1 /* Utilities for ipa analysis.
2 Copyright (C) 2005-2013 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
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
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/>. */
23 #include "coretypes.h"
27 #include "tree-inline.h"
29 #include "langhooks.h"
30 #include "pointer-set.h"
31 #include "splay-tree.h"
33 #include "ipa-utils.h"
34 #include "ipa-reference.h"
36 #include "diagnostic.h"
37 #include "langhooks.h"
38 #include "lto-streamer.h"
39 #include "ipa-inline.h"
41 /* Debugging function for postorder and inorder code. NOTE is a string
42 that is printed before the nodes are printed. ORDER is an array of
43 cgraph_nodes that has COUNT useful nodes in it. */
46 ipa_print_order (FILE* out
,
48 struct cgraph_node
** order
,
52 fprintf (out
, "\n\n ordered call graph: %s\n", note
);
54 for (i
= count
- 1; i
>= 0; i
--)
55 dump_cgraph_node (dump_file
, order
[i
]);
62 struct cgraph_node
**stack
;
64 struct cgraph_node
**result
;
66 splay_tree nodes_marked_new
;
68 bool allow_overwritable
;
72 /* This is an implementation of Tarjan's strongly connected region
73 finder as reprinted in Aho Hopcraft and Ullman's The Design and
74 Analysis of Computer Programs (1975) pages 192-193. This version
75 has been customized for cgraph_nodes. The env parameter is because
76 it is recursive and there are no nested functions here. This
77 function should only be called from itself or
78 ipa_reduced_postorder. ENV is a stack env and would be
79 unnecessary if C had nested functions. V is the node to start
83 searchc (struct searchc_env
* env
, struct cgraph_node
*v
,
84 bool (*ignore_edge
) (struct cgraph_edge
*))
86 struct cgraph_edge
*edge
;
87 struct ipa_dfs_info
*v_info
= (struct ipa_dfs_info
*) v
->aux
;
89 /* mark node as old */
90 v_info
->new_node
= false;
91 splay_tree_remove (env
->nodes_marked_new
, v
->uid
);
93 v_info
->dfn_number
= env
->count
;
94 v_info
->low_link
= env
->count
;
96 env
->stack
[(env
->stack_size
)++] = v
;
97 v_info
->on_stack
= true;
99 for (edge
= v
->callees
; edge
; edge
= edge
->next_callee
)
101 struct ipa_dfs_info
* w_info
;
102 enum availability avail
;
103 struct cgraph_node
*w
= cgraph_function_or_thunk_node (edge
->callee
, &avail
);
105 if (!w
|| (ignore_edge
&& ignore_edge (edge
)))
109 && (avail
> AVAIL_OVERWRITABLE
110 || (env
->allow_overwritable
&& avail
== AVAIL_OVERWRITABLE
)))
112 w_info
= (struct ipa_dfs_info
*) w
->aux
;
113 if (w_info
->new_node
)
115 searchc (env
, w
, ignore_edge
);
117 (v_info
->low_link
< w_info
->low_link
) ?
118 v_info
->low_link
: w_info
->low_link
;
121 if ((w_info
->dfn_number
< v_info
->dfn_number
)
122 && (w_info
->on_stack
))
124 (w_info
->dfn_number
< v_info
->low_link
) ?
125 w_info
->dfn_number
: v_info
->low_link
;
130 if (v_info
->low_link
== v_info
->dfn_number
)
132 struct cgraph_node
*last
= NULL
;
133 struct cgraph_node
*x
;
134 struct ipa_dfs_info
*x_info
;
136 x
= env
->stack
[--(env
->stack_size
)];
137 x_info
= (struct ipa_dfs_info
*) x
->aux
;
138 x_info
->on_stack
= false;
139 x_info
->scc_no
= v_info
->dfn_number
;
143 x_info
->next_cycle
= last
;
147 env
->result
[env
->order_pos
++] = x
;
151 env
->result
[env
->order_pos
++] = v
;
155 /* Topsort the call graph by caller relation. Put the result in ORDER.
157 The REDUCE flag is true if you want the cycles reduced to single nodes.
158 You can use ipa_get_nodes_in_cycle to obtain a vector containing all real
159 call graph nodes in a reduced node.
161 Set ALLOW_OVERWRITABLE if nodes with such availability should be included.
162 IGNORE_EDGE, if non-NULL is a hook that may make some edges insignificant
163 for the topological sort. */
166 ipa_reduced_postorder (struct cgraph_node
**order
,
167 bool reduce
, bool allow_overwritable
,
168 bool (*ignore_edge
) (struct cgraph_edge
*))
170 struct cgraph_node
*node
;
171 struct searchc_env env
;
172 splay_tree_node result
;
173 env
.stack
= XCNEWVEC (struct cgraph_node
*, cgraph_n_nodes
);
177 env
.nodes_marked_new
= splay_tree_new (splay_tree_compare_ints
, 0, 0);
180 env
.allow_overwritable
= allow_overwritable
;
182 FOR_EACH_DEFINED_FUNCTION (node
)
184 enum availability avail
= cgraph_function_body_availability (node
);
186 if (avail
> AVAIL_OVERWRITABLE
187 || (allow_overwritable
188 && (avail
== AVAIL_OVERWRITABLE
)))
190 /* Reuse the info if it is already there. */
191 struct ipa_dfs_info
*info
= (struct ipa_dfs_info
*) node
->aux
;
193 info
= XCNEW (struct ipa_dfs_info
);
194 info
->new_node
= true;
195 info
->on_stack
= false;
196 info
->next_cycle
= NULL
;
199 splay_tree_insert (env
.nodes_marked_new
,
200 (splay_tree_key
)node
->uid
,
201 (splay_tree_value
)node
);
206 result
= splay_tree_min (env
.nodes_marked_new
);
209 node
= (struct cgraph_node
*)result
->value
;
210 searchc (&env
, node
, ignore_edge
);
211 result
= splay_tree_min (env
.nodes_marked_new
);
213 splay_tree_delete (env
.nodes_marked_new
);
216 return env
.order_pos
;
219 /* Deallocate all ipa_dfs_info structures pointed to by the aux pointer of call
223 ipa_free_postorder_info (void)
225 struct cgraph_node
*node
;
226 FOR_EACH_DEFINED_FUNCTION (node
)
228 /* Get rid of the aux information. */
237 /* Get the set of nodes for the cycle in the reduced call graph starting
241 ipa_get_nodes_in_cycle (struct cgraph_node
*node
)
243 vec
<cgraph_node_ptr
> v
= vNULL
;
244 struct ipa_dfs_info
*node_dfs_info
;
248 node_dfs_info
= (struct ipa_dfs_info
*) node
->aux
;
249 node
= node_dfs_info
->next_cycle
;
254 /* Return true iff the CS is an edge within a strongly connected component as
255 computed by ipa_reduced_postorder. */
258 ipa_edge_within_scc (struct cgraph_edge
*cs
)
260 struct ipa_dfs_info
*caller_dfs
= (struct ipa_dfs_info
*) cs
->caller
->aux
;
261 struct ipa_dfs_info
*callee_dfs
;
262 struct cgraph_node
*callee
= cgraph_function_node (cs
->callee
, NULL
);
264 callee_dfs
= (struct ipa_dfs_info
*) callee
->aux
;
267 && caller_dfs
->scc_no
== callee_dfs
->scc_no
);
270 struct postorder_stack
272 struct cgraph_node
*node
;
273 struct cgraph_edge
*edge
;
277 /* Fill array order with all nodes with output flag set in the reverse
278 topological order. Return the number of elements in the array.
279 FIXME: While walking, consider aliases, too. */
282 ipa_reverse_postorder (struct cgraph_node
**order
)
284 struct cgraph_node
*node
, *node2
;
287 struct cgraph_edge
*edge
;
291 struct postorder_stack
*stack
=
292 XCNEWVEC (struct postorder_stack
, cgraph_n_nodes
);
294 /* We have to deal with cycles nicely, so use a depth first traversal
295 output algorithm. Ignore the fact that some functions won't need
296 to be output and put them into order as well, so we get dependencies
297 right through inline functions. */
298 FOR_EACH_FUNCTION (node
)
300 for (pass
= 0; pass
< 2; pass
++)
301 FOR_EACH_FUNCTION (node
)
304 || (!node
->address_taken
305 && !node
->global
.inlined_to
306 && !node
->alias
&& !node
->thunk
.thunk_p
307 && !cgraph_only_called_directly_p (node
))))
310 stack
[stack_size
].node
= node
;
311 stack
[stack_size
].edge
= node
->callers
;
312 stack
[stack_size
].ref
= 0;
313 node
->aux
= (void *)(size_t)1;
314 while (stack_size
>= 0)
319 while (stack
[stack_size
].edge
&& !node2
)
321 edge
= stack
[stack_size
].edge
;
322 node2
= edge
->caller
;
323 stack
[stack_size
].edge
= edge
->next_caller
;
324 /* Break possible cycles involving always-inline
325 functions by ignoring edges from always-inline
326 functions to non-always-inline functions. */
327 if (DECL_DISREGARD_INLINE_LIMITS (edge
->caller
->decl
)
328 && !DECL_DISREGARD_INLINE_LIMITS
329 (cgraph_function_node (edge
->callee
, NULL
)->decl
))
332 for (;ipa_ref_list_referring_iterate (&stack
[stack_size
].node
->ref_list
,
333 stack
[stack_size
].ref
,
335 stack
[stack_size
].ref
++)
337 if (ref
->use
== IPA_REF_ALIAS
)
338 node2
= ipa_ref_referring_node (ref
);
344 stack
[++stack_size
].node
= node2
;
345 stack
[stack_size
].edge
= node2
->callers
;
346 stack
[stack_size
].ref
= 0;
347 node2
->aux
= (void *)(size_t)1;
350 order
[order_pos
++] = stack
[stack_size
--].node
;
354 FOR_EACH_FUNCTION (node
)
361 /* Given a memory reference T, will return the variable at the bottom
362 of the access. Unlike get_base_address, this will recurse through
366 get_base_var (tree t
)
368 while (!SSA_VAR_P (t
)
369 && (!CONSTANT_CLASS_P (t
))
370 && TREE_CODE (t
) != LABEL_DECL
371 && TREE_CODE (t
) != FUNCTION_DECL
372 && TREE_CODE (t
) != CONST_DECL
373 && TREE_CODE (t
) != CONSTRUCTOR
)
375 t
= TREE_OPERAND (t
, 0);
381 /* Create a new cgraph node set. */
384 cgraph_node_set_new (void)
386 cgraph_node_set new_node_set
;
388 new_node_set
= XCNEW (struct cgraph_node_set_def
);
389 new_node_set
->map
= pointer_map_create ();
390 new_node_set
->nodes
.create (0);
395 /* Add cgraph_node NODE to cgraph_node_set SET. */
398 cgraph_node_set_add (cgraph_node_set set
, struct cgraph_node
*node
)
402 slot
= pointer_map_insert (set
->map
, node
);
406 int index
= (size_t) *slot
- 1;
407 gcc_checking_assert ((set
->nodes
[index
]
412 *slot
= (void *)(size_t) (set
->nodes
.length () + 1);
414 /* Insert into node vector. */
415 set
->nodes
.safe_push (node
);
419 /* Remove cgraph_node NODE from cgraph_node_set SET. */
422 cgraph_node_set_remove (cgraph_node_set set
, struct cgraph_node
*node
)
424 void **slot
, **last_slot
;
426 struct cgraph_node
*last_node
;
428 slot
= pointer_map_contains (set
->map
, node
);
429 if (slot
== NULL
|| !*slot
)
432 index
= (size_t) *slot
- 1;
433 gcc_checking_assert (set
->nodes
[index
]
436 /* Remove from vector. We do this by swapping node with the last element
438 last_node
= set
->nodes
.pop ();
439 if (last_node
!= node
)
441 last_slot
= pointer_map_contains (set
->map
, last_node
);
442 gcc_checking_assert (last_slot
&& *last_slot
);
443 *last_slot
= (void *)(size_t) (index
+ 1);
445 /* Move the last element to the original spot of NODE. */
446 set
->nodes
[index
] = last_node
;
449 /* Remove element from hash table. */
454 /* Find NODE in SET and return an iterator to it if found. A null iterator
455 is returned if NODE is not in SET. */
457 cgraph_node_set_iterator
458 cgraph_node_set_find (cgraph_node_set set
, struct cgraph_node
*node
)
461 cgraph_node_set_iterator csi
;
463 slot
= pointer_map_contains (set
->map
, node
);
464 if (slot
== NULL
|| !*slot
)
465 csi
.index
= (unsigned) ~0;
467 csi
.index
= (size_t)*slot
- 1;
474 /* Dump content of SET to file F. */
477 dump_cgraph_node_set (FILE *f
, cgraph_node_set set
)
479 cgraph_node_set_iterator iter
;
481 for (iter
= csi_start (set
); !csi_end_p (iter
); csi_next (&iter
))
483 struct cgraph_node
*node
= csi_node (iter
);
484 fprintf (f
, " %s/%i", cgraph_node_name (node
), node
->order
);
490 /* Dump content of SET to stderr. */
493 debug_cgraph_node_set (cgraph_node_set set
)
495 dump_cgraph_node_set (stderr
, set
);
499 /* Free varpool node set. */
502 free_cgraph_node_set (cgraph_node_set set
)
504 set
->nodes
.release ();
505 pointer_map_destroy (set
->map
);
510 /* Create a new varpool node set. */
513 varpool_node_set_new (void)
515 varpool_node_set new_node_set
;
517 new_node_set
= XCNEW (struct varpool_node_set_def
);
518 new_node_set
->map
= pointer_map_create ();
519 new_node_set
->nodes
.create (0);
524 /* Add varpool_node NODE to varpool_node_set SET. */
527 varpool_node_set_add (varpool_node_set set
, struct varpool_node
*node
)
531 slot
= pointer_map_insert (set
->map
, node
);
535 int index
= (size_t) *slot
- 1;
536 gcc_checking_assert ((set
->nodes
[index
]
541 *slot
= (void *)(size_t) (set
->nodes
.length () + 1);
543 /* Insert into node vector. */
544 set
->nodes
.safe_push (node
);
548 /* Remove varpool_node NODE from varpool_node_set SET. */
551 varpool_node_set_remove (varpool_node_set set
, struct varpool_node
*node
)
553 void **slot
, **last_slot
;
555 struct varpool_node
*last_node
;
557 slot
= pointer_map_contains (set
->map
, node
);
558 if (slot
== NULL
|| !*slot
)
561 index
= (size_t) *slot
- 1;
562 gcc_checking_assert (set
->nodes
[index
]
565 /* Remove from vector. We do this by swapping node with the last element
567 last_node
= set
->nodes
.pop ();
568 if (last_node
!= node
)
570 last_slot
= pointer_map_contains (set
->map
, last_node
);
571 gcc_checking_assert (last_slot
&& *last_slot
);
572 *last_slot
= (void *)(size_t) (index
+ 1);
574 /* Move the last element to the original spot of NODE. */
575 set
->nodes
[index
] = last_node
;
578 /* Remove element from hash table. */
583 /* Find NODE in SET and return an iterator to it if found. A null iterator
584 is returned if NODE is not in SET. */
586 varpool_node_set_iterator
587 varpool_node_set_find (varpool_node_set set
, struct varpool_node
*node
)
590 varpool_node_set_iterator vsi
;
592 slot
= pointer_map_contains (set
->map
, node
);
593 if (slot
== NULL
|| !*slot
)
594 vsi
.index
= (unsigned) ~0;
596 vsi
.index
= (size_t)*slot
- 1;
603 /* Dump content of SET to file F. */
606 dump_varpool_node_set (FILE *f
, varpool_node_set set
)
608 varpool_node_set_iterator iter
;
610 for (iter
= vsi_start (set
); !vsi_end_p (iter
); vsi_next (&iter
))
612 struct varpool_node
*node
= vsi_node (iter
);
613 fprintf (f
, " %s", varpool_node_name (node
));
619 /* Free varpool node set. */
622 free_varpool_node_set (varpool_node_set set
)
624 set
->nodes
.release ();
625 pointer_map_destroy (set
->map
);
630 /* Dump content of SET to stderr. */
633 debug_varpool_node_set (varpool_node_set set
)
635 dump_varpool_node_set (stderr
, set
);
639 /* SRC and DST are going to be merged. Take SRC's profile and merge it into
640 DST so it is not going to be lost. Destroy SRC's body on the way. */
643 ipa_merge_profiles (struct cgraph_node
*dst
,
644 struct cgraph_node
*src
)
646 tree oldsrcdecl
= src
->decl
;
647 struct function
*srccfun
, *dstcfun
;
653 if (src
->frequency
< dst
->frequency
)
654 src
->frequency
= dst
->frequency
;
657 if (cgraph_dump_file
)
659 fprintf (cgraph_dump_file
, "Merging profiles of %s/%i to %s/%i\n",
660 xstrdup (cgraph_node_name (src
)), src
->order
,
661 xstrdup (cgraph_node_name (dst
)), dst
->order
);
663 dst
->count
+= src
->count
;
665 /* This is ugly. We need to get both function bodies into memory.
666 If declaration is merged, we need to duplicate it to be able
667 to load body that is being replaced. This makes symbol table
668 temporarily inconsistent. */
669 if (src
->decl
== dst
->decl
)
672 struct lto_in_decl_state temp
;
673 struct lto_in_decl_state
*state
;
675 /* We are going to move the decl, we want to remove its file decl data.
676 and link these with the new decl. */
677 temp
.fn_decl
= src
->decl
;
678 slot
= htab_find_slot (src
->lto_file_data
->function_decl_states
,
680 state
= (lto_in_decl_state
*)*slot
;
681 htab_clear_slot (src
->lto_file_data
->function_decl_states
, slot
);
684 /* Duplicate the decl and be sure it does not link into body of DST. */
685 src
->decl
= copy_node (src
->decl
);
686 DECL_STRUCT_FUNCTION (src
->decl
) = NULL
;
687 DECL_ARGUMENTS (src
->decl
) = NULL
;
688 DECL_INITIAL (src
->decl
) = NULL
;
689 DECL_RESULT (src
->decl
) = NULL
;
691 /* Associate the decl state with new declaration, so LTO streamer
693 state
->fn_decl
= src
->decl
;
694 slot
= htab_find_slot (src
->lto_file_data
->function_decl_states
,
699 cgraph_get_body (src
);
700 cgraph_get_body (dst
);
701 srccfun
= DECL_STRUCT_FUNCTION (src
->decl
);
702 dstcfun
= DECL_STRUCT_FUNCTION (dst
->decl
);
703 if (n_basic_blocks_for_function (srccfun
)
704 != n_basic_blocks_for_function (dstcfun
))
706 if (cgraph_dump_file
)
707 fprintf (cgraph_dump_file
,
708 "Giving up; number of basic block mismatch.\n");
711 else if (last_basic_block_for_function (srccfun
)
712 != last_basic_block_for_function (dstcfun
))
714 if (cgraph_dump_file
)
715 fprintf (cgraph_dump_file
,
716 "Giving up; last block mismatch.\n");
721 basic_block srcbb
, dstbb
;
723 FOR_ALL_BB_FN (srcbb
, srccfun
)
727 dstbb
= BASIC_BLOCK_FOR_FUNCTION (dstcfun
, srcbb
->index
);
730 if (cgraph_dump_file
)
731 fprintf (cgraph_dump_file
,
732 "No matching block for bb %i.\n",
737 if (EDGE_COUNT (srcbb
->succs
) != EDGE_COUNT (dstbb
->succs
))
739 if (cgraph_dump_file
)
740 fprintf (cgraph_dump_file
,
741 "Edge count mistmatch for bb %i.\n",
746 for (i
= 0; i
< EDGE_COUNT (srcbb
->succs
); i
++)
748 edge srce
= EDGE_SUCC (srcbb
, i
);
749 edge dste
= EDGE_SUCC (dstbb
, i
);
750 if (srce
->dest
->index
!= dste
->dest
->index
)
752 if (cgraph_dump_file
)
753 fprintf (cgraph_dump_file
,
754 "Succ edge mistmatch for bb %i.\n",
764 struct cgraph_edge
*e
;
765 basic_block srcbb
, dstbb
;
767 /* TODO: merge also statement histograms. */
768 FOR_ALL_BB_FN (srcbb
, srccfun
)
772 dstbb
= BASIC_BLOCK_FOR_FUNCTION (dstcfun
, srcbb
->index
);
773 dstbb
->count
+= srcbb
->count
;
774 for (i
= 0; i
< EDGE_COUNT (srcbb
->succs
); i
++)
776 edge srce
= EDGE_SUCC (srcbb
, i
);
777 edge dste
= EDGE_SUCC (dstbb
, i
);
778 dste
->count
+= srce
->count
;
783 compute_function_frequency ();
785 for (e
= dst
->callees
; e
; e
= e
->next_callee
)
787 gcc_assert (!e
->speculative
);
788 e
->count
= gimple_bb (e
->call_stmt
)->count
;
789 e
->frequency
= compute_call_stmt_bb_frequency
791 gimple_bb (e
->call_stmt
));
793 for (e
= dst
->indirect_calls
; e
; e
= e
->next_callee
)
795 gcc_assert (!e
->speculative
);
796 e
->count
= gimple_bb (e
->call_stmt
)->count
;
797 e
->frequency
= compute_call_stmt_bb_frequency
799 gimple_bb (e
->call_stmt
));
801 cgraph_release_function_body (src
);
802 inline_update_overall_summary (dst
);
804 /* TODO: if there is no match, we can scale up. */
805 src
->decl
= oldsrcdecl
;
808 /* Return true if call to DEST is known to be self-recusive call withing FUNC. */
811 recursive_call_p (tree func
, tree dest
)
813 struct cgraph_node
*dest_node
= cgraph_get_create_node (dest
);
814 struct cgraph_node
*cnode
= cgraph_get_create_node (func
);
816 return symtab_semantically_equivalent_p (dest_node
,