Merge -r 127928:132243 from trunk
[official-gcc.git] / gcc / tree-inline.c
blobb683f20a70ea99b858b7107400d409c4ea4bac39
1 /* Tree inlining.
2 Copyright 2001, 2002, 2003, 2004, 2005, 2006, 2007, 2008
3 Free Software Foundation, Inc.
4 Contributed by Alexandre Oliva <aoliva@redhat.com>
6 This file is part of GCC.
8 GCC is free software; you can redistribute it and/or modify
9 it under the terms of the GNU General Public License as published by
10 the Free Software Foundation; either version 3, or (at your option)
11 any later version.
13 GCC is distributed in the hope that it will be useful,
14 but WITHOUT ANY WARRANTY; without even the implied warranty of
15 MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
16 GNU General Public License for more details.
18 You should have received a copy of the GNU General Public License
19 along with GCC; see the file COPYING3. If not see
20 <http://www.gnu.org/licenses/>. */
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 "langhooks.h"
38 #include "basic-block.h"
39 #include "tree-iterator.h"
40 #include "cgraph.h"
41 #include "intl.h"
42 #include "tree-mudflap.h"
43 #include "tree-flow.h"
44 #include "function.h"
45 #include "ggc.h"
46 #include "tree-flow.h"
47 #include "diagnostic.h"
48 #include "except.h"
49 #include "debug.h"
50 #include "pointer-set.h"
51 #include "ipa-prop.h"
52 #include "value-prof.h"
53 #include "tree-pass.h"
54 #include "target.h"
55 #include "integrate.h"
57 /* I'm not real happy about this, but we need to handle gimple and
58 non-gimple trees. */
59 #include "tree-gimple.h"
61 /* Inlining, Cloning, Versioning, Parallelization
63 Inlining: a function body is duplicated, but the PARM_DECLs are
64 remapped into VAR_DECLs, and non-void RETURN_EXPRs become
65 GIMPLE_MODIFY_STMTs that store to a dedicated returned-value variable.
66 The duplicated eh_region info of the copy will later be appended
67 to the info for the caller; the eh_region info in copied throwing
68 statements and RESX_EXPRs is adjusted accordingly.
70 Cloning: (only in C++) We have one body for a con/de/structor, and
71 multiple function decls, each with a unique parameter list.
72 Duplicate the body, using the given splay tree; some parameters
73 will become constants (like 0 or 1).
75 Versioning: a function body is duplicated and the result is a new
76 function rather than into blocks of an existing function as with
77 inlining. Some parameters will become constants.
79 Parallelization: a region of a function is duplicated resulting in
80 a new function. Variables may be replaced with complex expressions
81 to enable shared variable semantics.
83 All of these will simultaneously lookup any callgraph edges. If
84 we're going to inline the duplicated function body, and the given
85 function has some cloned callgraph nodes (one for each place this
86 function will be inlined) those callgraph edges will be duplicated.
87 If we're cloning the body, those callgraph edges will be
88 updated to point into the new body. (Note that the original
89 callgraph node and edge list will not be altered.)
91 See the CALL_EXPR handling case in copy_body_r (). */
93 /* 0 if we should not perform inlining.
94 1 if we should expand functions calls inline at the tree level.
95 2 if we should consider *all* functions to be inline
96 candidates. */
98 int flag_inline_trees = 0;
100 /* To Do:
102 o In order to make inlining-on-trees work, we pessimized
103 function-local static constants. In particular, they are now
104 always output, even when not addressed. Fix this by treating
105 function-local static constants just like global static
106 constants; the back-end already knows not to output them if they
107 are not needed.
109 o Provide heuristics to clamp inlining of recursive template
110 calls? */
113 /* Weights that estimate_num_insns uses for heuristics in inlining. */
115 eni_weights eni_inlining_weights;
117 /* Weights that estimate_num_insns uses to estimate the size of the
118 produced code. */
120 eni_weights eni_size_weights;
122 /* Weights that estimate_num_insns uses to estimate the time necessary
123 to execute the produced code. */
125 eni_weights eni_time_weights;
127 /* Prototypes. */
129 static tree declare_return_variable (copy_body_data *, tree, tree, tree *);
130 static tree copy_generic_body (copy_body_data *);
131 static bool inlinable_function_p (tree);
132 static void remap_block (tree *, copy_body_data *);
133 static tree remap_decls (tree, copy_body_data *);
134 static void copy_bind_expr (tree *, int *, copy_body_data *);
135 static tree mark_local_for_remap_r (tree *, int *, void *);
136 static void unsave_expr_1 (tree);
137 static tree unsave_r (tree *, int *, void *);
138 static void declare_inline_vars (tree, tree);
139 static void remap_save_expr (tree *, void *, int *);
140 static void add_lexical_block (tree current_block, tree new_block);
141 static tree copy_decl_to_var (tree, copy_body_data *);
142 static tree copy_result_decl_to_var (tree, copy_body_data *);
143 static tree copy_decl_no_change (tree, copy_body_data *);
144 static tree copy_decl_maybe_to_var (tree, copy_body_data *);
146 /* Insert a tree->tree mapping for ID. Despite the name suggests
147 that the trees should be variables, it is used for more than that. */
149 void
150 insert_decl_map (copy_body_data *id, tree key, tree value)
152 *pointer_map_insert (id->decl_map, key) = value;
154 /* Always insert an identity map as well. If we see this same new
155 node again, we won't want to duplicate it a second time. */
156 if (key != value)
157 *pointer_map_insert (id->decl_map, value) = value;
160 /* Construct new SSA name for old NAME. ID is the inline context. */
162 static tree
163 remap_ssa_name (tree name, copy_body_data *id)
165 tree new;
166 tree *n;
168 gcc_assert (TREE_CODE (name) == SSA_NAME);
170 n = (tree *) pointer_map_contains (id->decl_map, name);
171 if (n)
172 return *n;
174 /* Do not set DEF_STMT yet as statement is not copied yet. We do that
175 in copy_bb. */
176 new = remap_decl (SSA_NAME_VAR (name), id);
177 /* We might've substituted constant or another SSA_NAME for
178 the variable.
180 Replace the SSA name representing RESULT_DECL by variable during
181 inlining: this saves us from need to introduce PHI node in a case
182 return value is just partly initialized. */
183 if ((TREE_CODE (new) == VAR_DECL || TREE_CODE (new) == PARM_DECL)
184 && (TREE_CODE (SSA_NAME_VAR (name)) != RESULT_DECL
185 || !id->transform_return_to_modify))
187 new = make_ssa_name (new, NULL);
188 insert_decl_map (id, name, new);
189 SSA_NAME_OCCURS_IN_ABNORMAL_PHI (new)
190 = SSA_NAME_OCCURS_IN_ABNORMAL_PHI (name);
191 TREE_TYPE (new) = TREE_TYPE (SSA_NAME_VAR (new));
192 if (IS_EMPTY_STMT (SSA_NAME_DEF_STMT (name)))
194 /* By inlining function having uninitialized variable, we might
195 extend the lifetime (variable might get reused). This cause
196 ICE in the case we end up extending lifetime of SSA name across
197 abnormal edge, but also increase register presure.
199 We simply initialize all uninitialized vars by 0 except for case
200 we are inlining to very first BB. We can avoid this for all
201 BBs that are not withing strongly connected regions of the CFG,
202 but this is bit expensive to test.
204 if (id->entry_bb && is_gimple_reg (SSA_NAME_VAR (name))
205 && TREE_CODE (SSA_NAME_VAR (name)) != PARM_DECL
206 && (id->entry_bb != EDGE_SUCC (ENTRY_BLOCK_PTR, 0)->dest
207 || EDGE_COUNT (id->entry_bb->preds) != 1))
209 block_stmt_iterator bsi = bsi_last (id->entry_bb);
210 tree init_stmt
211 = build_gimple_modify_stmt (new,
212 fold_convert (TREE_TYPE (new),
213 integer_zero_node));
214 bsi_insert_after (&bsi, init_stmt, BSI_NEW_STMT);
215 SSA_NAME_DEF_STMT (new) = init_stmt;
216 SSA_NAME_IS_DEFAULT_DEF (new) = 0;
218 else
220 SSA_NAME_DEF_STMT (new) = build_empty_stmt ();
221 if (gimple_default_def (id->src_cfun, SSA_NAME_VAR (name)) == name)
222 set_default_def (SSA_NAME_VAR (new), new);
226 else
227 insert_decl_map (id, name, new);
228 return new;
231 /* Remap DECL during the copying of the BLOCK tree for the function. */
233 tree
234 remap_decl (tree decl, copy_body_data *id)
236 tree *n;
237 tree fn;
239 /* We only remap local variables in the current function. */
240 fn = id->src_fn;
242 /* See if we have remapped this declaration. */
244 n = (tree *) pointer_map_contains (id->decl_map, decl);
246 /* If we didn't already have an equivalent for this declaration,
247 create one now. */
248 if (!n)
250 /* Make a copy of the variable or label. */
251 tree t = id->copy_decl (decl, id);
253 /* Remember it, so that if we encounter this local entity again
254 we can reuse this copy. Do this early because remap_type may
255 need this decl for TYPE_STUB_DECL. */
256 insert_decl_map (id, decl, t);
258 if (!DECL_P (t))
259 return t;
261 /* Remap types, if necessary. */
262 TREE_TYPE (t) = remap_type (TREE_TYPE (t), id);
263 if (TREE_CODE (t) == TYPE_DECL)
264 DECL_ORIGINAL_TYPE (t) = remap_type (DECL_ORIGINAL_TYPE (t), id);
266 /* Remap sizes as necessary. */
267 walk_tree (&DECL_SIZE (t), copy_body_r, id, NULL);
268 walk_tree (&DECL_SIZE_UNIT (t), copy_body_r, id, NULL);
270 /* If fields, do likewise for offset and qualifier. */
271 if (TREE_CODE (t) == FIELD_DECL)
273 walk_tree (&DECL_FIELD_OFFSET (t), copy_body_r, id, NULL);
274 if (TREE_CODE (DECL_CONTEXT (t)) == QUAL_UNION_TYPE)
275 walk_tree (&DECL_QUALIFIER (t), copy_body_r, id, NULL);
278 if (cfun && gimple_in_ssa_p (cfun)
279 && (TREE_CODE (t) == VAR_DECL
280 || TREE_CODE (t) == RESULT_DECL || TREE_CODE (t) == PARM_DECL))
282 tree def = gimple_default_def (id->src_cfun, decl);
283 get_var_ann (t);
284 if (TREE_CODE (decl) != PARM_DECL && def)
286 tree map = remap_ssa_name (def, id);
287 /* Watch out RESULT_DECLs whose SSA names map directly
288 to them. */
289 if (TREE_CODE (map) == SSA_NAME
290 && IS_EMPTY_STMT (SSA_NAME_DEF_STMT (map)))
291 set_default_def (t, map);
293 add_referenced_var (t);
295 return t;
298 return unshare_expr (*n);
301 static tree
302 remap_type_1 (tree type, copy_body_data *id)
304 tree new, t;
306 /* We do need a copy. build and register it now. If this is a pointer or
307 reference type, remap the designated type and make a new pointer or
308 reference type. */
309 if (TREE_CODE (type) == POINTER_TYPE)
311 new = build_pointer_type_for_mode (remap_type (TREE_TYPE (type), id),
312 TYPE_MODE (type),
313 TYPE_REF_CAN_ALIAS_ALL (type));
314 insert_decl_map (id, type, new);
315 return new;
317 else if (TREE_CODE (type) == REFERENCE_TYPE)
319 new = build_reference_type_for_mode (remap_type (TREE_TYPE (type), id),
320 TYPE_MODE (type),
321 TYPE_REF_CAN_ALIAS_ALL (type));
322 insert_decl_map (id, type, new);
323 return new;
325 else
326 new = copy_node (type);
328 insert_decl_map (id, type, new);
330 /* This is a new type, not a copy of an old type. Need to reassociate
331 variants. We can handle everything except the main variant lazily. */
332 t = TYPE_MAIN_VARIANT (type);
333 if (type != t)
335 t = remap_type (t, id);
336 TYPE_MAIN_VARIANT (new) = t;
337 TYPE_NEXT_VARIANT (new) = TYPE_NEXT_VARIANT (t);
338 TYPE_NEXT_VARIANT (t) = new;
340 else
342 TYPE_MAIN_VARIANT (new) = new;
343 TYPE_NEXT_VARIANT (new) = NULL;
346 if (TYPE_STUB_DECL (type))
347 TYPE_STUB_DECL (new) = remap_decl (TYPE_STUB_DECL (type), id);
349 /* Lazily create pointer and reference types. */
350 TYPE_POINTER_TO (new) = NULL;
351 TYPE_REFERENCE_TO (new) = NULL;
353 switch (TREE_CODE (new))
355 case INTEGER_TYPE:
356 case REAL_TYPE:
357 case FIXED_POINT_TYPE:
358 case ENUMERAL_TYPE:
359 case BOOLEAN_TYPE:
360 t = TYPE_MIN_VALUE (new);
361 if (t && TREE_CODE (t) != INTEGER_CST)
362 walk_tree (&TYPE_MIN_VALUE (new), copy_body_r, id, NULL);
364 t = TYPE_MAX_VALUE (new);
365 if (t && TREE_CODE (t) != INTEGER_CST)
366 walk_tree (&TYPE_MAX_VALUE (new), copy_body_r, id, NULL);
367 return new;
369 case FUNCTION_TYPE:
370 TREE_TYPE (new) = remap_type (TREE_TYPE (new), id);
371 walk_tree (&TYPE_ARG_TYPES (new), copy_body_r, id, NULL);
372 return new;
374 case ARRAY_TYPE:
375 TREE_TYPE (new) = remap_type (TREE_TYPE (new), id);
376 TYPE_DOMAIN (new) = remap_type (TYPE_DOMAIN (new), id);
377 break;
379 case RECORD_TYPE:
380 case UNION_TYPE:
381 case QUAL_UNION_TYPE:
383 tree f, nf = NULL;
385 for (f = TYPE_FIELDS (new); f ; f = TREE_CHAIN (f))
387 t = remap_decl (f, id);
388 DECL_CONTEXT (t) = new;
389 TREE_CHAIN (t) = nf;
390 nf = t;
392 TYPE_FIELDS (new) = nreverse (nf);
394 break;
396 case OFFSET_TYPE:
397 default:
398 /* Shouldn't have been thought variable sized. */
399 gcc_unreachable ();
402 walk_tree (&TYPE_SIZE (new), copy_body_r, id, NULL);
403 walk_tree (&TYPE_SIZE_UNIT (new), copy_body_r, id, NULL);
405 return new;
408 tree
409 remap_type (tree type, copy_body_data *id)
411 tree *node;
412 tree tmp;
414 if (type == NULL)
415 return type;
417 /* See if we have remapped this type. */
418 node = (tree *) pointer_map_contains (id->decl_map, type);
419 if (node)
420 return *node;
422 /* The type only needs remapping if it's variably modified. */
423 if (! variably_modified_type_p (type, id->src_fn))
425 insert_decl_map (id, type, type);
426 return type;
429 id->remapping_type_depth++;
430 tmp = remap_type_1 (type, id);
431 id->remapping_type_depth--;
433 return tmp;
436 static tree
437 remap_decls (tree decls, copy_body_data *id)
439 tree old_var;
440 tree new_decls = NULL_TREE;
442 /* Remap its variables. */
443 for (old_var = decls; old_var; old_var = TREE_CHAIN (old_var))
445 tree new_var;
447 /* We can not chain the local static declarations into the unexpanded_var_list
448 as we can't duplicate them or break one decl rule. Go ahead and link
449 them into unexpanded_var_list. */
450 if (!auto_var_in_fn_p (old_var, id->src_fn)
451 && !DECL_EXTERNAL (old_var))
453 cfun->unexpanded_var_list = tree_cons (NULL_TREE, old_var,
454 cfun->unexpanded_var_list);
455 continue;
458 /* Remap the variable. */
459 new_var = remap_decl (old_var, id);
461 /* If we didn't remap this variable, so we can't mess with its
462 TREE_CHAIN. If we remapped this variable to the return slot, it's
463 already declared somewhere else, so don't declare it here. */
464 if (!new_var || new_var == id->retvar)
466 else
468 gcc_assert (DECL_P (new_var));
469 TREE_CHAIN (new_var) = new_decls;
470 new_decls = new_var;
474 return nreverse (new_decls);
477 /* Copy the BLOCK to contain remapped versions of the variables
478 therein. And hook the new block into the block-tree. */
480 static void
481 remap_block (tree *block, copy_body_data *id)
483 tree old_block;
484 tree new_block;
485 tree fn;
487 /* Make the new block. */
488 old_block = *block;
489 new_block = make_node (BLOCK);
490 TREE_USED (new_block) = TREE_USED (old_block);
491 BLOCK_ABSTRACT_ORIGIN (new_block) = old_block;
492 BLOCK_SOURCE_LOCATION (new_block) = BLOCK_SOURCE_LOCATION (old_block);
493 *block = new_block;
495 /* Remap its variables. */
496 BLOCK_VARS (new_block) = remap_decls (BLOCK_VARS (old_block), id);
498 fn = id->dst_fn;
500 if (id->transform_lang_insert_block)
501 lang_hooks.decls.insert_block (new_block);
503 /* Remember the remapped block. */
504 insert_decl_map (id, old_block, new_block);
507 /* Copy the whole block tree and root it in id->block. */
508 static tree
509 remap_blocks (tree block, copy_body_data *id)
511 tree t;
512 tree new = block;
514 if (!block)
515 return NULL;
517 remap_block (&new, id);
518 gcc_assert (new != block);
519 for (t = BLOCK_SUBBLOCKS (block); t ; t = BLOCK_CHAIN (t))
520 add_lexical_block (new, remap_blocks (t, id));
521 return new;
524 static void
525 copy_statement_list (tree *tp)
527 tree_stmt_iterator oi, ni;
528 tree new;
530 new = alloc_stmt_list ();
531 ni = tsi_start (new);
532 oi = tsi_start (*tp);
533 *tp = new;
535 for (; !tsi_end_p (oi); tsi_next (&oi))
536 tsi_link_after (&ni, tsi_stmt (oi), TSI_NEW_STMT);
539 static void
540 copy_bind_expr (tree *tp, int *walk_subtrees, copy_body_data *id)
542 tree block = BIND_EXPR_BLOCK (*tp);
543 /* Copy (and replace) the statement. */
544 copy_tree_r (tp, walk_subtrees, NULL);
545 if (block)
547 remap_block (&block, id);
548 BIND_EXPR_BLOCK (*tp) = block;
551 if (BIND_EXPR_VARS (*tp))
552 /* This will remap a lot of the same decls again, but this should be
553 harmless. */
554 BIND_EXPR_VARS (*tp) = remap_decls (BIND_EXPR_VARS (*tp), id);
557 /* Called from copy_body_id via walk_tree. DATA is really an
558 `copy_body_data *'. */
560 tree
561 copy_body_r (tree *tp, int *walk_subtrees, void *data)
563 copy_body_data *id = (copy_body_data *) data;
564 tree fn = id->src_fn;
565 tree new_block;
567 /* Begin by recognizing trees that we'll completely rewrite for the
568 inlining context. Our output for these trees is completely
569 different from out input (e.g. RETURN_EXPR is deleted, and morphs
570 into an edge). Further down, we'll handle trees that get
571 duplicated and/or tweaked. */
573 /* When requested, RETURN_EXPRs should be transformed to just the
574 contained GIMPLE_MODIFY_STMT. The branch semantics of the return will
575 be handled elsewhere by manipulating the CFG rather than a statement. */
576 if (TREE_CODE (*tp) == RETURN_EXPR && id->transform_return_to_modify)
578 tree assignment = TREE_OPERAND (*tp, 0);
580 /* If we're returning something, just turn that into an
581 assignment into the equivalent of the original RESULT_DECL.
582 If the "assignment" is just the result decl, the result
583 decl has already been set (e.g. a recent "foo (&result_decl,
584 ...)"); just toss the entire RETURN_EXPR. */
585 if (assignment && TREE_CODE (assignment) == GIMPLE_MODIFY_STMT)
587 /* Replace the RETURN_EXPR with (a copy of) the
588 GIMPLE_MODIFY_STMT hanging underneath. */
589 *tp = copy_node (assignment);
591 else /* Else the RETURN_EXPR returns no value. */
593 *tp = NULL;
594 return (tree) (void *)1;
597 else if (TREE_CODE (*tp) == SSA_NAME)
599 *tp = remap_ssa_name (*tp, id);
600 *walk_subtrees = 0;
601 return NULL;
604 /* Local variables and labels need to be replaced by equivalent
605 variables. We don't want to copy static variables; there's only
606 one of those, no matter how many times we inline the containing
607 function. Similarly for globals from an outer function. */
608 else if (auto_var_in_fn_p (*tp, fn))
610 tree new_decl;
612 /* Remap the declaration. */
613 new_decl = remap_decl (*tp, id);
614 gcc_assert (new_decl);
615 /* Replace this variable with the copy. */
616 STRIP_TYPE_NOPS (new_decl);
617 *tp = new_decl;
618 *walk_subtrees = 0;
620 else if (TREE_CODE (*tp) == STATEMENT_LIST)
621 copy_statement_list (tp);
622 else if (TREE_CODE (*tp) == SAVE_EXPR)
623 remap_save_expr (tp, id->decl_map, walk_subtrees);
624 else if (TREE_CODE (*tp) == LABEL_DECL
625 && (! DECL_CONTEXT (*tp)
626 || decl_function_context (*tp) == id->src_fn))
627 /* These may need to be remapped for EH handling. */
628 *tp = remap_decl (*tp, id);
629 else if (TREE_CODE (*tp) == BIND_EXPR)
630 copy_bind_expr (tp, walk_subtrees, id);
631 /* Types may need remapping as well. */
632 else if (TYPE_P (*tp))
633 *tp = remap_type (*tp, id);
635 /* If this is a constant, we have to copy the node iff the type will be
636 remapped. copy_tree_r will not copy a constant. */
637 else if (CONSTANT_CLASS_P (*tp))
639 tree new_type = remap_type (TREE_TYPE (*tp), id);
641 if (new_type == TREE_TYPE (*tp))
642 *walk_subtrees = 0;
644 else if (TREE_CODE (*tp) == INTEGER_CST)
645 *tp = build_int_cst_wide (new_type, TREE_INT_CST_LOW (*tp),
646 TREE_INT_CST_HIGH (*tp));
647 else
649 *tp = copy_node (*tp);
650 TREE_TYPE (*tp) = new_type;
654 /* Otherwise, just copy the node. Note that copy_tree_r already
655 knows not to copy VAR_DECLs, etc., so this is safe. */
656 else
658 /* Here we handle trees that are not completely rewritten.
659 First we detect some inlining-induced bogosities for
660 discarding. */
661 if (TREE_CODE (*tp) == GIMPLE_MODIFY_STMT
662 && GIMPLE_STMT_OPERAND (*tp, 0) == GIMPLE_STMT_OPERAND (*tp, 1)
663 && (auto_var_in_fn_p (GIMPLE_STMT_OPERAND (*tp, 0), fn)))
665 /* Some assignments VAR = VAR; don't generate any rtl code
666 and thus don't count as variable modification. Avoid
667 keeping bogosities like 0 = 0. */
668 tree decl = GIMPLE_STMT_OPERAND (*tp, 0), value;
669 tree *n;
671 n = (tree *) pointer_map_contains (id->decl_map, decl);
672 if (n)
674 value = *n;
675 STRIP_TYPE_NOPS (value);
676 if (TREE_CONSTANT (value) || TREE_READONLY_DECL_P (value))
678 *tp = build_empty_stmt ();
679 return copy_body_r (tp, walk_subtrees, data);
683 else if (TREE_CODE (*tp) == INDIRECT_REF)
685 /* Get rid of *& from inline substitutions that can happen when a
686 pointer argument is an ADDR_EXPR. */
687 tree decl = TREE_OPERAND (*tp, 0);
688 tree *n;
690 n = (tree *) pointer_map_contains (id->decl_map, decl);
691 if (n)
693 tree new;
694 tree old;
695 /* If we happen to get an ADDR_EXPR in n->value, strip
696 it manually here as we'll eventually get ADDR_EXPRs
697 which lie about their types pointed to. In this case
698 build_fold_indirect_ref wouldn't strip the INDIRECT_REF,
699 but we absolutely rely on that. As fold_indirect_ref
700 does other useful transformations, try that first, though. */
701 tree type = TREE_TYPE (TREE_TYPE (*n));
702 new = unshare_expr (*n);
703 old = *tp;
704 *tp = gimple_fold_indirect_ref (new);
705 if (! *tp)
707 if (TREE_CODE (new) == ADDR_EXPR)
709 *tp = fold_indirect_ref_1 (type, new);
710 /* ??? We should either assert here or build
711 a VIEW_CONVERT_EXPR instead of blindly leaking
712 incompatible types to our IL. */
713 if (! *tp)
714 *tp = TREE_OPERAND (new, 0);
716 else
718 *tp = build1 (INDIRECT_REF, type, new);
719 TREE_THIS_VOLATILE (*tp) = TREE_THIS_VOLATILE (old);
722 *walk_subtrees = 0;
723 return NULL;
727 /* Here is the "usual case". Copy this tree node, and then
728 tweak some special cases. */
729 copy_tree_r (tp, walk_subtrees, NULL);
731 /* Global variables we haven't seen yet needs to go into referenced
732 vars. If not referenced from types only. */
733 if (gimple_in_ssa_p (cfun) && TREE_CODE (*tp) == VAR_DECL
734 && id->remapping_type_depth == 0)
735 add_referenced_var (*tp);
737 /* If EXPR has block defined, map it to newly constructed block.
738 When inlining we want EXPRs without block appear in the block
739 of function call. */
740 if (EXPR_P (*tp) || GIMPLE_STMT_P (*tp))
742 new_block = id->block;
743 if (TREE_BLOCK (*tp))
745 tree *n;
746 n = (tree *) pointer_map_contains (id->decl_map,
747 TREE_BLOCK (*tp));
748 gcc_assert (n);
749 new_block = *n;
751 TREE_BLOCK (*tp) = new_block;
754 if (TREE_CODE (*tp) == RESX_EXPR && id->eh_region_offset)
755 TREE_OPERAND (*tp, 0) =
756 build_int_cst
757 (NULL_TREE,
758 id->eh_region_offset + TREE_INT_CST_LOW (TREE_OPERAND (*tp, 0)));
760 if (!GIMPLE_TUPLE_P (*tp) && TREE_CODE (*tp) != OMP_CLAUSE)
761 TREE_TYPE (*tp) = remap_type (TREE_TYPE (*tp), id);
763 /* The copied TARGET_EXPR has never been expanded, even if the
764 original node was expanded already. */
765 if (TREE_CODE (*tp) == TARGET_EXPR && TREE_OPERAND (*tp, 3))
767 TREE_OPERAND (*tp, 1) = TREE_OPERAND (*tp, 3);
768 TREE_OPERAND (*tp, 3) = NULL_TREE;
771 /* Variable substitution need not be simple. In particular, the
772 INDIRECT_REF substitution above. Make sure that TREE_CONSTANT
773 and friends are up-to-date. */
774 else if (TREE_CODE (*tp) == ADDR_EXPR)
776 int invariant = TREE_INVARIANT (*tp);
777 walk_tree (&TREE_OPERAND (*tp, 0), copy_body_r, id, NULL);
778 /* Handle the case where we substituted an INDIRECT_REF
779 into the operand of the ADDR_EXPR. */
780 if (TREE_CODE (TREE_OPERAND (*tp, 0)) == INDIRECT_REF)
781 *tp = TREE_OPERAND (TREE_OPERAND (*tp, 0), 0);
782 else
783 recompute_tree_invariant_for_addr_expr (*tp);
784 /* If this used to be invariant, but is not any longer,
785 then regimplification is probably needed. */
786 if (invariant && !TREE_INVARIANT (*tp))
787 id->regimplify = true;
788 *walk_subtrees = 0;
792 /* Keep iterating. */
793 return NULL_TREE;
796 /* Copy basic block, scale profile accordingly. Edges will be taken care of
797 later */
799 static basic_block
800 copy_bb (copy_body_data *id, basic_block bb, int frequency_scale, int count_scale)
802 block_stmt_iterator bsi, copy_bsi;
803 basic_block copy_basic_block;
805 /* create_basic_block() will append every new block to
806 basic_block_info automatically. */
807 copy_basic_block = create_basic_block (NULL, (void *) 0,
808 (basic_block) bb->prev_bb->aux);
809 copy_basic_block->count = bb->count * count_scale / REG_BR_PROB_BASE;
811 /* We are going to rebuild frequencies from scratch. These values have just
812 small importance to drive canonicalize_loop_headers. */
813 copy_basic_block->frequency = ((gcov_type)bb->frequency
814 * frequency_scale / REG_BR_PROB_BASE);
815 if (copy_basic_block->frequency > BB_FREQ_MAX)
816 copy_basic_block->frequency = BB_FREQ_MAX;
817 copy_bsi = bsi_start (copy_basic_block);
819 for (bsi = bsi_start (bb);
820 !bsi_end_p (bsi); bsi_next (&bsi))
822 tree stmt = bsi_stmt (bsi);
823 tree orig_stmt = stmt;
825 id->regimplify = false;
826 walk_tree (&stmt, copy_body_r, id, NULL);
828 /* RETURN_EXPR might be removed,
829 this is signalled by making stmt pointer NULL. */
830 if (stmt)
832 tree call, decl;
834 gimple_duplicate_stmt_histograms (cfun, stmt, id->src_cfun, orig_stmt);
836 /* With return slot optimization we can end up with
837 non-gimple (foo *)&this->m, fix that here. */
838 if ((TREE_CODE (stmt) == GIMPLE_MODIFY_STMT
839 && TREE_CODE (GIMPLE_STMT_OPERAND (stmt, 1)) == NOP_EXPR
840 && !is_gimple_val (TREE_OPERAND (GIMPLE_STMT_OPERAND (stmt, 1), 0)))
841 || id->regimplify)
842 gimplify_stmt (&stmt);
844 bsi_insert_after (&copy_bsi, stmt, BSI_NEW_STMT);
846 /* Process new statement. gimplify_stmt possibly turned statement
847 into multiple statements, we need to process all of them. */
848 while (!bsi_end_p (copy_bsi))
850 tree *stmtp = bsi_stmt_ptr (copy_bsi);
851 tree stmt = *stmtp;
852 call = get_call_expr_in (stmt);
854 if (call && CALL_EXPR_VA_ARG_PACK (call) && id->call_expr)
856 /* __builtin_va_arg_pack () should be replaced by
857 all arguments corresponding to ... in the caller. */
858 tree p, *argarray, new_call, *call_ptr;
859 int nargs = call_expr_nargs (id->call_expr);
861 for (p = DECL_ARGUMENTS (id->src_fn); p; p = TREE_CHAIN (p))
862 nargs--;
864 argarray = (tree *) alloca ((nargs + call_expr_nargs (call))
865 * sizeof (tree));
867 memcpy (argarray, CALL_EXPR_ARGP (call),
868 call_expr_nargs (call) * sizeof (*argarray));
869 memcpy (argarray + call_expr_nargs (call),
870 CALL_EXPR_ARGP (id->call_expr)
871 + (call_expr_nargs (id->call_expr) - nargs),
872 nargs * sizeof (*argarray));
874 new_call = build_call_array (TREE_TYPE (call),
875 CALL_EXPR_FN (call),
876 nargs + call_expr_nargs (call),
877 argarray);
878 /* Copy all CALL_EXPR flags, locus and block, except
879 CALL_EXPR_VA_ARG_PACK flag. */
880 CALL_EXPR_STATIC_CHAIN (new_call)
881 = CALL_EXPR_STATIC_CHAIN (call);
882 CALL_EXPR_TAILCALL (new_call) = CALL_EXPR_TAILCALL (call);
883 CALL_EXPR_RETURN_SLOT_OPT (new_call)
884 = CALL_EXPR_RETURN_SLOT_OPT (call);
885 CALL_FROM_THUNK_P (new_call) = CALL_FROM_THUNK_P (call);
886 CALL_CANNOT_INLINE_P (new_call)
887 = CALL_CANNOT_INLINE_P (call);
888 TREE_NOTHROW (new_call) = TREE_NOTHROW (call);
889 SET_EXPR_LOCUS (new_call, EXPR_LOCUS (call));
890 TREE_BLOCK (new_call) = TREE_BLOCK (call);
892 call_ptr = stmtp;
893 if (TREE_CODE (*call_ptr) == GIMPLE_MODIFY_STMT)
894 call_ptr = &GIMPLE_STMT_OPERAND (*call_ptr, 1);
895 if (TREE_CODE (*call_ptr) == WITH_SIZE_EXPR)
896 call_ptr = &TREE_OPERAND (*call_ptr, 0);
897 gcc_assert (*call_ptr == call);
898 if (call_ptr == stmtp)
900 bsi_replace (&copy_bsi, new_call, true);
901 stmtp = bsi_stmt_ptr (copy_bsi);
902 stmt = *stmtp;
904 else
906 *call_ptr = new_call;
907 stmt = *stmtp;
908 update_stmt (stmt);
911 else if (call
912 && id->call_expr
913 && (decl = get_callee_fndecl (call))
914 && DECL_BUILT_IN_CLASS (decl) == BUILT_IN_NORMAL
915 && DECL_FUNCTION_CODE (decl)
916 == BUILT_IN_VA_ARG_PACK_LEN)
918 /* __builtin_va_arg_pack_len () should be replaced by
919 the number of anonymous arguments. */
920 int nargs = call_expr_nargs (id->call_expr);
921 tree count, *call_ptr, p;
923 for (p = DECL_ARGUMENTS (id->src_fn); p; p = TREE_CHAIN (p))
924 nargs--;
926 count = build_int_cst (integer_type_node, nargs);
927 call_ptr = stmtp;
928 if (TREE_CODE (*call_ptr) == GIMPLE_MODIFY_STMT)
929 call_ptr = &GIMPLE_STMT_OPERAND (*call_ptr, 1);
930 if (TREE_CODE (*call_ptr) == WITH_SIZE_EXPR)
931 call_ptr = &TREE_OPERAND (*call_ptr, 0);
932 gcc_assert (*call_ptr == call && call_ptr != stmtp);
933 *call_ptr = count;
934 stmt = *stmtp;
935 update_stmt (stmt);
936 call = NULL_TREE;
939 /* Statements produced by inlining can be unfolded, especially
940 when we constant propagated some operands. We can't fold
941 them right now for two reasons:
942 1) folding require SSA_NAME_DEF_STMTs to be correct
943 2) we can't change function calls to builtins.
944 So we just mark statement for later folding. We mark
945 all new statements, instead just statements that has changed
946 by some nontrivial substitution so even statements made
947 foldable indirectly are updated. If this turns out to be
948 expensive, copy_body can be told to watch for nontrivial
949 changes. */
950 if (id->statements_to_fold)
951 pointer_set_insert (id->statements_to_fold, stmt);
952 /* We're duplicating a CALL_EXPR. Find any corresponding
953 callgraph edges and update or duplicate them. */
954 if (call && (decl = get_callee_fndecl (call)))
956 struct cgraph_node *node;
957 struct cgraph_edge *edge;
959 switch (id->transform_call_graph_edges)
961 case CB_CGE_DUPLICATE:
962 edge = cgraph_edge (id->src_node, orig_stmt);
963 if (edge)
964 cgraph_clone_edge (edge, id->dst_node, stmt,
965 REG_BR_PROB_BASE, 1, edge->frequency, true);
966 break;
968 case CB_CGE_MOVE_CLONES:
969 for (node = id->dst_node->next_clone;
970 node;
971 node = node->next_clone)
973 edge = cgraph_edge (node, orig_stmt);
974 gcc_assert (edge);
975 cgraph_set_call_stmt (edge, stmt);
977 /* FALLTHRU */
979 case CB_CGE_MOVE:
980 edge = cgraph_edge (id->dst_node, orig_stmt);
981 if (edge)
982 cgraph_set_call_stmt (edge, stmt);
983 break;
985 default:
986 gcc_unreachable ();
989 /* If you think we can abort here, you are wrong.
990 There is no region 0 in tree land. */
991 gcc_assert (lookup_stmt_eh_region_fn (id->src_cfun, orig_stmt)
992 != 0);
994 if (tree_could_throw_p (stmt)
995 /* When we are cloning for inlining, we are supposed to
996 construct a clone that calls precisely the same functions
997 as original. However IPA optimizers might've proved
998 earlier some function calls as non-trapping that might
999 render some basic blocks dead that might become
1000 unreachable.
1002 We can't update SSA with unreachable blocks in CFG and thus
1003 we prevent the scenario by preserving even the "dead" eh
1004 edges until the point they are later removed by
1005 fixup_cfg pass. */
1006 || (id->transform_call_graph_edges == CB_CGE_MOVE_CLONES
1007 && lookup_stmt_eh_region_fn (id->src_cfun, orig_stmt) > 0))
1009 int region = lookup_stmt_eh_region_fn (id->src_cfun, orig_stmt);
1010 /* Add an entry for the copied tree in the EH hashtable.
1011 When cloning or versioning, use the hashtable in
1012 cfun, and just copy the EH number. When inlining, use the
1013 hashtable in the caller, and adjust the region number. */
1014 if (region > 0)
1015 add_stmt_to_eh_region (stmt, region + id->eh_region_offset);
1017 /* If this tree doesn't have a region associated with it,
1018 and there is a "current region,"
1019 then associate this tree with the current region
1020 and add edges associated with this region. */
1021 if ((lookup_stmt_eh_region_fn (id->src_cfun,
1022 orig_stmt) <= 0
1023 && id->eh_region > 0)
1024 && tree_could_throw_p (stmt))
1025 add_stmt_to_eh_region (stmt, id->eh_region);
1027 if (gimple_in_ssa_p (cfun))
1029 ssa_op_iter i;
1030 tree def;
1032 find_new_referenced_vars (bsi_stmt_ptr (copy_bsi));
1033 FOR_EACH_SSA_TREE_OPERAND (def, stmt, i, SSA_OP_DEF)
1034 if (TREE_CODE (def) == SSA_NAME)
1035 SSA_NAME_DEF_STMT (def) = stmt;
1037 bsi_next (&copy_bsi);
1039 copy_bsi = bsi_last (copy_basic_block);
1042 return copy_basic_block;
1045 /* Inserting Single Entry Multiple Exit region in SSA form into code in SSA
1046 form is quite easy, since dominator relationship for old basic blocks does
1047 not change.
1049 There is however exception where inlining might change dominator relation
1050 across EH edges from basic block within inlined functions destinating
1051 to landing pads in function we inline into.
1053 The function fills in PHI_RESULTs of such PHI nodes if they refer
1054 to gimple regs. Otherwise, the function mark PHI_RESULT of such
1055 PHI nodes for renaming. For non-gimple regs, renaming is safe: the
1056 EH edges are abnormal and SSA_NAME_OCCURS_IN_ABNORMAL_PHI must be
1057 set, and this means that there will be no overlapping live ranges
1058 for the underlying symbol.
1060 This might change in future if we allow redirecting of EH edges and
1061 we might want to change way build CFG pre-inlining to include
1062 all the possible edges then. */
1063 static void
1064 update_ssa_across_abnormal_edges (basic_block bb, basic_block ret_bb,
1065 bool can_throw, bool nonlocal_goto)
1067 edge e;
1068 edge_iterator ei;
1070 FOR_EACH_EDGE (e, ei, bb->succs)
1071 if (!e->dest->aux
1072 || ((basic_block)e->dest->aux)->index == ENTRY_BLOCK)
1074 tree phi;
1076 gcc_assert (e->flags & EDGE_ABNORMAL);
1077 if (!nonlocal_goto)
1078 gcc_assert (e->flags & EDGE_EH);
1079 if (!can_throw)
1080 gcc_assert (!(e->flags & EDGE_EH));
1081 for (phi = phi_nodes (e->dest); phi; phi = PHI_CHAIN (phi))
1083 edge re;
1085 /* There shouldn't be any PHI nodes in the ENTRY_BLOCK. */
1086 gcc_assert (!e->dest->aux);
1088 gcc_assert (SSA_NAME_OCCURS_IN_ABNORMAL_PHI
1089 (PHI_RESULT (phi)));
1091 if (!is_gimple_reg (PHI_RESULT (phi)))
1093 mark_sym_for_renaming
1094 (SSA_NAME_VAR (PHI_RESULT (phi)));
1095 continue;
1098 re = find_edge (ret_bb, e->dest);
1099 gcc_assert (re);
1100 gcc_assert ((re->flags & (EDGE_EH | EDGE_ABNORMAL))
1101 == (e->flags & (EDGE_EH | EDGE_ABNORMAL)));
1103 SET_USE (PHI_ARG_DEF_PTR_FROM_EDGE (phi, e),
1104 USE_FROM_PTR (PHI_ARG_DEF_PTR_FROM_EDGE (phi, re)));
1109 /* Copy edges from BB into its copy constructed earlier, scale profile
1110 accordingly. Edges will be taken care of later. Assume aux
1111 pointers to point to the copies of each BB. */
1112 static void
1113 copy_edges_for_bb (basic_block bb, int count_scale, basic_block ret_bb)
1115 basic_block new_bb = (basic_block) bb->aux;
1116 edge_iterator ei;
1117 edge old_edge;
1118 block_stmt_iterator bsi;
1119 int flags;
1121 /* Use the indices from the original blocks to create edges for the
1122 new ones. */
1123 FOR_EACH_EDGE (old_edge, ei, bb->succs)
1124 if (!(old_edge->flags & EDGE_EH))
1126 edge new;
1128 flags = old_edge->flags;
1130 /* Return edges do get a FALLTHRU flag when the get inlined. */
1131 if (old_edge->dest->index == EXIT_BLOCK && !old_edge->flags
1132 && old_edge->dest->aux != EXIT_BLOCK_PTR)
1133 flags |= EDGE_FALLTHRU;
1134 new = make_edge (new_bb, (basic_block) old_edge->dest->aux, flags);
1135 new->count = old_edge->count * count_scale / REG_BR_PROB_BASE;
1136 new->probability = old_edge->probability;
1139 if (bb->index == ENTRY_BLOCK || bb->index == EXIT_BLOCK)
1140 return;
1142 for (bsi = bsi_start (new_bb); !bsi_end_p (bsi);)
1144 tree copy_stmt;
1145 bool can_throw, nonlocal_goto;
1147 copy_stmt = bsi_stmt (bsi);
1148 update_stmt (copy_stmt);
1149 if (gimple_in_ssa_p (cfun))
1150 mark_symbols_for_renaming (copy_stmt);
1151 /* Do this before the possible split_block. */
1152 bsi_next (&bsi);
1154 /* If this tree could throw an exception, there are two
1155 cases where we need to add abnormal edge(s): the
1156 tree wasn't in a region and there is a "current
1157 region" in the caller; or the original tree had
1158 EH edges. In both cases split the block after the tree,
1159 and add abnormal edge(s) as needed; we need both
1160 those from the callee and the caller.
1161 We check whether the copy can throw, because the const
1162 propagation can change an INDIRECT_REF which throws
1163 into a COMPONENT_REF which doesn't. If the copy
1164 can throw, the original could also throw. */
1166 can_throw = tree_can_throw_internal (copy_stmt);
1167 nonlocal_goto = tree_can_make_abnormal_goto (copy_stmt);
1169 if (can_throw || nonlocal_goto)
1171 if (!bsi_end_p (bsi))
1172 /* Note that bb's predecessor edges aren't necessarily
1173 right at this point; split_block doesn't care. */
1175 edge e = split_block (new_bb, copy_stmt);
1177 new_bb = e->dest;
1178 new_bb->aux = e->src->aux;
1179 bsi = bsi_start (new_bb);
1183 if (can_throw)
1184 make_eh_edges (copy_stmt);
1186 if (nonlocal_goto)
1187 make_abnormal_goto_edges (bb_for_stmt (copy_stmt), true);
1189 if ((can_throw || nonlocal_goto)
1190 && gimple_in_ssa_p (cfun))
1191 update_ssa_across_abnormal_edges (bb_for_stmt (copy_stmt), ret_bb,
1192 can_throw, nonlocal_goto);
1196 /* Copy the PHIs. All blocks and edges are copied, some blocks
1197 was possibly split and new outgoing EH edges inserted.
1198 BB points to the block of original function and AUX pointers links
1199 the original and newly copied blocks. */
1201 static void
1202 copy_phis_for_bb (basic_block bb, copy_body_data *id)
1204 basic_block new_bb = bb->aux;
1205 edge_iterator ei;
1206 tree phi;
1208 for (phi = phi_nodes (bb); phi; phi = PHI_CHAIN (phi))
1210 tree res = PHI_RESULT (phi);
1211 tree new_res = res;
1212 tree new_phi;
1213 edge new_edge;
1215 if (is_gimple_reg (res))
1217 walk_tree (&new_res, copy_body_r, id, NULL);
1218 SSA_NAME_DEF_STMT (new_res)
1219 = new_phi = create_phi_node (new_res, new_bb);
1220 FOR_EACH_EDGE (new_edge, ei, new_bb->preds)
1222 edge old_edge = find_edge (new_edge->src->aux, bb);
1223 tree arg = PHI_ARG_DEF_FROM_EDGE (phi, old_edge);
1224 tree new_arg = arg;
1226 walk_tree (&new_arg, copy_body_r, id, NULL);
1227 gcc_assert (new_arg);
1228 /* With return slot optimization we can end up with
1229 non-gimple (foo *)&this->m, fix that here. */
1230 if (TREE_CODE (new_arg) != SSA_NAME
1231 && TREE_CODE (new_arg) != FUNCTION_DECL
1232 && !is_gimple_val (new_arg))
1234 tree stmts = NULL_TREE;
1235 new_arg = force_gimple_operand (new_arg, &stmts,
1236 true, NULL);
1237 bsi_insert_on_edge_immediate (new_edge, stmts);
1239 add_phi_arg (new_phi, new_arg, new_edge);
1245 /* Wrapper for remap_decl so it can be used as a callback. */
1246 static tree
1247 remap_decl_1 (tree decl, void *data)
1249 return remap_decl (decl, (copy_body_data *) data);
1252 /* Build struct function and associated datastructures for the new clone
1253 NEW_FNDECL to be build. CALLEE_FNDECL is the original */
1255 static void
1256 initialize_cfun (tree new_fndecl, tree callee_fndecl, gcov_type count,
1257 int frequency)
1259 struct function *new_cfun
1260 = (struct function *) ggc_alloc_cleared (sizeof (struct function));
1261 struct function *src_cfun = DECL_STRUCT_FUNCTION (callee_fndecl);
1262 int count_scale, frequency_scale;
1264 if (ENTRY_BLOCK_PTR_FOR_FUNCTION (src_cfun)->count)
1265 count_scale = (REG_BR_PROB_BASE * count
1266 / ENTRY_BLOCK_PTR_FOR_FUNCTION (src_cfun)->count);
1267 else
1268 count_scale = 1;
1270 if (ENTRY_BLOCK_PTR_FOR_FUNCTION (src_cfun)->frequency)
1271 frequency_scale = (REG_BR_PROB_BASE * frequency
1273 ENTRY_BLOCK_PTR_FOR_FUNCTION (src_cfun)->frequency);
1274 else
1275 frequency_scale = count_scale;
1277 /* Register specific tree functions. */
1278 tree_register_cfg_hooks ();
1279 *new_cfun = *DECL_STRUCT_FUNCTION (callee_fndecl);
1280 new_cfun->funcdef_no = get_next_funcdef_no ();
1281 VALUE_HISTOGRAMS (new_cfun) = NULL;
1282 new_cfun->unexpanded_var_list = NULL;
1283 new_cfun->cfg = NULL;
1284 new_cfun->decl = new_fndecl /*= copy_node (callee_fndecl)*/;
1285 DECL_STRUCT_FUNCTION (new_fndecl) = new_cfun;
1286 push_cfun (new_cfun);
1287 init_empty_tree_cfg ();
1289 ENTRY_BLOCK_PTR->count =
1290 (ENTRY_BLOCK_PTR_FOR_FUNCTION (src_cfun)->count * count_scale /
1291 REG_BR_PROB_BASE);
1292 ENTRY_BLOCK_PTR->frequency =
1293 (ENTRY_BLOCK_PTR_FOR_FUNCTION (src_cfun)->frequency *
1294 frequency_scale / REG_BR_PROB_BASE);
1295 EXIT_BLOCK_PTR->count =
1296 (EXIT_BLOCK_PTR_FOR_FUNCTION (src_cfun)->count * count_scale /
1297 REG_BR_PROB_BASE);
1298 EXIT_BLOCK_PTR->frequency =
1299 (EXIT_BLOCK_PTR_FOR_FUNCTION (src_cfun)->frequency *
1300 frequency_scale / REG_BR_PROB_BASE);
1301 if (src_cfun->eh)
1302 init_eh_for_function ();
1304 if (src_cfun->gimple_df)
1306 init_tree_ssa ();
1307 cfun->gimple_df->in_ssa_p = true;
1308 init_ssa_operands ();
1310 pop_cfun ();
1313 /* Make a copy of the body of FN so that it can be inserted inline in
1314 another function. Walks FN via CFG, returns new fndecl. */
1316 static tree
1317 copy_cfg_body (copy_body_data * id, gcov_type count, int frequency,
1318 basic_block entry_block_map, basic_block exit_block_map)
1320 tree callee_fndecl = id->src_fn;
1321 /* Original cfun for the callee, doesn't change. */
1322 struct function *src_cfun = DECL_STRUCT_FUNCTION (callee_fndecl);
1323 struct function *cfun_to_copy;
1324 basic_block bb;
1325 tree new_fndecl = NULL;
1326 int count_scale, frequency_scale;
1327 int last;
1329 if (ENTRY_BLOCK_PTR_FOR_FUNCTION (src_cfun)->count)
1330 count_scale = (REG_BR_PROB_BASE * count
1331 / ENTRY_BLOCK_PTR_FOR_FUNCTION (src_cfun)->count);
1332 else
1333 count_scale = 1;
1335 if (ENTRY_BLOCK_PTR_FOR_FUNCTION (src_cfun)->frequency)
1336 frequency_scale = (REG_BR_PROB_BASE * frequency
1338 ENTRY_BLOCK_PTR_FOR_FUNCTION (src_cfun)->frequency);
1339 else
1340 frequency_scale = count_scale;
1342 /* Register specific tree functions. */
1343 tree_register_cfg_hooks ();
1345 /* Must have a CFG here at this point. */
1346 gcc_assert (ENTRY_BLOCK_PTR_FOR_FUNCTION
1347 (DECL_STRUCT_FUNCTION (callee_fndecl)));
1349 cfun_to_copy = id->src_cfun = DECL_STRUCT_FUNCTION (callee_fndecl);
1352 ENTRY_BLOCK_PTR_FOR_FUNCTION (cfun_to_copy)->aux = entry_block_map;
1353 EXIT_BLOCK_PTR_FOR_FUNCTION (cfun_to_copy)->aux = exit_block_map;
1354 entry_block_map->aux = ENTRY_BLOCK_PTR_FOR_FUNCTION (cfun_to_copy);
1355 exit_block_map->aux = EXIT_BLOCK_PTR_FOR_FUNCTION (cfun_to_copy);
1357 /* Duplicate any exception-handling regions. */
1358 if (cfun->eh)
1360 id->eh_region_offset
1361 = duplicate_eh_regions (cfun_to_copy, remap_decl_1, id,
1362 0, id->eh_region);
1364 /* Use aux pointers to map the original blocks to copy. */
1365 FOR_EACH_BB_FN (bb, cfun_to_copy)
1367 basic_block new = copy_bb (id, bb, frequency_scale, count_scale);
1368 bb->aux = new;
1369 new->aux = bb;
1372 last = last_basic_block;
1373 /* Now that we've duplicated the blocks, duplicate their edges. */
1374 FOR_ALL_BB_FN (bb, cfun_to_copy)
1375 copy_edges_for_bb (bb, count_scale, exit_block_map);
1376 if (gimple_in_ssa_p (cfun))
1377 FOR_ALL_BB_FN (bb, cfun_to_copy)
1378 copy_phis_for_bb (bb, id);
1379 FOR_ALL_BB_FN (bb, cfun_to_copy)
1381 ((basic_block)bb->aux)->aux = NULL;
1382 bb->aux = NULL;
1384 /* Zero out AUX fields of newly created block during EH edge
1385 insertion. */
1386 for (; last < last_basic_block; last++)
1387 BASIC_BLOCK (last)->aux = NULL;
1388 entry_block_map->aux = NULL;
1389 exit_block_map->aux = NULL;
1391 return new_fndecl;
1394 /* Make a copy of the body of FN so that it can be inserted inline in
1395 another function. */
1397 static tree
1398 copy_generic_body (copy_body_data *id)
1400 tree body;
1401 tree fndecl = id->src_fn;
1403 body = DECL_SAVED_TREE (fndecl);
1404 walk_tree (&body, copy_body_r, id, NULL);
1406 return body;
1409 static tree
1410 copy_body (copy_body_data *id, gcov_type count, int frequency,
1411 basic_block entry_block_map, basic_block exit_block_map)
1413 tree fndecl = id->src_fn;
1414 tree body;
1416 /* If this body has a CFG, walk CFG and copy. */
1417 gcc_assert (ENTRY_BLOCK_PTR_FOR_FUNCTION (DECL_STRUCT_FUNCTION (fndecl)));
1418 body = copy_cfg_body (id, count, frequency, entry_block_map, exit_block_map);
1420 return body;
1423 /* Return true if VALUE is an ADDR_EXPR of an automatic variable
1424 defined in function FN, or of a data member thereof. */
1426 static bool
1427 self_inlining_addr_expr (tree value, tree fn)
1429 tree var;
1431 if (TREE_CODE (value) != ADDR_EXPR)
1432 return false;
1434 var = get_base_address (TREE_OPERAND (value, 0));
1436 return var && auto_var_in_fn_p (var, fn);
1439 static void
1440 setup_one_parameter (copy_body_data *id, tree p, tree value, tree fn,
1441 basic_block bb, tree *vars)
1443 tree init_stmt;
1444 tree var;
1445 tree var_sub;
1446 tree rhs = value;
1447 tree def = (gimple_in_ssa_p (cfun)
1448 ? gimple_default_def (id->src_cfun, p) : NULL);
1450 if (value
1451 && value != error_mark_node
1452 && !useless_type_conversion_p (TREE_TYPE (p), TREE_TYPE (value)))
1454 if (fold_convertible_p (TREE_TYPE (p), value))
1455 rhs = fold_build1 (NOP_EXPR, TREE_TYPE (p), value);
1456 else
1457 /* ??? For valid (GIMPLE) programs we should not end up here.
1458 Still if something has gone wrong and we end up with truly
1459 mismatched types here, fall back to using a VIEW_CONVERT_EXPR
1460 to not leak invalid GIMPLE to the following passes. */
1461 rhs = fold_build1 (VIEW_CONVERT_EXPR, TREE_TYPE (p), value);
1464 /* If the parameter is never assigned to, has no SSA_NAMEs created,
1465 we may not need to create a new variable here at all. Instead, we may
1466 be able to just use the argument value. */
1467 if (TREE_READONLY (p)
1468 && !TREE_ADDRESSABLE (p)
1469 && value && !TREE_SIDE_EFFECTS (value)
1470 && !def)
1472 /* We may produce non-gimple trees by adding NOPs or introduce
1473 invalid sharing when operand is not really constant.
1474 It is not big deal to prohibit constant propagation here as
1475 we will constant propagate in DOM1 pass anyway. */
1476 if (is_gimple_min_invariant (value)
1477 && useless_type_conversion_p (TREE_TYPE (p),
1478 TREE_TYPE (value))
1479 /* We have to be very careful about ADDR_EXPR. Make sure
1480 the base variable isn't a local variable of the inlined
1481 function, e.g., when doing recursive inlining, direct or
1482 mutually-recursive or whatever, which is why we don't
1483 just test whether fn == current_function_decl. */
1484 && ! self_inlining_addr_expr (value, fn))
1486 insert_decl_map (id, p, value);
1487 return;
1491 /* Make an equivalent VAR_DECL. Note that we must NOT remap the type
1492 here since the type of this decl must be visible to the calling
1493 function. */
1494 var = copy_decl_to_var (p, id);
1495 if (gimple_in_ssa_p (cfun) && TREE_CODE (var) == VAR_DECL)
1497 get_var_ann (var);
1498 add_referenced_var (var);
1501 /* See if the frontend wants to pass this by invisible reference. If
1502 so, our new VAR_DECL will have REFERENCE_TYPE, and we need to
1503 replace uses of the PARM_DECL with dereferences. */
1504 if (TREE_TYPE (var) != TREE_TYPE (p)
1505 && POINTER_TYPE_P (TREE_TYPE (var))
1506 && TREE_TYPE (TREE_TYPE (var)) == TREE_TYPE (p))
1508 insert_decl_map (id, var, var);
1509 var_sub = build_fold_indirect_ref (var);
1511 else
1512 var_sub = var;
1514 /* Register the VAR_DECL as the equivalent for the PARM_DECL;
1515 that way, when the PARM_DECL is encountered, it will be
1516 automatically replaced by the VAR_DECL. */
1517 insert_decl_map (id, p, var_sub);
1519 /* Declare this new variable. */
1520 TREE_CHAIN (var) = *vars;
1521 *vars = var;
1523 /* Make gimplifier happy about this variable. */
1524 DECL_SEEN_IN_BIND_EXPR_P (var) = 1;
1526 /* Even if P was TREE_READONLY, the new VAR should not be.
1527 In the original code, we would have constructed a
1528 temporary, and then the function body would have never
1529 changed the value of P. However, now, we will be
1530 constructing VAR directly. The constructor body may
1531 change its value multiple times as it is being
1532 constructed. Therefore, it must not be TREE_READONLY;
1533 the back-end assumes that TREE_READONLY variable is
1534 assigned to only once. */
1535 if (TYPE_NEEDS_CONSTRUCTING (TREE_TYPE (p)))
1536 TREE_READONLY (var) = 0;
1538 /* If there is no setup required and we are in SSA, take the easy route
1539 replacing all SSA names representing the function parameter by the
1540 SSA name passed to function.
1542 We need to construct map for the variable anyway as it might be used
1543 in different SSA names when parameter is set in function.
1545 FIXME: This usually kills the last connection in between inlined
1546 function parameter and the actual value in debug info. Can we do
1547 better here? If we just inserted the statement, copy propagation
1548 would kill it anyway as it always did in older versions of GCC.
1550 We might want to introduce a notion that single SSA_NAME might
1551 represent multiple variables for purposes of debugging. */
1552 if (gimple_in_ssa_p (cfun) && rhs && def && is_gimple_reg (p)
1553 && (TREE_CODE (rhs) == SSA_NAME
1554 || is_gimple_min_invariant (rhs))
1555 && !SSA_NAME_OCCURS_IN_ABNORMAL_PHI (def))
1557 insert_decl_map (id, def, rhs);
1558 return;
1561 /* If the value of argument is never used, don't care about initializing
1562 it. */
1563 if (gimple_in_ssa_p (cfun) && !def && is_gimple_reg (p))
1565 gcc_assert (!value || !TREE_SIDE_EFFECTS (value));
1566 return;
1569 /* Initialize this VAR_DECL from the equivalent argument. Convert
1570 the argument to the proper type in case it was promoted. */
1571 if (value)
1573 block_stmt_iterator bsi = bsi_last (bb);
1575 if (rhs == error_mark_node)
1577 insert_decl_map (id, p, var_sub);
1578 return;
1581 STRIP_USELESS_TYPE_CONVERSION (rhs);
1583 /* We want to use GIMPLE_MODIFY_STMT, not INIT_EXPR here so that we
1584 keep our trees in gimple form. */
1585 if (def && gimple_in_ssa_p (cfun) && is_gimple_reg (p))
1587 def = remap_ssa_name (def, id);
1588 init_stmt = build_gimple_modify_stmt (def, rhs);
1589 SSA_NAME_DEF_STMT (def) = init_stmt;
1590 SSA_NAME_IS_DEFAULT_DEF (def) = 0;
1591 set_default_def (var, NULL);
1593 else
1594 init_stmt = build_gimple_modify_stmt (var, rhs);
1596 /* If we did not create a gimple value and we did not create a gimple
1597 cast of a gimple value, then we will need to gimplify INIT_STMTS
1598 at the end. Note that is_gimple_cast only checks the outer
1599 tree code, not its operand. Thus the explicit check that its
1600 operand is a gimple value. */
1601 if ((!is_gimple_val (rhs)
1602 && (!is_gimple_cast (rhs)
1603 || !is_gimple_val (TREE_OPERAND (rhs, 0))))
1604 || !is_gimple_reg (var))
1606 tree_stmt_iterator i;
1608 push_gimplify_context ();
1609 gimplify_stmt (&init_stmt);
1610 if (gimple_in_ssa_p (cfun)
1611 && init_stmt && TREE_CODE (init_stmt) == STATEMENT_LIST)
1613 /* The replacement can expose previously unreferenced
1614 variables. */
1615 for (i = tsi_start (init_stmt); !tsi_end_p (i); tsi_next (&i))
1616 find_new_referenced_vars (tsi_stmt_ptr (i));
1618 pop_gimplify_context (NULL);
1621 /* If VAR represents a zero-sized variable, it's possible that the
1622 assignment statment may result in no gimple statements. */
1623 if (init_stmt)
1624 bsi_insert_after (&bsi, init_stmt, BSI_NEW_STMT);
1625 if (gimple_in_ssa_p (cfun))
1626 for (;!bsi_end_p (bsi); bsi_next (&bsi))
1627 mark_symbols_for_renaming (bsi_stmt (bsi));
1631 /* Generate code to initialize the parameters of the function at the
1632 top of the stack in ID from the CALL_EXPR EXP. */
1634 static void
1635 initialize_inlined_parameters (copy_body_data *id, tree exp,
1636 tree fn, basic_block bb)
1638 tree parms;
1639 tree a;
1640 tree p;
1641 tree vars = NULL_TREE;
1642 call_expr_arg_iterator iter;
1643 tree static_chain = CALL_EXPR_STATIC_CHAIN (exp);
1645 /* Figure out what the parameters are. */
1646 parms = DECL_ARGUMENTS (fn);
1648 /* Loop through the parameter declarations, replacing each with an
1649 equivalent VAR_DECL, appropriately initialized. */
1650 for (p = parms, a = first_call_expr_arg (exp, &iter); p;
1651 a = next_call_expr_arg (&iter), p = TREE_CHAIN (p))
1652 setup_one_parameter (id, p, a, fn, bb, &vars);
1654 /* Initialize the static chain. */
1655 p = DECL_STRUCT_FUNCTION (fn)->static_chain_decl;
1656 gcc_assert (fn != current_function_decl);
1657 if (p)
1659 /* No static chain? Seems like a bug in tree-nested.c. */
1660 gcc_assert (static_chain);
1662 setup_one_parameter (id, p, static_chain, fn, bb, &vars);
1665 declare_inline_vars (id->block, vars);
1668 /* Declare a return variable to replace the RESULT_DECL for the
1669 function we are calling. An appropriate DECL_STMT is returned.
1670 The USE_STMT is filled to contain a use of the declaration to
1671 indicate the return value of the function.
1673 RETURN_SLOT, if non-null is place where to store the result. It
1674 is set only for CALL_EXPR_RETURN_SLOT_OPT. MODIFY_DEST, if non-null,
1675 was the LHS of the GIMPLE_MODIFY_STMT to which this call is the RHS.
1677 The return value is a (possibly null) value that is the result of the
1678 function as seen by the callee. *USE_P is a (possibly null) value that
1679 holds the result as seen by the caller. */
1681 static tree
1682 declare_return_variable (copy_body_data *id, tree return_slot, tree modify_dest,
1683 tree *use_p)
1685 tree callee = id->src_fn;
1686 tree caller = id->dst_fn;
1687 tree result = DECL_RESULT (callee);
1688 tree callee_type = TREE_TYPE (result);
1689 tree caller_type = TREE_TYPE (TREE_TYPE (callee));
1690 tree var, use;
1692 /* We don't need to do anything for functions that don't return
1693 anything. */
1694 if (!result || VOID_TYPE_P (callee_type))
1696 *use_p = NULL_TREE;
1697 return NULL_TREE;
1700 /* If there was a return slot, then the return value is the
1701 dereferenced address of that object. */
1702 if (return_slot)
1704 /* The front end shouldn't have used both return_slot and
1705 a modify expression. */
1706 gcc_assert (!modify_dest);
1707 if (DECL_BY_REFERENCE (result))
1709 tree return_slot_addr = build_fold_addr_expr (return_slot);
1710 STRIP_USELESS_TYPE_CONVERSION (return_slot_addr);
1712 /* We are going to construct *&return_slot and we can't do that
1713 for variables believed to be not addressable.
1715 FIXME: This check possibly can match, because values returned
1716 via return slot optimization are not believed to have address
1717 taken by alias analysis. */
1718 gcc_assert (TREE_CODE (return_slot) != SSA_NAME);
1719 if (gimple_in_ssa_p (cfun))
1721 HOST_WIDE_INT bitsize;
1722 HOST_WIDE_INT bitpos;
1723 tree offset;
1724 enum machine_mode mode;
1725 int unsignedp;
1726 int volatilep;
1727 tree base;
1728 base = get_inner_reference (return_slot, &bitsize, &bitpos,
1729 &offset,
1730 &mode, &unsignedp, &volatilep,
1731 false);
1732 if (TREE_CODE (base) == INDIRECT_REF)
1733 base = TREE_OPERAND (base, 0);
1734 if (TREE_CODE (base) == SSA_NAME)
1735 base = SSA_NAME_VAR (base);
1736 mark_sym_for_renaming (base);
1738 var = return_slot_addr;
1740 else
1742 var = return_slot;
1743 gcc_assert (TREE_CODE (var) != SSA_NAME);
1744 TREE_ADDRESSABLE (var) |= TREE_ADDRESSABLE (result);
1746 if ((TREE_CODE (TREE_TYPE (result)) == COMPLEX_TYPE
1747 || TREE_CODE (TREE_TYPE (result)) == VECTOR_TYPE)
1748 && !DECL_GIMPLE_REG_P (result)
1749 && DECL_P (var))
1750 DECL_GIMPLE_REG_P (var) = 0;
1751 use = NULL;
1752 goto done;
1755 /* All types requiring non-trivial constructors should have been handled. */
1756 gcc_assert (!TREE_ADDRESSABLE (callee_type));
1758 /* Attempt to avoid creating a new temporary variable. */
1759 if (modify_dest
1760 && TREE_CODE (modify_dest) != SSA_NAME)
1762 bool use_it = false;
1764 /* We can't use MODIFY_DEST if there's type promotion involved. */
1765 if (!useless_type_conversion_p (callee_type, caller_type))
1766 use_it = false;
1768 /* ??? If we're assigning to a variable sized type, then we must
1769 reuse the destination variable, because we've no good way to
1770 create variable sized temporaries at this point. */
1771 else if (TREE_CODE (TYPE_SIZE_UNIT (caller_type)) != INTEGER_CST)
1772 use_it = true;
1774 /* If the callee cannot possibly modify MODIFY_DEST, then we can
1775 reuse it as the result of the call directly. Don't do this if
1776 it would promote MODIFY_DEST to addressable. */
1777 else if (TREE_ADDRESSABLE (result))
1778 use_it = false;
1779 else
1781 tree base_m = get_base_address (modify_dest);
1783 /* If the base isn't a decl, then it's a pointer, and we don't
1784 know where that's going to go. */
1785 if (!DECL_P (base_m))
1786 use_it = false;
1787 else if (is_global_var (base_m))
1788 use_it = false;
1789 else if ((TREE_CODE (TREE_TYPE (result)) == COMPLEX_TYPE
1790 || TREE_CODE (TREE_TYPE (result)) == VECTOR_TYPE)
1791 && !DECL_GIMPLE_REG_P (result)
1792 && DECL_GIMPLE_REG_P (base_m))
1793 use_it = false;
1794 else if (!TREE_ADDRESSABLE (base_m))
1795 use_it = true;
1798 if (use_it)
1800 var = modify_dest;
1801 use = NULL;
1802 goto done;
1806 gcc_assert (TREE_CODE (TYPE_SIZE_UNIT (callee_type)) == INTEGER_CST);
1808 var = copy_result_decl_to_var (result, id);
1809 if (gimple_in_ssa_p (cfun))
1811 get_var_ann (var);
1812 add_referenced_var (var);
1815 DECL_SEEN_IN_BIND_EXPR_P (var) = 1;
1816 DECL_STRUCT_FUNCTION (caller)->unexpanded_var_list
1817 = tree_cons (NULL_TREE, var,
1818 DECL_STRUCT_FUNCTION (caller)->unexpanded_var_list);
1820 /* Do not have the rest of GCC warn about this variable as it should
1821 not be visible to the user. */
1822 TREE_NO_WARNING (var) = 1;
1824 declare_inline_vars (id->block, var);
1826 /* Build the use expr. If the return type of the function was
1827 promoted, convert it back to the expected type. */
1828 use = var;
1829 if (!useless_type_conversion_p (caller_type, TREE_TYPE (var)))
1830 use = fold_convert (caller_type, var);
1832 STRIP_USELESS_TYPE_CONVERSION (use);
1834 if (DECL_BY_REFERENCE (result))
1835 var = build_fold_addr_expr (var);
1837 done:
1838 /* Register the VAR_DECL as the equivalent for the RESULT_DECL; that
1839 way, when the RESULT_DECL is encountered, it will be
1840 automatically replaced by the VAR_DECL. */
1841 insert_decl_map (id, result, var);
1843 /* Remember this so we can ignore it in remap_decls. */
1844 id->retvar = var;
1846 *use_p = use;
1847 return var;
1850 /* Returns nonzero if a function can be inlined as a tree. */
1852 bool
1853 tree_inlinable_function_p (tree fn)
1855 return inlinable_function_p (fn);
1858 static const char *inline_forbidden_reason;
1860 static tree
1861 inline_forbidden_p_1 (tree *nodep, int *walk_subtrees ATTRIBUTE_UNUSED,
1862 void *fnp)
1864 tree node = *nodep;
1865 tree fn = (tree) fnp;
1866 tree t;
1868 switch (TREE_CODE (node))
1870 case CALL_EXPR:
1871 /* Refuse to inline alloca call unless user explicitly forced so as
1872 this may change program's memory overhead drastically when the
1873 function using alloca is called in loop. In GCC present in
1874 SPEC2000 inlining into schedule_block cause it to require 2GB of
1875 RAM instead of 256MB. */
1876 if (alloca_call_p (node)
1877 && !lookup_attribute ("always_inline", DECL_ATTRIBUTES (fn)))
1879 inline_forbidden_reason
1880 = G_("function %q+F can never be inlined because it uses "
1881 "alloca (override using the always_inline attribute)");
1882 return node;
1884 t = get_callee_fndecl (node);
1885 if (! t)
1886 break;
1888 /* We cannot inline functions that call setjmp. */
1889 if (setjmp_call_p (t))
1891 inline_forbidden_reason
1892 = G_("function %q+F can never be inlined because it uses setjmp");
1893 return node;
1896 if (DECL_BUILT_IN_CLASS (t) == BUILT_IN_NORMAL)
1897 switch (DECL_FUNCTION_CODE (t))
1899 /* We cannot inline functions that take a variable number of
1900 arguments. */
1901 case BUILT_IN_VA_START:
1902 case BUILT_IN_STDARG_START:
1903 case BUILT_IN_NEXT_ARG:
1904 case BUILT_IN_VA_END:
1905 inline_forbidden_reason
1906 = G_("function %q+F can never be inlined because it "
1907 "uses variable argument lists");
1908 return node;
1910 case BUILT_IN_LONGJMP:
1911 /* We can't inline functions that call __builtin_longjmp at
1912 all. The non-local goto machinery really requires the
1913 destination be in a different function. If we allow the
1914 function calling __builtin_longjmp to be inlined into the
1915 function calling __builtin_setjmp, Things will Go Awry. */
1916 inline_forbidden_reason
1917 = G_("function %q+F can never be inlined because "
1918 "it uses setjmp-longjmp exception handling");
1919 return node;
1921 case BUILT_IN_NONLOCAL_GOTO:
1922 /* Similarly. */
1923 inline_forbidden_reason
1924 = G_("function %q+F can never be inlined because "
1925 "it uses non-local goto");
1926 return node;
1928 case BUILT_IN_RETURN:
1929 case BUILT_IN_APPLY_ARGS:
1930 /* If a __builtin_apply_args caller would be inlined,
1931 it would be saving arguments of the function it has
1932 been inlined into. Similarly __builtin_return would
1933 return from the function the inline has been inlined into. */
1934 inline_forbidden_reason
1935 = G_("function %q+F can never be inlined because "
1936 "it uses __builtin_return or __builtin_apply_args");
1937 return node;
1939 default:
1940 break;
1942 break;
1944 case GOTO_EXPR:
1945 t = TREE_OPERAND (node, 0);
1947 /* We will not inline a function which uses computed goto. The
1948 addresses of its local labels, which may be tucked into
1949 global storage, are of course not constant across
1950 instantiations, which causes unexpected behavior. */
1951 if (TREE_CODE (t) != LABEL_DECL)
1953 inline_forbidden_reason
1954 = G_("function %q+F can never be inlined "
1955 "because it contains a computed goto");
1956 return node;
1958 break;
1960 case LABEL_EXPR:
1961 t = TREE_OPERAND (node, 0);
1962 if (DECL_NONLOCAL (t))
1964 /* We cannot inline a function that receives a non-local goto
1965 because we cannot remap the destination label used in the
1966 function that is performing the non-local goto. */
1967 inline_forbidden_reason
1968 = G_("function %q+F can never be inlined "
1969 "because it receives a non-local goto");
1970 return node;
1972 break;
1974 case RECORD_TYPE:
1975 case UNION_TYPE:
1976 /* We cannot inline a function of the form
1978 void F (int i) { struct S { int ar[i]; } s; }
1980 Attempting to do so produces a catch-22.
1981 If walk_tree examines the TYPE_FIELDS chain of RECORD_TYPE/
1982 UNION_TYPE nodes, then it goes into infinite recursion on a
1983 structure containing a pointer to its own type. If it doesn't,
1984 then the type node for S doesn't get adjusted properly when
1985 F is inlined.
1987 ??? This is likely no longer true, but it's too late in the 4.0
1988 cycle to try to find out. This should be checked for 4.1. */
1989 for (t = TYPE_FIELDS (node); t; t = TREE_CHAIN (t))
1990 if (variably_modified_type_p (TREE_TYPE (t), NULL))
1992 inline_forbidden_reason
1993 = G_("function %q+F can never be inlined "
1994 "because it uses variable sized variables");
1995 return node;
1998 default:
1999 break;
2002 return NULL_TREE;
2005 static tree
2006 inline_forbidden_p_2 (tree *nodep, int *walk_subtrees,
2007 void *fnp)
2009 tree node = *nodep;
2010 tree fn = (tree) fnp;
2012 if (TREE_CODE (node) == LABEL_DECL && DECL_CONTEXT (node) == fn)
2014 inline_forbidden_reason
2015 = G_("function %q+F can never be inlined "
2016 "because it saves address of local label in a static variable");
2017 return node;
2020 if (TYPE_P (node))
2021 *walk_subtrees = 0;
2023 return NULL_TREE;
2026 /* Return subexpression representing possible alloca call, if any. */
2027 static tree
2028 inline_forbidden_p (tree fndecl)
2030 location_t saved_loc = input_location;
2031 block_stmt_iterator bsi;
2032 basic_block bb;
2033 tree ret = NULL_TREE;
2034 struct function *fun = DECL_STRUCT_FUNCTION (fndecl);
2035 tree step;
2037 FOR_EACH_BB_FN (bb, fun)
2038 for (bsi = bsi_start (bb); !bsi_end_p (bsi); bsi_next (&bsi))
2040 ret = walk_tree_without_duplicates (bsi_stmt_ptr (bsi),
2041 inline_forbidden_p_1, fndecl);
2042 if (ret)
2043 goto egress;
2046 for (step = fun->unexpanded_var_list; step; step = TREE_CHAIN (step))
2048 tree decl = TREE_VALUE (step);
2049 if (TREE_CODE (decl) == VAR_DECL
2050 && TREE_STATIC (decl)
2051 && !DECL_EXTERNAL (decl)
2052 && DECL_INITIAL (decl))
2053 ret = walk_tree_without_duplicates (&DECL_INITIAL (decl),
2054 inline_forbidden_p_2, fndecl);
2055 if (ret)
2056 goto egress;
2059 egress:
2060 input_location = saved_loc;
2061 return ret;
2064 /* Returns nonzero if FN is a function that does not have any
2065 fundamental inline blocking properties. */
2067 static bool
2068 inlinable_function_p (tree fn)
2070 bool inlinable = true;
2071 bool do_warning;
2072 tree always_inline;
2074 /* If we've already decided this function shouldn't be inlined,
2075 there's no need to check again. */
2076 if (DECL_UNINLINABLE (fn))
2077 return false;
2079 /* We only warn for functions declared `inline' by the user. */
2080 do_warning = (warn_inline
2081 && DECL_INLINE (fn)
2082 && DECL_DECLARED_INLINE_P (fn)
2083 && !DECL_IN_SYSTEM_HEADER (fn));
2085 always_inline = lookup_attribute ("always_inline", DECL_ATTRIBUTES (fn));
2087 if (flag_really_no_inline
2088 && always_inline == NULL)
2090 if (do_warning)
2091 warning (OPT_Winline, "function %q+F can never be inlined because it "
2092 "is suppressed using -fno-inline", fn);
2093 inlinable = false;
2096 /* Don't auto-inline anything that might not be bound within
2097 this unit of translation. */
2098 else if (!DECL_DECLARED_INLINE_P (fn)
2099 && DECL_REPLACEABLE_P (fn))
2100 inlinable = false;
2102 else if (!function_attribute_inlinable_p (fn))
2104 if (do_warning)
2105 warning (OPT_Winline, "function %q+F can never be inlined because it "
2106 "uses attributes conflicting with inlining", fn);
2107 inlinable = false;
2110 /* If we don't have the function body available, we can't inline it.
2111 However, this should not be recorded since we also get here for
2112 forward declared inline functions. Therefore, return at once. */
2113 if (!DECL_SAVED_TREE (fn))
2114 return false;
2116 /* If we're not inlining at all, then we cannot inline this function. */
2117 else if (!flag_inline_trees)
2118 inlinable = false;
2120 /* Only try to inline functions if DECL_INLINE is set. This should be
2121 true for all functions declared `inline', and for all other functions
2122 as well with -finline-functions.
2124 Don't think of disregarding DECL_INLINE when flag_inline_trees == 2;
2125 it's the front-end that must set DECL_INLINE in this case, because
2126 dwarf2out loses if a function that does not have DECL_INLINE set is
2127 inlined anyway. That is why we have both DECL_INLINE and
2128 DECL_DECLARED_INLINE_P. */
2129 /* FIXME: When flag_inline_trees dies, the check for flag_unit_at_a_time
2130 here should be redundant. */
2131 else if (!DECL_INLINE (fn) && !flag_unit_at_a_time)
2132 inlinable = false;
2134 else if (inline_forbidden_p (fn))
2136 /* See if we should warn about uninlinable functions. Previously,
2137 some of these warnings would be issued while trying to expand
2138 the function inline, but that would cause multiple warnings
2139 about functions that would for example call alloca. But since
2140 this a property of the function, just one warning is enough.
2141 As a bonus we can now give more details about the reason why a
2142 function is not inlinable. */
2143 if (always_inline)
2144 sorry (inline_forbidden_reason, fn);
2145 else if (do_warning)
2146 warning (OPT_Winline, inline_forbidden_reason, fn);
2148 inlinable = false;
2151 /* Squirrel away the result so that we don't have to check again. */
2152 DECL_UNINLINABLE (fn) = !inlinable;
2154 return inlinable;
2157 /* Estimate the cost of a memory move. Use machine dependent
2158 word size and take possible memcpy call into account. */
2161 estimate_move_cost (tree type)
2163 HOST_WIDE_INT size;
2165 size = int_size_in_bytes (type);
2167 if (size < 0 || size > MOVE_MAX_PIECES * MOVE_RATIO)
2168 /* Cost of a memcpy call, 3 arguments and the call. */
2169 return 4;
2170 else
2171 return ((size + MOVE_MAX_PIECES - 1) / MOVE_MAX_PIECES);
2174 /* Arguments for estimate_num_insns_1. */
2176 struct eni_data
2178 /* Used to return the number of insns. */
2179 int count;
2181 /* Weights of various constructs. */
2182 eni_weights *weights;
2185 /* Used by estimate_num_insns. Estimate number of instructions seen
2186 by given statement. */
2188 static tree
2189 estimate_num_insns_1 (tree *tp, int *walk_subtrees, void *data)
2191 struct eni_data *d = data;
2192 tree x = *tp;
2193 unsigned cost;
2195 if (IS_TYPE_OR_DECL_P (x))
2197 *walk_subtrees = 0;
2198 return NULL;
2200 /* Assume that constants and references counts nothing. These should
2201 be majorized by amount of operations among them we count later
2202 and are common target of CSE and similar optimizations. */
2203 else if (CONSTANT_CLASS_P (x) || REFERENCE_CLASS_P (x))
2204 return NULL;
2206 switch (TREE_CODE (x))
2208 /* Containers have no cost. */
2209 case TREE_LIST:
2210 case TREE_VEC:
2211 case BLOCK:
2212 case COMPONENT_REF:
2213 case BIT_FIELD_REF:
2214 case INDIRECT_REF:
2215 case ALIGN_INDIRECT_REF:
2216 case MISALIGNED_INDIRECT_REF:
2217 case ARRAY_REF:
2218 case ARRAY_RANGE_REF:
2219 case OBJ_TYPE_REF:
2220 case EXC_PTR_EXPR: /* ??? */
2221 case FILTER_EXPR: /* ??? */
2222 case COMPOUND_EXPR:
2223 case BIND_EXPR:
2224 case WITH_CLEANUP_EXPR:
2225 case NOP_EXPR:
2226 case CONVERT_EXPR:
2227 case VIEW_CONVERT_EXPR:
2228 case SAVE_EXPR:
2229 case ADDR_EXPR:
2230 case COMPLEX_EXPR:
2231 case RANGE_EXPR:
2232 case CASE_LABEL_EXPR:
2233 case SSA_NAME:
2234 case CATCH_EXPR:
2235 case EH_FILTER_EXPR:
2236 case STATEMENT_LIST:
2237 case ERROR_MARK:
2238 case NON_LVALUE_EXPR:
2239 case FDESC_EXPR:
2240 case VA_ARG_EXPR:
2241 case TRY_CATCH_EXPR:
2242 case TRY_FINALLY_EXPR:
2243 case LABEL_EXPR:
2244 case GOTO_EXPR:
2245 case RETURN_EXPR:
2246 case EXIT_EXPR:
2247 case LOOP_EXPR:
2248 case PHI_NODE:
2249 case WITH_SIZE_EXPR:
2250 case OMP_CLAUSE:
2251 case OMP_RETURN:
2252 case OMP_CONTINUE:
2253 case OMP_SECTIONS_SWITCH:
2254 case OMP_ATOMIC_STORE:
2255 break;
2257 /* We don't account constants for now. Assume that the cost is amortized
2258 by operations that do use them. We may re-consider this decision once
2259 we are able to optimize the tree before estimating its size and break
2260 out static initializers. */
2261 case IDENTIFIER_NODE:
2262 case INTEGER_CST:
2263 case REAL_CST:
2264 case FIXED_CST:
2265 case COMPLEX_CST:
2266 case VECTOR_CST:
2267 case STRING_CST:
2268 *walk_subtrees = 0;
2269 return NULL;
2271 /* CHANGE_DYNAMIC_TYPE_EXPR explicitly expands to nothing. */
2272 case CHANGE_DYNAMIC_TYPE_EXPR:
2273 *walk_subtrees = 0;
2274 return NULL;
2276 /* Try to estimate the cost of assignments. We have three cases to
2277 deal with:
2278 1) Simple assignments to registers;
2279 2) Stores to things that must live in memory. This includes
2280 "normal" stores to scalars, but also assignments of large
2281 structures, or constructors of big arrays;
2282 3) TARGET_EXPRs.
2284 Let us look at the first two cases, assuming we have "a = b + C":
2285 <GIMPLE_MODIFY_STMT <var_decl "a">
2286 <plus_expr <var_decl "b"> <constant C>>
2287 If "a" is a GIMPLE register, the assignment to it is free on almost
2288 any target, because "a" usually ends up in a real register. Hence
2289 the only cost of this expression comes from the PLUS_EXPR, and we
2290 can ignore the GIMPLE_MODIFY_STMT.
2291 If "a" is not a GIMPLE register, the assignment to "a" will most
2292 likely be a real store, so the cost of the GIMPLE_MODIFY_STMT is the cost
2293 of moving something into "a", which we compute using the function
2294 estimate_move_cost.
2296 The third case deals with TARGET_EXPRs, for which the semantics are
2297 that a temporary is assigned, unless the TARGET_EXPR itself is being
2298 assigned to something else. In the latter case we do not need the
2299 temporary. E.g. in:
2300 <GIMPLE_MODIFY_STMT <var_decl "a"> <target_expr>>, the
2301 GIMPLE_MODIFY_STMT is free. */
2302 case INIT_EXPR:
2303 case GIMPLE_MODIFY_STMT:
2304 /* Is the right and side a TARGET_EXPR? */
2305 if (TREE_CODE (GENERIC_TREE_OPERAND (x, 1)) == TARGET_EXPR)
2306 break;
2307 /* ... fall through ... */
2309 case TARGET_EXPR:
2310 x = GENERIC_TREE_OPERAND (x, 0);
2311 /* Is this an assignments to a register? */
2312 if (is_gimple_reg (x))
2313 break;
2314 /* Otherwise it's a store, so fall through to compute the move cost. */
2316 case CONSTRUCTOR:
2317 d->count += estimate_move_cost (TREE_TYPE (x));
2318 break;
2320 /* Assign cost of 1 to usual operations.
2321 ??? We may consider mapping RTL costs to this. */
2322 case COND_EXPR:
2323 case VEC_COND_EXPR:
2325 case PLUS_EXPR:
2326 case POINTER_PLUS_EXPR:
2327 case MINUS_EXPR:
2328 case MULT_EXPR:
2330 case FIXED_CONVERT_EXPR:
2331 case FIX_TRUNC_EXPR:
2333 case NEGATE_EXPR:
2334 case FLOAT_EXPR:
2335 case MIN_EXPR:
2336 case MAX_EXPR:
2337 case ABS_EXPR:
2339 case LSHIFT_EXPR:
2340 case RSHIFT_EXPR:
2341 case LROTATE_EXPR:
2342 case RROTATE_EXPR:
2343 case VEC_LSHIFT_EXPR:
2344 case VEC_RSHIFT_EXPR:
2346 case BIT_IOR_EXPR:
2347 case BIT_XOR_EXPR:
2348 case BIT_AND_EXPR:
2349 case BIT_NOT_EXPR:
2351 case TRUTH_ANDIF_EXPR:
2352 case TRUTH_ORIF_EXPR:
2353 case TRUTH_AND_EXPR:
2354 case TRUTH_OR_EXPR:
2355 case TRUTH_XOR_EXPR:
2356 case TRUTH_NOT_EXPR:
2358 case LT_EXPR:
2359 case LE_EXPR:
2360 case GT_EXPR:
2361 case GE_EXPR:
2362 case EQ_EXPR:
2363 case NE_EXPR:
2364 case ORDERED_EXPR:
2365 case UNORDERED_EXPR:
2367 case UNLT_EXPR:
2368 case UNLE_EXPR:
2369 case UNGT_EXPR:
2370 case UNGE_EXPR:
2371 case UNEQ_EXPR:
2372 case LTGT_EXPR:
2374 case CONJ_EXPR:
2376 case PREDECREMENT_EXPR:
2377 case PREINCREMENT_EXPR:
2378 case POSTDECREMENT_EXPR:
2379 case POSTINCREMENT_EXPR:
2381 case ASM_EXPR:
2383 case REALIGN_LOAD_EXPR:
2385 case REDUC_MAX_EXPR:
2386 case REDUC_MIN_EXPR:
2387 case REDUC_PLUS_EXPR:
2388 case WIDEN_SUM_EXPR:
2389 case DOT_PROD_EXPR:
2390 case VEC_WIDEN_MULT_HI_EXPR:
2391 case VEC_WIDEN_MULT_LO_EXPR:
2392 case VEC_UNPACK_HI_EXPR:
2393 case VEC_UNPACK_LO_EXPR:
2394 case VEC_UNPACK_FLOAT_HI_EXPR:
2395 case VEC_UNPACK_FLOAT_LO_EXPR:
2396 case VEC_PACK_TRUNC_EXPR:
2397 case VEC_PACK_SAT_EXPR:
2398 case VEC_PACK_FIX_TRUNC_EXPR:
2400 case WIDEN_MULT_EXPR:
2402 case VEC_EXTRACT_EVEN_EXPR:
2403 case VEC_EXTRACT_ODD_EXPR:
2404 case VEC_INTERLEAVE_HIGH_EXPR:
2405 case VEC_INTERLEAVE_LOW_EXPR:
2407 case RESX_EXPR:
2408 d->count += 1;
2409 break;
2411 case SWITCH_EXPR:
2412 /* Take into account cost of the switch + guess 2 conditional jumps for
2413 each case label.
2415 TODO: once the switch expansion logic is sufficiently separated, we can
2416 do better job on estimating cost of the switch. */
2417 d->count += TREE_VEC_LENGTH (SWITCH_LABELS (x)) * 2;
2418 break;
2420 /* Few special cases of expensive operations. This is useful
2421 to avoid inlining on functions having too many of these. */
2422 case TRUNC_DIV_EXPR:
2423 case CEIL_DIV_EXPR:
2424 case FLOOR_DIV_EXPR:
2425 case ROUND_DIV_EXPR:
2426 case EXACT_DIV_EXPR:
2427 case TRUNC_MOD_EXPR:
2428 case CEIL_MOD_EXPR:
2429 case FLOOR_MOD_EXPR:
2430 case ROUND_MOD_EXPR:
2431 case RDIV_EXPR:
2432 d->count += d->weights->div_mod_cost;
2433 break;
2434 case CALL_EXPR:
2436 tree decl = get_callee_fndecl (x);
2438 if (decl && DECL_BUILT_IN_CLASS (decl) == BUILT_IN_MD)
2439 cost = d->weights->target_builtin_call_cost;
2440 else
2441 cost = d->weights->call_cost;
2443 if (decl && DECL_BUILT_IN_CLASS (decl) == BUILT_IN_NORMAL)
2444 switch (DECL_FUNCTION_CODE (decl))
2446 case BUILT_IN_CONSTANT_P:
2447 *walk_subtrees = 0;
2448 return NULL_TREE;
2449 case BUILT_IN_EXPECT:
2450 return NULL_TREE;
2451 /* Prefetch instruction is not expensive. */
2452 case BUILT_IN_PREFETCH:
2453 cost = 1;
2454 break;
2455 default:
2456 break;
2459 /* Our cost must be kept in sync with cgraph_estimate_size_after_inlining
2460 that does use function declaration to figure out the arguments. */
2461 if (!decl)
2463 tree a;
2464 call_expr_arg_iterator iter;
2465 FOR_EACH_CALL_EXPR_ARG (a, iter, x)
2466 d->count += estimate_move_cost (TREE_TYPE (a));
2468 else
2470 tree arg;
2471 for (arg = DECL_ARGUMENTS (decl); arg; arg = TREE_CHAIN (arg))
2472 d->count += estimate_move_cost (TREE_TYPE (arg));
2475 d->count += cost;
2476 break;
2479 case OMP_PARALLEL:
2480 case OMP_FOR:
2481 case OMP_SECTIONS:
2482 case OMP_SINGLE:
2483 case OMP_SECTION:
2484 case OMP_MASTER:
2485 case OMP_ORDERED:
2486 case OMP_CRITICAL:
2487 case OMP_ATOMIC:
2488 case OMP_ATOMIC_LOAD:
2489 /* OpenMP directives are generally very expensive. */
2490 d->count += d->weights->omp_cost;
2491 break;
2493 default:
2494 gcc_unreachable ();
2496 return NULL;
2499 /* Estimate number of instructions that will be created by expanding EXPR.
2500 WEIGHTS contains weights attributed to various constructs. */
2503 estimate_num_insns (tree expr, eni_weights *weights)
2505 struct pointer_set_t *visited_nodes;
2506 basic_block bb;
2507 block_stmt_iterator bsi;
2508 struct function *my_function;
2509 struct eni_data data;
2511 data.count = 0;
2512 data.weights = weights;
2514 /* If we're given an entire function, walk the CFG. */
2515 if (TREE_CODE (expr) == FUNCTION_DECL)
2517 my_function = DECL_STRUCT_FUNCTION (expr);
2518 gcc_assert (my_function && my_function->cfg);
2519 visited_nodes = pointer_set_create ();
2520 FOR_EACH_BB_FN (bb, my_function)
2522 for (bsi = bsi_start (bb);
2523 !bsi_end_p (bsi);
2524 bsi_next (&bsi))
2526 walk_tree (bsi_stmt_ptr (bsi), estimate_num_insns_1,
2527 &data, visited_nodes);
2530 pointer_set_destroy (visited_nodes);
2532 else
2533 walk_tree_without_duplicates (&expr, estimate_num_insns_1, &data);
2535 return data.count;
2538 /* Initializes weights used by estimate_num_insns. */
2540 void
2541 init_inline_once (void)
2543 eni_inlining_weights.call_cost = PARAM_VALUE (PARAM_INLINE_CALL_COST);
2544 eni_inlining_weights.target_builtin_call_cost = 1;
2545 eni_inlining_weights.div_mod_cost = 10;
2546 eni_inlining_weights.omp_cost = 40;
2548 eni_size_weights.call_cost = 1;
2549 eni_size_weights.target_builtin_call_cost = 1;
2550 eni_size_weights.div_mod_cost = 1;
2551 eni_size_weights.omp_cost = 40;
2553 /* Estimating time for call is difficult, since we have no idea what the
2554 called function does. In the current uses of eni_time_weights,
2555 underestimating the cost does less harm than overestimating it, so
2556 we choose a rather small value here. */
2557 eni_time_weights.call_cost = 10;
2558 eni_time_weights.target_builtin_call_cost = 10;
2559 eni_time_weights.div_mod_cost = 10;
2560 eni_time_weights.omp_cost = 40;
2563 /* Install new lexical TREE_BLOCK underneath 'current_block'. */
2564 static void
2565 add_lexical_block (tree current_block, tree new_block)
2567 tree *blk_p;
2569 /* Walk to the last sub-block. */
2570 for (blk_p = &BLOCK_SUBBLOCKS (current_block);
2571 *blk_p;
2572 blk_p = &BLOCK_CHAIN (*blk_p))
2574 *blk_p = new_block;
2575 BLOCK_SUPERCONTEXT (new_block) = current_block;
2578 /* If *TP is a CALL_EXPR, replace it with its inline expansion. */
2580 static bool
2581 expand_call_inline (basic_block bb, tree stmt, tree *tp, void *data)
2583 copy_body_data *id;
2584 tree t;
2585 tree retvar, use_retvar;
2586 tree fn;
2587 struct pointer_map_t *st;
2588 tree return_slot;
2589 tree modify_dest;
2590 location_t saved_location;
2591 struct cgraph_edge *cg_edge;
2592 const char *reason;
2593 basic_block return_block;
2594 edge e;
2595 block_stmt_iterator bsi, stmt_bsi;
2596 bool successfully_inlined = FALSE;
2597 bool purge_dead_abnormal_edges;
2598 tree t_step;
2599 tree var;
2601 /* See what we've got. */
2602 id = (copy_body_data *) data;
2603 t = *tp;
2605 /* Set input_location here so we get the right instantiation context
2606 if we call instantiate_decl from inlinable_function_p. */
2607 saved_location = input_location;
2608 if (EXPR_HAS_LOCATION (t))
2609 input_location = EXPR_LOCATION (t);
2611 /* From here on, we're only interested in CALL_EXPRs. */
2612 if (TREE_CODE (t) != CALL_EXPR)
2613 goto egress;
2615 /* First, see if we can figure out what function is being called.
2616 If we cannot, then there is no hope of inlining the function. */
2617 fn = get_callee_fndecl (t);
2618 if (!fn)
2619 goto egress;
2621 /* Turn forward declarations into real ones. */
2622 fn = cgraph_node (fn)->decl;
2624 /* If fn is a declaration of a function in a nested scope that was
2625 globally declared inline, we don't set its DECL_INITIAL.
2626 However, we can't blindly follow DECL_ABSTRACT_ORIGIN because the
2627 C++ front-end uses it for cdtors to refer to their internal
2628 declarations, that are not real functions. Fortunately those
2629 don't have trees to be saved, so we can tell by checking their
2630 DECL_SAVED_TREE. */
2631 if (! DECL_INITIAL (fn)
2632 && DECL_ABSTRACT_ORIGIN (fn)
2633 && DECL_SAVED_TREE (DECL_ABSTRACT_ORIGIN (fn)))
2634 fn = DECL_ABSTRACT_ORIGIN (fn);
2636 /* Objective C and fortran still calls tree_rest_of_compilation directly.
2637 Kill this check once this is fixed. */
2638 if (!id->dst_node->analyzed)
2639 goto egress;
2641 cg_edge = cgraph_edge (id->dst_node, stmt);
2643 /* Constant propagation on argument done during previous inlining
2644 may create new direct call. Produce an edge for it. */
2645 if (!cg_edge)
2647 struct cgraph_node *dest = cgraph_node (fn);
2649 /* We have missing edge in the callgraph. This can happen in one case
2650 where previous inlining turned indirect call into direct call by
2651 constant propagating arguments. In all other cases we hit a bug
2652 (incorrect node sharing is most common reason for missing edges. */
2653 gcc_assert (dest->needed || !flag_unit_at_a_time);
2654 cgraph_create_edge (id->dst_node, dest, stmt,
2655 bb->count, CGRAPH_FREQ_BASE,
2656 bb->loop_depth)->inline_failed
2657 = N_("originally indirect function call not considered for inlining");
2658 if (dump_file)
2660 fprintf (dump_file, "Created new direct edge to %s",
2661 cgraph_node_name (dest));
2663 goto egress;
2666 /* Don't try to inline functions that are not well-suited to
2667 inlining. */
2668 if (!cgraph_inline_p (cg_edge, &reason))
2670 if (lookup_attribute ("always_inline", DECL_ATTRIBUTES (fn))
2671 /* Avoid warnings during early inline pass. */
2672 && (!flag_unit_at_a_time || cgraph_global_info_ready))
2674 sorry ("inlining failed in call to %q+F: %s", fn, reason);
2675 sorry ("called from here");
2677 else if (warn_inline && DECL_DECLARED_INLINE_P (fn)
2678 && !DECL_IN_SYSTEM_HEADER (fn)
2679 && strlen (reason)
2680 && !lookup_attribute ("noinline", DECL_ATTRIBUTES (fn))
2681 /* Avoid warnings during early inline pass. */
2682 && (!flag_unit_at_a_time || cgraph_global_info_ready))
2684 warning (OPT_Winline, "inlining failed in call to %q+F: %s",
2685 fn, reason);
2686 warning (OPT_Winline, "called from here");
2688 goto egress;
2690 fn = cg_edge->callee->decl;
2692 #ifdef ENABLE_CHECKING
2693 if (cg_edge->callee->decl != id->dst_node->decl)
2694 verify_cgraph_node (cg_edge->callee);
2695 #endif
2697 /* We will be inlining this callee. */
2698 id->eh_region = lookup_stmt_eh_region (stmt);
2700 /* Split the block holding the CALL_EXPR. */
2701 e = split_block (bb, stmt);
2702 bb = e->src;
2703 return_block = e->dest;
2704 remove_edge (e);
2706 /* split_block splits after the statement; work around this by
2707 moving the call into the second block manually. Not pretty,
2708 but seems easier than doing the CFG manipulation by hand
2709 when the CALL_EXPR is in the last statement of BB. */
2710 stmt_bsi = bsi_last (bb);
2711 bsi_remove (&stmt_bsi, false);
2713 /* If the CALL_EXPR was in the last statement of BB, it may have
2714 been the source of abnormal edges. In this case, schedule
2715 the removal of dead abnormal edges. */
2716 bsi = bsi_start (return_block);
2717 if (bsi_end_p (bsi))
2719 bsi_insert_after (&bsi, stmt, BSI_NEW_STMT);
2720 purge_dead_abnormal_edges = true;
2722 else
2724 bsi_insert_before (&bsi, stmt, BSI_NEW_STMT);
2725 purge_dead_abnormal_edges = false;
2728 stmt_bsi = bsi_start (return_block);
2730 /* Build a block containing code to initialize the arguments, the
2731 actual inline expansion of the body, and a label for the return
2732 statements within the function to jump to. The type of the
2733 statement expression is the return type of the function call. */
2734 id->block = make_node (BLOCK);
2735 BLOCK_ABSTRACT_ORIGIN (id->block) = fn;
2736 BLOCK_SOURCE_LOCATION (id->block) = input_location;
2737 add_lexical_block (TREE_BLOCK (stmt), id->block);
2739 /* Local declarations will be replaced by their equivalents in this
2740 map. */
2741 st = id->decl_map;
2742 id->decl_map = pointer_map_create ();
2744 /* Record the function we are about to inline. */
2745 id->src_fn = fn;
2746 id->src_node = cg_edge->callee;
2747 id->src_cfun = DECL_STRUCT_FUNCTION (fn);
2748 id->call_expr = t;
2750 gcc_assert (!id->src_cfun->after_inlining);
2752 id->entry_bb = bb;
2753 initialize_inlined_parameters (id, t, fn, bb);
2755 if (DECL_INITIAL (fn))
2756 add_lexical_block (id->block, remap_blocks (DECL_INITIAL (fn), id));
2758 /* Return statements in the function body will be replaced by jumps
2759 to the RET_LABEL. */
2761 gcc_assert (DECL_INITIAL (fn));
2762 gcc_assert (TREE_CODE (DECL_INITIAL (fn)) == BLOCK);
2764 /* Find the lhs to which the result of this call is assigned. */
2765 return_slot = NULL;
2766 if (TREE_CODE (stmt) == GIMPLE_MODIFY_STMT)
2768 modify_dest = GIMPLE_STMT_OPERAND (stmt, 0);
2770 /* The function which we are inlining might not return a value,
2771 in which case we should issue a warning that the function
2772 does not return a value. In that case the optimizers will
2773 see that the variable to which the value is assigned was not
2774 initialized. We do not want to issue a warning about that
2775 uninitialized variable. */
2776 if (DECL_P (modify_dest))
2777 TREE_NO_WARNING (modify_dest) = 1;
2778 if (CALL_EXPR_RETURN_SLOT_OPT (t))
2780 return_slot = modify_dest;
2781 modify_dest = NULL;
2784 else
2785 modify_dest = NULL;
2787 /* If we are inlining a call to the C++ operator new, we don't want
2788 to use type based alias analysis on the return value. Otherwise
2789 we may get confused if the compiler sees that the inlined new
2790 function returns a pointer which was just deleted. See bug
2791 33407. */
2792 if (DECL_IS_OPERATOR_NEW (fn))
2794 return_slot = NULL;
2795 modify_dest = NULL;
2798 /* Declare the return variable for the function. */
2799 retvar = declare_return_variable (id, return_slot,
2800 modify_dest, &use_retvar);
2802 if (DECL_IS_OPERATOR_NEW (fn))
2804 gcc_assert (TREE_CODE (retvar) == VAR_DECL
2805 && POINTER_TYPE_P (TREE_TYPE (retvar)));
2806 DECL_NO_TBAA_P (retvar) = 1;
2809 /* This is it. Duplicate the callee body. Assume callee is
2810 pre-gimplified. Note that we must not alter the caller
2811 function in any way before this point, as this CALL_EXPR may be
2812 a self-referential call; if we're calling ourselves, we need to
2813 duplicate our body before altering anything. */
2814 copy_body (id, bb->count, bb->frequency, bb, return_block);
2816 /* Add local vars in this inlined callee to caller. */
2817 t_step = id->src_cfun->unexpanded_var_list;
2818 for (; t_step; t_step = TREE_CHAIN (t_step))
2820 var = TREE_VALUE (t_step);
2821 if (TREE_STATIC (var) && !TREE_ASM_WRITTEN (var))
2822 cfun->unexpanded_var_list = tree_cons (NULL_TREE, var,
2823 cfun->unexpanded_var_list);
2824 else
2825 cfun->unexpanded_var_list = tree_cons (NULL_TREE, remap_decl (var, id),
2826 cfun->unexpanded_var_list);
2829 /* Clean up. */
2830 pointer_map_destroy (id->decl_map);
2831 id->decl_map = st;
2833 /* If the inlined function returns a result that we care about,
2834 clobber the CALL_EXPR with a reference to the return variable. */
2835 if (use_retvar && (TREE_CODE (bsi_stmt (stmt_bsi)) != CALL_EXPR))
2837 *tp = use_retvar;
2838 if (gimple_in_ssa_p (cfun))
2840 update_stmt (stmt);
2841 mark_symbols_for_renaming (stmt);
2843 maybe_clean_or_replace_eh_stmt (stmt, stmt);
2845 else
2846 /* We're modifying a TSI owned by gimple_expand_calls_inline();
2847 tsi_delink() will leave the iterator in a sane state. */
2849 /* Handle case of inlining function that miss return statement so
2850 return value becomes undefined. */
2851 if (TREE_CODE (stmt) == GIMPLE_MODIFY_STMT
2852 && TREE_CODE (GIMPLE_STMT_OPERAND (stmt, 0)) == SSA_NAME)
2854 tree name = TREE_OPERAND (stmt, 0);
2855 tree var = SSA_NAME_VAR (TREE_OPERAND (stmt, 0));
2856 tree def = gimple_default_def (cfun, var);
2858 /* If the variable is used undefined, make this name undefined via
2859 move. */
2860 if (def)
2862 TREE_OPERAND (stmt, 1) = def;
2863 update_stmt (stmt);
2865 /* Otherwise make this variable undefined. */
2866 else
2868 bsi_remove (&stmt_bsi, true);
2869 set_default_def (var, name);
2870 SSA_NAME_DEF_STMT (name) = build_empty_stmt ();
2873 else
2874 bsi_remove (&stmt_bsi, true);
2877 if (purge_dead_abnormal_edges)
2878 tree_purge_dead_abnormal_call_edges (return_block);
2880 /* If the value of the new expression is ignored, that's OK. We
2881 don't warn about this for CALL_EXPRs, so we shouldn't warn about
2882 the equivalent inlined version either. */
2883 TREE_USED (*tp) = 1;
2885 /* Output the inlining info for this abstract function, since it has been
2886 inlined. If we don't do this now, we can lose the information about the
2887 variables in the function when the blocks get blown away as soon as we
2888 remove the cgraph node. */
2889 (*debug_hooks->outlining_inline_function) (cg_edge->callee->decl);
2891 /* Update callgraph if needed. */
2892 cgraph_remove_node (cg_edge->callee);
2894 id->block = NULL_TREE;
2895 successfully_inlined = TRUE;
2897 egress:
2898 input_location = saved_location;
2899 return successfully_inlined;
2902 /* Expand call statements reachable from STMT_P.
2903 We can only have CALL_EXPRs as the "toplevel" tree code or nested
2904 in a GIMPLE_MODIFY_STMT. See tree-gimple.c:get_call_expr_in(). We can
2905 unfortunately not use that function here because we need a pointer
2906 to the CALL_EXPR, not the tree itself. */
2908 static bool
2909 gimple_expand_calls_inline (basic_block bb, copy_body_data *id)
2911 block_stmt_iterator bsi;
2913 /* Register specific tree functions. */
2914 tree_register_cfg_hooks ();
2915 for (bsi = bsi_start (bb); !bsi_end_p (bsi); bsi_next (&bsi))
2917 tree *expr_p = bsi_stmt_ptr (bsi);
2918 tree stmt = *expr_p;
2920 if (TREE_CODE (*expr_p) == GIMPLE_MODIFY_STMT)
2921 expr_p = &GIMPLE_STMT_OPERAND (*expr_p, 1);
2922 if (TREE_CODE (*expr_p) == WITH_SIZE_EXPR)
2923 expr_p = &TREE_OPERAND (*expr_p, 0);
2924 if (TREE_CODE (*expr_p) == CALL_EXPR)
2925 if (expand_call_inline (bb, stmt, expr_p, id))
2926 return true;
2928 return false;
2931 /* Walk all basic blocks created after FIRST and try to fold every statement
2932 in the STATEMENTS pointer set. */
2933 static void
2934 fold_marked_statements (int first, struct pointer_set_t *statements)
2936 for (;first < n_basic_blocks;first++)
2937 if (BASIC_BLOCK (first))
2939 block_stmt_iterator bsi;
2940 for (bsi = bsi_start (BASIC_BLOCK (first));
2941 !bsi_end_p (bsi); bsi_next (&bsi))
2942 if (pointer_set_contains (statements, bsi_stmt (bsi)))
2944 tree old_stmt = bsi_stmt (bsi);
2945 tree old_call = get_call_expr_in (old_stmt);
2947 if (fold_stmt (bsi_stmt_ptr (bsi)))
2949 update_stmt (bsi_stmt (bsi));
2950 if (old_call)
2951 cgraph_update_edges_for_call_stmt (old_stmt, old_call,
2952 bsi_stmt (bsi));
2953 if (maybe_clean_or_replace_eh_stmt (old_stmt,
2954 bsi_stmt (bsi)))
2955 tree_purge_dead_eh_edges (BASIC_BLOCK (first));
2961 /* Return true if BB has at least one abnormal outgoing edge. */
2963 static inline bool
2964 has_abnormal_outgoing_edge_p (basic_block bb)
2966 edge e;
2967 edge_iterator ei;
2969 FOR_EACH_EDGE (e, ei, bb->succs)
2970 if (e->flags & EDGE_ABNORMAL)
2971 return true;
2973 return false;
2976 /* Expand calls to inline functions in the body of FN. */
2978 unsigned int
2979 optimize_inline_calls (tree fn)
2981 copy_body_data id;
2982 tree prev_fn;
2983 basic_block bb;
2984 int last = n_basic_blocks;
2985 /* There is no point in performing inlining if errors have already
2986 occurred -- and we might crash if we try to inline invalid
2987 code. */
2988 if (errorcount || sorrycount)
2989 return 0;
2991 /* Clear out ID. */
2992 memset (&id, 0, sizeof (id));
2994 id.src_node = id.dst_node = cgraph_node (fn);
2995 id.dst_fn = fn;
2996 /* Or any functions that aren't finished yet. */
2997 prev_fn = NULL_TREE;
2998 if (current_function_decl)
3000 id.dst_fn = current_function_decl;
3001 prev_fn = current_function_decl;
3004 id.copy_decl = copy_decl_maybe_to_var;
3005 id.transform_call_graph_edges = CB_CGE_DUPLICATE;
3006 id.transform_new_cfg = false;
3007 id.transform_return_to_modify = true;
3008 id.transform_lang_insert_block = false;
3009 id.statements_to_fold = pointer_set_create ();
3011 push_gimplify_context ();
3013 /* We make no attempts to keep dominance info up-to-date. */
3014 free_dominance_info (CDI_DOMINATORS);
3015 free_dominance_info (CDI_POST_DOMINATORS);
3017 /* Reach the trees by walking over the CFG, and note the
3018 enclosing basic-blocks in the call edges. */
3019 /* We walk the blocks going forward, because inlined function bodies
3020 will split id->current_basic_block, and the new blocks will
3021 follow it; we'll trudge through them, processing their CALL_EXPRs
3022 along the way. */
3023 FOR_EACH_BB (bb)
3024 gimple_expand_calls_inline (bb, &id);
3026 pop_gimplify_context (NULL);
3028 #ifdef ENABLE_CHECKING
3030 struct cgraph_edge *e;
3032 verify_cgraph_node (id.dst_node);
3034 /* Double check that we inlined everything we are supposed to inline. */
3035 for (e = id.dst_node->callees; e; e = e->next_callee)
3036 gcc_assert (e->inline_failed);
3038 #endif
3040 /* Fold the statements before compacting/renumbering the basic blocks. */
3041 fold_marked_statements (last, id.statements_to_fold);
3042 pointer_set_destroy (id.statements_to_fold);
3044 /* Renumber the (code) basic_blocks consecutively. */
3045 compact_blocks ();
3046 /* Renumber the lexical scoping (non-code) blocks consecutively. */
3047 number_blocks (fn);
3049 /* We are not going to maintain the cgraph edges up to date.
3050 Kill it so it won't confuse us. */
3051 cgraph_node_remove_callees (id.dst_node);
3053 fold_cond_expr_cond ();
3054 /* It would be nice to check SSA/CFG/statement consistency here, but it is
3055 not possible yet - the IPA passes might make various functions to not
3056 throw and they don't care to proactively update local EH info. This is
3057 done later in fixup_cfg pass that also execute the verification. */
3058 return (TODO_update_ssa | TODO_cleanup_cfg
3059 | (gimple_in_ssa_p (cfun) ? TODO_remove_unused_locals : 0)
3060 | (profile_status != PROFILE_ABSENT ? TODO_rebuild_frequencies : 0));
3063 /* FN is a function that has a complete body, and CLONE is a function whose
3064 body is to be set to a copy of FN, mapping argument declarations according
3065 to the ARG_MAP splay_tree. */
3067 void
3068 clone_body (tree clone, tree fn, void *arg_map)
3070 copy_body_data id;
3072 /* Clone the body, as if we were making an inline call. But, remap the
3073 parameters in the callee to the parameters of caller. */
3074 memset (&id, 0, sizeof (id));
3075 id.src_fn = fn;
3076 id.dst_fn = clone;
3077 id.src_cfun = DECL_STRUCT_FUNCTION (fn);
3078 id.decl_map = (struct pointer_map_t *)arg_map;
3080 id.copy_decl = copy_decl_no_change;
3081 id.transform_call_graph_edges = CB_CGE_DUPLICATE;
3082 id.transform_new_cfg = true;
3083 id.transform_return_to_modify = false;
3084 id.transform_lang_insert_block = true;
3086 /* We're not inside any EH region. */
3087 id.eh_region = -1;
3089 /* Actually copy the body. */
3090 append_to_statement_list_force (copy_generic_body (&id), &DECL_SAVED_TREE (clone));
3093 /* Passed to walk_tree. Copies the node pointed to, if appropriate. */
3095 tree
3096 copy_tree_r (tree *tp, int *walk_subtrees, void *data ATTRIBUTE_UNUSED)
3098 enum tree_code code = TREE_CODE (*tp);
3099 enum tree_code_class cl = TREE_CODE_CLASS (code);
3101 /* We make copies of most nodes. */
3102 if (IS_EXPR_CODE_CLASS (cl)
3103 || IS_GIMPLE_STMT_CODE_CLASS (cl)
3104 || code == TREE_LIST
3105 || code == TREE_VEC
3106 || code == TYPE_DECL
3107 || code == OMP_CLAUSE)
3109 /* Because the chain gets clobbered when we make a copy, we save it
3110 here. */
3111 tree chain = NULL_TREE, new;
3113 if (!GIMPLE_TUPLE_P (*tp))
3114 chain = TREE_CHAIN (*tp);
3116 /* Copy the node. */
3117 new = copy_node (*tp);
3119 /* Propagate mudflap marked-ness. */
3120 if (flag_mudflap && mf_marked_p (*tp))
3121 mf_mark (new);
3123 *tp = new;
3125 /* Now, restore the chain, if appropriate. That will cause
3126 walk_tree to walk into the chain as well. */
3127 if (code == PARM_DECL
3128 || code == TREE_LIST
3129 || code == OMP_CLAUSE)
3130 TREE_CHAIN (*tp) = chain;
3132 /* For now, we don't update BLOCKs when we make copies. So, we
3133 have to nullify all BIND_EXPRs. */
3134 if (TREE_CODE (*tp) == BIND_EXPR)
3135 BIND_EXPR_BLOCK (*tp) = NULL_TREE;
3137 else if (code == CONSTRUCTOR)
3139 /* CONSTRUCTOR nodes need special handling because
3140 we need to duplicate the vector of elements. */
3141 tree new;
3143 new = copy_node (*tp);
3145 /* Propagate mudflap marked-ness. */
3146 if (flag_mudflap && mf_marked_p (*tp))
3147 mf_mark (new);
3149 CONSTRUCTOR_ELTS (new) = VEC_copy (constructor_elt, gc,
3150 CONSTRUCTOR_ELTS (*tp));
3151 *tp = new;
3153 else if (TREE_CODE_CLASS (code) == tcc_type)
3154 *walk_subtrees = 0;
3155 else if (TREE_CODE_CLASS (code) == tcc_declaration)
3156 *walk_subtrees = 0;
3157 else if (TREE_CODE_CLASS (code) == tcc_constant)
3158 *walk_subtrees = 0;
3159 else
3160 gcc_assert (code != STATEMENT_LIST);
3161 return NULL_TREE;
3164 /* The SAVE_EXPR pointed to by TP is being copied. If ST contains
3165 information indicating to what new SAVE_EXPR this one should be mapped,
3166 use that one. Otherwise, create a new node and enter it in ST. FN is
3167 the function into which the copy will be placed. */
3169 static void
3170 remap_save_expr (tree *tp, void *st_, int *walk_subtrees)
3172 struct pointer_map_t *st = (struct pointer_map_t *) st_;
3173 tree *n;
3174 tree t;
3176 /* See if we already encountered this SAVE_EXPR. */
3177 n = (tree *) pointer_map_contains (st, *tp);
3179 /* If we didn't already remap this SAVE_EXPR, do so now. */
3180 if (!n)
3182 t = copy_node (*tp);
3184 /* Remember this SAVE_EXPR. */
3185 *pointer_map_insert (st, *tp) = t;
3186 /* Make sure we don't remap an already-remapped SAVE_EXPR. */
3187 *pointer_map_insert (st, t) = t;
3189 else
3191 /* We've already walked into this SAVE_EXPR; don't do it again. */
3192 *walk_subtrees = 0;
3193 t = *n;
3196 /* Replace this SAVE_EXPR with the copy. */
3197 *tp = t;
3200 /* Called via walk_tree. If *TP points to a DECL_STMT for a local label,
3201 copies the declaration and enters it in the splay_tree in DATA (which is
3202 really an `copy_body_data *'). */
3204 static tree
3205 mark_local_for_remap_r (tree *tp, int *walk_subtrees ATTRIBUTE_UNUSED,
3206 void *data)
3208 copy_body_data *id = (copy_body_data *) data;
3210 /* Don't walk into types. */
3211 if (TYPE_P (*tp))
3212 *walk_subtrees = 0;
3214 else if (TREE_CODE (*tp) == LABEL_EXPR)
3216 tree decl = TREE_OPERAND (*tp, 0);
3218 /* Copy the decl and remember the copy. */
3219 insert_decl_map (id, decl, id->copy_decl (decl, id));
3222 return NULL_TREE;
3225 /* Perform any modifications to EXPR required when it is unsaved. Does
3226 not recurse into EXPR's subtrees. */
3228 static void
3229 unsave_expr_1 (tree expr)
3231 switch (TREE_CODE (expr))
3233 case TARGET_EXPR:
3234 /* Don't mess with a TARGET_EXPR that hasn't been expanded.
3235 It's OK for this to happen if it was part of a subtree that
3236 isn't immediately expanded, such as operand 2 of another
3237 TARGET_EXPR. */
3238 if (TREE_OPERAND (expr, 1))
3239 break;
3241 TREE_OPERAND (expr, 1) = TREE_OPERAND (expr, 3);
3242 TREE_OPERAND (expr, 3) = NULL_TREE;
3243 break;
3245 default:
3246 break;
3250 /* Called via walk_tree when an expression is unsaved. Using the
3251 splay_tree pointed to by ST (which is really a `splay_tree'),
3252 remaps all local declarations to appropriate replacements. */
3254 static tree
3255 unsave_r (tree *tp, int *walk_subtrees, void *data)
3257 copy_body_data *id = (copy_body_data *) data;
3258 struct pointer_map_t *st = id->decl_map;
3259 tree *n;
3261 /* Only a local declaration (variable or label). */
3262 if ((TREE_CODE (*tp) == VAR_DECL && !TREE_STATIC (*tp))
3263 || TREE_CODE (*tp) == LABEL_DECL)
3265 /* Lookup the declaration. */
3266 n = (tree *) pointer_map_contains (st, *tp);
3268 /* If it's there, remap it. */
3269 if (n)
3270 *tp = *n;
3273 else if (TREE_CODE (*tp) == STATEMENT_LIST)
3274 copy_statement_list (tp);
3275 else if (TREE_CODE (*tp) == BIND_EXPR)
3276 copy_bind_expr (tp, walk_subtrees, id);
3277 else if (TREE_CODE (*tp) == SAVE_EXPR)
3278 remap_save_expr (tp, st, walk_subtrees);
3279 else
3281 copy_tree_r (tp, walk_subtrees, NULL);
3283 /* Do whatever unsaving is required. */
3284 unsave_expr_1 (*tp);
3287 /* Keep iterating. */
3288 return NULL_TREE;
3291 /* Copies everything in EXPR and replaces variables, labels
3292 and SAVE_EXPRs local to EXPR. */
3294 tree
3295 unsave_expr_now (tree expr)
3297 copy_body_data id;
3299 /* There's nothing to do for NULL_TREE. */
3300 if (expr == 0)
3301 return expr;
3303 /* Set up ID. */
3304 memset (&id, 0, sizeof (id));
3305 id.src_fn = current_function_decl;
3306 id.dst_fn = current_function_decl;
3307 id.decl_map = pointer_map_create ();
3309 id.copy_decl = copy_decl_no_change;
3310 id.transform_call_graph_edges = CB_CGE_DUPLICATE;
3311 id.transform_new_cfg = false;
3312 id.transform_return_to_modify = false;
3313 id.transform_lang_insert_block = false;
3315 /* Walk the tree once to find local labels. */
3316 walk_tree_without_duplicates (&expr, mark_local_for_remap_r, &id);
3318 /* Walk the tree again, copying, remapping, and unsaving. */
3319 walk_tree (&expr, unsave_r, &id, NULL);
3321 /* Clean up. */
3322 pointer_map_destroy (id.decl_map);
3324 return expr;
3327 /* Allow someone to determine if SEARCH is a child of TOP from gdb. */
3329 static tree
3330 debug_find_tree_1 (tree *tp, int *walk_subtrees ATTRIBUTE_UNUSED, void *data)
3332 if (*tp == data)
3333 return (tree) data;
3334 else
3335 return NULL;
3338 bool
3339 debug_find_tree (tree top, tree search)
3341 return walk_tree_without_duplicates (&top, debug_find_tree_1, search) != 0;
3345 /* Declare the variables created by the inliner. Add all the variables in
3346 VARS to BIND_EXPR. */
3348 static void
3349 declare_inline_vars (tree block, tree vars)
3351 tree t;
3352 for (t = vars; t; t = TREE_CHAIN (t))
3354 DECL_SEEN_IN_BIND_EXPR_P (t) = 1;
3355 gcc_assert (!TREE_STATIC (t) && !TREE_ASM_WRITTEN (t));
3356 cfun->unexpanded_var_list =
3357 tree_cons (NULL_TREE, t,
3358 cfun->unexpanded_var_list);
3361 if (block)
3362 BLOCK_VARS (block) = chainon (BLOCK_VARS (block), vars);
3366 /* Copy NODE (which must be a DECL). The DECL originally was in the FROM_FN,
3367 but now it will be in the TO_FN. PARM_TO_VAR means enable PARM_DECL to
3368 VAR_DECL translation. */
3370 static tree
3371 copy_decl_for_dup_finish (copy_body_data *id, tree decl, tree copy)
3373 /* Don't generate debug information for the copy if we wouldn't have
3374 generated it for the copy either. */
3375 DECL_ARTIFICIAL (copy) = DECL_ARTIFICIAL (decl);
3376 DECL_IGNORED_P (copy) = DECL_IGNORED_P (decl);
3378 /* Set the DECL_ABSTRACT_ORIGIN so the debugging routines know what
3379 declaration inspired this copy. */
3380 DECL_ABSTRACT_ORIGIN (copy) = DECL_ORIGIN (decl);
3382 /* The new variable/label has no RTL, yet. */
3383 if (CODE_CONTAINS_STRUCT (TREE_CODE (copy), TS_DECL_WRTL)
3384 && !TREE_STATIC (copy) && !DECL_EXTERNAL (copy))
3385 SET_DECL_RTL (copy, NULL_RTX);
3387 /* These args would always appear unused, if not for this. */
3388 TREE_USED (copy) = 1;
3390 /* Set the context for the new declaration. */
3391 if (!DECL_CONTEXT (decl))
3392 /* Globals stay global. */
3394 else if (DECL_CONTEXT (decl) != id->src_fn)
3395 /* Things that weren't in the scope of the function we're inlining
3396 from aren't in the scope we're inlining to, either. */
3398 else if (TREE_STATIC (decl))
3399 /* Function-scoped static variables should stay in the original
3400 function. */
3402 else
3403 /* Ordinary automatic local variables are now in the scope of the
3404 new function. */
3405 DECL_CONTEXT (copy) = id->dst_fn;
3407 return copy;
3410 static tree
3411 copy_decl_to_var (tree decl, copy_body_data *id)
3413 tree copy, type;
3415 gcc_assert (TREE_CODE (decl) == PARM_DECL
3416 || TREE_CODE (decl) == RESULT_DECL);
3418 type = TREE_TYPE (decl);
3420 copy = build_decl (VAR_DECL, DECL_NAME (decl), type);
3421 TREE_ADDRESSABLE (copy) = TREE_ADDRESSABLE (decl);
3422 TREE_READONLY (copy) = TREE_READONLY (decl);
3423 TREE_THIS_VOLATILE (copy) = TREE_THIS_VOLATILE (decl);
3424 DECL_GIMPLE_REG_P (copy) = DECL_GIMPLE_REG_P (decl);
3425 DECL_NO_TBAA_P (copy) = DECL_NO_TBAA_P (decl);
3427 return copy_decl_for_dup_finish (id, decl, copy);
3430 /* Like copy_decl_to_var, but create a return slot object instead of a
3431 pointer variable for return by invisible reference. */
3433 static tree
3434 copy_result_decl_to_var (tree decl, copy_body_data *id)
3436 tree copy, type;
3438 gcc_assert (TREE_CODE (decl) == PARM_DECL
3439 || TREE_CODE (decl) == RESULT_DECL);
3441 type = TREE_TYPE (decl);
3442 if (DECL_BY_REFERENCE (decl))
3443 type = TREE_TYPE (type);
3445 copy = build_decl (VAR_DECL, DECL_NAME (decl), type);
3446 TREE_READONLY (copy) = TREE_READONLY (decl);
3447 TREE_THIS_VOLATILE (copy) = TREE_THIS_VOLATILE (decl);
3448 if (!DECL_BY_REFERENCE (decl))
3450 TREE_ADDRESSABLE (copy) = TREE_ADDRESSABLE (decl);
3451 DECL_GIMPLE_REG_P (copy) = DECL_GIMPLE_REG_P (decl);
3452 DECL_NO_TBAA_P (copy) = DECL_NO_TBAA_P (decl);
3455 return copy_decl_for_dup_finish (id, decl, copy);
3459 static tree
3460 copy_decl_no_change (tree decl, copy_body_data *id)
3462 tree copy;
3464 copy = copy_node (decl);
3466 /* The COPY is not abstract; it will be generated in DST_FN. */
3467 DECL_ABSTRACT (copy) = 0;
3468 lang_hooks.dup_lang_specific_decl (copy);
3470 /* TREE_ADDRESSABLE isn't used to indicate that a label's address has
3471 been taken; it's for internal bookkeeping in expand_goto_internal. */
3472 if (TREE_CODE (copy) == LABEL_DECL)
3474 TREE_ADDRESSABLE (copy) = 0;
3475 LABEL_DECL_UID (copy) = -1;
3478 return copy_decl_for_dup_finish (id, decl, copy);
3481 static tree
3482 copy_decl_maybe_to_var (tree decl, copy_body_data *id)
3484 if (TREE_CODE (decl) == PARM_DECL || TREE_CODE (decl) == RESULT_DECL)
3485 return copy_decl_to_var (decl, id);
3486 else
3487 return copy_decl_no_change (decl, id);
3490 /* Return a copy of the function's argument tree. */
3491 static tree
3492 copy_arguments_for_versioning (tree orig_parm, copy_body_data * id)
3494 tree *arg_copy, *parg;
3496 arg_copy = &orig_parm;
3497 for (parg = arg_copy; *parg; parg = &TREE_CHAIN (*parg))
3499 tree new = remap_decl (*parg, id);
3500 lang_hooks.dup_lang_specific_decl (new);
3501 TREE_CHAIN (new) = TREE_CHAIN (*parg);
3502 *parg = new;
3504 return orig_parm;
3507 /* Return a copy of the function's static chain. */
3508 static tree
3509 copy_static_chain (tree static_chain, copy_body_data * id)
3511 tree *chain_copy, *pvar;
3513 chain_copy = &static_chain;
3514 for (pvar = chain_copy; *pvar; pvar = &TREE_CHAIN (*pvar))
3516 tree new = remap_decl (*pvar, id);
3517 lang_hooks.dup_lang_specific_decl (new);
3518 TREE_CHAIN (new) = TREE_CHAIN (*pvar);
3519 *pvar = new;
3521 return static_chain;
3524 /* Return true if the function is allowed to be versioned.
3525 This is a guard for the versioning functionality. */
3526 bool
3527 tree_versionable_function_p (tree fndecl)
3529 if (fndecl == NULL_TREE)
3530 return false;
3531 /* ??? There are cases where a function is
3532 uninlinable but can be versioned. */
3533 if (!tree_inlinable_function_p (fndecl))
3534 return false;
3536 return true;
3539 /* Create a copy of a function's tree.
3540 OLD_DECL and NEW_DECL are FUNCTION_DECL tree nodes
3541 of the original function and the new copied function
3542 respectively. In case we want to replace a DECL
3543 tree with another tree while duplicating the function's
3544 body, TREE_MAP represents the mapping between these
3545 trees. If UPDATE_CLONES is set, the call_stmt fields
3546 of edges of clones of the function will be updated. */
3547 void
3548 tree_function_versioning (tree old_decl, tree new_decl, varray_type tree_map,
3549 bool update_clones)
3551 struct cgraph_node *old_version_node;
3552 struct cgraph_node *new_version_node;
3553 copy_body_data id;
3554 tree p;
3555 unsigned i;
3556 struct ipa_replace_map *replace_info;
3557 basic_block old_entry_block;
3558 tree t_step;
3559 tree old_current_function_decl = current_function_decl;
3561 gcc_assert (TREE_CODE (old_decl) == FUNCTION_DECL
3562 && TREE_CODE (new_decl) == FUNCTION_DECL);
3563 DECL_POSSIBLY_INLINED (old_decl) = 1;
3565 old_version_node = cgraph_node (old_decl);
3566 new_version_node = cgraph_node (new_decl);
3568 DECL_ARTIFICIAL (new_decl) = 1;
3569 DECL_ABSTRACT_ORIGIN (new_decl) = DECL_ORIGIN (old_decl);
3571 /* Prepare the data structures for the tree copy. */
3572 memset (&id, 0, sizeof (id));
3574 /* Generate a new name for the new version. */
3575 if (!update_clones)
3577 DECL_NAME (new_decl) = create_tmp_var_name (NULL);
3578 SET_DECL_ASSEMBLER_NAME (new_decl, DECL_NAME (new_decl));
3579 SET_DECL_RTL (new_decl, NULL_RTX);
3580 id.statements_to_fold = pointer_set_create ();
3583 id.decl_map = pointer_map_create ();
3584 id.src_fn = old_decl;
3585 id.dst_fn = new_decl;
3586 id.src_node = old_version_node;
3587 id.dst_node = new_version_node;
3588 id.src_cfun = DECL_STRUCT_FUNCTION (old_decl);
3590 id.copy_decl = copy_decl_no_change;
3591 id.transform_call_graph_edges
3592 = update_clones ? CB_CGE_MOVE_CLONES : CB_CGE_MOVE;
3593 id.transform_new_cfg = true;
3594 id.transform_return_to_modify = false;
3595 id.transform_lang_insert_block = false;
3597 current_function_decl = new_decl;
3598 old_entry_block = ENTRY_BLOCK_PTR_FOR_FUNCTION
3599 (DECL_STRUCT_FUNCTION (old_decl));
3600 initialize_cfun (new_decl, old_decl,
3601 old_entry_block->count,
3602 old_entry_block->frequency);
3603 push_cfun (DECL_STRUCT_FUNCTION (new_decl));
3605 /* Copy the function's static chain. */
3606 p = DECL_STRUCT_FUNCTION (old_decl)->static_chain_decl;
3607 if (p)
3608 DECL_STRUCT_FUNCTION (new_decl)->static_chain_decl =
3609 copy_static_chain (DECL_STRUCT_FUNCTION (old_decl)->static_chain_decl,
3610 &id);
3611 /* Copy the function's arguments. */
3612 if (DECL_ARGUMENTS (old_decl) != NULL_TREE)
3613 DECL_ARGUMENTS (new_decl) =
3614 copy_arguments_for_versioning (DECL_ARGUMENTS (old_decl), &id);
3616 /* If there's a tree_map, prepare for substitution. */
3617 if (tree_map)
3618 for (i = 0; i < VARRAY_ACTIVE_SIZE (tree_map); i++)
3620 replace_info = VARRAY_GENERIC_PTR (tree_map, i);
3621 if (replace_info->replace_p)
3622 insert_decl_map (&id, replace_info->old_tree,
3623 replace_info->new_tree);
3626 DECL_INITIAL (new_decl) = remap_blocks (DECL_INITIAL (id.src_fn), &id);
3628 /* Renumber the lexical scoping (non-code) blocks consecutively. */
3629 number_blocks (id.dst_fn);
3631 if (DECL_STRUCT_FUNCTION (old_decl)->unexpanded_var_list != NULL_TREE)
3632 /* Add local vars. */
3633 for (t_step = DECL_STRUCT_FUNCTION (old_decl)->unexpanded_var_list;
3634 t_step; t_step = TREE_CHAIN (t_step))
3636 tree var = TREE_VALUE (t_step);
3637 if (TREE_STATIC (var) && !TREE_ASM_WRITTEN (var))
3638 cfun->unexpanded_var_list = tree_cons (NULL_TREE, var,
3639 cfun->unexpanded_var_list);
3640 else
3641 cfun->unexpanded_var_list =
3642 tree_cons (NULL_TREE, remap_decl (var, &id),
3643 cfun->unexpanded_var_list);
3646 /* Copy the Function's body. */
3647 copy_body (&id, old_entry_block->count, old_entry_block->frequency, ENTRY_BLOCK_PTR, EXIT_BLOCK_PTR);
3649 if (DECL_RESULT (old_decl) != NULL_TREE)
3651 tree *res_decl = &DECL_RESULT (old_decl);
3652 DECL_RESULT (new_decl) = remap_decl (*res_decl, &id);
3653 lang_hooks.dup_lang_specific_decl (DECL_RESULT (new_decl));
3656 /* Renumber the lexical scoping (non-code) blocks consecutively. */
3657 number_blocks (new_decl);
3659 /* Clean up. */
3660 pointer_map_destroy (id.decl_map);
3661 if (!update_clones)
3663 fold_marked_statements (0, id.statements_to_fold);
3664 pointer_set_destroy (id.statements_to_fold);
3665 fold_cond_expr_cond ();
3667 if (gimple_in_ssa_p (cfun))
3669 free_dominance_info (CDI_DOMINATORS);
3670 free_dominance_info (CDI_POST_DOMINATORS);
3671 if (!update_clones)
3672 delete_unreachable_blocks ();
3673 update_ssa (TODO_update_ssa);
3674 if (!update_clones)
3676 fold_cond_expr_cond ();
3677 if (need_ssa_update_p ())
3678 update_ssa (TODO_update_ssa);
3681 free_dominance_info (CDI_DOMINATORS);
3682 free_dominance_info (CDI_POST_DOMINATORS);
3683 pop_cfun ();
3684 current_function_decl = old_current_function_decl;
3685 gcc_assert (!current_function_decl
3686 || DECL_STRUCT_FUNCTION (current_function_decl) == cfun);
3687 return;
3690 /* Duplicate a type, fields and all. */
3692 tree
3693 build_duplicate_type (tree type)
3695 struct copy_body_data id;
3697 memset (&id, 0, sizeof (id));
3698 id.src_fn = current_function_decl;
3699 id.dst_fn = current_function_decl;
3700 id.src_cfun = cfun;
3701 id.decl_map = pointer_map_create ();
3702 id.copy_decl = copy_decl_no_change;
3704 type = remap_type_1 (type, &id);
3706 pointer_map_destroy (id.decl_map);
3708 return type;