Line data Source code
1 : /* Function summary pass.
2 : Copyright (C) 2003-2026 Free Software Foundation, Inc.
3 : Contributed by Jan Hubicka
4 :
5 : This file is part of GCC.
6 :
7 : GCC is free software; you can redistribute it and/or modify it under
8 : the terms of the GNU General Public License as published by the Free
9 : Software Foundation; either version 3, or (at your option) any later
10 : version.
11 :
12 : GCC is distributed in the hope that it will be useful, but WITHOUT ANY
13 : WARRANTY; without even the implied warranty of MERCHANTABILITY or
14 : FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
15 : for more details.
16 :
17 : You should have received a copy of the GNU General Public License
18 : along with GCC; see the file COPYING3. If not see
19 : <http://www.gnu.org/licenses/>. */
20 :
21 : /* Analysis of function bodies used by inter-procedural passes
22 :
23 : We estimate for each function
24 : - function body size and size after specializing into given context
25 : - average function execution time in a given context
26 : - function frame size
27 : For each call
28 : - call statement size, time and how often the parameters change
29 :
30 : ipa_fn_summary data structures store above information locally (i.e.
31 : parameters of the function itself) and globally (i.e. parameters of
32 : the function created by applying all the inline decisions already
33 : present in the callgraph).
34 :
35 : We provide access to the ipa_fn_summary data structure and
36 : basic logic updating the parameters when inlining is performed.
37 :
38 : The summaries are context sensitive. Context means
39 : 1) partial assignment of known constant values of operands
40 : 2) whether function is inlined into the call or not.
41 : It is easy to add more variants. To represent function size and time
42 : that depends on context (i.e. it is known to be optimized away when
43 : context is known either by inlining or from IP-CP and cloning),
44 : we use predicates.
45 :
46 : estimate_edge_size_and_time can be used to query
47 : function size/time in the given context. ipa_merge_fn_summary_after_inlining merges
48 : properties of caller and callee after inlining.
49 :
50 : Finally pass_inline_parameters is exported. This is used to drive
51 : computation of function parameters used by the early inliner. IPA
52 : inlined performs analysis via its analyze_function method. */
53 :
54 : #include "config.h"
55 : #define INCLUDE_VECTOR
56 : #include "system.h"
57 : #include "coretypes.h"
58 : #include "backend.h"
59 : #include "target.h"
60 : #include "tree.h"
61 : #include "gimple.h"
62 : #include "alloc-pool.h"
63 : #include "tree-pass.h"
64 : #include "ssa.h"
65 : #include "tree-streamer.h"
66 : #include "cgraph.h"
67 : #include "diagnostic.h"
68 : #include "fold-const.h"
69 : #include "print-tree.h"
70 : #include "tree-inline.h"
71 : #include "gimple-pretty-print.h"
72 : #include "cfganal.h"
73 : #include "gimple-iterator.h"
74 : #include "tree-cfg.h"
75 : #include "tree-ssa-loop-niter.h"
76 : #include "tree-ssa-loop.h"
77 : #include "symbol-summary.h"
78 : #include "sreal.h"
79 : #include "ipa-cp.h"
80 : #include "ipa-prop.h"
81 : #include "ipa-fnsummary.h"
82 : #include "cfgloop.h"
83 : #include "tree-scalar-evolution.h"
84 : #include "ipa-utils.h"
85 : #include "cfgexpand.h"
86 : #include "gimplify.h"
87 : #include "stringpool.h"
88 : #include "attribs.h"
89 : #include "tree-into-ssa.h"
90 : #include "symtab-clones.h"
91 : #include "gimple-range.h"
92 : #include "tree-dfa.h"
93 :
94 : /* Summaries. */
95 : fast_function_summary <ipa_fn_summary *, va_gc> *ipa_fn_summaries;
96 : fast_function_summary <ipa_size_summary *, va_heap> *ipa_size_summaries;
97 : fast_call_summary <ipa_call_summary *, va_heap> *ipa_call_summaries;
98 :
99 : /* Edge predicates goes here. */
100 : static object_allocator<ipa_predicate> edge_predicate_pool ("edge predicates");
101 :
102 :
103 : /* Dump IPA hints. */
104 : void
105 185 : ipa_dump_hints (FILE *f, ipa_hints hints)
106 : {
107 185 : if (!hints)
108 : return;
109 148 : fprintf (f, "IPA hints:");
110 148 : if (hints & INLINE_HINT_indirect_call)
111 : {
112 23 : hints &= ~INLINE_HINT_indirect_call;
113 23 : fprintf (f, " indirect_call");
114 : }
115 148 : if (hints & INLINE_HINT_loop_iterations)
116 : {
117 3 : hints &= ~INLINE_HINT_loop_iterations;
118 3 : fprintf (f, " loop_iterations");
119 : }
120 148 : if (hints & INLINE_HINT_loop_stride)
121 : {
122 3 : hints &= ~INLINE_HINT_loop_stride;
123 3 : fprintf (f, " loop_stride");
124 : }
125 148 : if (hints & INLINE_HINT_same_scc)
126 : {
127 4 : hints &= ~INLINE_HINT_same_scc;
128 4 : fprintf (f, " same_scc");
129 : }
130 148 : if (hints & INLINE_HINT_in_scc)
131 : {
132 11 : hints &= ~INLINE_HINT_in_scc;
133 11 : fprintf (f, " in_scc");
134 : }
135 148 : if (hints & INLINE_HINT_cross_module)
136 : {
137 2 : hints &= ~INLINE_HINT_cross_module;
138 2 : fprintf (f, " cross_module");
139 : }
140 148 : if (hints & INLINE_HINT_declared_inline)
141 : {
142 122 : hints &= ~INLINE_HINT_declared_inline;
143 122 : fprintf (f, " declared_inline");
144 : }
145 148 : if (hints & INLINE_HINT_known_hot)
146 : {
147 1 : hints &= ~INLINE_HINT_known_hot;
148 1 : fprintf (f, " known_hot");
149 : }
150 148 : if (hints & INLINE_HINT_builtin_constant_p)
151 : {
152 4 : hints &= ~INLINE_HINT_builtin_constant_p;
153 4 : fprintf (f, " builtin_constant_p");
154 : }
155 148 : gcc_assert (!hints);
156 : }
157 :
158 :
159 : /* Record SIZE and TIME to SUMMARY.
160 : The accounted code will be executed when EXEC_PRED is true.
161 : When NONCONST_PRED is false the code will evaluate to constant and
162 : will get optimized out in specialized clones of the function.
163 : If CALL is true account to call_size_time_table rather than
164 : size_time_table. */
165 :
166 : void
167 138915862 : ipa_fn_summary::account_size_time (int size, sreal time,
168 : const ipa_predicate &exec_pred,
169 : const ipa_predicate &nonconst_pred_in,
170 : bool call)
171 : {
172 138915862 : size_time_entry *e;
173 138915862 : bool found = false;
174 138915862 : int i;
175 138915862 : ipa_predicate nonconst_pred;
176 138915862 : vec<size_time_entry> *table = call ? &call_size_time_table : &size_time_table;
177 :
178 138915862 : if (exec_pred == false)
179 5100543 : return;
180 :
181 138620664 : nonconst_pred = nonconst_pred_in & exec_pred;
182 :
183 138620664 : if (nonconst_pred == false)
184 : return;
185 :
186 : /* We need to create initial empty unconditional clause, but otherwise
187 : we don't need to account empty times and sizes. */
188 162734833 : if (!size && time == 0 && table->length ())
189 4043844 : return;
190 :
191 : /* Only for calls we are unaccounting what we previously recorded. */
192 133815319 : gcc_checking_assert (time >= 0 || call);
193 :
194 486769530 : for (i = 0; table->iterate (i, &e); i++)
195 456411855 : if (e->exec_predicate == exec_pred
196 456411855 : && e->nonconst_predicate == nonconst_pred)
197 : {
198 : found = true;
199 : break;
200 : }
201 133815319 : if (i == max_size_time_table_size)
202 : {
203 4896 : i = 0;
204 4896 : found = true;
205 4896 : e = &(*table)[0];
206 4896 : if (dump_file && (dump_flags & TDF_DETAILS))
207 0 : fprintf (dump_file,
208 : "\t\tReached limit on number of entries, "
209 : "ignoring the predicate.");
210 : }
211 133819556 : if (dump_file && (dump_flags & TDF_DETAILS) && (time != 0 || size))
212 : {
213 3963 : fprintf (dump_file,
214 : "\t\tAccounting size:%3.2f, time:%3.2f on %spredicate exec:",
215 3730 : ((double) size) / ipa_fn_summary::size_scale,
216 : (time.to_double ()), found ? "" : "new ");
217 3730 : exec_pred.dump (dump_file, conds, 0);
218 3730 : if (exec_pred != nonconst_pred)
219 : {
220 84 : fprintf (dump_file, " nonconst:");
221 84 : nonconst_pred.dump (dump_file, conds);
222 : }
223 : else
224 3646 : fprintf (dump_file, "\n");
225 : }
226 133815319 : if (!found)
227 : {
228 30352779 : class size_time_entry new_entry;
229 30352779 : new_entry.size = size;
230 30352779 : new_entry.time = time;
231 30352779 : new_entry.exec_predicate = exec_pred;
232 30352779 : new_entry.nonconst_predicate = nonconst_pred;
233 30352779 : if (call)
234 1377346 : call_size_time_table.safe_push (new_entry);
235 : else
236 28975433 : size_time_table.safe_push (new_entry);
237 : }
238 : else
239 : {
240 103462540 : e->size += size;
241 103462540 : e->time += time;
242 : /* FIXME: PR bootstrap/92653 gcc_checking_assert (e->time >= -1); */
243 : /* Tolerate small roundoff issues. */
244 103462540 : if (e->time < 0)
245 216 : e->time = 0;
246 : }
247 : }
248 :
249 : /* We proved E to be unreachable, redirect it to __builtin_unreachable. */
250 :
251 : static struct cgraph_edge *
252 350766 : redirect_to_unreachable (struct cgraph_edge *e)
253 : {
254 350766 : struct cgraph_node *callee = !e->inline_failed ? e->callee : NULL;
255 350766 : struct cgraph_node *target
256 350766 : = cgraph_node::get_create (builtin_decl_unreachable ());
257 :
258 350766 : gcc_checking_assert (lookup_attribute ("cold",
259 : DECL_ATTRIBUTES (target->decl)));
260 :
261 350766 : if (e->speculative)
262 311 : e = cgraph_edge::resolve_speculation (e, target->decl);
263 350455 : else if (!e->callee)
264 785 : e = cgraph_edge::make_direct (e, target);
265 : else
266 349670 : e->redirect_callee (target);
267 350766 : class ipa_call_summary *es = ipa_call_summaries->get (e);
268 350766 : e->inline_failed = CIF_UNREACHABLE;
269 350766 : e->count = profile_count::zero ();
270 350766 : es->call_stmt_size = 0;
271 350766 : es->call_stmt_time = 0;
272 350766 : if (callee)
273 0 : callee->remove_symbol_and_inline_clones ();
274 350766 : if (e->has_callback)
275 6 : for (cgraph_edge *cbe = e->first_callback_edge (); cbe;
276 2 : cbe = cbe->next_callback_edge ())
277 : /* If the carrying edge is unreachable, so are the callback calls. */
278 2 : redirect_to_unreachable (cbe);
279 350766 : return e;
280 : }
281 :
282 : /* Set predicate for edge E. */
283 :
284 : static void
285 32070441 : edge_set_predicate (struct cgraph_edge *e, ipa_predicate *predicate)
286 : {
287 : /* If the edge is determined to be never executed, redirect it
288 : to BUILTIN_UNREACHABLE to make it clear to IPA passes the call will
289 : be optimized out. */
290 29093386 : if (predicate && *predicate == false
291 : /* When handling speculative edges, we need to do the redirection
292 : just once. Do it always on the direct edge, so we do not
293 : attempt to resolve speculation while duplicating the edge. */
294 32421282 : && (!e->speculative || e->callee))
295 350764 : e = redirect_to_unreachable (e);
296 :
297 32070441 : class ipa_call_summary *es = ipa_call_summaries->get (e);
298 61163827 : if (predicate && *predicate != true)
299 : {
300 4774520 : if (!es->predicate)
301 4248885 : es->predicate = edge_predicate_pool.allocate ();
302 4774520 : *es->predicate = *predicate;
303 : }
304 : else
305 : {
306 27295921 : if (es->predicate)
307 865717 : edge_predicate_pool.remove (es->predicate);
308 27295921 : es->predicate = NULL;
309 : }
310 32070441 : }
311 :
312 : /* Set predicate for hint *P. */
313 :
314 : static void
315 152549 : set_hint_predicate (ipa_predicate **p, ipa_predicate new_predicate)
316 : {
317 152549 : if (new_predicate == false || new_predicate == true)
318 : {
319 2947 : if (*p)
320 0 : edge_predicate_pool.remove (*p);
321 2947 : *p = NULL;
322 : }
323 : else
324 : {
325 149602 : if (!*p)
326 149602 : *p = edge_predicate_pool.allocate ();
327 149602 : **p = new_predicate;
328 : }
329 152549 : }
330 :
331 : /* Find if NEW_PREDICATE is already in V and if so, increment its freq.
332 : Otherwise add a new item to the vector with this predicate and frerq equal
333 : to add_freq, unless the number of predicates would exceed MAX_NUM_PREDICATES
334 : in which case the function does nothing. */
335 :
336 : static void
337 1131654 : add_freqcounting_predicate (vec<ipa_freqcounting_predicate, va_gc> **v,
338 : const ipa_predicate &new_predicate, sreal add_freq,
339 : unsigned max_num_predicates)
340 : {
341 1131654 : if (new_predicate == false || new_predicate == true)
342 : return;
343 : ipa_freqcounting_predicate *f;
344 143363 : for (int i = 0; vec_safe_iterate (*v, i, &f); i++)
345 72050 : if (new_predicate == f->predicate)
346 : {
347 0 : f->freq += add_freq;
348 0 : return;
349 : }
350 97484 : if (vec_safe_length (*v) >= max_num_predicates)
351 : /* Too many different predicates to account for. */
352 : return;
353 :
354 71041 : ipa_freqcounting_predicate fcp;
355 71041 : fcp.predicate = NULL;
356 71041 : set_hint_predicate (&fcp.predicate, new_predicate);
357 71041 : fcp.freq = add_freq;
358 71041 : vec_safe_push (*v, fcp);
359 71041 : return;
360 : }
361 :
362 : /* Compute what conditions may or may not hold given information about
363 : parameters. RET_CLAUSE returns truths that may hold in a specialized copy,
364 : while RET_NONSPEC_CLAUSE returns truths that may hold in an nonspecialized
365 : copy when called in a given context. It is a bitmask of conditions. Bit
366 : 0 means that condition is known to be false, while bit 1 means that condition
367 : may or may not be true. These differs - for example NOT_INLINED condition
368 : is always false in the second and also builtin_constant_p tests cannot use
369 : the fact that parameter is indeed a constant.
370 :
371 : When INLINE_P is true, assume that we are inlining. AVAL contains known
372 : information about argument values. The function does not modify its content
373 : and so AVALs could also be of type ipa_call_arg_values but so far all
374 : callers work with the auto version and so we avoid the conversion for
375 : convenience.
376 :
377 : ERROR_MARK value of an argument means compile time invariant. */
378 :
379 : static void
380 23049327 : evaluate_conditions_for_known_args (struct cgraph_node *node,
381 : bool inline_p,
382 : ipa_auto_call_arg_values *avals,
383 : clause_t *ret_clause,
384 : clause_t *ret_nonspec_clause,
385 : ipa_call_summary *es)
386 : {
387 23049327 : clause_t clause = inline_p ? 0 : 1 << ipa_predicate::not_inlined_condition;
388 23049327 : clause_t nonspec_clause = 1 << ipa_predicate::not_inlined_condition;
389 23049327 : class ipa_fn_summary *info = ipa_fn_summaries->get (node);
390 23049327 : int i;
391 23049327 : struct condition *c;
392 :
393 95617139 : for (i = 0; vec_safe_iterate (info->conds, i, &c); i++)
394 : {
395 72567812 : tree val = NULL;
396 72567812 : tree res;
397 72567812 : int j;
398 72567812 : struct expr_eval_op *op;
399 :
400 72567812 : if (c->code == ipa_predicate::not_sra_candidate)
401 : {
402 16899396 : if (!inline_p
403 16899396 : || !es
404 16742191 : || (int)es->param.length () <= c->operand_num
405 25270470 : || !es->param[c->operand_num].points_to_possible_sra_candidate)
406 12582921 : clause |= 1 << (i + ipa_predicate::first_dynamic_condition);
407 16899396 : nonspec_clause |= 1 << (i + ipa_predicate::first_dynamic_condition);
408 72567812 : continue;
409 : }
410 :
411 55668416 : if (c->agg_contents)
412 : {
413 30903662 : if (c->code == ipa_predicate::changed
414 19757356 : && !c->by_ref
415 33763487 : && (avals->safe_sval_at(c->operand_num) == error_mark_node))
416 3466 : continue;
417 :
418 30896730 : if (tree sval = avals->safe_sval_at (c->operand_num))
419 14445034 : val = ipa_find_agg_cst_from_init (sval, c->offset, c->by_ref);
420 : /* ipa_argagg_value_list is indexed by byte offsets, so a condition
421 : which does not start at a byte boundary (a bit-field) cannot be
422 : looked up in it; the containing byte would be reinterpreted as
423 : the whole field below. */
424 30896730 : if (!val && (c->offset % BITS_PER_UNIT) == 0)
425 : {
426 30859895 : ipa_argagg_value_list avs (avals);
427 30859895 : val = avs.get_value (c->operand_num, c->offset / BITS_PER_UNIT,
428 30859895 : c->by_ref);
429 : }
430 : }
431 : else
432 : {
433 24768220 : val = avals->safe_sval_at (c->operand_num);
434 24768220 : if (val && val == error_mark_node
435 2308949 : && c->code != ipa_predicate::changed)
436 : val = NULL_TREE;
437 : }
438 :
439 42239055 : if (!val
440 42118321 : && (c->code == ipa_predicate::changed
441 : || c->code == ipa_predicate::is_not_constant))
442 : {
443 26648153 : clause |= 1 << (i + ipa_predicate::first_dynamic_condition);
444 26648153 : nonspec_clause |= 1 << (i + ipa_predicate::first_dynamic_condition);
445 26648153 : continue;
446 : }
447 29016797 : if (c->code == ipa_predicate::changed)
448 : {
449 7800343 : nonspec_clause |= 1 << (i + ipa_predicate::first_dynamic_condition);
450 7800343 : continue;
451 : }
452 :
453 21216454 : if (c->code == ipa_predicate::is_not_constant)
454 : {
455 8693 : nonspec_clause |= 1 << (i + ipa_predicate::first_dynamic_condition);
456 8693 : continue;
457 : }
458 :
459 21207761 : if (val
460 21207761 : && (c->type == TREE_TYPE (val)
461 1023143 : || (TYPE_SIZE (c->type) == TYPE_SIZE (TREE_TYPE (val))
462 : /* Avoid precision mismatch like with bit-fields where the
463 : VIEW_CONVERT_EXPR does not truncate excess bits
464 : appropriately. */
465 942249 : && ((!INTEGRAL_TYPE_P (c->type)
466 238476 : || type_has_mode_precision_p (c->type))
467 940526 : && (!INTEGRAL_TYPE_P (TREE_TYPE (val))
468 236933 : || type_has_mode_precision_p (TREE_TYPE (val)))))))
469 : {
470 5654962 : if (c->type != TREE_TYPE (val))
471 940512 : val = fold_unary (VIEW_CONVERT_EXPR, c->type, val);
472 6181225 : for (j = 0; vec_safe_iterate (c->param_ops, j, &op); j++)
473 : {
474 526728 : if (!val)
475 : break;
476 526263 : if (!op->val[0])
477 225806 : val = fold_unary (op->code, op->type, val);
478 300457 : else if (!op->val[1])
479 600914 : val = fold_binary (op->code, op->type,
480 : op->index ? op->val[0] : val,
481 : op->index ? val : op->val[0]);
482 0 : else if (op->index == 0)
483 0 : val = fold_ternary (op->code, op->type,
484 : val, op->val[0], op->val[1]);
485 0 : else if (op->index == 1)
486 0 : val = fold_ternary (op->code, op->type,
487 : op->val[0], val, op->val[1]);
488 0 : else if (op->index == 2)
489 0 : val = fold_ternary (op->code, op->type,
490 : op->val[0], op->val[1], val);
491 : else
492 : val = NULL_TREE;
493 : }
494 :
495 5654962 : res = val
496 5654962 : ? fold_binary_to_constant (c->code, boolean_type_node, val, c->val)
497 : : NULL;
498 :
499 5654477 : if (res && integer_zerop (res))
500 2881667 : continue;
501 2773295 : if (res && integer_onep (res))
502 : {
503 2750623 : clause |= 1 << (i + ipa_predicate::first_dynamic_condition);
504 2750623 : nonspec_clause
505 2750623 : |= 1 << (i + ipa_predicate::first_dynamic_condition);
506 2750623 : continue;
507 : }
508 : }
509 15575471 : if (c->operand_num < (int) avals->m_known_value_ranges.length ()
510 9103823 : && !c->agg_contents
511 17779745 : && (!val || TREE_CODE (val) != INTEGER_CST))
512 : {
513 2203539 : value_range vr (avals->m_known_value_ranges[c->operand_num]);
514 2203539 : if (!vr.undefined_p ()
515 1547087 : && !vr.varying_p ()
516 3750626 : && (TYPE_SIZE (c->type) == TYPE_SIZE (vr.type ())))
517 : {
518 1546562 : if (!useless_type_conversion_p (c->type, vr.type ()))
519 512 : range_cast (vr, c->type);
520 :
521 1981739 : for (j = 0; vec_safe_iterate (c->param_ops, j, &op); j++)
522 : {
523 539600 : if (vr.varying_p () || vr.undefined_p ())
524 : break;
525 :
526 435177 : value_range res (op->type);
527 435177 : if (!op->val[0])
528 : {
529 155412 : value_range varying (op->type);
530 155412 : varying.set_varying (op->type);
531 155412 : range_op_handler handler (op->code);
532 155412 : if (!handler
533 155412 : || !res.supports_type_p (op->type)
534 310824 : || !handler.fold_range (res, op->type, vr, varying))
535 0 : res.set_varying (op->type);
536 155412 : }
537 279765 : else if (!op->val[1])
538 : {
539 279593 : value_range op0 (TREE_TYPE (op->val[0]));
540 279593 : range_op_handler handler (op->code);
541 :
542 279593 : ipa_get_range_from_ip_invariant (op0, op->val[0], node);
543 :
544 279593 : if (!handler
545 279593 : || !res.supports_type_p (op->type)
546 559186 : || !handler.fold_range (res, op->type,
547 279593 : op->index ? op0 : vr,
548 557418 : op->index ? vr : op0))
549 0 : res.set_varying (op->type);
550 279593 : }
551 : else
552 172 : res.set_varying (op->type);
553 435177 : vr = res;
554 435177 : }
555 1546562 : if (!vr.varying_p () && !vr.undefined_p ())
556 : {
557 1433122 : int_range<2> res;
558 1433122 : value_range val_vr (TREE_TYPE (c->val));
559 1433122 : range_op_handler handler (c->code);
560 :
561 1433122 : ipa_get_range_from_ip_invariant (val_vr, c->val, node);
562 :
563 1433122 : if (!handler
564 1433122 : || !val_vr.supports_type_p (TREE_TYPE (c->val))
565 2866244 : || !handler.fold_range (res, boolean_type_node, vr, val_vr))
566 0 : res.set_varying (boolean_type_node);
567 :
568 1433122 : if (res.zero_p ())
569 243838 : continue;
570 1433122 : }
571 : }
572 2203539 : }
573 :
574 15331633 : clause |= 1 << (i + ipa_predicate::first_dynamic_condition);
575 15331633 : nonspec_clause |= 1 << (i + ipa_predicate::first_dynamic_condition);
576 : }
577 23049327 : *ret_clause = clause;
578 23049327 : if (ret_nonspec_clause)
579 19954048 : *ret_nonspec_clause = nonspec_clause;
580 23049327 : }
581 :
582 : /* Return true if VRP will be executed on the function.
583 : We do not want to anticipate optimizations that will not happen.
584 :
585 : FIXME: This can be confused with -fdisable and debug counters and thus
586 : it should not be used for correctness (only to make heuristics work).
587 : This means that inliner should do its own optimizations of expressions
588 : that it predicts to be constant so wrong code can not be triggered by
589 : builtin_constant_p. */
590 :
591 : static bool
592 12653978 : vrp_will_run_p (struct cgraph_node *node)
593 : {
594 12653978 : return (opt_for_fn (node->decl, optimize)
595 12653978 : && !opt_for_fn (node->decl, optimize_debug)
596 25307188 : && opt_for_fn (node->decl, flag_tree_vrp));
597 : }
598 :
599 : /* Similarly about FRE. */
600 :
601 : static bool
602 14922564 : fre_will_run_p (struct cgraph_node *node)
603 : {
604 14922564 : return (opt_for_fn (node->decl, optimize)
605 14922564 : && !opt_for_fn (node->decl, optimize_debug)
606 29843486 : && opt_for_fn (node->decl, flag_tree_fre));
607 : }
608 :
609 : /* Work out what conditions might be true at invocation of E.
610 : Compute costs for inlined edge if INLINE_P is true.
611 :
612 : Return in CLAUSE_PTR the evaluated conditions and in NONSPEC_CLAUSE_PTR
613 : (if non-NULL) conditions evaluated for nonspecialized clone called
614 : in a given context.
615 :
616 : Vectors in AVALS will be populated with useful known information about
617 : argument values - information not known to have any uses will be omitted -
618 : except for m_known_contexts which will only be calculated if
619 : COMPUTE_CONTEXTS is true. */
620 :
621 : void
622 22803546 : evaluate_properties_for_edge (struct cgraph_edge *e, bool inline_p,
623 : clause_t *clause_ptr,
624 : clause_t *nonspec_clause_ptr,
625 : ipa_auto_call_arg_values *avals,
626 : bool compute_contexts)
627 : {
628 22803546 : struct cgraph_node *callee = e->callee->ultimate_alias_target ();
629 22803546 : class ipa_fn_summary *info = ipa_fn_summaries->get (callee);
630 22803546 : class ipa_edge_args *args;
631 22803546 : class ipa_call_summary *es = NULL;
632 :
633 22803546 : if (clause_ptr)
634 22803546 : *clause_ptr = inline_p ? 0 : 1 << ipa_predicate::not_inlined_condition;
635 :
636 22803546 : if (ipa_node_params_sum
637 10389742 : && !e->call_stmt_cannot_inline_p
638 10389742 : && (info->conds || compute_contexts)
639 33193288 : && (args = ipa_edge_args_sum->get (e)) != NULL)
640 : {
641 10335141 : struct cgraph_node *caller;
642 10335141 : class ipa_node_params *caller_parms_info, *callee_pi = NULL;
643 10335141 : int i, count = ipa_get_cs_argument_count (args);
644 10335141 : es = ipa_call_summaries->get (e);
645 :
646 10335141 : if (count)
647 : {
648 9621275 : if (e->caller->inlined_to)
649 : caller = e->caller->inlined_to;
650 : else
651 7670457 : caller = e->caller;
652 9621275 : caller_parms_info = ipa_node_params_sum->get (caller);
653 9621275 : callee_pi = ipa_node_params_sum->get (callee);
654 :
655 : /* Watch for thunks. */
656 9621275 : if (callee_pi)
657 : /* Watch for variadic functions. */
658 9620901 : count = MIN (count, ipa_get_param_count (callee_pi));
659 : }
660 :
661 9620901 : if (callee_pi)
662 31475589 : for (i = 0; i < count; i++)
663 : {
664 21854688 : struct ipa_jump_func *jf = ipa_get_ith_jump_func (args, i);
665 :
666 21854688 : if (ipa_is_param_used_by_indirect_call (callee_pi, i)
667 21854688 : || ipa_is_param_used_by_ipa_predicates (callee_pi, i))
668 : {
669 : /* Determine if we know constant value of the parameter. */
670 14922564 : tree type = ipa_get_type (callee_pi, i);
671 14922564 : tree cst = ipa_value_from_jfunc (caller_parms_info, jf, type);
672 :
673 11955963 : if (!cst && e->call_stmt
674 26825802 : && i < (int)gimple_call_num_args (e->call_stmt))
675 : {
676 11903238 : cst = gimple_call_arg (e->call_stmt, i);
677 11903238 : if (!is_gimple_min_invariant (cst))
678 : cst = NULL;
679 : }
680 3217956 : if (cst)
681 : {
682 3165231 : gcc_checking_assert (TREE_CODE (cst) != TREE_BINFO);
683 3165231 : if (!avals->m_known_vals.length ())
684 1828608 : avals->m_known_vals.safe_grow_cleared (count, true);
685 3165231 : avals->m_known_vals[i] = cst;
686 : }
687 11757333 : else if (inline_p && !es->param[i].change_prob)
688 : {
689 4350348 : if (!avals->m_known_vals.length ())
690 3559514 : avals->m_known_vals.safe_grow_cleared (count, true);
691 4350348 : avals->m_known_vals[i] = error_mark_node;
692 : }
693 :
694 : /* If we failed to get simple constant, try value range. */
695 3165231 : if ((!cst || TREE_CODE (cst) != INTEGER_CST)
696 12653978 : && vrp_will_run_p (caller)
697 27410658 : && ipa_is_param_used_by_ipa_predicates (callee_pi, i))
698 : {
699 12447268 : value_range vr (type);
700 :
701 12447268 : ipa_value_range_from_jfunc (vr, caller_parms_info, e, jf, type);
702 12447268 : if (!vr.undefined_p () && !vr.varying_p ())
703 : {
704 8859301 : if (!avals->m_known_value_ranges.length ())
705 : {
706 6478789 : avals->m_known_value_ranges.safe_grow_cleared (count,
707 : true);
708 26987043 : for (int i = 0; i < count; ++i)
709 14029465 : avals->m_known_value_ranges[i].set_range_class
710 14029465 : (void_type_node);
711 : }
712 8859301 : avals->m_known_value_ranges[i] = vr;
713 : }
714 12447268 : }
715 :
716 : /* Determine known aggregate values. */
717 14922564 : if (fre_will_run_p (caller))
718 14920518 : ipa_push_agg_values_from_jfunc (caller_parms_info,
719 : caller, &jf->agg, i,
720 : &avals->m_known_aggs);
721 : }
722 :
723 : /* For calls used in polymorphic calls we further determine
724 : polymorphic call context. */
725 21854688 : if (compute_contexts
726 21854688 : && ipa_is_param_used_by_polymorphic_call (callee_pi, i))
727 : {
728 176239 : ipa_polymorphic_call_context
729 176239 : ctx = ipa_context_from_jfunc (caller_parms_info, e, i, jf);
730 352478 : if (!ctx.useless_p ())
731 : {
732 174271 : if (!avals->m_known_contexts.length ())
733 174134 : avals->m_known_contexts.safe_grow_cleared (count, true);
734 174271 : avals->m_known_contexts[i]
735 348542 : = ipa_context_from_jfunc (caller_parms_info, e, i, jf);
736 : }
737 : }
738 : }
739 : else
740 714240 : gcc_assert (!count || callee->thunk);
741 : }
742 12468405 : else if (e->call_stmt && !e->call_stmt_cannot_inline_p && info->conds)
743 : {
744 9806831 : int i, count = (int)gimple_call_num_args (e->call_stmt);
745 :
746 27882922 : for (i = 0; i < count; i++)
747 : {
748 18076091 : tree cst = gimple_call_arg (e->call_stmt, i);
749 18076091 : if (!is_gimple_min_invariant (cst))
750 : cst = NULL;
751 6427505 : if (cst)
752 : {
753 6427505 : if (!avals->m_known_vals.length ())
754 4459968 : avals->m_known_vals.safe_grow_cleared (count, true);
755 6427505 : avals->m_known_vals[i] = cst;
756 : }
757 : }
758 : }
759 :
760 22803546 : evaluate_conditions_for_known_args (callee, inline_p, avals, clause_ptr,
761 : nonspec_clause_ptr, es);
762 22803546 : }
763 :
764 :
765 : /* Allocate the function summary. */
766 :
767 : static void
768 468121 : ipa_fn_summary_alloc (void)
769 : {
770 468121 : gcc_checking_assert (!ipa_fn_summaries);
771 468121 : ipa_size_summaries = new ipa_size_summary_t (symtab);
772 468121 : ipa_fn_summaries = ipa_fn_summary_t::create_ggc (symtab);
773 468121 : ipa_call_summaries = new ipa_call_summary_t (symtab);
774 468121 : }
775 :
776 28053247 : ipa_call_summary::~ipa_call_summary ()
777 : {
778 28053247 : if (predicate)
779 3383168 : edge_predicate_pool.remove (predicate);
780 :
781 28053247 : param.release ();
782 28053247 : }
783 :
784 10404745 : ipa_fn_summary::~ipa_fn_summary ()
785 : {
786 10404745 : unsigned len = vec_safe_length (loop_iterations);
787 10518652 : for (unsigned i = 0; i < len; i++)
788 113907 : edge_predicate_pool.remove ((*loop_iterations)[i].predicate);
789 10404745 : len = vec_safe_length (loop_strides);
790 10440440 : for (unsigned i = 0; i < len; i++)
791 35695 : edge_predicate_pool.remove ((*loop_strides)[i].predicate);
792 10404745 : vec_free (conds);
793 10404745 : call_size_time_table.release ();
794 10404745 : vec_free (loop_iterations);
795 10404745 : vec_free (loop_strides);
796 10404745 : builtin_constant_p_parms.release ();
797 10404745 : }
798 :
799 : void
800 7484735 : ipa_fn_summary_t::remove_callees (cgraph_node *node)
801 : {
802 7484735 : cgraph_edge *e;
803 30526095 : for (e = node->callees; e; e = e->next_callee)
804 23041360 : ipa_call_summaries->remove (e);
805 7967367 : for (e = node->indirect_calls; e; e = e->next_callee)
806 482632 : ipa_call_summaries->remove (e);
807 7484735 : }
808 :
809 : /* Duplicate predicates in loop hint vector, allocating memory for them and
810 : remove and deallocate any uninteresting (true or false) ones. Return the
811 : result. */
812 :
813 : static vec<ipa_freqcounting_predicate, va_gc> *
814 28516 : remap_freqcounting_preds_after_dup (vec<ipa_freqcounting_predicate, va_gc> *v,
815 : clause_t possible_truths)
816 : {
817 28516 : if (vec_safe_length (v) == 0)
818 : return NULL;
819 :
820 4691 : vec<ipa_freqcounting_predicate, va_gc> *res = v->copy ();
821 4691 : int len = res->length();
822 10871 : for (int i = len - 1; i >= 0; i--)
823 : {
824 6180 : ipa_predicate new_predicate
825 6180 : = (*res)[i].predicate->remap_after_duplication (possible_truths);
826 : /* We do not want to free previous predicate; it is used by node
827 : origin. */
828 6180 : (*res)[i].predicate = NULL;
829 6180 : set_hint_predicate (&(*res)[i].predicate, new_predicate);
830 :
831 6180 : if (!(*res)[i].predicate)
832 2947 : res->unordered_remove (i);
833 : }
834 :
835 : return res;
836 : }
837 :
838 :
839 : /* Hook that is called by cgraph.cc when a node is duplicated. */
840 : void
841 2831813 : ipa_fn_summary_t::duplicate (cgraph_node *src,
842 : cgraph_node *dst,
843 : ipa_fn_summary *src_info,
844 : ipa_fn_summary *info)
845 : {
846 2831813 : new (info) ipa_fn_summary (*src_info);
847 : /* TODO: as an optimization, we may avoid copying conditions
848 : that are known to be false or true. */
849 2831813 : info->conds = vec_safe_copy (info->conds);
850 :
851 2831813 : clone_info *cinfo = clone_info::get (dst);
852 : /* When there are any replacements in the function body, see if we can figure
853 : out that something was optimized out. */
854 2831813 : if (ipa_node_params_sum && cinfo && cinfo->tree_map)
855 : {
856 : /* Use SRC parm info since it may not be copied yet. */
857 14258 : ipa_node_params *parms_info = ipa_node_params_sum->get (src);
858 14258 : ipa_auto_call_arg_values avals;
859 14258 : int count = ipa_get_param_count (parms_info);
860 14258 : int i, j;
861 14258 : clause_t possible_truths;
862 14258 : ipa_predicate true_pred = true;
863 14258 : size_time_entry *e;
864 14258 : int optimized_out_size = 0;
865 14258 : bool inlined_to_p = false;
866 14258 : struct cgraph_edge *edge, *next;
867 :
868 14258 : info->size_time_table.release ();
869 14258 : avals.m_known_vals.safe_grow_cleared (count, true);
870 71587 : for (i = 0; i < count; i++)
871 : {
872 : struct ipa_replace_map *r;
873 :
874 140691 : for (j = 0; vec_safe_iterate (cinfo->tree_map, j, &r); j++)
875 : {
876 79172 : if (r->parm_num == i)
877 : {
878 24623 : avals.m_known_vals[i] = r->new_tree;
879 24623 : break;
880 : }
881 : }
882 : }
883 14258 : evaluate_conditions_for_known_args (dst, false,
884 : &avals,
885 : &possible_truths,
886 : /* We are going to specialize,
887 : so ignore nonspec truths. */
888 : NULL,
889 : NULL);
890 :
891 14258 : info->account_size_time (0, 0, true_pred, true_pred);
892 :
893 : /* Remap size_time vectors.
894 : Simplify the predicate by pruning out alternatives that are known
895 : to be false.
896 : TODO: as on optimization, we can also eliminate conditions known
897 : to be true. */
898 99323 : for (i = 0; src_info->size_time_table.iterate (i, &e); i++)
899 : {
900 85065 : ipa_predicate new_exec_pred;
901 85065 : ipa_predicate new_nonconst_pred;
902 85065 : new_exec_pred = e->exec_predicate.remap_after_duplication
903 85065 : (possible_truths);
904 85065 : new_nonconst_pred = e->nonconst_predicate.remap_after_duplication
905 85065 : (possible_truths);
906 85065 : if (new_exec_pred == false || new_nonconst_pred == false)
907 11811 : optimized_out_size += e->size;
908 : else
909 73254 : info->account_size_time (e->size, e->time, new_exec_pred,
910 : new_nonconst_pred);
911 : }
912 :
913 : /* Remap edge predicates with the same simplification as above.
914 : Also copy constantness arrays. */
915 183745 : for (edge = dst->callees; edge; edge = next)
916 : {
917 169487 : ipa_predicate new_predicate;
918 169487 : class ipa_call_summary *es = ipa_call_summaries->get (edge);
919 169487 : next = edge->next_callee;
920 :
921 169487 : if (!edge->inline_failed)
922 0 : inlined_to_p = true;
923 169487 : if (!es->predicate)
924 156993 : continue;
925 12494 : new_predicate = es->predicate->remap_after_duplication
926 12494 : (possible_truths);
927 12494 : if (new_predicate == false && *es->predicate != false)
928 3203 : optimized_out_size += es->call_stmt_size * ipa_fn_summary::size_scale;
929 12494 : edge_set_predicate (edge, &new_predicate);
930 : }
931 :
932 : /* Remap indirect edge predicates with the same simplification as above.
933 : Also copy constantness arrays. */
934 15154 : for (edge = dst->indirect_calls; edge; edge = next)
935 : {
936 896 : ipa_predicate new_predicate;
937 896 : class ipa_call_summary *es = ipa_call_summaries->get (edge);
938 896 : next = edge->next_callee;
939 :
940 896 : gcc_checking_assert (edge->inline_failed);
941 896 : if (!es->predicate)
942 792 : continue;
943 104 : new_predicate = es->predicate->remap_after_duplication
944 104 : (possible_truths);
945 104 : if (new_predicate == false && *es->predicate != false)
946 21 : optimized_out_size
947 21 : += es->call_stmt_size * ipa_fn_summary::size_scale;
948 104 : edge_set_predicate (edge, &new_predicate);
949 : }
950 14258 : info->loop_iterations
951 14258 : = remap_freqcounting_preds_after_dup (info->loop_iterations,
952 : possible_truths);
953 14258 : info->loop_strides
954 14258 : = remap_freqcounting_preds_after_dup (info->loop_strides,
955 : possible_truths);
956 14258 : if (info->builtin_constant_p_parms.length())
957 : {
958 22 : vec <int, va_heap, vl_ptr> parms = info->builtin_constant_p_parms;
959 22 : int ip;
960 22 : info->builtin_constant_p_parms = vNULL;
961 44 : for (i = 0; parms.iterate (i, &ip); i++)
962 22 : if (!avals.m_known_vals[ip])
963 8 : info->builtin_constant_p_parms.safe_push (ip);
964 : }
965 :
966 : /* If inliner or someone after inliner will ever start producing
967 : non-trivial clones, we will get trouble with lack of information
968 : about updating self sizes, because size vectors already contains
969 : sizes of the callees. */
970 14258 : gcc_assert (!inlined_to_p || !optimized_out_size);
971 14258 : }
972 : else
973 : {
974 2817555 : info->size_time_table = src_info->size_time_table.copy ();
975 2817555 : info->loop_iterations = vec_safe_copy (src_info->loop_iterations);
976 2817555 : info->loop_strides = vec_safe_copy (info->loop_strides);
977 :
978 2817555 : info->builtin_constant_p_parms
979 2817555 : = info->builtin_constant_p_parms.copy ();
980 :
981 2817555 : ipa_freqcounting_predicate *f;
982 5684506 : for (int i = 0; vec_safe_iterate (info->loop_iterations, i, &f); i++)
983 : {
984 49396 : ipa_predicate p = *f->predicate;
985 49396 : f->predicate = NULL;
986 49396 : set_hint_predicate (&f->predicate, p);
987 : }
988 2841807 : for (int i = 0; vec_safe_iterate (info->loop_strides, i, &f); i++)
989 : {
990 24252 : ipa_predicate p = *f->predicate;
991 24252 : f->predicate = NULL;
992 24252 : set_hint_predicate (&f->predicate, p);
993 : }
994 : }
995 2831813 : if (!dst->inlined_to)
996 148343 : ipa_update_overall_fn_summary (dst);
997 2831813 : }
998 :
999 :
1000 : /* Hook that is called by cgraph.cc when a node is duplicated. */
1001 :
1002 : void
1003 4180429 : ipa_call_summary_t::duplicate (struct cgraph_edge *src,
1004 : struct cgraph_edge *dst,
1005 : class ipa_call_summary *srcinfo,
1006 : class ipa_call_summary *info)
1007 : {
1008 4180429 : new (info) ipa_call_summary (*srcinfo);
1009 4180429 : info->predicate = NULL;
1010 4180429 : edge_set_predicate (dst, srcinfo->predicate);
1011 4180429 : info->param = srcinfo->param.copy ();
1012 4180429 : if (!dst->indirect_unknown_callee && src->indirect_unknown_callee
1013 : /* Don't subtract the size when dealing with callback pairs, since the
1014 : edge has no real size. */
1015 24001 : && !src->has_callback && !dst->callback)
1016 : {
1017 24001 : info->call_stmt_size -= (eni_size_weights.indirect_call_cost
1018 24001 : - eni_size_weights.call_cost);
1019 24001 : info->call_stmt_time -= (eni_time_weights.indirect_call_cost
1020 24001 : - eni_time_weights.call_cost);
1021 : }
1022 4180429 : }
1023 :
1024 : /* Dump edge summaries associated to NODE and recursively to all clones.
1025 : Indent by INDENT. */
1026 :
1027 : static void
1028 1969 : dump_ipa_call_summary (FILE *f, int indent, struct cgraph_node *node,
1029 : class ipa_fn_summary *info)
1030 : {
1031 1969 : struct cgraph_edge *edge;
1032 9787 : for (edge = node->callees; edge; edge = edge->next_callee)
1033 : {
1034 7818 : class ipa_call_summary *es = ipa_call_summaries->get (edge);
1035 7818 : struct cgraph_node *callee = edge->callee->ultimate_alias_target ();
1036 7818 : int i;
1037 :
1038 15636 : fprintf (f,
1039 : "%*s%s %s\n%*s freq:%4.2f",
1040 : indent, "", callee->dump_name (),
1041 7818 : !edge->inline_failed
1042 7265 : ? "inlined" : cgraph_inline_failed_string (edge-> inline_failed),
1043 7818 : indent, "", edge->sreal_frequency ().to_double ());
1044 :
1045 7818 : if (cross_module_call_p (edge))
1046 4 : fprintf (f, " cross module");
1047 :
1048 7818 : if (es)
1049 7265 : fprintf (f, " loop depth:%2i size:%2i time: %2i",
1050 : es->loop_depth, es->call_stmt_size, es->call_stmt_time);
1051 :
1052 7818 : ipa_fn_summary *s = ipa_fn_summaries->get (callee);
1053 7818 : ipa_size_summary *ss = ipa_size_summaries->get (callee);
1054 7818 : if (s != NULL)
1055 620 : fprintf (f, " callee size:%2i stack:%2i",
1056 620 : (int) (ss->size / ipa_fn_summary::size_scale),
1057 620 : (int) s->estimated_stack_size);
1058 :
1059 7818 : if (es && es->predicate)
1060 : {
1061 4627 : fprintf (f, " predicate: ");
1062 4627 : es->predicate->dump (f, info->conds);
1063 : }
1064 : else
1065 3191 : fprintf (f, "\n");
1066 7818 : if (es && es->param.exists ())
1067 5348 : for (i = 0; i < (int) es->param.length (); i++)
1068 : {
1069 1582 : int prob = es->param[i].change_prob;
1070 :
1071 1582 : if (!prob)
1072 620 : fprintf (f, "%*s op%i is compile time invariant\n",
1073 : indent + 2, "", i);
1074 962 : else if (prob != REG_BR_PROB_BASE)
1075 35 : fprintf (f, "%*s op%i change %f%% of time\n", indent + 2, "", i,
1076 35 : prob * 100.0 / REG_BR_PROB_BASE);
1077 1582 : if (es->param[i].points_to_local_or_readonly_memory)
1078 351 : fprintf (f, "%*s op%i points to local or readonly memory\n",
1079 : indent + 2, "", i);
1080 1582 : if (es->param[i].points_to_possible_sra_candidate)
1081 227 : fprintf (f, "%*s op%i points to possible sra candidate\n",
1082 : indent + 2, "", i);
1083 : }
1084 7818 : if (!edge->inline_failed)
1085 : {
1086 553 : ipa_size_summary *ss = ipa_size_summaries->get (callee);
1087 553 : fprintf (f, "%*sStack frame offset %i, callee self size %i\n",
1088 : indent + 2, "",
1089 553 : (int) ipa_get_stack_frame_offset (callee),
1090 553 : (int) ss->estimated_self_stack_size);
1091 553 : dump_ipa_call_summary (f, indent + 2, callee, info);
1092 : }
1093 : }
1094 2267 : for (edge = node->indirect_calls; edge; edge = edge->next_callee)
1095 : {
1096 298 : class ipa_call_summary *es = ipa_call_summaries->get (edge);
1097 298 : fprintf (f, "%*sindirect call loop depth:%2i freq:%4.2f size:%2i"
1098 : " time: %2i",
1099 : indent, "",
1100 : es->loop_depth,
1101 298 : edge->sreal_frequency ().to_double (), es->call_stmt_size,
1102 : es->call_stmt_time);
1103 298 : if (es->predicate)
1104 : {
1105 6 : fprintf (f, "predicate: ");
1106 6 : es->predicate->dump (f, info->conds);
1107 : }
1108 : else
1109 292 : fprintf (f, "\n");
1110 : }
1111 1969 : }
1112 :
1113 :
1114 : void
1115 1628 : ipa_dump_fn_summary (FILE *f, struct cgraph_node *node)
1116 : {
1117 1628 : if (node->definition)
1118 : {
1119 1628 : class ipa_fn_summary *s = ipa_fn_summaries->get (node);
1120 1628 : class ipa_size_summary *ss = ipa_size_summaries->get (node);
1121 1628 : if (s != NULL)
1122 : {
1123 1416 : size_time_entry *e;
1124 1416 : int i;
1125 1416 : fprintf (f, "IPA function summary for %s", node->dump_name ());
1126 1416 : if (DECL_DISREGARD_INLINE_LIMITS (node->decl))
1127 0 : fprintf (f, " always_inline");
1128 1416 : if (s->inlinable)
1129 1235 : fprintf (f, " inlinable");
1130 1416 : if (s->fp_expressions)
1131 44 : fprintf (f, " fp_expression");
1132 1416 : if (s->builtin_constant_p_parms.length ())
1133 : {
1134 2 : fprintf (f, " builtin_constant_p_parms");
1135 4 : for (unsigned int i = 0;
1136 4 : i < s->builtin_constant_p_parms.length (); i++)
1137 2 : fprintf (f, " %i", s->builtin_constant_p_parms[i]);
1138 : }
1139 1416 : fprintf (f, "\n global time: %f\n", s->time.to_double ());
1140 1416 : fprintf (f, " self size: %i\n", ss->self_size);
1141 1416 : fprintf (f, " global size: %i\n", ss->size);
1142 1416 : fprintf (f, " min size: %i\n", s->min_size);
1143 1416 : fprintf (f, " self stack: %i\n",
1144 1416 : (int) ss->estimated_self_stack_size);
1145 1416 : fprintf (f, " global stack: %i\n", (int) s->estimated_stack_size);
1146 1416 : if (s->growth)
1147 168 : fprintf (f, " estimated growth:%i\n", (int) s->growth);
1148 1416 : if (s->scc_no)
1149 5 : fprintf (f, " In SCC: %i\n", (int) s->scc_no);
1150 5908 : for (i = 0; s->size_time_table.iterate (i, &e); i++)
1151 : {
1152 4492 : fprintf (f, " size:%f, time:%f",
1153 4492 : (double) e->size / ipa_fn_summary::size_scale,
1154 : e->time.to_double ());
1155 4492 : if (e->exec_predicate != true)
1156 : {
1157 2623 : fprintf (f, ", executed if:");
1158 2623 : e->exec_predicate.dump (f, s->conds, 0);
1159 : }
1160 4492 : if (e->exec_predicate != e->nonconst_predicate)
1161 : {
1162 1167 : fprintf (f, ", nonconst if:");
1163 1167 : e->nonconst_predicate.dump (f, s->conds, 0);
1164 : }
1165 4492 : fprintf (f, "\n");
1166 : }
1167 : ipa_freqcounting_predicate *fcp;
1168 : bool first_fcp = true;
1169 1522 : for (int i = 0; vec_safe_iterate (s->loop_iterations, i, &fcp); i++)
1170 : {
1171 106 : if (first_fcp)
1172 : {
1173 78 : fprintf (f, " loop iterations:");
1174 78 : first_fcp = false;
1175 : }
1176 106 : fprintf (f, " %3.2f for ", fcp->freq.to_double ());
1177 106 : fcp->predicate->dump (f, s->conds);
1178 : }
1179 : first_fcp = true;
1180 1434 : for (int i = 0; vec_safe_iterate (s->loop_strides, i, &fcp); i++)
1181 : {
1182 18 : if (first_fcp)
1183 : {
1184 10 : fprintf (f, " loop strides:");
1185 10 : first_fcp = false;
1186 : }
1187 18 : fprintf (f, " %3.2f for :", fcp->freq.to_double ());
1188 18 : fcp->predicate->dump (f, s->conds);
1189 : }
1190 1416 : fprintf (f, " calls:\n");
1191 1416 : dump_ipa_call_summary (f, 4, node, s);
1192 1416 : fprintf (f, "\n");
1193 1416 : if (s->target_info)
1194 0 : fprintf (f, " target_info: %x\n", s->target_info);
1195 : }
1196 : else
1197 212 : fprintf (f, "IPA summary for %s is missing.\n", node->dump_name ());
1198 : }
1199 1628 : }
1200 :
1201 : DEBUG_FUNCTION void
1202 0 : ipa_debug_fn_summary (struct cgraph_node *node)
1203 : {
1204 0 : ipa_dump_fn_summary (stderr, node);
1205 0 : }
1206 :
1207 : void
1208 356 : ipa_dump_fn_summaries (FILE *f)
1209 : {
1210 356 : struct cgraph_node *node;
1211 :
1212 2288 : FOR_EACH_DEFINED_FUNCTION (node)
1213 1932 : if (!node->inlined_to)
1214 1379 : ipa_dump_fn_summary (f, node);
1215 356 : }
1216 :
1217 : /* Callback of walk_aliased_vdefs. Flags that it has been invoked to the
1218 : boolean variable pointed to by DATA. */
1219 :
1220 : static bool
1221 844858 : mark_modified (ao_ref *ao ATTRIBUTE_UNUSED, tree vdef ATTRIBUTE_UNUSED,
1222 : void *data)
1223 : {
1224 844858 : bool *b = (bool *) data;
1225 844858 : *b = true;
1226 844858 : return true;
1227 : }
1228 :
1229 : /* If OP refers to value of function parameter, return the corresponding
1230 : parameter. If non-NULL, the size of the memory load (or the SSA_NAME of the
1231 : PARM_DECL) will be stored to *SIZE_P in that case too. */
1232 :
1233 : static tree
1234 189618436 : unmodified_parm_1 (ipa_func_body_info *fbi, gimple *stmt, tree op,
1235 : poly_int64 *size_p)
1236 : {
1237 : /* SSA_NAME referring to parm default def? */
1238 189618436 : if (TREE_CODE (op) == SSA_NAME
1239 111309302 : && SSA_NAME_IS_DEFAULT_DEF (op)
1240 215955675 : && TREE_CODE (SSA_NAME_VAR (op)) == PARM_DECL)
1241 : {
1242 26123753 : if (size_p)
1243 0 : *size_p = tree_to_poly_int64 (TYPE_SIZE (TREE_TYPE (op)));
1244 26123753 : return SSA_NAME_VAR (op);
1245 : }
1246 : /* Non-SSA parm reference? */
1247 163494683 : if (TREE_CODE (op) == PARM_DECL
1248 1427599 : && fbi->aa_walk_budget > 0)
1249 : {
1250 1425053 : bool modified = false;
1251 :
1252 1425053 : ao_ref refd;
1253 1425053 : ao_ref_init (&refd, op);
1254 2850106 : int walked = walk_aliased_vdefs (&refd, gimple_vuse (stmt),
1255 : mark_modified, &modified, NULL, NULL,
1256 : fbi->aa_walk_budget);
1257 1425053 : if (walked < 0)
1258 : {
1259 8 : fbi->aa_walk_budget = 0;
1260 913969 : return NULL_TREE;
1261 : }
1262 1425045 : fbi->aa_walk_budget -= walked;
1263 1425045 : if (!modified)
1264 : {
1265 913961 : if (size_p)
1266 0 : *size_p = tree_to_poly_int64 (TYPE_SIZE (TREE_TYPE (op)));
1267 : return op;
1268 : }
1269 : }
1270 : return NULL_TREE;
1271 : }
1272 :
1273 : /* If OP refers to value of function parameter, return the corresponding
1274 : parameter. Also traverse chains of SSA register assignments. If non-NULL,
1275 : the size of the memory load (or the SSA_NAME of the PARM_DECL) will be
1276 : stored to *SIZE_P in that case too. */
1277 :
1278 : static tree
1279 124018633 : unmodified_parm (ipa_func_body_info *fbi, gimple *stmt, tree op,
1280 : poly_int64 *size_p)
1281 : {
1282 154167187 : tree res = unmodified_parm_1 (fbi, stmt, op, size_p);
1283 154167187 : if (res)
1284 : return res;
1285 :
1286 127987354 : if (TREE_CODE (op) == SSA_NAME
1287 72798890 : && !SSA_NAME_IS_DEFAULT_DEF (op)
1288 200577502 : && gimple_assign_single_p (SSA_NAME_DEF_STMT (op)))
1289 30148554 : return unmodified_parm (fbi, SSA_NAME_DEF_STMT (op),
1290 30148554 : gimple_assign_rhs1 (SSA_NAME_DEF_STMT (op)),
1291 30148554 : size_p);
1292 : return NULL_TREE;
1293 : }
1294 :
1295 : /* If OP refers to a value of a function parameter or value loaded from an
1296 : aggregate passed to a parameter (either by value or reference), return TRUE
1297 : and store the number of the parameter to *INDEX_P, the access size into
1298 : *SIZE_P, and information whether and how it has been loaded from an
1299 : aggregate into *AGGPOS. INFO describes the function parameters, STMT is the
1300 : statement in which OP is used or loaded. */
1301 :
1302 : static bool
1303 33993513 : unmodified_parm_or_parm_agg_item (struct ipa_func_body_info *fbi,
1304 : gimple *stmt, tree op, int *index_p,
1305 : poly_int64 *size_p,
1306 : struct agg_position_info *aggpos)
1307 : {
1308 35451249 : tree res = unmodified_parm_1 (fbi, stmt, op, size_p);
1309 :
1310 35451249 : gcc_checking_assert (aggpos);
1311 35451249 : if (res)
1312 : {
1313 857881 : *index_p = ipa_get_param_decl_index (fbi->info, res);
1314 857881 : if (*index_p < 0)
1315 : return false;
1316 857751 : aggpos->agg_contents = false;
1317 857751 : aggpos->by_ref = false;
1318 857751 : return true;
1319 : }
1320 :
1321 34593368 : if (TREE_CODE (op) == SSA_NAME)
1322 : {
1323 12386659 : if (SSA_NAME_IS_DEFAULT_DEF (op)
1324 12386659 : || !gimple_assign_single_p (SSA_NAME_DEF_STMT (op)))
1325 : return false;
1326 4947854 : stmt = SSA_NAME_DEF_STMT (op);
1327 4947854 : op = gimple_assign_rhs1 (stmt);
1328 4947854 : if (!REFERENCE_CLASS_P (op))
1329 : return unmodified_parm_or_parm_agg_item (fbi, stmt, op, index_p, size_p,
1330 : aggpos);
1331 : }
1332 :
1333 25696827 : aggpos->agg_contents = true;
1334 25696827 : return ipa_load_from_parm_agg (fbi, fbi->info->descriptors,
1335 : stmt, op, index_p, &aggpos->offset,
1336 25696827 : size_p, &aggpos->by_ref);
1337 : }
1338 :
1339 : /* If stmt is simple load or store of value pointed to by a function parameter,
1340 : return its index. */
1341 :
1342 : static int
1343 114053989 : load_or_store_of_ptr_parameter (ipa_func_body_info *fbi, gimple *stmt)
1344 : {
1345 114053989 : if (!optimize)
1346 : return -1;
1347 95686820 : gassign *assign = dyn_cast <gassign *> (stmt);
1348 56859745 : if (!assign)
1349 : return -1;
1350 56859745 : tree param;
1351 56859745 : if (gimple_assign_load_p (stmt))
1352 20837925 : param = gimple_assign_rhs1 (stmt);
1353 36021820 : else if (gimple_store_p (stmt))
1354 19106871 : param = gimple_assign_lhs (stmt);
1355 : else
1356 : return -1;
1357 39944796 : tree base = get_base_address (param);
1358 39944796 : if (TREE_CODE (base) != MEM_REF
1359 13922870 : || TREE_CODE (TREE_OPERAND (base, 0)) != SSA_NAME
1360 53863544 : || !SSA_NAME_IS_DEFAULT_DEF (TREE_OPERAND (base, 0)))
1361 : return -1;
1362 7540637 : tree p = SSA_NAME_VAR (TREE_OPERAND (base, 0));
1363 7540637 : if (TREE_CODE (p) != PARM_DECL)
1364 : return -1;
1365 7448986 : return ipa_get_param_decl_index (fbi->info, p);
1366 : }
1367 :
1368 : /* See if statement might disappear after inlining.
1369 : 0 - means not eliminated
1370 : 1 - half of statements goes away
1371 : 2 - for sure it is eliminated.
1372 : We are not terribly sophisticated, basically looking for simple abstraction
1373 : penalty wrappers. */
1374 :
1375 : static int
1376 114053989 : eliminated_by_inlining_prob (ipa_func_body_info *fbi, gimple *stmt)
1377 : {
1378 114053989 : enum gimple_code code = gimple_code (stmt);
1379 114053989 : enum tree_code rhs_code;
1380 :
1381 114053989 : if (!optimize)
1382 : return 0;
1383 :
1384 95686820 : switch (code)
1385 : {
1386 : case GIMPLE_RETURN:
1387 : return 2;
1388 56859745 : case GIMPLE_ASSIGN:
1389 56859745 : if (gimple_num_ops (stmt) != 2)
1390 : return 0;
1391 :
1392 40769167 : rhs_code = gimple_assign_rhs_code (stmt);
1393 :
1394 : /* Casts of parameters, loads from parameters passed by reference
1395 : and stores to return value or parameters are often free after
1396 : inlining due to SRA and further combining.
1397 : Assume that half of statements goes away. */
1398 40769167 : if (CONVERT_EXPR_CODE_P (rhs_code)
1399 : || rhs_code == VIEW_CONVERT_EXPR
1400 : || rhs_code == ADDR_EXPR
1401 37898611 : || gimple_assign_rhs_class (stmt) == GIMPLE_SINGLE_RHS)
1402 : {
1403 39944796 : tree rhs = gimple_assign_rhs1 (stmt);
1404 39944796 : tree lhs = gimple_assign_lhs (stmt);
1405 39944796 : tree inner_rhs = get_base_address (rhs);
1406 39944796 : tree inner_lhs = get_base_address (lhs);
1407 39944796 : bool rhs_free = false;
1408 39944796 : bool lhs_free = false;
1409 :
1410 39944796 : if (!inner_rhs)
1411 0 : inner_rhs = rhs;
1412 39944796 : if (!inner_lhs)
1413 0 : inner_lhs = lhs;
1414 :
1415 : /* Reads of parameter are expected to be free. */
1416 39944796 : if (unmodified_parm (fbi, stmt, inner_rhs, NULL))
1417 : rhs_free = true;
1418 : /* Match expressions of form &this->field. Those will most likely
1419 : combine with something upstream after inlining. */
1420 38625900 : else if (TREE_CODE (inner_rhs) == ADDR_EXPR)
1421 : {
1422 2582136 : tree op = get_base_address (TREE_OPERAND (inner_rhs, 0));
1423 2582136 : if (TREE_CODE (op) == PARM_DECL)
1424 : rhs_free = true;
1425 2537863 : else if (TREE_CODE (op) == MEM_REF
1426 2537863 : && unmodified_parm (fbi, stmt, TREE_OPERAND (op, 0),
1427 : NULL))
1428 : rhs_free = true;
1429 : }
1430 :
1431 : /* When parameter is not SSA register because its address is taken
1432 : and it is just copied into one, the statement will be completely
1433 : free after inlining (we will copy propagate backward). */
1434 1363169 : if (rhs_free && is_gimple_reg (lhs))
1435 : return 2;
1436 :
1437 : /* Reads of parameters passed by reference
1438 : expected to be free (i.e. optimized out after inlining). */
1439 39259861 : if (TREE_CODE (inner_rhs) == MEM_REF
1440 39259861 : && unmodified_parm (fbi, stmt, TREE_OPERAND (inner_rhs, 0), NULL))
1441 : rhs_free = true;
1442 :
1443 : /* Copying parameter passed by reference into gimple register is
1444 : probably also going to copy propagate, but we can't be quite
1445 : sure. */
1446 39259861 : if (rhs_free && is_gimple_reg (lhs))
1447 : lhs_free = true;
1448 :
1449 : /* Writes to parameters, parameters passed by value and return value
1450 : (either directly or passed via invisible reference) are free.
1451 :
1452 : TODO: We ought to handle testcase like
1453 : struct a {int a,b;};
1454 : struct a
1455 : returnstruct (void)
1456 : {
1457 : struct a a ={1,2};
1458 : return a;
1459 : }
1460 :
1461 : This translate into:
1462 :
1463 : returnstruct ()
1464 : {
1465 : int a$b;
1466 : int a$a;
1467 : struct a a;
1468 : struct a D.2739;
1469 :
1470 : <bb 2>:
1471 : D.2739.a = 1;
1472 : D.2739.b = 2;
1473 : return D.2739;
1474 :
1475 : }
1476 : For that we either need to copy ipa-split logic detecting writes
1477 : to return value. */
1478 39259861 : if (TREE_CODE (inner_lhs) == PARM_DECL
1479 39169996 : || TREE_CODE (inner_lhs) == RESULT_DECL
1480 77581925 : || (TREE_CODE (inner_lhs) == MEM_REF
1481 5206641 : && (unmodified_parm (fbi, stmt, TREE_OPERAND (inner_lhs, 0),
1482 : NULL)
1483 2680105 : || (TREE_CODE (TREE_OPERAND (inner_lhs, 0)) == SSA_NAME
1484 2678917 : && SSA_NAME_VAR (TREE_OPERAND (inner_lhs, 0))
1485 379099 : && TREE_CODE (SSA_NAME_VAR (TREE_OPERAND
1486 : (inner_lhs,
1487 : 0))) == RESULT_DECL))))
1488 : lhs_free = true;
1489 35704072 : if (lhs_free
1490 39259861 : && (is_gimple_reg (rhs) || is_gimple_min_invariant (rhs)))
1491 : rhs_free = true;
1492 39259861 : if (lhs_free && rhs_free)
1493 8226738 : return 1;
1494 : }
1495 : return 0;
1496 : default:
1497 : return 0;
1498 : }
1499 : }
1500 :
1501 : /* Analyze EXPR if it represents a series of simple operations performed on
1502 : a function parameter and return true if so. FBI, STMT, EXPR, INDEX_P and
1503 : AGGPOS have the same meaning like in unmodified_parm_or_parm_agg_item.
1504 : Type of the parameter or load from an aggregate via the parameter is
1505 : stored in *TYPE_P. Operations on the parameter are recorded to
1506 : PARAM_OPS_P if it is not NULL. */
1507 :
1508 : static bool
1509 27877261 : decompose_param_expr (struct ipa_func_body_info *fbi,
1510 : gimple *stmt, tree expr,
1511 : int *index_p, tree *type_p,
1512 : struct agg_position_info *aggpos,
1513 : expr_eval_ops *param_ops_p = NULL)
1514 : {
1515 27877261 : int op_limit = opt_for_fn (fbi->node->decl, param_ipa_max_param_expr_ops);
1516 27877261 : int op_count = 0;
1517 :
1518 27877261 : if (param_ops_p)
1519 9579862 : *param_ops_p = NULL;
1520 :
1521 33993513 : while (true)
1522 : {
1523 33993513 : expr_eval_op eval_op;
1524 33993513 : unsigned rhs_count;
1525 33993513 : unsigned cst_count = 0;
1526 :
1527 33993513 : if (unmodified_parm_or_parm_agg_item (fbi, stmt, expr, index_p, NULL,
1528 : aggpos))
1529 : {
1530 4761403 : tree type = TREE_TYPE (expr);
1531 :
1532 4761403 : if (aggpos->agg_contents)
1533 : {
1534 : /* Stop if containing bit-field. */
1535 3903652 : if (TREE_CODE (expr) == BIT_FIELD_REF
1536 3903652 : || contains_bitfld_component_ref_p (expr))
1537 : break;
1538 : }
1539 :
1540 4734174 : *type_p = type;
1541 4734174 : return true;
1542 : }
1543 :
1544 29232110 : if (TREE_CODE (expr) != SSA_NAME || SSA_NAME_IS_DEFAULT_DEF (expr))
1545 : break;
1546 11029321 : stmt = SSA_NAME_DEF_STMT (expr);
1547 :
1548 11029321 : if (gcall *call = dyn_cast <gcall *> (stmt))
1549 : {
1550 2596437 : int flags = gimple_call_return_flags (call);
1551 2596437 : if (!(flags & ERF_RETURNS_ARG))
1552 3192244 : goto fail;
1553 235366 : int arg = flags & ERF_RETURN_ARG_MASK;
1554 235366 : if (arg >= (int)gimple_call_num_args (call))
1555 0 : goto fail;
1556 235366 : expr = gimple_call_arg (stmt, arg);
1557 4174812 : continue;
1558 235366 : }
1559 :
1560 8432884 : if (!is_gimple_assign (stmt = SSA_NAME_DEF_STMT (expr)))
1561 : break;
1562 :
1563 6712146 : switch (gimple_assign_rhs_class (stmt))
1564 : {
1565 3939446 : case GIMPLE_SINGLE_RHS:
1566 3939446 : expr = gimple_assign_rhs1 (stmt);
1567 3939446 : continue;
1568 :
1569 : case GIMPLE_UNARY_RHS:
1570 : rhs_count = 1;
1571 : break;
1572 :
1573 1685097 : case GIMPLE_BINARY_RHS:
1574 1685097 : rhs_count = 2;
1575 1685097 : break;
1576 :
1577 696 : case GIMPLE_TERNARY_RHS:
1578 696 : rhs_count = 3;
1579 696 : break;
1580 :
1581 0 : default:
1582 0 : goto fail;
1583 : }
1584 :
1585 : /* Stop if expression is too complex. */
1586 2772700 : if (op_count++ == op_limit)
1587 : break;
1588 :
1589 2772613 : if (param_ops_p)
1590 : {
1591 2738316 : eval_op.code = gimple_assign_rhs_code (stmt);
1592 2738316 : eval_op.type = TREE_TYPE (gimple_assign_lhs (stmt));
1593 2738316 : eval_op.val[0] = NULL_TREE;
1594 2738316 : eval_op.val[1] = NULL_TREE;
1595 : }
1596 :
1597 2772613 : expr = NULL_TREE;
1598 6399852 : for (unsigned i = 0; i < rhs_count; i++)
1599 : {
1600 4458412 : tree op = gimple_op (stmt, i + 1);
1601 :
1602 4458412 : gcc_assert (op && !TYPE_P (op));
1603 4458412 : if (is_gimple_ip_invariant (op))
1604 : {
1605 856944 : if (++cst_count == rhs_count)
1606 1159 : goto fail;
1607 :
1608 855785 : eval_op.val[cst_count - 1] = op;
1609 : }
1610 3601468 : else if (!expr)
1611 : {
1612 : /* Found a non-constant operand, and record its index in rhs
1613 : operands. */
1614 2771454 : eval_op.index = i;
1615 2771454 : expr = op;
1616 : }
1617 : else
1618 : {
1619 : /* Found more than one non-constant operands. */
1620 830014 : goto fail;
1621 : }
1622 : }
1623 :
1624 1941440 : if (param_ops_p)
1625 1907763 : vec_safe_insert (*param_ops_p, 0, eval_op);
1626 : }
1627 :
1628 : /* Failed to decompose, free resource and return. */
1629 23143087 : fail:
1630 23143087 : if (param_ops_p)
1631 9382602 : vec_free (*param_ops_p);
1632 :
1633 : return false;
1634 : }
1635 :
1636 : /* Record to SUMMARY that PARM is used by builtin_constant_p. */
1637 :
1638 : static void
1639 12009 : add_builtin_constant_p_parm (class ipa_fn_summary *summary, int parm)
1640 : {
1641 12009 : int ip;
1642 :
1643 : /* Avoid duplicates. */
1644 12067 : for (unsigned int i = 0;
1645 12067 : summary->builtin_constant_p_parms.iterate (i, &ip); i++)
1646 5955 : if (ip == parm)
1647 12009 : return;
1648 6112 : summary->builtin_constant_p_parms.safe_push (parm);
1649 : }
1650 :
1651 : /* If BB ends by a conditional we can turn into predicates, attach corresponding
1652 : predicates to the CFG edges. */
1653 :
1654 : static void
1655 36025489 : set_cond_stmt_execution_predicate (struct ipa_func_body_info *fbi,
1656 : class ipa_fn_summary *summary,
1657 : class ipa_node_params *params_summary,
1658 : basic_block bb)
1659 : {
1660 36025489 : tree op, op2;
1661 36025489 : int index;
1662 36025489 : struct agg_position_info aggpos;
1663 36025489 : enum tree_code code, inverted_code;
1664 36025489 : edge e;
1665 36025489 : edge_iterator ei;
1666 36025489 : gimple *set_stmt;
1667 36025489 : tree param_type;
1668 36025489 : expr_eval_ops param_ops;
1669 :
1670 72050978 : gcond *last = safe_dyn_cast <gcond *> (*gsi_last_bb (bb));
1671 11953107 : if (!last)
1672 36012469 : return;
1673 11953107 : if (!is_gimple_ip_invariant (gimple_cond_rhs (last)))
1674 : return;
1675 9497976 : op = gimple_cond_lhs (last);
1676 :
1677 9497976 : if (decompose_param_expr (fbi, last, op, &index, ¶m_type, &aggpos,
1678 : ¶m_ops))
1679 : {
1680 1368747 : code = gimple_cond_code (last);
1681 1368747 : inverted_code = invert_tree_comparison (code, HONOR_NANS (op));
1682 :
1683 4106241 : FOR_EACH_EDGE (e, ei, bb->succs)
1684 : {
1685 1368747 : enum tree_code this_code = (e->flags & EDGE_TRUE_VALUE
1686 2737494 : ? code : inverted_code);
1687 : /* invert_tree_comparison will return ERROR_MARK on FP
1688 : comparisons that are not EQ/NE instead of returning proper
1689 : unordered one. Be sure it is not confused with NON_CONSTANT.
1690 :
1691 : And if the edge's target is the final block of diamond CFG graph
1692 : of this conditional statement, we do not need to compute
1693 : predicate for the edge because the final block's predicate must
1694 : be at least as that of the first block of the statement. */
1695 2737494 : if (this_code != ERROR_MARK
1696 2737494 : && !dominated_by_p (CDI_POST_DOMINATORS, bb, e->dest))
1697 : {
1698 2329130 : ipa_predicate p
1699 2329130 : = add_condition (summary, params_summary, index,
1700 : param_type, &aggpos,
1701 : this_code, gimple_cond_rhs (last), param_ops);
1702 2329130 : e->aux = edge_predicate_pool.allocate ();
1703 2329130 : *(ipa_predicate *) e->aux = p;
1704 : }
1705 : }
1706 36244071 : vec_free (param_ops);
1707 : return;
1708 : }
1709 :
1710 8129229 : if (TREE_CODE (op) != SSA_NAME)
1711 : return;
1712 : /* Special case
1713 : if (builtin_constant_p (op))
1714 : constant_code
1715 : else
1716 : nonconstant_code.
1717 : Here we can predicate nonconstant_code. We can't
1718 : really handle constant_code since we have no predicate
1719 : for this and also the constant code is not known to be
1720 : optimized away when inliner doesn't see operand is constant.
1721 : Other optimizers might think otherwise. */
1722 8128144 : if (gimple_cond_code (last) != NE_EXPR
1723 8128144 : || !integer_zerop (gimple_cond_rhs (last)))
1724 : return;
1725 3578163 : set_stmt = SSA_NAME_DEF_STMT (op);
1726 3578163 : if (!gimple_call_builtin_p (set_stmt, BUILT_IN_CONSTANT_P)
1727 3578163 : || gimple_call_num_args (set_stmt) != 1)
1728 : return;
1729 46400 : op2 = gimple_call_arg (set_stmt, 0);
1730 46400 : if (!decompose_param_expr (fbi, set_stmt, op2, &index, ¶m_type, &aggpos))
1731 : return;
1732 13020 : if (!aggpos.by_ref)
1733 11961 : add_builtin_constant_p_parm (summary, index);
1734 39060 : FOR_EACH_EDGE (e, ei, bb->succs) if (e->flags & EDGE_FALSE_VALUE)
1735 : {
1736 13020 : ipa_predicate p = add_condition (summary, params_summary, index,
1737 : param_type, &aggpos,
1738 : ipa_predicate::is_not_constant, NULL_TREE);
1739 13020 : e->aux = edge_predicate_pool.allocate ();
1740 13020 : *(ipa_predicate *) e->aux = p;
1741 : }
1742 : }
1743 :
1744 :
1745 : /* If BB ends by a switch we can turn into predicates, attach corresponding
1746 : predicates to the CFG edges. */
1747 :
1748 : static void
1749 36025489 : set_switch_stmt_execution_predicate (struct ipa_func_body_info *fbi,
1750 : class ipa_fn_summary *summary,
1751 : class ipa_node_params *params_summary,
1752 : basic_block bb)
1753 : {
1754 36025489 : tree op;
1755 36025489 : int index;
1756 36025489 : struct agg_position_info aggpos;
1757 36025489 : edge e;
1758 36025489 : edge_iterator ei;
1759 36025489 : size_t n;
1760 36025489 : size_t case_idx;
1761 36025489 : tree param_type;
1762 36025489 : expr_eval_ops param_ops;
1763 :
1764 72050978 : gswitch *last = safe_dyn_cast <gswitch *> (*gsi_last_bb (bb));
1765 81886 : if (!last)
1766 35995966 : return;
1767 81886 : op = gimple_switch_index (last);
1768 81886 : if (!decompose_param_expr (fbi, last, op, &index, ¶m_type, &aggpos,
1769 : ¶m_ops))
1770 : return;
1771 :
1772 44635 : auto_vec<std::pair<tree, tree> > ranges;
1773 44635 : tree type = TREE_TYPE (op);
1774 44635 : int bound_limit = opt_for_fn (fbi->node->decl,
1775 : param_ipa_max_switch_predicate_bounds);
1776 44635 : int bound_count = 0;
1777 : // This can safely be an integer range, as switches can only hold
1778 : // integers.
1779 44635 : int_range<2> vr;
1780 :
1781 89270 : get_range_query (cfun)->range_of_expr (vr, op);
1782 44635 : if (vr.undefined_p ())
1783 0 : vr.set_varying (TREE_TYPE (op));
1784 44635 : tree vr_min, vr_max;
1785 : // TODO: This entire function could use a rewrite to use the irange
1786 : // API, instead of trying to recreate its intersection/union logic.
1787 : // Any use of get_legacy_range() is a serious code smell.
1788 44635 : value_range_kind vr_type = get_legacy_range (vr, vr_min, vr_max);
1789 44635 : wide_int vr_wmin = wi::to_wide (vr_min);
1790 44635 : wide_int vr_wmax = wi::to_wide (vr_max);
1791 :
1792 308831 : FOR_EACH_EDGE (e, ei, bb->succs)
1793 : {
1794 264196 : e->aux = edge_predicate_pool.allocate ();
1795 264196 : *(ipa_predicate *) e->aux = false;
1796 : }
1797 :
1798 44635 : e = gimple_switch_edge (cfun, last, 0);
1799 : /* Set BOUND_COUNT to maximum count to bypass computing predicate for
1800 : default case if its target basic block is in convergence point of all
1801 : switch cases, which can be determined by checking whether it
1802 : post-dominates the switch statement. */
1803 44635 : if (dominated_by_p (CDI_POST_DOMINATORS, bb, e->dest))
1804 12603 : bound_count = INT_MAX;
1805 :
1806 44635 : n = gimple_switch_num_labels (last);
1807 295307 : for (case_idx = 1; case_idx < n; ++case_idx)
1808 : {
1809 250672 : tree cl = gimple_switch_label (last, case_idx);
1810 250672 : tree min = CASE_LOW (cl);
1811 250672 : tree max = CASE_HIGH (cl);
1812 250672 : ipa_predicate p;
1813 :
1814 250672 : e = gimple_switch_edge (cfun, last, case_idx);
1815 :
1816 : /* The case value might not have same type as switch expression,
1817 : extend the value based on the expression type. */
1818 250672 : if (TREE_TYPE (min) != type)
1819 44908 : min = wide_int_to_tree (type, wi::to_wide (min));
1820 :
1821 250672 : if (!max)
1822 : max = min;
1823 12477 : else if (TREE_TYPE (max) != type)
1824 3427 : max = wide_int_to_tree (type, wi::to_wide (max));
1825 :
1826 : /* The case's target basic block is in convergence point of all switch
1827 : cases, its predicate should be at least as that of the switch
1828 : statement. */
1829 250672 : if (dominated_by_p (CDI_POST_DOMINATORS, bb, e->dest))
1830 5189 : p = true;
1831 245483 : else if (min == max)
1832 234922 : p = add_condition (summary, params_summary, index, param_type,
1833 : &aggpos, EQ_EXPR, min, param_ops);
1834 : else
1835 : {
1836 10561 : ipa_predicate p1, p2;
1837 10561 : p1 = add_condition (summary, params_summary, index, param_type,
1838 : &aggpos, GE_EXPR, min, param_ops);
1839 10561 : p2 = add_condition (summary, params_summary,index, param_type,
1840 : &aggpos, LE_EXPR, max, param_ops);
1841 10561 : p = p1 & p2;
1842 : }
1843 250672 : *(ipa_predicate *) e->aux
1844 250672 : = p.or_with (summary->conds, *(ipa_predicate *) e->aux);
1845 :
1846 : /* If there are too many disjoint case ranges, predicate for default
1847 : case might become too complicated. So add a limit here. */
1848 250672 : if (bound_count > bound_limit)
1849 97925 : continue;
1850 :
1851 152747 : bool new_range = true;
1852 :
1853 152747 : if (!ranges.is_empty ())
1854 : {
1855 120715 : wide_int curr_wmin = wi::to_wide (min);
1856 120715 : wide_int last_wmax = wi::to_wide (ranges.last ().second);
1857 :
1858 : /* Merge case ranges if they are continuous. */
1859 120715 : if (curr_wmin == last_wmax + 1)
1860 : new_range = false;
1861 41874 : else if (vr_type == VR_ANTI_RANGE)
1862 : {
1863 : /* If two disjoint case ranges can be connected by anti-range
1864 : of switch index, combine them to one range. */
1865 141 : if (wi::lt_p (vr_wmax, curr_wmin - 1, TYPE_SIGN (type)))
1866 : vr_type = VR_UNDEFINED;
1867 0 : else if (wi::le_p (vr_wmin, last_wmax + 1, TYPE_SIGN (type)))
1868 0 : new_range = false;
1869 : }
1870 120715 : }
1871 :
1872 : /* Create/extend a case range. And we count endpoints of range set,
1873 : this number nearly equals to number of conditions that we will create
1874 : for predicate of default case. */
1875 120715 : if (new_range)
1876 : {
1877 73906 : bound_count += (min == max) ? 1 : 2;
1878 73906 : ranges.safe_push (std::make_pair (min, max));
1879 : }
1880 : else
1881 : {
1882 78841 : bound_count += (ranges.last ().first == ranges.last ().second);
1883 78841 : ranges.last ().second = max;
1884 : }
1885 : }
1886 :
1887 44635 : e = gimple_switch_edge (cfun, last, 0);
1888 44635 : if (bound_count > bound_limit)
1889 : {
1890 15112 : *(ipa_predicate *) e->aux = true;
1891 15112 : vec_free (param_ops);
1892 15112 : return;
1893 : }
1894 :
1895 29523 : ipa_predicate p_seg = true;
1896 29523 : ipa_predicate p_all = false;
1897 :
1898 29523 : if (vr_type != VR_RANGE)
1899 : {
1900 28923 : vr_wmin = wi::to_wide (TYPE_MIN_VALUE (type));
1901 28923 : vr_wmax = wi::to_wide (TYPE_MAX_VALUE (type));
1902 : }
1903 :
1904 : /* Construct predicate to represent default range set that is negation of
1905 : all case ranges. Case range is classified as containing single/non-single
1906 : values. Suppose a piece of case ranges in the following.
1907 :
1908 : [D1...D2] [S1] ... [Sn] [D3...D4]
1909 :
1910 : To represent default case's range sets between two non-single value
1911 : case ranges (From D2 to D3), we construct predicate as:
1912 :
1913 : D2 < x < D3 && x != S1 && ... && x != Sn
1914 : */
1915 90996 : for (size_t i = 0; i < ranges.length (); i++)
1916 : {
1917 61568 : tree min = ranges[i].first;
1918 61568 : tree max = ranges[i].second;
1919 :
1920 61568 : if (min == max)
1921 73378 : p_seg &= add_condition (summary, params_summary, index,
1922 : param_type, &aggpos, NE_EXPR,
1923 36689 : min, param_ops);
1924 : else
1925 : {
1926 : /* Do not create sub-predicate for range that is beyond low bound
1927 : of switch index. */
1928 24879 : if (wi::lt_p (vr_wmin, wi::to_wide (min), TYPE_SIGN (type)))
1929 : {
1930 17797 : p_seg &= add_condition (summary, params_summary, index,
1931 : param_type, &aggpos,
1932 17797 : LT_EXPR, min, param_ops);
1933 17797 : p_all = p_all.or_with (summary->conds, p_seg);
1934 : }
1935 :
1936 : /* Do not create sub-predicate for range that is beyond up bound
1937 : of switch index. */
1938 24879 : if (wi::le_p (vr_wmax, wi::to_wide (max), TYPE_SIGN (type)))
1939 : {
1940 95 : p_seg = false;
1941 95 : break;
1942 : }
1943 :
1944 24784 : p_seg = add_condition (summary, params_summary, index,
1945 : param_type, &aggpos, GT_EXPR,
1946 : max, param_ops);
1947 : }
1948 : }
1949 :
1950 29523 : p_all = p_all.or_with (summary->conds, p_seg);
1951 29523 : *(ipa_predicate *) e->aux
1952 29523 : = p_all.or_with (summary->conds, *(ipa_predicate *) e->aux);
1953 :
1954 36100 : vec_free (param_ops);
1955 44635 : }
1956 :
1957 :
1958 : /* For each BB in NODE attach to its AUX pointer predicate under
1959 : which it is executable. */
1960 :
1961 : static void
1962 6560932 : compute_bb_predicates (struct ipa_func_body_info *fbi,
1963 : struct cgraph_node *node,
1964 : class ipa_fn_summary *summary,
1965 : class ipa_node_params *params_summary)
1966 : {
1967 6560932 : struct function *my_function = DECL_STRUCT_FUNCTION (node->decl);
1968 6560932 : bool done = false;
1969 6560932 : basic_block bb;
1970 :
1971 42586421 : FOR_EACH_BB_FN (bb, my_function)
1972 : {
1973 36025489 : set_cond_stmt_execution_predicate (fbi, summary, params_summary, bb);
1974 36025489 : set_switch_stmt_execution_predicate (fbi, summary, params_summary, bb);
1975 : }
1976 :
1977 : /* Entry block is always executable. */
1978 6560932 : ENTRY_BLOCK_PTR_FOR_FN (my_function)->aux
1979 6560932 : = edge_predicate_pool.allocate ();
1980 6560932 : *(ipa_predicate *) ENTRY_BLOCK_PTR_FOR_FN (my_function)->aux = true;
1981 :
1982 : /* A simple dataflow propagation of predicates forward in the CFG.
1983 : TODO: work in reverse postorder. */
1984 19802445 : while (!done)
1985 : {
1986 13241513 : done = true;
1987 90387516 : FOR_EACH_BB_FN (bb, my_function)
1988 : {
1989 77146003 : ipa_predicate p = false;
1990 77146003 : edge e;
1991 77146003 : edge_iterator ei;
1992 95534551 : FOR_EACH_EDGE (e, ei, bb->preds)
1993 : {
1994 81990353 : if (e->src->aux)
1995 : {
1996 80257628 : ipa_predicate this_bb_predicate
1997 : = *(ipa_predicate *) e->src->aux;
1998 80257628 : if (e->aux)
1999 5416356 : this_bb_predicate &= (*(ipa_predicate *) e->aux);
2000 80257628 : p = p.or_with (summary->conds, this_bb_predicate);
2001 80257628 : if (p == true)
2002 : break;
2003 : }
2004 : }
2005 77146003 : if (p != false)
2006 : {
2007 76018979 : basic_block pdom_bb;
2008 :
2009 76018979 : if (!bb->aux)
2010 : {
2011 28316355 : done = false;
2012 28316355 : bb->aux = edge_predicate_pool.allocate ();
2013 28316355 : *((ipa_predicate *) bb->aux) = p;
2014 : }
2015 47702624 : else if (p != *(ipa_predicate *) bb->aux)
2016 : {
2017 : /* This OR operation is needed to ensure monotonous data flow
2018 : in the case we hit the limit on number of clauses and the
2019 : and/or operations above give approximate answers. */
2020 128793 : p = p.or_with (summary->conds, *(ipa_predicate *)bb->aux);
2021 128793 : if (p != *(ipa_predicate *)bb->aux)
2022 : {
2023 104192 : done = false;
2024 104192 : *((ipa_predicate *)bb->aux) = p;
2025 : }
2026 : }
2027 :
2028 : /* For switch/if statement, we can OR-combine predicates of all
2029 : its cases/branches to get predicate for basic block in their
2030 : convergence point, but sometimes this will generate very
2031 : complicated predicate. Actually, we can get simplified
2032 : predicate in another way by using the fact that predicate
2033 : for a basic block must also hold true for its post dominators.
2034 : To be specific, basic block in convergence point of
2035 : conditional statement should include predicate of the
2036 : statement. */
2037 76018979 : pdom_bb = get_immediate_dominator (CDI_POST_DOMINATORS, bb);
2038 76018979 : if (pdom_bb == EXIT_BLOCK_PTR_FOR_FN (my_function) || !pdom_bb)
2039 : ;
2040 38817682 : else if (!pdom_bb->aux)
2041 : {
2042 7709120 : done = false;
2043 7709120 : pdom_bb->aux = edge_predicate_pool.allocate ();
2044 7709120 : *((ipa_predicate *)pdom_bb->aux) = p;
2045 : }
2046 31108562 : else if (p != *(ipa_predicate *)pdom_bb->aux)
2047 : {
2048 3447539 : p = p.or_with (summary->conds,
2049 : *(ipa_predicate *)pdom_bb->aux);
2050 3447539 : if (p != *(ipa_predicate *)pdom_bb->aux)
2051 : {
2052 141600 : done = false;
2053 141600 : *((ipa_predicate *)pdom_bb->aux) = p;
2054 : }
2055 : }
2056 : }
2057 : }
2058 : }
2059 6560932 : }
2060 :
2061 :
2062 : /* Return predicate specifying when the STMT might have result that is not
2063 : a compile time constant. */
2064 :
2065 : static ipa_predicate
2066 4274359 : will_be_nonconstant_expr_predicate (ipa_func_body_info *fbi,
2067 : class ipa_fn_summary *summary,
2068 : class ipa_node_params *params_summary,
2069 : tree expr,
2070 : vec<ipa_predicate> nonconstant_names)
2071 : {
2072 4274359 : tree parm;
2073 4274359 : int index;
2074 :
2075 4519713 : while (UNARY_CLASS_P (expr))
2076 245354 : expr = TREE_OPERAND (expr, 0);
2077 :
2078 4274359 : parm = unmodified_parm (fbi, NULL, expr, NULL);
2079 4274359 : if (parm && (index = ipa_get_param_decl_index (fbi->info, parm)) >= 0)
2080 223775 : return add_condition (summary, params_summary, index, TREE_TYPE (parm), NULL,
2081 223775 : ipa_predicate::changed, NULL_TREE);
2082 4050584 : if (is_gimple_min_invariant (expr))
2083 63584 : return false;
2084 3987000 : if (TREE_CODE (expr) == SSA_NAME)
2085 3805643 : return nonconstant_names[SSA_NAME_VERSION (expr)];
2086 181357 : if (BINARY_CLASS_P (expr) || COMPARISON_CLASS_P (expr))
2087 : {
2088 180683 : ipa_predicate p1
2089 180683 : = will_be_nonconstant_expr_predicate (fbi, summary,
2090 : params_summary,
2091 180683 : TREE_OPERAND (expr, 0),
2092 : nonconstant_names);
2093 180683 : if (p1 == true)
2094 103933 : return p1;
2095 :
2096 76750 : ipa_predicate p2
2097 76750 : = will_be_nonconstant_expr_predicate (fbi, summary,
2098 : params_summary,
2099 76750 : TREE_OPERAND (expr, 1),
2100 : nonconstant_names);
2101 76750 : return p1.or_with (summary->conds, p2);
2102 : }
2103 674 : else if (TREE_CODE (expr) == COND_EXPR)
2104 : {
2105 188 : ipa_predicate p1
2106 188 : = will_be_nonconstant_expr_predicate (fbi, summary,
2107 : params_summary,
2108 188 : TREE_OPERAND (expr, 0),
2109 : nonconstant_names);
2110 188 : if (p1 == true)
2111 37 : return p1;
2112 :
2113 151 : ipa_predicate p2
2114 151 : = will_be_nonconstant_expr_predicate (fbi, summary,
2115 : params_summary,
2116 151 : TREE_OPERAND (expr, 1),
2117 : nonconstant_names);
2118 151 : if (p2 == true)
2119 151 : return p2;
2120 0 : p1 = p1.or_with (summary->conds, p2);
2121 0 : p2 = will_be_nonconstant_expr_predicate (fbi, summary,
2122 : params_summary,
2123 0 : TREE_OPERAND (expr, 2),
2124 : nonconstant_names);
2125 0 : return p2.or_with (summary->conds, p1);
2126 : }
2127 486 : else if (TREE_CODE (expr) == CALL_EXPR)
2128 486 : return true;
2129 : else
2130 : {
2131 0 : debug_tree (expr);
2132 0 : gcc_unreachable ();
2133 : }
2134 : }
2135 :
2136 :
2137 : /* Return predicate specifying when the STMT might have result that is not
2138 : a compile time constant. */
2139 :
2140 : static ipa_predicate
2141 117083517 : will_be_nonconstant_predicate (struct ipa_func_body_info *fbi,
2142 : class ipa_fn_summary *summary,
2143 : class ipa_node_params *params_summary,
2144 : gimple *stmt,
2145 : vec<ipa_predicate> nonconstant_names)
2146 : {
2147 117083517 : ipa_predicate p = true;
2148 117083517 : ssa_op_iter iter;
2149 117083517 : tree use;
2150 117083517 : tree param_type = NULL_TREE;
2151 117083517 : ipa_predicate op_non_const;
2152 117083517 : bool is_load;
2153 117083517 : int base_index;
2154 117083517 : struct agg_position_info aggpos;
2155 :
2156 : /* What statements might be optimized away
2157 : when their arguments are constant. */
2158 117083517 : if (gimple_code (stmt) != GIMPLE_ASSIGN
2159 : && gimple_code (stmt) != GIMPLE_COND
2160 : && gimple_code (stmt) != GIMPLE_SWITCH
2161 : && (gimple_code (stmt) != GIMPLE_CALL
2162 21398794 : || !(gimple_call_flags (stmt) & ECF_CONST)))
2163 26830624 : return p;
2164 :
2165 : /* Stores will stay anyway. */
2166 90252893 : if (gimple_store_p (stmt))
2167 26540469 : return p;
2168 :
2169 63712424 : is_load = gimple_assign_load_p (stmt);
2170 :
2171 : /* Loads can be optimized when the value is known. */
2172 63712424 : if (is_load)
2173 : {
2174 18250999 : tree op = gimple_assign_rhs1 (stmt);
2175 18250999 : if (!decompose_param_expr (fbi, stmt, op, &base_index, ¶m_type,
2176 : &aggpos))
2177 14943227 : return p;
2178 : }
2179 : else
2180 45461425 : base_index = -1;
2181 :
2182 : /* See if we understand all operands before we start
2183 : adding conditionals. */
2184 66534855 : FOR_EACH_SSA_TREE_OPERAND (use, stmt, iter, SSA_OP_USE)
2185 : {
2186 48613997 : tree parm = unmodified_parm (fbi, stmt, use, NULL);
2187 : /* For arguments we can build a condition. */
2188 48613997 : if (parm && ipa_get_param_decl_index (fbi->info, parm) >= 0)
2189 8701686 : continue;
2190 39912311 : if (TREE_CODE (use) != SSA_NAME)
2191 0 : return p;
2192 : /* If we know when operand is constant,
2193 : we still can say something useful. */
2194 39912311 : if (nonconstant_names[SSA_NAME_VERSION (use)] != true)
2195 9063972 : continue;
2196 30848339 : return p;
2197 : }
2198 :
2199 17920858 : if (is_load)
2200 3307691 : op_non_const =
2201 3307691 : add_condition (summary, params_summary,
2202 : base_index, param_type, &aggpos,
2203 : ipa_predicate::changed, NULL_TREE);
2204 : else
2205 14613167 : op_non_const = false;
2206 34640706 : FOR_EACH_SSA_TREE_OPERAND (use, stmt, iter, SSA_OP_USE)
2207 : {
2208 16719848 : tree parm = unmodified_parm (fbi, stmt, use, NULL);
2209 16719848 : int index;
2210 :
2211 16719848 : if (parm && (index = ipa_get_param_decl_index (fbi->info, parm)) >= 0)
2212 : {
2213 8202523 : if (index != base_index)
2214 5600177 : p = add_condition (summary, params_summary, index,
2215 5600177 : TREE_TYPE (parm), NULL,
2216 : ipa_predicate::changed, NULL_TREE);
2217 : else
2218 2602346 : continue;
2219 : }
2220 : else
2221 8517325 : p = nonconstant_names[SSA_NAME_VERSION (use)];
2222 14117502 : op_non_const = p.or_with (summary->conds, op_non_const);
2223 : }
2224 17920858 : if ((gimple_code (stmt) == GIMPLE_ASSIGN || gimple_code (stmt) == GIMPLE_CALL)
2225 15565400 : && gimple_op (stmt, 0)
2226 33181056 : && TREE_CODE (gimple_op (stmt, 0)) == SSA_NAME)
2227 15260198 : nonconstant_names[SSA_NAME_VERSION (gimple_op (stmt, 0))]
2228 15260198 : = op_non_const;
2229 17920858 : return op_non_const;
2230 : }
2231 :
2232 : struct record_modified_bb_info
2233 : {
2234 : tree op;
2235 : bitmap bb_set;
2236 : gimple *stmt;
2237 : };
2238 :
2239 : /* Value is initialized in INIT_BB and used in USE_BB. We want to compute
2240 : probability how often it changes between USE_BB.
2241 : INIT_BB->count/USE_BB->count is an estimate, but if INIT_BB
2242 : is in different loop nest, we can do better.
2243 : This is all just estimate. In theory we look for minimal cut separating
2244 : INIT_BB and USE_BB, but we only want to anticipate loop invariant motion
2245 : anyway. */
2246 :
2247 : static basic_block
2248 10693496 : get_minimal_bb (basic_block init_bb, basic_block use_bb)
2249 : {
2250 10693496 : class loop *l = find_common_loop (init_bb->loop_father, use_bb->loop_father);
2251 10693496 : if (l && l->header->count < init_bb->count)
2252 386097 : return l->header;
2253 : return init_bb;
2254 : }
2255 :
2256 : /* Callback of walk_aliased_vdefs. Records basic blocks where the value may be
2257 : set except for info->stmt. */
2258 :
2259 : static bool
2260 4952756 : record_modified (ao_ref *ao ATTRIBUTE_UNUSED, tree vdef, void *data)
2261 : {
2262 4952756 : struct record_modified_bb_info *info =
2263 : (struct record_modified_bb_info *) data;
2264 4952756 : if (SSA_NAME_DEF_STMT (vdef) == info->stmt)
2265 : return false;
2266 4924573 : if (gimple_clobber_p (SSA_NAME_DEF_STMT (vdef)))
2267 : return false;
2268 4874992 : bitmap_set_bit (info->bb_set,
2269 4874992 : SSA_NAME_IS_DEFAULT_DEF (vdef)
2270 0 : ? ENTRY_BLOCK_PTR_FOR_FN (cfun)->index
2271 : : get_minimal_bb
2272 4874992 : (gimple_bb (SSA_NAME_DEF_STMT (vdef)),
2273 4874992 : gimple_bb (info->stmt))->index);
2274 4874992 : if (dump_file)
2275 : {
2276 0 : fprintf (dump_file, " Param ");
2277 0 : print_generic_expr (dump_file, info->op, TDF_SLIM);
2278 0 : fprintf (dump_file, " changed at bb %i, minimal: %i stmt: ",
2279 0 : gimple_bb (SSA_NAME_DEF_STMT (vdef))->index,
2280 : get_minimal_bb
2281 0 : (gimple_bb (SSA_NAME_DEF_STMT (vdef)),
2282 0 : gimple_bb (info->stmt))->index);
2283 0 : print_gimple_stmt (dump_file, SSA_NAME_DEF_STMT (vdef), 0);
2284 : }
2285 : return false;
2286 : }
2287 :
2288 : /* Return probability (based on REG_BR_PROB_BASE) that I-th parameter of STMT
2289 : will change since last invocation of STMT.
2290 :
2291 : Value 0 is reserved for compile time invariants.
2292 : For common parameters it is REG_BR_PROB_BASE. For loop invariants it
2293 : ought to be REG_BR_PROB_BASE / estimated_iters. */
2294 :
2295 : static int
2296 39633232 : param_change_prob (ipa_func_body_info *fbi, gimple *stmt, int i)
2297 : {
2298 39633232 : tree op = gimple_call_arg (stmt, i);
2299 39633232 : basic_block bb = gimple_bb (stmt);
2300 :
2301 39633232 : if (TREE_CODE (op) == WITH_SIZE_EXPR)
2302 299 : op = TREE_OPERAND (op, 0);
2303 :
2304 39633232 : tree base = get_base_address (op);
2305 :
2306 : /* Global invariants never change. */
2307 39633232 : if (is_gimple_min_invariant (base))
2308 : return 0;
2309 :
2310 : /* We would have to do non-trivial analysis to really work out what
2311 : is the probability of value to change (i.e. when init statement
2312 : is in a sibling loop of the call).
2313 :
2314 : We do an conservative estimate: when call is executed N times more often
2315 : than the statement defining value, we take the frequency 1/N. */
2316 19903368 : if (TREE_CODE (base) == SSA_NAME)
2317 : {
2318 17506127 : profile_count init_count;
2319 :
2320 26857862 : if (!bb->count.nonzero_p ())
2321 : return REG_BR_PROB_BASE;
2322 :
2323 8945684 : if (SSA_NAME_IS_DEFAULT_DEF (base))
2324 3127180 : init_count = ENTRY_BLOCK_PTR_FOR_FN (cfun)->count;
2325 : else
2326 5818504 : init_count = get_minimal_bb
2327 5818504 : (gimple_bb (SSA_NAME_DEF_STMT (base)),
2328 : gimple_bb (stmt))->count;
2329 :
2330 8945684 : if (init_count < bb->count)
2331 461063 : return MAX ((init_count.to_sreal_scale (bb->count)
2332 : * REG_BR_PROB_BASE).to_int (), 1);
2333 : return REG_BR_PROB_BASE;
2334 : }
2335 : else
2336 : {
2337 2397241 : ao_ref refd;
2338 2397241 : profile_count max = ENTRY_BLOCK_PTR_FOR_FN (cfun)->count;
2339 2397241 : struct record_modified_bb_info info;
2340 2397241 : tree init = ctor_for_folding (base);
2341 :
2342 2397241 : if (init != error_mark_node)
2343 : return 0;
2344 5358791 : if (!bb->count.nonzero_p () || fbi->aa_walk_budget == 0)
2345 : return REG_BR_PROB_BASE;
2346 1469611 : if (dump_file)
2347 : {
2348 0 : fprintf (dump_file, " Analyzing param change probability of ");
2349 0 : print_generic_expr (dump_file, op, TDF_SLIM);
2350 0 : fprintf (dump_file, "\n");
2351 : }
2352 1469611 : ao_ref_init (&refd, op);
2353 1469611 : info.op = op;
2354 1469611 : info.stmt = stmt;
2355 1469611 : info.bb_set = BITMAP_ALLOC (NULL);
2356 1469611 : int walked
2357 2939222 : = walk_aliased_vdefs (&refd, gimple_vuse (stmt), record_modified, &info,
2358 : NULL, NULL, fbi->aa_walk_budget);
2359 1469611 : if (walked > 0)
2360 1335628 : fbi->aa_walk_budget -= walked;
2361 1469611 : if (walked < 0 || bitmap_bit_p (info.bb_set, bb->index))
2362 : {
2363 895365 : if (walked < 0)
2364 206 : fbi->aa_walk_budget = 0;
2365 895365 : if (dump_file)
2366 : {
2367 0 : if (walked < 0)
2368 0 : fprintf (dump_file, " Ran out of AA walking budget.\n");
2369 : else
2370 0 : fprintf (dump_file, " Set in same BB as used.\n");
2371 : }
2372 895365 : BITMAP_FREE (info.bb_set);
2373 895365 : return REG_BR_PROB_BASE;
2374 : }
2375 :
2376 574246 : bitmap_iterator bi;
2377 574246 : unsigned index;
2378 : /* Lookup the most frequent update of the value and believe that
2379 : it dominates all the other; precise analysis here is difficult. */
2380 1612694 : EXECUTE_IF_SET_IN_BITMAP (info.bb_set, 0, index, bi)
2381 1038448 : max = profile_count::max_prefer_initialized
2382 1038448 : (max, BASIC_BLOCK_FOR_FN (cfun, index)->count);
2383 574246 : if (dump_file)
2384 : {
2385 0 : fprintf (dump_file, " Set with count ");
2386 0 : max.dump (dump_file);
2387 0 : fprintf (dump_file, " and used with count ");
2388 0 : bb->count.dump (dump_file);
2389 0 : fprintf (dump_file, " freq %f\n",
2390 0 : max.to_sreal_scale (bb->count).to_double ());
2391 : }
2392 :
2393 574246 : BITMAP_FREE (info.bb_set);
2394 574246 : if (max < bb->count)
2395 95442 : return MAX ((max.to_sreal_scale (bb->count)
2396 : * REG_BR_PROB_BASE).to_int (), 1);
2397 : return REG_BR_PROB_BASE;
2398 : }
2399 : }
2400 :
2401 : /* Find whether a basic block BB is the final block of a (half) diamond CFG
2402 : sub-graph and if the predicate the condition depends on is known. If so,
2403 : return true and store the pointer the predicate in *P. */
2404 :
2405 : static bool
2406 6482553 : phi_result_unknown_predicate (ipa_func_body_info *fbi,
2407 : ipa_fn_summary *summary,
2408 : class ipa_node_params *params_summary,
2409 : basic_block bb,
2410 : ipa_predicate *p,
2411 : vec<ipa_predicate> nonconstant_names)
2412 : {
2413 6482553 : edge e;
2414 6482553 : edge_iterator ei;
2415 6482553 : basic_block first_bb = NULL;
2416 :
2417 6482553 : if (single_pred_p (bb))
2418 : {
2419 83158 : *p = false;
2420 83158 : return true;
2421 : }
2422 :
2423 14724759 : FOR_EACH_EDGE (e, ei, bb->preds)
2424 : {
2425 12603436 : if (single_succ_p (e->src))
2426 : {
2427 8429392 : if (!single_pred_p (e->src))
2428 : return false;
2429 7411810 : if (!first_bb)
2430 3257688 : first_bb = single_pred (e->src);
2431 4154122 : else if (single_pred (e->src) != first_bb)
2432 : return false;
2433 : }
2434 : else
2435 : {
2436 4174044 : if (!first_bb)
2437 : first_bb = e->src;
2438 1461533 : else if (e->src != first_bb)
2439 : return false;
2440 : }
2441 : }
2442 :
2443 2121323 : if (!first_bb)
2444 : return false;
2445 :
2446 4242646 : gcond *stmt = safe_dyn_cast <gcond *> (*gsi_last_bb (first_bb));
2447 2076538 : if (!stmt
2448 2076538 : || !is_gimple_ip_invariant (gimple_cond_rhs (stmt)))
2449 : return false;
2450 :
2451 1819653 : *p = will_be_nonconstant_expr_predicate (fbi, summary, params_summary,
2452 : gimple_cond_lhs (stmt),
2453 : nonconstant_names);
2454 1819653 : if (*p == true)
2455 : return false;
2456 : else
2457 607257 : return true;
2458 : }
2459 :
2460 : /* Given a PHI statement in a function described by inline properties SUMMARY
2461 : and *P being the predicate describing whether the selected PHI argument is
2462 : known, store a predicate for the result of the PHI statement into
2463 : NONCONSTANT_NAMES, if possible. */
2464 :
2465 : static void
2466 788438 : predicate_for_phi_result (class ipa_fn_summary *summary, gphi *phi,
2467 : ipa_predicate *p,
2468 : vec<ipa_predicate> nonconstant_names)
2469 : {
2470 788438 : unsigned i;
2471 :
2472 1109023 : for (i = 0; i < gimple_phi_num_args (phi); i++)
2473 : {
2474 950288 : tree arg = gimple_phi_arg (phi, i)->def;
2475 950288 : if (!is_gimple_min_invariant (arg))
2476 : {
2477 798231 : gcc_assert (TREE_CODE (arg) == SSA_NAME);
2478 1596462 : *p = p->or_with (summary->conds,
2479 798231 : nonconstant_names[SSA_NAME_VERSION (arg)]);
2480 798231 : if (*p == true)
2481 : return;
2482 : }
2483 : }
2484 :
2485 158735 : if (dump_file && (dump_flags & TDF_DETAILS))
2486 : {
2487 3 : fprintf (dump_file, "\t\tphi predicate: ");
2488 3 : p->dump (dump_file, summary->conds);
2489 : }
2490 158735 : nonconstant_names[SSA_NAME_VERSION (gimple_phi_result (phi))] = *p;
2491 : }
2492 :
2493 : /* For a typical usage of __builtin_expect (a<b, 1), we
2494 : may introduce an extra relation stmt:
2495 : With the builtin, we have
2496 : t1 = a <= b;
2497 : t2 = (long int) t1;
2498 : t3 = __builtin_expect (t2, 1);
2499 : if (t3 != 0)
2500 : goto ...
2501 : Without the builtin, we have
2502 : if (a<=b)
2503 : goto...
2504 : This affects the size/time estimation and may have
2505 : an impact on the earlier inlining.
2506 : Here find this pattern and fix it up later. */
2507 :
2508 : static gimple *
2509 42021189 : find_foldable_builtin_expect (basic_block bb)
2510 : {
2511 42021189 : gimple_stmt_iterator bsi;
2512 :
2513 295041495 : for (bsi = gsi_start_bb (bb); !gsi_end_p (bsi); gsi_next (&bsi))
2514 : {
2515 211186992 : gimple *stmt = gsi_stmt (bsi);
2516 211186992 : if (gimple_call_builtin_p (stmt, BUILT_IN_EXPECT)
2517 210946944 : || gimple_call_builtin_p (stmt, BUILT_IN_EXPECT_WITH_PROBABILITY)
2518 422133859 : || gimple_call_internal_p (stmt, IFN_BUILTIN_EXPECT))
2519 : {
2520 373190 : tree var = gimple_call_lhs (stmt);
2521 373190 : tree arg = gimple_call_arg (stmt, 0);
2522 373190 : use_operand_p use_p;
2523 373190 : gimple *use_stmt;
2524 373190 : bool match = false;
2525 373190 : bool done = false;
2526 :
2527 373190 : if (!var || !arg)
2528 3 : continue;
2529 373187 : gcc_assert (TREE_CODE (var) == SSA_NAME);
2530 :
2531 745007 : while (TREE_CODE (arg) == SSA_NAME)
2532 : {
2533 745007 : gimple *stmt_tmp = SSA_NAME_DEF_STMT (arg);
2534 745007 : if (!is_gimple_assign (stmt_tmp))
2535 : break;
2536 737659 : switch (gimple_assign_rhs_code (stmt_tmp))
2537 : {
2538 : case LT_EXPR:
2539 : case LE_EXPR:
2540 : case GT_EXPR:
2541 : case GE_EXPR:
2542 : case EQ_EXPR:
2543 : case NE_EXPR:
2544 : match = true;
2545 : done = true;
2546 : break;
2547 : CASE_CONVERT:
2548 : break;
2549 : default:
2550 : done = true;
2551 : break;
2552 : }
2553 371820 : if (done)
2554 : break;
2555 371820 : arg = gimple_assign_rhs1 (stmt_tmp);
2556 : }
2557 :
2558 336185 : if (match && single_imm_use (var, &use_p, &use_stmt)
2559 709054 : && gimple_code (use_stmt) == GIMPLE_COND)
2560 187875 : return use_stmt;
2561 : }
2562 : }
2563 : return NULL;
2564 : }
2565 :
2566 : /* Return true when the basic blocks contains only clobbers followed by RESX.
2567 : Such BBs are kept around to make removal of dead stores possible with
2568 : presence of EH and will be optimized out by optimize_clobbers later in the
2569 : game.
2570 :
2571 : NEED_EH is used to recurse in case the clobber has non-EH predecessors
2572 : that can be clobber only, too.. When it is false, the RESX is not necessary
2573 : on the end of basic block. */
2574 :
2575 : static bool
2576 42671662 : clobber_only_eh_bb_p (basic_block bb, bool need_eh = true)
2577 : {
2578 42671662 : gimple_stmt_iterator gsi = gsi_last_bb (bb);
2579 42671662 : edge_iterator ei;
2580 42671662 : edge e;
2581 :
2582 42671662 : if (need_eh)
2583 : {
2584 42601512 : if (gsi_end_p (gsi))
2585 : return false;
2586 41452738 : if (gimple_code (gsi_stmt (gsi)) != GIMPLE_RESX)
2587 : return false;
2588 1160009 : gsi_prev (&gsi);
2589 : }
2590 70150 : else if (!single_succ_p (bb))
2591 : return false;
2592 :
2593 3058363 : for (; !gsi_end_p (gsi); gsi_prev (&gsi))
2594 : {
2595 2981732 : gimple *stmt = gsi_stmt (gsi);
2596 2981732 : if (is_gimple_debug (stmt))
2597 996204 : continue;
2598 1985528 : if (gimple_clobber_p (stmt))
2599 939077 : continue;
2600 1046451 : if (gimple_code (stmt) == GIMPLE_LABEL)
2601 : break;
2602 : return false;
2603 : }
2604 :
2605 : /* See if all predecessors are either throws or clobber only BBs. */
2606 3121647 : FOR_EACH_EDGE (e, ei, bb->preds)
2607 2539822 : if (!(e->flags & EDGE_EH)
2608 2539822 : && !clobber_only_eh_bb_p (e->src, false))
2609 : return false;
2610 :
2611 : return true;
2612 : }
2613 :
2614 : /* Return true if STMT compute a floating point expression that may be affected
2615 : by -ffast-math and similar flags. */
2616 :
2617 : static bool
2618 96388256 : fp_expression_p (gimple *stmt)
2619 : {
2620 96388256 : ssa_op_iter i;
2621 96388256 : tree op;
2622 :
2623 216365377 : FOR_EACH_SSA_TREE_OPERAND (op, stmt, i, SSA_OP_DEF|SSA_OP_USE)
2624 120595055 : if (FLOAT_TYPE_P (TREE_TYPE (op)))
2625 : return true;
2626 : return false;
2627 : }
2628 :
2629 : /* Return true if T references memory location that is local
2630 : for the function (that means, dead after return) or read-only. */
2631 :
2632 : bool
2633 61014225 : refs_local_or_readonly_memory_p (tree t)
2634 : {
2635 : /* Non-escaping memory is fine. */
2636 61014225 : t = get_base_address (t);
2637 61014225 : if ((TREE_CODE (t) == MEM_REF
2638 61014225 : || TREE_CODE (t) == TARGET_MEM_REF))
2639 25789063 : return points_to_local_or_readonly_memory_p (TREE_OPERAND (t, 0));
2640 :
2641 : /* Automatic variables are fine. */
2642 35225162 : if (DECL_P (t)
2643 35225162 : && auto_var_in_fn_p (t, current_function_decl))
2644 : return true;
2645 :
2646 : /* Read-only variables are fine. */
2647 11352587 : if (DECL_P (t) && TREE_READONLY (t))
2648 1672555 : return true;
2649 :
2650 : return false;
2651 : }
2652 :
2653 : /* Return true if T is a pointer pointing to memory location that is local
2654 : for the function (that means, dead after return) or read-only. */
2655 :
2656 : bool
2657 74275840 : points_to_local_or_readonly_memory_p (tree t)
2658 : {
2659 : /* See if memory location is clearly invalid. */
2660 74275840 : if (integer_zerop (t))
2661 2486403 : return flag_delete_null_pointer_checks;
2662 71789437 : if (TREE_CODE (t) == SSA_NAME)
2663 : {
2664 : /* For IPA passes we can consinder accesses to return slot local
2665 : even if it is not local in the sense that memory is dead by
2666 : the end of founction.
2667 : The outer function will see a store in the call assignment
2668 : and thus this will do right thing for all uses of this
2669 : function in the current IPA passes (modref, pure/const discovery
2670 : and inlining heuristics). */
2671 45496197 : if (DECL_RESULT (current_function_decl)
2672 45496197 : && DECL_BY_REFERENCE (DECL_RESULT (current_function_decl))
2673 46629538 : && t == ssa_default_def (cfun, DECL_RESULT (current_function_decl)))
2674 : return true;
2675 45160416 : return !ptr_deref_may_alias_global_p (t, false);
2676 : }
2677 26293240 : if (TREE_CODE (t) == ADDR_EXPR
2678 26293240 : && (TREE_CODE (TREE_OPERAND (t, 0)) != TARGET_MEM_REF
2679 4692 : || TREE_CODE (TREE_OPERAND (TREE_OPERAND (t, 0), 0)) != INTEGER_CST))
2680 13631035 : return refs_local_or_readonly_memory_p (TREE_OPERAND (t, 0));
2681 : return false;
2682 : }
2683 :
2684 : /* Return true if T is a pointer pointing to memory location that is possible
2685 : sra candidate if all functions it is passed to are inlined. */
2686 :
2687 : static bool
2688 39633232 : points_to_possible_sra_candidate_p (tree t)
2689 : {
2690 39633232 : if (TREE_CODE (t) != ADDR_EXPR)
2691 : return false;
2692 :
2693 11837942 : t = get_base_address (TREE_OPERAND (t, 0));
2694 :
2695 : /* Automatic variables are fine. */
2696 11837942 : if (DECL_P (t)
2697 11837942 : && auto_var_in_fn_p (t, current_function_decl))
2698 : return true;
2699 : return false;
2700 : }
2701 :
2702 : /* Return true if BB only calls builtin_unreachable.
2703 : We skip empty basic blocks, debug statements, clobbers and predicts.
2704 : CACHE is used to memoize already analyzed blocks. */
2705 :
2706 : static bool
2707 27935183 : builtin_unreachable_bb_p (basic_block bb, vec<unsigned char> &cache)
2708 : {
2709 27935183 : if (cache[bb->index])
2710 2757374 : return cache[bb->index] - 1;
2711 25177809 : gimple_stmt_iterator si;
2712 25177809 : auto_vec <basic_block, 4> visited_bbs;
2713 25177809 : bool ret = false;
2714 25989892 : while (true)
2715 : {
2716 25989892 : bool empty_bb = true;
2717 25989892 : visited_bbs.safe_push (bb);
2718 25989892 : cache[bb->index] = 3;
2719 25989892 : for (si = gsi_start_nondebug_bb (bb);
2720 27510325 : !gsi_end_p (si) && empty_bb;
2721 1520433 : gsi_next_nondebug (&si))
2722 : {
2723 26134433 : if (gimple_code (gsi_stmt (si)) != GIMPLE_PREDICT
2724 25758727 : && !gimple_clobber_p (gsi_stmt (si))
2725 50781864 : && !gimple_nop_p (gsi_stmt (si)))
2726 : {
2727 : empty_bb = false;
2728 : break;
2729 : }
2730 : }
2731 25989892 : if (!empty_bb)
2732 : break;
2733 : else
2734 1375892 : bb = single_succ_edge (bb)->dest;
2735 1375892 : if (cache[bb->index])
2736 : {
2737 563809 : ret = cache[bb->index] == 3 ? false : cache[bb->index] - 1;
2738 563809 : goto done;
2739 : }
2740 : }
2741 24614000 : if (gimple_call_builtin_p (gsi_stmt (si), BUILT_IN_UNREACHABLE)
2742 24614000 : || gimple_call_builtin_p (gsi_stmt (si), BUILT_IN_UNREACHABLE_TRAP))
2743 : ret = true;
2744 25177809 : done:
2745 101523319 : for (basic_block vbb:visited_bbs)
2746 25989892 : cache[vbb->index] = (unsigned char)ret + 1;
2747 25177809 : return ret;
2748 25177809 : }
2749 :
2750 : static bool
2751 14096919 : guards_builtin_unreachable (basic_block bb, vec<unsigned char> &cache)
2752 : {
2753 14096919 : edge_iterator ei;
2754 14096919 : edge e;
2755 41758834 : FOR_EACH_EDGE (e, ei, bb->succs)
2756 27935183 : if (builtin_unreachable_bb_p (e->dest, cache))
2757 : {
2758 273268 : if (dump_file && (dump_flags & TDF_DETAILS))
2759 1 : fprintf (dump_file,
2760 : "BB %i ends with conditional guarding __builtin_unreachable;"
2761 : " conditinal is unnecessary\n", bb->index);
2762 : return true;
2763 : }
2764 : return false;
2765 : }
2766 :
2767 : #define STMT_NECESSARY GF_PLF_1
2768 :
2769 : /* If STMT is not already marked necessary, mark it, and add it to the
2770 : worklist if ADD_TO_WORKLIST is true. */
2771 :
2772 : static inline void
2773 175861636 : mark_stmt_necessary (gimple *stmt, auto_vec<gimple *> &worklist)
2774 : {
2775 175861636 : gcc_assert (stmt);
2776 :
2777 175861636 : if (gimple_plf (stmt, STMT_NECESSARY))
2778 : return;
2779 :
2780 146048216 : if (dump_file && (dump_flags & TDF_DETAILS))
2781 : {
2782 211 : fprintf (dump_file, "Marking useful stmt: ");
2783 211 : print_gimple_stmt (dump_file, stmt, 0, TDF_SLIM);
2784 211 : fprintf (dump_file, "\n");
2785 : }
2786 :
2787 146048216 : gimple_set_plf (stmt, STMT_NECESSARY, true);
2788 146048216 : worklist.safe_push (stmt);
2789 : }
2790 :
2791 : /* Mark the statement defining operand OP as necessary. */
2792 :
2793 : static inline void
2794 123839461 : mark_operand_necessary (tree op, auto_vec<gimple *> &worklist)
2795 : {
2796 123839461 : gimple *stmt = SSA_NAME_DEF_STMT (op);
2797 123839461 : if (gimple_nop_p (stmt))
2798 : return;
2799 100230723 : mark_stmt_necessary (stmt, worklist);
2800 : }
2801 :
2802 : /* Mark all statements that will remain in the body after optimizing out
2803 : conditionals guarding __builtin_unreachable which we keep to preserve
2804 : value ranges. */
2805 :
2806 : static void
2807 7483419 : find_necessary_statements (struct cgraph_node *node)
2808 : {
2809 7483419 : struct function *my_function = DECL_STRUCT_FUNCTION (node->decl);
2810 7483419 : auto_vec<unsigned char, 10> cache;
2811 7483419 : basic_block bb;
2812 7483419 : auto_vec<gimple *> worklist;
2813 :
2814 7483419 : cache.safe_grow_cleared (last_basic_block_for_fn (cfun));
2815 : /* Mark all obviously necessary statements. */
2816 50084931 : FOR_EACH_BB_FN (bb, my_function)
2817 : {
2818 42601512 : for (gimple_stmt_iterator gsi = gsi_start_phis (bb);
2819 54407319 : !gsi_end_p (gsi); gsi_next (&gsi))
2820 11805807 : gimple_set_plf (gsi_stmt (gsi), STMT_NECESSARY, false);
2821 :
2822 238639514 : for (gimple_stmt_iterator bsi = gsi_start_bb (bb); !gsi_end_p (bsi);
2823 153436490 : gsi_next_nondebug (&bsi))
2824 : {
2825 153436490 : gimple *stmt = gsi_stmt (bsi);
2826 :
2827 153436490 : gimple_set_plf (stmt, STMT_NECESSARY, false);
2828 153436490 : if (gimple_has_side_effects (stmt)
2829 125635530 : || (is_ctrl_stmt (stmt)
2830 22678020 : && (gimple_code (stmt) != GIMPLE_COND
2831 14096919 : || !guards_builtin_unreachable (bb, cache)))
2832 103230778 : || gimple_store_p (stmt)
2833 231263756 : || gimple_code (stmt) == GIMPLE_ASM)
2834 75630913 : mark_stmt_necessary (stmt, worklist);
2835 : }
2836 : }
2837 153531635 : while (worklist.length () > 0)
2838 : {
2839 146048216 : gimple *stmt = worklist.pop ();
2840 :
2841 146048216 : if (dump_file && (dump_flags & TDF_DETAILS))
2842 : {
2843 211 : fprintf (dump_file, "processing: ");
2844 211 : print_gimple_stmt (dump_file, stmt, 0, TDF_SLIM);
2845 211 : fprintf (dump_file, "\n");
2846 : }
2847 146048216 : if (gimple_code (stmt) == GIMPLE_PHI)
2848 18618257 : for (unsigned int k = 0; k < gimple_phi_num_args (stmt); k++)
2849 : {
2850 13174638 : tree arg = PHI_ARG_DEF (stmt, k);
2851 :
2852 13174638 : if (TREE_CODE (arg) == SSA_NAME)
2853 10080405 : mark_operand_necessary (arg, worklist);
2854 : }
2855 : else
2856 : {
2857 140604597 : ssa_op_iter iter;
2858 140604597 : tree use;
2859 :
2860 254363653 : FOR_EACH_SSA_TREE_OPERAND (use, stmt, iter, SSA_OP_USE)
2861 113759056 : mark_operand_necessary (use, worklist);
2862 : }
2863 : }
2864 7483419 : }
2865 :
2866 : /* Analyze function body for NODE.
2867 : EARLY indicates run from early optimization pipeline. */
2868 :
2869 : static void
2870 7483419 : analyze_function_body (struct cgraph_node *node, bool early)
2871 : {
2872 7483419 : sreal time = opt_for_fn (node->decl, param_uninlined_function_time);
2873 : /* Estimate static overhead for function prologue/epilogue and alignment. */
2874 7483419 : int size = opt_for_fn (node->decl, param_uninlined_function_insns);
2875 : /* Benefits are scaled by probability of elimination that is in range
2876 : <0,2>. */
2877 7483419 : basic_block bb;
2878 7483419 : struct function *my_function = DECL_STRUCT_FUNCTION (node->decl);
2879 7483419 : sreal freq;
2880 7483419 : class ipa_fn_summary *info = ipa_fn_summaries->get_create (node);
2881 7483419 : ipa_node_params *params_summary
2882 7483419 : = early ? NULL : ipa_node_params_sum->get (node);
2883 7483419 : ipa_predicate bb_predicate;
2884 7483419 : struct ipa_func_body_info fbi;
2885 7483419 : vec<ipa_predicate> nonconstant_names = vNULL;
2886 7483419 : int nblocks, n;
2887 7483419 : int *order;
2888 7483419 : gimple *fix_builtin_expect_stmt;
2889 :
2890 7483419 : gcc_assert (my_function && my_function->cfg);
2891 7483419 : gcc_assert (cfun == my_function);
2892 :
2893 7483419 : memset(&fbi, 0, sizeof(fbi));
2894 7483419 : vec_free (info->conds);
2895 7483419 : info->conds = NULL;
2896 7483419 : info->size_time_table.release ();
2897 7483419 : info->call_size_time_table.release ();
2898 :
2899 : /* When optimizing and analyzing for IPA inliner, initialize loop optimizer
2900 : so we can produce proper inline hints.
2901 :
2902 : When optimizing and analyzing for early inliner, initialize node params
2903 : so we can produce correct BB predicates. */
2904 :
2905 7483419 : if (opt_for_fn (node->decl, optimize))
2906 : {
2907 6560932 : calculate_dominance_info (CDI_DOMINATORS);
2908 6560932 : calculate_dominance_info (CDI_POST_DOMINATORS);
2909 6560932 : if (!early)
2910 1418182 : loop_optimizer_init (LOOPS_NORMAL | LOOPS_HAVE_RECORDED_EXITS);
2911 : else
2912 : {
2913 5142750 : ipa_check_create_node_params ();
2914 5142750 : ipa_initialize_node_params (node);
2915 : }
2916 :
2917 6560932 : if (ipa_node_params_sum)
2918 : {
2919 6560932 : fbi.node = node;
2920 6560932 : fbi.info = ipa_node_params_sum->get (node);
2921 6560932 : fbi.bb_infos = vNULL;
2922 6560932 : fbi.bb_infos.safe_grow_cleared (last_basic_block_for_fn (cfun), true);
2923 6560932 : fbi.param_count = count_formal_params (node->decl);
2924 6560932 : fbi.aa_walk_budget = opt_for_fn (node->decl, param_ipa_max_aa_steps);
2925 :
2926 6560932 : nonconstant_names.safe_grow_cleared
2927 6560932 : (SSANAMES (my_function)->length (), true);
2928 : }
2929 : }
2930 :
2931 7483419 : if (dump_file)
2932 249 : fprintf (dump_file, "\nAnalyzing function body size: %s\n",
2933 : node->dump_name ());
2934 :
2935 : /* When we run into maximal number of entries, we assign everything to the
2936 : constant truth case. Be sure to have it in list. */
2937 7483419 : bb_predicate = true;
2938 7483419 : info->account_size_time (0, 0, bb_predicate, bb_predicate);
2939 :
2940 7483419 : bb_predicate = ipa_predicate::not_inlined ();
2941 7483419 : info->account_size_time (opt_for_fn (node->decl,
2942 : param_uninlined_function_insns)
2943 : * ipa_fn_summary::size_scale,
2944 7483419 : opt_for_fn (node->decl,
2945 : param_uninlined_function_time),
2946 : bb_predicate,
2947 : bb_predicate);
2948 :
2949 : /* Only look for target information for inlinable functions. */
2950 7483419 : bool scan_for_target_info =
2951 7483419 : info->inlinable
2952 13455475 : && targetm.target_option.need_ipa_fn_target_info (node->decl,
2953 5972056 : info->target_info);
2954 :
2955 7483419 : if (fbi.info)
2956 6560932 : compute_bb_predicates (&fbi, node, info, params_summary);
2957 7483419 : find_necessary_statements (node);
2958 7483419 : const profile_count entry_count = ENTRY_BLOCK_PTR_FOR_FN (cfun)->count;
2959 7483419 : order = XNEWVEC (int, n_basic_blocks_for_fn (cfun));
2960 7483419 : nblocks = pre_and_rev_post_order_compute (NULL, order, false);
2961 57568350 : for (n = 0; n < nblocks; n++)
2962 : {
2963 42601512 : bb = BASIC_BLOCK_FOR_FN (cfun, order[n]);
2964 42601512 : freq = bb->count.to_sreal_scale (entry_count);
2965 42601512 : if (clobber_only_eh_bb_p (bb))
2966 : {
2967 580323 : if (dump_file && (dump_flags & TDF_DETAILS))
2968 0 : fprintf (dump_file, "\n Ignoring BB %i;"
2969 : " it will be optimized away by cleanup_clobbers\n",
2970 : bb->index);
2971 580323 : continue;
2972 : }
2973 :
2974 : /* TODO: Obviously predicates can be propagated down across CFG. */
2975 42021189 : if (fbi.info)
2976 : {
2977 35509929 : if (bb->aux)
2978 35509915 : bb_predicate = *(ipa_predicate *)bb->aux;
2979 : else
2980 14 : bb_predicate = false;
2981 : }
2982 : else
2983 6511260 : bb_predicate = true;
2984 :
2985 42021189 : if (dump_file && (dump_flags & TDF_DETAILS))
2986 : {
2987 69 : fprintf (dump_file, "\n BB %i predicate:", bb->index);
2988 69 : bb_predicate.dump (dump_file, info->conds);
2989 : }
2990 :
2991 42021189 : if (fbi.info && nonconstant_names.exists ())
2992 : {
2993 35509929 : ipa_predicate phi_predicate;
2994 35509929 : bool first_phi = true;
2995 :
2996 36298367 : for (gphi_iterator bsi = gsi_start_phis (bb); !gsi_end_p (bsi);
2997 788438 : gsi_next (&bsi))
2998 : {
2999 6580576 : if (first_phi
3000 6580576 : && !phi_result_unknown_predicate (&fbi, info,
3001 : params_summary,
3002 : bb,
3003 : &phi_predicate,
3004 : nonconstant_names))
3005 : break;
3006 788438 : first_phi = false;
3007 788438 : if (dump_file && (dump_flags & TDF_DETAILS))
3008 : {
3009 3 : fprintf (dump_file, " ");
3010 3 : print_gimple_stmt (dump_file, gsi_stmt (bsi), 0);
3011 : }
3012 788438 : predicate_for_phi_result (info, bsi.phi (), &phi_predicate,
3013 : nonconstant_names);
3014 : }
3015 : }
3016 :
3017 42021189 : fix_builtin_expect_stmt = find_foldable_builtin_expect (bb);
3018 :
3019 42021189 : for (gimple_stmt_iterator bsi = gsi_start_nondebug_bb (bb);
3020 186997531 : !gsi_end_p (bsi); gsi_next_nondebug (&bsi))
3021 : {
3022 144976342 : gimple *stmt = gsi_stmt (bsi);
3023 144976342 : if (!gimple_plf (stmt, STMT_NECESSARY))
3024 : {
3025 5737908 : if (dump_file && (dump_flags & TDF_DETAILS))
3026 : {
3027 19 : fprintf (dump_file, " skipping unnecessary stmt ");
3028 19 : print_gimple_stmt (dump_file, stmt, 0);
3029 : }
3030 : /* TODO: const calls used only to produce values for
3031 : builtion_unreachable guards should not be accounted. However
3032 : we still want to inline them and this does does not work well
3033 : with the cost model. For now account them as usual. */
3034 5737908 : if (!is_gimple_call (stmt)
3035 5737908 : || gimple_call_internal_p (stmt))
3036 5448987 : continue;
3037 : }
3038 139527355 : int this_size = estimate_num_insns (stmt, &eni_size_weights);
3039 139527355 : int this_time = estimate_num_insns (stmt, &eni_time_weights);
3040 139527355 : int prob;
3041 139527355 : ipa_predicate will_be_nonconstant;
3042 :
3043 : /* This relation stmt should be folded after we remove
3044 : __builtin_expect call. Adjust the cost here. */
3045 139527355 : if (stmt == fix_builtin_expect_stmt)
3046 : {
3047 187437 : this_size--;
3048 187437 : this_time--;
3049 : }
3050 :
3051 139527355 : if (dump_file && (dump_flags & TDF_DETAILS))
3052 : {
3053 205 : fprintf (dump_file, " ");
3054 205 : print_gimple_stmt (dump_file, stmt, 0);
3055 205 : fprintf (dump_file, "\t\tfreq:%3.2f size:%3i time:%3i\n",
3056 : freq.to_double (), this_size,
3057 : this_time);
3058 : }
3059 :
3060 139527355 : if (is_gimple_call (stmt)
3061 139527355 : && !gimple_call_internal_p (stmt))
3062 : {
3063 23507549 : struct cgraph_edge *edge = node->get_edge (stmt);
3064 23507549 : ipa_call_summary *es = ipa_call_summaries->get_create (edge);
3065 :
3066 : /* Special case: results of BUILT_IN_CONSTANT_P will be always
3067 : resolved as constant. We however don't want to optimize
3068 : out the cgraph edges. */
3069 23507549 : if (nonconstant_names.exists ()
3070 20578026 : && gimple_call_builtin_p (stmt, BUILT_IN_CONSTANT_P)
3071 48232 : && gimple_call_lhs (stmt)
3072 23555781 : && TREE_CODE (gimple_call_lhs (stmt)) == SSA_NAME)
3073 : {
3074 48232 : ipa_predicate false_p = false;
3075 48232 : nonconstant_names[SSA_NAME_VERSION (gimple_call_lhs (stmt))]
3076 48232 : = false_p;
3077 : }
3078 23507549 : if (ipa_node_params_sum)
3079 : {
3080 20590843 : int count = gimple_call_num_args (stmt);
3081 20590843 : int i;
3082 :
3083 20590843 : if (count)
3084 17700218 : es->param.safe_grow_cleared (count, true);
3085 60224075 : for (i = 0; i < count; i++)
3086 : {
3087 39633232 : int prob = param_change_prob (&fbi, stmt, i);
3088 39633232 : gcc_assert (prob >= 0 && prob <= REG_BR_PROB_BASE);
3089 39633232 : es->param[i].change_prob = prob;
3090 39633232 : es->param[i].points_to_local_or_readonly_memory
3091 39633232 : = points_to_local_or_readonly_memory_p
3092 39633232 : (gimple_call_arg (stmt, i));
3093 39633232 : es->param[i].points_to_possible_sra_candidate
3094 39633232 : = points_to_possible_sra_candidate_p
3095 39633232 : (gimple_call_arg (stmt, i));
3096 : }
3097 : }
3098 : /* We cannot setup VLA parameters during inlining. */
3099 69073331 : for (unsigned int i = 0; i < gimple_call_num_args (stmt); ++i)
3100 45566178 : if (TREE_CODE (gimple_call_arg (stmt, i)) == WITH_SIZE_EXPR)
3101 : {
3102 396 : edge->inline_failed = CIF_FUNCTION_NOT_INLINABLE;
3103 396 : break;
3104 : }
3105 23507549 : es->call_stmt_size = this_size;
3106 23507549 : es->call_stmt_time = this_time;
3107 23507549 : es->loop_depth = bb_loop_depth (bb);
3108 23507549 : edge_set_predicate (edge, &bb_predicate);
3109 23507549 : if (edge->speculative)
3110 : {
3111 0 : cgraph_edge *indirect
3112 0 : = edge->speculative_call_indirect_edge ();
3113 0 : ipa_call_summary *es2
3114 0 : = ipa_call_summaries->get_create (indirect);
3115 0 : ipa_call_summaries->duplicate (edge, indirect,
3116 : es, es2);
3117 :
3118 : /* Edge is the first direct call.
3119 : create and duplicate call summaries for multiple
3120 : speculative call targets. */
3121 0 : for (cgraph_edge *direct
3122 0 : = edge->next_speculative_call_target ();
3123 0 : direct;
3124 0 : direct = direct->next_speculative_call_target ())
3125 : {
3126 0 : ipa_call_summary *es3
3127 0 : = ipa_call_summaries->get_create (direct);
3128 0 : ipa_call_summaries->duplicate (edge, direct,
3129 : es, es3);
3130 : }
3131 : }
3132 :
3133 : /* If dealing with a carrying edge, copy its summary over to its
3134 : attached edges as well. */
3135 23507549 : if (edge->has_callback)
3136 : {
3137 15126 : cgraph_edge *cbe;
3138 30253 : for (cbe = edge->first_callback_edge (); cbe;
3139 15127 : cbe = cbe->next_callback_edge ())
3140 : {
3141 15127 : ipa_call_summary *es2 = ipa_call_summaries->get_create (cbe);
3142 15127 : ipa_call_summaries->duplicate (edge, cbe, es, es2);
3143 : /* Unlike speculative edges, callback edges have no real
3144 : size or time; the call doesn't exist. Reflect that in
3145 : their summaries. */
3146 15127 : es2->call_stmt_size = 0;
3147 15127 : es2->call_stmt_time = 0;
3148 : }
3149 : }
3150 : }
3151 :
3152 : /* TODO: When conditional jump or switch is known to be constant, but
3153 : we did not translate it into the predicates, we really can account
3154 : just maximum of the possible paths. */
3155 139527355 : if (fbi.info)
3156 117083517 : will_be_nonconstant
3157 117083517 : = will_be_nonconstant_predicate (&fbi, info, params_summary,
3158 : stmt, nonconstant_names);
3159 : else
3160 22443838 : will_be_nonconstant = true;
3161 139527355 : if (this_time || this_size)
3162 : {
3163 114053989 : sreal final_time = (sreal)this_time * freq;
3164 114053989 : prob = eliminated_by_inlining_prob (&fbi, stmt);
3165 114053989 : if (prob == 1 && dump_file && (dump_flags & TDF_DETAILS))
3166 11 : fprintf (dump_file,
3167 : "\t\t50%% will be eliminated by inlining\n");
3168 114053989 : if (prob == 2 && dump_file && (dump_flags & TDF_DETAILS))
3169 20 : fprintf (dump_file, "\t\tWill be eliminated by inlining\n");
3170 :
3171 114053989 : ipa_predicate p = bb_predicate & will_be_nonconstant;
3172 114053989 : int parm = load_or_store_of_ptr_parameter (&fbi, stmt);
3173 114053989 : ipa_predicate sra_predicate = true;
3174 114053989 : if (parm != -1)
3175 14581006 : sra_predicate &= add_condition (info, params_summary, parm,
3176 : ptr_type_node, NULL,
3177 7290503 : ipa_predicate::not_sra_candidate, NULL, 0);
3178 :
3179 : /* We can ignore statement when we proved it is never going
3180 : to happen, but we cannot do that for call statements
3181 : because edges are accounted specially. */
3182 :
3183 228107978 : if (*(is_gimple_call (stmt) ? &bb_predicate : &p) != false)
3184 : {
3185 113292488 : time += final_time;
3186 113292488 : size += this_size;
3187 : }
3188 :
3189 : /* We account everything but the calls. Calls have their own
3190 : size/time info attached to cgraph edges. This is necessary
3191 : in order to make the cost disappear after inlining. */
3192 114053989 : if (!is_gimple_call (stmt))
3193 : {
3194 91279903 : if (prob)
3195 : {
3196 15364809 : ipa_predicate ip
3197 15364809 : = bb_predicate & ipa_predicate::not_inlined () & sra_predicate;
3198 30729618 : info->account_size_time (this_size * prob,
3199 15364809 : (final_time * prob) / 2, ip,
3200 : p);
3201 : }
3202 15364809 : if (prob != 2)
3203 168283664 : info->account_size_time (this_size * (2 - prob),
3204 84141832 : (final_time * (2 - prob) / 2),
3205 168283664 : bb_predicate & sra_predicate,
3206 : p);
3207 : }
3208 :
3209 114053989 : if (!info->fp_expressions && fp_expression_p (stmt))
3210 : {
3211 617934 : info->fp_expressions = true;
3212 617934 : if (dump_file)
3213 9 : fprintf (dump_file, " fp_expression set\n");
3214 : }
3215 : }
3216 :
3217 : /* For target specific information, we want to scan all statements
3218 : rather than those statements with non-zero weights, to avoid
3219 : missing to scan something interesting for target information,
3220 : such as: internal function calls. */
3221 139527355 : if (scan_for_target_info)
3222 0 : scan_for_target_info =
3223 0 : targetm.target_option.update_ipa_fn_target_info
3224 0 : (info->target_info, stmt);
3225 :
3226 : /* Account cost of address calculations in the statements. */
3227 539681951 : for (unsigned int i = 0; i < gimple_num_ops (stmt); i++)
3228 : {
3229 400154596 : for (tree op = gimple_op (stmt, i);
3230 770286173 : op && handled_component_p (op);
3231 48354285 : op = TREE_OPERAND (op, 0))
3232 48354285 : if ((TREE_CODE (op) == ARRAY_REF
3233 48354285 : || TREE_CODE (op) == ARRAY_RANGE_REF)
3234 48354285 : && TREE_CODE (TREE_OPERAND (op, 1)) == SSA_NAME)
3235 : {
3236 2417114 : ipa_predicate p = bb_predicate;
3237 2417114 : if (fbi.info)
3238 1891006 : p = p & will_be_nonconstant_expr_predicate
3239 1891006 : (&fbi, info, params_summary,
3240 1891006 : TREE_OPERAND (op, 1),
3241 1891006 : nonconstant_names);
3242 2417114 : if (p != false)
3243 : {
3244 2411692 : time += freq;
3245 2411692 : size += 1;
3246 2411692 : if (dump_file)
3247 24 : fprintf (dump_file,
3248 : "\t\tAccounting address calculation.\n");
3249 2411692 : info->account_size_time (ipa_fn_summary::size_scale,
3250 : freq,
3251 : bb_predicate,
3252 : p);
3253 : }
3254 : }
3255 : }
3256 :
3257 : }
3258 : }
3259 7483419 : free (order);
3260 :
3261 7483419 : if (nonconstant_names.exists () && !early)
3262 : {
3263 1418182 : ipa_fn_summary *s = ipa_fn_summaries->get (node);
3264 1418182 : unsigned max_loop_predicates = opt_for_fn (node->decl,
3265 : param_ipa_max_loop_predicates);
3266 :
3267 1418182 : if (dump_file && (dump_flags & TDF_DETAILS))
3268 16 : flow_loops_dump (dump_file, NULL, 0);
3269 1418182 : scev_initialize ();
3270 4886735 : for (auto loop : loops_list (cfun, 0))
3271 : {
3272 632189 : ipa_predicate loop_iterations = true;
3273 632189 : sreal header_freq;
3274 632189 : edge ex;
3275 632189 : unsigned int j;
3276 632189 : class tree_niter_desc niter_desc;
3277 632189 : if (!loop->header->aux)
3278 0 : continue;
3279 :
3280 632189 : profile_count hdr_count = loop->header->count;
3281 632189 : sreal hdr_freq = hdr_count.to_sreal_scale (entry_count);
3282 :
3283 632189 : bb_predicate = *(ipa_predicate *)loop->header->aux;
3284 632189 : auto_vec<edge> exits = get_loop_exit_edges (loop);
3285 2293980 : FOR_EACH_VEC_ELT (exits, j, ex)
3286 1029602 : if (number_of_iterations_exit (loop, ex, &niter_desc, false)
3287 1029602 : && !is_gimple_min_invariant (niter_desc.niter))
3288 : {
3289 155501 : ipa_predicate will_be_nonconstant
3290 155501 : = will_be_nonconstant_expr_predicate (&fbi, info,
3291 : params_summary,
3292 : niter_desc.niter,
3293 : nonconstant_names);
3294 155501 : if (will_be_nonconstant != true)
3295 62506 : will_be_nonconstant = bb_predicate & will_be_nonconstant;
3296 155501 : if (will_be_nonconstant != true
3297 218007 : && will_be_nonconstant != false)
3298 61953 : loop_iterations &= will_be_nonconstant;
3299 : }
3300 632189 : add_freqcounting_predicate (&s->loop_iterations, loop_iterations,
3301 : hdr_freq, max_loop_predicates);
3302 632191 : }
3303 :
3304 : /* To avoid quadratic behavior we analyze stride predicates only
3305 : with respect to the containing loop. Thus we simply iterate
3306 : over all defs in the outermost loop body. */
3307 1418182 : for (class loop *loop = loops_for_fn (cfun)->tree_root->inner;
3308 1917647 : loop != NULL; loop = loop->next)
3309 : {
3310 499465 : ipa_predicate loop_stride = true;
3311 499465 : basic_block *body = get_loop_body (loop);
3312 499465 : profile_count hdr_count = loop->header->count;
3313 499465 : sreal hdr_freq = hdr_count.to_sreal_scale (entry_count);
3314 3432315 : for (unsigned i = 0; i < loop->num_nodes; i++)
3315 : {
3316 2433385 : gimple_stmt_iterator gsi;
3317 2433385 : if (!body[i]->aux)
3318 7 : continue;
3319 :
3320 2433378 : bb_predicate = *(ipa_predicate *)body[i]->aux;
3321 16465398 : for (gsi = gsi_start_bb (body[i]); !gsi_end_p (gsi);
3322 11598642 : gsi_next (&gsi))
3323 : {
3324 11598642 : gimple *stmt = gsi_stmt (gsi);
3325 :
3326 11598642 : if (!is_gimple_assign (stmt))
3327 11448215 : continue;
3328 :
3329 5488892 : tree def = gimple_assign_lhs (stmt);
3330 5488892 : if (TREE_CODE (def) != SSA_NAME)
3331 1047446 : continue;
3332 :
3333 4441446 : affine_iv iv;
3334 8882892 : if (!simple_iv (loop_containing_stmt (stmt),
3335 : loop_containing_stmt (stmt),
3336 : def, &iv, true)
3337 4441446 : || is_gimple_min_invariant (iv.step))
3338 4291019 : continue;
3339 :
3340 150427 : ipa_predicate will_be_nonconstant
3341 150427 : = will_be_nonconstant_expr_predicate (&fbi, info,
3342 : params_summary,
3343 : iv.step,
3344 : nonconstant_names);
3345 150427 : if (will_be_nonconstant != true)
3346 55290 : will_be_nonconstant = bb_predicate & will_be_nonconstant;
3347 150427 : if (will_be_nonconstant != true
3348 205717 : && will_be_nonconstant != false)
3349 45695 : loop_stride = loop_stride & will_be_nonconstant;
3350 : }
3351 : }
3352 499465 : add_freqcounting_predicate (&s->loop_strides, loop_stride,
3353 : hdr_freq, max_loop_predicates);
3354 499465 : free (body);
3355 : }
3356 1418182 : scev_finalize ();
3357 : }
3358 65051769 : FOR_ALL_BB_FN (bb, my_function)
3359 : {
3360 57568350 : edge e;
3361 57568350 : edge_iterator ei;
3362 :
3363 57568350 : if (bb->aux)
3364 42586407 : edge_predicate_pool.remove ((ipa_predicate *)bb->aux);
3365 57568350 : bb->aux = NULL;
3366 122314092 : FOR_EACH_EDGE (e, ei, bb->succs)
3367 : {
3368 64745742 : if (e->aux)
3369 2606346 : edge_predicate_pool.remove ((ipa_predicate *)e->aux);
3370 64745742 : e->aux = NULL;
3371 : }
3372 : }
3373 7483419 : ipa_fn_summary *s = ipa_fn_summaries->get (node);
3374 7483419 : ipa_size_summary *ss = ipa_size_summaries->get (node);
3375 7483419 : s->time = time;
3376 7483419 : ss->self_size = size;
3377 7483419 : nonconstant_names.release ();
3378 7483419 : ipa_release_body_info (&fbi);
3379 7483419 : if (opt_for_fn (node->decl, optimize))
3380 : {
3381 6560932 : if (!early)
3382 1418182 : loop_optimizer_finalize ();
3383 5142750 : else if (!ipa_edge_args_sum)
3384 5142736 : ipa_free_all_node_params ();
3385 6560932 : free_dominance_info (CDI_DOMINATORS);
3386 6560932 : free_dominance_info (CDI_POST_DOMINATORS);
3387 : }
3388 7483419 : if (dump_file)
3389 : {
3390 249 : fprintf (dump_file, "\n");
3391 249 : ipa_dump_fn_summary (dump_file, node);
3392 : }
3393 7483419 : }
3394 :
3395 :
3396 : /* Compute function summary.
3397 : EARLY is true when we compute parameters during early opts. */
3398 :
3399 : void
3400 7484735 : compute_fn_summary (struct cgraph_node *node, bool early)
3401 : {
3402 7484735 : HOST_WIDE_INT self_stack_size;
3403 7484735 : struct cgraph_edge *e;
3404 :
3405 7484735 : gcc_assert (!node->inlined_to);
3406 :
3407 7484735 : if (!ipa_fn_summaries)
3408 218693 : ipa_fn_summary_alloc ();
3409 :
3410 : /* Create a new ipa_fn_summary. */
3411 7484735 : ((ipa_fn_summary_t *)ipa_fn_summaries)->remove_callees (node);
3412 7484735 : ipa_fn_summaries->remove (node);
3413 7484735 : class ipa_fn_summary *info = ipa_fn_summaries->get_create (node);
3414 7484735 : class ipa_size_summary *size_info = ipa_size_summaries->get_create (node);
3415 :
3416 : /* Estimate the stack size for the function if we're optimizing. */
3417 13123164 : self_stack_size = optimize && !node->thunk
3418 14045667 : ? estimated_stack_frame_size (node) : 0;
3419 7484735 : size_info->estimated_self_stack_size = self_stack_size;
3420 7484735 : info->estimated_stack_size = self_stack_size;
3421 :
3422 7484735 : if (node->thunk)
3423 : {
3424 1316 : ipa_call_summary *es = ipa_call_summaries->get_create (node->callees);
3425 1316 : ipa_predicate t = true;
3426 :
3427 1316 : node->can_change_signature = false;
3428 1316 : es->call_stmt_size = eni_size_weights.call_cost;
3429 1316 : es->call_stmt_time = eni_time_weights.call_cost;
3430 3948 : info->account_size_time (ipa_fn_summary::size_scale
3431 1316 : * opt_for_fn (node->decl,
3432 : param_uninlined_function_thunk_insns),
3433 1316 : opt_for_fn (node->decl,
3434 : param_uninlined_function_thunk_time), t, t);
3435 1316 : t = ipa_predicate::not_inlined ();
3436 1316 : info->account_size_time (2 * ipa_fn_summary::size_scale, 0, t, t);
3437 1316 : ipa_update_overall_fn_summary (node);
3438 1316 : size_info->self_size = size_info->size;
3439 1316 : if (stdarg_p (TREE_TYPE (node->decl)))
3440 : {
3441 9 : info->inlinable = false;
3442 9 : node->callees->inline_failed = CIF_VARIADIC_THUNK;
3443 : }
3444 : else
3445 1307 : info->inlinable = true;
3446 : }
3447 : else
3448 : {
3449 : /* Even is_gimple_min_invariant rely on current_function_decl. */
3450 7483419 : push_cfun (DECL_STRUCT_FUNCTION (node->decl));
3451 :
3452 : /* During IPA profile merging we may be called w/o virtual SSA form
3453 : built. */
3454 7483419 : update_ssa (TODO_update_ssa_only_virtuals);
3455 :
3456 : /* Can this function be inlined at all? */
3457 7483419 : if (!opt_for_fn (node->decl, optimize)
3458 8405906 : && !lookup_attribute ("always_inline",
3459 922487 : DECL_ATTRIBUTES (node->decl)))
3460 851397 : info->inlinable = false;
3461 : else
3462 6632022 : info->inlinable = tree_inlinable_function_p (node->decl);
3463 :
3464 7483419 : bool no_signature = false;
3465 :
3466 : /* Don't allow signature changes for functions which have
3467 : [[gnu::musttail]] or [[clang::musttail]] calls. Sometimes
3468 : (more often on targets which pass everything on the stack)
3469 : signature changes can result in tail calls being impossible
3470 : even when without the signature changes they would be ok.
3471 : See PR121023. */
3472 7483419 : if (cfun->has_musttail)
3473 : {
3474 1441 : if (dump_file)
3475 0 : fprintf (dump_file, "No signature change:"
3476 : " function has calls with musttail attribute.\n");
3477 : no_signature = true;
3478 : }
3479 :
3480 : /* Type attributes can use parameter indices to describe them.
3481 : Special case fn spec since we can safely preserve them in
3482 : modref summaries. */
3483 7483419 : for (tree list = TYPE_ATTRIBUTES (TREE_TYPE (node->decl));
3484 7883422 : list && !no_signature; list = TREE_CHAIN (list))
3485 400003 : if (!ipa_param_adjustments::type_attribute_allowed_p
3486 400003 : (get_attribute_name (list)))
3487 : {
3488 159304 : if (dump_file)
3489 : {
3490 0 : fprintf (dump_file, "No signature change:"
3491 : " function type has unhandled attribute %s.\n",
3492 0 : IDENTIFIER_POINTER (get_attribute_name (list)));
3493 : }
3494 : no_signature = true;
3495 : }
3496 7483419 : for (tree parm = DECL_ARGUMENTS (node->decl);
3497 22321908 : parm && !no_signature; parm = DECL_CHAIN (parm))
3498 14838489 : if (variably_modified_type_p (TREE_TYPE (parm), node->decl))
3499 : {
3500 15944 : if (dump_file)
3501 : {
3502 0 : fprintf (dump_file, "No signature change:"
3503 : " has parameter with variably modified type.\n");
3504 : }
3505 : no_signature = true;
3506 : }
3507 :
3508 : /* Likewise for #pragma omp declare simd functions or functions
3509 : with simd attribute. */
3510 7483419 : if (no_signature
3511 14790149 : || lookup_attribute ("omp declare simd",
3512 7306730 : DECL_ATTRIBUTES (node->decl)))
3513 178383 : node->can_change_signature = false;
3514 : else
3515 : {
3516 : /* Otherwise, inlinable functions always can change signature. */
3517 7305036 : if (info->inlinable)
3518 5892181 : node->can_change_signature = true;
3519 : else
3520 : {
3521 : /* Functions calling builtin_apply cannot change signature. */
3522 5457303 : for (e = node->callees; e; e = e->next_callee)
3523 : {
3524 4077218 : tree cdecl = e->callee->decl;
3525 4077218 : if (fndecl_built_in_p (cdecl, BUILT_IN_APPLY_ARGS,
3526 : BUILT_IN_VA_START))
3527 : break;
3528 : }
3529 1412855 : node->can_change_signature = !e;
3530 : }
3531 : }
3532 7483419 : analyze_function_body (node, early);
3533 7483419 : pop_cfun ();
3534 : }
3535 :
3536 : /* Inlining characteristics are maintained by the cgraph_mark_inline. */
3537 7484735 : size_info->size = size_info->self_size;
3538 7484735 : info->estimated_stack_size = size_info->estimated_self_stack_size;
3539 :
3540 : /* Code above should compute exactly the same result as
3541 : ipa_update_overall_fn_summary except for case when speculative
3542 : edges are present since these are accounted to size but not
3543 : self_size. Do not compare time since different order the roundoff
3544 : errors result in slight changes. */
3545 7484735 : ipa_update_overall_fn_summary (node);
3546 7484735 : if (flag_checking)
3547 : {
3548 7967285 : for (e = node->indirect_calls; e; e = e->next_callee)
3549 482632 : if (e->speculative)
3550 : break;
3551 7484653 : gcc_assert (e || size_info->size == size_info->self_size);
3552 : }
3553 7484735 : }
3554 :
3555 :
3556 : /* Compute parameters of functions used by inliner using
3557 : current_function_decl. */
3558 :
3559 : static unsigned int
3560 6002610 : compute_fn_summary_for_current (void)
3561 : {
3562 6002610 : compute_fn_summary (cgraph_node::get (current_function_decl), true);
3563 6002610 : return 0;
3564 : }
3565 :
3566 : /* Estimate benefit devirtualizing indirect edge IE and return true if it can
3567 : be devirtualized and inlined, provided m_known_vals, m_known_contexts and
3568 : m_known_aggs in AVALS. Return false straight away if AVALS is NULL. */
3569 :
3570 : static bool
3571 4892000 : estimate_edge_devirt_benefit (struct cgraph_edge *ie,
3572 : int *size, int *time,
3573 : ipa_call_arg_values *avals)
3574 : {
3575 4892000 : tree target;
3576 4892000 : struct cgraph_node *callee;
3577 4892000 : class ipa_fn_summary *isummary;
3578 4892000 : enum availability avail;
3579 4892000 : bool speculative;
3580 :
3581 4892000 : if (!avals
3582 4892000 : || (!avals->m_known_vals.length() && !avals->m_known_contexts.length ()))
3583 : return false;
3584 1479281 : if (!opt_for_fn (ie->caller->decl, flag_indirect_inlining))
3585 : return false;
3586 :
3587 1472818 : target = ipa_get_indirect_edge_target
3588 1472818 : (ie->callee ? ie->speculative_call_indirect_edge () : ie,
3589 : avals, &speculative);
3590 1472818 : if (!target || speculative)
3591 : return false;
3592 :
3593 : /* If this is speculative call, turn its cost into 0; we will account
3594 : the call when processing the indirect call. */
3595 219121 : if (ie->callee)
3596 : {
3597 4402 : gcc_checking_assert (ie->speculative && *size > 0);
3598 4402 : *size = 0;
3599 4402 : *time = 0;
3600 : }
3601 : else
3602 : {
3603 : /* Account for difference in cost between indirect and direct calls. */
3604 214719 : *size -= (eni_size_weights.indirect_call_cost - eni_size_weights.call_cost);
3605 214719 : *time -= (eni_time_weights.indirect_call_cost - eni_time_weights.call_cost);
3606 : }
3607 219121 : gcc_checking_assert (*time >= 0);
3608 219121 : gcc_checking_assert (*size >= 0);
3609 :
3610 219121 : callee = cgraph_node::get (target);
3611 219121 : if (!callee || !callee->definition)
3612 : return false;
3613 198486 : callee = callee->function_symbol (&avail);
3614 198486 : if (avail < AVAIL_AVAILABLE)
3615 : return false;
3616 198460 : isummary = ipa_fn_summaries->get (callee);
3617 198460 : if (isummary == NULL)
3618 : return false;
3619 :
3620 198446 : return isummary->inlinable;
3621 : }
3622 :
3623 : /* Increase SIZE, MIN_SIZE (if non-NULL) and TIME for size and time needed to
3624 : handle edge E with probability PROB. Set HINTS accordingly if edge may be
3625 : devirtualized. AVALS, if non-NULL, describes the context of the call site
3626 : as far as values of parameters are concerned. */
3627 :
3628 : static inline void
3629 219482537 : estimate_edge_size_and_time (struct cgraph_edge *e, int *size, int *min_size,
3630 : sreal *time, ipa_call_arg_values *avals,
3631 : ipa_hints *hints)
3632 : {
3633 219482537 : class ipa_call_summary *es = ipa_call_summaries->get (e);
3634 219482537 : int call_size = es->call_stmt_size;
3635 219482537 : int call_time = es->call_stmt_time;
3636 219482537 : int cur_size;
3637 :
3638 215200127 : if ((!e->callee || e->speculative)
3639 220092127 : && estimate_edge_devirt_benefit (e, &call_size, &call_time, avals))
3640 : {
3641 198200 : if (hints && e->maybe_hot_p ())
3642 188982 : *hints |= INLINE_HINT_indirect_call;
3643 : }
3644 219482537 : cur_size = call_size * ipa_fn_summary::size_scale;
3645 219482537 : *size += cur_size;
3646 219482537 : if (min_size)
3647 29197213 : *min_size += cur_size;
3648 219482537 : if (time)
3649 209497479 : *time += ((sreal)call_time) * e->sreal_frequency ();
3650 219482537 : }
3651 :
3652 :
3653 : /* Increase SIZE, MIN_SIZE and TIME for size and time needed to handle all
3654 : calls in NODE. POSSIBLE_TRUTHS and AVALS describe the context of the call
3655 : site.
3656 :
3657 : Helper for estimate_calls_size_and_time which does the same but
3658 : (in most cases) faster. */
3659 :
3660 : static void
3661 79712958 : estimate_calls_size_and_time_1 (struct cgraph_node *node, int *size,
3662 : int *min_size, sreal *time,
3663 : ipa_hints *hints,
3664 : clause_t possible_truths,
3665 : ipa_call_arg_values *avals)
3666 : {
3667 79712958 : struct cgraph_edge *e;
3668 378377107 : for (e = node->callees; e; e = e->next_callee)
3669 : {
3670 298664149 : if (!e->inline_failed)
3671 : {
3672 58820099 : gcc_checking_assert (!ipa_call_summaries->get (e));
3673 58820099 : estimate_calls_size_and_time_1 (e->callee, size, min_size, time,
3674 : hints, possible_truths, avals);
3675 :
3676 58820099 : continue;
3677 : }
3678 239844050 : class ipa_call_summary *es = ipa_call_summaries->get (e);
3679 :
3680 : /* Do not care about zero sized builtins. */
3681 239844050 : if (!es->call_stmt_size)
3682 : {
3683 33501532 : gcc_checking_assert (!es->call_stmt_time);
3684 33501532 : continue;
3685 : }
3686 206342518 : if (!es->predicate
3687 206342518 : || es->predicate->evaluate (possible_truths))
3688 : {
3689 : /* Predicates of calls shall not use NOT_CHANGED codes,
3690 : so we do not need to compute probabilities. */
3691 203880387 : estimate_edge_size_and_time (e, size,
3692 203880387 : es->predicate ? NULL : min_size,
3693 : time, avals, hints);
3694 : }
3695 : }
3696 83710561 : for (e = node->indirect_calls; e; e = e->next_callee)
3697 : {
3698 3997603 : class ipa_call_summary *es = ipa_call_summaries->get (e);
3699 3997603 : if (!es->predicate
3700 3997603 : || es->predicate->evaluate (possible_truths))
3701 3965585 : estimate_edge_size_and_time (e, size,
3702 3965585 : es->predicate ? NULL : min_size,
3703 : time, avals, hints);
3704 : }
3705 79712958 : }
3706 :
3707 : /* Populate sum->call_size_time_table for edges from NODE. */
3708 :
3709 : static void
3710 3327922 : summarize_calls_size_and_time (struct cgraph_node *node,
3711 : ipa_fn_summary *sum)
3712 : {
3713 3327922 : struct cgraph_edge *e;
3714 15264722 : for (e = node->callees; e; e = e->next_callee)
3715 : {
3716 11936800 : if (!e->inline_failed)
3717 : {
3718 1524334 : gcc_checking_assert (!ipa_call_summaries->get (e));
3719 1524334 : summarize_calls_size_and_time (e->callee, sum);
3720 1524334 : continue;
3721 : }
3722 10412466 : int size = 0;
3723 10412466 : sreal time = 0;
3724 :
3725 10412466 : estimate_edge_size_and_time (e, &size, NULL, &time, NULL, NULL);
3726 :
3727 10412466 : ipa_predicate pred = true;
3728 10412466 : class ipa_call_summary *es = ipa_call_summaries->get (e);
3729 :
3730 10412466 : if (es->predicate)
3731 1915567 : pred = *es->predicate;
3732 10412466 : sum->account_size_time (size, time, pred, pred, true);
3733 : }
3734 3644747 : for (e = node->indirect_calls; e; e = e->next_callee)
3735 : {
3736 316825 : int size = 0;
3737 316825 : sreal time = 0;
3738 :
3739 316825 : estimate_edge_size_and_time (e, &size, NULL, &time, NULL, NULL);
3740 316825 : ipa_predicate pred = true;
3741 316825 : class ipa_call_summary *es = ipa_call_summaries->get (e);
3742 :
3743 316825 : if (es->predicate)
3744 81414 : pred = *es->predicate;
3745 316825 : sum->account_size_time (size, time, pred, pred, true);
3746 : }
3747 3327922 : }
3748 :
3749 : /* Increase SIZE, MIN_SIZE and TIME for size and time needed to handle all
3750 : calls in NODE. POSSIBLE_TRUTHS and AVALS (the latter if non-NULL) describe
3751 : context of the call site. */
3752 :
3753 : static void
3754 20892880 : estimate_calls_size_and_time (struct cgraph_node *node, int *size,
3755 : int *min_size, sreal *time,
3756 : ipa_hints *hints,
3757 : clause_t possible_truths,
3758 : ipa_call_arg_values *avals)
3759 : {
3760 20892880 : class ipa_fn_summary *sum = ipa_fn_summaries->get (node);
3761 20892880 : bool use_table = true;
3762 :
3763 20892880 : gcc_assert (node->callees || node->indirect_calls);
3764 :
3765 : /* During early inlining we do not calculate info for very
3766 : large functions and thus there is no need for producing
3767 : summaries. */
3768 20892880 : if (!ipa_node_params_sum)
3769 : use_table = false;
3770 : /* Do not calculate summaries for simple wrappers; it is waste
3771 : of memory. */
3772 11324783 : else if (node->callees && !node->indirect_calls
3773 10564268 : && node->callees->inline_failed && !node->callees->next_callee)
3774 : use_table = false;
3775 : /* If there is an indirect edge that may be optimized, we need
3776 : to go the slow way. */
3777 9098431 : else if (avals
3778 9098431 : && (avals->m_known_vals.length ()
3779 3399184 : || avals->m_known_contexts.length ()
3780 3252298 : || avals->m_known_aggs.length ()))
3781 : {
3782 4296568 : ipa_node_params *params_summary = ipa_node_params_sum->get (node);
3783 4296568 : unsigned int nargs = params_summary
3784 4296568 : ? ipa_get_param_count (params_summary) : 0;
3785 :
3786 14273334 : for (unsigned int i = 0; i < nargs && use_table; i++)
3787 : {
3788 9976766 : if (ipa_is_param_used_by_indirect_call (params_summary, i)
3789 9976766 : && (avals->safe_sval_at (i)
3790 270944 : || (ipa_argagg_value_list (avals).value_for_index_p (i))))
3791 : use_table = false;
3792 9880076 : else if (ipa_is_param_used_by_polymorphic_call (params_summary, i)
3793 9880076 : && (avals->m_known_contexts.length () > i
3794 10119447 : && !avals->m_known_contexts[i].useless_p ()))
3795 : use_table = false;
3796 : }
3797 : }
3798 :
3799 : /* Fast path is via the call size time table. */
3800 4296568 : if (use_table)
3801 : {
3802 : /* Build summary if it is absent. */
3803 8835974 : if (!sum->call_size_time_table.length ())
3804 : {
3805 896314 : ipa_predicate true_pred = true;
3806 896314 : sum->account_size_time (0, 0, true_pred, true_pred, true);
3807 896314 : summarize_calls_size_and_time (node, sum);
3808 : }
3809 :
3810 8835974 : int old_size = *size;
3811 8835974 : sreal old_time = time ? *time : 0;
3812 :
3813 8835974 : if (min_size)
3814 8835974 : *min_size += sum->call_size_time_table[0].size;
3815 :
3816 8835974 : unsigned int i;
3817 8835974 : size_time_entry *e;
3818 :
3819 : /* Walk the table and account sizes and times. */
3820 23293345 : for (i = 0; sum->call_size_time_table.iterate (i, &e);
3821 : i++)
3822 14457371 : if (e->exec_predicate.evaluate (possible_truths))
3823 : {
3824 13533780 : *size += e->size;
3825 13533780 : if (time)
3826 11359702 : *time += e->time;
3827 : }
3828 :
3829 : /* Be careful and see if both methods agree. */
3830 21 : if ((flag_checking || dump_file)
3831 : /* Do not try to sanity check when we know we lost some
3832 : precision. */
3833 8835974 : && sum->call_size_time_table.length ()
3834 : < ipa_fn_summary::max_size_time_table_size)
3835 : {
3836 8835953 : estimate_calls_size_and_time_1 (node, &old_size, NULL, &old_time, NULL,
3837 : possible_truths, avals);
3838 8835953 : gcc_assert (*size == old_size);
3839 16240399 : if (time && (*time - old_time > 1 || *time - old_time < -1)
3840 8835980 : && dump_file)
3841 0 : fprintf (dump_file, "Time mismatch in call summary %f!=%f\n",
3842 : old_time.to_double (),
3843 : time->to_double ());
3844 : }
3845 : }
3846 : /* Slow path by walking all edges. */
3847 : else
3848 12056906 : estimate_calls_size_and_time_1 (node, size, min_size, time, hints,
3849 : possible_truths, avals);
3850 20892880 : }
3851 :
3852 : /* Main constructor for ipa call context. Memory allocation of ARG_VALUES
3853 : is owned by the caller. INLINE_PARAM_SUMMARY is also owned by the
3854 : caller. */
3855 :
3856 19954048 : ipa_call_context::ipa_call_context (cgraph_node *node, clause_t possible_truths,
3857 : clause_t nonspec_possible_truths,
3858 : vec<inline_param_summary>
3859 : inline_param_summary,
3860 : ipa_auto_call_arg_values *arg_values)
3861 19954048 : : m_node (node), m_possible_truths (possible_truths),
3862 19954048 : m_nonspec_possible_truths (nonspec_possible_truths),
3863 19954048 : m_inline_param_summary (inline_param_summary),
3864 19954048 : m_avals (arg_values)
3865 : {
3866 19954048 : }
3867 :
3868 : /* Set THIS to be a duplicate of CTX. Copy all relevant info. */
3869 :
3870 : void
3871 2009410 : ipa_cached_call_context::duplicate_from (const ipa_call_context &ctx)
3872 : {
3873 2009410 : m_node = ctx.m_node;
3874 2009410 : m_possible_truths = ctx.m_possible_truths;
3875 2009410 : m_nonspec_possible_truths = ctx.m_nonspec_possible_truths;
3876 2009410 : ipa_node_params *params_summary = ipa_node_params_sum->get (m_node);
3877 2009410 : unsigned int nargs = params_summary
3878 2009410 : ? ipa_get_param_count (params_summary) : 0;
3879 :
3880 2009410 : m_inline_param_summary = vNULL;
3881 : /* Copy the info only if there is at least one useful entry. */
3882 2009410 : if (ctx.m_inline_param_summary.exists ())
3883 : {
3884 1790087 : unsigned int n = MIN (ctx.m_inline_param_summary.length (), nargs);
3885 :
3886 4099997 : for (unsigned int i = 0; i < n; i++)
3887 3349254 : if (ipa_is_param_used_by_ipa_predicates (params_summary, i)
3888 4780066 : && !ctx.m_inline_param_summary[i].useless_p ())
3889 : {
3890 1039344 : m_inline_param_summary
3891 1039344 : = ctx.m_inline_param_summary.copy ();
3892 1039344 : break;
3893 : }
3894 : }
3895 2009410 : m_avals.m_known_vals = vNULL;
3896 2009410 : if (ctx.m_avals.m_known_vals.exists ())
3897 : {
3898 2009410 : unsigned int n = MIN (ctx.m_avals.m_known_vals.length (), nargs);
3899 :
3900 4666531 : for (unsigned int i = 0; i < n; i++)
3901 2683948 : if (ipa_is_param_used_by_indirect_call (params_summary, i)
3902 2683948 : && ctx.m_avals.m_known_vals[i])
3903 : {
3904 26827 : m_avals.m_known_vals = ctx.m_avals.m_known_vals.copy ();
3905 26827 : break;
3906 : }
3907 : }
3908 :
3909 2009410 : m_avals.m_known_contexts = vNULL;
3910 2009410 : if (ctx.m_avals.m_known_contexts.exists ())
3911 : {
3912 2009410 : unsigned int n = MIN (ctx.m_avals.m_known_contexts.length (), nargs);
3913 :
3914 2011427 : for (unsigned int i = 0; i < n; i++)
3915 26720 : if (ipa_is_param_used_by_polymorphic_call (params_summary, i)
3916 26720 : && !ctx.m_avals.m_known_contexts[i].useless_p ())
3917 : {
3918 24703 : m_avals.m_known_contexts = ctx.m_avals.m_known_contexts.copy ();
3919 24703 : break;
3920 : }
3921 : }
3922 :
3923 2009410 : m_avals.m_known_aggs = vNULL;
3924 2009410 : if (ctx.m_avals.m_known_aggs.exists ())
3925 : {
3926 2009410 : const ipa_argagg_value_list avl (&ctx.m_avals);
3927 6678506 : for (unsigned int i = 0; i < nargs; i++)
3928 4672020 : if (ipa_is_param_used_by_indirect_call (params_summary, i)
3929 4672020 : && avl.value_for_index_p (i))
3930 : {
3931 2924 : m_avals.m_known_aggs = ctx.m_avals.m_known_aggs.copy ();
3932 2924 : break;
3933 : }
3934 : }
3935 :
3936 2009410 : m_avals.m_known_value_ranges = vNULL;
3937 2009410 : }
3938 :
3939 : /* Release memory used by known_vals/contexts/aggs vectors. and
3940 : inline_param_summary. */
3941 :
3942 : void
3943 3270908 : ipa_cached_call_context::release ()
3944 : {
3945 : /* See if context is initialized at first place. */
3946 3270908 : if (!m_node)
3947 : return;
3948 2009410 : m_avals.m_known_aggs.release ();
3949 2009410 : m_avals.m_known_vals.release ();
3950 2009410 : m_avals.m_known_contexts.release ();
3951 2009410 : m_inline_param_summary.release ();
3952 : }
3953 :
3954 : /* Return true if CTX describes the same call context as THIS. */
3955 :
3956 : bool
3957 7098152 : ipa_call_context::equal_to (const ipa_call_context &ctx)
3958 : {
3959 7098152 : if (m_node != ctx.m_node
3960 5836654 : || m_possible_truths != ctx.m_possible_truths
3961 5220533 : || m_nonspec_possible_truths != ctx.m_nonspec_possible_truths)
3962 : return false;
3963 :
3964 5220533 : ipa_node_params *params_summary = ipa_node_params_sum->get (m_node);
3965 5220533 : unsigned int nargs = params_summary
3966 5220533 : ? ipa_get_param_count (params_summary) : 0;
3967 :
3968 5220533 : if (m_inline_param_summary.exists () || ctx.m_inline_param_summary.exists ())
3969 : {
3970 15550030 : for (unsigned int i = 0; i < nargs; i++)
3971 : {
3972 10665883 : if (!ipa_is_param_used_by_ipa_predicates (params_summary, i))
3973 3348536 : continue;
3974 7317347 : if (i >= m_inline_param_summary.length ()
3975 5091902 : || m_inline_param_summary[i].useless_p ())
3976 : {
3977 3229647 : if (i < ctx.m_inline_param_summary.length ()
3978 3229647 : && !ctx.m_inline_param_summary[i].useless_p ())
3979 : return false;
3980 3184137 : continue;
3981 : }
3982 4087700 : if (i >= ctx.m_inline_param_summary.length ()
3983 4087697 : || ctx.m_inline_param_summary[i].useless_p ())
3984 : {
3985 44896 : if (i < m_inline_param_summary.length ()
3986 44896 : && !m_inline_param_summary[i].useless_p ())
3987 : return false;
3988 0 : continue;
3989 : }
3990 4042804 : if (!m_inline_param_summary[i].equal_to
3991 4042804 : (ctx.m_inline_param_summary[i]))
3992 : return false;
3993 : }
3994 : }
3995 5102286 : if (m_avals.m_known_vals.exists () || ctx.m_avals.m_known_vals.exists ())
3996 : {
3997 15570341 : for (unsigned int i = 0; i < nargs; i++)
3998 : {
3999 10478781 : if (!ipa_is_param_used_by_indirect_call (params_summary, i))
4000 10268481 : continue;
4001 210300 : if (i >= m_avals.m_known_vals.length () || !m_avals.m_known_vals[i])
4002 : {
4003 152811 : if (i < ctx.m_avals.m_known_vals.length ()
4004 152811 : && ctx.m_avals.m_known_vals[i])
4005 : return false;
4006 152722 : continue;
4007 : }
4008 57489 : if (i >= ctx.m_avals.m_known_vals.length ()
4009 57489 : || !ctx.m_avals.m_known_vals[i])
4010 : {
4011 : if (i < m_avals.m_known_vals.length () && m_avals.m_known_vals[i])
4012 : return false;
4013 : continue;
4014 : }
4015 57399 : if (m_avals.m_known_vals[i] != ctx.m_avals.m_known_vals[i])
4016 : return false;
4017 : }
4018 : }
4019 5091560 : if (m_avals.m_known_contexts.exists ()
4020 5091560 : || ctx.m_avals.m_known_contexts.exists ())
4021 : {
4022 15555091 : for (unsigned int i = 0; i < nargs; i++)
4023 : {
4024 10464944 : if (!ipa_is_param_used_by_polymorphic_call (params_summary, i))
4025 10328155 : continue;
4026 136789 : if (i >= m_avals.m_known_contexts.length ()
4027 135627 : || m_avals.m_known_contexts[i].useless_p ())
4028 : {
4029 1162 : if (i < ctx.m_avals.m_known_contexts.length ()
4030 1162 : && !ctx.m_avals.m_known_contexts[i].useless_p ())
4031 : return false;
4032 1150 : continue;
4033 : }
4034 135627 : if (i >= ctx.m_avals.m_known_contexts.length ()
4035 135627 : || ctx.m_avals.m_known_contexts[i].useless_p ())
4036 : {
4037 8 : if (i < m_avals.m_known_contexts.length ()
4038 8 : && !m_avals.m_known_contexts[i].useless_p ())
4039 : return false;
4040 0 : continue;
4041 : }
4042 135619 : if (!m_avals.m_known_contexts[i].equal_to
4043 135619 : (ctx.m_avals.m_known_contexts[i]))
4044 : return false;
4045 : }
4046 : }
4047 5090147 : if (m_avals.m_known_aggs.exists () || ctx.m_avals.m_known_aggs.exists ())
4048 : {
4049 : unsigned i = 0, j = 0;
4050 12060486 : while (i < m_avals.m_known_aggs.length ()
4051 6029484 : || j < ctx.m_avals.m_known_aggs.length ())
4052 : {
4053 940742 : if (i >= m_avals.m_known_aggs.length ())
4054 : {
4055 935844 : int idx2 = ctx.m_avals.m_known_aggs[j].index;
4056 935844 : if (ipa_is_param_used_by_indirect_call (params_summary, idx2))
4057 : return false;
4058 935290 : j++;
4059 935290 : continue;
4060 935290 : }
4061 4898 : if (j >= ctx.m_avals.m_known_aggs.length ())
4062 : {
4063 600 : int idx1 = m_avals.m_known_aggs[i].index;
4064 600 : if (ipa_is_param_used_by_indirect_call (params_summary, idx1))
4065 : return false;
4066 8 : i++;
4067 8 : continue;
4068 8 : }
4069 :
4070 4298 : int idx1 = m_avals.m_known_aggs[i].index;
4071 4298 : int idx2 = ctx.m_avals.m_known_aggs[j].index;
4072 4298 : if (idx1 < idx2)
4073 : {
4074 0 : if (ipa_is_param_used_by_indirect_call (params_summary, idx1))
4075 : return false;
4076 0 : i++;
4077 0 : continue;
4078 : }
4079 4298 : if (idx1 > idx2)
4080 : {
4081 0 : if (ipa_is_param_used_by_indirect_call (params_summary, idx2))
4082 : return false;
4083 0 : j++;
4084 0 : continue;
4085 : }
4086 4298 : if (!ipa_is_param_used_by_indirect_call (params_summary, idx1))
4087 : {
4088 482 : i++;
4089 482 : j++;
4090 482 : continue;
4091 : }
4092 :
4093 3816 : if ((m_avals.m_known_aggs[i].unit_offset
4094 3816 : != ctx.m_avals.m_known_aggs[j].unit_offset)
4095 3808 : || (m_avals.m_known_aggs[i].by_ref
4096 3808 : != ctx.m_avals.m_known_aggs[j].by_ref)
4097 7624 : || !operand_equal_p (m_avals.m_known_aggs[i].value,
4098 3808 : ctx.m_avals.m_known_aggs[j].value))
4099 : return false;
4100 3557 : i++;
4101 3557 : j++;
4102 : }
4103 : }
4104 : return true;
4105 : }
4106 :
4107 : /* Fill in the selected fields in ESTIMATES with value estimated for call in
4108 : this context. Always compute size and min_size. Only compute time and
4109 : nonspecialized_time if EST_TIMES is true. Only compute hints if EST_HINTS
4110 : is true. */
4111 :
4112 : void
4113 19922187 : ipa_call_context::estimate_size_and_time (ipa_call_estimates *estimates,
4114 : bool est_times, bool est_hints)
4115 : {
4116 19922187 : class ipa_fn_summary *info = ipa_fn_summaries->get (m_node);
4117 19922187 : size_time_entry *e;
4118 19922187 : int size = 0;
4119 19922187 : sreal time = 0;
4120 19922187 : int min_size = 0;
4121 19922187 : ipa_hints hints = 0;
4122 19922187 : sreal loops_with_known_iterations = 0;
4123 19922187 : sreal loops_with_known_strides = 0;
4124 19922187 : int i;
4125 :
4126 19922187 : if (dump_file && (dump_flags & TDF_DETAILS))
4127 : {
4128 1317 : bool found = false;
4129 1317 : fprintf (dump_file, " Estimating body: %s\n"
4130 : " Known to be false: ", m_node->dump_name ());
4131 :
4132 4267 : for (i = ipa_predicate::not_inlined_condition;
4133 8534 : i < (ipa_predicate::first_dynamic_condition
4134 6984 : + (int) vec_safe_length (info->conds)); i++)
4135 2950 : if (!(m_possible_truths & (1 << i)))
4136 : {
4137 1783 : if (found)
4138 519 : fprintf (dump_file, ", ");
4139 1783 : found = true;
4140 1783 : dump_condition (dump_file, info->conds, i);
4141 : }
4142 : }
4143 :
4144 19922187 : if (m_node->callees || m_node->indirect_calls)
4145 26412400 : estimate_calls_size_and_time (m_node, &size, &min_size,
4146 : est_times ? &time : NULL,
4147 : est_hints ? &hints : NULL, m_possible_truths,
4148 : &m_avals);
4149 :
4150 19922187 : sreal nonspecialized_time = time;
4151 :
4152 19922187 : min_size += info->size_time_table[0].size;
4153 136872298 : for (i = 0; info->size_time_table.iterate (i, &e); i++)
4154 : {
4155 116950111 : bool exec = e->exec_predicate.evaluate (m_nonspec_possible_truths);
4156 :
4157 : /* Because predicates are conservative, it can happen that nonconst is 1
4158 : but exec is 0. */
4159 116950111 : if (exec)
4160 : {
4161 114434874 : bool nonconst = e->nonconst_predicate.evaluate (m_possible_truths);
4162 :
4163 114434874 : gcc_checking_assert (e->time >= 0);
4164 114434874 : gcc_checking_assert (time >= 0);
4165 :
4166 : /* We compute specialized size only because size of nonspecialized
4167 : copy is context independent.
4168 :
4169 : The difference between nonspecialized execution and specialized is
4170 : that nonspecialized is not going to have optimized out computations
4171 : known to be constant in a specialized setting. */
4172 114434874 : if (nonconst)
4173 57614905 : size += e->size;
4174 114434874 : if (!est_times)
4175 59190740 : continue;
4176 55244134 : nonspecialized_time += e->time;
4177 55244134 : if (!nonconst)
4178 : ;
4179 26588190 : else if (!m_inline_param_summary.exists ())
4180 : {
4181 2424741 : if (nonconst)
4182 2424741 : time += e->time;
4183 : }
4184 : else
4185 : {
4186 24163449 : int prob = e->nonconst_predicate.probability
4187 24163449 : (info->conds, m_possible_truths,
4188 : m_inline_param_summary);
4189 24163449 : gcc_checking_assert (prob >= 0);
4190 24163449 : gcc_checking_assert (prob <= REG_BR_PROB_BASE);
4191 24163449 : if (prob == REG_BR_PROB_BASE)
4192 19971513 : time += e->time;
4193 : else
4194 4191936 : time += e->time * prob / REG_BR_PROB_BASE;
4195 : }
4196 55244134 : gcc_checking_assert (time >= 0);
4197 : }
4198 : }
4199 19922187 : gcc_checking_assert (info->size_time_table[0].exec_predicate == true);
4200 19922187 : gcc_checking_assert (info->size_time_table[0].nonconst_predicate == true);
4201 19922187 : gcc_checking_assert (min_size >= 0);
4202 19922187 : gcc_checking_assert (size >= 0);
4203 19922187 : gcc_checking_assert (time >= 0);
4204 : /* nonspecialized_time should be always bigger than specialized time.
4205 : Roundoff issues however may get into the way. */
4206 19922187 : gcc_checking_assert ((nonspecialized_time - time * 99 / 100) >= -1);
4207 :
4208 : /* Roundoff issues may make specialized time bigger than nonspecialized
4209 : time. We do not really want that to happen because some heuristics
4210 : may get confused by seeing negative speedups. */
4211 19922187 : if (time > nonspecialized_time)
4212 0 : time = nonspecialized_time;
4213 :
4214 19922187 : if (est_hints)
4215 : {
4216 7301615 : if (info->scc_no)
4217 233795 : hints |= INLINE_HINT_in_scc;
4218 7301615 : if (DECL_DECLARED_INLINE_P (m_node->decl))
4219 4656222 : hints |= INLINE_HINT_declared_inline;
4220 7301615 : if (info->builtin_constant_p_parms.length ()
4221 9549 : && DECL_DECLARED_INLINE_P (m_node->decl))
4222 9491 : hints |= INLINE_HINT_builtin_constant_p;
4223 :
4224 7301615 : ipa_freqcounting_predicate *fcp;
4225 7853829 : for (i = 0; vec_safe_iterate (info->loop_iterations, i, &fcp); i++)
4226 552214 : if (!fcp->predicate->evaluate (m_possible_truths))
4227 : {
4228 371070 : hints |= INLINE_HINT_loop_iterations;
4229 371070 : loops_with_known_iterations += fcp->freq;
4230 : }
4231 7301615 : estimates->loops_with_known_iterations = loops_with_known_iterations;
4232 :
4233 7593847 : for (i = 0; vec_safe_iterate (info->loop_strides, i, &fcp); i++)
4234 292232 : if (!fcp->predicate->evaluate (m_possible_truths))
4235 : {
4236 262468 : hints |= INLINE_HINT_loop_stride;
4237 262468 : loops_with_known_strides += fcp->freq;
4238 : }
4239 7301615 : estimates->loops_with_known_strides = loops_with_known_strides;
4240 : }
4241 :
4242 19922187 : size = RDIV (size, ipa_fn_summary::size_scale);
4243 19922187 : min_size = RDIV (min_size, ipa_fn_summary::size_scale);
4244 :
4245 19922187 : if (dump_file && (dump_flags & TDF_DETAILS))
4246 : {
4247 1317 : fprintf (dump_file, "\n size:%i", (int) size);
4248 1317 : if (est_times)
4249 1080 : fprintf (dump_file, " time:%f nonspec time:%f",
4250 : time.to_double (), nonspecialized_time.to_double ());
4251 1317 : if (est_hints)
4252 1080 : fprintf (dump_file, " loops with known iterations:%f "
4253 : "known strides:%f", loops_with_known_iterations.to_double (),
4254 : loops_with_known_strides.to_double ());
4255 1317 : fprintf (dump_file, "\n");
4256 : }
4257 19922187 : if (est_times)
4258 : {
4259 7301615 : estimates->time = time;
4260 7301615 : estimates->nonspecialized_time = nonspecialized_time;
4261 : }
4262 19922187 : estimates->size = size;
4263 19922187 : estimates->min_size = min_size;
4264 19922187 : if (est_hints)
4265 7301615 : estimates->hints = hints;
4266 19922187 : return;
4267 : }
4268 :
4269 :
4270 : /* Estimate size and time needed to execute callee of EDGE assuming that
4271 : parameters known to be constant at caller of EDGE are propagated.
4272 : KNOWN_VALS and KNOWN_CONTEXTS are vectors of assumed known constant values
4273 : and types for parameters. */
4274 :
4275 : void
4276 231523 : estimate_ipcp_clone_size_and_time (struct cgraph_node *node,
4277 : ipa_auto_call_arg_values *avals,
4278 : ipa_call_estimates *estimates)
4279 : {
4280 231523 : clause_t clause, nonspec_clause;
4281 :
4282 231523 : evaluate_conditions_for_known_args (node, false, avals, &clause,
4283 : &nonspec_clause, NULL);
4284 231523 : ipa_call_context ctx (node, clause, nonspec_clause, vNULL, avals);
4285 231523 : ctx.estimate_size_and_time (estimates);
4286 231523 : }
4287 :
4288 : /* Return stack frame offset where frame of NODE is supposed to start inside
4289 : of the function it is inlined to.
4290 : Return 0 for functions that are not inlined. */
4291 :
4292 : HOST_WIDE_INT
4293 5274487 : ipa_get_stack_frame_offset (struct cgraph_node *node)
4294 : {
4295 5274487 : HOST_WIDE_INT offset = 0;
4296 5274487 : if (!node->inlined_to)
4297 : return 0;
4298 4202530 : node = node->callers->caller;
4299 5444662 : while (true)
4300 : {
4301 4823596 : offset += ipa_size_summaries->get (node)->estimated_self_stack_size;
4302 4823596 : if (!node->inlined_to)
4303 : return offset;
4304 621066 : node = node->callers->caller;
4305 : }
4306 : }
4307 :
4308 :
4309 : /* Update summary information of inline clones after inlining.
4310 : Compute peak stack usage. */
4311 :
4312 : static void
4313 4960254 : inline_update_callee_summaries (struct cgraph_node *node, int depth)
4314 : {
4315 4960254 : struct cgraph_edge *e;
4316 :
4317 4960254 : ipa_propagate_frequency (node);
4318 9633658 : for (e = node->callees; e; e = e->next_callee)
4319 : {
4320 4673404 : if (!e->inline_failed)
4321 758277 : inline_update_callee_summaries (e->callee, depth);
4322 : else
4323 3915127 : ipa_call_summaries->get (e)->loop_depth += depth;
4324 : }
4325 5066359 : for (e = node->indirect_calls; e; e = e->next_callee)
4326 106105 : ipa_call_summaries->get (e)->loop_depth += depth;
4327 4960254 : }
4328 :
4329 : /* Update change_prob and points_to_local_or_readonly_memory of EDGE after
4330 : INLINED_EDGE has been inlined.
4331 :
4332 : When function A is inlined in B and A calls C with parameter that
4333 : changes with probability PROB1 and C is known to be passthrough
4334 : of argument if B that change with probability PROB2, the probability
4335 : of change is now PROB1*PROB2. */
4336 :
4337 : static void
4338 4021542 : remap_edge_params (struct cgraph_edge *inlined_edge,
4339 : struct cgraph_edge *edge)
4340 : {
4341 4021542 : if (ipa_node_params_sum)
4342 : {
4343 2833200 : int i;
4344 2833200 : ipa_edge_args *args = ipa_edge_args_sum->get (edge);
4345 2833200 : if (!args)
4346 : return;
4347 1610071 : class ipa_call_summary *es = ipa_call_summaries->get (edge);
4348 1610071 : class ipa_call_summary *inlined_es
4349 1610071 : = ipa_call_summaries->get (inlined_edge);
4350 :
4351 1610071 : if (es->param.length () == 0)
4352 : return;
4353 :
4354 9385858 : for (i = 0; i < ipa_get_cs_argument_count (args); i++)
4355 : {
4356 3162810 : struct ipa_jump_func *jfunc = ipa_get_ith_jump_func (args, i);
4357 3162810 : if (jfunc->type == IPA_JF_PASS_THROUGH
4358 2323547 : || jfunc->type == IPA_JF_ANCESTOR)
4359 : {
4360 985112 : int id = jfunc->type == IPA_JF_PASS_THROUGH
4361 985112 : ? ipa_get_jf_pass_through_formal_id (jfunc)
4362 145849 : : ipa_get_jf_ancestor_formal_id (jfunc);
4363 1970141 : if (id < (int) inlined_es->param.length ())
4364 : {
4365 985021 : int prob1 = es->param[i].change_prob;
4366 985021 : int prob2 = inlined_es->param[id].change_prob;
4367 985021 : int prob = combine_probabilities (prob1, prob2);
4368 :
4369 985021 : if (prob1 && prob2 && !prob)
4370 985021 : prob = 1;
4371 :
4372 985021 : es->param[i].change_prob = prob;
4373 :
4374 1970042 : if (inlined_es
4375 985021 : ->param[id].points_to_local_or_readonly_memory)
4376 215521 : es->param[i].points_to_local_or_readonly_memory = true;
4377 1970042 : if (inlined_es
4378 985021 : ->param[id].points_to_possible_sra_candidate)
4379 178846 : es->param[i].points_to_possible_sra_candidate = true;
4380 : }
4381 985112 : if (!es->param[i].points_to_local_or_readonly_memory
4382 : && jfunc->type == IPA_JF_CONST
4383 : && points_to_local_or_readonly_memory_p
4384 : (ipa_get_jf_constant (jfunc)))
4385 : es->param[i].points_to_local_or_readonly_memory = true;
4386 : }
4387 : }
4388 : }
4389 : }
4390 :
4391 : /* Update edge summaries of NODE after INLINED_EDGE has been inlined.
4392 :
4393 : Remap predicates of callees of NODE. Rest of arguments match
4394 : remap_predicate.
4395 :
4396 : Also update change probabilities. */
4397 :
4398 : static void
4399 4960254 : remap_edge_summaries (struct cgraph_edge *inlined_edge,
4400 : struct cgraph_node *node,
4401 : class ipa_fn_summary *info,
4402 : class ipa_node_params *params_summary,
4403 : class ipa_fn_summary *callee_info,
4404 : const vec<int> &operand_map,
4405 : const vec<HOST_WIDE_INT> &offset_map,
4406 : clause_t possible_truths,
4407 : ipa_predicate *toplev_predicate)
4408 : {
4409 4960254 : struct cgraph_edge *e, *next;
4410 9633205 : for (e = node->callees; e; e = next)
4411 : {
4412 4672951 : ipa_predicate p;
4413 4672951 : next = e->next_callee;
4414 :
4415 4672951 : if (e->inline_failed)
4416 : {
4417 3914674 : class ipa_call_summary *es = ipa_call_summaries->get (e);
4418 3914674 : remap_edge_params (inlined_edge, e);
4419 :
4420 3914674 : if (es->predicate)
4421 : {
4422 1351489 : p = es->predicate->remap_after_inlining
4423 1351489 : (info, params_summary,
4424 : callee_info, operand_map,
4425 : offset_map, possible_truths,
4426 : *toplev_predicate);
4427 1351489 : edge_set_predicate (e, &p);
4428 : }
4429 : else
4430 2563185 : edge_set_predicate (e, toplev_predicate);
4431 : }
4432 : else
4433 758277 : remap_edge_summaries (inlined_edge, e->callee, info,
4434 : params_summary, callee_info,
4435 : operand_map, offset_map, possible_truths,
4436 : toplev_predicate);
4437 : }
4438 5067122 : for (e = node->indirect_calls; e; e = next)
4439 : {
4440 106868 : class ipa_call_summary *es = ipa_call_summaries->get (e);
4441 106868 : ipa_predicate p;
4442 106868 : next = e->next_callee;
4443 :
4444 106868 : remap_edge_params (inlined_edge, e);
4445 106868 : if (es->predicate)
4446 : {
4447 27265 : p = es->predicate->remap_after_inlining
4448 27265 : (info, params_summary,
4449 : callee_info, operand_map, offset_map,
4450 : possible_truths, *toplev_predicate);
4451 27265 : edge_set_predicate (e, &p);
4452 : }
4453 : else
4454 79603 : edge_set_predicate (e, toplev_predicate);
4455 : }
4456 4960254 : }
4457 :
4458 : /* Run remap_after_inlining on each predicate in V. */
4459 :
4460 : static void
4461 8403954 : remap_freqcounting_predicate (class ipa_fn_summary *info,
4462 : class ipa_node_params *params_summary,
4463 : class ipa_fn_summary *callee_info,
4464 : vec<ipa_freqcounting_predicate, va_gc> *v,
4465 : const vec<int> &operand_map,
4466 : const vec<HOST_WIDE_INT> &offset_map,
4467 : clause_t possible_truths,
4468 : ipa_predicate *toplev_predicate)
4469 :
4470 : {
4471 8403954 : ipa_freqcounting_predicate *fcp;
4472 8431259 : for (int i = 0; vec_safe_iterate (v, i, &fcp); i++)
4473 : {
4474 27305 : ipa_predicate p
4475 27305 : = fcp->predicate->remap_after_inlining (info, params_summary,
4476 : callee_info, operand_map,
4477 : offset_map, possible_truths,
4478 : *toplev_predicate);
4479 37844 : if (p != false && p != true)
4480 2639 : *fcp->predicate &= p;
4481 : }
4482 8403954 : }
4483 :
4484 : /* We inlined EDGE. Update summary of the function we inlined into. */
4485 :
4486 : void
4487 4201977 : ipa_merge_fn_summary_after_inlining (struct cgraph_edge *edge)
4488 : {
4489 4201977 : ipa_fn_summary *callee_info = ipa_fn_summaries->get (edge->callee);
4490 3907063 : struct cgraph_node *to = (edge->caller->inlined_to
4491 4201977 : ? edge->caller->inlined_to : edge->caller);
4492 4201977 : class ipa_fn_summary *info = ipa_fn_summaries->get (to);
4493 4201977 : clause_t clause = 0; /* not_inline is known to be false. */
4494 4201977 : size_time_entry *e;
4495 4201977 : auto_vec<int, 8> operand_map;
4496 4201977 : auto_vec<HOST_WIDE_INT, 8> offset_map;
4497 4201977 : int i;
4498 4201977 : ipa_predicate toplev_predicate;
4499 4201977 : class ipa_call_summary *es = ipa_call_summaries->get (edge);
4500 4201977 : ipa_node_params *params_summary = (ipa_node_params_sum
4501 4201977 : ? ipa_node_params_sum->get (to) : NULL);
4502 :
4503 4201977 : if (es->predicate)
4504 359263 : toplev_predicate = *es->predicate;
4505 : else
4506 3842714 : toplev_predicate = true;
4507 :
4508 4201977 : info->fp_expressions |= callee_info->fp_expressions;
4509 4201977 : info->target_info |= callee_info->target_info;
4510 :
4511 4201977 : if (callee_info->conds)
4512 : {
4513 3081021 : ipa_auto_call_arg_values avals;
4514 3081021 : evaluate_properties_for_edge (edge, true, &clause, NULL, &avals, false);
4515 3081021 : }
4516 4201977 : if (ipa_node_params_sum && callee_info->conds)
4517 : {
4518 846213 : ipa_edge_args *args = ipa_edge_args_sum->get (edge);
4519 846213 : int count = args ? ipa_get_cs_argument_count (args) : 0;
4520 846074 : int i;
4521 :
4522 846074 : if (count)
4523 : {
4524 846074 : operand_map.safe_grow_cleared (count, true);
4525 846074 : offset_map.safe_grow_cleared (count, true);
4526 : }
4527 2709881 : for (i = 0; i < count; i++)
4528 : {
4529 1863668 : struct ipa_jump_func *jfunc = ipa_get_ith_jump_func (args, i);
4530 1863668 : int map = -1;
4531 :
4532 : /* TODO: handle non-NOPs when merging. */
4533 1863668 : if (jfunc->type == IPA_JF_PASS_THROUGH)
4534 : {
4535 300631 : if (ipa_get_jf_pass_through_operation (jfunc) == NOP_EXPR)
4536 297061 : map = ipa_get_jf_pass_through_formal_id (jfunc);
4537 300631 : if (!ipa_get_jf_pass_through_agg_preserved (jfunc))
4538 202188 : offset_map[i] = -1;
4539 : }
4540 1563037 : else if (jfunc->type == IPA_JF_ANCESTOR)
4541 : {
4542 95538 : HOST_WIDE_INT offset = ipa_get_jf_ancestor_offset (jfunc);
4543 95538 : if (offset >= 0 && offset < INT_MAX)
4544 : {
4545 95538 : map = ipa_get_jf_ancestor_formal_id (jfunc);
4546 95538 : if (!ipa_get_jf_ancestor_agg_preserved (jfunc))
4547 56426 : offset = -1;
4548 95538 : offset_map[i] = offset;
4549 : }
4550 : }
4551 1863668 : operand_map[i] = map;
4552 3375981 : gcc_assert (map < ipa_get_param_count (params_summary));
4553 : }
4554 :
4555 : int ip;
4556 849077 : for (i = 0; callee_info->builtin_constant_p_parms.iterate (i, &ip); i++)
4557 2864 : if (ip < count && operand_map[ip] >= 0)
4558 48 : add_builtin_constant_p_parm (info, operand_map[ip]);
4559 : }
4560 4201977 : sreal freq = edge->sreal_frequency ();
4561 27416045 : for (i = 0; callee_info->size_time_table.iterate (i, &e); i++)
4562 : {
4563 19012091 : ipa_predicate p;
4564 19012091 : p = e->exec_predicate.remap_after_inlining
4565 19012091 : (info, params_summary,
4566 : callee_info, operand_map,
4567 : offset_map, clause,
4568 : toplev_predicate);
4569 19012091 : ipa_predicate nonconstp;
4570 19012091 : nonconstp = e->nonconst_predicate.remap_after_inlining
4571 19012091 : (info, params_summary,
4572 : callee_info, operand_map,
4573 : offset_map, clause,
4574 : toplev_predicate);
4575 28755498 : if (p != false && nonconstp != false)
4576 : {
4577 9407668 : sreal add_time = ((sreal)e->time * freq);
4578 9407668 : int prob = e->nonconst_predicate.probability (callee_info->conds,
4579 : clause, es->param);
4580 9407668 : if (prob != REG_BR_PROB_BASE)
4581 1141378 : add_time = add_time * prob / REG_BR_PROB_BASE;
4582 1141378 : if (prob != REG_BR_PROB_BASE
4583 1141378 : && dump_file && (dump_flags & TDF_DETAILS))
4584 : {
4585 3 : fprintf (dump_file, "\t\tScaling time by probability:%f\n",
4586 3 : (double) prob / REG_BR_PROB_BASE);
4587 : }
4588 9407668 : info->account_size_time (e->size, add_time, p, nonconstp);
4589 : }
4590 : }
4591 4201977 : remap_edge_summaries (edge, edge->callee, info, params_summary,
4592 : callee_info, operand_map,
4593 : offset_map, clause, &toplev_predicate);
4594 4201977 : remap_freqcounting_predicate (info, params_summary, callee_info,
4595 : info->loop_iterations, operand_map,
4596 : offset_map, clause, &toplev_predicate);
4597 4201977 : remap_freqcounting_predicate (info, params_summary, callee_info,
4598 : info->loop_strides, operand_map,
4599 : offset_map, clause, &toplev_predicate);
4600 :
4601 4201977 : HOST_WIDE_INT stack_frame_offset = ipa_get_stack_frame_offset (edge->callee);
4602 4201977 : HOST_WIDE_INT peak = stack_frame_offset + callee_info->estimated_stack_size;
4603 :
4604 4201977 : if (info->estimated_stack_size < peak)
4605 124266 : info->estimated_stack_size = peak;
4606 :
4607 4201977 : inline_update_callee_summaries (edge->callee, es->loop_depth);
4608 4201977 : if (info->call_size_time_table.length ())
4609 : {
4610 907274 : int edge_size = 0;
4611 907274 : sreal edge_time = 0;
4612 :
4613 907274 : estimate_edge_size_and_time (edge, &edge_size, NULL, &edge_time, NULL, 0);
4614 : /* Unaccount size and time of the optimized out call. */
4615 907274 : info->account_size_time (-edge_size, -edge_time,
4616 907274 : es->predicate ? *es->predicate : true,
4617 907274 : es->predicate ? *es->predicate : true,
4618 : true);
4619 : /* Account new calls. */
4620 907274 : summarize_calls_size_and_time (edge->callee, info);
4621 : }
4622 :
4623 : /* Free summaries that are not maintained for inline clones/edges. */
4624 4201977 : ipa_call_summaries->remove (edge);
4625 4201977 : ipa_fn_summaries->remove (edge->callee);
4626 4201977 : ipa_remove_from_growth_caches (edge);
4627 4201977 : }
4628 :
4629 : /* For performance reasons ipa_merge_fn_summary_after_inlining is not updating
4630 : overall size and time. Recompute it.
4631 : If RESET is true also recompute call_time_size_table. */
4632 :
4633 : void
4634 9833359 : ipa_update_overall_fn_summary (struct cgraph_node *node, bool reset)
4635 : {
4636 9833359 : class ipa_fn_summary *info = ipa_fn_summaries->get (node);
4637 9833359 : class ipa_size_summary *size_info = ipa_size_summaries->get (node);
4638 9833359 : size_time_entry *e;
4639 9833359 : int i;
4640 :
4641 9833359 : size_info->size = 0;
4642 9833359 : info->time = 0;
4643 49881838 : for (i = 0; info->size_time_table.iterate (i, &e); i++)
4644 : {
4645 40048479 : size_info->size += e->size;
4646 40048479 : info->time += e->time;
4647 : }
4648 9833359 : info->min_size = info->size_time_table[0].size;
4649 9833359 : if (reset)
4650 8882442 : info->call_size_time_table.release ();
4651 9833359 : if (node->callees || node->indirect_calls)
4652 7489004 : estimate_calls_size_and_time (node, &size_info->size, &info->min_size,
4653 : &info->time, NULL,
4654 : ~(clause_t) (1 << ipa_predicate::false_condition),
4655 : NULL);
4656 9833359 : size_info->size = RDIV (size_info->size, ipa_fn_summary::size_scale);
4657 9833359 : info->min_size = RDIV (info->min_size, ipa_fn_summary::size_scale);
4658 9833359 : }
4659 :
4660 :
4661 : /* This function performs intraprocedural analysis in NODE that is required to
4662 : inline indirect calls. */
4663 :
4664 : static void
4665 1418179 : inline_indirect_intraprocedural_analysis (struct cgraph_node *node)
4666 : {
4667 1418179 : ipa_analyze_node (node);
4668 1418179 : if (dump_file && (dump_flags & TDF_DETAILS))
4669 : {
4670 16 : ipa_print_node_params (dump_file, node);
4671 16 : ipa_print_node_jump_functions (dump_file, node);
4672 : }
4673 1418179 : }
4674 :
4675 :
4676 : /* Note function body size. */
4677 :
4678 : void
4679 1428855 : inline_analyze_function (struct cgraph_node *node)
4680 : {
4681 1428855 : push_cfun (DECL_STRUCT_FUNCTION (node->decl));
4682 :
4683 1428855 : if (dump_file)
4684 96 : fprintf (dump_file, "\nAnalyzing function: %s\n", node->dump_name ());
4685 1428855 : if (opt_for_fn (node->decl, optimize) && !node->thunk)
4686 1418179 : inline_indirect_intraprocedural_analysis (node);
4687 1428855 : compute_fn_summary (node, false);
4688 1428855 : if (!optimize)
4689 : {
4690 9376 : struct cgraph_edge *e;
4691 22131 : for (e = node->callees; e; e = e->next_callee)
4692 12755 : e->inline_failed = CIF_FUNCTION_NOT_OPTIMIZED;
4693 9450 : for (e = node->indirect_calls; e; e = e->next_callee)
4694 74 : e->inline_failed = CIF_FUNCTION_NOT_OPTIMIZED;
4695 : }
4696 :
4697 1428855 : pop_cfun ();
4698 1428855 : }
4699 :
4700 :
4701 : /* Called when new function is inserted to callgraph late. */
4702 :
4703 : void
4704 19765 : ipa_fn_summary_t::insert (struct cgraph_node *node, ipa_fn_summary *)
4705 : {
4706 19765 : inline_analyze_function (node);
4707 19765 : }
4708 :
4709 : /* Note function body size. */
4710 :
4711 : static void
4712 236960 : ipa_fn_summary_generate (void)
4713 : {
4714 236960 : struct cgraph_node *node;
4715 :
4716 2157823 : FOR_EACH_DEFINED_FUNCTION (node)
4717 1920863 : if (DECL_STRUCT_FUNCTION (node->decl))
4718 1907169 : node->versionable = tree_versionable_function_p (node->decl);
4719 :
4720 236960 : ipa_fn_summary_alloc ();
4721 :
4722 236960 : ipa_fn_summaries->enable_insertion_hook ();
4723 :
4724 236960 : ipa_register_cgraph_hooks ();
4725 :
4726 2157823 : FOR_EACH_DEFINED_FUNCTION (node)
4727 1920863 : if (!node->alias
4728 1920863 : && (flag_generate_lto || flag_generate_offload|| flag_wpa
4729 1724838 : || opt_for_fn (node->decl, optimize)))
4730 1390485 : inline_analyze_function (node);
4731 236960 : }
4732 :
4733 :
4734 : /* Write inline summary for edge E to OB. */
4735 :
4736 : static void
4737 348372 : read_ipa_call_summary (class lto_input_block *ib, struct cgraph_edge *e,
4738 : bool prevails)
4739 : {
4740 348372 : class ipa_call_summary *es = prevails
4741 348372 : ? ipa_call_summaries->get_create (e) : NULL;
4742 348372 : ipa_predicate p;
4743 348372 : int length, i;
4744 :
4745 348372 : int size = streamer_read_uhwi (ib);
4746 348372 : int time = streamer_read_uhwi (ib);
4747 348372 : int depth = streamer_read_uhwi (ib);
4748 :
4749 348372 : if (es)
4750 : {
4751 348323 : es->call_stmt_size = size;
4752 348323 : es->call_stmt_time = time;
4753 348323 : es->loop_depth = depth;
4754 : }
4755 :
4756 348372 : bitpack_d bp = streamer_read_bitpack (ib);
4757 348372 : if (es)
4758 348323 : es->is_return_callee_uncaptured = bp_unpack_value (&bp, 1);
4759 : else
4760 49 : bp_unpack_value (&bp, 1);
4761 :
4762 348372 : p.stream_in (ib);
4763 348372 : if (es)
4764 348323 : edge_set_predicate (e, &p);
4765 348372 : length = streamer_read_uhwi (ib);
4766 348372 : if (length && es
4767 348372 : && (e->possibly_call_in_translation_unit_p ()
4768 : /* Also stream in jump functions to builtins in hope that they
4769 : will get fnspecs. */
4770 119094 : || fndecl_built_in_p (e->callee->decl, BUILT_IN_NORMAL)))
4771 : {
4772 233087 : es->param.safe_grow_cleared (length, true);
4773 1034898 : for (i = 0; i < length; i++)
4774 : {
4775 568724 : es->param[i].change_prob = streamer_read_uhwi (ib);
4776 568724 : bitpack_d bp = streamer_read_bitpack (ib);
4777 1706172 : es->param[i].points_to_local_or_readonly_memory
4778 568724 : = bp_unpack_value (&bp, 1);
4779 1706172 : es->param[i].points_to_possible_sra_candidate
4780 568724 : = bp_unpack_value (&bp, 1);
4781 : }
4782 : }
4783 : else
4784 : {
4785 138294 : for (i = 0; i < length; i++)
4786 : {
4787 23009 : streamer_read_uhwi (ib);
4788 23009 : streamer_read_uhwi (ib);
4789 : }
4790 : }
4791 348372 : }
4792 :
4793 :
4794 : /* Stream in inline summaries from the section. */
4795 :
4796 : static void
4797 13560 : inline_read_section (struct lto_file_decl_data *file_data, const char *data,
4798 : size_t len)
4799 : {
4800 13560 : const struct lto_function_header *header =
4801 : (const struct lto_function_header *) data;
4802 13560 : const int cfg_offset = sizeof (struct lto_function_header);
4803 13560 : const int main_offset = cfg_offset + header->cfg_size;
4804 13560 : const int string_offset = main_offset + header->main_size;
4805 13560 : class data_in *data_in;
4806 13560 : unsigned int i, count2, j;
4807 13560 : unsigned int f_count;
4808 :
4809 13560 : lto_input_block ib ((const char *) data + main_offset, header->main_size,
4810 13560 : file_data);
4811 :
4812 13560 : data_in =
4813 27120 : lto_data_in_create (file_data, (const char *) data + string_offset,
4814 13560 : header->string_size, vNULL);
4815 13560 : f_count = streamer_read_uhwi (&ib);
4816 100664 : for (i = 0; i < f_count; i++)
4817 : {
4818 87104 : unsigned int index;
4819 87104 : struct cgraph_node *node;
4820 87104 : class ipa_fn_summary *info;
4821 87104 : class ipa_node_params *params_summary;
4822 87104 : class ipa_size_summary *size_info;
4823 87104 : lto_symtab_encoder_t encoder;
4824 87104 : struct bitpack_d bp;
4825 87104 : struct cgraph_edge *e;
4826 87104 : ipa_predicate p;
4827 :
4828 87104 : index = streamer_read_uhwi (&ib);
4829 87104 : encoder = file_data->symtab_node_encoder;
4830 87104 : node = dyn_cast<cgraph_node *> (lto_symtab_encoder_deref (encoder,
4831 : index));
4832 87104 : info = node->prevailing_p () ? ipa_fn_summaries->get_create (node) : NULL;
4833 87104 : params_summary = node->prevailing_p ()
4834 87104 : ? ipa_node_params_sum->get (node) : NULL;
4835 87104 : size_info = node->prevailing_p ()
4836 87104 : ? ipa_size_summaries->get_create (node) : NULL;
4837 :
4838 87104 : int stack_size = streamer_read_uhwi (&ib);
4839 87104 : int size = streamer_read_uhwi (&ib);
4840 87104 : sreal time = sreal::stream_in (&ib);
4841 :
4842 87104 : if (info)
4843 : {
4844 87046 : info->estimated_stack_size
4845 87046 : = size_info->estimated_self_stack_size = stack_size;
4846 87046 : size_info->size = size_info->self_size = size;
4847 87046 : info->time = time;
4848 : }
4849 :
4850 87104 : bp = streamer_read_bitpack (&ib);
4851 87104 : if (info)
4852 : {
4853 87046 : info->inlinable = bp_unpack_value (&bp, 1);
4854 87046 : info->fp_expressions = bp_unpack_value (&bp, 1);
4855 87046 : if (!lto_stream_offload_p)
4856 87046 : info->target_info = streamer_read_uhwi (&ib);
4857 : }
4858 : else
4859 : {
4860 58 : bp_unpack_value (&bp, 1);
4861 58 : bp_unpack_value (&bp, 1);
4862 58 : if (!lto_stream_offload_p)
4863 58 : streamer_read_uhwi (&ib);
4864 : }
4865 :
4866 87104 : count2 = streamer_read_uhwi (&ib);
4867 87104 : gcc_assert (!info || !info->conds);
4868 87046 : if (info)
4869 87046 : vec_safe_reserve_exact (info->conds, count2);
4870 165205 : for (j = 0; j < count2; j++)
4871 : {
4872 78101 : struct condition c;
4873 78101 : unsigned int k, count3;
4874 78101 : c.operand_num = streamer_read_uhwi (&ib);
4875 78101 : c.code = (enum tree_code) streamer_read_uhwi (&ib);
4876 78101 : c.type = stream_read_tree (&ib, data_in);
4877 78101 : c.val = stream_read_tree (&ib, data_in);
4878 78101 : bp = streamer_read_bitpack (&ib);
4879 78101 : c.agg_contents = bp_unpack_value (&bp, 1);
4880 78101 : c.by_ref = bp_unpack_value (&bp, 1);
4881 78101 : if (c.agg_contents)
4882 12554 : c.offset = streamer_read_uhwi (&ib);
4883 78101 : count3 = streamer_read_uhwi (&ib);
4884 78101 : c.param_ops = NULL;
4885 78101 : if (info)
4886 78101 : vec_safe_reserve_exact (c.param_ops, count3);
4887 78101 : if (params_summary)
4888 78101 : ipa_set_param_used_by_ipa_predicates
4889 78101 : (params_summary, c.operand_num, true);
4890 81844 : for (k = 0; k < count3; k++)
4891 : {
4892 3743 : struct expr_eval_op op;
4893 3743 : enum gimple_rhs_class rhs_class;
4894 3743 : op.code = (enum tree_code) streamer_read_uhwi (&ib);
4895 3743 : op.type = stream_read_tree (&ib, data_in);
4896 3743 : switch (rhs_class = get_gimple_rhs_class (op.code))
4897 : {
4898 1562 : case GIMPLE_UNARY_RHS:
4899 1562 : op.index = 0;
4900 1562 : op.val[0] = NULL_TREE;
4901 1562 : op.val[1] = NULL_TREE;
4902 1562 : break;
4903 :
4904 2181 : case GIMPLE_BINARY_RHS:
4905 2181 : case GIMPLE_TERNARY_RHS:
4906 2181 : bp = streamer_read_bitpack (&ib);
4907 2181 : op.index = bp_unpack_value (&bp, 2);
4908 2181 : op.val[0] = stream_read_tree (&ib, data_in);
4909 2181 : if (rhs_class == GIMPLE_BINARY_RHS)
4910 2181 : op.val[1] = NULL_TREE;
4911 : else
4912 0 : op.val[1] = stream_read_tree (&ib, data_in);
4913 : break;
4914 :
4915 0 : default:
4916 0 : fatal_error (UNKNOWN_LOCATION,
4917 : "invalid fnsummary in LTO stream");
4918 : }
4919 3743 : if (info)
4920 3743 : c.param_ops->quick_push (op);
4921 : }
4922 78101 : if (info)
4923 78101 : info->conds->quick_push (c);
4924 : }
4925 87104 : count2 = streamer_read_uhwi (&ib);
4926 87104 : gcc_assert (!info || !info->size_time_table.length ());
4927 87104 : if (info && count2)
4928 87046 : info->size_time_table.reserve_exact (count2);
4929 331102 : for (j = 0; j < count2; j++)
4930 : {
4931 243998 : class size_time_entry e;
4932 :
4933 243998 : e.size = streamer_read_uhwi (&ib);
4934 243998 : e.time = sreal::stream_in (&ib);
4935 243998 : e.exec_predicate.stream_in (&ib);
4936 243998 : e.nonconst_predicate.stream_in (&ib);
4937 :
4938 243998 : if (info)
4939 243882 : info->size_time_table.quick_push (e);
4940 : }
4941 :
4942 87104 : count2 = streamer_read_uhwi (&ib);
4943 88494 : for (j = 0; j < count2; j++)
4944 : {
4945 1390 : p.stream_in (&ib);
4946 1390 : sreal fcp_freq = sreal::stream_in (&ib);
4947 1390 : if (info)
4948 : {
4949 1390 : ipa_freqcounting_predicate fcp;
4950 1390 : fcp.predicate = NULL;
4951 1390 : set_hint_predicate (&fcp.predicate, p);
4952 1390 : fcp.freq = fcp_freq;
4953 1390 : vec_safe_push (info->loop_iterations, fcp);
4954 : }
4955 : }
4956 87104 : count2 = streamer_read_uhwi (&ib);
4957 87394 : for (j = 0; j < count2; j++)
4958 : {
4959 290 : p.stream_in (&ib);
4960 290 : sreal fcp_freq = sreal::stream_in (&ib);
4961 290 : if (info)
4962 : {
4963 290 : ipa_freqcounting_predicate fcp;
4964 290 : fcp.predicate = NULL;
4965 290 : set_hint_predicate (&fcp.predicate, p);
4966 290 : fcp.freq = fcp_freq;
4967 290 : vec_safe_push (info->loop_strides, fcp);
4968 : }
4969 : }
4970 87104 : count2 = streamer_read_uhwi (&ib);
4971 87104 : if (info && count2)
4972 8 : info->builtin_constant_p_parms.reserve_exact (count2);
4973 87112 : for (j = 0; j < count2; j++)
4974 : {
4975 8 : int parm = streamer_read_uhwi (&ib);
4976 8 : if (info)
4977 8 : info->builtin_constant_p_parms.quick_push (parm);
4978 : }
4979 433981 : for (e = node->callees; e; e = e->next_callee)
4980 346877 : read_ipa_call_summary (&ib, e, info != NULL);
4981 88599 : for (e = node->indirect_calls; e; e = e->next_callee)
4982 1495 : read_ipa_call_summary (&ib, e, info != NULL);
4983 : }
4984 :
4985 13560 : lto_free_section_data (file_data, LTO_section_ipa_fn_summary, NULL, data,
4986 : len);
4987 13560 : lto_data_in_delete (data_in);
4988 13560 : }
4989 :
4990 :
4991 : /* Read inline summary. Jump functions are shared among ipa-cp
4992 : and inliner, so when ipa-cp is active, we don't need to write them
4993 : twice. */
4994 :
4995 : static void
4996 12468 : ipa_fn_summary_read (void)
4997 : {
4998 12468 : struct lto_file_decl_data **file_data_vec = lto_get_file_decl_data ();
4999 12468 : struct lto_file_decl_data *file_data;
5000 12468 : unsigned int j = 0;
5001 :
5002 12468 : ipa_prop_read_jump_functions ();
5003 12468 : ipa_fn_summary_alloc ();
5004 :
5005 38496 : while ((file_data = file_data_vec[j++]))
5006 : {
5007 13560 : size_t len;
5008 13560 : const char *data
5009 13560 : = lto_get_summary_section_data (file_data, LTO_section_ipa_fn_summary,
5010 : &len);
5011 13560 : if (data)
5012 13560 : inline_read_section (file_data, data, len);
5013 : else
5014 : /* Fatal error here. We do not want to support compiling ltrans units
5015 : with different version of compiler or different flags than the WPA
5016 : unit, so this should never happen. */
5017 0 : fatal_error (input_location,
5018 : "ipa inline summary is missing in input file");
5019 : }
5020 12468 : ipa_register_cgraph_hooks ();
5021 :
5022 12468 : gcc_assert (ipa_fn_summaries);
5023 12468 : ipa_fn_summaries->enable_insertion_hook ();
5024 12468 : }
5025 :
5026 :
5027 : /* Write inline summary for edge E to OB. */
5028 :
5029 : static void
5030 380101 : write_ipa_call_summary (struct output_block *ob, struct cgraph_edge *e)
5031 : {
5032 380101 : class ipa_call_summary *es = ipa_call_summaries->get (e);
5033 380101 : int i;
5034 :
5035 380101 : streamer_write_uhwi (ob, es->call_stmt_size);
5036 380101 : streamer_write_uhwi (ob, es->call_stmt_time);
5037 380101 : streamer_write_uhwi (ob, es->loop_depth);
5038 :
5039 380101 : bitpack_d bp = bitpack_create (ob->main_stream);
5040 380101 : bp_pack_value (&bp, es->is_return_callee_uncaptured, 1);
5041 380101 : streamer_write_bitpack (&bp);
5042 :
5043 380101 : if (es->predicate)
5044 10753 : es->predicate->stream_out (ob);
5045 : else
5046 369348 : streamer_write_uhwi (ob, 0);
5047 380101 : streamer_write_uhwi (ob, es->param.length ());
5048 2294106 : for (i = 0; i < (int) es->param.length (); i++)
5049 : {
5050 632482 : streamer_write_uhwi (ob, es->param[i].change_prob);
5051 632482 : bp = bitpack_create (ob->main_stream);
5052 632482 : bp_pack_value (&bp, es->param[i].points_to_local_or_readonly_memory, 1);
5053 632482 : bp_pack_value (&bp, es->param[i].points_to_possible_sra_candidate, 1);
5054 632482 : streamer_write_bitpack (&bp);
5055 : }
5056 380101 : }
5057 :
5058 :
5059 : /* Write inline summary for node in SET.
5060 : Jump functions are shared among ipa-cp and inliner, so when ipa-cp is
5061 : active, we don't need to write them twice. */
5062 :
5063 : static void
5064 23539 : ipa_fn_summary_write (void)
5065 : {
5066 23539 : struct output_block *ob = create_output_block (LTO_section_ipa_fn_summary);
5067 23539 : lto_symtab_encoder_iterator lsei;
5068 23539 : lto_symtab_encoder_t encoder = ob->decl_state->symtab_node_encoder;
5069 23539 : unsigned int count = 0;
5070 :
5071 129746 : for (lsei = lsei_start_function_in_partition (encoder); !lsei_end_p (lsei);
5072 106207 : lsei_next_function_in_partition (&lsei))
5073 : {
5074 106207 : cgraph_node *cnode = lsei_cgraph_node (lsei);
5075 106207 : if (cnode->definition && !cnode->alias)
5076 103739 : count++;
5077 : }
5078 23539 : streamer_write_uhwi (ob, count);
5079 :
5080 129746 : for (lsei = lsei_start_function_in_partition (encoder); !lsei_end_p (lsei);
5081 106207 : lsei_next_function_in_partition (&lsei))
5082 : {
5083 106207 : cgraph_node *cnode = lsei_cgraph_node (lsei);
5084 106207 : if (cnode->definition && !cnode->alias)
5085 : {
5086 103739 : class ipa_fn_summary *info = ipa_fn_summaries->get (cnode);
5087 103739 : class ipa_size_summary *size_info = ipa_size_summaries->get (cnode);
5088 103739 : struct bitpack_d bp;
5089 103739 : struct cgraph_edge *edge;
5090 103739 : int i;
5091 103739 : size_time_entry *e;
5092 103739 : struct condition *c;
5093 :
5094 103739 : streamer_write_uhwi (ob, lto_symtab_encoder_encode (encoder, cnode));
5095 103739 : streamer_write_hwi (ob, size_info->estimated_self_stack_size);
5096 103739 : streamer_write_hwi (ob, size_info->self_size);
5097 103739 : info->time.stream_out (ob);
5098 103739 : bp = bitpack_create (ob->main_stream);
5099 103739 : bp_pack_value (&bp, info->inlinable, 1);
5100 103739 : bp_pack_value (&bp, info->fp_expressions, 1);
5101 103739 : streamer_write_bitpack (&bp);
5102 103739 : if (!lto_stream_offload_p)
5103 103739 : streamer_write_uhwi (ob, info->target_info);
5104 103739 : streamer_write_uhwi (ob, vec_safe_length (info->conds));
5105 301692 : for (i = 0; vec_safe_iterate (info->conds, i, &c); i++)
5106 : {
5107 94214 : int j;
5108 94214 : struct expr_eval_op *op;
5109 :
5110 94214 : streamer_write_uhwi (ob, c->operand_num);
5111 94214 : streamer_write_uhwi (ob, c->code);
5112 94214 : stream_write_tree (ob, c->type, true);
5113 94214 : stream_write_tree (ob, c->val, true);
5114 94214 : bp = bitpack_create (ob->main_stream);
5115 94214 : bp_pack_value (&bp, c->agg_contents, 1);
5116 94214 : bp_pack_value (&bp, c->by_ref, 1);
5117 94214 : streamer_write_bitpack (&bp);
5118 94214 : if (c->agg_contents)
5119 15813 : streamer_write_uhwi (ob, c->offset);
5120 94214 : streamer_write_uhwi (ob, vec_safe_length (c->param_ops));
5121 195894 : for (j = 0; vec_safe_iterate (c->param_ops, j, &op); j++)
5122 : {
5123 4415 : streamer_write_uhwi (ob, op->code);
5124 4415 : stream_write_tree (ob, op->type, true);
5125 4415 : if (op->val[0])
5126 : {
5127 2579 : bp = bitpack_create (ob->main_stream);
5128 2579 : bp_pack_value (&bp, op->index, 2);
5129 2579 : streamer_write_bitpack (&bp);
5130 2579 : stream_write_tree (ob, op->val[0], true);
5131 2579 : if (op->val[1])
5132 4 : stream_write_tree (ob, op->val[1], true);
5133 : }
5134 : }
5135 : }
5136 103739 : streamer_write_uhwi (ob, info->size_time_table.length ());
5137 501773 : for (i = 0; info->size_time_table.iterate (i, &e); i++)
5138 : {
5139 294295 : streamer_write_uhwi (ob, e->size);
5140 294295 : e->time.stream_out (ob);
5141 294295 : e->exec_predicate.stream_out (ob);
5142 294295 : e->nonconst_predicate.stream_out (ob);
5143 : }
5144 103739 : ipa_freqcounting_predicate *fcp;
5145 103739 : streamer_write_uhwi (ob, vec_safe_length (info->loop_iterations));
5146 209515 : for (i = 0; vec_safe_iterate (info->loop_iterations, i, &fcp); i++)
5147 : {
5148 2037 : fcp->predicate->stream_out (ob);
5149 2037 : fcp->freq.stream_out (ob);
5150 : }
5151 103739 : streamer_write_uhwi (ob, vec_safe_length (info->loop_strides));
5152 207824 : for (i = 0; vec_safe_iterate (info->loop_strides, i, &fcp); i++)
5153 : {
5154 346 : fcp->predicate->stream_out (ob);
5155 346 : fcp->freq.stream_out (ob);
5156 : }
5157 103739 : streamer_write_uhwi (ob, info->builtin_constant_p_parms.length ());
5158 103739 : int ip;
5159 207494 : for (i = 0; info->builtin_constant_p_parms.iterate (i, &ip);
5160 : i++)
5161 16 : streamer_write_uhwi (ob, ip);
5162 481137 : for (edge = cnode->callees; edge; edge = edge->next_callee)
5163 377398 : write_ipa_call_summary (ob, edge);
5164 106442 : for (edge = cnode->indirect_calls; edge; edge = edge->next_callee)
5165 2703 : write_ipa_call_summary (ob, edge);
5166 : }
5167 : }
5168 23539 : streamer_write_char_stream (ob->main_stream, 0);
5169 23539 : produce_asm (ob);
5170 23539 : destroy_output_block (ob);
5171 :
5172 23539 : ipa_prop_write_jump_functions ();
5173 23539 : }
5174 :
5175 :
5176 : /* Release function summary. */
5177 :
5178 : void
5179 739834 : ipa_free_fn_summary (void)
5180 : {
5181 739834 : if (!ipa_call_summaries)
5182 : return;
5183 468116 : ggc_delete (ipa_fn_summaries);
5184 468116 : ipa_fn_summaries = NULL;
5185 468116 : delete ipa_call_summaries;
5186 468116 : ipa_call_summaries = NULL;
5187 468116 : edge_predicate_pool.release ();
5188 : /* During IPA this is one of largest datastructures to release. */
5189 468116 : if (flag_wpa)
5190 7948 : ggc_trim ();
5191 : }
5192 :
5193 : /* Release function summary. */
5194 :
5195 : void
5196 739834 : ipa_free_size_summary (void)
5197 : {
5198 739834 : if (!ipa_size_summaries)
5199 : return;
5200 468116 : delete ipa_size_summaries;
5201 468116 : ipa_size_summaries = NULL;
5202 : }
5203 :
5204 : namespace {
5205 :
5206 : const pass_data pass_data_local_fn_summary =
5207 : {
5208 : GIMPLE_PASS, /* type */
5209 : "local-fnsummary", /* name */
5210 : OPTGROUP_INLINE, /* optinfo_flags */
5211 : TV_INLINE_PARAMETERS, /* tv_id */
5212 : 0, /* properties_required */
5213 : 0, /* properties_provided */
5214 : 0, /* properties_destroyed */
5215 : 0, /* todo_flags_start */
5216 : 0, /* todo_flags_finish */
5217 : };
5218 :
5219 : class pass_local_fn_summary : public gimple_opt_pass
5220 : {
5221 : public:
5222 589174 : pass_local_fn_summary (gcc::context *ctxt)
5223 1178348 : : gimple_opt_pass (pass_data_local_fn_summary, ctxt)
5224 : {}
5225 :
5226 : /* opt_pass methods: */
5227 294587 : opt_pass * clone () final override
5228 : {
5229 294587 : return new pass_local_fn_summary (m_ctxt);
5230 : }
5231 6002610 : unsigned int execute (function *) final override
5232 : {
5233 6002610 : return compute_fn_summary_for_current ();
5234 : }
5235 :
5236 : }; // class pass_local_fn_summary
5237 :
5238 : } // anon namespace
5239 :
5240 : gimple_opt_pass *
5241 294587 : make_pass_local_fn_summary (gcc::context *ctxt)
5242 : {
5243 294587 : return new pass_local_fn_summary (ctxt);
5244 : }
5245 :
5246 :
5247 : /* Free inline summary. */
5248 :
5249 : namespace {
5250 :
5251 : const pass_data pass_data_ipa_free_fn_summary =
5252 : {
5253 : SIMPLE_IPA_PASS, /* type */
5254 : "free-fnsummary", /* name */
5255 : OPTGROUP_NONE, /* optinfo_flags */
5256 : TV_IPA_FREE_INLINE_SUMMARY, /* tv_id */
5257 : 0, /* properties_required */
5258 : 0, /* properties_provided */
5259 : 0, /* properties_destroyed */
5260 : 0, /* todo_flags_start */
5261 : 0, /* todo_flags_finish */
5262 : };
5263 :
5264 : class pass_ipa_free_fn_summary : public simple_ipa_opt_pass
5265 : {
5266 : public:
5267 589174 : pass_ipa_free_fn_summary (gcc::context *ctxt)
5268 : : simple_ipa_opt_pass (pass_data_ipa_free_fn_summary, ctxt),
5269 1178348 : small_p (false)
5270 : {}
5271 :
5272 : /* opt_pass methods: */
5273 294587 : opt_pass *clone () final override
5274 : {
5275 294587 : return new pass_ipa_free_fn_summary (m_ctxt);
5276 : }
5277 589174 : void set_pass_param (unsigned int n, bool param) final override
5278 : {
5279 589174 : gcc_assert (n == 0);
5280 589174 : small_p = param;
5281 589174 : }
5282 495204 : bool gate (function *) final override { return true; }
5283 474393 : unsigned int execute (function *) final override
5284 : {
5285 474393 : ipa_free_fn_summary ();
5286 : /* Free ipa-prop structures if they are no longer needed. */
5287 474393 : ipa_free_all_structures_after_iinln ();
5288 474393 : if (!flag_wpa)
5289 466445 : ipa_free_size_summary ();
5290 474393 : return 0;
5291 : }
5292 :
5293 : private:
5294 : bool small_p;
5295 : }; // class pass_ipa_free_fn_summary
5296 :
5297 : } // anon namespace
5298 :
5299 : simple_ipa_opt_pass *
5300 294587 : make_pass_ipa_free_fn_summary (gcc::context *ctxt)
5301 : {
5302 294587 : return new pass_ipa_free_fn_summary (ctxt);
5303 : }
5304 :
5305 : namespace {
5306 :
5307 : const pass_data pass_data_ipa_fn_summary =
5308 : {
5309 : IPA_PASS, /* type */
5310 : "fnsummary", /* name */
5311 : OPTGROUP_INLINE, /* optinfo_flags */
5312 : TV_IPA_FNSUMMARY, /* tv_id */
5313 : 0, /* properties_required */
5314 : 0, /* properties_provided */
5315 : 0, /* properties_destroyed */
5316 : 0, /* todo_flags_start */
5317 : ( TODO_dump_symtab ), /* todo_flags_finish */
5318 : };
5319 :
5320 : class pass_ipa_fn_summary : public ipa_opt_pass_d
5321 : {
5322 : public:
5323 294587 : pass_ipa_fn_summary (gcc::context *ctxt)
5324 : : ipa_opt_pass_d (pass_data_ipa_fn_summary, ctxt,
5325 : ipa_fn_summary_generate, /* generate_summary */
5326 : ipa_fn_summary_write, /* write_summary */
5327 : ipa_fn_summary_read, /* read_summary */
5328 : NULL, /* write_optimization_summary */
5329 : NULL, /* read_optimization_summary */
5330 : NULL, /* stmt_fixup */
5331 : 0, /* function_transform_todo_flags_start */
5332 : NULL, /* function_transform */
5333 294587 : NULL) /* variable_transform */
5334 294587 : {}
5335 :
5336 : /* opt_pass methods: */
5337 237284 : unsigned int execute (function *) final override { return 0; }
5338 :
5339 : }; // class pass_ipa_fn_summary
5340 :
5341 : } // anon namespace
5342 :
5343 : ipa_opt_pass_d *
5344 294587 : make_pass_ipa_fn_summary (gcc::context *ctxt)
5345 : {
5346 294587 : return new pass_ipa_fn_summary (ctxt);
5347 : }
5348 :
5349 : /* Reset all state within ipa-fnsummary.cc so that we can rerun the compiler
5350 : within the same process. For use by toplev::finalize. */
5351 :
5352 : void
5353 264541 : ipa_fnsummary_cc_finalize (void)
5354 : {
5355 264541 : ipa_free_fn_summary ();
5356 264541 : ipa_free_size_summary ();
5357 264541 : }
|
