1 /* Tree inlining.
2 Copyright 2001, 2002, 2003, 2004, 2005 Free Software Foundation, Inc.
3 Contributed by Alexandre Oliva <aoliva@redhat.com>
4
5 This file is part of GCC.
6
7 GCC is free software; you can redistribute it and/or modify
8 it under the terms of the GNU General Public License as published by
9 the Free Software Foundation; either version 2, or (at your option)
10 any later version.
11
12 GCC is distributed in the hope that it will be useful,
13 but WITHOUT ANY WARRANTY; without even the implied warranty of
14 MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
15 GNU General Public License for more details.
16
17 You should have received a copy of the GNU General Public License
18 along with GCC; see the file COPYING. If not, write to
19 the Free Software Foundation, 51 Franklin Street, Fifth Floor,
20 Boston, MA 02110-1301, USA. */
21
22 #include "config.h"
23 #include "system.h"
24 #include "coretypes.h"
25 #include "tm.h"
26 #include "toplev.h"
27 #include "tree.h"
28 #include "tree-inline.h"
29 #include "rtl.h"
30 #include "expr.h"
31 #include "flags.h"
32 #include "params.h"
33 #include "input.h"
34 #include "insn-config.h"
35 #include "varray.h"
36 #include "hashtab.h"
37 #include "splay-tree.h"
38 #include "langhooks.h"
39 #include "basic-block.h"
40 #include "tree-iterator.h"
41 #include "cgraph.h"
42 #include "intl.h"
43 #include "tree-mudflap.h"
44 #include "tree-flow.h"
45 #include "function.h"
46 #include "ggc.h"
47 #include "tree-flow.h"
48 #include "diagnostic.h"
49 #include "except.h"
50 #include "debug.h"
51 #include "pointer-set.h"
52 #include "ipa-prop.h"
53
54 /* I'm not real happy about this, but we need to handle gimple and
55 non-gimple trees. */
56 #include "tree-gimple.h"
57
58 /* Inlining, Saving, Cloning
59
60 Inlining: a function body is duplicated, but the PARM_DECLs are
61 remapped into VAR_DECLs, and non-void RETURN_EXPRs become
62 MODIFY_EXPRs that store to a dedicated returned-value variable.
63 The duplicated eh_region info of the copy will later be appended
64 to the info for the caller; the eh_region info in copied throwing
65 statements and RESX_EXPRs is adjusted accordingly.
66
67 Saving: make a semantically-identical copy of the function body.
68 Necessary when we want to generate code for the body (a destructive
69 operation), but we expect to need this body in the future (e.g. for
70 inlining into another function).
71
72 Cloning: (only in C++) We have one body for a con/de/structor, and
73 multiple function decls, each with a unique parameter list.
74 Duplicate the body, using the given splay tree; some parameters
75 will become constants (like 0 or 1).
76
77 All of these will simultaneously lookup any callgraph edges. If
78 we're going to inline the duplicated function body, and the given
79 function has some cloned callgraph nodes (one for each place this
80 function will be inlined) those callgraph edges will be duplicated.
81 If we're saving or cloning the body, those callgraph edges will be
82 updated to point into the new body. (Note that the original
83 callgraph node and edge list will not be altered.)
84
85 See the CALL_EXPR handling case in copy_body_r (). */
86
87 /* 0 if we should not perform inlining.
88 1 if we should expand functions calls inline at the tree level.
89 2 if we should consider *all* functions to be inline
90 candidates. */
91
92 int flag_inline_trees = 0;
93
94 /* To Do:
95
96 o In order to make inlining-on-trees work, we pessimized
97 function-local static constants. In particular, they are now
98 always output, even when not addressed. Fix this by treating
99 function-local static constants just like global static
100 constants; the back-end already knows not to output them if they
101 are not needed.
102
103 o Provide heuristics to clamp inlining of recursive template
104 calls? */
105
106 /* Data required for function inlining. */
107
108 typedef struct inline_data
109 {
110 /* FUNCTION_DECL for function being inlined. */
111 tree callee;
112 /* FUNCTION_DECL for function being inlined into. */
113 tree caller;
114 /* struct function for function being inlined. Usually this is the same
115 as DECL_STRUCT_FUNCTION (callee), but can be different if saved_cfg
116 and saved_eh are in use. */
117 struct function *callee_cfun;
118 /* The VAR_DECL for the return value. */
119 tree retvar;
120 /* The map from local declarations in the inlined function to
121 equivalents in the function into which it is being inlined. */
122 splay_tree decl_map;
123 /* We use the same mechanism to build clones that we do to perform
124 inlining. However, there are a few places where we need to
125 distinguish between those two situations. This flag is true if
126 we are cloning, rather than inlining. */
127 bool cloning_p;
128 /* Similarly for saving function body. */
129 bool saving_p;
130 /* Versioning function is slightly different from inlining. */
131 bool versioning_p;
132 /* Callgraph node of function we are inlining into. */
133 struct cgraph_node *node;
134 /* Callgraph node of currently inlined function. */
135 struct cgraph_node *current_node;
136 /* Current BLOCK. */
137 tree block;
138 varray_type ipa_info;
139 /* Exception region the inlined call lie in. */
140 int eh_region;
141 /* Take region number in the function being copied, add this value and
142 get eh region number of the duplicate in the function we inline into. */
143 int eh_region_offset;
144 } inline_data;
145
146 /* Prototypes. */
147
148 static tree declare_return_variable (inline_data *, tree, tree, tree *);
149 static tree copy_body_r (tree *, int *, void *);
150 static tree copy_generic_body (inline_data *);
151 static bool inlinable_function_p (tree);
152 static tree remap_decl (tree, inline_data *);
153 static tree remap_type (tree, inline_data *);
154 static void remap_block (tree *, inline_data *);
155 static tree remap_decl (tree, inline_data *);
156 static tree remap_decls (tree, inline_data *);
157 static void copy_bind_expr (tree *, int *, inline_data *);
158 static tree mark_local_for_remap_r (tree *, int *, void *);
159 static void unsave_expr_1 (tree);
160 static tree unsave_r (tree *, int *, void *);
161 static void declare_inline_vars (tree, tree);
162 static void remap_save_expr (tree *, void *, int *);
163 static bool replace_ref_tree (inline_data *, tree *);
164 static inline bool inlining_p (inline_data *);
165 static void add_lexical_block (tree current_block, tree new_block);
166
167 /* Insert a tree->tree mapping for ID. Despite the name suggests
168 that the trees should be variables, it is used for more than that. */
169
170 static void
insert_decl_map(inline_data * id,tree key,tree value)171 insert_decl_map (inline_data *id, tree key, tree value)
172 {
173 splay_tree_insert (id->decl_map, (splay_tree_key) key,
174 (splay_tree_value) value);
175
176 /* Always insert an identity map as well. If we see this same new
177 node again, we won't want to duplicate it a second time. */
178 if (key != value)
179 splay_tree_insert (id->decl_map, (splay_tree_key) value,
180 (splay_tree_value) value);
181 }
182
183 /* Remap DECL during the copying of the BLOCK tree for the function. */
184
185 static tree
remap_decl(tree decl,inline_data * id)186 remap_decl (tree decl, inline_data *id)
187 {
188 splay_tree_node n;
189 tree fn;
190
191 /* We only remap local variables in the current function. */
192 fn = id->callee;
193
194 /* See if we have remapped this declaration. */
195
196 n = splay_tree_lookup (id->decl_map, (splay_tree_key) decl);
197
198 /* If we didn't already have an equivalent for this declaration,
199 create one now. */
200 if (!n)
201 {
202 /* Make a copy of the variable or label. */
203 tree t;
204 t = copy_decl_for_dup (decl, fn, id->caller, id->versioning_p);
205
206 /* Remember it, so that if we encounter this local entity again
207 we can reuse this copy. Do this early because remap_type may
208 need this decl for TYPE_STUB_DECL. */
209 insert_decl_map (id, decl, t);
210
211 /* Remap types, if necessary. */
212 TREE_TYPE (t) = remap_type (TREE_TYPE (t), id);
213 if (TREE_CODE (t) == TYPE_DECL)
214 DECL_ORIGINAL_TYPE (t) = remap_type (DECL_ORIGINAL_TYPE (t), id);
215
216 /* Remap sizes as necessary. */
217 walk_tree (&DECL_SIZE (t), copy_body_r, id, NULL);
218 walk_tree (&DECL_SIZE_UNIT (t), copy_body_r, id, NULL);
219
220 /* If fields, do likewise for offset and qualifier. */
221 if (TREE_CODE (t) == FIELD_DECL)
222 {
223 walk_tree (&DECL_FIELD_OFFSET (t), copy_body_r, id, NULL);
224 if (TREE_CODE (DECL_CONTEXT (t)) == QUAL_UNION_TYPE)
225 walk_tree (&DECL_QUALIFIER (t), copy_body_r, id, NULL);
226 }
227
228 #if 0
229 /* FIXME handle anon aggrs. */
230 if (! DECL_NAME (t) && TREE_TYPE (t)
231 && lang_hooks.tree_inlining.anon_aggr_type_p (TREE_TYPE (t)))
232 {
233 /* For a VAR_DECL of anonymous type, we must also copy the
234 member VAR_DECLS here and rechain the DECL_ANON_UNION_ELEMS. */
235 tree members = NULL;
236 tree src;
237
238 for (src = DECL_ANON_UNION_ELEMS (t); src;
239 src = TREE_CHAIN (src))
240 {
241 tree member = remap_decl (TREE_VALUE (src), id);
242
243 gcc_assert (!TREE_PURPOSE (src));
244 members = tree_cons (NULL, member, members);
245 }
246 DECL_ANON_UNION_ELEMS (t) = nreverse (members);
247 }
248 #endif
249
250 /* Remember it, so that if we encounter this local entity
251 again we can reuse this copy. */
252 insert_decl_map (id, decl, t);
253 return t;
254 }
255
256 return unshare_expr ((tree) n->value);
257 }
258
259 static tree
remap_type_1(tree type,inline_data * id)260 remap_type_1 (tree type, inline_data *id)
261 {
262 tree new, t;
263
264 /* We do need a copy. build and register it now. If this is a pointer or
265 reference type, remap the designated type and make a new pointer or
266 reference type. */
267 if (TREE_CODE (type) == POINTER_TYPE)
268 {
269 new = build_pointer_type_for_mode (remap_type (TREE_TYPE (type), id),
270 TYPE_MODE (type),
271 TYPE_REF_CAN_ALIAS_ALL (type));
272 insert_decl_map (id, type, new);
273 return new;
274 }
275 else if (TREE_CODE (type) == REFERENCE_TYPE)
276 {
277 new = build_reference_type_for_mode (remap_type (TREE_TYPE (type), id),
278 TYPE_MODE (type),
279 TYPE_REF_CAN_ALIAS_ALL (type));
280 insert_decl_map (id, type, new);
281 return new;
282 }
283 else
284 new = copy_node (type);
285
286 insert_decl_map (id, type, new);
287
288 /* This is a new type, not a copy of an old type. Need to reassociate
289 variants. We can handle everything except the main variant lazily. */
290 t = TYPE_MAIN_VARIANT (type);
291 if (type != t)
292 {
293 t = remap_type (t, id);
294 TYPE_MAIN_VARIANT (new) = t;
295 TYPE_NEXT_VARIANT (new) = TYPE_MAIN_VARIANT (t);
296 TYPE_NEXT_VARIANT (t) = new;
297 }
298 else
299 {
300 TYPE_MAIN_VARIANT (new) = new;
301 TYPE_NEXT_VARIANT (new) = NULL;
302 }
303
304 if (TYPE_STUB_DECL (type))
305 TYPE_STUB_DECL (new) = remap_decl (TYPE_STUB_DECL (type), id);
306
307 /* Lazily create pointer and reference types. */
308 TYPE_POINTER_TO (new) = NULL;
309 TYPE_REFERENCE_TO (new) = NULL;
310
311 switch (TREE_CODE (new))
312 {
313 case INTEGER_TYPE:
314 case REAL_TYPE:
315 case ENUMERAL_TYPE:
316 case BOOLEAN_TYPE:
317 case CHAR_TYPE:
318 t = TYPE_MIN_VALUE (new);
319 if (t && TREE_CODE (t) != INTEGER_CST)
320 walk_tree (&TYPE_MIN_VALUE (new), copy_body_r, id, NULL);
321
322 t = TYPE_MAX_VALUE (new);
323 if (t && TREE_CODE (t) != INTEGER_CST)
324 walk_tree (&TYPE_MAX_VALUE (new), copy_body_r, id, NULL);
325 return new;
326
327 case FUNCTION_TYPE:
328 TREE_TYPE (new) = remap_type (TREE_TYPE (new), id);
329 walk_tree (&TYPE_ARG_TYPES (new), copy_body_r, id, NULL);
330 return new;
331
332 case ARRAY_TYPE:
333 TREE_TYPE (new) = remap_type (TREE_TYPE (new), id);
334 TYPE_DOMAIN (new) = remap_type (TYPE_DOMAIN (new), id);
335 break;
336
337 case RECORD_TYPE:
338 case UNION_TYPE:
339 case QUAL_UNION_TYPE:
340 {
341 tree f, nf = NULL;
342
343 for (f = TYPE_FIELDS (new); f ; f = TREE_CHAIN (f))
344 {
345 t = remap_decl (f, id);
346 DECL_CONTEXT (t) = new;
347 TREE_CHAIN (t) = nf;
348 nf = t;
349 }
350 TYPE_FIELDS (new) = nreverse (nf);
351 }
352 break;
353
354 case OFFSET_TYPE:
355 default:
356 /* Shouldn't have been thought variable sized. */
357 gcc_unreachable ();
358 }
359
360 walk_tree (&TYPE_SIZE (new), copy_body_r, id, NULL);
361 walk_tree (&TYPE_SIZE_UNIT (new), copy_body_r, id, NULL);
362
363 return new;
364 }
365
366 static tree
remap_type(tree type,inline_data * id)367 remap_type (tree type, inline_data *id)
368 {
369 splay_tree_node node;
370
371 if (type == NULL)
372 return type;
373
374 /* See if we have remapped this type. */
375 node = splay_tree_lookup (id->decl_map, (splay_tree_key) type);
376 if (node)
377 return (tree) node->value;
378
379 /* The type only needs remapping if it's variably modified. */
380 if (! variably_modified_type_p (type, id->callee))
381 {
382 insert_decl_map (id, type, type);
383 return type;
384 }
385
386 return remap_type_1 (type, id);
387 }
388
389 static tree
remap_decls(tree decls,inline_data * id)390 remap_decls (tree decls, inline_data *id)
391 {
392 tree old_var;
393 tree new_decls = NULL_TREE;
394
395 /* Remap its variables. */
396 for (old_var = decls; old_var; old_var = TREE_CHAIN (old_var))
397 {
398 tree new_var;
399
400 /* We can not chain the local static declarations into the unexpanded_var_list
401 as we can't duplicate them or break one decl rule. Go ahead and link
402 them into unexpanded_var_list. */
403 if (!lang_hooks.tree_inlining.auto_var_in_fn_p (old_var, id->callee)
404 && !DECL_EXTERNAL (old_var))
405 {
406 cfun->unexpanded_var_list = tree_cons (NULL_TREE, old_var,
407 cfun->unexpanded_var_list);
408 continue;
409 }
410
411 /* Remap the variable. */
412 new_var = remap_decl (old_var, id);
413
414 /* If we didn't remap this variable, so we can't mess with its
415 TREE_CHAIN. If we remapped this variable to the return slot, it's
416 already declared somewhere else, so don't declare it here. */
417 if (!new_var || new_var == id->retvar)
418 ;
419 else
420 {
421 gcc_assert (DECL_P (new_var));
422 TREE_CHAIN (new_var) = new_decls;
423 new_decls = new_var;
424 }
425 }
426
427 return nreverse (new_decls);
428 }
429
430 /* Copy the BLOCK to contain remapped versions of the variables
431 therein. And hook the new block into the block-tree. */
432
433 static void
remap_block(tree * block,inline_data * id)434 remap_block (tree *block, inline_data *id)
435 {
436 tree old_block;
437 tree new_block;
438 tree fn;
439
440 /* Make the new block. */
441 old_block = *block;
442 new_block = make_node (BLOCK);
443 TREE_USED (new_block) = TREE_USED (old_block);
444 BLOCK_ABSTRACT_ORIGIN (new_block) = old_block;
445 BLOCK_SOURCE_LOCATION (new_block) = BLOCK_SOURCE_LOCATION (old_block);
446 *block = new_block;
447
448 /* Remap its variables. */
449 BLOCK_VARS (new_block) = remap_decls (BLOCK_VARS (old_block), id);
450
451 fn = id->caller;
452 if (id->cloning_p)
453 /* We're building a clone; DECL_INITIAL is still
454 error_mark_node, and current_binding_level is the parm
455 binding level. */
456 lang_hooks.decls.insert_block (new_block);
457 /* Remember the remapped block. */
458 insert_decl_map (id, old_block, new_block);
459 }
460
461 /* Copy the whole block tree and root it in id->block. */
462 static tree
remap_blocks(tree block,inline_data * id)463 remap_blocks (tree block, inline_data *id)
464 {
465 tree t;
466 tree new = block;
467
468 if (!block)
469 return NULL;
470
471 remap_block (&new, id);
472 gcc_assert (new != block);
473 for (t = BLOCK_SUBBLOCKS (block); t ; t = BLOCK_CHAIN (t))
474 add_lexical_block (new, remap_blocks (t, id));
475 return new;
476 }
477
478 static void
copy_statement_list(tree * tp)479 copy_statement_list (tree *tp)
480 {
481 tree_stmt_iterator oi, ni;
482 tree new;
483
484 new = alloc_stmt_list ();
485 ni = tsi_start (new);
486 oi = tsi_start (*tp);
487 *tp = new;
488
489 for (; !tsi_end_p (oi); tsi_next (&oi))
490 tsi_link_after (&ni, tsi_stmt (oi), TSI_NEW_STMT);
491 }
492
493 static void
copy_bind_expr(tree * tp,int * walk_subtrees,inline_data * id)494 copy_bind_expr (tree *tp, int *walk_subtrees, inline_data *id)
495 {
496 tree block = BIND_EXPR_BLOCK (*tp);
497 /* Copy (and replace) the statement. */
498 copy_tree_r (tp, walk_subtrees, NULL);
499 if (block)
500 {
501 remap_block (&block, id);
502 BIND_EXPR_BLOCK (*tp) = block;
503 }
504
505 if (BIND_EXPR_VARS (*tp))
506 /* This will remap a lot of the same decls again, but this should be
507 harmless. */
508 BIND_EXPR_VARS (*tp) = remap_decls (BIND_EXPR_VARS (*tp), id);
509 }
510
511 /* Called from copy_body_id via walk_tree. DATA is really an
512 `inline_data *'. */
513
514 static tree
copy_body_r(tree * tp,int * walk_subtrees,void * data)515 copy_body_r (tree *tp, int *walk_subtrees, void *data)
516 {
517 inline_data *id = (inline_data *) data;
518 tree fn = id->callee;
519 tree new_block;
520
521 /* Begin by recognizing trees that we'll completely rewrite for the
522 inlining context. Our output for these trees is completely
523 different from out input (e.g. RETURN_EXPR is deleted, and morphs
524 into an edge). Further down, we'll handle trees that get
525 duplicated and/or tweaked. */
526
527 /* If this is a RETURN_STMT, change it into an EXPR_STMT and a
528 GOTO_STMT with the RET_LABEL as its target. */
529 if (TREE_CODE (*tp) == RETURN_EXPR && inlining_p (id))
530 {
531 tree assignment = TREE_OPERAND (*tp, 0);
532
533 /* If we're returning something, just turn that into an
534 assignment into the equivalent of the original RESULT_DECL.
535 If the "assignment" is just the result decl, the result
536 decl has already been set (e.g. a recent "foo (&result_decl,
537 ...)"); just toss the entire RETURN_EXPR. */
538 if (assignment && TREE_CODE (assignment) == MODIFY_EXPR)
539 {
540 /* Replace the RETURN_EXPR with (a copy of) the
541 MODIFY_EXPR hanging underneath. */
542 *tp = copy_node (assignment);
543 }
544 else /* Else the RETURN_EXPR returns no value. */
545 {
546 *tp = NULL;
547 return (void *)1;
548 }
549 }
550
551 /* Local variables and labels need to be replaced by equivalent
552 variables. We don't want to copy static variables; there's only
553 one of those, no matter how many times we inline the containing
554 function. Similarly for globals from an outer function. */
555 else if (lang_hooks.tree_inlining.auto_var_in_fn_p (*tp, fn))
556 {
557 tree new_decl;
558
559 /* Remap the declaration. */
560 new_decl = remap_decl (*tp, id);
561 gcc_assert (new_decl);
562 /* Replace this variable with the copy. */
563 STRIP_TYPE_NOPS (new_decl);
564 *tp = new_decl;
565 *walk_subtrees = 0;
566 }
567 else if (TREE_CODE (*tp) == STATEMENT_LIST)
568 copy_statement_list (tp);
569 else if (TREE_CODE (*tp) == SAVE_EXPR)
570 remap_save_expr (tp, id->decl_map, walk_subtrees);
571 else if (TREE_CODE (*tp) == LABEL_DECL
572 && (! DECL_CONTEXT (*tp)
573 || decl_function_context (*tp) == id->callee))
574 /* These may need to be remapped for EH handling. */
575 *tp = remap_decl (*tp, id);
576 else if (TREE_CODE (*tp) == BIND_EXPR)
577 copy_bind_expr (tp, walk_subtrees, id);
578 /* Types may need remapping as well. */
579 else if (TYPE_P (*tp))
580 *tp = remap_type (*tp, id);
581
582 /* If this is a constant, we have to copy the node iff the type will be
583 remapped. copy_tree_r will not copy a constant. */
584 else if (CONSTANT_CLASS_P (*tp))
585 {
586 tree new_type = remap_type (TREE_TYPE (*tp), id);
587
588 if (new_type == TREE_TYPE (*tp))
589 *walk_subtrees = 0;
590
591 else if (TREE_CODE (*tp) == INTEGER_CST)
592 *tp = build_int_cst_wide (new_type, TREE_INT_CST_LOW (*tp),
593 TREE_INT_CST_HIGH (*tp));
594 else
595 {
596 *tp = copy_node (*tp);
597 TREE_TYPE (*tp) = new_type;
598 }
599 }
600
601 /* Otherwise, just copy the node. Note that copy_tree_r already
602 knows not to copy VAR_DECLs, etc., so this is safe. */
603 else
604 {
605 /* Here we handle trees that are not completely rewritten.
606 First we detect some inlining-induced bogosities for
607 discarding. */
608 if (TREE_CODE (*tp) == MODIFY_EXPR
609 && TREE_OPERAND (*tp, 0) == TREE_OPERAND (*tp, 1)
610 && (lang_hooks.tree_inlining.auto_var_in_fn_p
611 (TREE_OPERAND (*tp, 0), fn)))
612 {
613 /* Some assignments VAR = VAR; don't generate any rtl code
614 and thus don't count as variable modification. Avoid
615 keeping bogosities like 0 = 0. */
616 tree decl = TREE_OPERAND (*tp, 0), value;
617 splay_tree_node n;
618
619 n = splay_tree_lookup (id->decl_map, (splay_tree_key) decl);
620 if (n)
621 {
622 value = (tree) n->value;
623 STRIP_TYPE_NOPS (value);
624 if (TREE_CONSTANT (value) || TREE_READONLY_DECL_P (value))
625 {
626 *tp = build_empty_stmt ();
627 return copy_body_r (tp, walk_subtrees, data);
628 }
629 }
630 }
631 else if (TREE_CODE (*tp) == INDIRECT_REF
632 && !id->versioning_p)
633 {
634 /* Get rid of *& from inline substitutions that can happen when a
635 pointer argument is an ADDR_EXPR. */
636 tree decl = TREE_OPERAND (*tp, 0);
637 splay_tree_node n;
638
639 n = splay_tree_lookup (id->decl_map, (splay_tree_key) decl);
640 if (n)
641 {
642 tree new;
643 tree old;
644 /* If we happen to get an ADDR_EXPR in n->value, strip
645 it manually here as we'll eventually get ADDR_EXPRs
646 which lie about their types pointed to. In this case
647 build_fold_indirect_ref wouldn't strip the INDIRECT_REF,
648 but we absolutely rely on that. As fold_indirect_ref
649 does other useful transformations, try that first, though. */
650 tree type = TREE_TYPE (TREE_TYPE ((tree)n->value));
651 new = unshare_expr ((tree)n->value);
652 old = *tp;
653 *tp = fold_indirect_ref_1 (type, new);
654 if (! *tp)
655 {
656 if (TREE_CODE (new) == ADDR_EXPR)
657 *tp = TREE_OPERAND (new, 0);
658 else
659 {
660 *tp = build1 (INDIRECT_REF, type, new);
661 TREE_THIS_VOLATILE (*tp) = TREE_THIS_VOLATILE (old);
662 }
663 }
664 *walk_subtrees = 0;
665 return NULL;
666 }
667 }
668
669 /* Here is the "usual case". Copy this tree node, and then
670 tweak some special cases. */
671 copy_tree_r (tp, walk_subtrees, id->versioning_p ? data : NULL);
672
673 /* If EXPR has block defined, map it to newly constructed block.
674 When inlining we want EXPRs without block appear in the block
675 of function call. */
676 if (IS_EXPR_CODE_CLASS (TREE_CODE_CLASS (TREE_CODE (*tp))))
677 {
678 new_block = id->block;
679 if (TREE_BLOCK (*tp))
680 {
681 splay_tree_node n;
682 n = splay_tree_lookup (id->decl_map,
683 (splay_tree_key) TREE_BLOCK (*tp));
684 gcc_assert (n);
685 new_block = (tree) n->value;
686 }
687 TREE_BLOCK (*tp) = new_block;
688 }
689
690 if (TREE_CODE (*tp) == RESX_EXPR && id->eh_region_offset)
691 TREE_OPERAND (*tp, 0) =
692 build_int_cst
693 (NULL_TREE,
694 id->eh_region_offset + TREE_INT_CST_LOW (TREE_OPERAND (*tp, 0)));
695
696 TREE_TYPE (*tp) = remap_type (TREE_TYPE (*tp), id);
697
698 /* The copied TARGET_EXPR has never been expanded, even if the
699 original node was expanded already. */
700 if (TREE_CODE (*tp) == TARGET_EXPR && TREE_OPERAND (*tp, 3))
701 {
702 TREE_OPERAND (*tp, 1) = TREE_OPERAND (*tp, 3);
703 TREE_OPERAND (*tp, 3) = NULL_TREE;
704 }
705
706 /* Variable substitution need not be simple. In particular, the
707 INDIRECT_REF substitution above. Make sure that TREE_CONSTANT
708 and friends are up-to-date. */
709 else if (TREE_CODE (*tp) == ADDR_EXPR)
710 {
711 walk_tree (&TREE_OPERAND (*tp, 0), copy_body_r, id, NULL);
712 /* Handle the case where we substituted an INDIRECT_REF
713 into the operand of the ADDR_EXPR. */
714 if (TREE_CODE (TREE_OPERAND (*tp, 0)) == INDIRECT_REF)
715 *tp = TREE_OPERAND (TREE_OPERAND (*tp, 0), 0);
716 else
717 recompute_tree_invarant_for_addr_expr (*tp);
718 *walk_subtrees = 0;
719 }
720 }
721
722 /* Keep iterating. */
723 return NULL_TREE;
724 }
725
726 /* Copy basic block, scale profile accordingly. Edges will be taken care of
727 later */
728
729 static basic_block
copy_bb(inline_data * id,basic_block bb,int frequency_scale,int count_scale)730 copy_bb (inline_data *id, basic_block bb, int frequency_scale, int count_scale)
731 {
732 block_stmt_iterator bsi, copy_bsi;
733 basic_block copy_basic_block;
734
735 /* create_basic_block() will append every new block to
736 basic_block_info automatically. */
737 copy_basic_block = create_basic_block (NULL, (void *) 0, bb->prev_bb->aux);
738 copy_basic_block->count = bb->count * count_scale / REG_BR_PROB_BASE;
739 copy_basic_block->frequency = (bb->frequency
740 * frequency_scale / REG_BR_PROB_BASE);
741 copy_bsi = bsi_start (copy_basic_block);
742
743 for (bsi = bsi_start (bb);
744 !bsi_end_p (bsi); bsi_next (&bsi))
745 {
746 tree stmt = bsi_stmt (bsi);
747 tree orig_stmt = stmt;
748
749 walk_tree (&stmt, copy_body_r, id, NULL);
750
751 /* RETURN_EXPR might be removed,
752 this is signalled by making stmt pointer NULL. */
753 if (stmt)
754 {
755 tree call, decl;
756
757 /* With return slot optimization we can end up with
758 non-gimple (foo *)&this->m, fix that here. */
759 if (TREE_CODE (stmt) == MODIFY_EXPR
760 && TREE_CODE (TREE_OPERAND (stmt, 1)) == NOP_EXPR
761 && !is_gimple_val (TREE_OPERAND (TREE_OPERAND (stmt, 1), 0)))
762 gimplify_stmt (&stmt);
763
764 bsi_insert_after (©_bsi, stmt, BSI_NEW_STMT);
765 call = get_call_expr_in (stmt);
766 /* We're duplicating a CALL_EXPR. Find any corresponding
767 callgraph edges and update or duplicate them. */
768 if (call && (decl = get_callee_fndecl (call)))
769 {
770 if (id->saving_p)
771 {
772 struct cgraph_node *node;
773 struct cgraph_edge *edge;
774
775 /* We're saving a copy of the body, so we'll update the
776 callgraph nodes in place. Note that we avoid
777 altering the original callgraph node; we begin with
778 the first clone. */
779 for (node = id->node->next_clone;
780 node;
781 node = node->next_clone)
782 {
783 edge = cgraph_edge (node, orig_stmt);
784 gcc_assert (edge);
785 edge->call_stmt = stmt;
786 }
787 }
788 else
789 {
790 struct cgraph_edge *edge;
791
792 /* We're cloning or inlining this body; duplicate the
793 associate callgraph nodes. */
794 if (!id->versioning_p)
795 {
796 edge = cgraph_edge (id->current_node, orig_stmt);
797 if (edge)
798 cgraph_clone_edge (edge, id->node, stmt,
799 REG_BR_PROB_BASE, 1, true);
800 }
801 }
802 if (id->versioning_p)
803 {
804 /* Update the call_expr on the edges from the new version
805 to its callees. */
806 struct cgraph_edge *edge;
807 edge = cgraph_edge (id->node, orig_stmt);
808 if (edge)
809 edge->call_stmt = stmt;
810 }
811 }
812 /* If you think we can abort here, you are wrong.
813 There is no region 0 in tree land. */
814 gcc_assert (lookup_stmt_eh_region_fn (id->callee_cfun, orig_stmt)
815 != 0);
816
817 if (tree_could_throw_p (stmt))
818 {
819 int region = lookup_stmt_eh_region_fn (id->callee_cfun, orig_stmt);
820 /* Add an entry for the copied tree in the EH hashtable.
821 When saving or cloning or versioning, use the hashtable in
822 cfun, and just copy the EH number. When inlining, use the
823 hashtable in the caller, and adjust the region number. */
824 if (region > 0)
825 add_stmt_to_eh_region (stmt, region + id->eh_region_offset);
826
827 /* If this tree doesn't have a region associated with it,
828 and there is a "current region,"
829 then associate this tree with the current region
830 and add edges associated with this region. */
831 if ((lookup_stmt_eh_region_fn (id->callee_cfun,
832 orig_stmt) <= 0
833 && id->eh_region > 0)
834 && tree_could_throw_p (stmt))
835 add_stmt_to_eh_region (stmt, id->eh_region);
836 }
837 }
838 }
839 return copy_basic_block;
840 }
841
842 /* Copy edges from BB into its copy constructed earlier, scale profile
843 accordingly. Edges will be taken care of later. Assume aux
844 pointers to point to the copies of each BB. */
845 static void
copy_edges_for_bb(basic_block bb,int count_scale)846 copy_edges_for_bb (basic_block bb, int count_scale)
847 {
848 basic_block new_bb = bb->aux;
849 edge_iterator ei;
850 edge old_edge;
851 block_stmt_iterator bsi;
852 int flags;
853
854 /* Use the indices from the original blocks to create edges for the
855 new ones. */
856 FOR_EACH_EDGE (old_edge, ei, bb->succs)
857 if (!(old_edge->flags & EDGE_EH))
858 {
859 edge new;
860
861 flags = old_edge->flags;
862
863 /* Return edges do get a FALLTHRU flag when the get inlined. */
864 if (old_edge->dest->index == EXIT_BLOCK && !old_edge->flags
865 && old_edge->dest->aux != EXIT_BLOCK_PTR)
866 flags |= EDGE_FALLTHRU;
867 new = make_edge (new_bb, old_edge->dest->aux, flags);
868 new->count = old_edge->count * count_scale / REG_BR_PROB_BASE;
869 new->probability = old_edge->probability;
870 }
871
872 if (bb->index == ENTRY_BLOCK || bb->index == EXIT_BLOCK)
873 return;
874
875 for (bsi = bsi_start (new_bb); !bsi_end_p (bsi);)
876 {
877 tree copy_stmt;
878
879 copy_stmt = bsi_stmt (bsi);
880 update_stmt (copy_stmt);
881 /* Do this before the possible split_block. */
882 bsi_next (&bsi);
883
884 /* If this tree could throw an exception, there are two
885 cases where we need to add abnormal edge(s): the
886 tree wasn't in a region and there is a "current
887 region" in the caller; or the original tree had
888 EH edges. In both cases split the block after the tree,
889 and add abnormal edge(s) as needed; we need both
890 those from the callee and the caller.
891 We check whether the copy can throw, because the const
892 propagation can change an INDIRECT_REF which throws
893 into a COMPONENT_REF which doesn't. If the copy
894 can throw, the original could also throw. */
895
896 if (tree_can_throw_internal (copy_stmt))
897 {
898 if (!bsi_end_p (bsi))
899 /* Note that bb's predecessor edges aren't necessarily
900 right at this point; split_block doesn't care. */
901 {
902 edge e = split_block (new_bb, copy_stmt);
903 new_bb = e->dest;
904 bsi = bsi_start (new_bb);
905 }
906
907 make_eh_edges (copy_stmt);
908 }
909 }
910 }
911
912 /* Wrapper for remap_decl so it can be used as a callback. */
913 static tree
remap_decl_1(tree decl,void * data)914 remap_decl_1 (tree decl, void *data)
915 {
916 return remap_decl (decl, data);
917 }
918
919 /* Make a copy of the body of FN so that it can be inserted inline in
920 another function. Walks FN via CFG, returns new fndecl. */
921
922 static tree
copy_cfg_body(inline_data * id,gcov_type count,int frequency,basic_block entry_block_map,basic_block exit_block_map)923 copy_cfg_body (inline_data * id, gcov_type count, int frequency,
924 basic_block entry_block_map, basic_block exit_block_map)
925 {
926 tree callee_fndecl = id->callee;
927 /* Original cfun for the callee, doesn't change. */
928 struct function *callee_cfun = DECL_STRUCT_FUNCTION (callee_fndecl);
929 /* Copy, built by this function. */
930 struct function *new_cfun;
931 /* Place to copy from; when a copy of the function was saved off earlier,
932 use that instead of the main copy. */
933 struct function *cfun_to_copy =
934 (struct function *) ggc_alloc_cleared (sizeof (struct function));
935 basic_block bb;
936 tree new_fndecl = NULL;
937 bool saving_or_cloning;
938 int count_scale, frequency_scale;
939
940 if (ENTRY_BLOCK_PTR_FOR_FUNCTION (callee_cfun)->count)
941 count_scale = (REG_BR_PROB_BASE * count
942 / ENTRY_BLOCK_PTR_FOR_FUNCTION (callee_cfun)->count);
943 else
944 count_scale = 1;
945
946 if (ENTRY_BLOCK_PTR_FOR_FUNCTION (callee_cfun)->frequency)
947 frequency_scale = (REG_BR_PROB_BASE * frequency
948 /
949 ENTRY_BLOCK_PTR_FOR_FUNCTION (callee_cfun)->frequency);
950 else
951 frequency_scale = count_scale;
952
953 /* Register specific tree functions. */
954 tree_register_cfg_hooks ();
955
956 /* Must have a CFG here at this point. */
957 gcc_assert (ENTRY_BLOCK_PTR_FOR_FUNCTION
958 (DECL_STRUCT_FUNCTION (callee_fndecl)));
959
960 *cfun_to_copy = *DECL_STRUCT_FUNCTION (callee_fndecl);
961
962 /* If there is a saved_cfg+saved_args lurking in the
963 struct function, a copy of the callee body was saved there, and
964 the 'struct cgraph edge' nodes have been fudged to point into the
965 saved body. Accordingly, we want to copy that saved body so the
966 callgraph edges will be recognized and cloned properly. */
967 if (cfun_to_copy->saved_cfg)
968 {
969 cfun_to_copy->cfg = cfun_to_copy->saved_cfg;
970 cfun_to_copy->eh = cfun_to_copy->saved_eh;
971 }
972 id->callee_cfun = cfun_to_copy;
973
974 /* If saving or cloning a function body, create new basic_block_info
975 and label_to_block_maps. Otherwise, we're duplicating a function
976 body for inlining; insert our new blocks and labels into the
977 existing varrays. */
978 saving_or_cloning = (id->saving_p || id->cloning_p || id->versioning_p);
979 if (saving_or_cloning)
980 {
981 new_cfun =
982 (struct function *) ggc_alloc_cleared (sizeof (struct function));
983 *new_cfun = *DECL_STRUCT_FUNCTION (callee_fndecl);
984 new_cfun->cfg = NULL;
985 new_cfun->decl = new_fndecl = copy_node (callee_fndecl);
986 new_cfun->ib_boundaries_block = (varray_type) 0;
987 DECL_STRUCT_FUNCTION (new_fndecl) = new_cfun;
988 push_cfun (new_cfun);
989 init_empty_tree_cfg ();
990
991 ENTRY_BLOCK_PTR->count =
992 (ENTRY_BLOCK_PTR_FOR_FUNCTION (callee_cfun)->count * count_scale /
993 REG_BR_PROB_BASE);
994 ENTRY_BLOCK_PTR->frequency =
995 (ENTRY_BLOCK_PTR_FOR_FUNCTION (callee_cfun)->frequency *
996 frequency_scale / REG_BR_PROB_BASE);
997 EXIT_BLOCK_PTR->count =
998 (EXIT_BLOCK_PTR_FOR_FUNCTION (callee_cfun)->count * count_scale /
999 REG_BR_PROB_BASE);
1000 EXIT_BLOCK_PTR->frequency =
1001 (EXIT_BLOCK_PTR_FOR_FUNCTION (callee_cfun)->frequency *
1002 frequency_scale / REG_BR_PROB_BASE);
1003
1004 entry_block_map = ENTRY_BLOCK_PTR;
1005 exit_block_map = EXIT_BLOCK_PTR;
1006 }
1007
1008 ENTRY_BLOCK_PTR_FOR_FUNCTION (cfun_to_copy)->aux = entry_block_map;
1009 EXIT_BLOCK_PTR_FOR_FUNCTION (cfun_to_copy)->aux = exit_block_map;
1010
1011
1012 /* Duplicate any exception-handling regions. */
1013 if (cfun->eh)
1014 {
1015 if (saving_or_cloning)
1016 init_eh_for_function ();
1017 id->eh_region_offset = duplicate_eh_regions (cfun_to_copy,
1018 remap_decl_1,
1019 id, id->eh_region);
1020 gcc_assert (inlining_p (id) || !id->eh_region_offset);
1021 }
1022 /* Use aux pointers to map the original blocks to copy. */
1023 FOR_EACH_BB_FN (bb, cfun_to_copy)
1024 bb->aux = copy_bb (id, bb, frequency_scale, count_scale);
1025 /* Now that we've duplicated the blocks, duplicate their edges. */
1026 FOR_ALL_BB_FN (bb, cfun_to_copy)
1027 copy_edges_for_bb (bb, count_scale);
1028 FOR_ALL_BB_FN (bb, cfun_to_copy)
1029 bb->aux = NULL;
1030
1031 if (saving_or_cloning)
1032 pop_cfun ();
1033
1034 return new_fndecl;
1035 }
1036
1037 /* Make a copy of the body of FN so that it can be inserted inline in
1038 another function. */
1039
1040 static tree
copy_generic_body(inline_data * id)1041 copy_generic_body (inline_data *id)
1042 {
1043 tree body;
1044 tree fndecl = id->callee;
1045
1046 body = DECL_SAVED_TREE (fndecl);
1047 walk_tree (&body, copy_body_r, id, NULL);
1048
1049 return body;
1050 }
1051
1052 static tree
copy_body(inline_data * id,gcov_type count,int frequency,basic_block entry_block_map,basic_block exit_block_map)1053 copy_body (inline_data *id, gcov_type count, int frequency,
1054 basic_block entry_block_map, basic_block exit_block_map)
1055 {
1056 tree fndecl = id->callee;
1057 tree body;
1058
1059 /* If this body has a CFG, walk CFG and copy. */
1060 gcc_assert (ENTRY_BLOCK_PTR_FOR_FUNCTION (DECL_STRUCT_FUNCTION (fndecl)));
1061 body = copy_cfg_body (id, count, frequency, entry_block_map, exit_block_map);
1062
1063 return body;
1064 }
1065
1066 /* Return true if VALUE is an ADDR_EXPR of an automatic variable
1067 defined in function FN, or of a data member thereof. */
1068
1069 static bool
self_inlining_addr_expr(tree value,tree fn)1070 self_inlining_addr_expr (tree value, tree fn)
1071 {
1072 tree var;
1073
1074 if (TREE_CODE (value) != ADDR_EXPR)
1075 return false;
1076
1077 var = get_base_address (TREE_OPERAND (value, 0));
1078
1079 return var && lang_hooks.tree_inlining.auto_var_in_fn_p (var, fn);
1080 }
1081
1082 static void
setup_one_parameter(inline_data * id,tree p,tree value,tree fn,basic_block bb,tree * vars)1083 setup_one_parameter (inline_data *id, tree p, tree value, tree fn,
1084 basic_block bb, tree *vars)
1085 {
1086 tree init_stmt;
1087 tree var;
1088 tree var_sub;
1089
1090 /* If the parameter is never assigned to, we may not need to
1091 create a new variable here at all. Instead, we may be able
1092 to just use the argument value. */
1093 if (TREE_READONLY (p)
1094 && !TREE_ADDRESSABLE (p)
1095 && value && !TREE_SIDE_EFFECTS (value))
1096 {
1097 /* We may produce non-gimple trees by adding NOPs or introduce
1098 invalid sharing when operand is not really constant.
1099 It is not big deal to prohibit constant propagation here as
1100 we will constant propagate in DOM1 pass anyway. */
1101 if (is_gimple_min_invariant (value)
1102 && lang_hooks.types_compatible_p (TREE_TYPE (value), TREE_TYPE (p))
1103 /* We have to be very careful about ADDR_EXPR. Make sure
1104 the base variable isn't a local variable of the inlined
1105 function, e.g., when doing recursive inlining, direct or
1106 mutually-recursive or whatever, which is why we don't
1107 just test whether fn == current_function_decl. */
1108 && ! self_inlining_addr_expr (value, fn))
1109 {
1110 insert_decl_map (id, p, value);
1111 return;
1112 }
1113 }
1114
1115 /* Make an equivalent VAR_DECL. Note that we must NOT remap the type
1116 here since the type of this decl must be visible to the calling
1117 function. */
1118 var = copy_decl_for_dup (p, fn, id->caller, /*versioning=*/false);
1119
1120 /* See if the frontend wants to pass this by invisible reference. If
1121 so, our new VAR_DECL will have REFERENCE_TYPE, and we need to
1122 replace uses of the PARM_DECL with dereferences. */
1123 if (TREE_TYPE (var) != TREE_TYPE (p)
1124 && POINTER_TYPE_P (TREE_TYPE (var))
1125 && TREE_TYPE (TREE_TYPE (var)) == TREE_TYPE (p))
1126 {
1127 insert_decl_map (id, var, var);
1128 var_sub = build_fold_indirect_ref (var);
1129 }
1130 else
1131 var_sub = var;
1132
1133 /* Register the VAR_DECL as the equivalent for the PARM_DECL;
1134 that way, when the PARM_DECL is encountered, it will be
1135 automatically replaced by the VAR_DECL. */
1136 insert_decl_map (id, p, var_sub);
1137
1138 /* Declare this new variable. */
1139 TREE_CHAIN (var) = *vars;
1140 *vars = var;
1141
1142 /* Make gimplifier happy about this variable. */
1143 DECL_SEEN_IN_BIND_EXPR_P (var) = 1;
1144
1145 /* Even if P was TREE_READONLY, the new VAR should not be.
1146 In the original code, we would have constructed a
1147 temporary, and then the function body would have never
1148 changed the value of P. However, now, we will be
1149 constructing VAR directly. The constructor body may
1150 change its value multiple times as it is being
1151 constructed. Therefore, it must not be TREE_READONLY;
1152 the back-end assumes that TREE_READONLY variable is
1153 assigned to only once. */
1154 if (TYPE_NEEDS_CONSTRUCTING (TREE_TYPE (p)))
1155 TREE_READONLY (var) = 0;
1156
1157 /* Initialize this VAR_DECL from the equivalent argument. Convert
1158 the argument to the proper type in case it was promoted. */
1159 if (value)
1160 {
1161 tree rhs = fold_convert (TREE_TYPE (var), value);
1162 block_stmt_iterator bsi = bsi_last (bb);
1163
1164 if (rhs == error_mark_node)
1165 return;
1166
1167 /* We want to use MODIFY_EXPR, not INIT_EXPR here so that we
1168 keep our trees in gimple form. */
1169 init_stmt = build (MODIFY_EXPR, TREE_TYPE (var), var, rhs);
1170
1171 /* If we did not create a gimple value and we did not create a gimple
1172 cast of a gimple value, then we will need to gimplify INIT_STMTS
1173 at the end. Note that is_gimple_cast only checks the outer
1174 tree code, not its operand. Thus the explicit check that its
1175 operand is a gimple value. */
1176 if (!is_gimple_val (rhs)
1177 && (!is_gimple_cast (rhs)
1178 || !is_gimple_val (TREE_OPERAND (rhs, 0))))
1179 gimplify_stmt (&init_stmt);
1180
1181 /* If VAR represents a zero-sized variable, it's possible that the
1182 assignment statment may result in no gimple statements. */
1183 if (init_stmt)
1184 bsi_insert_after (&bsi, init_stmt, BSI_NEW_STMT);
1185 }
1186 }
1187
1188 /* Generate code to initialize the parameters of the function at the
1189 top of the stack in ID from the ARGS (presented as a TREE_LIST). */
1190
1191 static void
initialize_inlined_parameters(inline_data * id,tree args,tree static_chain,tree fn,basic_block bb)1192 initialize_inlined_parameters (inline_data *id, tree args, tree static_chain,
1193 tree fn, basic_block bb)
1194 {
1195 tree parms;
1196 tree a;
1197 tree p;
1198 tree vars = NULL_TREE;
1199 int argnum = 0;
1200
1201 /* Figure out what the parameters are. */
1202 parms = DECL_ARGUMENTS (fn);
1203 if (fn == current_function_decl)
1204 parms = cfun->saved_args;
1205
1206 /* Loop through the parameter declarations, replacing each with an
1207 equivalent VAR_DECL, appropriately initialized. */
1208 for (p = parms, a = args; p;
1209 a = a ? TREE_CHAIN (a) : a, p = TREE_CHAIN (p))
1210 {
1211 tree value;
1212
1213 ++argnum;
1214
1215 /* Find the initializer. */
1216 value = lang_hooks.tree_inlining.convert_parm_for_inlining
1217 (p, a ? TREE_VALUE (a) : NULL_TREE, fn, argnum);
1218
1219 setup_one_parameter (id, p, value, fn, bb, &vars);
1220 }
1221
1222 /* Initialize the static chain. */
1223 p = DECL_STRUCT_FUNCTION (fn)->static_chain_decl;
1224 if (fn == current_function_decl)
1225 p = DECL_STRUCT_FUNCTION (fn)->saved_static_chain_decl;
1226 if (p)
1227 {
1228 /* No static chain? Seems like a bug in tree-nested.c. */
1229 gcc_assert (static_chain);
1230
1231 setup_one_parameter (id, p, static_chain, fn, bb, &vars);
1232 }
1233
1234 declare_inline_vars (id->block, vars);
1235 }
1236
1237 /* Declare a return variable to replace the RESULT_DECL for the
1238 function we are calling. An appropriate DECL_STMT is returned.
1239 The USE_STMT is filled to contain a use of the declaration to
1240 indicate the return value of the function.
1241
1242 RETURN_SLOT_ADDR, if non-null, was a fake parameter that
1243 took the address of the result. MODIFY_DEST, if non-null, was the LHS of
1244 the MODIFY_EXPR to which this call is the RHS.
1245
1246 The return value is a (possibly null) value that is the result of the
1247 function as seen by the callee. *USE_P is a (possibly null) value that
1248 holds the result as seen by the caller. */
1249
1250 static tree
declare_return_variable(inline_data * id,tree return_slot_addr,tree modify_dest,tree * use_p)1251 declare_return_variable (inline_data *id, tree return_slot_addr,
1252 tree modify_dest, tree *use_p)
1253 {
1254 tree callee = id->callee;
1255 tree caller = id->caller;
1256 tree result = DECL_RESULT (callee);
1257 tree callee_type = TREE_TYPE (result);
1258 tree caller_type = TREE_TYPE (TREE_TYPE (callee));
1259 tree var, use;
1260
1261 /* We don't need to do anything for functions that don't return
1262 anything. */
1263 if (!result || VOID_TYPE_P (callee_type))
1264 {
1265 *use_p = NULL_TREE;
1266 return NULL_TREE;
1267 }
1268
1269 /* If there was a return slot, then the return value is the
1270 dereferenced address of that object. */
1271 if (return_slot_addr)
1272 {
1273 /* The front end shouldn't have used both return_slot_addr and
1274 a modify expression. */
1275 gcc_assert (!modify_dest);
1276 if (DECL_BY_REFERENCE (result))
1277 var = return_slot_addr;
1278 else
1279 var = build_fold_indirect_ref (return_slot_addr);
1280 if (TREE_CODE (TREE_TYPE (result)) == COMPLEX_TYPE
1281 && !DECL_COMPLEX_GIMPLE_REG_P (result)
1282 && DECL_P (var))
1283 DECL_COMPLEX_GIMPLE_REG_P (var) = 0;
1284 use = NULL;
1285 goto done;
1286 }
1287
1288 /* All types requiring non-trivial constructors should have been handled. */
1289 gcc_assert (!TREE_ADDRESSABLE (callee_type));
1290
1291 /* Attempt to avoid creating a new temporary variable. */
1292 if (modify_dest)
1293 {
1294 bool use_it = false;
1295
1296 /* We can't use MODIFY_DEST if there's type promotion involved. */
1297 if (!lang_hooks.types_compatible_p (caller_type, callee_type))
1298 use_it = false;
1299
1300 /* ??? If we're assigning to a variable sized type, then we must
1301 reuse the destination variable, because we've no good way to
1302 create variable sized temporaries at this point. */
1303 else if (TREE_CODE (TYPE_SIZE_UNIT (caller_type)) != INTEGER_CST)
1304 use_it = true;
1305
1306 /* If the callee cannot possibly modify MODIFY_DEST, then we can
1307 reuse it as the result of the call directly. Don't do this if
1308 it would promote MODIFY_DEST to addressable. */
1309 else if (TREE_ADDRESSABLE (result))
1310 use_it = false;
1311 else
1312 {
1313 tree base_m = get_base_address (modify_dest);
1314
1315 /* If the base isn't a decl, then it's a pointer, and we don't
1316 know where that's going to go. */
1317 if (!DECL_P (base_m))
1318 use_it = false;
1319 else if (is_global_var (base_m))
1320 use_it = false;
1321 else if (TREE_CODE (TREE_TYPE (result)) == COMPLEX_TYPE
1322 && !DECL_COMPLEX_GIMPLE_REG_P (result)
1323 && DECL_COMPLEX_GIMPLE_REG_P (base_m))
1324 use_it = false;
1325 else if (!TREE_ADDRESSABLE (base_m))
1326 use_it = true;
1327 }
1328
1329 if (use_it)
1330 {
1331 var = modify_dest;
1332 use = NULL;
1333 goto done;
1334 }
1335 }
1336
1337 gcc_assert (TREE_CODE (TYPE_SIZE_UNIT (callee_type)) == INTEGER_CST);
1338
1339 var = copy_decl_for_dup (result, callee, caller, /*versioning=*/false);
1340
1341 DECL_SEEN_IN_BIND_EXPR_P (var) = 1;
1342 DECL_STRUCT_FUNCTION (caller)->unexpanded_var_list
1343 = tree_cons (NULL_TREE, var,
1344 DECL_STRUCT_FUNCTION (caller)->unexpanded_var_list);
1345
1346 /* Do not have the rest of GCC warn about this variable as it should
1347 not be visible to the user. */
1348 TREE_NO_WARNING (var) = 1;
1349
1350 /* Build the use expr. If the return type of the function was
1351 promoted, convert it back to the expected type. */
1352 use = var;
1353 if (!lang_hooks.types_compatible_p (TREE_TYPE (var), caller_type))
1354 use = fold_convert (caller_type, var);
1355
1356 done:
1357 /* Register the VAR_DECL as the equivalent for the RESULT_DECL; that
1358 way, when the RESULT_DECL is encountered, it will be
1359 automatically replaced by the VAR_DECL. */
1360 insert_decl_map (id, result, var);
1361
1362 /* Remember this so we can ignore it in remap_decls. */
1363 id->retvar = var;
1364
1365 *use_p = use;
1366 return var;
1367 }
1368
1369 /* Returns nonzero if a function can be inlined as a tree. */
1370
1371 bool
tree_inlinable_function_p(tree fn)1372 tree_inlinable_function_p (tree fn)
1373 {
1374 return inlinable_function_p (fn);
1375 }
1376
1377 static const char *inline_forbidden_reason;
1378
1379 static tree
inline_forbidden_p_1(tree * nodep,int * walk_subtrees ATTRIBUTE_UNUSED,void * fnp)1380 inline_forbidden_p_1 (tree *nodep, int *walk_subtrees ATTRIBUTE_UNUSED,
1381 void *fnp)
1382 {
1383 tree node = *nodep;
1384 tree fn = (tree) fnp;
1385 tree t;
1386
1387 switch (TREE_CODE (node))
1388 {
1389 case CALL_EXPR:
1390 /* Refuse to inline alloca call unless user explicitly forced so as
1391 this may change program's memory overhead drastically when the
1392 function using alloca is called in loop. In GCC present in
1393 SPEC2000 inlining into schedule_block cause it to require 2GB of
1394 RAM instead of 256MB. */
1395 if (alloca_call_p (node)
1396 && !lookup_attribute ("always_inline", DECL_ATTRIBUTES (fn)))
1397 {
1398 inline_forbidden_reason
1399 = G_("function %q+F can never be inlined because it uses "
1400 "alloca (override using the always_inline attribute)");
1401 return node;
1402 }
1403 t = get_callee_fndecl (node);
1404 if (! t)
1405 break;
1406
1407 /* We cannot inline functions that call setjmp. */
1408 if (setjmp_call_p (t))
1409 {
1410 inline_forbidden_reason
1411 = G_("function %q+F can never be inlined because it uses setjmp");
1412 return node;
1413 }
1414
1415 if (DECL_BUILT_IN_CLASS (t) == BUILT_IN_NORMAL)
1416 switch (DECL_FUNCTION_CODE (t))
1417 {
1418 /* We cannot inline functions that take a variable number of
1419 arguments. */
1420 case BUILT_IN_VA_START:
1421 case BUILT_IN_STDARG_START:
1422 case BUILT_IN_NEXT_ARG:
1423 case BUILT_IN_VA_END:
1424 inline_forbidden_reason
1425 = G_("function %q+F can never be inlined because it "
1426 "uses variable argument lists");
1427 return node;
1428
1429 case BUILT_IN_LONGJMP:
1430 /* We can't inline functions that call __builtin_longjmp at
1431 all. The non-local goto machinery really requires the
1432 destination be in a different function. If we allow the
1433 function calling __builtin_longjmp to be inlined into the
1434 function calling __builtin_setjmp, Things will Go Awry. */
1435 inline_forbidden_reason
1436 = G_("function %q+F can never be inlined because "
1437 "it uses setjmp-longjmp exception handling");
1438 return node;
1439
1440 case BUILT_IN_NONLOCAL_GOTO:
1441 /* Similarly. */
1442 inline_forbidden_reason
1443 = G_("function %q+F can never be inlined because "
1444 "it uses non-local goto");
1445 return node;
1446
1447 case BUILT_IN_RETURN:
1448 case BUILT_IN_APPLY_ARGS:
1449 /* If a __builtin_apply_args caller would be inlined,
1450 it would be saving arguments of the function it has
1451 been inlined into. Similarly __builtin_return would
1452 return from the function the inline has been inlined into. */
1453 inline_forbidden_reason
1454 = G_("function %q+F can never be inlined because "
1455 "it uses __builtin_return or __builtin_apply_args");
1456 return node;
1457
1458 default:
1459 break;
1460 }
1461 break;
1462
1463 case GOTO_EXPR:
1464 t = TREE_OPERAND (node, 0);
1465
1466 /* We will not inline a function which uses computed goto. The
1467 addresses of its local labels, which may be tucked into
1468 global storage, are of course not constant across
1469 instantiations, which causes unexpected behavior. */
1470 if (TREE_CODE (t) != LABEL_DECL)
1471 {
1472 inline_forbidden_reason
1473 = G_("function %q+F can never be inlined "
1474 "because it contains a computed goto");
1475 return node;
1476 }
1477 break;
1478
1479 case LABEL_EXPR:
1480 t = TREE_OPERAND (node, 0);
1481 if (DECL_NONLOCAL (t))
1482 {
1483 /* We cannot inline a function that receives a non-local goto
1484 because we cannot remap the destination label used in the
1485 function that is performing the non-local goto. */
1486 inline_forbidden_reason
1487 = G_("function %q+F can never be inlined "
1488 "because it receives a non-local goto");
1489 return node;
1490 }
1491 break;
1492
1493 case RECORD_TYPE:
1494 case UNION_TYPE:
1495 /* We cannot inline a function of the form
1496
1497 void F (int i) { struct S { int ar[i]; } s; }
1498
1499 Attempting to do so produces a catch-22.
1500 If walk_tree examines the TYPE_FIELDS chain of RECORD_TYPE/
1501 UNION_TYPE nodes, then it goes into infinite recursion on a
1502 structure containing a pointer to its own type. If it doesn't,
1503 then the type node for S doesn't get adjusted properly when
1504 F is inlined.
1505
1506 ??? This is likely no longer true, but it's too late in the 4.0
1507 cycle to try to find out. This should be checked for 4.1. */
1508 for (t = TYPE_FIELDS (node); t; t = TREE_CHAIN (t))
1509 if (variably_modified_type_p (TREE_TYPE (t), NULL))
1510 {
1511 inline_forbidden_reason
1512 = G_("function %q+F can never be inlined "
1513 "because it uses variable sized variables");
1514 return node;
1515 }
1516
1517 default:
1518 break;
1519 }
1520
1521 return NULL_TREE;
1522 }
1523
1524 /* Return subexpression representing possible alloca call, if any. */
1525 static tree
inline_forbidden_p(tree fndecl)1526 inline_forbidden_p (tree fndecl)
1527 {
1528 location_t saved_loc = input_location;
1529 block_stmt_iterator bsi;
1530 basic_block bb;
1531 tree ret = NULL_TREE;
1532
1533 FOR_EACH_BB_FN (bb, DECL_STRUCT_FUNCTION (fndecl))
1534 for (bsi = bsi_start (bb); !bsi_end_p (bsi); bsi_next (&bsi))
1535 {
1536 ret = walk_tree_without_duplicates (bsi_stmt_ptr (bsi),
1537 inline_forbidden_p_1, fndecl);
1538 if (ret)
1539 goto egress;
1540 }
1541
1542 egress:
1543 input_location = saved_loc;
1544 return ret;
1545 }
1546
1547 /* Returns nonzero if FN is a function that does not have any
1548 fundamental inline blocking properties. */
1549
1550 static bool
inlinable_function_p(tree fn)1551 inlinable_function_p (tree fn)
1552 {
1553 bool inlinable = true;
1554
1555 /* If we've already decided this function shouldn't be inlined,
1556 there's no need to check again. */
1557 if (DECL_UNINLINABLE (fn))
1558 return false;
1559
1560 /* See if there is any language-specific reason it cannot be
1561 inlined. (It is important that this hook be called early because
1562 in C++ it may result in template instantiation.)
1563 If the function is not inlinable for language-specific reasons,
1564 it is left up to the langhook to explain why. */
1565 inlinable = !lang_hooks.tree_inlining.cannot_inline_tree_fn (&fn);
1566
1567 /* If we don't have the function body available, we can't inline it.
1568 However, this should not be recorded since we also get here for
1569 forward declared inline functions. Therefore, return at once. */
1570 if (!DECL_SAVED_TREE (fn))
1571 return false;
1572
1573 /* If we're not inlining at all, then we cannot inline this function. */
1574 else if (!flag_inline_trees)
1575 inlinable = false;
1576
1577 /* Only try to inline functions if DECL_INLINE is set. This should be
1578 true for all functions declared `inline', and for all other functions
1579 as well with -finline-functions.
1580
1581 Don't think of disregarding DECL_INLINE when flag_inline_trees == 2;
1582 it's the front-end that must set DECL_INLINE in this case, because
1583 dwarf2out loses if a function that does not have DECL_INLINE set is
1584 inlined anyway. That is why we have both DECL_INLINE and
1585 DECL_DECLARED_INLINE_P. */
1586 /* FIXME: When flag_inline_trees dies, the check for flag_unit_at_a_time
1587 here should be redundant. */
1588 else if (!DECL_INLINE (fn) && !flag_unit_at_a_time)
1589 inlinable = false;
1590
1591 else if (inline_forbidden_p (fn))
1592 {
1593 /* See if we should warn about uninlinable functions. Previously,
1594 some of these warnings would be issued while trying to expand
1595 the function inline, but that would cause multiple warnings
1596 about functions that would for example call alloca. But since
1597 this a property of the function, just one warning is enough.
1598 As a bonus we can now give more details about the reason why a
1599 function is not inlinable.
1600 We only warn for functions declared `inline' by the user. */
1601 bool do_warning = (warn_inline
1602 && DECL_INLINE (fn)
1603 && DECL_DECLARED_INLINE_P (fn)
1604 && !DECL_IN_SYSTEM_HEADER (fn));
1605
1606 if (lookup_attribute ("always_inline", DECL_ATTRIBUTES (fn)))
1607 sorry (inline_forbidden_reason, fn);
1608 else if (do_warning)
1609 warning (OPT_Winline, inline_forbidden_reason, fn);
1610
1611 inlinable = false;
1612 }
1613
1614 /* Squirrel away the result so that we don't have to check again. */
1615 DECL_UNINLINABLE (fn) = !inlinable;
1616
1617 return inlinable;
1618 }
1619
1620 /* Estimate the cost of a memory move. Use machine dependent
1621 word size and take possible memcpy call into account. */
1622
1623 int
estimate_move_cost(tree type)1624 estimate_move_cost (tree type)
1625 {
1626 HOST_WIDE_INT size;
1627
1628 size = int_size_in_bytes (type);
1629
1630 if (size < 0 || size > MOVE_MAX_PIECES * MOVE_RATIO)
1631 /* Cost of a memcpy call, 3 arguments and the call. */
1632 return 4;
1633 else
1634 return ((size + MOVE_MAX_PIECES - 1) / MOVE_MAX_PIECES);
1635 }
1636
1637 /* Used by estimate_num_insns. Estimate number of instructions seen
1638 by given statement. */
1639
1640 static tree
estimate_num_insns_1(tree * tp,int * walk_subtrees,void * data)1641 estimate_num_insns_1 (tree *tp, int *walk_subtrees, void *data)
1642 {
1643 int *count = data;
1644 tree x = *tp;
1645
1646 if (IS_TYPE_OR_DECL_P (x))
1647 {
1648 *walk_subtrees = 0;
1649 return NULL;
1650 }
1651 /* Assume that constants and references counts nothing. These should
1652 be majorized by amount of operations among them we count later
1653 and are common target of CSE and similar optimizations. */
1654 else if (CONSTANT_CLASS_P (x) || REFERENCE_CLASS_P (x))
1655 return NULL;
1656
1657 switch (TREE_CODE (x))
1658 {
1659 /* Containers have no cost. */
1660 case TREE_LIST:
1661 case TREE_VEC:
1662 case BLOCK:
1663 case COMPONENT_REF:
1664 case BIT_FIELD_REF:
1665 case INDIRECT_REF:
1666 case ALIGN_INDIRECT_REF:
1667 case MISALIGNED_INDIRECT_REF:
1668 case ARRAY_REF:
1669 case ARRAY_RANGE_REF:
1670 case OBJ_TYPE_REF:
1671 case EXC_PTR_EXPR: /* ??? */
1672 case FILTER_EXPR: /* ??? */
1673 case COMPOUND_EXPR:
1674 case BIND_EXPR:
1675 case WITH_CLEANUP_EXPR:
1676 case NOP_EXPR:
1677 case VIEW_CONVERT_EXPR:
1678 case SAVE_EXPR:
1679 case ADDR_EXPR:
1680 case COMPLEX_EXPR:
1681 case RANGE_EXPR:
1682 case CASE_LABEL_EXPR:
1683 case SSA_NAME:
1684 case CATCH_EXPR:
1685 case EH_FILTER_EXPR:
1686 case STATEMENT_LIST:
1687 case ERROR_MARK:
1688 case NON_LVALUE_EXPR:
1689 case FDESC_EXPR:
1690 case VA_ARG_EXPR:
1691 case TRY_CATCH_EXPR:
1692 case TRY_FINALLY_EXPR:
1693 case LABEL_EXPR:
1694 case GOTO_EXPR:
1695 case RETURN_EXPR:
1696 case EXIT_EXPR:
1697 case LOOP_EXPR:
1698 case PHI_NODE:
1699 case WITH_SIZE_EXPR:
1700 break;
1701
1702 /* We don't account constants for now. Assume that the cost is amortized
1703 by operations that do use them. We may re-consider this decision once
1704 we are able to optimize the tree before estimating its size and break
1705 out static initializers. */
1706 case IDENTIFIER_NODE:
1707 case INTEGER_CST:
1708 case REAL_CST:
1709 case COMPLEX_CST:
1710 case VECTOR_CST:
1711 case STRING_CST:
1712 *walk_subtrees = 0;
1713 return NULL;
1714
1715 /* Try to estimate the cost of assignments. We have three cases to
1716 deal with:
1717 1) Simple assignments to registers;
1718 2) Stores to things that must live in memory. This includes
1719 "normal" stores to scalars, but also assignments of large
1720 structures, or constructors of big arrays;
1721 3) TARGET_EXPRs.
1722
1723 Let us look at the first two cases, assuming we have "a = b + C":
1724 <modify_expr <var_decl "a"> <plus_expr <var_decl "b"> <constant C>>
1725 If "a" is a GIMPLE register, the assignment to it is free on almost
1726 any target, because "a" usually ends up in a real register. Hence
1727 the only cost of this expression comes from the PLUS_EXPR, and we
1728 can ignore the MODIFY_EXPR.
1729 If "a" is not a GIMPLE register, the assignment to "a" will most
1730 likely be a real store, so the cost of the MODIFY_EXPR is the cost
1731 of moving something into "a", which we compute using the function
1732 estimate_move_cost.
1733
1734 The third case deals with TARGET_EXPRs, for which the semantics are
1735 that a temporary is assigned, unless the TARGET_EXPR itself is being
1736 assigned to something else. In the latter case we do not need the
1737 temporary. E.g. in <modify_expr <var_decl "a"> <target_expr>>, the
1738 MODIFY_EXPR is free. */
1739 case INIT_EXPR:
1740 case MODIFY_EXPR:
1741 /* Is the right and side a TARGET_EXPR? */
1742 if (TREE_CODE (TREE_OPERAND (x, 1)) == TARGET_EXPR)
1743 break;
1744 /* ... fall through ... */
1745
1746 case TARGET_EXPR:
1747 x = TREE_OPERAND (x, 0);
1748 /* Is this an assignments to a register? */
1749 if (is_gimple_reg (x))
1750 break;
1751 /* Otherwise it's a store, so fall through to compute the move cost. */
1752
1753 case CONSTRUCTOR:
1754 *count += estimate_move_cost (TREE_TYPE (x));
1755 break;
1756
1757 /* Assign cost of 1 to usual operations.
1758 ??? We may consider mapping RTL costs to this. */
1759 case COND_EXPR:
1760 case VEC_COND_EXPR:
1761
1762 case PLUS_EXPR:
1763 case MINUS_EXPR:
1764 case MULT_EXPR:
1765
1766 case FIX_TRUNC_EXPR:
1767 case FIX_CEIL_EXPR:
1768 case FIX_FLOOR_EXPR:
1769 case FIX_ROUND_EXPR:
1770
1771 case NEGATE_EXPR:
1772 case FLOAT_EXPR:
1773 case MIN_EXPR:
1774 case MAX_EXPR:
1775 case ABS_EXPR:
1776
1777 case LSHIFT_EXPR:
1778 case RSHIFT_EXPR:
1779 case LROTATE_EXPR:
1780 case RROTATE_EXPR:
1781 case VEC_LSHIFT_EXPR:
1782 case VEC_RSHIFT_EXPR:
1783
1784 case BIT_IOR_EXPR:
1785 case BIT_XOR_EXPR:
1786 case BIT_AND_EXPR:
1787 case BIT_NOT_EXPR:
1788
1789 case TRUTH_ANDIF_EXPR:
1790 case TRUTH_ORIF_EXPR:
1791 case TRUTH_AND_EXPR:
1792 case TRUTH_OR_EXPR:
1793 case TRUTH_XOR_EXPR:
1794 case TRUTH_NOT_EXPR:
1795
1796 case LT_EXPR:
1797 case LE_EXPR:
1798 case GT_EXPR:
1799 case GE_EXPR:
1800 case EQ_EXPR:
1801 case NE_EXPR:
1802 case ORDERED_EXPR:
1803 case UNORDERED_EXPR:
1804
1805 case UNLT_EXPR:
1806 case UNLE_EXPR:
1807 case UNGT_EXPR:
1808 case UNGE_EXPR:
1809 case UNEQ_EXPR:
1810 case LTGT_EXPR:
1811
1812 case CONVERT_EXPR:
1813
1814 case CONJ_EXPR:
1815
1816 case PREDECREMENT_EXPR:
1817 case PREINCREMENT_EXPR:
1818 case POSTDECREMENT_EXPR:
1819 case POSTINCREMENT_EXPR:
1820
1821 case SWITCH_EXPR:
1822
1823 case ASM_EXPR:
1824
1825 case REALIGN_LOAD_EXPR:
1826
1827 case REDUC_MAX_EXPR:
1828 case REDUC_MIN_EXPR:
1829 case REDUC_PLUS_EXPR:
1830
1831 case RESX_EXPR:
1832 *count += 1;
1833 break;
1834
1835 /* Few special cases of expensive operations. This is useful
1836 to avoid inlining on functions having too many of these. */
1837 case TRUNC_DIV_EXPR:
1838 case CEIL_DIV_EXPR:
1839 case FLOOR_DIV_EXPR:
1840 case ROUND_DIV_EXPR:
1841 case EXACT_DIV_EXPR:
1842 case TRUNC_MOD_EXPR:
1843 case CEIL_MOD_EXPR:
1844 case FLOOR_MOD_EXPR:
1845 case ROUND_MOD_EXPR:
1846 case RDIV_EXPR:
1847 *count += 10;
1848 break;
1849 case CALL_EXPR:
1850 {
1851 tree decl = get_callee_fndecl (x);
1852 tree arg;
1853
1854 if (decl && DECL_BUILT_IN_CLASS (decl) == BUILT_IN_NORMAL)
1855 switch (DECL_FUNCTION_CODE (decl))
1856 {
1857 case BUILT_IN_CONSTANT_P:
1858 *walk_subtrees = 0;
1859 return NULL_TREE;
1860 case BUILT_IN_EXPECT:
1861 return NULL_TREE;
1862 default:
1863 break;
1864 }
1865
1866 /* Our cost must be kept in sync with cgraph_estimate_size_after_inlining
1867 that does use function declaration to figure out the arguments. */
1868 if (!decl)
1869 {
1870 for (arg = TREE_OPERAND (x, 1); arg; arg = TREE_CHAIN (arg))
1871 *count += estimate_move_cost (TREE_TYPE (TREE_VALUE (arg)));
1872 }
1873 else
1874 {
1875 for (arg = DECL_ARGUMENTS (decl); arg; arg = TREE_CHAIN (arg))
1876 *count += estimate_move_cost (TREE_TYPE (arg));
1877 }
1878
1879 *count += PARAM_VALUE (PARAM_INLINE_CALL_COST);
1880 break;
1881 }
1882 default:
1883 gcc_unreachable ();
1884 }
1885 return NULL;
1886 }
1887
1888 /* Estimate number of instructions that will be created by expanding EXPR. */
1889
1890 int
estimate_num_insns(tree expr)1891 estimate_num_insns (tree expr)
1892 {
1893 int num = 0;
1894 struct pointer_set_t *visited_nodes;
1895 basic_block bb;
1896 block_stmt_iterator bsi;
1897 struct function *my_function;
1898
1899 /* If we're given an entire function, walk the CFG. */
1900 if (TREE_CODE (expr) == FUNCTION_DECL)
1901 {
1902 my_function = DECL_STRUCT_FUNCTION (expr);
1903 gcc_assert (my_function && my_function->cfg);
1904 visited_nodes = pointer_set_create ();
1905 FOR_EACH_BB_FN (bb, my_function)
1906 {
1907 for (bsi = bsi_start (bb);
1908 !bsi_end_p (bsi);
1909 bsi_next (&bsi))
1910 {
1911 walk_tree (bsi_stmt_ptr (bsi), estimate_num_insns_1,
1912 &num, visited_nodes);
1913 }
1914 }
1915 pointer_set_destroy (visited_nodes);
1916 }
1917 else
1918 walk_tree_without_duplicates (&expr, estimate_num_insns_1, &num);
1919
1920 return num;
1921 }
1922
1923 typedef struct function *function_p;
1924
1925 DEF_VEC_P(function_p);
1926 DEF_VEC_ALLOC_P(function_p,heap);
1927
1928 /* Initialized with NOGC, making this poisonous to the garbage collector. */
VEC(function_p,heap)1929 static VEC(function_p,heap) *cfun_stack;
1930
1931 void
1932 push_cfun (struct function *new_cfun)
1933 {
1934 VEC_safe_push (function_p, heap, cfun_stack, cfun);
1935 cfun = new_cfun;
1936 }
1937
1938 void
pop_cfun(void)1939 pop_cfun (void)
1940 {
1941 cfun = VEC_pop (function_p, cfun_stack);
1942 }
1943
1944 /* Install new lexical TREE_BLOCK underneath 'current_block'. */
1945 static void
add_lexical_block(tree current_block,tree new_block)1946 add_lexical_block (tree current_block, tree new_block)
1947 {
1948 tree *blk_p;
1949
1950 /* Walk to the last sub-block. */
1951 for (blk_p = &BLOCK_SUBBLOCKS (current_block);
1952 *blk_p;
1953 blk_p = &TREE_CHAIN (*blk_p))
1954 ;
1955 *blk_p = new_block;
1956 BLOCK_SUPERCONTEXT (new_block) = current_block;
1957 }
1958
1959 /* If *TP is a CALL_EXPR, replace it with its inline expansion. */
1960
1961 static bool
expand_call_inline(basic_block bb,tree stmt,tree * tp,void * data)1962 expand_call_inline (basic_block bb, tree stmt, tree *tp, void *data)
1963 {
1964 inline_data *id;
1965 tree t;
1966 tree use_retvar;
1967 tree fn;
1968 splay_tree st;
1969 tree args;
1970 tree return_slot_addr;
1971 tree modify_dest;
1972 location_t saved_location;
1973 struct cgraph_edge *cg_edge;
1974 const char *reason;
1975 basic_block return_block;
1976 edge e;
1977 block_stmt_iterator bsi, stmt_bsi;
1978 bool successfully_inlined = FALSE;
1979 tree t_step;
1980 tree var;
1981 struct cgraph_node *old_node;
1982 tree decl;
1983
1984 /* See what we've got. */
1985 id = (inline_data *) data;
1986 t = *tp;
1987
1988 /* Set input_location here so we get the right instantiation context
1989 if we call instantiate_decl from inlinable_function_p. */
1990 saved_location = input_location;
1991 if (EXPR_HAS_LOCATION (t))
1992 input_location = EXPR_LOCATION (t);
1993
1994 /* From here on, we're only interested in CALL_EXPRs. */
1995 if (TREE_CODE (t) != CALL_EXPR)
1996 goto egress;
1997
1998 /* First, see if we can figure out what function is being called.
1999 If we cannot, then there is no hope of inlining the function. */
2000 fn = get_callee_fndecl (t);
2001 if (!fn)
2002 goto egress;
2003
2004 /* Turn forward declarations into real ones. */
2005 fn = cgraph_node (fn)->decl;
2006
2007 /* If fn is a declaration of a function in a nested scope that was
2008 globally declared inline, we don't set its DECL_INITIAL.
2009 However, we can't blindly follow DECL_ABSTRACT_ORIGIN because the
2010 C++ front-end uses it for cdtors to refer to their internal
2011 declarations, that are not real functions. Fortunately those
2012 don't have trees to be saved, so we can tell by checking their
2013 DECL_SAVED_TREE. */
2014 if (! DECL_INITIAL (fn)
2015 && DECL_ABSTRACT_ORIGIN (fn)
2016 && DECL_SAVED_TREE (DECL_ABSTRACT_ORIGIN (fn)))
2017 fn = DECL_ABSTRACT_ORIGIN (fn);
2018
2019 /* Objective C and fortran still calls tree_rest_of_compilation directly.
2020 Kill this check once this is fixed. */
2021 if (!id->current_node->analyzed)
2022 goto egress;
2023
2024 cg_edge = cgraph_edge (id->current_node, stmt);
2025
2026 /* Constant propagation on argument done during previous inlining
2027 may create new direct call. Produce an edge for it. */
2028 if (!cg_edge)
2029 {
2030 struct cgraph_node *dest = cgraph_node (fn);
2031
2032 /* We have missing edge in the callgraph. This can happen in one case
2033 where previous inlining turned indirect call into direct call by
2034 constant propagating arguments. In all other cases we hit a bug
2035 (incorrect node sharing is most common reason for missing edges. */
2036 gcc_assert (dest->needed || !flag_unit_at_a_time);
2037 cgraph_create_edge (id->node, dest, stmt,
2038 bb->count, bb->loop_depth)->inline_failed
2039 = N_("originally indirect function call not considered for inlining");
2040 goto egress;
2041 }
2042
2043 /* Don't try to inline functions that are not well-suited to
2044 inlining. */
2045 if (!cgraph_inline_p (cg_edge, &reason))
2046 {
2047 if (lookup_attribute ("always_inline", DECL_ATTRIBUTES (fn))
2048 /* Avoid warnings during early inline pass. */
2049 && (!flag_unit_at_a_time || cgraph_global_info_ready))
2050 {
2051 sorry ("inlining failed in call to %q+F: %s", fn, reason);
2052 sorry ("called from here");
2053 }
2054 else if (warn_inline && DECL_DECLARED_INLINE_P (fn)
2055 && !DECL_IN_SYSTEM_HEADER (fn)
2056 && strlen (reason)
2057 && !lookup_attribute ("noinline", DECL_ATTRIBUTES (fn))
2058 /* Avoid warnings during early inline pass. */
2059 && (!flag_unit_at_a_time || cgraph_global_info_ready))
2060 {
2061 warning (OPT_Winline, "inlining failed in call to %q+F: %s",
2062 fn, reason);
2063 warning (OPT_Winline, "called from here");
2064 }
2065 goto egress;
2066 }
2067
2068 #ifdef ENABLE_CHECKING
2069 if (cg_edge->callee->decl != id->node->decl)
2070 verify_cgraph_node (cg_edge->callee);
2071 #endif
2072
2073 /* (TIGCC 20040926) The following code by Eric Botcazou fixes an ICE when
2074 inlining tries to change the mode of parameters or the return value. Eric
2075 Botcazou's comments explain the details. -- Kevin Kofler */
2076 /* We can't inline functions at a calling point where they are viewed
2077 with too different a prototype than the actual one, because the
2078 calling convention may not be the same on both sides. */
2079 if (TREE_CODE (TREE_OPERAND (t, 0)) == NOP_EXPR)
2080 {
2081 tree from_ftype = TREE_TYPE (TREE_TYPE (TREE_OPERAND (t, 0)));
2082 tree to_ftype = TREE_TYPE (fn);
2083
2084 if (from_ftype != to_ftype)
2085 {
2086 tree from_arg, to_arg;
2087
2088 /* If the calling point expects a return value and it is too
2089 different from the one actually returned, don't inline. */
2090 if (! VOID_TYPE_P (TREE_TYPE (from_ftype))
2091 && TYPE_MODE (TREE_TYPE (from_ftype))
2092 != TYPE_MODE (TREE_TYPE (to_ftype)))
2093 goto egress;
2094
2095 /* If the calling point doesn't pass at least the correct
2096 number of arguments with the correct modes, don't inline.
2097 Objective-C appears to add a trailing void parameter at
2098 the calling point under certain circumstances. */
2099 from_arg = TYPE_ARG_TYPES (from_ftype);
2100 to_arg = TYPE_ARG_TYPES (to_ftype);
2101
2102 while (to_arg)
2103 {
2104 if (! from_arg
2105 || TYPE_MODE (TREE_VALUE (from_arg))
2106 != TYPE_MODE (TREE_VALUE (to_arg)))
2107 goto egress;
2108
2109 from_arg = TREE_CHAIN (from_arg);
2110 to_arg = TREE_CHAIN (to_arg);
2111 }
2112 }
2113 }
2114
2115 /* We will be inlining this callee. */
2116
2117 id->eh_region = lookup_stmt_eh_region (stmt);
2118
2119 /* Split the block holding the CALL_EXPR. */
2120
2121 e = split_block (bb, stmt);
2122 bb = e->src;
2123 return_block = e->dest;
2124 remove_edge (e);
2125
2126 /* split_block splits before the statement, work around this by moving
2127 the call into the first half_bb. Not pretty, but seems easier than
2128 doing the CFG manipulation by hand when the CALL_EXPR is in the last
2129 statement in BB. */
2130 stmt_bsi = bsi_last (bb);
2131 bsi = bsi_start (return_block);
2132 if (!bsi_end_p (bsi))
2133 bsi_move_before (&stmt_bsi, &bsi);
2134 else
2135 {
2136 tree stmt = bsi_stmt (stmt_bsi);
2137 bsi_remove (&stmt_bsi);
2138 bsi_insert_after (&bsi, stmt, BSI_NEW_STMT);
2139 }
2140 stmt_bsi = bsi_start (return_block);
2141
2142 /* Build a block containing code to initialize the arguments, the
2143 actual inline expansion of the body, and a label for the return
2144 statements within the function to jump to. The type of the
2145 statement expression is the return type of the function call. */
2146 id->block = make_node (BLOCK);
2147 BLOCK_ABSTRACT_ORIGIN (id->block) = fn;
2148 BLOCK_SOURCE_LOCATION (id->block) = input_location;
2149 add_lexical_block (TREE_BLOCK (stmt), id->block);
2150
2151 /* Local declarations will be replaced by their equivalents in this
2152 map. */
2153 st = id->decl_map;
2154 id->decl_map = splay_tree_new (splay_tree_compare_pointers,
2155 NULL, NULL);
2156
2157 /* Initialize the parameters. */
2158 args = TREE_OPERAND (t, 1);
2159
2160 initialize_inlined_parameters (id, args, TREE_OPERAND (t, 2), fn, bb);
2161
2162 /* Record the function we are about to inline. */
2163 id->callee = fn;
2164
2165 if (DECL_STRUCT_FUNCTION (fn)->saved_blocks)
2166 add_lexical_block (id->block, remap_blocks (DECL_STRUCT_FUNCTION (fn)->saved_blocks, id));
2167 else if (DECL_INITIAL (fn))
2168 add_lexical_block (id->block, remap_blocks (DECL_INITIAL (fn), id));
2169
2170 /* Return statements in the function body will be replaced by jumps
2171 to the RET_LABEL. */
2172
2173 gcc_assert (DECL_INITIAL (fn));
2174 gcc_assert (TREE_CODE (DECL_INITIAL (fn)) == BLOCK);
2175
2176 /* Find the lhs to which the result of this call is assigned. */
2177 return_slot_addr = NULL;
2178 if (TREE_CODE (stmt) == MODIFY_EXPR)
2179 {
2180 modify_dest = TREE_OPERAND (stmt, 0);
2181
2182 /* The function which we are inlining might not return a value,
2183 in which case we should issue a warning that the function
2184 does not return a value. In that case the optimizers will
2185 see that the variable to which the value is assigned was not
2186 initialized. We do not want to issue a warning about that
2187 uninitialized variable. */
2188 if (DECL_P (modify_dest))
2189 TREE_NO_WARNING (modify_dest) = 1;
2190 if (CALL_EXPR_RETURN_SLOT_OPT (t))
2191 {
2192 return_slot_addr = build_fold_addr_expr (modify_dest);
2193 STRIP_USELESS_TYPE_CONVERSION (return_slot_addr);
2194 modify_dest = NULL;
2195 }
2196 }
2197 else
2198 modify_dest = NULL;
2199
2200 /* Declare the return variable for the function. */
2201 decl = declare_return_variable (id, return_slot_addr,
2202 modify_dest, &use_retvar);
2203 /* Do this only if declare_return_variable created a new one. */
2204 if (decl && !return_slot_addr && decl != modify_dest)
2205 declare_inline_vars (id->block, decl);
2206
2207 /* After we've initialized the parameters, we insert the body of the
2208 function itself. */
2209 old_node = id->current_node;
2210
2211 /* Anoint the callee-to-be-duplicated as the "current_node." When
2212 CALL_EXPRs within callee are duplicated, the edges from callee to
2213 callee's callees (caller's grandchildren) will be cloned. */
2214 id->current_node = cg_edge->callee;
2215
2216 /* This is it. Duplicate the callee body. Assume callee is
2217 pre-gimplified. Note that we must not alter the caller
2218 function in any way before this point, as this CALL_EXPR may be
2219 a self-referential call; if we're calling ourselves, we need to
2220 duplicate our body before altering anything. */
2221 copy_body (id, bb->count, bb->frequency, bb, return_block);
2222 id->current_node = old_node;
2223
2224 /* Add local vars in this inlined callee to caller. */
2225 t_step = id->callee_cfun->unexpanded_var_list;
2226 if (id->callee_cfun->saved_unexpanded_var_list)
2227 t_step = id->callee_cfun->saved_unexpanded_var_list;
2228 for (; t_step; t_step = TREE_CHAIN (t_step))
2229 {
2230 var = TREE_VALUE (t_step);
2231 if (TREE_STATIC (var) && !TREE_ASM_WRITTEN (var))
2232 cfun->unexpanded_var_list = tree_cons (NULL_TREE, var,
2233 cfun->unexpanded_var_list);
2234 else
2235 cfun->unexpanded_var_list = tree_cons (NULL_TREE, remap_decl (var, id),
2236 cfun->unexpanded_var_list);
2237 }
2238
2239 /* Clean up. */
2240 splay_tree_delete (id->decl_map);
2241 id->decl_map = st;
2242
2243 /* If the inlined function returns a result that we care about,
2244 clobber the CALL_EXPR with a reference to the return variable. */
2245 if (use_retvar && (TREE_CODE (bsi_stmt (stmt_bsi)) != CALL_EXPR))
2246 {
2247 *tp = use_retvar;
2248 maybe_clean_or_replace_eh_stmt (stmt, stmt);
2249 }
2250 else
2251 /* We're modifying a TSI owned by gimple_expand_calls_inline();
2252 tsi_delink() will leave the iterator in a sane state. */
2253 bsi_remove (&stmt_bsi);
2254
2255 bsi_next (&bsi);
2256 if (bsi_end_p (bsi))
2257 tree_purge_dead_eh_edges (return_block);
2258
2259 /* If the value of the new expression is ignored, that's OK. We
2260 don't warn about this for CALL_EXPRs, so we shouldn't warn about
2261 the equivalent inlined version either. */
2262 TREE_USED (*tp) = 1;
2263
2264 /* Output the inlining info for this abstract function, since it has been
2265 inlined. If we don't do this now, we can lose the information about the
2266 variables in the function when the blocks get blown away as soon as we
2267 remove the cgraph node. */
2268 (*debug_hooks->outlining_inline_function) (cg_edge->callee->decl);
2269
2270 /* Update callgraph if needed. */
2271 cgraph_remove_node (cg_edge->callee);
2272
2273 /* Declare the 'auto' variables added with this inlined body. */
2274 record_vars (BLOCK_VARS (id->block));
2275 id->block = NULL_TREE;
2276 successfully_inlined = TRUE;
2277
2278 egress:
2279 input_location = saved_location;
2280 return successfully_inlined;
2281 }
2282
2283 /* Expand call statements reachable from STMT_P.
2284 We can only have CALL_EXPRs as the "toplevel" tree code or nested
2285 in a MODIFY_EXPR. See tree-gimple.c:get_call_expr_in(). We can
2286 unfortunately not use that function here because we need a pointer
2287 to the CALL_EXPR, not the tree itself. */
2288
2289 static bool
gimple_expand_calls_inline(basic_block bb,inline_data * id)2290 gimple_expand_calls_inline (basic_block bb, inline_data *id)
2291 {
2292 block_stmt_iterator bsi;
2293
2294 /* Register specific tree functions. */
2295 tree_register_cfg_hooks ();
2296 for (bsi = bsi_start (bb); !bsi_end_p (bsi); bsi_next (&bsi))
2297 {
2298 tree *expr_p = bsi_stmt_ptr (bsi);
2299 tree stmt = *expr_p;
2300
2301 if (TREE_CODE (*expr_p) == MODIFY_EXPR)
2302 expr_p = &TREE_OPERAND (*expr_p, 1);
2303 if (TREE_CODE (*expr_p) == WITH_SIZE_EXPR)
2304 expr_p = &TREE_OPERAND (*expr_p, 0);
2305 if (TREE_CODE (*expr_p) == CALL_EXPR)
2306 if (expand_call_inline (bb, stmt, expr_p, id))
2307 return true;
2308 }
2309 return false;
2310 }
2311
2312 /* Expand calls to inline functions in the body of FN. */
2313
2314 void
optimize_inline_calls(tree fn)2315 optimize_inline_calls (tree fn)
2316 {
2317 inline_data id;
2318 tree prev_fn;
2319 basic_block bb;
2320 /* There is no point in performing inlining if errors have already
2321 occurred -- and we might crash if we try to inline invalid
2322 code. */
2323 if (errorcount || sorrycount)
2324 return;
2325
2326 /* Clear out ID. */
2327 memset (&id, 0, sizeof (id));
2328
2329 id.current_node = id.node = cgraph_node (fn);
2330 id.caller = fn;
2331 /* Or any functions that aren't finished yet. */
2332 prev_fn = NULL_TREE;
2333 if (current_function_decl)
2334 {
2335 id.caller = current_function_decl;
2336 prev_fn = current_function_decl;
2337 }
2338 push_gimplify_context ();
2339
2340 /* Reach the trees by walking over the CFG, and note the
2341 enclosing basic-blocks in the call edges. */
2342 /* We walk the blocks going forward, because inlined function bodies
2343 will split id->current_basic_block, and the new blocks will
2344 follow it; we'll trudge through them, processing their CALL_EXPRs
2345 along the way. */
2346 FOR_EACH_BB (bb)
2347 gimple_expand_calls_inline (bb, &id);
2348
2349
2350 pop_gimplify_context (NULL);
2351 /* Renumber the (code) basic_blocks consecutively. */
2352 compact_blocks ();
2353 /* Renumber the lexical scoping (non-code) blocks consecutively. */
2354 number_blocks (fn);
2355
2356 #ifdef ENABLE_CHECKING
2357 {
2358 struct cgraph_edge *e;
2359
2360 verify_cgraph_node (id.node);
2361
2362 /* Double check that we inlined everything we are supposed to inline. */
2363 for (e = id.node->callees; e; e = e->next_callee)
2364 gcc_assert (e->inline_failed);
2365 }
2366 #endif
2367 /* We need to rescale frequencies again to peak at REG_BR_PROB_BASE
2368 as inlining loops might increase the maximum. */
2369 if (ENTRY_BLOCK_PTR->count)
2370 counts_to_freqs ();
2371 fold_cond_expr_cond ();
2372 }
2373
2374 /* FN is a function that has a complete body, and CLONE is a function whose
2375 body is to be set to a copy of FN, mapping argument declarations according
2376 to the ARG_MAP splay_tree. */
2377
2378 void
clone_body(tree clone,tree fn,void * arg_map)2379 clone_body (tree clone, tree fn, void *arg_map)
2380 {
2381 inline_data id;
2382
2383 /* Clone the body, as if we were making an inline call. But, remap the
2384 parameters in the callee to the parameters of caller. */
2385 memset (&id, 0, sizeof (id));
2386 id.caller = clone;
2387 id.callee = fn;
2388 id.callee_cfun = DECL_STRUCT_FUNCTION (fn);
2389 id.decl_map = (splay_tree)arg_map;
2390
2391 /* Cloning is treated slightly differently from inlining. Set
2392 CLONING_P so that it's clear which operation we're performing. */
2393 id.cloning_p = true;
2394
2395 /* We're not inside any EH region. */
2396 id.eh_region = -1;
2397
2398 /* Actually copy the body. */
2399 append_to_statement_list_force (copy_generic_body (&id), &DECL_SAVED_TREE (clone));
2400 }
2401
2402 /* Save duplicate body in FN. MAP is used to pass around splay tree
2403 used to update arguments in restore_body. */
2404
2405 /* Make and return duplicate of body in FN. Put copies of DECL_ARGUMENTS
2406 in *arg_copy and of the static chain, if any, in *sc_copy. */
2407
2408 void
save_body(tree fn,tree * arg_copy,tree * sc_copy)2409 save_body (tree fn, tree *arg_copy, tree *sc_copy)
2410 {
2411 inline_data id;
2412 tree newdecl, *parg;
2413 basic_block fn_entry_block;
2414 tree t_step;
2415
2416 memset (&id, 0, sizeof (id));
2417 id.callee = fn;
2418 id.callee_cfun = DECL_STRUCT_FUNCTION (fn);
2419 id.caller = fn;
2420 id.node = cgraph_node (fn);
2421 id.saving_p = true;
2422 id.decl_map = splay_tree_new (splay_tree_compare_pointers, NULL, NULL);
2423 *arg_copy = DECL_ARGUMENTS (fn);
2424
2425 for (parg = arg_copy; *parg; parg = &TREE_CHAIN (*parg))
2426 {
2427 tree new = copy_node (*parg);
2428
2429 lang_hooks.dup_lang_specific_decl (new);
2430 DECL_ABSTRACT_ORIGIN (new) = DECL_ORIGIN (*parg);
2431 insert_decl_map (&id, *parg, new);
2432 TREE_CHAIN (new) = TREE_CHAIN (*parg);
2433 *parg = new;
2434 }
2435
2436 *sc_copy = DECL_STRUCT_FUNCTION (fn)->static_chain_decl;
2437 if (*sc_copy)
2438 {
2439 tree new = copy_node (*sc_copy);
2440
2441 lang_hooks.dup_lang_specific_decl (new);
2442 DECL_ABSTRACT_ORIGIN (new) = DECL_ORIGIN (*sc_copy);
2443 insert_decl_map (&id, *sc_copy, new);
2444 TREE_CHAIN (new) = TREE_CHAIN (*sc_copy);
2445 *sc_copy = new;
2446 }
2447
2448 /* We're not inside any EH region. */
2449 id.eh_region = -1;
2450
2451 insert_decl_map (&id, DECL_RESULT (fn), DECL_RESULT (fn));
2452
2453 DECL_STRUCT_FUNCTION (fn)->saved_blocks
2454 = remap_blocks (DECL_INITIAL (fn), &id);
2455 for (t_step = id.callee_cfun->unexpanded_var_list;
2456 t_step;
2457 t_step = TREE_CHAIN (t_step))
2458 {
2459 tree var = TREE_VALUE (t_step);
2460 if (TREE_STATIC (var) && !TREE_ASM_WRITTEN (var))
2461 cfun->saved_unexpanded_var_list
2462 = tree_cons (NULL_TREE, var, cfun->saved_unexpanded_var_list);
2463 else
2464 cfun->saved_unexpanded_var_list
2465 = tree_cons (NULL_TREE, remap_decl (var, &id),
2466 cfun->saved_unexpanded_var_list);
2467 }
2468
2469 /* Actually copy the body, including a new (struct function *) and CFG.
2470 EH info is also duplicated so its labels point into the copied
2471 CFG, not the original. */
2472 fn_entry_block = ENTRY_BLOCK_PTR_FOR_FUNCTION (DECL_STRUCT_FUNCTION (fn));
2473 newdecl = copy_body (&id, fn_entry_block->count, fn_entry_block->frequency,
2474 NULL, NULL);
2475 DECL_STRUCT_FUNCTION (fn)->saved_cfg = DECL_STRUCT_FUNCTION (newdecl)->cfg;
2476 DECL_STRUCT_FUNCTION (fn)->saved_eh = DECL_STRUCT_FUNCTION (newdecl)->eh;
2477
2478 /* Clean up. */
2479 splay_tree_delete (id.decl_map);
2480 }
2481
2482 /* Passed to walk_tree. Copies the node pointed to, if appropriate. */
2483
2484 tree
copy_tree_r(tree * tp,int * walk_subtrees,void * data ATTRIBUTE_UNUSED)2485 copy_tree_r (tree *tp, int *walk_subtrees, void *data ATTRIBUTE_UNUSED)
2486 {
2487 enum tree_code code = TREE_CODE (*tp);
2488 inline_data *id = (inline_data *) data;
2489
2490 /* We make copies of most nodes. */
2491 if (IS_EXPR_CODE_CLASS (TREE_CODE_CLASS (code))
2492 || code == TREE_LIST
2493 || code == TREE_VEC
2494 || code == TYPE_DECL)
2495 {
2496 /* Because the chain gets clobbered when we make a copy, we save it
2497 here. */
2498 tree chain = TREE_CHAIN (*tp);
2499 tree new;
2500
2501 if (id && id->versioning_p && replace_ref_tree (id, tp))
2502 {
2503 *walk_subtrees = 0;
2504 return NULL_TREE;
2505 }
2506 /* Copy the node. */
2507 new = copy_node (*tp);
2508
2509 /* Propagate mudflap marked-ness. */
2510 if (flag_mudflap && mf_marked_p (*tp))
2511 mf_mark (new);
2512
2513 *tp = new;
2514
2515 /* Now, restore the chain, if appropriate. That will cause
2516 walk_tree to walk into the chain as well. */
2517 if (code == PARM_DECL || code == TREE_LIST)
2518 TREE_CHAIN (*tp) = chain;
2519
2520 /* For now, we don't update BLOCKs when we make copies. So, we
2521 have to nullify all BIND_EXPRs. */
2522 if (TREE_CODE (*tp) == BIND_EXPR)
2523 BIND_EXPR_BLOCK (*tp) = NULL_TREE;
2524 }
2525 else if (code == CONSTRUCTOR)
2526 {
2527 /* CONSTRUCTOR nodes need special handling because
2528 we need to duplicate the vector of elements. */
2529 tree new;
2530
2531 new = copy_node (*tp);
2532
2533 /* Propagate mudflap marked-ness. */
2534 if (flag_mudflap && mf_marked_p (*tp))
2535 mf_mark (new);
2536
2537 CONSTRUCTOR_ELTS (new) = VEC_copy (constructor_elt, gc,
2538 CONSTRUCTOR_ELTS (*tp));
2539 *tp = new;
2540 }
2541 else if (TREE_CODE_CLASS (code) == tcc_type)
2542 *walk_subtrees = 0;
2543 else if (TREE_CODE_CLASS (code) == tcc_declaration)
2544 *walk_subtrees = 0;
2545 else if (TREE_CODE_CLASS (code) == tcc_constant)
2546 *walk_subtrees = 0;
2547 else
2548 gcc_assert (code != STATEMENT_LIST);
2549 return NULL_TREE;
2550 }
2551
2552 /* The SAVE_EXPR pointed to by TP is being copied. If ST contains
2553 information indicating to what new SAVE_EXPR this one should be mapped,
2554 use that one. Otherwise, create a new node and enter it in ST. FN is
2555 the function into which the copy will be placed. */
2556
2557 static void
remap_save_expr(tree * tp,void * st_,int * walk_subtrees)2558 remap_save_expr (tree *tp, void *st_, int *walk_subtrees)
2559 {
2560 splay_tree st = (splay_tree) st_;
2561 splay_tree_node n;
2562 tree t;
2563
2564 /* See if we already encountered this SAVE_EXPR. */
2565 n = splay_tree_lookup (st, (splay_tree_key) *tp);
2566
2567 /* If we didn't already remap this SAVE_EXPR, do so now. */
2568 if (!n)
2569 {
2570 t = copy_node (*tp);
2571
2572 /* Remember this SAVE_EXPR. */
2573 splay_tree_insert (st, (splay_tree_key) *tp, (splay_tree_value) t);
2574 /* Make sure we don't remap an already-remapped SAVE_EXPR. */
2575 splay_tree_insert (st, (splay_tree_key) t, (splay_tree_value) t);
2576 }
2577 else
2578 {
2579 /* We've already walked into this SAVE_EXPR; don't do it again. */
2580 *walk_subtrees = 0;
2581 t = (tree) n->value;
2582 }
2583
2584 /* Replace this SAVE_EXPR with the copy. */
2585 *tp = t;
2586 }
2587
2588 /* Called via walk_tree. If *TP points to a DECL_STMT for a local label,
2589 copies the declaration and enters it in the splay_tree in DATA (which is
2590 really an `inline_data *'). */
2591
2592 static tree
mark_local_for_remap_r(tree * tp,int * walk_subtrees ATTRIBUTE_UNUSED,void * data)2593 mark_local_for_remap_r (tree *tp, int *walk_subtrees ATTRIBUTE_UNUSED,
2594 void *data)
2595 {
2596 inline_data *id = (inline_data *) data;
2597
2598 /* Don't walk into types. */
2599 if (TYPE_P (*tp))
2600 *walk_subtrees = 0;
2601
2602 else if (TREE_CODE (*tp) == LABEL_EXPR)
2603 {
2604 tree decl = TREE_OPERAND (*tp, 0);
2605
2606 /* Copy the decl and remember the copy. */
2607 insert_decl_map (id, decl,
2608 copy_decl_for_dup (decl, DECL_CONTEXT (decl),
2609 DECL_CONTEXT (decl), /*versioning=*/false));
2610 }
2611
2612 return NULL_TREE;
2613 }
2614
2615 /* Perform any modifications to EXPR required when it is unsaved. Does
2616 not recurse into EXPR's subtrees. */
2617
2618 static void
unsave_expr_1(tree expr)2619 unsave_expr_1 (tree expr)
2620 {
2621 switch (TREE_CODE (expr))
2622 {
2623 case TARGET_EXPR:
2624 /* Don't mess with a TARGET_EXPR that hasn't been expanded.
2625 It's OK for this to happen if it was part of a subtree that
2626 isn't immediately expanded, such as operand 2 of another
2627 TARGET_EXPR. */
2628 if (TREE_OPERAND (expr, 1))
2629 break;
2630
2631 TREE_OPERAND (expr, 1) = TREE_OPERAND (expr, 3);
2632 TREE_OPERAND (expr, 3) = NULL_TREE;
2633 break;
2634
2635 default:
2636 break;
2637 }
2638 }
2639
2640 /* Called via walk_tree when an expression is unsaved. Using the
2641 splay_tree pointed to by ST (which is really a `splay_tree'),
2642 remaps all local declarations to appropriate replacements. */
2643
2644 static tree
unsave_r(tree * tp,int * walk_subtrees,void * data)2645 unsave_r (tree *tp, int *walk_subtrees, void *data)
2646 {
2647 inline_data *id = (inline_data *) data;
2648 splay_tree st = id->decl_map;
2649 splay_tree_node n;
2650
2651 /* Only a local declaration (variable or label). */
2652 if ((TREE_CODE (*tp) == VAR_DECL && !TREE_STATIC (*tp))
2653 || TREE_CODE (*tp) == LABEL_DECL)
2654 {
2655 /* Lookup the declaration. */
2656 n = splay_tree_lookup (st, (splay_tree_key) *tp);
2657
2658 /* If it's there, remap it. */
2659 if (n)
2660 *tp = (tree) n->value;
2661 }
2662
2663 else if (TREE_CODE (*tp) == STATEMENT_LIST)
2664 copy_statement_list (tp);
2665 else if (TREE_CODE (*tp) == BIND_EXPR)
2666 copy_bind_expr (tp, walk_subtrees, id);
2667 else if (TREE_CODE (*tp) == SAVE_EXPR)
2668 remap_save_expr (tp, st, walk_subtrees);
2669 else
2670 {
2671 copy_tree_r (tp, walk_subtrees, NULL);
2672
2673 /* Do whatever unsaving is required. */
2674 unsave_expr_1 (*tp);
2675 }
2676
2677 /* Keep iterating. */
2678 return NULL_TREE;
2679 }
2680
2681 /* Copies everything in EXPR and replaces variables, labels
2682 and SAVE_EXPRs local to EXPR. */
2683
2684 tree
unsave_expr_now(tree expr)2685 unsave_expr_now (tree expr)
2686 {
2687 inline_data id;
2688
2689 /* There's nothing to do for NULL_TREE. */
2690 if (expr == 0)
2691 return expr;
2692
2693 /* Set up ID. */
2694 memset (&id, 0, sizeof (id));
2695 id.callee = current_function_decl;
2696 id.caller = current_function_decl;
2697 id.decl_map = splay_tree_new (splay_tree_compare_pointers, NULL, NULL);
2698
2699 /* Walk the tree once to find local labels. */
2700 walk_tree_without_duplicates (&expr, mark_local_for_remap_r, &id);
2701
2702 /* Walk the tree again, copying, remapping, and unsaving. */
2703 walk_tree (&expr, unsave_r, &id, NULL);
2704
2705 /* Clean up. */
2706 splay_tree_delete (id.decl_map);
2707
2708 return expr;
2709 }
2710
2711 /* Allow someone to determine if SEARCH is a child of TOP from gdb. */
2712
2713 static tree
debug_find_tree_1(tree * tp,int * walk_subtrees ATTRIBUTE_UNUSED,void * data)2714 debug_find_tree_1 (tree *tp, int *walk_subtrees ATTRIBUTE_UNUSED, void *data)
2715 {
2716 if (*tp == data)
2717 return (tree) data;
2718 else
2719 return NULL;
2720 }
2721
2722 bool
debug_find_tree(tree top,tree search)2723 debug_find_tree (tree top, tree search)
2724 {
2725 return walk_tree_without_duplicates (&top, debug_find_tree_1, search) != 0;
2726 }
2727
2728
2729 /* Declare the variables created by the inliner. Add all the variables in
2730 VARS to BIND_EXPR. */
2731
2732 static void
declare_inline_vars(tree block,tree vars)2733 declare_inline_vars (tree block, tree vars)
2734 {
2735 tree t;
2736 for (t = vars; t; t = TREE_CHAIN (t))
2737 DECL_SEEN_IN_BIND_EXPR_P (t) = 1;
2738
2739 if (block)
2740 BLOCK_VARS (block) = chainon (BLOCK_VARS (block), vars);
2741 }
2742
2743
2744 /* Copy NODE (which must be a DECL). The DECL originally was in the FROM_FN,
2745 but now it will be in the TO_FN. VERSIONING means that this function
2746 is used by the versioning utility (not inlining or cloning). */
2747
2748 tree
copy_decl_for_dup(tree decl,tree from_fn,tree to_fn,bool versioning)2749 copy_decl_for_dup (tree decl, tree from_fn, tree to_fn, bool versioning)
2750 {
2751 tree copy;
2752
2753 gcc_assert (DECL_P (decl));
2754 /* Copy the declaration. */
2755 if (!versioning
2756 && (TREE_CODE (decl) == PARM_DECL
2757 || TREE_CODE (decl) == RESULT_DECL))
2758 {
2759 tree type = TREE_TYPE (decl);
2760
2761 /* For a parameter or result, we must make an equivalent VAR_DECL,
2762 not a new PARM_DECL. */
2763 copy = build_decl (VAR_DECL, DECL_NAME (decl), type);
2764 TREE_ADDRESSABLE (copy) = TREE_ADDRESSABLE (decl);
2765 TREE_READONLY (copy) = TREE_READONLY (decl);
2766 TREE_THIS_VOLATILE (copy) = TREE_THIS_VOLATILE (decl);
2767 DECL_COMPLEX_GIMPLE_REG_P (copy) = DECL_COMPLEX_GIMPLE_REG_P (decl);
2768 }
2769 else
2770 {
2771 copy = copy_node (decl);
2772 /* The COPY is not abstract; it will be generated in TO_FN. */
2773 DECL_ABSTRACT (copy) = 0;
2774 lang_hooks.dup_lang_specific_decl (copy);
2775
2776 /* TREE_ADDRESSABLE isn't used to indicate that a label's
2777 address has been taken; it's for internal bookkeeping in
2778 expand_goto_internal. */
2779 if (TREE_CODE (copy) == LABEL_DECL)
2780 {
2781 TREE_ADDRESSABLE (copy) = 0;
2782 LABEL_DECL_UID (copy) = -1;
2783 }
2784 }
2785
2786 /* Don't generate debug information for the copy if we wouldn't have
2787 generated it for the copy either. */
2788 DECL_ARTIFICIAL (copy) = DECL_ARTIFICIAL (decl);
2789 DECL_IGNORED_P (copy) = DECL_IGNORED_P (decl);
2790
2791 /* Set the DECL_ABSTRACT_ORIGIN so the debugging routines know what
2792 declaration inspired this copy. */
2793 DECL_ABSTRACT_ORIGIN (copy) = DECL_ORIGIN (decl);
2794
2795 /* The new variable/label has no RTL, yet. */
2796 if (CODE_CONTAINS_STRUCT (TREE_CODE (copy), TS_DECL_WRTL)
2797 && !TREE_STATIC (copy) && !DECL_EXTERNAL (copy))
2798 SET_DECL_RTL (copy, NULL_RTX);
2799
2800 /* These args would always appear unused, if not for this. */
2801 TREE_USED (copy) = 1;
2802
2803 /* Set the context for the new declaration. */
2804 if (!DECL_CONTEXT (decl))
2805 /* Globals stay global. */
2806 ;
2807 else if (DECL_CONTEXT (decl) != from_fn)
2808 /* Things that weren't in the scope of the function we're inlining
2809 from aren't in the scope we're inlining to, either. */
2810 ;
2811 else if (TREE_STATIC (decl))
2812 /* Function-scoped static variables should stay in the original
2813 function. */
2814 ;
2815 else
2816 /* Ordinary automatic local variables are now in the scope of the
2817 new function. */
2818 DECL_CONTEXT (copy) = to_fn;
2819
2820 return copy;
2821 }
2822
2823 /* Return a copy of the function's argument tree. */
2824 static tree
copy_arguments_for_versioning(tree orig_parm,inline_data * id)2825 copy_arguments_for_versioning (tree orig_parm, inline_data * id)
2826 {
2827 tree *arg_copy, *parg;
2828
2829 arg_copy = &orig_parm;
2830 for (parg = arg_copy; *parg; parg = &TREE_CHAIN (*parg))
2831 {
2832 tree new = remap_decl (*parg, id);
2833 lang_hooks.dup_lang_specific_decl (new);
2834 TREE_CHAIN (new) = TREE_CHAIN (*parg);
2835 *parg = new;
2836 }
2837 return orig_parm;
2838 }
2839
2840 /* Return a copy of the function's static chain. */
2841 static tree
copy_static_chain(tree static_chain,inline_data * id)2842 copy_static_chain (tree static_chain, inline_data * id)
2843 {
2844 tree *chain_copy, *pvar;
2845
2846 chain_copy = &static_chain;
2847 for (pvar = chain_copy; *pvar; pvar = &TREE_CHAIN (*pvar))
2848 {
2849 tree new = remap_decl (*pvar, id);
2850 lang_hooks.dup_lang_specific_decl (new);
2851 TREE_CHAIN (new) = TREE_CHAIN (*pvar);
2852 *pvar = new;
2853 }
2854 return static_chain;
2855 }
2856
2857 /* Return true if the function is allowed to be versioned.
2858 This is a guard for the versioning functionality. */
2859 bool
tree_versionable_function_p(tree fndecl)2860 tree_versionable_function_p (tree fndecl)
2861 {
2862 if (fndecl == NULL_TREE)
2863 return false;
2864 /* ??? There are cases where a function is
2865 uninlinable but can be versioned. */
2866 if (!tree_inlinable_function_p (fndecl))
2867 return false;
2868
2869 return true;
2870 }
2871
2872 /* Create a copy of a function's tree.
2873 OLD_DECL and NEW_DECL are FUNCTION_DECL tree nodes
2874 of the original function and the new copied function
2875 respectively. In case we want to replace a DECL
2876 tree with another tree while duplicating the function's
2877 body, TREE_MAP represents the mapping between these
2878 trees. */
2879 void
tree_function_versioning(tree old_decl,tree new_decl,varray_type tree_map)2880 tree_function_versioning (tree old_decl, tree new_decl, varray_type tree_map)
2881 {
2882 struct cgraph_node *old_version_node;
2883 struct cgraph_node *new_version_node;
2884 inline_data id;
2885 tree p, new_fndecl;
2886 unsigned i;
2887 struct ipa_replace_map *replace_info;
2888 basic_block old_entry_block;
2889 tree t_step;
2890
2891 gcc_assert (TREE_CODE (old_decl) == FUNCTION_DECL
2892 && TREE_CODE (new_decl) == FUNCTION_DECL);
2893 DECL_POSSIBLY_INLINED (old_decl) = 1;
2894
2895 old_version_node = cgraph_node (old_decl);
2896 new_version_node = cgraph_node (new_decl);
2897
2898 allocate_struct_function (new_decl);
2899 /* Cfun points to the new allocated function struct at this point. */
2900 cfun->function_end_locus = DECL_SOURCE_LOCATION (new_decl);
2901
2902 DECL_ARTIFICIAL (new_decl) = 1;
2903 DECL_ABSTRACT_ORIGIN (new_decl) = DECL_ORIGIN (old_decl);
2904
2905 /* Generate a new name for the new version. */
2906 DECL_NAME (new_decl) =
2907 create_tmp_var_name (NULL);
2908 /* Create a new SYMBOL_REF rtx for the new name. */
2909 if (DECL_RTL (old_decl) != NULL)
2910 {
2911 SET_DECL_RTL (new_decl, copy_rtx (DECL_RTL (old_decl)));
2912 XEXP (DECL_RTL (new_decl), 0) =
2913 gen_rtx_SYMBOL_REF (GET_MODE (XEXP (DECL_RTL (old_decl), 0)),
2914 IDENTIFIER_POINTER (DECL_NAME (new_decl)));
2915 }
2916
2917 /* Prepare the data structures for the tree copy. */
2918 memset (&id, 0, sizeof (id));
2919
2920 /* The new version. */
2921 id.node = new_version_node;
2922
2923 /* The old version. */
2924 id.current_node = cgraph_node (old_decl);
2925
2926 id.versioning_p = true;
2927 id.decl_map = splay_tree_new (splay_tree_compare_pointers, NULL, NULL);
2928 id.caller = new_decl;
2929 id.callee = old_decl;
2930 id.callee_cfun = DECL_STRUCT_FUNCTION (old_decl);
2931
2932 current_function_decl = new_decl;
2933
2934 /* Copy the function's static chain. */
2935 p = DECL_STRUCT_FUNCTION (old_decl)->static_chain_decl;
2936 if (p)
2937 DECL_STRUCT_FUNCTION (new_decl)->static_chain_decl =
2938 copy_static_chain (DECL_STRUCT_FUNCTION (old_decl)->static_chain_decl,
2939 &id);
2940 /* Copy the function's arguments. */
2941 if (DECL_ARGUMENTS (old_decl) != NULL_TREE)
2942 DECL_ARGUMENTS (new_decl) =
2943 copy_arguments_for_versioning (DECL_ARGUMENTS (old_decl), &id);
2944
2945 /* If there's a tree_map, prepare for substitution. */
2946 if (tree_map)
2947 for (i = 0; i < VARRAY_ACTIVE_SIZE (tree_map); i++)
2948 {
2949 replace_info = VARRAY_GENERIC_PTR (tree_map, i);
2950 if (replace_info->replace_p && !replace_info->ref_p)
2951 insert_decl_map (&id, replace_info->old_tree,
2952 replace_info->new_tree);
2953 else if (replace_info->replace_p && replace_info->ref_p)
2954 id.ipa_info = tree_map;
2955 }
2956
2957 DECL_INITIAL (new_decl) = remap_blocks (DECL_INITIAL (id.callee), &id);
2958
2959 /* Renumber the lexical scoping (non-code) blocks consecutively. */
2960 number_blocks (id.caller);
2961
2962 if (DECL_STRUCT_FUNCTION (old_decl)->unexpanded_var_list != NULL_TREE)
2963 /* Add local vars. */
2964 for (t_step = DECL_STRUCT_FUNCTION (old_decl)->unexpanded_var_list;
2965 t_step; t_step = TREE_CHAIN (t_step))
2966 {
2967 tree var = TREE_VALUE (t_step);
2968 if (TREE_STATIC (var) && !TREE_ASM_WRITTEN (var))
2969 cfun->unexpanded_var_list = tree_cons (NULL_TREE, var,
2970 cfun->unexpanded_var_list);
2971 else
2972 cfun->unexpanded_var_list =
2973 tree_cons (NULL_TREE, remap_decl (var, &id),
2974 cfun->unexpanded_var_list);
2975 }
2976
2977 /* Copy the Function's body. */
2978 old_entry_block = ENTRY_BLOCK_PTR_FOR_FUNCTION
2979 (DECL_STRUCT_FUNCTION (old_decl));
2980 new_fndecl = copy_body (&id,
2981 old_entry_block->count,
2982 old_entry_block->frequency, NULL, NULL);
2983
2984 DECL_SAVED_TREE (new_decl) = DECL_SAVED_TREE (new_fndecl);
2985
2986 DECL_STRUCT_FUNCTION (new_decl)->cfg =
2987 DECL_STRUCT_FUNCTION (new_fndecl)->cfg;
2988 DECL_STRUCT_FUNCTION (new_decl)->eh = DECL_STRUCT_FUNCTION (new_fndecl)->eh;
2989 DECL_STRUCT_FUNCTION (new_decl)->ib_boundaries_block =
2990 DECL_STRUCT_FUNCTION (new_fndecl)->ib_boundaries_block;
2991 DECL_STRUCT_FUNCTION (new_decl)->last_label_uid =
2992 DECL_STRUCT_FUNCTION (new_fndecl)->last_label_uid;
2993
2994 if (DECL_RESULT (old_decl) != NULL_TREE)
2995 {
2996 tree *res_decl = &DECL_RESULT (old_decl);
2997 DECL_RESULT (new_decl) = remap_decl (*res_decl, &id);
2998 lang_hooks.dup_lang_specific_decl (DECL_RESULT (new_decl));
2999 }
3000
3001 current_function_decl = NULL;
3002 /* Renumber the lexical scoping (non-code) blocks consecutively. */
3003 number_blocks (new_decl);
3004
3005 /* Clean up. */
3006 splay_tree_delete (id.decl_map);
3007 fold_cond_expr_cond ();
3008 return;
3009 }
3010
3011 /* Replace an INDIRECT_REF tree of a given DECL tree with a new
3012 given tree.
3013 ID->ipa_info keeps the old tree and the new tree.
3014 TP points to the INDIRECT REF tree. Return true if
3015 the trees were replaced. */
3016 static bool
replace_ref_tree(inline_data * id,tree * tp)3017 replace_ref_tree (inline_data * id, tree * tp)
3018 {
3019 bool replaced = false;
3020 tree new;
3021
3022 if (id->ipa_info && VARRAY_ACTIVE_SIZE (id->ipa_info) > 0)
3023 {
3024 unsigned i;
3025
3026 for (i = 0; i < VARRAY_ACTIVE_SIZE (id->ipa_info); i++)
3027 {
3028 struct ipa_replace_map *replace_info;
3029 replace_info = VARRAY_GENERIC_PTR (id->ipa_info, i);
3030
3031 if (replace_info->replace_p && replace_info->ref_p)
3032 {
3033 tree old_tree = replace_info->old_tree;
3034 tree new_tree = replace_info->new_tree;
3035
3036 if (TREE_CODE (*tp) == INDIRECT_REF
3037 && TREE_OPERAND (*tp, 0) == old_tree)
3038 {
3039 new = copy_node (new_tree);
3040 *tp = new;
3041 replaced = true;
3042 }
3043 }
3044 }
3045 }
3046 return replaced;
3047 }
3048
3049 /* Return true if we are inlining. */
3050 static inline bool
inlining_p(inline_data * id)3051 inlining_p (inline_data * id)
3052 {
3053 return (!id->saving_p && !id->cloning_p && !id->versioning_p);
3054 }
3055
3056 /* Duplicate a type, fields and all. */
3057
3058 tree
build_duplicate_type(tree type)3059 build_duplicate_type (tree type)
3060 {
3061 inline_data id;
3062
3063 memset (&id, 0, sizeof (id));
3064 id.callee = current_function_decl;
3065 id.caller = current_function_decl;
3066 id.callee_cfun = cfun;
3067 id.decl_map = splay_tree_new (splay_tree_compare_pointers, NULL, NULL);
3068
3069 type = remap_type_1 (type, &id);
3070
3071 splay_tree_delete (id.decl_map);
3072
3073 return type;
3074 }
3075