2 Copyright 2001, 2002, 2003, 2004, 2005, 2006, 2007
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)
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/>. */
24 #include "coretypes.h"
28 #include "tree-inline.h"
34 #include "insn-config.h"
37 #include "langhooks.h"
38 #include "basic-block.h"
39 #include "tree-iterator.h"
42 #include "tree-mudflap.h"
43 #include "tree-flow.h"
46 #include "tree-flow.h"
47 #include "diagnostic.h"
50 #include "pointer-set.h"
52 #include "value-prof.h"
53 #include "tree-pass.h"
55 #include "integrate.h"
57 /* I'm not real happy about this, but we need to handle gimple and
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
98 int flag_inline_trees = 0;
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
109 o Provide heuristics to clamp inlining of recursive template
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
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;
129 static tree declare_return_variable (copy_body_data *, tree, tree, tree *);
130 static tree copy_generic_body (copy_body_data *);
131 static bool inlinable_function_p (tree);
132 static void remap_block (tree *, copy_body_data *);
133 static tree remap_decls (tree, copy_body_data *);
134 static void copy_bind_expr (tree *, int *, copy_body_data *);
135 static tree mark_local_for_remap_r (tree *, int *, void *);
136 static void unsave_expr_1 (tree);
137 static tree unsave_r (tree *, int *, void *);
138 static void declare_inline_vars (tree, tree);
139 static void remap_save_expr (tree *, void *, int *);
140 static void add_lexical_block (tree current_block, tree new_block);
141 static tree copy_decl_to_var (tree, copy_body_data *);
142 static tree copy_result_decl_to_var (tree, copy_body_data *);
143 static tree copy_decl_no_change (tree, copy_body_data *);
144 static tree copy_decl_maybe_to_var (tree, copy_body_data *);
146 /* Insert a tree->tree mapping for ID. Despite the name suggests
147 that the trees should be variables, it is used for more than that. */
150 insert_decl_map (copy_body_data *id, tree key, tree value)
152 *pointer_map_insert (id->decl_map, key) = value;
154 /* Always insert an identity map as well. If we see this same new
155 node again, we won't want to duplicate it a second time. */
157 *pointer_map_insert (id->decl_map, value) = value;
160 /* Construct new SSA name for old NAME. ID is the inline context. */
163 remap_ssa_name (tree name, copy_body_data *id)
168 gcc_assert (TREE_CODE (name) == SSA_NAME);
170 n = (tree *) pointer_map_contains (id->decl_map, name);
174 /* Do not set DEF_STMT yet as statement is not copied yet. We do that
176 new = remap_decl (SSA_NAME_VAR (name), id);
177 /* We might've substituted constant or another SSA_NAME for
180 Replace the SSA name representing RESULT_DECL by variable during
181 inlining: this saves us from need to introduce PHI node in a case
182 return value is just partly initialized. */
183 if ((TREE_CODE (new) == VAR_DECL || TREE_CODE (new) == PARM_DECL)
184 && (TREE_CODE (SSA_NAME_VAR (name)) != RESULT_DECL
185 || !id->transform_return_to_modify))
187 new = make_ssa_name (new, NULL);
188 insert_decl_map (id, name, new);
189 if (IS_EMPTY_STMT (SSA_NAME_DEF_STMT (name)))
191 SSA_NAME_DEF_STMT (new) = build_empty_stmt ();
192 if (gimple_default_def (id->src_cfun, SSA_NAME_VAR (name)) == name)
193 set_default_def (SSA_NAME_VAR (new), new);
195 SSA_NAME_OCCURS_IN_ABNORMAL_PHI (new)
196 = SSA_NAME_OCCURS_IN_ABNORMAL_PHI (name);
197 TREE_TYPE (new) = TREE_TYPE (SSA_NAME_VAR (new));
200 insert_decl_map (id, name, new);
204 /* Remap DECL during the copying of the BLOCK tree for the function. */
207 remap_decl (tree decl, copy_body_data *id)
212 /* We only remap local variables in the current function. */
215 /* See if we have remapped this declaration. */
217 n = (tree *) pointer_map_contains (id->decl_map, decl);
219 /* If we didn't already have an equivalent for this declaration,
223 /* Make a copy of the variable or label. */
224 tree t = id->copy_decl (decl, id);
226 /* Remember it, so that if we encounter this local entity again
227 we can reuse this copy. Do this early because remap_type may
228 need this decl for TYPE_STUB_DECL. */
229 insert_decl_map (id, decl, t);
234 /* Remap types, if necessary. */
235 TREE_TYPE (t) = remap_type (TREE_TYPE (t), id);
236 if (TREE_CODE (t) == TYPE_DECL)
237 DECL_ORIGINAL_TYPE (t) = remap_type (DECL_ORIGINAL_TYPE (t), id);
239 /* Remap sizes as necessary. */
240 walk_tree (&DECL_SIZE (t), copy_body_r, id, NULL);
241 walk_tree (&DECL_SIZE_UNIT (t), copy_body_r, id, NULL);
243 /* If fields, do likewise for offset and qualifier. */
244 if (TREE_CODE (t) == FIELD_DECL)
246 walk_tree (&DECL_FIELD_OFFSET (t), copy_body_r, id, NULL);
247 if (TREE_CODE (DECL_CONTEXT (t)) == QUAL_UNION_TYPE)
248 walk_tree (&DECL_QUALIFIER (t), copy_body_r, id, NULL);
251 if (cfun && gimple_in_ssa_p (cfun)
252 && (TREE_CODE (t) == VAR_DECL
253 || TREE_CODE (t) == RESULT_DECL || TREE_CODE (t) == PARM_DECL))
255 tree def = gimple_default_def (id->src_cfun, decl);
257 if (TREE_CODE (decl) != PARM_DECL && def)
259 tree map = remap_ssa_name (def, id);
260 /* Watch out RESULT_DECLs whose SSA names map directly
262 if (TREE_CODE (map) == SSA_NAME)
263 set_default_def (t, map);
265 add_referenced_var (t);
270 return unshare_expr (*n);
274 remap_type_1 (tree type, copy_body_data *id)
278 /* We do need a copy. build and register it now. If this is a pointer or
279 reference type, remap the designated type and make a new pointer or
281 if (TREE_CODE (type) == POINTER_TYPE)
283 new = build_pointer_type_for_mode (remap_type (TREE_TYPE (type), id),
285 TYPE_REF_CAN_ALIAS_ALL (type));
286 insert_decl_map (id, type, new);
289 else if (TREE_CODE (type) == REFERENCE_TYPE)
291 new = build_reference_type_for_mode (remap_type (TREE_TYPE (type), id),
293 TYPE_REF_CAN_ALIAS_ALL (type));
294 insert_decl_map (id, type, new);
298 new = copy_node (type);
300 insert_decl_map (id, type, new);
302 /* This is a new type, not a copy of an old type. Need to reassociate
303 variants. We can handle everything except the main variant lazily. */
304 t = TYPE_MAIN_VARIANT (type);
307 t = remap_type (t, id);
308 TYPE_MAIN_VARIANT (new) = t;
309 TYPE_NEXT_VARIANT (new) = TYPE_NEXT_VARIANT (t);
310 TYPE_NEXT_VARIANT (t) = new;
314 TYPE_MAIN_VARIANT (new) = new;
315 TYPE_NEXT_VARIANT (new) = NULL;
318 if (TYPE_STUB_DECL (type))
319 TYPE_STUB_DECL (new) = remap_decl (TYPE_STUB_DECL (type), id);
321 /* Lazily create pointer and reference types. */
322 TYPE_POINTER_TO (new) = NULL;
323 TYPE_REFERENCE_TO (new) = NULL;
325 switch (TREE_CODE (new))
329 case FIXED_POINT_TYPE:
332 t = TYPE_MIN_VALUE (new);
333 if (t && TREE_CODE (t) != INTEGER_CST)
334 walk_tree (&TYPE_MIN_VALUE (new), copy_body_r, id, NULL);
336 t = TYPE_MAX_VALUE (new);
337 if (t && TREE_CODE (t) != INTEGER_CST)
338 walk_tree (&TYPE_MAX_VALUE (new), copy_body_r, id, NULL);
342 TREE_TYPE (new) = remap_type (TREE_TYPE (new), id);
343 walk_tree (&TYPE_ARG_TYPES (new), copy_body_r, id, NULL);
347 TREE_TYPE (new) = remap_type (TREE_TYPE (new), id);
348 TYPE_DOMAIN (new) = remap_type (TYPE_DOMAIN (new), id);
353 case QUAL_UNION_TYPE:
357 for (f = TYPE_FIELDS (new); f ; f = TREE_CHAIN (f))
359 t = remap_decl (f, id);
360 DECL_CONTEXT (t) = new;
364 TYPE_FIELDS (new) = nreverse (nf);
370 /* Shouldn't have been thought variable sized. */
374 walk_tree (&TYPE_SIZE (new), copy_body_r, id, NULL);
375 walk_tree (&TYPE_SIZE_UNIT (new), copy_body_r, id, NULL);
381 remap_type (tree type, copy_body_data *id)
388 /* See if we have remapped this type. */
389 node = (tree *) pointer_map_contains (id->decl_map, type);
393 /* The type only needs remapping if it's variably modified. */
394 if (! variably_modified_type_p (type, id->src_fn))
396 insert_decl_map (id, type, type);
400 return remap_type_1 (type, id);
404 remap_decls (tree decls, copy_body_data *id)
407 tree new_decls = NULL_TREE;
409 /* Remap its variables. */
410 for (old_var = decls; old_var; old_var = TREE_CHAIN (old_var))
414 /* We can not chain the local static declarations into the unexpanded_var_list
415 as we can't duplicate them or break one decl rule. Go ahead and link
416 them into unexpanded_var_list. */
417 if (!auto_var_in_fn_p (old_var, id->src_fn)
418 && !DECL_EXTERNAL (old_var))
420 cfun->unexpanded_var_list = tree_cons (NULL_TREE, old_var,
421 cfun->unexpanded_var_list);
425 /* Remap the variable. */
426 new_var = remap_decl (old_var, id);
428 /* If we didn't remap this variable, so we can't mess with its
429 TREE_CHAIN. If we remapped this variable to the return slot, it's
430 already declared somewhere else, so don't declare it here. */
431 if (!new_var || new_var == id->retvar)
435 gcc_assert (DECL_P (new_var));
436 TREE_CHAIN (new_var) = new_decls;
441 return nreverse (new_decls);
444 /* Copy the BLOCK to contain remapped versions of the variables
445 therein. And hook the new block into the block-tree. */
448 remap_block (tree *block, copy_body_data *id)
454 /* Make the new block. */
456 new_block = make_node (BLOCK);
457 TREE_USED (new_block) = TREE_USED (old_block);
458 BLOCK_ABSTRACT_ORIGIN (new_block) = old_block;
459 BLOCK_SOURCE_LOCATION (new_block) = BLOCK_SOURCE_LOCATION (old_block);
462 /* Remap its variables. */
463 BLOCK_VARS (new_block) = remap_decls (BLOCK_VARS (old_block), id);
467 if (id->transform_lang_insert_block)
468 lang_hooks.decls.insert_block (new_block);
470 /* Remember the remapped block. */
471 insert_decl_map (id, old_block, new_block);
474 /* Copy the whole block tree and root it in id->block. */
476 remap_blocks (tree block, copy_body_data *id)
484 remap_block (&new, id);
485 gcc_assert (new != block);
486 for (t = BLOCK_SUBBLOCKS (block); t ; t = BLOCK_CHAIN (t))
487 add_lexical_block (new, remap_blocks (t, id));
492 copy_statement_list (tree *tp)
494 tree_stmt_iterator oi, ni;
497 new = alloc_stmt_list ();
498 ni = tsi_start (new);
499 oi = tsi_start (*tp);
502 for (; !tsi_end_p (oi); tsi_next (&oi))
503 tsi_link_after (&ni, tsi_stmt (oi), TSI_NEW_STMT);
507 copy_bind_expr (tree *tp, int *walk_subtrees, copy_body_data *id)
509 tree block = BIND_EXPR_BLOCK (*tp);
510 /* Copy (and replace) the statement. */
511 copy_tree_r (tp, walk_subtrees, NULL);
514 remap_block (&block, id);
515 BIND_EXPR_BLOCK (*tp) = block;
518 if (BIND_EXPR_VARS (*tp))
519 /* This will remap a lot of the same decls again, but this should be
521 BIND_EXPR_VARS (*tp) = remap_decls (BIND_EXPR_VARS (*tp), id);
524 /* Called from copy_body_id via walk_tree. DATA is really an
525 `copy_body_data *'. */
528 copy_body_r (tree *tp, int *walk_subtrees, void *data)
530 copy_body_data *id = (copy_body_data *) data;
531 tree fn = id->src_fn;
534 /* Begin by recognizing trees that we'll completely rewrite for the
535 inlining context. Our output for these trees is completely
536 different from out input (e.g. RETURN_EXPR is deleted, and morphs
537 into an edge). Further down, we'll handle trees that get
538 duplicated and/or tweaked. */
540 /* When requested, RETURN_EXPRs should be transformed to just the
541 contained GIMPLE_MODIFY_STMT. The branch semantics of the return will
542 be handled elsewhere by manipulating the CFG rather than a statement. */
543 if (TREE_CODE (*tp) == RETURN_EXPR && id->transform_return_to_modify)
545 tree assignment = TREE_OPERAND (*tp, 0);
547 /* If we're returning something, just turn that into an
548 assignment into the equivalent of the original RESULT_DECL.
549 If the "assignment" is just the result decl, the result
550 decl has already been set (e.g. a recent "foo (&result_decl,
551 ...)"); just toss the entire RETURN_EXPR. */
552 if (assignment && TREE_CODE (assignment) == GIMPLE_MODIFY_STMT)
554 /* Replace the RETURN_EXPR with (a copy of) the
555 GIMPLE_MODIFY_STMT hanging underneath. */
556 *tp = copy_node (assignment);
558 else /* Else the RETURN_EXPR returns no value. */
561 return (tree) (void *)1;
564 else if (TREE_CODE (*tp) == SSA_NAME)
566 *tp = remap_ssa_name (*tp, id);
571 /* Local variables and labels need to be replaced by equivalent
572 variables. We don't want to copy static variables; there's only
573 one of those, no matter how many times we inline the containing
574 function. Similarly for globals from an outer function. */
575 else if (auto_var_in_fn_p (*tp, fn))
579 /* Remap the declaration. */
580 new_decl = remap_decl (*tp, id);
581 gcc_assert (new_decl);
582 /* Replace this variable with the copy. */
583 STRIP_TYPE_NOPS (new_decl);
587 else if (TREE_CODE (*tp) == STATEMENT_LIST)
588 copy_statement_list (tp);
589 else if (TREE_CODE (*tp) == SAVE_EXPR)
590 remap_save_expr (tp, id->decl_map, walk_subtrees);
591 else if (TREE_CODE (*tp) == LABEL_DECL
592 && (! DECL_CONTEXT (*tp)
593 || decl_function_context (*tp) == id->src_fn))
594 /* These may need to be remapped for EH handling. */
595 *tp = remap_decl (*tp, id);
596 else if (TREE_CODE (*tp) == BIND_EXPR)
597 copy_bind_expr (tp, walk_subtrees, id);
598 /* Types may need remapping as well. */
599 else if (TYPE_P (*tp))
600 *tp = remap_type (*tp, id);
602 /* If this is a constant, we have to copy the node iff the type will be
603 remapped. copy_tree_r will not copy a constant. */
604 else if (CONSTANT_CLASS_P (*tp))
606 tree new_type = remap_type (TREE_TYPE (*tp), id);
608 if (new_type == TREE_TYPE (*tp))
611 else if (TREE_CODE (*tp) == INTEGER_CST)
612 *tp = build_int_cst_wide (new_type, TREE_INT_CST_LOW (*tp),
613 TREE_INT_CST_HIGH (*tp));
616 *tp = copy_node (*tp);
617 TREE_TYPE (*tp) = new_type;
621 /* Otherwise, just copy the node. Note that copy_tree_r already
622 knows not to copy VAR_DECLs, etc., so this is safe. */
625 /* Here we handle trees that are not completely rewritten.
626 First we detect some inlining-induced bogosities for
628 if (TREE_CODE (*tp) == GIMPLE_MODIFY_STMT
629 && GIMPLE_STMT_OPERAND (*tp, 0) == GIMPLE_STMT_OPERAND (*tp, 1)
630 && (auto_var_in_fn_p (GIMPLE_STMT_OPERAND (*tp, 0), fn)))
632 /* Some assignments VAR = VAR; don't generate any rtl code
633 and thus don't count as variable modification. Avoid
634 keeping bogosities like 0 = 0. */
635 tree decl = GIMPLE_STMT_OPERAND (*tp, 0), value;
638 n = (tree *) pointer_map_contains (id->decl_map, decl);
642 STRIP_TYPE_NOPS (value);
643 if (TREE_CONSTANT (value) || TREE_READONLY_DECL_P (value))
645 *tp = build_empty_stmt ();
646 return copy_body_r (tp, walk_subtrees, data);
650 else if (TREE_CODE (*tp) == INDIRECT_REF)
652 /* Get rid of *& from inline substitutions that can happen when a
653 pointer argument is an ADDR_EXPR. */
654 tree decl = TREE_OPERAND (*tp, 0);
657 n = (tree *) pointer_map_contains (id->decl_map, decl);
662 /* If we happen to get an ADDR_EXPR in n->value, strip
663 it manually here as we'll eventually get ADDR_EXPRs
664 which lie about their types pointed to. In this case
665 build_fold_indirect_ref wouldn't strip the INDIRECT_REF,
666 but we absolutely rely on that. As fold_indirect_ref
667 does other useful transformations, try that first, though. */
668 tree type = TREE_TYPE (TREE_TYPE (*n));
669 new = unshare_expr (*n);
671 *tp = fold_indirect_ref_1 (type, new);
674 if (TREE_CODE (new) == ADDR_EXPR)
675 *tp = TREE_OPERAND (new, 0);
678 *tp = build1 (INDIRECT_REF, type, new);
679 TREE_THIS_VOLATILE (*tp) = TREE_THIS_VOLATILE (old);
687 /* Here is the "usual case". Copy this tree node, and then
688 tweak some special cases. */
689 copy_tree_r (tp, walk_subtrees, NULL);
691 /* Global variables we didn't seen yet needs to go into referenced
693 if (gimple_in_ssa_p (cfun) && TREE_CODE (*tp) == VAR_DECL)
694 add_referenced_var (*tp);
696 /* If EXPR has block defined, map it to newly constructed block.
697 When inlining we want EXPRs without block appear in the block
699 if (EXPR_P (*tp) || GIMPLE_STMT_P (*tp))
701 new_block = id->block;
702 if (TREE_BLOCK (*tp))
705 n = (tree *) pointer_map_contains (id->decl_map,
710 TREE_BLOCK (*tp) = new_block;
713 if (TREE_CODE (*tp) == RESX_EXPR && id->eh_region_offset)
714 TREE_OPERAND (*tp, 0) =
717 id->eh_region_offset + TREE_INT_CST_LOW (TREE_OPERAND (*tp, 0)));
719 if (!GIMPLE_TUPLE_P (*tp) && TREE_CODE (*tp) != OMP_CLAUSE)
720 TREE_TYPE (*tp) = remap_type (TREE_TYPE (*tp), id);
722 /* The copied TARGET_EXPR has never been expanded, even if the
723 original node was expanded already. */
724 if (TREE_CODE (*tp) == TARGET_EXPR && TREE_OPERAND (*tp, 3))
726 TREE_OPERAND (*tp, 1) = TREE_OPERAND (*tp, 3);
727 TREE_OPERAND (*tp, 3) = NULL_TREE;
730 /* Variable substitution need not be simple. In particular, the
731 INDIRECT_REF substitution above. Make sure that TREE_CONSTANT
732 and friends are up-to-date. */
733 else if (TREE_CODE (*tp) == ADDR_EXPR)
735 int invariant = TREE_INVARIANT (*tp);
736 walk_tree (&TREE_OPERAND (*tp, 0), copy_body_r, id, NULL);
737 /* Handle the case where we substituted an INDIRECT_REF
738 into the operand of the ADDR_EXPR. */
739 if (TREE_CODE (TREE_OPERAND (*tp, 0)) == INDIRECT_REF)
740 *tp = TREE_OPERAND (TREE_OPERAND (*tp, 0), 0);
742 recompute_tree_invariant_for_addr_expr (*tp);
743 /* If this used to be invariant, but is not any longer,
744 then regimplification is probably needed. */
745 if (invariant && !TREE_INVARIANT (*tp))
746 id->regimplify = true;
751 /* Keep iterating. */
755 /* Copy basic block, scale profile accordingly. Edges will be taken care of
759 copy_bb (copy_body_data *id, basic_block bb, int frequency_scale, int count_scale)
761 block_stmt_iterator bsi, copy_bsi;
762 basic_block copy_basic_block;
764 /* create_basic_block() will append every new block to
765 basic_block_info automatically. */
766 copy_basic_block = create_basic_block (NULL, (void *) 0,
767 (basic_block) bb->prev_bb->aux);
768 copy_basic_block->count = bb->count * count_scale / REG_BR_PROB_BASE;
770 /* We are going to rebuild frequencies from scratch. These values have just
771 small importance to drive canonicalize_loop_headers. */
772 copy_basic_block->frequency = ((gcov_type)bb->frequency
773 * frequency_scale / REG_BR_PROB_BASE);
774 if (copy_basic_block->frequency > BB_FREQ_MAX)
775 copy_basic_block->frequency = BB_FREQ_MAX;
776 copy_bsi = bsi_start (copy_basic_block);
778 for (bsi = bsi_start (bb);
779 !bsi_end_p (bsi); bsi_next (&bsi))
781 tree stmt = bsi_stmt (bsi);
782 tree orig_stmt = stmt;
784 id->regimplify = false;
785 walk_tree (&stmt, copy_body_r, id, NULL);
787 /* RETURN_EXPR might be removed,
788 this is signalled by making stmt pointer NULL. */
793 gimple_duplicate_stmt_histograms (cfun, stmt, id->src_cfun, orig_stmt);
795 /* With return slot optimization we can end up with
796 non-gimple (foo *)&this->m, fix that here. */
797 if ((TREE_CODE (stmt) == GIMPLE_MODIFY_STMT
798 && TREE_CODE (GIMPLE_STMT_OPERAND (stmt, 1)) == NOP_EXPR
799 && !is_gimple_val (TREE_OPERAND (GIMPLE_STMT_OPERAND (stmt, 1), 0)))
801 gimplify_stmt (&stmt);
803 bsi_insert_after (©_bsi, stmt, BSI_NEW_STMT);
805 /* Process new statement. gimplify_stmt possibly turned statement
806 into multiple statements, we need to process all of them. */
807 while (!bsi_end_p (copy_bsi))
809 tree *stmtp = bsi_stmt_ptr (copy_bsi);
811 call = get_call_expr_in (stmt);
813 if (call && CALL_EXPR_VA_ARG_PACK (call) && id->call_expr)
815 /* __builtin_va_arg_pack () should be replaced by
816 all arguments corresponding to ... in the caller. */
817 tree p, *argarray, new_call, *call_ptr;
818 int nargs = call_expr_nargs (id->call_expr);
820 for (p = DECL_ARGUMENTS (id->src_fn); p; p = TREE_CHAIN (p))
823 argarray = (tree *) alloca ((nargs + call_expr_nargs (call))
826 memcpy (argarray, CALL_EXPR_ARGP (call),
827 call_expr_nargs (call) * sizeof (*argarray));
828 memcpy (argarray + call_expr_nargs (call),
829 CALL_EXPR_ARGP (id->call_expr)
830 + (call_expr_nargs (id->call_expr) - nargs),
831 nargs * sizeof (*argarray));
833 new_call = build_call_array (TREE_TYPE (call),
835 nargs + call_expr_nargs (call),
837 /* Copy all CALL_EXPR flags, locus and block, except
838 CALL_EXPR_VA_ARG_PACK flag. */
839 CALL_EXPR_STATIC_CHAIN (new_call)
840 = CALL_EXPR_STATIC_CHAIN (call);
841 CALL_EXPR_TAILCALL (new_call) = CALL_EXPR_TAILCALL (call);
842 CALL_EXPR_RETURN_SLOT_OPT (new_call)
843 = CALL_EXPR_RETURN_SLOT_OPT (call);
844 CALL_FROM_THUNK_P (new_call) = CALL_FROM_THUNK_P (call);
845 CALL_CANNOT_INLINE_P (new_call)
846 = CALL_CANNOT_INLINE_P (call);
847 TREE_NOTHROW (new_call) = TREE_NOTHROW (call);
848 SET_EXPR_LOCUS (new_call, EXPR_LOCUS (call));
849 TREE_BLOCK (new_call) = TREE_BLOCK (call);
852 if (TREE_CODE (*call_ptr) == GIMPLE_MODIFY_STMT)
853 call_ptr = &GIMPLE_STMT_OPERAND (*call_ptr, 1);
854 if (TREE_CODE (*call_ptr) == WITH_SIZE_EXPR)
855 call_ptr = &TREE_OPERAND (*call_ptr, 0);
856 gcc_assert (*call_ptr == call);
857 if (call_ptr == stmtp)
859 bsi_replace (©_bsi, new_call, true);
860 stmtp = bsi_stmt_ptr (copy_bsi);
865 *call_ptr = new_call;
872 && (decl = get_callee_fndecl (call))
873 && DECL_BUILT_IN_CLASS (decl) == BUILT_IN_NORMAL
874 && DECL_FUNCTION_CODE (decl)
875 == BUILT_IN_VA_ARG_PACK_LEN)
877 /* __builtin_va_arg_pack_len () should be replaced by
878 the number of anonymous arguments. */
879 int nargs = call_expr_nargs (id->call_expr);
880 tree count, *call_ptr, p;
882 for (p = DECL_ARGUMENTS (id->src_fn); p; p = TREE_CHAIN (p))
885 count = build_int_cst (integer_type_node, nargs);
887 if (TREE_CODE (*call_ptr) == GIMPLE_MODIFY_STMT)
888 call_ptr = &GIMPLE_STMT_OPERAND (*call_ptr, 1);
889 if (TREE_CODE (*call_ptr) == WITH_SIZE_EXPR)
890 call_ptr = &TREE_OPERAND (*call_ptr, 0);
891 gcc_assert (*call_ptr == call && call_ptr != stmtp);
898 /* Statements produced by inlining can be unfolded, especially
899 when we constant propagated some operands. We can't fold
900 them right now for two reasons:
901 1) folding require SSA_NAME_DEF_STMTs to be correct
902 2) we can't change function calls to builtins.
903 So we just mark statement for later folding. We mark
904 all new statements, instead just statements that has changed
905 by some nontrivial substitution so even statements made
906 foldable indirectly are updated. If this turns out to be
907 expensive, copy_body can be told to watch for nontrivial
909 if (id->statements_to_fold)
910 pointer_set_insert (id->statements_to_fold, stmt);
911 /* We're duplicating a CALL_EXPR. Find any corresponding
912 callgraph edges and update or duplicate them. */
913 if (call && (decl = get_callee_fndecl (call)))
915 struct cgraph_node *node;
916 struct cgraph_edge *edge;
918 switch (id->transform_call_graph_edges)
920 case CB_CGE_DUPLICATE:
921 edge = cgraph_edge (id->src_node, orig_stmt);
923 cgraph_clone_edge (edge, id->dst_node, stmt,
924 REG_BR_PROB_BASE, 1, edge->frequency, true);
927 case CB_CGE_MOVE_CLONES:
928 for (node = id->dst_node->next_clone;
930 node = node->next_clone)
932 edge = cgraph_edge (node, orig_stmt);
934 cgraph_set_call_stmt (edge, stmt);
939 edge = cgraph_edge (id->dst_node, orig_stmt);
941 cgraph_set_call_stmt (edge, stmt);
948 /* If you think we can abort here, you are wrong.
949 There is no region 0 in tree land. */
950 gcc_assert (lookup_stmt_eh_region_fn (id->src_cfun, orig_stmt)
953 if (tree_could_throw_p (stmt)
954 /* When we are cloning for inlining, we are supposed to
955 construct a clone that calls precisely the same functions
956 as original. However IPA optimizers might've proved
957 earlier some function calls as non-trapping that might
958 render some basic blocks dead that might become
961 We can't update SSA with unreachable blocks in CFG and thus
962 we prevent the scenario by preserving even the "dead" eh
963 edges until the point they are later removed by
965 || (id->transform_call_graph_edges == CB_CGE_MOVE_CLONES
966 && lookup_stmt_eh_region_fn (id->src_cfun, orig_stmt) > 0))
968 int region = lookup_stmt_eh_region_fn (id->src_cfun, orig_stmt);
969 /* Add an entry for the copied tree in the EH hashtable.
970 When cloning or versioning, use the hashtable in
971 cfun, and just copy the EH number. When inlining, use the
972 hashtable in the caller, and adjust the region number. */
974 add_stmt_to_eh_region (stmt, region + id->eh_region_offset);
976 /* If this tree doesn't have a region associated with it,
977 and there is a "current region,"
978 then associate this tree with the current region
979 and add edges associated with this region. */
980 if ((lookup_stmt_eh_region_fn (id->src_cfun,
982 && id->eh_region > 0)
983 && tree_could_throw_p (stmt))
984 add_stmt_to_eh_region (stmt, id->eh_region);
986 if (gimple_in_ssa_p (cfun))
991 find_new_referenced_vars (bsi_stmt_ptr (copy_bsi));
992 FOR_EACH_SSA_TREE_OPERAND (def, stmt, i, SSA_OP_DEF)
993 if (TREE_CODE (def) == SSA_NAME)
994 SSA_NAME_DEF_STMT (def) = stmt;
996 bsi_next (©_bsi);
998 copy_bsi = bsi_last (copy_basic_block);
1001 return copy_basic_block;
1004 /* Inserting Single Entry Multiple Exit region in SSA form into code in SSA
1005 form is quite easy, since dominator relationship for old basic blocks does
1008 There is however exception where inlining might change dominator relation
1009 across EH edges from basic block within inlined functions destinating
1010 to landing pads in function we inline into.
1012 The function fills in PHI_RESULTs of such PHI nodes if they refer
1013 to gimple regs. Otherwise, the function mark PHI_RESULT of such
1014 PHI nodes for renaming. For non-gimple regs, renaming is safe: the
1015 EH edges are abnormal and SSA_NAME_OCCURS_IN_ABNORMAL_PHI must be
1016 set, and this means that there will be no overlapping live ranges
1017 for the underlying symbol.
1019 This might change in future if we allow redirecting of EH edges and
1020 we might want to change way build CFG pre-inlining to include
1021 all the possible edges then. */
1023 update_ssa_across_abnormal_edges (basic_block bb, basic_block ret_bb,
1024 bool can_throw, bool nonlocal_goto)
1029 FOR_EACH_EDGE (e, ei, bb->succs)
1031 || ((basic_block)e->dest->aux)->index == ENTRY_BLOCK)
1035 gcc_assert (e->flags & EDGE_ABNORMAL);
1037 gcc_assert (e->flags & EDGE_EH);
1039 gcc_assert (!(e->flags & EDGE_EH));
1040 for (phi = phi_nodes (e->dest); phi; phi = PHI_CHAIN (phi))
1044 /* There shouldn't be any PHI nodes in the ENTRY_BLOCK. */
1045 gcc_assert (!e->dest->aux);
1047 gcc_assert (SSA_NAME_OCCURS_IN_ABNORMAL_PHI
1048 (PHI_RESULT (phi)));
1050 if (!is_gimple_reg (PHI_RESULT (phi)))
1052 mark_sym_for_renaming
1053 (SSA_NAME_VAR (PHI_RESULT (phi)));
1057 re = find_edge (ret_bb, e->dest);
1059 gcc_assert ((re->flags & (EDGE_EH | EDGE_ABNORMAL))
1060 == (e->flags & (EDGE_EH | EDGE_ABNORMAL)));
1062 SET_USE (PHI_ARG_DEF_PTR_FROM_EDGE (phi, e),
1063 USE_FROM_PTR (PHI_ARG_DEF_PTR_FROM_EDGE (phi, re)));
1068 /* Copy edges from BB into its copy constructed earlier, scale profile
1069 accordingly. Edges will be taken care of later. Assume aux
1070 pointers to point to the copies of each BB. */
1072 copy_edges_for_bb (basic_block bb, int count_scale, basic_block ret_bb)
1074 basic_block new_bb = (basic_block) bb->aux;
1077 block_stmt_iterator bsi;
1080 /* Use the indices from the original blocks to create edges for the
1082 FOR_EACH_EDGE (old_edge, ei, bb->succs)
1083 if (!(old_edge->flags & EDGE_EH))
1087 flags = old_edge->flags;
1089 /* Return edges do get a FALLTHRU flag when the get inlined. */
1090 if (old_edge->dest->index == EXIT_BLOCK && !old_edge->flags
1091 && old_edge->dest->aux != EXIT_BLOCK_PTR)
1092 flags |= EDGE_FALLTHRU;
1093 new = make_edge (new_bb, (basic_block) old_edge->dest->aux, flags);
1094 new->count = old_edge->count * count_scale / REG_BR_PROB_BASE;
1095 new->probability = old_edge->probability;
1098 if (bb->index == ENTRY_BLOCK || bb->index == EXIT_BLOCK)
1101 for (bsi = bsi_start (new_bb); !bsi_end_p (bsi);)
1104 bool can_throw, nonlocal_goto;
1106 copy_stmt = bsi_stmt (bsi);
1107 update_stmt (copy_stmt);
1108 if (gimple_in_ssa_p (cfun))
1109 mark_symbols_for_renaming (copy_stmt);
1110 /* Do this before the possible split_block. */
1113 /* If this tree could throw an exception, there are two
1114 cases where we need to add abnormal edge(s): the
1115 tree wasn't in a region and there is a "current
1116 region" in the caller; or the original tree had
1117 EH edges. In both cases split the block after the tree,
1118 and add abnormal edge(s) as needed; we need both
1119 those from the callee and the caller.
1120 We check whether the copy can throw, because the const
1121 propagation can change an INDIRECT_REF which throws
1122 into a COMPONENT_REF which doesn't. If the copy
1123 can throw, the original could also throw. */
1125 can_throw = tree_can_throw_internal (copy_stmt);
1126 nonlocal_goto = tree_can_make_abnormal_goto (copy_stmt);
1128 if (can_throw || nonlocal_goto)
1130 if (!bsi_end_p (bsi))
1131 /* Note that bb's predecessor edges aren't necessarily
1132 right at this point; split_block doesn't care. */
1134 edge e = split_block (new_bb, copy_stmt);
1137 new_bb->aux = e->src->aux;
1138 bsi = bsi_start (new_bb);
1143 make_eh_edges (copy_stmt);
1146 make_abnormal_goto_edges (bb_for_stmt (copy_stmt), true);
1148 if ((can_throw || nonlocal_goto)
1149 && gimple_in_ssa_p (cfun))
1150 update_ssa_across_abnormal_edges (bb_for_stmt (copy_stmt), ret_bb,
1151 can_throw, nonlocal_goto);
1155 /* Copy the PHIs. All blocks and edges are copied, some blocks
1156 was possibly split and new outgoing EH edges inserted.
1157 BB points to the block of original function and AUX pointers links
1158 the original and newly copied blocks. */
1161 copy_phis_for_bb (basic_block bb, copy_body_data *id)
1163 basic_block new_bb = bb->aux;
1167 for (phi = phi_nodes (bb); phi; phi = PHI_CHAIN (phi))
1169 tree res = PHI_RESULT (phi);
1174 if (is_gimple_reg (res))
1176 walk_tree (&new_res, copy_body_r, id, NULL);
1177 SSA_NAME_DEF_STMT (new_res)
1178 = new_phi = create_phi_node (new_res, new_bb);
1179 FOR_EACH_EDGE (new_edge, ei, new_bb->preds)
1181 edge old_edge = find_edge (new_edge->src->aux, bb);
1182 tree arg = PHI_ARG_DEF_FROM_EDGE (phi, old_edge);
1185 walk_tree (&new_arg, copy_body_r, id, NULL);
1186 gcc_assert (new_arg);
1187 /* With return slot optimization we can end up with
1188 non-gimple (foo *)&this->m, fix that here. */
1189 if (TREE_CODE (new_arg) != SSA_NAME
1190 && TREE_CODE (new_arg) != FUNCTION_DECL
1191 && !is_gimple_val (new_arg))
1193 tree stmts = NULL_TREE;
1194 new_arg = force_gimple_operand (new_arg, &stmts,
1196 bsi_insert_on_edge_immediate (new_edge, stmts);
1198 add_phi_arg (new_phi, new_arg, new_edge);
1204 /* Wrapper for remap_decl so it can be used as a callback. */
1206 remap_decl_1 (tree decl, void *data)
1208 return remap_decl (decl, (copy_body_data *) data);
1211 /* Build struct function and associated datastructures for the new clone
1212 NEW_FNDECL to be build. CALLEE_FNDECL is the original */
1215 initialize_cfun (tree new_fndecl, tree callee_fndecl, gcov_type count,
1218 struct function *new_cfun
1219 = (struct function *) ggc_alloc_cleared (sizeof (struct function));
1220 struct function *src_cfun = DECL_STRUCT_FUNCTION (callee_fndecl);
1221 int count_scale, frequency_scale;
1223 if (ENTRY_BLOCK_PTR_FOR_FUNCTION (src_cfun)->count)
1224 count_scale = (REG_BR_PROB_BASE * count
1225 / ENTRY_BLOCK_PTR_FOR_FUNCTION (src_cfun)->count);
1229 if (ENTRY_BLOCK_PTR_FOR_FUNCTION (src_cfun)->frequency)
1230 frequency_scale = (REG_BR_PROB_BASE * frequency
1232 ENTRY_BLOCK_PTR_FOR_FUNCTION (src_cfun)->frequency);
1234 frequency_scale = count_scale;
1236 /* Register specific tree functions. */
1237 tree_register_cfg_hooks ();
1238 *new_cfun = *DECL_STRUCT_FUNCTION (callee_fndecl);
1239 new_cfun->funcdef_no = get_next_funcdef_no ();
1240 VALUE_HISTOGRAMS (new_cfun) = NULL;
1241 new_cfun->unexpanded_var_list = NULL;
1242 new_cfun->cfg = NULL;
1243 new_cfun->decl = new_fndecl /*= copy_node (callee_fndecl)*/;
1244 DECL_STRUCT_FUNCTION (new_fndecl) = new_cfun;
1245 push_cfun (new_cfun);
1246 init_empty_tree_cfg ();
1248 ENTRY_BLOCK_PTR->count =
1249 (ENTRY_BLOCK_PTR_FOR_FUNCTION (src_cfun)->count * count_scale /
1251 ENTRY_BLOCK_PTR->frequency =
1252 (ENTRY_BLOCK_PTR_FOR_FUNCTION (src_cfun)->frequency *
1253 frequency_scale / REG_BR_PROB_BASE);
1254 EXIT_BLOCK_PTR->count =
1255 (EXIT_BLOCK_PTR_FOR_FUNCTION (src_cfun)->count * count_scale /
1257 EXIT_BLOCK_PTR->frequency =
1258 (EXIT_BLOCK_PTR_FOR_FUNCTION (src_cfun)->frequency *
1259 frequency_scale / REG_BR_PROB_BASE);
1261 init_eh_for_function ();
1263 if (src_cfun->gimple_df)
1266 cfun->gimple_df->in_ssa_p = true;
1267 init_ssa_operands ();
1272 /* Make a copy of the body of FN so that it can be inserted inline in
1273 another function. Walks FN via CFG, returns new fndecl. */
1276 copy_cfg_body (copy_body_data * id, gcov_type count, int frequency,
1277 basic_block entry_block_map, basic_block exit_block_map)
1279 tree callee_fndecl = id->src_fn;
1280 /* Original cfun for the callee, doesn't change. */
1281 struct function *src_cfun = DECL_STRUCT_FUNCTION (callee_fndecl);
1282 struct function *cfun_to_copy;
1284 tree new_fndecl = NULL;
1285 int count_scale, frequency_scale;
1288 if (ENTRY_BLOCK_PTR_FOR_FUNCTION (src_cfun)->count)
1289 count_scale = (REG_BR_PROB_BASE * count
1290 / ENTRY_BLOCK_PTR_FOR_FUNCTION (src_cfun)->count);
1294 if (ENTRY_BLOCK_PTR_FOR_FUNCTION (src_cfun)->frequency)
1295 frequency_scale = (REG_BR_PROB_BASE * frequency
1297 ENTRY_BLOCK_PTR_FOR_FUNCTION (src_cfun)->frequency);
1299 frequency_scale = count_scale;
1301 /* Register specific tree functions. */
1302 tree_register_cfg_hooks ();
1304 /* Must have a CFG here at this point. */
1305 gcc_assert (ENTRY_BLOCK_PTR_FOR_FUNCTION
1306 (DECL_STRUCT_FUNCTION (callee_fndecl)));
1308 cfun_to_copy = id->src_cfun = DECL_STRUCT_FUNCTION (callee_fndecl);
1311 ENTRY_BLOCK_PTR_FOR_FUNCTION (cfun_to_copy)->aux = entry_block_map;
1312 EXIT_BLOCK_PTR_FOR_FUNCTION (cfun_to_copy)->aux = exit_block_map;
1313 entry_block_map->aux = ENTRY_BLOCK_PTR_FOR_FUNCTION (cfun_to_copy);
1314 exit_block_map->aux = EXIT_BLOCK_PTR_FOR_FUNCTION (cfun_to_copy);
1316 /* Duplicate any exception-handling regions. */
1319 id->eh_region_offset
1320 = duplicate_eh_regions (cfun_to_copy, remap_decl_1, id,
1323 /* Use aux pointers to map the original blocks to copy. */
1324 FOR_EACH_BB_FN (bb, cfun_to_copy)
1326 basic_block new = copy_bb (id, bb, frequency_scale, count_scale);
1331 last = last_basic_block;
1332 /* Now that we've duplicated the blocks, duplicate their edges. */
1333 FOR_ALL_BB_FN (bb, cfun_to_copy)
1334 copy_edges_for_bb (bb, count_scale, exit_block_map);
1335 if (gimple_in_ssa_p (cfun))
1336 FOR_ALL_BB_FN (bb, cfun_to_copy)
1337 copy_phis_for_bb (bb, id);
1338 FOR_ALL_BB_FN (bb, cfun_to_copy)
1340 ((basic_block)bb->aux)->aux = NULL;
1343 /* Zero out AUX fields of newly created block during EH edge
1345 for (; last < last_basic_block; last++)
1346 BASIC_BLOCK (last)->aux = NULL;
1347 entry_block_map->aux = NULL;
1348 exit_block_map->aux = NULL;
1353 /* Make a copy of the body of FN so that it can be inserted inline in
1354 another function. */
1357 copy_generic_body (copy_body_data *id)
1360 tree fndecl = id->src_fn;
1362 body = DECL_SAVED_TREE (fndecl);
1363 walk_tree (&body, copy_body_r, id, NULL);
1369 copy_body (copy_body_data *id, gcov_type count, int frequency,
1370 basic_block entry_block_map, basic_block exit_block_map)
1372 tree fndecl = id->src_fn;
1375 /* If this body has a CFG, walk CFG and copy. */
1376 gcc_assert (ENTRY_BLOCK_PTR_FOR_FUNCTION (DECL_STRUCT_FUNCTION (fndecl)));
1377 body = copy_cfg_body (id, count, frequency, entry_block_map, exit_block_map);
1382 /* Return true if VALUE is an ADDR_EXPR of an automatic variable
1383 defined in function FN, or of a data member thereof. */
1386 self_inlining_addr_expr (tree value, tree fn)
1390 if (TREE_CODE (value) != ADDR_EXPR)
1393 var = get_base_address (TREE_OPERAND (value, 0));
1395 return var && auto_var_in_fn_p (var, fn);
1399 setup_one_parameter (copy_body_data *id, tree p, tree value, tree fn,
1400 basic_block bb, tree *vars)
1406 tree def = (gimple_in_ssa_p (cfun)
1407 ? gimple_default_def (id->src_cfun, p) : NULL);
1410 && value != error_mark_node
1411 && !useless_type_conversion_p (TREE_TYPE (p), TREE_TYPE (value)))
1412 rhs = fold_build1 (NOP_EXPR, TREE_TYPE (p), value);
1414 /* If the parameter is never assigned to, has no SSA_NAMEs created,
1415 we may not need to create a new variable here at all. Instead, we may
1416 be able to just use the argument value. */
1417 if (TREE_READONLY (p)
1418 && !TREE_ADDRESSABLE (p)
1419 && value && !TREE_SIDE_EFFECTS (value)
1422 /* We may produce non-gimple trees by adding NOPs or introduce
1423 invalid sharing when operand is not really constant.
1424 It is not big deal to prohibit constant propagation here as
1425 we will constant propagate in DOM1 pass anyway. */
1426 if (is_gimple_min_invariant (value)
1427 && useless_type_conversion_p (TREE_TYPE (p),
1429 /* We have to be very careful about ADDR_EXPR. Make sure
1430 the base variable isn't a local variable of the inlined
1431 function, e.g., when doing recursive inlining, direct or
1432 mutually-recursive or whatever, which is why we don't
1433 just test whether fn == current_function_decl. */
1434 && ! self_inlining_addr_expr (value, fn))
1436 insert_decl_map (id, p, value);
1441 /* Make an equivalent VAR_DECL. Note that we must NOT remap the type
1442 here since the type of this decl must be visible to the calling
1444 var = copy_decl_to_var (p, id);
1445 if (gimple_in_ssa_p (cfun) && TREE_CODE (var) == VAR_DECL)
1448 add_referenced_var (var);
1451 /* See if the frontend wants to pass this by invisible reference. If
1452 so, our new VAR_DECL will have REFERENCE_TYPE, and we need to
1453 replace uses of the PARM_DECL with dereferences. */
1454 if (TREE_TYPE (var) != TREE_TYPE (p)
1455 && POINTER_TYPE_P (TREE_TYPE (var))
1456 && TREE_TYPE (TREE_TYPE (var)) == TREE_TYPE (p))
1458 insert_decl_map (id, var, var);
1459 var_sub = build_fold_indirect_ref (var);
1464 /* Register the VAR_DECL as the equivalent for the PARM_DECL;
1465 that way, when the PARM_DECL is encountered, it will be
1466 automatically replaced by the VAR_DECL. */
1467 insert_decl_map (id, p, var_sub);
1469 /* Declare this new variable. */
1470 TREE_CHAIN (var) = *vars;
1473 /* Make gimplifier happy about this variable. */
1474 DECL_SEEN_IN_BIND_EXPR_P (var) = 1;
1476 /* Even if P was TREE_READONLY, the new VAR should not be.
1477 In the original code, we would have constructed a
1478 temporary, and then the function body would have never
1479 changed the value of P. However, now, we will be
1480 constructing VAR directly. The constructor body may
1481 change its value multiple times as it is being
1482 constructed. Therefore, it must not be TREE_READONLY;
1483 the back-end assumes that TREE_READONLY variable is
1484 assigned to only once. */
1485 if (TYPE_NEEDS_CONSTRUCTING (TREE_TYPE (p)))
1486 TREE_READONLY (var) = 0;
1488 /* If there is no setup required and we are in SSA, take the easy route
1489 replacing all SSA names representing the function parameter by the
1490 SSA name passed to function.
1492 We need to construct map for the variable anyway as it might be used
1493 in different SSA names when parameter is set in function.
1495 FIXME: This usually kills the last connection in between inlined
1496 function parameter and the actual value in debug info. Can we do
1497 better here? If we just inserted the statement, copy propagation
1498 would kill it anyway as it always did in older versions of GCC.
1500 We might want to introduce a notion that single SSA_NAME might
1501 represent multiple variables for purposes of debugging. */
1502 if (gimple_in_ssa_p (cfun) && rhs && def && is_gimple_reg (p)
1503 && (TREE_CODE (rhs) == SSA_NAME
1504 || is_gimple_min_invariant (rhs))
1505 && !SSA_NAME_OCCURS_IN_ABNORMAL_PHI (def))
1507 insert_decl_map (id, def, rhs);
1511 /* If the value of argument is never used, don't care about initializing
1513 if (gimple_in_ssa_p (cfun) && !def && is_gimple_reg (p))
1515 gcc_assert (!value || !TREE_SIDE_EFFECTS (value));
1519 /* Initialize this VAR_DECL from the equivalent argument. Convert
1520 the argument to the proper type in case it was promoted. */
1523 block_stmt_iterator bsi = bsi_last (bb);
1525 if (rhs == error_mark_node)
1527 insert_decl_map (id, p, var_sub);
1531 STRIP_USELESS_TYPE_CONVERSION (rhs);
1533 /* We want to use GIMPLE_MODIFY_STMT, not INIT_EXPR here so that we
1534 keep our trees in gimple form. */
1535 if (def && gimple_in_ssa_p (cfun) && is_gimple_reg (p))
1537 def = remap_ssa_name (def, id);
1538 init_stmt = build_gimple_modify_stmt (def, rhs);
1539 SSA_NAME_DEF_STMT (def) = init_stmt;
1540 SSA_NAME_IS_DEFAULT_DEF (def) = 0;
1541 set_default_def (var, NULL);
1544 init_stmt = build_gimple_modify_stmt (var, rhs);
1546 /* If we did not create a gimple value and we did not create a gimple
1547 cast of a gimple value, then we will need to gimplify INIT_STMTS
1548 at the end. Note that is_gimple_cast only checks the outer
1549 tree code, not its operand. Thus the explicit check that its
1550 operand is a gimple value. */
1551 if ((!is_gimple_val (rhs)
1552 && (!is_gimple_cast (rhs)
1553 || !is_gimple_val (TREE_OPERAND (rhs, 0))))
1554 || !is_gimple_reg (var))
1556 tree_stmt_iterator i;
1558 push_gimplify_context ();
1559 gimplify_stmt (&init_stmt);
1560 if (gimple_in_ssa_p (cfun)
1561 && init_stmt && TREE_CODE (init_stmt) == STATEMENT_LIST)
1563 /* The replacement can expose previously unreferenced
1565 for (i = tsi_start (init_stmt); !tsi_end_p (i); tsi_next (&i))
1566 find_new_referenced_vars (tsi_stmt_ptr (i));
1568 pop_gimplify_context (NULL);
1571 /* If VAR represents a zero-sized variable, it's possible that the
1572 assignment statment may result in no gimple statements. */
1574 bsi_insert_after (&bsi, init_stmt, BSI_NEW_STMT);
1575 if (gimple_in_ssa_p (cfun))
1576 for (;!bsi_end_p (bsi); bsi_next (&bsi))
1577 mark_symbols_for_renaming (bsi_stmt (bsi));
1581 /* Generate code to initialize the parameters of the function at the
1582 top of the stack in ID from the CALL_EXPR EXP. */
1585 initialize_inlined_parameters (copy_body_data *id, tree exp,
1586 tree fn, basic_block bb)
1591 tree vars = NULL_TREE;
1592 call_expr_arg_iterator iter;
1593 tree static_chain = CALL_EXPR_STATIC_CHAIN (exp);
1595 /* Figure out what the parameters are. */
1596 parms = DECL_ARGUMENTS (fn);
1598 /* Loop through the parameter declarations, replacing each with an
1599 equivalent VAR_DECL, appropriately initialized. */
1600 for (p = parms, a = first_call_expr_arg (exp, &iter); p;
1601 a = next_call_expr_arg (&iter), p = TREE_CHAIN (p))
1602 setup_one_parameter (id, p, a, fn, bb, &vars);
1604 /* Initialize the static chain. */
1605 p = DECL_STRUCT_FUNCTION (fn)->static_chain_decl;
1606 gcc_assert (fn != current_function_decl);
1609 /* No static chain? Seems like a bug in tree-nested.c. */
1610 gcc_assert (static_chain);
1612 setup_one_parameter (id, p, static_chain, fn, bb, &vars);
1615 declare_inline_vars (id->block, vars);
1618 /* Declare a return variable to replace the RESULT_DECL for the
1619 function we are calling. An appropriate DECL_STMT is returned.
1620 The USE_STMT is filled to contain a use of the declaration to
1621 indicate the return value of the function.
1623 RETURN_SLOT, if non-null is place where to store the result. It
1624 is set only for CALL_EXPR_RETURN_SLOT_OPT. MODIFY_DEST, if non-null,
1625 was the LHS of the GIMPLE_MODIFY_STMT to which this call is the RHS.
1627 The return value is a (possibly null) value that is the result of the
1628 function as seen by the callee. *USE_P is a (possibly null) value that
1629 holds the result as seen by the caller. */
1632 declare_return_variable (copy_body_data *id, tree return_slot, tree modify_dest,
1635 tree callee = id->src_fn;
1636 tree caller = id->dst_fn;
1637 tree result = DECL_RESULT (callee);
1638 tree callee_type = TREE_TYPE (result);
1639 tree caller_type = TREE_TYPE (TREE_TYPE (callee));
1642 /* We don't need to do anything for functions that don't return
1644 if (!result || VOID_TYPE_P (callee_type))
1650 /* If there was a return slot, then the return value is the
1651 dereferenced address of that object. */
1654 /* The front end shouldn't have used both return_slot and
1655 a modify expression. */
1656 gcc_assert (!modify_dest);
1657 if (DECL_BY_REFERENCE (result))
1659 tree return_slot_addr = build_fold_addr_expr (return_slot);
1660 STRIP_USELESS_TYPE_CONVERSION (return_slot_addr);
1662 /* We are going to construct *&return_slot and we can't do that
1663 for variables believed to be not addressable.
1665 FIXME: This check possibly can match, because values returned
1666 via return slot optimization are not believed to have address
1667 taken by alias analysis. */
1668 gcc_assert (TREE_CODE (return_slot) != SSA_NAME);
1669 if (gimple_in_ssa_p (cfun))
1671 HOST_WIDE_INT bitsize;
1672 HOST_WIDE_INT bitpos;
1674 enum machine_mode mode;
1678 base = get_inner_reference (return_slot, &bitsize, &bitpos,
1680 &mode, &unsignedp, &volatilep,
1682 if (TREE_CODE (base) == INDIRECT_REF)
1683 base = TREE_OPERAND (base, 0);
1684 if (TREE_CODE (base) == SSA_NAME)
1685 base = SSA_NAME_VAR (base);
1686 mark_sym_for_renaming (base);
1688 var = return_slot_addr;
1693 gcc_assert (TREE_CODE (var) != SSA_NAME);
1695 if ((TREE_CODE (TREE_TYPE (result)) == COMPLEX_TYPE
1696 || TREE_CODE (TREE_TYPE (result)) == VECTOR_TYPE)
1697 && !DECL_GIMPLE_REG_P (result)
1699 DECL_GIMPLE_REG_P (var) = 0;
1704 /* All types requiring non-trivial constructors should have been handled. */
1705 gcc_assert (!TREE_ADDRESSABLE (callee_type));
1707 /* Attempt to avoid creating a new temporary variable. */
1709 && TREE_CODE (modify_dest) != SSA_NAME)
1711 bool use_it = false;
1713 /* We can't use MODIFY_DEST if there's type promotion involved. */
1714 if (!useless_type_conversion_p (callee_type, caller_type))
1717 /* ??? If we're assigning to a variable sized type, then we must
1718 reuse the destination variable, because we've no good way to
1719 create variable sized temporaries at this point. */
1720 else if (TREE_CODE (TYPE_SIZE_UNIT (caller_type)) != INTEGER_CST)
1723 /* If the callee cannot possibly modify MODIFY_DEST, then we can
1724 reuse it as the result of the call directly. Don't do this if
1725 it would promote MODIFY_DEST to addressable. */
1726 else if (TREE_ADDRESSABLE (result))
1730 tree base_m = get_base_address (modify_dest);
1732 /* If the base isn't a decl, then it's a pointer, and we don't
1733 know where that's going to go. */
1734 if (!DECL_P (base_m))
1736 else if (is_global_var (base_m))
1738 else if ((TREE_CODE (TREE_TYPE (result)) == COMPLEX_TYPE
1739 || TREE_CODE (TREE_TYPE (result)) == VECTOR_TYPE)
1740 && !DECL_GIMPLE_REG_P (result)
1741 && DECL_GIMPLE_REG_P (base_m))
1743 else if (!TREE_ADDRESSABLE (base_m))
1755 gcc_assert (TREE_CODE (TYPE_SIZE_UNIT (callee_type)) == INTEGER_CST);
1757 var = copy_result_decl_to_var (result, id);
1758 if (gimple_in_ssa_p (cfun))
1761 add_referenced_var (var);
1764 DECL_SEEN_IN_BIND_EXPR_P (var) = 1;
1765 DECL_STRUCT_FUNCTION (caller)->unexpanded_var_list
1766 = tree_cons (NULL_TREE, var,
1767 DECL_STRUCT_FUNCTION (caller)->unexpanded_var_list);
1769 /* Do not have the rest of GCC warn about this variable as it should
1770 not be visible to the user. */
1771 TREE_NO_WARNING (var) = 1;
1773 declare_inline_vars (id->block, var);
1775 /* Build the use expr. If the return type of the function was
1776 promoted, convert it back to the expected type. */
1778 if (!useless_type_conversion_p (caller_type, TREE_TYPE (var)))
1779 use = fold_convert (caller_type, var);
1781 STRIP_USELESS_TYPE_CONVERSION (use);
1783 if (DECL_BY_REFERENCE (result))
1784 var = build_fold_addr_expr (var);
1787 /* Register the VAR_DECL as the equivalent for the RESULT_DECL; that
1788 way, when the RESULT_DECL is encountered, it will be
1789 automatically replaced by the VAR_DECL. */
1790 insert_decl_map (id, result, var);
1792 /* Remember this so we can ignore it in remap_decls. */
1799 /* Returns nonzero if a function can be inlined as a tree. */
1802 tree_inlinable_function_p (tree fn)
1804 return inlinable_function_p (fn);
1807 static const char *inline_forbidden_reason;
1810 inline_forbidden_p_1 (tree *nodep, int *walk_subtrees ATTRIBUTE_UNUSED,
1814 tree fn = (tree) fnp;
1817 switch (TREE_CODE (node))
1820 /* Refuse to inline alloca call unless user explicitly forced so as
1821 this may change program's memory overhead drastically when the
1822 function using alloca is called in loop. In GCC present in
1823 SPEC2000 inlining into schedule_block cause it to require 2GB of
1824 RAM instead of 256MB. */
1825 if (alloca_call_p (node)
1826 && !lookup_attribute ("always_inline", DECL_ATTRIBUTES (fn)))
1828 inline_forbidden_reason
1829 = G_("function %q+F can never be inlined because it uses "
1830 "alloca (override using the always_inline attribute)");
1833 t = get_callee_fndecl (node);
1837 /* We cannot inline functions that call setjmp. */
1838 if (setjmp_call_p (t))
1840 inline_forbidden_reason
1841 = G_("function %q+F can never be inlined because it uses setjmp");
1845 if (DECL_BUILT_IN_CLASS (t) == BUILT_IN_NORMAL)
1846 switch (DECL_FUNCTION_CODE (t))
1848 /* We cannot inline functions that take a variable number of
1850 case BUILT_IN_VA_START:
1851 case BUILT_IN_STDARG_START:
1852 case BUILT_IN_NEXT_ARG:
1853 case BUILT_IN_VA_END:
1854 inline_forbidden_reason
1855 = G_("function %q+F can never be inlined because it "
1856 "uses variable argument lists");
1859 case BUILT_IN_LONGJMP:
1860 /* We can't inline functions that call __builtin_longjmp at
1861 all. The non-local goto machinery really requires the
1862 destination be in a different function. If we allow the
1863 function calling __builtin_longjmp to be inlined into the
1864 function calling __builtin_setjmp, Things will Go Awry. */
1865 inline_forbidden_reason
1866 = G_("function %q+F can never be inlined because "
1867 "it uses setjmp-longjmp exception handling");
1870 case BUILT_IN_NONLOCAL_GOTO:
1872 inline_forbidden_reason
1873 = G_("function %q+F can never be inlined because "
1874 "it uses non-local goto");
1877 case BUILT_IN_RETURN:
1878 case BUILT_IN_APPLY_ARGS:
1879 /* If a __builtin_apply_args caller would be inlined,
1880 it would be saving arguments of the function it has
1881 been inlined into. Similarly __builtin_return would
1882 return from the function the inline has been inlined into. */
1883 inline_forbidden_reason
1884 = G_("function %q+F can never be inlined because "
1885 "it uses __builtin_return or __builtin_apply_args");
1894 t = TREE_OPERAND (node, 0);
1896 /* We will not inline a function which uses computed goto. The
1897 addresses of its local labels, which may be tucked into
1898 global storage, are of course not constant across
1899 instantiations, which causes unexpected behavior. */
1900 if (TREE_CODE (t) != LABEL_DECL)
1902 inline_forbidden_reason
1903 = G_("function %q+F can never be inlined "
1904 "because it contains a computed goto");
1910 t = TREE_OPERAND (node, 0);
1911 if (DECL_NONLOCAL (t))
1913 /* We cannot inline a function that receives a non-local goto
1914 because we cannot remap the destination label used in the
1915 function that is performing the non-local goto. */
1916 inline_forbidden_reason
1917 = G_("function %q+F can never be inlined "
1918 "because it receives a non-local goto");
1925 /* We cannot inline a function of the form
1927 void F (int i) { struct S { int ar[i]; } s; }
1929 Attempting to do so produces a catch-22.
1930 If walk_tree examines the TYPE_FIELDS chain of RECORD_TYPE/
1931 UNION_TYPE nodes, then it goes into infinite recursion on a
1932 structure containing a pointer to its own type. If it doesn't,
1933 then the type node for S doesn't get adjusted properly when
1936 ??? This is likely no longer true, but it's too late in the 4.0
1937 cycle to try to find out. This should be checked for 4.1. */
1938 for (t = TYPE_FIELDS (node); t; t = TREE_CHAIN (t))
1939 if (variably_modified_type_p (TREE_TYPE (t), NULL))
1941 inline_forbidden_reason
1942 = G_("function %q+F can never be inlined "
1943 "because it uses variable sized variables");
1955 inline_forbidden_p_2 (tree *nodep, int *walk_subtrees,
1959 tree fn = (tree) fnp;
1961 if (TREE_CODE (node) == LABEL_DECL && DECL_CONTEXT (node) == fn)
1963 inline_forbidden_reason
1964 = G_("function %q+F can never be inlined "
1965 "because it saves address of local label in a static variable");
1975 /* Return subexpression representing possible alloca call, if any. */
1977 inline_forbidden_p (tree fndecl)
1979 location_t saved_loc = input_location;
1980 block_stmt_iterator bsi;
1982 tree ret = NULL_TREE;
1983 struct function *fun = DECL_STRUCT_FUNCTION (fndecl);
1986 FOR_EACH_BB_FN (bb, fun)
1987 for (bsi = bsi_start (bb); !bsi_end_p (bsi); bsi_next (&bsi))
1989 ret = walk_tree_without_duplicates (bsi_stmt_ptr (bsi),
1990 inline_forbidden_p_1, fndecl);
1995 for (step = fun->unexpanded_var_list; step; step = TREE_CHAIN (step))
1997 tree decl = TREE_VALUE (step);
1998 if (TREE_CODE (decl) == VAR_DECL
1999 && TREE_STATIC (decl)
2000 && !DECL_EXTERNAL (decl)
2001 && DECL_INITIAL (decl))
2002 ret = walk_tree_without_duplicates (&DECL_INITIAL (decl),
2003 inline_forbidden_p_2, fndecl);
2009 input_location = saved_loc;
2013 /* Returns nonzero if FN is a function that does not have any
2014 fundamental inline blocking properties. */
2017 inlinable_function_p (tree fn)
2019 bool inlinable = true;
2023 /* If we've already decided this function shouldn't be inlined,
2024 there's no need to check again. */
2025 if (DECL_UNINLINABLE (fn))
2028 /* We only warn for functions declared `inline' by the user. */
2029 do_warning = (warn_inline
2031 && DECL_DECLARED_INLINE_P (fn)
2032 && !DECL_IN_SYSTEM_HEADER (fn));
2034 always_inline = lookup_attribute ("always_inline", DECL_ATTRIBUTES (fn));
2036 if (flag_really_no_inline
2037 && always_inline == NULL)
2040 warning (OPT_Winline, "function %q+F can never be inlined because it "
2041 "is suppressed using -fno-inline", fn);
2045 /* Don't auto-inline anything that might not be bound within
2046 this unit of translation. */
2047 else if (!DECL_DECLARED_INLINE_P (fn)
2048 && DECL_REPLACEABLE_P (fn))
2051 else if (!function_attribute_inlinable_p (fn))
2054 warning (OPT_Winline, "function %q+F can never be inlined because it "
2055 "uses attributes conflicting with inlining", fn);
2059 /* If we don't have the function body available, we can't inline it.
2060 However, this should not be recorded since we also get here for
2061 forward declared inline functions. Therefore, return at once. */
2062 if (!DECL_SAVED_TREE (fn))
2065 /* If we're not inlining at all, then we cannot inline this function. */
2066 else if (!flag_inline_trees)
2069 /* Only try to inline functions if DECL_INLINE is set. This should be
2070 true for all functions declared `inline', and for all other functions
2071 as well with -finline-functions.
2073 Don't think of disregarding DECL_INLINE when flag_inline_trees == 2;
2074 it's the front-end that must set DECL_INLINE in this case, because
2075 dwarf2out loses if a function that does not have DECL_INLINE set is
2076 inlined anyway. That is why we have both DECL_INLINE and
2077 DECL_DECLARED_INLINE_P. */
2078 /* FIXME: When flag_inline_trees dies, the check for flag_unit_at_a_time
2079 here should be redundant. */
2080 else if (!DECL_INLINE (fn) && !flag_unit_at_a_time)
2083 else if (inline_forbidden_p (fn))
2085 /* See if we should warn about uninlinable functions. Previously,
2086 some of these warnings would be issued while trying to expand
2087 the function inline, but that would cause multiple warnings
2088 about functions that would for example call alloca. But since
2089 this a property of the function, just one warning is enough.
2090 As a bonus we can now give more details about the reason why a
2091 function is not inlinable. */
2093 sorry (inline_forbidden_reason, fn);
2094 else if (do_warning)
2095 warning (OPT_Winline, inline_forbidden_reason, fn);
2100 /* Squirrel away the result so that we don't have to check again. */
2101 DECL_UNINLINABLE (fn) = !inlinable;
2106 /* Estimate the cost of a memory move. Use machine dependent
2107 word size and take possible memcpy call into account. */
2110 estimate_move_cost (tree type)
2114 size = int_size_in_bytes (type);
2116 if (size < 0 || size > MOVE_MAX_PIECES * MOVE_RATIO)
2117 /* Cost of a memcpy call, 3 arguments and the call. */
2120 return ((size + MOVE_MAX_PIECES - 1) / MOVE_MAX_PIECES);
2123 /* Arguments for estimate_num_insns_1. */
2127 /* Used to return the number of insns. */
2130 /* Weights of various constructs. */
2131 eni_weights *weights;
2134 /* Used by estimate_num_insns. Estimate number of instructions seen
2135 by given statement. */
2138 estimate_num_insns_1 (tree *tp, int *walk_subtrees, void *data)
2140 struct eni_data *d = data;
2144 if (IS_TYPE_OR_DECL_P (x))
2149 /* Assume that constants and references counts nothing. These should
2150 be majorized by amount of operations among them we count later
2151 and are common target of CSE and similar optimizations. */
2152 else if (CONSTANT_CLASS_P (x) || REFERENCE_CLASS_P (x))
2155 switch (TREE_CODE (x))
2157 /* Containers have no cost. */
2164 case ALIGN_INDIRECT_REF:
2165 case MISALIGNED_INDIRECT_REF:
2167 case ARRAY_RANGE_REF:
2169 case EXC_PTR_EXPR: /* ??? */
2170 case FILTER_EXPR: /* ??? */
2173 case WITH_CLEANUP_EXPR:
2176 case VIEW_CONVERT_EXPR:
2181 case CASE_LABEL_EXPR:
2184 case EH_FILTER_EXPR:
2185 case STATEMENT_LIST:
2187 case NON_LVALUE_EXPR:
2190 case TRY_CATCH_EXPR:
2191 case TRY_FINALLY_EXPR:
2198 case WITH_SIZE_EXPR:
2202 case OMP_SECTIONS_SWITCH:
2203 case OMP_ATOMIC_STORE:
2206 /* We don't account constants for now. Assume that the cost is amortized
2207 by operations that do use them. We may re-consider this decision once
2208 we are able to optimize the tree before estimating its size and break
2209 out static initializers. */
2210 case IDENTIFIER_NODE:
2220 /* CHANGE_DYNAMIC_TYPE_EXPR explicitly expands to nothing. */
2221 case CHANGE_DYNAMIC_TYPE_EXPR:
2225 /* Try to estimate the cost of assignments. We have three cases to
2227 1) Simple assignments to registers;
2228 2) Stores to things that must live in memory. This includes
2229 "normal" stores to scalars, but also assignments of large
2230 structures, or constructors of big arrays;
2233 Let us look at the first two cases, assuming we have "a = b + C":
2234 <GIMPLE_MODIFY_STMT <var_decl "a">
2235 <plus_expr <var_decl "b"> <constant C>>
2236 If "a" is a GIMPLE register, the assignment to it is free on almost
2237 any target, because "a" usually ends up in a real register. Hence
2238 the only cost of this expression comes from the PLUS_EXPR, and we
2239 can ignore the GIMPLE_MODIFY_STMT.
2240 If "a" is not a GIMPLE register, the assignment to "a" will most
2241 likely be a real store, so the cost of the GIMPLE_MODIFY_STMT is the cost
2242 of moving something into "a", which we compute using the function
2245 The third case deals with TARGET_EXPRs, for which the semantics are
2246 that a temporary is assigned, unless the TARGET_EXPR itself is being
2247 assigned to something else. In the latter case we do not need the
2249 <GIMPLE_MODIFY_STMT <var_decl "a"> <target_expr>>, the
2250 GIMPLE_MODIFY_STMT is free. */
2252 case GIMPLE_MODIFY_STMT:
2253 /* Is the right and side a TARGET_EXPR? */
2254 if (TREE_CODE (GENERIC_TREE_OPERAND (x, 1)) == TARGET_EXPR)
2256 /* ... fall through ... */
2259 x = GENERIC_TREE_OPERAND (x, 0);
2260 /* Is this an assignments to a register? */
2261 if (is_gimple_reg (x))
2263 /* Otherwise it's a store, so fall through to compute the move cost. */
2266 d->count += estimate_move_cost (TREE_TYPE (x));
2269 /* Assign cost of 1 to usual operations.
2270 ??? We may consider mapping RTL costs to this. */
2275 case POINTER_PLUS_EXPR:
2279 case FIXED_CONVERT_EXPR:
2280 case FIX_TRUNC_EXPR:
2292 case VEC_LSHIFT_EXPR:
2293 case VEC_RSHIFT_EXPR:
2300 case TRUTH_ANDIF_EXPR:
2301 case TRUTH_ORIF_EXPR:
2302 case TRUTH_AND_EXPR:
2304 case TRUTH_XOR_EXPR:
2305 case TRUTH_NOT_EXPR:
2314 case UNORDERED_EXPR:
2325 case PREDECREMENT_EXPR:
2326 case PREINCREMENT_EXPR:
2327 case POSTDECREMENT_EXPR:
2328 case POSTINCREMENT_EXPR:
2332 case REALIGN_LOAD_EXPR:
2334 case REDUC_MAX_EXPR:
2335 case REDUC_MIN_EXPR:
2336 case REDUC_PLUS_EXPR:
2337 case WIDEN_SUM_EXPR:
2339 case VEC_WIDEN_MULT_HI_EXPR:
2340 case VEC_WIDEN_MULT_LO_EXPR:
2341 case VEC_UNPACK_HI_EXPR:
2342 case VEC_UNPACK_LO_EXPR:
2343 case VEC_UNPACK_FLOAT_HI_EXPR:
2344 case VEC_UNPACK_FLOAT_LO_EXPR:
2345 case VEC_PACK_TRUNC_EXPR:
2346 case VEC_PACK_SAT_EXPR:
2347 case VEC_PACK_FIX_TRUNC_EXPR:
2349 case WIDEN_MULT_EXPR:
2351 case VEC_EXTRACT_EVEN_EXPR:
2352 case VEC_EXTRACT_ODD_EXPR:
2353 case VEC_INTERLEAVE_HIGH_EXPR:
2354 case VEC_INTERLEAVE_LOW_EXPR:
2361 /* TODO: Cost of a switch should be derived from the number of
2363 d->count += d->weights->switch_cost;
2366 /* Few special cases of expensive operations. This is useful
2367 to avoid inlining on functions having too many of these. */
2368 case TRUNC_DIV_EXPR:
2370 case FLOOR_DIV_EXPR:
2371 case ROUND_DIV_EXPR:
2372 case EXACT_DIV_EXPR:
2373 case TRUNC_MOD_EXPR:
2375 case FLOOR_MOD_EXPR:
2376 case ROUND_MOD_EXPR:
2378 d->count += d->weights->div_mod_cost;
2382 tree decl = get_callee_fndecl (x);
2384 if (decl && DECL_BUILT_IN_CLASS (decl) == BUILT_IN_MD)
2385 cost = d->weights->target_builtin_call_cost;
2387 cost = d->weights->call_cost;
2389 if (decl && DECL_BUILT_IN_CLASS (decl) == BUILT_IN_NORMAL)
2390 switch (DECL_FUNCTION_CODE (decl))
2392 case BUILT_IN_CONSTANT_P:
2395 case BUILT_IN_EXPECT:
2397 /* Prefetch instruction is not expensive. */
2398 case BUILT_IN_PREFETCH:
2405 /* Our cost must be kept in sync with cgraph_estimate_size_after_inlining
2406 that does use function declaration to figure out the arguments. */
2410 call_expr_arg_iterator iter;
2411 FOR_EACH_CALL_EXPR_ARG (a, iter, x)
2412 d->count += estimate_move_cost (TREE_TYPE (a));
2417 for (arg = DECL_ARGUMENTS (decl); arg; arg = TREE_CHAIN (arg))
2418 d->count += estimate_move_cost (TREE_TYPE (arg));
2434 case OMP_ATOMIC_LOAD:
2435 /* OpenMP directives are generally very expensive. */
2436 d->count += d->weights->omp_cost;
2445 /* Estimate number of instructions that will be created by expanding EXPR.
2446 WEIGHTS contains weights attributed to various constructs. */
2449 estimate_num_insns (tree expr, eni_weights *weights)
2451 struct pointer_set_t *visited_nodes;
2453 block_stmt_iterator bsi;
2454 struct function *my_function;
2455 struct eni_data data;
2458 data.weights = weights;
2460 /* If we're given an entire function, walk the CFG. */
2461 if (TREE_CODE (expr) == FUNCTION_DECL)
2463 my_function = DECL_STRUCT_FUNCTION (expr);
2464 gcc_assert (my_function && my_function->cfg);
2465 visited_nodes = pointer_set_create ();
2466 FOR_EACH_BB_FN (bb, my_function)
2468 for (bsi = bsi_start (bb);
2472 walk_tree (bsi_stmt_ptr (bsi), estimate_num_insns_1,
2473 &data, visited_nodes);
2476 pointer_set_destroy (visited_nodes);
2479 walk_tree_without_duplicates (&expr, estimate_num_insns_1, &data);
2484 /* Initializes weights used by estimate_num_insns. */
2487 init_inline_once (void)
2489 eni_inlining_weights.call_cost = PARAM_VALUE (PARAM_INLINE_CALL_COST);
2490 eni_inlining_weights.target_builtin_call_cost = 1;
2491 eni_inlining_weights.div_mod_cost = 10;
2492 eni_inlining_weights.switch_cost = 1;
2493 eni_inlining_weights.omp_cost = 40;
2495 eni_size_weights.call_cost = 1;
2496 eni_size_weights.target_builtin_call_cost = 1;
2497 eni_size_weights.div_mod_cost = 1;
2498 eni_size_weights.switch_cost = 10;
2499 eni_size_weights.omp_cost = 40;
2501 /* Estimating time for call is difficult, since we have no idea what the
2502 called function does. In the current uses of eni_time_weights,
2503 underestimating the cost does less harm than overestimating it, so
2504 we choose a rather small value here. */
2505 eni_time_weights.call_cost = 10;
2506 eni_time_weights.target_builtin_call_cost = 10;
2507 eni_time_weights.div_mod_cost = 10;
2508 eni_time_weights.switch_cost = 4;
2509 eni_time_weights.omp_cost = 40;
2512 /* Install new lexical TREE_BLOCK underneath 'current_block'. */
2514 add_lexical_block (tree current_block, tree new_block)
2518 /* Walk to the last sub-block. */
2519 for (blk_p = &BLOCK_SUBBLOCKS (current_block);
2521 blk_p = &BLOCK_CHAIN (*blk_p))
2524 BLOCK_SUPERCONTEXT (new_block) = current_block;
2527 /* If *TP is a CALL_EXPR, replace it with its inline expansion. */
2530 expand_call_inline (basic_block bb, tree stmt, tree *tp, void *data)
2536 struct pointer_map_t *st;
2539 location_t saved_location;
2540 struct cgraph_edge *cg_edge;
2542 basic_block return_block;
2544 block_stmt_iterator bsi, stmt_bsi;
2545 bool successfully_inlined = FALSE;
2546 bool purge_dead_abnormal_edges;
2550 /* See what we've got. */
2551 id = (copy_body_data *) data;
2554 /* Set input_location here so we get the right instantiation context
2555 if we call instantiate_decl from inlinable_function_p. */
2556 saved_location = input_location;
2557 if (EXPR_HAS_LOCATION (t))
2558 input_location = EXPR_LOCATION (t);
2560 /* From here on, we're only interested in CALL_EXPRs. */
2561 if (TREE_CODE (t) != CALL_EXPR)
2564 /* First, see if we can figure out what function is being called.
2565 If we cannot, then there is no hope of inlining the function. */
2566 fn = get_callee_fndecl (t);
2570 /* Turn forward declarations into real ones. */
2571 fn = cgraph_node (fn)->decl;
2573 /* If fn is a declaration of a function in a nested scope that was
2574 globally declared inline, we don't set its DECL_INITIAL.
2575 However, we can't blindly follow DECL_ABSTRACT_ORIGIN because the
2576 C++ front-end uses it for cdtors to refer to their internal
2577 declarations, that are not real functions. Fortunately those
2578 don't have trees to be saved, so we can tell by checking their
2580 if (! DECL_INITIAL (fn)
2581 && DECL_ABSTRACT_ORIGIN (fn)
2582 && DECL_SAVED_TREE (DECL_ABSTRACT_ORIGIN (fn)))
2583 fn = DECL_ABSTRACT_ORIGIN (fn);
2585 /* Objective C and fortran still calls tree_rest_of_compilation directly.
2586 Kill this check once this is fixed. */
2587 if (!id->dst_node->analyzed)
2590 cg_edge = cgraph_edge (id->dst_node, stmt);
2592 /* Constant propagation on argument done during previous inlining
2593 may create new direct call. Produce an edge for it. */
2596 struct cgraph_node *dest = cgraph_node (fn);
2598 /* We have missing edge in the callgraph. This can happen in one case
2599 where previous inlining turned indirect call into direct call by
2600 constant propagating arguments. In all other cases we hit a bug
2601 (incorrect node sharing is most common reason for missing edges. */
2602 gcc_assert (dest->needed || !flag_unit_at_a_time);
2603 cgraph_create_edge (id->dst_node, dest, stmt,
2604 bb->count, CGRAPH_FREQ_BASE,
2605 bb->loop_depth)->inline_failed
2606 = N_("originally indirect function call not considered for inlining");
2609 fprintf (dump_file, "Created new direct edge to %s",
2610 cgraph_node_name (dest));
2615 /* Don't try to inline functions that are not well-suited to
2617 if (!cgraph_inline_p (cg_edge, &reason))
2619 if (lookup_attribute ("always_inline", DECL_ATTRIBUTES (fn))
2620 /* Avoid warnings during early inline pass. */
2621 && (!flag_unit_at_a_time || cgraph_global_info_ready))
2623 sorry ("inlining failed in call to %q+F: %s", fn, reason);
2624 sorry ("called from here");
2626 else if (warn_inline && DECL_DECLARED_INLINE_P (fn)
2627 && !DECL_IN_SYSTEM_HEADER (fn)
2629 && !lookup_attribute ("noinline", DECL_ATTRIBUTES (fn))
2630 /* Avoid warnings during early inline pass. */
2631 && (!flag_unit_at_a_time || cgraph_global_info_ready))
2633 warning (OPT_Winline, "inlining failed in call to %q+F: %s",
2635 warning (OPT_Winline, "called from here");
2639 fn = cg_edge->callee->decl;
2641 #ifdef ENABLE_CHECKING
2642 if (cg_edge->callee->decl != id->dst_node->decl)
2643 verify_cgraph_node (cg_edge->callee);
2646 /* We will be inlining this callee. */
2647 id->eh_region = lookup_stmt_eh_region (stmt);
2649 /* Split the block holding the CALL_EXPR. */
2650 e = split_block (bb, stmt);
2652 return_block = e->dest;
2655 /* split_block splits after the statement; work around this by
2656 moving the call into the second block manually. Not pretty,
2657 but seems easier than doing the CFG manipulation by hand
2658 when the CALL_EXPR is in the last statement of BB. */
2659 stmt_bsi = bsi_last (bb);
2660 bsi_remove (&stmt_bsi, false);
2662 /* If the CALL_EXPR was in the last statement of BB, it may have
2663 been the source of abnormal edges. In this case, schedule
2664 the removal of dead abnormal edges. */
2665 bsi = bsi_start (return_block);
2666 if (bsi_end_p (bsi))
2668 bsi_insert_after (&bsi, stmt, BSI_NEW_STMT);
2669 purge_dead_abnormal_edges = true;
2673 bsi_insert_before (&bsi, stmt, BSI_NEW_STMT);
2674 purge_dead_abnormal_edges = false;
2677 stmt_bsi = bsi_start (return_block);
2679 /* Build a block containing code to initialize the arguments, the
2680 actual inline expansion of the body, and a label for the return
2681 statements within the function to jump to. The type of the
2682 statement expression is the return type of the function call. */
2683 id->block = make_node (BLOCK);
2684 BLOCK_ABSTRACT_ORIGIN (id->block) = fn;
2685 BLOCK_SOURCE_LOCATION (id->block) = input_location;
2686 add_lexical_block (TREE_BLOCK (stmt), id->block);
2688 /* Local declarations will be replaced by their equivalents in this
2691 id->decl_map = pointer_map_create ();
2693 /* Record the function we are about to inline. */
2695 id->src_node = cg_edge->callee;
2696 id->src_cfun = DECL_STRUCT_FUNCTION (fn);
2699 gcc_assert (!id->src_cfun->after_inlining);
2701 initialize_inlined_parameters (id, t, fn, bb);
2703 if (DECL_INITIAL (fn))
2704 add_lexical_block (id->block, remap_blocks (DECL_INITIAL (fn), id));
2706 /* Return statements in the function body will be replaced by jumps
2707 to the RET_LABEL. */
2709 gcc_assert (DECL_INITIAL (fn));
2710 gcc_assert (TREE_CODE (DECL_INITIAL (fn)) == BLOCK);
2712 /* Find the lhs to which the result of this call is assigned. */
2714 if (TREE_CODE (stmt) == GIMPLE_MODIFY_STMT)
2716 modify_dest = GIMPLE_STMT_OPERAND (stmt, 0);
2718 /* The function which we are inlining might not return a value,
2719 in which case we should issue a warning that the function
2720 does not return a value. In that case the optimizers will
2721 see that the variable to which the value is assigned was not
2722 initialized. We do not want to issue a warning about that
2723 uninitialized variable. */
2724 if (DECL_P (modify_dest))
2725 TREE_NO_WARNING (modify_dest) = 1;
2726 if (CALL_EXPR_RETURN_SLOT_OPT (t))
2728 return_slot = modify_dest;
2735 /* Declare the return variable for the function. */
2736 declare_return_variable (id, return_slot,
2737 modify_dest, &use_retvar);
2739 /* This is it. Duplicate the callee body. Assume callee is
2740 pre-gimplified. Note that we must not alter the caller
2741 function in any way before this point, as this CALL_EXPR may be
2742 a self-referential call; if we're calling ourselves, we need to
2743 duplicate our body before altering anything. */
2744 copy_body (id, bb->count, bb->frequency, bb, return_block);
2746 /* Add local vars in this inlined callee to caller. */
2747 t_step = id->src_cfun->unexpanded_var_list;
2748 for (; t_step; t_step = TREE_CHAIN (t_step))
2750 var = TREE_VALUE (t_step);
2751 if (TREE_STATIC (var) && !TREE_ASM_WRITTEN (var))
2752 cfun->unexpanded_var_list = tree_cons (NULL_TREE, var,
2753 cfun->unexpanded_var_list);
2755 cfun->unexpanded_var_list = tree_cons (NULL_TREE, remap_decl (var, id),
2756 cfun->unexpanded_var_list);
2760 pointer_map_destroy (id->decl_map);
2763 /* If the inlined function returns a result that we care about,
2764 clobber the CALL_EXPR with a reference to the return variable. */
2765 if (use_retvar && (TREE_CODE (bsi_stmt (stmt_bsi)) != CALL_EXPR))
2768 if (gimple_in_ssa_p (cfun))
2771 mark_symbols_for_renaming (stmt);
2773 maybe_clean_or_replace_eh_stmt (stmt, stmt);
2776 /* We're modifying a TSI owned by gimple_expand_calls_inline();
2777 tsi_delink() will leave the iterator in a sane state. */
2779 /* Handle case of inlining function that miss return statement so
2780 return value becomes undefined. */
2781 if (TREE_CODE (stmt) == GIMPLE_MODIFY_STMT
2782 && TREE_CODE (GIMPLE_STMT_OPERAND (stmt, 0)) == SSA_NAME)
2784 tree name = TREE_OPERAND (stmt, 0);
2785 tree var = SSA_NAME_VAR (TREE_OPERAND (stmt, 0));
2786 tree def = gimple_default_def (cfun, var);
2788 /* If the variable is used undefined, make this name undefined via
2792 TREE_OPERAND (stmt, 1) = def;
2795 /* Otherwise make this variable undefined. */
2798 bsi_remove (&stmt_bsi, true);
2799 set_default_def (var, name);
2800 SSA_NAME_DEF_STMT (name) = build_empty_stmt ();
2804 bsi_remove (&stmt_bsi, true);
2807 if (purge_dead_abnormal_edges)
2808 tree_purge_dead_abnormal_call_edges (return_block);
2810 /* If the value of the new expression is ignored, that's OK. We
2811 don't warn about this for CALL_EXPRs, so we shouldn't warn about
2812 the equivalent inlined version either. */
2813 TREE_USED (*tp) = 1;
2815 /* Output the inlining info for this abstract function, since it has been
2816 inlined. If we don't do this now, we can lose the information about the
2817 variables in the function when the blocks get blown away as soon as we
2818 remove the cgraph node. */
2819 (*debug_hooks->outlining_inline_function) (cg_edge->callee->decl);
2821 /* Update callgraph if needed. */
2822 cgraph_remove_node (cg_edge->callee);
2824 id->block = NULL_TREE;
2825 successfully_inlined = TRUE;
2828 input_location = saved_location;
2829 return successfully_inlined;
2832 /* Expand call statements reachable from STMT_P.
2833 We can only have CALL_EXPRs as the "toplevel" tree code or nested
2834 in a GIMPLE_MODIFY_STMT. See tree-gimple.c:get_call_expr_in(). We can
2835 unfortunately not use that function here because we need a pointer
2836 to the CALL_EXPR, not the tree itself. */
2839 gimple_expand_calls_inline (basic_block bb, copy_body_data *id)
2841 block_stmt_iterator bsi;
2843 /* Register specific tree functions. */
2844 tree_register_cfg_hooks ();
2845 for (bsi = bsi_start (bb); !bsi_end_p (bsi); bsi_next (&bsi))
2847 tree *expr_p = bsi_stmt_ptr (bsi);
2848 tree stmt = *expr_p;
2850 if (TREE_CODE (*expr_p) == GIMPLE_MODIFY_STMT)
2851 expr_p = &GIMPLE_STMT_OPERAND (*expr_p, 1);
2852 if (TREE_CODE (*expr_p) == WITH_SIZE_EXPR)
2853 expr_p = &TREE_OPERAND (*expr_p, 0);
2854 if (TREE_CODE (*expr_p) == CALL_EXPR)
2855 if (expand_call_inline (bb, stmt, expr_p, id))
2861 /* Walk all basic blocks created after FIRST and try to fold every statement
2862 in the STATEMENTS pointer set. */
2864 fold_marked_statements (int first, struct pointer_set_t *statements)
2866 for (;first < n_basic_blocks;first++)
2867 if (BASIC_BLOCK (first))
2869 block_stmt_iterator bsi;
2870 for (bsi = bsi_start (BASIC_BLOCK (first));
2871 !bsi_end_p (bsi); bsi_next (&bsi))
2872 if (pointer_set_contains (statements, bsi_stmt (bsi)))
2874 tree old_stmt = bsi_stmt (bsi);
2875 if (fold_stmt (bsi_stmt_ptr (bsi)))
2877 update_stmt (bsi_stmt (bsi));
2878 if (maybe_clean_or_replace_eh_stmt (old_stmt, bsi_stmt (bsi)))
2879 tree_purge_dead_eh_edges (BASIC_BLOCK (first));
2885 /* Return true if BB has at least one abnormal outgoing edge. */
2888 has_abnormal_outgoing_edge_p (basic_block bb)
2893 FOR_EACH_EDGE (e, ei, bb->succs)
2894 if (e->flags & EDGE_ABNORMAL)
2900 /* Expand calls to inline functions in the body of FN. */
2903 optimize_inline_calls (tree fn)
2908 int last = n_basic_blocks;
2909 /* There is no point in performing inlining if errors have already
2910 occurred -- and we might crash if we try to inline invalid
2912 if (errorcount || sorrycount)
2916 memset (&id, 0, sizeof (id));
2918 id.src_node = id.dst_node = cgraph_node (fn);
2920 /* Or any functions that aren't finished yet. */
2921 prev_fn = NULL_TREE;
2922 if (current_function_decl)
2924 id.dst_fn = current_function_decl;
2925 prev_fn = current_function_decl;
2928 id.copy_decl = copy_decl_maybe_to_var;
2929 id.transform_call_graph_edges = CB_CGE_DUPLICATE;
2930 id.transform_new_cfg = false;
2931 id.transform_return_to_modify = true;
2932 id.transform_lang_insert_block = false;
2933 id.statements_to_fold = pointer_set_create ();
2935 push_gimplify_context ();
2937 /* We make no attempts to keep dominance info up-to-date. */
2938 free_dominance_info (CDI_DOMINATORS);
2939 free_dominance_info (CDI_POST_DOMINATORS);
2941 /* Reach the trees by walking over the CFG, and note the
2942 enclosing basic-blocks in the call edges. */
2943 /* We walk the blocks going forward, because inlined function bodies
2944 will split id->current_basic_block, and the new blocks will
2945 follow it; we'll trudge through them, processing their CALL_EXPRs
2948 gimple_expand_calls_inline (bb, &id);
2950 pop_gimplify_context (NULL);
2952 #ifdef ENABLE_CHECKING
2954 struct cgraph_edge *e;
2956 verify_cgraph_node (id.dst_node);
2958 /* Double check that we inlined everything we are supposed to inline. */
2959 for (e = id.dst_node->callees; e; e = e->next_callee)
2960 gcc_assert (e->inline_failed);
2964 /* Fold the statements before compacting/renumbering the basic blocks. */
2965 fold_marked_statements (last, id.statements_to_fold);
2966 pointer_set_destroy (id.statements_to_fold);
2968 /* Renumber the (code) basic_blocks consecutively. */
2970 /* Renumber the lexical scoping (non-code) blocks consecutively. */
2973 /* We are not going to maintain the cgraph edges up to date.
2974 Kill it so it won't confuse us. */
2975 cgraph_node_remove_callees (id.dst_node);
2977 fold_cond_expr_cond ();
2978 /* It would be nice to check SSA/CFG/statement consistency here, but it is
2979 not possible yet - the IPA passes might make various functions to not
2980 throw and they don't care to proactively update local EH info. This is
2981 done later in fixup_cfg pass that also execute the verification. */
2982 return (TODO_update_ssa | TODO_cleanup_cfg
2983 | (gimple_in_ssa_p (cfun) ? TODO_remove_unused_locals : 0)
2984 | (profile_status != PROFILE_ABSENT ? TODO_rebuild_frequencies : 0));
2987 /* FN is a function that has a complete body, and CLONE is a function whose
2988 body is to be set to a copy of FN, mapping argument declarations according
2989 to the ARG_MAP splay_tree. */
2992 clone_body (tree clone, tree fn, void *arg_map)
2996 /* Clone the body, as if we were making an inline call. But, remap the
2997 parameters in the callee to the parameters of caller. */
2998 memset (&id, 0, sizeof (id));
3001 id.src_cfun = DECL_STRUCT_FUNCTION (fn);
3002 id.decl_map = (struct pointer_map_t *)arg_map;
3004 id.copy_decl = copy_decl_no_change;
3005 id.transform_call_graph_edges = CB_CGE_DUPLICATE;
3006 id.transform_new_cfg = true;
3007 id.transform_return_to_modify = false;
3008 id.transform_lang_insert_block = true;
3010 /* We're not inside any EH region. */
3013 /* Actually copy the body. */
3014 append_to_statement_list_force (copy_generic_body (&id), &DECL_SAVED_TREE (clone));
3017 /* Passed to walk_tree. Copies the node pointed to, if appropriate. */
3020 copy_tree_r (tree *tp, int *walk_subtrees, void *data ATTRIBUTE_UNUSED)
3022 enum tree_code code = TREE_CODE (*tp);
3023 enum tree_code_class cl = TREE_CODE_CLASS (code);
3025 /* We make copies of most nodes. */
3026 if (IS_EXPR_CODE_CLASS (cl)
3027 || IS_GIMPLE_STMT_CODE_CLASS (cl)
3028 || code == TREE_LIST
3030 || code == TYPE_DECL
3031 || code == OMP_CLAUSE)
3033 /* Because the chain gets clobbered when we make a copy, we save it
3035 tree chain = NULL_TREE, new;
3037 if (!GIMPLE_TUPLE_P (*tp))
3038 chain = TREE_CHAIN (*tp);
3040 /* Copy the node. */
3041 new = copy_node (*tp);
3043 /* Propagate mudflap marked-ness. */
3044 if (flag_mudflap && mf_marked_p (*tp))
3049 /* Now, restore the chain, if appropriate. That will cause
3050 walk_tree to walk into the chain as well. */
3051 if (code == PARM_DECL
3052 || code == TREE_LIST
3053 || code == OMP_CLAUSE)
3054 TREE_CHAIN (*tp) = chain;
3056 /* For now, we don't update BLOCKs when we make copies. So, we
3057 have to nullify all BIND_EXPRs. */
3058 if (TREE_CODE (*tp) == BIND_EXPR)
3059 BIND_EXPR_BLOCK (*tp) = NULL_TREE;
3061 else if (code == CONSTRUCTOR)
3063 /* CONSTRUCTOR nodes need special handling because
3064 we need to duplicate the vector of elements. */
3067 new = copy_node (*tp);
3069 /* Propagate mudflap marked-ness. */
3070 if (flag_mudflap && mf_marked_p (*tp))
3073 CONSTRUCTOR_ELTS (new) = VEC_copy (constructor_elt, gc,
3074 CONSTRUCTOR_ELTS (*tp));
3077 else if (TREE_CODE_CLASS (code) == tcc_type)
3079 else if (TREE_CODE_CLASS (code) == tcc_declaration)
3081 else if (TREE_CODE_CLASS (code) == tcc_constant)
3084 gcc_assert (code != STATEMENT_LIST);
3088 /* The SAVE_EXPR pointed to by TP is being copied. If ST contains
3089 information indicating to what new SAVE_EXPR this one should be mapped,
3090 use that one. Otherwise, create a new node and enter it in ST. FN is
3091 the function into which the copy will be placed. */
3094 remap_save_expr (tree *tp, void *st_, int *walk_subtrees)
3096 struct pointer_map_t *st = (struct pointer_map_t *) st_;
3100 /* See if we already encountered this SAVE_EXPR. */
3101 n = (tree *) pointer_map_contains (st, *tp);
3103 /* If we didn't already remap this SAVE_EXPR, do so now. */
3106 t = copy_node (*tp);
3108 /* Remember this SAVE_EXPR. */
3109 *pointer_map_insert (st, *tp) = t;
3110 /* Make sure we don't remap an already-remapped SAVE_EXPR. */
3111 *pointer_map_insert (st, t) = t;
3115 /* We've already walked into this SAVE_EXPR; don't do it again. */
3120 /* Replace this SAVE_EXPR with the copy. */
3124 /* Called via walk_tree. If *TP points to a DECL_STMT for a local label,
3125 copies the declaration and enters it in the splay_tree in DATA (which is
3126 really an `copy_body_data *'). */
3129 mark_local_for_remap_r (tree *tp, int *walk_subtrees ATTRIBUTE_UNUSED,
3132 copy_body_data *id = (copy_body_data *) data;
3134 /* Don't walk into types. */
3138 else if (TREE_CODE (*tp) == LABEL_EXPR)
3140 tree decl = TREE_OPERAND (*tp, 0);
3142 /* Copy the decl and remember the copy. */
3143 insert_decl_map (id, decl, id->copy_decl (decl, id));
3149 /* Perform any modifications to EXPR required when it is unsaved. Does
3150 not recurse into EXPR's subtrees. */
3153 unsave_expr_1 (tree expr)
3155 switch (TREE_CODE (expr))
3158 /* Don't mess with a TARGET_EXPR that hasn't been expanded.
3159 It's OK for this to happen if it was part of a subtree that
3160 isn't immediately expanded, such as operand 2 of another
3162 if (TREE_OPERAND (expr, 1))
3165 TREE_OPERAND (expr, 1) = TREE_OPERAND (expr, 3);
3166 TREE_OPERAND (expr, 3) = NULL_TREE;
3174 /* Called via walk_tree when an expression is unsaved. Using the
3175 splay_tree pointed to by ST (which is really a `splay_tree'),
3176 remaps all local declarations to appropriate replacements. */
3179 unsave_r (tree *tp, int *walk_subtrees, void *data)
3181 copy_body_data *id = (copy_body_data *) data;
3182 struct pointer_map_t *st = id->decl_map;
3185 /* Only a local declaration (variable or label). */
3186 if ((TREE_CODE (*tp) == VAR_DECL && !TREE_STATIC (*tp))
3187 || TREE_CODE (*tp) == LABEL_DECL)
3189 /* Lookup the declaration. */
3190 n = (tree *) pointer_map_contains (st, *tp);
3192 /* If it's there, remap it. */
3197 else if (TREE_CODE (*tp) == STATEMENT_LIST)
3198 copy_statement_list (tp);
3199 else if (TREE_CODE (*tp) == BIND_EXPR)
3200 copy_bind_expr (tp, walk_subtrees, id);
3201 else if (TREE_CODE (*tp) == SAVE_EXPR)
3202 remap_save_expr (tp, st, walk_subtrees);
3205 copy_tree_r (tp, walk_subtrees, NULL);
3207 /* Do whatever unsaving is required. */
3208 unsave_expr_1 (*tp);
3211 /* Keep iterating. */
3215 /* Copies everything in EXPR and replaces variables, labels
3216 and SAVE_EXPRs local to EXPR. */
3219 unsave_expr_now (tree expr)
3223 /* There's nothing to do for NULL_TREE. */
3228 memset (&id, 0, sizeof (id));
3229 id.src_fn = current_function_decl;
3230 id.dst_fn = current_function_decl;
3231 id.decl_map = pointer_map_create ();
3233 id.copy_decl = copy_decl_no_change;
3234 id.transform_call_graph_edges = CB_CGE_DUPLICATE;
3235 id.transform_new_cfg = false;
3236 id.transform_return_to_modify = false;
3237 id.transform_lang_insert_block = false;
3239 /* Walk the tree once to find local labels. */
3240 walk_tree_without_duplicates (&expr, mark_local_for_remap_r, &id);
3242 /* Walk the tree again, copying, remapping, and unsaving. */
3243 walk_tree (&expr, unsave_r, &id, NULL);
3246 pointer_map_destroy (id.decl_map);
3251 /* Allow someone to determine if SEARCH is a child of TOP from gdb. */
3254 debug_find_tree_1 (tree *tp, int *walk_subtrees ATTRIBUTE_UNUSED, void *data)
3263 debug_find_tree (tree top, tree search)
3265 return walk_tree_without_duplicates (&top, debug_find_tree_1, search) != 0;
3269 /* Declare the variables created by the inliner. Add all the variables in
3270 VARS to BIND_EXPR. */
3273 declare_inline_vars (tree block, tree vars)
3276 for (t = vars; t; t = TREE_CHAIN (t))
3278 DECL_SEEN_IN_BIND_EXPR_P (t) = 1;
3279 gcc_assert (!TREE_STATIC (t) && !TREE_ASM_WRITTEN (t));
3280 cfun->unexpanded_var_list =
3281 tree_cons (NULL_TREE, t,
3282 cfun->unexpanded_var_list);
3286 BLOCK_VARS (block) = chainon (BLOCK_VARS (block), vars);
3290 /* Copy NODE (which must be a DECL). The DECL originally was in the FROM_FN,
3291 but now it will be in the TO_FN. PARM_TO_VAR means enable PARM_DECL to
3292 VAR_DECL translation. */
3295 copy_decl_for_dup_finish (copy_body_data *id, tree decl, tree copy)
3297 /* Don't generate debug information for the copy if we wouldn't have
3298 generated it for the copy either. */
3299 DECL_ARTIFICIAL (copy) = DECL_ARTIFICIAL (decl);
3300 DECL_IGNORED_P (copy) = DECL_IGNORED_P (decl);
3302 /* Set the DECL_ABSTRACT_ORIGIN so the debugging routines know what
3303 declaration inspired this copy. */
3304 DECL_ABSTRACT_ORIGIN (copy) = DECL_ORIGIN (decl);
3306 /* The new variable/label has no RTL, yet. */
3307 if (CODE_CONTAINS_STRUCT (TREE_CODE (copy), TS_DECL_WRTL)
3308 && !TREE_STATIC (copy) && !DECL_EXTERNAL (copy))
3309 SET_DECL_RTL (copy, NULL_RTX);
3311 /* These args would always appear unused, if not for this. */
3312 TREE_USED (copy) = 1;
3314 /* Set the context for the new declaration. */
3315 if (!DECL_CONTEXT (decl))
3316 /* Globals stay global. */
3318 else if (DECL_CONTEXT (decl) != id->src_fn)
3319 /* Things that weren't in the scope of the function we're inlining
3320 from aren't in the scope we're inlining to, either. */
3322 else if (TREE_STATIC (decl))
3323 /* Function-scoped static variables should stay in the original
3327 /* Ordinary automatic local variables are now in the scope of the
3329 DECL_CONTEXT (copy) = id->dst_fn;
3335 copy_decl_to_var (tree decl, copy_body_data *id)
3339 gcc_assert (TREE_CODE (decl) == PARM_DECL
3340 || TREE_CODE (decl) == RESULT_DECL);
3342 type = TREE_TYPE (decl);
3344 copy = build_decl (VAR_DECL, DECL_NAME (decl), type);
3345 TREE_ADDRESSABLE (copy) = TREE_ADDRESSABLE (decl);
3346 TREE_READONLY (copy) = TREE_READONLY (decl);
3347 TREE_THIS_VOLATILE (copy) = TREE_THIS_VOLATILE (decl);
3348 DECL_GIMPLE_REG_P (copy) = DECL_GIMPLE_REG_P (decl);
3349 DECL_NO_TBAA_P (copy) = DECL_NO_TBAA_P (decl);
3351 return copy_decl_for_dup_finish (id, decl, copy);
3354 /* Like copy_decl_to_var, but create a return slot object instead of a
3355 pointer variable for return by invisible reference. */
3358 copy_result_decl_to_var (tree decl, copy_body_data *id)
3362 gcc_assert (TREE_CODE (decl) == PARM_DECL
3363 || TREE_CODE (decl) == RESULT_DECL);
3365 type = TREE_TYPE (decl);
3366 if (DECL_BY_REFERENCE (decl))
3367 type = TREE_TYPE (type);
3369 copy = build_decl (VAR_DECL, DECL_NAME (decl), type);
3370 TREE_READONLY (copy) = TREE_READONLY (decl);
3371 TREE_THIS_VOLATILE (copy) = TREE_THIS_VOLATILE (decl);
3372 if (!DECL_BY_REFERENCE (decl))
3374 TREE_ADDRESSABLE (copy) = TREE_ADDRESSABLE (decl);
3375 DECL_GIMPLE_REG_P (copy) = DECL_GIMPLE_REG_P (decl);
3376 DECL_NO_TBAA_P (copy) = DECL_NO_TBAA_P (decl);
3379 return copy_decl_for_dup_finish (id, decl, copy);
3384 copy_decl_no_change (tree decl, copy_body_data *id)
3388 copy = copy_node (decl);
3390 /* The COPY is not abstract; it will be generated in DST_FN. */
3391 DECL_ABSTRACT (copy) = 0;
3392 lang_hooks.dup_lang_specific_decl (copy);
3394 /* TREE_ADDRESSABLE isn't used to indicate that a label's address has
3395 been taken; it's for internal bookkeeping in expand_goto_internal. */
3396 if (TREE_CODE (copy) == LABEL_DECL)
3398 TREE_ADDRESSABLE (copy) = 0;
3399 LABEL_DECL_UID (copy) = -1;
3402 return copy_decl_for_dup_finish (id, decl, copy);
3406 copy_decl_maybe_to_var (tree decl, copy_body_data *id)
3408 if (TREE_CODE (decl) == PARM_DECL || TREE_CODE (decl) == RESULT_DECL)
3409 return copy_decl_to_var (decl, id);
3411 return copy_decl_no_change (decl, id);
3414 /* Return a copy of the function's argument tree. */
3416 copy_arguments_for_versioning (tree orig_parm, copy_body_data * id)
3418 tree *arg_copy, *parg;
3420 arg_copy = &orig_parm;
3421 for (parg = arg_copy; *parg; parg = &TREE_CHAIN (*parg))
3423 tree new = remap_decl (*parg, id);
3424 lang_hooks.dup_lang_specific_decl (new);
3425 TREE_CHAIN (new) = TREE_CHAIN (*parg);
3431 /* Return a copy of the function's static chain. */
3433 copy_static_chain (tree static_chain, copy_body_data * id)
3435 tree *chain_copy, *pvar;
3437 chain_copy = &static_chain;
3438 for (pvar = chain_copy; *pvar; pvar = &TREE_CHAIN (*pvar))
3440 tree new = remap_decl (*pvar, id);
3441 lang_hooks.dup_lang_specific_decl (new);
3442 TREE_CHAIN (new) = TREE_CHAIN (*pvar);
3445 return static_chain;
3448 /* Return true if the function is allowed to be versioned.
3449 This is a guard for the versioning functionality. */
3451 tree_versionable_function_p (tree fndecl)
3453 if (fndecl == NULL_TREE)
3455 /* ??? There are cases where a function is
3456 uninlinable but can be versioned. */
3457 if (!tree_inlinable_function_p (fndecl))
3463 /* Create a copy of a function's tree.
3464 OLD_DECL and NEW_DECL are FUNCTION_DECL tree nodes
3465 of the original function and the new copied function
3466 respectively. In case we want to replace a DECL
3467 tree with another tree while duplicating the function's
3468 body, TREE_MAP represents the mapping between these
3469 trees. If UPDATE_CLONES is set, the call_stmt fields
3470 of edges of clones of the function will be updated. */
3472 tree_function_versioning (tree old_decl, tree new_decl, varray_type tree_map,
3475 struct cgraph_node *old_version_node;
3476 struct cgraph_node *new_version_node;
3480 struct ipa_replace_map *replace_info;
3481 basic_block old_entry_block;
3483 tree old_current_function_decl = current_function_decl;
3485 gcc_assert (TREE_CODE (old_decl) == FUNCTION_DECL
3486 && TREE_CODE (new_decl) == FUNCTION_DECL);
3487 DECL_POSSIBLY_INLINED (old_decl) = 1;
3489 old_version_node = cgraph_node (old_decl);
3490 new_version_node = cgraph_node (new_decl);
3492 DECL_ARTIFICIAL (new_decl) = 1;
3493 DECL_ABSTRACT_ORIGIN (new_decl) = DECL_ORIGIN (old_decl);
3495 /* Prepare the data structures for the tree copy. */
3496 memset (&id, 0, sizeof (id));
3498 /* Generate a new name for the new version. */
3501 DECL_NAME (new_decl) = create_tmp_var_name (NULL);
3502 SET_DECL_ASSEMBLER_NAME (new_decl, DECL_NAME (new_decl));
3503 SET_DECL_RTL (new_decl, NULL_RTX);
3504 id.statements_to_fold = pointer_set_create ();
3507 id.decl_map = pointer_map_create ();
3508 id.src_fn = old_decl;
3509 id.dst_fn = new_decl;
3510 id.src_node = old_version_node;
3511 id.dst_node = new_version_node;
3512 id.src_cfun = DECL_STRUCT_FUNCTION (old_decl);
3514 id.copy_decl = copy_decl_no_change;
3515 id.transform_call_graph_edges
3516 = update_clones ? CB_CGE_MOVE_CLONES : CB_CGE_MOVE;
3517 id.transform_new_cfg = true;
3518 id.transform_return_to_modify = false;
3519 id.transform_lang_insert_block = false;
3521 current_function_decl = new_decl;
3522 old_entry_block = ENTRY_BLOCK_PTR_FOR_FUNCTION
3523 (DECL_STRUCT_FUNCTION (old_decl));
3524 initialize_cfun (new_decl, old_decl,
3525 old_entry_block->count,
3526 old_entry_block->frequency);
3527 push_cfun (DECL_STRUCT_FUNCTION (new_decl));
3529 /* Copy the function's static chain. */
3530 p = DECL_STRUCT_FUNCTION (old_decl)->static_chain_decl;
3532 DECL_STRUCT_FUNCTION (new_decl)->static_chain_decl =
3533 copy_static_chain (DECL_STRUCT_FUNCTION (old_decl)->static_chain_decl,
3535 /* Copy the function's arguments. */
3536 if (DECL_ARGUMENTS (old_decl) != NULL_TREE)
3537 DECL_ARGUMENTS (new_decl) =
3538 copy_arguments_for_versioning (DECL_ARGUMENTS (old_decl), &id);
3540 /* If there's a tree_map, prepare for substitution. */
3542 for (i = 0; i < VARRAY_ACTIVE_SIZE (tree_map); i++)
3544 replace_info = VARRAY_GENERIC_PTR (tree_map, i);
3545 if (replace_info->replace_p)
3546 insert_decl_map (&id, replace_info->old_tree,
3547 replace_info->new_tree);
3550 DECL_INITIAL (new_decl) = remap_blocks (DECL_INITIAL (id.src_fn), &id);
3552 /* Renumber the lexical scoping (non-code) blocks consecutively. */
3553 number_blocks (id.dst_fn);
3555 if (DECL_STRUCT_FUNCTION (old_decl)->unexpanded_var_list != NULL_TREE)
3556 /* Add local vars. */
3557 for (t_step = DECL_STRUCT_FUNCTION (old_decl)->unexpanded_var_list;
3558 t_step; t_step = TREE_CHAIN (t_step))
3560 tree var = TREE_VALUE (t_step);
3561 if (TREE_STATIC (var) && !TREE_ASM_WRITTEN (var))
3562 cfun->unexpanded_var_list = tree_cons (NULL_TREE, var,
3563 cfun->unexpanded_var_list);
3565 cfun->unexpanded_var_list =
3566 tree_cons (NULL_TREE, remap_decl (var, &id),
3567 cfun->unexpanded_var_list);
3570 /* Copy the Function's body. */
3571 copy_body (&id, old_entry_block->count, old_entry_block->frequency, ENTRY_BLOCK_PTR, EXIT_BLOCK_PTR);
3573 if (DECL_RESULT (old_decl) != NULL_TREE)
3575 tree *res_decl = &DECL_RESULT (old_decl);
3576 DECL_RESULT (new_decl) = remap_decl (*res_decl, &id);
3577 lang_hooks.dup_lang_specific_decl (DECL_RESULT (new_decl));
3580 /* Renumber the lexical scoping (non-code) blocks consecutively. */
3581 number_blocks (new_decl);
3584 pointer_map_destroy (id.decl_map);
3587 fold_marked_statements (0, id.statements_to_fold);
3588 pointer_set_destroy (id.statements_to_fold);
3589 fold_cond_expr_cond ();
3591 if (gimple_in_ssa_p (cfun))
3593 free_dominance_info (CDI_DOMINATORS);
3594 free_dominance_info (CDI_POST_DOMINATORS);
3596 delete_unreachable_blocks ();
3597 update_ssa (TODO_update_ssa);
3600 fold_cond_expr_cond ();
3601 if (need_ssa_update_p ())
3602 update_ssa (TODO_update_ssa);
3605 free_dominance_info (CDI_DOMINATORS);
3606 free_dominance_info (CDI_POST_DOMINATORS);
3608 current_function_decl = old_current_function_decl;
3609 gcc_assert (!current_function_decl
3610 || DECL_STRUCT_FUNCTION (current_function_decl) == cfun);
3614 /* Duplicate a type, fields and all. */
3617 build_duplicate_type (tree type)
3619 struct copy_body_data id;
3621 memset (&id, 0, sizeof (id));
3622 id.src_fn = current_function_decl;
3623 id.dst_fn = current_function_decl;
3625 id.decl_map = pointer_map_create ();
3626 id.copy_decl = copy_decl_no_change;
3628 type = remap_type_1 (type, &id);
3630 pointer_map_destroy (id.decl_map);