge_brings_all_scalars_for_node (cs, val->spec_node)
4439 && cgraph_edge_brings_all_agg_vals_for_node (cs, val->spec_node))
4441 if (dump_file)
4442 fprintf (dump_file, " - adding an extra caller %s/%i"
4443 " of %s/%i\n",
4444 xstrdup_for_dump (cs->caller->name ()),
4445 cs->caller->order,
4446 xstrdup_for_dump (val->spec_node->name ()),
4447 val->spec_node->order);
4449 cs->redirect_callee_duplicating_thunks (val->spec_node);
4450 val->spec_node->expand_all_artificial_thunks ();
4451 redirected_sum += cs->count;
4453 cs = get_next_cgraph_edge_clone (cs);
4457 if (redirected_sum)
4458 update_specialized_profile (val->spec_node, node, redirected_sum);
4461 /* Return true if KNOWN_CONTEXTS contain at least one useful context. */
4463 static bool
4464 known_contexts_useful_p (vec<ipa_polymorphic_call_context> known_contexts)
4466 ipa_polymorphic_call_context *ctx;
4467 int i;
4469 FOR_EACH_VEC_ELT (known_contexts, i, ctx)
4470 if (!ctx->useless_p ())
4471 return true;
4472 return false;
4475 /* Return a copy of KNOWN_CSTS if it is not empty, otherwise return vNULL. */
4477 static vec<ipa_polymorphic_call_context>
4478 copy_useful_known_contexts (vec<ipa_polymorphic_call_context> known_contexts)
4480 if (known_contexts_useful_p (known_contexts))
4481 return known_contexts.copy ();
4482 else
4483 return vNULL;
4486 /* Copy KNOWN_CSTS and modify the copy according to VAL and INDEX. If
4487 non-empty, replace KNOWN_CONTEXTS with its copy too. */
4489 static void
4490 modify_known_vectors_with_val (vec<tree> *known_csts,
4491 vec<ipa_polymorphic_call_context> *known_contexts,
4492 ipcp_value<tree> *val,
4493 int index)
4495 *known_csts = known_csts->copy ();
4496 *known_contexts = copy_useful_known_contexts (*known_contexts);
4497 (*known_csts)[index] = val->value;
4500 /* Replace KNOWN_CSTS with its copy. Also copy KNOWN_CONTEXTS and modify the
4501 copy according to VAL and INDEX. */
4503 static void
4504 modify_known_vectors_with_val (vec<tree> *known_csts,
4505 vec<ipa_polymorphic_call_context> *known_contexts,
4506 ipcp_value<ipa_polymorphic_call_context> *val,
4507 int index)
4509 *known_csts = known_csts->copy ();
4510 *known_contexts = known_contexts->copy ();
4511 (*known_contexts)[index] = val->value;
4514 /* Return true if OFFSET indicates this was not an aggregate value or there is
4515 a replacement equivalent to VALUE, INDEX and OFFSET among those in the
4516 AGGVALS list. */
4518 DEBUG_FUNCTION bool
4519 ipcp_val_agg_replacement_ok_p (ipa_agg_replacement_value *aggvals,
4520 int index, HOST_WIDE_INT offset, tree value)
4522 if (offset == -1)
4523 return true;
4525 while (aggvals)
4527 if (aggvals->index == index
4528 && aggvals->offset == offset
4529 && values_equal_for_ipcp_p (aggvals->value, value))
4530 return true;
4531 aggvals = aggvals->next;
4533 return false;
4536 /* Return true if offset is minus one because source of a polymorphic contect
4537 cannot be an aggregate value. */
4539 DEBUG_FUNCTION bool
4540 ipcp_val_agg_replacement_ok_p (ipa_agg_replacement_value *,
4541 int , HOST_WIDE_INT offset,
4542 ipa_polymorphic_call_context)
4544 return offset == -1;
4547 /* Decide wheter to create a special version of NODE for value VAL of parameter
4548 at the given INDEX. If OFFSET is -1, the value is for the parameter itself,
4549 otherwise it is stored at the given OFFSET of the parameter. KNOWN_CSTS,
4550 KNOWN_CONTEXTS and KNOWN_AGGS describe the other already known values. */
4552 template <typename valtype>
4553 static bool
4554 decide_about_value (struct cgraph_node *node, int index, HOST_WIDE_INT offset,
4555 ipcp_value<valtype> *val, vec<tree> known_csts,
4556 vec<ipa_polymorphic_call_context> known_contexts)
4558 struct ipa_agg_replacement_value *aggvals;
4559 int freq_sum, caller_count;
4560 gcov_type count_sum;
4561 vec<cgraph_edge *> callers;
4563 if (val->spec_node)
4565 perhaps_add_new_callers (node, val);
4566 return false;
4568 else if (val->local_size_cost + overall_size > max_new_size)
4570 if (dump_file && (dump_flags & TDF_DETAILS))
4571 fprintf (dump_file, " Ignoring candidate value because "
4572 "max_new_size would be reached with %li.\n",
4573 val->local_size_cost + overall_size);
4574 return false;
4576 else if (!get_info_about_necessary_edges (val, node, &freq_sum, &count_sum,
4577 &caller_count))
4578 return false;
4580 if (dump_file && (dump_flags & TDF_DETAILS))
4582 fprintf (dump_file, " - considering value ");
4583 print_ipcp_constant_value (dump_file, val->value);
4584 fprintf (dump_file, " for ");
4585 ipa_dump_param (dump_file, IPA_NODE_REF (node), index);
4586 if (offset != -1)
4587 fprintf (dump_file, ", offset: " HOST_WIDE_INT_PRINT_DEC, offset);
4588 fprintf (dump_file, " (caller_count: %i)\n", caller_count);
4591 if (!good_cloning_opportunity_p (node, val->local_time_benefit,
4592 freq_sum, count_sum,
4593 val->local_size_cost)
4594 && !good_cloning_opportunity_p (node,
4595 val->local_time_benefit
4596 + val->prop_time_benefit,
4597 freq_sum, count_sum,
4598 val->local_size_cost
4599 + val->prop_size_cost))
4600 return false;
4602 if (dump_file)
4603 fprintf (dump_file, " Creating a specialized node of %s/%i.\n",
4604 node->name (), node->order);
4606 callers = gather_edges_for_value (val, node, caller_count);
4607 if (offset == -1)
4608 modify_known_vectors_with_val (&known_csts, &known_contexts, val, index);
4609 else
4611 known_csts = known_csts.copy ();
4612 known_contexts = copy_useful_known_contexts (known_contexts);
4614 find_more_scalar_values_for_callers_subset (node, known_csts, callers);
4615 find_more_contexts_for_caller_subset (node, &known_contexts, callers);
4616 aggvals = find_aggregate_values_for_callers_subset (node, callers);
4617 gcc_checking_assert (ipcp_val_agg_replacement_ok_p (aggvals, index,
4618 offset, val->value));
4619 val->spec_node = create_specialized_node (node, known_csts, known_contexts,
4620 aggvals, callers);
4621 overall_size += val->local_size_cost;
4623 /* TODO: If for some lattice there is only one other known value
4624 left, make a special node for it too. */
4626 return true;
4629 /* Decide whether and what specialized clones of NODE should be created. */
4631 static bool
4632 decide_whether_version_node (struct cgraph_node *node)
4634 struct ipa_node_params *info = IPA_NODE_REF (node);
4635 int i, count = ipa_get_param_count (info);
4636 vec<tree> known_csts;
4637 vec<ipa_polymorphic_call_context> known_contexts;
4638 vec<ipa_agg_jump_function> known_aggs = vNULL;
4639 bool ret = false;
4641 if (count == 0)
4642 return false;
4644 if (dump_file && (dump_flags & TDF_DETAILS))
4645 fprintf (dump_file, "\nEvaluating opportunities for %s/%i.\n",
4646 node->name (), node->order);
4648 gather_context_independent_values (info, &known_csts, &known_contexts,
4649 info->do_clone_for_all_contexts ? &known_aggs
4650 : NULL, NULL);
4652 for (i = 0; i < count;i++)
4654 struct ipcp_param_lattices *plats = ipa_get_parm_lattices (info, i);
4655 ipcp_lattice<tree> *lat = &plats->itself;
4656 ipcp_lattice<ipa_polymorphic_call_context> *ctxlat = &plats->ctxlat;
4658 if (!lat->bottom
4659 && !known_csts[i])
4661 ipcp_value<tree> *val;
4662 for (val = lat->values; val; val = val->next)
4663 ret |= decide_about_value (node, i, -1, val, known_csts,
4664 known_contexts);
4667 if (!plats->aggs_bottom)
4669 struct ipcp_agg_lattice *aglat;
4670 ipcp_value<tree> *val;
4671 for (aglat = plats->aggs; aglat; aglat = aglat->next)
4672 if (!aglat->bottom && aglat->values
4673 /* If the following is false, the one value is in
4674 known_aggs. */
4675 && (plats->aggs_contain_variable
4676 || !aglat->is_single_const ()))
4677 for (val = aglat->values; val; val = val->next)
4678 ret |= decide_about_value (node, i, aglat->offset, val,
4679 known_csts, known_contexts);
4682 if (!ctxlat->bottom
4683 && known_contexts[i].useless_p ())
4685 ipcp_value<ipa_polymorphic_call_context> *val;
4686 for (val = ctxlat->values; val; val = val->next)
4687 ret |= decide_about_value (node, i, -1, val, known_csts,
4688 known_contexts);
4691 info = IPA_NODE_REF (node);
4694 if (info->do_clone_for_all_contexts)
4696 struct cgraph_node *clone;
4697 vec<cgraph_edge *> callers;
4699 if (dump_file)
4700 fprintf (dump_file, " - Creating a specialized node of %s/%i "
4701 "for all known contexts.\n", node->name (),
4702 node->order);
4704 callers = node->collect_callers ();
4706 if (!known_contexts_useful_p (known_contexts))
4708 known_contexts.release ();
4709 known_contexts = vNULL;
4711 clone = create_specialized_node (node, known_csts, known_contexts,
4712 known_aggs_to_agg_replacement_list (known_aggs),
4713 callers);
4714 info = IPA_NODE_REF (node);
4715 info->do_clone_for_all_contexts = false;
4716 IPA_NODE_REF (clone)->is_all_contexts_clone = true;
4717 for (i = 0; i < count; i++)
4718 vec_free (known_aggs[i].items);
4719 known_aggs.release ();
4720 ret = true;
4722 else
4724 known_csts.release ();
4725 known_contexts.release ();
4728 return ret;
4731 /* Transitively mark all callees of NODE within the same SCC as not dead. */
4733 static void
4734 spread_undeadness (struct cgraph_node *node)
4736 struct cgraph_edge *cs;
4738 for (cs = node->callees; cs; cs = cs->next_callee)
4739 if (ipa_edge_within_scc (cs))
4741 struct cgraph_node *callee;
4742 struct ipa_node_params *info;
4744 callee = cs->callee->function_symbol (NULL);
4745 info = IPA_NODE_REF (callee);
4747 if (info->node_dead)
4749 info->node_dead = 0;
4750 spread_undeadness (callee);
4755 /* Return true if NODE has a caller from outside of its SCC that is not
4756 dead. Worker callback for cgraph_for_node_and_aliases. */
4758 static bool
4759 has_undead_caller_from_outside_scc_p (struct cgraph_node *node,
4760 void *data ATTRIBUTE_UNUSED)
4762 struct cgraph_edge *cs;
4764 for (cs = node->callers; cs; cs = cs->next_caller)
4765 if (cs->caller->thunk.thunk_p
4766 && cs->caller->call_for_symbol_thunks_and_aliases
4767 (has_undead_caller_from_outside_scc_p, NULL, true))
4768 return true;
4769 else if (!ipa_edge_within_scc (cs)
4770 && !IPA_NODE_REF (cs->caller)->node_dead)
4771 return true;
4772 return false;
4776 /* Identify nodes within the same SCC as NODE which are no longer needed
4777 because of new clones and will be removed as unreachable. */
4779 static void
4780 identify_dead_nodes (struct cgraph_node *node)
4782 struct cgraph_node *v;
4783 for (v = node; v; v = ((struct ipa_dfs_info *) v->aux)->next_cycle)
4784 if (v->local.local
4785 && !v->call_for_symbol_thunks_and_aliases
4786 (has_undead_caller_from_outside_scc_p, NULL, true))
4787 IPA_NODE_REF (v)->node_dead = 1;
4789 for (v = node; v; v = ((struct ipa_dfs_info *) v->aux)->next_cycle)
4790 if (!IPA_NODE_REF (v)->node_dead)
4791 spread_undeadness (v);
4793 if (dump_file && (dump_flags & TDF_DETAILS))
4795 for (v = node; v; v = ((struct ipa_dfs_info *) v->aux)->next_cycle)
4796 if (IPA_NODE_REF (v)->node_dead)
4797 fprintf (dump_file, " Marking node as dead: %s/%i.\n",
4798 v->name (), v->order);
4802 /* The decision stage. Iterate over the topological order of call graph nodes
4803 TOPO and make specialized clones if deemed beneficial. */
4805 static void
4806 ipcp_decision_stage (struct ipa_topo_info *topo)
4808 int i;
4810 if (dump_file)
4811 fprintf (dump_file, "\nIPA decision stage:\n\n");
4813 for (i = topo->nnodes - 1; i >= 0; i--)
4815 struct cgraph_node *node = topo->order[i];
4816 bool change = false, iterate = true;
4818 while (iterate)
4820 struct cgraph_node *v;
4821 iterate = false;
4822 for (v = node; v; v = ((struct ipa_dfs_info *) v->aux)->next_cycle)
4823 if (v->has_gimple_body_p ()
4824 && ipcp_versionable_function_p (v))
4825 iterate |= decide_whether_version_node (v);
4827 change |= iterate;
4829 if (change)
4830 identify_dead_nodes (node);
4834 /* Look up all the bits information that we have discovered and copy it over
4835 to the transformation summary. */
4837 static void
4838 ipcp_store_bits_results (void)
4840 cgraph_node *node;
4842 FOR_EACH_FUNCTION_WITH_GIMPLE_BODY (node)
4844 ipa_node_params *info = IPA_NODE_REF (node);
4845 bool dumped_sth = false;
4846 bool found_useful_result = false;
4848 if (!opt_for_fn (node->decl, flag_ipa_bit_cp))
4850 if (dump_file)
4851 fprintf (dump_file, "Not considering %s for ipa bitwise propagation "
4852 "; -fipa-bit-cp: disabled.\n",
4853 node->name ());
4854 continue;
4857 if (info->ipcp_orig_node)
4858 info = IPA_NODE_REF (info->ipcp_orig_node);
4860 unsigned count = ipa_get_param_count (info);
4861 for (unsigned i = 0; i < count; i++)
4863 ipcp_param_lattices *plats = ipa_get_parm_lattices (info, i);
4864 if (plats->bits_lattice.constant_p ())
4866 found_useful_result = true;
4867 break;
4871 if (!found_useful_result)
4872 continue;
4874 ipcp_grow_transformations_if_necessary ();
4875 ipcp_transformation_summary *ts = ipcp_get_transformation_summary (node);
4876 vec_safe_reserve_exact (ts->bits, count);
4878 for (unsigned i = 0; i < count; i++)
4880 ipcp_param_lattices *plats = ipa_get_parm_lattices (info, i);
4881 ipa_bits *jfbits;
4883 if (plats->bits_lattice.constant_p ())
4884 jfbits
4885 = ipa_get_ipa_bits_for_value (plats->bits_lattice.get_value (),
4886 plats->bits_lattice.get_mask ());
4887 else
4888 jfbits = NULL;
4890 ts->bits->quick_push (jfbits);
4891 if (!dump_file || !jfbits)
4892 continue;
4893 if (!dumped_sth)
4895 fprintf (dump_file, "Propagated bits info for function %s/%i:\n",
4896 node->name (), node->order);
4897 dumped_sth = true;
4899 fprintf (dump_file, " param %i: value = ", i);
4900 print_hex (jfbits->value, dump_file);
4901 fprintf (dump_file, ", mask = ");
4902 print_hex (jfbits->mask, dump_file);
4903 fprintf (dump_file, "\n");
4908 /* Look up all VR information that we have discovered and copy it over
4909 to the transformation summary. */
4911 static void
4912 ipcp_store_vr_results (void)
4914 cgraph_node *node;
4916 FOR_EACH_FUNCTION_WITH_GIMPLE_BODY (node)
4918 ipa_node_params *info = IPA_NODE_REF (node);
4919 bool found_useful_result = false;
4921 if (!opt_for_fn (node->decl, flag_ipa_vrp))
4923 if (dump_file)
4924 fprintf (dump_file, "Not considering %s for VR discovery "
4925 "and propagate; -fipa-ipa-vrp: disabled.\n",
4926 node->name ());
4927 continue;
4930 if (info->ipcp_orig_node)
4931 info = IPA_NODE_REF (info->ipcp_orig_node);
4933 unsigned count = ipa_get_param_count (info);
4934 for (unsigned i = 0; i < count; i++)
4936 ipcp_param_lattices *plats = ipa_get_parm_lattices (info, i);
4937 if (!plats->m_value_range.bottom_p ()
4938 && !plats->m_value_range.top_p ())
4940 found_useful_result = true;
4941 break;
4944 if (!found_useful_result)
4945 continue;
4947 ipcp_grow_transformations_if_necessary ();
4948 ipcp_transformation_summary *ts = ipcp_get_transformation_summary (node);
4949 vec_safe_reserve_exact (ts->m_vr, count);
4951 for (unsigned i = 0; i < count; i++)
4953 ipcp_param_lattices *plats = ipa_get_parm_lattices (info, i);
4954 ipa_vr vr;
4956 if (!plats->m_value_range.bottom_p ()
4957 && !plats->m_value_range.top_p ())
4959 vr.known = true;
4960 vr.type = plats->m_value_range.m_vr.type;
4961 vr.min = plats->m_value_range.m_vr.min;
4962 vr.max = plats->m_value_range.m_vr.max;
4964 else
4966 vr.known = false;
4967 vr.type = VR_VARYING;
4968 vr.min = vr.max = wi::zero (INT_TYPE_SIZE);
4970 ts->m_vr->quick_push (vr);
4975 /* The IPCP driver. */
4977 static unsigned int
4978 ipcp_driver (void)
4980 struct cgraph_2edge_hook_list *edge_duplication_hook_holder;
4981 struct cgraph_edge_hook_list *edge_removal_hook_holder;
4982 struct ipa_topo_info topo;
4984 ipa_check_create_node_params ();
4985 ipa_check_create_edge_args ();
4986 grow_edge_clone_vectors ();
4987 edge_duplication_hook_holder
4988 = symtab->add_edge_duplication_hook (&ipcp_edge_duplication_hook, NULL);
4989 edge_removal_hook_holder
4990 = symtab->add_edge_removal_hook (&ipcp_edge_removal_hook, NULL);
4992 if (dump_file)
4994 fprintf (dump_file, "\nIPA structures before propagation:\n");
4995 if (dump_flags & TDF_DETAILS)
4996 ipa_print_all_params (dump_file);
4997 ipa_print_all_jump_functions (dump_file);
5000 /* Topological sort. */
5001 build_toporder_info (&topo);
5002 /* Do the interprocedural propagation. */
5003 ipcp_propagate_stage (&topo);
5004 /* Decide what constant propagation and cloning should be performed. */
5005 ipcp_decision_stage (&topo);
5006 /* Store results of bits propagation. */
5007 ipcp_store_bits_results ();
5008 /* Store results of value range propagation. */
5009 ipcp_store_vr_results ();
5011 /* Free all IPCP structures. */
5012 free_toporder_info (&topo);
5013 next_edge_clone.release ();
5014 prev_edge_clone.release ();
5015 symtab->remove_edge_removal_hook (edge_removal_hook_holder);
5016 symtab->remove_edge_duplication_hook (edge_duplication_hook_holder);
5017 ipa_free_all_structures_after_ipa_cp ();
5018 if (dump_file)
5019 fprintf (dump_file, "\nIPA constant propagation end\n");
5020 return 0;
5023 /* Initialization and computation of IPCP data structures. This is the initial
5024 intraprocedural analysis of functions, which gathers information to be
5025 propagated later on. */
5027 static void
5028 ipcp_generate_summary (void)
5030 struct cgraph_node *node;
5032 if (dump_file)
5033 fprintf (dump_file, "\nIPA constant propagation start:\n");
5034 ipa_register_cgraph_hooks ();
5036 FOR_EACH_FUNCTION_WITH_GIMPLE_BODY (node)
5037 ipa_analyze_node (node);
5040 /* Write ipcp summary for nodes in SET. */
5042 static void
5043 ipcp_write_summary (void)
5045 ipa_prop_write_jump_functions ();
5048 /* Read ipcp summary. */
5050 static void
5051 ipcp_read_summary (void)
5053 ipa_prop_read_jump_functions ();
5056 namespace {
5058 const pass_data pass_data_ipa_cp =
5060 IPA_PASS, /* type */
5061 "cp", /* name */
5062 OPTGROUP_NONE, /* optinfo_flags */
5063 TV_IPA_CONSTANT_PROP, /* tv_id */
5064 0, /* properties_required */
5065 0, /* properties_provided */
5066 0, /* properties_destroyed */
5067 0, /* todo_flags_start */
5068 ( TODO_dump_symtab | TODO_remove_functions ), /* todo_flags_finish */
5071 class pass_ipa_cp : public ipa_opt_pass_d
5073 public:
5074 pass_ipa_cp (gcc::context *ctxt)
5075 : ipa_opt_pass_d (pass_data_ipa_cp, ctxt,
5076 ipcp_generate_summary, /* generate_summary */
5077 ipcp_write_summary, /* write_summary */
5078 ipcp_read_summary, /* read_summary */
5079 ipcp_write_transformation_summaries, /*
5080 write_optimization_summary */
5081 ipcp_read_transformation_summaries, /*
5082 read_optimization_summary */
5083 NULL, /* stmt_fixup */
5084 0, /* function_transform_todo_flags_start */
5085 ipcp_transform_function, /* function_transform */
5086 NULL) /* variable_transform */
5089 /* opt_pass methods: */
5090 virtual bool gate (function *)
5092 /* FIXME: We should remove the optimize check after we ensure we never run
5093 IPA passes when not optimizing. */
5094 return (flag_ipa_cp && optimize) || in_lto_p;
5097 virtual unsigned int execute (function *) { return ipcp_driver (); }
5099 }; // class pass_ipa_cp
5101 } // anon namespace
5103 ipa_opt_pass_d *
5104 make_pass_ipa_cp (gcc::context *ctxt)
5106 return new pass_ipa_cp (ctxt);
5109 /* Reset all state within ipa-cp.c so that we can rerun the compiler
5110 within the same process. For use by toplev::finalize. */
5112 void
5113 ipa_cp_c_finalize (void)
5115 max_count = 0;
5116 overall_size = 0;
5117 max_new_size = 0;