Line data Source code
1 : /* Interprocedural constant propagation
2 : Copyright (C) 2005-2026 Free Software Foundation, Inc.
3 :
4 : Contributed by Razya Ladelsky <RAZYA@il.ibm.com> and Martin Jambor
5 : <mjambor@suse.cz>
6 :
7 : This file is part of GCC.
8 :
9 : GCC is free software; you can redistribute it and/or modify it under
10 : the terms of the GNU General Public License as published by the Free
11 : Software Foundation; either version 3, or (at your option) any later
12 : version.
13 :
14 : GCC is distributed in the hope that it will be useful, but WITHOUT ANY
15 : WARRANTY; without even the implied warranty of MERCHANTABILITY or
16 : FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
17 : for more details.
18 :
19 : You should have received a copy of the GNU General Public License
20 : along with GCC; see the file COPYING3. If not see
21 : <http://www.gnu.org/licenses/>. */
22 :
23 : /* Interprocedural constant propagation (IPA-CP).
24 :
25 : The goal of this transformation is to
26 :
27 : 1) discover functions which are always invoked with some arguments with the
28 : same known constant values and modify the functions so that the
29 : subsequent optimizations can take advantage of the knowledge, and
30 :
31 : 2) partial specialization - create specialized versions of functions
32 : transformed in this way if some parameters are known constants only in
33 : certain contexts but the estimated tradeoff between speedup and cost size
34 : is deemed good.
35 :
36 : The algorithm also propagates types and attempts to perform type based
37 : devirtualization. Types are propagated much like constants.
38 :
39 : The algorithm basically consists of three stages. In the first, functions
40 : are analyzed one at a time and jump functions are constructed for all known
41 : call-sites. In the second phase, the pass propagates information from the
42 : jump functions across the call to reveal what values are available at what
43 : call sites, performs estimations of effects of known values on functions and
44 : their callees, and finally decides what specialized extra versions should be
45 : created. In the third, the special versions materialize and appropriate
46 : calls are redirected.
47 :
48 : The algorithm used is to a certain extent based on "Interprocedural Constant
49 : Propagation", by David Callahan, Keith D Cooper, Ken Kennedy, Linda Torczon,
50 : Comp86, pg 152-161 and "A Methodology for Procedure Cloning" by Keith D
51 : Cooper, Mary W. Hall, and Ken Kennedy.
52 :
53 :
54 : First stage - intraprocedural analysis
55 : =======================================
56 :
57 : This phase computes jump_function and modification flags.
58 :
59 : A jump function for a call-site represents the values passed as an actual
60 : arguments of a given call-site. In principle, there are three types of
61 : values:
62 :
63 : Pass through - the caller's formal parameter is passed as an actual
64 : argument, plus an operation on it can be performed.
65 : Constant - a constant is passed as an actual argument.
66 : Unknown - neither of the above.
67 :
68 : All jump function types are described in detail in ipa-prop.h, together with
69 : the data structures that represent them and methods of accessing them.
70 :
71 : ipcp_generate_summary() is the main function of the first stage.
72 :
73 : Second stage - interprocedural analysis
74 : ========================================
75 :
76 : This stage is itself divided into two phases. In the first, we propagate
77 : known values over the call graph, in the second, we make cloning decisions.
78 : It uses a different algorithm than the original Callahan's paper.
79 :
80 : First, we traverse the functions topologically from callers to callees and,
81 : for each strongly connected component (SCC), we propagate constants
82 : according to previously computed jump functions. We also record what known
83 : values depend on other known values and estimate local effects. Finally, we
84 : propagate cumulative information about these effects from dependent values
85 : to those on which they depend.
86 :
87 : Second, we again traverse the call graph in the same topological order and
88 : make clones for functions which we know are called with the same values in
89 : all contexts and decide about extra specialized clones of functions just for
90 : some contexts - these decisions are based on both local estimates and
91 : cumulative estimates propagated from callees.
92 :
93 : ipcp_propagate_stage() and ipcp_decision_stage() together constitute the
94 : third stage.
95 :
96 : Third phase - materialization of clones, call statement updates.
97 : ============================================
98 :
99 : This stage is currently performed by call graph code (mainly in cgraphunit.cc
100 : and tree-inline.cc) according to instructions inserted to the call graph by
101 : the second stage. */
102 :
103 : #define INCLUDE_ALGORITHM
104 : #include "config.h"
105 : #include "system.h"
106 : #include "coretypes.h"
107 : #include "backend.h"
108 : #include "tree.h"
109 : #include "gimple-expr.h"
110 : #include "gimple.h"
111 : #include "predict.h"
112 : #include "sreal.h"
113 : #include "alloc-pool.h"
114 : #include "tree-pass.h"
115 : #include "cgraph.h"
116 : #include "diagnostic.h"
117 : #include "fold-const.h"
118 : #include "gimple-iterator.h"
119 : #include "gimple-fold.h"
120 : #include "symbol-summary.h"
121 : #include "tree-vrp.h"
122 : #include "ipa-cp.h"
123 : #include "ipa-prop.h"
124 : #include "tree-pretty-print.h"
125 : #include "tree-inline.h"
126 : #include "ipa-fnsummary.h"
127 : #include "ipa-utils.h"
128 : #include "tree-ssa-ccp.h"
129 : #include "stringpool.h"
130 : #include "attribs.h"
131 : #include "dbgcnt.h"
132 : #include "symtab-clones.h"
133 : #include "gimple-range.h"
134 : #include "attr-callback.h"
135 :
136 : /* Allocation pools for values and their sources in ipa-cp. */
137 :
138 : object_allocator<ipcp_value<tree> > ipcp_cst_values_pool
139 : ("IPA-CP constant values");
140 :
141 : object_allocator<ipcp_value<ipa_polymorphic_call_context> >
142 : ipcp_poly_ctx_values_pool ("IPA-CP polymorphic contexts");
143 :
144 : object_allocator<ipcp_value_source<tree> > ipcp_sources_pool
145 : ("IPA-CP value sources");
146 :
147 : object_allocator<ipcp_agg_lattice> ipcp_agg_lattice_pool
148 : ("IPA_CP aggregate lattices");
149 :
150 : /* Original overall size of the program. */
151 :
152 : static long overall_size, orig_overall_size;
153 :
154 : /* The maximum number of IPA-CP decision sweeps that any node requested in its
155 : param. */
156 : static int max_number_sweeps;
157 :
158 : /* Node name to unique clone suffix number map. */
159 : static hash_map<const char *, unsigned> *clone_num_suffixes;
160 :
161 : /* Return the param lattices structure corresponding to the Ith formal
162 : parameter of the function described by INFO. */
163 : static inline class ipcp_param_lattices *
164 33647381 : ipa_get_parm_lattices (class ipa_node_params *info, int i)
165 : {
166 67294762 : gcc_assert (i >= 0 && i < ipa_get_param_count (info));
167 33647381 : gcc_checking_assert (!info->ipcp_orig_node);
168 33647381 : return &(info->lattices[i]);
169 : }
170 :
171 : /* Return the lattice corresponding to the scalar value of the Ith formal
172 : parameter of the function described by INFO. */
173 : static inline ipcp_lattice<tree> *
174 5935509 : ipa_get_scalar_lat (class ipa_node_params *info, int i)
175 : {
176 6120034 : class ipcp_param_lattices *plats = ipa_get_parm_lattices (info, i);
177 5935509 : return &plats->itself;
178 : }
179 :
180 : /* Return the lattice corresponding to the scalar value of the Ith formal
181 : parameter of the function described by INFO. */
182 : static inline ipcp_lattice<ipa_polymorphic_call_context> *
183 798013 : ipa_get_poly_ctx_lat (class ipa_node_params *info, int i)
184 : {
185 798013 : class ipcp_param_lattices *plats = ipa_get_parm_lattices (info, i);
186 798013 : return &plats->ctxlat;
187 : }
188 :
189 : /* Return whether LAT is a lattice with a single constant and without an
190 : undefined value. */
191 :
192 : template <typename valtype>
193 : inline bool
194 15057625 : ipcp_lattice<valtype>::is_single_const ()
195 : {
196 3225947 : if (bottom || contains_variable || values_count != 1)
197 : return false;
198 : else
199 : return true;
200 : }
201 :
202 : /* Return true iff X and Y should be considered equal values by IPA-CP. */
203 :
204 : bool
205 1400573 : values_equal_for_ipcp_p (tree x, tree y)
206 : {
207 1400573 : gcc_checking_assert (x != NULL_TREE && y != NULL_TREE);
208 :
209 1400573 : if (x == y)
210 : return true;
211 :
212 621625 : if (TREE_CODE (x) == ADDR_EXPR
213 221300 : && TREE_CODE (y) == ADDR_EXPR
214 220668 : && (TREE_CODE (TREE_OPERAND (x, 0)) == CONST_DECL
215 173295 : || (TREE_CODE (TREE_OPERAND (x, 0)) == VAR_DECL
216 92356 : && DECL_IN_CONSTANT_POOL (TREE_OPERAND (x, 0))))
217 668998 : && (TREE_CODE (TREE_OPERAND (y, 0)) == CONST_DECL
218 13 : || (TREE_CODE (TREE_OPERAND (y, 0)) == VAR_DECL
219 8 : && DECL_IN_CONSTANT_POOL (TREE_OPERAND (y, 0)))))
220 47360 : return TREE_OPERAND (x, 0) == TREE_OPERAND (y, 0)
221 94464 : || operand_equal_p (DECL_INITIAL (TREE_OPERAND (x, 0)),
222 47104 : DECL_INITIAL (TREE_OPERAND (y, 0)), 0);
223 : else
224 574265 : return operand_equal_p (x, y, 0);
225 : }
226 :
227 : /* Print V which is extracted from a value in a lattice to F. This overloaded
228 : function is used to print tree constants. */
229 :
230 : static void
231 715 : print_ipcp_constant_value (FILE * f, tree v)
232 : {
233 0 : ipa_print_constant_value (f, v);
234 44 : }
235 :
236 : /* Print V which is extracted from a value in a lattice to F. This overloaded
237 : function is used to print constant polymorphic call contexts. */
238 :
239 : static void
240 214 : print_ipcp_constant_value (FILE * f, ipa_polymorphic_call_context v)
241 : {
242 214 : v.dump(f, false);
243 0 : }
244 :
245 : /* Print a lattice LAT to F. */
246 :
247 : template <typename valtype>
248 : void
249 1985 : ipcp_lattice<valtype>::print (FILE * f, bool dump_sources, bool dump_benefits)
250 : {
251 : ipcp_value<valtype> *val;
252 1985 : bool prev = false;
253 :
254 1985 : if (bottom)
255 : {
256 844 : fprintf (f, "BOTTOM\n");
257 844 : return;
258 : }
259 :
260 1141 : if (!values_count && !contains_variable)
261 : {
262 0 : fprintf (f, "TOP\n");
263 0 : return;
264 : }
265 :
266 1141 : if (contains_variable)
267 : {
268 861 : fprintf (f, "VARIABLE");
269 861 : prev = true;
270 861 : if (dump_benefits)
271 861 : fprintf (f, "\n");
272 : }
273 :
274 1770 : for (val = values; val; val = val->next)
275 : {
276 629 : if (dump_benefits && prev)
277 349 : fprintf (f, " ");
278 280 : else if (!dump_benefits && prev)
279 0 : fprintf (f, ", ");
280 : else
281 : prev = true;
282 :
283 629 : print_ipcp_constant_value (f, val->value);
284 :
285 629 : if (dump_sources)
286 : {
287 : ipcp_value_source<valtype> *s;
288 :
289 174 : if (val->self_recursion_generated_p ())
290 27 : fprintf (f, " [self_gen(%i), from:",
291 : val->self_recursion_generated_level);
292 : else
293 147 : fprintf (f, " [scc: %i, from:", val->scc_no);
294 366 : for (s = val->sources; s; s = s->next)
295 192 : fprintf (f, " %i(%f)", s->cs->caller->get_uid (),
296 384 : s->cs->sreal_frequency ().to_double ());
297 174 : fprintf (f, "]");
298 : }
299 :
300 629 : if (dump_benefits)
301 629 : fprintf (f, " [loc_time: %g, loc_size: %i, "
302 : "prop_time: %g, prop_size: %i]\n",
303 : val->local_time_benefit.to_double (), val->local_size_cost,
304 : val->prop_time_benefit.to_double (), val->prop_size_cost);
305 : }
306 1141 : if (!dump_benefits)
307 0 : fprintf (f, "\n");
308 : }
309 :
310 : /* Print VALUE to F in a form which in usual cases does not take thousands of
311 : characters. */
312 :
313 : static void
314 1450 : ipcp_print_widest_int (FILE *f, const widest_int &value)
315 : {
316 1450 : if (value == -1)
317 0 : fprintf (f, "-1");
318 1450 : else if (wi::arshift (value, 128) == -1)
319 : {
320 330 : char buf[35], *p = buf + 2;
321 330 : widest_int v = wi::zext (value, 128);
322 330 : size_t len;
323 330 : print_hex (v, buf);
324 330 : len = strlen (p);
325 330 : if (len == 32)
326 : {
327 330 : fprintf (f, "0xf..f");
328 9795 : while (*p == 'f')
329 9135 : ++p;
330 : }
331 : else
332 0 : fprintf (f, "0xf..f%0*d", (int) (32 - len), 0);
333 330 : fputs (p, f);
334 330 : }
335 : else
336 1120 : print_hex (value, f);
337 1450 : }
338 :
339 : void
340 917 : ipcp_bits_lattice::print (FILE *f)
341 : {
342 917 : if (bottom_p ())
343 : {
344 607 : fprintf (f, " Bits unusable (BOTTOM)\n");
345 607 : return;
346 : }
347 :
348 310 : if (top_p ())
349 0 : fprintf (f, " Bits unknown (TOP)");
350 : else
351 : {
352 310 : fprintf (f, " Bits: value = ");
353 310 : ipcp_print_widest_int (f, get_value ());
354 310 : fprintf (f, ", mask = ");
355 310 : ipcp_print_widest_int (f, get_mask ());
356 : }
357 :
358 310 : if (m_recipient_only)
359 136 : fprintf (f, " (recipient only)");
360 310 : fprintf (f, "\n");
361 : }
362 :
363 : /* Print value range lattice to F. */
364 :
365 : void
366 917 : ipcp_vr_lattice::print (FILE * f)
367 : {
368 917 : if (m_recipient_only)
369 263 : fprintf (f, "(recipient only) ");
370 917 : m_vr.dump (f);
371 917 : }
372 :
373 : /* Print all ipcp_lattices of all functions to F. */
374 :
375 : static void
376 161 : print_all_lattices (FILE * f, bool dump_sources, bool dump_benefits)
377 : {
378 161 : struct cgraph_node *node;
379 161 : int i, count;
380 :
381 161 : fprintf (f, "\nLattices:\n");
382 889 : FOR_EACH_FUNCTION_WITH_GIMPLE_BODY (node)
383 : {
384 728 : class ipa_node_params *info;
385 :
386 728 : info = ipa_node_params_sum->get (node);
387 : /* Skip unoptimized functions and constprop clones since we don't make
388 : lattices for them. */
389 728 : if (!info || info->ipcp_orig_node)
390 0 : continue;
391 728 : fprintf (f, " Node: %s:\n", node->dump_name ());
392 728 : count = ipa_get_param_count (info);
393 1645 : for (i = 0; i < count; i++)
394 : {
395 917 : struct ipcp_agg_lattice *aglat;
396 917 : class ipcp_param_lattices *plats = ipa_get_parm_lattices (info, i);
397 917 : fprintf (f, " param [%d]: ", i);
398 917 : plats->itself.print (f, dump_sources, dump_benefits);
399 917 : fprintf (f, " ctxs: ");
400 917 : plats->ctxlat.print (f, dump_sources, dump_benefits);
401 917 : plats->bits_lattice.print (f);
402 917 : fprintf (f, " ");
403 917 : plats->m_value_range.print (f);
404 917 : fprintf (f, "\n");
405 917 : if (plats->virt_call)
406 75 : fprintf (f, " virt_call flag set\n");
407 :
408 917 : if (plats->aggs_bottom)
409 : {
410 442 : fprintf (f, " AGGS BOTTOM\n");
411 442 : continue;
412 : }
413 475 : if (plats->aggs_contain_variable)
414 437 : fprintf (f, " AGGS VARIABLE\n");
415 626 : for (aglat = plats->aggs; aglat; aglat = aglat->next)
416 : {
417 151 : fprintf (f, " %soffset " HOST_WIDE_INT_PRINT_DEC ": ",
418 151 : plats->aggs_by_ref ? "ref " : "", aglat->offset);
419 151 : aglat->print (f, dump_sources, dump_benefits);
420 : }
421 : }
422 : }
423 161 : }
424 :
425 : /* Determine whether it is at all technically possible to create clones of NODE
426 : and store this information in the ipa_node_params structure associated
427 : with NODE. */
428 :
429 : static void
430 1269638 : determine_versionability (struct cgraph_node *node,
431 : class ipa_node_params *info)
432 : {
433 1269638 : const char *reason = NULL;
434 :
435 : /* There are a number of generic reasons functions cannot be versioned. We
436 : also cannot remove parameters if there are type attributes such as fnspec
437 : present. */
438 1269638 : if (node->alias || node->thunk)
439 : reason = "alias or thunk";
440 1269638 : else if (!node->versionable)
441 : reason = "not a tree_versionable_function";
442 1140708 : else if (node->get_availability () <= AVAIL_INTERPOSABLE)
443 : reason = "insufficient body availability";
444 1073723 : else if (!opt_for_fn (node->decl, optimize)
445 1073723 : || !opt_for_fn (node->decl, flag_ipa_cp))
446 : reason = "non-optimized function";
447 1073723 : else if (lookup_attribute ("omp declare simd", DECL_ATTRIBUTES (node->decl)))
448 : {
449 : /* Ideally we should clone the SIMD clones themselves and create
450 : vector copies of them, so IPA-cp and SIMD clones can happily
451 : coexist, but that may not be worth the effort. */
452 : reason = "function has SIMD clones";
453 : }
454 1073363 : else if (lookup_attribute ("target_clones", DECL_ATTRIBUTES (node->decl)))
455 : {
456 : /* Ideally we should clone the target clones themselves and create
457 : copies of them, so IPA-cp and target clones can happily
458 : coexist, but that may not be worth the effort. */
459 : reason = "function target_clones attribute";
460 : }
461 : /* Don't clone decls local to a comdat group; it breaks and for C++
462 : decloned constructors, inlining is always better anyway. */
463 1073363 : else if (node->comdat_local_p ())
464 : reason = "comdat-local function";
465 1071208 : else if (node->calls_comdat_local)
466 : {
467 : /* TODO: call is versionable if we make sure that all
468 : callers are inside of a comdat group. */
469 2257 : reason = "calls comdat-local function";
470 : }
471 :
472 : /* Functions calling BUILT_IN_VA_ARG_PACK and BUILT_IN_VA_ARG_PACK_LEN
473 : work only when inlined. Cloning them may still lead to better code
474 : because ipa-cp will not give up on cloning further. If the function is
475 : external this however leads to wrong code because we may end up producing
476 : offline copy of the function. */
477 1269638 : if (DECL_EXTERNAL (node->decl))
478 175471 : for (cgraph_edge *edge = node->callees; !reason && edge;
479 129050 : edge = edge->next_callee)
480 129050 : if (fndecl_built_in_p (edge->callee->decl, BUILT_IN_NORMAL))
481 : {
482 36366 : if (DECL_FUNCTION_CODE (edge->callee->decl) == BUILT_IN_VA_ARG_PACK)
483 0 : reason = "external function which calls va_arg_pack";
484 36366 : if (DECL_FUNCTION_CODE (edge->callee->decl)
485 : == BUILT_IN_VA_ARG_PACK_LEN)
486 0 : reason = "external function which calls va_arg_pack_len";
487 : }
488 :
489 1269638 : if (reason && dump_file && !node->alias && !node->thunk)
490 56 : fprintf (dump_file, "Function %s is not versionable, reason: %s.\n",
491 : node->dump_name (), reason);
492 :
493 1269638 : info->versionable = (reason == NULL);
494 1269638 : }
495 :
496 : /* Return true if it is at all technically possible to create clones of a
497 : NODE. */
498 :
499 : static bool
500 6079596 : ipcp_versionable_function_p (struct cgraph_node *node)
501 : {
502 6079596 : ipa_node_params *info = ipa_node_params_sum->get (node);
503 6079596 : return info && info->versionable;
504 : }
505 :
506 : /* Structure holding accumulated information about callers of a node. */
507 :
508 3405795 : struct caller_statistics
509 : {
510 : /* If requested (see below), self-recursive call counts are summed into this
511 : field. */
512 : profile_count rec_count_sum;
513 : /* The sum of all ipa counts of all the other (non-recursive) calls. */
514 : profile_count count_sum;
515 : /* Sum of all frequencies for all calls. */
516 : sreal freq_sum;
517 : /* Number of calls and calls considered interesting respectively. */
518 : int n_calls, n_interesting_calls;
519 : /* If itself is set up, also count the number of non-self-recursive
520 : calls. */
521 : int n_nonrec_calls;
522 : /* If non-NULL, this is the node itself and calls from it should have their
523 : counts included in rec_count_sum and not count_sum. */
524 : cgraph_node *itself;
525 : /* True if there is a caller that has no IPA profile. */
526 : bool called_without_ipa_profile;
527 : };
528 :
529 : /* Initialize fields of STAT to zeroes and optionally set it up so that edges
530 : from IGNORED_CALLER are not counted. */
531 :
532 : static inline void
533 2660995 : init_caller_stats (caller_statistics *stats, cgraph_node *itself = NULL)
534 : {
535 2660995 : stats->rec_count_sum = profile_count::zero ();
536 2660995 : stats->count_sum = profile_count::zero ();
537 2660995 : stats->n_calls = 0;
538 2660995 : stats->n_interesting_calls = 0;
539 2660995 : stats->n_nonrec_calls = 0;
540 2660995 : stats->freq_sum = 0;
541 2660995 : stats->itself = itself;
542 2660995 : stats->called_without_ipa_profile = false;
543 2660995 : }
544 :
545 : /* We want to propagate across edges that may be executed, however
546 : we do not want to check maybe_hot, since call itself may be cold
547 : while calee contains some heavy loop which makes propagation still
548 : relevant.
549 :
550 : In particular, even edge called once may lead to significant
551 : improvement. */
552 :
553 : static bool
554 4715177 : cs_interesting_for_ipcp_p (cgraph_edge *e)
555 : {
556 : /* If profile says the edge is executed, we want to optimize. */
557 4715177 : if (e->count.ipa ().nonzero_p ())
558 899 : return true;
559 : /* If local (possibly guseed or adjusted 0 profile) claims edge is
560 : not executed, do not propagate.
561 : Do not trust AFDO since branch needs to be executed multiple
562 : time to count while we want to propagate even call called
563 : once during the train run if callee is important. */
564 4714278 : if (e->count.initialized_p () && !e->count.nonzero_p ()
565 5381403 : && e->count.quality () != AFDO)
566 : return false;
567 : /* If we have zero IPA profile, still consider edge for cloning
568 : in case we do partial training. */
569 4047153 : if (e->count.ipa ().initialized_p ()
570 4047153 : && e->count.ipa ().quality () != AFDO
571 4047168 : && !opt_for_fn (e->callee->decl,flag_profile_partial_training))
572 15 : return false;
573 : return true;
574 : }
575 :
576 : /* Worker callback of cgraph_for_node_and_aliases accumulating statistics of
577 : non-thunk incoming edges to NODE. */
578 :
579 : static bool
580 2837205 : gather_caller_stats (struct cgraph_node *node, void *data)
581 : {
582 2837205 : struct caller_statistics *stats = (struct caller_statistics *) data;
583 2837205 : struct cgraph_edge *cs;
584 :
585 7344420 : for (cs = node->callers; cs; cs = cs->next_caller)
586 4507215 : if (!cs->caller->thunk)
587 : {
588 4503714 : ipa_node_params *info = ipa_node_params_sum->get (cs->caller);
589 4503714 : if (info && info->node_dead)
590 162036 : continue;
591 :
592 4341678 : if (cs->count.ipa ().initialized_p ())
593 : {
594 336320 : if (stats->itself && stats->itself == cs->caller)
595 0 : stats->rec_count_sum += cs->count.ipa ();
596 : else
597 336320 : stats->count_sum += cs->count.ipa ();
598 : }
599 : else
600 4005358 : stats->called_without_ipa_profile = true;
601 4341678 : stats->freq_sum += cs->sreal_frequency ();
602 4341678 : stats->n_calls++;
603 4341678 : if (stats->itself && stats->itself != cs->caller)
604 8 : stats->n_nonrec_calls++;
605 :
606 : /* If profile known to be zero, we do not want to clone for performance.
607 : However if call is cold, the called function may still contain
608 : important hot loops. */
609 4341678 : if (cs_interesting_for_ipcp_p (cs))
610 3719930 : stats->n_interesting_calls++;
611 : }
612 2837205 : return false;
613 :
614 : }
615 :
616 : /* Return true if this NODE is viable candidate for cloning. */
617 :
618 : static bool
619 791857 : ipcp_cloning_candidate_p (struct cgraph_node *node)
620 : {
621 791857 : struct caller_statistics stats;
622 :
623 791857 : gcc_checking_assert (node->has_gimple_body_p ());
624 :
625 791857 : if (!opt_for_fn (node->decl, flag_ipa_cp_clone))
626 : {
627 740621 : if (dump_file)
628 31 : fprintf (dump_file, "Not considering %s for cloning; "
629 : "-fipa-cp-clone disabled.\n",
630 : node->dump_name ());
631 740621 : return false;
632 : }
633 :
634 : /* Do not use profile here since cold wrapper wrap
635 : hot function. */
636 51236 : if (opt_for_fn (node->decl, optimize_size))
637 : {
638 10 : if (dump_file)
639 0 : fprintf (dump_file, "Not considering %s for cloning; "
640 : "optimizing it for size.\n",
641 : node->dump_name ());
642 10 : return false;
643 : }
644 :
645 51226 : init_caller_stats (&stats);
646 51226 : node->call_for_symbol_thunks_and_aliases (gather_caller_stats, &stats, false);
647 :
648 51226 : if (ipa_size_summaries->get (node)->self_size < stats.n_calls)
649 : {
650 303 : if (dump_file)
651 0 : fprintf (dump_file, "Considering %s for cloning; code might shrink.\n",
652 : node->dump_name ());
653 303 : return true;
654 : }
655 50923 : if (!stats.n_interesting_calls)
656 : {
657 39062 : if (dump_file)
658 201 : fprintf (dump_file, "Not considering %s for cloning; "
659 : "no calls considered interesting by profile.\n",
660 : node->dump_name ());
661 39062 : return false;
662 : }
663 11861 : if (dump_file)
664 186 : fprintf (dump_file, "Considering %s for cloning.\n",
665 : node->dump_name ());
666 : return true;
667 : }
668 :
669 : template <typename valtype>
670 : class value_topo_info
671 : {
672 : public:
673 : /* Head of the linked list of topologically sorted values. */
674 : ipcp_value<valtype> *values_topo;
675 : /* Stack for creating SCCs, represented by a linked list too. */
676 : ipcp_value<valtype> *stack;
677 : /* Counter driving the algorithm in add_val_to_toposort. */
678 : int dfs_counter;
679 :
680 129027 : value_topo_info () : values_topo (NULL), stack (NULL), dfs_counter (0)
681 : {}
682 : void add_val (ipcp_value<valtype> *cur_val);
683 : void propagate_effects ();
684 : };
685 :
686 : /* Arrays representing a topological ordering of call graph nodes and a stack
687 : of nodes used during constant propagation and also data required to perform
688 : topological sort of values and propagation of benefits in the determined
689 : order. */
690 :
691 : class ipa_topo_info
692 : {
693 : public:
694 : /* Array with obtained topological order of cgraph nodes. */
695 : struct cgraph_node **order;
696 : /* Stack of cgraph nodes used during propagation within SCC until all values
697 : in the SCC stabilize. */
698 : struct cgraph_node **stack;
699 : int nnodes, stack_top;
700 :
701 : value_topo_info<tree> constants;
702 : value_topo_info<ipa_polymorphic_call_context> contexts;
703 :
704 129027 : ipa_topo_info () : order(NULL), stack(NULL), nnodes(0), stack_top(0),
705 129027 : constants ()
706 : {}
707 : };
708 :
709 : /* Skip edges from and to nodes without ipa_cp enabled.
710 : Ignore not available symbols. */
711 :
712 : static bool
713 5269093 : ignore_edge_p (cgraph_edge *e)
714 : {
715 5269093 : enum availability avail;
716 5269093 : cgraph_node *ultimate_target
717 5269093 : = e->callee->function_or_virtual_thunk_symbol (&avail, e->caller);
718 :
719 5269093 : return (avail <= AVAIL_INTERPOSABLE
720 1857677 : || !opt_for_fn (ultimate_target->decl, optimize)
721 7118044 : || !opt_for_fn (ultimate_target->decl, flag_ipa_cp));
722 : }
723 :
724 : /* Allocate the arrays in TOPO and topologically sort the nodes into order. */
725 :
726 : static void
727 129027 : build_toporder_info (class ipa_topo_info *topo)
728 : {
729 129027 : topo->order = XCNEWVEC (struct cgraph_node *, symtab->cgraph_count);
730 129027 : topo->stack = XCNEWVEC (struct cgraph_node *, symtab->cgraph_count);
731 :
732 129027 : gcc_checking_assert (topo->stack_top == 0);
733 129027 : topo->nnodes = ipa_reduced_postorder (topo->order, true,
734 : ignore_edge_p);
735 129027 : }
736 :
737 : /* Free information about strongly connected components and the arrays in
738 : TOPO. */
739 :
740 : static void
741 129027 : free_toporder_info (class ipa_topo_info *topo)
742 : {
743 129027 : ipa_free_postorder_info ();
744 129027 : free (topo->order);
745 129027 : free (topo->stack);
746 129027 : }
747 :
748 : /* Add NODE to the stack in TOPO, unless it is already there. */
749 :
750 : static inline void
751 1273791 : push_node_to_stack (class ipa_topo_info *topo, struct cgraph_node *node)
752 : {
753 1273791 : ipa_node_params *info = ipa_node_params_sum->get (node);
754 1273791 : if (info->node_enqueued)
755 : return;
756 1272822 : info->node_enqueued = 1;
757 1272822 : topo->stack[topo->stack_top++] = node;
758 : }
759 :
760 : /* Pop a node from the stack in TOPO and return it or return NULL if the stack
761 : is empty. */
762 :
763 : static struct cgraph_node *
764 2623424 : pop_node_from_stack (class ipa_topo_info *topo)
765 : {
766 2623424 : if (topo->stack_top)
767 : {
768 1272822 : struct cgraph_node *node;
769 1272822 : topo->stack_top--;
770 1272822 : node = topo->stack[topo->stack_top];
771 1272822 : ipa_node_params_sum->get (node)->node_enqueued = 0;
772 1272822 : return node;
773 : }
774 : else
775 : return NULL;
776 : }
777 :
778 : /* Set lattice LAT to bottom and return true if it previously was not set as
779 : such. */
780 :
781 : template <typename valtype>
782 : inline bool
783 2112135 : ipcp_lattice<valtype>::set_to_bottom ()
784 : {
785 2112135 : bool ret = !bottom;
786 2112135 : bottom = true;
787 : return ret;
788 : }
789 :
790 : /* Mark lattice as containing an unknown value and return true if it previously
791 : was not marked as such. */
792 :
793 : template <typename valtype>
794 : inline bool
795 1539476 : ipcp_lattice<valtype>::set_contains_variable ()
796 : {
797 1539476 : bool ret = !contains_variable;
798 1539476 : contains_variable = true;
799 : return ret;
800 : }
801 :
802 : /* Set all aggregate lattices in PLATS to bottom and return true if they were
803 : not previously set as such. */
804 :
805 : static inline bool
806 2111842 : set_agg_lats_to_bottom (class ipcp_param_lattices *plats)
807 : {
808 2111842 : bool ret = !plats->aggs_bottom;
809 2111842 : plats->aggs_bottom = true;
810 2111842 : return ret;
811 : }
812 :
813 : /* Mark all aggregate lattices in PLATS as containing an unknown value and
814 : return true if they were not previously marked as such. */
815 :
816 : static inline bool
817 1043895 : set_agg_lats_contain_variable (class ipcp_param_lattices *plats)
818 : {
819 1043895 : bool ret = !plats->aggs_contain_variable;
820 1043895 : plats->aggs_contain_variable = true;
821 1043895 : return ret;
822 : }
823 :
824 : bool
825 0 : ipcp_vr_lattice::meet_with (const ipcp_vr_lattice &other)
826 : {
827 0 : return meet_with_1 (other.m_vr);
828 : }
829 :
830 : /* Meet the current value of the lattice with the range described by
831 : P_VR. */
832 :
833 : bool
834 491549 : ipcp_vr_lattice::meet_with (const vrange &p_vr)
835 : {
836 491549 : return meet_with_1 (p_vr);
837 : }
838 :
839 : /* Meet the current value of the lattice with the range described by
840 : OTHER_VR. Return TRUE if anything changed. */
841 :
842 : bool
843 491549 : ipcp_vr_lattice::meet_with_1 (const vrange &other_vr)
844 : {
845 491549 : if (bottom_p ())
846 : return false;
847 :
848 491549 : if (other_vr.varying_p ())
849 0 : return set_to_bottom ();
850 :
851 491549 : bool res;
852 491549 : if (flag_checking)
853 : {
854 491549 : value_range save (m_vr);
855 491549 : res = m_vr.union_ (other_vr);
856 491549 : gcc_assert (res == (m_vr != save));
857 491549 : }
858 : else
859 0 : res = m_vr.union_ (other_vr);
860 : return res;
861 : }
862 :
863 : /* Return true if value range information in the lattice is yet unknown. */
864 :
865 : bool
866 : ipcp_vr_lattice::top_p () const
867 : {
868 167545 : return m_vr.undefined_p ();
869 : }
870 :
871 : /* Return true if value range information in the lattice is known to be
872 : unusable. */
873 :
874 : bool
875 4833372 : ipcp_vr_lattice::bottom_p () const
876 : {
877 491549 : return m_vr.varying_p ();
878 : }
879 :
880 : /* Set value range information in the lattice to bottom. Return true if it
881 : previously was in a different state. */
882 :
883 : bool
884 2385480 : ipcp_vr_lattice::set_to_bottom ()
885 : {
886 2385480 : if (m_vr.varying_p ())
887 : return false;
888 :
889 : /* Setting an unsupported type here forces the temporary to default
890 : to unsupported_range, which can handle VARYING/DEFINED ranges,
891 : but nothing else (union, intersect, etc). This allows us to set
892 : bottoms on any ranges, and is safe as all users of the lattice
893 : check for bottom first. */
894 2244452 : m_vr.set_range_class (void_type_node);
895 2244452 : m_vr.set_varying (void_type_node);
896 :
897 2244452 : return true;
898 : }
899 :
900 : /* Set the flag that this lattice is a recipient only, return true if it was
901 : not set before. */
902 :
903 : bool
904 29238 : ipcp_vr_lattice::set_recipient_only ()
905 : {
906 29238 : if (m_recipient_only)
907 : return false;
908 29238 : m_recipient_only = true;
909 29238 : return true;
910 : }
911 :
912 : /* Set lattice value to bottom, if it already isn't the case. */
913 :
914 : bool
915 2404562 : ipcp_bits_lattice::set_to_bottom ()
916 : {
917 2404562 : if (bottom_p ())
918 : return false;
919 2264062 : m_lattice_val = IPA_BITS_VARYING;
920 2264062 : m_value = 0;
921 2264062 : m_mask = -1;
922 2264062 : return true;
923 : }
924 :
925 : /* Set to constant if it isn't already. Only meant to be called
926 : when switching state from TOP. */
927 :
928 : bool
929 76357 : ipcp_bits_lattice::set_to_constant (widest_int value, widest_int mask)
930 : {
931 76357 : gcc_assert (top_p ());
932 76357 : m_lattice_val = IPA_BITS_CONSTANT;
933 76357 : m_value = wi::bit_and (wi::bit_not (mask), value);
934 76357 : m_mask = mask;
935 76357 : return true;
936 : }
937 :
938 : /* Return true if any of the known bits are non-zero. */
939 :
940 : bool
941 472 : ipcp_bits_lattice::known_nonzero_p () const
942 : {
943 472 : if (!constant_p ())
944 : return false;
945 472 : return wi::ne_p (wi::bit_and (wi::bit_not (m_mask), m_value), 0);
946 : }
947 :
948 : /* Set the flag that this lattice is a recipient only, return true if it was not
949 : set before. */
950 :
951 : bool
952 29238 : ipcp_bits_lattice::set_recipient_only ()
953 : {
954 29238 : if (m_recipient_only)
955 : return false;
956 29238 : m_recipient_only = true;
957 29238 : return true;
958 : }
959 :
960 : /* Convert operand to value, mask form. */
961 :
962 : void
963 2063 : ipcp_bits_lattice::get_value_and_mask (tree operand, widest_int *valuep, widest_int *maskp)
964 : {
965 2063 : wide_int get_nonzero_bits (const_tree);
966 :
967 2063 : if (TREE_CODE (operand) == INTEGER_CST)
968 : {
969 2063 : *valuep = wi::to_widest (operand);
970 2063 : *maskp = 0;
971 : }
972 : else
973 : {
974 0 : *valuep = 0;
975 0 : *maskp = -1;
976 : }
977 2063 : }
978 :
979 : /* Meet operation, similar to ccp_lattice_meet, we xor values
980 : if this->value, value have different values at same bit positions, we want
981 : to drop that bit to varying. Return true if mask is changed.
982 : This function assumes that the lattice value is in CONSTANT state. If
983 : DROP_ALL_ONES, mask out any known bits with value one afterwards. */
984 :
985 : bool
986 302239 : ipcp_bits_lattice::meet_with_1 (widest_int value, widest_int mask,
987 : unsigned precision, bool drop_all_ones)
988 : {
989 302239 : gcc_assert (constant_p ());
990 :
991 302239 : widest_int old_mask = m_mask;
992 302239 : m_mask = (m_mask | mask) | (m_value ^ value);
993 302239 : if (drop_all_ones)
994 199 : m_mask |= m_value;
995 :
996 302239 : widest_int cap_mask = wi::shifted_mask <widest_int> (0, precision, true);
997 302239 : m_mask |= cap_mask;
998 302239 : if (wi::sext (m_mask, precision) == -1)
999 3504 : return set_to_bottom ();
1000 :
1001 298735 : m_value &= ~m_mask;
1002 298735 : return m_mask != old_mask;
1003 302239 : }
1004 :
1005 : /* Meet the bits lattice with operand
1006 : described by <value, mask, sgn, precision. */
1007 :
1008 : bool
1009 410217 : ipcp_bits_lattice::meet_with (widest_int value, widest_int mask,
1010 : unsigned precision)
1011 : {
1012 410217 : if (bottom_p ())
1013 : return false;
1014 :
1015 410217 : if (top_p ())
1016 : {
1017 120599 : if (wi::sext (mask, precision) == -1)
1018 49499 : return set_to_bottom ();
1019 71100 : return set_to_constant (value, mask);
1020 : }
1021 :
1022 289618 : return meet_with_1 (value, mask, precision, false);
1023 : }
1024 :
1025 : /* Meet bits lattice with the result of bit_value_binop (other, operand)
1026 : if code is binary operation or bit_value_unop (other) if code is unary op.
1027 : In the case when code is nop_expr, no adjustment is required. If
1028 : DROP_ALL_ONES, mask out any known bits with value one afterwards. */
1029 :
1030 : bool
1031 21452 : ipcp_bits_lattice::meet_with (ipcp_bits_lattice& other, unsigned precision,
1032 : signop sgn, enum tree_code code, tree operand,
1033 : bool drop_all_ones)
1034 : {
1035 21452 : if (other.bottom_p ())
1036 0 : return set_to_bottom ();
1037 :
1038 21452 : if (bottom_p () || other.top_p ())
1039 : return false;
1040 :
1041 17994 : widest_int adjusted_value, adjusted_mask;
1042 :
1043 17994 : if (TREE_CODE_CLASS (code) == tcc_binary)
1044 : {
1045 2063 : tree type = TREE_TYPE (operand);
1046 2063 : widest_int o_value, o_mask;
1047 2063 : get_value_and_mask (operand, &o_value, &o_mask);
1048 :
1049 2063 : bit_value_binop (code, sgn, precision, &adjusted_value, &adjusted_mask,
1050 4126 : sgn, precision, other.get_value (), other.get_mask (),
1051 2063 : TYPE_SIGN (type), TYPE_PRECISION (type), o_value, o_mask);
1052 :
1053 2063 : if (wi::sext (adjusted_mask, precision) == -1)
1054 87 : return set_to_bottom ();
1055 2063 : }
1056 :
1057 15931 : else if (TREE_CODE_CLASS (code) == tcc_unary)
1058 : {
1059 31812 : bit_value_unop (code, sgn, precision, &adjusted_value,
1060 31812 : &adjusted_mask, sgn, precision, other.get_value (),
1061 15906 : other.get_mask ());
1062 :
1063 15906 : if (wi::sext (adjusted_mask, precision) == -1)
1064 4 : return set_to_bottom ();
1065 : }
1066 :
1067 : else
1068 25 : return set_to_bottom ();
1069 :
1070 17878 : if (top_p ())
1071 : {
1072 5257 : if (drop_all_ones)
1073 : {
1074 273 : adjusted_mask |= adjusted_value;
1075 273 : adjusted_value &= ~adjusted_mask;
1076 : }
1077 5257 : widest_int cap_mask = wi::shifted_mask <widest_int> (0, precision, true);
1078 5257 : adjusted_mask |= cap_mask;
1079 5257 : if (wi::sext (adjusted_mask, precision) == -1)
1080 0 : return set_to_bottom ();
1081 5257 : return set_to_constant (adjusted_value, adjusted_mask);
1082 5257 : }
1083 : else
1084 12621 : return meet_with_1 (adjusted_value, adjusted_mask, precision,
1085 : drop_all_ones);
1086 17994 : }
1087 :
1088 : /* Dump the contents of the list to FILE. */
1089 :
1090 : void
1091 120 : ipa_argagg_value_list::dump (FILE *f)
1092 : {
1093 120 : bool comma = false;
1094 333 : for (const ipa_argagg_value &av : m_elts)
1095 : {
1096 213 : fprintf (f, "%s %i[%u]=", comma ? "," : "",
1097 213 : av.index, av.unit_offset);
1098 213 : print_generic_expr (f, av.value);
1099 213 : if (av.by_ref)
1100 186 : fprintf (f, "(by_ref)");
1101 213 : if (av.killed)
1102 1 : fprintf (f, "(killed)");
1103 213 : comma = true;
1104 : }
1105 120 : fprintf (f, "\n");
1106 120 : }
1107 :
1108 : /* Dump the contents of the list to stderr. */
1109 :
1110 : void
1111 0 : ipa_argagg_value_list::debug ()
1112 : {
1113 0 : dump (stderr);
1114 0 : }
1115 :
1116 : /* Return the item describing a constant stored for INDEX at UNIT_OFFSET or
1117 : NULL if there is no such constant. */
1118 :
1119 : const ipa_argagg_value *
1120 29138191 : ipa_argagg_value_list::get_elt (int index, unsigned unit_offset) const
1121 : {
1122 29138191 : ipa_argagg_value key;
1123 29138191 : key.index = index;
1124 29138191 : key.unit_offset = unit_offset;
1125 29138191 : const ipa_argagg_value *res
1126 29138191 : = std::lower_bound (m_elts.begin (), m_elts.end (), key,
1127 6772998 : [] (const ipa_argagg_value &elt,
1128 : const ipa_argagg_value &val)
1129 : {
1130 6772998 : if (elt.index < val.index)
1131 : return true;
1132 5763984 : if (elt.index > val.index)
1133 : return false;
1134 4622009 : if (elt.unit_offset < val.unit_offset)
1135 : return true;
1136 : return false;
1137 : });
1138 :
1139 29138191 : if (res == m_elts.end ()
1140 2908700 : || res->index != index
1141 31497220 : || res->unit_offset != unit_offset)
1142 : res = nullptr;
1143 :
1144 : /* TODO: perhaps remove the check (that the underlying array is indeed
1145 : sorted) if it turns out it can be too slow? */
1146 29138191 : if (!flag_checking)
1147 : return res;
1148 :
1149 : const ipa_argagg_value *slow_res = NULL;
1150 : int prev_index = -1;
1151 : unsigned prev_unit_offset = 0;
1152 45469402 : for (const ipa_argagg_value &av : m_elts)
1153 : {
1154 16331211 : gcc_assert (prev_index < 0
1155 : || prev_index < av.index
1156 : || prev_unit_offset < av.unit_offset);
1157 16331211 : prev_index = av.index;
1158 16331211 : prev_unit_offset = av.unit_offset;
1159 16331211 : if (av.index == index
1160 7494321 : && av.unit_offset == unit_offset)
1161 16331211 : slow_res = &av;
1162 : }
1163 29138191 : gcc_assert (res == slow_res);
1164 :
1165 : return res;
1166 : }
1167 :
1168 : /* Return the first item describing a constant stored for parameter with INDEX,
1169 : regardless of offset or reference, or NULL if there is no such constant. */
1170 :
1171 : const ipa_argagg_value *
1172 229354 : ipa_argagg_value_list::get_elt_for_index (int index) const
1173 : {
1174 229354 : const ipa_argagg_value *res
1175 229354 : = std::lower_bound (m_elts.begin (), m_elts.end (), index,
1176 18426 : [] (const ipa_argagg_value &elt, unsigned idx)
1177 : {
1178 18426 : return elt.index < idx;
1179 : });
1180 229354 : if (res == m_elts.end ()
1181 229354 : || res->index != index)
1182 : res = nullptr;
1183 229354 : return res;
1184 : }
1185 :
1186 : /* Return the aggregate constant stored for INDEX at UNIT_OFFSET, not
1187 : performing any check of whether value is passed by reference, or NULL_TREE
1188 : if there is no such constant. */
1189 :
1190 : tree
1191 36156 : ipa_argagg_value_list::get_value (int index, unsigned unit_offset) const
1192 : {
1193 36156 : const ipa_argagg_value *av = get_elt (index, unit_offset);
1194 36156 : return av ? av->value : NULL_TREE;
1195 : }
1196 :
1197 : /* Return the aggregate constant stored for INDEX at UNIT_OFFSET, if it is
1198 : passed by reference or not according to BY_REF, or NULL_TREE if there is
1199 : no such constant. */
1200 :
1201 : tree
1202 29092312 : ipa_argagg_value_list::get_value (int index, unsigned unit_offset,
1203 : bool by_ref) const
1204 : {
1205 29092312 : const ipa_argagg_value *av = get_elt (index, unit_offset);
1206 29092312 : if (av && av->by_ref == by_ref)
1207 1869199 : return av->value;
1208 : return NULL_TREE;
1209 : }
1210 :
1211 : /* Return true if all elements present in OTHER are also present in this
1212 : list. */
1213 :
1214 : bool
1215 45 : ipa_argagg_value_list::superset_of_p (const ipa_argagg_value_list &other) const
1216 : {
1217 45 : unsigned j = 0;
1218 200 : for (unsigned i = 0; i < other.m_elts.size (); i++)
1219 : {
1220 172 : unsigned other_index = other.m_elts[i].index;
1221 172 : unsigned other_offset = other.m_elts[i].unit_offset;
1222 :
1223 172 : while (j < m_elts.size ()
1224 327 : && (m_elts[j].index < other_index
1225 311 : || (m_elts[j].index == other_index
1226 311 : && m_elts[j].unit_offset < other_offset)))
1227 155 : j++;
1228 :
1229 172 : if (j >= m_elts.size ()
1230 159 : || m_elts[j].index != other_index
1231 159 : || m_elts[j].unit_offset != other_offset
1232 159 : || m_elts[j].by_ref != other.m_elts[i].by_ref
1233 159 : || !m_elts[j].value
1234 331 : || !values_equal_for_ipcp_p (m_elts[j].value, other.m_elts[i].value))
1235 17 : return false;
1236 : }
1237 : return true;
1238 : }
1239 :
1240 : /* Push all items in this list that describe parameter SRC_INDEX into RES as
1241 : ones describing DST_INDEX while subtracting UNIT_DELTA from their unit
1242 : offsets but skip those which would end up with a negative offset. */
1243 :
1244 : void
1245 3145 : ipa_argagg_value_list::push_adjusted_values (unsigned src_index,
1246 : unsigned dest_index,
1247 : unsigned unit_delta,
1248 : vec<ipa_argagg_value> *res) const
1249 : {
1250 3145 : const ipa_argagg_value *av = get_elt_for_index (src_index);
1251 3145 : if (!av)
1252 : return;
1253 : unsigned prev_unit_offset = 0;
1254 : bool first = true;
1255 11981 : for (; av < m_elts.end (); ++av)
1256 : {
1257 9558 : if (av->index > src_index)
1258 : return;
1259 8948 : if (av->index == src_index
1260 8948 : && (av->unit_offset >= unit_delta)
1261 8804 : && av->value)
1262 : {
1263 8804 : ipa_argagg_value new_av;
1264 8804 : gcc_checking_assert (av->value);
1265 8804 : new_av.value = av->value;
1266 8804 : new_av.unit_offset = av->unit_offset - unit_delta;
1267 8804 : new_av.index = dest_index;
1268 8804 : new_av.by_ref = av->by_ref;
1269 8804 : gcc_assert (!av->killed);
1270 8804 : new_av.killed = false;
1271 :
1272 : /* Quick check that the offsets we push are indeed increasing. */
1273 8804 : gcc_assert (first
1274 : || new_av.unit_offset > prev_unit_offset);
1275 8804 : prev_unit_offset = new_av.unit_offset;
1276 8804 : first = false;
1277 :
1278 8804 : res->safe_push (new_av);
1279 : }
1280 : }
1281 : }
1282 :
1283 : /* Push to RES information about single lattices describing aggregate values in
1284 : PLATS as those describing parameter DEST_INDEX and the original offset minus
1285 : UNIT_DELTA. Return true if any item has been pushed to RES. */
1286 :
1287 : static bool
1288 4554707 : push_agg_values_from_plats (ipcp_param_lattices *plats, int dest_index,
1289 : unsigned unit_delta,
1290 : vec<ipa_argagg_value> *res)
1291 : {
1292 4554707 : if (plats->aggs_contain_variable)
1293 : return false;
1294 :
1295 3867357 : bool pushed_sth = false;
1296 3867357 : bool first = true;
1297 3867357 : unsigned prev_unit_offset = 0;
1298 3932873 : for (struct ipcp_agg_lattice *aglat = plats->aggs; aglat; aglat = aglat->next)
1299 129824 : if (aglat->is_single_const ()
1300 40321 : && (aglat->offset / BITS_PER_UNIT - unit_delta) >= 0)
1301 : {
1302 40321 : ipa_argagg_value iav;
1303 40321 : iav.value = aglat->values->value;
1304 40321 : iav.unit_offset = aglat->offset / BITS_PER_UNIT - unit_delta;
1305 40321 : iav.index = dest_index;
1306 40321 : iav.by_ref = plats->aggs_by_ref;
1307 40321 : iav.killed = false;
1308 :
1309 40321 : gcc_assert (first
1310 : || iav.unit_offset > prev_unit_offset);
1311 40321 : prev_unit_offset = iav.unit_offset;
1312 40321 : first = false;
1313 :
1314 40321 : pushed_sth = true;
1315 40321 : res->safe_push (iav);
1316 : }
1317 : return pushed_sth;
1318 : }
1319 :
1320 : /* Turn all values in LIST that are not present in OTHER into NULL_TREEs.
1321 : Return the number of remaining valid entries. */
1322 :
1323 : static unsigned
1324 55087 : intersect_argaggs_with (vec<ipa_argagg_value> &elts,
1325 : const vec<ipa_argagg_value> &other)
1326 : {
1327 55087 : unsigned valid_entries = 0;
1328 55087 : unsigned j = 0;
1329 400705 : for (unsigned i = 0; i < elts.length (); i++)
1330 : {
1331 345618 : if (!elts[i].value)
1332 49282 : continue;
1333 :
1334 296336 : unsigned this_index = elts[i].index;
1335 296336 : unsigned this_offset = elts[i].unit_offset;
1336 :
1337 296336 : while (j < other.length ()
1338 1106944 : && (other[j].index < this_index
1339 521373 : || (other[j].index == this_index
1340 518009 : && other[j].unit_offset < this_offset)))
1341 261331 : j++;
1342 :
1343 296336 : if (j >= other.length ())
1344 : {
1345 8390 : elts[i].value = NULL_TREE;
1346 8390 : continue;
1347 : }
1348 :
1349 287946 : if (other[j].index == this_index
1350 284582 : && other[j].unit_offset == this_offset
1351 279695 : && other[j].by_ref == elts[i].by_ref
1352 279695 : && other[j].value
1353 567641 : && values_equal_for_ipcp_p (other[j].value, elts[i].value))
1354 261709 : valid_entries++;
1355 : else
1356 26237 : elts[i].value = NULL_TREE;
1357 : }
1358 55087 : return valid_entries;
1359 : }
1360 :
1361 : /* Mark bot aggregate and scalar lattices as containing an unknown variable,
1362 : return true is any of them has not been marked as such so far. If if
1363 : MAKE_SIMPLE_RECIPIENTS is true, set the lattices that can only hold one
1364 : value to being recipients only, otherwise also set them to bottom. */
1365 :
1366 : static inline bool
1367 169993 : set_all_contains_variable (class ipcp_param_lattices *plats,
1368 : bool make_simple_recipients = false)
1369 : {
1370 169993 : bool ret;
1371 169993 : ret = plats->itself.set_contains_variable ();
1372 169993 : ret |= plats->ctxlat.set_contains_variable ();
1373 169993 : ret |= set_agg_lats_contain_variable (plats);
1374 169993 : if (make_simple_recipients)
1375 : {
1376 29238 : ret |= plats->bits_lattice.set_recipient_only ();
1377 29238 : ret |= plats->m_value_range.set_recipient_only ();
1378 : }
1379 : else
1380 : {
1381 140755 : ret |= plats->bits_lattice.set_to_bottom ();
1382 140755 : ret |= plats->m_value_range.set_to_bottom ();
1383 : }
1384 169993 : return ret;
1385 : }
1386 :
1387 : /* Worker of call_for_symbol_thunks_and_aliases, increment the integer DATA
1388 : points to by the number of callers to NODE. */
1389 :
1390 : static bool
1391 99460 : count_callers (cgraph_node *node, void *data)
1392 : {
1393 99460 : int *caller_count = (int *) data;
1394 :
1395 405047 : for (cgraph_edge *cs = node->callers; cs; cs = cs->next_caller)
1396 : /* Local thunks can be handled transparently, but if the thunk cannot
1397 : be optimized out, count it as a real use. */
1398 305587 : if (!cs->caller->thunk || !cs->caller->local)
1399 305587 : ++*caller_count;
1400 99460 : return false;
1401 : }
1402 :
1403 : /* Worker of call_for_symbol_thunks_and_aliases, it is supposed to be called on
1404 : the one caller of some other node. Set the caller's corresponding flag. */
1405 :
1406 : static bool
1407 56684 : set_single_call_flag (cgraph_node *node, void *)
1408 : {
1409 56684 : cgraph_edge *cs = node->callers;
1410 : /* Local thunks can be handled transparently, skip them. */
1411 56684 : while (cs && cs->caller->thunk && cs->caller->local)
1412 0 : cs = cs->next_caller;
1413 56684 : if (cs)
1414 56114 : if (ipa_node_params* info = ipa_node_params_sum->get (cs->caller))
1415 : {
1416 56113 : info->node_calling_single_call = true;
1417 56113 : return true;
1418 : }
1419 : return false;
1420 : }
1421 :
1422 : /* Initialize ipcp_lattices. */
1423 :
1424 : static void
1425 1269638 : initialize_node_lattices (struct cgraph_node *node)
1426 : {
1427 1269638 : ipa_node_params *info = ipa_node_params_sum->get (node);
1428 1269638 : struct cgraph_edge *ie;
1429 1269638 : bool disable = false, variable = false;
1430 1269638 : int i;
1431 :
1432 1269638 : gcc_checking_assert (node->has_gimple_body_p ());
1433 :
1434 1269638 : if (!ipa_get_param_count (info))
1435 : disable = true;
1436 1040531 : else if (node->local)
1437 : {
1438 88513 : int caller_count = 0;
1439 88513 : node->call_for_symbol_thunks_and_aliases (count_callers, &caller_count,
1440 : true);
1441 88513 : if (caller_count == 1)
1442 56114 : node->call_for_symbol_thunks_and_aliases (set_single_call_flag,
1443 : NULL, true);
1444 32399 : else if (caller_count == 0)
1445 : {
1446 1 : gcc_checking_assert (!opt_for_fn (node->decl, flag_toplevel_reorder));
1447 : variable = true;
1448 : }
1449 : }
1450 : else
1451 : {
1452 : /* When cloning is allowed, we can assume that externally visible
1453 : functions are not called. We will compensate this by cloning
1454 : later. */
1455 952018 : if (ipcp_versionable_function_p (node)
1456 952018 : && ipcp_cloning_candidate_p (node))
1457 : variable = true;
1458 : else
1459 : disable = true;
1460 : }
1461 :
1462 728 : if (dump_file && (dump_flags & TDF_DETAILS)
1463 1269805 : && !node->alias && !node->thunk)
1464 : {
1465 167 : fprintf (dump_file, "Initializing lattices of %s\n",
1466 : node->dump_name ());
1467 167 : if (disable || variable)
1468 132 : fprintf (dump_file, " Marking all lattices as %s\n",
1469 : disable ? "BOTTOM" : "VARIABLE");
1470 : }
1471 :
1472 1269638 : auto_vec<bool, 16> surviving_params;
1473 1269638 : bool pre_modified = false;
1474 :
1475 1269638 : clone_info *cinfo = clone_info::get (node);
1476 :
1477 1269638 : if (!disable && cinfo && cinfo->param_adjustments)
1478 : {
1479 : /* At the moment all IPA optimizations should use the number of
1480 : parameters of the prevailing decl as the m_always_copy_start.
1481 : Handling any other value would complicate the code below, so for the
1482 : time bing let's only assert it is so. */
1483 0 : gcc_assert ((cinfo->param_adjustments->m_always_copy_start
1484 : == ipa_get_param_count (info))
1485 : || cinfo->param_adjustments->m_always_copy_start < 0);
1486 :
1487 0 : pre_modified = true;
1488 0 : cinfo->param_adjustments->get_surviving_params (&surviving_params);
1489 :
1490 0 : if (dump_file && (dump_flags & TDF_DETAILS)
1491 0 : && !node->alias && !node->thunk)
1492 : {
1493 : bool first = true;
1494 0 : for (int j = 0; j < ipa_get_param_count (info); j++)
1495 : {
1496 0 : if (j < (int) surviving_params.length ()
1497 0 : && surviving_params[j])
1498 0 : continue;
1499 0 : if (first)
1500 : {
1501 0 : fprintf (dump_file,
1502 : " The following parameters are dead on arrival:");
1503 0 : first = false;
1504 : }
1505 0 : fprintf (dump_file, " %u", j);
1506 : }
1507 0 : if (!first)
1508 0 : fprintf (dump_file, "\n");
1509 : }
1510 : }
1511 :
1512 6977753 : for (i = 0; i < ipa_get_param_count (info); i++)
1513 : {
1514 2333792 : ipcp_param_lattices *plats = ipa_get_parm_lattices (info, i);
1515 2333792 : tree type = ipa_get_type (info, i);
1516 2333792 : if (disable
1517 223023 : || !ipa_get_type (info, i)
1518 2556815 : || (pre_modified && (surviving_params.length () <= (unsigned) i
1519 0 : || !surviving_params[i])))
1520 : {
1521 2110769 : plats->itself.set_to_bottom ();
1522 2110769 : plats->ctxlat.set_to_bottom ();
1523 2110769 : set_agg_lats_to_bottom (plats);
1524 2110769 : plats->bits_lattice.set_to_bottom ();
1525 2110769 : plats->m_value_range.init (type);
1526 2110769 : plats->m_value_range.set_to_bottom ();
1527 : }
1528 : else
1529 : {
1530 223023 : plats->m_value_range.init (type);
1531 223023 : if (variable)
1532 29238 : set_all_contains_variable (plats, true);
1533 : }
1534 : }
1535 :
1536 1403516 : for (ie = node->indirect_calls; ie; ie = ie->next_callee)
1537 133878 : if (ie->indirect_info->param_index >= 0
1538 142973 : && is_a <cgraph_polymorphic_indirect_info *> (ie->indirect_info))
1539 9095 : ipa_get_parm_lattices (info,
1540 9095 : ie->indirect_info->param_index)->virt_call = 1;
1541 1269638 : }
1542 :
1543 : /* Return VALUE if it is NULL_TREE or if it can be directly safely IPA-CP
1544 : propagated to a parameter of type PARAM_TYPE, or return a fold-converted
1545 : VALUE to PARAM_TYPE if that is possible. Return NULL_TREE otherwise. */
1546 :
1547 : tree
1548 5369012 : ipacp_value_safe_for_type (tree param_type, tree value)
1549 : {
1550 5369012 : if (!value)
1551 : return NULL_TREE;
1552 5368678 : tree val_type = TREE_TYPE (value);
1553 5368678 : if (param_type == val_type
1554 5368678 : || useless_type_conversion_p (param_type, val_type))
1555 5365310 : return value;
1556 3368 : if (fold_convertible_p (param_type, value))
1557 3163 : return fold_convert (param_type, value);
1558 : else
1559 : return NULL_TREE;
1560 : }
1561 :
1562 : /* Return the result of a (possibly arithmetic) operation determined by OPCODE
1563 : on the constant value INPUT. OPERAND is 2nd operand for binary operation
1564 : and is required for binary operations. RES_TYPE, required when opcode is
1565 : not NOP_EXPR, is the type in which any operation is to be performed. Return
1566 : NULL_TREE if that cannot be determined or be considered an interprocedural
1567 : invariant. */
1568 :
1569 : static tree
1570 68752 : ipa_get_jf_arith_result (enum tree_code opcode, tree input, tree operand,
1571 : tree res_type)
1572 : {
1573 68752 : tree res;
1574 :
1575 68752 : if (opcode == NOP_EXPR)
1576 : return input;
1577 6533 : if (!is_gimple_ip_invariant (input))
1578 : return NULL_TREE;
1579 :
1580 6533 : if (opcode == ASSERT_EXPR)
1581 : {
1582 3542 : if (values_equal_for_ipcp_p (input, operand))
1583 : return input;
1584 : else
1585 : return NULL_TREE;
1586 : }
1587 :
1588 2991 : if (TREE_CODE_CLASS (opcode) == tcc_unary)
1589 91 : res = fold_unary (opcode, res_type, input);
1590 : else
1591 2900 : res = fold_binary (opcode, res_type, input, operand);
1592 :
1593 2991 : if (res && !is_gimple_ip_invariant (res))
1594 : return NULL_TREE;
1595 :
1596 : return res;
1597 : }
1598 :
1599 : /* Return the result of an ancestor jump function JFUNC on the constant value
1600 : INPUT. Return NULL_TREE if that cannot be determined. */
1601 :
1602 : static tree
1603 1270 : ipa_get_jf_ancestor_result (struct ipa_jump_func *jfunc, tree input)
1604 : {
1605 1270 : gcc_checking_assert (TREE_CODE (input) != TREE_BINFO);
1606 1270 : if (TREE_CODE (input) == ADDR_EXPR)
1607 : {
1608 1188 : gcc_checking_assert (is_gimple_ip_invariant_address (input));
1609 1188 : poly_int64 off = ipa_get_jf_ancestor_offset (jfunc);
1610 1188 : if (known_eq (off, 0))
1611 : return input;
1612 1066 : poly_int64 byte_offset = exact_div (off, BITS_PER_UNIT);
1613 2132 : return build1 (ADDR_EXPR, TREE_TYPE (input),
1614 1066 : fold_build2 (MEM_REF, TREE_TYPE (TREE_TYPE (input)), input,
1615 1066 : build_int_cst (ptr_type_node, byte_offset)));
1616 : }
1617 82 : else if (ipa_get_jf_ancestor_keep_null (jfunc)
1618 82 : && zerop (input))
1619 : return input;
1620 : else
1621 78 : return NULL_TREE;
1622 : }
1623 :
1624 : /* Determine whether JFUNC evaluates to a single known constant value and if
1625 : so, return it. Otherwise return NULL. INFO describes the caller node or
1626 : the one it is inlined to, so that pass-through jump functions can be
1627 : evaluated. PARM_TYPE is the type of the parameter to which the result is
1628 : passed. */
1629 :
1630 : tree
1631 18134753 : ipa_value_from_jfunc (class ipa_node_params *info, struct ipa_jump_func *jfunc,
1632 : tree parm_type)
1633 : {
1634 18134753 : if (!parm_type)
1635 : return NULL_TREE;
1636 17897583 : if (jfunc->type == IPA_JF_CONST)
1637 4768265 : return ipacp_value_safe_for_type (parm_type, ipa_get_jf_constant (jfunc));
1638 13129318 : else if (jfunc->type == IPA_JF_PASS_THROUGH
1639 10316574 : || jfunc->type == IPA_JF_ANCESTOR)
1640 : {
1641 3612427 : tree input;
1642 3612427 : int idx;
1643 :
1644 3612427 : if (jfunc->type == IPA_JF_PASS_THROUGH)
1645 2812744 : idx = ipa_get_jf_pass_through_formal_id (jfunc);
1646 : else
1647 799683 : idx = ipa_get_jf_ancestor_formal_id (jfunc);
1648 :
1649 3612427 : if (info->ipcp_orig_node)
1650 56510 : input = info->known_csts[idx];
1651 : else
1652 : {
1653 3555917 : ipcp_lattice<tree> *lat;
1654 :
1655 6409252 : if (info->lattices.is_empty ()
1656 2853335 : || idx >= ipa_get_param_count (info))
1657 : return NULL_TREE;
1658 2853335 : lat = ipa_get_scalar_lat (info, idx);
1659 2853335 : if (!lat->is_single_const ())
1660 : return NULL_TREE;
1661 147 : input = lat->values->value;
1662 : }
1663 :
1664 56657 : if (!input)
1665 : return NULL_TREE;
1666 :
1667 19316 : if (jfunc->type == IPA_JF_PASS_THROUGH)
1668 : {
1669 18357 : enum tree_code opcode = ipa_get_jf_pass_through_operation (jfunc);
1670 18357 : tree op2 = ipa_get_jf_pass_through_operand (jfunc);
1671 18357 : tree op_type
1672 18357 : = (opcode == NOP_EXPR) ? NULL_TREE
1673 918 : : ipa_get_jf_pass_through_op_type (jfunc);
1674 18357 : tree cstval = ipa_get_jf_arith_result (opcode, input, op2, op_type);
1675 18357 : return ipacp_value_safe_for_type (parm_type, cstval);
1676 : }
1677 : else
1678 959 : return ipacp_value_safe_for_type (parm_type,
1679 : ipa_get_jf_ancestor_result (jfunc,
1680 959 : input));
1681 : }
1682 : else
1683 : return NULL_TREE;
1684 : }
1685 :
1686 : /* Determine whether JFUNC evaluates to single known polymorphic context, given
1687 : that INFO describes the caller node or the one it is inlined to, CS is the
1688 : call graph edge corresponding to JFUNC and CSIDX index of the described
1689 : parameter. */
1690 :
1691 : ipa_polymorphic_call_context
1692 878342 : ipa_context_from_jfunc (ipa_node_params *info, cgraph_edge *cs, int csidx,
1693 : ipa_jump_func *jfunc)
1694 : {
1695 878342 : ipa_edge_args *args = ipa_edge_args_sum->get (cs);
1696 878342 : ipa_polymorphic_call_context ctx;
1697 878342 : ipa_polymorphic_call_context *edge_ctx
1698 878342 : = cs ? ipa_get_ith_polymorhic_call_context (args, csidx) : NULL;
1699 :
1700 351756 : if (edge_ctx && !edge_ctx->useless_p ())
1701 346508 : ctx = *edge_ctx;
1702 :
1703 878342 : if (jfunc->type == IPA_JF_PASS_THROUGH
1704 786553 : || jfunc->type == IPA_JF_ANCESTOR)
1705 : {
1706 99848 : ipa_polymorphic_call_context srcctx;
1707 99848 : int srcidx;
1708 99848 : bool type_preserved = true;
1709 99848 : if (jfunc->type == IPA_JF_PASS_THROUGH)
1710 : {
1711 91789 : if (ipa_get_jf_pass_through_operation (jfunc) != NOP_EXPR)
1712 1842 : return ctx;
1713 89947 : type_preserved = ipa_get_jf_pass_through_type_preserved (jfunc);
1714 89947 : srcidx = ipa_get_jf_pass_through_formal_id (jfunc);
1715 : }
1716 : else
1717 : {
1718 8059 : type_preserved = ipa_get_jf_ancestor_type_preserved (jfunc);
1719 8059 : srcidx = ipa_get_jf_ancestor_formal_id (jfunc);
1720 : }
1721 98006 : if (info->ipcp_orig_node)
1722 : {
1723 11462 : if (info->known_contexts.exists ())
1724 1358 : srcctx = info->known_contexts[srcidx];
1725 : }
1726 : else
1727 : {
1728 171453 : if (info->lattices.is_empty ()
1729 84909 : || srcidx >= ipa_get_param_count (info))
1730 1635 : return ctx;
1731 84909 : ipcp_lattice<ipa_polymorphic_call_context> *lat;
1732 84909 : lat = ipa_get_poly_ctx_lat (info, srcidx);
1733 84909 : if (!lat->is_single_const ())
1734 80964 : return ctx;
1735 3945 : srcctx = lat->values->value;
1736 : }
1737 15407 : if (srcctx.useless_p ())
1738 10550 : return ctx;
1739 4857 : if (jfunc->type == IPA_JF_ANCESTOR)
1740 253 : srcctx.offset_by (ipa_get_jf_ancestor_offset (jfunc));
1741 4857 : if (!type_preserved)
1742 2889 : srcctx.possible_dynamic_type_change (cs->in_polymorphic_cdtor);
1743 4857 : srcctx.combine_with (ctx);
1744 4857 : return srcctx;
1745 : }
1746 :
1747 778494 : return ctx;
1748 : }
1749 :
1750 : /* Emulate effects of unary OPERATION and/or conversion from SRC_TYPE to
1751 : DST_TYPE on value range in SRC_VR and store it to DST_VR. Return true if
1752 : the result is a range that is not VARYING nor UNDEFINED. */
1753 :
1754 : bool
1755 9514566 : ipa_vr_operation_and_type_effects (vrange &dst_vr,
1756 : const vrange &src_vr,
1757 : enum tree_code operation,
1758 : tree dst_type, tree src_type)
1759 : {
1760 17871324 : if (!ipa_vr_supported_type_p (dst_type)
1761 0 : || !ipa_vr_supported_type_p (src_type))
1762 : return false;
1763 :
1764 9514566 : range_op_handler handler (operation);
1765 9514566 : if (!handler)
1766 : return false;
1767 :
1768 9514566 : value_range varying (dst_type);
1769 9514566 : varying.set_varying (dst_type);
1770 :
1771 9514566 : return (handler.operand_check_p (dst_type, src_type, dst_type)
1772 9514566 : && handler.fold_range (dst_vr, dst_type, src_vr, varying)
1773 9514564 : && !dst_vr.varying_p ()
1774 19029070 : && !dst_vr.undefined_p ());
1775 9514566 : }
1776 :
1777 : /* Same as above, but the SRC_VR argument is an IPA_VR which must
1778 : first be extracted onto a vrange. */
1779 :
1780 : bool
1781 9422209 : ipa_vr_operation_and_type_effects (vrange &dst_vr,
1782 : const ipa_vr &src_vr,
1783 : enum tree_code operation,
1784 : tree dst_type, tree src_type)
1785 : {
1786 9422209 : value_range tmp;
1787 9422209 : src_vr.get_vrange (tmp);
1788 9422209 : return ipa_vr_operation_and_type_effects (dst_vr, tmp, operation,
1789 9422209 : dst_type, src_type);
1790 9422209 : }
1791 :
1792 : /* Given a PASS_THROUGH jump function JFUNC that takes as its source SRC_VR of
1793 : SRC_TYPE and the result needs to be DST_TYPE, if any value range information
1794 : can be deduced at all, intersect VR with it. CONTEXT_NODE is the call graph
1795 : node representing the function for which optimization flags should be
1796 : evaluated. */
1797 :
1798 : static void
1799 92757 : ipa_vr_intersect_with_arith_jfunc (vrange &vr,
1800 : ipa_jump_func *jfunc,
1801 : cgraph_node *context_node,
1802 : const value_range &src_vr,
1803 : tree src_type,
1804 : tree dst_type)
1805 : {
1806 92757 : if (src_vr.undefined_p () || src_vr.varying_p ())
1807 91481 : return;
1808 :
1809 92298 : enum tree_code operation = ipa_get_jf_pass_through_operation (jfunc);
1810 92298 : if (TREE_CODE_CLASS (operation) == tcc_unary)
1811 : {
1812 91022 : value_range op_res;
1813 91022 : const value_range *inter_vr;
1814 91022 : if (operation != NOP_EXPR)
1815 : {
1816 89 : tree operation_type = ipa_get_jf_pass_through_op_type (jfunc);
1817 89 : op_res.set_varying (operation_type);
1818 89 : if (!ipa_vr_operation_and_type_effects (op_res, src_vr, operation,
1819 : operation_type, src_type))
1820 : return;
1821 89 : if (src_type == dst_type)
1822 : {
1823 30 : vr.intersect (op_res);
1824 30 : return;
1825 : }
1826 : inter_vr = &op_res;
1827 : src_type = operation_type;
1828 : }
1829 : else
1830 : inter_vr = &src_vr;
1831 :
1832 90992 : value_range tmp_res (dst_type);
1833 90992 : if (ipa_vr_operation_and_type_effects (tmp_res, *inter_vr, NOP_EXPR,
1834 : dst_type, src_type))
1835 90992 : vr.intersect (tmp_res);
1836 90992 : return;
1837 91022 : }
1838 :
1839 1276 : tree operand = ipa_get_jf_pass_through_operand (jfunc);
1840 1276 : range_op_handler handler (operation);
1841 1276 : if (!handler)
1842 : return;
1843 1276 : value_range op_vr (TREE_TYPE (operand));
1844 1276 : ipa_get_range_from_ip_invariant (op_vr, operand, context_node);
1845 :
1846 1276 : tree operation_type = ipa_get_jf_pass_through_op_type (jfunc);
1847 1276 : value_range op_res (operation_type);
1848 1716 : if (!ipa_vr_supported_type_p (operation_type)
1849 1276 : || !handler.operand_check_p (operation_type, src_type, op_vr.type ())
1850 1276 : || !handler.fold_range (op_res, operation_type, src_vr, op_vr))
1851 0 : return;
1852 :
1853 1276 : value_range tmp_res (dst_type);
1854 1276 : if (ipa_vr_operation_and_type_effects (tmp_res, op_res, NOP_EXPR, dst_type,
1855 : operation_type))
1856 1228 : vr.intersect (tmp_res);
1857 1276 : }
1858 :
1859 : /* Determine range of JFUNC given that INFO describes the caller node or
1860 : the one it is inlined to, CS is the call graph edge corresponding to JFUNC
1861 : and PARM_TYPE of the parameter. */
1862 :
1863 : void
1864 12105647 : ipa_value_range_from_jfunc (vrange &vr,
1865 : ipa_node_params *info, cgraph_edge *cs,
1866 : ipa_jump_func *jfunc, tree parm_type)
1867 : {
1868 12105647 : vr.set_varying (parm_type);
1869 :
1870 12105647 : if (jfunc->m_vr && jfunc->m_vr->known_p ())
1871 8606280 : ipa_vr_operation_and_type_effects (vr,
1872 : *jfunc->m_vr,
1873 : NOP_EXPR, parm_type,
1874 8606280 : jfunc->m_vr->type ());
1875 12105647 : if (vr.singleton_p ())
1876 : return;
1877 :
1878 12105507 : if (jfunc->type == IPA_JF_PASS_THROUGH)
1879 : {
1880 2245575 : ipcp_transformation *sum
1881 2245575 : = ipcp_get_transformation_summary (cs->caller->inlined_to
1882 : ? cs->caller->inlined_to
1883 : : cs->caller);
1884 2245575 : if (!sum || !sum->m_vr)
1885 2168020 : return;
1886 :
1887 118883 : int idx = ipa_get_jf_pass_through_formal_id (jfunc);
1888 :
1889 118883 : if (!(*sum->m_vr)[idx].known_p ())
1890 : return;
1891 77555 : tree src_type = ipa_get_type (info, idx);
1892 77555 : value_range srcvr;
1893 77555 : (*sum->m_vr)[idx].get_vrange (srcvr);
1894 :
1895 77555 : ipa_vr_intersect_with_arith_jfunc (vr, jfunc, cs->caller, srcvr, src_type,
1896 : parm_type);
1897 77555 : }
1898 : }
1899 :
1900 : /* Determine whether ITEM, jump function for an aggregate part, evaluates to a
1901 : single known constant value and if so, return it. Otherwise return NULL.
1902 : NODE and INFO describes the caller node or the one it is inlined to, and
1903 : its related info. */
1904 :
1905 : tree
1906 3184432 : ipa_agg_value_from_jfunc (ipa_node_params *info, cgraph_node *node,
1907 : const ipa_agg_jf_item *item)
1908 : {
1909 3184432 : tree value = NULL_TREE;
1910 3184432 : int src_idx;
1911 :
1912 3184432 : if (item->offset < 0
1913 3134375 : || item->jftype == IPA_JF_UNKNOWN
1914 2985428 : || item->offset >= (HOST_WIDE_INT) UINT_MAX * BITS_PER_UNIT)
1915 : return NULL_TREE;
1916 :
1917 2985428 : if (item->jftype == IPA_JF_CONST)
1918 2651285 : return item->value.constant;
1919 :
1920 334143 : gcc_checking_assert (item->jftype == IPA_JF_PASS_THROUGH
1921 : || item->jftype == IPA_JF_LOAD_AGG);
1922 :
1923 334143 : src_idx = item->value.pass_through.formal_id;
1924 :
1925 334143 : if (info->ipcp_orig_node)
1926 : {
1927 15573 : if (item->jftype == IPA_JF_PASS_THROUGH)
1928 3772 : value = info->known_csts[src_idx];
1929 11801 : else if (ipcp_transformation *ts = ipcp_get_transformation_summary (node))
1930 : {
1931 11801 : ipa_argagg_value_list avl (ts);
1932 11801 : value = avl.get_value (src_idx,
1933 11801 : item->value.load_agg.offset / BITS_PER_UNIT,
1934 11801 : item->value.load_agg.by_ref);
1935 : }
1936 : }
1937 318570 : else if (!info->lattices.is_empty ())
1938 : {
1939 217999 : class ipcp_param_lattices *src_plats
1940 217999 : = ipa_get_parm_lattices (info, src_idx);
1941 :
1942 217999 : if (item->jftype == IPA_JF_PASS_THROUGH)
1943 : {
1944 131455 : struct ipcp_lattice<tree> *lat = &src_plats->itself;
1945 :
1946 531836 : if (!lat->is_single_const ())
1947 : return NULL_TREE;
1948 :
1949 0 : value = lat->values->value;
1950 : }
1951 86544 : else if (src_plats->aggs
1952 9829 : && !src_plats->aggs_bottom
1953 9829 : && !src_plats->aggs_contain_variable
1954 1499 : && src_plats->aggs_by_ref == item->value.load_agg.by_ref)
1955 : {
1956 : struct ipcp_agg_lattice *aglat;
1957 :
1958 2366 : for (aglat = src_plats->aggs; aglat; aglat = aglat->next)
1959 : {
1960 2366 : if (aglat->offset > item->value.load_agg.offset)
1961 : break;
1962 :
1963 2334 : if (aglat->offset == item->value.load_agg.offset)
1964 : {
1965 1467 : if (aglat->is_single_const ())
1966 7 : value = aglat->values->value;
1967 : break;
1968 : }
1969 : }
1970 : }
1971 : }
1972 :
1973 15612 : if (!value)
1974 192862 : return NULL_TREE;
1975 :
1976 9826 : if (item->jftype == IPA_JF_LOAD_AGG)
1977 : {
1978 7516 : tree load_type = item->value.load_agg.type;
1979 7516 : tree value_type = TREE_TYPE (value);
1980 :
1981 : /* Ensure value type is compatible with load type. */
1982 7516 : if (!useless_type_conversion_p (load_type, value_type))
1983 : return NULL_TREE;
1984 : }
1985 :
1986 19652 : tree cstval = ipa_get_jf_arith_result (item->value.pass_through.operation,
1987 : value,
1988 9826 : item->value.pass_through.operand,
1989 9826 : item->value.pass_through.op_type);
1990 9826 : return ipacp_value_safe_for_type (item->type, cstval);
1991 : }
1992 :
1993 : /* Process all items in AGG_JFUNC relative to caller (or the node the original
1994 : caller is inlined to) NODE which described by INFO and push the results to
1995 : RES as describing values passed in parameter DST_INDEX. */
1996 :
1997 : void
1998 14591007 : ipa_push_agg_values_from_jfunc (ipa_node_params *info, cgraph_node *node,
1999 : ipa_agg_jump_function *agg_jfunc,
2000 : unsigned dst_index,
2001 : vec<ipa_argagg_value> *res)
2002 : {
2003 14591007 : unsigned prev_unit_offset = 0;
2004 14591007 : bool first = true;
2005 :
2006 19063693 : for (const ipa_agg_jf_item &item : agg_jfunc->items)
2007 : {
2008 2209476 : tree value = ipa_agg_value_from_jfunc (info, node, &item);
2009 2209476 : if (!value)
2010 500679 : continue;
2011 :
2012 1708797 : ipa_argagg_value iav;
2013 1708797 : iav.value = value;
2014 1708797 : iav.unit_offset = item.offset / BITS_PER_UNIT;
2015 1708797 : iav.index = dst_index;
2016 1708797 : iav.by_ref = agg_jfunc->by_ref;
2017 1708797 : iav.killed = 0;
2018 :
2019 1708797 : gcc_assert (first
2020 : || iav.unit_offset > prev_unit_offset);
2021 1708797 : prev_unit_offset = iav.unit_offset;
2022 1708797 : first = false;
2023 :
2024 1708797 : res->safe_push (iav);
2025 : }
2026 14591007 : }
2027 :
2028 : /* If checking is enabled, verify that no lattice is in the TOP state, i.e. not
2029 : bottom, not containing a variable component and without any known value at
2030 : the same time. */
2031 :
2032 : DEBUG_FUNCTION void
2033 129019 : ipcp_verify_propagated_values (void)
2034 : {
2035 129019 : struct cgraph_node *node;
2036 :
2037 1407482 : FOR_EACH_FUNCTION_WITH_GIMPLE_BODY (node)
2038 : {
2039 1278463 : ipa_node_params *info = ipa_node_params_sum->get (node);
2040 1278463 : if (!opt_for_fn (node->decl, flag_ipa_cp)
2041 1278463 : || !opt_for_fn (node->decl, optimize))
2042 8842 : continue;
2043 1269621 : int i, count = ipa_get_param_count (info);
2044 :
2045 3603399 : for (i = 0; i < count; i++)
2046 : {
2047 2333778 : ipcp_lattice<tree> *lat = ipa_get_scalar_lat (info, i);
2048 :
2049 2333778 : if (!lat->bottom
2050 222017 : && !lat->contains_variable
2051 31712 : && lat->values_count == 0)
2052 : {
2053 0 : if (dump_file)
2054 : {
2055 0 : symtab->dump (dump_file);
2056 0 : fprintf (dump_file, "\nIPA lattices after constant "
2057 : "propagation, before gcc_unreachable:\n");
2058 0 : print_all_lattices (dump_file, true, false);
2059 : }
2060 :
2061 0 : gcc_unreachable ();
2062 : }
2063 : }
2064 : }
2065 129019 : }
2066 :
2067 : /* Return true iff X and Y should be considered equal contexts by IPA-CP. */
2068 :
2069 : static bool
2070 2685 : values_equal_for_ipcp_p (ipa_polymorphic_call_context x,
2071 : ipa_polymorphic_call_context y)
2072 : {
2073 2173 : return x.equal_to (y);
2074 : }
2075 :
2076 :
2077 : /* Add a new value source to the value represented by THIS, marking that a
2078 : value comes from edge CS and (if the underlying jump function is a
2079 : pass-through or an ancestor one) from a caller value SRC_VAL of a caller
2080 : parameter described by SRC_INDEX. OFFSET is negative if the source was the
2081 : scalar value of the parameter itself or the offset within an aggregate. */
2082 :
2083 : template <typename valtype>
2084 : void
2085 486227 : ipcp_value<valtype>::add_source (cgraph_edge *cs, ipcp_value *src_val,
2086 : int src_idx, HOST_WIDE_INT offset)
2087 : {
2088 : ipcp_value_source<valtype> *src;
2089 :
2090 486227 : src = new (ipcp_sources_pool.allocate ()) ipcp_value_source<valtype>;
2091 486227 : src->offset = offset;
2092 486227 : src->cs = cs;
2093 486227 : src->val = src_val;
2094 486227 : src->index = src_idx;
2095 :
2096 486227 : src->next = sources;
2097 486227 : sources = src;
2098 486227 : }
2099 :
2100 : /* Allocate a new ipcp_value holding a tree constant, initialize its value to
2101 : SOURCE and clear all other fields. */
2102 :
2103 : static ipcp_value<tree> *
2104 139155 : allocate_and_init_ipcp_value (tree cst, unsigned same_lat_gen_level)
2105 : {
2106 139155 : ipcp_value<tree> *val;
2107 :
2108 139155 : val = new (ipcp_cst_values_pool.allocate ()) ipcp_value<tree>();
2109 139155 : val->value = cst;
2110 139155 : val->self_recursion_generated_level = same_lat_gen_level;
2111 139155 : return val;
2112 : }
2113 :
2114 : /* Allocate a new ipcp_value holding a polymorphic context, initialize its
2115 : value to SOURCE and clear all other fields. */
2116 :
2117 : static ipcp_value<ipa_polymorphic_call_context> *
2118 7743 : allocate_and_init_ipcp_value (ipa_polymorphic_call_context ctx,
2119 : unsigned same_lat_gen_level)
2120 : {
2121 7743 : ipcp_value<ipa_polymorphic_call_context> *val;
2122 :
2123 7743 : val = new (ipcp_poly_ctx_values_pool.allocate ())
2124 7743 : ipcp_value<ipa_polymorphic_call_context>();
2125 7743 : val->value = ctx;
2126 7743 : val->self_recursion_generated_level = same_lat_gen_level;
2127 7743 : return val;
2128 : }
2129 :
2130 : /* Try to add NEWVAL to LAT, potentially creating a new ipcp_value for it. CS,
2131 : SRC_VAL SRC_INDEX and OFFSET are meant for add_source and have the same
2132 : meaning. OFFSET -1 means the source is scalar and not a part of an
2133 : aggregate. If non-NULL, VAL_P records address of existing or newly added
2134 : ipcp_value.
2135 :
2136 : If the value is generated for a self-recursive call as a result of an
2137 : arithmetic pass-through jump-function acting on a value in the same lattice,
2138 : SAME_LAT_GEN_LEVEL must be the length of such chain, otherwise it must be
2139 : zero. If it is non-zero, PARAM_IPA_CP_VALUE_LIST_SIZE limit is ignored. */
2140 :
2141 : template <typename valtype>
2142 : bool
2143 498606 : ipcp_lattice<valtype>::add_value (valtype newval, cgraph_edge *cs,
2144 : ipcp_value<valtype> *src_val,
2145 : int src_idx, HOST_WIDE_INT offset,
2146 : ipcp_value<valtype> **val_p,
2147 : unsigned same_lat_gen_level)
2148 : {
2149 498606 : ipcp_value<valtype> *val, *last_val = NULL;
2150 :
2151 498606 : if (val_p)
2152 1257 : *val_p = NULL;
2153 :
2154 498606 : if (bottom)
2155 : return false;
2156 :
2157 967830 : for (val = values; val; last_val = val, val = val->next)
2158 819566 : if (values_equal_for_ipcp_p (val->value, newval))
2159 : {
2160 346984 : if (val_p)
2161 416 : *val_p = val;
2162 :
2163 346984 : if (val->self_recursion_generated_level < same_lat_gen_level)
2164 179 : val->self_recursion_generated_level = same_lat_gen_level;
2165 :
2166 346984 : if (ipa_edge_within_scc (cs))
2167 : {
2168 : ipcp_value_source<valtype> *s;
2169 48562 : for (s = val->sources; s; s = s->next)
2170 44373 : if (s->cs == cs && s->val == src_val)
2171 : break;
2172 11844 : if (s)
2173 : return false;
2174 : }
2175 :
2176 339329 : val->add_source (cs, src_val, src_idx, offset);
2177 339329 : return false;
2178 : }
2179 :
2180 148264 : if (!same_lat_gen_level && values_count >= opt_for_fn (cs->callee->decl,
2181 : param_ipa_cp_value_list_size))
2182 : {
2183 : /* We can only free sources, not the values themselves, because sources
2184 : of other values in this SCC might point to them. */
2185 12276 : for (val = values; val; val = val->next)
2186 : {
2187 40439 : while (val->sources)
2188 : {
2189 29529 : ipcp_value_source<valtype> *src = val->sources;
2190 29529 : val->sources = src->next;
2191 29529 : ipcp_sources_pool.remove ((ipcp_value_source<tree>*)src);
2192 : }
2193 : }
2194 1366 : values = NULL;
2195 1366 : return set_to_bottom ();
2196 : }
2197 :
2198 146898 : values_count++;
2199 146898 : val = allocate_and_init_ipcp_value (newval, same_lat_gen_level);
2200 146898 : val->add_source (cs, src_val, src_idx, offset);
2201 146898 : val->next = NULL;
2202 :
2203 : /* Add the new value to end of value list, which can reduce iterations
2204 : of propagation stage for recursive function. */
2205 146898 : if (last_val)
2206 45331 : last_val->next = val;
2207 : else
2208 101567 : values = val;
2209 :
2210 146898 : if (val_p)
2211 841 : *val_p = val;
2212 :
2213 : return true;
2214 : }
2215 :
2216 : /* A helper function that returns result of operation specified by OPCODE on
2217 : the value of SRC_VAL. If non-NULL, OPND1_TYPE is expected type for the
2218 : value of SRC_VAL. If the operation is binary, OPND2 is a constant value
2219 : acting as its second operand. OP_TYPE is the type in which the operation is
2220 : performed. */
2221 :
2222 : static tree
2223 21392 : get_val_across_arith_op (enum tree_code opcode,
2224 : tree opnd1_type,
2225 : tree opnd2,
2226 : ipcp_value<tree> *src_val,
2227 : tree op_type)
2228 : {
2229 21392 : tree opnd1 = src_val->value;
2230 :
2231 : /* Skip source values that is incompatible with specified type. */
2232 21392 : if (opnd1_type
2233 21392 : && !useless_type_conversion_p (opnd1_type, TREE_TYPE (opnd1)))
2234 : return NULL_TREE;
2235 :
2236 21392 : return ipa_get_jf_arith_result (opcode, opnd1, opnd2, op_type);
2237 : }
2238 :
2239 : /* Propagate values through an arithmetic transformation described by a jump
2240 : function associated with edge CS, taking values from SRC_LAT and putting
2241 : them into DEST_LAT. OPND1_TYPE, if non-NULL, is the expected type for the
2242 : values in SRC_LAT. OPND2 is a constant value if transformation is a binary
2243 : operation. SRC_OFFSET specifies offset in an aggregate if SRC_LAT describes
2244 : lattice of a part of an aggregate, otherwise it should be -1. SRC_IDX is
2245 : the index of the source parameter. OP_TYPE is the type in which the
2246 : operation is performed and can be NULL when OPCODE is NOP_EXPR. RES_TYPE is
2247 : the value type of result being propagated into. Return true if DEST_LAT
2248 : changed. */
2249 :
2250 : static bool
2251 76745 : propagate_vals_across_arith_jfunc (cgraph_edge *cs,
2252 : enum tree_code opcode,
2253 : tree opnd1_type,
2254 : tree opnd2,
2255 : ipcp_lattice<tree> *src_lat,
2256 : ipcp_lattice<tree> *dest_lat,
2257 : HOST_WIDE_INT src_offset,
2258 : int src_idx,
2259 : tree op_type,
2260 : tree res_type)
2261 : {
2262 76745 : ipcp_value<tree> *src_val;
2263 76745 : bool ret = false;
2264 :
2265 : /* Due to circular dependencies, propagating within an SCC through arithmetic
2266 : transformation would create infinite number of values. But for
2267 : self-feeding recursive function, we could allow propagation in a limited
2268 : count, and this can enable a simple kind of recursive function versioning.
2269 : For other scenario, we would just make lattices bottom. */
2270 76745 : if (opcode != NOP_EXPR && ipa_edge_within_scc (cs))
2271 : {
2272 2184 : int i;
2273 :
2274 2184 : int max_recursive_depth = opt_for_fn(cs->caller->decl,
2275 : param_ipa_cp_max_recursive_depth);
2276 2184 : if (src_lat != dest_lat || max_recursive_depth < 1)
2277 1666 : return dest_lat->set_contains_variable ();
2278 :
2279 : /* No benefit if recursive execution is in low probability. */
2280 1300 : if (cs->sreal_frequency () * 100
2281 2600 : <= ((sreal) 1) * opt_for_fn (cs->caller->decl,
2282 : param_ipa_cp_min_recursive_probability))
2283 89 : return dest_lat->set_contains_variable ();
2284 :
2285 1211 : auto_vec<ipcp_value<tree> *, 8> val_seeds;
2286 :
2287 2258 : for (src_val = src_lat->values; src_val; src_val = src_val->next)
2288 : {
2289 : /* Now we do not use self-recursively generated value as propagation
2290 : source, this is absolutely conservative, but could avoid explosion
2291 : of lattice's value space, especially when one recursive function
2292 : calls another recursive. */
2293 1740 : if (src_val->self_recursion_generated_p ())
2294 : {
2295 909 : ipcp_value_source<tree> *s;
2296 :
2297 : /* If the lattice has already been propagated for the call site,
2298 : no need to do that again. */
2299 1422 : for (s = src_val->sources; s; s = s->next)
2300 1206 : if (s->cs == cs)
2301 693 : return dest_lat->set_contains_variable ();
2302 : }
2303 : else
2304 831 : val_seeds.safe_push (src_val);
2305 : }
2306 :
2307 1036 : gcc_assert ((int) val_seeds.length () <= param_ipa_cp_value_list_size);
2308 :
2309 : /* Recursively generate lattice values with a limited count. */
2310 836 : FOR_EACH_VEC_ELT (val_seeds, i, src_val)
2311 : {
2312 1416 : for (int j = 1; j < max_recursive_depth; j++)
2313 : {
2314 1261 : tree cstval = get_val_across_arith_op (opcode, opnd1_type, opnd2,
2315 : src_val, op_type);
2316 1261 : cstval = ipacp_value_safe_for_type (res_type, cstval);
2317 1261 : if (!cstval)
2318 : break;
2319 :
2320 1257 : ret |= dest_lat->add_value (cstval, cs, src_val, src_idx,
2321 : src_offset, &src_val, j);
2322 1257 : gcc_checking_assert (src_val);
2323 : }
2324 : }
2325 518 : ret |= dest_lat->set_contains_variable ();
2326 1211 : }
2327 : else
2328 94817 : for (src_val = src_lat->values; src_val; src_val = src_val->next)
2329 : {
2330 : /* Now we do not use self-recursively generated value as propagation
2331 : source, otherwise it is easy to make value space of normal lattice
2332 : overflow. */
2333 20256 : if (src_val->self_recursion_generated_p ())
2334 : {
2335 125 : ret |= dest_lat->set_contains_variable ();
2336 125 : continue;
2337 : }
2338 :
2339 20131 : tree cstval = get_val_across_arith_op (opcode, opnd1_type, opnd2,
2340 : src_val, op_type);
2341 20131 : cstval = ipacp_value_safe_for_type (res_type, cstval);
2342 20131 : if (cstval)
2343 19930 : ret |= dest_lat->add_value (cstval, cs, src_val, src_idx,
2344 : src_offset);
2345 : else
2346 201 : ret |= dest_lat->set_contains_variable ();
2347 : }
2348 :
2349 : return ret;
2350 : }
2351 :
2352 : /* Propagate values through a pass-through jump function JFUNC associated with
2353 : edge CS, taking values from SRC_LAT and putting them into DEST_LAT. SRC_IDX
2354 : is the index of the source parameter. PARM_TYPE is the type of the
2355 : parameter to which the result is passed. */
2356 :
2357 : static bool
2358 72130 : propagate_vals_across_pass_through (cgraph_edge *cs, ipa_jump_func *jfunc,
2359 : ipcp_lattice<tree> *src_lat,
2360 : ipcp_lattice<tree> *dest_lat, int src_idx,
2361 : tree parm_type)
2362 : {
2363 72130 : gcc_checking_assert (parm_type);
2364 72130 : enum tree_code opcode = ipa_get_jf_pass_through_operation (jfunc);
2365 72130 : tree op_type = (opcode == NOP_EXPR) ? NULL_TREE
2366 2429 : : ipa_get_jf_pass_through_op_type (jfunc);
2367 72130 : return propagate_vals_across_arith_jfunc (cs, opcode, NULL_TREE,
2368 : ipa_get_jf_pass_through_operand (jfunc),
2369 : src_lat, dest_lat, -1, src_idx, op_type,
2370 72130 : parm_type);
2371 : }
2372 :
2373 : /* Propagate values through an ancestor jump function JFUNC associated with
2374 : edge CS, taking values from SRC_LAT and putting them into DEST_LAT. SRC_IDX
2375 : is the index of the source parameter. */
2376 :
2377 : static bool
2378 2203 : propagate_vals_across_ancestor (struct cgraph_edge *cs,
2379 : struct ipa_jump_func *jfunc,
2380 : ipcp_lattice<tree> *src_lat,
2381 : ipcp_lattice<tree> *dest_lat, int src_idx,
2382 : tree param_type)
2383 : {
2384 2203 : ipcp_value<tree> *src_val;
2385 2203 : bool ret = false;
2386 :
2387 2203 : if (ipa_edge_within_scc (cs))
2388 14 : return dest_lat->set_contains_variable ();
2389 :
2390 2500 : for (src_val = src_lat->values; src_val; src_val = src_val->next)
2391 : {
2392 311 : tree t = ipa_get_jf_ancestor_result (jfunc, src_val->value);
2393 311 : t = ipacp_value_safe_for_type (param_type, t);
2394 311 : if (t)
2395 253 : ret |= dest_lat->add_value (t, cs, src_val, src_idx);
2396 : else
2397 58 : ret |= dest_lat->set_contains_variable ();
2398 : }
2399 :
2400 : return ret;
2401 : }
2402 :
2403 : /* Propagate scalar values across jump function JFUNC that is associated with
2404 : edge CS and put the values into DEST_LAT. PARM_TYPE is the type of the
2405 : parameter to which the result is passed. */
2406 :
2407 : static bool
2408 3845318 : propagate_scalar_across_jump_function (struct cgraph_edge *cs,
2409 : struct ipa_jump_func *jfunc,
2410 : ipcp_lattice<tree> *dest_lat,
2411 : tree param_type)
2412 : {
2413 3845318 : if (dest_lat->bottom)
2414 : return false;
2415 :
2416 819190 : if (jfunc->type == IPA_JF_CONST)
2417 : {
2418 369970 : tree val = ipa_get_jf_constant (jfunc);
2419 369970 : val = ipacp_value_safe_for_type (param_type, val);
2420 369970 : if (val)
2421 369953 : return dest_lat->add_value (val, cs, NULL, 0);
2422 : else
2423 17 : return dest_lat->set_contains_variable ();
2424 : }
2425 449220 : else if (jfunc->type == IPA_JF_PASS_THROUGH
2426 269134 : || jfunc->type == IPA_JF_ANCESTOR)
2427 : {
2428 184525 : ipa_node_params *caller_info = ipa_node_params_sum->get (cs->caller);
2429 184525 : ipcp_lattice<tree> *src_lat;
2430 184525 : int src_idx;
2431 184525 : bool ret;
2432 :
2433 184525 : if (jfunc->type == IPA_JF_PASS_THROUGH)
2434 180086 : src_idx = ipa_get_jf_pass_through_formal_id (jfunc);
2435 : else
2436 4439 : src_idx = ipa_get_jf_ancestor_formal_id (jfunc);
2437 :
2438 184525 : src_lat = ipa_get_scalar_lat (caller_info, src_idx);
2439 184525 : if (src_lat->bottom)
2440 110044 : return dest_lat->set_contains_variable ();
2441 :
2442 : /* If we would need to clone the caller and cannot, do not propagate. */
2443 74481 : if (!ipcp_versionable_function_p (cs->caller)
2444 74481 : && (src_lat->contains_variable
2445 132 : || (src_lat->values_count > 1)))
2446 148 : return dest_lat->set_contains_variable ();
2447 :
2448 74333 : if (jfunc->type == IPA_JF_PASS_THROUGH)
2449 72130 : ret = propagate_vals_across_pass_through (cs, jfunc, src_lat,
2450 : dest_lat, src_idx,
2451 : param_type);
2452 : else
2453 2203 : ret = propagate_vals_across_ancestor (cs, jfunc, src_lat, dest_lat,
2454 : src_idx, param_type);
2455 :
2456 74333 : if (src_lat->contains_variable)
2457 64950 : ret |= dest_lat->set_contains_variable ();
2458 :
2459 74333 : return ret;
2460 : }
2461 :
2462 : /* TODO: We currently do not handle member method pointers in IPA-CP (we only
2463 : use it for indirect inlining), we should propagate them too. */
2464 264695 : return dest_lat->set_contains_variable ();
2465 : }
2466 :
2467 : /* Propagate scalar values across jump function JFUNC that is associated with
2468 : edge CS and describes argument IDX and put the values into DEST_LAT. */
2469 :
2470 : static bool
2471 3845318 : propagate_context_across_jump_function (cgraph_edge *cs,
2472 : ipa_jump_func *jfunc, int idx,
2473 : ipcp_lattice<ipa_polymorphic_call_context> *dest_lat)
2474 : {
2475 3845318 : if (dest_lat->bottom)
2476 : return false;
2477 915672 : ipa_edge_args *args = ipa_edge_args_sum->get (cs);
2478 915672 : bool ret = false;
2479 915672 : bool added_sth = false;
2480 915672 : bool type_preserved = true;
2481 :
2482 915672 : ipa_polymorphic_call_context edge_ctx, *edge_ctx_ptr
2483 929567 : = ipa_get_ith_polymorhic_call_context (args, idx);
2484 :
2485 13895 : if (edge_ctx_ptr)
2486 13895 : edge_ctx = *edge_ctx_ptr;
2487 :
2488 915672 : if (jfunc->type == IPA_JF_PASS_THROUGH
2489 735095 : || jfunc->type == IPA_JF_ANCESTOR)
2490 : {
2491 185112 : ipa_node_params *caller_info = ipa_node_params_sum->get (cs->caller);
2492 185112 : int src_idx;
2493 185112 : ipcp_lattice<ipa_polymorphic_call_context> *src_lat;
2494 :
2495 : /* TODO: Once we figure out how to propagate speculations, it will
2496 : probably be a good idea to switch to speculation if type_preserved is
2497 : not set instead of punting. */
2498 185112 : if (jfunc->type == IPA_JF_PASS_THROUGH)
2499 : {
2500 180577 : if (ipa_get_jf_pass_through_operation (jfunc) != NOP_EXPR)
2501 7580 : goto prop_fail;
2502 172997 : type_preserved = ipa_get_jf_pass_through_type_preserved (jfunc);
2503 172997 : src_idx = ipa_get_jf_pass_through_formal_id (jfunc);
2504 : }
2505 : else
2506 : {
2507 4535 : type_preserved = ipa_get_jf_ancestor_type_preserved (jfunc);
2508 4535 : src_idx = ipa_get_jf_ancestor_formal_id (jfunc);
2509 : }
2510 :
2511 177532 : src_lat = ipa_get_poly_ctx_lat (caller_info, src_idx);
2512 : /* If we would need to clone the caller and cannot, do not propagate. */
2513 177532 : if (!ipcp_versionable_function_p (cs->caller)
2514 177532 : && (src_lat->contains_variable
2515 14005 : || (src_lat->values_count > 1)))
2516 2485 : goto prop_fail;
2517 :
2518 175047 : ipcp_value<ipa_polymorphic_call_context> *src_val;
2519 176353 : for (src_val = src_lat->values; src_val; src_val = src_val->next)
2520 : {
2521 1306 : ipa_polymorphic_call_context cur = src_val->value;
2522 :
2523 1306 : if (!type_preserved)
2524 871 : cur.possible_dynamic_type_change (cs->in_polymorphic_cdtor);
2525 1306 : if (jfunc->type == IPA_JF_ANCESTOR)
2526 325 : cur.offset_by (ipa_get_jf_ancestor_offset (jfunc));
2527 : /* TODO: In cases we know how the context is going to be used,
2528 : we can improve the result by passing proper OTR_TYPE. */
2529 1306 : cur.combine_with (edge_ctx);
2530 2612 : if (!cur.useless_p ())
2531 : {
2532 847 : if (src_lat->contains_variable
2533 847 : && !edge_ctx.equal_to (cur))
2534 269 : ret |= dest_lat->set_contains_variable ();
2535 847 : ret |= dest_lat->add_value (cur, cs, src_val, src_idx);
2536 847 : added_sth = true;
2537 : }
2538 : }
2539 : }
2540 :
2541 730560 : prop_fail:
2542 185112 : if (!added_sth)
2543 : {
2544 914888 : if (!edge_ctx.useless_p ())
2545 8487 : ret |= dest_lat->add_value (edge_ctx, cs);
2546 : else
2547 906401 : ret |= dest_lat->set_contains_variable ();
2548 : }
2549 :
2550 : return ret;
2551 : }
2552 :
2553 : /* Propagate bits across jfunc that is associated with
2554 : edge cs and update dest_lattice accordingly. */
2555 :
2556 : bool
2557 3845318 : propagate_bits_across_jump_function (cgraph_edge *cs, int idx,
2558 : ipa_jump_func *jfunc,
2559 : ipcp_bits_lattice *dest_lattice)
2560 : {
2561 3845318 : if (dest_lattice->bottom_p ())
2562 : return false;
2563 :
2564 531588 : enum availability availability;
2565 531588 : cgraph_node *callee = cs->callee->function_symbol (&availability);
2566 531588 : ipa_node_params *callee_info = ipa_node_params_sum->get (callee);
2567 531588 : tree parm_type = ipa_get_type (callee_info, idx);
2568 :
2569 : /* For K&R C programs, ipa_get_type() could return NULL_TREE. Avoid the
2570 : transform for these cases. Similarly, we can have bad type mismatches
2571 : with LTO, avoid doing anything with those too. */
2572 531588 : if (!parm_type
2573 531588 : || (!INTEGRAL_TYPE_P (parm_type) && !POINTER_TYPE_P (parm_type)))
2574 : {
2575 29210 : if (dump_file && (dump_flags & TDF_DETAILS))
2576 11 : fprintf (dump_file, "Setting dest_lattice to bottom, because type of "
2577 : "param %i of %s is NULL or unsuitable for bits propagation\n",
2578 11 : idx, cs->callee->dump_name ());
2579 :
2580 29210 : return dest_lattice->set_to_bottom ();
2581 : }
2582 :
2583 502378 : if (jfunc->type == IPA_JF_PASS_THROUGH
2584 404710 : || jfunc->type == IPA_JF_ANCESTOR)
2585 : {
2586 100145 : ipa_node_params *caller_info = ipa_node_params_sum->get (cs->caller);
2587 100145 : tree operand = NULL_TREE;
2588 100145 : tree op_type = NULL_TREE;
2589 100145 : enum tree_code code;
2590 100145 : unsigned src_idx;
2591 100145 : bool keep_null = false;
2592 :
2593 100145 : if (jfunc->type == IPA_JF_PASS_THROUGH)
2594 : {
2595 97668 : code = ipa_get_jf_pass_through_operation (jfunc);
2596 97668 : src_idx = ipa_get_jf_pass_through_formal_id (jfunc);
2597 97668 : if (code != NOP_EXPR)
2598 : {
2599 2069 : operand = ipa_get_jf_pass_through_operand (jfunc);
2600 2069 : op_type = ipa_get_jf_pass_through_op_type (jfunc);
2601 : }
2602 : }
2603 : else
2604 : {
2605 2477 : code = POINTER_PLUS_EXPR;
2606 2477 : src_idx = ipa_get_jf_ancestor_formal_id (jfunc);
2607 2477 : unsigned HOST_WIDE_INT offset
2608 2477 : = ipa_get_jf_ancestor_offset (jfunc) / BITS_PER_UNIT;
2609 2477 : keep_null = (ipa_get_jf_ancestor_keep_null (jfunc) || !offset);
2610 2477 : operand = build_int_cstu (size_type_node, offset);
2611 : }
2612 :
2613 100145 : class ipcp_param_lattices *src_lats
2614 100145 : = ipa_get_parm_lattices (caller_info, src_idx);
2615 :
2616 : /* Try to propagate bits if src_lattice is bottom, but jfunc is known.
2617 : for eg consider:
2618 : int f(int x)
2619 : {
2620 : g (x & 0xff);
2621 : }
2622 : Assume lattice for x is bottom, however we can still propagate
2623 : result of x & 0xff == 0xff, which gets computed during ccp1 pass
2624 : and we store it in jump function during analysis stage. */
2625 :
2626 100145 : if (!src_lats->bits_lattice.bottom_p ()
2627 100145 : && !src_lats->bits_lattice.recipient_only_p ())
2628 : {
2629 21452 : if (!op_type)
2630 20277 : op_type = ipa_get_type (caller_info, src_idx);
2631 :
2632 21452 : unsigned precision = TYPE_PRECISION (op_type);
2633 21452 : signop sgn = TYPE_SIGN (op_type);
2634 21452 : bool drop_all_ones
2635 21452 : = keep_null && !src_lats->bits_lattice.known_nonzero_p ();
2636 :
2637 21452 : return dest_lattice->meet_with (src_lats->bits_lattice, precision,
2638 21452 : sgn, code, operand, drop_all_ones);
2639 : }
2640 : }
2641 :
2642 480926 : value_range vr (parm_type);
2643 480926 : if (jfunc->m_vr)
2644 : {
2645 410217 : jfunc->m_vr->get_vrange (vr);
2646 410217 : if (!vr.undefined_p () && !vr.varying_p ())
2647 : {
2648 410217 : irange_bitmask bm = vr.get_bitmask ();
2649 410217 : widest_int mask
2650 410217 : = widest_int::from (bm.mask (), TYPE_SIGN (parm_type));
2651 410217 : widest_int value
2652 410217 : = widest_int::from (bm.value (), TYPE_SIGN (parm_type));
2653 410217 : return dest_lattice->meet_with (value, mask,
2654 410217 : TYPE_PRECISION (parm_type));
2655 410217 : }
2656 : }
2657 70709 : return dest_lattice->set_to_bottom ();
2658 480926 : }
2659 :
2660 : /* Propagate value range across jump function JFUNC that is associated with
2661 : edge CS with param of callee of PARAM_TYPE and update DEST_PLATS
2662 : accordingly. */
2663 :
2664 : static bool
2665 3844473 : propagate_vr_across_jump_function (cgraph_edge *cs, ipa_jump_func *jfunc,
2666 : class ipcp_param_lattices *dest_plats,
2667 : tree param_type)
2668 : {
2669 3844473 : ipcp_vr_lattice *dest_lat = &dest_plats->m_value_range;
2670 :
2671 3844473 : if (dest_lat->bottom_p ())
2672 : return false;
2673 :
2674 624660 : if (!param_type
2675 624660 : || !ipa_vr_supported_type_p (param_type))
2676 29150 : return dest_lat->set_to_bottom ();
2677 :
2678 595510 : value_range vr (param_type);
2679 595510 : vr.set_varying (param_type);
2680 595510 : if (jfunc->m_vr)
2681 514018 : ipa_vr_operation_and_type_effects (vr, *jfunc->m_vr, NOP_EXPR,
2682 : param_type,
2683 514018 : jfunc->m_vr->type ());
2684 :
2685 595510 : if (jfunc->type == IPA_JF_PASS_THROUGH)
2686 : {
2687 91815 : ipa_node_params *caller_info = ipa_node_params_sum->get (cs->caller);
2688 91815 : int src_idx = ipa_get_jf_pass_through_formal_id (jfunc);
2689 91815 : class ipcp_param_lattices *src_lats
2690 91815 : = ipa_get_parm_lattices (caller_info, src_idx);
2691 91815 : tree operand_type = ipa_get_type (caller_info, src_idx);
2692 :
2693 91815 : if (src_lats->m_value_range.bottom_p ()
2694 91815 : || src_lats->m_value_range.recipient_only_p ())
2695 76118 : return dest_lat->set_to_bottom ();
2696 :
2697 15697 : if (ipa_get_jf_pass_through_operation (jfunc) == NOP_EXPR
2698 15697 : || !ipa_edge_within_scc (cs))
2699 15202 : ipa_vr_intersect_with_arith_jfunc (vr, jfunc, cs->caller,
2700 15202 : src_lats->m_value_range.m_vr,
2701 : operand_type, param_type);
2702 : }
2703 :
2704 519392 : if (!vr.undefined_p () && !vr.varying_p ())
2705 491549 : return dest_lat->meet_with (vr);
2706 : else
2707 27843 : return dest_lat->set_to_bottom ();
2708 595510 : }
2709 :
2710 : /* If DEST_PLATS already has aggregate items, check that aggs_by_ref matches
2711 : NEW_AGGS_BY_REF and if not, mark all aggs as bottoms and return true (in all
2712 : other cases, return false). If there are no aggregate items, set
2713 : aggs_by_ref to NEW_AGGS_BY_REF. */
2714 :
2715 : static bool
2716 41868 : set_check_aggs_by_ref (class ipcp_param_lattices *dest_plats,
2717 : bool new_aggs_by_ref)
2718 : {
2719 0 : if (dest_plats->aggs)
2720 : {
2721 22307 : if (dest_plats->aggs_by_ref != new_aggs_by_ref)
2722 : {
2723 0 : set_agg_lats_to_bottom (dest_plats);
2724 0 : return true;
2725 : }
2726 : }
2727 : else
2728 19561 : dest_plats->aggs_by_ref = new_aggs_by_ref;
2729 : return false;
2730 : }
2731 :
2732 : /* Walk aggregate lattices in DEST_PLATS from ***AGLAT on, until ***aglat is an
2733 : already existing lattice for the given OFFSET and SIZE, marking all skipped
2734 : lattices as containing variable and checking for overlaps. If there is no
2735 : already existing lattice for the OFFSET and VAL_SIZE, create one, initialize
2736 : it with offset, size and contains_variable to PRE_EXISTING, and return true,
2737 : unless there are too many already. If there are two many, return false. If
2738 : there are overlaps turn whole DEST_PLATS to bottom and return false. If any
2739 : skipped lattices were newly marked as containing variable, set *CHANGE to
2740 : true. MAX_AGG_ITEMS is the maximum number of lattices. */
2741 :
2742 : static bool
2743 112701 : merge_agg_lats_step (class ipcp_param_lattices *dest_plats,
2744 : HOST_WIDE_INT offset, HOST_WIDE_INT val_size,
2745 : struct ipcp_agg_lattice ***aglat,
2746 : bool pre_existing, bool *change, int max_agg_items)
2747 : {
2748 112701 : gcc_checking_assert (offset >= 0);
2749 :
2750 116809 : while (**aglat && (**aglat)->offset < offset)
2751 : {
2752 4108 : if ((**aglat)->offset + (**aglat)->size > offset)
2753 : {
2754 0 : set_agg_lats_to_bottom (dest_plats);
2755 0 : return false;
2756 : }
2757 4108 : *change |= (**aglat)->set_contains_variable ();
2758 4108 : *aglat = &(**aglat)->next;
2759 : }
2760 :
2761 112701 : if (**aglat && (**aglat)->offset == offset)
2762 : {
2763 56240 : if ((**aglat)->size != val_size)
2764 : {
2765 13 : set_agg_lats_to_bottom (dest_plats);
2766 13 : return false;
2767 : }
2768 56227 : gcc_assert (!(**aglat)->next
2769 : || (**aglat)->next->offset >= offset + val_size);
2770 : return true;
2771 : }
2772 : else
2773 : {
2774 56461 : struct ipcp_agg_lattice *new_al;
2775 :
2776 56461 : if (**aglat && (**aglat)->offset < offset + val_size)
2777 : {
2778 3 : set_agg_lats_to_bottom (dest_plats);
2779 3 : return false;
2780 : }
2781 56458 : if (dest_plats->aggs_count == max_agg_items)
2782 : return false;
2783 56419 : dest_plats->aggs_count++;
2784 56419 : new_al = ipcp_agg_lattice_pool.allocate ();
2785 :
2786 56419 : new_al->offset = offset;
2787 56419 : new_al->size = val_size;
2788 56419 : new_al->contains_variable = pre_existing;
2789 :
2790 56419 : new_al->next = **aglat;
2791 56419 : **aglat = new_al;
2792 56419 : return true;
2793 : }
2794 : }
2795 :
2796 : /* Set all AGLAT and all other aggregate lattices reachable by next pointers as
2797 : containing an unknown value. */
2798 :
2799 : static bool
2800 41850 : set_chain_of_aglats_contains_variable (struct ipcp_agg_lattice *aglat)
2801 : {
2802 41850 : bool ret = false;
2803 44308 : while (aglat)
2804 : {
2805 2458 : ret |= aglat->set_contains_variable ();
2806 2458 : aglat = aglat->next;
2807 : }
2808 41850 : return ret;
2809 : }
2810 :
2811 : /* Merge existing aggregate lattices in SRC_PLATS to DEST_PLATS, subtracting
2812 : DELTA_OFFSET. CS is the call graph edge and SRC_IDX the index of the source
2813 : parameter used for lattice value sources. Return true if DEST_PLATS changed
2814 : in any way. */
2815 :
2816 : static bool
2817 3984 : merge_aggregate_lattices (struct cgraph_edge *cs,
2818 : class ipcp_param_lattices *dest_plats,
2819 : class ipcp_param_lattices *src_plats,
2820 : int src_idx, HOST_WIDE_INT offset_delta)
2821 : {
2822 3984 : bool pre_existing = dest_plats->aggs != NULL;
2823 3984 : struct ipcp_agg_lattice **dst_aglat;
2824 3984 : bool ret = false;
2825 :
2826 3984 : if (set_check_aggs_by_ref (dest_plats, src_plats->aggs_by_ref))
2827 0 : return true;
2828 3984 : if (src_plats->aggs_bottom)
2829 2 : return set_agg_lats_contain_variable (dest_plats);
2830 3982 : if (src_plats->aggs_contain_variable)
2831 2314 : ret |= set_agg_lats_contain_variable (dest_plats);
2832 3982 : dst_aglat = &dest_plats->aggs;
2833 :
2834 3982 : int max_agg_items = opt_for_fn (cs->callee->function_symbol ()->decl,
2835 : param_ipa_max_agg_items);
2836 3982 : for (struct ipcp_agg_lattice *src_aglat = src_plats->aggs;
2837 11663 : src_aglat;
2838 7681 : src_aglat = src_aglat->next)
2839 : {
2840 7681 : HOST_WIDE_INT new_offset = src_aglat->offset - offset_delta;
2841 :
2842 7681 : if (new_offset < 0)
2843 51 : continue;
2844 7630 : if (merge_agg_lats_step (dest_plats, new_offset, src_aglat->size,
2845 : &dst_aglat, pre_existing, &ret, max_agg_items))
2846 : {
2847 7626 : struct ipcp_agg_lattice *new_al = *dst_aglat;
2848 :
2849 7626 : dst_aglat = &(*dst_aglat)->next;
2850 7626 : if (src_aglat->bottom)
2851 : {
2852 0 : ret |= new_al->set_contains_variable ();
2853 0 : continue;
2854 : }
2855 7626 : if (src_aglat->contains_variable)
2856 4444 : ret |= new_al->set_contains_variable ();
2857 7626 : for (ipcp_value<tree> *val = src_aglat->values;
2858 11827 : val;
2859 4201 : val = val->next)
2860 4201 : ret |= new_al->add_value (val->value, cs, val, src_idx,
2861 : src_aglat->offset);
2862 : }
2863 4 : else if (dest_plats->aggs_bottom)
2864 : return true;
2865 : }
2866 3982 : ret |= set_chain_of_aglats_contains_variable (*dst_aglat);
2867 3982 : return ret;
2868 : }
2869 :
2870 : /* Determine whether there is anything to propagate FROM SRC_PLATS through a
2871 : pass-through JFUNC and if so, whether it has conform and conforms to the
2872 : rules about propagating values passed by reference. */
2873 :
2874 : static bool
2875 172838 : agg_pass_through_permissible_p (class ipcp_param_lattices *src_plats,
2876 : struct ipa_jump_func *jfunc)
2877 : {
2878 172838 : return src_plats->aggs
2879 172838 : && (!src_plats->aggs_by_ref
2880 5019 : || ipa_get_jf_pass_through_agg_preserved (jfunc));
2881 : }
2882 :
2883 : /* Propagate values through ITEM, jump function for a part of an aggregate,
2884 : into corresponding aggregate lattice AGLAT. CS is the call graph edge
2885 : associated with the jump function. Return true if AGLAT changed in any
2886 : way. */
2887 :
2888 : static bool
2889 105020 : propagate_aggregate_lattice (struct cgraph_edge *cs,
2890 : struct ipa_agg_jf_item *item,
2891 : struct ipcp_agg_lattice *aglat)
2892 : {
2893 105020 : class ipa_node_params *caller_info;
2894 105020 : class ipcp_param_lattices *src_plats;
2895 105020 : struct ipcp_lattice<tree> *src_lat;
2896 105020 : HOST_WIDE_INT src_offset;
2897 105020 : int src_idx;
2898 105020 : tree load_type;
2899 105020 : bool ret;
2900 :
2901 105020 : if (item->jftype == IPA_JF_CONST)
2902 : {
2903 93678 : tree value = item->value.constant;
2904 :
2905 93678 : gcc_checking_assert (is_gimple_ip_invariant (value));
2906 93678 : return aglat->add_value (value, cs, NULL, 0);
2907 : }
2908 :
2909 11342 : gcc_checking_assert (item->jftype == IPA_JF_PASS_THROUGH
2910 : || item->jftype == IPA_JF_LOAD_AGG);
2911 :
2912 11342 : caller_info = ipa_node_params_sum->get (cs->caller);
2913 11342 : src_idx = item->value.pass_through.formal_id;
2914 11342 : src_plats = ipa_get_parm_lattices (caller_info, src_idx);
2915 :
2916 11342 : if (item->jftype == IPA_JF_PASS_THROUGH)
2917 : {
2918 3294 : load_type = NULL_TREE;
2919 3294 : src_lat = &src_plats->itself;
2920 3294 : src_offset = -1;
2921 : }
2922 : else
2923 : {
2924 8048 : HOST_WIDE_INT load_offset = item->value.load_agg.offset;
2925 8048 : struct ipcp_agg_lattice *src_aglat;
2926 :
2927 12452 : for (src_aglat = src_plats->aggs; src_aglat; src_aglat = src_aglat->next)
2928 8178 : if (src_aglat->offset >= load_offset)
2929 : break;
2930 :
2931 8048 : load_type = item->value.load_agg.type;
2932 8048 : if (!src_aglat
2933 3774 : || src_aglat->offset > load_offset
2934 3432 : || src_aglat->size != tree_to_shwi (TYPE_SIZE (load_type))
2935 11480 : || src_plats->aggs_by_ref != item->value.load_agg.by_ref)
2936 4616 : return aglat->set_contains_variable ();
2937 :
2938 : src_lat = src_aglat;
2939 : src_offset = load_offset;
2940 : }
2941 :
2942 6726 : if (src_lat->bottom
2943 6726 : || (!ipcp_versionable_function_p (cs->caller)
2944 6726 : && !src_lat->is_single_const ()))
2945 2111 : return aglat->set_contains_variable ();
2946 :
2947 4615 : ret = propagate_vals_across_arith_jfunc (cs,
2948 : item->value.pass_through.operation,
2949 : load_type,
2950 : item->value.pass_through.operand,
2951 : src_lat, aglat,
2952 : src_offset,
2953 : src_idx,
2954 : item->value.pass_through.op_type,
2955 : item->type);
2956 :
2957 4615 : if (src_lat->contains_variable)
2958 2640 : ret |= aglat->set_contains_variable ();
2959 :
2960 : return ret;
2961 : }
2962 :
2963 : /* Propagate scalar values across jump function JFUNC that is associated with
2964 : edge CS and put the values into DEST_LAT. */
2965 :
2966 : static bool
2967 3845318 : propagate_aggs_across_jump_function (struct cgraph_edge *cs,
2968 : struct ipa_jump_func *jfunc,
2969 : class ipcp_param_lattices *dest_plats)
2970 : {
2971 3845318 : bool ret = false;
2972 :
2973 3845318 : if (dest_plats->aggs_bottom)
2974 : return false;
2975 :
2976 914511 : if (jfunc->type == IPA_JF_PASS_THROUGH
2977 914511 : && ipa_get_jf_pass_through_operation (jfunc) == NOP_EXPR)
2978 : {
2979 172838 : ipa_node_params *caller_info = ipa_node_params_sum->get (cs->caller);
2980 172838 : int src_idx = ipa_get_jf_pass_through_formal_id (jfunc);
2981 172838 : class ipcp_param_lattices *src_plats;
2982 :
2983 172838 : src_plats = ipa_get_parm_lattices (caller_info, src_idx);
2984 172838 : if (agg_pass_through_permissible_p (src_plats, jfunc))
2985 : {
2986 : /* Currently we do not produce clobber aggregate jump
2987 : functions, replace with merging when we do. */
2988 3854 : gcc_assert (!jfunc->agg.items);
2989 3854 : ret |= merge_aggregate_lattices (cs, dest_plats, src_plats,
2990 : src_idx, 0);
2991 3854 : return ret;
2992 : }
2993 : }
2994 741673 : else if (jfunc->type == IPA_JF_ANCESTOR
2995 741673 : && ipa_get_jf_ancestor_agg_preserved (jfunc))
2996 : {
2997 1191 : ipa_node_params *caller_info = ipa_node_params_sum->get (cs->caller);
2998 1191 : int src_idx = ipa_get_jf_ancestor_formal_id (jfunc);
2999 1191 : class ipcp_param_lattices *src_plats;
3000 :
3001 1191 : src_plats = ipa_get_parm_lattices (caller_info, src_idx);
3002 1191 : if (src_plats->aggs && src_plats->aggs_by_ref)
3003 : {
3004 : /* Currently we do not produce clobber aggregate jump
3005 : functions, replace with merging when we do. */
3006 130 : gcc_assert (!jfunc->agg.items);
3007 130 : ret |= merge_aggregate_lattices (cs, dest_plats, src_plats, src_idx,
3008 : ipa_get_jf_ancestor_offset (jfunc));
3009 : }
3010 1061 : else if (!src_plats->aggs_by_ref)
3011 1057 : ret |= set_agg_lats_to_bottom (dest_plats);
3012 : else
3013 4 : ret |= set_agg_lats_contain_variable (dest_plats);
3014 1191 : return ret;
3015 : }
3016 :
3017 909466 : if (jfunc->agg.items)
3018 : {
3019 37884 : bool pre_existing = dest_plats->aggs != NULL;
3020 37884 : struct ipcp_agg_lattice **aglat = &dest_plats->aggs;
3021 37884 : struct ipa_agg_jf_item *item;
3022 37884 : int i;
3023 :
3024 37884 : if (set_check_aggs_by_ref (dest_plats, jfunc->agg.by_ref))
3025 16 : return true;
3026 :
3027 37884 : int max_agg_items = opt_for_fn (cs->callee->function_symbol ()->decl,
3028 : param_ipa_max_agg_items);
3029 143260 : FOR_EACH_VEC_ELT (*jfunc->agg.items, i, item)
3030 : {
3031 105392 : HOST_WIDE_INT val_size;
3032 :
3033 105392 : if (item->offset < 0 || item->jftype == IPA_JF_UNKNOWN)
3034 321 : continue;
3035 105071 : val_size = tree_to_shwi (TYPE_SIZE (item->type));
3036 :
3037 105071 : if (merge_agg_lats_step (dest_plats, item->offset, val_size,
3038 : &aglat, pre_existing, &ret, max_agg_items))
3039 : {
3040 105020 : ret |= propagate_aggregate_lattice (cs, item, *aglat);
3041 105020 : aglat = &(*aglat)->next;
3042 : }
3043 51 : else if (dest_plats->aggs_bottom)
3044 : return true;
3045 : }
3046 :
3047 75736 : ret |= set_chain_of_aglats_contains_variable (*aglat);
3048 : }
3049 : else
3050 871582 : ret |= set_agg_lats_contain_variable (dest_plats);
3051 :
3052 909450 : return ret;
3053 : }
3054 :
3055 : /* Return true if on the way cfrom CS->caller to the final (non-alias and
3056 : non-thunk) destination, the call passes through a thunk. */
3057 :
3058 : static bool
3059 1915750 : call_passes_through_thunk (cgraph_edge *cs)
3060 : {
3061 1915750 : cgraph_node *alias_or_thunk = cs->callee;
3062 2054454 : while (alias_or_thunk->alias)
3063 138704 : alias_or_thunk = alias_or_thunk->get_alias_target ();
3064 1915750 : return alias_or_thunk->thunk;
3065 : }
3066 :
3067 : /* Propagate constants from the caller to the callee of CS. INFO describes the
3068 : caller. */
3069 :
3070 : static bool
3071 5275314 : propagate_constants_across_call (struct cgraph_edge *cs)
3072 : {
3073 5275314 : class ipa_node_params *callee_info;
3074 5275314 : enum availability availability;
3075 5275314 : cgraph_node *callee;
3076 5275314 : class ipa_edge_args *args;
3077 5275314 : bool ret = false;
3078 5275314 : int i, args_count, parms_count;
3079 :
3080 5275314 : callee = cs->callee->function_symbol (&availability);
3081 5275314 : if (!callee->definition)
3082 : return false;
3083 1938168 : gcc_checking_assert (callee->has_gimple_body_p ());
3084 1938168 : callee_info = ipa_node_params_sum->get (callee);
3085 1938168 : if (!callee_info)
3086 : return false;
3087 :
3088 1929682 : args = ipa_edge_args_sum->get (cs);
3089 1929682 : parms_count = ipa_get_param_count (callee_info);
3090 1736733 : if (parms_count == 0)
3091 : return false;
3092 1736733 : if (!args
3093 1736447 : || !opt_for_fn (cs->caller->decl, flag_ipa_cp)
3094 3473180 : || !opt_for_fn (cs->caller->decl, optimize))
3095 : {
3096 857 : for (i = 0; i < parms_count; i++)
3097 571 : ret |= set_all_contains_variable (ipa_get_parm_lattices (callee_info,
3098 : i));
3099 : return ret;
3100 : }
3101 1736447 : args_count = ipa_get_cs_argument_count (args);
3102 :
3103 : /* If this call goes through a thunk we must not propagate to the first (0th)
3104 : parameter. However, we might need to uncover a thunk from below a series
3105 : of aliases first. */
3106 1736447 : if (call_passes_through_thunk (cs))
3107 : {
3108 227 : ret |= set_all_contains_variable (ipa_get_parm_lattices (callee_info,
3109 : 0));
3110 227 : i = 1;
3111 : }
3112 : else
3113 : i = 0;
3114 :
3115 5721539 : for (; (i < args_count) && (i < parms_count); i++)
3116 : {
3117 3985092 : struct ipa_jump_func *jump_func = ipa_get_ith_jump_func (args, i);
3118 3985092 : class ipcp_param_lattices *dest_plats;
3119 3985092 : tree param_type = ipa_get_type (callee_info, i);
3120 :
3121 3985092 : dest_plats = ipa_get_parm_lattices (callee_info, i);
3122 3985092 : if (availability == AVAIL_INTERPOSABLE)
3123 139774 : ret |= set_all_contains_variable (dest_plats);
3124 : else
3125 : {
3126 3845318 : ret |= propagate_scalar_across_jump_function (cs, jump_func,
3127 : &dest_plats->itself,
3128 : param_type);
3129 3845318 : ret |= propagate_context_across_jump_function (cs, jump_func, i,
3130 : &dest_plats->ctxlat);
3131 3845318 : ret
3132 3845318 : |= propagate_bits_across_jump_function (cs, i, jump_func,
3133 : &dest_plats->bits_lattice);
3134 3845318 : ret |= propagate_aggs_across_jump_function (cs, jump_func,
3135 : dest_plats);
3136 3845318 : if (opt_for_fn (callee->decl, flag_ipa_vrp))
3137 3844473 : ret |= propagate_vr_across_jump_function (cs, jump_func,
3138 : dest_plats, param_type);
3139 : else
3140 845 : ret |= dest_plats->m_value_range.set_to_bottom ();
3141 : }
3142 : }
3143 1736630 : for (; i < parms_count; i++)
3144 183 : ret |= set_all_contains_variable (ipa_get_parm_lattices (callee_info, i));
3145 :
3146 : return ret;
3147 : }
3148 :
3149 : /* If an indirect edge IE can be turned into a direct one based on KNOWN_VALS
3150 : KNOWN_CONTEXTS, and known aggregates either in AVS or KNOWN_AGGS return
3151 : the destination. The latter three can be NULL. If AGG_REPS is not NULL,
3152 : KNOWN_AGGS is ignored. */
3153 :
3154 : static tree
3155 1565975 : ipa_get_indirect_edge_target_1 (struct cgraph_edge *ie,
3156 : const vec<tree> &known_csts,
3157 : const vec<ipa_polymorphic_call_context> &known_contexts,
3158 : const ipa_argagg_value_list &avs,
3159 : bool *speculative)
3160 : {
3161 1565975 : int param_index = ie->indirect_info->param_index;
3162 1565975 : *speculative = false;
3163 :
3164 1565975 : if (param_index == -1)
3165 : return NULL_TREE;
3166 :
3167 599778 : if (cgraph_simple_indirect_info *sii
3168 599778 : = dyn_cast <cgraph_simple_indirect_info *> (ie->indirect_info))
3169 : {
3170 308174 : tree t = NULL;
3171 :
3172 308174 : if (sii->agg_contents)
3173 : {
3174 72015 : t = NULL;
3175 72015 : if ((unsigned) param_index < known_csts.length ()
3176 72015 : && known_csts[param_index])
3177 61478 : t = ipa_find_agg_cst_from_init (known_csts[param_index],
3178 : sii->offset,
3179 : sii->by_ref);
3180 :
3181 72015 : if (!t && sii->guaranteed_unmodified)
3182 65873 : t = avs.get_value (param_index, sii->offset / BITS_PER_UNIT,
3183 : sii->by_ref);
3184 : }
3185 236159 : else if ((unsigned) param_index < known_csts.length ())
3186 236159 : t = known_csts[param_index];
3187 :
3188 308104 : if (t
3189 204956 : && TREE_CODE (t) == ADDR_EXPR
3190 512841 : && TREE_CODE (TREE_OPERAND (t, 0)) == FUNCTION_DECL)
3191 204737 : return TREE_OPERAND (t, 0);
3192 : else
3193 103437 : return NULL_TREE;
3194 : }
3195 :
3196 291604 : if (!opt_for_fn (ie->caller->decl, flag_devirtualize))
3197 : return NULL_TREE;
3198 :
3199 291604 : cgraph_polymorphic_indirect_info *pii
3200 291604 : = as_a <cgraph_polymorphic_indirect_info *> (ie->indirect_info);
3201 291604 : if (!pii->usable_p ())
3202 : return NULL_TREE;
3203 :
3204 291604 : HOST_WIDE_INT anc_offset = pii->offset;
3205 291604 : tree t = NULL;
3206 291604 : tree target = NULL;
3207 291604 : if ((unsigned) param_index < known_csts.length ()
3208 291604 : && known_csts[param_index])
3209 16977 : t = ipa_find_agg_cst_from_init (known_csts[param_index], anc_offset, true);
3210 :
3211 : /* Try to work out value of virtual table pointer value in replacements. */
3212 : /* or known aggregate values. */
3213 16977 : if (!t)
3214 291595 : t = avs.get_value (param_index, anc_offset / BITS_PER_UNIT, true);
3215 :
3216 : /* If we found the virtual table pointer, lookup the target. */
3217 291595 : if (t)
3218 : {
3219 7211 : tree vtable;
3220 7211 : unsigned HOST_WIDE_INT offset;
3221 7211 : if (vtable_pointer_value_to_vtable (t, &vtable, &offset))
3222 : {
3223 7211 : bool can_refer;
3224 7211 : target = gimple_get_virt_method_for_vtable (pii->otr_token, vtable,
3225 : offset, &can_refer);
3226 7211 : if (can_refer)
3227 : {
3228 7148 : if (!target
3229 7148 : || fndecl_built_in_p (target, BUILT_IN_UNREACHABLE)
3230 14176 : || !possible_polymorphic_call_target_p
3231 7028 : (ie, cgraph_node::get (target)))
3232 : {
3233 : /* Do not speculate builtin_unreachable, it is stupid! */
3234 237 : if (pii->vptr_changed)
3235 5679 : return NULL;
3236 237 : target = ipa_impossible_devirt_target (ie, target);
3237 : }
3238 7148 : *speculative = pii->vptr_changed;
3239 7148 : if (!*speculative)
3240 : return target;
3241 : }
3242 : }
3243 : }
3244 :
3245 : /* Do we know the constant value of pointer? */
3246 285925 : if (!t && (unsigned) param_index < known_csts.length ())
3247 52902 : t = known_csts[param_index];
3248 :
3249 285925 : ipa_polymorphic_call_context context;
3250 285925 : if (known_contexts.length () > (unsigned int) param_index)
3251 : {
3252 285553 : context = known_contexts[param_index];
3253 285553 : context.offset_by (anc_offset);
3254 285553 : if (pii->vptr_changed)
3255 48929 : context.possible_dynamic_type_change (ie->in_polymorphic_cdtor,
3256 : pii->otr_type);
3257 285553 : if (t)
3258 : {
3259 12020 : ipa_polymorphic_call_context ctx2
3260 12020 : = ipa_polymorphic_call_context (t, pii->otr_type, anc_offset);
3261 24040 : if (!ctx2.useless_p ())
3262 10493 : context.combine_with (ctx2, pii->otr_type);
3263 : }
3264 : }
3265 372 : else if (t)
3266 : {
3267 21 : context = ipa_polymorphic_call_context (t, pii->otr_type, anc_offset);
3268 21 : if (pii->vptr_changed)
3269 6 : context.possible_dynamic_type_change (ie->in_polymorphic_cdtor,
3270 : pii->otr_type);
3271 : }
3272 : else
3273 : return NULL_TREE;
3274 :
3275 285574 : vec <cgraph_node *>targets;
3276 285574 : bool final;
3277 :
3278 285574 : targets = possible_polymorphic_call_targets (pii->otr_type, pii->otr_token,
3279 : context, &final);
3280 297159 : if (!final || targets.length () > 1)
3281 : {
3282 274687 : struct cgraph_node *node;
3283 274687 : if (*speculative)
3284 : return target;
3285 274655 : if (!opt_for_fn (ie->caller->decl, flag_devirtualize_speculatively)
3286 274655 : || ie->speculative || !ie->maybe_hot_p ())
3287 188420 : return NULL;
3288 86235 : node = try_speculative_devirtualization (pii->otr_type, pii->otr_token,
3289 : context);
3290 86235 : if (node)
3291 : {
3292 653 : *speculative = true;
3293 653 : target = node->decl;
3294 : }
3295 : else
3296 : return NULL;
3297 : }
3298 : else
3299 : {
3300 10887 : *speculative = false;
3301 10887 : if (targets.length () == 1)
3302 10848 : target = targets[0]->decl;
3303 : else
3304 39 : target = ipa_impossible_devirt_target (ie, NULL_TREE);
3305 : }
3306 :
3307 11540 : if (target && !possible_polymorphic_call_target_p (ie,
3308 : cgraph_node::get (target)))
3309 : {
3310 54 : if (*speculative)
3311 : return NULL;
3312 40 : target = ipa_impossible_devirt_target (ie, target);
3313 : }
3314 :
3315 : return target;
3316 : }
3317 :
3318 : /* If an indirect edge IE can be turned into a direct one based on data in
3319 : AVALS, return the destination. Store into *SPECULATIVE a boolean determinig
3320 : whether the discovered target is only speculative guess. */
3321 :
3322 : tree
3323 1396545 : ipa_get_indirect_edge_target (struct cgraph_edge *ie,
3324 : ipa_call_arg_values *avals,
3325 : bool *speculative)
3326 : {
3327 1396545 : ipa_argagg_value_list avl (avals);
3328 1396545 : return ipa_get_indirect_edge_target_1 (ie, avals->m_known_vals,
3329 1396545 : avals->m_known_contexts,
3330 1396545 : avl, speculative);
3331 : }
3332 :
3333 : /* Calculate devirtualization time bonus for NODE, assuming we know information
3334 : about arguments stored in AVALS.
3335 :
3336 : FIXME: This function will also consider devirtualization of calls that are
3337 : known to be dead in the clone. */
3338 :
3339 : static sreal
3340 1530947 : devirtualization_time_bonus (struct cgraph_node *node,
3341 : ipa_auto_call_arg_values *avals)
3342 : {
3343 1530947 : struct cgraph_edge *ie;
3344 1530947 : sreal res = 0;
3345 :
3346 1698492 : for (ie = node->indirect_calls; ie; ie = ie->next_callee)
3347 : {
3348 167545 : struct cgraph_node *callee;
3349 167545 : class ipa_fn_summary *isummary;
3350 167545 : enum availability avail;
3351 167545 : tree target;
3352 167545 : bool speculative;
3353 :
3354 167545 : ipa_argagg_value_list avl (avals);
3355 167545 : target = ipa_get_indirect_edge_target_1 (ie, avals->m_known_vals,
3356 : avals->m_known_contexts,
3357 : avl, &speculative);
3358 167545 : if (!target)
3359 166563 : continue;
3360 :
3361 : /* Only bare minimum benefit for clearly un-inlineable targets. */
3362 3244 : res = res + ie->combined_sreal_frequency ();
3363 3244 : callee = cgraph_node::get (target);
3364 3244 : if (!callee || !callee->definition)
3365 636 : continue;
3366 2608 : callee = callee->function_symbol (&avail);
3367 2608 : if (avail < AVAIL_AVAILABLE)
3368 0 : continue;
3369 2608 : isummary = ipa_fn_summaries->get (callee);
3370 2608 : if (!isummary || !isummary->inlinable)
3371 65 : continue;
3372 :
3373 2543 : int savings = 0;
3374 2543 : int size = ipa_size_summaries->get (callee)->size;
3375 : /* FIXME: The values below need re-considering and perhaps also
3376 : integrating into the cost metrics, at lest in some very basic way. */
3377 2543 : int max_inline_insns_auto
3378 2543 : = opt_for_fn (callee->decl, param_max_inline_insns_auto);
3379 2543 : if (size <= max_inline_insns_auto / 4)
3380 402 : savings = 31 / ((int)speculative + 1);
3381 2141 : else if (size <= max_inline_insns_auto / 2)
3382 390 : savings = 15 / ((int)speculative + 1);
3383 3312 : else if (size <= max_inline_insns_auto
3384 1751 : || DECL_DECLARED_INLINE_P (callee->decl))
3385 190 : savings = 7 / ((int)speculative + 1);
3386 : else
3387 1561 : continue;
3388 982 : res = res + ie->combined_sreal_frequency () * (sreal) savings;
3389 : }
3390 :
3391 1530947 : return res;
3392 : }
3393 :
3394 : /* Return time bonus incurred because of hints stored in ESTIMATES. */
3395 :
3396 : static sreal
3397 287895 : hint_time_bonus (cgraph_node *node, const ipa_call_estimates &estimates)
3398 : {
3399 287895 : sreal result = 0;
3400 287895 : ipa_hints hints = estimates.hints;
3401 287895 : if (hints & (INLINE_HINT_loop_iterations | INLINE_HINT_loop_stride))
3402 27278 : result += opt_for_fn (node->decl, param_ipa_cp_loop_hint_bonus);
3403 :
3404 287895 : sreal bonus_for_one = opt_for_fn (node->decl, param_ipa_cp_loop_hint_bonus);
3405 :
3406 287895 : if (hints & INLINE_HINT_loop_iterations)
3407 19018 : result += estimates.loops_with_known_iterations * bonus_for_one;
3408 :
3409 287895 : if (hints & INLINE_HINT_loop_stride)
3410 11003 : result += estimates.loops_with_known_strides * bonus_for_one;
3411 :
3412 287895 : return result;
3413 : }
3414 :
3415 : /* If there is a reason to penalize the function described by INFO in the
3416 : cloning goodness evaluation, do so. */
3417 :
3418 : static inline sreal
3419 103210 : incorporate_penalties (cgraph_node *node, ipa_node_params *info,
3420 : sreal evaluation)
3421 : {
3422 103210 : if (info->node_within_scc && !info->node_is_self_scc)
3423 1744 : evaluation = (evaluation
3424 1744 : * (100 - opt_for_fn (node->decl,
3425 3488 : param_ipa_cp_recursion_penalty))) / 100;
3426 :
3427 103210 : if (info->node_calling_single_call)
3428 6869 : evaluation = (evaluation
3429 6869 : * (100 - opt_for_fn (node->decl,
3430 6869 : param_ipa_cp_single_call_penalty)))
3431 13738 : / 100;
3432 :
3433 103210 : return evaluation;
3434 : }
3435 :
3436 : /* Return true if cloning NODE is a good idea, given the estimated TIME_BENEFIT
3437 : and SIZE_COST and with the sum of frequencies of incoming edges to the
3438 : potential new clone in FREQUENCIES. CUR_SWEEP is the number of the current
3439 : sweep of IPA-CP over the call-graph in the decision stage. */
3440 :
3441 : static bool
3442 343303 : good_cloning_opportunity_p (struct cgraph_node *node, sreal time_benefit,
3443 : sreal freq_sum, profile_count count_sum,
3444 : int size_cost, bool called_without_ipa_profile,
3445 : int cur_sweep)
3446 : {
3447 343303 : gcc_assert (count_sum.ipa () == count_sum);
3448 343303 : if (count_sum.quality () == AFDO)
3449 0 : count_sum = count_sum.force_nonzero ();
3450 343303 : if (time_benefit == 0
3451 290397 : || !opt_for_fn (node->decl, flag_ipa_cp_clone)
3452 : /* If there is no call which was executed in profiling or where
3453 : profile is missing, we do not want to clone. */
3454 103302 : || (!called_without_ipa_profile && !count_sum.nonzero_p ()))
3455 : {
3456 240093 : if (dump_file && (dump_flags & TDF_DETAILS))
3457 24 : fprintf (dump_file, " good_cloning_opportunity_p (time: %g, "
3458 : "size: %i): Definitely not good or prohibited.\n",
3459 : time_benefit.to_double (), size_cost);
3460 240093 : return false;
3461 : }
3462 :
3463 103210 : gcc_assert (size_cost > 0);
3464 :
3465 103210 : ipa_node_params *info = ipa_node_params_sum->get (node);
3466 103210 : int num_sweeps = opt_for_fn (node->decl, param_ipa_cp_sweeps);
3467 103210 : int eval_threshold = opt_for_fn (node->decl, param_ipa_cp_eval_threshold);
3468 103210 : eval_threshold = (eval_threshold * num_sweeps) / cur_sweep;
3469 : /* If we know the execution IPA execution counts, we can estimate overall
3470 : speedup of the program. */
3471 103210 : if (count_sum.nonzero_p ())
3472 : {
3473 403 : profile_count saved_time = count_sum * time_benefit;
3474 403 : sreal evaluation = saved_time.to_sreal_scale (profile_count::one ())
3475 806 : / size_cost;
3476 403 : evaluation = incorporate_penalties (node, info, evaluation);
3477 :
3478 403 : if (dump_file && (dump_flags & TDF_DETAILS))
3479 : {
3480 0 : fprintf (dump_file, " good_cloning_opportunity_p (time: %g, "
3481 : "size: %i, count_sum: ", time_benefit.to_double (),
3482 : size_cost);
3483 0 : count_sum.dump (dump_file);
3484 0 : fprintf (dump_file, ", overall time saved: ");
3485 0 : saved_time.dump (dump_file);
3486 0 : fprintf (dump_file, "%s%s) -> evaluation: %.2f, threshold: %i\n",
3487 0 : info->node_within_scc
3488 0 : ? (info->node_is_self_scc ? ", self_scc" : ", scc") : "",
3489 0 : info->node_calling_single_call ? ", single_call" : "",
3490 : evaluation.to_double (), eval_threshold);
3491 : }
3492 403 : gcc_checking_assert (saved_time == saved_time.ipa ());
3493 403 : if (!maybe_hot_count_p (NULL, saved_time))
3494 : {
3495 27 : if (dump_file && (dump_flags & TDF_DETAILS))
3496 0 : fprintf (dump_file, " not cloning: time saved is not hot\n");
3497 : }
3498 : /* Evaulation approximately corresponds to time saved per instruction
3499 : introduced. This is likely almost always going to be true, since we
3500 : already checked that time saved is large enough to be considered
3501 : hot. */
3502 376 : else if (evaluation >= (sreal)eval_threshold)
3503 403 : return true;
3504 : /* If all call sites have profile known; we know we do not want t clone.
3505 : If there are calls with unknown profile; try local heuristics. */
3506 377 : if (!called_without_ipa_profile)
3507 : return false;
3508 : }
3509 102807 : sreal evaluation = (time_benefit * freq_sum) / size_cost;
3510 102807 : evaluation = incorporate_penalties (node, info, evaluation);
3511 102807 : evaluation *= 1000;
3512 :
3513 102807 : if (dump_file && (dump_flags & TDF_DETAILS))
3514 356 : fprintf (dump_file, " good_cloning_opportunity_p (time: %g, "
3515 : "size: %i, freq_sum: %g%s%s) -> evaluation: %.2f, "
3516 : "threshold: %i\n",
3517 : time_benefit.to_double (), size_cost, freq_sum.to_double (),
3518 178 : info->node_within_scc
3519 26 : ? (info->node_is_self_scc ? ", self_scc" : ", scc") : "",
3520 178 : info->node_calling_single_call ? ", single_call" : "",
3521 : evaluation.to_double (), eval_threshold);
3522 :
3523 102807 : return evaluation >= eval_threshold;
3524 : }
3525 :
3526 : /* Grow vectors in AVALS and fill them with information about values of
3527 : parameters that are known to be independent of the context. INFO describes
3528 : the function. If REMOVABLE_PARAMS_COST is non-NULL, the movement cost of
3529 : all removable parameters will be stored in it.
3530 :
3531 : TODO: Also grow context independent value range vectors. */
3532 :
3533 : static bool
3534 2179866 : gather_context_independent_values (class ipa_node_params *info,
3535 : ipa_auto_call_arg_values *avals,
3536 : int *removable_params_cost)
3537 : {
3538 2179866 : int i, count = ipa_get_param_count (info);
3539 2179866 : bool ret = false;
3540 :
3541 2179866 : avals->m_known_vals.safe_grow_cleared (count, true);
3542 2179866 : avals->m_known_contexts.safe_grow_cleared (count, true);
3543 :
3544 2179866 : if (removable_params_cost)
3545 2179866 : *removable_params_cost = 0;
3546 :
3547 7224574 : for (i = 0; i < count; i++)
3548 : {
3549 5044708 : class ipcp_param_lattices *plats = ipa_get_parm_lattices (info, i);
3550 5044708 : ipcp_lattice<tree> *lat = &plats->itself;
3551 :
3552 5044708 : if (lat->is_single_const ())
3553 : {
3554 34267 : ipcp_value<tree> *val = lat->values;
3555 34267 : gcc_checking_assert (TREE_CODE (val->value) != TREE_BINFO);
3556 34267 : avals->m_known_vals[i] = val->value;
3557 34267 : if (removable_params_cost)
3558 68534 : *removable_params_cost
3559 34267 : += estimate_move_cost (TREE_TYPE (val->value), false);
3560 : ret = true;
3561 : }
3562 5010441 : else if (removable_params_cost
3563 5010441 : && !ipa_is_param_used (info, i))
3564 969892 : *removable_params_cost
3565 484946 : += ipa_get_param_move_cost (info, i);
3566 :
3567 5044708 : if (!ipa_is_param_used (info, i))
3568 490081 : continue;
3569 :
3570 4554627 : ipcp_lattice<ipa_polymorphic_call_context> *ctxlat = &plats->ctxlat;
3571 : /* Do not account known context as reason for cloning. We can see
3572 : if it permits devirtualization. */
3573 4554627 : if (ctxlat->is_single_const ())
3574 23385 : avals->m_known_contexts[i] = ctxlat->values->value;
3575 :
3576 4554627 : ret |= push_agg_values_from_plats (plats, i, 0, &avals->m_known_aggs);
3577 : }
3578 :
3579 2179866 : return ret;
3580 : }
3581 :
3582 : /* Perform time and size measurement of NODE with the context given in AVALS,
3583 : calculate the benefit compared to the node without specialization and store
3584 : it into VAL. Take into account REMOVABLE_PARAMS_COST of all
3585 : context-independent or unused removable parameters and EST_MOVE_COST, the
3586 : estimated movement of the considered parameter. */
3587 :
3588 : static void
3589 78094 : perform_estimation_of_a_value (cgraph_node *node,
3590 : ipa_auto_call_arg_values *avals,
3591 : int removable_params_cost, int est_move_cost,
3592 : ipcp_value_base *val)
3593 : {
3594 78094 : sreal time_benefit;
3595 78094 : ipa_call_estimates estimates;
3596 :
3597 78094 : estimate_ipcp_clone_size_and_time (node, avals, &estimates);
3598 :
3599 : /* Extern inline functions have no cloning local time benefits because they
3600 : will be inlined anyway. The only reason to clone them is if it enables
3601 : optimization in any of the functions they call. */
3602 78094 : if (DECL_EXTERNAL (node->decl) && DECL_DECLARED_INLINE_P (node->decl))
3603 114 : time_benefit = 0;
3604 : else
3605 77980 : time_benefit = (estimates.nonspecialized_time - estimates.time)
3606 155960 : + hint_time_bonus (node, estimates)
3607 155960 : + (devirtualization_time_bonus (node, avals)
3608 155960 : + removable_params_cost + est_move_cost);
3609 :
3610 78094 : int size = estimates.size;
3611 78094 : gcc_checking_assert (size >=0);
3612 : /* The inliner-heuristics based estimates may think that in certain
3613 : contexts some functions do not have any size at all but we want
3614 : all specializations to have at least a tiny cost, not least not to
3615 : divide by zero. */
3616 78094 : if (size == 0)
3617 0 : size = 1;
3618 :
3619 78094 : val->local_time_benefit = time_benefit;
3620 78094 : val->local_size_cost = size;
3621 78094 : }
3622 :
3623 : /* Get the overall limit of growth based on parameters extracted from NODE. It
3624 : does not really make sense to mix functions with different overall growth
3625 : limits or even number of sweeps but it is possible and if it happens, we do
3626 : not want to select one limit at random, so get the limits from NODE. */
3627 :
3628 : static long
3629 211506 : get_max_overall_size (cgraph_node *node)
3630 : {
3631 211506 : long max_new_size = orig_overall_size;
3632 211506 : long large_unit = opt_for_fn (node->decl, param_ipa_cp_large_unit_insns);
3633 211506 : if (max_new_size < large_unit)
3634 : max_new_size = large_unit;
3635 211506 : int unit_growth = opt_for_fn (node->decl, param_ipa_cp_unit_growth);
3636 211506 : max_new_size += max_new_size * unit_growth / 100 + 1;
3637 :
3638 211506 : return max_new_size;
3639 : }
3640 :
3641 : /* Return true if NODE should be cloned just for a parameter removal, possibly
3642 : dumping a reason if not. */
3643 :
3644 : static bool
3645 193326 : clone_for_param_removal_p (cgraph_node *node)
3646 : {
3647 193326 : if (!node->can_change_signature)
3648 : {
3649 5017 : if (dump_file && (dump_flags & TDF_DETAILS))
3650 0 : fprintf (dump_file, " Not considering cloning to remove parameters, "
3651 : "function cannot change signature.\n");
3652 5017 : return false;
3653 : }
3654 188309 : if (node->can_be_local_p ())
3655 : {
3656 139218 : if (dump_file && (dump_flags & TDF_DETAILS))
3657 0 : fprintf (dump_file, " Not considering cloning to remove parameters, "
3658 : "IPA-SRA can do it potentially better.\n");
3659 139218 : return false;
3660 : }
3661 : return true;
3662 : }
3663 :
3664 : /* Iterate over known values of parameters of NODE and estimate the local
3665 : effects in terms of time and size they have. */
3666 :
3667 : static void
3668 1269638 : estimate_local_effects (struct cgraph_node *node)
3669 : {
3670 1269638 : ipa_node_params *info = ipa_node_params_sum->get (node);
3671 1269638 : int count = ipa_get_param_count (info);
3672 1040531 : int removable_params_cost;
3673 :
3674 1040531 : if (!count || !ipcp_versionable_function_p (node))
3675 396466 : return;
3676 :
3677 873172 : if (dump_file && (dump_flags & TDF_DETAILS))
3678 117 : fprintf (dump_file, "\nEstimating effects for %s.\n", node->dump_name ());
3679 :
3680 873172 : ipa_auto_call_arg_values avals;
3681 873172 : gather_context_independent_values (info, &avals, &removable_params_cost);
3682 :
3683 2912970 : for (int i = 0; i < count; i++)
3684 : {
3685 2039798 : class ipcp_param_lattices *plats = ipa_get_parm_lattices (info, i);
3686 2039798 : ipcp_lattice<tree> *lat = &plats->itself;
3687 2039798 : ipcp_value<tree> *val;
3688 :
3689 4058337 : if (lat->bottom
3690 211309 : || !lat->values
3691 2078267 : || avals.m_known_vals[i])
3692 2018539 : continue;
3693 :
3694 65158 : for (val = lat->values; val; val = val->next)
3695 : {
3696 43899 : gcc_checking_assert (TREE_CODE (val->value) != TREE_BINFO);
3697 43899 : avals.m_known_vals[i] = val->value;
3698 :
3699 43899 : int emc = estimate_move_cost (TREE_TYPE (val->value), true);
3700 43899 : perform_estimation_of_a_value (node, &avals, removable_params_cost,
3701 : emc, val);
3702 :
3703 43899 : if (dump_file && (dump_flags & TDF_DETAILS))
3704 : {
3705 44 : fprintf (dump_file, " - estimates for value ");
3706 44 : print_ipcp_constant_value (dump_file, val->value);
3707 44 : fprintf (dump_file, " for ");
3708 44 : ipa_dump_param (dump_file, info, i);
3709 44 : fprintf (dump_file, ": time_benefit: %g, size: %i\n",
3710 : val->local_time_benefit.to_double (),
3711 : val->local_size_cost);
3712 : }
3713 : }
3714 21259 : avals.m_known_vals[i] = NULL_TREE;
3715 : }
3716 :
3717 2912970 : for (int i = 0; i < count; i++)
3718 : {
3719 2039798 : class ipcp_param_lattices *plats = ipa_get_parm_lattices (info, i);
3720 :
3721 2039798 : if (!plats->virt_call)
3722 2032076 : continue;
3723 :
3724 7722 : ipcp_lattice<ipa_polymorphic_call_context> *ctxlat = &plats->ctxlat;
3725 7722 : ipcp_value<ipa_polymorphic_call_context> *val;
3726 :
3727 15286 : if (ctxlat->bottom
3728 2843 : || !ctxlat->values
3729 10559 : || !avals.m_known_contexts[i].useless_p ())
3730 7564 : continue;
3731 :
3732 384 : for (val = ctxlat->values; val; val = val->next)
3733 : {
3734 226 : avals.m_known_contexts[i] = val->value;
3735 226 : perform_estimation_of_a_value (node, &avals, removable_params_cost,
3736 : 0, val);
3737 :
3738 226 : if (dump_file && (dump_flags & TDF_DETAILS))
3739 : {
3740 0 : fprintf (dump_file, " - estimates for polymorphic context ");
3741 0 : print_ipcp_constant_value (dump_file, val->value);
3742 0 : fprintf (dump_file, " for ");
3743 0 : ipa_dump_param (dump_file, info, i);
3744 0 : fprintf (dump_file, ": time_benefit: %g, size: %i\n",
3745 : val->local_time_benefit.to_double (),
3746 : val->local_size_cost);
3747 : }
3748 : }
3749 158 : avals.m_known_contexts[i] = ipa_polymorphic_call_context ();
3750 : }
3751 :
3752 873172 : unsigned all_ctx_len = avals.m_known_aggs.length ();
3753 873172 : auto_vec<ipa_argagg_value, 32> all_ctx;
3754 873172 : all_ctx.reserve_exact (all_ctx_len);
3755 873172 : all_ctx.splice (avals.m_known_aggs);
3756 873172 : avals.m_known_aggs.safe_grow_cleared (all_ctx_len + 1);
3757 :
3758 873172 : unsigned j = 0;
3759 2912970 : for (int index = 0; index < count; index++)
3760 : {
3761 2039798 : class ipcp_param_lattices *plats = ipa_get_parm_lattices (info, index);
3762 :
3763 2039798 : if (plats->aggs_bottom || !plats->aggs)
3764 2020782 : continue;
3765 :
3766 72895 : for (ipcp_agg_lattice *aglat = plats->aggs; aglat; aglat = aglat->next)
3767 : {
3768 53879 : ipcp_value<tree> *val;
3769 53523 : if (aglat->bottom || !aglat->values
3770 : /* If the following is true, the one value is already part of all
3771 : context estimations. */
3772 100017 : || (!plats->aggs_contain_variable
3773 25056 : && aglat->is_single_const ()))
3774 28442 : continue;
3775 :
3776 25437 : unsigned unit_offset = aglat->offset / BITS_PER_UNIT;
3777 25437 : while (j < all_ctx_len
3778 33753 : && (all_ctx[j].index < index
3779 3300 : || (all_ctx[j].index == index
3780 2333 : && all_ctx[j].unit_offset < unit_offset)))
3781 : {
3782 3209 : avals.m_known_aggs[j] = all_ctx[j];
3783 3209 : j++;
3784 : }
3785 :
3786 34484 : for (unsigned k = j; k < all_ctx_len; k++)
3787 9047 : avals.m_known_aggs[k+1] = all_ctx[k];
3788 :
3789 59406 : for (val = aglat->values; val; val = val->next)
3790 : {
3791 33969 : avals.m_known_aggs[j].value = val->value;
3792 33969 : avals.m_known_aggs[j].unit_offset = unit_offset;
3793 33969 : avals.m_known_aggs[j].index = index;
3794 33969 : avals.m_known_aggs[j].by_ref = plats->aggs_by_ref;
3795 33969 : avals.m_known_aggs[j].killed = false;
3796 :
3797 33969 : perform_estimation_of_a_value (node, &avals,
3798 : removable_params_cost, 0, val);
3799 :
3800 33969 : if (dump_file && (dump_flags & TDF_DETAILS))
3801 : {
3802 79 : fprintf (dump_file, " - estimates for value ");
3803 79 : print_ipcp_constant_value (dump_file, val->value);
3804 79 : fprintf (dump_file, " for ");
3805 79 : ipa_dump_param (dump_file, info, index);
3806 158 : fprintf (dump_file, "[%soffset: " HOST_WIDE_INT_PRINT_DEC
3807 : "]: time_benefit: %g, size: %i\n",
3808 79 : plats->aggs_by_ref ? "ref " : "",
3809 : aglat->offset,
3810 : val->local_time_benefit.to_double (),
3811 : val->local_size_cost);
3812 : }
3813 : }
3814 : }
3815 : }
3816 873172 : }
3817 :
3818 :
3819 : /* Add value CUR_VAL and all yet-unsorted values it is dependent on to the
3820 : topological sort of values. */
3821 :
3822 : template <typename valtype>
3823 : void
3824 136141 : value_topo_info<valtype>::add_val (ipcp_value<valtype> *cur_val)
3825 : {
3826 : ipcp_value_source<valtype> *src;
3827 :
3828 136141 : if (cur_val->dfs)
3829 : return;
3830 :
3831 135979 : dfs_counter++;
3832 135979 : cur_val->dfs = dfs_counter;
3833 135979 : cur_val->low_link = dfs_counter;
3834 :
3835 135979 : cur_val->topo_next = stack;
3836 135979 : stack = cur_val;
3837 135979 : cur_val->on_stack = true;
3838 :
3839 592644 : for (src = cur_val->sources; src; src = src->next)
3840 456665 : if (src->val)
3841 : {
3842 20947 : if (src->val->dfs == 0)
3843 : {
3844 186 : add_val (src->val);
3845 186 : if (src->val->low_link < cur_val->low_link)
3846 19 : cur_val->low_link = src->val->low_link;
3847 : }
3848 20761 : else if (src->val->on_stack
3849 1587 : && src->val->dfs < cur_val->low_link)
3850 73 : cur_val->low_link = src->val->dfs;
3851 : }
3852 :
3853 135979 : if (cur_val->dfs == cur_val->low_link)
3854 : {
3855 : ipcp_value<valtype> *v, *scc_list = NULL;
3856 :
3857 : do
3858 : {
3859 135979 : v = stack;
3860 135979 : stack = v->topo_next;
3861 135979 : v->on_stack = false;
3862 135979 : v->scc_no = cur_val->dfs;
3863 :
3864 135979 : v->scc_next = scc_list;
3865 135979 : scc_list = v;
3866 : }
3867 135979 : while (v != cur_val);
3868 :
3869 135891 : cur_val->topo_next = values_topo;
3870 135891 : values_topo = cur_val;
3871 : }
3872 : }
3873 :
3874 : /* Add all values in lattices associated with NODE to the topological sort if
3875 : they are not there yet. */
3876 :
3877 : static void
3878 1269638 : add_all_node_vals_to_toposort (cgraph_node *node, ipa_topo_info *topo)
3879 : {
3880 1269638 : ipa_node_params *info = ipa_node_params_sum->get (node);
3881 1269638 : int i, count = ipa_get_param_count (info);
3882 :
3883 3603430 : for (i = 0; i < count; i++)
3884 : {
3885 2333792 : class ipcp_param_lattices *plats = ipa_get_parm_lattices (info, i);
3886 2333792 : ipcp_lattice<tree> *lat = &plats->itself;
3887 2333792 : struct ipcp_agg_lattice *aglat;
3888 :
3889 2333792 : if (!lat->bottom)
3890 : {
3891 222017 : ipcp_value<tree> *val;
3892 293903 : for (val = lat->values; val; val = val->next)
3893 71886 : topo->constants.add_val (val);
3894 : }
3895 :
3896 2333792 : if (!plats->aggs_bottom)
3897 278334 : for (aglat = plats->aggs; aglat; aglat = aglat->next)
3898 56384 : if (!aglat->bottom)
3899 : {
3900 56028 : ipcp_value<tree> *val;
3901 112386 : for (val = aglat->values; val; val = val->next)
3902 56358 : topo->constants.add_val (val);
3903 : }
3904 :
3905 2333792 : ipcp_lattice<ipa_polymorphic_call_context> *ctxlat = &plats->ctxlat;
3906 2333792 : if (!ctxlat->bottom)
3907 : {
3908 223019 : ipcp_value<ipa_polymorphic_call_context> *ctxval;
3909 230730 : for (ctxval = ctxlat->values; ctxval; ctxval = ctxval->next)
3910 7711 : topo->contexts.add_val (ctxval);
3911 : }
3912 : }
3913 1269638 : }
3914 :
3915 : /* One pass of constants propagation along the call graph edges, from callers
3916 : to callees (requires topological ordering in TOPO), iterate over strongly
3917 : connected components. */
3918 :
3919 : static void
3920 129027 : propagate_constants_topo (class ipa_topo_info *topo)
3921 : {
3922 129027 : int i;
3923 :
3924 1479629 : for (i = topo->nnodes - 1; i >= 0; i--)
3925 : {
3926 1350602 : unsigned j;
3927 1350602 : struct cgraph_node *v, *node = topo->order[i];
3928 1350602 : vec<cgraph_node *> cycle_nodes = ipa_get_nodes_in_cycle (node);
3929 :
3930 : /* First, iteratively propagate within the strongly connected component
3931 : until all lattices stabilize. */
3932 2706058 : FOR_EACH_VEC_ELT (cycle_nodes, j, v)
3933 1355456 : if (v->has_gimple_body_p ())
3934 : {
3935 1278480 : if (opt_for_fn (v->decl, flag_ipa_cp)
3936 1278480 : && opt_for_fn (v->decl, optimize))
3937 1269638 : push_node_to_stack (topo, v);
3938 : /* When V is not optimized, we can not push it to stack, but
3939 : still we need to set all its callees lattices to bottom. */
3940 : else
3941 : {
3942 21855 : for (cgraph_edge *cs = v->callees; cs; cs = cs->next_callee)
3943 13013 : propagate_constants_across_call (cs);
3944 : }
3945 : }
3946 :
3947 1350602 : v = pop_node_from_stack (topo);
3948 3974026 : while (v)
3949 : {
3950 1272822 : struct cgraph_edge *cs;
3951 1272822 : class ipa_node_params *info = NULL;
3952 1272822 : bool self_scc = true;
3953 :
3954 6561314 : for (cs = v->callees; cs; cs = cs->next_callee)
3955 5288492 : if (ipa_edge_within_scc (cs))
3956 : {
3957 29596 : cgraph_node *callee = cs->callee->function_symbol ();
3958 :
3959 29596 : if (v != callee)
3960 17987 : self_scc = false;
3961 :
3962 29596 : if (!info)
3963 : {
3964 13888 : info = ipa_node_params_sum->get (v);
3965 13888 : info->node_within_scc = true;
3966 : }
3967 :
3968 29596 : if (propagate_constants_across_call (cs))
3969 4153 : push_node_to_stack (topo, callee);
3970 : }
3971 :
3972 1272822 : if (info)
3973 13888 : info->node_is_self_scc = self_scc;
3974 :
3975 1272822 : v = pop_node_from_stack (topo);
3976 : }
3977 :
3978 : /* Afterwards, propagate along edges leading out of the SCC, calculates
3979 : the local effects of the discovered constants and all valid values to
3980 : their topological sort. */
3981 2706058 : FOR_EACH_VEC_ELT (cycle_nodes, j, v)
3982 1355456 : if (v->has_gimple_body_p ()
3983 1278480 : && opt_for_fn (v->decl, flag_ipa_cp)
3984 2625094 : && opt_for_fn (v->decl, optimize))
3985 : {
3986 1269638 : struct cgraph_edge *cs;
3987 :
3988 1269638 : estimate_local_effects (v);
3989 1269638 : add_all_node_vals_to_toposort (v, topo);
3990 6524571 : for (cs = v->callees; cs; cs = cs->next_callee)
3991 5254933 : if (!ipa_edge_within_scc (cs))
3992 5232705 : propagate_constants_across_call (cs);
3993 : }
3994 1350602 : cycle_nodes.release ();
3995 : }
3996 129027 : }
3997 :
3998 : /* Propagate the estimated effects of individual values along the topological
3999 : from the dependent values to those they depend on. */
4000 :
4001 : template <typename valtype>
4002 : void
4003 258054 : value_topo_info<valtype>::propagate_effects ()
4004 : {
4005 : ipcp_value<valtype> *base;
4006 258054 : hash_set<ipcp_value<valtype> *> processed_srcvals;
4007 :
4008 393945 : for (base = values_topo; base; base = base->topo_next)
4009 : {
4010 : ipcp_value_source<valtype> *src;
4011 : ipcp_value<valtype> *val;
4012 135891 : sreal time = 0;
4013 135891 : HOST_WIDE_INT size = 0;
4014 :
4015 271870 : for (val = base; val; val = val->scc_next)
4016 : {
4017 135979 : time = time + val->local_time_benefit + val->prop_time_benefit;
4018 135979 : size = size + val->local_size_cost + val->prop_size_cost;
4019 : }
4020 :
4021 271870 : for (val = base; val; val = val->scc_next)
4022 : {
4023 135979 : processed_srcvals.empty ();
4024 592644 : for (src = val->sources; src; src = src->next)
4025 456665 : if (src->val
4026 456665 : && cs_interesting_for_ipcp_p (src->cs))
4027 : {
4028 20907 : if (!processed_srcvals.add (src->val))
4029 : {
4030 16740 : HOST_WIDE_INT prop_size = size + src->val->prop_size_cost;
4031 16740 : if (prop_size < INT_MAX)
4032 16740 : src->val->prop_size_cost = prop_size;
4033 : else
4034 0 : continue;
4035 : }
4036 :
4037 20907 : int special_factor = 1;
4038 20907 : if (val->same_scc (src->val))
4039 : special_factor
4040 1675 : = opt_for_fn(src->cs->caller->decl,
4041 : param_ipa_cp_recursive_freq_factor);
4042 19232 : else if (val->self_recursion_generated_p ()
4043 19232 : && (src->cs->callee->function_symbol ()
4044 822 : == src->cs->caller))
4045 : {
4046 822 : int max_recur_gen_depth
4047 822 : = opt_for_fn(src->cs->caller->decl,
4048 : param_ipa_cp_max_recursive_depth);
4049 822 : special_factor = max_recur_gen_depth
4050 822 : - val->self_recursion_generated_level + 1;
4051 : }
4052 :
4053 20907 : src->val->prop_time_benefit
4054 41814 : += time * special_factor * src->cs->sreal_frequency ();
4055 : }
4056 :
4057 135979 : if (size < INT_MAX)
4058 : {
4059 135979 : val->prop_time_benefit = time;
4060 135979 : val->prop_size_cost = size;
4061 : }
4062 : else
4063 : {
4064 0 : val->prop_time_benefit = 0;
4065 0 : val->prop_size_cost = 0;
4066 : }
4067 : }
4068 : }
4069 258054 : }
4070 :
4071 :
4072 : /* Propagate constants, polymorphic contexts and their effects from the
4073 : summaries interprocedurally. */
4074 :
4075 : static void
4076 129027 : ipcp_propagate_stage (class ipa_topo_info *topo)
4077 : {
4078 129027 : struct cgraph_node *node;
4079 :
4080 129027 : if (dump_file)
4081 161 : fprintf (dump_file, "\n Propagating constants:\n\n");
4082 :
4083 1484487 : FOR_EACH_DEFINED_FUNCTION (node)
4084 : {
4085 1355460 : if (node->has_gimple_body_p ()
4086 1278480 : && opt_for_fn (node->decl, flag_ipa_cp)
4087 2625098 : && opt_for_fn (node->decl, optimize))
4088 : {
4089 1269638 : ipa_node_params *info = ipa_node_params_sum->get (node);
4090 1269638 : determine_versionability (node, info);
4091 :
4092 1269638 : unsigned nlattices = ipa_get_param_count (info);
4093 1269638 : info->lattices.safe_grow_cleared (nlattices, true);
4094 1269638 : initialize_node_lattices (node);
4095 :
4096 1269638 : int num_sweeps = opt_for_fn (node->decl, param_ipa_cp_sweeps);
4097 1269638 : if (max_number_sweeps < num_sweeps)
4098 121228 : max_number_sweeps = num_sweeps;
4099 : }
4100 1355460 : ipa_size_summary *s = ipa_size_summaries->get (node);
4101 1355460 : if (node->definition && !node->alias && s != NULL)
4102 1279413 : overall_size += s->self_size;
4103 : }
4104 :
4105 129027 : orig_overall_size = overall_size;
4106 :
4107 129027 : if (dump_file)
4108 161 : fprintf (dump_file, "\noverall_size: %li\n", overall_size);
4109 :
4110 129027 : propagate_constants_topo (topo);
4111 129027 : if (flag_checking)
4112 129019 : ipcp_verify_propagated_values ();
4113 129027 : topo->constants.propagate_effects ();
4114 129027 : topo->contexts.propagate_effects ();
4115 :
4116 129027 : if (dump_file)
4117 : {
4118 161 : fprintf (dump_file, "\nIPA lattices after all propagation:\n");
4119 161 : print_all_lattices (dump_file, (dump_flags & TDF_DETAILS), true);
4120 : }
4121 129027 : }
4122 :
4123 : /* Discover newly direct outgoing edges from NODE which is a new clone with
4124 : known KNOWN_CSTS and make them direct. */
4125 :
4126 : static void
4127 21316 : ipcp_discover_new_direct_edges (struct cgraph_node *node,
4128 : vec<tree> known_csts,
4129 : vec<ipa_polymorphic_call_context>
4130 : known_contexts,
4131 : vec<ipa_argagg_value, va_gc> *aggvals)
4132 : {
4133 21316 : struct cgraph_edge *ie, *next_ie;
4134 21316 : bool found = false;
4135 :
4136 23201 : for (ie = node->indirect_calls; ie; ie = next_ie)
4137 : {
4138 1885 : tree target;
4139 1885 : bool speculative;
4140 :
4141 1885 : next_ie = ie->next_callee;
4142 1885 : ipa_argagg_value_list avs (aggvals);
4143 1885 : target = ipa_get_indirect_edge_target_1 (ie, known_csts, known_contexts,
4144 : avs, &speculative);
4145 1885 : if (target)
4146 : {
4147 555 : cgraph_polymorphic_indirect_info *pii
4148 555 : = dyn_cast <cgraph_polymorphic_indirect_info *> (ie->indirect_info);
4149 555 : cgraph_simple_indirect_info *sii
4150 1043 : = dyn_cast <cgraph_simple_indirect_info *> (ie->indirect_info);
4151 416 : bool agg_contents = sii && sii->agg_contents;
4152 555 : bool polymorphic = !!pii;
4153 555 : int param_index = ie->indirect_info->param_index;
4154 555 : struct cgraph_edge *cs = ipa_make_edge_direct_to_target (ie, target,
4155 : speculative);
4156 555 : found = true;
4157 :
4158 555 : if (cs && !agg_contents && !polymorphic)
4159 : {
4160 349 : ipa_node_params *info = ipa_node_params_sum->get (node);
4161 349 : int c = ipa_get_controlled_uses (info, param_index);
4162 349 : if (c != IPA_UNDESCRIBED_USE
4163 349 : && !ipa_get_param_load_dereferenced (info, param_index))
4164 : {
4165 345 : struct ipa_ref *to_del;
4166 :
4167 345 : c--;
4168 345 : ipa_set_controlled_uses (info, param_index, c);
4169 345 : if (dump_file && (dump_flags & TDF_DETAILS))
4170 3 : fprintf (dump_file, " controlled uses count of param "
4171 : "%i bumped down to %i\n", param_index, c);
4172 345 : if (c == 0
4173 345 : && (to_del = node->find_reference (cs->callee, NULL, 0,
4174 : IPA_REF_ADDR)))
4175 : {
4176 281 : if (dump_file && (dump_flags & TDF_DETAILS))
4177 3 : fprintf (dump_file, " and even removing its "
4178 : "cloning-created reference\n");
4179 281 : to_del->remove_reference ();
4180 : }
4181 : }
4182 : }
4183 : }
4184 : }
4185 : /* Turning calls to direct calls will improve overall summary. */
4186 21316 : if (found)
4187 462 : ipa_update_overall_fn_summary (node);
4188 21316 : }
4189 :
4190 : class edge_clone_summary;
4191 : static call_summary <edge_clone_summary *> *edge_clone_summaries = NULL;
4192 :
4193 : /* Edge clone summary. */
4194 :
4195 : class edge_clone_summary
4196 : {
4197 : public:
4198 : /* Default constructor. */
4199 377360 : edge_clone_summary (): prev_clone (NULL), next_clone (NULL) {}
4200 :
4201 : /* Default destructor. */
4202 377360 : ~edge_clone_summary ()
4203 : {
4204 377360 : if (prev_clone)
4205 34243 : edge_clone_summaries->get (prev_clone)->next_clone = next_clone;
4206 377360 : if (next_clone)
4207 157698 : edge_clone_summaries->get (next_clone)->prev_clone = prev_clone;
4208 377360 : }
4209 :
4210 : cgraph_edge *prev_clone;
4211 : cgraph_edge *next_clone;
4212 : };
4213 :
4214 : class edge_clone_summary_t:
4215 : public call_summary <edge_clone_summary *>
4216 : {
4217 : public:
4218 129027 : edge_clone_summary_t (symbol_table *symtab):
4219 258054 : call_summary <edge_clone_summary *> (symtab)
4220 : {
4221 129027 : m_initialize_when_cloning = true;
4222 : }
4223 :
4224 : void duplicate (cgraph_edge *src_edge, cgraph_edge *dst_edge,
4225 : edge_clone_summary *src_data,
4226 : edge_clone_summary *dst_data) final override;
4227 : };
4228 :
4229 : /* Edge duplication hook. */
4230 :
4231 : void
4232 191328 : edge_clone_summary_t::duplicate (cgraph_edge *src_edge, cgraph_edge *dst_edge,
4233 : edge_clone_summary *src_data,
4234 : edge_clone_summary *dst_data)
4235 : {
4236 191328 : if (src_data->next_clone)
4237 5287 : edge_clone_summaries->get (src_data->next_clone)->prev_clone = dst_edge;
4238 191328 : dst_data->prev_clone = src_edge;
4239 191328 : dst_data->next_clone = src_data->next_clone;
4240 191328 : src_data->next_clone = dst_edge;
4241 191328 : }
4242 :
4243 : /* Return true is CS calls DEST or its clone for all contexts. When
4244 : ALLOW_RECURSION_TO_CLONE is false, also return false for self-recursive
4245 : edges from/to an all-context clone. */
4246 :
4247 : static bool
4248 1812843 : calls_same_node_or_its_all_contexts_clone_p (cgraph_edge *cs, cgraph_node *dest,
4249 : bool allow_recursion_to_clone)
4250 : {
4251 1812843 : enum availability availability;
4252 1812843 : cgraph_node *callee = cs->callee->function_symbol (&availability);
4253 :
4254 1812843 : if (availability <= AVAIL_INTERPOSABLE)
4255 : return false;
4256 1806976 : if (callee == dest)
4257 : return true;
4258 617279 : if (!allow_recursion_to_clone && cs->caller == callee)
4259 : return false;
4260 :
4261 617122 : ipa_node_params *info = ipa_node_params_sum->get (callee);
4262 617122 : return info->is_all_contexts_clone && info->ipcp_orig_node == dest;
4263 : }
4264 :
4265 : /* Return true if edge CS does bring about the value described by SRC to
4266 : DEST_VAL of node DEST or its clone for all contexts. */
4267 :
4268 : static bool
4269 1802873 : cgraph_edge_brings_value_p (cgraph_edge *cs, ipcp_value_source<tree> *src,
4270 : cgraph_node *dest, ipcp_value<tree> *dest_val)
4271 : {
4272 1802873 : ipa_node_params *caller_info = ipa_node_params_sum->get (cs->caller);
4273 :
4274 1802873 : if (!calls_same_node_or_its_all_contexts_clone_p (cs, dest, !src->val)
4275 1802873 : || caller_info->node_dead)
4276 : return false;
4277 :
4278 740157 : if (!src->val)
4279 : return true;
4280 :
4281 61193 : if (caller_info->ipcp_orig_node)
4282 : {
4283 19086 : tree t = NULL_TREE;
4284 19086 : if (src->offset == -1)
4285 13618 : t = caller_info->known_csts[src->index];
4286 5468 : else if (ipcp_transformation *ts
4287 5468 : = ipcp_get_transformation_summary (cs->caller))
4288 : {
4289 5468 : ipa_argagg_value_list avl (ts);
4290 5468 : t = avl.get_value (src->index, src->offset / BITS_PER_UNIT);
4291 : }
4292 19086 : return (t != NULL_TREE
4293 19086 : && values_equal_for_ipcp_p (src->val->value, t));
4294 : }
4295 : else
4296 : {
4297 42107 : if (src->val == dest_val)
4298 : return true;
4299 :
4300 36389 : struct ipcp_agg_lattice *aglat;
4301 36389 : class ipcp_param_lattices *plats = ipa_get_parm_lattices (caller_info,
4302 : src->index);
4303 36389 : if (src->offset == -1)
4304 26895 : return (plats->itself.is_single_const ()
4305 20 : && values_equal_for_ipcp_p (src->val->value,
4306 20 : plats->itself.values->value));
4307 : else
4308 : {
4309 9494 : if (plats->aggs_bottom || plats->aggs_contain_variable)
4310 : return false;
4311 3867 : for (aglat = plats->aggs; aglat; aglat = aglat->next)
4312 3867 : if (aglat->offset == src->offset)
4313 1739 : return (aglat->is_single_const ()
4314 8 : && values_equal_for_ipcp_p (src->val->value,
4315 8 : aglat->values->value));
4316 : }
4317 : return false;
4318 : }
4319 : }
4320 :
4321 : /* Return true if edge CS does bring about the value described by SRC to
4322 : DST_VAL of node DEST or its clone for all contexts. */
4323 :
4324 : static bool
4325 9970 : cgraph_edge_brings_value_p (cgraph_edge *cs,
4326 : ipcp_value_source<ipa_polymorphic_call_context> *src,
4327 : cgraph_node *dest,
4328 : ipcp_value<ipa_polymorphic_call_context> *)
4329 : {
4330 9970 : ipa_node_params *caller_info = ipa_node_params_sum->get (cs->caller);
4331 :
4332 9970 : if (!calls_same_node_or_its_all_contexts_clone_p (cs, dest, true)
4333 9970 : || caller_info->node_dead)
4334 : return false;
4335 8987 : if (!src->val)
4336 : return true;
4337 :
4338 1744 : if (caller_info->ipcp_orig_node)
4339 2475 : return (caller_info->known_contexts.length () > (unsigned) src->index)
4340 484 : && values_equal_for_ipcp_p (src->val->value,
4341 242 : caller_info->known_contexts[src->index]);
4342 :
4343 1487 : class ipcp_param_lattices *plats = ipa_get_parm_lattices (caller_info,
4344 : src->index);
4345 1487 : return plats->ctxlat.is_single_const ()
4346 270 : && values_equal_for_ipcp_p (src->val->value,
4347 270 : plats->ctxlat.values->value);
4348 : }
4349 :
4350 : /* Get the next clone in the linked list of clones of an edge. */
4351 :
4352 : static inline struct cgraph_edge *
4353 1813132 : get_next_cgraph_edge_clone (struct cgraph_edge *cs)
4354 : {
4355 1813132 : edge_clone_summary *s = edge_clone_summaries->get (cs);
4356 1813132 : return s != NULL ? s->next_clone : NULL;
4357 : }
4358 :
4359 : /* Given VAL that is intended for DEST, iterate over all its sources and if any
4360 : of them is viable and hot, return true. In that case, for those that still
4361 : hold, add their edge frequency and their number and cumulative profile
4362 : counts of self-ecursive and other edges into *FREQUENCY, *CALLER_COUNT,
4363 : REC_COUNT_SUM and NONREC_COUNT_SUM respectively. */
4364 :
4365 : template <typename valtype>
4366 : static bool
4367 210863 : get_info_about_necessary_edges (ipcp_value<valtype> *val, cgraph_node *dest,
4368 : sreal *freq_sum, int *caller_count,
4369 : profile_count *rec_count_sum,
4370 : profile_count *nonrec_count_sum,
4371 : bool *called_without_ipa_profile)
4372 : {
4373 : ipcp_value_source<valtype> *src;
4374 210863 : sreal freq = 0;
4375 210863 : int count = 0;
4376 210863 : profile_count rec_cnt = profile_count::zero ();
4377 210863 : profile_count nonrec_cnt = profile_count::zero ();
4378 210863 : bool interesting = false;
4379 210863 : bool non_self_recursive = false;
4380 210863 : *called_without_ipa_profile = false;
4381 :
4382 967229 : for (src = val->sources; src; src = src->next)
4383 : {
4384 756366 : struct cgraph_edge *cs = src->cs;
4385 1876702 : while (cs)
4386 : {
4387 1120336 : if (cgraph_edge_brings_value_p (cs, src, dest, val))
4388 : {
4389 352552 : count++;
4390 352552 : freq += cs->sreal_frequency ();
4391 352552 : interesting |= cs_interesting_for_ipcp_p (cs);
4392 352552 : if (cs->caller != dest)
4393 : {
4394 345741 : non_self_recursive = true;
4395 345741 : if (cs->count.ipa ().initialized_p ())
4396 953 : rec_cnt += cs->count.ipa ();
4397 : else
4398 344788 : *called_without_ipa_profile = true;
4399 : }
4400 6811 : else if (cs->count.ipa ().initialized_p ())
4401 0 : nonrec_cnt += cs->count.ipa ();
4402 : else
4403 6811 : *called_without_ipa_profile = true;
4404 : }
4405 1120336 : cs = get_next_cgraph_edge_clone (cs);
4406 : }
4407 : }
4408 :
4409 : /* If the only edges bringing a value are self-recursive ones, do not bother
4410 : evaluating it. */
4411 210863 : if (!non_self_recursive)
4412 : return false;
4413 :
4414 150641 : *freq_sum = freq;
4415 150641 : *caller_count = count;
4416 150641 : *rec_count_sum = rec_cnt;
4417 150641 : *nonrec_count_sum = nonrec_cnt;
4418 :
4419 150641 : return interesting;
4420 : }
4421 :
4422 : /* Given a NODE, and a set of its CALLERS, try to adjust order of the callers
4423 : to let a non-self-recursive caller be the first element. Thus, we can
4424 : simplify intersecting operations on values that arrive from all of these
4425 : callers, especially when there exists self-recursive call. Return true if
4426 : this kind of adjustment is possible. */
4427 :
4428 : static bool
4429 57481 : adjust_callers_for_value_intersection (vec<cgraph_edge *> &callers,
4430 : cgraph_node *node)
4431 : {
4432 61694 : for (unsigned i = 0; i < callers.length (); i++)
4433 : {
4434 61598 : cgraph_edge *cs = callers[i];
4435 :
4436 61598 : if (cs->caller != node)
4437 : {
4438 57385 : if (i > 0)
4439 : {
4440 1967 : callers[i] = callers[0];
4441 1967 : callers[0] = cs;
4442 : }
4443 57385 : return true;
4444 : }
4445 : }
4446 : return false;
4447 : }
4448 :
4449 : /* Return a vector of incoming edges that do bring value VAL to node DEST. It
4450 : is assumed their number is known and equal to CALLER_COUNT. */
4451 :
4452 : template <typename valtype>
4453 : static auto_vec<cgraph_edge *>
4454 150288 : gather_edges_for_value (ipcp_value<valtype> *val, cgraph_node *dest,
4455 : int caller_count)
4456 : {
4457 : ipcp_value_source<valtype> *src;
4458 150288 : auto_vec<cgraph_edge *> ret (caller_count);
4459 :
4460 515954 : for (src = val->sources; src; src = src->next)
4461 : {
4462 365666 : struct cgraph_edge *cs = src->cs;
4463 818265 : while (cs)
4464 : {
4465 452599 : if (cgraph_edge_brings_value_p (cs, src, dest, val))
4466 349461 : ret.quick_push (cs);
4467 452599 : cs = get_next_cgraph_edge_clone (cs);
4468 : }
4469 : }
4470 :
4471 150288 : if (caller_count > 1)
4472 40388 : adjust_callers_for_value_intersection (ret, dest);
4473 :
4474 150288 : return ret;
4475 : }
4476 :
4477 : /* Construct a replacement map for a know VALUE for a formal parameter PARAM.
4478 : Return it or NULL if for some reason it cannot be created. FORCE_LOAD_REF
4479 : should be set to true when the reference created for the constant should be
4480 : a load one and not an address one because the corresponding parameter p is
4481 : only used as *p. */
4482 :
4483 : static struct ipa_replace_map *
4484 24293 : get_replacement_map (class ipa_node_params *info, tree value, int parm_num,
4485 : bool force_load_ref)
4486 : {
4487 24293 : struct ipa_replace_map *replace_map;
4488 :
4489 24293 : replace_map = ggc_alloc<ipa_replace_map> ();
4490 24293 : if (dump_file)
4491 : {
4492 171 : fprintf (dump_file, " replacing ");
4493 171 : ipa_dump_param (dump_file, info, parm_num);
4494 :
4495 171 : fprintf (dump_file, " with const ");
4496 171 : print_generic_expr (dump_file, value);
4497 :
4498 171 : if (force_load_ref)
4499 11 : fprintf (dump_file, " - forcing load reference\n");
4500 : else
4501 160 : fprintf (dump_file, "\n");
4502 : }
4503 24293 : replace_map->parm_num = parm_num;
4504 24293 : replace_map->new_tree = value;
4505 24293 : replace_map->force_load_ref = force_load_ref;
4506 24293 : return replace_map;
4507 : }
4508 :
4509 : /* Dump new profiling counts of NODE. SPEC is true when NODE is a specialzied
4510 : one, otherwise it will be referred to as the original node. */
4511 :
4512 : static void
4513 4 : dump_profile_updates (cgraph_node *node, bool spec)
4514 : {
4515 4 : if (spec)
4516 2 : fprintf (dump_file, " setting count of the specialized node %s to ",
4517 : node->dump_name ());
4518 : else
4519 2 : fprintf (dump_file, " setting count of the original node %s to ",
4520 : node->dump_name ());
4521 :
4522 4 : node->count.dump (dump_file);
4523 4 : fprintf (dump_file, "\n");
4524 6 : for (cgraph_edge *cs = node->callees; cs; cs = cs->next_callee)
4525 : {
4526 2 : fprintf (dump_file, " edge to %s has count ",
4527 2 : cs->callee->dump_name ());
4528 2 : cs->count.dump (dump_file);
4529 2 : fprintf (dump_file, "\n");
4530 : }
4531 4 : }
4532 :
4533 : /* With partial train run we do not want to assume that original's count is
4534 : zero whenever we redurect all executed edges to clone. Simply drop profile
4535 : to local one in this case. In eany case, return the new value. ORIG_NODE
4536 : is the original node and its count has not been updaed yet. */
4537 :
4538 : profile_count
4539 16 : lenient_count_portion_handling (profile_count remainder, cgraph_node *orig_node)
4540 : {
4541 32 : if (remainder.ipa_p () && !remainder.ipa ().nonzero_p ()
4542 24 : && orig_node->count.ipa_p () && orig_node->count.ipa ().nonzero_p ()
4543 4 : && opt_for_fn (orig_node->decl, flag_profile_partial_training))
4544 0 : remainder = orig_node->count.guessed_local ();
4545 :
4546 16 : return remainder;
4547 : }
4548 :
4549 : /* Structure to sum counts coming from nodes other than the original node and
4550 : its clones. */
4551 :
4552 : struct gather_other_count_struct
4553 : {
4554 : cgraph_node *orig;
4555 : profile_count other_count;
4556 : };
4557 :
4558 : /* Worker callback of call_for_symbol_thunks_and_aliases summing the number of
4559 : counts that come from non-self-recursive calls.. */
4560 :
4561 : static bool
4562 8 : gather_count_of_non_rec_edges (cgraph_node *node, void *data)
4563 : {
4564 8 : gather_other_count_struct *desc = (gather_other_count_struct *) data;
4565 20 : for (cgraph_edge *cs = node->callers; cs; cs = cs->next_caller)
4566 12 : if (cs->caller != desc->orig && cs->caller->clone_of != desc->orig)
4567 0 : if (cs->count.ipa ().initialized_p ())
4568 0 : desc->other_count += cs->count.ipa ();
4569 8 : return false;
4570 : }
4571 :
4572 : /* Structure to help analyze if we need to boost counts of some clones of some
4573 : non-recursive edges to match the new callee count. */
4574 :
4575 : struct desc_incoming_count_struct
4576 : {
4577 : cgraph_node *orig;
4578 : hash_set <cgraph_edge *> *processed_edges;
4579 : profile_count count;
4580 : unsigned unproc_orig_rec_edges;
4581 : };
4582 :
4583 : /* Go over edges calling NODE and its thunks and gather information about
4584 : incoming counts so that we know if we need to make any adjustments. */
4585 :
4586 : static void
4587 8 : analyze_clone_icoming_counts (cgraph_node *node,
4588 : desc_incoming_count_struct *desc)
4589 : {
4590 20 : for (cgraph_edge *cs = node->callers; cs; cs = cs->next_caller)
4591 12 : if (cs->caller->thunk)
4592 : {
4593 0 : analyze_clone_icoming_counts (cs->caller, desc);
4594 0 : continue;
4595 : }
4596 : else
4597 : {
4598 12 : if (cs->count.initialized_p ())
4599 12 : desc->count += cs->count.ipa ();
4600 12 : if (!desc->processed_edges->contains (cs)
4601 12 : && cs->caller->clone_of == desc->orig)
4602 4 : desc->unproc_orig_rec_edges++;
4603 : }
4604 8 : }
4605 :
4606 : /* If caller edge counts of a clone created for a self-recursive arithmetic
4607 : jump function must be adjusted because it is coming from a the "seed" clone
4608 : for the first value and so has been excessively scaled back as if it was not
4609 : a recursive call, adjust it so that the incoming counts of NODE match its
4610 : count. NODE is the node or its thunk. */
4611 :
4612 : static void
4613 0 : adjust_clone_incoming_counts (cgraph_node *node,
4614 : desc_incoming_count_struct *desc)
4615 : {
4616 0 : for (cgraph_edge *cs = node->callers; cs; cs = cs->next_caller)
4617 0 : if (cs->caller->thunk)
4618 : {
4619 0 : adjust_clone_incoming_counts (cs->caller, desc);
4620 0 : profile_count sum = profile_count::zero ();
4621 0 : for (cgraph_edge *e = cs->caller->callers; e; e = e->next_caller)
4622 0 : if (e->count.initialized_p ())
4623 0 : sum += e->count.ipa ();
4624 0 : cs->count = cs->count.combine_with_ipa_count (sum);
4625 : }
4626 0 : else if (!desc->processed_edges->contains (cs)
4627 0 : && cs->caller->clone_of == desc->orig
4628 0 : && cs->count.compatible_p (desc->count))
4629 : {
4630 0 : cs->count += desc->count;
4631 0 : if (dump_file)
4632 : {
4633 0 : fprintf (dump_file, " Adjusted count of an incoming edge of "
4634 0 : "a clone %s -> %s to ", cs->caller->dump_name (),
4635 0 : cs->callee->dump_name ());
4636 0 : cs->count.dump (dump_file);
4637 0 : fprintf (dump_file, "\n");
4638 : }
4639 : }
4640 0 : }
4641 :
4642 : /* When ORIG_NODE has been cloned for values which have been generated fora
4643 : self-recursive call as a result of an arithmetic pass-through
4644 : jump-functions, adjust its count together with counts of all such clones in
4645 : SELF_GEN_CLONES which also at this point contains ORIG_NODE itself.
4646 :
4647 : The function sums the counts of the original node and all its clones that
4648 : cannot be attributed to a specific clone because it comes from a
4649 : non-recursive edge. This sum is then evenly divided between the clones and
4650 : on top of that each one gets all the counts which can be attributed directly
4651 : to it. */
4652 :
4653 : static void
4654 33 : update_counts_for_self_gen_clones (cgraph_node *orig_node,
4655 : const vec<cgraph_node *> &self_gen_clones)
4656 : {
4657 33 : profile_count redist_sum = orig_node->count.ipa ();
4658 33 : if (!redist_sum.nonzero_p ())
4659 : return;
4660 :
4661 4 : if (dump_file)
4662 0 : fprintf (dump_file, " Updating profile of self recursive clone "
4663 : "series\n");
4664 :
4665 4 : gather_other_count_struct gocs;
4666 4 : gocs.orig = orig_node;
4667 4 : gocs.other_count = profile_count::zero ();
4668 :
4669 4 : auto_vec <profile_count, 8> other_edges_count;
4670 20 : for (cgraph_node *n : self_gen_clones)
4671 : {
4672 8 : gocs.other_count = profile_count::zero ();
4673 8 : n->call_for_symbol_thunks_and_aliases (gather_count_of_non_rec_edges,
4674 : &gocs, false);
4675 8 : other_edges_count.safe_push (gocs.other_count);
4676 8 : redist_sum -= gocs.other_count;
4677 : }
4678 :
4679 4 : hash_set<cgraph_edge *> processed_edges;
4680 4 : unsigned i = 0;
4681 20 : for (cgraph_node *n : self_gen_clones)
4682 : {
4683 8 : profile_count new_count
4684 16 : = (redist_sum / self_gen_clones.length () + other_edges_count[i]);
4685 8 : new_count = lenient_count_portion_handling (new_count, orig_node);
4686 8 : n->scale_profile_to (new_count);
4687 16 : for (cgraph_edge *cs = n->callees; cs; cs = cs->next_callee)
4688 8 : processed_edges.add (cs);
4689 :
4690 8 : i++;
4691 : }
4692 :
4693 : /* There are still going to be edges to ORIG_NODE that have one or more
4694 : clones coming from another node clone in SELF_GEN_CLONES and which we
4695 : scaled by the same amount, which means that the total incoming sum of
4696 : counts to ORIG_NODE will be too high, scale such edges back. */
4697 8 : for (cgraph_edge *cs = orig_node->callees; cs; cs = cs->next_callee)
4698 : {
4699 4 : if (cs->callee->ultimate_alias_target () == orig_node)
4700 : {
4701 4 : unsigned den = 0;
4702 18 : for (cgraph_edge *e = cs; e; e = get_next_cgraph_edge_clone (e))
4703 14 : if (e->callee->ultimate_alias_target () == orig_node
4704 14 : && processed_edges.contains (e))
4705 8 : den++;
4706 4 : if (den > 0)
4707 18 : for (cgraph_edge *e = cs; e; e = get_next_cgraph_edge_clone (e))
4708 14 : if (e->callee->ultimate_alias_target () == orig_node
4709 8 : && processed_edges.contains (e)
4710 : /* If count is not IPA, this adjustment makes verifier
4711 : unhappy, since we expect bb->count to match e->count.
4712 : We may add a flag to mark edge conts that has been
4713 : modified by IPA code, but so far it does not seem
4714 : to be worth the effort. With local counts the profile
4715 : will not propagate at IPA level. */
4716 30 : && e->count.ipa_p ())
4717 8 : e->count /= den;
4718 : }
4719 : }
4720 :
4721 : /* Edges from the seeds of the valus generated for arithmetic jump-functions
4722 : along self-recursive edges are likely to have fairly low count and so
4723 : edges from them to nodes in the self_gen_clones do not correspond to the
4724 : artificially distributed count of the nodes, the total sum of incoming
4725 : edges to some clones might be too low. Detect this situation and correct
4726 : it. */
4727 20 : for (cgraph_node *n : self_gen_clones)
4728 : {
4729 8 : if (!n->count.ipa ().nonzero_p ())
4730 0 : continue;
4731 :
4732 8 : desc_incoming_count_struct desc;
4733 8 : desc.orig = orig_node;
4734 8 : desc.processed_edges = &processed_edges;
4735 8 : desc.count = profile_count::zero ();
4736 8 : desc.unproc_orig_rec_edges = 0;
4737 8 : analyze_clone_icoming_counts (n, &desc);
4738 :
4739 8 : if (n->count.differs_from_p (desc.count))
4740 : {
4741 0 : if (n->count > desc.count
4742 0 : && desc.unproc_orig_rec_edges > 0)
4743 : {
4744 0 : desc.count = n->count - desc.count;
4745 0 : desc.count = desc.count /= desc.unproc_orig_rec_edges;
4746 0 : adjust_clone_incoming_counts (n, &desc);
4747 : }
4748 0 : else if (dump_file)
4749 0 : fprintf (dump_file,
4750 : " Unable to fix up incoming counts for %s.\n",
4751 : n->dump_name ());
4752 : }
4753 : }
4754 :
4755 4 : if (dump_file)
4756 0 : for (cgraph_node *n : self_gen_clones)
4757 0 : dump_profile_updates (n, n != orig_node);
4758 4 : return;
4759 4 : }
4760 :
4761 : /* After a specialized NEW_NODE version of ORIG_NODE has been created, update
4762 : their profile information to reflect this. This function should not be used
4763 : for clones generated for arithmetic pass-through jump functions on a
4764 : self-recursive call graph edge, that situation is handled by
4765 : update_counts_for_self_gen_clones. */
4766 :
4767 : static void
4768 4177 : update_profiling_info (struct cgraph_node *orig_node,
4769 : struct cgraph_node *new_node)
4770 : {
4771 4177 : struct caller_statistics stats;
4772 4177 : profile_count new_sum;
4773 4177 : profile_count remainder, orig_node_count = orig_node->count.ipa ();
4774 :
4775 4177 : if (!orig_node_count.nonzero_p ())
4776 4169 : return;
4777 :
4778 8 : if (dump_file)
4779 : {
4780 2 : fprintf (dump_file, " Updating profile from original count: ");
4781 2 : orig_node_count.dump (dump_file);
4782 2 : fprintf (dump_file, "\n");
4783 : }
4784 :
4785 8 : init_caller_stats (&stats, new_node);
4786 8 : new_node->call_for_symbol_thunks_and_aliases (gather_caller_stats, &stats,
4787 : false);
4788 8 : new_sum = stats.count_sum;
4789 :
4790 8 : bool orig_edges_processed = false;
4791 8 : if (new_sum > orig_node_count)
4792 : {
4793 : /* Profile has alreay gone astray, keep what we have but lower it
4794 : to global0adjusted or to local if we have partial training. */
4795 0 : if (opt_for_fn (orig_node->decl, flag_profile_partial_training))
4796 0 : orig_node->make_profile_local ();
4797 0 : if (new_sum.quality () == AFDO)
4798 0 : orig_node->make_profile_global0 (GUESSED_GLOBAL0_AFDO);
4799 : else
4800 0 : orig_node->make_profile_global0 (GUESSED_GLOBAL0_ADJUSTED);
4801 : orig_edges_processed = true;
4802 : }
4803 8 : else if (stats.rec_count_sum.nonzero_p ())
4804 : {
4805 0 : int new_nonrec_calls = stats.n_nonrec_calls;
4806 : /* There are self-recursive edges which are likely to bring in the
4807 : majority of calls but which we must divide in between the original and
4808 : new node. */
4809 0 : init_caller_stats (&stats, orig_node);
4810 0 : orig_node->call_for_symbol_thunks_and_aliases (gather_caller_stats,
4811 : &stats, false);
4812 0 : int orig_nonrec_calls = stats.n_nonrec_calls;
4813 0 : profile_count orig_nonrec_call_count = stats.count_sum;
4814 :
4815 0 : if (orig_node->local)
4816 : {
4817 0 : if (!orig_nonrec_call_count.nonzero_p ())
4818 : {
4819 0 : if (dump_file)
4820 0 : fprintf (dump_file, " The original is local and the only "
4821 : "incoming edges from non-dead callers with nonzero "
4822 : "counts are self-recursive, assuming it is cold.\n");
4823 : /* The NEW_NODE count and counts of all its outgoing edges
4824 : are still unmodified copies of ORIG_NODE's. Just clear
4825 : the latter and bail out. */
4826 0 : if (opt_for_fn (orig_node->decl, flag_profile_partial_training))
4827 0 : orig_node->make_profile_local ();
4828 0 : else if (orig_nonrec_call_count.quality () == AFDO)
4829 0 : orig_node->make_profile_global0 (GUESSED_GLOBAL0_AFDO);
4830 : else
4831 0 : orig_node->make_profile_global0 (GUESSED_GLOBAL0_ADJUSTED);
4832 0 : return;
4833 : }
4834 : }
4835 : else
4836 : {
4837 : /* Let's behave as if there was another caller that accounts for all
4838 : the calls that were either indirect or from other compilation
4839 : units. */
4840 0 : orig_nonrec_calls++;
4841 0 : profile_count pretend_caller_count
4842 0 : = (orig_node_count - new_sum - orig_nonrec_call_count
4843 0 : - stats.rec_count_sum);
4844 0 : orig_nonrec_call_count += pretend_caller_count;
4845 : }
4846 :
4847 : /* Divide all "unexplained" counts roughly proportionally to sums of
4848 : counts of non-recursive calls.
4849 :
4850 : We put rather arbitrary limits on how many counts we claim because the
4851 : number of non-self-recursive incoming count is only a rough guideline
4852 : and there are cases (such as mcf) where using it blindly just takes
4853 : too many. And if lattices are considered in the opposite order we
4854 : could also take too few. */
4855 0 : profile_count unexp = orig_node_count - new_sum - orig_nonrec_call_count;
4856 :
4857 0 : int limit_den = 2 * (orig_nonrec_calls + new_nonrec_calls);
4858 0 : profile_count new_part = unexp.apply_scale (limit_den - 1, limit_den);
4859 0 : profile_count den = new_sum + orig_nonrec_call_count;
4860 0 : if (den.nonzero_p ())
4861 0 : new_part = MIN (unexp.apply_scale (new_sum, den), new_part);
4862 0 : new_part = MAX (new_part,
4863 : unexp.apply_scale (new_nonrec_calls, limit_den));
4864 0 : if (dump_file)
4865 : {
4866 0 : fprintf (dump_file, " Claiming ");
4867 0 : new_part.dump (dump_file);
4868 0 : fprintf (dump_file, " of unexplained ");
4869 0 : unexp.dump (dump_file);
4870 0 : fprintf (dump_file, " counts because of self-recursive "
4871 : "calls\n");
4872 : }
4873 0 : new_sum += new_part;
4874 0 : remainder = lenient_count_portion_handling (orig_node_count - new_sum,
4875 : orig_node);
4876 : }
4877 : else
4878 8 : remainder = lenient_count_portion_handling (orig_node_count - new_sum,
4879 : orig_node);
4880 :
4881 8 : new_node->scale_profile_to (new_sum);
4882 :
4883 8 : if (!orig_edges_processed)
4884 8 : orig_node->scale_profile_to (remainder);
4885 :
4886 8 : if (dump_file)
4887 : {
4888 2 : dump_profile_updates (new_node, true);
4889 2 : dump_profile_updates (orig_node, false);
4890 : }
4891 : }
4892 :
4893 : /* Update the respective profile of specialized NEW_NODE and the original
4894 : ORIG_NODE after additional edges with cumulative count sum REDIRECTED_SUM
4895 : have been redirected to the specialized version. */
4896 :
4897 : static void
4898 0 : update_specialized_profile (struct cgraph_node *new_node,
4899 : struct cgraph_node *orig_node,
4900 : profile_count redirected_sum)
4901 : {
4902 0 : if (dump_file)
4903 : {
4904 0 : fprintf (dump_file, " the sum of counts of redirected edges is ");
4905 0 : redirected_sum.dump (dump_file);
4906 0 : fprintf (dump_file, "\n old ipa count of the original node is ");
4907 0 : orig_node->count.dump (dump_file);
4908 0 : fprintf (dump_file, "\n");
4909 : }
4910 0 : if (!orig_node->count.ipa ().nonzero_p ()
4911 0 : || !redirected_sum.nonzero_p ())
4912 0 : return;
4913 :
4914 0 : orig_node->scale_profile_to
4915 0 : (lenient_count_portion_handling (orig_node->count.ipa () - redirected_sum,
4916 : orig_node));
4917 :
4918 0 : new_node->scale_profile_to (new_node->count.ipa () + redirected_sum);
4919 :
4920 0 : if (dump_file)
4921 : {
4922 0 : dump_profile_updates (new_node, true);
4923 0 : dump_profile_updates (orig_node, false);
4924 : }
4925 : }
4926 :
4927 : static void adjust_references_in_caller (cgraph_edge *cs,
4928 : symtab_node *symbol, int index);
4929 :
4930 : /* Simple structure to pass a symbol and index (with same meaning as parameters
4931 : of adjust_references_in_caller) through a void* parameter of a
4932 : call_for_symbol_thunks_and_aliases callback. */
4933 : struct symbol_and_index_together
4934 : {
4935 : symtab_node *symbol;
4936 : int index;
4937 : };
4938 :
4939 : /* Worker callback of call_for_symbol_thunks_and_aliases to recursively call
4940 : adjust_references_in_caller on edges up in the call-graph, if necessary. */
4941 : static bool
4942 8 : adjust_refs_in_act_callers (struct cgraph_node *node, void *data)
4943 : {
4944 8 : symbol_and_index_together *pack = (symbol_and_index_together *) data;
4945 38 : for (cgraph_edge *cs = node->callers; cs; cs = cs->next_caller)
4946 30 : if (!cs->caller->thunk)
4947 30 : adjust_references_in_caller (cs, pack->symbol, pack->index);
4948 8 : return false;
4949 : }
4950 :
4951 : /* At INDEX of a function being called by CS there is an ADDR_EXPR of a
4952 : variable which is only dereferenced and which is represented by SYMBOL. See
4953 : if we can remove ADDR reference in callers assosiated witht the call. */
4954 :
4955 : static void
4956 400 : adjust_references_in_caller (cgraph_edge *cs, symtab_node *symbol, int index)
4957 : {
4958 400 : ipa_edge_args *args = ipa_edge_args_sum->get (cs);
4959 400 : ipa_jump_func *jfunc = ipa_get_ith_jump_func (args, index);
4960 400 : if (jfunc->type == IPA_JF_CONST)
4961 : {
4962 382 : ipa_ref *to_del = cs->caller->find_reference (symbol, cs->call_stmt,
4963 : cs->lto_stmt_uid,
4964 : IPA_REF_ADDR);
4965 382 : if (!to_del)
4966 392 : return;
4967 382 : to_del->remove_reference ();
4968 382 : ipa_zap_jf_refdesc (jfunc);
4969 382 : if (dump_file)
4970 22 : fprintf (dump_file, " Removed a reference from %s to %s.\n",
4971 11 : cs->caller->dump_name (), symbol->dump_name ());
4972 382 : return;
4973 : }
4974 :
4975 18 : if (jfunc->type != IPA_JF_PASS_THROUGH
4976 18 : || ipa_get_jf_pass_through_operation (jfunc) != NOP_EXPR
4977 36 : || ipa_get_jf_pass_through_refdesc_decremented (jfunc))
4978 : return;
4979 :
4980 18 : int fidx = ipa_get_jf_pass_through_formal_id (jfunc);
4981 18 : cgraph_node *caller = cs->caller;
4982 18 : ipa_node_params *caller_info = ipa_node_params_sum->get (caller);
4983 : /* TODO: This consistency check may be too big and not really
4984 : that useful. Consider removing it. */
4985 18 : tree cst;
4986 18 : if (caller_info->ipcp_orig_node)
4987 17 : cst = caller_info->known_csts[fidx];
4988 : else
4989 : {
4990 1 : ipcp_lattice<tree> *lat = ipa_get_scalar_lat (caller_info, fidx);
4991 1 : gcc_assert (lat->is_single_const ());
4992 1 : cst = lat->values->value;
4993 : }
4994 18 : gcc_assert (TREE_CODE (cst) == ADDR_EXPR
4995 : && (symtab_node::get (get_base_address (TREE_OPERAND (cst, 0)))
4996 : == symbol));
4997 :
4998 18 : int cuses = ipa_get_controlled_uses (caller_info, fidx);
4999 18 : if (cuses == IPA_UNDESCRIBED_USE)
5000 : return;
5001 18 : gcc_assert (cuses > 0);
5002 18 : cuses--;
5003 18 : ipa_set_controlled_uses (caller_info, fidx, cuses);
5004 18 : ipa_set_jf_pass_through_refdesc_decremented (jfunc, true);
5005 18 : if (dump_file && (dump_flags & TDF_DETAILS))
5006 3 : fprintf (dump_file, " Controlled uses of parameter %i of %s dropped "
5007 : "to %i.\n", fidx, caller->dump_name (), cuses);
5008 18 : if (cuses)
5009 : return;
5010 :
5011 8 : if (caller_info->ipcp_orig_node)
5012 : {
5013 : /* Cloning machinery has created a reference here, we need to either
5014 : remove it or change it to a read one. */
5015 7 : ipa_ref *to_del = caller->find_reference (symbol, NULL, 0, IPA_REF_ADDR);
5016 7 : if (to_del)
5017 : {
5018 7 : to_del->remove_reference ();
5019 7 : if (dump_file)
5020 6 : fprintf (dump_file, " Removed a reference from %s to %s.\n",
5021 3 : cs->caller->dump_name (), symbol->dump_name ());
5022 7 : if (ipa_get_param_load_dereferenced (caller_info, fidx))
5023 : {
5024 3 : caller->create_reference (symbol, IPA_REF_LOAD, NULL);
5025 3 : if (dump_file)
5026 2 : fprintf (dump_file,
5027 : " ...and replaced it with LOAD one.\n");
5028 : }
5029 : }
5030 : }
5031 :
5032 8 : symbol_and_index_together pack;
5033 8 : pack.symbol = symbol;
5034 8 : pack.index = fidx;
5035 8 : if (caller->can_change_signature)
5036 8 : caller->call_for_symbol_thunks_and_aliases (adjust_refs_in_act_callers,
5037 : &pack, true);
5038 : }
5039 :
5040 :
5041 : /* Return true if we would like to remove a parameter from NODE when cloning it
5042 : with KNOWN_CSTS scalar constants. */
5043 :
5044 : static bool
5045 19912 : want_remove_some_param_p (cgraph_node *node, vec<tree> known_csts)
5046 : {
5047 19912 : auto_vec<bool, 16> surviving;
5048 19912 : bool filled_vec = false;
5049 19912 : ipa_node_params *info = ipa_node_params_sum->get (node);
5050 19912 : int i, count = ipa_get_param_count (info);
5051 :
5052 39555 : for (i = 0; i < count; i++)
5053 : {
5054 35041 : if (!known_csts[i] && ipa_is_param_used (info, i))
5055 19643 : continue;
5056 :
5057 15398 : if (!filled_vec)
5058 : {
5059 15398 : clone_info *info = clone_info::get (node);
5060 15398 : if (!info || !info->param_adjustments)
5061 : return true;
5062 0 : info->param_adjustments->get_surviving_params (&surviving);
5063 0 : filled_vec = true;
5064 : }
5065 0 : if (surviving.length() < (unsigned) i && surviving[i])
5066 : return true;
5067 : }
5068 : return false;
5069 19912 : }
5070 :
5071 : /* Create a specialized version of NODE with known constants in KNOWN_CSTS,
5072 : known contexts in KNOWN_CONTEXTS and known aggregate values in AGGVALS and
5073 : redirect all edges in CALLERS to it. */
5074 :
5075 : static struct cgraph_node *
5076 21316 : create_specialized_node (struct cgraph_node *node,
5077 : vec<tree> known_csts,
5078 : vec<ipa_polymorphic_call_context> known_contexts,
5079 : vec<ipa_argagg_value, va_gc> *aggvals,
5080 : vec<cgraph_edge *> &callers)
5081 : {
5082 21316 : ipa_node_params *new_info, *info = ipa_node_params_sum->get (node);
5083 21316 : vec<ipa_replace_map *, va_gc> *replace_trees = NULL;
5084 21316 : vec<ipa_adjusted_param, va_gc> *new_params = NULL;
5085 21316 : struct cgraph_node *new_node;
5086 21316 : int i, count = ipa_get_param_count (info);
5087 21316 : clone_info *cinfo = clone_info::get (node);
5088 0 : ipa_param_adjustments *old_adjustments = cinfo
5089 21316 : ? cinfo->param_adjustments : NULL;
5090 21316 : ipa_param_adjustments *new_adjustments;
5091 21316 : gcc_assert (!info->ipcp_orig_node);
5092 21316 : gcc_assert (node->can_change_signature
5093 : || !old_adjustments);
5094 :
5095 19912 : if (old_adjustments)
5096 : {
5097 : /* At the moment all IPA optimizations should use the number of
5098 : parameters of the prevailing decl as the m_always_copy_start.
5099 : Handling any other value would complicate the code below, so for the
5100 : time bing let's only assert it is so. */
5101 0 : gcc_assert (old_adjustments->m_always_copy_start == count
5102 : || old_adjustments->m_always_copy_start < 0);
5103 0 : int old_adj_count = vec_safe_length (old_adjustments->m_adj_params);
5104 0 : for (i = 0; i < old_adj_count; i++)
5105 : {
5106 0 : ipa_adjusted_param *old_adj = &(*old_adjustments->m_adj_params)[i];
5107 0 : if (!node->can_change_signature
5108 0 : || old_adj->op != IPA_PARAM_OP_COPY
5109 0 : || (!known_csts[old_adj->base_index]
5110 0 : && ipa_is_param_used (info, old_adj->base_index)))
5111 : {
5112 0 : ipa_adjusted_param new_adj = *old_adj;
5113 :
5114 0 : new_adj.prev_clone_adjustment = true;
5115 0 : new_adj.prev_clone_index = i;
5116 0 : vec_safe_push (new_params, new_adj);
5117 : }
5118 : }
5119 0 : bool skip_return = old_adjustments->m_skip_return;
5120 0 : new_adjustments = (new (ggc_alloc <ipa_param_adjustments> ())
5121 : ipa_param_adjustments (new_params, count,
5122 0 : skip_return));
5123 : }
5124 21316 : else if (node->can_change_signature
5125 21316 : && want_remove_some_param_p (node, known_csts))
5126 : {
5127 15398 : ipa_adjusted_param adj;
5128 15398 : memset (&adj, 0, sizeof (adj));
5129 15398 : adj.op = IPA_PARAM_OP_COPY;
5130 59836 : for (i = 0; i < count; i++)
5131 44438 : if (!known_csts[i] && ipa_is_param_used (info, i))
5132 : {
5133 15929 : adj.base_index = i;
5134 15929 : adj.prev_clone_index = i;
5135 15929 : vec_safe_push (new_params, adj);
5136 : }
5137 15398 : new_adjustments = (new (ggc_alloc <ipa_param_adjustments> ())
5138 15398 : ipa_param_adjustments (new_params, count, false));
5139 : }
5140 : else
5141 : new_adjustments = NULL;
5142 :
5143 21316 : auto_vec<cgraph_edge *, 2> self_recursive_calls;
5144 186694 : for (i = callers.length () - 1; i >= 0; i--)
5145 : {
5146 144062 : cgraph_edge *cs = callers[i];
5147 144062 : if (cs->caller == node)
5148 : {
5149 131 : self_recursive_calls.safe_push (cs);
5150 131 : callers.unordered_remove (i);
5151 : }
5152 : }
5153 21316 : replace_trees = cinfo ? vec_safe_copy (cinfo->tree_map) : NULL;
5154 81076 : for (i = 0; i < count; i++)
5155 : {
5156 59760 : tree t = known_csts[i];
5157 59760 : if (!t)
5158 35467 : continue;
5159 :
5160 24293 : gcc_checking_assert (TREE_CODE (t) != TREE_BINFO);
5161 :
5162 24293 : bool load_ref = false;
5163 24293 : symtab_node *ref_symbol;
5164 24293 : if (TREE_CODE (t) == ADDR_EXPR)
5165 : {
5166 6493 : tree base = get_base_address (TREE_OPERAND (t, 0));
5167 6493 : if (TREE_CODE (base) == VAR_DECL
5168 3107 : && ipa_get_controlled_uses (info, i) == 0
5169 930 : && ipa_get_param_load_dereferenced (info, i)
5170 6878 : && (ref_symbol = symtab_node::get (base)))
5171 : {
5172 385 : load_ref = true;
5173 385 : if (node->can_change_signature)
5174 1399 : for (cgraph_edge *caller : callers)
5175 370 : adjust_references_in_caller (caller, ref_symbol, i);
5176 : }
5177 : }
5178 :
5179 24293 : ipa_replace_map *replace_map = get_replacement_map (info, t, i, load_ref);
5180 24293 : if (replace_map)
5181 24293 : vec_safe_push (replace_trees, replace_map);
5182 : }
5183 :
5184 63948 : unsigned &suffix_counter = clone_num_suffixes->get_or_insert (
5185 21316 : IDENTIFIER_POINTER (DECL_ASSEMBLER_NAME (
5186 : node->decl)));
5187 21316 : new_node = node->create_virtual_clone (callers, replace_trees,
5188 : new_adjustments, "constprop",
5189 : suffix_counter);
5190 21316 : suffix_counter++;
5191 :
5192 21316 : bool have_self_recursive_calls = !self_recursive_calls.is_empty ();
5193 21447 : for (unsigned j = 0; j < self_recursive_calls.length (); j++)
5194 : {
5195 131 : cgraph_edge *cs = get_next_cgraph_edge_clone (self_recursive_calls[j]);
5196 : /* Cloned edges can disappear during cloning as speculation can be
5197 : resolved, check that we have one and that it comes from the last
5198 : cloning. */
5199 131 : if (cs && cs->caller == new_node)
5200 130 : cs->redirect_callee_duplicating_thunks (new_node);
5201 : /* Any future code that would make more than one clone of an outgoing
5202 : edge would confuse this mechanism, so let's check that does not
5203 : happen. */
5204 130 : gcc_checking_assert (!cs
5205 : || !get_next_cgraph_edge_clone (cs)
5206 : || get_next_cgraph_edge_clone (cs)->caller != new_node);
5207 : }
5208 21316 : if (have_self_recursive_calls)
5209 121 : new_node->expand_all_artificial_thunks ();
5210 :
5211 21316 : ipa_set_node_agg_value_chain (new_node, aggvals);
5212 50899 : for (const ipa_argagg_value &av : aggvals)
5213 29583 : new_node->maybe_create_reference (av.value, NULL);
5214 :
5215 21316 : if (dump_file && (dump_flags & TDF_DETAILS))
5216 : {
5217 90 : fprintf (dump_file, " the new node is %s.\n", new_node->dump_name ());
5218 90 : if (known_contexts.exists ())
5219 : {
5220 0 : for (i = 0; i < count; i++)
5221 0 : if (!known_contexts[i].useless_p ())
5222 : {
5223 0 : fprintf (dump_file, " known ctx %i is ", i);
5224 0 : known_contexts[i].dump (dump_file);
5225 : }
5226 : }
5227 90 : if (aggvals)
5228 : {
5229 48 : fprintf (dump_file, " Aggregate replacements:");
5230 48 : ipa_argagg_value_list avs (aggvals);
5231 48 : avs.dump (dump_file);
5232 : }
5233 : }
5234 :
5235 21316 : new_info = ipa_node_params_sum->get (new_node);
5236 21316 : new_info->ipcp_orig_node = node;
5237 21316 : new_node->ipcp_clone = true;
5238 21316 : new_info->known_csts = known_csts;
5239 21316 : new_info->known_contexts = known_contexts;
5240 :
5241 21316 : ipcp_discover_new_direct_edges (new_node, known_csts, known_contexts,
5242 : aggvals);
5243 :
5244 21316 : return new_node;
5245 21316 : }
5246 :
5247 : /* Return true if JFUNC, which describes a i-th parameter of call CS, is a
5248 : pass-through function to itself when the cgraph_node involved is not an
5249 : IPA-CP clone. When SIMPLE is true, further check if JFUNC is a simple
5250 : no-operation pass-through. */
5251 :
5252 : static bool
5253 776991 : self_recursive_pass_through_p (cgraph_edge *cs, ipa_jump_func *jfunc, int i,
5254 : bool simple = true)
5255 : {
5256 776991 : enum availability availability;
5257 776991 : if (jfunc->type == IPA_JF_PASS_THROUGH
5258 80027 : && cs->caller == cs->callee->function_symbol (&availability)
5259 19350 : && availability > AVAIL_INTERPOSABLE
5260 19350 : && (!simple || ipa_get_jf_pass_through_operation (jfunc) == NOP_EXPR)
5261 19350 : && ipa_get_jf_pass_through_formal_id (jfunc) == i
5262 19350 : && ipa_node_params_sum->get (cs->caller)
5263 796341 : && !ipa_node_params_sum->get (cs->caller)->ipcp_orig_node)
5264 : return true;
5265 : return false;
5266 : }
5267 :
5268 : /* Return true if JFUNC, which describes the i-th parameter of call CS, is an
5269 : ancestor function with zero offset to itself when the cgraph_node involved
5270 : is not an IPA-CP clone. */
5271 :
5272 : static bool
5273 757653 : self_recursive_ancestor_p (cgraph_edge *cs, ipa_jump_func *jfunc, int i)
5274 : {
5275 757653 : enum availability availability;
5276 757653 : if (jfunc->type == IPA_JF_ANCESTOR
5277 3226 : && cs->caller == cs->callee->function_symbol (&availability)
5278 1 : && availability > AVAIL_INTERPOSABLE
5279 1 : && ipa_get_jf_ancestor_offset (jfunc) == 0
5280 1 : && ipa_get_jf_ancestor_formal_id (jfunc) == i
5281 1 : && ipa_node_params_sum->get (cs->caller)
5282 757654 : && !ipa_node_params_sum->get (cs->caller)->ipcp_orig_node)
5283 : return true;
5284 : return false;
5285 : }
5286 :
5287 : /* Return true if JFUNC, which describes a part of an aggregate represented or
5288 : pointed to by the i-th parameter of call CS, is a pass-through function to
5289 : itself when the cgraph_node involved is not an IPA-CP clone.. When
5290 : SIMPLE is true, further check if JFUNC is a simple no-operation
5291 : pass-through. */
5292 :
5293 : static bool
5294 333298 : self_recursive_agg_pass_through_p (const cgraph_edge *cs,
5295 : const ipa_agg_jf_item *jfunc,
5296 : int i, bool simple = true)
5297 : {
5298 333298 : enum availability availability;
5299 333298 : if (cs->caller == cs->callee->function_symbol (&availability)
5300 3690 : && availability > AVAIL_INTERPOSABLE
5301 3690 : && jfunc->jftype == IPA_JF_LOAD_AGG
5302 487 : && jfunc->offset == jfunc->value.load_agg.offset
5303 487 : && (!simple || jfunc->value.pass_through.operation == NOP_EXPR)
5304 487 : && jfunc->value.pass_through.formal_id == i
5305 481 : && useless_type_conversion_p (jfunc->value.load_agg.type, jfunc->type)
5306 481 : && ipa_node_params_sum->get (cs->caller)
5307 333779 : && !ipa_node_params_sum->get (cs->caller)->ipcp_orig_node)
5308 : return true;
5309 : return false;
5310 : }
5311 :
5312 : /* Given a NODE, and a subset of its CALLERS, try to populate blanks slots in
5313 : KNOWN_CSTS with constants that are also known for all of the CALLERS. */
5314 :
5315 : static void
5316 167285 : find_scalar_values_for_callers_subset (vec<tree> &known_csts,
5317 : ipa_node_params *info,
5318 : const vec<cgraph_edge *> &callers)
5319 : {
5320 167285 : int i, count = ipa_get_param_count (info);
5321 :
5322 731155 : for (i = 0; i < count; i++)
5323 : {
5324 563870 : ipcp_lattice<tree> *lat = ipa_get_scalar_lat (info, i);
5325 563870 : if (lat->bottom)
5326 563870 : continue;
5327 550253 : if (lat->is_single_const ())
5328 : {
5329 29202 : known_csts[i] = lat->values->value;
5330 29202 : continue;
5331 : }
5332 :
5333 521051 : struct cgraph_edge *cs;
5334 521051 : tree newval = NULL_TREE;
5335 521051 : int j;
5336 521051 : bool first = true;
5337 521051 : tree type = ipa_get_type (info, i);
5338 :
5339 971803 : FOR_EACH_VEC_ELT (callers, j, cs)
5340 : {
5341 776449 : struct ipa_jump_func *jump_func;
5342 776449 : tree t;
5343 :
5344 776449 : ipa_edge_args *args = ipa_edge_args_sum->get (cs);
5345 776449 : if (!args
5346 776449 : || i >= ipa_get_cs_argument_count (args)
5347 1552867 : || (i == 0
5348 179303 : && call_passes_through_thunk (cs)))
5349 : {
5350 : newval = NULL_TREE;
5351 : break;
5352 : }
5353 776348 : jump_func = ipa_get_ith_jump_func (args, i);
5354 :
5355 : /* Besides simple pass-through jump function, arithmetic jump
5356 : function could also introduce argument-direct-pass-through for
5357 : self-feeding recursive call. For example,
5358 :
5359 : fn (int i)
5360 : {
5361 : fn (i & 1);
5362 : }
5363 :
5364 : Given that i is 0, recursive propagation via (i & 1) also gets
5365 : 0. */
5366 776348 : if (self_recursive_pass_through_p (cs, jump_func, i, false))
5367 : {
5368 18702 : gcc_assert (newval);
5369 18702 : enum tree_code opcode
5370 18702 : = ipa_get_jf_pass_through_operation (jump_func);
5371 18702 : tree op_type = (opcode == NOP_EXPR) ? NULL_TREE
5372 49 : : ipa_get_jf_pass_through_op_type (jump_func);
5373 18702 : t = ipa_get_jf_arith_result (opcode, newval,
5374 : ipa_get_jf_pass_through_operand (jump_func),
5375 : op_type);
5376 18702 : t = ipacp_value_safe_for_type (type, t);
5377 : }
5378 757646 : else if (self_recursive_ancestor_p (cs, jump_func, i))
5379 0 : continue;
5380 : else
5381 757646 : t = ipa_value_from_jfunc (ipa_node_params_sum->get (cs->caller),
5382 : jump_func, type);
5383 776348 : if (!t
5384 471955 : || (newval
5385 250880 : && !values_equal_for_ipcp_p (t, newval))
5386 1227100 : || (!first && !newval))
5387 : {
5388 : newval = NULL_TREE;
5389 : break;
5390 : }
5391 : else
5392 : newval = t;
5393 : first = false;
5394 : }
5395 :
5396 521051 : if (newval)
5397 195354 : known_csts[i] = newval;
5398 : }
5399 167285 : }
5400 :
5401 : /* Given a NODE and a subset of its CALLERS, try to populate plank slots in
5402 : KNOWN_CONTEXTS with polymorphic contexts that are also known for all of the
5403 : CALLERS. */
5404 :
5405 : static void
5406 167285 : find_contexts_for_caller_subset (vec<ipa_polymorphic_call_context>
5407 : &known_contexts,
5408 : ipa_node_params *info,
5409 : const vec<cgraph_edge *> &callers)
5410 : {
5411 167285 : int i, count = ipa_get_param_count (info);
5412 :
5413 731136 : for (i = 0; i < count; i++)
5414 : {
5415 563864 : if (!ipa_is_param_used (info, i))
5416 30387 : continue;
5417 :
5418 535572 : ipcp_lattice<ipa_polymorphic_call_context> *ctxlat
5419 535572 : = ipa_get_poly_ctx_lat (info, i);
5420 535572 : if (ctxlat->bottom)
5421 978 : continue;
5422 534594 : if (ctxlat->is_single_const ())
5423 : {
5424 1117 : if (!ctxlat->values->value.useless_p ())
5425 : {
5426 1117 : if (known_contexts.is_empty ())
5427 1056 : known_contexts.safe_grow_cleared (count, true);
5428 1117 : known_contexts[i] = ctxlat->values->value;
5429 : }
5430 1117 : continue;
5431 : }
5432 :
5433 533477 : cgraph_edge *cs;
5434 533477 : ipa_polymorphic_call_context newval;
5435 533477 : bool first = true;
5436 533477 : int j;
5437 :
5438 538660 : FOR_EACH_VEC_ELT (callers, j, cs)
5439 : {
5440 534916 : ipa_edge_args *args = ipa_edge_args_sum->get (cs);
5441 534916 : if (!args
5442 1069832 : || i >= ipa_get_cs_argument_count (args))
5443 13 : return;
5444 534903 : ipa_jump_func *jfunc = ipa_get_ith_jump_func (args, i);
5445 534903 : ipa_polymorphic_call_context ctx;
5446 534903 : ctx = ipa_context_from_jfunc (ipa_node_params_sum->get (cs->caller),
5447 : cs, i, jfunc);
5448 534903 : if (first)
5449 : {
5450 533464 : newval = ctx;
5451 533464 : first = false;
5452 : }
5453 : else
5454 1439 : newval.meet_with (ctx);
5455 1067115 : if (newval.useless_p ())
5456 : break;
5457 : }
5458 :
5459 1066928 : if (!newval.useless_p ())
5460 : {
5461 3744 : if (known_contexts.is_empty ())
5462 3521 : known_contexts.safe_grow_cleared (count, true);
5463 3744 : known_contexts[i] = newval;
5464 : }
5465 :
5466 : }
5467 : }
5468 :
5469 : /* Push all aggregate values coming along edge CS for parameter number INDEX to
5470 : RES. If INTERIM is non-NULL, it contains the current interim state of
5471 : collected aggregate values which can be used to compute values passed over
5472 : self-recursive edges.
5473 :
5474 : This basically one iteration of push_agg_values_from_edge over one
5475 : parameter, which allows for simpler early returns. */
5476 :
5477 : static void
5478 619701 : push_agg_values_for_index_from_edge (struct cgraph_edge *cs, int index,
5479 : vec<ipa_argagg_value> *res,
5480 : const ipa_argagg_value_list *interim)
5481 : {
5482 619701 : bool agg_values_from_caller = false;
5483 619701 : bool agg_jf_preserved = false;
5484 619701 : unsigned unit_delta = UINT_MAX;
5485 619701 : int src_idx = -1;
5486 619701 : ipa_jump_func *jfunc = ipa_get_ith_jump_func (ipa_edge_args_sum->get (cs),
5487 : index);
5488 :
5489 619701 : if (jfunc->type == IPA_JF_PASS_THROUGH
5490 619701 : && ipa_get_jf_pass_through_operation (jfunc) == NOP_EXPR)
5491 : {
5492 57056 : agg_values_from_caller = true;
5493 57056 : agg_jf_preserved = ipa_get_jf_pass_through_agg_preserved (jfunc);
5494 57056 : src_idx = ipa_get_jf_pass_through_formal_id (jfunc);
5495 57056 : unit_delta = 0;
5496 : }
5497 562645 : else if (jfunc->type == IPA_JF_ANCESTOR
5498 562645 : && ipa_get_jf_ancestor_agg_preserved (jfunc))
5499 : {
5500 407 : agg_values_from_caller = true;
5501 407 : agg_jf_preserved = true;
5502 407 : src_idx = ipa_get_jf_ancestor_formal_id (jfunc);
5503 407 : unit_delta = ipa_get_jf_ancestor_offset (jfunc) / BITS_PER_UNIT;
5504 : }
5505 :
5506 619701 : ipa_node_params *caller_info = ipa_node_params_sum->get (cs->caller);
5507 619701 : if (agg_values_from_caller)
5508 : {
5509 57463 : if (caller_info->ipcp_orig_node)
5510 : {
5511 11048 : struct cgraph_node *orig_node = caller_info->ipcp_orig_node;
5512 11048 : ipcp_transformation *ts
5513 11048 : = ipcp_get_transformation_summary (cs->caller);
5514 11048 : ipa_node_params *orig_info = ipa_node_params_sum->get (orig_node);
5515 11048 : ipcp_param_lattices *orig_plats
5516 11048 : = ipa_get_parm_lattices (orig_info, src_idx);
5517 11048 : if (ts
5518 11048 : && orig_plats->aggs
5519 2983 : && (agg_jf_preserved || !orig_plats->aggs_by_ref))
5520 : {
5521 2508 : ipa_argagg_value_list src (ts);
5522 2508 : src.push_adjusted_values (src_idx, index, unit_delta, res);
5523 2508 : return;
5524 : }
5525 : }
5526 : else
5527 : {
5528 46415 : ipcp_param_lattices *src_plats
5529 46415 : = ipa_get_parm_lattices (caller_info, src_idx);
5530 46415 : if (src_plats->aggs
5531 2440 : && !src_plats->aggs_bottom
5532 2440 : && (agg_jf_preserved || !src_plats->aggs_by_ref))
5533 : {
5534 1450 : if (interim && (self_recursive_pass_through_p (cs, jfunc, index)
5535 7 : || self_recursive_ancestor_p (cs, jfunc, index)))
5536 : {
5537 637 : interim->push_adjusted_values (src_idx, index, unit_delta,
5538 : res);
5539 637 : return;
5540 : }
5541 813 : if (!src_plats->aggs_contain_variable)
5542 : {
5543 80 : push_agg_values_from_plats (src_plats, index, unit_delta,
5544 : res);
5545 80 : return;
5546 : }
5547 : }
5548 : }
5549 : }
5550 :
5551 616476 : if (!jfunc->agg.items)
5552 : return;
5553 219405 : bool first = true;
5554 219405 : unsigned prev_unit_offset = 0;
5555 1193737 : for (const ipa_agg_jf_item &agg_jf : *jfunc->agg.items)
5556 : {
5557 974332 : tree value, srcvalue;
5558 : /* Besides simple pass-through aggregate jump function, arithmetic
5559 : aggregate jump function could also bring same aggregate value as
5560 : parameter passed-in for self-feeding recursive call. For example,
5561 :
5562 : fn (int *i)
5563 : {
5564 : int j = *i & 1;
5565 : fn (&j);
5566 : }
5567 :
5568 : Given that *i is 0, recursive propagation via (*i & 1) also gets 0. */
5569 974332 : if (interim
5570 333298 : && self_recursive_agg_pass_through_p (cs, &agg_jf, index, false)
5571 974813 : && (srcvalue = interim->get_value(index,
5572 481 : agg_jf.offset / BITS_PER_UNIT)))
5573 : {
5574 950 : value = ipa_get_jf_arith_result (agg_jf.value.pass_through.operation,
5575 : srcvalue,
5576 475 : agg_jf.value.pass_through.operand,
5577 475 : agg_jf.value.pass_through.op_type);
5578 475 : value = ipacp_value_safe_for_type (agg_jf.type, value);
5579 : }
5580 : else
5581 973857 : value = ipa_agg_value_from_jfunc (caller_info, cs->caller,
5582 : &agg_jf);
5583 974332 : if (value)
5584 : {
5585 951665 : struct ipa_argagg_value iav;
5586 951665 : iav.value = value;
5587 951665 : iav.unit_offset = agg_jf.offset / BITS_PER_UNIT;
5588 951665 : iav.index = index;
5589 951665 : iav.by_ref = jfunc->agg.by_ref;
5590 951665 : iav.killed = false;
5591 :
5592 951665 : gcc_assert (first
5593 : || iav.unit_offset > prev_unit_offset);
5594 951665 : prev_unit_offset = iav.unit_offset;
5595 951665 : first = false;
5596 :
5597 951665 : res->safe_push (iav);
5598 : }
5599 : }
5600 : return;
5601 : }
5602 :
5603 : /* Push all aggregate values coming along edge CS to RES. DEST_INFO is the
5604 : description of ultimate callee of CS or the one it was cloned from (the
5605 : summary where lattices are). If INTERIM is non-NULL, it contains the
5606 : current interim state of collected aggregate values which can be used to
5607 : compute values passed over self-recursive edges (if OPTIMIZE_SELF_RECURSION
5608 : is true) and to skip values which clearly will not be part of intersection
5609 : with INTERIM. */
5610 :
5611 : static void
5612 222417 : push_agg_values_from_edge (struct cgraph_edge *cs,
5613 : ipa_node_params *dest_info,
5614 : vec<ipa_argagg_value> *res,
5615 : const ipa_argagg_value_list *interim,
5616 : bool optimize_self_recursion)
5617 : {
5618 222417 : ipa_edge_args *args = ipa_edge_args_sum->get (cs);
5619 222417 : if (!args)
5620 : return;
5621 :
5622 444834 : int count = MIN (ipa_get_param_count (dest_info),
5623 : ipa_get_cs_argument_count (args));
5624 :
5625 222417 : unsigned interim_index = 0;
5626 929287 : for (int index = 0; index < count; index++)
5627 : {
5628 706870 : if (interim)
5629 : {
5630 273099 : while (interim_index < interim->m_elts.size ()
5631 249390 : && interim->m_elts[interim_index].value
5632 479716 : && interim->m_elts[interim_index].index < index)
5633 130068 : interim_index++;
5634 197843 : if (interim_index >= interim->m_elts.size ()
5635 143031 : || interim->m_elts[interim_index].index > index)
5636 54812 : continue;
5637 : }
5638 :
5639 652058 : ipcp_param_lattices *plats = ipa_get_parm_lattices (dest_info, index);
5640 652058 : if (!ipa_is_param_used (dest_info, index)
5641 652058 : || plats->aggs_bottom)
5642 32357 : continue;
5643 619749 : push_agg_values_for_index_from_edge (cs, index, res,
5644 : optimize_self_recursion ? interim
5645 : : NULL);
5646 : }
5647 : }
5648 :
5649 :
5650 : /* Look at edges in CALLERS and collect all known aggregate values that arrive
5651 : from all of them into INTERIM. Return how many there are. */
5652 :
5653 : static unsigned int
5654 167285 : find_aggregate_values_for_callers_subset_1 (vec<ipa_argagg_value> &interim,
5655 : struct cgraph_node *node,
5656 : const vec<cgraph_edge *> &callers)
5657 : {
5658 167285 : ipa_node_params *dest_info = ipa_node_params_sum->get (node);
5659 167285 : if (dest_info->ipcp_orig_node)
5660 0 : dest_info = ipa_node_params_sum->get (dest_info->ipcp_orig_node);
5661 :
5662 : /* gather_edges_for_value puts a non-recursive call into the first element of
5663 : callers if it can. */
5664 167285 : push_agg_values_from_edge (callers[0], dest_info, &interim, NULL, true);
5665 :
5666 252292 : unsigned valid_entries = interim.length ();
5667 167285 : if (!valid_entries)
5668 : return 0;
5669 :
5670 83957 : unsigned caller_count = callers.length();
5671 137365 : for (unsigned i = 1; i < caller_count; i++)
5672 : {
5673 55087 : auto_vec<ipa_argagg_value, 32> last;
5674 55087 : ipa_argagg_value_list avs (&interim);
5675 55087 : push_agg_values_from_edge (callers[i], dest_info, &last, &avs, true);
5676 :
5677 55087 : valid_entries = intersect_argaggs_with (interim, last);
5678 55087 : if (!valid_entries)
5679 1679 : return 0;
5680 55087 : }
5681 :
5682 : return valid_entries;
5683 : }
5684 :
5685 : /* Look at edges in CALLERS and collect all known aggregate values that arrive
5686 : from all of them and return them in a garbage-collected vector. Return
5687 : nullptr if there are none. */
5688 :
5689 : static void
5690 150288 : find_aggregate_values_for_callers_subset (vec<ipa_argagg_value> &res,
5691 : struct cgraph_node *node,
5692 : const vec<cgraph_edge *> &callers)
5693 : {
5694 150288 : auto_vec<ipa_argagg_value, 32> interim;
5695 150288 : unsigned valid_entries
5696 150288 : = find_aggregate_values_for_callers_subset_1 (interim, node, callers);
5697 150288 : if (!valid_entries)
5698 : return;
5699 :
5700 833876 : for (const ipa_argagg_value &av : interim)
5701 602600 : if (av.value)
5702 573019 : res.safe_push(av);
5703 : return;
5704 150288 : }
5705 :
5706 : /* Look at edges in CALLERS and collect all known aggregate values that arrive
5707 : from all of them and return them in a garbage-collected vector. Return
5708 : nullptr if there are none. */
5709 :
5710 : static struct vec<ipa_argagg_value, va_gc> *
5711 16997 : find_aggregate_values_for_callers_subset_gc (struct cgraph_node *node,
5712 : const vec<cgraph_edge *> &callers)
5713 : {
5714 16997 : auto_vec<ipa_argagg_value, 32> interim;
5715 16997 : unsigned valid_entries
5716 16997 : = find_aggregate_values_for_callers_subset_1 (interim, node, callers);
5717 16997 : if (!valid_entries)
5718 : return nullptr;
5719 :
5720 5186 : vec<ipa_argagg_value, va_gc> *res = NULL;
5721 5186 : vec_safe_reserve_exact (res, valid_entries);
5722 36527 : for (const ipa_argagg_value &av : interim)
5723 20969 : if (av.value)
5724 19568 : res->quick_push(av);
5725 5186 : gcc_checking_assert (res->length () == valid_entries);
5726 : return res;
5727 16997 : }
5728 :
5729 : /* Determine whether CS also brings all scalar values that the NODE is
5730 : specialized for. */
5731 :
5732 : static bool
5733 77 : cgraph_edge_brings_all_scalars_for_node (struct cgraph_edge *cs,
5734 : struct cgraph_node *node)
5735 : {
5736 77 : ipa_node_params *dest_info = ipa_node_params_sum->get (node);
5737 77 : int count = ipa_get_param_count (dest_info);
5738 77 : class ipa_node_params *caller_info;
5739 77 : class ipa_edge_args *args;
5740 77 : int i;
5741 :
5742 77 : caller_info = ipa_node_params_sum->get (cs->caller);
5743 77 : args = ipa_edge_args_sum->get (cs);
5744 177 : for (i = 0; i < count; i++)
5745 : {
5746 122 : struct ipa_jump_func *jump_func;
5747 122 : tree val, t;
5748 :
5749 122 : val = dest_info->known_csts[i];
5750 122 : if (!val)
5751 72 : continue;
5752 :
5753 100 : if (i >= ipa_get_cs_argument_count (args))
5754 : return false;
5755 50 : jump_func = ipa_get_ith_jump_func (args, i);
5756 50 : t = ipa_value_from_jfunc (caller_info, jump_func,
5757 : ipa_get_type (dest_info, i));
5758 50 : if (!t || !values_equal_for_ipcp_p (val, t))
5759 22 : return false;
5760 : }
5761 : return true;
5762 : }
5763 :
5764 : /* Determine whether CS also brings all aggregate values that NODE is
5765 : specialized for. */
5766 :
5767 : static bool
5768 55 : cgraph_edge_brings_all_agg_vals_for_node (struct cgraph_edge *cs,
5769 : struct cgraph_node *node)
5770 : {
5771 55 : ipcp_transformation *ts = ipcp_get_transformation_summary (node);
5772 55 : if (!ts || vec_safe_is_empty (ts->m_agg_values))
5773 : return true;
5774 :
5775 45 : const ipa_argagg_value_list existing (ts->m_agg_values);
5776 45 : auto_vec<ipa_argagg_value, 32> edge_values;
5777 45 : ipa_node_params *dest_info = ipa_node_params_sum->get (node);
5778 45 : gcc_checking_assert (dest_info->ipcp_orig_node);
5779 45 : dest_info = ipa_node_params_sum->get (dest_info->ipcp_orig_node);
5780 45 : push_agg_values_from_edge (cs, dest_info, &edge_values, &existing, false);
5781 45 : const ipa_argagg_value_list avl (&edge_values);
5782 45 : return avl.superset_of_p (existing);
5783 45 : }
5784 :
5785 : /* Given an original NODE and a VAL for which we have already created a
5786 : specialized clone, look whether there are incoming edges that still lead
5787 : into the old node but now also bring the requested value and also conform to
5788 : all other criteria such that they can be redirected the special node.
5789 : This function can therefore redirect the final edge in a SCC. */
5790 :
5791 : template <typename valtype>
5792 : static void
5793 8758 : perhaps_add_new_callers (cgraph_node *node, ipcp_value<valtype> *val)
5794 : {
5795 : ipcp_value_source<valtype> *src;
5796 8758 : profile_count redirected_sum = profile_count::zero ();
5797 :
5798 122814 : for (src = val->sources; src; src = src->next)
5799 : {
5800 114056 : struct cgraph_edge *cs = src->cs;
5801 353964 : while (cs)
5802 : {
5803 239908 : if (cgraph_edge_brings_value_p (cs, src, node, val)
5804 77 : && cgraph_edge_brings_all_scalars_for_node (cs, val->spec_node)
5805 239963 : && cgraph_edge_brings_all_agg_vals_for_node (cs, val->spec_node))
5806 : {
5807 38 : if (dump_file)
5808 3 : fprintf (dump_file, " - adding an extra caller %s of %s\n",
5809 3 : cs->caller->dump_name (),
5810 3 : val->spec_node->dump_name ());
5811 :
5812 38 : cs->redirect_callee_duplicating_thunks (val->spec_node);
5813 38 : val->spec_node->expand_all_artificial_thunks ();
5814 38 : if (cs->count.ipa ().initialized_p ())
5815 0 : redirected_sum = redirected_sum + cs->count.ipa ();
5816 : }
5817 239908 : cs = get_next_cgraph_edge_clone (cs);
5818 : }
5819 : }
5820 :
5821 8758 : if (redirected_sum.nonzero_p ())
5822 0 : update_specialized_profile (val->spec_node, node, redirected_sum);
5823 8758 : }
5824 :
5825 : /* Return true if KNOWN_CONTEXTS contain at least one useful context. */
5826 :
5827 : static bool
5828 4319 : known_contexts_useful_p (vec<ipa_polymorphic_call_context> known_contexts)
5829 : {
5830 4319 : ipa_polymorphic_call_context *ctx;
5831 4319 : int i;
5832 :
5833 4319 : FOR_EACH_VEC_ELT (known_contexts, i, ctx)
5834 100 : if (!ctx->useless_p ())
5835 : return true;
5836 : return false;
5837 : }
5838 :
5839 : /* Return a copy of KNOWN_CSTS if it is not empty, otherwise return vNULL. */
5840 :
5841 : static vec<ipa_polymorphic_call_context>
5842 4319 : copy_useful_known_contexts (const vec<ipa_polymorphic_call_context> &known_contexts)
5843 : {
5844 4319 : if (known_contexts_useful_p (known_contexts))
5845 100 : return known_contexts.copy ();
5846 : else
5847 4219 : return vNULL;
5848 : }
5849 :
5850 : /* Return true if the VALUE is represented in KNOWN_CSTS at INDEX if OFFSET is
5851 : minus one or in AGGVALS for INDEX and OFFSET otherwise. */
5852 :
5853 : DEBUG_FUNCTION bool
5854 4268 : ipcp_val_replacement_ok_p (vec<tree> &known_csts,
5855 : vec<ipa_polymorphic_call_context> &,
5856 : vec<ipa_argagg_value, va_gc> *aggvals,
5857 : int index, HOST_WIDE_INT offset, tree value)
5858 : {
5859 4268 : tree v;
5860 4268 : if (offset == -1)
5861 3128 : v = known_csts[index];
5862 : else
5863 : {
5864 1140 : const ipa_argagg_value_list avl (aggvals);
5865 1140 : v = avl.get_value (index, offset / BITS_PER_UNIT);
5866 : }
5867 :
5868 4268 : return v && values_equal_for_ipcp_p (v, value);
5869 : }
5870 :
5871 : /* Dump to F all the values in AVALS for which we are re-evaluating the effects
5872 : on the function represented b INFO. */
5873 :
5874 : DEBUG_FUNCTION void
5875 68 : dump_reestimation_message (FILE *f, ipa_node_params *info,
5876 : const ipa_auto_call_arg_values &avals)
5877 : {
5878 68 : fprintf (f, " Re-estimating effects with\n"
5879 : " Scalar constants:");
5880 68 : int param_count = ipa_get_param_count (info);
5881 168 : for (int i = 0; i < param_count; i++)
5882 100 : if (avals.m_known_vals[i])
5883 : {
5884 44 : fprintf (f, " %i:", i);
5885 44 : print_ipcp_constant_value (f, avals.m_known_vals[i]);
5886 : }
5887 68 : fprintf (f, "\n");
5888 68 : if (!avals.m_known_contexts.is_empty ())
5889 : {
5890 0 : fprintf (f, " Pol. contexts:");
5891 0 : for (int i = 0; i < param_count; i++)
5892 0 : if (!avals.m_known_contexts[i].useless_p ())
5893 : {
5894 0 : fprintf (f, " %i:", i);
5895 0 : avals.m_known_contexts[i].dump (f);
5896 : }
5897 0 : fprintf (f, "\n");
5898 : }
5899 68 : if (!avals.m_known_aggs.is_empty ())
5900 : {
5901 24 : fprintf (f, " Aggregate replacements:");
5902 24 : ipa_argagg_value_list avs (&avals);
5903 24 : avs.dump (f);
5904 : }
5905 68 : }
5906 :
5907 : /* Return true if the VALUE is represented in KNOWN_CONTEXTS at INDEX and that
5908 : if OFFSET is is equal to minus one (because source of a polymorphic context
5909 : cannot be an aggregate value). */
5910 :
5911 : DEBUG_FUNCTION bool
5912 51 : ipcp_val_replacement_ok_p (vec<tree> &,
5913 : vec<ipa_polymorphic_call_context> &known_contexts,
5914 : vec<ipa_argagg_value, va_gc> *,
5915 : int index, HOST_WIDE_INT offset,
5916 : ipa_polymorphic_call_context value)
5917 : {
5918 51 : if (offset != -1
5919 51 : || known_contexts.length () <= (unsigned) index
5920 102 : || known_contexts[index].useless_p ())
5921 : return false;
5922 :
5923 51 : if (known_contexts[index].equal_to (value))
5924 : return true;
5925 :
5926 : /* In some corner cases, the final gathering of contexts can figure out that
5927 : the available context is actually more precise than what we wanted to
5928 : clone for. Allow it. */
5929 0 : value.combine_with (known_contexts[index]);
5930 0 : return known_contexts[index].equal_to (value);
5931 : }
5932 :
5933 : /* Decide whether to create a special version of NODE for value VAL of
5934 : parameter at the given INDEX. If OFFSET is -1, the value is for the
5935 : parameter itself, otherwise it is stored at the given OFFSET of the
5936 : parameter. AVALS describes the other already known values. SELF_GEN_CLONES
5937 : is a vector which contains clones created for self-recursive calls with an
5938 : arithmetic pass-through jump function. CUR_SWEEP is the number of the
5939 : current sweep of the call-graph during the decision stage. */
5940 :
5941 : template <typename valtype>
5942 : static bool
5943 220071 : decide_about_value (struct cgraph_node *node, int index, HOST_WIDE_INT offset,
5944 : ipcp_value<valtype> *val,
5945 : vec<cgraph_node *> *self_gen_clones, int cur_sweep)
5946 : {
5947 : int caller_count;
5948 220071 : sreal freq_sum;
5949 : profile_count count_sum, rec_count_sum;
5950 : bool called_without_ipa_profile;
5951 :
5952 220071 : if (val->spec_node)
5953 : {
5954 8758 : perhaps_add_new_callers (node, val);
5955 8758 : return false;
5956 : }
5957 211313 : else if (val->local_size_cost + overall_size > get_max_overall_size (node))
5958 : {
5959 450 : if (dump_file && (dump_flags & TDF_DETAILS))
5960 0 : fprintf (dump_file, " Ignoring candidate value because "
5961 : "maximum unit size would be reached with %li.\n",
5962 : val->local_size_cost + overall_size);
5963 450 : return false;
5964 : }
5965 210863 : else if (!get_info_about_necessary_edges (val, node, &freq_sum, &caller_count,
5966 : &rec_count_sum, &count_sum,
5967 : &called_without_ipa_profile))
5968 : return false;
5969 :
5970 150288 : if (!dbg_cnt (ipa_cp_values))
5971 : return false;
5972 :
5973 150288 : if (val->self_recursion_generated_p ())
5974 : {
5975 : /* The edge counts in this case might not have been adjusted yet.
5976 : Nevertleless, even if they were it would be only a guesswork which we
5977 : can do now. The recursive part of the counts can be derived from the
5978 : count of the original node anyway. */
5979 293 : if (node->count.ipa ().nonzero_p ())
5980 : {
5981 14 : unsigned dem = self_gen_clones->length () + 1;
5982 14 : rec_count_sum = node->count.ipa () / dem;
5983 : }
5984 : else
5985 265 : rec_count_sum = profile_count::zero ();
5986 : }
5987 :
5988 : /* get_info_about_necessary_edges only sums up ipa counts. */
5989 150288 : count_sum += rec_count_sum;
5990 :
5991 150288 : if (dump_file && (dump_flags & TDF_DETAILS))
5992 : {
5993 133 : fprintf (dump_file, " - considering value ");
5994 133 : print_ipcp_constant_value (dump_file, val->value);
5995 133 : fprintf (dump_file, " for ");
5996 133 : ipa_dump_param (dump_file, ipa_node_params_sum->get (node), index);
5997 133 : if (offset != -1)
5998 61 : fprintf (dump_file, ", offset: " HOST_WIDE_INT_PRINT_DEC, offset);
5999 133 : fprintf (dump_file, " (caller_count: %i)\n", caller_count);
6000 : }
6001 :
6002 150288 : auto_vec<cgraph_edge *> callers
6003 : = gather_edges_for_value (val, node, caller_count);
6004 150288 : ipa_node_params *info = ipa_node_params_sum->get (node);
6005 150288 : ipa_auto_call_arg_values avals;
6006 150288 : avals.m_known_vals.safe_grow_cleared (ipa_get_param_count (info), true);
6007 150288 : find_scalar_values_for_callers_subset (avals.m_known_vals, info, callers);
6008 150288 : find_contexts_for_caller_subset (avals.m_known_contexts, info, callers);
6009 150288 : find_aggregate_values_for_callers_subset (avals.m_known_aggs, node, callers);
6010 :
6011 :
6012 150288 : if (good_cloning_opportunity_p (node, val->prop_time_benefit,
6013 : freq_sum, count_sum, val->prop_size_cost,
6014 : called_without_ipa_profile, cur_sweep))
6015 : ;
6016 : else
6017 : {
6018 : /* Extern inline functions are only meaningful to clione to propagate
6019 : values to their callees. */
6020 148417 : if (DECL_EXTERNAL (node->decl) && DECL_DECLARED_INLINE_P (node->decl))
6021 : {
6022 345 : if (dump_file && (dump_flags & TDF_DETAILS))
6023 0 : fprintf (dump_file, " Skipping extern inline.\n");
6024 145969 : return false;
6025 : }
6026 148072 : if (dump_file && (dump_flags & TDF_DETAILS))
6027 68 : dump_reestimation_message (dump_file, info, avals);
6028 :
6029 148072 : ipa_call_estimates estimates;
6030 148072 : estimate_ipcp_clone_size_and_time (node, &avals, &estimates);
6031 148072 : int removable_params_cost = 0;
6032 955846 : for (tree t : avals.m_known_vals)
6033 511630 : if (t)
6034 205371 : removable_params_cost += estimate_move_cost (TREE_TYPE (t), true);
6035 :
6036 148072 : int size = estimates.size - caller_count * removable_params_cost;
6037 :
6038 148072 : if (size <= 0)
6039 : {
6040 1799 : if (dump_file)
6041 0 : fprintf (dump_file, " Code not going to grow.\n");
6042 : }
6043 : else
6044 : {
6045 : sreal time_benefit
6046 146273 : = ((estimates.nonspecialized_time - estimates.time)
6047 292546 : + hint_time_bonus (node, estimates)
6048 146273 : + (devirtualization_time_bonus (node, &avals)
6049 146273 : + removable_params_cost));
6050 :
6051 146273 : if (!good_cloning_opportunity_p (node, time_benefit, freq_sum,
6052 : count_sum, size,
6053 : called_without_ipa_profile,
6054 : cur_sweep))
6055 145624 : return false;
6056 : }
6057 : }
6058 :
6059 4319 : if (dump_file)
6060 137 : fprintf (dump_file, " Creating a specialized node of %s.\n",
6061 : node->dump_name ());
6062 :
6063 4319 : vec<tree> known_csts = avals.m_known_vals.copy ();
6064 : vec<ipa_polymorphic_call_context> known_contexts
6065 4319 : = copy_useful_known_contexts (avals.m_known_contexts);
6066 :
6067 4319 : vec<ipa_argagg_value, va_gc> *aggvals = NULL;
6068 4319 : vec_safe_reserve_exact (aggvals, avals.m_known_aggs.length ());
6069 22972 : for (const ipa_argagg_value &av : avals.m_known_aggs)
6070 10015 : aggvals->quick_push (av);
6071 4319 : gcc_checking_assert (ipcp_val_replacement_ok_p (known_csts, known_contexts,
6072 : aggvals, index,
6073 : offset, val->value));
6074 4319 : val->spec_node = create_specialized_node (node, known_csts, known_contexts,
6075 : aggvals, callers);
6076 :
6077 4319 : if (val->self_recursion_generated_p ())
6078 142 : self_gen_clones->safe_push (val->spec_node);
6079 : else
6080 4177 : update_profiling_info (node, val->spec_node);
6081 :
6082 4319 : overall_size += val->local_size_cost;
6083 4319 : if (dump_file && (dump_flags & TDF_DETAILS))
6084 66 : fprintf (dump_file, " overall size reached %li\n",
6085 : overall_size);
6086 :
6087 : /* TODO: If for some lattice there is only one other known value
6088 : left, make a special node for it too. */
6089 :
6090 : return true;
6091 150288 : }
6092 :
6093 : /* Like irange::contains_p(), but convert VAL to the range of R if
6094 : necessary. */
6095 :
6096 : static inline bool
6097 47690 : ipa_range_contains_p (const vrange &r, tree val)
6098 : {
6099 47690 : if (r.undefined_p ())
6100 : return false;
6101 :
6102 47690 : tree type = r.type ();
6103 47690 : if (!wi::fits_to_tree_p (wi::to_wide (val), type))
6104 : return false;
6105 :
6106 47690 : val = fold_convert (type, val);
6107 47690 : return r.contains_p (val);
6108 : }
6109 :
6110 : /* Structure holding opportunitties so that they can be pre-sorted. */
6111 :
6112 220071 : struct cloning_opportunity_ranking
6113 : {
6114 : /* A very rough evaluation of likely benefit. */
6115 : sreal eval;
6116 : /* In the case of aggregate constants, a non-negative offset within their
6117 : aggregates. -1 for scalar constants, -2 for polymorphic contexts. */
6118 : HOST_WIDE_INT offset;
6119 : /* The value being considered for evaluation for cloning. */
6120 : ipcp_value_base *val;
6121 : /* Index of the formal parameter the value is coming in. */
6122 : int index;
6123 : };
6124 :
6125 : /* Helper function to qsort a vecotr of cloning opportunities. */
6126 :
6127 : static int
6128 2072060 : compare_cloning_opportunities (const void *a, const void *b)
6129 : {
6130 2072060 : const cloning_opportunity_ranking *o1
6131 : = (const cloning_opportunity_ranking *) a;
6132 2072060 : const cloning_opportunity_ranking *o2
6133 : = (const cloning_opportunity_ranking *) b;
6134 2072060 : if (o1->eval < o2->eval)
6135 : return 1;
6136 1625885 : if (o1->eval > o2->eval)
6137 522225 : return -1;
6138 : return 0;
6139 : }
6140 :
6141 : /* Use the estimations in VAL to determine how good a candidate it represents
6142 : for the purposes of ordering real evaluation of opportunities (which
6143 : includes information about incoming edges, among other things). */
6144 :
6145 : static sreal
6146 220071 : cloning_opportunity_ranking_evaluation (const ipcp_value_base *val)
6147 : {
6148 220071 : sreal e1 = (val->local_time_benefit * 1000) / MAX (val->local_size_cost, 1);
6149 220071 : sreal e2 = (val->prop_time_benefit * 1000) / MAX (val->prop_size_cost, 1);
6150 220071 : if (e2 > e1)
6151 15323 : return e2;
6152 : else
6153 204748 : return e1;
6154 : }
6155 :
6156 : /* Decide whether and what specialized clones of NODE should be created.
6157 : CUR_SWEEP is the number of the current sweep of the call-graph during the
6158 : decision stage. */
6159 :
6160 : static bool
6161 3227265 : decide_whether_version_node (struct cgraph_node *node, int cur_sweep)
6162 : {
6163 3227265 : ipa_node_params *info = ipa_node_params_sum->get (node);
6164 3227265 : int count = ipa_get_param_count (info);
6165 3227265 : bool ret = false;
6166 :
6167 3227265 : if (info->node_dead || count == 0)
6168 : return false;
6169 :
6170 2609761 : if (dump_file && (dump_flags & TDF_DETAILS))
6171 344 : fprintf (dump_file, "\nEvaluating opportunities for %s.\n",
6172 : node->dump_name ());
6173 :
6174 2609761 : auto_vec <cloning_opportunity_ranking, 32> opp_ranking;
6175 8704742 : for (int i = 0; i < count;i++)
6176 : {
6177 6094981 : if (!ipa_is_param_used (info, i))
6178 687092 : continue;
6179 :
6180 5407889 : class ipcp_param_lattices *plats = ipa_get_parm_lattices (info, i);
6181 5407889 : ipcp_lattice<tree> *lat = &plats->itself;
6182 5407889 : ipcp_lattice<ipa_polymorphic_call_context> *ctxlat = &plats->ctxlat;
6183 :
6184 5407889 : if (!lat->bottom
6185 5407889 : && !lat->is_single_const ())
6186 : {
6187 534134 : ipcp_value<tree> *val;
6188 649823 : for (val = lat->values; val; val = val->next)
6189 : {
6190 : /* If some values generated for self-recursive calls with
6191 : arithmetic jump functions fall outside of the known
6192 : range for the parameter, we can skip them. */
6193 115751 : if (TREE_CODE (val->value) == INTEGER_CST
6194 70113 : && !plats->m_value_range.bottom_p ()
6195 163379 : && !ipa_range_contains_p (plats->m_value_range.m_vr,
6196 : val->value))
6197 : {
6198 : /* This can happen also if a constant present in the source
6199 : code falls outside of the range of parameter's type, so we
6200 : cannot assert. */
6201 62 : if (dump_file && (dump_flags & TDF_DETAILS))
6202 : {
6203 0 : fprintf (dump_file, " - skipping%s value ",
6204 0 : val->self_recursion_generated_p ()
6205 : ? " self_recursion_generated" : "");
6206 0 : print_ipcp_constant_value (dump_file, val->value);
6207 0 : fprintf (dump_file, " because it is outside known "
6208 : "value range.\n");
6209 : }
6210 62 : continue;
6211 : }
6212 115627 : cloning_opportunity_ranking opp;
6213 115627 : opp.eval = cloning_opportunity_ranking_evaluation (val);
6214 115627 : opp.offset = -1;
6215 115627 : opp.val = val;
6216 115627 : opp.index = i;
6217 115627 : opp_ranking.safe_push (opp);
6218 : }
6219 : }
6220 :
6221 5407889 : if (!plats->aggs_bottom)
6222 : {
6223 563123 : struct ipcp_agg_lattice *aglat;
6224 563123 : ipcp_value<tree> *val;
6225 703480 : for (aglat = plats->aggs; aglat; aglat = aglat->next)
6226 139271 : if (!aglat->bottom && aglat->values
6227 : /* If the following is false, the one value has been considered
6228 : for cloning for all contexts. */
6229 258536 : && (plats->aggs_contain_variable
6230 192483 : || !aglat->is_single_const ()))
6231 177541 : for (val = aglat->values; val; val = val->next)
6232 : {
6233 100648 : cloning_opportunity_ranking opp;
6234 100648 : opp.eval = cloning_opportunity_ranking_evaluation (val);
6235 100648 : opp.offset = aglat->offset;
6236 100648 : opp.val = val;
6237 100648 : opp.index = i;
6238 100648 : opp_ranking.safe_push (opp);
6239 : }
6240 : }
6241 :
6242 5407889 : if (!ctxlat->bottom
6243 6660983 : && !ctxlat->is_single_const ())
6244 : {
6245 548211 : ipcp_value<ipa_polymorphic_call_context> *val;
6246 552007 : for (val = ctxlat->values; val; val = val->next)
6247 7592 : if (!val->value.useless_p ())
6248 : {
6249 3796 : cloning_opportunity_ranking opp;
6250 3796 : opp.eval = cloning_opportunity_ranking_evaluation (val);
6251 3796 : opp.offset = -2;
6252 3796 : opp.val = val;
6253 3796 : opp.index = i;
6254 3796 : opp_ranking.safe_push (opp);
6255 : }
6256 : }
6257 : }
6258 :
6259 2609761 : if (!opp_ranking.is_empty ())
6260 : {
6261 51396 : opp_ranking.qsort (compare_cloning_opportunities);
6262 51396 : auto_vec <cgraph_node *, 9> self_gen_clones;
6263 374259 : for (const cloning_opportunity_ranking &opp : opp_ranking)
6264 220071 : if (opp.offset == -2)
6265 : {
6266 3796 : ipcp_value<ipa_polymorphic_call_context> *val
6267 : = static_cast <ipcp_value<ipa_polymorphic_call_context> *>
6268 : (opp.val);
6269 3796 : ret |= decide_about_value (node, opp.index, -1, val,
6270 : &self_gen_clones, cur_sweep);
6271 : }
6272 : else
6273 : {
6274 216275 : ipcp_value<tree> *val = static_cast<ipcp_value<tree> *> (opp.val);
6275 216275 : ret |= decide_about_value (node, opp.index, opp.offset, val,
6276 : &self_gen_clones, cur_sweep);
6277 : }
6278 :
6279 102792 : if (!self_gen_clones.is_empty ())
6280 : {
6281 33 : self_gen_clones.safe_push (node);
6282 33 : update_counts_for_self_gen_clones (node, self_gen_clones);
6283 : }
6284 51396 : }
6285 :
6286 2609761 : struct caller_statistics stats;
6287 2609761 : init_caller_stats (&stats);
6288 2609761 : node->call_for_symbol_thunks_and_aliases (gather_caller_stats, &stats,
6289 : false);
6290 2609761 : if (!stats.n_calls)
6291 : {
6292 1303067 : if (dump_file)
6293 942 : fprintf (dump_file, " Not cloning for all contexts because "
6294 : "there are no callers of the original node (any more).\n");
6295 1303067 : return ret;
6296 : }
6297 :
6298 1306694 : bool do_clone_for_all_contexts = false;
6299 1306694 : ipa_auto_call_arg_values avals;
6300 1306694 : int removable_params_cost;
6301 1306694 : bool ctx_independent_const
6302 1306694 : = gather_context_independent_values (info, &avals, &removable_params_cost);
6303 1306694 : sreal devirt_bonus = devirtualization_time_bonus (node, &avals);
6304 1292149 : if (ctx_independent_const || devirt_bonus > 0
6305 2598837 : || (removable_params_cost && clone_for_param_removal_p (node)))
6306 : {
6307 63642 : ipa_call_estimates estimates;
6308 :
6309 63642 : estimate_ipcp_clone_size_and_time (node, &avals, &estimates);
6310 63642 : sreal time = estimates.nonspecialized_time - estimates.time;
6311 63642 : time += devirt_bonus;
6312 63642 : time += hint_time_bonus (node, estimates);
6313 63642 : time += removable_params_cost;
6314 63642 : int size = estimates.size - stats.n_calls * removable_params_cost;
6315 :
6316 63642 : if (dump_file && (dump_flags & TDF_DETAILS))
6317 24 : fprintf (dump_file, " - context independent values, size: %i, "
6318 : "time_benefit: %f\n", size, (time).to_double ());
6319 :
6320 63642 : if (size <= 0 || node->local)
6321 : {
6322 16900 : if (!dbg_cnt (ipa_cp_values))
6323 0 : return ret;
6324 :
6325 16900 : do_clone_for_all_contexts = true;
6326 16900 : if (dump_file)
6327 106 : fprintf (dump_file, " Decided to specialize for all "
6328 : "known contexts, code not going to grow.\n");
6329 : }
6330 46742 : else if (good_cloning_opportunity_p (node, time, stats.freq_sum,
6331 : stats.count_sum, size,
6332 46742 : stats.called_without_ipa_profile,
6333 : cur_sweep))
6334 : {
6335 193 : if (size + overall_size <= get_max_overall_size (node))
6336 : {
6337 193 : if (!dbg_cnt (ipa_cp_values))
6338 : return ret;
6339 :
6340 193 : do_clone_for_all_contexts = true;
6341 193 : overall_size += size;
6342 193 : if (dump_file)
6343 11 : fprintf (dump_file, " Decided to specialize for all "
6344 : "known contexts, growth (to %li) deemed "
6345 : "beneficial.\n", overall_size);
6346 : }
6347 0 : else if (dump_file && (dump_flags & TDF_DETAILS))
6348 0 : fprintf (dump_file, " Not cloning for all contexts because "
6349 : "maximum unit size would be reached with %li.\n",
6350 : size + overall_size);
6351 : }
6352 46549 : else if (dump_file && (dump_flags & TDF_DETAILS))
6353 0 : fprintf (dump_file, " Not cloning for all contexts because "
6354 : "!good_cloning_opportunity_p.\n");
6355 : }
6356 :
6357 1306694 : if (do_clone_for_all_contexts)
6358 : {
6359 17093 : auto_vec<cgraph_edge *> callers = node->collect_callers ();
6360 :
6361 87453 : for (int i = callers.length () - 1; i >= 0; i--)
6362 : {
6363 53267 : cgraph_edge *cs = callers[i];
6364 53267 : ipa_node_params *caller_info = ipa_node_params_sum->get (cs->caller);
6365 :
6366 53267 : if (caller_info && caller_info->node_dead)
6367 2703 : callers.unordered_remove (i);
6368 : }
6369 :
6370 17093 : if (!adjust_callers_for_value_intersection (callers, node))
6371 : /* If node is not called by anyone, or all its caller edges are
6372 : self-recursive, the node is not really in use, no need to do
6373 : cloning. */
6374 96 : return ret;
6375 :
6376 16997 : if (dump_file)
6377 115 : fprintf (dump_file, " Creating a specialized node of %s "
6378 : "for all known contexts.\n", node->dump_name ());
6379 :
6380 16997 : vec<tree> known_csts = vNULL;
6381 16997 : known_csts.safe_grow_cleared (count, true);
6382 16997 : find_scalar_values_for_callers_subset (known_csts, info, callers);
6383 16997 : vec<ipa_polymorphic_call_context> known_contexts = vNULL;
6384 16997 : find_contexts_for_caller_subset (known_contexts, info, callers);
6385 16997 : vec<ipa_argagg_value, va_gc> *aggvals
6386 16997 : = find_aggregate_values_for_callers_subset_gc (node, callers);
6387 :
6388 16997 : struct cgraph_node *clone = create_specialized_node (node, known_csts,
6389 : known_contexts,
6390 : aggvals, callers);
6391 16997 : ipa_node_params_sum->get (clone)->is_all_contexts_clone = true;
6392 16997 : ret = true;
6393 17093 : }
6394 :
6395 : return ret;
6396 3916455 : }
6397 :
6398 : /* Transitively mark all callees of NODE within the same SCC as not dead. */
6399 :
6400 : static void
6401 4627 : spread_undeadness (struct cgraph_node *node)
6402 : {
6403 4627 : struct cgraph_edge *cs;
6404 :
6405 15622 : for (cs = node->callees; cs; cs = cs->next_callee)
6406 10995 : if (ipa_edge_within_scc (cs))
6407 : {
6408 832 : struct cgraph_node *callee;
6409 832 : class ipa_node_params *info;
6410 :
6411 832 : callee = cs->callee->function_symbol (NULL);
6412 832 : info = ipa_node_params_sum->get (callee);
6413 :
6414 832 : if (info && info->node_dead)
6415 : {
6416 68 : info->node_dead = 0;
6417 68 : spread_undeadness (callee);
6418 : }
6419 : }
6420 4627 : }
6421 :
6422 : /* Return true if NODE has a caller from outside of its SCC that is not
6423 : dead. Worker callback for cgraph_for_node_and_aliases. */
6424 :
6425 : static bool
6426 15835 : has_undead_caller_from_outside_scc_p (struct cgraph_node *node,
6427 : void *data ATTRIBUTE_UNUSED)
6428 : {
6429 15835 : struct cgraph_edge *cs;
6430 :
6431 81123 : for (cs = node->callers; cs; cs = cs->next_caller)
6432 65741 : if (cs->caller->thunk
6433 65741 : && cs->caller->call_for_symbol_thunks_and_aliases
6434 0 : (has_undead_caller_from_outside_scc_p, NULL, true))
6435 : return true;
6436 65741 : else if (!ipa_edge_within_scc (cs))
6437 : {
6438 65489 : ipa_node_params *caller_info = ipa_node_params_sum->get (cs->caller);
6439 65489 : if (!caller_info /* Unoptimized caller are like dead ones. */
6440 65487 : || !caller_info->node_dead)
6441 : return true;
6442 : }
6443 : return false;
6444 : }
6445 :
6446 :
6447 : /* Identify nodes within the same SCC as NODE which are no longer needed
6448 : because of new clones and will be removed as unreachable. */
6449 :
6450 : static void
6451 19363 : identify_dead_nodes (struct cgraph_node *node)
6452 : {
6453 19363 : struct cgraph_node *v;
6454 39006 : for (v = node; v; v = ((struct ipa_dfs_info *) v->aux)->next_cycle)
6455 19643 : if (v->local)
6456 : {
6457 15596 : ipa_node_params *info = ipa_node_params_sum->get (v);
6458 15596 : if (info
6459 31192 : && !v->call_for_symbol_thunks_and_aliases
6460 15596 : (has_undead_caller_from_outside_scc_p, NULL, true))
6461 15143 : info->node_dead = 1;
6462 : }
6463 :
6464 39006 : for (v = node; v; v = ((struct ipa_dfs_info *) v->aux)->next_cycle)
6465 : {
6466 19643 : ipa_node_params *info = ipa_node_params_sum->get (v);
6467 19643 : if (info && !info->node_dead)
6468 4559 : spread_undeadness (v);
6469 : }
6470 :
6471 19363 : if (dump_file && (dump_flags & TDF_DETAILS))
6472 : {
6473 105 : for (v = node; v; v = ((struct ipa_dfs_info *) v->aux)->next_cycle)
6474 54 : if (ipa_node_params_sum->get (v)
6475 54 : && ipa_node_params_sum->get (v)->node_dead)
6476 32 : fprintf (dump_file, " Marking node as dead: %s.\n",
6477 : v->dump_name ());
6478 : }
6479 19363 : }
6480 :
6481 : /* Removes all useless callback edges from the callgraph. Useless callback
6482 : edges might mess up the callgraph, because they might be impossible to
6483 : redirect and so on, leading to crashes. Their usefulness is evaluated
6484 : through callback_edge_useful_p. */
6485 :
6486 : static void
6487 129027 : purge_useless_callback_edges ()
6488 : {
6489 129027 : if (dump_file)
6490 161 : fprintf (dump_file, "\nPurging useless callback edges:\n");
6491 :
6492 129027 : cgraph_edge *e;
6493 129027 : cgraph_node *node;
6494 1428823 : FOR_EACH_FUNCTION_WITH_GIMPLE_BODY (node)
6495 : {
6496 6757435 : for (e = node->callees; e; e = e->next_callee)
6497 : {
6498 5457639 : if (e->has_callback)
6499 : {
6500 13648 : if (dump_file)
6501 3 : fprintf (dump_file, "\tExamining callbacks of edge %s -> %s:\n",
6502 3 : e->caller->dump_name (), e->callee->dump_name ());
6503 13648 : if (!lookup_attribute (CALLBACK_ATTR_IDENT,
6504 13648 : DECL_ATTRIBUTES (e->callee->decl))
6505 13648 : && !callback_is_special_cased (e->callee->decl, e->call_stmt))
6506 : {
6507 1 : if (dump_file)
6508 0 : fprintf (
6509 : dump_file,
6510 : "\t\tPurging callbacks, because the callback-dispatching"
6511 : "function no longer has any callback attributes.\n");
6512 1 : e->purge_callback_edges ();
6513 1 : continue;
6514 : }
6515 13647 : cgraph_edge *cbe, *next;
6516 27294 : for (cbe = e->first_callback_edge (); cbe; cbe = next)
6517 : {
6518 13647 : next = cbe->next_callback_edge ();
6519 13647 : if (!callback_edge_useful_p (cbe))
6520 : {
6521 13285 : if (dump_file)
6522 1 : fprintf (dump_file,
6523 : "\t\tCallback edge %s -> %s not deemed "
6524 : "useful, removing.\n",
6525 1 : cbe->caller->dump_name (),
6526 1 : cbe->callee->dump_name ());
6527 13285 : cgraph_edge::remove (cbe);
6528 : }
6529 : else
6530 : {
6531 362 : if (dump_file)
6532 2 : fprintf (dump_file,
6533 : "\t\tKept callback edge %s -> %s "
6534 : "because it looks useful.\n",
6535 2 : cbe->caller->dump_name (),
6536 2 : cbe->callee->dump_name ());
6537 : }
6538 : }
6539 : }
6540 : }
6541 : }
6542 :
6543 129027 : if (dump_file)
6544 161 : fprintf (dump_file, "\n");
6545 129027 : }
6546 :
6547 : /* The decision stage. Iterate over the topological order of call graph nodes
6548 : TOPO and make specialized clones if deemed beneficial. */
6549 :
6550 : static void
6551 129027 : ipcp_decision_stage (class ipa_topo_info *topo)
6552 : {
6553 129027 : int i;
6554 :
6555 129027 : if (dump_file)
6556 161 : fprintf (dump_file, "\nIPA decision stage (%i sweeps):\n",
6557 : max_number_sweeps);
6558 :
6559 495323 : for (int cur_sweep = 1; cur_sweep <= max_number_sweeps; cur_sweep++)
6560 : {
6561 366296 : if (dump_file && (dump_flags & TDF_DETAILS))
6562 144 : fprintf (dump_file, "\nIPA decision sweep number %i (out of %i):\n",
6563 : cur_sweep, max_number_sweeps);
6564 :
6565 4390176 : for (i = topo->nnodes - 1; i >= 0; i--)
6566 : {
6567 4023880 : struct cgraph_node *node = topo->order[i];
6568 4023880 : bool change = false, iterate = true;
6569 :
6570 8067130 : while (iterate)
6571 : {
6572 : struct cgraph_node *v;
6573 : iterate = false;
6574 4058100 : for (v = node;
6575 8101350 : v;
6576 4058100 : v = ((struct ipa_dfs_info *) v->aux)->next_cycle)
6577 4058100 : if (v->has_gimple_body_p ()
6578 3830418 : && ipcp_versionable_function_p (v)
6579 4058100 : && (cur_sweep
6580 3227265 : <= opt_for_fn (node->decl, param_ipa_cp_sweeps)))
6581 3227265 : iterate |= decide_whether_version_node (v, cur_sweep);
6582 :
6583 4043250 : change |= iterate;
6584 : }
6585 4023880 : if (change)
6586 19363 : identify_dead_nodes (node);
6587 : }
6588 : }
6589 :
6590 : /* Currently, the primary use of callback edges is constant propagation.
6591 : Constant propagation is now over, so we have to remove unused callback
6592 : edges. */
6593 129027 : purge_useless_callback_edges ();
6594 129027 : }
6595 :
6596 : /* Look up all VR and bits information that we have discovered and copy it
6597 : over to the transformation summary. */
6598 :
6599 : static void
6600 129027 : ipcp_store_vr_results (void)
6601 : {
6602 129027 : cgraph_node *node;
6603 :
6604 1428823 : FOR_EACH_FUNCTION_WITH_GIMPLE_BODY (node)
6605 : {
6606 1299796 : ipa_node_params *info = ipa_node_params_sum->get (node);
6607 1299796 : bool dumped_sth = false;
6608 1299796 : bool found_useful_result = false;
6609 1299796 : bool do_vr = true;
6610 1299796 : bool do_bits = true;
6611 :
6612 : /* If the function is not local, the gathered information is only useful
6613 : for clones. */
6614 1299796 : if (!node->local)
6615 1134174 : continue;
6616 :
6617 165622 : if (!info || !opt_for_fn (node->decl, flag_ipa_vrp))
6618 : {
6619 4819 : if (dump_file)
6620 6 : fprintf (dump_file, "Not considering %s for VR discovery "
6621 : "and propagate; -fipa-ipa-vrp: disabled.\n",
6622 : node->dump_name ());
6623 : do_vr = false;
6624 : }
6625 165622 : if (!info || !opt_for_fn (node->decl, flag_ipa_bit_cp))
6626 : {
6627 4790 : if (dump_file)
6628 2 : fprintf (dump_file, "Not considering %s for ipa bitwise "
6629 : "propagation ; -fipa-bit-cp: disabled.\n",
6630 : node->dump_name ());
6631 4790 : do_bits = false;
6632 : }
6633 4790 : if (!do_bits && !do_vr)
6634 4784 : continue;
6635 :
6636 160838 : if (info->ipcp_orig_node)
6637 21123 : info = ipa_node_params_sum->get (info->ipcp_orig_node);
6638 160838 : if (info->lattices.is_empty ())
6639 : /* Newly expanded artificial thunks do not have lattices. */
6640 51230 : continue;
6641 :
6642 109608 : unsigned count = ipa_get_param_count (info);
6643 226903 : for (unsigned i = 0; i < count; i++)
6644 : {
6645 177288 : ipcp_param_lattices *plats = ipa_get_parm_lattices (info, i);
6646 177288 : if (do_vr
6647 177262 : && !plats->m_value_range.bottom_p ()
6648 235112 : && !plats->m_value_range.top_p ())
6649 : {
6650 : found_useful_result = true;
6651 : break;
6652 : }
6653 119465 : if (do_bits && plats->bits_lattice.constant_p ())
6654 : {
6655 : found_useful_result = true;
6656 : break;
6657 : }
6658 : }
6659 109608 : if (!found_useful_result)
6660 49615 : continue;
6661 :
6662 59993 : ipcp_transformation_initialize ();
6663 59993 : ipcp_transformation *ts = ipcp_transformation_sum->get_create (node);
6664 59993 : vec_safe_reserve_exact (ts->m_vr, count);
6665 :
6666 218177 : for (unsigned i = 0; i < count; i++)
6667 : {
6668 158184 : ipcp_param_lattices *plats = ipa_get_parm_lattices (info, i);
6669 158184 : ipcp_bits_lattice *bits = NULL;
6670 :
6671 158184 : if (do_bits
6672 158180 : && plats->bits_lattice.constant_p ()
6673 249408 : && dbg_cnt (ipa_cp_bits))
6674 91224 : bits = &plats->bits_lattice;
6675 :
6676 158184 : if (do_vr
6677 158160 : && !plats->m_value_range.bottom_p ()
6678 109721 : && !plats->m_value_range.top_p ()
6679 267905 : && dbg_cnt (ipa_cp_vr))
6680 : {
6681 109721 : if (bits)
6682 : {
6683 86014 : value_range tmp = plats->m_value_range.m_vr;
6684 86014 : tree type = ipa_get_type (info, i);
6685 172028 : irange_bitmask bm (wide_int::from (bits->get_value (),
6686 86014 : TYPE_PRECISION (type),
6687 86014 : TYPE_SIGN (type)),
6688 172028 : wide_int::from (bits->get_mask (),
6689 86014 : TYPE_PRECISION (type),
6690 172028 : TYPE_SIGN (type)));
6691 86014 : tmp.update_bitmask (bm);
6692 : // Reflecting the bitmask on the ranges can sometime
6693 : // produce an UNDEFINED value if the the bitmask update
6694 : // was previously deferred. See PR 120048.
6695 86014 : if (tmp.undefined_p ())
6696 0 : tmp.set_varying (type);
6697 86014 : ipa_vr vr (tmp);
6698 86014 : ts->m_vr->quick_push (vr);
6699 86014 : }
6700 : else
6701 : {
6702 23707 : ipa_vr vr (plats->m_value_range.m_vr);
6703 23707 : ts->m_vr->quick_push (vr);
6704 : }
6705 : }
6706 48463 : else if (bits)
6707 : {
6708 5210 : tree type = ipa_get_type (info, i);
6709 5210 : value_range tmp;
6710 5210 : tmp.set_varying (type);
6711 10420 : irange_bitmask bm (wide_int::from (bits->get_value (),
6712 5210 : TYPE_PRECISION (type),
6713 5210 : TYPE_SIGN (type)),
6714 10420 : wide_int::from (bits->get_mask (),
6715 5210 : TYPE_PRECISION (type),
6716 10420 : TYPE_SIGN (type)));
6717 5210 : tmp.update_bitmask (bm);
6718 : // Reflecting the bitmask on the ranges can sometime
6719 : // produce an UNDEFINED value if the the bitmask update
6720 : // was previously deferred. See PR 120048.
6721 5210 : if (tmp.undefined_p ())
6722 0 : tmp.set_varying (type);
6723 5210 : ipa_vr vr (tmp);
6724 5210 : ts->m_vr->quick_push (vr);
6725 5210 : }
6726 : else
6727 : {
6728 43253 : ipa_vr vr;
6729 43253 : ts->m_vr->quick_push (vr);
6730 : }
6731 :
6732 158184 : if (!dump_file || !bits)
6733 157769 : continue;
6734 :
6735 415 : if (!dumped_sth)
6736 : {
6737 296 : fprintf (dump_file, "Propagated bits info for function %s:\n",
6738 : node->dump_name ());
6739 296 : dumped_sth = true;
6740 : }
6741 415 : fprintf (dump_file, " param %i: value = ", i);
6742 415 : ipcp_print_widest_int (dump_file, bits->get_value ());
6743 415 : fprintf (dump_file, ", mask = ");
6744 415 : ipcp_print_widest_int (dump_file, bits->get_mask ());
6745 415 : fprintf (dump_file, "\n");
6746 : }
6747 : }
6748 129027 : }
6749 :
6750 : /* The IPCP driver. */
6751 :
6752 : static unsigned int
6753 129027 : ipcp_driver (void)
6754 : {
6755 129027 : class ipa_topo_info topo;
6756 :
6757 129027 : if (edge_clone_summaries == NULL)
6758 129027 : edge_clone_summaries = new edge_clone_summary_t (symtab);
6759 :
6760 129027 : ipa_check_create_node_params ();
6761 129027 : ipa_check_create_edge_args ();
6762 129027 : clone_num_suffixes = new hash_map<const char *, unsigned>;
6763 :
6764 129027 : if (dump_file)
6765 : {
6766 161 : fprintf (dump_file, "\nIPA structures before propagation:\n");
6767 161 : if (dump_flags & TDF_DETAILS)
6768 48 : ipa_print_all_params (dump_file);
6769 161 : ipa_print_all_jump_functions (dump_file);
6770 : }
6771 :
6772 : /* Topological sort. */
6773 129027 : build_toporder_info (&topo);
6774 : /* Do the interprocedural propagation. */
6775 129027 : ipcp_propagate_stage (&topo);
6776 : /* Decide what constant propagation and cloning should be performed. */
6777 129027 : ipcp_decision_stage (&topo);
6778 : /* Store results of value range and bits propagation. */
6779 129027 : ipcp_store_vr_results ();
6780 :
6781 : /* Free all IPCP structures. */
6782 258054 : delete clone_num_suffixes;
6783 129027 : free_toporder_info (&topo);
6784 129027 : delete edge_clone_summaries;
6785 129027 : edge_clone_summaries = NULL;
6786 129027 : ipa_free_all_structures_after_ipa_cp ();
6787 129027 : if (dump_file)
6788 161 : fprintf (dump_file, "\nIPA constant propagation end\n");
6789 129027 : return 0;
6790 : }
6791 :
6792 : /* Initialization and computation of IPCP data structures. This is the initial
6793 : intraprocedural analysis of functions, which gathers information to be
6794 : propagated later on. */
6795 :
6796 : static void
6797 125886 : ipcp_generate_summary (void)
6798 : {
6799 125886 : struct cgraph_node *node;
6800 :
6801 125886 : if (dump_file)
6802 163 : fprintf (dump_file, "\nIPA constant propagation start:\n");
6803 125886 : ipa_register_cgraph_hooks ();
6804 :
6805 1380770 : FOR_EACH_FUNCTION_WITH_GIMPLE_BODY (node)
6806 1254884 : ipa_analyze_node (node);
6807 :
6808 125886 : varpool_node *vnode;
6809 1799698 : FOR_EACH_STATIC_INITIALIZER (vnode)
6810 1673812 : ipa_analyze_var_static_initializer (vnode);
6811 125886 : }
6812 :
6813 : namespace {
6814 :
6815 : const pass_data pass_data_ipa_cp =
6816 : {
6817 : IPA_PASS, /* type */
6818 : "cp", /* name */
6819 : OPTGROUP_NONE, /* optinfo_flags */
6820 : TV_IPA_CONSTANT_PROP, /* tv_id */
6821 : 0, /* properties_required */
6822 : 0, /* properties_provided */
6823 : 0, /* properties_destroyed */
6824 : 0, /* todo_flags_start */
6825 : ( TODO_dump_symtab | TODO_remove_functions ), /* todo_flags_finish */
6826 : };
6827 :
6828 : class pass_ipa_cp : public ipa_opt_pass_d
6829 : {
6830 : public:
6831 288767 : pass_ipa_cp (gcc::context *ctxt)
6832 : : ipa_opt_pass_d (pass_data_ipa_cp, ctxt,
6833 : ipcp_generate_summary, /* generate_summary */
6834 : NULL, /* write_summary */
6835 : NULL, /* read_summary */
6836 : ipcp_write_transformation_summaries, /*
6837 : write_optimization_summary */
6838 : ipcp_read_transformation_summaries, /*
6839 : read_optimization_summary */
6840 : NULL, /* stmt_fixup */
6841 : 0, /* function_transform_todo_flags_start */
6842 : ipcp_transform_function, /* function_transform */
6843 288767 : NULL) /* variable_transform */
6844 288767 : {}
6845 :
6846 : /* opt_pass methods: */
6847 577224 : bool gate (function *) final override
6848 : {
6849 : /* FIXME: We should remove the optimize check after we ensure we never run
6850 : IPA passes when not optimizing. */
6851 577224 : return (flag_ipa_cp && optimize) || in_lto_p;
6852 : }
6853 :
6854 129027 : unsigned int execute (function *) final override { return ipcp_driver (); }
6855 :
6856 : }; // class pass_ipa_cp
6857 :
6858 : } // anon namespace
6859 :
6860 : ipa_opt_pass_d *
6861 288767 : make_pass_ipa_cp (gcc::context *ctxt)
6862 : {
6863 288767 : return new pass_ipa_cp (ctxt);
6864 : }
6865 :
6866 : /* Reset all state within ipa-cp.cc so that we can rerun the compiler
6867 : within the same process. For use by toplev::finalize. */
6868 :
6869 : void
6870 259439 : ipa_cp_cc_finalize (void)
6871 : {
6872 259439 : overall_size = 0;
6873 259439 : orig_overall_size = 0;
6874 259439 : ipcp_free_transformation_sum ();
6875 259439 : }
6876 :
6877 : /* Given PARAM which must be a parameter of function FNDECL described by THIS,
6878 : return its index in the DECL_ARGUMENTS chain, using a pre-computed
6879 : DECL_UID-sorted vector if available (which is pre-computed only if there are
6880 : many parameters). Can return -1 if param is static chain not represented
6881 : among DECL_ARGUMENTS. */
6882 :
6883 : int
6884 124963 : ipcp_transformation::get_param_index (const_tree fndecl, const_tree param) const
6885 : {
6886 124963 : gcc_assert (TREE_CODE (param) == PARM_DECL);
6887 124963 : if (m_uid_to_idx)
6888 : {
6889 0 : unsigned puid = DECL_UID (param);
6890 0 : const ipa_uid_to_idx_map_elt *res
6891 0 : = std::lower_bound (m_uid_to_idx->begin(), m_uid_to_idx->end (), puid,
6892 0 : [] (const ipa_uid_to_idx_map_elt &elt, unsigned uid)
6893 : {
6894 0 : return elt.uid < uid;
6895 : });
6896 0 : if (res == m_uid_to_idx->end ()
6897 0 : || res->uid != puid)
6898 : {
6899 0 : gcc_assert (DECL_STATIC_CHAIN (fndecl));
6900 : return -1;
6901 : }
6902 0 : return res->index;
6903 : }
6904 :
6905 124963 : unsigned index = 0;
6906 284524 : for (tree p = DECL_ARGUMENTS (fndecl); p; p = DECL_CHAIN (p), index++)
6907 283070 : if (p == param)
6908 123509 : return (int) index;
6909 :
6910 1454 : gcc_assert (DECL_STATIC_CHAIN (fndecl));
6911 : return -1;
6912 : }
6913 :
6914 : /* Helper function to qsort a vector of ipa_uid_to_idx_map_elt elements
6915 : according to the uid. */
6916 :
6917 : static int
6918 0 : compare_uids (const void *a, const void *b)
6919 : {
6920 0 : const ipa_uid_to_idx_map_elt *e1 = (const ipa_uid_to_idx_map_elt *) a;
6921 0 : const ipa_uid_to_idx_map_elt *e2 = (const ipa_uid_to_idx_map_elt *) b;
6922 0 : if (e1->uid < e2->uid)
6923 : return -1;
6924 0 : if (e1->uid > e2->uid)
6925 : return 1;
6926 0 : gcc_unreachable ();
6927 : }
6928 :
6929 : /* Assuming THIS describes FNDECL and it has sufficiently many parameters to
6930 : justify the overhead, create a DECL_UID-sorted vector to speed up mapping
6931 : from parameters to their indices in DECL_ARGUMENTS chain. */
6932 :
6933 : void
6934 22912 : ipcp_transformation::maybe_create_parm_idx_map (tree fndecl)
6935 : {
6936 22912 : int c = count_formal_params (fndecl);
6937 22912 : if (c < 32)
6938 : return;
6939 :
6940 0 : m_uid_to_idx = NULL;
6941 0 : vec_safe_reserve (m_uid_to_idx, c, true);
6942 0 : unsigned index = 0;
6943 0 : for (tree p = DECL_ARGUMENTS (fndecl); p; p = DECL_CHAIN (p), index++)
6944 : {
6945 0 : ipa_uid_to_idx_map_elt elt;
6946 0 : elt.uid = DECL_UID (p);
6947 0 : elt.index = index;
6948 0 : m_uid_to_idx->quick_push (elt);
6949 : }
6950 0 : m_uid_to_idx->qsort (compare_uids);
6951 : }
|
