Daily bump.
[official-gcc.git] / gcc / ipa-utils.c
blob708710d61353a3c921dc61a5190c308e4bab1a00
1 /* Utilities for ipa analysis.
2 Copyright (C) 2005-2017 Free Software Foundation, Inc.
3 Contributed by Kenneth Zadeck <zadeck@naturalbridge.com>
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 #include "config.h"
22 #include "system.h"
23 #include "coretypes.h"
24 #include "backend.h"
25 #include "tree.h"
26 #include "gimple.h"
27 #include "predict.h"
28 #include "alloc-pool.h"
29 #include "cgraph.h"
30 #include "lto-streamer.h"
31 #include "dumpfile.h"
32 #include "splay-tree.h"
33 #include "ipa-utils.h"
34 #include "symbol-summary.h"
35 #include "tree-vrp.h"
36 #include "ipa-prop.h"
37 #include "ipa-fnsummary.h"
39 /* Debugging function for postorder and inorder code. NOTE is a string
40 that is printed before the nodes are printed. ORDER is an array of
41 cgraph_nodes that has COUNT useful nodes in it. */
43 void
44 ipa_print_order (FILE* out,
45 const char * note,
46 struct cgraph_node** order,
47 int count)
49 int i;
50 fprintf (out, "\n\n ordered call graph: %s\n", note);
52 for (i = count - 1; i >= 0; i--)
53 order[i]->dump (out);
54 fprintf (out, "\n");
55 fflush (out);
59 struct searchc_env {
60 struct cgraph_node **stack;
61 struct cgraph_node **result;
62 int stack_size;
63 int order_pos;
64 splay_tree nodes_marked_new;
65 bool reduce;
66 bool allow_overwritable;
67 int count;
70 /* This is an implementation of Tarjan's strongly connected region
71 finder as reprinted in Aho Hopcraft and Ullman's The Design and
72 Analysis of Computer Programs (1975) pages 192-193. This version
73 has been customized for cgraph_nodes. The env parameter is because
74 it is recursive and there are no nested functions here. This
75 function should only be called from itself or
76 ipa_reduced_postorder. ENV is a stack env and would be
77 unnecessary if C had nested functions. V is the node to start
78 searching from. */
80 static void
81 searchc (struct searchc_env* env, struct cgraph_node *v,
82 bool (*ignore_edge) (struct cgraph_edge *))
84 struct cgraph_edge *edge;
85 struct ipa_dfs_info *v_info = (struct ipa_dfs_info *) v->aux;
87 /* mark node as old */
88 v_info->new_node = false;
89 splay_tree_remove (env->nodes_marked_new, v->uid);
91 v_info->dfn_number = env->count;
92 v_info->low_link = env->count;
93 env->count++;
94 env->stack[(env->stack_size)++] = v;
95 v_info->on_stack = true;
97 for (edge = v->callees; edge; edge = edge->next_callee)
99 struct ipa_dfs_info * w_info;
100 enum availability avail;
101 struct cgraph_node *w = edge->callee->ultimate_alias_target (&avail);
103 if (!w || (ignore_edge && ignore_edge (edge)))
104 continue;
106 if (w->aux
107 && (avail > AVAIL_INTERPOSABLE
108 || (env->allow_overwritable && avail == AVAIL_INTERPOSABLE)))
110 w_info = (struct ipa_dfs_info *) w->aux;
111 if (w_info->new_node)
113 searchc (env, w, ignore_edge);
114 v_info->low_link =
115 (v_info->low_link < w_info->low_link) ?
116 v_info->low_link : w_info->low_link;
118 else
119 if ((w_info->dfn_number < v_info->dfn_number)
120 && (w_info->on_stack))
121 v_info->low_link =
122 (w_info->dfn_number < v_info->low_link) ?
123 w_info->dfn_number : v_info->low_link;
128 if (v_info->low_link == v_info->dfn_number)
130 struct cgraph_node *last = NULL;
131 struct cgraph_node *x;
132 struct ipa_dfs_info *x_info;
133 do {
134 x = env->stack[--(env->stack_size)];
135 x_info = (struct ipa_dfs_info *) x->aux;
136 x_info->on_stack = false;
137 x_info->scc_no = v_info->dfn_number;
139 if (env->reduce)
141 x_info->next_cycle = last;
142 last = x;
144 else
145 env->result[env->order_pos++] = x;
147 while (v != x);
148 if (env->reduce)
149 env->result[env->order_pos++] = v;
153 /* Topsort the call graph by caller relation. Put the result in ORDER.
155 The REDUCE flag is true if you want the cycles reduced to single nodes.
156 You can use ipa_get_nodes_in_cycle to obtain a vector containing all real
157 call graph nodes in a reduced node.
159 Set ALLOW_OVERWRITABLE if nodes with such availability should be included.
160 IGNORE_EDGE, if non-NULL is a hook that may make some edges insignificant
161 for the topological sort. */
164 ipa_reduced_postorder (struct cgraph_node **order,
165 bool reduce, bool allow_overwritable,
166 bool (*ignore_edge) (struct cgraph_edge *))
168 struct cgraph_node *node;
169 struct searchc_env env;
170 splay_tree_node result;
171 env.stack = XCNEWVEC (struct cgraph_node *, symtab->cgraph_count);
172 env.stack_size = 0;
173 env.result = order;
174 env.order_pos = 0;
175 env.nodes_marked_new = splay_tree_new (splay_tree_compare_ints, 0, 0);
176 env.count = 1;
177 env.reduce = reduce;
178 env.allow_overwritable = allow_overwritable;
180 FOR_EACH_DEFINED_FUNCTION (node)
182 enum availability avail = node->get_availability ();
184 if (avail > AVAIL_INTERPOSABLE
185 || (allow_overwritable
186 && (avail == AVAIL_INTERPOSABLE)))
188 /* Reuse the info if it is already there. */
189 struct ipa_dfs_info *info = (struct ipa_dfs_info *) node->aux;
190 if (!info)
191 info = XCNEW (struct ipa_dfs_info);
192 info->new_node = true;
193 info->on_stack = false;
194 info->next_cycle = NULL;
195 node->aux = info;
197 splay_tree_insert (env.nodes_marked_new,
198 (splay_tree_key)node->uid,
199 (splay_tree_value)node);
201 else
202 node->aux = NULL;
204 result = splay_tree_min (env.nodes_marked_new);
205 while (result)
207 node = (struct cgraph_node *)result->value;
208 searchc (&env, node, ignore_edge);
209 result = splay_tree_min (env.nodes_marked_new);
211 splay_tree_delete (env.nodes_marked_new);
212 free (env.stack);
214 return env.order_pos;
217 /* Deallocate all ipa_dfs_info structures pointed to by the aux pointer of call
218 graph nodes. */
220 void
221 ipa_free_postorder_info (void)
223 struct cgraph_node *node;
224 FOR_EACH_DEFINED_FUNCTION (node)
226 /* Get rid of the aux information. */
227 if (node->aux)
229 free (node->aux);
230 node->aux = NULL;
235 /* Get the set of nodes for the cycle in the reduced call graph starting
236 from NODE. */
238 vec<cgraph_node *>
239 ipa_get_nodes_in_cycle (struct cgraph_node *node)
241 vec<cgraph_node *> v = vNULL;
242 struct ipa_dfs_info *node_dfs_info;
243 while (node)
245 v.safe_push (node);
246 node_dfs_info = (struct ipa_dfs_info *) node->aux;
247 node = node_dfs_info->next_cycle;
249 return v;
252 /* Return true iff the CS is an edge within a strongly connected component as
253 computed by ipa_reduced_postorder. */
255 bool
256 ipa_edge_within_scc (struct cgraph_edge *cs)
258 struct ipa_dfs_info *caller_dfs = (struct ipa_dfs_info *) cs->caller->aux;
259 struct ipa_dfs_info *callee_dfs;
260 struct cgraph_node *callee = cs->callee->function_symbol ();
262 callee_dfs = (struct ipa_dfs_info *) callee->aux;
263 return (caller_dfs
264 && callee_dfs
265 && caller_dfs->scc_no == callee_dfs->scc_no);
268 struct postorder_stack
270 struct cgraph_node *node;
271 struct cgraph_edge *edge;
272 int ref;
275 /* Fill array order with all nodes with output flag set in the reverse
276 topological order. Return the number of elements in the array.
277 FIXME: While walking, consider aliases, too. */
280 ipa_reverse_postorder (struct cgraph_node **order)
282 struct cgraph_node *node, *node2;
283 int stack_size = 0;
284 int order_pos = 0;
285 struct cgraph_edge *edge;
286 int pass;
287 struct ipa_ref *ref = NULL;
289 struct postorder_stack *stack =
290 XCNEWVEC (struct postorder_stack, symtab->cgraph_count);
292 /* We have to deal with cycles nicely, so use a depth first traversal
293 output algorithm. Ignore the fact that some functions won't need
294 to be output and put them into order as well, so we get dependencies
295 right through inline functions. */
296 FOR_EACH_FUNCTION (node)
297 node->aux = NULL;
298 for (pass = 0; pass < 2; pass++)
299 FOR_EACH_FUNCTION (node)
300 if (!node->aux
301 && (pass
302 || (!node->address_taken
303 && !node->global.inlined_to
304 && !node->alias && !node->thunk.thunk_p
305 && !node->only_called_directly_p ())))
307 stack_size = 0;
308 stack[stack_size].node = node;
309 stack[stack_size].edge = node->callers;
310 stack[stack_size].ref = 0;
311 node->aux = (void *)(size_t)1;
312 while (stack_size >= 0)
314 while (true)
316 node2 = NULL;
317 while (stack[stack_size].edge && !node2)
319 edge = stack[stack_size].edge;
320 node2 = edge->caller;
321 stack[stack_size].edge = edge->next_caller;
322 /* Break possible cycles involving always-inline
323 functions by ignoring edges from always-inline
324 functions to non-always-inline functions. */
325 if (DECL_DISREGARD_INLINE_LIMITS (edge->caller->decl)
326 && !DECL_DISREGARD_INLINE_LIMITS
327 (edge->callee->function_symbol ()->decl))
328 node2 = NULL;
330 for (; stack[stack_size].node->iterate_referring (
331 stack[stack_size].ref,
332 ref) && !node2;
333 stack[stack_size].ref++)
335 if (ref->use == IPA_REF_ALIAS)
336 node2 = dyn_cast <cgraph_node *> (ref->referring);
338 if (!node2)
339 break;
340 if (!node2->aux)
342 stack[++stack_size].node = node2;
343 stack[stack_size].edge = node2->callers;
344 stack[stack_size].ref = 0;
345 node2->aux = (void *)(size_t)1;
348 order[order_pos++] = stack[stack_size--].node;
351 free (stack);
352 FOR_EACH_FUNCTION (node)
353 node->aux = NULL;
354 return order_pos;
359 /* Given a memory reference T, will return the variable at the bottom
360 of the access. Unlike get_base_address, this will recurse through
361 INDIRECT_REFS. */
363 tree
364 get_base_var (tree t)
366 while (!SSA_VAR_P (t)
367 && (!CONSTANT_CLASS_P (t))
368 && TREE_CODE (t) != LABEL_DECL
369 && TREE_CODE (t) != FUNCTION_DECL
370 && TREE_CODE (t) != CONST_DECL
371 && TREE_CODE (t) != CONSTRUCTOR)
373 t = TREE_OPERAND (t, 0);
375 return t;
379 /* SRC and DST are going to be merged. Take SRC's profile and merge it into
380 DST so it is not going to be lost. Possibly destroy SRC's body on the way
381 unless PRESERVE_BODY is set. */
383 void
384 ipa_merge_profiles (struct cgraph_node *dst,
385 struct cgraph_node *src,
386 bool preserve_body)
388 tree oldsrcdecl = src->decl;
389 struct function *srccfun, *dstcfun;
390 bool match = true;
392 if (!src->definition
393 || !dst->definition)
394 return;
395 if (src->frequency < dst->frequency)
396 src->frequency = dst->frequency;
398 /* Time profiles are merged. */
399 if (dst->tp_first_run > src->tp_first_run && src->tp_first_run)
400 dst->tp_first_run = src->tp_first_run;
402 if (src->profile_id && !dst->profile_id)
403 dst->profile_id = src->profile_id;
405 /* FIXME when we merge in unknown profile, we ought to set counts as
406 unsafe. */
407 if (!src->count.initialized_p ())
408 return;
409 if (symtab->dump_file)
411 fprintf (symtab->dump_file, "Merging profiles of %s to %s\n",
412 src->dump_name (), dst->dump_name ());
414 if (dst->count.initialized_p ())
415 dst->count += src->count;
416 else
417 dst->count = src->count;
419 /* This is ugly. We need to get both function bodies into memory.
420 If declaration is merged, we need to duplicate it to be able
421 to load body that is being replaced. This makes symbol table
422 temporarily inconsistent. */
423 if (src->decl == dst->decl)
425 struct lto_in_decl_state temp;
426 struct lto_in_decl_state *state;
428 /* We are going to move the decl, we want to remove its file decl data.
429 and link these with the new decl. */
430 temp.fn_decl = src->decl;
431 lto_in_decl_state **slot
432 = src->lto_file_data->function_decl_states->find_slot (&temp,
433 NO_INSERT);
434 state = *slot;
435 src->lto_file_data->function_decl_states->clear_slot (slot);
436 gcc_assert (state);
438 /* Duplicate the decl and be sure it does not link into body of DST. */
439 src->decl = copy_node (src->decl);
440 DECL_STRUCT_FUNCTION (src->decl) = NULL;
441 DECL_ARGUMENTS (src->decl) = NULL;
442 DECL_INITIAL (src->decl) = NULL;
443 DECL_RESULT (src->decl) = NULL;
445 /* Associate the decl state with new declaration, so LTO streamer
446 can look it up. */
447 state->fn_decl = src->decl;
448 slot
449 = src->lto_file_data->function_decl_states->find_slot (state, INSERT);
450 gcc_assert (!*slot);
451 *slot = state;
453 src->get_untransformed_body ();
454 dst->get_untransformed_body ();
455 srccfun = DECL_STRUCT_FUNCTION (src->decl);
456 dstcfun = DECL_STRUCT_FUNCTION (dst->decl);
457 if (n_basic_blocks_for_fn (srccfun)
458 != n_basic_blocks_for_fn (dstcfun))
460 if (symtab->dump_file)
461 fprintf (symtab->dump_file,
462 "Giving up; number of basic block mismatch.\n");
463 match = false;
465 else if (last_basic_block_for_fn (srccfun)
466 != last_basic_block_for_fn (dstcfun))
468 if (symtab->dump_file)
469 fprintf (symtab->dump_file,
470 "Giving up; last block mismatch.\n");
471 match = false;
473 else
475 basic_block srcbb, dstbb;
477 FOR_ALL_BB_FN (srcbb, srccfun)
479 unsigned int i;
481 dstbb = BASIC_BLOCK_FOR_FN (dstcfun, srcbb->index);
482 if (dstbb == NULL)
484 if (symtab->dump_file)
485 fprintf (symtab->dump_file,
486 "No matching block for bb %i.\n",
487 srcbb->index);
488 match = false;
489 break;
491 if (EDGE_COUNT (srcbb->succs) != EDGE_COUNT (dstbb->succs))
493 if (symtab->dump_file)
494 fprintf (symtab->dump_file,
495 "Edge count mistmatch for bb %i.\n",
496 srcbb->index);
497 match = false;
498 break;
500 for (i = 0; i < EDGE_COUNT (srcbb->succs); i++)
502 edge srce = EDGE_SUCC (srcbb, i);
503 edge dste = EDGE_SUCC (dstbb, i);
504 if (srce->dest->index != dste->dest->index)
506 if (symtab->dump_file)
507 fprintf (symtab->dump_file,
508 "Succ edge mistmatch for bb %i.\n",
509 srce->dest->index);
510 match = false;
511 break;
516 if (match)
518 struct cgraph_edge *e, *e2;
519 basic_block srcbb, dstbb;
521 /* TODO: merge also statement histograms. */
522 FOR_ALL_BB_FN (srcbb, srccfun)
524 unsigned int i;
526 dstbb = BASIC_BLOCK_FOR_FN (dstcfun, srcbb->index);
527 if (dstbb->count.initialized_p ())
528 dstbb->count += srcbb->count;
529 else
530 dstbb->count = srcbb->count;
531 for (i = 0; i < EDGE_COUNT (srcbb->succs); i++)
533 edge srce = EDGE_SUCC (srcbb, i);
534 edge dste = EDGE_SUCC (dstbb, i);
535 if (dstbb->count.initialized_p ())
536 dste->count += srce->count;
537 else
538 dste->count = srce->count;
539 if (dstbb->count > 0 && dste->count.initialized_p ())
540 dste->probability = dste->count.probability_in (dstbb->count);
543 push_cfun (dstcfun);
544 counts_to_freqs ();
545 compute_function_frequency ();
546 pop_cfun ();
547 for (e = dst->callees; e; e = e->next_callee)
549 if (e->speculative)
550 continue;
551 e->count = gimple_bb (e->call_stmt)->count;
552 e->frequency = compute_call_stmt_bb_frequency
553 (dst->decl,
554 gimple_bb (e->call_stmt));
556 for (e = dst->indirect_calls, e2 = src->indirect_calls; e;
557 e2 = (e2 ? e2->next_callee : NULL), e = e->next_callee)
559 profile_count count = gimple_bb (e->call_stmt)->count;
560 int freq = compute_call_stmt_bb_frequency
561 (dst->decl,
562 gimple_bb (e->call_stmt));
563 /* When call is speculative, we need to re-distribute probabilities
564 the same way as they was. This is not really correct because
565 in the other copy the speculation may differ; but probably it
566 is not really worth the effort. */
567 if (e->speculative)
569 cgraph_edge *direct, *indirect;
570 cgraph_edge *direct2 = NULL, *indirect2 = NULL;
571 ipa_ref *ref;
573 e->speculative_call_info (direct, indirect, ref);
574 gcc_assert (e == indirect);
575 if (e2 && e2->speculative)
576 e2->speculative_call_info (direct2, indirect2, ref);
577 if (indirect->count > profile_count::zero ()
578 || direct->count > profile_count::zero ())
580 /* We should mismatch earlier if there is no matching
581 indirect edge. */
582 if (!e2)
584 if (dump_file)
585 fprintf (dump_file,
586 "Mismatch in merging indirect edges\n");
588 else if (!e2->speculative)
589 indirect->count += e2->count;
590 else if (e2->speculative)
592 if (DECL_ASSEMBLER_NAME (direct2->callee->decl)
593 != DECL_ASSEMBLER_NAME (direct->callee->decl))
595 if (direct2->count >= direct->count)
597 direct->redirect_callee (direct2->callee);
598 indirect->count += indirect2->count
599 + direct->count;
600 direct->count = direct2->count;
602 else
603 indirect->count += indirect2->count + direct2->count;
605 else
607 direct->count += direct2->count;
608 indirect->count += indirect2->count;
611 int prob = direct->count.probability_in (direct->count
612 + indirect->count).
613 to_reg_br_prob_base ();
614 direct->frequency = RDIV (freq * prob, REG_BR_PROB_BASE);
615 indirect->frequency = RDIV (freq * (REG_BR_PROB_BASE - prob),
616 REG_BR_PROB_BASE);
618 else
619 /* At the moment we should have only profile feedback based
620 speculations when merging. */
621 gcc_unreachable ();
623 else if (e2 && e2->speculative)
625 cgraph_edge *direct, *indirect;
626 ipa_ref *ref;
628 e2->speculative_call_info (direct, indirect, ref);
629 e->count = count;
630 e->frequency = freq;
631 int prob = direct->count.probability_in (e->count)
632 .to_reg_br_prob_base ();
633 e->make_speculative (direct->callee, direct->count,
634 RDIV (freq * prob, REG_BR_PROB_BASE));
636 else
638 e->count = count;
639 e->frequency = freq;
642 if (!preserve_body)
643 src->release_body ();
644 ipa_update_overall_fn_summary (dst);
646 /* TODO: if there is no match, we can scale up. */
647 src->decl = oldsrcdecl;
650 /* Return true if call to DEST is known to be self-recusive call withing FUNC. */
652 bool
653 recursive_call_p (tree func, tree dest)
655 struct cgraph_node *dest_node = cgraph_node::get_create (dest);
656 struct cgraph_node *cnode = cgraph_node::get_create (func);
657 ipa_ref *alias;
658 enum availability avail;
660 gcc_assert (!cnode->alias);
661 if (cnode != dest_node->ultimate_alias_target (&avail))
662 return false;
663 if (avail >= AVAIL_AVAILABLE)
664 return true;
665 if (!dest_node->semantically_equivalent_p (cnode))
666 return false;
667 /* If there is only one way to call the fuction or we know all of them
668 are semantically equivalent, we still can consider call recursive. */
669 FOR_EACH_ALIAS (cnode, alias)
670 if (!dest_node->semantically_equivalent_p (alias->referring))
671 return false;
672 return true;