Add ability to set target options (ix86 only) and optimization options on a function...
[official-gcc.git] / gcc / tree-inline.c
blob9d827ff92615f2bbbd28196a4cad2de3a5bf0e26
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 bool inlinable_function_p (tree);
131 static void remap_block (tree *, copy_body_data *);
132 static tree remap_decls (tree, copy_body_data *);
133 static void copy_bind_expr (tree *, int *, copy_body_data *);
134 static tree mark_local_for_remap_r (tree *, int *, void *);
135 static void unsave_expr_1 (tree);
136 static tree unsave_r (tree *, int *, void *);
137 static void declare_inline_vars (tree, tree);
138 static void remap_save_expr (tree *, void *, int *);
139 static void add_lexical_block (tree current_block, tree new_block);
140 static tree copy_decl_to_var (tree, copy_body_data *);
141 static tree copy_result_decl_to_var (tree, copy_body_data *);
142 static tree copy_decl_maybe_to_var (tree, copy_body_data *);
144 /* Insert a tree->tree mapping for ID. Despite the name suggests
145 that the trees should be variables, it is used for more than that. */
147 void
148 insert_decl_map (copy_body_data *id, tree key, tree value)
150 *pointer_map_insert (id->decl_map, key) = value;
152 /* Always insert an identity map as well. If we see this same new
153 node again, we won't want to duplicate it a second time. */
154 if (key != value)
155 *pointer_map_insert (id->decl_map, value) = value;
158 /* Construct new SSA name for old NAME. ID is the inline context. */
160 static tree
161 remap_ssa_name (tree name, copy_body_data *id)
163 tree new;
164 tree *n;
166 gcc_assert (TREE_CODE (name) == SSA_NAME);
168 n = (tree *) pointer_map_contains (id->decl_map, name);
169 if (n)
170 return *n;
172 /* Do not set DEF_STMT yet as statement is not copied yet. We do that
173 in copy_bb. */
174 new = remap_decl (SSA_NAME_VAR (name), id);
175 /* We might've substituted constant or another SSA_NAME for
176 the variable.
178 Replace the SSA name representing RESULT_DECL by variable during
179 inlining: this saves us from need to introduce PHI node in a case
180 return value is just partly initialized. */
181 if ((TREE_CODE (new) == VAR_DECL || TREE_CODE (new) == PARM_DECL)
182 && (TREE_CODE (SSA_NAME_VAR (name)) != RESULT_DECL
183 || !id->transform_return_to_modify))
185 new = make_ssa_name (new, NULL);
186 insert_decl_map (id, name, new);
187 SSA_NAME_OCCURS_IN_ABNORMAL_PHI (new)
188 = SSA_NAME_OCCURS_IN_ABNORMAL_PHI (name);
189 TREE_TYPE (new) = TREE_TYPE (SSA_NAME_VAR (new));
190 if (IS_EMPTY_STMT (SSA_NAME_DEF_STMT (name)))
192 /* By inlining function having uninitialized variable, we might
193 extend the lifetime (variable might get reused). This cause
194 ICE in the case we end up extending lifetime of SSA name across
195 abnormal edge, but also increase register pressure.
197 We simply initialize all uninitialized vars by 0 except for case
198 we are inlining to very first BB. We can avoid this for all
199 BBs that are not withing strongly connected regions of the CFG,
200 but this is bit expensive to test.
202 if (id->entry_bb && is_gimple_reg (SSA_NAME_VAR (name))
203 && TREE_CODE (SSA_NAME_VAR (name)) != PARM_DECL
204 && (id->entry_bb != EDGE_SUCC (ENTRY_BLOCK_PTR, 0)->dest
205 || EDGE_COUNT (id->entry_bb->preds) != 1))
207 block_stmt_iterator bsi = bsi_last (id->entry_bb);
208 tree init_stmt
209 = build_gimple_modify_stmt (new,
210 fold_convert (TREE_TYPE (new),
211 integer_zero_node));
212 bsi_insert_after (&bsi, init_stmt, BSI_NEW_STMT);
213 SSA_NAME_DEF_STMT (new) = init_stmt;
214 SSA_NAME_IS_DEFAULT_DEF (new) = 0;
216 else
218 SSA_NAME_DEF_STMT (new) = build_empty_stmt ();
219 if (gimple_default_def (id->src_cfun, SSA_NAME_VAR (name)) == name)
220 set_default_def (SSA_NAME_VAR (new), new);
224 else
225 insert_decl_map (id, name, new);
226 return new;
229 /* Remap DECL during the copying of the BLOCK tree for the function. */
231 tree
232 remap_decl (tree decl, copy_body_data *id)
234 tree *n;
235 tree fn;
237 /* We only remap local variables in the current function. */
238 fn = id->src_fn;
240 /* See if we have remapped this declaration. */
242 n = (tree *) pointer_map_contains (id->decl_map, decl);
244 /* If we didn't already have an equivalent for this declaration,
245 create one now. */
246 if (!n)
248 /* Make a copy of the variable or label. */
249 tree t = id->copy_decl (decl, id);
251 /* Remember it, so that if we encounter this local entity again
252 we can reuse this copy. Do this early because remap_type may
253 need this decl for TYPE_STUB_DECL. */
254 insert_decl_map (id, decl, t);
256 if (!DECL_P (t))
257 return t;
259 /* Remap types, if necessary. */
260 TREE_TYPE (t) = remap_type (TREE_TYPE (t), id);
261 if (TREE_CODE (t) == TYPE_DECL)
262 DECL_ORIGINAL_TYPE (t) = remap_type (DECL_ORIGINAL_TYPE (t), id);
264 /* Remap sizes as necessary. */
265 walk_tree (&DECL_SIZE (t), copy_body_r, id, NULL);
266 walk_tree (&DECL_SIZE_UNIT (t), copy_body_r, id, NULL);
268 /* If fields, do likewise for offset and qualifier. */
269 if (TREE_CODE (t) == FIELD_DECL)
271 walk_tree (&DECL_FIELD_OFFSET (t), copy_body_r, id, NULL);
272 if (TREE_CODE (DECL_CONTEXT (t)) == QUAL_UNION_TYPE)
273 walk_tree (&DECL_QUALIFIER (t), copy_body_r, id, NULL);
276 if (cfun && gimple_in_ssa_p (cfun)
277 && (TREE_CODE (t) == VAR_DECL
278 || TREE_CODE (t) == RESULT_DECL || TREE_CODE (t) == PARM_DECL))
280 tree def = gimple_default_def (id->src_cfun, decl);
281 get_var_ann (t);
282 if (TREE_CODE (decl) != PARM_DECL && def)
284 tree map = remap_ssa_name (def, id);
285 /* Watch out RESULT_DECLs whose SSA names map directly
286 to them. */
287 if (TREE_CODE (map) == SSA_NAME
288 && IS_EMPTY_STMT (SSA_NAME_DEF_STMT (map)))
289 set_default_def (t, map);
291 add_referenced_var (t);
293 return t;
296 return unshare_expr (*n);
299 static tree
300 remap_type_1 (tree type, copy_body_data *id)
302 tree new, t;
304 /* We do need a copy. build and register it now. If this is a pointer or
305 reference type, remap the designated type and make a new pointer or
306 reference type. */
307 if (TREE_CODE (type) == POINTER_TYPE)
309 new = build_pointer_type_for_mode (remap_type (TREE_TYPE (type), id),
310 TYPE_MODE (type),
311 TYPE_REF_CAN_ALIAS_ALL (type));
312 insert_decl_map (id, type, new);
313 return new;
315 else if (TREE_CODE (type) == REFERENCE_TYPE)
317 new = build_reference_type_for_mode (remap_type (TREE_TYPE (type), id),
318 TYPE_MODE (type),
319 TYPE_REF_CAN_ALIAS_ALL (type));
320 insert_decl_map (id, type, new);
321 return new;
323 else
324 new = copy_node (type);
326 insert_decl_map (id, type, new);
328 /* This is a new type, not a copy of an old type. Need to reassociate
329 variants. We can handle everything except the main variant lazily. */
330 t = TYPE_MAIN_VARIANT (type);
331 if (type != t)
333 t = remap_type (t, id);
334 TYPE_MAIN_VARIANT (new) = t;
335 TYPE_NEXT_VARIANT (new) = TYPE_NEXT_VARIANT (t);
336 TYPE_NEXT_VARIANT (t) = new;
338 else
340 TYPE_MAIN_VARIANT (new) = new;
341 TYPE_NEXT_VARIANT (new) = NULL;
344 if (TYPE_STUB_DECL (type))
345 TYPE_STUB_DECL (new) = remap_decl (TYPE_STUB_DECL (type), id);
347 /* Lazily create pointer and reference types. */
348 TYPE_POINTER_TO (new) = NULL;
349 TYPE_REFERENCE_TO (new) = NULL;
351 switch (TREE_CODE (new))
353 case INTEGER_TYPE:
354 case REAL_TYPE:
355 case FIXED_POINT_TYPE:
356 case ENUMERAL_TYPE:
357 case BOOLEAN_TYPE:
358 t = TYPE_MIN_VALUE (new);
359 if (t && TREE_CODE (t) != INTEGER_CST)
360 walk_tree (&TYPE_MIN_VALUE (new), copy_body_r, id, NULL);
362 t = TYPE_MAX_VALUE (new);
363 if (t && TREE_CODE (t) != INTEGER_CST)
364 walk_tree (&TYPE_MAX_VALUE (new), copy_body_r, id, NULL);
365 return new;
367 case FUNCTION_TYPE:
368 TREE_TYPE (new) = remap_type (TREE_TYPE (new), id);
369 walk_tree (&TYPE_ARG_TYPES (new), copy_body_r, id, NULL);
370 return new;
372 case ARRAY_TYPE:
373 TREE_TYPE (new) = remap_type (TREE_TYPE (new), id);
374 TYPE_DOMAIN (new) = remap_type (TYPE_DOMAIN (new), id);
375 break;
377 case RECORD_TYPE:
378 case UNION_TYPE:
379 case QUAL_UNION_TYPE:
381 tree f, nf = NULL;
383 for (f = TYPE_FIELDS (new); f ; f = TREE_CHAIN (f))
385 t = remap_decl (f, id);
386 DECL_CONTEXT (t) = new;
387 TREE_CHAIN (t) = nf;
388 nf = t;
390 TYPE_FIELDS (new) = nreverse (nf);
392 break;
394 case OFFSET_TYPE:
395 default:
396 /* Shouldn't have been thought variable sized. */
397 gcc_unreachable ();
400 walk_tree (&TYPE_SIZE (new), copy_body_r, id, NULL);
401 walk_tree (&TYPE_SIZE_UNIT (new), copy_body_r, id, NULL);
403 return new;
406 tree
407 remap_type (tree type, copy_body_data *id)
409 tree *node;
410 tree tmp;
412 if (type == NULL)
413 return type;
415 /* See if we have remapped this type. */
416 node = (tree *) pointer_map_contains (id->decl_map, type);
417 if (node)
418 return *node;
420 /* The type only needs remapping if it's variably modified. */
421 if (! variably_modified_type_p (type, id->src_fn))
423 insert_decl_map (id, type, type);
424 return type;
427 id->remapping_type_depth++;
428 tmp = remap_type_1 (type, id);
429 id->remapping_type_depth--;
431 return tmp;
434 static tree
435 remap_decls (tree decls, copy_body_data *id)
437 tree old_var;
438 tree new_decls = NULL_TREE;
440 /* Remap its variables. */
441 for (old_var = decls; old_var; old_var = TREE_CHAIN (old_var))
443 tree new_var;
445 /* We can not chain the local static declarations into the local_decls
446 as we can't duplicate them or break one decl rule. Go ahead and link
447 them into local_decls. */
448 if (!auto_var_in_fn_p (old_var, id->src_fn)
449 && !DECL_EXTERNAL (old_var))
451 cfun->local_decls = tree_cons (NULL_TREE, old_var,
452 cfun->local_decls);
453 continue;
456 /* Remap the variable. */
457 new_var = remap_decl (old_var, id);
459 /* If we didn't remap this variable, so we can't mess with its
460 TREE_CHAIN. If we remapped this variable to the return slot, it's
461 already declared somewhere else, so don't declare it here. */
462 if (!new_var || new_var == id->retvar)
464 else
466 gcc_assert (DECL_P (new_var));
467 TREE_CHAIN (new_var) = new_decls;
468 new_decls = new_var;
472 return nreverse (new_decls);
475 /* Copy the BLOCK to contain remapped versions of the variables
476 therein. And hook the new block into the block-tree. */
478 static void
479 remap_block (tree *block, copy_body_data *id)
481 tree old_block;
482 tree new_block;
483 tree fn;
485 /* Make the new block. */
486 old_block = *block;
487 new_block = make_node (BLOCK);
488 TREE_USED (new_block) = TREE_USED (old_block);
489 BLOCK_ABSTRACT_ORIGIN (new_block) = old_block;
490 BLOCK_SOURCE_LOCATION (new_block) = BLOCK_SOURCE_LOCATION (old_block);
491 *block = new_block;
493 /* Remap its variables. */
494 BLOCK_VARS (new_block) = remap_decls (BLOCK_VARS (old_block), id);
496 fn = id->dst_fn;
498 if (id->transform_lang_insert_block)
499 id->transform_lang_insert_block (new_block);
501 /* Remember the remapped block. */
502 insert_decl_map (id, old_block, new_block);
505 /* Copy the whole block tree and root it in id->block. */
506 static tree
507 remap_blocks (tree block, copy_body_data *id)
509 tree t;
510 tree new = block;
512 if (!block)
513 return NULL;
515 remap_block (&new, id);
516 gcc_assert (new != block);
517 for (t = BLOCK_SUBBLOCKS (block); t ; t = BLOCK_CHAIN (t))
518 add_lexical_block (new, remap_blocks (t, id));
519 return new;
522 static void
523 copy_statement_list (tree *tp)
525 tree_stmt_iterator oi, ni;
526 tree new;
528 new = alloc_stmt_list ();
529 ni = tsi_start (new);
530 oi = tsi_start (*tp);
531 *tp = new;
533 for (; !tsi_end_p (oi); tsi_next (&oi))
534 tsi_link_after (&ni, tsi_stmt (oi), TSI_NEW_STMT);
537 static void
538 copy_bind_expr (tree *tp, int *walk_subtrees, copy_body_data *id)
540 tree block = BIND_EXPR_BLOCK (*tp);
541 /* Copy (and replace) the statement. */
542 copy_tree_r (tp, walk_subtrees, NULL);
543 if (block)
545 remap_block (&block, id);
546 BIND_EXPR_BLOCK (*tp) = block;
549 if (BIND_EXPR_VARS (*tp))
550 /* This will remap a lot of the same decls again, but this should be
551 harmless. */
552 BIND_EXPR_VARS (*tp) = remap_decls (BIND_EXPR_VARS (*tp), id);
555 /* Called from copy_body_id via walk_tree. DATA is really an
556 `copy_body_data *'. */
558 tree
559 copy_body_r (tree *tp, int *walk_subtrees, void *data)
561 copy_body_data *id = (copy_body_data *) data;
562 tree fn = id->src_fn;
563 tree new_block;
565 /* Begin by recognizing trees that we'll completely rewrite for the
566 inlining context. Our output for these trees is completely
567 different from out input (e.g. RETURN_EXPR is deleted, and morphs
568 into an edge). Further down, we'll handle trees that get
569 duplicated and/or tweaked. */
571 /* When requested, RETURN_EXPRs should be transformed to just the
572 contained GIMPLE_MODIFY_STMT. The branch semantics of the return will
573 be handled elsewhere by manipulating the CFG rather than a statement. */
574 if (TREE_CODE (*tp) == RETURN_EXPR && id->transform_return_to_modify)
576 tree assignment = TREE_OPERAND (*tp, 0);
578 /* If we're returning something, just turn that into an
579 assignment into the equivalent of the original RESULT_DECL.
580 If the "assignment" is just the result decl, the result
581 decl has already been set (e.g. a recent "foo (&result_decl,
582 ...)"); just toss the entire RETURN_EXPR. */
583 if (assignment && TREE_CODE (assignment) == GIMPLE_MODIFY_STMT)
585 /* Replace the RETURN_EXPR with (a copy of) the
586 GIMPLE_MODIFY_STMT hanging underneath. */
587 *tp = copy_node (assignment);
589 else /* Else the RETURN_EXPR returns no value. */
591 *tp = NULL;
592 return (tree) (void *)1;
595 else if (TREE_CODE (*tp) == SSA_NAME)
597 *tp = remap_ssa_name (*tp, id);
598 *walk_subtrees = 0;
599 return NULL;
602 /* Local variables and labels need to be replaced by equivalent
603 variables. We don't want to copy static variables; there's only
604 one of those, no matter how many times we inline the containing
605 function. Similarly for globals from an outer function. */
606 else if (auto_var_in_fn_p (*tp, fn))
608 tree new_decl;
610 /* Remap the declaration. */
611 new_decl = remap_decl (*tp, id);
612 gcc_assert (new_decl);
613 /* Replace this variable with the copy. */
614 STRIP_TYPE_NOPS (new_decl);
615 *tp = new_decl;
616 *walk_subtrees = 0;
618 else if (TREE_CODE (*tp) == STATEMENT_LIST)
619 copy_statement_list (tp);
620 else if (TREE_CODE (*tp) == SAVE_EXPR)
621 remap_save_expr (tp, id->decl_map, walk_subtrees);
622 else if (TREE_CODE (*tp) == LABEL_DECL
623 && (! DECL_CONTEXT (*tp)
624 || decl_function_context (*tp) == id->src_fn))
625 /* These may need to be remapped for EH handling. */
626 *tp = remap_decl (*tp, id);
627 else if (TREE_CODE (*tp) == BIND_EXPR)
628 copy_bind_expr (tp, walk_subtrees, id);
629 /* Types may need remapping as well. */
630 else if (TYPE_P (*tp))
631 *tp = remap_type (*tp, id);
633 /* If this is a constant, we have to copy the node iff the type will be
634 remapped. copy_tree_r will not copy a constant. */
635 else if (CONSTANT_CLASS_P (*tp))
637 tree new_type = remap_type (TREE_TYPE (*tp), id);
639 if (new_type == TREE_TYPE (*tp))
640 *walk_subtrees = 0;
642 else if (TREE_CODE (*tp) == INTEGER_CST)
643 *tp = build_int_cst_wide (new_type, TREE_INT_CST_LOW (*tp),
644 TREE_INT_CST_HIGH (*tp));
645 else
647 *tp = copy_node (*tp);
648 TREE_TYPE (*tp) = new_type;
652 /* Otherwise, just copy the node. Note that copy_tree_r already
653 knows not to copy VAR_DECLs, etc., so this is safe. */
654 else
656 /* Here we handle trees that are not completely rewritten.
657 First we detect some inlining-induced bogosities for
658 discarding. */
659 if (TREE_CODE (*tp) == GIMPLE_MODIFY_STMT
660 && GIMPLE_STMT_OPERAND (*tp, 0) == GIMPLE_STMT_OPERAND (*tp, 1)
661 && (auto_var_in_fn_p (GIMPLE_STMT_OPERAND (*tp, 0), fn)))
663 /* Some assignments VAR = VAR; don't generate any rtl code
664 and thus don't count as variable modification. Avoid
665 keeping bogosities like 0 = 0. */
666 tree decl = GIMPLE_STMT_OPERAND (*tp, 0), value;
667 tree *n;
669 n = (tree *) pointer_map_contains (id->decl_map, decl);
670 if (n)
672 value = *n;
673 STRIP_TYPE_NOPS (value);
674 if (TREE_CONSTANT (value) || TREE_READONLY (value))
676 *tp = build_empty_stmt ();
677 return copy_body_r (tp, walk_subtrees, data);
681 else if (TREE_CODE (*tp) == INDIRECT_REF)
683 /* Get rid of *& from inline substitutions that can happen when a
684 pointer argument is an ADDR_EXPR. */
685 tree decl = TREE_OPERAND (*tp, 0);
686 tree *n;
688 n = (tree *) pointer_map_contains (id->decl_map, decl);
689 if (n)
691 tree new;
692 tree old;
693 /* If we happen to get an ADDR_EXPR in n->value, strip
694 it manually here as we'll eventually get ADDR_EXPRs
695 which lie about their types pointed to. In this case
696 build_fold_indirect_ref wouldn't strip the INDIRECT_REF,
697 but we absolutely rely on that. As fold_indirect_ref
698 does other useful transformations, try that first, though. */
699 tree type = TREE_TYPE (TREE_TYPE (*n));
700 new = unshare_expr (*n);
701 old = *tp;
702 *tp = gimple_fold_indirect_ref (new);
703 if (! *tp)
705 if (TREE_CODE (new) == ADDR_EXPR)
707 *tp = fold_indirect_ref_1 (type, new);
708 /* ??? We should either assert here or build
709 a VIEW_CONVERT_EXPR instead of blindly leaking
710 incompatible types to our IL. */
711 if (! *tp)
712 *tp = TREE_OPERAND (new, 0);
714 else
716 *tp = build1 (INDIRECT_REF, type, new);
717 TREE_THIS_VOLATILE (*tp) = TREE_THIS_VOLATILE (old);
718 TREE_SIDE_EFFECTS (*tp) = TREE_SIDE_EFFECTS (old);
721 *walk_subtrees = 0;
722 return NULL;
726 /* Here is the "usual case". Copy this tree node, and then
727 tweak some special cases. */
728 copy_tree_r (tp, walk_subtrees, NULL);
730 /* Global variables we haven't seen yet needs to go into referenced
731 vars. If not referenced from types only. */
732 if (gimple_in_ssa_p (cfun) && TREE_CODE (*tp) == VAR_DECL
733 && id->remapping_type_depth == 0)
734 add_referenced_var (*tp);
736 /* If EXPR has block defined, map it to newly constructed block.
737 When inlining we want EXPRs without block appear in the block
738 of function call. */
739 if (EXPR_P (*tp) || GIMPLE_STMT_P (*tp))
741 new_block = id->block;
742 if (TREE_BLOCK (*tp))
744 tree *n;
745 n = (tree *) pointer_map_contains (id->decl_map,
746 TREE_BLOCK (*tp));
747 gcc_assert (n);
748 new_block = *n;
750 TREE_BLOCK (*tp) = new_block;
753 if (TREE_CODE (*tp) == RESX_EXPR && id->eh_region_offset)
754 TREE_OPERAND (*tp, 0) =
755 build_int_cst
756 (NULL_TREE,
757 id->eh_region_offset + TREE_INT_CST_LOW (TREE_OPERAND (*tp, 0)));
759 if (!GIMPLE_TUPLE_P (*tp) && TREE_CODE (*tp) != OMP_CLAUSE)
760 TREE_TYPE (*tp) = remap_type (TREE_TYPE (*tp), id);
762 /* The copied TARGET_EXPR has never been expanded, even if the
763 original node was expanded already. */
764 if (TREE_CODE (*tp) == TARGET_EXPR && TREE_OPERAND (*tp, 3))
766 TREE_OPERAND (*tp, 1) = TREE_OPERAND (*tp, 3);
767 TREE_OPERAND (*tp, 3) = NULL_TREE;
770 /* Variable substitution need not be simple. In particular, the
771 INDIRECT_REF substitution above. Make sure that TREE_CONSTANT
772 and friends are up-to-date. */
773 else if (TREE_CODE (*tp) == ADDR_EXPR)
775 int invariant = is_gimple_min_invariant (*tp);
776 walk_tree (&TREE_OPERAND (*tp, 0), copy_body_r, id, NULL);
777 /* Handle the case where we substituted an INDIRECT_REF
778 into the operand of the ADDR_EXPR. */
779 if (TREE_CODE (TREE_OPERAND (*tp, 0)) == INDIRECT_REF)
780 *tp = TREE_OPERAND (TREE_OPERAND (*tp, 0), 0);
781 else
782 recompute_tree_invariant_for_addr_expr (*tp);
783 /* If this used to be invariant, but is not any longer,
784 then regimplification is probably needed. */
785 if (invariant && !is_gimple_min_invariant (*tp))
786 id->regimplify = true;
787 *walk_subtrees = 0;
791 /* Keep iterating. */
792 return NULL_TREE;
795 /* Copy basic block, scale profile accordingly. Edges will be taken care of
796 later */
798 static basic_block
799 copy_bb (copy_body_data *id, basic_block bb, int frequency_scale,
800 gcov_type 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, gcov_type 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 const new_bb = (basic_block) 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 const old_edge = find_edge ((basic_block) 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 gcov_type 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->local_decls = 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 (cfun);
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 gcov_type 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 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 rhs = value;
1446 tree def = (gimple_in_ssa_p (cfun)
1447 ? gimple_default_def (id->src_cfun, p) : NULL);
1449 if (value
1450 && value != error_mark_node
1451 && !useless_type_conversion_p (TREE_TYPE (p), TREE_TYPE (value)))
1453 if (fold_convertible_p (TREE_TYPE (p), value))
1454 rhs = fold_build1 (NOP_EXPR, TREE_TYPE (p), value);
1455 else
1456 /* ??? For valid (GIMPLE) programs we should not end up here.
1457 Still if something has gone wrong and we end up with truly
1458 mismatched types here, fall back to using a VIEW_CONVERT_EXPR
1459 to not leak invalid GIMPLE to the following passes. */
1460 rhs = fold_build1 (VIEW_CONVERT_EXPR, TREE_TYPE (p), value);
1463 /* If the parameter is never assigned to, has no SSA_NAMEs created,
1464 we may not need to create a new variable here at all. Instead, we may
1465 be able to just use the argument value. */
1466 if (TREE_READONLY (p)
1467 && !TREE_ADDRESSABLE (p)
1468 && value && !TREE_SIDE_EFFECTS (value)
1469 && !def)
1471 /* We may produce non-gimple trees by adding NOPs or introduce
1472 invalid sharing when operand is not really constant.
1473 It is not big deal to prohibit constant propagation here as
1474 we will constant propagate in DOM1 pass anyway. */
1475 if (is_gimple_min_invariant (value)
1476 && useless_type_conversion_p (TREE_TYPE (p),
1477 TREE_TYPE (value))
1478 /* We have to be very careful about ADDR_EXPR. Make sure
1479 the base variable isn't a local variable of the inlined
1480 function, e.g., when doing recursive inlining, direct or
1481 mutually-recursive or whatever, which is why we don't
1482 just test whether fn == current_function_decl. */
1483 && ! self_inlining_addr_expr (value, fn))
1485 insert_decl_map (id, p, value);
1486 return;
1490 /* Make an equivalent VAR_DECL. Note that we must NOT remap the type
1491 here since the type of this decl must be visible to the calling
1492 function. */
1493 var = copy_decl_to_var (p, id);
1494 if (gimple_in_ssa_p (cfun) && TREE_CODE (var) == VAR_DECL)
1496 get_var_ann (var);
1497 add_referenced_var (var);
1500 /* Register the VAR_DECL as the equivalent for the PARM_DECL;
1501 that way, when the PARM_DECL is encountered, it will be
1502 automatically replaced by the VAR_DECL. */
1503 insert_decl_map (id, p, var);
1505 /* Declare this new variable. */
1506 TREE_CHAIN (var) = *vars;
1507 *vars = var;
1509 /* Make gimplifier happy about this variable. */
1510 DECL_SEEN_IN_BIND_EXPR_P (var) = 1;
1512 /* Even if P was TREE_READONLY, the new VAR should not be.
1513 In the original code, we would have constructed a
1514 temporary, and then the function body would have never
1515 changed the value of P. However, now, we will be
1516 constructing VAR directly. The constructor body may
1517 change its value multiple times as it is being
1518 constructed. Therefore, it must not be TREE_READONLY;
1519 the back-end assumes that TREE_READONLY variable is
1520 assigned to only once. */
1521 if (TYPE_NEEDS_CONSTRUCTING (TREE_TYPE (p)))
1522 TREE_READONLY (var) = 0;
1524 /* If there is no setup required and we are in SSA, take the easy route
1525 replacing all SSA names representing the function parameter by the
1526 SSA name passed to function.
1528 We need to construct map for the variable anyway as it might be used
1529 in different SSA names when parameter is set in function.
1531 FIXME: This usually kills the last connection in between inlined
1532 function parameter and the actual value in debug info. Can we do
1533 better here? If we just inserted the statement, copy propagation
1534 would kill it anyway as it always did in older versions of GCC.
1536 We might want to introduce a notion that single SSA_NAME might
1537 represent multiple variables for purposes of debugging. */
1538 if (gimple_in_ssa_p (cfun) && rhs && def && is_gimple_reg (p)
1539 && (TREE_CODE (rhs) == SSA_NAME
1540 || is_gimple_min_invariant (rhs))
1541 && !SSA_NAME_OCCURS_IN_ABNORMAL_PHI (def))
1543 insert_decl_map (id, def, rhs);
1544 return;
1547 /* If the value of argument is never used, don't care about initializing
1548 it. */
1549 if (gimple_in_ssa_p (cfun) && !def && is_gimple_reg (p))
1551 gcc_assert (!value || !TREE_SIDE_EFFECTS (value));
1552 return;
1555 /* Initialize this VAR_DECL from the equivalent argument. Convert
1556 the argument to the proper type in case it was promoted. */
1557 if (value)
1559 block_stmt_iterator bsi = bsi_last (bb);
1561 if (rhs == error_mark_node)
1563 insert_decl_map (id, p, var);
1564 return;
1567 STRIP_USELESS_TYPE_CONVERSION (rhs);
1569 /* We want to use GIMPLE_MODIFY_STMT, not INIT_EXPR here so that we
1570 keep our trees in gimple form. */
1571 if (def && gimple_in_ssa_p (cfun) && is_gimple_reg (p))
1573 def = remap_ssa_name (def, id);
1574 init_stmt = build_gimple_modify_stmt (def, rhs);
1575 SSA_NAME_DEF_STMT (def) = init_stmt;
1576 SSA_NAME_IS_DEFAULT_DEF (def) = 0;
1577 set_default_def (var, NULL);
1579 else
1580 init_stmt = build_gimple_modify_stmt (var, rhs);
1582 /* If we did not create a gimple value and we did not create a gimple
1583 cast of a gimple value, then we will need to gimplify INIT_STMTS
1584 at the end. Note that is_gimple_cast only checks the outer
1585 tree code, not its operand. Thus the explicit check that its
1586 operand is a gimple value. */
1587 if ((!is_gimple_val (rhs)
1588 && (!is_gimple_cast (rhs)
1589 || !is_gimple_val (TREE_OPERAND (rhs, 0))))
1590 || !is_gimple_reg (var))
1592 tree_stmt_iterator i;
1593 struct gimplify_ctx gctx;
1595 push_gimplify_context (&gctx);
1596 gimplify_stmt (&init_stmt);
1597 if (gimple_in_ssa_p (cfun)
1598 && init_stmt && TREE_CODE (init_stmt) == STATEMENT_LIST)
1600 /* The replacement can expose previously unreferenced
1601 variables. */
1602 for (i = tsi_start (init_stmt); !tsi_end_p (i); tsi_next (&i))
1603 find_new_referenced_vars (tsi_stmt_ptr (i));
1605 pop_gimplify_context (NULL);
1608 /* If VAR represents a zero-sized variable, it's possible that the
1609 assignment statement may result in no gimple statements. */
1610 if (init_stmt)
1611 bsi_insert_after (&bsi, init_stmt, BSI_NEW_STMT);
1612 if (gimple_in_ssa_p (cfun))
1613 for (;!bsi_end_p (bsi); bsi_next (&bsi))
1614 mark_symbols_for_renaming (bsi_stmt (bsi));
1618 /* Generate code to initialize the parameters of the function at the
1619 top of the stack in ID from the CALL_EXPR EXP. */
1621 static void
1622 initialize_inlined_parameters (copy_body_data *id, tree exp,
1623 tree fn, basic_block bb)
1625 tree parms;
1626 tree a;
1627 tree p;
1628 tree vars = NULL_TREE;
1629 call_expr_arg_iterator iter;
1630 tree static_chain = CALL_EXPR_STATIC_CHAIN (exp);
1632 /* Figure out what the parameters are. */
1633 parms = DECL_ARGUMENTS (fn);
1635 /* Loop through the parameter declarations, replacing each with an
1636 equivalent VAR_DECL, appropriately initialized. */
1637 for (p = parms, a = first_call_expr_arg (exp, &iter); p;
1638 a = next_call_expr_arg (&iter), p = TREE_CHAIN (p))
1639 setup_one_parameter (id, p, a, fn, bb, &vars);
1641 /* Initialize the static chain. */
1642 p = DECL_STRUCT_FUNCTION (fn)->static_chain_decl;
1643 gcc_assert (fn != current_function_decl);
1644 if (p)
1646 /* No static chain? Seems like a bug in tree-nested.c. */
1647 gcc_assert (static_chain);
1649 setup_one_parameter (id, p, static_chain, fn, bb, &vars);
1652 declare_inline_vars (id->block, vars);
1655 /* Declare a return variable to replace the RESULT_DECL for the
1656 function we are calling. An appropriate DECL_STMT is returned.
1657 The USE_STMT is filled to contain a use of the declaration to
1658 indicate the return value of the function.
1660 RETURN_SLOT, if non-null is place where to store the result. It
1661 is set only for CALL_EXPR_RETURN_SLOT_OPT. MODIFY_DEST, if non-null,
1662 was the LHS of the GIMPLE_MODIFY_STMT to which this call is the RHS.
1664 The return value is a (possibly null) value that is the result of the
1665 function as seen by the callee. *USE_P is a (possibly null) value that
1666 holds the result as seen by the caller. */
1668 static tree
1669 declare_return_variable (copy_body_data *id, tree return_slot, tree modify_dest,
1670 tree *use_p)
1672 tree callee = id->src_fn;
1673 tree caller = id->dst_fn;
1674 tree result = DECL_RESULT (callee);
1675 tree callee_type = TREE_TYPE (result);
1676 tree caller_type = TREE_TYPE (TREE_TYPE (callee));
1677 tree var, use;
1679 /* We don't need to do anything for functions that don't return
1680 anything. */
1681 if (!result || VOID_TYPE_P (callee_type))
1683 *use_p = NULL_TREE;
1684 return NULL_TREE;
1687 /* If there was a return slot, then the return value is the
1688 dereferenced address of that object. */
1689 if (return_slot)
1691 /* The front end shouldn't have used both return_slot and
1692 a modify expression. */
1693 gcc_assert (!modify_dest);
1694 if (DECL_BY_REFERENCE (result))
1696 tree return_slot_addr = build_fold_addr_expr (return_slot);
1697 STRIP_USELESS_TYPE_CONVERSION (return_slot_addr);
1699 /* We are going to construct *&return_slot and we can't do that
1700 for variables believed to be not addressable.
1702 FIXME: This check possibly can match, because values returned
1703 via return slot optimization are not believed to have address
1704 taken by alias analysis. */
1705 gcc_assert (TREE_CODE (return_slot) != SSA_NAME);
1706 if (gimple_in_ssa_p (cfun))
1708 HOST_WIDE_INT bitsize;
1709 HOST_WIDE_INT bitpos;
1710 tree offset;
1711 enum machine_mode mode;
1712 int unsignedp;
1713 int volatilep;
1714 tree base;
1715 base = get_inner_reference (return_slot, &bitsize, &bitpos,
1716 &offset,
1717 &mode, &unsignedp, &volatilep,
1718 false);
1719 if (TREE_CODE (base) == INDIRECT_REF)
1720 base = TREE_OPERAND (base, 0);
1721 if (TREE_CODE (base) == SSA_NAME)
1722 base = SSA_NAME_VAR (base);
1723 mark_sym_for_renaming (base);
1725 var = return_slot_addr;
1727 else
1729 var = return_slot;
1730 gcc_assert (TREE_CODE (var) != SSA_NAME);
1731 TREE_ADDRESSABLE (var) |= TREE_ADDRESSABLE (result);
1733 if ((TREE_CODE (TREE_TYPE (result)) == COMPLEX_TYPE
1734 || TREE_CODE (TREE_TYPE (result)) == VECTOR_TYPE)
1735 && !DECL_GIMPLE_REG_P (result)
1736 && DECL_P (var))
1737 DECL_GIMPLE_REG_P (var) = 0;
1738 use = NULL;
1739 goto done;
1742 /* All types requiring non-trivial constructors should have been handled. */
1743 gcc_assert (!TREE_ADDRESSABLE (callee_type));
1745 /* Attempt to avoid creating a new temporary variable. */
1746 if (modify_dest
1747 && TREE_CODE (modify_dest) != SSA_NAME)
1749 bool use_it = false;
1751 /* We can't use MODIFY_DEST if there's type promotion involved. */
1752 if (!useless_type_conversion_p (callee_type, caller_type))
1753 use_it = false;
1755 /* ??? If we're assigning to a variable sized type, then we must
1756 reuse the destination variable, because we've no good way to
1757 create variable sized temporaries at this point. */
1758 else if (TREE_CODE (TYPE_SIZE_UNIT (caller_type)) != INTEGER_CST)
1759 use_it = true;
1761 /* If the callee cannot possibly modify MODIFY_DEST, then we can
1762 reuse it as the result of the call directly. Don't do this if
1763 it would promote MODIFY_DEST to addressable. */
1764 else if (TREE_ADDRESSABLE (result))
1765 use_it = false;
1766 else
1768 tree base_m = get_base_address (modify_dest);
1770 /* If the base isn't a decl, then it's a pointer, and we don't
1771 know where that's going to go. */
1772 if (!DECL_P (base_m))
1773 use_it = false;
1774 else if (is_global_var (base_m))
1775 use_it = false;
1776 else if ((TREE_CODE (TREE_TYPE (result)) == COMPLEX_TYPE
1777 || TREE_CODE (TREE_TYPE (result)) == VECTOR_TYPE)
1778 && !DECL_GIMPLE_REG_P (result)
1779 && DECL_GIMPLE_REG_P (base_m))
1780 use_it = false;
1781 else if (!TREE_ADDRESSABLE (base_m))
1782 use_it = true;
1785 if (use_it)
1787 var = modify_dest;
1788 use = NULL;
1789 goto done;
1793 gcc_assert (TREE_CODE (TYPE_SIZE_UNIT (callee_type)) == INTEGER_CST);
1795 var = copy_result_decl_to_var (result, id);
1796 if (gimple_in_ssa_p (cfun))
1798 get_var_ann (var);
1799 add_referenced_var (var);
1802 DECL_SEEN_IN_BIND_EXPR_P (var) = 1;
1803 DECL_STRUCT_FUNCTION (caller)->local_decls
1804 = tree_cons (NULL_TREE, var,
1805 DECL_STRUCT_FUNCTION (caller)->local_decls);
1807 /* Do not have the rest of GCC warn about this variable as it should
1808 not be visible to the user. */
1809 TREE_NO_WARNING (var) = 1;
1811 declare_inline_vars (id->block, var);
1813 /* Build the use expr. If the return type of the function was
1814 promoted, convert it back to the expected type. */
1815 use = var;
1816 if (!useless_type_conversion_p (caller_type, TREE_TYPE (var)))
1817 use = fold_convert (caller_type, var);
1819 STRIP_USELESS_TYPE_CONVERSION (use);
1821 if (DECL_BY_REFERENCE (result))
1822 var = build_fold_addr_expr (var);
1824 done:
1825 /* Register the VAR_DECL as the equivalent for the RESULT_DECL; that
1826 way, when the RESULT_DECL is encountered, it will be
1827 automatically replaced by the VAR_DECL. */
1828 insert_decl_map (id, result, var);
1830 /* Remember this so we can ignore it in remap_decls. */
1831 id->retvar = var;
1833 *use_p = use;
1834 return var;
1837 /* Returns nonzero if a function can be inlined as a tree. */
1839 bool
1840 tree_inlinable_function_p (tree fn)
1842 return inlinable_function_p (fn);
1845 static const char *inline_forbidden_reason;
1847 static tree
1848 inline_forbidden_p_1 (tree *nodep, int *walk_subtrees ATTRIBUTE_UNUSED,
1849 void *fnp)
1851 tree node = *nodep;
1852 tree fn = (tree) fnp;
1853 tree t;
1855 switch (TREE_CODE (node))
1857 case CALL_EXPR:
1858 /* Refuse to inline alloca call unless user explicitly forced so as
1859 this may change program's memory overhead drastically when the
1860 function using alloca is called in loop. In GCC present in
1861 SPEC2000 inlining into schedule_block cause it to require 2GB of
1862 RAM instead of 256MB. */
1863 if (alloca_call_p (node)
1864 && !lookup_attribute ("always_inline", DECL_ATTRIBUTES (fn)))
1866 inline_forbidden_reason
1867 = G_("function %q+F can never be inlined because it uses "
1868 "alloca (override using the always_inline attribute)");
1869 return node;
1871 t = get_callee_fndecl (node);
1872 if (! t)
1873 break;
1875 /* We cannot inline functions that call setjmp. */
1876 if (setjmp_call_p (t))
1878 inline_forbidden_reason
1879 = G_("function %q+F can never be inlined because it uses setjmp");
1880 return node;
1883 if (DECL_BUILT_IN_CLASS (t) == BUILT_IN_NORMAL)
1884 switch (DECL_FUNCTION_CODE (t))
1886 /* We cannot inline functions that take a variable number of
1887 arguments. */
1888 case BUILT_IN_VA_START:
1889 case BUILT_IN_NEXT_ARG:
1890 case BUILT_IN_VA_END:
1891 inline_forbidden_reason
1892 = G_("function %q+F can never be inlined because it "
1893 "uses variable argument lists");
1894 return node;
1896 case BUILT_IN_LONGJMP:
1897 /* We can't inline functions that call __builtin_longjmp at
1898 all. The non-local goto machinery really requires the
1899 destination be in a different function. If we allow the
1900 function calling __builtin_longjmp to be inlined into the
1901 function calling __builtin_setjmp, Things will Go Awry. */
1902 inline_forbidden_reason
1903 = G_("function %q+F can never be inlined because "
1904 "it uses setjmp-longjmp exception handling");
1905 return node;
1907 case BUILT_IN_NONLOCAL_GOTO:
1908 /* Similarly. */
1909 inline_forbidden_reason
1910 = G_("function %q+F can never be inlined because "
1911 "it uses non-local goto");
1912 return node;
1914 case BUILT_IN_RETURN:
1915 case BUILT_IN_APPLY_ARGS:
1916 /* If a __builtin_apply_args caller would be inlined,
1917 it would be saving arguments of the function it has
1918 been inlined into. Similarly __builtin_return would
1919 return from the function the inline has been inlined into. */
1920 inline_forbidden_reason
1921 = G_("function %q+F can never be inlined because "
1922 "it uses __builtin_return or __builtin_apply_args");
1923 return node;
1925 default:
1926 break;
1928 break;
1930 case GOTO_EXPR:
1931 t = TREE_OPERAND (node, 0);
1933 /* We will not inline a function which uses computed goto. The
1934 addresses of its local labels, which may be tucked into
1935 global storage, are of course not constant across
1936 instantiations, which causes unexpected behavior. */
1937 if (TREE_CODE (t) != LABEL_DECL)
1939 inline_forbidden_reason
1940 = G_("function %q+F can never be inlined "
1941 "because it contains a computed goto");
1942 return node;
1944 break;
1946 case LABEL_EXPR:
1947 t = TREE_OPERAND (node, 0);
1948 if (DECL_NONLOCAL (t))
1950 /* We cannot inline a function that receives a non-local goto
1951 because we cannot remap the destination label used in the
1952 function that is performing the non-local goto. */
1953 inline_forbidden_reason
1954 = G_("function %q+F can never be inlined "
1955 "because it receives a non-local goto");
1956 return node;
1958 break;
1960 case RECORD_TYPE:
1961 case UNION_TYPE:
1962 /* We cannot inline a function of the form
1964 void F (int i) { struct S { int ar[i]; } s; }
1966 Attempting to do so produces a catch-22.
1967 If walk_tree examines the TYPE_FIELDS chain of RECORD_TYPE/
1968 UNION_TYPE nodes, then it goes into infinite recursion on a
1969 structure containing a pointer to its own type. If it doesn't,
1970 then the type node for S doesn't get adjusted properly when
1971 F is inlined.
1973 ??? This is likely no longer true, but it's too late in the 4.0
1974 cycle to try to find out. This should be checked for 4.1. */
1975 for (t = TYPE_FIELDS (node); t; t = TREE_CHAIN (t))
1976 if (variably_modified_type_p (TREE_TYPE (t), NULL))
1978 inline_forbidden_reason
1979 = G_("function %q+F can never be inlined "
1980 "because it uses variable sized variables");
1981 return node;
1984 default:
1985 break;
1988 return NULL_TREE;
1991 static tree
1992 inline_forbidden_p_2 (tree *nodep, int *walk_subtrees,
1993 void *fnp)
1995 tree node = *nodep;
1996 tree fn = (tree) fnp;
1998 if (TREE_CODE (node) == LABEL_DECL && DECL_CONTEXT (node) == fn)
2000 inline_forbidden_reason
2001 = G_("function %q+F can never be inlined "
2002 "because it saves address of local label in a static variable");
2003 return node;
2006 if (TYPE_P (node))
2007 *walk_subtrees = 0;
2009 return NULL_TREE;
2012 /* Return subexpression representing possible alloca call, if any. */
2013 static tree
2014 inline_forbidden_p (tree fndecl)
2016 location_t saved_loc = input_location;
2017 block_stmt_iterator bsi;
2018 basic_block bb;
2019 tree ret = NULL_TREE;
2020 struct function *fun = DECL_STRUCT_FUNCTION (fndecl);
2021 tree step;
2023 FOR_EACH_BB_FN (bb, fun)
2024 for (bsi = bsi_start (bb); !bsi_end_p (bsi); bsi_next (&bsi))
2026 ret = walk_tree_without_duplicates (bsi_stmt_ptr (bsi),
2027 inline_forbidden_p_1, fndecl);
2028 if (ret)
2029 goto egress;
2032 for (step = fun->local_decls; step; step = TREE_CHAIN (step))
2034 tree decl = TREE_VALUE (step);
2035 if (TREE_CODE (decl) == VAR_DECL
2036 && TREE_STATIC (decl)
2037 && !DECL_EXTERNAL (decl)
2038 && DECL_INITIAL (decl))
2039 ret = walk_tree_without_duplicates (&DECL_INITIAL (decl),
2040 inline_forbidden_p_2, fndecl);
2041 if (ret)
2042 goto egress;
2045 egress:
2046 input_location = saved_loc;
2047 return ret;
2050 /* Returns nonzero if FN is a function that does not have any
2051 fundamental inline blocking properties. */
2053 static bool
2054 inlinable_function_p (tree fn)
2056 bool inlinable = true;
2057 bool do_warning;
2058 tree always_inline;
2060 /* If we've already decided this function shouldn't be inlined,
2061 there's no need to check again. */
2062 if (DECL_UNINLINABLE (fn))
2063 return false;
2065 /* We only warn for functions declared `inline' by the user. */
2066 do_warning = (warn_inline
2067 && DECL_INLINE (fn)
2068 && DECL_DECLARED_INLINE_P (fn)
2069 && !DECL_IN_SYSTEM_HEADER (fn));
2071 always_inline = lookup_attribute ("always_inline", DECL_ATTRIBUTES (fn));
2073 if (flag_really_no_inline
2074 && always_inline == NULL)
2076 if (do_warning)
2077 warning (OPT_Winline, "function %q+F can never be inlined because it "
2078 "is suppressed using -fno-inline", fn);
2079 inlinable = false;
2082 /* Don't auto-inline anything that might not be bound within
2083 this unit of translation. */
2084 else if (!DECL_DECLARED_INLINE_P (fn)
2085 && DECL_REPLACEABLE_P (fn))
2086 inlinable = false;
2088 else if (!function_attribute_inlinable_p (fn))
2090 if (do_warning)
2091 warning (OPT_Winline, "function %q+F can never be inlined because it "
2092 "uses attributes conflicting with inlining", fn);
2093 inlinable = false;
2096 /* If we don't have the function body available, we can't inline it.
2097 However, this should not be recorded since we also get here for
2098 forward declared inline functions. Therefore, return at once. */
2099 if (!DECL_SAVED_TREE (fn))
2100 return false;
2102 /* If we're not inlining at all, then we cannot inline this function. */
2103 else if (!flag_inline_trees)
2104 inlinable = false;
2106 /* Only try to inline functions if DECL_INLINE is set. This should be
2107 true for all functions declared `inline', and for all other functions
2108 as well with -finline-functions.
2110 Don't think of disregarding DECL_INLINE when flag_inline_trees == 2;
2111 it's the front-end that must set DECL_INLINE in this case, because
2112 dwarf2out loses if a function that does not have DECL_INLINE set is
2113 inlined anyway. That is why we have both DECL_INLINE and
2114 DECL_DECLARED_INLINE_P. */
2115 /* FIXME: When flag_inline_trees dies, the check for flag_unit_at_a_time
2116 here should be redundant. */
2117 else if (!DECL_INLINE (fn) && !flag_unit_at_a_time)
2118 inlinable = false;
2120 else if (inline_forbidden_p (fn))
2122 /* See if we should warn about uninlinable functions. Previously,
2123 some of these warnings would be issued while trying to expand
2124 the function inline, but that would cause multiple warnings
2125 about functions that would for example call alloca. But since
2126 this a property of the function, just one warning is enough.
2127 As a bonus we can now give more details about the reason why a
2128 function is not inlinable. */
2129 if (always_inline)
2130 sorry (inline_forbidden_reason, fn);
2131 else if (do_warning)
2132 warning (OPT_Winline, inline_forbidden_reason, fn);
2134 inlinable = false;
2137 /* Squirrel away the result so that we don't have to check again. */
2138 DECL_UNINLINABLE (fn) = !inlinable;
2140 return inlinable;
2143 /* Estimate the cost of a memory move. Use machine dependent
2144 word size and take possible memcpy call into account. */
2147 estimate_move_cost (tree type)
2149 HOST_WIDE_INT size;
2151 size = int_size_in_bytes (type);
2153 if (size < 0 || size > MOVE_MAX_PIECES * MOVE_RATIO)
2154 /* Cost of a memcpy call, 3 arguments and the call. */
2155 return 4;
2156 else
2157 return ((size + MOVE_MAX_PIECES - 1) / MOVE_MAX_PIECES);
2160 /* Arguments for estimate_num_insns_1. */
2162 struct eni_data
2164 /* Used to return the number of insns. */
2165 int count;
2167 /* Weights of various constructs. */
2168 eni_weights *weights;
2171 /* Used by estimate_num_insns. Estimate number of instructions seen
2172 by given statement. */
2174 static tree
2175 estimate_num_insns_1 (tree *tp, int *walk_subtrees, void *data)
2177 struct eni_data *const d = (struct eni_data *) data;
2178 tree x = *tp;
2179 unsigned cost;
2181 if (IS_TYPE_OR_DECL_P (x))
2183 *walk_subtrees = 0;
2184 return NULL;
2186 /* Assume that constants and references counts nothing. These should
2187 be majorized by amount of operations among them we count later
2188 and are common target of CSE and similar optimizations. */
2189 else if (CONSTANT_CLASS_P (x) || REFERENCE_CLASS_P (x))
2190 return NULL;
2192 switch (TREE_CODE (x))
2194 /* Containers have no cost. */
2195 case TREE_LIST:
2196 case TREE_VEC:
2197 case BLOCK:
2198 case COMPONENT_REF:
2199 case BIT_FIELD_REF:
2200 case INDIRECT_REF:
2201 case ALIGN_INDIRECT_REF:
2202 case MISALIGNED_INDIRECT_REF:
2203 case ARRAY_REF:
2204 case ARRAY_RANGE_REF:
2205 case OBJ_TYPE_REF:
2206 case EXC_PTR_EXPR: /* ??? */
2207 case FILTER_EXPR: /* ??? */
2208 case COMPOUND_EXPR:
2209 case BIND_EXPR:
2210 case WITH_CLEANUP_EXPR:
2211 case PAREN_EXPR:
2212 CASE_CONVERT:
2213 case VIEW_CONVERT_EXPR:
2214 case SAVE_EXPR:
2215 case ADDR_EXPR:
2216 case COMPLEX_EXPR:
2217 case RANGE_EXPR:
2218 case CASE_LABEL_EXPR:
2219 case SSA_NAME:
2220 case CATCH_EXPR:
2221 case EH_FILTER_EXPR:
2222 case STATEMENT_LIST:
2223 case ERROR_MARK:
2224 case FDESC_EXPR:
2225 case VA_ARG_EXPR:
2226 case TRY_CATCH_EXPR:
2227 case TRY_FINALLY_EXPR:
2228 case LABEL_EXPR:
2229 case GOTO_EXPR:
2230 case RETURN_EXPR:
2231 case EXIT_EXPR:
2232 case LOOP_EXPR:
2233 case PHI_NODE:
2234 case WITH_SIZE_EXPR:
2235 case OMP_CLAUSE:
2236 case OMP_RETURN:
2237 case OMP_CONTINUE:
2238 case OMP_SECTIONS_SWITCH:
2239 case OMP_ATOMIC_STORE:
2240 break;
2242 /* We don't account constants for now. Assume that the cost is amortized
2243 by operations that do use them. We may re-consider this decision once
2244 we are able to optimize the tree before estimating its size and break
2245 out static initializers. */
2246 case IDENTIFIER_NODE:
2247 case INTEGER_CST:
2248 case REAL_CST:
2249 case FIXED_CST:
2250 case COMPLEX_CST:
2251 case VECTOR_CST:
2252 case STRING_CST:
2253 case PREDICT_EXPR:
2254 *walk_subtrees = 0;
2255 return NULL;
2257 /* CHANGE_DYNAMIC_TYPE_EXPR explicitly expands to nothing. */
2258 case CHANGE_DYNAMIC_TYPE_EXPR:
2259 *walk_subtrees = 0;
2260 return NULL;
2262 /* Try to estimate the cost of assignments. We have three cases to
2263 deal with:
2264 1) Simple assignments to registers;
2265 2) Stores to things that must live in memory. This includes
2266 "normal" stores to scalars, but also assignments of large
2267 structures, or constructors of big arrays;
2268 3) TARGET_EXPRs.
2270 Let us look at the first two cases, assuming we have "a = b + C":
2271 <GIMPLE_MODIFY_STMT <var_decl "a">
2272 <plus_expr <var_decl "b"> <constant C>>
2273 If "a" is a GIMPLE register, the assignment to it is free on almost
2274 any target, because "a" usually ends up in a real register. Hence
2275 the only cost of this expression comes from the PLUS_EXPR, and we
2276 can ignore the GIMPLE_MODIFY_STMT.
2277 If "a" is not a GIMPLE register, the assignment to "a" will most
2278 likely be a real store, so the cost of the GIMPLE_MODIFY_STMT is the cost
2279 of moving something into "a", which we compute using the function
2280 estimate_move_cost.
2282 The third case deals with TARGET_EXPRs, for which the semantics are
2283 that a temporary is assigned, unless the TARGET_EXPR itself is being
2284 assigned to something else. In the latter case we do not need the
2285 temporary. E.g. in:
2286 <GIMPLE_MODIFY_STMT <var_decl "a"> <target_expr>>, the
2287 GIMPLE_MODIFY_STMT is free. */
2288 case INIT_EXPR:
2289 case GIMPLE_MODIFY_STMT:
2290 /* Is the right and side a TARGET_EXPR? */
2291 if (TREE_CODE (GENERIC_TREE_OPERAND (x, 1)) == TARGET_EXPR)
2292 break;
2293 /* ... fall through ... */
2295 case TARGET_EXPR:
2296 x = GENERIC_TREE_OPERAND (x, 0);
2297 /* Is this an assignments to a register? */
2298 if (is_gimple_reg (x))
2299 break;
2300 /* Otherwise it's a store, so fall through to compute the move cost. */
2302 case CONSTRUCTOR:
2303 d->count += estimate_move_cost (TREE_TYPE (x));
2304 break;
2306 /* Assign cost of 1 to usual operations.
2307 ??? We may consider mapping RTL costs to this. */
2308 case COND_EXPR:
2309 case VEC_COND_EXPR:
2311 case PLUS_EXPR:
2312 case POINTER_PLUS_EXPR:
2313 case MINUS_EXPR:
2314 case MULT_EXPR:
2316 case FIXED_CONVERT_EXPR:
2317 case FIX_TRUNC_EXPR:
2319 case NEGATE_EXPR:
2320 case FLOAT_EXPR:
2321 case MIN_EXPR:
2322 case MAX_EXPR:
2323 case ABS_EXPR:
2325 case LSHIFT_EXPR:
2326 case RSHIFT_EXPR:
2327 case LROTATE_EXPR:
2328 case RROTATE_EXPR:
2329 case VEC_LSHIFT_EXPR:
2330 case VEC_RSHIFT_EXPR:
2332 case BIT_IOR_EXPR:
2333 case BIT_XOR_EXPR:
2334 case BIT_AND_EXPR:
2335 case BIT_NOT_EXPR:
2337 case TRUTH_ANDIF_EXPR:
2338 case TRUTH_ORIF_EXPR:
2339 case TRUTH_AND_EXPR:
2340 case TRUTH_OR_EXPR:
2341 case TRUTH_XOR_EXPR:
2342 case TRUTH_NOT_EXPR:
2344 case LT_EXPR:
2345 case LE_EXPR:
2346 case GT_EXPR:
2347 case GE_EXPR:
2348 case EQ_EXPR:
2349 case NE_EXPR:
2350 case ORDERED_EXPR:
2351 case UNORDERED_EXPR:
2353 case UNLT_EXPR:
2354 case UNLE_EXPR:
2355 case UNGT_EXPR:
2356 case UNGE_EXPR:
2357 case UNEQ_EXPR:
2358 case LTGT_EXPR:
2360 case CONJ_EXPR:
2362 case PREDECREMENT_EXPR:
2363 case PREINCREMENT_EXPR:
2364 case POSTDECREMENT_EXPR:
2365 case POSTINCREMENT_EXPR:
2367 case ASM_EXPR:
2369 case REALIGN_LOAD_EXPR:
2371 case REDUC_MAX_EXPR:
2372 case REDUC_MIN_EXPR:
2373 case REDUC_PLUS_EXPR:
2374 case WIDEN_SUM_EXPR:
2375 case DOT_PROD_EXPR:
2376 case VEC_WIDEN_MULT_HI_EXPR:
2377 case VEC_WIDEN_MULT_LO_EXPR:
2378 case VEC_UNPACK_HI_EXPR:
2379 case VEC_UNPACK_LO_EXPR:
2380 case VEC_UNPACK_FLOAT_HI_EXPR:
2381 case VEC_UNPACK_FLOAT_LO_EXPR:
2382 case VEC_PACK_TRUNC_EXPR:
2383 case VEC_PACK_SAT_EXPR:
2384 case VEC_PACK_FIX_TRUNC_EXPR:
2386 case WIDEN_MULT_EXPR:
2388 case VEC_EXTRACT_EVEN_EXPR:
2389 case VEC_EXTRACT_ODD_EXPR:
2390 case VEC_INTERLEAVE_HIGH_EXPR:
2391 case VEC_INTERLEAVE_LOW_EXPR:
2393 case RESX_EXPR:
2394 d->count += 1;
2395 break;
2397 case SWITCH_EXPR:
2398 /* Take into account cost of the switch + guess 2 conditional jumps for
2399 each case label.
2401 TODO: once the switch expansion logic is sufficiently separated, we can
2402 do better job on estimating cost of the switch. */
2403 d->count += TREE_VEC_LENGTH (SWITCH_LABELS (x)) * 2;
2404 break;
2406 /* Few special cases of expensive operations. This is useful
2407 to avoid inlining on functions having too many of these. */
2408 case TRUNC_DIV_EXPR:
2409 case CEIL_DIV_EXPR:
2410 case FLOOR_DIV_EXPR:
2411 case ROUND_DIV_EXPR:
2412 case EXACT_DIV_EXPR:
2413 case TRUNC_MOD_EXPR:
2414 case CEIL_MOD_EXPR:
2415 case FLOOR_MOD_EXPR:
2416 case ROUND_MOD_EXPR:
2417 case RDIV_EXPR:
2418 d->count += d->weights->div_mod_cost;
2419 break;
2420 case CALL_EXPR:
2422 tree decl = get_callee_fndecl (x);
2423 tree addr = CALL_EXPR_FN (x);
2424 tree funtype = TREE_TYPE (addr);
2426 gcc_assert (POINTER_TYPE_P (funtype));
2427 funtype = TREE_TYPE (funtype);
2429 if (decl && DECL_BUILT_IN_CLASS (decl) == BUILT_IN_MD)
2430 cost = d->weights->target_builtin_call_cost;
2431 else
2432 cost = d->weights->call_cost;
2434 if (decl && DECL_BUILT_IN_CLASS (decl) == BUILT_IN_NORMAL)
2435 switch (DECL_FUNCTION_CODE (decl))
2437 case BUILT_IN_CONSTANT_P:
2438 *walk_subtrees = 0;
2439 return NULL_TREE;
2440 case BUILT_IN_EXPECT:
2441 return NULL_TREE;
2442 /* Prefetch instruction is not expensive. */
2443 case BUILT_IN_PREFETCH:
2444 cost = 1;
2445 break;
2446 default:
2447 break;
2450 if (decl)
2451 funtype = TREE_TYPE (decl);
2453 /* Our cost must be kept in sync with cgraph_estimate_size_after_inlining
2454 that does use function declaration to figure out the arguments.
2456 When we deal with function with no body nor prototype, base estimates on
2457 actual parameters of the call expression. Otherwise use either the actual
2458 arguments types or function declaration for more precise answer. */
2459 if (decl && DECL_ARGUMENTS (decl))
2461 tree arg;
2462 for (arg = DECL_ARGUMENTS (decl); arg; arg = TREE_CHAIN (arg))
2463 d->count += estimate_move_cost (TREE_TYPE (arg));
2465 else if (funtype && prototype_p (funtype))
2467 tree t;
2468 for (t = TYPE_ARG_TYPES (funtype); t; t = TREE_CHAIN (t))
2469 d->count += estimate_move_cost (TREE_VALUE (t));
2471 else
2473 tree a;
2474 call_expr_arg_iterator iter;
2475 FOR_EACH_CALL_EXPR_ARG (a, iter, x)
2476 d->count += estimate_move_cost (TREE_TYPE (a));
2479 d->count += cost;
2480 break;
2483 case OMP_PARALLEL:
2484 case OMP_TASK:
2485 case OMP_FOR:
2486 case OMP_SECTIONS:
2487 case OMP_SINGLE:
2488 case OMP_SECTION:
2489 case OMP_MASTER:
2490 case OMP_ORDERED:
2491 case OMP_CRITICAL:
2492 case OMP_ATOMIC:
2493 case OMP_ATOMIC_LOAD:
2494 /* OpenMP directives are generally very expensive. */
2495 d->count += d->weights->omp_cost;
2496 break;
2498 default:
2499 gcc_unreachable ();
2501 return NULL;
2504 /* Estimate number of instructions that will be created by expanding EXPR.
2505 WEIGHTS contains weights attributed to various constructs. */
2508 estimate_num_insns (tree expr, eni_weights *weights)
2510 struct pointer_set_t *visited_nodes;
2511 basic_block bb;
2512 block_stmt_iterator bsi;
2513 struct function *my_function;
2514 struct eni_data data;
2516 data.count = 0;
2517 data.weights = weights;
2519 /* If we're given an entire function, walk the CFG. */
2520 if (TREE_CODE (expr) == FUNCTION_DECL)
2522 my_function = DECL_STRUCT_FUNCTION (expr);
2523 gcc_assert (my_function && my_function->cfg);
2524 visited_nodes = pointer_set_create ();
2525 FOR_EACH_BB_FN (bb, my_function)
2527 for (bsi = bsi_start (bb);
2528 !bsi_end_p (bsi);
2529 bsi_next (&bsi))
2531 walk_tree (bsi_stmt_ptr (bsi), estimate_num_insns_1,
2532 &data, visited_nodes);
2535 pointer_set_destroy (visited_nodes);
2537 else
2538 walk_tree_without_duplicates (&expr, estimate_num_insns_1, &data);
2540 return data.count;
2543 /* Initializes weights used by estimate_num_insns. */
2545 void
2546 init_inline_once (void)
2548 eni_inlining_weights.call_cost = PARAM_VALUE (PARAM_INLINE_CALL_COST);
2549 eni_inlining_weights.target_builtin_call_cost = 1;
2550 eni_inlining_weights.div_mod_cost = 10;
2551 eni_inlining_weights.omp_cost = 40;
2553 eni_size_weights.call_cost = 1;
2554 eni_size_weights.target_builtin_call_cost = 1;
2555 eni_size_weights.div_mod_cost = 1;
2556 eni_size_weights.omp_cost = 40;
2558 /* Estimating time for call is difficult, since we have no idea what the
2559 called function does. In the current uses of eni_time_weights,
2560 underestimating the cost does less harm than overestimating it, so
2561 we choose a rather small value here. */
2562 eni_time_weights.call_cost = 10;
2563 eni_time_weights.target_builtin_call_cost = 10;
2564 eni_time_weights.div_mod_cost = 10;
2565 eni_time_weights.omp_cost = 40;
2568 /* Install new lexical TREE_BLOCK underneath 'current_block'. */
2569 static void
2570 add_lexical_block (tree current_block, tree new_block)
2572 tree *blk_p;
2574 /* Walk to the last sub-block. */
2575 for (blk_p = &BLOCK_SUBBLOCKS (current_block);
2576 *blk_p;
2577 blk_p = &BLOCK_CHAIN (*blk_p))
2579 *blk_p = new_block;
2580 BLOCK_SUPERCONTEXT (new_block) = current_block;
2583 /* If *TP is a CALL_EXPR, replace it with its inline expansion. */
2585 static bool
2586 expand_call_inline (basic_block bb, tree stmt, tree *tp, void *data)
2588 copy_body_data *id;
2589 tree t;
2590 tree retvar, use_retvar;
2591 tree fn;
2592 struct pointer_map_t *st;
2593 tree return_slot;
2594 tree modify_dest;
2595 location_t saved_location;
2596 struct cgraph_edge *cg_edge;
2597 const char *reason;
2598 basic_block return_block;
2599 edge e;
2600 block_stmt_iterator bsi, stmt_bsi;
2601 bool successfully_inlined = FALSE;
2602 bool purge_dead_abnormal_edges;
2603 tree t_step;
2604 tree var;
2606 /* See what we've got. */
2607 id = (copy_body_data *) data;
2608 t = *tp;
2610 /* Set input_location here so we get the right instantiation context
2611 if we call instantiate_decl from inlinable_function_p. */
2612 saved_location = input_location;
2613 if (EXPR_HAS_LOCATION (t))
2614 input_location = EXPR_LOCATION (t);
2616 /* From here on, we're only interested in CALL_EXPRs. */
2617 if (TREE_CODE (t) != CALL_EXPR)
2618 goto egress;
2620 /* First, see if we can figure out what function is being called.
2621 If we cannot, then there is no hope of inlining the function. */
2622 fn = get_callee_fndecl (t);
2623 if (!fn)
2624 goto egress;
2626 /* Turn forward declarations into real ones. */
2627 fn = cgraph_node (fn)->decl;
2629 /* If fn is a declaration of a function in a nested scope that was
2630 globally declared inline, we don't set its DECL_INITIAL.
2631 However, we can't blindly follow DECL_ABSTRACT_ORIGIN because the
2632 C++ front-end uses it for cdtors to refer to their internal
2633 declarations, that are not real functions. Fortunately those
2634 don't have trees to be saved, so we can tell by checking their
2635 DECL_SAVED_TREE. */
2636 if (! DECL_INITIAL (fn)
2637 && DECL_ABSTRACT_ORIGIN (fn)
2638 && DECL_SAVED_TREE (DECL_ABSTRACT_ORIGIN (fn)))
2639 fn = DECL_ABSTRACT_ORIGIN (fn);
2641 /* Objective C and fortran still calls tree_rest_of_compilation directly.
2642 Kill this check once this is fixed. */
2643 if (!id->dst_node->analyzed)
2644 goto egress;
2646 cg_edge = cgraph_edge (id->dst_node, stmt);
2648 /* Constant propagation on argument done during previous inlining
2649 may create new direct call. Produce an edge for it. */
2650 if (!cg_edge)
2652 struct cgraph_node *dest = cgraph_node (fn);
2654 /* We have missing edge in the callgraph. This can happen in one case
2655 where previous inlining turned indirect call into direct call by
2656 constant propagating arguments. In all other cases we hit a bug
2657 (incorrect node sharing is most common reason for missing edges. */
2658 gcc_assert (dest->needed || !flag_unit_at_a_time);
2659 cgraph_create_edge (id->dst_node, dest, stmt,
2660 bb->count, CGRAPH_FREQ_BASE,
2661 bb->loop_depth)->inline_failed
2662 = N_("originally indirect function call not considered for inlining");
2663 if (dump_file)
2665 fprintf (dump_file, "Created new direct edge to %s",
2666 cgraph_node_name (dest));
2668 goto egress;
2671 /* Don't try to inline functions that are not well-suited to
2672 inlining. */
2673 if (!cgraph_inline_p (cg_edge, &reason))
2675 if (lookup_attribute ("always_inline", DECL_ATTRIBUTES (fn))
2676 /* Avoid warnings during early inline pass. */
2677 && (!flag_unit_at_a_time || cgraph_global_info_ready))
2679 sorry ("inlining failed in call to %q+F: %s", fn, reason);
2680 sorry ("called from here");
2682 else if (warn_inline && DECL_DECLARED_INLINE_P (fn)
2683 && !DECL_IN_SYSTEM_HEADER (fn)
2684 && strlen (reason)
2685 && !lookup_attribute ("noinline", DECL_ATTRIBUTES (fn))
2686 /* Avoid warnings during early inline pass. */
2687 && (!flag_unit_at_a_time || cgraph_global_info_ready))
2689 warning (OPT_Winline, "inlining failed in call to %q+F: %s",
2690 fn, reason);
2691 warning (OPT_Winline, "called from here");
2693 goto egress;
2695 fn = cg_edge->callee->decl;
2697 #ifdef ENABLE_CHECKING
2698 if (cg_edge->callee->decl != id->dst_node->decl)
2699 verify_cgraph_node (cg_edge->callee);
2700 #endif
2702 /* We will be inlining this callee. */
2703 id->eh_region = lookup_stmt_eh_region (stmt);
2705 /* Split the block holding the CALL_EXPR. */
2706 e = split_block (bb, stmt);
2707 bb = e->src;
2708 return_block = e->dest;
2709 remove_edge (e);
2711 /* split_block splits after the statement; work around this by
2712 moving the call into the second block manually. Not pretty,
2713 but seems easier than doing the CFG manipulation by hand
2714 when the CALL_EXPR is in the last statement of BB. */
2715 stmt_bsi = bsi_last (bb);
2716 bsi_remove (&stmt_bsi, false);
2718 /* If the CALL_EXPR was in the last statement of BB, it may have
2719 been the source of abnormal edges. In this case, schedule
2720 the removal of dead abnormal edges. */
2721 bsi = bsi_start (return_block);
2722 if (bsi_end_p (bsi))
2724 bsi_insert_after (&bsi, stmt, BSI_NEW_STMT);
2725 purge_dead_abnormal_edges = true;
2727 else
2729 bsi_insert_before (&bsi, stmt, BSI_NEW_STMT);
2730 purge_dead_abnormal_edges = false;
2733 stmt_bsi = bsi_start (return_block);
2735 /* Build a block containing code to initialize the arguments, the
2736 actual inline expansion of the body, and a label for the return
2737 statements within the function to jump to. The type of the
2738 statement expression is the return type of the function call. */
2739 id->block = make_node (BLOCK);
2740 BLOCK_ABSTRACT_ORIGIN (id->block) = fn;
2741 BLOCK_SOURCE_LOCATION (id->block) = input_location;
2742 add_lexical_block (TREE_BLOCK (stmt), id->block);
2744 /* Local declarations will be replaced by their equivalents in this
2745 map. */
2746 st = id->decl_map;
2747 id->decl_map = pointer_map_create ();
2749 /* Record the function we are about to inline. */
2750 id->src_fn = fn;
2751 id->src_node = cg_edge->callee;
2752 id->src_cfun = DECL_STRUCT_FUNCTION (fn);
2753 id->call_expr = t;
2755 gcc_assert (!id->src_cfun->after_inlining);
2757 id->entry_bb = bb;
2758 initialize_inlined_parameters (id, t, fn, bb);
2760 if (DECL_INITIAL (fn))
2761 add_lexical_block (id->block, remap_blocks (DECL_INITIAL (fn), id));
2763 /* Return statements in the function body will be replaced by jumps
2764 to the RET_LABEL. */
2766 gcc_assert (DECL_INITIAL (fn));
2767 gcc_assert (TREE_CODE (DECL_INITIAL (fn)) == BLOCK);
2769 /* Find the lhs to which the result of this call is assigned. */
2770 return_slot = NULL;
2771 if (TREE_CODE (stmt) == GIMPLE_MODIFY_STMT)
2773 modify_dest = GIMPLE_STMT_OPERAND (stmt, 0);
2775 /* The function which we are inlining might not return a value,
2776 in which case we should issue a warning that the function
2777 does not return a value. In that case the optimizers will
2778 see that the variable to which the value is assigned was not
2779 initialized. We do not want to issue a warning about that
2780 uninitialized variable. */
2781 if (DECL_P (modify_dest))
2782 TREE_NO_WARNING (modify_dest) = 1;
2783 if (CALL_EXPR_RETURN_SLOT_OPT (t))
2785 return_slot = modify_dest;
2786 modify_dest = NULL;
2789 else
2790 modify_dest = NULL;
2792 /* If we are inlining a call to the C++ operator new, we don't want
2793 to use type based alias analysis on the return value. Otherwise
2794 we may get confused if the compiler sees that the inlined new
2795 function returns a pointer which was just deleted. See bug
2796 33407. */
2797 if (DECL_IS_OPERATOR_NEW (fn))
2799 return_slot = NULL;
2800 modify_dest = NULL;
2803 /* Declare the return variable for the function. */
2804 retvar = declare_return_variable (id, return_slot,
2805 modify_dest, &use_retvar);
2807 if (DECL_IS_OPERATOR_NEW (fn))
2809 gcc_assert (TREE_CODE (retvar) == VAR_DECL
2810 && POINTER_TYPE_P (TREE_TYPE (retvar)));
2811 DECL_NO_TBAA_P (retvar) = 1;
2814 /* This is it. Duplicate the callee body. Assume callee is
2815 pre-gimplified. Note that we must not alter the caller
2816 function in any way before this point, as this CALL_EXPR may be
2817 a self-referential call; if we're calling ourselves, we need to
2818 duplicate our body before altering anything. */
2819 copy_body (id, bb->count, bb->frequency, bb, return_block);
2821 /* Add local vars in this inlined callee to caller. */
2822 t_step = id->src_cfun->local_decls;
2823 for (; t_step; t_step = TREE_CHAIN (t_step))
2825 var = TREE_VALUE (t_step);
2826 if (TREE_STATIC (var) && !TREE_ASM_WRITTEN (var))
2827 cfun->local_decls = tree_cons (NULL_TREE, var,
2828 cfun->local_decls);
2829 else
2830 cfun->local_decls = tree_cons (NULL_TREE, remap_decl (var, id),
2831 cfun->local_decls);
2834 /* Clean up. */
2835 pointer_map_destroy (id->decl_map);
2836 id->decl_map = st;
2838 /* If the inlined function returns a result that we care about,
2839 clobber the CALL_EXPR with a reference to the return variable. */
2840 if (use_retvar && (TREE_CODE (bsi_stmt (stmt_bsi)) != CALL_EXPR))
2842 *tp = use_retvar;
2843 if (gimple_in_ssa_p (cfun))
2845 update_stmt (stmt);
2846 mark_symbols_for_renaming (stmt);
2848 maybe_clean_or_replace_eh_stmt (stmt, stmt);
2850 else
2851 /* We're modifying a TSI owned by gimple_expand_calls_inline();
2852 tsi_delink() will leave the iterator in a sane state. */
2854 /* Handle case of inlining function that miss return statement so
2855 return value becomes undefined. */
2856 if (TREE_CODE (stmt) == GIMPLE_MODIFY_STMT
2857 && TREE_CODE (GIMPLE_STMT_OPERAND (stmt, 0)) == SSA_NAME)
2859 tree name = GIMPLE_STMT_OPERAND (stmt, 0);
2860 tree var = SSA_NAME_VAR (GIMPLE_STMT_OPERAND (stmt, 0));
2861 tree def = gimple_default_def (cfun, var);
2863 /* If the variable is used undefined, make this name undefined via
2864 move. */
2865 if (def)
2867 GIMPLE_STMT_OPERAND (stmt, 1) = def;
2868 update_stmt (stmt);
2870 /* Otherwise make this variable undefined. */
2871 else
2873 bsi_remove (&stmt_bsi, true);
2874 set_default_def (var, name);
2875 SSA_NAME_DEF_STMT (name) = build_empty_stmt ();
2878 else
2879 bsi_remove (&stmt_bsi, true);
2882 if (purge_dead_abnormal_edges)
2883 tree_purge_dead_abnormal_call_edges (return_block);
2885 /* If the value of the new expression is ignored, that's OK. We
2886 don't warn about this for CALL_EXPRs, so we shouldn't warn about
2887 the equivalent inlined version either. */
2888 TREE_USED (*tp) = 1;
2890 /* Output the inlining info for this abstract function, since it has been
2891 inlined. If we don't do this now, we can lose the information about the
2892 variables in the function when the blocks get blown away as soon as we
2893 remove the cgraph node. */
2894 (*debug_hooks->outlining_inline_function) (cg_edge->callee->decl);
2896 /* Update callgraph if needed. */
2897 cgraph_remove_node (cg_edge->callee);
2899 id->block = NULL_TREE;
2900 successfully_inlined = TRUE;
2902 egress:
2903 input_location = saved_location;
2904 return successfully_inlined;
2907 /* Expand call statements reachable from STMT_P.
2908 We can only have CALL_EXPRs as the "toplevel" tree code or nested
2909 in a GIMPLE_MODIFY_STMT. See tree-gimple.c:get_call_expr_in(). We can
2910 unfortunately not use that function here because we need a pointer
2911 to the CALL_EXPR, not the tree itself. */
2913 static bool
2914 gimple_expand_calls_inline (basic_block bb, copy_body_data *id)
2916 block_stmt_iterator bsi;
2918 /* Register specific tree functions. */
2919 tree_register_cfg_hooks ();
2920 for (bsi = bsi_start (bb); !bsi_end_p (bsi); bsi_next (&bsi))
2922 tree *expr_p = bsi_stmt_ptr (bsi);
2923 tree stmt = *expr_p;
2925 if (TREE_CODE (*expr_p) == GIMPLE_MODIFY_STMT)
2926 expr_p = &GIMPLE_STMT_OPERAND (*expr_p, 1);
2927 if (TREE_CODE (*expr_p) == WITH_SIZE_EXPR)
2928 expr_p = &TREE_OPERAND (*expr_p, 0);
2929 if (TREE_CODE (*expr_p) == CALL_EXPR)
2930 if (expand_call_inline (bb, stmt, expr_p, id))
2931 return true;
2933 return false;
2936 /* Walk all basic blocks created after FIRST and try to fold every statement
2937 in the STATEMENTS pointer set. */
2938 static void
2939 fold_marked_statements (int first, struct pointer_set_t *statements)
2941 for (;first < n_basic_blocks;first++)
2942 if (BASIC_BLOCK (first))
2944 block_stmt_iterator bsi;
2945 for (bsi = bsi_start (BASIC_BLOCK (first));
2946 !bsi_end_p (bsi); bsi_next (&bsi))
2947 if (pointer_set_contains (statements, bsi_stmt (bsi)))
2949 tree old_stmt = bsi_stmt (bsi);
2950 tree old_call = get_call_expr_in (old_stmt);
2952 if (fold_stmt (bsi_stmt_ptr (bsi)))
2954 update_stmt (bsi_stmt (bsi));
2955 if (old_call)
2956 cgraph_update_edges_for_call_stmt (old_stmt, old_call,
2957 bsi_stmt (bsi));
2958 if (maybe_clean_or_replace_eh_stmt (old_stmt,
2959 bsi_stmt (bsi)))
2960 tree_purge_dead_eh_edges (BASIC_BLOCK (first));
2966 /* Return true if BB has at least one abnormal outgoing edge. */
2968 static inline bool
2969 has_abnormal_outgoing_edge_p (basic_block bb)
2971 edge e;
2972 edge_iterator ei;
2974 FOR_EACH_EDGE (e, ei, bb->succs)
2975 if (e->flags & EDGE_ABNORMAL)
2976 return true;
2978 return false;
2981 /* Expand calls to inline functions in the body of FN. */
2983 unsigned int
2984 optimize_inline_calls (tree fn)
2986 copy_body_data id;
2987 tree prev_fn;
2988 basic_block bb;
2989 int last = n_basic_blocks;
2990 struct gimplify_ctx gctx;
2992 /* There is no point in performing inlining if errors have already
2993 occurred -- and we might crash if we try to inline invalid
2994 code. */
2995 if (errorcount || sorrycount)
2996 return 0;
2998 /* Clear out ID. */
2999 memset (&id, 0, sizeof (id));
3001 id.src_node = id.dst_node = cgraph_node (fn);
3002 id.dst_fn = fn;
3003 /* Or any functions that aren't finished yet. */
3004 prev_fn = NULL_TREE;
3005 if (current_function_decl)
3007 id.dst_fn = current_function_decl;
3008 prev_fn = current_function_decl;
3011 id.copy_decl = copy_decl_maybe_to_var;
3012 id.transform_call_graph_edges = CB_CGE_DUPLICATE;
3013 id.transform_new_cfg = false;
3014 id.transform_return_to_modify = true;
3015 id.transform_lang_insert_block = NULL;
3016 id.statements_to_fold = pointer_set_create ();
3018 push_gimplify_context (&gctx);
3020 /* We make no attempts to keep dominance info up-to-date. */
3021 free_dominance_info (CDI_DOMINATORS);
3022 free_dominance_info (CDI_POST_DOMINATORS);
3024 /* Reach the trees by walking over the CFG, and note the
3025 enclosing basic-blocks in the call edges. */
3026 /* We walk the blocks going forward, because inlined function bodies
3027 will split id->current_basic_block, and the new blocks will
3028 follow it; we'll trudge through them, processing their CALL_EXPRs
3029 along the way. */
3030 FOR_EACH_BB (bb)
3031 gimple_expand_calls_inline (bb, &id);
3033 pop_gimplify_context (NULL);
3035 #ifdef ENABLE_CHECKING
3037 struct cgraph_edge *e;
3039 verify_cgraph_node (id.dst_node);
3041 /* Double check that we inlined everything we are supposed to inline. */
3042 for (e = id.dst_node->callees; e; e = e->next_callee)
3043 gcc_assert (e->inline_failed);
3045 #endif
3047 /* Fold the statements before compacting/renumbering the basic blocks. */
3048 fold_marked_statements (last, id.statements_to_fold);
3049 pointer_set_destroy (id.statements_to_fold);
3051 /* Renumber the (code) basic_blocks consecutively. */
3052 compact_blocks ();
3053 /* Renumber the lexical scoping (non-code) blocks consecutively. */
3054 number_blocks (fn);
3056 /* We are not going to maintain the cgraph edges up to date.
3057 Kill it so it won't confuse us. */
3058 cgraph_node_remove_callees (id.dst_node);
3060 fold_cond_expr_cond ();
3061 /* It would be nice to check SSA/CFG/statement consistency here, but it is
3062 not possible yet - the IPA passes might make various functions to not
3063 throw and they don't care to proactively update local EH info. This is
3064 done later in fixup_cfg pass that also execute the verification. */
3065 return (TODO_update_ssa | TODO_cleanup_cfg
3066 | (gimple_in_ssa_p (cfun) ? TODO_remove_unused_locals : 0)
3067 | (profile_status != PROFILE_ABSENT ? TODO_rebuild_frequencies : 0));
3070 /* Passed to walk_tree. Copies the node pointed to, if appropriate. */
3072 tree
3073 copy_tree_r (tree *tp, int *walk_subtrees, void *data ATTRIBUTE_UNUSED)
3075 enum tree_code code = TREE_CODE (*tp);
3076 enum tree_code_class cl = TREE_CODE_CLASS (code);
3078 /* We make copies of most nodes. */
3079 if (IS_EXPR_CODE_CLASS (cl)
3080 || IS_GIMPLE_STMT_CODE_CLASS (cl)
3081 || code == TREE_LIST
3082 || code == TREE_VEC
3083 || code == TYPE_DECL
3084 || code == OMP_CLAUSE)
3086 /* Because the chain gets clobbered when we make a copy, we save it
3087 here. */
3088 tree chain = NULL_TREE, new;
3090 if (!GIMPLE_TUPLE_P (*tp))
3091 chain = TREE_CHAIN (*tp);
3093 /* Copy the node. */
3094 new = copy_node (*tp);
3096 /* Propagate mudflap marked-ness. */
3097 if (flag_mudflap && mf_marked_p (*tp))
3098 mf_mark (new);
3100 *tp = new;
3102 /* Now, restore the chain, if appropriate. That will cause
3103 walk_tree to walk into the chain as well. */
3104 if (code == PARM_DECL
3105 || code == TREE_LIST
3106 || code == OMP_CLAUSE)
3107 TREE_CHAIN (*tp) = chain;
3109 /* For now, we don't update BLOCKs when we make copies. So, we
3110 have to nullify all BIND_EXPRs. */
3111 if (TREE_CODE (*tp) == BIND_EXPR)
3112 BIND_EXPR_BLOCK (*tp) = NULL_TREE;
3114 else if (code == CONSTRUCTOR)
3116 /* CONSTRUCTOR nodes need special handling because
3117 we need to duplicate the vector of elements. */
3118 tree new;
3120 new = copy_node (*tp);
3122 /* Propagate mudflap marked-ness. */
3123 if (flag_mudflap && mf_marked_p (*tp))
3124 mf_mark (new);
3126 CONSTRUCTOR_ELTS (new) = VEC_copy (constructor_elt, gc,
3127 CONSTRUCTOR_ELTS (*tp));
3128 *tp = new;
3130 else if (TREE_CODE_CLASS (code) == tcc_type)
3131 *walk_subtrees = 0;
3132 else if (TREE_CODE_CLASS (code) == tcc_declaration)
3133 *walk_subtrees = 0;
3134 else if (TREE_CODE_CLASS (code) == tcc_constant)
3135 *walk_subtrees = 0;
3136 else
3137 gcc_assert (code != STATEMENT_LIST);
3138 return NULL_TREE;
3141 /* The SAVE_EXPR pointed to by TP is being copied. If ST contains
3142 information indicating to what new SAVE_EXPR this one should be mapped,
3143 use that one. Otherwise, create a new node and enter it in ST. FN is
3144 the function into which the copy will be placed. */
3146 static void
3147 remap_save_expr (tree *tp, void *st_, int *walk_subtrees)
3149 struct pointer_map_t *st = (struct pointer_map_t *) st_;
3150 tree *n;
3151 tree t;
3153 /* See if we already encountered this SAVE_EXPR. */
3154 n = (tree *) pointer_map_contains (st, *tp);
3156 /* If we didn't already remap this SAVE_EXPR, do so now. */
3157 if (!n)
3159 t = copy_node (*tp);
3161 /* Remember this SAVE_EXPR. */
3162 *pointer_map_insert (st, *tp) = t;
3163 /* Make sure we don't remap an already-remapped SAVE_EXPR. */
3164 *pointer_map_insert (st, t) = t;
3166 else
3168 /* We've already walked into this SAVE_EXPR; don't do it again. */
3169 *walk_subtrees = 0;
3170 t = *n;
3173 /* Replace this SAVE_EXPR with the copy. */
3174 *tp = t;
3177 /* Called via walk_tree. If *TP points to a DECL_STMT for a local label,
3178 copies the declaration and enters it in the splay_tree in DATA (which is
3179 really an `copy_body_data *'). */
3181 static tree
3182 mark_local_for_remap_r (tree *tp, int *walk_subtrees ATTRIBUTE_UNUSED,
3183 void *data)
3185 copy_body_data *id = (copy_body_data *) data;
3187 /* Don't walk into types. */
3188 if (TYPE_P (*tp))
3189 *walk_subtrees = 0;
3191 else if (TREE_CODE (*tp) == LABEL_EXPR)
3193 tree decl = TREE_OPERAND (*tp, 0);
3195 /* Copy the decl and remember the copy. */
3196 insert_decl_map (id, decl, id->copy_decl (decl, id));
3199 return NULL_TREE;
3202 /* Perform any modifications to EXPR required when it is unsaved. Does
3203 not recurse into EXPR's subtrees. */
3205 static void
3206 unsave_expr_1 (tree expr)
3208 switch (TREE_CODE (expr))
3210 case TARGET_EXPR:
3211 /* Don't mess with a TARGET_EXPR that hasn't been expanded.
3212 It's OK for this to happen if it was part of a subtree that
3213 isn't immediately expanded, such as operand 2 of another
3214 TARGET_EXPR. */
3215 if (TREE_OPERAND (expr, 1))
3216 break;
3218 TREE_OPERAND (expr, 1) = TREE_OPERAND (expr, 3);
3219 TREE_OPERAND (expr, 3) = NULL_TREE;
3220 break;
3222 default:
3223 break;
3227 /* Called via walk_tree when an expression is unsaved. Using the
3228 splay_tree pointed to by ST (which is really a `splay_tree'),
3229 remaps all local declarations to appropriate replacements. */
3231 static tree
3232 unsave_r (tree *tp, int *walk_subtrees, void *data)
3234 copy_body_data *id = (copy_body_data *) data;
3235 struct pointer_map_t *st = id->decl_map;
3236 tree *n;
3238 /* Only a local declaration (variable or label). */
3239 if ((TREE_CODE (*tp) == VAR_DECL && !TREE_STATIC (*tp))
3240 || TREE_CODE (*tp) == LABEL_DECL)
3242 /* Lookup the declaration. */
3243 n = (tree *) pointer_map_contains (st, *tp);
3245 /* If it's there, remap it. */
3246 if (n)
3247 *tp = *n;
3250 else if (TREE_CODE (*tp) == STATEMENT_LIST)
3251 copy_statement_list (tp);
3252 else if (TREE_CODE (*tp) == BIND_EXPR)
3253 copy_bind_expr (tp, walk_subtrees, id);
3254 else if (TREE_CODE (*tp) == SAVE_EXPR)
3255 remap_save_expr (tp, st, walk_subtrees);
3256 else
3258 copy_tree_r (tp, walk_subtrees, NULL);
3260 /* Do whatever unsaving is required. */
3261 unsave_expr_1 (*tp);
3264 /* Keep iterating. */
3265 return NULL_TREE;
3268 /* Copies everything in EXPR and replaces variables, labels
3269 and SAVE_EXPRs local to EXPR. */
3271 tree
3272 unsave_expr_now (tree expr)
3274 copy_body_data id;
3276 /* There's nothing to do for NULL_TREE. */
3277 if (expr == 0)
3278 return expr;
3280 /* Set up ID. */
3281 memset (&id, 0, sizeof (id));
3282 id.src_fn = current_function_decl;
3283 id.dst_fn = current_function_decl;
3284 id.decl_map = pointer_map_create ();
3286 id.copy_decl = copy_decl_no_change;
3287 id.transform_call_graph_edges = CB_CGE_DUPLICATE;
3288 id.transform_new_cfg = false;
3289 id.transform_return_to_modify = false;
3290 id.transform_lang_insert_block = NULL;
3292 /* Walk the tree once to find local labels. */
3293 walk_tree_without_duplicates (&expr, mark_local_for_remap_r, &id);
3295 /* Walk the tree again, copying, remapping, and unsaving. */
3296 walk_tree (&expr, unsave_r, &id, NULL);
3298 /* Clean up. */
3299 pointer_map_destroy (id.decl_map);
3301 return expr;
3304 /* Allow someone to determine if SEARCH is a child of TOP from gdb. */
3306 static tree
3307 debug_find_tree_1 (tree *tp, int *walk_subtrees ATTRIBUTE_UNUSED, void *data)
3309 if (*tp == data)
3310 return (tree) data;
3311 else
3312 return NULL;
3315 bool
3316 debug_find_tree (tree top, tree search)
3318 return walk_tree_without_duplicates (&top, debug_find_tree_1, search) != 0;
3322 /* Declare the variables created by the inliner. Add all the variables in
3323 VARS to BIND_EXPR. */
3325 static void
3326 declare_inline_vars (tree block, tree vars)
3328 tree t;
3329 for (t = vars; t; t = TREE_CHAIN (t))
3331 DECL_SEEN_IN_BIND_EXPR_P (t) = 1;
3332 gcc_assert (!TREE_STATIC (t) && !TREE_ASM_WRITTEN (t));
3333 cfun->local_decls = tree_cons (NULL_TREE, t, cfun->local_decls);
3336 if (block)
3337 BLOCK_VARS (block) = chainon (BLOCK_VARS (block), vars);
3341 /* Copy NODE (which must be a DECL). The DECL originally was in the FROM_FN,
3342 but now it will be in the TO_FN. PARM_TO_VAR means enable PARM_DECL to
3343 VAR_DECL translation. */
3345 static tree
3346 copy_decl_for_dup_finish (copy_body_data *id, tree decl, tree copy)
3348 /* Don't generate debug information for the copy if we wouldn't have
3349 generated it for the copy either. */
3350 DECL_ARTIFICIAL (copy) = DECL_ARTIFICIAL (decl);
3351 DECL_IGNORED_P (copy) = DECL_IGNORED_P (decl);
3353 /* Set the DECL_ABSTRACT_ORIGIN so the debugging routines know what
3354 declaration inspired this copy. */
3355 DECL_ABSTRACT_ORIGIN (copy) = DECL_ORIGIN (decl);
3357 /* The new variable/label has no RTL, yet. */
3358 if (CODE_CONTAINS_STRUCT (TREE_CODE (copy), TS_DECL_WRTL)
3359 && !TREE_STATIC (copy) && !DECL_EXTERNAL (copy))
3360 SET_DECL_RTL (copy, NULL_RTX);
3362 /* These args would always appear unused, if not for this. */
3363 TREE_USED (copy) = 1;
3365 /* Set the context for the new declaration. */
3366 if (!DECL_CONTEXT (decl))
3367 /* Globals stay global. */
3369 else if (DECL_CONTEXT (decl) != id->src_fn)
3370 /* Things that weren't in the scope of the function we're inlining
3371 from aren't in the scope we're inlining to, either. */
3373 else if (TREE_STATIC (decl))
3374 /* Function-scoped static variables should stay in the original
3375 function. */
3377 else
3378 /* Ordinary automatic local variables are now in the scope of the
3379 new function. */
3380 DECL_CONTEXT (copy) = id->dst_fn;
3382 return copy;
3385 static tree
3386 copy_decl_to_var (tree decl, copy_body_data *id)
3388 tree copy, type;
3390 gcc_assert (TREE_CODE (decl) == PARM_DECL
3391 || TREE_CODE (decl) == RESULT_DECL);
3393 type = TREE_TYPE (decl);
3395 copy = build_decl (VAR_DECL, DECL_NAME (decl), type);
3396 TREE_ADDRESSABLE (copy) = TREE_ADDRESSABLE (decl);
3397 TREE_READONLY (copy) = TREE_READONLY (decl);
3398 TREE_THIS_VOLATILE (copy) = TREE_THIS_VOLATILE (decl);
3399 DECL_GIMPLE_REG_P (copy) = DECL_GIMPLE_REG_P (decl);
3400 DECL_NO_TBAA_P (copy) = DECL_NO_TBAA_P (decl);
3402 return copy_decl_for_dup_finish (id, decl, copy);
3405 /* Like copy_decl_to_var, but create a return slot object instead of a
3406 pointer variable for return by invisible reference. */
3408 static tree
3409 copy_result_decl_to_var (tree decl, copy_body_data *id)
3411 tree copy, type;
3413 gcc_assert (TREE_CODE (decl) == PARM_DECL
3414 || TREE_CODE (decl) == RESULT_DECL);
3416 type = TREE_TYPE (decl);
3417 if (DECL_BY_REFERENCE (decl))
3418 type = TREE_TYPE (type);
3420 copy = build_decl (VAR_DECL, DECL_NAME (decl), type);
3421 TREE_READONLY (copy) = TREE_READONLY (decl);
3422 TREE_THIS_VOLATILE (copy) = TREE_THIS_VOLATILE (decl);
3423 if (!DECL_BY_REFERENCE (decl))
3425 TREE_ADDRESSABLE (copy) = TREE_ADDRESSABLE (decl);
3426 DECL_GIMPLE_REG_P (copy) = DECL_GIMPLE_REG_P (decl);
3427 DECL_NO_TBAA_P (copy) = DECL_NO_TBAA_P (decl);
3430 return copy_decl_for_dup_finish (id, decl, copy);
3434 tree
3435 copy_decl_no_change (tree decl, copy_body_data *id)
3437 tree copy;
3439 copy = copy_node (decl);
3441 /* The COPY is not abstract; it will be generated in DST_FN. */
3442 DECL_ABSTRACT (copy) = 0;
3443 lang_hooks.dup_lang_specific_decl (copy);
3445 /* TREE_ADDRESSABLE isn't used to indicate that a label's address has
3446 been taken; it's for internal bookkeeping in expand_goto_internal. */
3447 if (TREE_CODE (copy) == LABEL_DECL)
3449 TREE_ADDRESSABLE (copy) = 0;
3450 LABEL_DECL_UID (copy) = -1;
3453 return copy_decl_for_dup_finish (id, decl, copy);
3456 static tree
3457 copy_decl_maybe_to_var (tree decl, copy_body_data *id)
3459 if (TREE_CODE (decl) == PARM_DECL || TREE_CODE (decl) == RESULT_DECL)
3460 return copy_decl_to_var (decl, id);
3461 else
3462 return copy_decl_no_change (decl, id);
3465 /* Return a copy of the function's argument tree. */
3466 static tree
3467 copy_arguments_for_versioning (tree orig_parm, copy_body_data * id)
3469 tree *arg_copy, *parg;
3471 arg_copy = &orig_parm;
3472 for (parg = arg_copy; *parg; parg = &TREE_CHAIN (*parg))
3474 tree new = remap_decl (*parg, id);
3475 lang_hooks.dup_lang_specific_decl (new);
3476 TREE_CHAIN (new) = TREE_CHAIN (*parg);
3477 *parg = new;
3479 return orig_parm;
3482 /* Return a copy of the function's static chain. */
3483 static tree
3484 copy_static_chain (tree static_chain, copy_body_data * id)
3486 tree *chain_copy, *pvar;
3488 chain_copy = &static_chain;
3489 for (pvar = chain_copy; *pvar; pvar = &TREE_CHAIN (*pvar))
3491 tree new = remap_decl (*pvar, id);
3492 lang_hooks.dup_lang_specific_decl (new);
3493 TREE_CHAIN (new) = TREE_CHAIN (*pvar);
3494 *pvar = new;
3496 return static_chain;
3499 /* Return true if the function is allowed to be versioned.
3500 This is a guard for the versioning functionality. */
3501 bool
3502 tree_versionable_function_p (tree fndecl)
3504 if (fndecl == NULL_TREE)
3505 return false;
3506 /* ??? There are cases where a function is
3507 uninlinable but can be versioned. */
3508 if (!tree_inlinable_function_p (fndecl))
3509 return false;
3511 return true;
3514 /* Create a copy of a function's tree.
3515 OLD_DECL and NEW_DECL are FUNCTION_DECL tree nodes
3516 of the original function and the new copied function
3517 respectively. In case we want to replace a DECL
3518 tree with another tree while duplicating the function's
3519 body, TREE_MAP represents the mapping between these
3520 trees. If UPDATE_CLONES is set, the call_stmt fields
3521 of edges of clones of the function will be updated. */
3522 void
3523 tree_function_versioning (tree old_decl, tree new_decl, varray_type tree_map,
3524 bool update_clones)
3526 struct cgraph_node *old_version_node;
3527 struct cgraph_node *new_version_node;
3528 copy_body_data id;
3529 tree p;
3530 unsigned i;
3531 struct ipa_replace_map *replace_info;
3532 basic_block old_entry_block;
3533 tree t_step;
3534 tree old_current_function_decl = current_function_decl;
3536 gcc_assert (TREE_CODE (old_decl) == FUNCTION_DECL
3537 && TREE_CODE (new_decl) == FUNCTION_DECL);
3538 DECL_POSSIBLY_INLINED (old_decl) = 1;
3540 old_version_node = cgraph_node (old_decl);
3541 new_version_node = cgraph_node (new_decl);
3543 DECL_ARTIFICIAL (new_decl) = 1;
3544 DECL_ABSTRACT_ORIGIN (new_decl) = DECL_ORIGIN (old_decl);
3546 /* Prepare the data structures for the tree copy. */
3547 memset (&id, 0, sizeof (id));
3549 /* Generate a new name for the new version. */
3550 if (!update_clones)
3552 DECL_NAME (new_decl) = create_tmp_var_name (NULL);
3553 SET_DECL_ASSEMBLER_NAME (new_decl, DECL_NAME (new_decl));
3554 SET_DECL_RTL (new_decl, NULL_RTX);
3555 id.statements_to_fold = pointer_set_create ();
3558 id.decl_map = pointer_map_create ();
3559 id.src_fn = old_decl;
3560 id.dst_fn = new_decl;
3561 id.src_node = old_version_node;
3562 id.dst_node = new_version_node;
3563 id.src_cfun = DECL_STRUCT_FUNCTION (old_decl);
3565 id.copy_decl = copy_decl_no_change;
3566 id.transform_call_graph_edges
3567 = update_clones ? CB_CGE_MOVE_CLONES : CB_CGE_MOVE;
3568 id.transform_new_cfg = true;
3569 id.transform_return_to_modify = false;
3570 id.transform_lang_insert_block = NULL;
3572 current_function_decl = new_decl;
3573 old_entry_block = ENTRY_BLOCK_PTR_FOR_FUNCTION
3574 (DECL_STRUCT_FUNCTION (old_decl));
3575 initialize_cfun (new_decl, old_decl,
3576 old_entry_block->count,
3577 old_entry_block->frequency);
3578 push_cfun (DECL_STRUCT_FUNCTION (new_decl));
3580 /* Copy the function's static chain. */
3581 p = DECL_STRUCT_FUNCTION (old_decl)->static_chain_decl;
3582 if (p)
3583 DECL_STRUCT_FUNCTION (new_decl)->static_chain_decl =
3584 copy_static_chain (DECL_STRUCT_FUNCTION (old_decl)->static_chain_decl,
3585 &id);
3586 /* Copy the function's arguments. */
3587 if (DECL_ARGUMENTS (old_decl) != NULL_TREE)
3588 DECL_ARGUMENTS (new_decl) =
3589 copy_arguments_for_versioning (DECL_ARGUMENTS (old_decl), &id);
3591 /* If there's a tree_map, prepare for substitution. */
3592 if (tree_map)
3593 for (i = 0; i < VARRAY_ACTIVE_SIZE (tree_map); i++)
3595 replace_info = (struct ipa_replace_map *) VARRAY_GENERIC_PTR (tree_map, i);
3596 if (replace_info->replace_p)
3597 insert_decl_map (&id, replace_info->old_tree,
3598 replace_info->new_tree);
3601 DECL_INITIAL (new_decl) = remap_blocks (DECL_INITIAL (id.src_fn), &id);
3603 /* Renumber the lexical scoping (non-code) blocks consecutively. */
3604 number_blocks (id.dst_fn);
3606 if (DECL_STRUCT_FUNCTION (old_decl)->local_decls != NULL_TREE)
3607 /* Add local vars. */
3608 for (t_step = DECL_STRUCT_FUNCTION (old_decl)->local_decls;
3609 t_step; t_step = TREE_CHAIN (t_step))
3611 tree var = TREE_VALUE (t_step);
3612 if (TREE_STATIC (var) && !TREE_ASM_WRITTEN (var))
3613 cfun->local_decls = tree_cons (NULL_TREE, var, cfun->local_decls);
3614 else
3615 cfun->local_decls =
3616 tree_cons (NULL_TREE, remap_decl (var, &id),
3617 cfun->local_decls);
3620 /* Copy the Function's body. */
3621 copy_body (&id, old_entry_block->count, old_entry_block->frequency, ENTRY_BLOCK_PTR, EXIT_BLOCK_PTR);
3623 if (DECL_RESULT (old_decl) != NULL_TREE)
3625 tree *res_decl = &DECL_RESULT (old_decl);
3626 DECL_RESULT (new_decl) = remap_decl (*res_decl, &id);
3627 lang_hooks.dup_lang_specific_decl (DECL_RESULT (new_decl));
3630 /* Renumber the lexical scoping (non-code) blocks consecutively. */
3631 number_blocks (new_decl);
3633 /* Clean up. */
3634 pointer_map_destroy (id.decl_map);
3635 if (!update_clones)
3637 fold_marked_statements (0, id.statements_to_fold);
3638 pointer_set_destroy (id.statements_to_fold);
3639 fold_cond_expr_cond ();
3641 if (gimple_in_ssa_p (cfun))
3643 free_dominance_info (CDI_DOMINATORS);
3644 free_dominance_info (CDI_POST_DOMINATORS);
3645 if (!update_clones)
3646 delete_unreachable_blocks ();
3647 update_ssa (TODO_update_ssa);
3648 if (!update_clones)
3650 fold_cond_expr_cond ();
3651 if (need_ssa_update_p ())
3652 update_ssa (TODO_update_ssa);
3655 free_dominance_info (CDI_DOMINATORS);
3656 free_dominance_info (CDI_POST_DOMINATORS);
3657 pop_cfun ();
3658 current_function_decl = old_current_function_decl;
3659 gcc_assert (!current_function_decl
3660 || DECL_STRUCT_FUNCTION (current_function_decl) == cfun);
3661 return;
3664 /* Duplicate a type, fields and all. */
3666 tree
3667 build_duplicate_type (tree type)
3669 struct copy_body_data id;
3671 memset (&id, 0, sizeof (id));
3672 id.src_fn = current_function_decl;
3673 id.dst_fn = current_function_decl;
3674 id.src_cfun = cfun;
3675 id.decl_map = pointer_map_create ();
3676 id.copy_decl = copy_decl_no_change;
3678 type = remap_type_1 (type, &id);
3680 pointer_map_destroy (id.decl_map);
3682 TYPE_CANONICAL (type) = type;
3684 return type;
3687 /* Return whether it is safe to inline a function because it used different
3688 target specific options or different optimization options. */
3689 bool
3690 tree_can_inline_p (tree caller, tree callee)
3692 /* Don't inline a function with a higher optimization level than the
3693 caller, or with different space constraints (hot/cold functions). */
3694 tree caller_tree = DECL_FUNCTION_SPECIFIC_OPTIMIZATION (caller);
3695 tree callee_tree = DECL_FUNCTION_SPECIFIC_OPTIMIZATION (callee);
3697 if (caller_tree != callee_tree)
3699 struct cl_optimization *caller_opt
3700 = TREE_OPTIMIZATION ((caller_tree)
3701 ? caller_tree
3702 : optimization_default_node);
3704 struct cl_optimization *callee_opt
3705 = TREE_OPTIMIZATION ((callee_tree)
3706 ? callee_tree
3707 : optimization_default_node);
3709 if ((caller_opt->optimize > callee_opt->optimize)
3710 || (caller_opt->optimize_size != callee_opt->optimize_size))
3711 return false;
3714 /* Allow the backend to decide if inlining is ok. */
3715 return targetm.target_option.can_inline_p (caller, callee);