Add testcase of PR c++/92542, already fixed.
[official-gcc.git] / gcc / ipa-fnsummary.c
blobdbd53f12a405cf1c8a045a2e979311f8411c46b2
1 /* Function summary pass.
2 Copyright (C) 2003-2020 Free Software Foundation, Inc.
3 Contributed by Jan Hubicka
5 This file is part of GCC.
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.
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.
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/>. */
21 /* Analysis of function bodies used by inter-procedural passes
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
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).
35 We provide access to the ipa_fn_summary data structure and
36 basic logic updating the parameters when inlining is performed.
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.
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.
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. */
54 #include "config.h"
55 #include "system.h"
56 #include "coretypes.h"
57 #include "backend.h"
58 #include "tree.h"
59 #include "gimple.h"
60 #include "alloc-pool.h"
61 #include "tree-pass.h"
62 #include "ssa.h"
63 #include "tree-streamer.h"
64 #include "cgraph.h"
65 #include "diagnostic.h"
66 #include "fold-const.h"
67 #include "print-tree.h"
68 #include "tree-inline.h"
69 #include "gimple-pretty-print.h"
70 #include "cfganal.h"
71 #include "gimple-iterator.h"
72 #include "tree-cfg.h"
73 #include "tree-ssa-loop-niter.h"
74 #include "tree-ssa-loop.h"
75 #include "symbol-summary.h"
76 #include "ipa-prop.h"
77 #include "ipa-fnsummary.h"
78 #include "cfgloop.h"
79 #include "tree-scalar-evolution.h"
80 #include "ipa-utils.h"
81 #include "cfgexpand.h"
82 #include "gimplify.h"
83 #include "stringpool.h"
84 #include "attribs.h"
85 #include "tree-into-ssa.h"
87 /* Summaries. */
88 fast_function_summary <ipa_fn_summary *, va_gc> *ipa_fn_summaries;
89 fast_function_summary <ipa_size_summary *, va_heap> *ipa_size_summaries;
90 fast_call_summary <ipa_call_summary *, va_heap> *ipa_call_summaries;
92 /* Edge predicates goes here. */
93 static object_allocator<predicate> edge_predicate_pool ("edge predicates");
96 /* Dump IPA hints. */
97 void
98 ipa_dump_hints (FILE *f, ipa_hints hints)
100 if (!hints)
101 return;
102 fprintf (f, "IPA hints:");
103 if (hints & INLINE_HINT_indirect_call)
105 hints &= ~INLINE_HINT_indirect_call;
106 fprintf (f, " indirect_call");
108 if (hints & INLINE_HINT_loop_iterations)
110 hints &= ~INLINE_HINT_loop_iterations;
111 fprintf (f, " loop_iterations");
113 if (hints & INLINE_HINT_loop_stride)
115 hints &= ~INLINE_HINT_loop_stride;
116 fprintf (f, " loop_stride");
118 if (hints & INLINE_HINT_same_scc)
120 hints &= ~INLINE_HINT_same_scc;
121 fprintf (f, " same_scc");
123 if (hints & INLINE_HINT_in_scc)
125 hints &= ~INLINE_HINT_in_scc;
126 fprintf (f, " in_scc");
128 if (hints & INLINE_HINT_cross_module)
130 hints &= ~INLINE_HINT_cross_module;
131 fprintf (f, " cross_module");
133 if (hints & INLINE_HINT_declared_inline)
135 hints &= ~INLINE_HINT_declared_inline;
136 fprintf (f, " declared_inline");
138 if (hints & INLINE_HINT_known_hot)
140 hints &= ~INLINE_HINT_known_hot;
141 fprintf (f, " known_hot");
143 gcc_assert (!hints);
147 /* Record SIZE and TIME to SUMMARY.
148 The accounted code will be executed when EXEC_PRED is true.
149 When NONCONST_PRED is false the code will evaluate to constant and
150 will get optimized out in specialized clones of the function.
151 If CALL is true account to call_size_time_table rather than
152 size_time_table. */
154 void
155 ipa_fn_summary::account_size_time (int size, sreal time,
156 const predicate &exec_pred,
157 const predicate &nonconst_pred_in,
158 bool call)
160 size_time_entry *e;
161 bool found = false;
162 int i;
163 predicate nonconst_pred;
164 vec<size_time_entry, va_gc> *table = call
165 ? call_size_time_table : size_time_table;
167 if (exec_pred == false)
168 return;
170 nonconst_pred = nonconst_pred_in & exec_pred;
172 if (nonconst_pred == false)
173 return;
175 /* We need to create initial empty unconditional clause, but otherwise
176 we don't need to account empty times and sizes. */
177 if (!size && time == 0 && table)
178 return;
180 /* Only for calls we are unaccounting what we previously recorded. */
181 gcc_checking_assert (time >= 0 || call);
183 for (i = 0; vec_safe_iterate (table, i, &e); i++)
184 if (e->exec_predicate == exec_pred
185 && e->nonconst_predicate == nonconst_pred)
187 found = true;
188 break;
190 if (i == max_size_time_table_size)
192 i = 0;
193 found = true;
194 e = &(*table)[0];
195 if (dump_file && (dump_flags & TDF_DETAILS))
196 fprintf (dump_file,
197 "\t\tReached limit on number of entries, "
198 "ignoring the predicate.");
200 if (dump_file && (dump_flags & TDF_DETAILS) && (time != 0 || size))
202 fprintf (dump_file,
203 "\t\tAccounting size:%3.2f, time:%3.2f on %spredicate exec:",
204 ((double) size) / ipa_fn_summary::size_scale,
205 (time.to_double ()), found ? "" : "new ");
206 exec_pred.dump (dump_file, conds, 0);
207 if (exec_pred != nonconst_pred)
209 fprintf (dump_file, " nonconst:");
210 nonconst_pred.dump (dump_file, conds);
212 else
213 fprintf (dump_file, "\n");
215 if (!found)
217 class size_time_entry new_entry;
218 new_entry.size = size;
219 new_entry.time = time;
220 new_entry.exec_predicate = exec_pred;
221 new_entry.nonconst_predicate = nonconst_pred;
222 if (call)
223 vec_safe_push (call_size_time_table, new_entry);
224 else
225 vec_safe_push (size_time_table, new_entry);
227 else
229 e->size += size;
230 e->time += time;
231 /* FIXME: PR bootstrap/92653 gcc_checking_assert (e->time >= -1); */
232 /* Tolerate small roundoff issues. */
233 if (e->time < 0)
234 e->time = 0;
238 /* We proved E to be unreachable, redirect it to __builtin_unreachable. */
240 static struct cgraph_edge *
241 redirect_to_unreachable (struct cgraph_edge *e)
243 struct cgraph_node *callee = !e->inline_failed ? e->callee : NULL;
244 struct cgraph_node *target = cgraph_node::get_create
245 (builtin_decl_implicit (BUILT_IN_UNREACHABLE));
247 if (e->speculative)
248 e = cgraph_edge::resolve_speculation (e, target->decl);
249 else if (!e->callee)
250 e = cgraph_edge::make_direct (e, target);
251 else
252 e->redirect_callee (target);
253 class ipa_call_summary *es = ipa_call_summaries->get (e);
254 e->inline_failed = CIF_UNREACHABLE;
255 e->count = profile_count::zero ();
256 es->call_stmt_size = 0;
257 es->call_stmt_time = 0;
258 if (callee)
259 callee->remove_symbol_and_inline_clones ();
260 return e;
263 /* Set predicate for edge E. */
265 static void
266 edge_set_predicate (struct cgraph_edge *e, predicate *predicate)
268 /* If the edge is determined to be never executed, redirect it
269 to BUILTIN_UNREACHABLE to make it clear to IPA passes the call will
270 be optimized out. */
271 if (predicate && *predicate == false
272 /* When handling speculative edges, we need to do the redirection
273 just once. Do it always on the direct edge, so we do not
274 attempt to resolve speculation while duplicating the edge. */
275 && (!e->speculative || e->callee))
276 e = redirect_to_unreachable (e);
278 class ipa_call_summary *es = ipa_call_summaries->get (e);
279 if (predicate && *predicate != true)
281 if (!es->predicate)
282 es->predicate = edge_predicate_pool.allocate ();
283 *es->predicate = *predicate;
285 else
287 if (es->predicate)
288 edge_predicate_pool.remove (es->predicate);
289 es->predicate = NULL;
293 /* Set predicate for hint *P. */
295 static void
296 set_hint_predicate (predicate **p, predicate new_predicate)
298 if (new_predicate == false || new_predicate == true)
300 if (*p)
301 edge_predicate_pool.remove (*p);
302 *p = NULL;
304 else
306 if (!*p)
307 *p = edge_predicate_pool.allocate ();
308 **p = new_predicate;
313 /* Compute what conditions may or may not hold given information about
314 parameters. RET_CLAUSE returns truths that may hold in a specialized copy,
315 while RET_NONSPEC_CLAUSE returns truths that may hold in an nonspecialized
316 copy when called in a given context. It is a bitmask of conditions. Bit
317 0 means that condition is known to be false, while bit 1 means that condition
318 may or may not be true. These differs - for example NOT_INLINED condition
319 is always false in the second and also builtin_constant_p tests cannot use
320 the fact that parameter is indeed a constant.
322 KNOWN_VALS is partial mapping of parameters of NODE to constant values.
323 KNOWN_AGGS is a vector of aggregate known offset/value set for each
324 parameter. Return clause of possible truths. When INLINE_P is true, assume
325 that we are inlining.
327 ERROR_MARK means compile time invariant. */
329 static void
330 evaluate_conditions_for_known_args (struct cgraph_node *node,
331 bool inline_p,
332 vec<tree> known_vals,
333 vec<value_range> known_value_ranges,
334 vec<ipa_agg_value_set> known_aggs,
335 clause_t *ret_clause,
336 clause_t *ret_nonspec_clause)
338 clause_t clause = inline_p ? 0 : 1 << predicate::not_inlined_condition;
339 clause_t nonspec_clause = 1 << predicate::not_inlined_condition;
340 class ipa_fn_summary *info = ipa_fn_summaries->get (node);
341 int i;
342 struct condition *c;
344 for (i = 0; vec_safe_iterate (info->conds, i, &c); i++)
346 tree val = NULL;
347 tree res;
348 int j;
349 struct expr_eval_op *op;
351 /* We allow call stmt to have fewer arguments than the callee function
352 (especially for K&R style programs). So bound check here (we assume
353 known_aggs vector, if non-NULL, has the same length as
354 known_vals). */
355 gcc_checking_assert (!known_aggs.length () || !known_vals.length ()
356 || (known_vals.length () == known_aggs.length ()));
358 if (c->agg_contents)
360 struct ipa_agg_value_set *agg;
362 if (c->code == predicate::changed
363 && !c->by_ref
364 && c->operand_num < (int)known_vals.length ()
365 && (known_vals[c->operand_num] == error_mark_node))
366 continue;
368 if (c->operand_num < (int)known_aggs.length ())
370 agg = &known_aggs[c->operand_num];
371 val = ipa_find_agg_cst_for_param (agg,
372 c->operand_num
373 < (int) known_vals.length ()
374 ? known_vals[c->operand_num]
375 : NULL,
376 c->offset, c->by_ref);
378 else
379 val = NULL_TREE;
381 else if (c->operand_num < (int) known_vals.length ())
383 val = known_vals[c->operand_num];
384 if (val == error_mark_node && c->code != predicate::changed)
385 val = NULL_TREE;
388 if (!val
389 && (c->code == predicate::changed
390 || c->code == predicate::is_not_constant))
392 clause |= 1 << (i + predicate::first_dynamic_condition);
393 nonspec_clause |= 1 << (i + predicate::first_dynamic_condition);
394 continue;
396 if (c->code == predicate::changed)
398 nonspec_clause |= 1 << (i + predicate::first_dynamic_condition);
399 continue;
402 if (c->code == predicate::is_not_constant)
404 nonspec_clause |= 1 << (i + predicate::first_dynamic_condition);
405 continue;
408 if (val && TYPE_SIZE (c->type) == TYPE_SIZE (TREE_TYPE (val)))
410 if (c->type != TREE_TYPE (val))
411 val = fold_unary (VIEW_CONVERT_EXPR, c->type, val);
412 for (j = 0; vec_safe_iterate (c->param_ops, j, &op); j++)
414 if (!val)
415 break;
416 if (!op->val[0])
417 val = fold_unary (op->code, op->type, val);
418 else if (!op->val[1])
419 val = fold_binary (op->code, op->type,
420 op->index ? op->val[0] : val,
421 op->index ? val : op->val[0]);
422 else if (op->index == 0)
423 val = fold_ternary (op->code, op->type,
424 val, op->val[0], op->val[1]);
425 else if (op->index == 1)
426 val = fold_ternary (op->code, op->type,
427 op->val[0], val, op->val[1]);
428 else if (op->index == 2)
429 val = fold_ternary (op->code, op->type,
430 op->val[0], op->val[1], val);
431 else
432 val = NULL_TREE;
435 res = val
436 ? fold_binary_to_constant (c->code, boolean_type_node, val, c->val)
437 : NULL;
439 if (res && integer_zerop (res))
440 continue;
441 if (res && integer_onep (res))
443 clause |= 1 << (i + predicate::first_dynamic_condition);
444 nonspec_clause |= 1 << (i + predicate::first_dynamic_condition);
445 continue;
448 if (c->operand_num < (int) known_value_ranges.length ()
449 && !c->agg_contents
450 && !known_value_ranges[c->operand_num].undefined_p ()
451 && !known_value_ranges[c->operand_num].varying_p ()
452 && TYPE_SIZE (c->type)
453 == TYPE_SIZE (known_value_ranges[c->operand_num].type ())
454 && (!val || TREE_CODE (val) != INTEGER_CST))
456 value_range vr = known_value_ranges[c->operand_num];
457 if (!useless_type_conversion_p (c->type, vr.type ()))
459 value_range res;
460 range_fold_unary_expr (&res, NOP_EXPR,
461 c->type, &vr, vr.type ());
462 vr = res;
464 tree type = c->type;
466 for (j = 0; vec_safe_iterate (c->param_ops, j, &op); j++)
468 if (vr.varying_p () || vr.undefined_p ())
469 break;
471 value_range res;
472 if (!op->val[0])
473 range_fold_unary_expr (&res, op->code, op->type, &vr, type);
474 else if (!op->val[1])
476 value_range op0 (op->val[0], op->val[0]);
477 range_fold_binary_expr (&res, op->code, op->type,
478 op->index ? &op0 : &vr,
479 op->index ? &vr : &op0);
481 else
482 gcc_unreachable ();
483 type = op->type;
484 vr = res;
486 if (!vr.varying_p () && !vr.undefined_p ())
488 value_range res;
489 value_range val_vr (c->val, c->val);
490 range_fold_binary_expr (&res, c->code, boolean_type_node,
491 &vr,
492 &val_vr);
493 if (res.zero_p ())
494 continue;
498 clause |= 1 << (i + predicate::first_dynamic_condition);
499 nonspec_clause |= 1 << (i + predicate::first_dynamic_condition);
501 *ret_clause = clause;
502 if (ret_nonspec_clause)
503 *ret_nonspec_clause = nonspec_clause;
507 /* Work out what conditions might be true at invocation of E.
508 Compute costs for inlined edge if INLINE_P is true.
510 Return in CLAUSE_PTR the evaluated conditions and in NONSPEC_CLAUSE_PTR
511 (if non-NULL) conditions evaluated for nonspecialized clone called
512 in a given context.
514 KNOWN_VALS_PTR and KNOWN_AGGS_PTR must be non-NULL and will be filled by
515 known constant and aggregate values of parameters.
517 KNOWN_CONTEXT_PTR, if non-NULL, will be filled by polymorphic call contexts
518 of parameter used by a polymorphic call. */
520 void
521 evaluate_properties_for_edge (struct cgraph_edge *e, bool inline_p,
522 clause_t *clause_ptr,
523 clause_t *nonspec_clause_ptr,
524 vec<tree> *known_vals_ptr,
525 vec<ipa_polymorphic_call_context>
526 *known_contexts_ptr,
527 vec<ipa_agg_value_set> *known_aggs_ptr)
529 struct cgraph_node *callee = e->callee->ultimate_alias_target ();
530 class ipa_fn_summary *info = ipa_fn_summaries->get (callee);
531 auto_vec<value_range, 32> known_value_ranges;
532 class ipa_edge_args *args;
534 if (clause_ptr)
535 *clause_ptr = inline_p ? 0 : 1 << predicate::not_inlined_condition;
537 if (ipa_node_params_sum
538 && !e->call_stmt_cannot_inline_p
539 && (info->conds || known_contexts_ptr)
540 && (args = IPA_EDGE_REF (e)) != NULL)
542 struct cgraph_node *caller;
543 class ipa_node_params *caller_parms_info, *callee_pi = NULL;
544 class ipa_call_summary *es = ipa_call_summaries->get (e);
545 int i, count = ipa_get_cs_argument_count (args);
547 if (count)
549 if (e->caller->inlined_to)
550 caller = e->caller->inlined_to;
551 else
552 caller = e->caller;
553 caller_parms_info = IPA_NODE_REF (caller);
554 callee_pi = IPA_NODE_REF (callee);
556 /* Watch for thunks. */
557 if (callee_pi)
558 /* Watch for variadic functions. */
559 count = MIN (count, ipa_get_param_count (callee_pi));
562 if (callee_pi)
563 for (i = 0; i < count; i++)
565 struct ipa_jump_func *jf = ipa_get_ith_jump_func (args, i);
567 if (ipa_is_param_used_by_indirect_call (callee_pi, i)
568 || ipa_is_param_used_by_ipa_predicates (callee_pi, i))
570 /* Determine if we know constant value of the parameter. */
571 tree cst = ipa_value_from_jfunc (caller_parms_info, jf,
572 ipa_get_type (callee_pi, i));
574 if (!cst && e->call_stmt
575 && i < (int)gimple_call_num_args (e->call_stmt))
577 cst = gimple_call_arg (e->call_stmt, i);
578 if (!is_gimple_min_invariant (cst))
579 cst = NULL;
581 if (cst)
583 gcc_checking_assert (TREE_CODE (cst) != TREE_BINFO);
584 if (!known_vals_ptr->length ())
585 vec_safe_grow_cleared (known_vals_ptr, count);
586 (*known_vals_ptr)[i] = cst;
588 else if (inline_p && !es->param[i].change_prob)
590 if (!known_vals_ptr->length ())
591 vec_safe_grow_cleared (known_vals_ptr, count);
592 (*known_vals_ptr)[i] = error_mark_node;
595 /* If we failed to get simple constant, try value range. */
596 if ((!cst || TREE_CODE (cst) != INTEGER_CST)
597 && ipa_is_param_used_by_ipa_predicates (callee_pi, i))
599 value_range vr
600 = ipa_value_range_from_jfunc (caller_parms_info, e, jf,
601 ipa_get_type (callee_pi,
602 i));
603 if (!vr.undefined_p () && !vr.varying_p ())
605 if (!known_value_ranges.length ())
606 known_value_ranges.safe_grow_cleared (count);
607 known_value_ranges[i] = vr;
611 /* Determine known aggregate values. */
612 ipa_agg_value_set agg
613 = ipa_agg_value_set_from_jfunc (caller_parms_info,
614 caller, &jf->agg);
615 if (agg.items.length ())
617 if (!known_aggs_ptr->length ())
618 vec_safe_grow_cleared (known_aggs_ptr, count);
619 (*known_aggs_ptr)[i] = agg;
623 /* For calls used in polymorphic calls we further determine
624 polymorphic call context. */
625 if (known_contexts_ptr
626 && ipa_is_param_used_by_polymorphic_call (callee_pi, i))
628 ipa_polymorphic_call_context
629 ctx = ipa_context_from_jfunc (caller_parms_info, e, i, jf);
630 if (!ctx.useless_p ())
632 if (!known_contexts_ptr->length ())
633 known_contexts_ptr->safe_grow_cleared (count);
634 (*known_contexts_ptr)[i]
635 = ipa_context_from_jfunc (caller_parms_info, e, i, jf);
639 else
640 gcc_assert (!count || callee->thunk.thunk_p);
642 else if (e->call_stmt && !e->call_stmt_cannot_inline_p && info->conds)
644 int i, count = (int)gimple_call_num_args (e->call_stmt);
646 for (i = 0; i < count; i++)
648 tree cst = gimple_call_arg (e->call_stmt, i);
649 if (!is_gimple_min_invariant (cst))
650 cst = NULL;
651 if (cst)
653 if (!known_vals_ptr->length ())
654 vec_safe_grow_cleared (known_vals_ptr, count);
655 (*known_vals_ptr)[i] = cst;
660 evaluate_conditions_for_known_args (callee, inline_p,
661 *known_vals_ptr,
662 known_value_ranges,
663 *known_aggs_ptr,
664 clause_ptr,
665 nonspec_clause_ptr);
669 /* Allocate the function summary. */
671 static void
672 ipa_fn_summary_alloc (void)
674 gcc_checking_assert (!ipa_fn_summaries);
675 ipa_size_summaries = new ipa_size_summary_t (symtab);
676 ipa_fn_summaries = ipa_fn_summary_t::create_ggc (symtab);
677 ipa_call_summaries = new ipa_call_summary_t (symtab);
680 ipa_call_summary::~ipa_call_summary ()
682 if (predicate)
683 edge_predicate_pool.remove (predicate);
685 param.release ();
688 ipa_fn_summary::~ipa_fn_summary ()
690 if (loop_iterations)
691 edge_predicate_pool.remove (loop_iterations);
692 if (loop_stride)
693 edge_predicate_pool.remove (loop_stride);
694 vec_free (conds);
695 vec_free (size_time_table);
696 vec_free (call_size_time_table);
699 void
700 ipa_fn_summary_t::remove_callees (cgraph_node *node)
702 cgraph_edge *e;
703 for (e = node->callees; e; e = e->next_callee)
704 ipa_call_summaries->remove (e);
705 for (e = node->indirect_calls; e; e = e->next_callee)
706 ipa_call_summaries->remove (e);
709 /* Same as remap_predicate_after_duplication but handle hint predicate *P.
710 Additionally care about allocating new memory slot for updated predicate
711 and set it to NULL when it becomes true or false (and thus uninteresting).
714 static void
715 remap_hint_predicate_after_duplication (predicate **p,
716 clause_t possible_truths)
718 predicate new_predicate;
720 if (!*p)
721 return;
723 new_predicate = (*p)->remap_after_duplication (possible_truths);
724 /* We do not want to free previous predicate; it is used by node origin. */
725 *p = NULL;
726 set_hint_predicate (p, new_predicate);
730 /* Hook that is called by cgraph.c when a node is duplicated. */
731 void
732 ipa_fn_summary_t::duplicate (cgraph_node *src,
733 cgraph_node *dst,
734 ipa_fn_summary *,
735 ipa_fn_summary *info)
737 new (info) ipa_fn_summary (*ipa_fn_summaries->get (src));
738 /* TODO: as an optimization, we may avoid copying conditions
739 that are known to be false or true. */
740 info->conds = vec_safe_copy (info->conds);
742 /* When there are any replacements in the function body, see if we can figure
743 out that something was optimized out. */
744 if (ipa_node_params_sum && dst->clone.tree_map)
746 vec<size_time_entry, va_gc> *entry = info->size_time_table;
747 /* Use SRC parm info since it may not be copied yet. */
748 class ipa_node_params *parms_info = IPA_NODE_REF (src);
749 vec<tree> known_vals = vNULL;
750 int count = ipa_get_param_count (parms_info);
751 int i, j;
752 clause_t possible_truths;
753 predicate true_pred = true;
754 size_time_entry *e;
755 int optimized_out_size = 0;
756 bool inlined_to_p = false;
757 struct cgraph_edge *edge, *next;
759 info->size_time_table = 0;
760 known_vals.safe_grow_cleared (count);
761 for (i = 0; i < count; i++)
763 struct ipa_replace_map *r;
765 for (j = 0; vec_safe_iterate (dst->clone.tree_map, j, &r); j++)
767 if (r->parm_num == i)
769 known_vals[i] = r->new_tree;
770 break;
774 evaluate_conditions_for_known_args (dst, false,
775 known_vals,
776 vNULL,
777 vNULL,
778 &possible_truths,
779 /* We are going to specialize,
780 so ignore nonspec truths. */
781 NULL);
782 known_vals.release ();
784 info->account_size_time (0, 0, true_pred, true_pred);
786 /* Remap size_time vectors.
787 Simplify the predicate by pruning out alternatives that are known
788 to be false.
789 TODO: as on optimization, we can also eliminate conditions known
790 to be true. */
791 for (i = 0; vec_safe_iterate (entry, i, &e); i++)
793 predicate new_exec_pred;
794 predicate new_nonconst_pred;
795 new_exec_pred = e->exec_predicate.remap_after_duplication
796 (possible_truths);
797 new_nonconst_pred = e->nonconst_predicate.remap_after_duplication
798 (possible_truths);
799 if (new_exec_pred == false || new_nonconst_pred == false)
800 optimized_out_size += e->size;
801 else
802 info->account_size_time (e->size, e->time, new_exec_pred,
803 new_nonconst_pred);
806 /* Remap edge predicates with the same simplification as above.
807 Also copy constantness arrays. */
808 for (edge = dst->callees; edge; edge = next)
810 predicate new_predicate;
811 class ipa_call_summary *es = ipa_call_summaries->get (edge);
812 next = edge->next_callee;
814 if (!edge->inline_failed)
815 inlined_to_p = true;
816 if (!es->predicate)
817 continue;
818 new_predicate = es->predicate->remap_after_duplication
819 (possible_truths);
820 if (new_predicate == false && *es->predicate != false)
821 optimized_out_size += es->call_stmt_size * ipa_fn_summary::size_scale;
822 edge_set_predicate (edge, &new_predicate);
825 /* Remap indirect edge predicates with the same simplification as above.
826 Also copy constantness arrays. */
827 for (edge = dst->indirect_calls; edge; edge = next)
829 predicate new_predicate;
830 class ipa_call_summary *es = ipa_call_summaries->get (edge);
831 next = edge->next_callee;
833 gcc_checking_assert (edge->inline_failed);
834 if (!es->predicate)
835 continue;
836 new_predicate = es->predicate->remap_after_duplication
837 (possible_truths);
838 if (new_predicate == false && *es->predicate != false)
839 optimized_out_size += es->call_stmt_size * ipa_fn_summary::size_scale;
840 edge_set_predicate (edge, &new_predicate);
842 remap_hint_predicate_after_duplication (&info->loop_iterations,
843 possible_truths);
844 remap_hint_predicate_after_duplication (&info->loop_stride,
845 possible_truths);
847 /* If inliner or someone after inliner will ever start producing
848 non-trivial clones, we will get trouble with lack of information
849 about updating self sizes, because size vectors already contains
850 sizes of the callees. */
851 gcc_assert (!inlined_to_p || !optimized_out_size);
853 else
855 info->size_time_table = vec_safe_copy (info->size_time_table);
856 if (info->loop_iterations)
858 predicate p = *info->loop_iterations;
859 info->loop_iterations = NULL;
860 set_hint_predicate (&info->loop_iterations, p);
862 if (info->loop_stride)
864 predicate p = *info->loop_stride;
865 info->loop_stride = NULL;
866 set_hint_predicate (&info->loop_stride, p);
869 if (!dst->inlined_to)
870 ipa_update_overall_fn_summary (dst);
874 /* Hook that is called by cgraph.c when a node is duplicated. */
876 void
877 ipa_call_summary_t::duplicate (struct cgraph_edge *src,
878 struct cgraph_edge *dst,
879 class ipa_call_summary *srcinfo,
880 class ipa_call_summary *info)
882 new (info) ipa_call_summary (*srcinfo);
883 info->predicate = NULL;
884 edge_set_predicate (dst, srcinfo->predicate);
885 info->param = srcinfo->param.copy ();
886 if (!dst->indirect_unknown_callee && src->indirect_unknown_callee)
888 info->call_stmt_size -= (eni_size_weights.indirect_call_cost
889 - eni_size_weights.call_cost);
890 info->call_stmt_time -= (eni_time_weights.indirect_call_cost
891 - eni_time_weights.call_cost);
895 /* Dump edge summaries associated to NODE and recursively to all clones.
896 Indent by INDENT. */
898 static void
899 dump_ipa_call_summary (FILE *f, int indent, struct cgraph_node *node,
900 class ipa_fn_summary *info)
902 struct cgraph_edge *edge;
903 for (edge = node->callees; edge; edge = edge->next_callee)
905 class ipa_call_summary *es = ipa_call_summaries->get (edge);
906 struct cgraph_node *callee = edge->callee->ultimate_alias_target ();
907 int i;
909 fprintf (f,
910 "%*s%s %s\n%*s freq:%4.2f",
911 indent, "", callee->dump_name (),
912 !edge->inline_failed
913 ? "inlined" : cgraph_inline_failed_string (edge-> inline_failed),
914 indent, "", edge->sreal_frequency ().to_double ());
916 if (cross_module_call_p (edge))
917 fprintf (f, " cross module");
919 if (es)
920 fprintf (f, " loop depth:%2i size:%2i time: %2i",
921 es->loop_depth, es->call_stmt_size, es->call_stmt_time);
923 ipa_fn_summary *s = ipa_fn_summaries->get (callee);
924 ipa_size_summary *ss = ipa_size_summaries->get (callee);
925 if (s != NULL)
926 fprintf (f, " callee size:%2i stack:%2i",
927 (int) (ss->size / ipa_fn_summary::size_scale),
928 (int) s->estimated_stack_size);
930 if (es && es->predicate)
932 fprintf (f, " predicate: ");
933 es->predicate->dump (f, info->conds);
935 else
936 fprintf (f, "\n");
937 if (es && es->param.exists ())
938 for (i = 0; i < (int) es->param.length (); i++)
940 int prob = es->param[i].change_prob;
942 if (!prob)
943 fprintf (f, "%*s op%i is compile time invariant\n",
944 indent + 2, "", i);
945 else if (prob != REG_BR_PROB_BASE)
946 fprintf (f, "%*s op%i change %f%% of time\n", indent + 2, "", i,
947 prob * 100.0 / REG_BR_PROB_BASE);
949 if (!edge->inline_failed)
951 ipa_size_summary *ss = ipa_size_summaries->get (callee);
952 fprintf (f, "%*sStack frame offset %i, callee self size %i\n",
953 indent + 2, "",
954 (int) ipa_get_stack_frame_offset (callee),
955 (int) ss->estimated_self_stack_size);
956 dump_ipa_call_summary (f, indent + 2, callee, info);
959 for (edge = node->indirect_calls; edge; edge = edge->next_callee)
961 class ipa_call_summary *es = ipa_call_summaries->get (edge);
962 fprintf (f, "%*sindirect call loop depth:%2i freq:%4.2f size:%2i"
963 " time: %2i",
964 indent, "",
965 es->loop_depth,
966 edge->sreal_frequency ().to_double (), es->call_stmt_size,
967 es->call_stmt_time);
968 if (es->predicate)
970 fprintf (f, "predicate: ");
971 es->predicate->dump (f, info->conds);
973 else
974 fprintf (f, "\n");
979 void
980 ipa_dump_fn_summary (FILE *f, struct cgraph_node *node)
982 if (node->definition)
984 class ipa_fn_summary *s = ipa_fn_summaries->get (node);
985 class ipa_size_summary *ss = ipa_size_summaries->get (node);
986 if (s != NULL)
988 size_time_entry *e;
989 int i;
990 fprintf (f, "IPA function summary for %s", node->dump_name ());
991 if (DECL_DISREGARD_INLINE_LIMITS (node->decl))
992 fprintf (f, " always_inline");
993 if (s->inlinable)
994 fprintf (f, " inlinable");
995 if (s->fp_expressions)
996 fprintf (f, " fp_expression");
997 fprintf (f, "\n global time: %f\n", s->time.to_double ());
998 fprintf (f, " self size: %i\n", ss->self_size);
999 fprintf (f, " global size: %i\n", ss->size);
1000 fprintf (f, " min size: %i\n", s->min_size);
1001 fprintf (f, " self stack: %i\n",
1002 (int) ss->estimated_self_stack_size);
1003 fprintf (f, " global stack: %i\n", (int) s->estimated_stack_size);
1004 if (s->growth)
1005 fprintf (f, " estimated growth:%i\n", (int) s->growth);
1006 if (s->scc_no)
1007 fprintf (f, " In SCC: %i\n", (int) s->scc_no);
1008 for (i = 0; vec_safe_iterate (s->size_time_table, i, &e); i++)
1010 fprintf (f, " size:%f, time:%f",
1011 (double) e->size / ipa_fn_summary::size_scale,
1012 e->time.to_double ());
1013 if (e->exec_predicate != true)
1015 fprintf (f, ", executed if:");
1016 e->exec_predicate.dump (f, s->conds, 0);
1018 if (e->exec_predicate != e->nonconst_predicate)
1020 fprintf (f, ", nonconst if:");
1021 e->nonconst_predicate.dump (f, s->conds, 0);
1023 fprintf (f, "\n");
1025 if (s->loop_iterations)
1027 fprintf (f, " loop iterations:");
1028 s->loop_iterations->dump (f, s->conds);
1030 if (s->loop_stride)
1032 fprintf (f, " loop stride:");
1033 s->loop_stride->dump (f, s->conds);
1035 fprintf (f, " calls:\n");
1036 dump_ipa_call_summary (f, 4, node, s);
1037 fprintf (f, "\n");
1039 else
1040 fprintf (f, "IPA summary for %s is missing.\n", node->dump_name ());
1044 DEBUG_FUNCTION void
1045 ipa_debug_fn_summary (struct cgraph_node *node)
1047 ipa_dump_fn_summary (stderr, node);
1050 void
1051 ipa_dump_fn_summaries (FILE *f)
1053 struct cgraph_node *node;
1055 FOR_EACH_DEFINED_FUNCTION (node)
1056 if (!node->inlined_to)
1057 ipa_dump_fn_summary (f, node);
1060 /* Callback of walk_aliased_vdefs. Flags that it has been invoked to the
1061 boolean variable pointed to by DATA. */
1063 static bool
1064 mark_modified (ao_ref *ao ATTRIBUTE_UNUSED, tree vdef ATTRIBUTE_UNUSED,
1065 void *data)
1067 bool *b = (bool *) data;
1068 *b = true;
1069 return true;
1072 /* If OP refers to value of function parameter, return the corresponding
1073 parameter. If non-NULL, the size of the memory load (or the SSA_NAME of the
1074 PARM_DECL) will be stored to *SIZE_P in that case too. */
1076 static tree
1077 unmodified_parm_1 (ipa_func_body_info *fbi, gimple *stmt, tree op,
1078 poly_int64 *size_p)
1080 /* SSA_NAME referring to parm default def? */
1081 if (TREE_CODE (op) == SSA_NAME
1082 && SSA_NAME_IS_DEFAULT_DEF (op)
1083 && TREE_CODE (SSA_NAME_VAR (op)) == PARM_DECL)
1085 if (size_p)
1086 *size_p = tree_to_poly_int64 (TYPE_SIZE (TREE_TYPE (op)));
1087 return SSA_NAME_VAR (op);
1089 /* Non-SSA parm reference? */
1090 if (TREE_CODE (op) == PARM_DECL)
1092 bool modified = false;
1094 ao_ref refd;
1095 ao_ref_init (&refd, op);
1096 int walked = walk_aliased_vdefs (&refd, gimple_vuse (stmt),
1097 mark_modified, &modified, NULL, NULL,
1098 fbi->aa_walk_budget + 1);
1099 if (walked < 0)
1101 fbi->aa_walk_budget = 0;
1102 return NULL_TREE;
1104 if (!modified)
1106 if (size_p)
1107 *size_p = tree_to_poly_int64 (TYPE_SIZE (TREE_TYPE (op)));
1108 return op;
1111 return NULL_TREE;
1114 /* If OP refers to value of function parameter, return the corresponding
1115 parameter. Also traverse chains of SSA register assignments. If non-NULL,
1116 the size of the memory load (or the SSA_NAME of the PARM_DECL) will be
1117 stored to *SIZE_P in that case too. */
1119 static tree
1120 unmodified_parm (ipa_func_body_info *fbi, gimple *stmt, tree op,
1121 poly_int64 *size_p)
1123 tree res = unmodified_parm_1 (fbi, stmt, op, size_p);
1124 if (res)
1125 return res;
1127 if (TREE_CODE (op) == SSA_NAME
1128 && !SSA_NAME_IS_DEFAULT_DEF (op)
1129 && gimple_assign_single_p (SSA_NAME_DEF_STMT (op)))
1130 return unmodified_parm (fbi, SSA_NAME_DEF_STMT (op),
1131 gimple_assign_rhs1 (SSA_NAME_DEF_STMT (op)),
1132 size_p);
1133 return NULL_TREE;
1136 /* If OP refers to a value of a function parameter or value loaded from an
1137 aggregate passed to a parameter (either by value or reference), return TRUE
1138 and store the number of the parameter to *INDEX_P, the access size into
1139 *SIZE_P, and information whether and how it has been loaded from an
1140 aggregate into *AGGPOS. INFO describes the function parameters, STMT is the
1141 statement in which OP is used or loaded. */
1143 static bool
1144 unmodified_parm_or_parm_agg_item (struct ipa_func_body_info *fbi,
1145 gimple *stmt, tree op, int *index_p,
1146 poly_int64 *size_p,
1147 struct agg_position_info *aggpos)
1149 tree res = unmodified_parm_1 (fbi, stmt, op, size_p);
1151 gcc_checking_assert (aggpos);
1152 if (res)
1154 *index_p = ipa_get_param_decl_index (fbi->info, res);
1155 if (*index_p < 0)
1156 return false;
1157 aggpos->agg_contents = false;
1158 aggpos->by_ref = false;
1159 return true;
1162 if (TREE_CODE (op) == SSA_NAME)
1164 if (SSA_NAME_IS_DEFAULT_DEF (op)
1165 || !gimple_assign_single_p (SSA_NAME_DEF_STMT (op)))
1166 return false;
1167 stmt = SSA_NAME_DEF_STMT (op);
1168 op = gimple_assign_rhs1 (stmt);
1169 if (!REFERENCE_CLASS_P (op))
1170 return unmodified_parm_or_parm_agg_item (fbi, stmt, op, index_p, size_p,
1171 aggpos);
1174 aggpos->agg_contents = true;
1175 return ipa_load_from_parm_agg (fbi, fbi->info->descriptors,
1176 stmt, op, index_p, &aggpos->offset,
1177 size_p, &aggpos->by_ref);
1180 /* See if statement might disappear after inlining.
1181 0 - means not eliminated
1182 1 - half of statements goes away
1183 2 - for sure it is eliminated.
1184 We are not terribly sophisticated, basically looking for simple abstraction
1185 penalty wrappers. */
1187 static int
1188 eliminated_by_inlining_prob (ipa_func_body_info *fbi, gimple *stmt)
1190 enum gimple_code code = gimple_code (stmt);
1191 enum tree_code rhs_code;
1193 if (!optimize)
1194 return 0;
1196 switch (code)
1198 case GIMPLE_RETURN:
1199 return 2;
1200 case GIMPLE_ASSIGN:
1201 if (gimple_num_ops (stmt) != 2)
1202 return 0;
1204 rhs_code = gimple_assign_rhs_code (stmt);
1206 /* Casts of parameters, loads from parameters passed by reference
1207 and stores to return value or parameters are often free after
1208 inlining due to SRA and further combining.
1209 Assume that half of statements goes away. */
1210 if (CONVERT_EXPR_CODE_P (rhs_code)
1211 || rhs_code == VIEW_CONVERT_EXPR
1212 || rhs_code == ADDR_EXPR
1213 || gimple_assign_rhs_class (stmt) == GIMPLE_SINGLE_RHS)
1215 tree rhs = gimple_assign_rhs1 (stmt);
1216 tree lhs = gimple_assign_lhs (stmt);
1217 tree inner_rhs = get_base_address (rhs);
1218 tree inner_lhs = get_base_address (lhs);
1219 bool rhs_free = false;
1220 bool lhs_free = false;
1222 if (!inner_rhs)
1223 inner_rhs = rhs;
1224 if (!inner_lhs)
1225 inner_lhs = lhs;
1227 /* Reads of parameter are expected to be free. */
1228 if (unmodified_parm (fbi, stmt, inner_rhs, NULL))
1229 rhs_free = true;
1230 /* Match expressions of form &this->field. Those will most likely
1231 combine with something upstream after inlining. */
1232 else if (TREE_CODE (inner_rhs) == ADDR_EXPR)
1234 tree op = get_base_address (TREE_OPERAND (inner_rhs, 0));
1235 if (TREE_CODE (op) == PARM_DECL)
1236 rhs_free = true;
1237 else if (TREE_CODE (op) == MEM_REF
1238 && unmodified_parm (fbi, stmt, TREE_OPERAND (op, 0),
1239 NULL))
1240 rhs_free = true;
1243 /* When parameter is not SSA register because its address is taken
1244 and it is just copied into one, the statement will be completely
1245 free after inlining (we will copy propagate backward). */
1246 if (rhs_free && is_gimple_reg (lhs))
1247 return 2;
1249 /* Reads of parameters passed by reference
1250 expected to be free (i.e. optimized out after inlining). */
1251 if (TREE_CODE (inner_rhs) == MEM_REF
1252 && unmodified_parm (fbi, stmt, TREE_OPERAND (inner_rhs, 0), NULL))
1253 rhs_free = true;
1255 /* Copying parameter passed by reference into gimple register is
1256 probably also going to copy propagate, but we can't be quite
1257 sure. */
1258 if (rhs_free && is_gimple_reg (lhs))
1259 lhs_free = true;
1261 /* Writes to parameters, parameters passed by value and return value
1262 (either directly or passed via invisible reference) are free.
1264 TODO: We ought to handle testcase like
1265 struct a {int a,b;};
1266 struct a
1267 returnstruct (void)
1269 struct a a ={1,2};
1270 return a;
1273 This translate into:
1275 returnstruct ()
1277 int a$b;
1278 int a$a;
1279 struct a a;
1280 struct a D.2739;
1282 <bb 2>:
1283 D.2739.a = 1;
1284 D.2739.b = 2;
1285 return D.2739;
1288 For that we either need to copy ipa-split logic detecting writes
1289 to return value. */
1290 if (TREE_CODE (inner_lhs) == PARM_DECL
1291 || TREE_CODE (inner_lhs) == RESULT_DECL
1292 || (TREE_CODE (inner_lhs) == MEM_REF
1293 && (unmodified_parm (fbi, stmt, TREE_OPERAND (inner_lhs, 0),
1294 NULL)
1295 || (TREE_CODE (TREE_OPERAND (inner_lhs, 0)) == SSA_NAME
1296 && SSA_NAME_VAR (TREE_OPERAND (inner_lhs, 0))
1297 && TREE_CODE (SSA_NAME_VAR (TREE_OPERAND
1298 (inner_lhs,
1299 0))) == RESULT_DECL))))
1300 lhs_free = true;
1301 if (lhs_free
1302 && (is_gimple_reg (rhs) || is_gimple_min_invariant (rhs)))
1303 rhs_free = true;
1304 if (lhs_free && rhs_free)
1305 return 1;
1307 return 0;
1308 default:
1309 return 0;
1313 /* Analyze EXPR if it represents a series of simple operations performed on
1314 a function parameter and return true if so. FBI, STMT, EXPR, INDEX_P and
1315 AGGPOS have the same meaning like in unmodified_parm_or_parm_agg_item.
1316 Type of the parameter or load from an aggregate via the parameter is
1317 stored in *TYPE_P. Operations on the parameter are recorded to
1318 PARAM_OPS_P if it is not NULL. */
1320 static bool
1321 decompose_param_expr (struct ipa_func_body_info *fbi,
1322 gimple *stmt, tree expr,
1323 int *index_p, tree *type_p,
1324 struct agg_position_info *aggpos,
1325 expr_eval_ops *param_ops_p = NULL)
1327 int op_limit = opt_for_fn (fbi->node->decl, param_ipa_max_param_expr_ops);
1328 int op_count = 0;
1330 if (param_ops_p)
1331 *param_ops_p = NULL;
1333 while (true)
1335 expr_eval_op eval_op;
1336 unsigned rhs_count;
1337 unsigned cst_count = 0;
1339 if (unmodified_parm_or_parm_agg_item (fbi, stmt, expr, index_p, NULL,
1340 aggpos))
1342 tree type = TREE_TYPE (expr);
1344 if (aggpos->agg_contents)
1346 /* Stop if containing bit-field. */
1347 if (TREE_CODE (expr) == BIT_FIELD_REF
1348 || contains_bitfld_component_ref_p (expr))
1349 break;
1352 *type_p = type;
1353 return true;
1356 if (TREE_CODE (expr) != SSA_NAME || SSA_NAME_IS_DEFAULT_DEF (expr))
1357 break;
1359 if (!is_gimple_assign (stmt = SSA_NAME_DEF_STMT (expr)))
1360 break;
1362 switch (gimple_assign_rhs_class (stmt))
1364 case GIMPLE_SINGLE_RHS:
1365 expr = gimple_assign_rhs1 (stmt);
1366 continue;
1368 case GIMPLE_UNARY_RHS:
1369 rhs_count = 1;
1370 break;
1372 case GIMPLE_BINARY_RHS:
1373 rhs_count = 2;
1374 break;
1376 case GIMPLE_TERNARY_RHS:
1377 rhs_count = 3;
1378 break;
1380 default:
1381 goto fail;
1384 /* Stop if expression is too complex. */
1385 if (op_count++ == op_limit)
1386 break;
1388 if (param_ops_p)
1390 eval_op.code = gimple_assign_rhs_code (stmt);
1391 eval_op.type = TREE_TYPE (gimple_assign_lhs (stmt));
1392 eval_op.val[0] = NULL_TREE;
1393 eval_op.val[1] = NULL_TREE;
1396 expr = NULL_TREE;
1397 for (unsigned i = 0; i < rhs_count; i++)
1399 tree op = gimple_op (stmt, i + 1);
1401 gcc_assert (op && !TYPE_P (op));
1402 if (is_gimple_ip_invariant (op))
1404 if (++cst_count == rhs_count)
1405 goto fail;
1407 eval_op.val[cst_count - 1] = op;
1409 else if (!expr)
1411 /* Found a non-constant operand, and record its index in rhs
1412 operands. */
1413 eval_op.index = i;
1414 expr = op;
1416 else
1418 /* Found more than one non-constant operands. */
1419 goto fail;
1423 if (param_ops_p)
1424 vec_safe_insert (*param_ops_p, 0, eval_op);
1427 /* Failed to decompose, free resource and return. */
1428 fail:
1429 if (param_ops_p)
1430 vec_free (*param_ops_p);
1432 return false;
1435 /* If BB ends by a conditional we can turn into predicates, attach corresponding
1436 predicates to the CFG edges. */
1438 static void
1439 set_cond_stmt_execution_predicate (struct ipa_func_body_info *fbi,
1440 class ipa_fn_summary *summary,
1441 class ipa_node_params *params_summary,
1442 basic_block bb)
1444 gimple *last;
1445 tree op, op2;
1446 int index;
1447 struct agg_position_info aggpos;
1448 enum tree_code code, inverted_code;
1449 edge e;
1450 edge_iterator ei;
1451 gimple *set_stmt;
1452 tree param_type;
1453 expr_eval_ops param_ops;
1455 last = last_stmt (bb);
1456 if (!last || gimple_code (last) != GIMPLE_COND)
1457 return;
1458 if (!is_gimple_ip_invariant (gimple_cond_rhs (last)))
1459 return;
1460 op = gimple_cond_lhs (last);
1462 if (decompose_param_expr (fbi, last, op, &index, &param_type, &aggpos,
1463 &param_ops))
1465 code = gimple_cond_code (last);
1466 inverted_code = invert_tree_comparison (code, HONOR_NANS (op));
1468 FOR_EACH_EDGE (e, ei, bb->succs)
1470 enum tree_code this_code = (e->flags & EDGE_TRUE_VALUE
1471 ? code : inverted_code);
1472 /* invert_tree_comparison will return ERROR_MARK on FP
1473 comparisons that are not EQ/NE instead of returning proper
1474 unordered one. Be sure it is not confused with NON_CONSTANT.
1476 And if the edge's target is the final block of diamond CFG graph
1477 of this conditional statement, we do not need to compute
1478 predicate for the edge because the final block's predicate must
1479 be at least as that of the first block of the statement. */
1480 if (this_code != ERROR_MARK
1481 && !dominated_by_p (CDI_POST_DOMINATORS, bb, e->dest))
1483 predicate p
1484 = add_condition (summary, params_summary, index,
1485 param_type, &aggpos,
1486 this_code, gimple_cond_rhs (last), param_ops);
1487 e->aux = edge_predicate_pool.allocate ();
1488 *(predicate *) e->aux = p;
1491 vec_free (param_ops);
1494 if (TREE_CODE (op) != SSA_NAME)
1495 return;
1496 /* Special case
1497 if (builtin_constant_p (op))
1498 constant_code
1499 else
1500 nonconstant_code.
1501 Here we can predicate nonconstant_code. We can't
1502 really handle constant_code since we have no predicate
1503 for this and also the constant code is not known to be
1504 optimized away when inliner doesn't see operand is constant.
1505 Other optimizers might think otherwise. */
1506 if (gimple_cond_code (last) != NE_EXPR
1507 || !integer_zerop (gimple_cond_rhs (last)))
1508 return;
1509 set_stmt = SSA_NAME_DEF_STMT (op);
1510 if (!gimple_call_builtin_p (set_stmt, BUILT_IN_CONSTANT_P)
1511 || gimple_call_num_args (set_stmt) != 1)
1512 return;
1513 op2 = gimple_call_arg (set_stmt, 0);
1514 if (!decompose_param_expr (fbi, set_stmt, op2, &index, &param_type, &aggpos))
1515 return;
1516 FOR_EACH_EDGE (e, ei, bb->succs) if (e->flags & EDGE_FALSE_VALUE)
1518 predicate p = add_condition (summary, params_summary, index,
1519 param_type, &aggpos,
1520 predicate::is_not_constant, NULL_TREE);
1521 e->aux = edge_predicate_pool.allocate ();
1522 *(predicate *) e->aux = p;
1527 /* If BB ends by a switch we can turn into predicates, attach corresponding
1528 predicates to the CFG edges. */
1530 static void
1531 set_switch_stmt_execution_predicate (struct ipa_func_body_info *fbi,
1532 class ipa_fn_summary *summary,
1533 class ipa_node_params *params_summary,
1534 basic_block bb)
1536 gimple *lastg;
1537 tree op;
1538 int index;
1539 struct agg_position_info aggpos;
1540 edge e;
1541 edge_iterator ei;
1542 size_t n;
1543 size_t case_idx;
1544 tree param_type;
1545 expr_eval_ops param_ops;
1547 lastg = last_stmt (bb);
1548 if (!lastg || gimple_code (lastg) != GIMPLE_SWITCH)
1549 return;
1550 gswitch *last = as_a <gswitch *> (lastg);
1551 op = gimple_switch_index (last);
1552 if (!decompose_param_expr (fbi, last, op, &index, &param_type, &aggpos,
1553 &param_ops))
1554 return;
1556 auto_vec<std::pair<tree, tree> > ranges;
1557 tree type = TREE_TYPE (op);
1558 int bound_limit = opt_for_fn (fbi->node->decl,
1559 param_ipa_max_switch_predicate_bounds);
1560 int bound_count = 0;
1561 wide_int vr_wmin, vr_wmax;
1562 value_range_kind vr_type = get_range_info (op, &vr_wmin, &vr_wmax);
1564 FOR_EACH_EDGE (e, ei, bb->succs)
1566 e->aux = edge_predicate_pool.allocate ();
1567 *(predicate *) e->aux = false;
1570 e = gimple_switch_edge (cfun, last, 0);
1571 /* Set BOUND_COUNT to maximum count to bypass computing predicate for
1572 default case if its target basic block is in convergence point of all
1573 switch cases, which can be determined by checking whether it
1574 post-dominates the switch statement. */
1575 if (dominated_by_p (CDI_POST_DOMINATORS, bb, e->dest))
1576 bound_count = INT_MAX;
1578 n = gimple_switch_num_labels (last);
1579 for (case_idx = 1; case_idx < n; ++case_idx)
1581 tree cl = gimple_switch_label (last, case_idx);
1582 tree min = CASE_LOW (cl);
1583 tree max = CASE_HIGH (cl);
1584 predicate p;
1586 e = gimple_switch_edge (cfun, last, case_idx);
1588 /* The case value might not have same type as switch expression,
1589 extend the value based on the expression type. */
1590 if (TREE_TYPE (min) != type)
1591 min = wide_int_to_tree (type, wi::to_wide (min));
1593 if (!max)
1594 max = min;
1595 else if (TREE_TYPE (max) != type)
1596 max = wide_int_to_tree (type, wi::to_wide (max));
1598 /* The case's target basic block is in convergence point of all switch
1599 cases, its predicate should be at least as that of the switch
1600 statement. */
1601 if (dominated_by_p (CDI_POST_DOMINATORS, bb, e->dest))
1602 p = true;
1603 else if (min == max)
1604 p = add_condition (summary, params_summary, index, param_type,
1605 &aggpos, EQ_EXPR, min, param_ops);
1606 else
1608 predicate p1, p2;
1609 p1 = add_condition (summary, params_summary, index, param_type,
1610 &aggpos, GE_EXPR, min, param_ops);
1611 p2 = add_condition (summary, params_summary,index, param_type,
1612 &aggpos, LE_EXPR, max, param_ops);
1613 p = p1 & p2;
1615 *(class predicate *) e->aux
1616 = p.or_with (summary->conds, *(class predicate *) e->aux);
1618 /* If there are too many disjoint case ranges, predicate for default
1619 case might become too complicated. So add a limit here. */
1620 if (bound_count > bound_limit)
1621 continue;
1623 bool new_range = true;
1625 if (!ranges.is_empty ())
1627 wide_int curr_wmin = wi::to_wide (min);
1628 wide_int last_wmax = wi::to_wide (ranges.last ().second);
1630 /* Merge case ranges if they are continuous. */
1631 if (curr_wmin == last_wmax + 1)
1632 new_range = false;
1633 else if (vr_type == VR_ANTI_RANGE)
1635 /* If two disjoint case ranges can be connected by anti-range
1636 of switch index, combine them to one range. */
1637 if (wi::lt_p (vr_wmax, curr_wmin - 1, TYPE_SIGN (type)))
1638 vr_type = VR_UNDEFINED;
1639 else if (wi::le_p (vr_wmin, last_wmax + 1, TYPE_SIGN (type)))
1640 new_range = false;
1644 /* Create/extend a case range. And we count endpoints of range set,
1645 this number nearly equals to number of conditions that we will create
1646 for predicate of default case. */
1647 if (new_range)
1649 bound_count += (min == max) ? 1 : 2;
1650 ranges.safe_push (std::make_pair (min, max));
1652 else
1654 bound_count += (ranges.last ().first == ranges.last ().second);
1655 ranges.last ().second = max;
1659 e = gimple_switch_edge (cfun, last, 0);
1660 if (bound_count > bound_limit)
1662 *(class predicate *) e->aux = true;
1663 vec_free (param_ops);
1664 return;
1667 predicate p_seg = true;
1668 predicate p_all = false;
1670 if (vr_type != VR_RANGE)
1672 vr_wmin = wi::to_wide (TYPE_MIN_VALUE (type));
1673 vr_wmax = wi::to_wide (TYPE_MAX_VALUE (type));
1676 /* Construct predicate to represent default range set that is negation of
1677 all case ranges. Case range is classified as containing single/non-single
1678 values. Suppose a piece of case ranges in the following.
1680 [D1...D2] [S1] ... [Sn] [D3...D4]
1682 To represent default case's range sets between two non-single value
1683 case ranges (From D2 to D3), we construct predicate as:
1685 D2 < x < D3 && x != S1 && ... && x != Sn
1687 for (size_t i = 0; i < ranges.length (); i++)
1689 tree min = ranges[i].first;
1690 tree max = ranges[i].second;
1692 if (min == max)
1693 p_seg &= add_condition (summary, params_summary, index,
1694 param_type, &aggpos, NE_EXPR,
1695 min, param_ops);
1696 else
1698 /* Do not create sub-predicate for range that is beyond low bound
1699 of switch index. */
1700 if (wi::lt_p (vr_wmin, wi::to_wide (min), TYPE_SIGN (type)))
1702 p_seg &= add_condition (summary, params_summary, index,
1703 param_type, &aggpos,
1704 LT_EXPR, min, param_ops);
1705 p_all = p_all.or_with (summary->conds, p_seg);
1708 /* Do not create sub-predicate for range that is beyond up bound
1709 of switch index. */
1710 if (wi::le_p (vr_wmax, wi::to_wide (max), TYPE_SIGN (type)))
1712 p_seg = false;
1713 break;
1716 p_seg = add_condition (summary, params_summary, index,
1717 param_type, &aggpos, GT_EXPR,
1718 max, param_ops);
1722 p_all = p_all.or_with (summary->conds, p_seg);
1723 *(class predicate *) e->aux
1724 = p_all.or_with (summary->conds, *(class predicate *) e->aux);
1726 vec_free (param_ops);
1730 /* For each BB in NODE attach to its AUX pointer predicate under
1731 which it is executable. */
1733 static void
1734 compute_bb_predicates (struct ipa_func_body_info *fbi,
1735 struct cgraph_node *node,
1736 class ipa_fn_summary *summary,
1737 class ipa_node_params *params_summary)
1739 struct function *my_function = DECL_STRUCT_FUNCTION (node->decl);
1740 bool done = false;
1741 basic_block bb;
1743 FOR_EACH_BB_FN (bb, my_function)
1745 set_cond_stmt_execution_predicate (fbi, summary, params_summary, bb);
1746 set_switch_stmt_execution_predicate (fbi, summary, params_summary, bb);
1749 /* Entry block is always executable. */
1750 ENTRY_BLOCK_PTR_FOR_FN (my_function)->aux
1751 = edge_predicate_pool.allocate ();
1752 *(predicate *) ENTRY_BLOCK_PTR_FOR_FN (my_function)->aux = true;
1754 /* A simple dataflow propagation of predicates forward in the CFG.
1755 TODO: work in reverse postorder. */
1756 while (!done)
1758 done = true;
1759 FOR_EACH_BB_FN (bb, my_function)
1761 predicate p = false;
1762 edge e;
1763 edge_iterator ei;
1764 FOR_EACH_EDGE (e, ei, bb->preds)
1766 if (e->src->aux)
1768 predicate this_bb_predicate
1769 = *(predicate *) e->src->aux;
1770 if (e->aux)
1771 this_bb_predicate &= (*(class predicate *) e->aux);
1772 p = p.or_with (summary->conds, this_bb_predicate);
1773 if (p == true)
1774 break;
1777 if (p != false)
1779 basic_block pdom_bb;
1781 if (!bb->aux)
1783 done = false;
1784 bb->aux = edge_predicate_pool.allocate ();
1785 *((predicate *) bb->aux) = p;
1787 else if (p != *(predicate *) bb->aux)
1789 /* This OR operation is needed to ensure monotonous data flow
1790 in the case we hit the limit on number of clauses and the
1791 and/or operations above give approximate answers. */
1792 p = p.or_with (summary->conds, *(predicate *)bb->aux);
1793 if (p != *(predicate *) bb->aux)
1795 done = false;
1796 *((predicate *) bb->aux) = p;
1800 /* For switch/if statement, we can OR-combine predicates of all
1801 its cases/branches to get predicate for basic block in their
1802 convergence point, but sometimes this will generate very
1803 complicated predicate. Actually, we can get simplified
1804 predicate in another way by using the fact that predicate
1805 for a basic block must also hold true for its post dominators.
1806 To be specific, basic block in convergence point of
1807 conditional statement should include predicate of the
1808 statement. */
1809 pdom_bb = get_immediate_dominator (CDI_POST_DOMINATORS, bb);
1810 if (pdom_bb == EXIT_BLOCK_PTR_FOR_FN (my_function) || !pdom_bb)
1812 else if (!pdom_bb->aux)
1814 done = false;
1815 pdom_bb->aux = edge_predicate_pool.allocate ();
1816 *((predicate *) pdom_bb->aux) = p;
1818 else if (p != *(predicate *) pdom_bb->aux)
1820 p = p.or_with (summary->conds, *(predicate *)pdom_bb->aux);
1821 if (p != *(predicate *) pdom_bb->aux)
1823 done = false;
1824 *((predicate *) pdom_bb->aux) = p;
1833 /* Return predicate specifying when the STMT might have result that is not
1834 a compile time constant. */
1836 static predicate
1837 will_be_nonconstant_expr_predicate (ipa_func_body_info *fbi,
1838 class ipa_fn_summary *summary,
1839 class ipa_node_params *params_summary,
1840 tree expr,
1841 vec<predicate> nonconstant_names)
1843 tree parm;
1844 int index;
1846 while (UNARY_CLASS_P (expr))
1847 expr = TREE_OPERAND (expr, 0);
1849 parm = unmodified_parm (fbi, NULL, expr, NULL);
1850 if (parm && (index = ipa_get_param_decl_index (fbi->info, parm)) >= 0)
1851 return add_condition (summary, params_summary, index, TREE_TYPE (parm), NULL,
1852 predicate::changed, NULL_TREE);
1853 if (is_gimple_min_invariant (expr))
1854 return false;
1855 if (TREE_CODE (expr) == SSA_NAME)
1856 return nonconstant_names[SSA_NAME_VERSION (expr)];
1857 if (BINARY_CLASS_P (expr) || COMPARISON_CLASS_P (expr))
1859 predicate p1
1860 = will_be_nonconstant_expr_predicate (fbi, summary,
1861 params_summary,
1862 TREE_OPERAND (expr, 0),
1863 nonconstant_names);
1864 if (p1 == true)
1865 return p1;
1867 predicate p2
1868 = will_be_nonconstant_expr_predicate (fbi, summary,
1869 params_summary,
1870 TREE_OPERAND (expr, 1),
1871 nonconstant_names);
1872 return p1.or_with (summary->conds, p2);
1874 else if (TREE_CODE (expr) == COND_EXPR)
1876 predicate p1
1877 = will_be_nonconstant_expr_predicate (fbi, summary,
1878 params_summary,
1879 TREE_OPERAND (expr, 0),
1880 nonconstant_names);
1881 if (p1 == true)
1882 return p1;
1884 predicate p2
1885 = will_be_nonconstant_expr_predicate (fbi, summary,
1886 params_summary,
1887 TREE_OPERAND (expr, 1),
1888 nonconstant_names);
1889 if (p2 == true)
1890 return p2;
1891 p1 = p1.or_with (summary->conds, p2);
1892 p2 = will_be_nonconstant_expr_predicate (fbi, summary,
1893 params_summary,
1894 TREE_OPERAND (expr, 2),
1895 nonconstant_names);
1896 return p2.or_with (summary->conds, p1);
1898 else if (TREE_CODE (expr) == CALL_EXPR)
1899 return true;
1900 else
1902 debug_tree (expr);
1903 gcc_unreachable ();
1905 return false;
1909 /* Return predicate specifying when the STMT might have result that is not
1910 a compile time constant. */
1912 static predicate
1913 will_be_nonconstant_predicate (struct ipa_func_body_info *fbi,
1914 class ipa_fn_summary *summary,
1915 class ipa_node_params *params_summary,
1916 gimple *stmt,
1917 vec<predicate> nonconstant_names)
1919 predicate p = true;
1920 ssa_op_iter iter;
1921 tree use;
1922 tree param_type = NULL_TREE;
1923 predicate op_non_const;
1924 bool is_load;
1925 int base_index;
1926 struct agg_position_info aggpos;
1928 /* What statements might be optimized away
1929 when their arguments are constant. */
1930 if (gimple_code (stmt) != GIMPLE_ASSIGN
1931 && gimple_code (stmt) != GIMPLE_COND
1932 && gimple_code (stmt) != GIMPLE_SWITCH
1933 && (gimple_code (stmt) != GIMPLE_CALL
1934 || !(gimple_call_flags (stmt) & ECF_CONST)))
1935 return p;
1937 /* Stores will stay anyway. */
1938 if (gimple_store_p (stmt))
1939 return p;
1941 is_load = gimple_assign_load_p (stmt);
1943 /* Loads can be optimized when the value is known. */
1944 if (is_load)
1946 tree op = gimple_assign_rhs1 (stmt);
1947 if (!decompose_param_expr (fbi, stmt, op, &base_index, &param_type,
1948 &aggpos))
1949 return p;
1951 else
1952 base_index = -1;
1954 /* See if we understand all operands before we start
1955 adding conditionals. */
1956 FOR_EACH_SSA_TREE_OPERAND (use, stmt, iter, SSA_OP_USE)
1958 tree parm = unmodified_parm (fbi, stmt, use, NULL);
1959 /* For arguments we can build a condition. */
1960 if (parm && ipa_get_param_decl_index (fbi->info, parm) >= 0)
1961 continue;
1962 if (TREE_CODE (use) != SSA_NAME)
1963 return p;
1964 /* If we know when operand is constant,
1965 we still can say something useful. */
1966 if (nonconstant_names[SSA_NAME_VERSION (use)] != true)
1967 continue;
1968 return p;
1971 if (is_load)
1972 op_non_const =
1973 add_condition (summary, params_summary,
1974 base_index, param_type, &aggpos,
1975 predicate::changed, NULL_TREE);
1976 else
1977 op_non_const = false;
1978 FOR_EACH_SSA_TREE_OPERAND (use, stmt, iter, SSA_OP_USE)
1980 tree parm = unmodified_parm (fbi, stmt, use, NULL);
1981 int index;
1983 if (parm && (index = ipa_get_param_decl_index (fbi->info, parm)) >= 0)
1985 if (index != base_index)
1986 p = add_condition (summary, params_summary, index,
1987 TREE_TYPE (parm), NULL,
1988 predicate::changed, NULL_TREE);
1989 else
1990 continue;
1992 else
1993 p = nonconstant_names[SSA_NAME_VERSION (use)];
1994 op_non_const = p.or_with (summary->conds, op_non_const);
1996 if ((gimple_code (stmt) == GIMPLE_ASSIGN || gimple_code (stmt) == GIMPLE_CALL)
1997 && gimple_op (stmt, 0)
1998 && TREE_CODE (gimple_op (stmt, 0)) == SSA_NAME)
1999 nonconstant_names[SSA_NAME_VERSION (gimple_op (stmt, 0))]
2000 = op_non_const;
2001 return op_non_const;
2004 struct record_modified_bb_info
2006 tree op;
2007 bitmap bb_set;
2008 gimple *stmt;
2011 /* Value is initialized in INIT_BB and used in USE_BB. We want to compute
2012 probability how often it changes between USE_BB.
2013 INIT_BB->count/USE_BB->count is an estimate, but if INIT_BB
2014 is in different loop nest, we can do better.
2015 This is all just estimate. In theory we look for minimal cut separating
2016 INIT_BB and USE_BB, but we only want to anticipate loop invariant motion
2017 anyway. */
2019 static basic_block
2020 get_minimal_bb (basic_block init_bb, basic_block use_bb)
2022 class loop *l = find_common_loop (init_bb->loop_father, use_bb->loop_father);
2023 if (l && l->header->count < init_bb->count)
2024 return l->header;
2025 return init_bb;
2028 /* Callback of walk_aliased_vdefs. Records basic blocks where the value may be
2029 set except for info->stmt. */
2031 static bool
2032 record_modified (ao_ref *ao ATTRIBUTE_UNUSED, tree vdef, void *data)
2034 struct record_modified_bb_info *info =
2035 (struct record_modified_bb_info *) data;
2036 if (SSA_NAME_DEF_STMT (vdef) == info->stmt)
2037 return false;
2038 if (gimple_clobber_p (SSA_NAME_DEF_STMT (vdef)))
2039 return false;
2040 bitmap_set_bit (info->bb_set,
2041 SSA_NAME_IS_DEFAULT_DEF (vdef)
2042 ? ENTRY_BLOCK_PTR_FOR_FN (cfun)->index
2043 : get_minimal_bb
2044 (gimple_bb (SSA_NAME_DEF_STMT (vdef)),
2045 gimple_bb (info->stmt))->index);
2046 if (dump_file)
2048 fprintf (dump_file, " Param ");
2049 print_generic_expr (dump_file, info->op, TDF_SLIM);
2050 fprintf (dump_file, " changed at bb %i, minimal: %i stmt: ",
2051 gimple_bb (SSA_NAME_DEF_STMT (vdef))->index,
2052 get_minimal_bb
2053 (gimple_bb (SSA_NAME_DEF_STMT (vdef)),
2054 gimple_bb (info->stmt))->index);
2055 print_gimple_stmt (dump_file, SSA_NAME_DEF_STMT (vdef), 0);
2057 return false;
2060 /* Return probability (based on REG_BR_PROB_BASE) that I-th parameter of STMT
2061 will change since last invocation of STMT.
2063 Value 0 is reserved for compile time invariants.
2064 For common parameters it is REG_BR_PROB_BASE. For loop invariants it
2065 ought to be REG_BR_PROB_BASE / estimated_iters. */
2067 static int
2068 param_change_prob (ipa_func_body_info *fbi, gimple *stmt, int i)
2070 tree op = gimple_call_arg (stmt, i);
2071 basic_block bb = gimple_bb (stmt);
2073 if (TREE_CODE (op) == WITH_SIZE_EXPR)
2074 op = TREE_OPERAND (op, 0);
2076 tree base = get_base_address (op);
2078 /* Global invariants never change. */
2079 if (is_gimple_min_invariant (base))
2080 return 0;
2082 /* We would have to do non-trivial analysis to really work out what
2083 is the probability of value to change (i.e. when init statement
2084 is in a sibling loop of the call).
2086 We do an conservative estimate: when call is executed N times more often
2087 than the statement defining value, we take the frequency 1/N. */
2088 if (TREE_CODE (base) == SSA_NAME)
2090 profile_count init_count;
2092 if (!bb->count.nonzero_p ())
2093 return REG_BR_PROB_BASE;
2095 if (SSA_NAME_IS_DEFAULT_DEF (base))
2096 init_count = ENTRY_BLOCK_PTR_FOR_FN (cfun)->count;
2097 else
2098 init_count = get_minimal_bb
2099 (gimple_bb (SSA_NAME_DEF_STMT (base)),
2100 gimple_bb (stmt))->count;
2102 if (init_count < bb->count)
2103 return MAX ((init_count.to_sreal_scale (bb->count)
2104 * REG_BR_PROB_BASE).to_int (), 1);
2105 return REG_BR_PROB_BASE;
2107 else
2109 ao_ref refd;
2110 profile_count max = ENTRY_BLOCK_PTR_FOR_FN (cfun)->count;
2111 struct record_modified_bb_info info;
2112 tree init = ctor_for_folding (base);
2114 if (init != error_mark_node)
2115 return 0;
2116 if (!bb->count.nonzero_p ())
2117 return REG_BR_PROB_BASE;
2118 if (dump_file)
2120 fprintf (dump_file, " Analyzing param change probability of ");
2121 print_generic_expr (dump_file, op, TDF_SLIM);
2122 fprintf (dump_file, "\n");
2124 ao_ref_init (&refd, op);
2125 info.op = op;
2126 info.stmt = stmt;
2127 info.bb_set = BITMAP_ALLOC (NULL);
2128 int walked
2129 = walk_aliased_vdefs (&refd, gimple_vuse (stmt), record_modified, &info,
2130 NULL, NULL, fbi->aa_walk_budget);
2131 if (walked < 0 || bitmap_bit_p (info.bb_set, bb->index))
2133 if (dump_file)
2135 if (walked < 0)
2136 fprintf (dump_file, " Ran out of AA walking budget.\n");
2137 else
2138 fprintf (dump_file, " Set in same BB as used.\n");
2140 BITMAP_FREE (info.bb_set);
2141 return REG_BR_PROB_BASE;
2144 bitmap_iterator bi;
2145 unsigned index;
2146 /* Lookup the most frequent update of the value and believe that
2147 it dominates all the other; precise analysis here is difficult. */
2148 EXECUTE_IF_SET_IN_BITMAP (info.bb_set, 0, index, bi)
2149 max = max.max (BASIC_BLOCK_FOR_FN (cfun, index)->count);
2150 if (dump_file)
2152 fprintf (dump_file, " Set with count ");
2153 max.dump (dump_file);
2154 fprintf (dump_file, " and used with count ");
2155 bb->count.dump (dump_file);
2156 fprintf (dump_file, " freq %f\n",
2157 max.to_sreal_scale (bb->count).to_double ());
2160 BITMAP_FREE (info.bb_set);
2161 if (max < bb->count)
2162 return MAX ((max.to_sreal_scale (bb->count)
2163 * REG_BR_PROB_BASE).to_int (), 1);
2164 return REG_BR_PROB_BASE;
2168 /* Find whether a basic block BB is the final block of a (half) diamond CFG
2169 sub-graph and if the predicate the condition depends on is known. If so,
2170 return true and store the pointer the predicate in *P. */
2172 static bool
2173 phi_result_unknown_predicate (ipa_func_body_info *fbi,
2174 ipa_fn_summary *summary,
2175 class ipa_node_params *params_summary,
2176 basic_block bb,
2177 predicate *p,
2178 vec<predicate> nonconstant_names)
2180 edge e;
2181 edge_iterator ei;
2182 basic_block first_bb = NULL;
2183 gimple *stmt;
2185 if (single_pred_p (bb))
2187 *p = false;
2188 return true;
2191 FOR_EACH_EDGE (e, ei, bb->preds)
2193 if (single_succ_p (e->src))
2195 if (!single_pred_p (e->src))
2196 return false;
2197 if (!first_bb)
2198 first_bb = single_pred (e->src);
2199 else if (single_pred (e->src) != first_bb)
2200 return false;
2202 else
2204 if (!first_bb)
2205 first_bb = e->src;
2206 else if (e->src != first_bb)
2207 return false;
2211 if (!first_bb)
2212 return false;
2214 stmt = last_stmt (first_bb);
2215 if (!stmt
2216 || gimple_code (stmt) != GIMPLE_COND
2217 || !is_gimple_ip_invariant (gimple_cond_rhs (stmt)))
2218 return false;
2220 *p = will_be_nonconstant_expr_predicate (fbi, summary, params_summary,
2221 gimple_cond_lhs (stmt),
2222 nonconstant_names);
2223 if (*p == true)
2224 return false;
2225 else
2226 return true;
2229 /* Given a PHI statement in a function described by inline properties SUMMARY
2230 and *P being the predicate describing whether the selected PHI argument is
2231 known, store a predicate for the result of the PHI statement into
2232 NONCONSTANT_NAMES, if possible. */
2234 static void
2235 predicate_for_phi_result (class ipa_fn_summary *summary, gphi *phi,
2236 predicate *p,
2237 vec<predicate> nonconstant_names)
2239 unsigned i;
2241 for (i = 0; i < gimple_phi_num_args (phi); i++)
2243 tree arg = gimple_phi_arg (phi, i)->def;
2244 if (!is_gimple_min_invariant (arg))
2246 gcc_assert (TREE_CODE (arg) == SSA_NAME);
2247 *p = p->or_with (summary->conds,
2248 nonconstant_names[SSA_NAME_VERSION (arg)]);
2249 if (*p == true)
2250 return;
2254 if (dump_file && (dump_flags & TDF_DETAILS))
2256 fprintf (dump_file, "\t\tphi predicate: ");
2257 p->dump (dump_file, summary->conds);
2259 nonconstant_names[SSA_NAME_VERSION (gimple_phi_result (phi))] = *p;
2262 /* For a typical usage of __builtin_expect (a<b, 1), we
2263 may introduce an extra relation stmt:
2264 With the builtin, we have
2265 t1 = a <= b;
2266 t2 = (long int) t1;
2267 t3 = __builtin_expect (t2, 1);
2268 if (t3 != 0)
2269 goto ...
2270 Without the builtin, we have
2271 if (a<=b)
2272 goto...
2273 This affects the size/time estimation and may have
2274 an impact on the earlier inlining.
2275 Here find this pattern and fix it up later. */
2277 static gimple *
2278 find_foldable_builtin_expect (basic_block bb)
2280 gimple_stmt_iterator bsi;
2282 for (bsi = gsi_start_bb (bb); !gsi_end_p (bsi); gsi_next (&bsi))
2284 gimple *stmt = gsi_stmt (bsi);
2285 if (gimple_call_builtin_p (stmt, BUILT_IN_EXPECT)
2286 || gimple_call_builtin_p (stmt, BUILT_IN_EXPECT_WITH_PROBABILITY)
2287 || gimple_call_internal_p (stmt, IFN_BUILTIN_EXPECT))
2289 tree var = gimple_call_lhs (stmt);
2290 tree arg = gimple_call_arg (stmt, 0);
2291 use_operand_p use_p;
2292 gimple *use_stmt;
2293 bool match = false;
2294 bool done = false;
2296 if (!var || !arg)
2297 continue;
2298 gcc_assert (TREE_CODE (var) == SSA_NAME);
2300 while (TREE_CODE (arg) == SSA_NAME)
2302 gimple *stmt_tmp = SSA_NAME_DEF_STMT (arg);
2303 if (!is_gimple_assign (stmt_tmp))
2304 break;
2305 switch (gimple_assign_rhs_code (stmt_tmp))
2307 case LT_EXPR:
2308 case LE_EXPR:
2309 case GT_EXPR:
2310 case GE_EXPR:
2311 case EQ_EXPR:
2312 case NE_EXPR:
2313 match = true;
2314 done = true;
2315 break;
2316 CASE_CONVERT:
2317 break;
2318 default:
2319 done = true;
2320 break;
2322 if (done)
2323 break;
2324 arg = gimple_assign_rhs1 (stmt_tmp);
2327 if (match && single_imm_use (var, &use_p, &use_stmt)
2328 && gimple_code (use_stmt) == GIMPLE_COND)
2329 return use_stmt;
2332 return NULL;
2335 /* Return true when the basic blocks contains only clobbers followed by RESX.
2336 Such BBs are kept around to make removal of dead stores possible with
2337 presence of EH and will be optimized out by optimize_clobbers later in the
2338 game.
2340 NEED_EH is used to recurse in case the clobber has non-EH predecessors
2341 that can be clobber only, too.. When it is false, the RESX is not necessary
2342 on the end of basic block. */
2344 static bool
2345 clobber_only_eh_bb_p (basic_block bb, bool need_eh = true)
2347 gimple_stmt_iterator gsi = gsi_last_bb (bb);
2348 edge_iterator ei;
2349 edge e;
2351 if (need_eh)
2353 if (gsi_end_p (gsi))
2354 return false;
2355 if (gimple_code (gsi_stmt (gsi)) != GIMPLE_RESX)
2356 return false;
2357 gsi_prev (&gsi);
2359 else if (!single_succ_p (bb))
2360 return false;
2362 for (; !gsi_end_p (gsi); gsi_prev (&gsi))
2364 gimple *stmt = gsi_stmt (gsi);
2365 if (is_gimple_debug (stmt))
2366 continue;
2367 if (gimple_clobber_p (stmt))
2368 continue;
2369 if (gimple_code (stmt) == GIMPLE_LABEL)
2370 break;
2371 return false;
2374 /* See if all predecessors are either throws or clobber only BBs. */
2375 FOR_EACH_EDGE (e, ei, bb->preds)
2376 if (!(e->flags & EDGE_EH)
2377 && !clobber_only_eh_bb_p (e->src, false))
2378 return false;
2380 return true;
2383 /* Return true if STMT compute a floating point expression that may be affected
2384 by -ffast-math and similar flags. */
2386 static bool
2387 fp_expression_p (gimple *stmt)
2389 ssa_op_iter i;
2390 tree op;
2392 FOR_EACH_SSA_TREE_OPERAND (op, stmt, i, SSA_OP_DEF|SSA_OP_USE)
2393 if (FLOAT_TYPE_P (TREE_TYPE (op)))
2394 return true;
2395 return false;
2398 /* Analyze function body for NODE.
2399 EARLY indicates run from early optimization pipeline. */
2401 static void
2402 analyze_function_body (struct cgraph_node *node, bool early)
2404 sreal time = opt_for_fn (node->decl, param_uninlined_function_time);
2405 /* Estimate static overhead for function prologue/epilogue and alignment. */
2406 int size = opt_for_fn (node->decl, param_uninlined_function_insns);
2407 /* Benefits are scaled by probability of elimination that is in range
2408 <0,2>. */
2409 basic_block bb;
2410 struct function *my_function = DECL_STRUCT_FUNCTION (node->decl);
2411 sreal freq;
2412 class ipa_fn_summary *info = ipa_fn_summaries->get_create (node);
2413 class ipa_node_params *params_summary = early ? NULL : IPA_NODE_REF (node);
2414 predicate bb_predicate;
2415 struct ipa_func_body_info fbi;
2416 vec<predicate> nonconstant_names = vNULL;
2417 int nblocks, n;
2418 int *order;
2419 gimple *fix_builtin_expect_stmt;
2421 gcc_assert (my_function && my_function->cfg);
2422 gcc_assert (cfun == my_function);
2424 memset(&fbi, 0, sizeof(fbi));
2425 vec_free (info->conds);
2426 info->conds = NULL;
2427 vec_free (info->size_time_table);
2428 info->size_time_table = NULL;
2430 /* When optimizing and analyzing for IPA inliner, initialize loop optimizer
2431 so we can produce proper inline hints.
2433 When optimizing and analyzing for early inliner, initialize node params
2434 so we can produce correct BB predicates. */
2436 if (opt_for_fn (node->decl, optimize))
2438 calculate_dominance_info (CDI_DOMINATORS);
2439 calculate_dominance_info (CDI_POST_DOMINATORS);
2440 if (!early)
2441 loop_optimizer_init (LOOPS_NORMAL | LOOPS_HAVE_RECORDED_EXITS);
2442 else
2444 ipa_check_create_node_params ();
2445 ipa_initialize_node_params (node);
2448 if (ipa_node_params_sum)
2450 fbi.node = node;
2451 fbi.info = IPA_NODE_REF (node);
2452 fbi.bb_infos = vNULL;
2453 fbi.bb_infos.safe_grow_cleared (last_basic_block_for_fn (cfun));
2454 fbi.param_count = count_formal_params (node->decl);
2455 fbi.aa_walk_budget = opt_for_fn (node->decl, param_ipa_max_aa_steps);
2457 nonconstant_names.safe_grow_cleared
2458 (SSANAMES (my_function)->length ());
2462 if (dump_file)
2463 fprintf (dump_file, "\nAnalyzing function body size: %s\n",
2464 node->dump_name ());
2466 /* When we run into maximal number of entries, we assign everything to the
2467 constant truth case. Be sure to have it in list. */
2468 bb_predicate = true;
2469 info->account_size_time (0, 0, bb_predicate, bb_predicate);
2471 bb_predicate = predicate::not_inlined ();
2472 info->account_size_time (opt_for_fn (node->decl,
2473 param_uninlined_function_insns)
2474 * ipa_fn_summary::size_scale,
2475 opt_for_fn (node->decl,
2476 param_uninlined_function_time),
2477 bb_predicate,
2478 bb_predicate);
2480 if (fbi.info)
2481 compute_bb_predicates (&fbi, node, info, params_summary);
2482 order = XNEWVEC (int, n_basic_blocks_for_fn (cfun));
2483 nblocks = pre_and_rev_post_order_compute (NULL, order, false);
2484 for (n = 0; n < nblocks; n++)
2486 bb = BASIC_BLOCK_FOR_FN (cfun, order[n]);
2487 freq = bb->count.to_sreal_scale (ENTRY_BLOCK_PTR_FOR_FN (cfun)->count);
2488 if (clobber_only_eh_bb_p (bb))
2490 if (dump_file && (dump_flags & TDF_DETAILS))
2491 fprintf (dump_file, "\n Ignoring BB %i;"
2492 " it will be optimized away by cleanup_clobbers\n",
2493 bb->index);
2494 continue;
2497 /* TODO: Obviously predicates can be propagated down across CFG. */
2498 if (fbi.info)
2500 if (bb->aux)
2501 bb_predicate = *(predicate *) bb->aux;
2502 else
2503 bb_predicate = false;
2505 else
2506 bb_predicate = true;
2508 if (dump_file && (dump_flags & TDF_DETAILS))
2510 fprintf (dump_file, "\n BB %i predicate:", bb->index);
2511 bb_predicate.dump (dump_file, info->conds);
2514 if (fbi.info && nonconstant_names.exists ())
2516 predicate phi_predicate;
2517 bool first_phi = true;
2519 for (gphi_iterator bsi = gsi_start_phis (bb); !gsi_end_p (bsi);
2520 gsi_next (&bsi))
2522 if (first_phi
2523 && !phi_result_unknown_predicate (&fbi, info,
2524 params_summary,
2526 &phi_predicate,
2527 nonconstant_names))
2528 break;
2529 first_phi = false;
2530 if (dump_file && (dump_flags & TDF_DETAILS))
2532 fprintf (dump_file, " ");
2533 print_gimple_stmt (dump_file, gsi_stmt (bsi), 0);
2535 predicate_for_phi_result (info, bsi.phi (), &phi_predicate,
2536 nonconstant_names);
2540 fix_builtin_expect_stmt = find_foldable_builtin_expect (bb);
2542 for (gimple_stmt_iterator bsi = gsi_start_nondebug_bb (bb);
2543 !gsi_end_p (bsi); gsi_next_nondebug (&bsi))
2545 gimple *stmt = gsi_stmt (bsi);
2546 int this_size = estimate_num_insns (stmt, &eni_size_weights);
2547 int this_time = estimate_num_insns (stmt, &eni_time_weights);
2548 int prob;
2549 predicate will_be_nonconstant;
2551 /* This relation stmt should be folded after we remove
2552 __builtin_expect call. Adjust the cost here. */
2553 if (stmt == fix_builtin_expect_stmt)
2555 this_size--;
2556 this_time--;
2559 if (dump_file && (dump_flags & TDF_DETAILS))
2561 fprintf (dump_file, " ");
2562 print_gimple_stmt (dump_file, stmt, 0);
2563 fprintf (dump_file, "\t\tfreq:%3.2f size:%3i time:%3i\n",
2564 freq.to_double (), this_size,
2565 this_time);
2568 if (is_gimple_call (stmt)
2569 && !gimple_call_internal_p (stmt))
2571 struct cgraph_edge *edge = node->get_edge (stmt);
2572 ipa_call_summary *es = ipa_call_summaries->get_create (edge);
2574 /* Special case: results of BUILT_IN_CONSTANT_P will be always
2575 resolved as constant. We however don't want to optimize
2576 out the cgraph edges. */
2577 if (nonconstant_names.exists ()
2578 && gimple_call_builtin_p (stmt, BUILT_IN_CONSTANT_P)
2579 && gimple_call_lhs (stmt)
2580 && TREE_CODE (gimple_call_lhs (stmt)) == SSA_NAME)
2582 predicate false_p = false;
2583 nonconstant_names[SSA_NAME_VERSION (gimple_call_lhs (stmt))]
2584 = false_p;
2586 if (ipa_node_params_sum)
2588 int count = gimple_call_num_args (stmt);
2589 int i;
2591 if (count)
2592 es->param.safe_grow_cleared (count);
2593 for (i = 0; i < count; i++)
2595 int prob = param_change_prob (&fbi, stmt, i);
2596 gcc_assert (prob >= 0 && prob <= REG_BR_PROB_BASE);
2597 es->param[i].change_prob = prob;
2601 es->call_stmt_size = this_size;
2602 es->call_stmt_time = this_time;
2603 es->loop_depth = bb_loop_depth (bb);
2604 edge_set_predicate (edge, &bb_predicate);
2605 if (edge->speculative)
2607 cgraph_edge *direct, *indirect, *next_direct;
2608 ipa_ref *ref;
2609 edge->speculative_call_info (direct, indirect, ref);
2610 gcc_assert (direct == edge);
2611 ipa_call_summary *es2
2612 = ipa_call_summaries->get_create (indirect);
2613 ipa_call_summaries->duplicate (edge, indirect,
2614 es, es2);
2616 /* Create and duplicate call summaries for multiple
2617 speculative call targets. */
2618 int num_specs = indirect->num_speculative_call_targets_p ();
2619 if (num_specs > 1)
2620 for (next_direct = edge->next_callee;
2621 next_direct && --num_specs;
2622 next_direct = next_direct->next_callee)
2624 next_direct->speculative_call_info (direct, indirect,
2625 ref);
2626 if (direct == next_direct && next_direct->speculative
2627 && edge->call_stmt == stmt)
2629 ipa_call_summary *es3
2630 = ipa_call_summaries->get_create (next_direct);
2631 ipa_call_summaries->duplicate (edge, next_direct,
2632 es, es3);
2638 /* TODO: When conditional jump or switch is known to be constant, but
2639 we did not translate it into the predicates, we really can account
2640 just maximum of the possible paths. */
2641 if (fbi.info)
2642 will_be_nonconstant
2643 = will_be_nonconstant_predicate (&fbi, info, params_summary,
2644 stmt, nonconstant_names);
2645 else
2646 will_be_nonconstant = true;
2647 if (this_time || this_size)
2649 sreal final_time = (sreal)this_time * freq;
2651 prob = eliminated_by_inlining_prob (&fbi, stmt);
2652 if (prob == 1 && dump_file && (dump_flags & TDF_DETAILS))
2653 fprintf (dump_file,
2654 "\t\t50%% will be eliminated by inlining\n");
2655 if (prob == 2 && dump_file && (dump_flags & TDF_DETAILS))
2656 fprintf (dump_file, "\t\tWill be eliminated by inlining\n");
2658 class predicate p = bb_predicate & will_be_nonconstant;
2660 /* We can ignore statement when we proved it is never going
2661 to happen, but we cannot do that for call statements
2662 because edges are accounted specially. */
2664 if (*(is_gimple_call (stmt) ? &bb_predicate : &p) != false)
2666 time += final_time;
2667 size += this_size;
2670 /* We account everything but the calls. Calls have their own
2671 size/time info attached to cgraph edges. This is necessary
2672 in order to make the cost disappear after inlining. */
2673 if (!is_gimple_call (stmt))
2675 if (prob)
2677 predicate ip = bb_predicate & predicate::not_inlined ();
2678 info->account_size_time (this_size * prob,
2679 (final_time * prob) / 2, ip,
2682 if (prob != 2)
2683 info->account_size_time (this_size * (2 - prob),
2684 (final_time * (2 - prob) / 2),
2685 bb_predicate,
2689 if (!info->fp_expressions && fp_expression_p (stmt))
2691 info->fp_expressions = true;
2692 if (dump_file)
2693 fprintf (dump_file, " fp_expression set\n");
2697 /* Account cost of address calculations in the statements. */
2698 for (unsigned int i = 0; i < gimple_num_ops (stmt); i++)
2700 for (tree op = gimple_op (stmt, i);
2701 op && handled_component_p (op);
2702 op = TREE_OPERAND (op, 0))
2703 if ((TREE_CODE (op) == ARRAY_REF
2704 || TREE_CODE (op) == ARRAY_RANGE_REF)
2705 && TREE_CODE (TREE_OPERAND (op, 1)) == SSA_NAME)
2707 predicate p = bb_predicate;
2708 if (fbi.info)
2709 p = p & will_be_nonconstant_expr_predicate
2710 (&fbi, info, params_summary,
2711 TREE_OPERAND (op, 1),
2712 nonconstant_names);
2713 if (p != false)
2715 time += freq;
2716 size += 1;
2717 if (dump_file)
2718 fprintf (dump_file,
2719 "\t\tAccounting address calculation.\n");
2720 info->account_size_time (ipa_fn_summary::size_scale,
2721 freq,
2722 bb_predicate,
2730 free (order);
2732 if (nonconstant_names.exists () && !early)
2734 class loop *loop;
2735 predicate loop_iterations = true;
2736 predicate loop_stride = true;
2738 if (dump_file && (dump_flags & TDF_DETAILS))
2739 flow_loops_dump (dump_file, NULL, 0);
2740 scev_initialize ();
2741 FOR_EACH_LOOP (loop, 0)
2743 vec<edge> exits;
2744 edge ex;
2745 unsigned int j;
2746 class tree_niter_desc niter_desc;
2747 bb_predicate = *(predicate *) loop->header->aux;
2749 exits = get_loop_exit_edges (loop);
2750 FOR_EACH_VEC_ELT (exits, j, ex)
2751 if (number_of_iterations_exit (loop, ex, &niter_desc, false)
2752 && !is_gimple_min_invariant (niter_desc.niter))
2754 predicate will_be_nonconstant
2755 = will_be_nonconstant_expr_predicate (&fbi, info,
2756 params_summary,
2757 niter_desc.niter,
2758 nonconstant_names);
2759 if (will_be_nonconstant != true)
2760 will_be_nonconstant = bb_predicate & will_be_nonconstant;
2761 if (will_be_nonconstant != true
2762 && will_be_nonconstant != false)
2763 /* This is slightly inprecise. We may want to represent each
2764 loop with independent predicate. */
2765 loop_iterations &= will_be_nonconstant;
2767 exits.release ();
2770 /* To avoid quadratic behavior we analyze stride predicates only
2771 with respect to the containing loop. Thus we simply iterate
2772 over all defs in the outermost loop body. */
2773 for (loop = loops_for_fn (cfun)->tree_root->inner;
2774 loop != NULL; loop = loop->next)
2776 basic_block *body = get_loop_body (loop);
2777 for (unsigned i = 0; i < loop->num_nodes; i++)
2779 gimple_stmt_iterator gsi;
2780 bb_predicate = *(predicate *) body[i]->aux;
2781 for (gsi = gsi_start_bb (body[i]); !gsi_end_p (gsi);
2782 gsi_next (&gsi))
2784 gimple *stmt = gsi_stmt (gsi);
2786 if (!is_gimple_assign (stmt))
2787 continue;
2789 tree def = gimple_assign_lhs (stmt);
2790 if (TREE_CODE (def) != SSA_NAME)
2791 continue;
2793 affine_iv iv;
2794 if (!simple_iv (loop_containing_stmt (stmt),
2795 loop_containing_stmt (stmt),
2796 def, &iv, true)
2797 || is_gimple_min_invariant (iv.step))
2798 continue;
2800 predicate will_be_nonconstant
2801 = will_be_nonconstant_expr_predicate (&fbi, info,
2802 params_summary,
2803 iv.step,
2804 nonconstant_names);
2805 if (will_be_nonconstant != true)
2806 will_be_nonconstant = bb_predicate & will_be_nonconstant;
2807 if (will_be_nonconstant != true
2808 && will_be_nonconstant != false)
2809 /* This is slightly inprecise. We may want to represent
2810 each loop with independent predicate. */
2811 loop_stride = loop_stride & will_be_nonconstant;
2814 free (body);
2816 ipa_fn_summary *s = ipa_fn_summaries->get (node);
2817 set_hint_predicate (&s->loop_iterations, loop_iterations);
2818 set_hint_predicate (&s->loop_stride, loop_stride);
2819 scev_finalize ();
2821 FOR_ALL_BB_FN (bb, my_function)
2823 edge e;
2824 edge_iterator ei;
2826 if (bb->aux)
2827 edge_predicate_pool.remove ((predicate *)bb->aux);
2828 bb->aux = NULL;
2829 FOR_EACH_EDGE (e, ei, bb->succs)
2831 if (e->aux)
2832 edge_predicate_pool.remove ((predicate *) e->aux);
2833 e->aux = NULL;
2836 ipa_fn_summary *s = ipa_fn_summaries->get (node);
2837 ipa_size_summary *ss = ipa_size_summaries->get (node);
2838 s->time = time;
2839 ss->self_size = size;
2840 nonconstant_names.release ();
2841 ipa_release_body_info (&fbi);
2842 if (opt_for_fn (node->decl, optimize))
2844 if (!early)
2845 loop_optimizer_finalize ();
2846 else if (!ipa_edge_args_sum)
2847 ipa_free_all_node_params ();
2848 free_dominance_info (CDI_DOMINATORS);
2849 free_dominance_info (CDI_POST_DOMINATORS);
2851 if (dump_file)
2853 fprintf (dump_file, "\n");
2854 ipa_dump_fn_summary (dump_file, node);
2859 /* Compute function summary.
2860 EARLY is true when we compute parameters during early opts. */
2862 void
2863 compute_fn_summary (struct cgraph_node *node, bool early)
2865 HOST_WIDE_INT self_stack_size;
2866 struct cgraph_edge *e;
2868 gcc_assert (!node->inlined_to);
2870 if (!ipa_fn_summaries)
2871 ipa_fn_summary_alloc ();
2873 /* Create a new ipa_fn_summary. */
2874 ((ipa_fn_summary_t *)ipa_fn_summaries)->remove_callees (node);
2875 ipa_fn_summaries->remove (node);
2876 class ipa_fn_summary *info = ipa_fn_summaries->get_create (node);
2877 class ipa_size_summary *size_info = ipa_size_summaries->get_create (node);
2879 /* Estimate the stack size for the function if we're optimizing. */
2880 self_stack_size = optimize && !node->thunk.thunk_p
2881 ? estimated_stack_frame_size (node) : 0;
2882 size_info->estimated_self_stack_size = self_stack_size;
2883 info->estimated_stack_size = self_stack_size;
2885 if (node->thunk.thunk_p)
2887 ipa_call_summary *es = ipa_call_summaries->get_create (node->callees);
2888 predicate t = true;
2890 node->can_change_signature = false;
2891 es->call_stmt_size = eni_size_weights.call_cost;
2892 es->call_stmt_time = eni_time_weights.call_cost;
2893 info->account_size_time (ipa_fn_summary::size_scale
2894 * opt_for_fn (node->decl,
2895 param_uninlined_function_thunk_insns),
2896 opt_for_fn (node->decl,
2897 param_uninlined_function_thunk_time), t, t);
2898 t = predicate::not_inlined ();
2899 info->account_size_time (2 * ipa_fn_summary::size_scale, 0, t, t);
2900 ipa_update_overall_fn_summary (node);
2901 size_info->self_size = size_info->size;
2902 if (stdarg_p (TREE_TYPE (node->decl)))
2904 info->inlinable = false;
2905 node->callees->inline_failed = CIF_VARIADIC_THUNK;
2907 else
2908 info->inlinable = true;
2910 else
2912 /* Even is_gimple_min_invariant rely on current_function_decl. */
2913 push_cfun (DECL_STRUCT_FUNCTION (node->decl));
2915 /* During IPA profile merging we may be called w/o virtual SSA form
2916 built. */
2917 update_ssa (TODO_update_ssa_only_virtuals);
2919 /* Can this function be inlined at all? */
2920 if (!opt_for_fn (node->decl, optimize)
2921 && !lookup_attribute ("always_inline",
2922 DECL_ATTRIBUTES (node->decl)))
2923 info->inlinable = false;
2924 else
2925 info->inlinable = tree_inlinable_function_p (node->decl);
2927 /* Type attributes can use parameter indices to describe them. */
2928 if (TYPE_ATTRIBUTES (TREE_TYPE (node->decl))
2929 /* Likewise for #pragma omp declare simd functions or functions
2930 with simd attribute. */
2931 || lookup_attribute ("omp declare simd",
2932 DECL_ATTRIBUTES (node->decl)))
2933 node->can_change_signature = false;
2934 else
2936 /* Otherwise, inlinable functions always can change signature. */
2937 if (info->inlinable)
2938 node->can_change_signature = true;
2939 else
2941 /* Functions calling builtin_apply cannot change signature. */
2942 for (e = node->callees; e; e = e->next_callee)
2944 tree cdecl = e->callee->decl;
2945 if (fndecl_built_in_p (cdecl, BUILT_IN_APPLY_ARGS)
2946 || fndecl_built_in_p (cdecl, BUILT_IN_VA_START))
2947 break;
2949 node->can_change_signature = !e;
2952 analyze_function_body (node, early);
2953 pop_cfun ();
2955 for (e = node->callees; e; e = e->next_callee)
2956 if (e->callee->comdat_local_p ())
2957 break;
2958 node->calls_comdat_local = (e != NULL);
2960 /* Inlining characteristics are maintained by the cgraph_mark_inline. */
2961 size_info->size = size_info->self_size;
2962 info->estimated_stack_size = size_info->estimated_self_stack_size;
2964 /* Code above should compute exactly the same result as
2965 ipa_update_overall_fn_summary but because computation happens in
2966 different order the roundoff errors result in slight changes. */
2967 ipa_update_overall_fn_summary (node);
2968 /* In LTO mode we may have speculative edges set. */
2969 gcc_assert (in_lto_p || size_info->size == size_info->self_size);
2973 /* Compute parameters of functions used by inliner using
2974 current_function_decl. */
2976 static unsigned int
2977 compute_fn_summary_for_current (void)
2979 compute_fn_summary (cgraph_node::get (current_function_decl), true);
2980 return 0;
2983 /* Estimate benefit devirtualizing indirect edge IE, provided KNOWN_VALS,
2984 KNOWN_CONTEXTS and KNOWN_AGGS. */
2986 static bool
2987 estimate_edge_devirt_benefit (struct cgraph_edge *ie,
2988 int *size, int *time,
2989 vec<tree> known_vals,
2990 vec<ipa_polymorphic_call_context> known_contexts,
2991 vec<ipa_agg_value_set> known_aggs)
2993 tree target;
2994 struct cgraph_node *callee;
2995 class ipa_fn_summary *isummary;
2996 enum availability avail;
2997 bool speculative;
2999 if (!known_vals.length () && !known_contexts.length ())
3000 return false;
3001 if (!opt_for_fn (ie->caller->decl, flag_indirect_inlining))
3002 return false;
3004 target = ipa_get_indirect_edge_target (ie, known_vals, known_contexts,
3005 known_aggs, &speculative);
3006 if (!target || speculative)
3007 return false;
3009 /* Account for difference in cost between indirect and direct calls. */
3010 *size -= (eni_size_weights.indirect_call_cost - eni_size_weights.call_cost);
3011 *time -= (eni_time_weights.indirect_call_cost - eni_time_weights.call_cost);
3012 gcc_checking_assert (*time >= 0);
3013 gcc_checking_assert (*size >= 0);
3015 callee = cgraph_node::get (target);
3016 if (!callee || !callee->definition)
3017 return false;
3018 callee = callee->function_symbol (&avail);
3019 if (avail < AVAIL_AVAILABLE)
3020 return false;
3021 isummary = ipa_fn_summaries->get (callee);
3022 if (isummary == NULL)
3023 return false;
3025 return isummary->inlinable;
3028 /* Increase SIZE, MIN_SIZE (if non-NULL) and TIME for size and time needed to
3029 handle edge E with probability PROB.
3030 Set HINTS if edge may be devirtualized.
3031 KNOWN_VALS, KNOWN_AGGS and KNOWN_CONTEXTS describe context of the call
3032 site. */
3034 static inline void
3035 estimate_edge_size_and_time (struct cgraph_edge *e, int *size, int *min_size,
3036 sreal *time,
3037 vec<tree> known_vals,
3038 vec<ipa_polymorphic_call_context> known_contexts,
3039 vec<ipa_agg_value_set> known_aggs,
3040 ipa_hints *hints)
3042 class ipa_call_summary *es = ipa_call_summaries->get (e);
3043 int call_size = es->call_stmt_size;
3044 int call_time = es->call_stmt_time;
3045 int cur_size;
3047 if (!e->callee && hints && e->maybe_hot_p ()
3048 && estimate_edge_devirt_benefit (e, &call_size, &call_time,
3049 known_vals, known_contexts, known_aggs))
3050 *hints |= INLINE_HINT_indirect_call;
3051 cur_size = call_size * ipa_fn_summary::size_scale;
3052 *size += cur_size;
3053 if (min_size)
3054 *min_size += cur_size;
3055 if (time)
3056 *time += ((sreal)call_time) * e->sreal_frequency ();
3060 /* Increase SIZE, MIN_SIZE and TIME for size and time needed to handle all
3061 calls in NODE. POSSIBLE_TRUTHS, KNOWN_VALS, KNOWN_AGGS and KNOWN_CONTEXTS
3062 describe context of the call site.
3064 Helper for estimate_calls_size_and_time which does the same but
3065 (in most cases) faster. */
3067 static void
3068 estimate_calls_size_and_time_1 (struct cgraph_node *node, int *size,
3069 int *min_size, sreal *time,
3070 ipa_hints *hints,
3071 clause_t possible_truths,
3072 vec<tree> known_vals,
3073 vec<ipa_polymorphic_call_context> known_contexts,
3074 vec<ipa_agg_value_set> known_aggs)
3076 struct cgraph_edge *e;
3077 for (e = node->callees; e; e = e->next_callee)
3079 if (!e->inline_failed)
3081 gcc_checking_assert (!ipa_call_summaries->get (e));
3082 estimate_calls_size_and_time_1 (e->callee, size, min_size, time,
3083 hints,
3084 possible_truths,
3085 known_vals, known_contexts,
3086 known_aggs);
3087 continue;
3089 class ipa_call_summary *es = ipa_call_summaries->get (e);
3091 /* Do not care about zero sized builtins. */
3092 if (!es->call_stmt_size)
3094 gcc_checking_assert (!es->call_stmt_time);
3095 continue;
3097 if (!es->predicate
3098 || es->predicate->evaluate (possible_truths))
3100 /* Predicates of calls shall not use NOT_CHANGED codes,
3101 so we do not need to compute probabilities. */
3102 estimate_edge_size_and_time (e, size,
3103 es->predicate ? NULL : min_size,
3104 time,
3105 known_vals, known_contexts,
3106 known_aggs, hints);
3109 for (e = node->indirect_calls; e; e = e->next_callee)
3111 class ipa_call_summary *es = ipa_call_summaries->get (e);
3112 if (!es->predicate
3113 || es->predicate->evaluate (possible_truths))
3114 estimate_edge_size_and_time (e, size,
3115 es->predicate ? NULL : min_size,
3116 time,
3117 known_vals, known_contexts, known_aggs,
3118 hints);
3122 /* Populate sum->call_size_time_table for edges from NODE. */
3124 static void
3125 summarize_calls_size_and_time (struct cgraph_node *node,
3126 ipa_fn_summary *sum)
3128 struct cgraph_edge *e;
3129 for (e = node->callees; e; e = e->next_callee)
3131 if (!e->inline_failed)
3133 gcc_checking_assert (!ipa_call_summaries->get (e));
3134 summarize_calls_size_and_time (e->callee, sum);
3135 continue;
3137 int size = 0;
3138 sreal time = 0;
3140 estimate_edge_size_and_time (e, &size, NULL, &time,
3141 vNULL, vNULL, vNULL, NULL);
3143 struct predicate pred = true;
3144 class ipa_call_summary *es = ipa_call_summaries->get (e);
3146 if (es->predicate)
3147 pred = *es->predicate;
3148 sum->account_size_time (size, time, pred, pred, true);
3150 for (e = node->indirect_calls; e; e = e->next_callee)
3152 int size = 0;
3153 sreal time = 0;
3155 estimate_edge_size_and_time (e, &size, NULL, &time,
3156 vNULL, vNULL, vNULL, NULL);
3157 struct predicate pred = true;
3158 class ipa_call_summary *es = ipa_call_summaries->get (e);
3160 if (es->predicate)
3161 pred = *es->predicate;
3162 sum->account_size_time (size, time, pred, pred, true);
3166 /* Increase SIZE, MIN_SIZE and TIME for size and time needed to handle all
3167 calls in NODE. POSSIBLE_TRUTHS, KNOWN_VALS, KNOWN_AGGS and KNOWN_CONTEXTS
3168 describe context of the call site. */
3170 static void
3171 estimate_calls_size_and_time (struct cgraph_node *node, int *size,
3172 int *min_size, sreal *time,
3173 ipa_hints *hints,
3174 clause_t possible_truths,
3175 vec<tree> known_vals,
3176 vec<ipa_polymorphic_call_context> known_contexts,
3177 vec<ipa_agg_value_set> known_aggs)
3179 class ipa_fn_summary *sum = ipa_fn_summaries->get (node);
3180 bool use_table = true;
3182 gcc_assert (node->callees || node->indirect_calls);
3184 /* During early inlining we do not calculate info for very
3185 large functions and thus there is no need for producing
3186 summaries. */
3187 if (!ipa_node_params_sum)
3188 use_table = false;
3189 /* Do not calculate summaries for simple wrappers; it is waste
3190 of memory. */
3191 else if (node->callees && node->indirect_calls
3192 && node->callees->inline_failed && !node->callees->next_callee)
3193 use_table = false;
3194 /* If there is an indirect edge that may be optimized, we need
3195 to go the slow way. */
3196 else if ((known_vals.length ()
3197 || known_contexts.length ()
3198 || known_aggs.length ()) && hints)
3200 class ipa_node_params *params_summary = IPA_NODE_REF (node);
3201 unsigned int nargs = params_summary
3202 ? ipa_get_param_count (params_summary) : 0;
3204 for (unsigned int i = 0; i < nargs && use_table; i++)
3206 if (ipa_is_param_used_by_indirect_call (params_summary, i)
3207 && ((known_vals.length () > i && known_vals[i])
3208 || (known_aggs.length () > i
3209 && known_aggs[i].items.length ())))
3210 use_table = false;
3211 else if (ipa_is_param_used_by_polymorphic_call (params_summary, i)
3212 && (known_contexts.length () > i
3213 && !known_contexts[i].useless_p ()))
3214 use_table = false;
3218 /* Fast path is via the call size time table. */
3219 if (use_table)
3221 /* Build summary if it is absent. */
3222 if (!sum->call_size_time_table)
3224 predicate true_pred = true;
3225 sum->account_size_time (0, 0, true_pred, true_pred, true);
3226 summarize_calls_size_and_time (node, sum);
3229 int old_size = *size;
3230 sreal old_time = time ? *time : 0;
3232 if (min_size)
3233 *min_size += (*sum->call_size_time_table)[0].size;
3235 unsigned int i;
3236 size_time_entry *e;
3238 /* Walk the table and account sizes and times. */
3239 for (i = 0; vec_safe_iterate (sum->call_size_time_table, i, &e);
3240 i++)
3241 if (e->exec_predicate.evaluate (possible_truths))
3243 *size += e->size;
3244 if (time)
3245 *time += e->time;
3248 /* Be careful and see if both methods agree. */
3249 if ((flag_checking || dump_file)
3250 /* Do not try to sanity check when we know we lost some
3251 precision. */
3252 && sum->call_size_time_table->length ()
3253 < ipa_fn_summary::max_size_time_table_size)
3255 estimate_calls_size_and_time_1 (node, &old_size, NULL, &old_time, NULL,
3256 possible_truths, known_vals,
3257 known_contexts, known_aggs);
3258 gcc_assert (*size == old_size);
3259 if (time && (*time - old_time > 1 || *time - old_time < -1)
3260 && dump_file)
3261 fprintf (dump_file, "Time mismatch in call summary %f!=%f\n",
3262 old_time.to_double (),
3263 time->to_double ());
3266 /* Slow path by walking all edges. */
3267 else
3268 estimate_calls_size_and_time_1 (node, size, min_size, time, hints,
3269 possible_truths, known_vals, known_contexts,
3270 known_aggs);
3273 /* Default constructor for ipa call context.
3274 Memory allocation of known_vals, known_contexts
3275 and known_aggs vectors is owned by the caller, but can
3276 be release by ipa_call_context::release.
3278 inline_param_summary is owned by the caller. */
3279 ipa_call_context::ipa_call_context (cgraph_node *node,
3280 clause_t possible_truths,
3281 clause_t nonspec_possible_truths,
3282 vec<tree> known_vals,
3283 vec<ipa_polymorphic_call_context>
3284 known_contexts,
3285 vec<ipa_agg_value_set> known_aggs,
3286 vec<inline_param_summary>
3287 inline_param_summary)
3288 : m_node (node), m_possible_truths (possible_truths),
3289 m_nonspec_possible_truths (nonspec_possible_truths),
3290 m_inline_param_summary (inline_param_summary),
3291 m_known_vals (known_vals),
3292 m_known_contexts (known_contexts),
3293 m_known_aggs (known_aggs)
3297 /* Set THIS to be a duplicate of CTX. Copy all relevant info. */
3299 void
3300 ipa_call_context::duplicate_from (const ipa_call_context &ctx)
3302 m_node = ctx.m_node;
3303 m_possible_truths = ctx.m_possible_truths;
3304 m_nonspec_possible_truths = ctx.m_nonspec_possible_truths;
3305 class ipa_node_params *params_summary = IPA_NODE_REF (m_node);
3306 unsigned int nargs = params_summary
3307 ? ipa_get_param_count (params_summary) : 0;
3309 m_inline_param_summary = vNULL;
3310 /* Copy the info only if there is at least one useful entry. */
3311 if (ctx.m_inline_param_summary.exists ())
3313 unsigned int n = MIN (ctx.m_inline_param_summary.length (), nargs);
3315 for (unsigned int i = 0; i < n; i++)
3316 if (ipa_is_param_used_by_ipa_predicates (params_summary, i)
3317 && !ctx.m_inline_param_summary[i].useless_p ())
3319 m_inline_param_summary
3320 = ctx.m_inline_param_summary.copy ();
3321 break;
3324 m_known_vals = vNULL;
3325 if (ctx.m_known_vals.exists ())
3327 unsigned int n = MIN (ctx.m_known_vals.length (), nargs);
3329 for (unsigned int i = 0; i < n; i++)
3330 if (ipa_is_param_used_by_indirect_call (params_summary, i)
3331 && ctx.m_known_vals[i])
3333 m_known_vals = ctx.m_known_vals.copy ();
3334 break;
3338 m_known_contexts = vNULL;
3339 if (ctx.m_known_contexts.exists ())
3341 unsigned int n = MIN (ctx.m_known_contexts.length (), nargs);
3343 for (unsigned int i = 0; i < n; i++)
3344 if (ipa_is_param_used_by_polymorphic_call (params_summary, i)
3345 && !ctx.m_known_contexts[i].useless_p ())
3347 m_known_contexts = ctx.m_known_contexts.copy ();
3348 break;
3352 m_known_aggs = vNULL;
3353 if (ctx.m_known_aggs.exists ())
3355 unsigned int n = MIN (ctx.m_known_aggs.length (), nargs);
3357 for (unsigned int i = 0; i < n; i++)
3358 if (ipa_is_param_used_by_indirect_call (params_summary, i)
3359 && !ctx.m_known_aggs[i].is_empty ())
3361 m_known_aggs = ipa_copy_agg_values (ctx.m_known_aggs);
3362 break;
3367 /* Release memory used by known_vals/contexts/aggs vectors.
3368 If ALL is true release also inline_param_summary.
3369 This happens when context was previously duplicated to be stored
3370 into cache. */
3372 void
3373 ipa_call_context::release (bool all)
3375 /* See if context is initialized at first place. */
3376 if (!m_node)
3377 return;
3378 ipa_release_agg_values (m_known_aggs, all);
3379 if (all)
3381 m_known_vals.release ();
3382 m_known_contexts.release ();
3383 m_inline_param_summary.release ();
3387 /* Return true if CTX describes the same call context as THIS. */
3389 bool
3390 ipa_call_context::equal_to (const ipa_call_context &ctx)
3392 if (m_node != ctx.m_node
3393 || m_possible_truths != ctx.m_possible_truths
3394 || m_nonspec_possible_truths != ctx.m_nonspec_possible_truths)
3395 return false;
3397 class ipa_node_params *params_summary = IPA_NODE_REF (m_node);
3398 unsigned int nargs = params_summary
3399 ? ipa_get_param_count (params_summary) : 0;
3401 if (m_inline_param_summary.exists () || ctx.m_inline_param_summary.exists ())
3403 for (unsigned int i = 0; i < nargs; i++)
3405 if (!ipa_is_param_used_by_ipa_predicates (params_summary, i))
3406 continue;
3407 if (i >= m_inline_param_summary.length ()
3408 || m_inline_param_summary[i].useless_p ())
3410 if (i < ctx.m_inline_param_summary.length ()
3411 && !ctx.m_inline_param_summary[i].useless_p ())
3412 return false;
3413 continue;
3415 if (i >= ctx.m_inline_param_summary.length ()
3416 || ctx.m_inline_param_summary[i].useless_p ())
3418 if (i < m_inline_param_summary.length ()
3419 && !m_inline_param_summary[i].useless_p ())
3420 return false;
3421 continue;
3423 if (!m_inline_param_summary[i].equal_to
3424 (ctx.m_inline_param_summary[i]))
3425 return false;
3428 if (m_known_vals.exists () || ctx.m_known_vals.exists ())
3430 for (unsigned int i = 0; i < nargs; i++)
3432 if (!ipa_is_param_used_by_indirect_call (params_summary, i))
3433 continue;
3434 if (i >= m_known_vals.length () || !m_known_vals[i])
3436 if (i < ctx.m_known_vals.length () && ctx.m_known_vals[i])
3437 return false;
3438 continue;
3440 if (i >= ctx.m_known_vals.length () || !ctx.m_known_vals[i])
3442 if (i < m_known_vals.length () && m_known_vals[i])
3443 return false;
3444 continue;
3446 if (m_known_vals[i] != ctx.m_known_vals[i])
3447 return false;
3450 if (m_known_contexts.exists () || ctx.m_known_contexts.exists ())
3452 for (unsigned int i = 0; i < nargs; i++)
3454 if (!ipa_is_param_used_by_polymorphic_call (params_summary, i))
3455 continue;
3456 if (i >= m_known_contexts.length ()
3457 || m_known_contexts[i].useless_p ())
3459 if (i < ctx.m_known_contexts.length ()
3460 && !ctx.m_known_contexts[i].useless_p ())
3461 return false;
3462 continue;
3464 if (i >= ctx.m_known_contexts.length ()
3465 || ctx.m_known_contexts[i].useless_p ())
3467 if (i < m_known_contexts.length ()
3468 && !m_known_contexts[i].useless_p ())
3469 return false;
3470 continue;
3472 if (!m_known_contexts[i].equal_to
3473 (ctx.m_known_contexts[i]))
3474 return false;
3477 if (m_known_aggs.exists () || ctx.m_known_aggs.exists ())
3479 for (unsigned int i = 0; i < nargs; i++)
3481 if (!ipa_is_param_used_by_indirect_call (params_summary, i))
3482 continue;
3483 if (i >= m_known_aggs.length () || m_known_aggs[i].is_empty ())
3485 if (i < ctx.m_known_aggs.length ()
3486 && !ctx.m_known_aggs[i].is_empty ())
3487 return false;
3488 continue;
3490 if (i >= ctx.m_known_aggs.length ()
3491 || ctx.m_known_aggs[i].is_empty ())
3493 if (i < m_known_aggs.length ()
3494 && !m_known_aggs[i].is_empty ())
3495 return false;
3496 continue;
3498 if (!m_known_aggs[i].equal_to (ctx.m_known_aggs[i]))
3499 return false;
3502 return true;
3505 /* Estimate size and time needed to execute call in the given context.
3506 Additionally determine hints determined by the context. Finally compute
3507 minimal size needed for the call that is independent on the call context and
3508 can be used for fast estimates. Return the values in RET_SIZE,
3509 RET_MIN_SIZE, RET_TIME and RET_HINTS. */
3511 void
3512 ipa_call_context::estimate_size_and_time (int *ret_size,
3513 int *ret_min_size,
3514 sreal *ret_time,
3515 sreal *ret_nonspecialized_time,
3516 ipa_hints *ret_hints)
3518 class ipa_fn_summary *info = ipa_fn_summaries->get (m_node);
3519 size_time_entry *e;
3520 int size = 0;
3521 sreal time = 0;
3522 int min_size = 0;
3523 ipa_hints hints = 0;
3524 int i;
3526 if (dump_file && (dump_flags & TDF_DETAILS))
3528 bool found = false;
3529 fprintf (dump_file, " Estimating body: %s\n"
3530 " Known to be false: ", m_node->dump_name ());
3532 for (i = predicate::not_inlined_condition;
3533 i < (predicate::first_dynamic_condition
3534 + (int) vec_safe_length (info->conds)); i++)
3535 if (!(m_possible_truths & (1 << i)))
3537 if (found)
3538 fprintf (dump_file, ", ");
3539 found = true;
3540 dump_condition (dump_file, info->conds, i);
3544 if (m_node->callees || m_node->indirect_calls)
3545 estimate_calls_size_and_time (m_node, &size, &min_size,
3546 ret_time ? &time : NULL,
3547 ret_hints ? &hints : NULL, m_possible_truths,
3548 m_known_vals, m_known_contexts, m_known_aggs);
3550 sreal nonspecialized_time = time;
3552 min_size += (*info->size_time_table)[0].size;
3553 for (i = 0; vec_safe_iterate (info->size_time_table, i, &e); i++)
3555 bool exec = e->exec_predicate.evaluate (m_nonspec_possible_truths);
3557 /* Because predicates are conservative, it can happen that nonconst is 1
3558 but exec is 0. */
3559 if (exec)
3561 bool nonconst = e->nonconst_predicate.evaluate (m_possible_truths);
3563 gcc_checking_assert (e->time >= 0);
3564 gcc_checking_assert (time >= 0);
3566 /* We compute specialized size only because size of nonspecialized
3567 copy is context independent.
3569 The difference between nonspecialized execution and specialized is
3570 that nonspecialized is not going to have optimized out computations
3571 known to be constant in a specialized setting. */
3572 if (nonconst)
3573 size += e->size;
3574 if (!ret_time)
3575 continue;
3576 nonspecialized_time += e->time;
3577 if (!nonconst)
3579 else if (!m_inline_param_summary.exists ())
3581 if (nonconst)
3582 time += e->time;
3584 else
3586 int prob = e->nonconst_predicate.probability
3587 (info->conds, m_possible_truths,
3588 m_inline_param_summary);
3589 gcc_checking_assert (prob >= 0);
3590 gcc_checking_assert (prob <= REG_BR_PROB_BASE);
3591 if (prob == REG_BR_PROB_BASE)
3592 time += e->time;
3593 else
3594 time += e->time * prob / REG_BR_PROB_BASE;
3596 gcc_checking_assert (time >= 0);
3599 gcc_checking_assert ((*info->size_time_table)[0].exec_predicate == true);
3600 gcc_checking_assert ((*info->size_time_table)[0].nonconst_predicate == true);
3601 gcc_checking_assert (min_size >= 0);
3602 gcc_checking_assert (size >= 0);
3603 gcc_checking_assert (time >= 0);
3604 /* nonspecialized_time should be always bigger than specialized time.
3605 Roundoff issues however may get into the way. */
3606 gcc_checking_assert ((nonspecialized_time - time * 99 / 100) >= -1);
3608 /* Roundoff issues may make specialized time bigger than nonspecialized
3609 time. We do not really want that to happen because some heuristics
3610 may get confused by seeing negative speedups. */
3611 if (time > nonspecialized_time)
3612 time = nonspecialized_time;
3614 if (ret_hints)
3616 if (info->loop_iterations
3617 && !info->loop_iterations->evaluate (m_possible_truths))
3618 hints |= INLINE_HINT_loop_iterations;
3619 if (info->loop_stride
3620 && !info->loop_stride->evaluate (m_possible_truths))
3621 hints |= INLINE_HINT_loop_stride;
3622 if (info->scc_no)
3623 hints |= INLINE_HINT_in_scc;
3624 if (DECL_DECLARED_INLINE_P (m_node->decl))
3625 hints |= INLINE_HINT_declared_inline;
3628 size = RDIV (size, ipa_fn_summary::size_scale);
3629 min_size = RDIV (min_size, ipa_fn_summary::size_scale);
3631 if (dump_file && (dump_flags & TDF_DETAILS))
3632 fprintf (dump_file, "\n size:%i time:%f nonspec time:%f\n", (int) size,
3633 time.to_double (), nonspecialized_time.to_double ());
3634 if (ret_time)
3635 *ret_time = time;
3636 if (ret_nonspecialized_time)
3637 *ret_nonspecialized_time = nonspecialized_time;
3638 if (ret_size)
3639 *ret_size = size;
3640 if (ret_min_size)
3641 *ret_min_size = min_size;
3642 if (ret_hints)
3643 *ret_hints = hints;
3644 return;
3648 /* Estimate size and time needed to execute callee of EDGE assuming that
3649 parameters known to be constant at caller of EDGE are propagated.
3650 KNOWN_VALS and KNOWN_CONTEXTS are vectors of assumed known constant values
3651 and types for parameters. */
3653 void
3654 estimate_ipcp_clone_size_and_time (struct cgraph_node *node,
3655 vec<tree> known_vals,
3656 vec<ipa_polymorphic_call_context>
3657 known_contexts,
3658 vec<ipa_agg_value_set> known_aggs,
3659 int *ret_size, sreal *ret_time,
3660 sreal *ret_nonspec_time,
3661 ipa_hints *hints)
3663 clause_t clause, nonspec_clause;
3665 /* TODO: Also pass known value ranges. */
3666 evaluate_conditions_for_known_args (node, false, known_vals, vNULL,
3667 known_aggs, &clause, &nonspec_clause);
3668 ipa_call_context ctx (node, clause, nonspec_clause,
3669 known_vals, known_contexts,
3670 known_aggs, vNULL);
3671 ctx.estimate_size_and_time (ret_size, NULL, ret_time,
3672 ret_nonspec_time, hints);
3675 /* Return stack frame offset where frame of NODE is supposed to start inside
3676 of the function it is inlined to.
3677 Return 0 for functions that are not inlined. */
3679 HOST_WIDE_INT
3680 ipa_get_stack_frame_offset (struct cgraph_node *node)
3682 HOST_WIDE_INT offset = 0;
3683 if (!node->inlined_to)
3684 return 0;
3685 node = node->callers->caller;
3686 while (true)
3688 offset += ipa_size_summaries->get (node)->estimated_self_stack_size;
3689 if (!node->inlined_to)
3690 return offset;
3691 node = node->callers->caller;
3696 /* Update summary information of inline clones after inlining.
3697 Compute peak stack usage. */
3699 static void
3700 inline_update_callee_summaries (struct cgraph_node *node, int depth)
3702 struct cgraph_edge *e;
3704 ipa_propagate_frequency (node);
3705 for (e = node->callees; e; e = e->next_callee)
3707 if (!e->inline_failed)
3708 inline_update_callee_summaries (e->callee, depth);
3709 else
3710 ipa_call_summaries->get (e)->loop_depth += depth;
3712 for (e = node->indirect_calls; e; e = e->next_callee)
3713 ipa_call_summaries->get (e)->loop_depth += depth;
3716 /* Update change_prob of EDGE after INLINED_EDGE has been inlined.
3717 When function A is inlined in B and A calls C with parameter that
3718 changes with probability PROB1 and C is known to be passthrough
3719 of argument if B that change with probability PROB2, the probability
3720 of change is now PROB1*PROB2. */
3722 static void
3723 remap_edge_change_prob (struct cgraph_edge *inlined_edge,
3724 struct cgraph_edge *edge)
3726 if (ipa_node_params_sum)
3728 int i;
3729 class ipa_edge_args *args = IPA_EDGE_REF (edge);
3730 if (!args)
3731 return;
3732 class ipa_call_summary *es = ipa_call_summaries->get (edge);
3733 class ipa_call_summary *inlined_es
3734 = ipa_call_summaries->get (inlined_edge);
3736 if (es->param.length () == 0)
3737 return;
3739 for (i = 0; i < ipa_get_cs_argument_count (args); i++)
3741 struct ipa_jump_func *jfunc = ipa_get_ith_jump_func (args, i);
3742 if (jfunc->type == IPA_JF_PASS_THROUGH
3743 || jfunc->type == IPA_JF_ANCESTOR)
3745 int id = jfunc->type == IPA_JF_PASS_THROUGH
3746 ? ipa_get_jf_pass_through_formal_id (jfunc)
3747 : ipa_get_jf_ancestor_formal_id (jfunc);
3748 if (id < (int) inlined_es->param.length ())
3750 int prob1 = es->param[i].change_prob;
3751 int prob2 = inlined_es->param[id].change_prob;
3752 int prob = combine_probabilities (prob1, prob2);
3754 if (prob1 && prob2 && !prob)
3755 prob = 1;
3757 es->param[i].change_prob = prob;
3764 /* Update edge summaries of NODE after INLINED_EDGE has been inlined.
3766 Remap predicates of callees of NODE. Rest of arguments match
3767 remap_predicate.
3769 Also update change probabilities. */
3771 static void
3772 remap_edge_summaries (struct cgraph_edge *inlined_edge,
3773 struct cgraph_node *node,
3774 class ipa_fn_summary *info,
3775 class ipa_node_params *params_summary,
3776 class ipa_fn_summary *callee_info,
3777 vec<int> operand_map,
3778 vec<int> offset_map,
3779 clause_t possible_truths,
3780 predicate *toplev_predicate)
3782 struct cgraph_edge *e, *next;
3783 for (e = node->callees; e; e = next)
3785 predicate p;
3786 next = e->next_callee;
3788 if (e->inline_failed)
3790 class ipa_call_summary *es = ipa_call_summaries->get (e);
3791 remap_edge_change_prob (inlined_edge, e);
3793 if (es->predicate)
3795 p = es->predicate->remap_after_inlining
3796 (info, params_summary,
3797 callee_info, operand_map,
3798 offset_map, possible_truths,
3799 *toplev_predicate);
3800 edge_set_predicate (e, &p);
3802 else
3803 edge_set_predicate (e, toplev_predicate);
3805 else
3806 remap_edge_summaries (inlined_edge, e->callee, info,
3807 params_summary, callee_info,
3808 operand_map, offset_map, possible_truths,
3809 toplev_predicate);
3811 for (e = node->indirect_calls; e; e = next)
3813 class ipa_call_summary *es = ipa_call_summaries->get (e);
3814 predicate p;
3815 next = e->next_callee;
3817 remap_edge_change_prob (inlined_edge, e);
3818 if (es->predicate)
3820 p = es->predicate->remap_after_inlining
3821 (info, params_summary,
3822 callee_info, operand_map, offset_map,
3823 possible_truths, *toplev_predicate);
3824 edge_set_predicate (e, &p);
3826 else
3827 edge_set_predicate (e, toplev_predicate);
3831 /* Same as remap_predicate, but set result into hint *HINT. */
3833 static void
3834 remap_hint_predicate (class ipa_fn_summary *info,
3835 class ipa_node_params *params_summary,
3836 class ipa_fn_summary *callee_info,
3837 predicate **hint,
3838 vec<int> operand_map,
3839 vec<int> offset_map,
3840 clause_t possible_truths,
3841 predicate *toplev_predicate)
3843 predicate p;
3845 if (!*hint)
3846 return;
3847 p = (*hint)->remap_after_inlining
3848 (info, params_summary, callee_info,
3849 operand_map, offset_map,
3850 possible_truths, *toplev_predicate);
3851 if (p != false && p != true)
3853 if (!*hint)
3854 set_hint_predicate (hint, p);
3855 else
3856 **hint &= p;
3860 /* We inlined EDGE. Update summary of the function we inlined into. */
3862 void
3863 ipa_merge_fn_summary_after_inlining (struct cgraph_edge *edge)
3865 ipa_fn_summary *callee_info = ipa_fn_summaries->get (edge->callee);
3866 struct cgraph_node *to = (edge->caller->inlined_to
3867 ? edge->caller->inlined_to : edge->caller);
3868 class ipa_fn_summary *info = ipa_fn_summaries->get (to);
3869 clause_t clause = 0; /* not_inline is known to be false. */
3870 size_time_entry *e;
3871 auto_vec<int, 8> operand_map;
3872 auto_vec<int, 8> offset_map;
3873 int i;
3874 predicate toplev_predicate;
3875 class ipa_call_summary *es = ipa_call_summaries->get (edge);
3876 class ipa_node_params *params_summary = (ipa_node_params_sum
3877 ? IPA_NODE_REF (to) : NULL);
3879 if (es->predicate)
3880 toplev_predicate = *es->predicate;
3881 else
3882 toplev_predicate = true;
3884 info->fp_expressions |= callee_info->fp_expressions;
3886 if (callee_info->conds)
3888 auto_vec<tree, 32> known_vals;
3889 auto_vec<ipa_agg_value_set, 32> known_aggs;
3890 evaluate_properties_for_edge (edge, true, &clause, NULL,
3891 &known_vals, NULL, &known_aggs);
3893 if (ipa_node_params_sum && callee_info->conds)
3895 class ipa_edge_args *args = IPA_EDGE_REF (edge);
3896 int count = args ? ipa_get_cs_argument_count (args) : 0;
3897 int i;
3899 if (count)
3901 operand_map.safe_grow_cleared (count);
3902 offset_map.safe_grow_cleared (count);
3904 for (i = 0; i < count; i++)
3906 struct ipa_jump_func *jfunc = ipa_get_ith_jump_func (args, i);
3907 int map = -1;
3909 /* TODO: handle non-NOPs when merging. */
3910 if (jfunc->type == IPA_JF_PASS_THROUGH)
3912 if (ipa_get_jf_pass_through_operation (jfunc) == NOP_EXPR)
3913 map = ipa_get_jf_pass_through_formal_id (jfunc);
3914 if (!ipa_get_jf_pass_through_agg_preserved (jfunc))
3915 offset_map[i] = -1;
3917 else if (jfunc->type == IPA_JF_ANCESTOR)
3919 HOST_WIDE_INT offset = ipa_get_jf_ancestor_offset (jfunc);
3920 if (offset >= 0 && offset < INT_MAX)
3922 map = ipa_get_jf_ancestor_formal_id (jfunc);
3923 if (!ipa_get_jf_ancestor_agg_preserved (jfunc))
3924 offset = -1;
3925 offset_map[i] = offset;
3928 operand_map[i] = map;
3929 gcc_assert (map < ipa_get_param_count (params_summary));
3932 sreal freq = edge->sreal_frequency ();
3933 for (i = 0; vec_safe_iterate (callee_info->size_time_table, i, &e); i++)
3935 predicate p;
3936 p = e->exec_predicate.remap_after_inlining
3937 (info, params_summary,
3938 callee_info, operand_map,
3939 offset_map, clause,
3940 toplev_predicate);
3941 predicate nonconstp;
3942 nonconstp = e->nonconst_predicate.remap_after_inlining
3943 (info, params_summary,
3944 callee_info, operand_map,
3945 offset_map, clause,
3946 toplev_predicate);
3947 if (p != false && nonconstp != false)
3949 sreal add_time = ((sreal)e->time * freq);
3950 int prob = e->nonconst_predicate.probability (callee_info->conds,
3951 clause, es->param);
3952 if (prob != REG_BR_PROB_BASE)
3953 add_time = add_time * prob / REG_BR_PROB_BASE;
3954 if (prob != REG_BR_PROB_BASE
3955 && dump_file && (dump_flags & TDF_DETAILS))
3957 fprintf (dump_file, "\t\tScaling time by probability:%f\n",
3958 (double) prob / REG_BR_PROB_BASE);
3960 info->account_size_time (e->size, add_time, p, nonconstp);
3963 remap_edge_summaries (edge, edge->callee, info, params_summary,
3964 callee_info, operand_map,
3965 offset_map, clause, &toplev_predicate);
3966 remap_hint_predicate (info, params_summary, callee_info,
3967 &callee_info->loop_iterations,
3968 operand_map, offset_map, clause, &toplev_predicate);
3969 remap_hint_predicate (info, params_summary, callee_info,
3970 &callee_info->loop_stride,
3971 operand_map, offset_map, clause, &toplev_predicate);
3973 HOST_WIDE_INT stack_frame_offset = ipa_get_stack_frame_offset (edge->callee);
3974 HOST_WIDE_INT peak = stack_frame_offset + callee_info->estimated_stack_size;
3976 if (info->estimated_stack_size < peak)
3977 info->estimated_stack_size = peak;
3979 inline_update_callee_summaries (edge->callee, es->loop_depth);
3980 if (info->call_size_time_table)
3982 int edge_size = 0;
3983 sreal edge_time = 0;
3985 estimate_edge_size_and_time (edge, &edge_size, NULL, &edge_time, vNULL,
3986 vNULL, vNULL, 0);
3987 /* Unaccount size and time of the optimized out call. */
3988 info->account_size_time (-edge_size, -edge_time,
3989 es->predicate ? *es->predicate : true,
3990 es->predicate ? *es->predicate : true,
3991 true);
3992 /* Account new calls. */
3993 summarize_calls_size_and_time (edge->callee, info);
3996 /* Free summaries that are not maintained for inline clones/edges. */
3997 ipa_call_summaries->remove (edge);
3998 ipa_fn_summaries->remove (edge->callee);
3999 ipa_remove_from_growth_caches (edge);
4002 /* For performance reasons ipa_merge_fn_summary_after_inlining is not updating
4003 overall size and time. Recompute it.
4004 If RESET is true also recompute call_time_size_table. */
4006 void
4007 ipa_update_overall_fn_summary (struct cgraph_node *node, bool reset)
4009 class ipa_fn_summary *info = ipa_fn_summaries->get (node);
4010 class ipa_size_summary *size_info = ipa_size_summaries->get (node);
4011 size_time_entry *e;
4012 int i;
4014 size_info->size = 0;
4015 info->time = 0;
4016 for (i = 0; vec_safe_iterate (info->size_time_table, i, &e); i++)
4018 size_info->size += e->size;
4019 info->time += e->time;
4021 info->min_size = (*info->size_time_table)[0].size;
4022 if (reset)
4023 vec_free (info->call_size_time_table);
4024 if (node->callees || node->indirect_calls)
4025 estimate_calls_size_and_time (node, &size_info->size, &info->min_size,
4026 &info->time, NULL,
4027 ~(clause_t) (1 << predicate::false_condition),
4028 vNULL, vNULL, vNULL);
4029 size_info->size = RDIV (size_info->size, ipa_fn_summary::size_scale);
4030 info->min_size = RDIV (info->min_size, ipa_fn_summary::size_scale);
4034 /* This function performs intraprocedural analysis in NODE that is required to
4035 inline indirect calls. */
4037 static void
4038 inline_indirect_intraprocedural_analysis (struct cgraph_node *node)
4040 ipa_analyze_node (node);
4041 if (dump_file && (dump_flags & TDF_DETAILS))
4043 ipa_print_node_params (dump_file, node);
4044 ipa_print_node_jump_functions (dump_file, node);
4049 /* Note function body size. */
4051 void
4052 inline_analyze_function (struct cgraph_node *node)
4054 push_cfun (DECL_STRUCT_FUNCTION (node->decl));
4056 if (dump_file)
4057 fprintf (dump_file, "\nAnalyzing function: %s\n", node->dump_name ());
4058 if (opt_for_fn (node->decl, optimize) && !node->thunk.thunk_p)
4059 inline_indirect_intraprocedural_analysis (node);
4060 compute_fn_summary (node, false);
4061 if (!optimize)
4063 struct cgraph_edge *e;
4064 for (e = node->callees; e; e = e->next_callee)
4065 e->inline_failed = CIF_FUNCTION_NOT_OPTIMIZED;
4066 for (e = node->indirect_calls; e; e = e->next_callee)
4067 e->inline_failed = CIF_FUNCTION_NOT_OPTIMIZED;
4070 pop_cfun ();
4074 /* Called when new function is inserted to callgraph late. */
4076 void
4077 ipa_fn_summary_t::insert (struct cgraph_node *node, ipa_fn_summary *)
4079 inline_analyze_function (node);
4082 /* Note function body size. */
4084 static void
4085 ipa_fn_summary_generate (void)
4087 struct cgraph_node *node;
4089 FOR_EACH_DEFINED_FUNCTION (node)
4090 if (DECL_STRUCT_FUNCTION (node->decl))
4091 node->versionable = tree_versionable_function_p (node->decl);
4093 ipa_fn_summary_alloc ();
4095 ipa_fn_summaries->enable_insertion_hook ();
4097 ipa_register_cgraph_hooks ();
4099 FOR_EACH_DEFINED_FUNCTION (node)
4100 if (!node->alias
4101 && (flag_generate_lto || flag_generate_offload|| flag_wpa
4102 || opt_for_fn (node->decl, optimize)))
4103 inline_analyze_function (node);
4107 /* Write inline summary for edge E to OB. */
4109 static void
4110 read_ipa_call_summary (class lto_input_block *ib, struct cgraph_edge *e,
4111 bool prevails)
4113 class ipa_call_summary *es = prevails
4114 ? ipa_call_summaries->get_create (e) : NULL;
4115 predicate p;
4116 int length, i;
4118 int size = streamer_read_uhwi (ib);
4119 int time = streamer_read_uhwi (ib);
4120 int depth = streamer_read_uhwi (ib);
4122 if (es)
4124 es->call_stmt_size = size;
4125 es->call_stmt_time = time;
4126 es->loop_depth = depth;
4129 bitpack_d bp = streamer_read_bitpack (ib);
4130 if (es)
4131 es->is_return_callee_uncaptured = bp_unpack_value (&bp, 1);
4132 else
4133 bp_unpack_value (&bp, 1);
4135 p.stream_in (ib);
4136 if (es)
4137 edge_set_predicate (e, &p);
4138 length = streamer_read_uhwi (ib);
4139 if (length && es && e->possibly_call_in_translation_unit_p ())
4141 es->param.safe_grow_cleared (length);
4142 for (i = 0; i < length; i++)
4143 es->param[i].change_prob = streamer_read_uhwi (ib);
4145 else
4147 for (i = 0; i < length; i++)
4148 streamer_read_uhwi (ib);
4153 /* Stream in inline summaries from the section. */
4155 static void
4156 inline_read_section (struct lto_file_decl_data *file_data, const char *data,
4157 size_t len)
4159 const struct lto_function_header *header =
4160 (const struct lto_function_header *) data;
4161 const int cfg_offset = sizeof (struct lto_function_header);
4162 const int main_offset = cfg_offset + header->cfg_size;
4163 const int string_offset = main_offset + header->main_size;
4164 class data_in *data_in;
4165 unsigned int i, count2, j;
4166 unsigned int f_count;
4168 lto_input_block ib ((const char *) data + main_offset, header->main_size,
4169 file_data->mode_table);
4171 data_in =
4172 lto_data_in_create (file_data, (const char *) data + string_offset,
4173 header->string_size, vNULL);
4174 f_count = streamer_read_uhwi (&ib);
4175 for (i = 0; i < f_count; i++)
4177 unsigned int index;
4178 struct cgraph_node *node;
4179 class ipa_fn_summary *info;
4180 class ipa_node_params *params_summary;
4181 class ipa_size_summary *size_info;
4182 lto_symtab_encoder_t encoder;
4183 struct bitpack_d bp;
4184 struct cgraph_edge *e;
4185 predicate p;
4187 index = streamer_read_uhwi (&ib);
4188 encoder = file_data->symtab_node_encoder;
4189 node = dyn_cast<cgraph_node *> (lto_symtab_encoder_deref (encoder,
4190 index));
4191 info = node->prevailing_p () ? ipa_fn_summaries->get_create (node) : NULL;
4192 params_summary = node->prevailing_p () ? IPA_NODE_REF (node) : NULL;
4193 size_info = node->prevailing_p ()
4194 ? ipa_size_summaries->get_create (node) : NULL;
4196 int stack_size = streamer_read_uhwi (&ib);
4197 int size = streamer_read_uhwi (&ib);
4198 sreal time = sreal::stream_in (&ib);
4200 if (info)
4202 info->estimated_stack_size
4203 = size_info->estimated_self_stack_size = stack_size;
4204 size_info->size = size_info->self_size = size;
4205 info->time = time;
4208 bp = streamer_read_bitpack (&ib);
4209 if (info)
4211 info->inlinable = bp_unpack_value (&bp, 1);
4212 info->fp_expressions = bp_unpack_value (&bp, 1);
4214 else
4216 bp_unpack_value (&bp, 1);
4217 bp_unpack_value (&bp, 1);
4220 count2 = streamer_read_uhwi (&ib);
4221 gcc_assert (!info || !info->conds);
4222 if (info)
4223 vec_safe_reserve_exact (info->conds, count2);
4224 for (j = 0; j < count2; j++)
4226 struct condition c;
4227 unsigned int k, count3;
4228 c.operand_num = streamer_read_uhwi (&ib);
4229 c.code = (enum tree_code) streamer_read_uhwi (&ib);
4230 c.type = stream_read_tree (&ib, data_in);
4231 c.val = stream_read_tree (&ib, data_in);
4232 bp = streamer_read_bitpack (&ib);
4233 c.agg_contents = bp_unpack_value (&bp, 1);
4234 c.by_ref = bp_unpack_value (&bp, 1);
4235 if (c.agg_contents)
4236 c.offset = streamer_read_uhwi (&ib);
4237 count3 = streamer_read_uhwi (&ib);
4238 c.param_ops = NULL;
4239 if (info)
4240 vec_safe_reserve_exact (c.param_ops, count3);
4241 if (params_summary)
4242 ipa_set_param_used_by_ipa_predicates
4243 (params_summary, c.operand_num, true);
4244 for (k = 0; k < count3; k++)
4246 struct expr_eval_op op;
4247 enum gimple_rhs_class rhs_class;
4248 op.code = (enum tree_code) streamer_read_uhwi (&ib);
4249 op.type = stream_read_tree (&ib, data_in);
4250 switch (rhs_class = get_gimple_rhs_class (op.code))
4252 case GIMPLE_UNARY_RHS:
4253 op.index = 0;
4254 op.val[0] = NULL_TREE;
4255 op.val[1] = NULL_TREE;
4256 break;
4258 case GIMPLE_BINARY_RHS:
4259 case GIMPLE_TERNARY_RHS:
4260 bp = streamer_read_bitpack (&ib);
4261 op.index = bp_unpack_value (&bp, 2);
4262 op.val[0] = stream_read_tree (&ib, data_in);
4263 if (rhs_class == GIMPLE_BINARY_RHS)
4264 op.val[1] = NULL_TREE;
4265 else
4266 op.val[1] = stream_read_tree (&ib, data_in);
4267 break;
4269 default:
4270 fatal_error (UNKNOWN_LOCATION,
4271 "invalid fnsummary in LTO stream");
4273 if (info)
4274 c.param_ops->quick_push (op);
4276 if (info)
4277 info->conds->quick_push (c);
4279 count2 = streamer_read_uhwi (&ib);
4280 gcc_assert (!info || !info->size_time_table);
4281 if (info && count2)
4282 vec_safe_reserve_exact (info->size_time_table, count2);
4283 for (j = 0; j < count2; j++)
4285 class size_time_entry e;
4287 e.size = streamer_read_uhwi (&ib);
4288 e.time = sreal::stream_in (&ib);
4289 e.exec_predicate.stream_in (&ib);
4290 e.nonconst_predicate.stream_in (&ib);
4292 if (info)
4293 info->size_time_table->quick_push (e);
4296 p.stream_in (&ib);
4297 if (info)
4298 set_hint_predicate (&info->loop_iterations, p);
4299 p.stream_in (&ib);
4300 if (info)
4301 set_hint_predicate (&info->loop_stride, p);
4302 for (e = node->callees; e; e = e->next_callee)
4303 read_ipa_call_summary (&ib, e, info != NULL);
4304 for (e = node->indirect_calls; e; e = e->next_callee)
4305 read_ipa_call_summary (&ib, e, info != NULL);
4308 lto_free_section_data (file_data, LTO_section_ipa_fn_summary, NULL, data,
4309 len);
4310 lto_data_in_delete (data_in);
4314 /* Read inline summary. Jump functions are shared among ipa-cp
4315 and inliner, so when ipa-cp is active, we don't need to write them
4316 twice. */
4318 static void
4319 ipa_fn_summary_read (void)
4321 struct lto_file_decl_data **file_data_vec = lto_get_file_decl_data ();
4322 struct lto_file_decl_data *file_data;
4323 unsigned int j = 0;
4325 ipa_fn_summary_alloc ();
4327 while ((file_data = file_data_vec[j++]))
4329 size_t len;
4330 const char *data
4331 = lto_get_summary_section_data (file_data, LTO_section_ipa_fn_summary,
4332 &len);
4333 if (data)
4334 inline_read_section (file_data, data, len);
4335 else
4336 /* Fatal error here. We do not want to support compiling ltrans units
4337 with different version of compiler or different flags than the WPA
4338 unit, so this should never happen. */
4339 fatal_error (input_location,
4340 "ipa inline summary is missing in input file");
4342 ipa_register_cgraph_hooks ();
4343 if (!flag_ipa_cp)
4344 ipa_prop_read_jump_functions ();
4346 gcc_assert (ipa_fn_summaries);
4347 ipa_fn_summaries->enable_insertion_hook ();
4351 /* Write inline summary for edge E to OB. */
4353 static void
4354 write_ipa_call_summary (struct output_block *ob, struct cgraph_edge *e)
4356 class ipa_call_summary *es = ipa_call_summaries->get (e);
4357 int i;
4359 streamer_write_uhwi (ob, es->call_stmt_size);
4360 streamer_write_uhwi (ob, es->call_stmt_time);
4361 streamer_write_uhwi (ob, es->loop_depth);
4363 bitpack_d bp = bitpack_create (ob->main_stream);
4364 bp_pack_value (&bp, es->is_return_callee_uncaptured, 1);
4365 streamer_write_bitpack (&bp);
4367 if (es->predicate)
4368 es->predicate->stream_out (ob);
4369 else
4370 streamer_write_uhwi (ob, 0);
4371 streamer_write_uhwi (ob, es->param.length ());
4372 for (i = 0; i < (int) es->param.length (); i++)
4373 streamer_write_uhwi (ob, es->param[i].change_prob);
4377 /* Write inline summary for node in SET.
4378 Jump functions are shared among ipa-cp and inliner, so when ipa-cp is
4379 active, we don't need to write them twice. */
4381 static void
4382 ipa_fn_summary_write (void)
4384 struct output_block *ob = create_output_block (LTO_section_ipa_fn_summary);
4385 lto_symtab_encoder_iterator lsei;
4386 lto_symtab_encoder_t encoder = ob->decl_state->symtab_node_encoder;
4387 unsigned int count = 0;
4389 for (lsei = lsei_start_function_in_partition (encoder); !lsei_end_p (lsei);
4390 lsei_next_function_in_partition (&lsei))
4392 cgraph_node *cnode = lsei_cgraph_node (lsei);
4393 if (cnode->definition && !cnode->alias)
4394 count++;
4396 streamer_write_uhwi (ob, count);
4398 for (lsei = lsei_start_function_in_partition (encoder); !lsei_end_p (lsei);
4399 lsei_next_function_in_partition (&lsei))
4401 cgraph_node *cnode = lsei_cgraph_node (lsei);
4402 if (cnode->definition && !cnode->alias)
4404 class ipa_fn_summary *info = ipa_fn_summaries->get (cnode);
4405 class ipa_size_summary *size_info = ipa_size_summaries->get (cnode);
4406 struct bitpack_d bp;
4407 struct cgraph_edge *edge;
4408 int i;
4409 size_time_entry *e;
4410 struct condition *c;
4412 streamer_write_uhwi (ob, lto_symtab_encoder_encode (encoder, cnode));
4413 streamer_write_hwi (ob, size_info->estimated_self_stack_size);
4414 streamer_write_hwi (ob, size_info->self_size);
4415 info->time.stream_out (ob);
4416 bp = bitpack_create (ob->main_stream);
4417 bp_pack_value (&bp, info->inlinable, 1);
4418 bp_pack_value (&bp, false, 1);
4419 bp_pack_value (&bp, info->fp_expressions, 1);
4420 streamer_write_bitpack (&bp);
4421 streamer_write_uhwi (ob, vec_safe_length (info->conds));
4422 for (i = 0; vec_safe_iterate (info->conds, i, &c); i++)
4424 int j;
4425 struct expr_eval_op *op;
4427 streamer_write_uhwi (ob, c->operand_num);
4428 streamer_write_uhwi (ob, c->code);
4429 stream_write_tree (ob, c->type, true);
4430 stream_write_tree (ob, c->val, true);
4431 bp = bitpack_create (ob->main_stream);
4432 bp_pack_value (&bp, c->agg_contents, 1);
4433 bp_pack_value (&bp, c->by_ref, 1);
4434 streamer_write_bitpack (&bp);
4435 if (c->agg_contents)
4436 streamer_write_uhwi (ob, c->offset);
4437 streamer_write_uhwi (ob, vec_safe_length (c->param_ops));
4438 for (j = 0; vec_safe_iterate (c->param_ops, j, &op); j++)
4440 streamer_write_uhwi (ob, op->code);
4441 stream_write_tree (ob, op->type, true);
4442 if (op->val[0])
4444 bp = bitpack_create (ob->main_stream);
4445 bp_pack_value (&bp, op->index, 2);
4446 streamer_write_bitpack (&bp);
4447 stream_write_tree (ob, op->val[0], true);
4448 if (op->val[1])
4449 stream_write_tree (ob, op->val[1], true);
4453 streamer_write_uhwi (ob, vec_safe_length (info->size_time_table));
4454 for (i = 0; vec_safe_iterate (info->size_time_table, i, &e); i++)
4456 streamer_write_uhwi (ob, e->size);
4457 e->time.stream_out (ob);
4458 e->exec_predicate.stream_out (ob);
4459 e->nonconst_predicate.stream_out (ob);
4461 if (info->loop_iterations)
4462 info->loop_iterations->stream_out (ob);
4463 else
4464 streamer_write_uhwi (ob, 0);
4465 if (info->loop_stride)
4466 info->loop_stride->stream_out (ob);
4467 else
4468 streamer_write_uhwi (ob, 0);
4469 for (edge = cnode->callees; edge; edge = edge->next_callee)
4470 write_ipa_call_summary (ob, edge);
4471 for (edge = cnode->indirect_calls; edge; edge = edge->next_callee)
4472 write_ipa_call_summary (ob, edge);
4475 streamer_write_char_stream (ob->main_stream, 0);
4476 produce_asm (ob, NULL);
4477 destroy_output_block (ob);
4479 if (!flag_ipa_cp)
4480 ipa_prop_write_jump_functions ();
4484 /* Release function summary. */
4486 void
4487 ipa_free_fn_summary (void)
4489 if (!ipa_call_summaries)
4490 return;
4491 ggc_delete (ipa_fn_summaries);
4492 ipa_fn_summaries = NULL;
4493 delete ipa_call_summaries;
4494 ipa_call_summaries = NULL;
4495 edge_predicate_pool.release ();
4496 /* During IPA this is one of largest datastructures to release. */
4497 if (flag_wpa)
4498 ggc_trim ();
4501 /* Release function summary. */
4503 void
4504 ipa_free_size_summary (void)
4506 if (!ipa_size_summaries)
4507 return;
4508 delete ipa_size_summaries;
4509 ipa_size_summaries = NULL;
4512 namespace {
4514 const pass_data pass_data_local_fn_summary =
4516 GIMPLE_PASS, /* type */
4517 "local-fnsummary", /* name */
4518 OPTGROUP_INLINE, /* optinfo_flags */
4519 TV_INLINE_PARAMETERS, /* tv_id */
4520 0, /* properties_required */
4521 0, /* properties_provided */
4522 0, /* properties_destroyed */
4523 0, /* todo_flags_start */
4524 0, /* todo_flags_finish */
4527 class pass_local_fn_summary : public gimple_opt_pass
4529 public:
4530 pass_local_fn_summary (gcc::context *ctxt)
4531 : gimple_opt_pass (pass_data_local_fn_summary, ctxt)
4534 /* opt_pass methods: */
4535 opt_pass * clone () { return new pass_local_fn_summary (m_ctxt); }
4536 virtual unsigned int execute (function *)
4538 return compute_fn_summary_for_current ();
4541 }; // class pass_local_fn_summary
4543 } // anon namespace
4545 gimple_opt_pass *
4546 make_pass_local_fn_summary (gcc::context *ctxt)
4548 return new pass_local_fn_summary (ctxt);
4552 /* Free inline summary. */
4554 namespace {
4556 const pass_data pass_data_ipa_free_fn_summary =
4558 SIMPLE_IPA_PASS, /* type */
4559 "free-fnsummary", /* name */
4560 OPTGROUP_NONE, /* optinfo_flags */
4561 TV_IPA_FREE_INLINE_SUMMARY, /* tv_id */
4562 0, /* properties_required */
4563 0, /* properties_provided */
4564 0, /* properties_destroyed */
4565 0, /* todo_flags_start */
4566 0, /* todo_flags_finish */
4569 class pass_ipa_free_fn_summary : public simple_ipa_opt_pass
4571 public:
4572 pass_ipa_free_fn_summary (gcc::context *ctxt)
4573 : simple_ipa_opt_pass (pass_data_ipa_free_fn_summary, ctxt),
4574 small_p (false)
4577 /* opt_pass methods: */
4578 opt_pass *clone () { return new pass_ipa_free_fn_summary (m_ctxt); }
4579 void set_pass_param (unsigned int n, bool param)
4581 gcc_assert (n == 0);
4582 small_p = param;
4584 virtual bool gate (function *) { return true; }
4585 virtual unsigned int execute (function *)
4587 ipa_free_fn_summary ();
4588 if (!flag_wpa)
4589 ipa_free_size_summary ();
4590 return 0;
4593 private:
4594 bool small_p;
4595 }; // class pass_ipa_free_fn_summary
4597 } // anon namespace
4599 simple_ipa_opt_pass *
4600 make_pass_ipa_free_fn_summary (gcc::context *ctxt)
4602 return new pass_ipa_free_fn_summary (ctxt);
4605 namespace {
4607 const pass_data pass_data_ipa_fn_summary =
4609 IPA_PASS, /* type */
4610 "fnsummary", /* name */
4611 OPTGROUP_INLINE, /* optinfo_flags */
4612 TV_IPA_FNSUMMARY, /* tv_id */
4613 0, /* properties_required */
4614 0, /* properties_provided */
4615 0, /* properties_destroyed */
4616 0, /* todo_flags_start */
4617 ( TODO_dump_symtab ), /* todo_flags_finish */
4620 class pass_ipa_fn_summary : public ipa_opt_pass_d
4622 public:
4623 pass_ipa_fn_summary (gcc::context *ctxt)
4624 : ipa_opt_pass_d (pass_data_ipa_fn_summary, ctxt,
4625 ipa_fn_summary_generate, /* generate_summary */
4626 ipa_fn_summary_write, /* write_summary */
4627 ipa_fn_summary_read, /* read_summary */
4628 NULL, /* write_optimization_summary */
4629 NULL, /* read_optimization_summary */
4630 NULL, /* stmt_fixup */
4631 0, /* function_transform_todo_flags_start */
4632 NULL, /* function_transform */
4633 NULL) /* variable_transform */
4636 /* opt_pass methods: */
4637 virtual unsigned int execute (function *) { return 0; }
4639 }; // class pass_ipa_fn_summary
4641 } // anon namespace
4643 ipa_opt_pass_d *
4644 make_pass_ipa_fn_summary (gcc::context *ctxt)
4646 return new pass_ipa_fn_summary (ctxt);
4649 /* Reset all state within ipa-fnsummary.c so that we can rerun the compiler
4650 within the same process. For use by toplev::finalize. */
4652 void
4653 ipa_fnsummary_c_finalize (void)
4655 ipa_free_fn_summary ();